2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE
= 1 << 1, /* die die die */
75 WORKER_IDLE
= 1 << 2, /* is idle */
76 WORKER_PREP
= 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
79 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
82 WORKER_UNBOUND
| WORKER_REBOUND
,
84 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
96 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL
= MIN_NICE
,
103 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * WQ: wq->mutex protected.
132 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
134 * MD: wq_mayday_lock protected.
137 /* struct worker is defined in workqueue_internal.h */
140 spinlock_t lock
; /* the pool lock */
141 int cpu
; /* I: the associated cpu */
142 int node
; /* I: the associated node ID */
143 int id
; /* I: pool ID */
144 unsigned int flags
; /* X: flags */
146 struct list_head worklist
; /* L: list of pending works */
147 int nr_workers
; /* L: total number of workers */
149 /* nr_idle includes the ones off idle_list for rebinding */
150 int nr_idle
; /* L: currently idle ones */
152 struct list_head idle_list
; /* X: list of idle workers */
153 struct timer_list idle_timer
; /* L: worker idle timeout */
154 struct timer_list mayday_timer
; /* L: SOS timer for workers */
156 /* a workers is either on busy_hash or idle_list, or the manager */
157 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
158 /* L: hash of busy workers */
160 /* see manage_workers() for details on the two manager mutexes */
161 struct mutex manager_arb
; /* manager arbitration */
162 struct mutex attach_mutex
; /* attach/detach exclusion */
163 struct list_head workers
; /* A: attached workers */
164 struct completion
*detach_completion
; /* all workers detached */
166 struct ida worker_ida
; /* worker IDs for task name */
168 struct workqueue_attrs
*attrs
; /* I: worker attributes */
169 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
170 int refcnt
; /* PL: refcnt for unbound pools */
173 * The current concurrency level. As it's likely to be accessed
174 * from other CPUs during try_to_wake_up(), put it in a separate
177 atomic_t nr_running ____cacheline_aligned_in_smp
;
180 * Destruction of pool is sched-RCU protected to allow dereferences
181 * from get_work_pool().
184 } ____cacheline_aligned_in_smp
;
187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 * of work_struct->data are used for flags and the remaining high bits
189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 * number of flag bits.
192 struct pool_workqueue
{
193 struct worker_pool
*pool
; /* I: the associated pool */
194 struct workqueue_struct
*wq
; /* I: the owning workqueue */
195 int work_color
; /* L: current color */
196 int flush_color
; /* L: flushing color */
197 int refcnt
; /* L: reference count */
198 int nr_in_flight
[WORK_NR_COLORS
];
199 /* L: nr of in_flight works */
200 int nr_active
; /* L: nr of active works */
201 int max_active
; /* L: max active works */
202 struct list_head delayed_works
; /* L: delayed works */
203 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
204 struct list_head mayday_node
; /* MD: node on wq->maydays */
207 * Release of unbound pwq is punted to system_wq. See put_pwq()
208 * and pwq_unbound_release_workfn() for details. pool_workqueue
209 * itself is also sched-RCU protected so that the first pwq can be
210 * determined without grabbing wq->mutex.
212 struct work_struct unbound_release_work
;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
217 * Structure used to wait for workqueue flush.
220 struct list_head list
; /* WQ: list of flushers */
221 int flush_color
; /* WQ: flush color waiting for */
222 struct completion done
; /* flush completion */
228 * The externally visible workqueue. It relays the issued work items to
229 * the appropriate worker_pool through its pool_workqueues.
231 struct workqueue_struct
{
232 struct list_head pwqs
; /* WR: all pwqs of this wq */
233 struct list_head list
; /* PL: list of all workqueues */
235 struct mutex mutex
; /* protects this wq */
236 int work_color
; /* WQ: current work color */
237 int flush_color
; /* WQ: current flush color */
238 atomic_t nr_pwqs_to_flush
; /* flush in progress */
239 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
240 struct list_head flusher_queue
; /* WQ: flush waiters */
241 struct list_head flusher_overflow
; /* WQ: flush overflow list */
243 struct list_head maydays
; /* MD: pwqs requesting rescue */
244 struct worker
*rescuer
; /* I: rescue worker */
246 int nr_drainers
; /* WQ: drain in progress */
247 int saved_max_active
; /* WQ: saved pwq max_active */
249 struct workqueue_attrs
*unbound_attrs
; /* WQ: only for unbound wqs */
250 struct pool_workqueue
*dfl_pwq
; /* WQ: only for unbound wqs */
253 struct wq_device
*wq_dev
; /* I: for sysfs interface */
255 #ifdef CONFIG_LOCKDEP
256 struct lockdep_map lockdep_map
;
258 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
260 /* hot fields used during command issue, aligned to cacheline */
261 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
262 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
263 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* FR: unbound pwqs indexed by node */
266 static struct kmem_cache
*pwq_cache
;
268 static int wq_numa_tbl_len
; /* highest possible NUMA node id + 1 */
269 static cpumask_var_t
*wq_numa_possible_cpumask
;
270 /* possible CPUs of each node */
272 static bool wq_disable_numa
;
273 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
275 /* see the comment above the definition of WQ_POWER_EFFICIENT */
276 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
277 static bool wq_power_efficient
= true;
279 static bool wq_power_efficient
;
282 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
284 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
286 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
287 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
289 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
290 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
292 static LIST_HEAD(workqueues
); /* PL: list of all workqueues */
293 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
295 /* the per-cpu worker pools */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
299 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
301 /* PL: hash of all unbound pools keyed by pool->attrs */
302 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
304 /* I: attributes used when instantiating standard unbound pools on demand */
305 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
307 /* I: attributes used when instantiating ordered pools on demand */
308 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
310 struct workqueue_struct
*system_wq __read_mostly
;
311 EXPORT_SYMBOL(system_wq
);
312 struct workqueue_struct
*system_highpri_wq __read_mostly
;
313 EXPORT_SYMBOL_GPL(system_highpri_wq
);
314 struct workqueue_struct
*system_long_wq __read_mostly
;
315 EXPORT_SYMBOL_GPL(system_long_wq
);
316 struct workqueue_struct
*system_unbound_wq __read_mostly
;
317 EXPORT_SYMBOL_GPL(system_unbound_wq
);
318 struct workqueue_struct
*system_freezable_wq __read_mostly
;
319 EXPORT_SYMBOL_GPL(system_freezable_wq
);
320 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
321 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
322 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
323 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
325 static int worker_thread(void *__worker
);
326 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
327 const struct workqueue_attrs
*from
);
329 #define CREATE_TRACE_POINTS
330 #include <trace/events/workqueue.h>
332 #define assert_rcu_or_pool_mutex() \
333 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
334 lockdep_is_held(&wq_pool_mutex), \
335 "sched RCU or wq_pool_mutex should be held")
337 #define assert_rcu_or_wq_mutex(wq) \
338 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
339 lockdep_is_held(&wq->mutex), \
340 "sched RCU or wq->mutex should be held")
342 #define for_each_cpu_worker_pool(pool, cpu) \
343 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
344 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
348 * for_each_pool - iterate through all worker_pools in the system
349 * @pool: iteration cursor
350 * @pi: integer used for iteration
352 * This must be called either with wq_pool_mutex held or sched RCU read
353 * locked. If the pool needs to be used beyond the locking in effect, the
354 * caller is responsible for guaranteeing that the pool stays online.
356 * The if/else clause exists only for the lockdep assertion and can be
359 #define for_each_pool(pool, pi) \
360 idr_for_each_entry(&worker_pool_idr, pool, pi) \
361 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
365 * for_each_pool_worker - iterate through all workers of a worker_pool
366 * @worker: iteration cursor
367 * @pool: worker_pool to iterate workers of
369 * This must be called with @pool->attach_mutex.
371 * The if/else clause exists only for the lockdep assertion and can be
374 #define for_each_pool_worker(worker, pool) \
375 list_for_each_entry((worker), &(pool)->workers, node) \
376 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
380 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
381 * @pwq: iteration cursor
382 * @wq: the target workqueue
384 * This must be called either with wq->mutex held or sched RCU read locked.
385 * If the pwq needs to be used beyond the locking in effect, the caller is
386 * responsible for guaranteeing that the pwq stays online.
388 * The if/else clause exists only for the lockdep assertion and can be
391 #define for_each_pwq(pwq, wq) \
392 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
393 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
396 #ifdef CONFIG_DEBUG_OBJECTS_WORK
398 static struct debug_obj_descr work_debug_descr
;
400 static void *work_debug_hint(void *addr
)
402 return ((struct work_struct
*) addr
)->func
;
406 * fixup_init is called when:
407 * - an active object is initialized
409 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
411 struct work_struct
*work
= addr
;
414 case ODEBUG_STATE_ACTIVE
:
415 cancel_work_sync(work
);
416 debug_object_init(work
, &work_debug_descr
);
424 * fixup_activate is called when:
425 * - an active object is activated
426 * - an unknown object is activated (might be a statically initialized object)
428 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
430 struct work_struct
*work
= addr
;
434 case ODEBUG_STATE_NOTAVAILABLE
:
436 * This is not really a fixup. The work struct was
437 * statically initialized. We just make sure that it
438 * is tracked in the object tracker.
440 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
441 debug_object_init(work
, &work_debug_descr
);
442 debug_object_activate(work
, &work_debug_descr
);
448 case ODEBUG_STATE_ACTIVE
:
457 * fixup_free is called when:
458 * - an active object is freed
460 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
462 struct work_struct
*work
= addr
;
465 case ODEBUG_STATE_ACTIVE
:
466 cancel_work_sync(work
);
467 debug_object_free(work
, &work_debug_descr
);
474 static struct debug_obj_descr work_debug_descr
= {
475 .name
= "work_struct",
476 .debug_hint
= work_debug_hint
,
477 .fixup_init
= work_fixup_init
,
478 .fixup_activate
= work_fixup_activate
,
479 .fixup_free
= work_fixup_free
,
482 static inline void debug_work_activate(struct work_struct
*work
)
484 debug_object_activate(work
, &work_debug_descr
);
487 static inline void debug_work_deactivate(struct work_struct
*work
)
489 debug_object_deactivate(work
, &work_debug_descr
);
492 void __init_work(struct work_struct
*work
, int onstack
)
495 debug_object_init_on_stack(work
, &work_debug_descr
);
497 debug_object_init(work
, &work_debug_descr
);
499 EXPORT_SYMBOL_GPL(__init_work
);
501 void destroy_work_on_stack(struct work_struct
*work
)
503 debug_object_free(work
, &work_debug_descr
);
505 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
507 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
509 destroy_timer_on_stack(&work
->timer
);
510 debug_object_free(&work
->work
, &work_debug_descr
);
512 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
515 static inline void debug_work_activate(struct work_struct
*work
) { }
516 static inline void debug_work_deactivate(struct work_struct
*work
) { }
520 * worker_pool_assign_id - allocate ID and assing it to @pool
521 * @pool: the pool pointer of interest
523 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
524 * successfully, -errno on failure.
526 static int worker_pool_assign_id(struct worker_pool
*pool
)
530 lockdep_assert_held(&wq_pool_mutex
);
532 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
542 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
543 * @wq: the target workqueue
546 * This must be called either with pwq_lock held or sched RCU read locked.
547 * If the pwq needs to be used beyond the locking in effect, the caller is
548 * responsible for guaranteeing that the pwq stays online.
550 * Return: The unbound pool_workqueue for @node.
552 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
555 assert_rcu_or_wq_mutex(wq
);
556 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
559 static unsigned int work_color_to_flags(int color
)
561 return color
<< WORK_STRUCT_COLOR_SHIFT
;
564 static int get_work_color(struct work_struct
*work
)
566 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
567 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
570 static int work_next_color(int color
)
572 return (color
+ 1) % WORK_NR_COLORS
;
576 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
577 * contain the pointer to the queued pwq. Once execution starts, the flag
578 * is cleared and the high bits contain OFFQ flags and pool ID.
580 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
581 * and clear_work_data() can be used to set the pwq, pool or clear
582 * work->data. These functions should only be called while the work is
583 * owned - ie. while the PENDING bit is set.
585 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
586 * corresponding to a work. Pool is available once the work has been
587 * queued anywhere after initialization until it is sync canceled. pwq is
588 * available only while the work item is queued.
590 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
591 * canceled. While being canceled, a work item may have its PENDING set
592 * but stay off timer and worklist for arbitrarily long and nobody should
593 * try to steal the PENDING bit.
595 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
598 WARN_ON_ONCE(!work_pending(work
));
599 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
602 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
603 unsigned long extra_flags
)
605 set_work_data(work
, (unsigned long)pwq
,
606 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
609 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
612 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
613 WORK_STRUCT_PENDING
);
616 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
620 * The following wmb is paired with the implied mb in
621 * test_and_set_bit(PENDING) and ensures all updates to @work made
622 * here are visible to and precede any updates by the next PENDING
626 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
629 static void clear_work_data(struct work_struct
*work
)
631 smp_wmb(); /* see set_work_pool_and_clear_pending() */
632 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
635 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
637 unsigned long data
= atomic_long_read(&work
->data
);
639 if (data
& WORK_STRUCT_PWQ
)
640 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
646 * get_work_pool - return the worker_pool a given work was associated with
647 * @work: the work item of interest
649 * Pools are created and destroyed under wq_pool_mutex, and allows read
650 * access under sched-RCU read lock. As such, this function should be
651 * called under wq_pool_mutex or with preemption disabled.
653 * All fields of the returned pool are accessible as long as the above
654 * mentioned locking is in effect. If the returned pool needs to be used
655 * beyond the critical section, the caller is responsible for ensuring the
656 * returned pool is and stays online.
658 * Return: The worker_pool @work was last associated with. %NULL if none.
660 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
662 unsigned long data
= atomic_long_read(&work
->data
);
665 assert_rcu_or_pool_mutex();
667 if (data
& WORK_STRUCT_PWQ
)
668 return ((struct pool_workqueue
*)
669 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
671 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
672 if (pool_id
== WORK_OFFQ_POOL_NONE
)
675 return idr_find(&worker_pool_idr
, pool_id
);
679 * get_work_pool_id - return the worker pool ID a given work is associated with
680 * @work: the work item of interest
682 * Return: The worker_pool ID @work was last associated with.
683 * %WORK_OFFQ_POOL_NONE if none.
685 static int get_work_pool_id(struct work_struct
*work
)
687 unsigned long data
= atomic_long_read(&work
->data
);
689 if (data
& WORK_STRUCT_PWQ
)
690 return ((struct pool_workqueue
*)
691 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
693 return data
>> WORK_OFFQ_POOL_SHIFT
;
696 static void mark_work_canceling(struct work_struct
*work
)
698 unsigned long pool_id
= get_work_pool_id(work
);
700 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
701 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
704 static bool work_is_canceling(struct work_struct
*work
)
706 unsigned long data
= atomic_long_read(&work
->data
);
708 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
712 * Policy functions. These define the policies on how the global worker
713 * pools are managed. Unless noted otherwise, these functions assume that
714 * they're being called with pool->lock held.
717 static bool __need_more_worker(struct worker_pool
*pool
)
719 return !atomic_read(&pool
->nr_running
);
723 * Need to wake up a worker? Called from anything but currently
726 * Note that, because unbound workers never contribute to nr_running, this
727 * function will always return %true for unbound pools as long as the
728 * worklist isn't empty.
730 static bool need_more_worker(struct worker_pool
*pool
)
732 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
735 /* Can I start working? Called from busy but !running workers. */
736 static bool may_start_working(struct worker_pool
*pool
)
738 return pool
->nr_idle
;
741 /* Do I need to keep working? Called from currently running workers. */
742 static bool keep_working(struct worker_pool
*pool
)
744 return !list_empty(&pool
->worklist
) &&
745 atomic_read(&pool
->nr_running
) <= 1;
748 /* Do we need a new worker? Called from manager. */
749 static bool need_to_create_worker(struct worker_pool
*pool
)
751 return need_more_worker(pool
) && !may_start_working(pool
);
754 /* Do we have too many workers and should some go away? */
755 static bool too_many_workers(struct worker_pool
*pool
)
757 bool managing
= mutex_is_locked(&pool
->manager_arb
);
758 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
759 int nr_busy
= pool
->nr_workers
- nr_idle
;
761 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
768 /* Return the first idle worker. Safe with preemption disabled */
769 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
771 if (unlikely(list_empty(&pool
->idle_list
)))
774 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
778 * wake_up_worker - wake up an idle worker
779 * @pool: worker pool to wake worker from
781 * Wake up the first idle worker of @pool.
784 * spin_lock_irq(pool->lock).
786 static void wake_up_worker(struct worker_pool
*pool
)
788 struct worker
*worker
= first_idle_worker(pool
);
791 wake_up_process(worker
->task
);
795 * wq_worker_waking_up - a worker is waking up
796 * @task: task waking up
797 * @cpu: CPU @task is waking up to
799 * This function is called during try_to_wake_up() when a worker is
803 * spin_lock_irq(rq->lock)
805 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
807 struct worker
*worker
= kthread_data(task
);
809 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
810 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
811 atomic_inc(&worker
->pool
->nr_running
);
816 * wq_worker_sleeping - a worker is going to sleep
817 * @task: task going to sleep
818 * @cpu: CPU in question, must be the current CPU number
820 * This function is called during schedule() when a busy worker is
821 * going to sleep. Worker on the same cpu can be woken up by
822 * returning pointer to its task.
825 * spin_lock_irq(rq->lock)
828 * Worker task on @cpu to wake up, %NULL if none.
830 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
832 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
833 struct worker_pool
*pool
;
836 * Rescuers, which may not have all the fields set up like normal
837 * workers, also reach here, let's not access anything before
838 * checking NOT_RUNNING.
840 if (worker
->flags
& WORKER_NOT_RUNNING
)
845 /* this can only happen on the local cpu */
846 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id() || pool
->cpu
!= cpu
))
850 * The counterpart of the following dec_and_test, implied mb,
851 * worklist not empty test sequence is in insert_work().
852 * Please read comment there.
854 * NOT_RUNNING is clear. This means that we're bound to and
855 * running on the local cpu w/ rq lock held and preemption
856 * disabled, which in turn means that none else could be
857 * manipulating idle_list, so dereferencing idle_list without pool
860 if (atomic_dec_and_test(&pool
->nr_running
) &&
861 !list_empty(&pool
->worklist
))
862 to_wakeup
= first_idle_worker(pool
);
863 return to_wakeup
? to_wakeup
->task
: NULL
;
867 * worker_set_flags - set worker flags and adjust nr_running accordingly
869 * @flags: flags to set
871 * Set @flags in @worker->flags and adjust nr_running accordingly.
874 * spin_lock_irq(pool->lock)
876 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
878 struct worker_pool
*pool
= worker
->pool
;
880 WARN_ON_ONCE(worker
->task
!= current
);
882 /* If transitioning into NOT_RUNNING, adjust nr_running. */
883 if ((flags
& WORKER_NOT_RUNNING
) &&
884 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
885 atomic_dec(&pool
->nr_running
);
888 worker
->flags
|= flags
;
892 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
894 * @flags: flags to clear
896 * Clear @flags in @worker->flags and adjust nr_running accordingly.
899 * spin_lock_irq(pool->lock)
901 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
903 struct worker_pool
*pool
= worker
->pool
;
904 unsigned int oflags
= worker
->flags
;
906 WARN_ON_ONCE(worker
->task
!= current
);
908 worker
->flags
&= ~flags
;
911 * If transitioning out of NOT_RUNNING, increment nr_running. Note
912 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
913 * of multiple flags, not a single flag.
915 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
916 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
917 atomic_inc(&pool
->nr_running
);
921 * find_worker_executing_work - find worker which is executing a work
922 * @pool: pool of interest
923 * @work: work to find worker for
925 * Find a worker which is executing @work on @pool by searching
926 * @pool->busy_hash which is keyed by the address of @work. For a worker
927 * to match, its current execution should match the address of @work and
928 * its work function. This is to avoid unwanted dependency between
929 * unrelated work executions through a work item being recycled while still
932 * This is a bit tricky. A work item may be freed once its execution
933 * starts and nothing prevents the freed area from being recycled for
934 * another work item. If the same work item address ends up being reused
935 * before the original execution finishes, workqueue will identify the
936 * recycled work item as currently executing and make it wait until the
937 * current execution finishes, introducing an unwanted dependency.
939 * This function checks the work item address and work function to avoid
940 * false positives. Note that this isn't complete as one may construct a
941 * work function which can introduce dependency onto itself through a
942 * recycled work item. Well, if somebody wants to shoot oneself in the
943 * foot that badly, there's only so much we can do, and if such deadlock
944 * actually occurs, it should be easy to locate the culprit work function.
947 * spin_lock_irq(pool->lock).
950 * Pointer to worker which is executing @work if found, %NULL
953 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
954 struct work_struct
*work
)
956 struct worker
*worker
;
958 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
960 if (worker
->current_work
== work
&&
961 worker
->current_func
== work
->func
)
968 * move_linked_works - move linked works to a list
969 * @work: start of series of works to be scheduled
970 * @head: target list to append @work to
971 * @nextp: out paramter for nested worklist walking
973 * Schedule linked works starting from @work to @head. Work series to
974 * be scheduled starts at @work and includes any consecutive work with
975 * WORK_STRUCT_LINKED set in its predecessor.
977 * If @nextp is not NULL, it's updated to point to the next work of
978 * the last scheduled work. This allows move_linked_works() to be
979 * nested inside outer list_for_each_entry_safe().
982 * spin_lock_irq(pool->lock).
984 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
985 struct work_struct
**nextp
)
987 struct work_struct
*n
;
990 * Linked worklist will always end before the end of the list,
991 * use NULL for list head.
993 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
994 list_move_tail(&work
->entry
, head
);
995 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1000 * If we're already inside safe list traversal and have moved
1001 * multiple works to the scheduled queue, the next position
1002 * needs to be updated.
1009 * get_pwq - get an extra reference on the specified pool_workqueue
1010 * @pwq: pool_workqueue to get
1012 * Obtain an extra reference on @pwq. The caller should guarantee that
1013 * @pwq has positive refcnt and be holding the matching pool->lock.
1015 static void get_pwq(struct pool_workqueue
*pwq
)
1017 lockdep_assert_held(&pwq
->pool
->lock
);
1018 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1023 * put_pwq - put a pool_workqueue reference
1024 * @pwq: pool_workqueue to put
1026 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1027 * destruction. The caller should be holding the matching pool->lock.
1029 static void put_pwq(struct pool_workqueue
*pwq
)
1031 lockdep_assert_held(&pwq
->pool
->lock
);
1032 if (likely(--pwq
->refcnt
))
1034 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1037 * @pwq can't be released under pool->lock, bounce to
1038 * pwq_unbound_release_workfn(). This never recurses on the same
1039 * pool->lock as this path is taken only for unbound workqueues and
1040 * the release work item is scheduled on a per-cpu workqueue. To
1041 * avoid lockdep warning, unbound pool->locks are given lockdep
1042 * subclass of 1 in get_unbound_pool().
1044 schedule_work(&pwq
->unbound_release_work
);
1048 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1049 * @pwq: pool_workqueue to put (can be %NULL)
1051 * put_pwq() with locking. This function also allows %NULL @pwq.
1053 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1057 * As both pwqs and pools are sched-RCU protected, the
1058 * following lock operations are safe.
1060 spin_lock_irq(&pwq
->pool
->lock
);
1062 spin_unlock_irq(&pwq
->pool
->lock
);
1066 static void pwq_activate_delayed_work(struct work_struct
*work
)
1068 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1070 trace_workqueue_activate_work(work
);
1071 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1072 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1076 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1078 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1079 struct work_struct
, entry
);
1081 pwq_activate_delayed_work(work
);
1085 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1086 * @pwq: pwq of interest
1087 * @color: color of work which left the queue
1089 * A work either has completed or is removed from pending queue,
1090 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1093 * spin_lock_irq(pool->lock).
1095 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1097 /* uncolored work items don't participate in flushing or nr_active */
1098 if (color
== WORK_NO_COLOR
)
1101 pwq
->nr_in_flight
[color
]--;
1104 if (!list_empty(&pwq
->delayed_works
)) {
1105 /* one down, submit a delayed one */
1106 if (pwq
->nr_active
< pwq
->max_active
)
1107 pwq_activate_first_delayed(pwq
);
1110 /* is flush in progress and are we at the flushing tip? */
1111 if (likely(pwq
->flush_color
!= color
))
1114 /* are there still in-flight works? */
1115 if (pwq
->nr_in_flight
[color
])
1118 /* this pwq is done, clear flush_color */
1119 pwq
->flush_color
= -1;
1122 * If this was the last pwq, wake up the first flusher. It
1123 * will handle the rest.
1125 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1126 complete(&pwq
->wq
->first_flusher
->done
);
1132 * try_to_grab_pending - steal work item from worklist and disable irq
1133 * @work: work item to steal
1134 * @is_dwork: @work is a delayed_work
1135 * @flags: place to store irq state
1137 * Try to grab PENDING bit of @work. This function can handle @work in any
1138 * stable state - idle, on timer or on worklist.
1141 * 1 if @work was pending and we successfully stole PENDING
1142 * 0 if @work was idle and we claimed PENDING
1143 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1144 * -ENOENT if someone else is canceling @work, this state may persist
1145 * for arbitrarily long
1148 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1149 * interrupted while holding PENDING and @work off queue, irq must be
1150 * disabled on entry. This, combined with delayed_work->timer being
1151 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1153 * On successful return, >= 0, irq is disabled and the caller is
1154 * responsible for releasing it using local_irq_restore(*@flags).
1156 * This function is safe to call from any context including IRQ handler.
1158 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1159 unsigned long *flags
)
1161 struct worker_pool
*pool
;
1162 struct pool_workqueue
*pwq
;
1164 local_irq_save(*flags
);
1166 /* try to steal the timer if it exists */
1168 struct delayed_work
*dwork
= to_delayed_work(work
);
1171 * dwork->timer is irqsafe. If del_timer() fails, it's
1172 * guaranteed that the timer is not queued anywhere and not
1173 * running on the local CPU.
1175 if (likely(del_timer(&dwork
->timer
)))
1179 /* try to claim PENDING the normal way */
1180 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1184 * The queueing is in progress, or it is already queued. Try to
1185 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1187 pool
= get_work_pool(work
);
1191 spin_lock(&pool
->lock
);
1193 * work->data is guaranteed to point to pwq only while the work
1194 * item is queued on pwq->wq, and both updating work->data to point
1195 * to pwq on queueing and to pool on dequeueing are done under
1196 * pwq->pool->lock. This in turn guarantees that, if work->data
1197 * points to pwq which is associated with a locked pool, the work
1198 * item is currently queued on that pool.
1200 pwq
= get_work_pwq(work
);
1201 if (pwq
&& pwq
->pool
== pool
) {
1202 debug_work_deactivate(work
);
1205 * A delayed work item cannot be grabbed directly because
1206 * it might have linked NO_COLOR work items which, if left
1207 * on the delayed_list, will confuse pwq->nr_active
1208 * management later on and cause stall. Make sure the work
1209 * item is activated before grabbing.
1211 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1212 pwq_activate_delayed_work(work
);
1214 list_del_init(&work
->entry
);
1215 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1217 /* work->data points to pwq iff queued, point to pool */
1218 set_work_pool_and_keep_pending(work
, pool
->id
);
1220 spin_unlock(&pool
->lock
);
1223 spin_unlock(&pool
->lock
);
1225 local_irq_restore(*flags
);
1226 if (work_is_canceling(work
))
1233 * insert_work - insert a work into a pool
1234 * @pwq: pwq @work belongs to
1235 * @work: work to insert
1236 * @head: insertion point
1237 * @extra_flags: extra WORK_STRUCT_* flags to set
1239 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1240 * work_struct flags.
1243 * spin_lock_irq(pool->lock).
1245 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1246 struct list_head
*head
, unsigned int extra_flags
)
1248 struct worker_pool
*pool
= pwq
->pool
;
1250 /* we own @work, set data and link */
1251 set_work_pwq(work
, pwq
, extra_flags
);
1252 list_add_tail(&work
->entry
, head
);
1256 * Ensure either wq_worker_sleeping() sees the above
1257 * list_add_tail() or we see zero nr_running to avoid workers lying
1258 * around lazily while there are works to be processed.
1262 if (__need_more_worker(pool
))
1263 wake_up_worker(pool
);
1267 * Test whether @work is being queued from another work executing on the
1270 static bool is_chained_work(struct workqueue_struct
*wq
)
1272 struct worker
*worker
;
1274 worker
= current_wq_worker();
1276 * Return %true iff I'm a worker execuing a work item on @wq. If
1277 * I'm @worker, it's safe to dereference it without locking.
1279 return worker
&& worker
->current_pwq
->wq
== wq
;
1282 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1283 struct work_struct
*work
)
1285 struct pool_workqueue
*pwq
;
1286 struct worker_pool
*last_pool
;
1287 struct list_head
*worklist
;
1288 unsigned int work_flags
;
1289 unsigned int req_cpu
= cpu
;
1292 * While a work item is PENDING && off queue, a task trying to
1293 * steal the PENDING will busy-loop waiting for it to either get
1294 * queued or lose PENDING. Grabbing PENDING and queueing should
1295 * happen with IRQ disabled.
1297 WARN_ON_ONCE(!irqs_disabled());
1299 debug_work_activate(work
);
1301 /* if draining, only works from the same workqueue are allowed */
1302 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1303 WARN_ON_ONCE(!is_chained_work(wq
)))
1306 if (req_cpu
== WORK_CPU_UNBOUND
)
1307 cpu
= raw_smp_processor_id();
1309 /* pwq which will be used unless @work is executing elsewhere */
1310 if (!(wq
->flags
& WQ_UNBOUND
))
1311 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1313 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1316 * If @work was previously on a different pool, it might still be
1317 * running there, in which case the work needs to be queued on that
1318 * pool to guarantee non-reentrancy.
1320 last_pool
= get_work_pool(work
);
1321 if (last_pool
&& last_pool
!= pwq
->pool
) {
1322 struct worker
*worker
;
1324 spin_lock(&last_pool
->lock
);
1326 worker
= find_worker_executing_work(last_pool
, work
);
1328 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1329 pwq
= worker
->current_pwq
;
1331 /* meh... not running there, queue here */
1332 spin_unlock(&last_pool
->lock
);
1333 spin_lock(&pwq
->pool
->lock
);
1336 spin_lock(&pwq
->pool
->lock
);
1340 * pwq is determined and locked. For unbound pools, we could have
1341 * raced with pwq release and it could already be dead. If its
1342 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1343 * without another pwq replacing it in the numa_pwq_tbl or while
1344 * work items are executing on it, so the retrying is guaranteed to
1345 * make forward-progress.
1347 if (unlikely(!pwq
->refcnt
)) {
1348 if (wq
->flags
& WQ_UNBOUND
) {
1349 spin_unlock(&pwq
->pool
->lock
);
1354 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1358 /* pwq determined, queue */
1359 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1361 if (WARN_ON(!list_empty(&work
->entry
))) {
1362 spin_unlock(&pwq
->pool
->lock
);
1366 pwq
->nr_in_flight
[pwq
->work_color
]++;
1367 work_flags
= work_color_to_flags(pwq
->work_color
);
1369 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1370 trace_workqueue_activate_work(work
);
1372 worklist
= &pwq
->pool
->worklist
;
1374 work_flags
|= WORK_STRUCT_DELAYED
;
1375 worklist
= &pwq
->delayed_works
;
1378 insert_work(pwq
, work
, worklist
, work_flags
);
1380 spin_unlock(&pwq
->pool
->lock
);
1384 * queue_work_on - queue work on specific cpu
1385 * @cpu: CPU number to execute work on
1386 * @wq: workqueue to use
1387 * @work: work to queue
1389 * We queue the work to a specific CPU, the caller must ensure it
1392 * Return: %false if @work was already on a queue, %true otherwise.
1394 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1395 struct work_struct
*work
)
1398 unsigned long flags
;
1400 local_irq_save(flags
);
1402 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1403 __queue_work(cpu
, wq
, work
);
1407 local_irq_restore(flags
);
1410 EXPORT_SYMBOL(queue_work_on
);
1412 void delayed_work_timer_fn(unsigned long __data
)
1414 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1416 /* should have been called from irqsafe timer with irq already off */
1417 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1419 EXPORT_SYMBOL(delayed_work_timer_fn
);
1421 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1422 struct delayed_work
*dwork
, unsigned long delay
)
1424 struct timer_list
*timer
= &dwork
->timer
;
1425 struct work_struct
*work
= &dwork
->work
;
1427 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1428 timer
->data
!= (unsigned long)dwork
);
1429 WARN_ON_ONCE(timer_pending(timer
));
1430 WARN_ON_ONCE(!list_empty(&work
->entry
));
1433 * If @delay is 0, queue @dwork->work immediately. This is for
1434 * both optimization and correctness. The earliest @timer can
1435 * expire is on the closest next tick and delayed_work users depend
1436 * on that there's no such delay when @delay is 0.
1439 __queue_work(cpu
, wq
, &dwork
->work
);
1443 timer_stats_timer_set_start_info(&dwork
->timer
);
1447 timer
->expires
= jiffies
+ delay
;
1449 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1450 add_timer_on(timer
, cpu
);
1456 * queue_delayed_work_on - queue work on specific CPU after delay
1457 * @cpu: CPU number to execute work on
1458 * @wq: workqueue to use
1459 * @dwork: work to queue
1460 * @delay: number of jiffies to wait before queueing
1462 * Return: %false if @work was already on a queue, %true otherwise. If
1463 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1466 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1467 struct delayed_work
*dwork
, unsigned long delay
)
1469 struct work_struct
*work
= &dwork
->work
;
1471 unsigned long flags
;
1473 /* read the comment in __queue_work() */
1474 local_irq_save(flags
);
1476 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1477 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1481 local_irq_restore(flags
);
1484 EXPORT_SYMBOL(queue_delayed_work_on
);
1487 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1488 * @cpu: CPU number to execute work on
1489 * @wq: workqueue to use
1490 * @dwork: work to queue
1491 * @delay: number of jiffies to wait before queueing
1493 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1494 * modify @dwork's timer so that it expires after @delay. If @delay is
1495 * zero, @work is guaranteed to be scheduled immediately regardless of its
1498 * Return: %false if @dwork was idle and queued, %true if @dwork was
1499 * pending and its timer was modified.
1501 * This function is safe to call from any context including IRQ handler.
1502 * See try_to_grab_pending() for details.
1504 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1505 struct delayed_work
*dwork
, unsigned long delay
)
1507 unsigned long flags
;
1511 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1512 } while (unlikely(ret
== -EAGAIN
));
1514 if (likely(ret
>= 0)) {
1515 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1516 local_irq_restore(flags
);
1519 /* -ENOENT from try_to_grab_pending() becomes %true */
1522 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1525 * worker_enter_idle - enter idle state
1526 * @worker: worker which is entering idle state
1528 * @worker is entering idle state. Update stats and idle timer if
1532 * spin_lock_irq(pool->lock).
1534 static void worker_enter_idle(struct worker
*worker
)
1536 struct worker_pool
*pool
= worker
->pool
;
1538 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1539 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1540 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1543 /* can't use worker_set_flags(), also called from create_worker() */
1544 worker
->flags
|= WORKER_IDLE
;
1546 worker
->last_active
= jiffies
;
1548 /* idle_list is LIFO */
1549 list_add(&worker
->entry
, &pool
->idle_list
);
1551 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1552 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1555 * Sanity check nr_running. Because wq_unbind_fn() releases
1556 * pool->lock between setting %WORKER_UNBOUND and zapping
1557 * nr_running, the warning may trigger spuriously. Check iff
1558 * unbind is not in progress.
1560 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1561 pool
->nr_workers
== pool
->nr_idle
&&
1562 atomic_read(&pool
->nr_running
));
1566 * worker_leave_idle - leave idle state
1567 * @worker: worker which is leaving idle state
1569 * @worker is leaving idle state. Update stats.
1572 * spin_lock_irq(pool->lock).
1574 static void worker_leave_idle(struct worker
*worker
)
1576 struct worker_pool
*pool
= worker
->pool
;
1578 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1580 worker_clr_flags(worker
, WORKER_IDLE
);
1582 list_del_init(&worker
->entry
);
1585 static struct worker
*alloc_worker(int node
)
1587 struct worker
*worker
;
1589 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1591 INIT_LIST_HEAD(&worker
->entry
);
1592 INIT_LIST_HEAD(&worker
->scheduled
);
1593 INIT_LIST_HEAD(&worker
->node
);
1594 /* on creation a worker is in !idle && prep state */
1595 worker
->flags
= WORKER_PREP
;
1601 * worker_attach_to_pool() - attach a worker to a pool
1602 * @worker: worker to be attached
1603 * @pool: the target pool
1605 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1606 * cpu-binding of @worker are kept coordinated with the pool across
1609 static void worker_attach_to_pool(struct worker
*worker
,
1610 struct worker_pool
*pool
)
1612 mutex_lock(&pool
->attach_mutex
);
1615 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1616 * online CPUs. It'll be re-applied when any of the CPUs come up.
1618 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1621 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1622 * stable across this function. See the comments above the
1623 * flag definition for details.
1625 if (pool
->flags
& POOL_DISASSOCIATED
)
1626 worker
->flags
|= WORKER_UNBOUND
;
1628 list_add_tail(&worker
->node
, &pool
->workers
);
1630 mutex_unlock(&pool
->attach_mutex
);
1634 * worker_detach_from_pool() - detach a worker from its pool
1635 * @worker: worker which is attached to its pool
1636 * @pool: the pool @worker is attached to
1638 * Undo the attaching which had been done in worker_attach_to_pool(). The
1639 * caller worker shouldn't access to the pool after detached except it has
1640 * other reference to the pool.
1642 static void worker_detach_from_pool(struct worker
*worker
,
1643 struct worker_pool
*pool
)
1645 struct completion
*detach_completion
= NULL
;
1647 mutex_lock(&pool
->attach_mutex
);
1648 list_del(&worker
->node
);
1649 if (list_empty(&pool
->workers
))
1650 detach_completion
= pool
->detach_completion
;
1651 mutex_unlock(&pool
->attach_mutex
);
1653 /* clear leftover flags without pool->lock after it is detached */
1654 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1656 if (detach_completion
)
1657 complete(detach_completion
);
1661 * create_worker - create a new workqueue worker
1662 * @pool: pool the new worker will belong to
1664 * Create and start a new worker which is attached to @pool.
1667 * Might sleep. Does GFP_KERNEL allocations.
1670 * Pointer to the newly created worker.
1672 static struct worker
*create_worker(struct worker_pool
*pool
)
1674 struct worker
*worker
= NULL
;
1678 /* ID is needed to determine kthread name */
1679 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1683 worker
= alloc_worker(pool
->node
);
1687 worker
->pool
= pool
;
1691 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1692 pool
->attrs
->nice
< 0 ? "H" : "");
1694 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1696 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1697 "kworker/%s", id_buf
);
1698 if (IS_ERR(worker
->task
))
1701 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1703 /* prevent userland from meddling with cpumask of workqueue workers */
1704 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1706 /* successful, attach the worker to the pool */
1707 worker_attach_to_pool(worker
, pool
);
1709 /* start the newly created worker */
1710 spin_lock_irq(&pool
->lock
);
1711 worker
->pool
->nr_workers
++;
1712 worker_enter_idle(worker
);
1713 wake_up_process(worker
->task
);
1714 spin_unlock_irq(&pool
->lock
);
1720 ida_simple_remove(&pool
->worker_ida
, id
);
1726 * destroy_worker - destroy a workqueue worker
1727 * @worker: worker to be destroyed
1729 * Destroy @worker and adjust @pool stats accordingly. The worker should
1733 * spin_lock_irq(pool->lock).
1735 static void destroy_worker(struct worker
*worker
)
1737 struct worker_pool
*pool
= worker
->pool
;
1739 lockdep_assert_held(&pool
->lock
);
1741 /* sanity check frenzy */
1742 if (WARN_ON(worker
->current_work
) ||
1743 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1744 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1750 list_del_init(&worker
->entry
);
1751 worker
->flags
|= WORKER_DIE
;
1752 wake_up_process(worker
->task
);
1755 static void idle_worker_timeout(unsigned long __pool
)
1757 struct worker_pool
*pool
= (void *)__pool
;
1759 spin_lock_irq(&pool
->lock
);
1761 while (too_many_workers(pool
)) {
1762 struct worker
*worker
;
1763 unsigned long expires
;
1765 /* idle_list is kept in LIFO order, check the last one */
1766 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1767 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1769 if (time_before(jiffies
, expires
)) {
1770 mod_timer(&pool
->idle_timer
, expires
);
1774 destroy_worker(worker
);
1777 spin_unlock_irq(&pool
->lock
);
1780 static void send_mayday(struct work_struct
*work
)
1782 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1783 struct workqueue_struct
*wq
= pwq
->wq
;
1785 lockdep_assert_held(&wq_mayday_lock
);
1790 /* mayday mayday mayday */
1791 if (list_empty(&pwq
->mayday_node
)) {
1793 * If @pwq is for an unbound wq, its base ref may be put at
1794 * any time due to an attribute change. Pin @pwq until the
1795 * rescuer is done with it.
1798 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1799 wake_up_process(wq
->rescuer
->task
);
1803 static void pool_mayday_timeout(unsigned long __pool
)
1805 struct worker_pool
*pool
= (void *)__pool
;
1806 struct work_struct
*work
;
1808 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1809 spin_lock(&pool
->lock
);
1811 if (need_to_create_worker(pool
)) {
1813 * We've been trying to create a new worker but
1814 * haven't been successful. We might be hitting an
1815 * allocation deadlock. Send distress signals to
1818 list_for_each_entry(work
, &pool
->worklist
, entry
)
1822 spin_unlock(&pool
->lock
);
1823 spin_unlock_irq(&wq_mayday_lock
);
1825 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1829 * maybe_create_worker - create a new worker if necessary
1830 * @pool: pool to create a new worker for
1832 * Create a new worker for @pool if necessary. @pool is guaranteed to
1833 * have at least one idle worker on return from this function. If
1834 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1835 * sent to all rescuers with works scheduled on @pool to resolve
1836 * possible allocation deadlock.
1838 * On return, need_to_create_worker() is guaranteed to be %false and
1839 * may_start_working() %true.
1842 * spin_lock_irq(pool->lock) which may be released and regrabbed
1843 * multiple times. Does GFP_KERNEL allocations. Called only from
1847 * %false if no action was taken and pool->lock stayed locked, %true
1850 static bool maybe_create_worker(struct worker_pool
*pool
)
1851 __releases(&pool
->lock
)
1852 __acquires(&pool
->lock
)
1854 if (!need_to_create_worker(pool
))
1857 spin_unlock_irq(&pool
->lock
);
1859 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1860 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1863 if (create_worker(pool
) || !need_to_create_worker(pool
))
1866 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1868 if (!need_to_create_worker(pool
))
1872 del_timer_sync(&pool
->mayday_timer
);
1873 spin_lock_irq(&pool
->lock
);
1875 * This is necessary even after a new worker was just successfully
1876 * created as @pool->lock was dropped and the new worker might have
1877 * already become busy.
1879 if (need_to_create_worker(pool
))
1885 * manage_workers - manage worker pool
1888 * Assume the manager role and manage the worker pool @worker belongs
1889 * to. At any given time, there can be only zero or one manager per
1890 * pool. The exclusion is handled automatically by this function.
1892 * The caller can safely start processing works on false return. On
1893 * true return, it's guaranteed that need_to_create_worker() is false
1894 * and may_start_working() is true.
1897 * spin_lock_irq(pool->lock) which may be released and regrabbed
1898 * multiple times. Does GFP_KERNEL allocations.
1901 * %false if the pool don't need management and the caller can safely start
1902 * processing works, %true indicates that the function released pool->lock
1903 * and reacquired it to perform some management function and that the
1904 * conditions that the caller verified while holding the lock before
1905 * calling the function might no longer be true.
1907 static bool manage_workers(struct worker
*worker
)
1909 struct worker_pool
*pool
= worker
->pool
;
1913 * Anyone who successfully grabs manager_arb wins the arbitration
1914 * and becomes the manager. mutex_trylock() on pool->manager_arb
1915 * failure while holding pool->lock reliably indicates that someone
1916 * else is managing the pool and the worker which failed trylock
1917 * can proceed to executing work items. This means that anyone
1918 * grabbing manager_arb is responsible for actually performing
1919 * manager duties. If manager_arb is grabbed and released without
1920 * actual management, the pool may stall indefinitely.
1922 if (!mutex_trylock(&pool
->manager_arb
))
1925 ret
|= maybe_create_worker(pool
);
1927 mutex_unlock(&pool
->manager_arb
);
1932 * process_one_work - process single work
1934 * @work: work to process
1936 * Process @work. This function contains all the logics necessary to
1937 * process a single work including synchronization against and
1938 * interaction with other workers on the same cpu, queueing and
1939 * flushing. As long as context requirement is met, any worker can
1940 * call this function to process a work.
1943 * spin_lock_irq(pool->lock) which is released and regrabbed.
1945 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1946 __releases(&pool
->lock
)
1947 __acquires(&pool
->lock
)
1949 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1950 struct worker_pool
*pool
= worker
->pool
;
1951 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1953 struct worker
*collision
;
1954 #ifdef CONFIG_LOCKDEP
1956 * It is permissible to free the struct work_struct from
1957 * inside the function that is called from it, this we need to
1958 * take into account for lockdep too. To avoid bogus "held
1959 * lock freed" warnings as well as problems when looking into
1960 * work->lockdep_map, make a copy and use that here.
1962 struct lockdep_map lockdep_map
;
1964 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
1966 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1967 raw_smp_processor_id() != pool
->cpu
);
1970 * A single work shouldn't be executed concurrently by
1971 * multiple workers on a single cpu. Check whether anyone is
1972 * already processing the work. If so, defer the work to the
1973 * currently executing one.
1975 collision
= find_worker_executing_work(pool
, work
);
1976 if (unlikely(collision
)) {
1977 move_linked_works(work
, &collision
->scheduled
, NULL
);
1981 /* claim and dequeue */
1982 debug_work_deactivate(work
);
1983 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
1984 worker
->current_work
= work
;
1985 worker
->current_func
= work
->func
;
1986 worker
->current_pwq
= pwq
;
1987 work_color
= get_work_color(work
);
1989 list_del_init(&work
->entry
);
1992 * CPU intensive works don't participate in concurrency management.
1993 * They're the scheduler's responsibility. This takes @worker out
1994 * of concurrency management and the next code block will chain
1995 * execution of the pending work items.
1997 if (unlikely(cpu_intensive
))
1998 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2001 * Wake up another worker if necessary. The condition is always
2002 * false for normal per-cpu workers since nr_running would always
2003 * be >= 1 at this point. This is used to chain execution of the
2004 * pending work items for WORKER_NOT_RUNNING workers such as the
2005 * UNBOUND and CPU_INTENSIVE ones.
2007 if (need_more_worker(pool
))
2008 wake_up_worker(pool
);
2011 * Record the last pool and clear PENDING which should be the last
2012 * update to @work. Also, do this inside @pool->lock so that
2013 * PENDING and queued state changes happen together while IRQ is
2016 set_work_pool_and_clear_pending(work
, pool
->id
);
2018 spin_unlock_irq(&pool
->lock
);
2020 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2021 lock_map_acquire(&lockdep_map
);
2022 trace_workqueue_execute_start(work
);
2023 worker
->current_func(work
);
2025 * While we must be careful to not use "work" after this, the trace
2026 * point will only record its address.
2028 trace_workqueue_execute_end(work
);
2029 lock_map_release(&lockdep_map
);
2030 lock_map_release(&pwq
->wq
->lockdep_map
);
2032 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2033 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2034 " last function: %pf\n",
2035 current
->comm
, preempt_count(), task_pid_nr(current
),
2036 worker
->current_func
);
2037 debug_show_held_locks(current
);
2042 * The following prevents a kworker from hogging CPU on !PREEMPT
2043 * kernels, where a requeueing work item waiting for something to
2044 * happen could deadlock with stop_machine as such work item could
2045 * indefinitely requeue itself while all other CPUs are trapped in
2050 spin_lock_irq(&pool
->lock
);
2052 /* clear cpu intensive status */
2053 if (unlikely(cpu_intensive
))
2054 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2056 /* we're done with it, release */
2057 hash_del(&worker
->hentry
);
2058 worker
->current_work
= NULL
;
2059 worker
->current_func
= NULL
;
2060 worker
->current_pwq
= NULL
;
2061 worker
->desc_valid
= false;
2062 pwq_dec_nr_in_flight(pwq
, work_color
);
2066 * process_scheduled_works - process scheduled works
2069 * Process all scheduled works. Please note that the scheduled list
2070 * may change while processing a work, so this function repeatedly
2071 * fetches a work from the top and executes it.
2074 * spin_lock_irq(pool->lock) which may be released and regrabbed
2077 static void process_scheduled_works(struct worker
*worker
)
2079 while (!list_empty(&worker
->scheduled
)) {
2080 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2081 struct work_struct
, entry
);
2082 process_one_work(worker
, work
);
2087 * worker_thread - the worker thread function
2090 * The worker thread function. All workers belong to a worker_pool -
2091 * either a per-cpu one or dynamic unbound one. These workers process all
2092 * work items regardless of their specific target workqueue. The only
2093 * exception is work items which belong to workqueues with a rescuer which
2094 * will be explained in rescuer_thread().
2098 static int worker_thread(void *__worker
)
2100 struct worker
*worker
= __worker
;
2101 struct worker_pool
*pool
= worker
->pool
;
2103 /* tell the scheduler that this is a workqueue worker */
2104 worker
->task
->flags
|= PF_WQ_WORKER
;
2106 spin_lock_irq(&pool
->lock
);
2108 /* am I supposed to die? */
2109 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2110 spin_unlock_irq(&pool
->lock
);
2111 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2112 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2114 set_task_comm(worker
->task
, "kworker/dying");
2115 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2116 worker_detach_from_pool(worker
, pool
);
2121 worker_leave_idle(worker
);
2123 /* no more worker necessary? */
2124 if (!need_more_worker(pool
))
2127 /* do we need to manage? */
2128 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2132 * ->scheduled list can only be filled while a worker is
2133 * preparing to process a work or actually processing it.
2134 * Make sure nobody diddled with it while I was sleeping.
2136 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2139 * Finish PREP stage. We're guaranteed to have at least one idle
2140 * worker or that someone else has already assumed the manager
2141 * role. This is where @worker starts participating in concurrency
2142 * management if applicable and concurrency management is restored
2143 * after being rebound. See rebind_workers() for details.
2145 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2148 struct work_struct
*work
=
2149 list_first_entry(&pool
->worklist
,
2150 struct work_struct
, entry
);
2152 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2153 /* optimization path, not strictly necessary */
2154 process_one_work(worker
, work
);
2155 if (unlikely(!list_empty(&worker
->scheduled
)))
2156 process_scheduled_works(worker
);
2158 move_linked_works(work
, &worker
->scheduled
, NULL
);
2159 process_scheduled_works(worker
);
2161 } while (keep_working(pool
));
2163 worker_set_flags(worker
, WORKER_PREP
);
2166 * pool->lock is held and there's no work to process and no need to
2167 * manage, sleep. Workers are woken up only while holding
2168 * pool->lock or from local cpu, so setting the current state
2169 * before releasing pool->lock is enough to prevent losing any
2172 worker_enter_idle(worker
);
2173 __set_current_state(TASK_INTERRUPTIBLE
);
2174 spin_unlock_irq(&pool
->lock
);
2180 * rescuer_thread - the rescuer thread function
2183 * Workqueue rescuer thread function. There's one rescuer for each
2184 * workqueue which has WQ_MEM_RECLAIM set.
2186 * Regular work processing on a pool may block trying to create a new
2187 * worker which uses GFP_KERNEL allocation which has slight chance of
2188 * developing into deadlock if some works currently on the same queue
2189 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2190 * the problem rescuer solves.
2192 * When such condition is possible, the pool summons rescuers of all
2193 * workqueues which have works queued on the pool and let them process
2194 * those works so that forward progress can be guaranteed.
2196 * This should happen rarely.
2200 static int rescuer_thread(void *__rescuer
)
2202 struct worker
*rescuer
= __rescuer
;
2203 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2204 struct list_head
*scheduled
= &rescuer
->scheduled
;
2207 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2210 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2211 * doesn't participate in concurrency management.
2213 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2215 set_current_state(TASK_INTERRUPTIBLE
);
2218 * By the time the rescuer is requested to stop, the workqueue
2219 * shouldn't have any work pending, but @wq->maydays may still have
2220 * pwq(s) queued. This can happen by non-rescuer workers consuming
2221 * all the work items before the rescuer got to them. Go through
2222 * @wq->maydays processing before acting on should_stop so that the
2223 * list is always empty on exit.
2225 should_stop
= kthread_should_stop();
2227 /* see whether any pwq is asking for help */
2228 spin_lock_irq(&wq_mayday_lock
);
2230 while (!list_empty(&wq
->maydays
)) {
2231 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2232 struct pool_workqueue
, mayday_node
);
2233 struct worker_pool
*pool
= pwq
->pool
;
2234 struct work_struct
*work
, *n
;
2236 __set_current_state(TASK_RUNNING
);
2237 list_del_init(&pwq
->mayday_node
);
2239 spin_unlock_irq(&wq_mayday_lock
);
2241 worker_attach_to_pool(rescuer
, pool
);
2243 spin_lock_irq(&pool
->lock
);
2244 rescuer
->pool
= pool
;
2247 * Slurp in all works issued via this workqueue and
2250 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2251 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2252 if (get_work_pwq(work
) == pwq
)
2253 move_linked_works(work
, scheduled
, &n
);
2255 process_scheduled_works(rescuer
);
2258 * Put the reference grabbed by send_mayday(). @pool won't
2259 * go away while we're still attached to it.
2264 * Leave this pool. If need_more_worker() is %true, notify a
2265 * regular worker; otherwise, we end up with 0 concurrency
2266 * and stalling the execution.
2268 if (need_more_worker(pool
))
2269 wake_up_worker(pool
);
2271 rescuer
->pool
= NULL
;
2272 spin_unlock_irq(&pool
->lock
);
2274 worker_detach_from_pool(rescuer
, pool
);
2276 spin_lock_irq(&wq_mayday_lock
);
2279 spin_unlock_irq(&wq_mayday_lock
);
2282 __set_current_state(TASK_RUNNING
);
2283 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2287 /* rescuers should never participate in concurrency management */
2288 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2294 struct work_struct work
;
2295 struct completion done
;
2298 static void wq_barrier_func(struct work_struct
*work
)
2300 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2301 complete(&barr
->done
);
2305 * insert_wq_barrier - insert a barrier work
2306 * @pwq: pwq to insert barrier into
2307 * @barr: wq_barrier to insert
2308 * @target: target work to attach @barr to
2309 * @worker: worker currently executing @target, NULL if @target is not executing
2311 * @barr is linked to @target such that @barr is completed only after
2312 * @target finishes execution. Please note that the ordering
2313 * guarantee is observed only with respect to @target and on the local
2316 * Currently, a queued barrier can't be canceled. This is because
2317 * try_to_grab_pending() can't determine whether the work to be
2318 * grabbed is at the head of the queue and thus can't clear LINKED
2319 * flag of the previous work while there must be a valid next work
2320 * after a work with LINKED flag set.
2322 * Note that when @worker is non-NULL, @target may be modified
2323 * underneath us, so we can't reliably determine pwq from @target.
2326 * spin_lock_irq(pool->lock).
2328 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2329 struct wq_barrier
*barr
,
2330 struct work_struct
*target
, struct worker
*worker
)
2332 struct list_head
*head
;
2333 unsigned int linked
= 0;
2336 * debugobject calls are safe here even with pool->lock locked
2337 * as we know for sure that this will not trigger any of the
2338 * checks and call back into the fixup functions where we
2341 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2342 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2343 init_completion(&barr
->done
);
2346 * If @target is currently being executed, schedule the
2347 * barrier to the worker; otherwise, put it after @target.
2350 head
= worker
->scheduled
.next
;
2352 unsigned long *bits
= work_data_bits(target
);
2354 head
= target
->entry
.next
;
2355 /* there can already be other linked works, inherit and set */
2356 linked
= *bits
& WORK_STRUCT_LINKED
;
2357 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2360 debug_work_activate(&barr
->work
);
2361 insert_work(pwq
, &barr
->work
, head
,
2362 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2366 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2367 * @wq: workqueue being flushed
2368 * @flush_color: new flush color, < 0 for no-op
2369 * @work_color: new work color, < 0 for no-op
2371 * Prepare pwqs for workqueue flushing.
2373 * If @flush_color is non-negative, flush_color on all pwqs should be
2374 * -1. If no pwq has in-flight commands at the specified color, all
2375 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2376 * has in flight commands, its pwq->flush_color is set to
2377 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2378 * wakeup logic is armed and %true is returned.
2380 * The caller should have initialized @wq->first_flusher prior to
2381 * calling this function with non-negative @flush_color. If
2382 * @flush_color is negative, no flush color update is done and %false
2385 * If @work_color is non-negative, all pwqs should have the same
2386 * work_color which is previous to @work_color and all will be
2387 * advanced to @work_color.
2390 * mutex_lock(wq->mutex).
2393 * %true if @flush_color >= 0 and there's something to flush. %false
2396 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2397 int flush_color
, int work_color
)
2400 struct pool_workqueue
*pwq
;
2402 if (flush_color
>= 0) {
2403 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2404 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2407 for_each_pwq(pwq
, wq
) {
2408 struct worker_pool
*pool
= pwq
->pool
;
2410 spin_lock_irq(&pool
->lock
);
2412 if (flush_color
>= 0) {
2413 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2415 if (pwq
->nr_in_flight
[flush_color
]) {
2416 pwq
->flush_color
= flush_color
;
2417 atomic_inc(&wq
->nr_pwqs_to_flush
);
2422 if (work_color
>= 0) {
2423 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2424 pwq
->work_color
= work_color
;
2427 spin_unlock_irq(&pool
->lock
);
2430 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2431 complete(&wq
->first_flusher
->done
);
2437 * flush_workqueue - ensure that any scheduled work has run to completion.
2438 * @wq: workqueue to flush
2440 * This function sleeps until all work items which were queued on entry
2441 * have finished execution, but it is not livelocked by new incoming ones.
2443 void flush_workqueue(struct workqueue_struct
*wq
)
2445 struct wq_flusher this_flusher
= {
2446 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2448 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2452 lock_map_acquire(&wq
->lockdep_map
);
2453 lock_map_release(&wq
->lockdep_map
);
2455 mutex_lock(&wq
->mutex
);
2458 * Start-to-wait phase
2460 next_color
= work_next_color(wq
->work_color
);
2462 if (next_color
!= wq
->flush_color
) {
2464 * Color space is not full. The current work_color
2465 * becomes our flush_color and work_color is advanced
2468 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2469 this_flusher
.flush_color
= wq
->work_color
;
2470 wq
->work_color
= next_color
;
2472 if (!wq
->first_flusher
) {
2473 /* no flush in progress, become the first flusher */
2474 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2476 wq
->first_flusher
= &this_flusher
;
2478 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2480 /* nothing to flush, done */
2481 wq
->flush_color
= next_color
;
2482 wq
->first_flusher
= NULL
;
2487 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2488 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2489 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2493 * Oops, color space is full, wait on overflow queue.
2494 * The next flush completion will assign us
2495 * flush_color and transfer to flusher_queue.
2497 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2500 mutex_unlock(&wq
->mutex
);
2502 wait_for_completion(&this_flusher
.done
);
2505 * Wake-up-and-cascade phase
2507 * First flushers are responsible for cascading flushes and
2508 * handling overflow. Non-first flushers can simply return.
2510 if (wq
->first_flusher
!= &this_flusher
)
2513 mutex_lock(&wq
->mutex
);
2515 /* we might have raced, check again with mutex held */
2516 if (wq
->first_flusher
!= &this_flusher
)
2519 wq
->first_flusher
= NULL
;
2521 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2522 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2525 struct wq_flusher
*next
, *tmp
;
2527 /* complete all the flushers sharing the current flush color */
2528 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2529 if (next
->flush_color
!= wq
->flush_color
)
2531 list_del_init(&next
->list
);
2532 complete(&next
->done
);
2535 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2536 wq
->flush_color
!= work_next_color(wq
->work_color
));
2538 /* this flush_color is finished, advance by one */
2539 wq
->flush_color
= work_next_color(wq
->flush_color
);
2541 /* one color has been freed, handle overflow queue */
2542 if (!list_empty(&wq
->flusher_overflow
)) {
2544 * Assign the same color to all overflowed
2545 * flushers, advance work_color and append to
2546 * flusher_queue. This is the start-to-wait
2547 * phase for these overflowed flushers.
2549 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2550 tmp
->flush_color
= wq
->work_color
;
2552 wq
->work_color
= work_next_color(wq
->work_color
);
2554 list_splice_tail_init(&wq
->flusher_overflow
,
2555 &wq
->flusher_queue
);
2556 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2559 if (list_empty(&wq
->flusher_queue
)) {
2560 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2565 * Need to flush more colors. Make the next flusher
2566 * the new first flusher and arm pwqs.
2568 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2569 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2571 list_del_init(&next
->list
);
2572 wq
->first_flusher
= next
;
2574 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2578 * Meh... this color is already done, clear first
2579 * flusher and repeat cascading.
2581 wq
->first_flusher
= NULL
;
2585 mutex_unlock(&wq
->mutex
);
2587 EXPORT_SYMBOL_GPL(flush_workqueue
);
2590 * drain_workqueue - drain a workqueue
2591 * @wq: workqueue to drain
2593 * Wait until the workqueue becomes empty. While draining is in progress,
2594 * only chain queueing is allowed. IOW, only currently pending or running
2595 * work items on @wq can queue further work items on it. @wq is flushed
2596 * repeatedly until it becomes empty. The number of flushing is detemined
2597 * by the depth of chaining and should be relatively short. Whine if it
2600 void drain_workqueue(struct workqueue_struct
*wq
)
2602 unsigned int flush_cnt
= 0;
2603 struct pool_workqueue
*pwq
;
2606 * __queue_work() needs to test whether there are drainers, is much
2607 * hotter than drain_workqueue() and already looks at @wq->flags.
2608 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2610 mutex_lock(&wq
->mutex
);
2611 if (!wq
->nr_drainers
++)
2612 wq
->flags
|= __WQ_DRAINING
;
2613 mutex_unlock(&wq
->mutex
);
2615 flush_workqueue(wq
);
2617 mutex_lock(&wq
->mutex
);
2619 for_each_pwq(pwq
, wq
) {
2622 spin_lock_irq(&pwq
->pool
->lock
);
2623 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2624 spin_unlock_irq(&pwq
->pool
->lock
);
2629 if (++flush_cnt
== 10 ||
2630 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2631 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2632 wq
->name
, flush_cnt
);
2634 mutex_unlock(&wq
->mutex
);
2638 if (!--wq
->nr_drainers
)
2639 wq
->flags
&= ~__WQ_DRAINING
;
2640 mutex_unlock(&wq
->mutex
);
2642 EXPORT_SYMBOL_GPL(drain_workqueue
);
2644 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2646 struct worker
*worker
= NULL
;
2647 struct worker_pool
*pool
;
2648 struct pool_workqueue
*pwq
;
2652 local_irq_disable();
2653 pool
= get_work_pool(work
);
2659 spin_lock(&pool
->lock
);
2660 /* see the comment in try_to_grab_pending() with the same code */
2661 pwq
= get_work_pwq(work
);
2663 if (unlikely(pwq
->pool
!= pool
))
2666 worker
= find_worker_executing_work(pool
, work
);
2669 pwq
= worker
->current_pwq
;
2672 insert_wq_barrier(pwq
, barr
, work
, worker
);
2673 spin_unlock_irq(&pool
->lock
);
2676 * If @max_active is 1 or rescuer is in use, flushing another work
2677 * item on the same workqueue may lead to deadlock. Make sure the
2678 * flusher is not running on the same workqueue by verifying write
2681 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2682 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2684 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2685 lock_map_release(&pwq
->wq
->lockdep_map
);
2689 spin_unlock_irq(&pool
->lock
);
2694 * flush_work - wait for a work to finish executing the last queueing instance
2695 * @work: the work to flush
2697 * Wait until @work has finished execution. @work is guaranteed to be idle
2698 * on return if it hasn't been requeued since flush started.
2701 * %true if flush_work() waited for the work to finish execution,
2702 * %false if it was already idle.
2704 bool flush_work(struct work_struct
*work
)
2706 struct wq_barrier barr
;
2708 lock_map_acquire(&work
->lockdep_map
);
2709 lock_map_release(&work
->lockdep_map
);
2711 if (start_flush_work(work
, &barr
)) {
2712 wait_for_completion(&barr
.done
);
2713 destroy_work_on_stack(&barr
.work
);
2719 EXPORT_SYMBOL_GPL(flush_work
);
2721 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2723 unsigned long flags
;
2727 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2729 * If someone else is canceling, wait for the same event it
2730 * would be waiting for before retrying.
2732 if (unlikely(ret
== -ENOENT
))
2734 } while (unlikely(ret
< 0));
2736 /* tell other tasks trying to grab @work to back off */
2737 mark_work_canceling(work
);
2738 local_irq_restore(flags
);
2741 clear_work_data(work
);
2746 * cancel_work_sync - cancel a work and wait for it to finish
2747 * @work: the work to cancel
2749 * Cancel @work and wait for its execution to finish. This function
2750 * can be used even if the work re-queues itself or migrates to
2751 * another workqueue. On return from this function, @work is
2752 * guaranteed to be not pending or executing on any CPU.
2754 * cancel_work_sync(&delayed_work->work) must not be used for
2755 * delayed_work's. Use cancel_delayed_work_sync() instead.
2757 * The caller must ensure that the workqueue on which @work was last
2758 * queued can't be destroyed before this function returns.
2761 * %true if @work was pending, %false otherwise.
2763 bool cancel_work_sync(struct work_struct
*work
)
2765 return __cancel_work_timer(work
, false);
2767 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2770 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2771 * @dwork: the delayed work to flush
2773 * Delayed timer is cancelled and the pending work is queued for
2774 * immediate execution. Like flush_work(), this function only
2775 * considers the last queueing instance of @dwork.
2778 * %true if flush_work() waited for the work to finish execution,
2779 * %false if it was already idle.
2781 bool flush_delayed_work(struct delayed_work
*dwork
)
2783 local_irq_disable();
2784 if (del_timer_sync(&dwork
->timer
))
2785 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2787 return flush_work(&dwork
->work
);
2789 EXPORT_SYMBOL(flush_delayed_work
);
2792 * cancel_delayed_work - cancel a delayed work
2793 * @dwork: delayed_work to cancel
2795 * Kill off a pending delayed_work.
2797 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2801 * The work callback function may still be running on return, unless
2802 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2803 * use cancel_delayed_work_sync() to wait on it.
2805 * This function is safe to call from any context including IRQ handler.
2807 bool cancel_delayed_work(struct delayed_work
*dwork
)
2809 unsigned long flags
;
2813 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2814 } while (unlikely(ret
== -EAGAIN
));
2816 if (unlikely(ret
< 0))
2819 set_work_pool_and_clear_pending(&dwork
->work
,
2820 get_work_pool_id(&dwork
->work
));
2821 local_irq_restore(flags
);
2824 EXPORT_SYMBOL(cancel_delayed_work
);
2827 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2828 * @dwork: the delayed work cancel
2830 * This is cancel_work_sync() for delayed works.
2833 * %true if @dwork was pending, %false otherwise.
2835 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2837 return __cancel_work_timer(&dwork
->work
, true);
2839 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2842 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2843 * @func: the function to call
2845 * schedule_on_each_cpu() executes @func on each online CPU using the
2846 * system workqueue and blocks until all CPUs have completed.
2847 * schedule_on_each_cpu() is very slow.
2850 * 0 on success, -errno on failure.
2852 int schedule_on_each_cpu(work_func_t func
)
2855 struct work_struct __percpu
*works
;
2857 works
= alloc_percpu(struct work_struct
);
2863 for_each_online_cpu(cpu
) {
2864 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2866 INIT_WORK(work
, func
);
2867 schedule_work_on(cpu
, work
);
2870 for_each_online_cpu(cpu
)
2871 flush_work(per_cpu_ptr(works
, cpu
));
2879 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2881 * Forces execution of the kernel-global workqueue and blocks until its
2884 * Think twice before calling this function! It's very easy to get into
2885 * trouble if you don't take great care. Either of the following situations
2886 * will lead to deadlock:
2888 * One of the work items currently on the workqueue needs to acquire
2889 * a lock held by your code or its caller.
2891 * Your code is running in the context of a work routine.
2893 * They will be detected by lockdep when they occur, but the first might not
2894 * occur very often. It depends on what work items are on the workqueue and
2895 * what locks they need, which you have no control over.
2897 * In most situations flushing the entire workqueue is overkill; you merely
2898 * need to know that a particular work item isn't queued and isn't running.
2899 * In such cases you should use cancel_delayed_work_sync() or
2900 * cancel_work_sync() instead.
2902 void flush_scheduled_work(void)
2904 flush_workqueue(system_wq
);
2906 EXPORT_SYMBOL(flush_scheduled_work
);
2909 * execute_in_process_context - reliably execute the routine with user context
2910 * @fn: the function to execute
2911 * @ew: guaranteed storage for the execute work structure (must
2912 * be available when the work executes)
2914 * Executes the function immediately if process context is available,
2915 * otherwise schedules the function for delayed execution.
2917 * Return: 0 - function was executed
2918 * 1 - function was scheduled for execution
2920 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
2922 if (!in_interrupt()) {
2927 INIT_WORK(&ew
->work
, fn
);
2928 schedule_work(&ew
->work
);
2932 EXPORT_SYMBOL_GPL(execute_in_process_context
);
2936 * Workqueues with WQ_SYSFS flag set is visible to userland via
2937 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
2938 * following attributes.
2940 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
2941 * max_active RW int : maximum number of in-flight work items
2943 * Unbound workqueues have the following extra attributes.
2945 * id RO int : the associated pool ID
2946 * nice RW int : nice value of the workers
2947 * cpumask RW mask : bitmask of allowed CPUs for the workers
2950 struct workqueue_struct
*wq
;
2954 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
2956 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
2961 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
2964 struct workqueue_struct
*wq
= dev_to_wq(dev
);
2966 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
2968 static DEVICE_ATTR_RO(per_cpu
);
2970 static ssize_t
max_active_show(struct device
*dev
,
2971 struct device_attribute
*attr
, char *buf
)
2973 struct workqueue_struct
*wq
= dev_to_wq(dev
);
2975 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
2978 static ssize_t
max_active_store(struct device
*dev
,
2979 struct device_attribute
*attr
, const char *buf
,
2982 struct workqueue_struct
*wq
= dev_to_wq(dev
);
2985 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
2988 workqueue_set_max_active(wq
, val
);
2991 static DEVICE_ATTR_RW(max_active
);
2993 static struct attribute
*wq_sysfs_attrs
[] = {
2994 &dev_attr_per_cpu
.attr
,
2995 &dev_attr_max_active
.attr
,
2998 ATTRIBUTE_GROUPS(wq_sysfs
);
3000 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3001 struct device_attribute
*attr
, char *buf
)
3003 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3004 const char *delim
= "";
3005 int node
, written
= 0;
3007 rcu_read_lock_sched();
3008 for_each_node(node
) {
3009 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3010 "%s%d:%d", delim
, node
,
3011 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3014 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3015 rcu_read_unlock_sched();
3020 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3023 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3026 mutex_lock(&wq
->mutex
);
3027 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3028 mutex_unlock(&wq
->mutex
);
3033 /* prepare workqueue_attrs for sysfs store operations */
3034 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3036 struct workqueue_attrs
*attrs
;
3038 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3042 mutex_lock(&wq
->mutex
);
3043 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3044 mutex_unlock(&wq
->mutex
);
3048 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3049 const char *buf
, size_t count
)
3051 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3052 struct workqueue_attrs
*attrs
;
3055 attrs
= wq_sysfs_prep_attrs(wq
);
3059 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3060 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
3061 ret
= apply_workqueue_attrs(wq
, attrs
);
3065 free_workqueue_attrs(attrs
);
3066 return ret
?: count
;
3069 static ssize_t
wq_cpumask_show(struct device
*dev
,
3070 struct device_attribute
*attr
, char *buf
)
3072 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3075 mutex_lock(&wq
->mutex
);
3076 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3077 mutex_unlock(&wq
->mutex
);
3079 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3083 static ssize_t
wq_cpumask_store(struct device
*dev
,
3084 struct device_attribute
*attr
,
3085 const char *buf
, size_t count
)
3087 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3088 struct workqueue_attrs
*attrs
;
3091 attrs
= wq_sysfs_prep_attrs(wq
);
3095 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3097 ret
= apply_workqueue_attrs(wq
, attrs
);
3099 free_workqueue_attrs(attrs
);
3100 return ret
?: count
;
3103 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3106 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3109 mutex_lock(&wq
->mutex
);
3110 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3111 !wq
->unbound_attrs
->no_numa
);
3112 mutex_unlock(&wq
->mutex
);
3117 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3118 const char *buf
, size_t count
)
3120 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3121 struct workqueue_attrs
*attrs
;
3124 attrs
= wq_sysfs_prep_attrs(wq
);
3129 if (sscanf(buf
, "%d", &v
) == 1) {
3130 attrs
->no_numa
= !v
;
3131 ret
= apply_workqueue_attrs(wq
, attrs
);
3134 free_workqueue_attrs(attrs
);
3135 return ret
?: count
;
3138 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3139 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3140 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3141 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3142 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3146 static struct bus_type wq_subsys
= {
3147 .name
= "workqueue",
3148 .dev_groups
= wq_sysfs_groups
,
3151 static int __init
wq_sysfs_init(void)
3153 return subsys_virtual_register(&wq_subsys
, NULL
);
3155 core_initcall(wq_sysfs_init
);
3157 static void wq_device_release(struct device
*dev
)
3159 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3165 * workqueue_sysfs_register - make a workqueue visible in sysfs
3166 * @wq: the workqueue to register
3168 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3169 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3170 * which is the preferred method.
3172 * Workqueue user should use this function directly iff it wants to apply
3173 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3174 * apply_workqueue_attrs() may race against userland updating the
3177 * Return: 0 on success, -errno on failure.
3179 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3181 struct wq_device
*wq_dev
;
3185 * Adjusting max_active or creating new pwqs by applyting
3186 * attributes breaks ordering guarantee. Disallow exposing ordered
3189 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3192 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3197 wq_dev
->dev
.bus
= &wq_subsys
;
3198 wq_dev
->dev
.init_name
= wq
->name
;
3199 wq_dev
->dev
.release
= wq_device_release
;
3202 * unbound_attrs are created separately. Suppress uevent until
3203 * everything is ready.
3205 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3207 ret
= device_register(&wq_dev
->dev
);
3214 if (wq
->flags
& WQ_UNBOUND
) {
3215 struct device_attribute
*attr
;
3217 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3218 ret
= device_create_file(&wq_dev
->dev
, attr
);
3220 device_unregister(&wq_dev
->dev
);
3227 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3232 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3233 * @wq: the workqueue to unregister
3235 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3237 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3239 struct wq_device
*wq_dev
= wq
->wq_dev
;
3245 device_unregister(&wq_dev
->dev
);
3247 #else /* CONFIG_SYSFS */
3248 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3249 #endif /* CONFIG_SYSFS */
3252 * free_workqueue_attrs - free a workqueue_attrs
3253 * @attrs: workqueue_attrs to free
3255 * Undo alloc_workqueue_attrs().
3257 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3260 free_cpumask_var(attrs
->cpumask
);
3266 * alloc_workqueue_attrs - allocate a workqueue_attrs
3267 * @gfp_mask: allocation mask to use
3269 * Allocate a new workqueue_attrs, initialize with default settings and
3272 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3274 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3276 struct workqueue_attrs
*attrs
;
3278 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3281 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3284 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3287 free_workqueue_attrs(attrs
);
3291 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3292 const struct workqueue_attrs
*from
)
3294 to
->nice
= from
->nice
;
3295 cpumask_copy(to
->cpumask
, from
->cpumask
);
3297 * Unlike hash and equality test, this function doesn't ignore
3298 * ->no_numa as it is used for both pool and wq attrs. Instead,
3299 * get_unbound_pool() explicitly clears ->no_numa after copying.
3301 to
->no_numa
= from
->no_numa
;
3304 /* hash value of the content of @attr */
3305 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3309 hash
= jhash_1word(attrs
->nice
, hash
);
3310 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3311 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3315 /* content equality test */
3316 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3317 const struct workqueue_attrs
*b
)
3319 if (a
->nice
!= b
->nice
)
3321 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3327 * init_worker_pool - initialize a newly zalloc'd worker_pool
3328 * @pool: worker_pool to initialize
3330 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3332 * Return: 0 on success, -errno on failure. Even on failure, all fields
3333 * inside @pool proper are initialized and put_unbound_pool() can be called
3334 * on @pool safely to release it.
3336 static int init_worker_pool(struct worker_pool
*pool
)
3338 spin_lock_init(&pool
->lock
);
3341 pool
->node
= NUMA_NO_NODE
;
3342 pool
->flags
|= POOL_DISASSOCIATED
;
3343 INIT_LIST_HEAD(&pool
->worklist
);
3344 INIT_LIST_HEAD(&pool
->idle_list
);
3345 hash_init(pool
->busy_hash
);
3347 init_timer_deferrable(&pool
->idle_timer
);
3348 pool
->idle_timer
.function
= idle_worker_timeout
;
3349 pool
->idle_timer
.data
= (unsigned long)pool
;
3351 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3352 (unsigned long)pool
);
3354 mutex_init(&pool
->manager_arb
);
3355 mutex_init(&pool
->attach_mutex
);
3356 INIT_LIST_HEAD(&pool
->workers
);
3358 ida_init(&pool
->worker_ida
);
3359 INIT_HLIST_NODE(&pool
->hash_node
);
3362 /* shouldn't fail above this point */
3363 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3369 static void rcu_free_pool(struct rcu_head
*rcu
)
3371 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3373 ida_destroy(&pool
->worker_ida
);
3374 free_workqueue_attrs(pool
->attrs
);
3379 * put_unbound_pool - put a worker_pool
3380 * @pool: worker_pool to put
3382 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3383 * safe manner. get_unbound_pool() calls this function on its failure path
3384 * and this function should be able to release pools which went through,
3385 * successfully or not, init_worker_pool().
3387 * Should be called with wq_pool_mutex held.
3389 static void put_unbound_pool(struct worker_pool
*pool
)
3391 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3392 struct worker
*worker
;
3394 lockdep_assert_held(&wq_pool_mutex
);
3400 if (WARN_ON(!(pool
->cpu
< 0)) ||
3401 WARN_ON(!list_empty(&pool
->worklist
)))
3404 /* release id and unhash */
3406 idr_remove(&worker_pool_idr
, pool
->id
);
3407 hash_del(&pool
->hash_node
);
3410 * Become the manager and destroy all workers. Grabbing
3411 * manager_arb prevents @pool's workers from blocking on
3414 mutex_lock(&pool
->manager_arb
);
3416 spin_lock_irq(&pool
->lock
);
3417 while ((worker
= first_idle_worker(pool
)))
3418 destroy_worker(worker
);
3419 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3420 spin_unlock_irq(&pool
->lock
);
3422 mutex_lock(&pool
->attach_mutex
);
3423 if (!list_empty(&pool
->workers
))
3424 pool
->detach_completion
= &detach_completion
;
3425 mutex_unlock(&pool
->attach_mutex
);
3427 if (pool
->detach_completion
)
3428 wait_for_completion(pool
->detach_completion
);
3430 mutex_unlock(&pool
->manager_arb
);
3432 /* shut down the timers */
3433 del_timer_sync(&pool
->idle_timer
);
3434 del_timer_sync(&pool
->mayday_timer
);
3436 /* sched-RCU protected to allow dereferences from get_work_pool() */
3437 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3441 * get_unbound_pool - get a worker_pool with the specified attributes
3442 * @attrs: the attributes of the worker_pool to get
3444 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3445 * reference count and return it. If there already is a matching
3446 * worker_pool, it will be used; otherwise, this function attempts to
3449 * Should be called with wq_pool_mutex held.
3451 * Return: On success, a worker_pool with the same attributes as @attrs.
3452 * On failure, %NULL.
3454 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3456 u32 hash
= wqattrs_hash(attrs
);
3457 struct worker_pool
*pool
;
3460 lockdep_assert_held(&wq_pool_mutex
);
3462 /* do we already have a matching pool? */
3463 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3464 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3470 /* nope, create a new one */
3471 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3472 if (!pool
|| init_worker_pool(pool
) < 0)
3475 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3476 copy_workqueue_attrs(pool
->attrs
, attrs
);
3479 * no_numa isn't a worker_pool attribute, always clear it. See
3480 * 'struct workqueue_attrs' comments for detail.
3482 pool
->attrs
->no_numa
= false;
3484 /* if cpumask is contained inside a NUMA node, we belong to that node */
3485 if (wq_numa_enabled
) {
3486 for_each_node(node
) {
3487 if (cpumask_subset(pool
->attrs
->cpumask
,
3488 wq_numa_possible_cpumask
[node
])) {
3495 if (worker_pool_assign_id(pool
) < 0)
3498 /* create and start the initial worker */
3499 if (!create_worker(pool
))
3503 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3508 put_unbound_pool(pool
);
3512 static void rcu_free_pwq(struct rcu_head
*rcu
)
3514 kmem_cache_free(pwq_cache
,
3515 container_of(rcu
, struct pool_workqueue
, rcu
));
3519 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3520 * and needs to be destroyed.
3522 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3524 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3525 unbound_release_work
);
3526 struct workqueue_struct
*wq
= pwq
->wq
;
3527 struct worker_pool
*pool
= pwq
->pool
;
3530 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3533 mutex_lock(&wq
->mutex
);
3534 list_del_rcu(&pwq
->pwqs_node
);
3535 is_last
= list_empty(&wq
->pwqs
);
3536 mutex_unlock(&wq
->mutex
);
3538 mutex_lock(&wq_pool_mutex
);
3539 put_unbound_pool(pool
);
3540 mutex_unlock(&wq_pool_mutex
);
3542 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3545 * If we're the last pwq going away, @wq is already dead and no one
3546 * is gonna access it anymore. Free it.
3549 free_workqueue_attrs(wq
->unbound_attrs
);
3555 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3556 * @pwq: target pool_workqueue
3558 * If @pwq isn't freezing, set @pwq->max_active to the associated
3559 * workqueue's saved_max_active and activate delayed work items
3560 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3562 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3564 struct workqueue_struct
*wq
= pwq
->wq
;
3565 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3567 /* for @wq->saved_max_active */
3568 lockdep_assert_held(&wq
->mutex
);
3570 /* fast exit for non-freezable wqs */
3571 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3574 spin_lock_irq(&pwq
->pool
->lock
);
3577 * During [un]freezing, the caller is responsible for ensuring that
3578 * this function is called at least once after @workqueue_freezing
3579 * is updated and visible.
3581 if (!freezable
|| !workqueue_freezing
) {
3582 pwq
->max_active
= wq
->saved_max_active
;
3584 while (!list_empty(&pwq
->delayed_works
) &&
3585 pwq
->nr_active
< pwq
->max_active
)
3586 pwq_activate_first_delayed(pwq
);
3589 * Need to kick a worker after thawed or an unbound wq's
3590 * max_active is bumped. It's a slow path. Do it always.
3592 wake_up_worker(pwq
->pool
);
3594 pwq
->max_active
= 0;
3597 spin_unlock_irq(&pwq
->pool
->lock
);
3600 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3601 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3602 struct worker_pool
*pool
)
3604 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3606 memset(pwq
, 0, sizeof(*pwq
));
3610 pwq
->flush_color
= -1;
3612 INIT_LIST_HEAD(&pwq
->delayed_works
);
3613 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3614 INIT_LIST_HEAD(&pwq
->mayday_node
);
3615 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3618 /* sync @pwq with the current state of its associated wq and link it */
3619 static void link_pwq(struct pool_workqueue
*pwq
)
3621 struct workqueue_struct
*wq
= pwq
->wq
;
3623 lockdep_assert_held(&wq
->mutex
);
3625 /* may be called multiple times, ignore if already linked */
3626 if (!list_empty(&pwq
->pwqs_node
))
3629 /* set the matching work_color */
3630 pwq
->work_color
= wq
->work_color
;
3632 /* sync max_active to the current setting */
3633 pwq_adjust_max_active(pwq
);
3636 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3639 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3640 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3641 const struct workqueue_attrs
*attrs
)
3643 struct worker_pool
*pool
;
3644 struct pool_workqueue
*pwq
;
3646 lockdep_assert_held(&wq_pool_mutex
);
3648 pool
= get_unbound_pool(attrs
);
3652 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3654 put_unbound_pool(pool
);
3658 init_pwq(pwq
, wq
, pool
);
3662 /* undo alloc_unbound_pwq(), used only in the error path */
3663 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3665 lockdep_assert_held(&wq_pool_mutex
);
3668 put_unbound_pool(pwq
->pool
);
3669 kmem_cache_free(pwq_cache
, pwq
);
3674 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3675 * @attrs: the wq_attrs of interest
3676 * @node: the target NUMA node
3677 * @cpu_going_down: if >= 0, the CPU to consider as offline
3678 * @cpumask: outarg, the resulting cpumask
3680 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3681 * @cpu_going_down is >= 0, that cpu is considered offline during
3682 * calculation. The result is stored in @cpumask.
3684 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3685 * enabled and @node has online CPUs requested by @attrs, the returned
3686 * cpumask is the intersection of the possible CPUs of @node and
3689 * The caller is responsible for ensuring that the cpumask of @node stays
3692 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3695 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3696 int cpu_going_down
, cpumask_t
*cpumask
)
3698 if (!wq_numa_enabled
|| attrs
->no_numa
)
3701 /* does @node have any online CPUs @attrs wants? */
3702 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3703 if (cpu_going_down
>= 0)
3704 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3706 if (cpumask_empty(cpumask
))
3709 /* yeap, return possible CPUs in @node that @attrs wants */
3710 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3711 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3714 cpumask_copy(cpumask
, attrs
->cpumask
);
3718 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3719 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3721 struct pool_workqueue
*pwq
)
3723 struct pool_workqueue
*old_pwq
;
3725 lockdep_assert_held(&wq
->mutex
);
3727 /* link_pwq() can handle duplicate calls */
3730 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3731 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3736 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3737 * @wq: the target workqueue
3738 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3740 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3741 * machines, this function maps a separate pwq to each NUMA node with
3742 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3743 * NUMA node it was issued on. Older pwqs are released as in-flight work
3744 * items finish. Note that a work item which repeatedly requeues itself
3745 * back-to-back will stay on its current pwq.
3747 * Performs GFP_KERNEL allocations.
3749 * Return: 0 on success and -errno on failure.
3751 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3752 const struct workqueue_attrs
*attrs
)
3754 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3755 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3758 /* only unbound workqueues can change attributes */
3759 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3762 /* creating multiple pwqs breaks ordering guarantee */
3763 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3766 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3767 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3768 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3769 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3772 /* make a copy of @attrs and sanitize it */
3773 copy_workqueue_attrs(new_attrs
, attrs
);
3774 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3777 * We may create multiple pwqs with differing cpumasks. Make a
3778 * copy of @new_attrs which will be modified and used to obtain
3781 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3784 * CPUs should stay stable across pwq creations and installations.
3785 * Pin CPUs, determine the target cpumask for each node and create
3790 mutex_lock(&wq_pool_mutex
);
3793 * If something goes wrong during CPU up/down, we'll fall back to
3794 * the default pwq covering whole @attrs->cpumask. Always create
3795 * it even if we don't use it immediately.
3797 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3801 for_each_node(node
) {
3802 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3803 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3808 pwq_tbl
[node
] = dfl_pwq
;
3812 mutex_unlock(&wq_pool_mutex
);
3814 /* all pwqs have been created successfully, let's install'em */
3815 mutex_lock(&wq
->mutex
);
3817 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3819 /* save the previous pwq and install the new one */
3821 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3823 /* @dfl_pwq might not have been used, ensure it's linked */
3825 swap(wq
->dfl_pwq
, dfl_pwq
);
3827 mutex_unlock(&wq
->mutex
);
3829 /* put the old pwqs */
3831 put_pwq_unlocked(pwq_tbl
[node
]);
3832 put_pwq_unlocked(dfl_pwq
);
3838 free_workqueue_attrs(tmp_attrs
);
3839 free_workqueue_attrs(new_attrs
);
3844 free_unbound_pwq(dfl_pwq
);
3846 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
3847 free_unbound_pwq(pwq_tbl
[node
]);
3848 mutex_unlock(&wq_pool_mutex
);
3856 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3857 * @wq: the target workqueue
3858 * @cpu: the CPU coming up or going down
3859 * @online: whether @cpu is coming up or going down
3861 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3862 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3865 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3866 * falls back to @wq->dfl_pwq which may not be optimal but is always
3869 * Note that when the last allowed CPU of a NUMA node goes offline for a
3870 * workqueue with a cpumask spanning multiple nodes, the workers which were
3871 * already executing the work items for the workqueue will lose their CPU
3872 * affinity and may execute on any CPU. This is similar to how per-cpu
3873 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3874 * affinity, it's the user's responsibility to flush the work item from
3877 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3880 int node
= cpu_to_node(cpu
);
3881 int cpu_off
= online
? -1 : cpu
;
3882 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3883 struct workqueue_attrs
*target_attrs
;
3886 lockdep_assert_held(&wq_pool_mutex
);
3888 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
3892 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3893 * Let's use a preallocated one. The following buf is protected by
3894 * CPU hotplug exclusion.
3896 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3897 cpumask
= target_attrs
->cpumask
;
3899 mutex_lock(&wq
->mutex
);
3900 if (wq
->unbound_attrs
->no_numa
)
3903 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3904 pwq
= unbound_pwq_by_node(wq
, node
);
3907 * Let's determine what needs to be done. If the target cpumask is
3908 * different from wq's, we need to compare it to @pwq's and create
3909 * a new one if they don't match. If the target cpumask equals
3910 * wq's, the default pwq should be used.
3912 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
3913 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3919 mutex_unlock(&wq
->mutex
);
3921 /* create a new pwq */
3922 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3924 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3926 mutex_lock(&wq
->mutex
);
3931 * Install the new pwq. As this function is called only from CPU
3932 * hotplug callbacks and applying a new attrs is wrapped with
3933 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3936 mutex_lock(&wq
->mutex
);
3937 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3941 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3942 get_pwq(wq
->dfl_pwq
);
3943 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3944 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
3946 mutex_unlock(&wq
->mutex
);
3947 put_pwq_unlocked(old_pwq
);
3950 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3952 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3955 if (!(wq
->flags
& WQ_UNBOUND
)) {
3956 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3960 for_each_possible_cpu(cpu
) {
3961 struct pool_workqueue
*pwq
=
3962 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3963 struct worker_pool
*cpu_pools
=
3964 per_cpu(cpu_worker_pools
, cpu
);
3966 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
3968 mutex_lock(&wq
->mutex
);
3970 mutex_unlock(&wq
->mutex
);
3973 } else if (wq
->flags
& __WQ_ORDERED
) {
3974 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
3975 /* there should only be single pwq for ordering guarantee */
3976 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
3977 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
3978 "ordering guarantee broken for workqueue %s\n", wq
->name
);
3981 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3985 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3988 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3990 if (max_active
< 1 || max_active
> lim
)
3991 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3992 max_active
, name
, 1, lim
);
3994 return clamp_val(max_active
, 1, lim
);
3997 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4000 struct lock_class_key
*key
,
4001 const char *lock_name
, ...)
4003 size_t tbl_size
= 0;
4005 struct workqueue_struct
*wq
;
4006 struct pool_workqueue
*pwq
;
4008 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4009 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4010 flags
|= WQ_UNBOUND
;
4012 /* allocate wq and format name */
4013 if (flags
& WQ_UNBOUND
)
4014 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4016 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4020 if (flags
& WQ_UNBOUND
) {
4021 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4022 if (!wq
->unbound_attrs
)
4026 va_start(args
, lock_name
);
4027 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4030 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4031 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4035 wq
->saved_max_active
= max_active
;
4036 mutex_init(&wq
->mutex
);
4037 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4038 INIT_LIST_HEAD(&wq
->pwqs
);
4039 INIT_LIST_HEAD(&wq
->flusher_queue
);
4040 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4041 INIT_LIST_HEAD(&wq
->maydays
);
4043 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4044 INIT_LIST_HEAD(&wq
->list
);
4046 if (alloc_and_link_pwqs(wq
) < 0)
4050 * Workqueues which may be used during memory reclaim should
4051 * have a rescuer to guarantee forward progress.
4053 if (flags
& WQ_MEM_RECLAIM
) {
4054 struct worker
*rescuer
;
4056 rescuer
= alloc_worker(NUMA_NO_NODE
);
4060 rescuer
->rescue_wq
= wq
;
4061 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4063 if (IS_ERR(rescuer
->task
)) {
4068 wq
->rescuer
= rescuer
;
4069 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4070 wake_up_process(rescuer
->task
);
4073 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4077 * wq_pool_mutex protects global freeze state and workqueues list.
4078 * Grab it, adjust max_active and add the new @wq to workqueues
4081 mutex_lock(&wq_pool_mutex
);
4083 mutex_lock(&wq
->mutex
);
4084 for_each_pwq(pwq
, wq
)
4085 pwq_adjust_max_active(pwq
);
4086 mutex_unlock(&wq
->mutex
);
4088 list_add(&wq
->list
, &workqueues
);
4090 mutex_unlock(&wq_pool_mutex
);
4095 free_workqueue_attrs(wq
->unbound_attrs
);
4099 destroy_workqueue(wq
);
4102 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4105 * destroy_workqueue - safely terminate a workqueue
4106 * @wq: target workqueue
4108 * Safely destroy a workqueue. All work currently pending will be done first.
4110 void destroy_workqueue(struct workqueue_struct
*wq
)
4112 struct pool_workqueue
*pwq
;
4115 /* drain it before proceeding with destruction */
4116 drain_workqueue(wq
);
4119 mutex_lock(&wq
->mutex
);
4120 for_each_pwq(pwq
, wq
) {
4123 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4124 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4125 mutex_unlock(&wq
->mutex
);
4130 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4131 WARN_ON(pwq
->nr_active
) ||
4132 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4133 mutex_unlock(&wq
->mutex
);
4137 mutex_unlock(&wq
->mutex
);
4140 * wq list is used to freeze wq, remove from list after
4141 * flushing is complete in case freeze races us.
4143 mutex_lock(&wq_pool_mutex
);
4144 list_del_init(&wq
->list
);
4145 mutex_unlock(&wq_pool_mutex
);
4147 workqueue_sysfs_unregister(wq
);
4150 kthread_stop(wq
->rescuer
->task
);
4155 if (!(wq
->flags
& WQ_UNBOUND
)) {
4157 * The base ref is never dropped on per-cpu pwqs. Directly
4158 * free the pwqs and wq.
4160 free_percpu(wq
->cpu_pwqs
);
4164 * We're the sole accessor of @wq at this point. Directly
4165 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4166 * @wq will be freed when the last pwq is released.
4168 for_each_node(node
) {
4169 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4170 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4171 put_pwq_unlocked(pwq
);
4175 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4176 * put. Don't access it afterwards.
4180 put_pwq_unlocked(pwq
);
4183 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4186 * workqueue_set_max_active - adjust max_active of a workqueue
4187 * @wq: target workqueue
4188 * @max_active: new max_active value.
4190 * Set max_active of @wq to @max_active.
4193 * Don't call from IRQ context.
4195 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4197 struct pool_workqueue
*pwq
;
4199 /* disallow meddling with max_active for ordered workqueues */
4200 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4203 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4205 mutex_lock(&wq
->mutex
);
4207 wq
->saved_max_active
= max_active
;
4209 for_each_pwq(pwq
, wq
)
4210 pwq_adjust_max_active(pwq
);
4212 mutex_unlock(&wq
->mutex
);
4214 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4217 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4219 * Determine whether %current is a workqueue rescuer. Can be used from
4220 * work functions to determine whether it's being run off the rescuer task.
4222 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4224 bool current_is_workqueue_rescuer(void)
4226 struct worker
*worker
= current_wq_worker();
4228 return worker
&& worker
->rescue_wq
;
4232 * workqueue_congested - test whether a workqueue is congested
4233 * @cpu: CPU in question
4234 * @wq: target workqueue
4236 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4237 * no synchronization around this function and the test result is
4238 * unreliable and only useful as advisory hints or for debugging.
4240 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4241 * Note that both per-cpu and unbound workqueues may be associated with
4242 * multiple pool_workqueues which have separate congested states. A
4243 * workqueue being congested on one CPU doesn't mean the workqueue is also
4244 * contested on other CPUs / NUMA nodes.
4247 * %true if congested, %false otherwise.
4249 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4251 struct pool_workqueue
*pwq
;
4254 rcu_read_lock_sched();
4256 if (cpu
== WORK_CPU_UNBOUND
)
4257 cpu
= smp_processor_id();
4259 if (!(wq
->flags
& WQ_UNBOUND
))
4260 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4262 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4264 ret
= !list_empty(&pwq
->delayed_works
);
4265 rcu_read_unlock_sched();
4269 EXPORT_SYMBOL_GPL(workqueue_congested
);
4272 * work_busy - test whether a work is currently pending or running
4273 * @work: the work to be tested
4275 * Test whether @work is currently pending or running. There is no
4276 * synchronization around this function and the test result is
4277 * unreliable and only useful as advisory hints or for debugging.
4280 * OR'd bitmask of WORK_BUSY_* bits.
4282 unsigned int work_busy(struct work_struct
*work
)
4284 struct worker_pool
*pool
;
4285 unsigned long flags
;
4286 unsigned int ret
= 0;
4288 if (work_pending(work
))
4289 ret
|= WORK_BUSY_PENDING
;
4291 local_irq_save(flags
);
4292 pool
= get_work_pool(work
);
4294 spin_lock(&pool
->lock
);
4295 if (find_worker_executing_work(pool
, work
))
4296 ret
|= WORK_BUSY_RUNNING
;
4297 spin_unlock(&pool
->lock
);
4299 local_irq_restore(flags
);
4303 EXPORT_SYMBOL_GPL(work_busy
);
4306 * set_worker_desc - set description for the current work item
4307 * @fmt: printf-style format string
4308 * @...: arguments for the format string
4310 * This function can be called by a running work function to describe what
4311 * the work item is about. If the worker task gets dumped, this
4312 * information will be printed out together to help debugging. The
4313 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4315 void set_worker_desc(const char *fmt
, ...)
4317 struct worker
*worker
= current_wq_worker();
4321 va_start(args
, fmt
);
4322 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4324 worker
->desc_valid
= true;
4329 * print_worker_info - print out worker information and description
4330 * @log_lvl: the log level to use when printing
4331 * @task: target task
4333 * If @task is a worker and currently executing a work item, print out the
4334 * name of the workqueue being serviced and worker description set with
4335 * set_worker_desc() by the currently executing work item.
4337 * This function can be safely called on any task as long as the
4338 * task_struct itself is accessible. While safe, this function isn't
4339 * synchronized and may print out mixups or garbages of limited length.
4341 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4343 work_func_t
*fn
= NULL
;
4344 char name
[WQ_NAME_LEN
] = { };
4345 char desc
[WORKER_DESC_LEN
] = { };
4346 struct pool_workqueue
*pwq
= NULL
;
4347 struct workqueue_struct
*wq
= NULL
;
4348 bool desc_valid
= false;
4349 struct worker
*worker
;
4351 if (!(task
->flags
& PF_WQ_WORKER
))
4355 * This function is called without any synchronization and @task
4356 * could be in any state. Be careful with dereferences.
4358 worker
= probe_kthread_data(task
);
4361 * Carefully copy the associated workqueue's workfn and name. Keep
4362 * the original last '\0' in case the original contains garbage.
4364 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4365 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4366 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4367 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4369 /* copy worker description */
4370 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4372 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4374 if (fn
|| name
[0] || desc
[0]) {
4375 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4377 pr_cont(" (%s)", desc
);
4385 * There are two challenges in supporting CPU hotplug. Firstly, there
4386 * are a lot of assumptions on strong associations among work, pwq and
4387 * pool which make migrating pending and scheduled works very
4388 * difficult to implement without impacting hot paths. Secondly,
4389 * worker pools serve mix of short, long and very long running works making
4390 * blocked draining impractical.
4392 * This is solved by allowing the pools to be disassociated from the CPU
4393 * running as an unbound one and allowing it to be reattached later if the
4394 * cpu comes back online.
4397 static void wq_unbind_fn(struct work_struct
*work
)
4399 int cpu
= smp_processor_id();
4400 struct worker_pool
*pool
;
4401 struct worker
*worker
;
4403 for_each_cpu_worker_pool(pool
, cpu
) {
4404 mutex_lock(&pool
->attach_mutex
);
4405 spin_lock_irq(&pool
->lock
);
4408 * We've blocked all attach/detach operations. Make all workers
4409 * unbound and set DISASSOCIATED. Before this, all workers
4410 * except for the ones which are still executing works from
4411 * before the last CPU down must be on the cpu. After
4412 * this, they may become diasporas.
4414 for_each_pool_worker(worker
, pool
)
4415 worker
->flags
|= WORKER_UNBOUND
;
4417 pool
->flags
|= POOL_DISASSOCIATED
;
4419 spin_unlock_irq(&pool
->lock
);
4420 mutex_unlock(&pool
->attach_mutex
);
4423 * Call schedule() so that we cross rq->lock and thus can
4424 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4425 * This is necessary as scheduler callbacks may be invoked
4431 * Sched callbacks are disabled now. Zap nr_running.
4432 * After this, nr_running stays zero and need_more_worker()
4433 * and keep_working() are always true as long as the
4434 * worklist is not empty. This pool now behaves as an
4435 * unbound (in terms of concurrency management) pool which
4436 * are served by workers tied to the pool.
4438 atomic_set(&pool
->nr_running
, 0);
4441 * With concurrency management just turned off, a busy
4442 * worker blocking could lead to lengthy stalls. Kick off
4443 * unbound chain execution of currently pending work items.
4445 spin_lock_irq(&pool
->lock
);
4446 wake_up_worker(pool
);
4447 spin_unlock_irq(&pool
->lock
);
4452 * rebind_workers - rebind all workers of a pool to the associated CPU
4453 * @pool: pool of interest
4455 * @pool->cpu is coming online. Rebind all workers to the CPU.
4457 static void rebind_workers(struct worker_pool
*pool
)
4459 struct worker
*worker
;
4461 lockdep_assert_held(&pool
->attach_mutex
);
4464 * Restore CPU affinity of all workers. As all idle workers should
4465 * be on the run-queue of the associated CPU before any local
4466 * wake-ups for concurrency management happen, restore CPU affinty
4467 * of all workers first and then clear UNBOUND. As we're called
4468 * from CPU_ONLINE, the following shouldn't fail.
4470 for_each_pool_worker(worker
, pool
)
4471 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4472 pool
->attrs
->cpumask
) < 0);
4474 spin_lock_irq(&pool
->lock
);
4475 pool
->flags
&= ~POOL_DISASSOCIATED
;
4477 for_each_pool_worker(worker
, pool
) {
4478 unsigned int worker_flags
= worker
->flags
;
4481 * A bound idle worker should actually be on the runqueue
4482 * of the associated CPU for local wake-ups targeting it to
4483 * work. Kick all idle workers so that they migrate to the
4484 * associated CPU. Doing this in the same loop as
4485 * replacing UNBOUND with REBOUND is safe as no worker will
4486 * be bound before @pool->lock is released.
4488 if (worker_flags
& WORKER_IDLE
)
4489 wake_up_process(worker
->task
);
4492 * We want to clear UNBOUND but can't directly call
4493 * worker_clr_flags() or adjust nr_running. Atomically
4494 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4495 * @worker will clear REBOUND using worker_clr_flags() when
4496 * it initiates the next execution cycle thus restoring
4497 * concurrency management. Note that when or whether
4498 * @worker clears REBOUND doesn't affect correctness.
4500 * ACCESS_ONCE() is necessary because @worker->flags may be
4501 * tested without holding any lock in
4502 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4503 * fail incorrectly leading to premature concurrency
4504 * management operations.
4506 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4507 worker_flags
|= WORKER_REBOUND
;
4508 worker_flags
&= ~WORKER_UNBOUND
;
4509 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4512 spin_unlock_irq(&pool
->lock
);
4516 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4517 * @pool: unbound pool of interest
4518 * @cpu: the CPU which is coming up
4520 * An unbound pool may end up with a cpumask which doesn't have any online
4521 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4522 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4523 * online CPU before, cpus_allowed of all its workers should be restored.
4525 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4527 static cpumask_t cpumask
;
4528 struct worker
*worker
;
4530 lockdep_assert_held(&pool
->attach_mutex
);
4532 /* is @cpu allowed for @pool? */
4533 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4536 /* is @cpu the only online CPU? */
4537 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4538 if (cpumask_weight(&cpumask
) != 1)
4541 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4542 for_each_pool_worker(worker
, pool
)
4543 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4544 pool
->attrs
->cpumask
) < 0);
4548 * Workqueues should be brought up before normal priority CPU notifiers.
4549 * This will be registered high priority CPU notifier.
4551 static int workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4552 unsigned long action
,
4555 int cpu
= (unsigned long)hcpu
;
4556 struct worker_pool
*pool
;
4557 struct workqueue_struct
*wq
;
4560 switch (action
& ~CPU_TASKS_FROZEN
) {
4561 case CPU_UP_PREPARE
:
4562 for_each_cpu_worker_pool(pool
, cpu
) {
4563 if (pool
->nr_workers
)
4565 if (!create_worker(pool
))
4570 case CPU_DOWN_FAILED
:
4572 mutex_lock(&wq_pool_mutex
);
4574 for_each_pool(pool
, pi
) {
4575 mutex_lock(&pool
->attach_mutex
);
4577 if (pool
->cpu
== cpu
) {
4578 rebind_workers(pool
);
4579 } else if (pool
->cpu
< 0) {
4580 restore_unbound_workers_cpumask(pool
, cpu
);
4583 mutex_unlock(&pool
->attach_mutex
);
4586 /* update NUMA affinity of unbound workqueues */
4587 list_for_each_entry(wq
, &workqueues
, list
)
4588 wq_update_unbound_numa(wq
, cpu
, true);
4590 mutex_unlock(&wq_pool_mutex
);
4597 * Workqueues should be brought down after normal priority CPU notifiers.
4598 * This will be registered as low priority CPU notifier.
4600 static int workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4601 unsigned long action
,
4604 int cpu
= (unsigned long)hcpu
;
4605 struct work_struct unbind_work
;
4606 struct workqueue_struct
*wq
;
4608 switch (action
& ~CPU_TASKS_FROZEN
) {
4609 case CPU_DOWN_PREPARE
:
4610 /* unbinding per-cpu workers should happen on the local CPU */
4611 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4612 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4614 /* update NUMA affinity of unbound workqueues */
4615 mutex_lock(&wq_pool_mutex
);
4616 list_for_each_entry(wq
, &workqueues
, list
)
4617 wq_update_unbound_numa(wq
, cpu
, false);
4618 mutex_unlock(&wq_pool_mutex
);
4620 /* wait for per-cpu unbinding to finish */
4621 flush_work(&unbind_work
);
4622 destroy_work_on_stack(&unbind_work
);
4630 struct work_for_cpu
{
4631 struct work_struct work
;
4637 static void work_for_cpu_fn(struct work_struct
*work
)
4639 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4641 wfc
->ret
= wfc
->fn(wfc
->arg
);
4645 * work_on_cpu - run a function in user context on a particular cpu
4646 * @cpu: the cpu to run on
4647 * @fn: the function to run
4648 * @arg: the function arg
4650 * It is up to the caller to ensure that the cpu doesn't go offline.
4651 * The caller must not hold any locks which would prevent @fn from completing.
4653 * Return: The value @fn returns.
4655 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4657 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4659 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4660 schedule_work_on(cpu
, &wfc
.work
);
4661 flush_work(&wfc
.work
);
4662 destroy_work_on_stack(&wfc
.work
);
4665 EXPORT_SYMBOL_GPL(work_on_cpu
);
4666 #endif /* CONFIG_SMP */
4668 #ifdef CONFIG_FREEZER
4671 * freeze_workqueues_begin - begin freezing workqueues
4673 * Start freezing workqueues. After this function returns, all freezable
4674 * workqueues will queue new works to their delayed_works list instead of
4678 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4680 void freeze_workqueues_begin(void)
4682 struct workqueue_struct
*wq
;
4683 struct pool_workqueue
*pwq
;
4685 mutex_lock(&wq_pool_mutex
);
4687 WARN_ON_ONCE(workqueue_freezing
);
4688 workqueue_freezing
= true;
4690 list_for_each_entry(wq
, &workqueues
, list
) {
4691 mutex_lock(&wq
->mutex
);
4692 for_each_pwq(pwq
, wq
)
4693 pwq_adjust_max_active(pwq
);
4694 mutex_unlock(&wq
->mutex
);
4697 mutex_unlock(&wq_pool_mutex
);
4701 * freeze_workqueues_busy - are freezable workqueues still busy?
4703 * Check whether freezing is complete. This function must be called
4704 * between freeze_workqueues_begin() and thaw_workqueues().
4707 * Grabs and releases wq_pool_mutex.
4710 * %true if some freezable workqueues are still busy. %false if freezing
4713 bool freeze_workqueues_busy(void)
4716 struct workqueue_struct
*wq
;
4717 struct pool_workqueue
*pwq
;
4719 mutex_lock(&wq_pool_mutex
);
4721 WARN_ON_ONCE(!workqueue_freezing
);
4723 list_for_each_entry(wq
, &workqueues
, list
) {
4724 if (!(wq
->flags
& WQ_FREEZABLE
))
4727 * nr_active is monotonically decreasing. It's safe
4728 * to peek without lock.
4730 rcu_read_lock_sched();
4731 for_each_pwq(pwq
, wq
) {
4732 WARN_ON_ONCE(pwq
->nr_active
< 0);
4733 if (pwq
->nr_active
) {
4735 rcu_read_unlock_sched();
4739 rcu_read_unlock_sched();
4742 mutex_unlock(&wq_pool_mutex
);
4747 * thaw_workqueues - thaw workqueues
4749 * Thaw workqueues. Normal queueing is restored and all collected
4750 * frozen works are transferred to their respective pool worklists.
4753 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4755 void thaw_workqueues(void)
4757 struct workqueue_struct
*wq
;
4758 struct pool_workqueue
*pwq
;
4760 mutex_lock(&wq_pool_mutex
);
4762 if (!workqueue_freezing
)
4765 workqueue_freezing
= false;
4767 /* restore max_active and repopulate worklist */
4768 list_for_each_entry(wq
, &workqueues
, list
) {
4769 mutex_lock(&wq
->mutex
);
4770 for_each_pwq(pwq
, wq
)
4771 pwq_adjust_max_active(pwq
);
4772 mutex_unlock(&wq
->mutex
);
4776 mutex_unlock(&wq_pool_mutex
);
4778 #endif /* CONFIG_FREEZER */
4780 static void __init
wq_numa_init(void)
4785 /* determine NUMA pwq table len - highest node id + 1 */
4787 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4789 if (num_possible_nodes() <= 1)
4792 if (wq_disable_numa
) {
4793 pr_info("workqueue: NUMA affinity support disabled\n");
4797 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4798 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4801 * We want masks of possible CPUs of each node which isn't readily
4802 * available. Build one from cpu_to_node() which should have been
4803 * fully initialized by now.
4805 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
4809 BUG_ON(!alloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
4810 node_online(node
) ? node
: NUMA_NO_NODE
));
4812 for_each_possible_cpu(cpu
) {
4813 node
= cpu_to_node(cpu
);
4814 if (WARN_ON(node
== NUMA_NO_NODE
)) {
4815 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
4816 /* happens iff arch is bonkers, let's just proceed */
4819 cpumask_set_cpu(cpu
, tbl
[node
]);
4822 wq_numa_possible_cpumask
= tbl
;
4823 wq_numa_enabled
= true;
4826 static int __init
init_workqueues(void)
4828 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4831 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4833 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4835 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4836 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4840 /* initialize CPU pools */
4841 for_each_possible_cpu(cpu
) {
4842 struct worker_pool
*pool
;
4845 for_each_cpu_worker_pool(pool
, cpu
) {
4846 BUG_ON(init_worker_pool(pool
));
4848 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4849 pool
->attrs
->nice
= std_nice
[i
++];
4850 pool
->node
= cpu_to_node(cpu
);
4853 mutex_lock(&wq_pool_mutex
);
4854 BUG_ON(worker_pool_assign_id(pool
));
4855 mutex_unlock(&wq_pool_mutex
);
4859 /* create the initial worker */
4860 for_each_online_cpu(cpu
) {
4861 struct worker_pool
*pool
;
4863 for_each_cpu_worker_pool(pool
, cpu
) {
4864 pool
->flags
&= ~POOL_DISASSOCIATED
;
4865 BUG_ON(!create_worker(pool
));
4869 /* create default unbound and ordered wq attrs */
4870 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4871 struct workqueue_attrs
*attrs
;
4873 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4874 attrs
->nice
= std_nice
[i
];
4875 unbound_std_wq_attrs
[i
] = attrs
;
4878 * An ordered wq should have only one pwq as ordering is
4879 * guaranteed by max_active which is enforced by pwqs.
4880 * Turn off NUMA so that dfl_pwq is used for all nodes.
4882 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4883 attrs
->nice
= std_nice
[i
];
4884 attrs
->no_numa
= true;
4885 ordered_wq_attrs
[i
] = attrs
;
4888 system_wq
= alloc_workqueue("events", 0, 0);
4889 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4890 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4891 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4892 WQ_UNBOUND_MAX_ACTIVE
);
4893 system_freezable_wq
= alloc_workqueue("events_freezable",
4895 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
4896 WQ_POWER_EFFICIENT
, 0);
4897 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
4898 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
4900 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
4901 !system_unbound_wq
|| !system_freezable_wq
||
4902 !system_power_efficient_wq
||
4903 !system_freezable_power_efficient_wq
);
4906 early_initcall(init_workqueues
);