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 should be flipped only while holding
64 * manager_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 * MG: pool->manager_mutex and pool->lock protected. Writes require both
123 * locks. Reads can happen under either lock.
125 * WQ: wq_mutex protected.
127 * WR: wq_mutex protected for writes. Sched-RCU protected for reads.
129 * PW: pwq_lock protected.
131 * FR: wq->flush_mutex and pwq_lock protected for writes. Sched-RCU
132 * protected for reads.
134 * MD: wq_mayday_lock protected.
137 /* struct worker is defined in workqueue_internal.h */
140 spinlock_t lock
; /* the pool lock */
141 int cpu
; /* I: the associated cpu */
142 int id
; /* I: pool ID */
143 unsigned int flags
; /* X: flags */
145 struct list_head worklist
; /* L: list of pending works */
146 int nr_workers
; /* L: total number of workers */
148 /* nr_idle includes the ones off idle_list for rebinding */
149 int nr_idle
; /* L: currently idle ones */
151 struct list_head idle_list
; /* X: list of idle workers */
152 struct timer_list idle_timer
; /* L: worker idle timeout */
153 struct timer_list mayday_timer
; /* L: SOS timer for workers */
155 /* a workers is either on busy_hash or idle_list, or the manager */
156 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
157 /* L: hash of busy workers */
159 /* see manage_workers() for details on the two manager mutexes */
160 struct mutex manager_arb
; /* manager arbitration */
161 struct mutex manager_mutex
; /* manager exclusion */
162 struct idr worker_idr
; /* MG: worker IDs and iteration */
164 struct workqueue_attrs
*attrs
; /* I: worker attributes */
165 struct hlist_node hash_node
; /* WQ: unbound_pool_hash node */
166 int refcnt
; /* WQ: refcnt for unbound pools */
169 * The current concurrency level. As it's likely to be accessed
170 * from other CPUs during try_to_wake_up(), put it in a separate
173 atomic_t nr_running ____cacheline_aligned_in_smp
;
176 * Destruction of pool is sched-RCU protected to allow dereferences
177 * from get_work_pool().
180 } ____cacheline_aligned_in_smp
;
183 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
184 * of work_struct->data are used for flags and the remaining high bits
185 * point to the pwq; thus, pwqs need to be aligned at two's power of the
186 * number of flag bits.
188 struct pool_workqueue
{
189 struct worker_pool
*pool
; /* I: the associated pool */
190 struct workqueue_struct
*wq
; /* I: the owning workqueue */
191 int work_color
; /* L: current color */
192 int flush_color
; /* L: flushing color */
193 int refcnt
; /* L: reference count */
194 int nr_in_flight
[WORK_NR_COLORS
];
195 /* L: nr of in_flight works */
196 int nr_active
; /* L: nr of active works */
197 int max_active
; /* L: max active works */
198 struct list_head delayed_works
; /* L: delayed works */
199 struct list_head pwqs_node
; /* FR: node on wq->pwqs */
200 struct list_head mayday_node
; /* MD: node on wq->maydays */
203 * Release of unbound pwq is punted to system_wq. See put_pwq()
204 * and pwq_unbound_release_workfn() for details. pool_workqueue
205 * itself is also sched-RCU protected so that the first pwq can be
206 * determined without grabbing pwq_lock.
208 struct work_struct unbound_release_work
;
210 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
213 * Structure used to wait for workqueue flush.
216 struct list_head list
; /* F: list of flushers */
217 int flush_color
; /* F: flush color waiting for */
218 struct completion done
; /* flush completion */
224 * The externally visible workqueue. It relays the issued work items to
225 * the appropriate worker_pool through its pool_workqueues.
227 struct workqueue_struct
{
228 unsigned int flags
; /* WQ: WQ_* flags */
229 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwq's */
230 struct list_head pwqs
; /* FR: all pwqs of this wq */
231 struct list_head list
; /* WQ: list of all workqueues */
233 struct mutex flush_mutex
; /* protects wq flushing */
234 int work_color
; /* F: current work color */
235 int flush_color
; /* F: current flush color */
236 atomic_t nr_pwqs_to_flush
; /* flush in progress */
237 struct wq_flusher
*first_flusher
; /* F: first flusher */
238 struct list_head flusher_queue
; /* F: flush waiters */
239 struct list_head flusher_overflow
; /* F: flush overflow list */
241 struct list_head maydays
; /* MD: pwqs requesting rescue */
242 struct worker
*rescuer
; /* I: rescue worker */
244 int nr_drainers
; /* WQ: drain in progress */
245 int saved_max_active
; /* PW: saved pwq max_active */
248 struct wq_device
*wq_dev
; /* I: for sysfs interface */
250 #ifdef CONFIG_LOCKDEP
251 struct lockdep_map lockdep_map
;
253 char name
[]; /* I: workqueue name */
256 static struct kmem_cache
*pwq_cache
;
258 static DEFINE_MUTEX(wq_mutex
); /* protects workqueues and pools */
259 static DEFINE_SPINLOCK(pwq_lock
); /* protects pool_workqueues */
260 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
262 static LIST_HEAD(workqueues
); /* WQ: list of all workqueues */
263 static bool workqueue_freezing
; /* WQ: have wqs started freezing? */
265 /* the per-cpu worker pools */
266 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
269 static DEFINE_IDR(worker_pool_idr
); /* WR: idr of all pools */
271 /* WQ: hash of all unbound pools keyed by pool->attrs */
272 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
274 /* I: attributes used when instantiating standard unbound pools on demand */
275 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
277 struct workqueue_struct
*system_wq __read_mostly
;
278 EXPORT_SYMBOL_GPL(system_wq
);
279 struct workqueue_struct
*system_highpri_wq __read_mostly
;
280 EXPORT_SYMBOL_GPL(system_highpri_wq
);
281 struct workqueue_struct
*system_long_wq __read_mostly
;
282 EXPORT_SYMBOL_GPL(system_long_wq
);
283 struct workqueue_struct
*system_unbound_wq __read_mostly
;
284 EXPORT_SYMBOL_GPL(system_unbound_wq
);
285 struct workqueue_struct
*system_freezable_wq __read_mostly
;
286 EXPORT_SYMBOL_GPL(system_freezable_wq
);
288 static int worker_thread(void *__worker
);
289 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
290 const struct workqueue_attrs
*from
);
292 #define CREATE_TRACE_POINTS
293 #include <trace/events/workqueue.h>
295 #define assert_rcu_or_wq_mutex() \
296 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
297 lockdep_is_held(&wq_mutex), \
298 "sched RCU or wq_mutex should be held")
300 #define assert_rcu_or_pwq_lock() \
301 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
302 lockdep_is_held(&pwq_lock), \
303 "sched RCU or pwq_lock should be held")
305 #ifdef CONFIG_LOCKDEP
306 #define assert_manager_or_pool_lock(pool) \
307 WARN_ONCE(!lockdep_is_held(&(pool)->manager_mutex) && \
308 !lockdep_is_held(&(pool)->lock), \
309 "pool->manager_mutex or ->lock should be held")
311 #define assert_manager_or_pool_lock(pool) do { } while (0)
314 #define for_each_cpu_worker_pool(pool, cpu) \
315 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
316 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
319 #define for_each_busy_worker(worker, i, pool) \
320 hash_for_each(pool->busy_hash, i, worker, hentry)
323 * for_each_pool - iterate through all worker_pools in the system
324 * @pool: iteration cursor
325 * @pi: integer used for iteration
327 * This must be called either with wq_mutex held or sched RCU read locked.
328 * If the pool needs to be used beyond the locking in effect, the caller is
329 * responsible for guaranteeing that the pool stays online.
331 * The if/else clause exists only for the lockdep assertion and can be
334 #define for_each_pool(pool, pi) \
335 idr_for_each_entry(&worker_pool_idr, pool, pi) \
336 if (({ assert_rcu_or_wq_mutex(); false; })) { } \
340 * for_each_pool_worker - iterate through all workers of a worker_pool
341 * @worker: iteration cursor
342 * @wi: integer used for iteration
343 * @pool: worker_pool to iterate workers of
345 * This must be called with either @pool->manager_mutex or ->lock held.
347 * The if/else clause exists only for the lockdep assertion and can be
350 #define for_each_pool_worker(worker, wi, pool) \
351 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
352 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
356 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
357 * @pwq: iteration cursor
358 * @wq: the target workqueue
360 * This must be called either with pwq_lock held or sched RCU read locked.
361 * If the pwq needs to be used beyond the locking in effect, the caller is
362 * responsible for guaranteeing that the pwq stays online.
364 * The if/else clause exists only for the lockdep assertion and can be
367 #define for_each_pwq(pwq, wq) \
368 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
369 if (({ assert_rcu_or_pwq_lock(); false; })) { } \
372 #ifdef CONFIG_DEBUG_OBJECTS_WORK
374 static struct debug_obj_descr work_debug_descr
;
376 static void *work_debug_hint(void *addr
)
378 return ((struct work_struct
*) addr
)->func
;
382 * fixup_init is called when:
383 * - an active object is initialized
385 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
387 struct work_struct
*work
= addr
;
390 case ODEBUG_STATE_ACTIVE
:
391 cancel_work_sync(work
);
392 debug_object_init(work
, &work_debug_descr
);
400 * fixup_activate is called when:
401 * - an active object is activated
402 * - an unknown object is activated (might be a statically initialized object)
404 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
406 struct work_struct
*work
= addr
;
410 case ODEBUG_STATE_NOTAVAILABLE
:
412 * This is not really a fixup. The work struct was
413 * statically initialized. We just make sure that it
414 * is tracked in the object tracker.
416 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
417 debug_object_init(work
, &work_debug_descr
);
418 debug_object_activate(work
, &work_debug_descr
);
424 case ODEBUG_STATE_ACTIVE
:
433 * fixup_free is called when:
434 * - an active object is freed
436 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
438 struct work_struct
*work
= addr
;
441 case ODEBUG_STATE_ACTIVE
:
442 cancel_work_sync(work
);
443 debug_object_free(work
, &work_debug_descr
);
450 static struct debug_obj_descr work_debug_descr
= {
451 .name
= "work_struct",
452 .debug_hint
= work_debug_hint
,
453 .fixup_init
= work_fixup_init
,
454 .fixup_activate
= work_fixup_activate
,
455 .fixup_free
= work_fixup_free
,
458 static inline void debug_work_activate(struct work_struct
*work
)
460 debug_object_activate(work
, &work_debug_descr
);
463 static inline void debug_work_deactivate(struct work_struct
*work
)
465 debug_object_deactivate(work
, &work_debug_descr
);
468 void __init_work(struct work_struct
*work
, int onstack
)
471 debug_object_init_on_stack(work
, &work_debug_descr
);
473 debug_object_init(work
, &work_debug_descr
);
475 EXPORT_SYMBOL_GPL(__init_work
);
477 void destroy_work_on_stack(struct work_struct
*work
)
479 debug_object_free(work
, &work_debug_descr
);
481 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
484 static inline void debug_work_activate(struct work_struct
*work
) { }
485 static inline void debug_work_deactivate(struct work_struct
*work
) { }
488 /* allocate ID and assign it to @pool */
489 static int worker_pool_assign_id(struct worker_pool
*pool
)
493 lockdep_assert_held(&wq_mutex
);
496 if (!idr_pre_get(&worker_pool_idr
, GFP_KERNEL
))
498 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
499 } while (ret
== -EAGAIN
);
505 * first_pwq - return the first pool_workqueue of the specified workqueue
506 * @wq: the target workqueue
508 * This must be called either with pwq_lock held or sched RCU read locked.
509 * If the pwq needs to be used beyond the locking in effect, the caller is
510 * responsible for guaranteeing that the pwq stays online.
512 static struct pool_workqueue
*first_pwq(struct workqueue_struct
*wq
)
514 assert_rcu_or_pwq_lock();
515 return list_first_or_null_rcu(&wq
->pwqs
, struct pool_workqueue
,
519 static unsigned int work_color_to_flags(int color
)
521 return color
<< WORK_STRUCT_COLOR_SHIFT
;
524 static int get_work_color(struct work_struct
*work
)
526 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
527 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
530 static int work_next_color(int color
)
532 return (color
+ 1) % WORK_NR_COLORS
;
536 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
537 * contain the pointer to the queued pwq. Once execution starts, the flag
538 * is cleared and the high bits contain OFFQ flags and pool ID.
540 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
541 * and clear_work_data() can be used to set the pwq, pool or clear
542 * work->data. These functions should only be called while the work is
543 * owned - ie. while the PENDING bit is set.
545 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
546 * corresponding to a work. Pool is available once the work has been
547 * queued anywhere after initialization until it is sync canceled. pwq is
548 * available only while the work item is queued.
550 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
551 * canceled. While being canceled, a work item may have its PENDING set
552 * but stay off timer and worklist for arbitrarily long and nobody should
553 * try to steal the PENDING bit.
555 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
558 WARN_ON_ONCE(!work_pending(work
));
559 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
562 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
563 unsigned long extra_flags
)
565 set_work_data(work
, (unsigned long)pwq
,
566 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
569 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
572 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
573 WORK_STRUCT_PENDING
);
576 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
580 * The following wmb is paired with the implied mb in
581 * test_and_set_bit(PENDING) and ensures all updates to @work made
582 * here are visible to and precede any updates by the next PENDING
586 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
589 static void clear_work_data(struct work_struct
*work
)
591 smp_wmb(); /* see set_work_pool_and_clear_pending() */
592 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
595 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
597 unsigned long data
= atomic_long_read(&work
->data
);
599 if (data
& WORK_STRUCT_PWQ
)
600 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
606 * get_work_pool - return the worker_pool a given work was associated with
607 * @work: the work item of interest
609 * Return the worker_pool @work was last associated with. %NULL if none.
611 * Pools are created and destroyed under wq_mutex, and allows read access
612 * under sched-RCU read lock. As such, this function should be called
613 * under wq_mutex or with preemption disabled.
615 * All fields of the returned pool are accessible as long as the above
616 * mentioned locking is in effect. If the returned pool needs to be used
617 * beyond the critical section, the caller is responsible for ensuring the
618 * returned pool is and stays online.
620 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
622 unsigned long data
= atomic_long_read(&work
->data
);
625 assert_rcu_or_wq_mutex();
627 if (data
& WORK_STRUCT_PWQ
)
628 return ((struct pool_workqueue
*)
629 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
631 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
632 if (pool_id
== WORK_OFFQ_POOL_NONE
)
635 return idr_find(&worker_pool_idr
, pool_id
);
639 * get_work_pool_id - return the worker pool ID a given work is associated with
640 * @work: the work item of interest
642 * Return the worker_pool ID @work was last associated with.
643 * %WORK_OFFQ_POOL_NONE if none.
645 static int get_work_pool_id(struct work_struct
*work
)
647 unsigned long data
= atomic_long_read(&work
->data
);
649 if (data
& WORK_STRUCT_PWQ
)
650 return ((struct pool_workqueue
*)
651 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
653 return data
>> WORK_OFFQ_POOL_SHIFT
;
656 static void mark_work_canceling(struct work_struct
*work
)
658 unsigned long pool_id
= get_work_pool_id(work
);
660 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
661 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
664 static bool work_is_canceling(struct work_struct
*work
)
666 unsigned long data
= atomic_long_read(&work
->data
);
668 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
672 * Policy functions. These define the policies on how the global worker
673 * pools are managed. Unless noted otherwise, these functions assume that
674 * they're being called with pool->lock held.
677 static bool __need_more_worker(struct worker_pool
*pool
)
679 return !atomic_read(&pool
->nr_running
);
683 * Need to wake up a worker? Called from anything but currently
686 * Note that, because unbound workers never contribute to nr_running, this
687 * function will always return %true for unbound pools as long as the
688 * worklist isn't empty.
690 static bool need_more_worker(struct worker_pool
*pool
)
692 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
695 /* Can I start working? Called from busy but !running workers. */
696 static bool may_start_working(struct worker_pool
*pool
)
698 return pool
->nr_idle
;
701 /* Do I need to keep working? Called from currently running workers. */
702 static bool keep_working(struct worker_pool
*pool
)
704 return !list_empty(&pool
->worklist
) &&
705 atomic_read(&pool
->nr_running
) <= 1;
708 /* Do we need a new worker? Called from manager. */
709 static bool need_to_create_worker(struct worker_pool
*pool
)
711 return need_more_worker(pool
) && !may_start_working(pool
);
714 /* Do I need to be the manager? */
715 static bool need_to_manage_workers(struct worker_pool
*pool
)
717 return need_to_create_worker(pool
) ||
718 (pool
->flags
& POOL_MANAGE_WORKERS
);
721 /* Do we have too many workers and should some go away? */
722 static bool too_many_workers(struct worker_pool
*pool
)
724 bool managing
= mutex_is_locked(&pool
->manager_arb
);
725 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
726 int nr_busy
= pool
->nr_workers
- nr_idle
;
729 * nr_idle and idle_list may disagree if idle rebinding is in
730 * progress. Never return %true if idle_list is empty.
732 if (list_empty(&pool
->idle_list
))
735 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
742 /* Return the first worker. Safe with preemption disabled */
743 static struct worker
*first_worker(struct worker_pool
*pool
)
745 if (unlikely(list_empty(&pool
->idle_list
)))
748 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
752 * wake_up_worker - wake up an idle worker
753 * @pool: worker pool to wake worker from
755 * Wake up the first idle worker of @pool.
758 * spin_lock_irq(pool->lock).
760 static void wake_up_worker(struct worker_pool
*pool
)
762 struct worker
*worker
= first_worker(pool
);
765 wake_up_process(worker
->task
);
769 * wq_worker_waking_up - a worker is waking up
770 * @task: task waking up
771 * @cpu: CPU @task is waking up to
773 * This function is called during try_to_wake_up() when a worker is
777 * spin_lock_irq(rq->lock)
779 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
781 struct worker
*worker
= kthread_data(task
);
783 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
784 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
785 atomic_inc(&worker
->pool
->nr_running
);
790 * wq_worker_sleeping - a worker is going to sleep
791 * @task: task going to sleep
792 * @cpu: CPU in question, must be the current CPU number
794 * This function is called during schedule() when a busy worker is
795 * going to sleep. Worker on the same cpu can be woken up by
796 * returning pointer to its task.
799 * spin_lock_irq(rq->lock)
802 * Worker task on @cpu to wake up, %NULL if none.
804 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
806 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
807 struct worker_pool
*pool
;
810 * Rescuers, which may not have all the fields set up like normal
811 * workers, also reach here, let's not access anything before
812 * checking NOT_RUNNING.
814 if (worker
->flags
& WORKER_NOT_RUNNING
)
819 /* this can only happen on the local cpu */
820 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
824 * The counterpart of the following dec_and_test, implied mb,
825 * worklist not empty test sequence is in insert_work().
826 * Please read comment there.
828 * NOT_RUNNING is clear. This means that we're bound to and
829 * running on the local cpu w/ rq lock held and preemption
830 * disabled, which in turn means that none else could be
831 * manipulating idle_list, so dereferencing idle_list without pool
834 if (atomic_dec_and_test(&pool
->nr_running
) &&
835 !list_empty(&pool
->worklist
))
836 to_wakeup
= first_worker(pool
);
837 return to_wakeup
? to_wakeup
->task
: NULL
;
841 * worker_set_flags - set worker flags and adjust nr_running accordingly
843 * @flags: flags to set
844 * @wakeup: wakeup an idle worker if necessary
846 * Set @flags in @worker->flags and adjust nr_running accordingly. If
847 * nr_running becomes zero and @wakeup is %true, an idle worker is
851 * spin_lock_irq(pool->lock)
853 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
856 struct worker_pool
*pool
= worker
->pool
;
858 WARN_ON_ONCE(worker
->task
!= current
);
861 * If transitioning into NOT_RUNNING, adjust nr_running and
862 * wake up an idle worker as necessary if requested by
865 if ((flags
& WORKER_NOT_RUNNING
) &&
866 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
868 if (atomic_dec_and_test(&pool
->nr_running
) &&
869 !list_empty(&pool
->worklist
))
870 wake_up_worker(pool
);
872 atomic_dec(&pool
->nr_running
);
875 worker
->flags
|= flags
;
879 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
881 * @flags: flags to clear
883 * Clear @flags in @worker->flags and adjust nr_running accordingly.
886 * spin_lock_irq(pool->lock)
888 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
890 struct worker_pool
*pool
= worker
->pool
;
891 unsigned int oflags
= worker
->flags
;
893 WARN_ON_ONCE(worker
->task
!= current
);
895 worker
->flags
&= ~flags
;
898 * If transitioning out of NOT_RUNNING, increment nr_running. Note
899 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
900 * of multiple flags, not a single flag.
902 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
903 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
904 atomic_inc(&pool
->nr_running
);
908 * find_worker_executing_work - find worker which is executing a work
909 * @pool: pool of interest
910 * @work: work to find worker for
912 * Find a worker which is executing @work on @pool by searching
913 * @pool->busy_hash which is keyed by the address of @work. For a worker
914 * to match, its current execution should match the address of @work and
915 * its work function. This is to avoid unwanted dependency between
916 * unrelated work executions through a work item being recycled while still
919 * This is a bit tricky. A work item may be freed once its execution
920 * starts and nothing prevents the freed area from being recycled for
921 * another work item. If the same work item address ends up being reused
922 * before the original execution finishes, workqueue will identify the
923 * recycled work item as currently executing and make it wait until the
924 * current execution finishes, introducing an unwanted dependency.
926 * This function checks the work item address and work function to avoid
927 * false positives. Note that this isn't complete as one may construct a
928 * work function which can introduce dependency onto itself through a
929 * recycled work item. Well, if somebody wants to shoot oneself in the
930 * foot that badly, there's only so much we can do, and if such deadlock
931 * actually occurs, it should be easy to locate the culprit work function.
934 * spin_lock_irq(pool->lock).
937 * Pointer to worker which is executing @work if found, NULL
940 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
941 struct work_struct
*work
)
943 struct worker
*worker
;
945 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
947 if (worker
->current_work
== work
&&
948 worker
->current_func
== work
->func
)
955 * move_linked_works - move linked works to a list
956 * @work: start of series of works to be scheduled
957 * @head: target list to append @work to
958 * @nextp: out paramter for nested worklist walking
960 * Schedule linked works starting from @work to @head. Work series to
961 * be scheduled starts at @work and includes any consecutive work with
962 * WORK_STRUCT_LINKED set in its predecessor.
964 * If @nextp is not NULL, it's updated to point to the next work of
965 * the last scheduled work. This allows move_linked_works() to be
966 * nested inside outer list_for_each_entry_safe().
969 * spin_lock_irq(pool->lock).
971 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
972 struct work_struct
**nextp
)
974 struct work_struct
*n
;
977 * Linked worklist will always end before the end of the list,
978 * use NULL for list head.
980 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
981 list_move_tail(&work
->entry
, head
);
982 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
987 * If we're already inside safe list traversal and have moved
988 * multiple works to the scheduled queue, the next position
989 * needs to be updated.
996 * get_pwq - get an extra reference on the specified pool_workqueue
997 * @pwq: pool_workqueue to get
999 * Obtain an extra reference on @pwq. The caller should guarantee that
1000 * @pwq has positive refcnt and be holding the matching pool->lock.
1002 static void get_pwq(struct pool_workqueue
*pwq
)
1004 lockdep_assert_held(&pwq
->pool
->lock
);
1005 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1010 * put_pwq - put a pool_workqueue reference
1011 * @pwq: pool_workqueue to put
1013 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1014 * destruction. The caller should be holding the matching pool->lock.
1016 static void put_pwq(struct pool_workqueue
*pwq
)
1018 lockdep_assert_held(&pwq
->pool
->lock
);
1019 if (likely(--pwq
->refcnt
))
1021 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1024 * @pwq can't be released under pool->lock, bounce to
1025 * pwq_unbound_release_workfn(). This never recurses on the same
1026 * pool->lock as this path is taken only for unbound workqueues and
1027 * the release work item is scheduled on a per-cpu workqueue. To
1028 * avoid lockdep warning, unbound pool->locks are given lockdep
1029 * subclass of 1 in get_unbound_pool().
1031 schedule_work(&pwq
->unbound_release_work
);
1034 static void pwq_activate_delayed_work(struct work_struct
*work
)
1036 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1038 trace_workqueue_activate_work(work
);
1039 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1040 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1044 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1046 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1047 struct work_struct
, entry
);
1049 pwq_activate_delayed_work(work
);
1053 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1054 * @pwq: pwq of interest
1055 * @color: color of work which left the queue
1057 * A work either has completed or is removed from pending queue,
1058 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1061 * spin_lock_irq(pool->lock).
1063 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1065 /* uncolored work items don't participate in flushing or nr_active */
1066 if (color
== WORK_NO_COLOR
)
1069 pwq
->nr_in_flight
[color
]--;
1072 if (!list_empty(&pwq
->delayed_works
)) {
1073 /* one down, submit a delayed one */
1074 if (pwq
->nr_active
< pwq
->max_active
)
1075 pwq_activate_first_delayed(pwq
);
1078 /* is flush in progress and are we at the flushing tip? */
1079 if (likely(pwq
->flush_color
!= color
))
1082 /* are there still in-flight works? */
1083 if (pwq
->nr_in_flight
[color
])
1086 /* this pwq is done, clear flush_color */
1087 pwq
->flush_color
= -1;
1090 * If this was the last pwq, wake up the first flusher. It
1091 * will handle the rest.
1093 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1094 complete(&pwq
->wq
->first_flusher
->done
);
1100 * try_to_grab_pending - steal work item from worklist and disable irq
1101 * @work: work item to steal
1102 * @is_dwork: @work is a delayed_work
1103 * @flags: place to store irq state
1105 * Try to grab PENDING bit of @work. This function can handle @work in any
1106 * stable state - idle, on timer or on worklist. Return values are
1108 * 1 if @work was pending and we successfully stole PENDING
1109 * 0 if @work was idle and we claimed PENDING
1110 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1111 * -ENOENT if someone else is canceling @work, this state may persist
1112 * for arbitrarily long
1114 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1115 * interrupted while holding PENDING and @work off queue, irq must be
1116 * disabled on entry. This, combined with delayed_work->timer being
1117 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1119 * On successful return, >= 0, irq is disabled and the caller is
1120 * responsible for releasing it using local_irq_restore(*@flags).
1122 * This function is safe to call from any context including IRQ handler.
1124 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1125 unsigned long *flags
)
1127 struct worker_pool
*pool
;
1128 struct pool_workqueue
*pwq
;
1130 local_irq_save(*flags
);
1132 /* try to steal the timer if it exists */
1134 struct delayed_work
*dwork
= to_delayed_work(work
);
1137 * dwork->timer is irqsafe. If del_timer() fails, it's
1138 * guaranteed that the timer is not queued anywhere and not
1139 * running on the local CPU.
1141 if (likely(del_timer(&dwork
->timer
)))
1145 /* try to claim PENDING the normal way */
1146 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1150 * The queueing is in progress, or it is already queued. Try to
1151 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1153 pool
= get_work_pool(work
);
1157 spin_lock(&pool
->lock
);
1159 * work->data is guaranteed to point to pwq only while the work
1160 * item is queued on pwq->wq, and both updating work->data to point
1161 * to pwq on queueing and to pool on dequeueing are done under
1162 * pwq->pool->lock. This in turn guarantees that, if work->data
1163 * points to pwq which is associated with a locked pool, the work
1164 * item is currently queued on that pool.
1166 pwq
= get_work_pwq(work
);
1167 if (pwq
&& pwq
->pool
== pool
) {
1168 debug_work_deactivate(work
);
1171 * A delayed work item cannot be grabbed directly because
1172 * it might have linked NO_COLOR work items which, if left
1173 * on the delayed_list, will confuse pwq->nr_active
1174 * management later on and cause stall. Make sure the work
1175 * item is activated before grabbing.
1177 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1178 pwq_activate_delayed_work(work
);
1180 list_del_init(&work
->entry
);
1181 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1183 /* work->data points to pwq iff queued, point to pool */
1184 set_work_pool_and_keep_pending(work
, pool
->id
);
1186 spin_unlock(&pool
->lock
);
1189 spin_unlock(&pool
->lock
);
1191 local_irq_restore(*flags
);
1192 if (work_is_canceling(work
))
1199 * insert_work - insert a work into a pool
1200 * @pwq: pwq @work belongs to
1201 * @work: work to insert
1202 * @head: insertion point
1203 * @extra_flags: extra WORK_STRUCT_* flags to set
1205 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1206 * work_struct flags.
1209 * spin_lock_irq(pool->lock).
1211 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1212 struct list_head
*head
, unsigned int extra_flags
)
1214 struct worker_pool
*pool
= pwq
->pool
;
1216 /* we own @work, set data and link */
1217 set_work_pwq(work
, pwq
, extra_flags
);
1218 list_add_tail(&work
->entry
, head
);
1222 * Ensure either wq_worker_sleeping() sees the above
1223 * list_add_tail() or we see zero nr_running to avoid workers lying
1224 * around lazily while there are works to be processed.
1228 if (__need_more_worker(pool
))
1229 wake_up_worker(pool
);
1233 * Test whether @work is being queued from another work executing on the
1236 static bool is_chained_work(struct workqueue_struct
*wq
)
1238 struct worker
*worker
;
1240 worker
= current_wq_worker();
1242 * Return %true iff I'm a worker execuing a work item on @wq. If
1243 * I'm @worker, it's safe to dereference it without locking.
1245 return worker
&& worker
->current_pwq
->wq
== wq
;
1248 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1249 struct work_struct
*work
)
1251 struct pool_workqueue
*pwq
;
1252 struct worker_pool
*last_pool
;
1253 struct list_head
*worklist
;
1254 unsigned int work_flags
;
1255 unsigned int req_cpu
= cpu
;
1258 * While a work item is PENDING && off queue, a task trying to
1259 * steal the PENDING will busy-loop waiting for it to either get
1260 * queued or lose PENDING. Grabbing PENDING and queueing should
1261 * happen with IRQ disabled.
1263 WARN_ON_ONCE(!irqs_disabled());
1265 debug_work_activate(work
);
1267 /* if dying, only works from the same workqueue are allowed */
1268 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1269 WARN_ON_ONCE(!is_chained_work(wq
)))
1272 /* pwq which will be used unless @work is executing elsewhere */
1273 if (!(wq
->flags
& WQ_UNBOUND
)) {
1274 if (cpu
== WORK_CPU_UNBOUND
)
1275 cpu
= raw_smp_processor_id();
1276 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1278 pwq
= first_pwq(wq
);
1282 * If @work was previously on a different pool, it might still be
1283 * running there, in which case the work needs to be queued on that
1284 * pool to guarantee non-reentrancy.
1286 last_pool
= get_work_pool(work
);
1287 if (last_pool
&& last_pool
!= pwq
->pool
) {
1288 struct worker
*worker
;
1290 spin_lock(&last_pool
->lock
);
1292 worker
= find_worker_executing_work(last_pool
, work
);
1294 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1295 pwq
= worker
->current_pwq
;
1297 /* meh... not running there, queue here */
1298 spin_unlock(&last_pool
->lock
);
1299 spin_lock(&pwq
->pool
->lock
);
1302 spin_lock(&pwq
->pool
->lock
);
1306 * pwq is determined and locked. For unbound pools, we could have
1307 * raced with pwq release and it could already be dead. If its
1308 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1309 * without another pwq replacing it as the first pwq or while a
1310 * work item is executing on it, so the retying is guaranteed to
1311 * make forward-progress.
1313 if (unlikely(!pwq
->refcnt
)) {
1314 if (wq
->flags
& WQ_UNBOUND
) {
1315 spin_unlock(&pwq
->pool
->lock
);
1320 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1324 /* pwq determined, queue */
1325 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1327 if (WARN_ON(!list_empty(&work
->entry
))) {
1328 spin_unlock(&pwq
->pool
->lock
);
1332 pwq
->nr_in_flight
[pwq
->work_color
]++;
1333 work_flags
= work_color_to_flags(pwq
->work_color
);
1335 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1336 trace_workqueue_activate_work(work
);
1338 worklist
= &pwq
->pool
->worklist
;
1340 work_flags
|= WORK_STRUCT_DELAYED
;
1341 worklist
= &pwq
->delayed_works
;
1344 insert_work(pwq
, work
, worklist
, work_flags
);
1346 spin_unlock(&pwq
->pool
->lock
);
1350 * queue_work_on - queue work on specific cpu
1351 * @cpu: CPU number to execute work on
1352 * @wq: workqueue to use
1353 * @work: work to queue
1355 * Returns %false if @work was already on a queue, %true otherwise.
1357 * We queue the work to a specific CPU, the caller must ensure it
1360 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1361 struct work_struct
*work
)
1364 unsigned long flags
;
1366 local_irq_save(flags
);
1368 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1369 __queue_work(cpu
, wq
, work
);
1373 local_irq_restore(flags
);
1376 EXPORT_SYMBOL_GPL(queue_work_on
);
1378 void delayed_work_timer_fn(unsigned long __data
)
1380 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1382 /* should have been called from irqsafe timer with irq already off */
1383 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1385 EXPORT_SYMBOL(delayed_work_timer_fn
);
1387 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1388 struct delayed_work
*dwork
, unsigned long delay
)
1390 struct timer_list
*timer
= &dwork
->timer
;
1391 struct work_struct
*work
= &dwork
->work
;
1393 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1394 timer
->data
!= (unsigned long)dwork
);
1395 WARN_ON_ONCE(timer_pending(timer
));
1396 WARN_ON_ONCE(!list_empty(&work
->entry
));
1399 * If @delay is 0, queue @dwork->work immediately. This is for
1400 * both optimization and correctness. The earliest @timer can
1401 * expire is on the closest next tick and delayed_work users depend
1402 * on that there's no such delay when @delay is 0.
1405 __queue_work(cpu
, wq
, &dwork
->work
);
1409 timer_stats_timer_set_start_info(&dwork
->timer
);
1413 timer
->expires
= jiffies
+ delay
;
1415 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1416 add_timer_on(timer
, cpu
);
1422 * queue_delayed_work_on - queue work on specific CPU after delay
1423 * @cpu: CPU number to execute work on
1424 * @wq: workqueue to use
1425 * @dwork: work to queue
1426 * @delay: number of jiffies to wait before queueing
1428 * Returns %false if @work was already on a queue, %true otherwise. If
1429 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1432 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1433 struct delayed_work
*dwork
, unsigned long delay
)
1435 struct work_struct
*work
= &dwork
->work
;
1437 unsigned long flags
;
1439 /* read the comment in __queue_work() */
1440 local_irq_save(flags
);
1442 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1443 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1447 local_irq_restore(flags
);
1450 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1453 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1454 * @cpu: CPU number to execute work on
1455 * @wq: workqueue to use
1456 * @dwork: work to queue
1457 * @delay: number of jiffies to wait before queueing
1459 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1460 * modify @dwork's timer so that it expires after @delay. If @delay is
1461 * zero, @work is guaranteed to be scheduled immediately regardless of its
1464 * Returns %false if @dwork was idle and queued, %true if @dwork was
1465 * pending and its timer was modified.
1467 * This function is safe to call from any context including IRQ handler.
1468 * See try_to_grab_pending() for details.
1470 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1471 struct delayed_work
*dwork
, unsigned long delay
)
1473 unsigned long flags
;
1477 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1478 } while (unlikely(ret
== -EAGAIN
));
1480 if (likely(ret
>= 0)) {
1481 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1482 local_irq_restore(flags
);
1485 /* -ENOENT from try_to_grab_pending() becomes %true */
1488 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1491 * worker_enter_idle - enter idle state
1492 * @worker: worker which is entering idle state
1494 * @worker is entering idle state. Update stats and idle timer if
1498 * spin_lock_irq(pool->lock).
1500 static void worker_enter_idle(struct worker
*worker
)
1502 struct worker_pool
*pool
= worker
->pool
;
1504 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1505 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1506 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1509 /* can't use worker_set_flags(), also called from start_worker() */
1510 worker
->flags
|= WORKER_IDLE
;
1512 worker
->last_active
= jiffies
;
1514 /* idle_list is LIFO */
1515 list_add(&worker
->entry
, &pool
->idle_list
);
1517 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1518 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1521 * Sanity check nr_running. Because wq_unbind_fn() releases
1522 * pool->lock between setting %WORKER_UNBOUND and zapping
1523 * nr_running, the warning may trigger spuriously. Check iff
1524 * unbind is not in progress.
1526 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1527 pool
->nr_workers
== pool
->nr_idle
&&
1528 atomic_read(&pool
->nr_running
));
1532 * worker_leave_idle - leave idle state
1533 * @worker: worker which is leaving idle state
1535 * @worker is leaving idle state. Update stats.
1538 * spin_lock_irq(pool->lock).
1540 static void worker_leave_idle(struct worker
*worker
)
1542 struct worker_pool
*pool
= worker
->pool
;
1544 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1546 worker_clr_flags(worker
, WORKER_IDLE
);
1548 list_del_init(&worker
->entry
);
1552 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1553 * @pool: target worker_pool
1555 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1557 * Works which are scheduled while the cpu is online must at least be
1558 * scheduled to a worker which is bound to the cpu so that if they are
1559 * flushed from cpu callbacks while cpu is going down, they are
1560 * guaranteed to execute on the cpu.
1562 * This function is to be used by unbound workers and rescuers to bind
1563 * themselves to the target cpu and may race with cpu going down or
1564 * coming online. kthread_bind() can't be used because it may put the
1565 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1566 * verbatim as it's best effort and blocking and pool may be
1567 * [dis]associated in the meantime.
1569 * This function tries set_cpus_allowed() and locks pool and verifies the
1570 * binding against %POOL_DISASSOCIATED which is set during
1571 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1572 * enters idle state or fetches works without dropping lock, it can
1573 * guarantee the scheduling requirement described in the first paragraph.
1576 * Might sleep. Called without any lock but returns with pool->lock
1580 * %true if the associated pool is online (@worker is successfully
1581 * bound), %false if offline.
1583 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1584 __acquires(&pool
->lock
)
1588 * The following call may fail, succeed or succeed
1589 * without actually migrating the task to the cpu if
1590 * it races with cpu hotunplug operation. Verify
1591 * against POOL_DISASSOCIATED.
1593 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1594 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1596 spin_lock_irq(&pool
->lock
);
1597 if (pool
->flags
& POOL_DISASSOCIATED
)
1599 if (task_cpu(current
) == pool
->cpu
&&
1600 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1602 spin_unlock_irq(&pool
->lock
);
1605 * We've raced with CPU hot[un]plug. Give it a breather
1606 * and retry migration. cond_resched() is required here;
1607 * otherwise, we might deadlock against cpu_stop trying to
1608 * bring down the CPU on non-preemptive kernel.
1616 * Rebind an idle @worker to its CPU. worker_thread() will test
1617 * list_empty(@worker->entry) before leaving idle and call this function.
1619 static void idle_worker_rebind(struct worker
*worker
)
1621 /* CPU may go down again inbetween, clear UNBOUND only on success */
1622 if (worker_maybe_bind_and_lock(worker
->pool
))
1623 worker_clr_flags(worker
, WORKER_UNBOUND
);
1625 /* rebind complete, become available again */
1626 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1627 spin_unlock_irq(&worker
->pool
->lock
);
1631 * Function for @worker->rebind.work used to rebind unbound busy workers to
1632 * the associated cpu which is coming back online. This is scheduled by
1633 * cpu up but can race with other cpu hotplug operations and may be
1634 * executed twice without intervening cpu down.
1636 static void busy_worker_rebind_fn(struct work_struct
*work
)
1638 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1640 if (worker_maybe_bind_and_lock(worker
->pool
))
1641 worker_clr_flags(worker
, WORKER_UNBOUND
);
1643 spin_unlock_irq(&worker
->pool
->lock
);
1647 * rebind_workers - rebind all workers of a pool to the associated CPU
1648 * @pool: pool of interest
1650 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1651 * is different for idle and busy ones.
1653 * Idle ones will be removed from the idle_list and woken up. They will
1654 * add themselves back after completing rebind. This ensures that the
1655 * idle_list doesn't contain any unbound workers when re-bound busy workers
1656 * try to perform local wake-ups for concurrency management.
1658 * Busy workers can rebind after they finish their current work items.
1659 * Queueing the rebind work item at the head of the scheduled list is
1660 * enough. Note that nr_running will be properly bumped as busy workers
1663 * On return, all non-manager workers are scheduled for rebind - see
1664 * manage_workers() for the manager special case. Any idle worker
1665 * including the manager will not appear on @idle_list until rebind is
1666 * complete, making local wake-ups safe.
1668 static void rebind_workers(struct worker_pool
*pool
)
1670 struct worker
*worker
, *n
;
1673 lockdep_assert_held(&pool
->manager_mutex
);
1674 lockdep_assert_held(&pool
->lock
);
1676 /* dequeue and kick idle ones */
1677 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1679 * idle workers should be off @pool->idle_list until rebind
1680 * is complete to avoid receiving premature local wake-ups.
1682 list_del_init(&worker
->entry
);
1685 * worker_thread() will see the above dequeuing and call
1686 * idle_worker_rebind().
1688 wake_up_process(worker
->task
);
1691 /* rebind busy workers */
1692 for_each_busy_worker(worker
, i
, pool
) {
1693 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1694 struct workqueue_struct
*wq
;
1696 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1697 work_data_bits(rebind_work
)))
1700 debug_work_activate(rebind_work
);
1703 * wq doesn't really matter but let's keep @worker->pool
1704 * and @pwq->pool consistent for sanity.
1706 if (worker
->pool
->attrs
->nice
< 0)
1707 wq
= system_highpri_wq
;
1711 insert_work(per_cpu_ptr(wq
->cpu_pwqs
, pool
->cpu
), rebind_work
,
1712 worker
->scheduled
.next
,
1713 work_color_to_flags(WORK_NO_COLOR
));
1717 static struct worker
*alloc_worker(void)
1719 struct worker
*worker
;
1721 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1723 INIT_LIST_HEAD(&worker
->entry
);
1724 INIT_LIST_HEAD(&worker
->scheduled
);
1725 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1726 /* on creation a worker is in !idle && prep state */
1727 worker
->flags
= WORKER_PREP
;
1733 * create_worker - create a new workqueue worker
1734 * @pool: pool the new worker will belong to
1736 * Create a new worker which is bound to @pool. The returned worker
1737 * can be started by calling start_worker() or destroyed using
1741 * Might sleep. Does GFP_KERNEL allocations.
1744 * Pointer to the newly created worker.
1746 static struct worker
*create_worker(struct worker_pool
*pool
)
1748 const char *pri
= pool
->attrs
->nice
< 0 ? "H" : "";
1749 struct worker
*worker
= NULL
;
1752 lockdep_assert_held(&pool
->manager_mutex
);
1755 * ID is needed to determine kthread name. Allocate ID first
1756 * without installing the pointer.
1758 idr_preload(GFP_KERNEL
);
1759 spin_lock_irq(&pool
->lock
);
1761 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1763 spin_unlock_irq(&pool
->lock
);
1768 worker
= alloc_worker();
1772 worker
->pool
= pool
;
1776 worker
->task
= kthread_create_on_node(worker_thread
,
1777 worker
, cpu_to_node(pool
->cpu
),
1778 "kworker/%d:%d%s", pool
->cpu
, id
, pri
);
1780 worker
->task
= kthread_create(worker_thread
, worker
,
1783 if (IS_ERR(worker
->task
))
1787 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1788 * online CPUs. It'll be re-applied when any of the CPUs come up.
1790 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1791 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1793 /* prevent userland from meddling with cpumask of workqueue workers */
1794 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1797 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1798 * remains stable across this function. See the comments above the
1799 * flag definition for details.
1801 if (pool
->flags
& POOL_DISASSOCIATED
)
1802 worker
->flags
|= WORKER_UNBOUND
;
1804 /* successful, commit the pointer to idr */
1805 spin_lock_irq(&pool
->lock
);
1806 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1807 spin_unlock_irq(&pool
->lock
);
1813 spin_lock_irq(&pool
->lock
);
1814 idr_remove(&pool
->worker_idr
, id
);
1815 spin_unlock_irq(&pool
->lock
);
1822 * start_worker - start a newly created worker
1823 * @worker: worker to start
1825 * Make the pool aware of @worker and start it.
1828 * spin_lock_irq(pool->lock).
1830 static void start_worker(struct worker
*worker
)
1832 worker
->flags
|= WORKER_STARTED
;
1833 worker
->pool
->nr_workers
++;
1834 worker_enter_idle(worker
);
1835 wake_up_process(worker
->task
);
1839 * create_and_start_worker - create and start a worker for a pool
1840 * @pool: the target pool
1842 * Grab the managership of @pool and create and start a new worker for it.
1844 static int create_and_start_worker(struct worker_pool
*pool
)
1846 struct worker
*worker
;
1848 mutex_lock(&pool
->manager_mutex
);
1850 worker
= create_worker(pool
);
1852 spin_lock_irq(&pool
->lock
);
1853 start_worker(worker
);
1854 spin_unlock_irq(&pool
->lock
);
1857 mutex_unlock(&pool
->manager_mutex
);
1859 return worker
? 0 : -ENOMEM
;
1863 * destroy_worker - destroy a workqueue worker
1864 * @worker: worker to be destroyed
1866 * Destroy @worker and adjust @pool stats accordingly.
1869 * spin_lock_irq(pool->lock) which is released and regrabbed.
1871 static void destroy_worker(struct worker
*worker
)
1873 struct worker_pool
*pool
= worker
->pool
;
1875 lockdep_assert_held(&pool
->manager_mutex
);
1876 lockdep_assert_held(&pool
->lock
);
1878 /* sanity check frenzy */
1879 if (WARN_ON(worker
->current_work
) ||
1880 WARN_ON(!list_empty(&worker
->scheduled
)))
1883 if (worker
->flags
& WORKER_STARTED
)
1885 if (worker
->flags
& WORKER_IDLE
)
1888 list_del_init(&worker
->entry
);
1889 worker
->flags
|= WORKER_DIE
;
1891 idr_remove(&pool
->worker_idr
, worker
->id
);
1893 spin_unlock_irq(&pool
->lock
);
1895 kthread_stop(worker
->task
);
1898 spin_lock_irq(&pool
->lock
);
1901 static void idle_worker_timeout(unsigned long __pool
)
1903 struct worker_pool
*pool
= (void *)__pool
;
1905 spin_lock_irq(&pool
->lock
);
1907 if (too_many_workers(pool
)) {
1908 struct worker
*worker
;
1909 unsigned long expires
;
1911 /* idle_list is kept in LIFO order, check the last one */
1912 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1913 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1915 if (time_before(jiffies
, expires
))
1916 mod_timer(&pool
->idle_timer
, expires
);
1918 /* it's been idle for too long, wake up manager */
1919 pool
->flags
|= POOL_MANAGE_WORKERS
;
1920 wake_up_worker(pool
);
1924 spin_unlock_irq(&pool
->lock
);
1927 static void send_mayday(struct work_struct
*work
)
1929 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1930 struct workqueue_struct
*wq
= pwq
->wq
;
1932 lockdep_assert_held(&wq_mayday_lock
);
1937 /* mayday mayday mayday */
1938 if (list_empty(&pwq
->mayday_node
)) {
1939 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1940 wake_up_process(wq
->rescuer
->task
);
1944 static void pool_mayday_timeout(unsigned long __pool
)
1946 struct worker_pool
*pool
= (void *)__pool
;
1947 struct work_struct
*work
;
1949 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1950 spin_lock(&pool
->lock
);
1952 if (need_to_create_worker(pool
)) {
1954 * We've been trying to create a new worker but
1955 * haven't been successful. We might be hitting an
1956 * allocation deadlock. Send distress signals to
1959 list_for_each_entry(work
, &pool
->worklist
, entry
)
1963 spin_unlock(&pool
->lock
);
1964 spin_unlock_irq(&wq_mayday_lock
);
1966 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1970 * maybe_create_worker - create a new worker if necessary
1971 * @pool: pool to create a new worker for
1973 * Create a new worker for @pool if necessary. @pool is guaranteed to
1974 * have at least one idle worker on return from this function. If
1975 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1976 * sent to all rescuers with works scheduled on @pool to resolve
1977 * possible allocation deadlock.
1979 * On return, need_to_create_worker() is guaranteed to be %false and
1980 * may_start_working() %true.
1983 * spin_lock_irq(pool->lock) which may be released and regrabbed
1984 * multiple times. Does GFP_KERNEL allocations. Called only from
1988 * %false if no action was taken and pool->lock stayed locked, %true
1991 static bool maybe_create_worker(struct worker_pool
*pool
)
1992 __releases(&pool
->lock
)
1993 __acquires(&pool
->lock
)
1995 if (!need_to_create_worker(pool
))
1998 spin_unlock_irq(&pool
->lock
);
2000 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2001 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
2004 struct worker
*worker
;
2006 worker
= create_worker(pool
);
2008 del_timer_sync(&pool
->mayday_timer
);
2009 spin_lock_irq(&pool
->lock
);
2010 start_worker(worker
);
2011 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
2016 if (!need_to_create_worker(pool
))
2019 __set_current_state(TASK_INTERRUPTIBLE
);
2020 schedule_timeout(CREATE_COOLDOWN
);
2022 if (!need_to_create_worker(pool
))
2026 del_timer_sync(&pool
->mayday_timer
);
2027 spin_lock_irq(&pool
->lock
);
2028 if (need_to_create_worker(pool
))
2034 * maybe_destroy_worker - destroy workers which have been idle for a while
2035 * @pool: pool to destroy workers for
2037 * Destroy @pool workers which have been idle for longer than
2038 * IDLE_WORKER_TIMEOUT.
2041 * spin_lock_irq(pool->lock) which may be released and regrabbed
2042 * multiple times. Called only from manager.
2045 * %false if no action was taken and pool->lock stayed locked, %true
2048 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2052 while (too_many_workers(pool
)) {
2053 struct worker
*worker
;
2054 unsigned long expires
;
2056 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2057 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2059 if (time_before(jiffies
, expires
)) {
2060 mod_timer(&pool
->idle_timer
, expires
);
2064 destroy_worker(worker
);
2072 * manage_workers - manage worker pool
2075 * Assume the manager role and manage the worker pool @worker belongs
2076 * to. At any given time, there can be only zero or one manager per
2077 * pool. The exclusion is handled automatically by this function.
2079 * The caller can safely start processing works on false return. On
2080 * true return, it's guaranteed that need_to_create_worker() is false
2081 * and may_start_working() is true.
2084 * spin_lock_irq(pool->lock) which may be released and regrabbed
2085 * multiple times. Does GFP_KERNEL allocations.
2088 * spin_lock_irq(pool->lock) which may be released and regrabbed
2089 * multiple times. Does GFP_KERNEL allocations.
2091 static bool manage_workers(struct worker
*worker
)
2093 struct worker_pool
*pool
= worker
->pool
;
2097 * Managership is governed by two mutexes - manager_arb and
2098 * manager_mutex. manager_arb handles arbitration of manager role.
2099 * Anyone who successfully grabs manager_arb wins the arbitration
2100 * and becomes the manager. mutex_trylock() on pool->manager_arb
2101 * failure while holding pool->lock reliably indicates that someone
2102 * else is managing the pool and the worker which failed trylock
2103 * can proceed to executing work items. This means that anyone
2104 * grabbing manager_arb is responsible for actually performing
2105 * manager duties. If manager_arb is grabbed and released without
2106 * actual management, the pool may stall indefinitely.
2108 * manager_mutex is used for exclusion of actual management
2109 * operations. The holder of manager_mutex can be sure that none
2110 * of management operations, including creation and destruction of
2111 * workers, won't take place until the mutex is released. Because
2112 * manager_mutex doesn't interfere with manager role arbitration,
2113 * it is guaranteed that the pool's management, while may be
2114 * delayed, won't be disturbed by someone else grabbing
2117 if (!mutex_trylock(&pool
->manager_arb
))
2121 * With manager arbitration won, manager_mutex would be free in
2122 * most cases. trylock first without dropping @pool->lock.
2124 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2125 spin_unlock_irq(&pool
->lock
);
2126 mutex_lock(&pool
->manager_mutex
);
2128 * CPU hotplug could have happened while we were waiting
2129 * for assoc_mutex. Hotplug itself can't handle us
2130 * because manager isn't either on idle or busy list, and
2131 * @pool's state and ours could have deviated.
2133 * As hotplug is now excluded via manager_mutex, we can
2134 * simply try to bind. It will succeed or fail depending
2135 * on @pool's current state. Try it and adjust
2136 * %WORKER_UNBOUND accordingly.
2138 if (worker_maybe_bind_and_lock(pool
))
2139 worker
->flags
&= ~WORKER_UNBOUND
;
2141 worker
->flags
|= WORKER_UNBOUND
;
2146 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2149 * Destroy and then create so that may_start_working() is true
2152 ret
|= maybe_destroy_workers(pool
);
2153 ret
|= maybe_create_worker(pool
);
2155 mutex_unlock(&pool
->manager_mutex
);
2156 mutex_unlock(&pool
->manager_arb
);
2161 * process_one_work - process single work
2163 * @work: work to process
2165 * Process @work. This function contains all the logics necessary to
2166 * process a single work including synchronization against and
2167 * interaction with other workers on the same cpu, queueing and
2168 * flushing. As long as context requirement is met, any worker can
2169 * call this function to process a work.
2172 * spin_lock_irq(pool->lock) which is released and regrabbed.
2174 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2175 __releases(&pool
->lock
)
2176 __acquires(&pool
->lock
)
2178 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2179 struct worker_pool
*pool
= worker
->pool
;
2180 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2182 struct worker
*collision
;
2183 #ifdef CONFIG_LOCKDEP
2185 * It is permissible to free the struct work_struct from
2186 * inside the function that is called from it, this we need to
2187 * take into account for lockdep too. To avoid bogus "held
2188 * lock freed" warnings as well as problems when looking into
2189 * work->lockdep_map, make a copy and use that here.
2191 struct lockdep_map lockdep_map
;
2193 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2196 * Ensure we're on the correct CPU. DISASSOCIATED test is
2197 * necessary to avoid spurious warnings from rescuers servicing the
2198 * unbound or a disassociated pool.
2200 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2201 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2202 raw_smp_processor_id() != pool
->cpu
);
2205 * A single work shouldn't be executed concurrently by
2206 * multiple workers on a single cpu. Check whether anyone is
2207 * already processing the work. If so, defer the work to the
2208 * currently executing one.
2210 collision
= find_worker_executing_work(pool
, work
);
2211 if (unlikely(collision
)) {
2212 move_linked_works(work
, &collision
->scheduled
, NULL
);
2216 /* claim and dequeue */
2217 debug_work_deactivate(work
);
2218 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2219 worker
->current_work
= work
;
2220 worker
->current_func
= work
->func
;
2221 worker
->current_pwq
= pwq
;
2222 work_color
= get_work_color(work
);
2224 list_del_init(&work
->entry
);
2227 * CPU intensive works don't participate in concurrency
2228 * management. They're the scheduler's responsibility.
2230 if (unlikely(cpu_intensive
))
2231 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2234 * Unbound pool isn't concurrency managed and work items should be
2235 * executed ASAP. Wake up another worker if necessary.
2237 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2238 wake_up_worker(pool
);
2241 * Record the last pool and clear PENDING which should be the last
2242 * update to @work. Also, do this inside @pool->lock so that
2243 * PENDING and queued state changes happen together while IRQ is
2246 set_work_pool_and_clear_pending(work
, pool
->id
);
2248 spin_unlock_irq(&pool
->lock
);
2250 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2251 lock_map_acquire(&lockdep_map
);
2252 trace_workqueue_execute_start(work
);
2253 worker
->current_func(work
);
2255 * While we must be careful to not use "work" after this, the trace
2256 * point will only record its address.
2258 trace_workqueue_execute_end(work
);
2259 lock_map_release(&lockdep_map
);
2260 lock_map_release(&pwq
->wq
->lockdep_map
);
2262 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2263 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2264 " last function: %pf\n",
2265 current
->comm
, preempt_count(), task_pid_nr(current
),
2266 worker
->current_func
);
2267 debug_show_held_locks(current
);
2271 spin_lock_irq(&pool
->lock
);
2273 /* clear cpu intensive status */
2274 if (unlikely(cpu_intensive
))
2275 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2277 /* we're done with it, release */
2278 hash_del(&worker
->hentry
);
2279 worker
->current_work
= NULL
;
2280 worker
->current_func
= NULL
;
2281 worker
->current_pwq
= NULL
;
2282 pwq_dec_nr_in_flight(pwq
, work_color
);
2286 * process_scheduled_works - process scheduled works
2289 * Process all scheduled works. Please note that the scheduled list
2290 * may change while processing a work, so this function repeatedly
2291 * fetches a work from the top and executes it.
2294 * spin_lock_irq(pool->lock) which may be released and regrabbed
2297 static void process_scheduled_works(struct worker
*worker
)
2299 while (!list_empty(&worker
->scheduled
)) {
2300 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2301 struct work_struct
, entry
);
2302 process_one_work(worker
, work
);
2307 * worker_thread - the worker thread function
2310 * The worker thread function. All workers belong to a worker_pool -
2311 * either a per-cpu one or dynamic unbound one. These workers process all
2312 * work items regardless of their specific target workqueue. The only
2313 * exception is work items which belong to workqueues with a rescuer which
2314 * will be explained in rescuer_thread().
2316 static int worker_thread(void *__worker
)
2318 struct worker
*worker
= __worker
;
2319 struct worker_pool
*pool
= worker
->pool
;
2321 /* tell the scheduler that this is a workqueue worker */
2322 worker
->task
->flags
|= PF_WQ_WORKER
;
2324 spin_lock_irq(&pool
->lock
);
2326 /* we are off idle list if destruction or rebind is requested */
2327 if (unlikely(list_empty(&worker
->entry
))) {
2328 spin_unlock_irq(&pool
->lock
);
2330 /* if DIE is set, destruction is requested */
2331 if (worker
->flags
& WORKER_DIE
) {
2332 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2336 /* otherwise, rebind */
2337 idle_worker_rebind(worker
);
2341 worker_leave_idle(worker
);
2343 /* no more worker necessary? */
2344 if (!need_more_worker(pool
))
2347 /* do we need to manage? */
2348 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2352 * ->scheduled list can only be filled while a worker is
2353 * preparing to process a work or actually processing it.
2354 * Make sure nobody diddled with it while I was sleeping.
2356 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2359 * When control reaches this point, we're guaranteed to have
2360 * at least one idle worker or that someone else has already
2361 * assumed the manager role.
2363 worker_clr_flags(worker
, WORKER_PREP
);
2366 struct work_struct
*work
=
2367 list_first_entry(&pool
->worklist
,
2368 struct work_struct
, entry
);
2370 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2371 /* optimization path, not strictly necessary */
2372 process_one_work(worker
, work
);
2373 if (unlikely(!list_empty(&worker
->scheduled
)))
2374 process_scheduled_works(worker
);
2376 move_linked_works(work
, &worker
->scheduled
, NULL
);
2377 process_scheduled_works(worker
);
2379 } while (keep_working(pool
));
2381 worker_set_flags(worker
, WORKER_PREP
, false);
2383 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2387 * pool->lock is held and there's no work to process and no need to
2388 * manage, sleep. Workers are woken up only while holding
2389 * pool->lock or from local cpu, so setting the current state
2390 * before releasing pool->lock is enough to prevent losing any
2393 worker_enter_idle(worker
);
2394 __set_current_state(TASK_INTERRUPTIBLE
);
2395 spin_unlock_irq(&pool
->lock
);
2401 * rescuer_thread - the rescuer thread function
2404 * Workqueue rescuer thread function. There's one rescuer for each
2405 * workqueue which has WQ_MEM_RECLAIM set.
2407 * Regular work processing on a pool may block trying to create a new
2408 * worker which uses GFP_KERNEL allocation which has slight chance of
2409 * developing into deadlock if some works currently on the same queue
2410 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2411 * the problem rescuer solves.
2413 * When such condition is possible, the pool summons rescuers of all
2414 * workqueues which have works queued on the pool and let them process
2415 * those works so that forward progress can be guaranteed.
2417 * This should happen rarely.
2419 static int rescuer_thread(void *__rescuer
)
2421 struct worker
*rescuer
= __rescuer
;
2422 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2423 struct list_head
*scheduled
= &rescuer
->scheduled
;
2425 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2428 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2429 * doesn't participate in concurrency management.
2431 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2433 set_current_state(TASK_INTERRUPTIBLE
);
2435 if (kthread_should_stop()) {
2436 __set_current_state(TASK_RUNNING
);
2437 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2441 /* see whether any pwq is asking for help */
2442 spin_lock_irq(&wq_mayday_lock
);
2444 while (!list_empty(&wq
->maydays
)) {
2445 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2446 struct pool_workqueue
, mayday_node
);
2447 struct worker_pool
*pool
= pwq
->pool
;
2448 struct work_struct
*work
, *n
;
2450 __set_current_state(TASK_RUNNING
);
2451 list_del_init(&pwq
->mayday_node
);
2453 spin_unlock_irq(&wq_mayday_lock
);
2455 /* migrate to the target cpu if possible */
2456 worker_maybe_bind_and_lock(pool
);
2457 rescuer
->pool
= pool
;
2460 * Slurp in all works issued via this workqueue and
2463 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2464 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2465 if (get_work_pwq(work
) == pwq
)
2466 move_linked_works(work
, scheduled
, &n
);
2468 process_scheduled_works(rescuer
);
2471 * Leave this pool. If keep_working() is %true, notify a
2472 * regular worker; otherwise, we end up with 0 concurrency
2473 * and stalling the execution.
2475 if (keep_working(pool
))
2476 wake_up_worker(pool
);
2478 rescuer
->pool
= NULL
;
2479 spin_unlock(&pool
->lock
);
2480 spin_lock(&wq_mayday_lock
);
2483 spin_unlock_irq(&wq_mayday_lock
);
2485 /* rescuers should never participate in concurrency management */
2486 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2492 struct work_struct work
;
2493 struct completion done
;
2496 static void wq_barrier_func(struct work_struct
*work
)
2498 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2499 complete(&barr
->done
);
2503 * insert_wq_barrier - insert a barrier work
2504 * @pwq: pwq to insert barrier into
2505 * @barr: wq_barrier to insert
2506 * @target: target work to attach @barr to
2507 * @worker: worker currently executing @target, NULL if @target is not executing
2509 * @barr is linked to @target such that @barr is completed only after
2510 * @target finishes execution. Please note that the ordering
2511 * guarantee is observed only with respect to @target and on the local
2514 * Currently, a queued barrier can't be canceled. This is because
2515 * try_to_grab_pending() can't determine whether the work to be
2516 * grabbed is at the head of the queue and thus can't clear LINKED
2517 * flag of the previous work while there must be a valid next work
2518 * after a work with LINKED flag set.
2520 * Note that when @worker is non-NULL, @target may be modified
2521 * underneath us, so we can't reliably determine pwq from @target.
2524 * spin_lock_irq(pool->lock).
2526 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2527 struct wq_barrier
*barr
,
2528 struct work_struct
*target
, struct worker
*worker
)
2530 struct list_head
*head
;
2531 unsigned int linked
= 0;
2534 * debugobject calls are safe here even with pool->lock locked
2535 * as we know for sure that this will not trigger any of the
2536 * checks and call back into the fixup functions where we
2539 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2540 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2541 init_completion(&barr
->done
);
2544 * If @target is currently being executed, schedule the
2545 * barrier to the worker; otherwise, put it after @target.
2548 head
= worker
->scheduled
.next
;
2550 unsigned long *bits
= work_data_bits(target
);
2552 head
= target
->entry
.next
;
2553 /* there can already be other linked works, inherit and set */
2554 linked
= *bits
& WORK_STRUCT_LINKED
;
2555 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2558 debug_work_activate(&barr
->work
);
2559 insert_work(pwq
, &barr
->work
, head
,
2560 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2564 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2565 * @wq: workqueue being flushed
2566 * @flush_color: new flush color, < 0 for no-op
2567 * @work_color: new work color, < 0 for no-op
2569 * Prepare pwqs for workqueue flushing.
2571 * If @flush_color is non-negative, flush_color on all pwqs should be
2572 * -1. If no pwq has in-flight commands at the specified color, all
2573 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2574 * has in flight commands, its pwq->flush_color is set to
2575 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2576 * wakeup logic is armed and %true is returned.
2578 * The caller should have initialized @wq->first_flusher prior to
2579 * calling this function with non-negative @flush_color. If
2580 * @flush_color is negative, no flush color update is done and %false
2583 * If @work_color is non-negative, all pwqs should have the same
2584 * work_color which is previous to @work_color and all will be
2585 * advanced to @work_color.
2588 * mutex_lock(wq->flush_mutex).
2591 * %true if @flush_color >= 0 and there's something to flush. %false
2594 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2595 int flush_color
, int work_color
)
2598 struct pool_workqueue
*pwq
;
2600 if (flush_color
>= 0) {
2601 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2602 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2605 local_irq_disable();
2607 for_each_pwq(pwq
, wq
) {
2608 struct worker_pool
*pool
= pwq
->pool
;
2610 spin_lock(&pool
->lock
);
2612 if (flush_color
>= 0) {
2613 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2615 if (pwq
->nr_in_flight
[flush_color
]) {
2616 pwq
->flush_color
= flush_color
;
2617 atomic_inc(&wq
->nr_pwqs_to_flush
);
2622 if (work_color
>= 0) {
2623 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2624 pwq
->work_color
= work_color
;
2627 spin_unlock(&pool
->lock
);
2632 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2633 complete(&wq
->first_flusher
->done
);
2639 * flush_workqueue - ensure that any scheduled work has run to completion.
2640 * @wq: workqueue to flush
2642 * This function sleeps until all work items which were queued on entry
2643 * have finished execution, but it is not livelocked by new incoming ones.
2645 void flush_workqueue(struct workqueue_struct
*wq
)
2647 struct wq_flusher this_flusher
= {
2648 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2650 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2654 lock_map_acquire(&wq
->lockdep_map
);
2655 lock_map_release(&wq
->lockdep_map
);
2657 mutex_lock(&wq
->flush_mutex
);
2660 * Start-to-wait phase
2662 next_color
= work_next_color(wq
->work_color
);
2664 if (next_color
!= wq
->flush_color
) {
2666 * Color space is not full. The current work_color
2667 * becomes our flush_color and work_color is advanced
2670 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2671 this_flusher
.flush_color
= wq
->work_color
;
2672 wq
->work_color
= next_color
;
2674 if (!wq
->first_flusher
) {
2675 /* no flush in progress, become the first flusher */
2676 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2678 wq
->first_flusher
= &this_flusher
;
2680 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2682 /* nothing to flush, done */
2683 wq
->flush_color
= next_color
;
2684 wq
->first_flusher
= NULL
;
2689 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2690 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2691 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2695 * Oops, color space is full, wait on overflow queue.
2696 * The next flush completion will assign us
2697 * flush_color and transfer to flusher_queue.
2699 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2702 mutex_unlock(&wq
->flush_mutex
);
2704 wait_for_completion(&this_flusher
.done
);
2707 * Wake-up-and-cascade phase
2709 * First flushers are responsible for cascading flushes and
2710 * handling overflow. Non-first flushers can simply return.
2712 if (wq
->first_flusher
!= &this_flusher
)
2715 mutex_lock(&wq
->flush_mutex
);
2717 /* we might have raced, check again with mutex held */
2718 if (wq
->first_flusher
!= &this_flusher
)
2721 wq
->first_flusher
= NULL
;
2723 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2724 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2727 struct wq_flusher
*next
, *tmp
;
2729 /* complete all the flushers sharing the current flush color */
2730 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2731 if (next
->flush_color
!= wq
->flush_color
)
2733 list_del_init(&next
->list
);
2734 complete(&next
->done
);
2737 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2738 wq
->flush_color
!= work_next_color(wq
->work_color
));
2740 /* this flush_color is finished, advance by one */
2741 wq
->flush_color
= work_next_color(wq
->flush_color
);
2743 /* one color has been freed, handle overflow queue */
2744 if (!list_empty(&wq
->flusher_overflow
)) {
2746 * Assign the same color to all overflowed
2747 * flushers, advance work_color and append to
2748 * flusher_queue. This is the start-to-wait
2749 * phase for these overflowed flushers.
2751 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2752 tmp
->flush_color
= wq
->work_color
;
2754 wq
->work_color
= work_next_color(wq
->work_color
);
2756 list_splice_tail_init(&wq
->flusher_overflow
,
2757 &wq
->flusher_queue
);
2758 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2761 if (list_empty(&wq
->flusher_queue
)) {
2762 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2767 * Need to flush more colors. Make the next flusher
2768 * the new first flusher and arm pwqs.
2770 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2771 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2773 list_del_init(&next
->list
);
2774 wq
->first_flusher
= next
;
2776 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2780 * Meh... this color is already done, clear first
2781 * flusher and repeat cascading.
2783 wq
->first_flusher
= NULL
;
2787 mutex_unlock(&wq
->flush_mutex
);
2789 EXPORT_SYMBOL_GPL(flush_workqueue
);
2792 * drain_workqueue - drain a workqueue
2793 * @wq: workqueue to drain
2795 * Wait until the workqueue becomes empty. While draining is in progress,
2796 * only chain queueing is allowed. IOW, only currently pending or running
2797 * work items on @wq can queue further work items on it. @wq is flushed
2798 * repeatedly until it becomes empty. The number of flushing is detemined
2799 * by the depth of chaining and should be relatively short. Whine if it
2802 void drain_workqueue(struct workqueue_struct
*wq
)
2804 unsigned int flush_cnt
= 0;
2805 struct pool_workqueue
*pwq
;
2808 * __queue_work() needs to test whether there are drainers, is much
2809 * hotter than drain_workqueue() and already looks at @wq->flags.
2810 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2812 mutex_lock(&wq_mutex
);
2813 if (!wq
->nr_drainers
++)
2814 wq
->flags
|= __WQ_DRAINING
;
2815 mutex_unlock(&wq_mutex
);
2817 flush_workqueue(wq
);
2819 local_irq_disable();
2821 for_each_pwq(pwq
, wq
) {
2824 spin_lock(&pwq
->pool
->lock
);
2825 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2826 spin_unlock(&pwq
->pool
->lock
);
2831 if (++flush_cnt
== 10 ||
2832 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2833 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2834 wq
->name
, flush_cnt
);
2842 mutex_lock(&wq_mutex
);
2843 if (!--wq
->nr_drainers
)
2844 wq
->flags
&= ~__WQ_DRAINING
;
2845 mutex_unlock(&wq_mutex
);
2847 EXPORT_SYMBOL_GPL(drain_workqueue
);
2849 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2851 struct worker
*worker
= NULL
;
2852 struct worker_pool
*pool
;
2853 struct pool_workqueue
*pwq
;
2857 local_irq_disable();
2858 pool
= get_work_pool(work
);
2864 spin_lock(&pool
->lock
);
2865 /* see the comment in try_to_grab_pending() with the same code */
2866 pwq
= get_work_pwq(work
);
2868 if (unlikely(pwq
->pool
!= pool
))
2871 worker
= find_worker_executing_work(pool
, work
);
2874 pwq
= worker
->current_pwq
;
2877 insert_wq_barrier(pwq
, barr
, work
, worker
);
2878 spin_unlock_irq(&pool
->lock
);
2881 * If @max_active is 1 or rescuer is in use, flushing another work
2882 * item on the same workqueue may lead to deadlock. Make sure the
2883 * flusher is not running on the same workqueue by verifying write
2886 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2887 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2889 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2890 lock_map_release(&pwq
->wq
->lockdep_map
);
2894 spin_unlock_irq(&pool
->lock
);
2899 * flush_work - wait for a work to finish executing the last queueing instance
2900 * @work: the work to flush
2902 * Wait until @work has finished execution. @work is guaranteed to be idle
2903 * on return if it hasn't been requeued since flush started.
2906 * %true if flush_work() waited for the work to finish execution,
2907 * %false if it was already idle.
2909 bool flush_work(struct work_struct
*work
)
2911 struct wq_barrier barr
;
2913 lock_map_acquire(&work
->lockdep_map
);
2914 lock_map_release(&work
->lockdep_map
);
2916 if (start_flush_work(work
, &barr
)) {
2917 wait_for_completion(&barr
.done
);
2918 destroy_work_on_stack(&barr
.work
);
2924 EXPORT_SYMBOL_GPL(flush_work
);
2926 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2928 unsigned long flags
;
2932 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2934 * If someone else is canceling, wait for the same event it
2935 * would be waiting for before retrying.
2937 if (unlikely(ret
== -ENOENT
))
2939 } while (unlikely(ret
< 0));
2941 /* tell other tasks trying to grab @work to back off */
2942 mark_work_canceling(work
);
2943 local_irq_restore(flags
);
2946 clear_work_data(work
);
2951 * cancel_work_sync - cancel a work and wait for it to finish
2952 * @work: the work to cancel
2954 * Cancel @work and wait for its execution to finish. This function
2955 * can be used even if the work re-queues itself or migrates to
2956 * another workqueue. On return from this function, @work is
2957 * guaranteed to be not pending or executing on any CPU.
2959 * cancel_work_sync(&delayed_work->work) must not be used for
2960 * delayed_work's. Use cancel_delayed_work_sync() instead.
2962 * The caller must ensure that the workqueue on which @work was last
2963 * queued can't be destroyed before this function returns.
2966 * %true if @work was pending, %false otherwise.
2968 bool cancel_work_sync(struct work_struct
*work
)
2970 return __cancel_work_timer(work
, false);
2972 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2975 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2976 * @dwork: the delayed work to flush
2978 * Delayed timer is cancelled and the pending work is queued for
2979 * immediate execution. Like flush_work(), this function only
2980 * considers the last queueing instance of @dwork.
2983 * %true if flush_work() waited for the work to finish execution,
2984 * %false if it was already idle.
2986 bool flush_delayed_work(struct delayed_work
*dwork
)
2988 local_irq_disable();
2989 if (del_timer_sync(&dwork
->timer
))
2990 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2992 return flush_work(&dwork
->work
);
2994 EXPORT_SYMBOL(flush_delayed_work
);
2997 * cancel_delayed_work - cancel a delayed work
2998 * @dwork: delayed_work to cancel
3000 * Kill off a pending delayed_work. Returns %true if @dwork was pending
3001 * and canceled; %false if wasn't pending. Note that the work callback
3002 * function may still be running on return, unless it returns %true and the
3003 * work doesn't re-arm itself. Explicitly flush or use
3004 * cancel_delayed_work_sync() to wait on it.
3006 * This function is safe to call from any context including IRQ handler.
3008 bool cancel_delayed_work(struct delayed_work
*dwork
)
3010 unsigned long flags
;
3014 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
3015 } while (unlikely(ret
== -EAGAIN
));
3017 if (unlikely(ret
< 0))
3020 set_work_pool_and_clear_pending(&dwork
->work
,
3021 get_work_pool_id(&dwork
->work
));
3022 local_irq_restore(flags
);
3025 EXPORT_SYMBOL(cancel_delayed_work
);
3028 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3029 * @dwork: the delayed work cancel
3031 * This is cancel_work_sync() for delayed works.
3034 * %true if @dwork was pending, %false otherwise.
3036 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3038 return __cancel_work_timer(&dwork
->work
, true);
3040 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3043 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3044 * @func: the function to call
3046 * schedule_on_each_cpu() executes @func on each online CPU using the
3047 * system workqueue and blocks until all CPUs have completed.
3048 * schedule_on_each_cpu() is very slow.
3051 * 0 on success, -errno on failure.
3053 int schedule_on_each_cpu(work_func_t func
)
3056 struct work_struct __percpu
*works
;
3058 works
= alloc_percpu(struct work_struct
);
3064 for_each_online_cpu(cpu
) {
3065 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3067 INIT_WORK(work
, func
);
3068 schedule_work_on(cpu
, work
);
3071 for_each_online_cpu(cpu
)
3072 flush_work(per_cpu_ptr(works
, cpu
));
3080 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3082 * Forces execution of the kernel-global workqueue and blocks until its
3085 * Think twice before calling this function! It's very easy to get into
3086 * trouble if you don't take great care. Either of the following situations
3087 * will lead to deadlock:
3089 * One of the work items currently on the workqueue needs to acquire
3090 * a lock held by your code or its caller.
3092 * Your code is running in the context of a work routine.
3094 * They will be detected by lockdep when they occur, but the first might not
3095 * occur very often. It depends on what work items are on the workqueue and
3096 * what locks they need, which you have no control over.
3098 * In most situations flushing the entire workqueue is overkill; you merely
3099 * need to know that a particular work item isn't queued and isn't running.
3100 * In such cases you should use cancel_delayed_work_sync() or
3101 * cancel_work_sync() instead.
3103 void flush_scheduled_work(void)
3105 flush_workqueue(system_wq
);
3107 EXPORT_SYMBOL(flush_scheduled_work
);
3110 * execute_in_process_context - reliably execute the routine with user context
3111 * @fn: the function to execute
3112 * @ew: guaranteed storage for the execute work structure (must
3113 * be available when the work executes)
3115 * Executes the function immediately if process context is available,
3116 * otherwise schedules the function for delayed execution.
3118 * Returns: 0 - function was executed
3119 * 1 - function was scheduled for execution
3121 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3123 if (!in_interrupt()) {
3128 INIT_WORK(&ew
->work
, fn
);
3129 schedule_work(&ew
->work
);
3133 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3137 * Workqueues with WQ_SYSFS flag set is visible to userland via
3138 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3139 * following attributes.
3141 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3142 * max_active RW int : maximum number of in-flight work items
3144 * Unbound workqueues have the following extra attributes.
3146 * id RO int : the associated pool ID
3147 * nice RW int : nice value of the workers
3148 * cpumask RW mask : bitmask of allowed CPUs for the workers
3151 struct workqueue_struct
*wq
;
3155 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3157 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3162 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3163 struct device_attribute
*attr
, char *buf
)
3165 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3167 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3170 static ssize_t
wq_max_active_show(struct device
*dev
,
3171 struct device_attribute
*attr
, char *buf
)
3173 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3175 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3178 static ssize_t
wq_max_active_store(struct device
*dev
,
3179 struct device_attribute
*attr
,
3180 const char *buf
, size_t count
)
3182 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3185 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3188 workqueue_set_max_active(wq
, val
);
3192 static struct device_attribute wq_sysfs_attrs
[] = {
3193 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3194 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3198 static ssize_t
wq_pool_id_show(struct device
*dev
,
3199 struct device_attribute
*attr
, char *buf
)
3201 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3202 struct worker_pool
*pool
;
3205 rcu_read_lock_sched();
3206 pool
= first_pwq(wq
)->pool
;
3207 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", pool
->id
);
3208 rcu_read_unlock_sched();
3213 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3216 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3219 rcu_read_lock_sched();
3220 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3221 first_pwq(wq
)->pool
->attrs
->nice
);
3222 rcu_read_unlock_sched();
3227 /* prepare workqueue_attrs for sysfs store operations */
3228 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3230 struct workqueue_attrs
*attrs
;
3232 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3236 rcu_read_lock_sched();
3237 copy_workqueue_attrs(attrs
, first_pwq(wq
)->pool
->attrs
);
3238 rcu_read_unlock_sched();
3242 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3243 const char *buf
, size_t count
)
3245 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3246 struct workqueue_attrs
*attrs
;
3249 attrs
= wq_sysfs_prep_attrs(wq
);
3253 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3254 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3255 ret
= apply_workqueue_attrs(wq
, attrs
);
3259 free_workqueue_attrs(attrs
);
3260 return ret
?: count
;
3263 static ssize_t
wq_cpumask_show(struct device
*dev
,
3264 struct device_attribute
*attr
, char *buf
)
3266 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3269 rcu_read_lock_sched();
3270 written
= cpumask_scnprintf(buf
, PAGE_SIZE
,
3271 first_pwq(wq
)->pool
->attrs
->cpumask
);
3272 rcu_read_unlock_sched();
3274 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3278 static ssize_t
wq_cpumask_store(struct device
*dev
,
3279 struct device_attribute
*attr
,
3280 const char *buf
, size_t count
)
3282 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3283 struct workqueue_attrs
*attrs
;
3286 attrs
= wq_sysfs_prep_attrs(wq
);
3290 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3292 ret
= apply_workqueue_attrs(wq
, attrs
);
3294 free_workqueue_attrs(attrs
);
3295 return ret
?: count
;
3298 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3299 __ATTR(pool_id
, 0444, wq_pool_id_show
, NULL
),
3300 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3301 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3305 static struct bus_type wq_subsys
= {
3306 .name
= "workqueue",
3307 .dev_attrs
= wq_sysfs_attrs
,
3310 static int __init
wq_sysfs_init(void)
3312 return subsys_virtual_register(&wq_subsys
, NULL
);
3314 core_initcall(wq_sysfs_init
);
3316 static void wq_device_release(struct device
*dev
)
3318 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3324 * workqueue_sysfs_register - make a workqueue visible in sysfs
3325 * @wq: the workqueue to register
3327 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3328 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3329 * which is the preferred method.
3331 * Workqueue user should use this function directly iff it wants to apply
3332 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3333 * apply_workqueue_attrs() may race against userland updating the
3336 * Returns 0 on success, -errno on failure.
3338 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3340 struct wq_device
*wq_dev
;
3344 * Adjusting max_active or creating new pwqs by applyting
3345 * attributes breaks ordering guarantee. Disallow exposing ordered
3348 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3351 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3356 wq_dev
->dev
.bus
= &wq_subsys
;
3357 wq_dev
->dev
.init_name
= wq
->name
;
3358 wq_dev
->dev
.release
= wq_device_release
;
3361 * unbound_attrs are created separately. Suppress uevent until
3362 * everything is ready.
3364 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3366 ret
= device_register(&wq_dev
->dev
);
3373 if (wq
->flags
& WQ_UNBOUND
) {
3374 struct device_attribute
*attr
;
3376 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3377 ret
= device_create_file(&wq_dev
->dev
, attr
);
3379 device_unregister(&wq_dev
->dev
);
3386 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3391 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3392 * @wq: the workqueue to unregister
3394 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3396 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3398 struct wq_device
*wq_dev
= wq
->wq_dev
;
3404 device_unregister(&wq_dev
->dev
);
3406 #else /* CONFIG_SYSFS */
3407 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3408 #endif /* CONFIG_SYSFS */
3411 * free_workqueue_attrs - free a workqueue_attrs
3412 * @attrs: workqueue_attrs to free
3414 * Undo alloc_workqueue_attrs().
3416 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3419 free_cpumask_var(attrs
->cpumask
);
3425 * alloc_workqueue_attrs - allocate a workqueue_attrs
3426 * @gfp_mask: allocation mask to use
3428 * Allocate a new workqueue_attrs, initialize with default settings and
3429 * return it. Returns NULL on failure.
3431 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3433 struct workqueue_attrs
*attrs
;
3435 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3438 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3441 cpumask_setall(attrs
->cpumask
);
3444 free_workqueue_attrs(attrs
);
3448 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3449 const struct workqueue_attrs
*from
)
3451 to
->nice
= from
->nice
;
3452 cpumask_copy(to
->cpumask
, from
->cpumask
);
3456 * Hacky implementation of jhash of bitmaps which only considers the
3457 * specified number of bits. We probably want a proper implementation in
3458 * include/linux/jhash.h.
3460 static u32
jhash_bitmap(const unsigned long *bitmap
, int bits
, u32 hash
)
3462 int nr_longs
= bits
/ BITS_PER_LONG
;
3463 int nr_leftover
= bits
% BITS_PER_LONG
;
3464 unsigned long leftover
= 0;
3467 hash
= jhash(bitmap
, nr_longs
* sizeof(long), hash
);
3469 bitmap_copy(&leftover
, bitmap
+ nr_longs
, nr_leftover
);
3470 hash
= jhash(&leftover
, sizeof(long), hash
);
3475 /* hash value of the content of @attr */
3476 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3480 hash
= jhash_1word(attrs
->nice
, hash
);
3481 hash
= jhash_bitmap(cpumask_bits(attrs
->cpumask
), nr_cpu_ids
, hash
);
3485 /* content equality test */
3486 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3487 const struct workqueue_attrs
*b
)
3489 if (a
->nice
!= b
->nice
)
3491 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3497 * init_worker_pool - initialize a newly zalloc'd worker_pool
3498 * @pool: worker_pool to initialize
3500 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3501 * Returns 0 on success, -errno on failure. Even on failure, all fields
3502 * inside @pool proper are initialized and put_unbound_pool() can be called
3503 * on @pool safely to release it.
3505 static int init_worker_pool(struct worker_pool
*pool
)
3507 spin_lock_init(&pool
->lock
);
3510 pool
->flags
|= POOL_DISASSOCIATED
;
3511 INIT_LIST_HEAD(&pool
->worklist
);
3512 INIT_LIST_HEAD(&pool
->idle_list
);
3513 hash_init(pool
->busy_hash
);
3515 init_timer_deferrable(&pool
->idle_timer
);
3516 pool
->idle_timer
.function
= idle_worker_timeout
;
3517 pool
->idle_timer
.data
= (unsigned long)pool
;
3519 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3520 (unsigned long)pool
);
3522 mutex_init(&pool
->manager_arb
);
3523 mutex_init(&pool
->manager_mutex
);
3524 idr_init(&pool
->worker_idr
);
3526 INIT_HLIST_NODE(&pool
->hash_node
);
3529 /* shouldn't fail above this point */
3530 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3536 static void rcu_free_pool(struct rcu_head
*rcu
)
3538 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3540 idr_destroy(&pool
->worker_idr
);
3541 free_workqueue_attrs(pool
->attrs
);
3546 * put_unbound_pool - put a worker_pool
3547 * @pool: worker_pool to put
3549 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3550 * safe manner. get_unbound_pool() calls this function on its failure path
3551 * and this function should be able to release pools which went through,
3552 * successfully or not, init_worker_pool().
3554 static void put_unbound_pool(struct worker_pool
*pool
)
3556 struct worker
*worker
;
3558 mutex_lock(&wq_mutex
);
3559 if (--pool
->refcnt
) {
3560 mutex_unlock(&wq_mutex
);
3565 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3566 WARN_ON(!list_empty(&pool
->worklist
))) {
3567 mutex_unlock(&wq_mutex
);
3571 /* release id and unhash */
3573 idr_remove(&worker_pool_idr
, pool
->id
);
3574 hash_del(&pool
->hash_node
);
3576 mutex_unlock(&wq_mutex
);
3579 * Become the manager and destroy all workers. Grabbing
3580 * manager_arb prevents @pool's workers from blocking on
3583 mutex_lock(&pool
->manager_arb
);
3584 mutex_lock(&pool
->manager_mutex
);
3585 spin_lock_irq(&pool
->lock
);
3587 while ((worker
= first_worker(pool
)))
3588 destroy_worker(worker
);
3589 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3591 spin_unlock_irq(&pool
->lock
);
3592 mutex_unlock(&pool
->manager_mutex
);
3593 mutex_unlock(&pool
->manager_arb
);
3595 /* shut down the timers */
3596 del_timer_sync(&pool
->idle_timer
);
3597 del_timer_sync(&pool
->mayday_timer
);
3599 /* sched-RCU protected to allow dereferences from get_work_pool() */
3600 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3604 * get_unbound_pool - get a worker_pool with the specified attributes
3605 * @attrs: the attributes of the worker_pool to get
3607 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3608 * reference count and return it. If there already is a matching
3609 * worker_pool, it will be used; otherwise, this function attempts to
3610 * create a new one. On failure, returns NULL.
3612 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3614 u32 hash
= wqattrs_hash(attrs
);
3615 struct worker_pool
*pool
;
3617 mutex_lock(&wq_mutex
);
3619 /* do we already have a matching pool? */
3620 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3621 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3627 /* nope, create a new one */
3628 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3629 if (!pool
|| init_worker_pool(pool
) < 0)
3632 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3633 copy_workqueue_attrs(pool
->attrs
, attrs
);
3635 if (worker_pool_assign_id(pool
) < 0)
3638 /* create and start the initial worker */
3639 if (create_and_start_worker(pool
) < 0)
3643 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3645 mutex_unlock(&wq_mutex
);
3648 mutex_unlock(&wq_mutex
);
3650 put_unbound_pool(pool
);
3654 static void rcu_free_pwq(struct rcu_head
*rcu
)
3656 kmem_cache_free(pwq_cache
,
3657 container_of(rcu
, struct pool_workqueue
, rcu
));
3661 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3662 * and needs to be destroyed.
3664 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3666 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3667 unbound_release_work
);
3668 struct workqueue_struct
*wq
= pwq
->wq
;
3669 struct worker_pool
*pool
= pwq
->pool
;
3671 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3675 * Unlink @pwq. Synchronization against flush_mutex isn't strictly
3676 * necessary on release but do it anyway. It's easier to verify
3677 * and consistent with the linking path.
3679 mutex_lock(&wq
->flush_mutex
);
3680 spin_lock_irq(&pwq_lock
);
3681 list_del_rcu(&pwq
->pwqs_node
);
3682 spin_unlock_irq(&pwq_lock
);
3683 mutex_unlock(&wq
->flush_mutex
);
3685 put_unbound_pool(pool
);
3686 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3689 * If we're the last pwq going away, @wq is already dead and no one
3690 * is gonna access it anymore. Free it.
3692 if (list_empty(&wq
->pwqs
))
3697 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3698 * @pwq: target pool_workqueue
3700 * If @pwq isn't freezing, set @pwq->max_active to the associated
3701 * workqueue's saved_max_active and activate delayed work items
3702 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3704 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3706 struct workqueue_struct
*wq
= pwq
->wq
;
3707 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3709 /* for @wq->saved_max_active */
3710 lockdep_assert_held(&pwq_lock
);
3712 /* fast exit for non-freezable wqs */
3713 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3716 spin_lock(&pwq
->pool
->lock
);
3718 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3719 pwq
->max_active
= wq
->saved_max_active
;
3721 while (!list_empty(&pwq
->delayed_works
) &&
3722 pwq
->nr_active
< pwq
->max_active
)
3723 pwq_activate_first_delayed(pwq
);
3725 pwq
->max_active
= 0;
3728 spin_unlock(&pwq
->pool
->lock
);
3731 static void init_and_link_pwq(struct pool_workqueue
*pwq
,
3732 struct workqueue_struct
*wq
,
3733 struct worker_pool
*pool
,
3734 struct pool_workqueue
**p_last_pwq
)
3736 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3740 pwq
->flush_color
= -1;
3742 INIT_LIST_HEAD(&pwq
->delayed_works
);
3743 INIT_LIST_HEAD(&pwq
->mayday_node
);
3744 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3746 mutex_lock(&wq
->flush_mutex
);
3747 spin_lock_irq(&pwq_lock
);
3750 * Set the matching work_color. This is synchronized with
3751 * flush_mutex to avoid confusing flush_workqueue().
3754 *p_last_pwq
= first_pwq(wq
);
3755 pwq
->work_color
= wq
->work_color
;
3757 /* sync max_active to the current setting */
3758 pwq_adjust_max_active(pwq
);
3761 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3763 spin_unlock_irq(&pwq_lock
);
3764 mutex_unlock(&wq
->flush_mutex
);
3768 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3769 * @wq: the target workqueue
3770 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3772 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3773 * current attributes, a new pwq is created and made the first pwq which
3774 * will serve all new work items. Older pwqs are released as in-flight
3775 * work items finish. Note that a work item which repeatedly requeues
3776 * itself back-to-back will stay on its current pwq.
3778 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3781 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3782 const struct workqueue_attrs
*attrs
)
3784 struct pool_workqueue
*pwq
, *last_pwq
;
3785 struct worker_pool
*pool
;
3787 /* only unbound workqueues can change attributes */
3788 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3791 /* creating multiple pwqs breaks ordering guarantee */
3792 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3795 pwq
= kmem_cache_zalloc(pwq_cache
, GFP_KERNEL
);
3799 pool
= get_unbound_pool(attrs
);
3801 kmem_cache_free(pwq_cache
, pwq
);
3805 init_and_link_pwq(pwq
, wq
, pool
, &last_pwq
);
3807 spin_lock_irq(&last_pwq
->pool
->lock
);
3809 spin_unlock_irq(&last_pwq
->pool
->lock
);
3815 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3817 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3820 if (!(wq
->flags
& WQ_UNBOUND
)) {
3821 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3825 for_each_possible_cpu(cpu
) {
3826 struct pool_workqueue
*pwq
=
3827 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3828 struct worker_pool
*cpu_pools
=
3829 per_cpu(cpu_worker_pools
, cpu
);
3831 init_and_link_pwq(pwq
, wq
, &cpu_pools
[highpri
], NULL
);
3835 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3839 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3842 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3844 if (max_active
< 1 || max_active
> lim
)
3845 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3846 max_active
, name
, 1, lim
);
3848 return clamp_val(max_active
, 1, lim
);
3851 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3854 struct lock_class_key
*key
,
3855 const char *lock_name
, ...)
3857 va_list args
, args1
;
3858 struct workqueue_struct
*wq
;
3859 struct pool_workqueue
*pwq
;
3862 /* determine namelen, allocate wq and format name */
3863 va_start(args
, lock_name
);
3864 va_copy(args1
, args
);
3865 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3867 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3871 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3875 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3876 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3880 wq
->saved_max_active
= max_active
;
3881 mutex_init(&wq
->flush_mutex
);
3882 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3883 INIT_LIST_HEAD(&wq
->pwqs
);
3884 INIT_LIST_HEAD(&wq
->flusher_queue
);
3885 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3886 INIT_LIST_HEAD(&wq
->maydays
);
3888 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3889 INIT_LIST_HEAD(&wq
->list
);
3891 if (alloc_and_link_pwqs(wq
) < 0)
3895 * Workqueues which may be used during memory reclaim should
3896 * have a rescuer to guarantee forward progress.
3898 if (flags
& WQ_MEM_RECLAIM
) {
3899 struct worker
*rescuer
;
3901 rescuer
= alloc_worker();
3905 rescuer
->rescue_wq
= wq
;
3906 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3908 if (IS_ERR(rescuer
->task
)) {
3913 wq
->rescuer
= rescuer
;
3914 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
3915 wake_up_process(rescuer
->task
);
3918 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
3922 * wq_mutex protects global freeze state and workqueues list. Grab
3923 * it, adjust max_active and add the new @wq to workqueues list.
3925 mutex_lock(&wq_mutex
);
3927 spin_lock_irq(&pwq_lock
);
3928 for_each_pwq(pwq
, wq
)
3929 pwq_adjust_max_active(pwq
);
3930 spin_unlock_irq(&pwq_lock
);
3932 list_add(&wq
->list
, &workqueues
);
3934 mutex_unlock(&wq_mutex
);
3942 destroy_workqueue(wq
);
3945 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3948 * destroy_workqueue - safely terminate a workqueue
3949 * @wq: target workqueue
3951 * Safely destroy a workqueue. All work currently pending will be done first.
3953 void destroy_workqueue(struct workqueue_struct
*wq
)
3955 struct pool_workqueue
*pwq
;
3957 /* drain it before proceeding with destruction */
3958 drain_workqueue(wq
);
3961 spin_lock_irq(&pwq_lock
);
3962 for_each_pwq(pwq
, wq
) {
3965 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
3966 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
3967 spin_unlock_irq(&pwq_lock
);
3972 if (WARN_ON(pwq
->refcnt
> 1) ||
3973 WARN_ON(pwq
->nr_active
) ||
3974 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
3975 spin_unlock_irq(&pwq_lock
);
3979 spin_unlock_irq(&pwq_lock
);
3982 * wq list is used to freeze wq, remove from list after
3983 * flushing is complete in case freeze races us.
3985 mutex_lock(&wq_mutex
);
3986 list_del_init(&wq
->list
);
3987 mutex_unlock(&wq_mutex
);
3989 workqueue_sysfs_unregister(wq
);
3992 kthread_stop(wq
->rescuer
->task
);
3997 if (!(wq
->flags
& WQ_UNBOUND
)) {
3999 * The base ref is never dropped on per-cpu pwqs. Directly
4000 * free the pwqs and wq.
4002 free_percpu(wq
->cpu_pwqs
);
4006 * We're the sole accessor of @wq at this point. Directly
4007 * access the first pwq and put the base ref. As both pwqs
4008 * and pools are sched-RCU protected, the lock operations
4009 * are safe. @wq will be freed when the last pwq is
4012 pwq
= list_first_entry(&wq
->pwqs
, struct pool_workqueue
,
4014 spin_lock_irq(&pwq
->pool
->lock
);
4016 spin_unlock_irq(&pwq
->pool
->lock
);
4019 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4022 * workqueue_set_max_active - adjust max_active of a workqueue
4023 * @wq: target workqueue
4024 * @max_active: new max_active value.
4026 * Set max_active of @wq to @max_active.
4029 * Don't call from IRQ context.
4031 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4033 struct pool_workqueue
*pwq
;
4035 /* disallow meddling with max_active for ordered workqueues */
4036 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4039 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4041 spin_lock_irq(&pwq_lock
);
4043 wq
->saved_max_active
= max_active
;
4045 for_each_pwq(pwq
, wq
)
4046 pwq_adjust_max_active(pwq
);
4048 spin_unlock_irq(&pwq_lock
);
4050 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4053 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4055 * Determine whether %current is a workqueue rescuer. Can be used from
4056 * work functions to determine whether it's being run off the rescuer task.
4058 bool current_is_workqueue_rescuer(void)
4060 struct worker
*worker
= current_wq_worker();
4062 return worker
&& worker
== worker
->current_pwq
->wq
->rescuer
;
4066 * workqueue_congested - test whether a workqueue is congested
4067 * @cpu: CPU in question
4068 * @wq: target workqueue
4070 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4071 * no synchronization around this function and the test result is
4072 * unreliable and only useful as advisory hints or for debugging.
4075 * %true if congested, %false otherwise.
4077 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4079 struct pool_workqueue
*pwq
;
4084 if (!(wq
->flags
& WQ_UNBOUND
))
4085 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4087 pwq
= first_pwq(wq
);
4089 ret
= !list_empty(&pwq
->delayed_works
);
4094 EXPORT_SYMBOL_GPL(workqueue_congested
);
4097 * work_busy - test whether a work is currently pending or running
4098 * @work: the work to be tested
4100 * Test whether @work is currently pending or running. There is no
4101 * synchronization around this function and the test result is
4102 * unreliable and only useful as advisory hints or for debugging.
4105 * OR'd bitmask of WORK_BUSY_* bits.
4107 unsigned int work_busy(struct work_struct
*work
)
4109 struct worker_pool
*pool
;
4110 unsigned long flags
;
4111 unsigned int ret
= 0;
4113 if (work_pending(work
))
4114 ret
|= WORK_BUSY_PENDING
;
4116 local_irq_save(flags
);
4117 pool
= get_work_pool(work
);
4119 spin_lock(&pool
->lock
);
4120 if (find_worker_executing_work(pool
, work
))
4121 ret
|= WORK_BUSY_RUNNING
;
4122 spin_unlock(&pool
->lock
);
4124 local_irq_restore(flags
);
4128 EXPORT_SYMBOL_GPL(work_busy
);
4133 * There are two challenges in supporting CPU hotplug. Firstly, there
4134 * are a lot of assumptions on strong associations among work, pwq and
4135 * pool which make migrating pending and scheduled works very
4136 * difficult to implement without impacting hot paths. Secondly,
4137 * worker pools serve mix of short, long and very long running works making
4138 * blocked draining impractical.
4140 * This is solved by allowing the pools to be disassociated from the CPU
4141 * running as an unbound one and allowing it to be reattached later if the
4142 * cpu comes back online.
4145 static void wq_unbind_fn(struct work_struct
*work
)
4147 int cpu
= smp_processor_id();
4148 struct worker_pool
*pool
;
4149 struct worker
*worker
;
4152 for_each_cpu_worker_pool(pool
, cpu
) {
4153 WARN_ON_ONCE(cpu
!= smp_processor_id());
4155 mutex_lock(&pool
->manager_mutex
);
4156 spin_lock_irq(&pool
->lock
);
4159 * We've blocked all manager operations. Make all workers
4160 * unbound and set DISASSOCIATED. Before this, all workers
4161 * except for the ones which are still executing works from
4162 * before the last CPU down must be on the cpu. After
4163 * this, they may become diasporas.
4165 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
4166 worker
->flags
|= WORKER_UNBOUND
;
4168 for_each_busy_worker(worker
, i
, pool
)
4169 worker
->flags
|= WORKER_UNBOUND
;
4171 pool
->flags
|= POOL_DISASSOCIATED
;
4173 spin_unlock_irq(&pool
->lock
);
4174 mutex_unlock(&pool
->manager_mutex
);
4178 * Call schedule() so that we cross rq->lock and thus can guarantee
4179 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4180 * as scheduler callbacks may be invoked from other cpus.
4185 * Sched callbacks are disabled now. Zap nr_running. After this,
4186 * nr_running stays zero and need_more_worker() and keep_working()
4187 * are always true as long as the worklist is not empty. Pools on
4188 * @cpu now behave as unbound (in terms of concurrency management)
4189 * pools which are served by workers tied to the CPU.
4191 * On return from this function, the current worker would trigger
4192 * unbound chain execution of pending work items if other workers
4195 for_each_cpu_worker_pool(pool
, cpu
)
4196 atomic_set(&pool
->nr_running
, 0);
4200 * Workqueues should be brought up before normal priority CPU notifiers.
4201 * This will be registered high priority CPU notifier.
4203 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4204 unsigned long action
,
4207 int cpu
= (unsigned long)hcpu
;
4208 struct worker_pool
*pool
;
4210 switch (action
& ~CPU_TASKS_FROZEN
) {
4211 case CPU_UP_PREPARE
:
4212 for_each_cpu_worker_pool(pool
, cpu
) {
4213 if (pool
->nr_workers
)
4215 if (create_and_start_worker(pool
) < 0)
4220 case CPU_DOWN_FAILED
:
4222 for_each_cpu_worker_pool(pool
, cpu
) {
4223 mutex_lock(&pool
->manager_mutex
);
4224 spin_lock_irq(&pool
->lock
);
4226 pool
->flags
&= ~POOL_DISASSOCIATED
;
4227 rebind_workers(pool
);
4229 spin_unlock_irq(&pool
->lock
);
4230 mutex_unlock(&pool
->manager_mutex
);
4238 * Workqueues should be brought down after normal priority CPU notifiers.
4239 * This will be registered as low priority CPU notifier.
4241 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4242 unsigned long action
,
4245 int cpu
= (unsigned long)hcpu
;
4246 struct work_struct unbind_work
;
4248 switch (action
& ~CPU_TASKS_FROZEN
) {
4249 case CPU_DOWN_PREPARE
:
4250 /* unbinding should happen on the local CPU */
4251 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4252 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4253 flush_work(&unbind_work
);
4261 struct work_for_cpu
{
4262 struct work_struct work
;
4268 static void work_for_cpu_fn(struct work_struct
*work
)
4270 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4272 wfc
->ret
= wfc
->fn(wfc
->arg
);
4276 * work_on_cpu - run a function in user context on a particular cpu
4277 * @cpu: the cpu to run on
4278 * @fn: the function to run
4279 * @arg: the function arg
4281 * This will return the value @fn returns.
4282 * It is up to the caller to ensure that the cpu doesn't go offline.
4283 * The caller must not hold any locks which would prevent @fn from completing.
4285 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4287 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4289 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4290 schedule_work_on(cpu
, &wfc
.work
);
4291 flush_work(&wfc
.work
);
4294 EXPORT_SYMBOL_GPL(work_on_cpu
);
4295 #endif /* CONFIG_SMP */
4297 #ifdef CONFIG_FREEZER
4300 * freeze_workqueues_begin - begin freezing workqueues
4302 * Start freezing workqueues. After this function returns, all freezable
4303 * workqueues will queue new works to their delayed_works list instead of
4307 * Grabs and releases wq_mutex, pwq_lock and pool->lock's.
4309 void freeze_workqueues_begin(void)
4311 struct worker_pool
*pool
;
4312 struct workqueue_struct
*wq
;
4313 struct pool_workqueue
*pwq
;
4316 mutex_lock(&wq_mutex
);
4318 WARN_ON_ONCE(workqueue_freezing
);
4319 workqueue_freezing
= true;
4322 for_each_pool(pool
, pi
) {
4323 spin_lock_irq(&pool
->lock
);
4324 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4325 pool
->flags
|= POOL_FREEZING
;
4326 spin_unlock_irq(&pool
->lock
);
4329 /* suppress further executions by setting max_active to zero */
4330 spin_lock_irq(&pwq_lock
);
4331 list_for_each_entry(wq
, &workqueues
, list
) {
4332 for_each_pwq(pwq
, wq
)
4333 pwq_adjust_max_active(pwq
);
4335 spin_unlock_irq(&pwq_lock
);
4337 mutex_unlock(&wq_mutex
);
4341 * freeze_workqueues_busy - are freezable workqueues still busy?
4343 * Check whether freezing is complete. This function must be called
4344 * between freeze_workqueues_begin() and thaw_workqueues().
4347 * Grabs and releases wq_mutex.
4350 * %true if some freezable workqueues are still busy. %false if freezing
4353 bool freeze_workqueues_busy(void)
4356 struct workqueue_struct
*wq
;
4357 struct pool_workqueue
*pwq
;
4359 mutex_lock(&wq_mutex
);
4361 WARN_ON_ONCE(!workqueue_freezing
);
4363 list_for_each_entry(wq
, &workqueues
, list
) {
4364 if (!(wq
->flags
& WQ_FREEZABLE
))
4367 * nr_active is monotonically decreasing. It's safe
4368 * to peek without lock.
4371 for_each_pwq(pwq
, wq
) {
4372 WARN_ON_ONCE(pwq
->nr_active
< 0);
4373 if (pwq
->nr_active
) {
4382 mutex_unlock(&wq_mutex
);
4387 * thaw_workqueues - thaw workqueues
4389 * Thaw workqueues. Normal queueing is restored and all collected
4390 * frozen works are transferred to their respective pool worklists.
4393 * Grabs and releases wq_mutex, pwq_lock and pool->lock's.
4395 void thaw_workqueues(void)
4397 struct workqueue_struct
*wq
;
4398 struct pool_workqueue
*pwq
;
4399 struct worker_pool
*pool
;
4402 mutex_lock(&wq_mutex
);
4404 if (!workqueue_freezing
)
4407 /* clear FREEZING */
4408 for_each_pool(pool
, pi
) {
4409 spin_lock_irq(&pool
->lock
);
4410 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4411 pool
->flags
&= ~POOL_FREEZING
;
4412 spin_unlock_irq(&pool
->lock
);
4415 /* restore max_active and repopulate worklist */
4416 spin_lock_irq(&pwq_lock
);
4417 list_for_each_entry(wq
, &workqueues
, list
) {
4418 for_each_pwq(pwq
, wq
)
4419 pwq_adjust_max_active(pwq
);
4421 spin_unlock_irq(&pwq_lock
);
4424 for_each_pool(pool
, pi
) {
4425 spin_lock_irq(&pool
->lock
);
4426 wake_up_worker(pool
);
4427 spin_unlock_irq(&pool
->lock
);
4430 workqueue_freezing
= false;
4432 mutex_unlock(&wq_mutex
);
4434 #endif /* CONFIG_FREEZER */
4436 static int __init
init_workqueues(void)
4438 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4441 /* make sure we have enough bits for OFFQ pool ID */
4442 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
4443 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
4445 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4447 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4449 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4450 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4452 /* initialize CPU pools */
4453 for_each_possible_cpu(cpu
) {
4454 struct worker_pool
*pool
;
4457 for_each_cpu_worker_pool(pool
, cpu
) {
4458 BUG_ON(init_worker_pool(pool
));
4460 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4461 pool
->attrs
->nice
= std_nice
[i
++];
4464 mutex_lock(&wq_mutex
);
4465 BUG_ON(worker_pool_assign_id(pool
));
4466 mutex_unlock(&wq_mutex
);
4470 /* create the initial worker */
4471 for_each_online_cpu(cpu
) {
4472 struct worker_pool
*pool
;
4474 for_each_cpu_worker_pool(pool
, cpu
) {
4475 pool
->flags
&= ~POOL_DISASSOCIATED
;
4476 BUG_ON(create_and_start_worker(pool
) < 0);
4480 /* create default unbound wq attrs */
4481 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4482 struct workqueue_attrs
*attrs
;
4484 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4486 attrs
->nice
= std_nice
[i
];
4487 cpumask_setall(attrs
->cpumask
);
4489 unbound_std_wq_attrs
[i
] = attrs
;
4492 system_wq
= alloc_workqueue("events", 0, 0);
4493 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4494 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4495 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4496 WQ_UNBOUND_MAX_ACTIVE
);
4497 system_freezable_wq
= alloc_workqueue("events_freezable",
4499 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
4500 !system_unbound_wq
|| !system_freezable_wq
);
4503 early_initcall(init_workqueues
);