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
870 * @wakeup: wakeup an idle worker if necessary
872 * Set @flags in @worker->flags and adjust nr_running accordingly. If
873 * nr_running becomes zero and @wakeup is %true, an idle worker is
877 * spin_lock_irq(pool->lock)
879 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
882 struct worker_pool
*pool
= worker
->pool
;
884 WARN_ON_ONCE(worker
->task
!= current
);
887 * If transitioning into NOT_RUNNING, adjust nr_running and
888 * wake up an idle worker as necessary if requested by
891 if ((flags
& WORKER_NOT_RUNNING
) &&
892 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
894 if (atomic_dec_and_test(&pool
->nr_running
) &&
895 !list_empty(&pool
->worklist
))
896 wake_up_worker(pool
);
898 atomic_dec(&pool
->nr_running
);
901 worker
->flags
|= flags
;
905 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
907 * @flags: flags to clear
909 * Clear @flags in @worker->flags and adjust nr_running accordingly.
912 * spin_lock_irq(pool->lock)
914 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
916 struct worker_pool
*pool
= worker
->pool
;
917 unsigned int oflags
= worker
->flags
;
919 WARN_ON_ONCE(worker
->task
!= current
);
921 worker
->flags
&= ~flags
;
924 * If transitioning out of NOT_RUNNING, increment nr_running. Note
925 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
926 * of multiple flags, not a single flag.
928 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
929 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
930 atomic_inc(&pool
->nr_running
);
934 * find_worker_executing_work - find worker which is executing a work
935 * @pool: pool of interest
936 * @work: work to find worker for
938 * Find a worker which is executing @work on @pool by searching
939 * @pool->busy_hash which is keyed by the address of @work. For a worker
940 * to match, its current execution should match the address of @work and
941 * its work function. This is to avoid unwanted dependency between
942 * unrelated work executions through a work item being recycled while still
945 * This is a bit tricky. A work item may be freed once its execution
946 * starts and nothing prevents the freed area from being recycled for
947 * another work item. If the same work item address ends up being reused
948 * before the original execution finishes, workqueue will identify the
949 * recycled work item as currently executing and make it wait until the
950 * current execution finishes, introducing an unwanted dependency.
952 * This function checks the work item address and work function to avoid
953 * false positives. Note that this isn't complete as one may construct a
954 * work function which can introduce dependency onto itself through a
955 * recycled work item. Well, if somebody wants to shoot oneself in the
956 * foot that badly, there's only so much we can do, and if such deadlock
957 * actually occurs, it should be easy to locate the culprit work function.
960 * spin_lock_irq(pool->lock).
963 * Pointer to worker which is executing @work if found, %NULL
966 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
967 struct work_struct
*work
)
969 struct worker
*worker
;
971 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
973 if (worker
->current_work
== work
&&
974 worker
->current_func
== work
->func
)
981 * move_linked_works - move linked works to a list
982 * @work: start of series of works to be scheduled
983 * @head: target list to append @work to
984 * @nextp: out paramter for nested worklist walking
986 * Schedule linked works starting from @work to @head. Work series to
987 * be scheduled starts at @work and includes any consecutive work with
988 * WORK_STRUCT_LINKED set in its predecessor.
990 * If @nextp is not NULL, it's updated to point to the next work of
991 * the last scheduled work. This allows move_linked_works() to be
992 * nested inside outer list_for_each_entry_safe().
995 * spin_lock_irq(pool->lock).
997 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
998 struct work_struct
**nextp
)
1000 struct work_struct
*n
;
1003 * Linked worklist will always end before the end of the list,
1004 * use NULL for list head.
1006 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1007 list_move_tail(&work
->entry
, head
);
1008 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1013 * If we're already inside safe list traversal and have moved
1014 * multiple works to the scheduled queue, the next position
1015 * needs to be updated.
1022 * get_pwq - get an extra reference on the specified pool_workqueue
1023 * @pwq: pool_workqueue to get
1025 * Obtain an extra reference on @pwq. The caller should guarantee that
1026 * @pwq has positive refcnt and be holding the matching pool->lock.
1028 static void get_pwq(struct pool_workqueue
*pwq
)
1030 lockdep_assert_held(&pwq
->pool
->lock
);
1031 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1036 * put_pwq - put a pool_workqueue reference
1037 * @pwq: pool_workqueue to put
1039 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1040 * destruction. The caller should be holding the matching pool->lock.
1042 static void put_pwq(struct pool_workqueue
*pwq
)
1044 lockdep_assert_held(&pwq
->pool
->lock
);
1045 if (likely(--pwq
->refcnt
))
1047 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1050 * @pwq can't be released under pool->lock, bounce to
1051 * pwq_unbound_release_workfn(). This never recurses on the same
1052 * pool->lock as this path is taken only for unbound workqueues and
1053 * the release work item is scheduled on a per-cpu workqueue. To
1054 * avoid lockdep warning, unbound pool->locks are given lockdep
1055 * subclass of 1 in get_unbound_pool().
1057 schedule_work(&pwq
->unbound_release_work
);
1061 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1062 * @pwq: pool_workqueue to put (can be %NULL)
1064 * put_pwq() with locking. This function also allows %NULL @pwq.
1066 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1070 * As both pwqs and pools are sched-RCU protected, the
1071 * following lock operations are safe.
1073 spin_lock_irq(&pwq
->pool
->lock
);
1075 spin_unlock_irq(&pwq
->pool
->lock
);
1079 static void pwq_activate_delayed_work(struct work_struct
*work
)
1081 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1083 trace_workqueue_activate_work(work
);
1084 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1085 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1089 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1091 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1092 struct work_struct
, entry
);
1094 pwq_activate_delayed_work(work
);
1098 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1099 * @pwq: pwq of interest
1100 * @color: color of work which left the queue
1102 * A work either has completed or is removed from pending queue,
1103 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1106 * spin_lock_irq(pool->lock).
1108 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1110 /* uncolored work items don't participate in flushing or nr_active */
1111 if (color
== WORK_NO_COLOR
)
1114 pwq
->nr_in_flight
[color
]--;
1117 if (!list_empty(&pwq
->delayed_works
)) {
1118 /* one down, submit a delayed one */
1119 if (pwq
->nr_active
< pwq
->max_active
)
1120 pwq_activate_first_delayed(pwq
);
1123 /* is flush in progress and are we at the flushing tip? */
1124 if (likely(pwq
->flush_color
!= color
))
1127 /* are there still in-flight works? */
1128 if (pwq
->nr_in_flight
[color
])
1131 /* this pwq is done, clear flush_color */
1132 pwq
->flush_color
= -1;
1135 * If this was the last pwq, wake up the first flusher. It
1136 * will handle the rest.
1138 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1139 complete(&pwq
->wq
->first_flusher
->done
);
1145 * try_to_grab_pending - steal work item from worklist and disable irq
1146 * @work: work item to steal
1147 * @is_dwork: @work is a delayed_work
1148 * @flags: place to store irq state
1150 * Try to grab PENDING bit of @work. This function can handle @work in any
1151 * stable state - idle, on timer or on worklist.
1154 * 1 if @work was pending and we successfully stole PENDING
1155 * 0 if @work was idle and we claimed PENDING
1156 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1157 * -ENOENT if someone else is canceling @work, this state may persist
1158 * for arbitrarily long
1161 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1162 * interrupted while holding PENDING and @work off queue, irq must be
1163 * disabled on entry. This, combined with delayed_work->timer being
1164 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1166 * On successful return, >= 0, irq is disabled and the caller is
1167 * responsible for releasing it using local_irq_restore(*@flags).
1169 * This function is safe to call from any context including IRQ handler.
1171 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1172 unsigned long *flags
)
1174 struct worker_pool
*pool
;
1175 struct pool_workqueue
*pwq
;
1177 local_irq_save(*flags
);
1179 /* try to steal the timer if it exists */
1181 struct delayed_work
*dwork
= to_delayed_work(work
);
1184 * dwork->timer is irqsafe. If del_timer() fails, it's
1185 * guaranteed that the timer is not queued anywhere and not
1186 * running on the local CPU.
1188 if (likely(del_timer(&dwork
->timer
)))
1192 /* try to claim PENDING the normal way */
1193 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1197 * The queueing is in progress, or it is already queued. Try to
1198 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1200 pool
= get_work_pool(work
);
1204 spin_lock(&pool
->lock
);
1206 * work->data is guaranteed to point to pwq only while the work
1207 * item is queued on pwq->wq, and both updating work->data to point
1208 * to pwq on queueing and to pool on dequeueing are done under
1209 * pwq->pool->lock. This in turn guarantees that, if work->data
1210 * points to pwq which is associated with a locked pool, the work
1211 * item is currently queued on that pool.
1213 pwq
= get_work_pwq(work
);
1214 if (pwq
&& pwq
->pool
== pool
) {
1215 debug_work_deactivate(work
);
1218 * A delayed work item cannot be grabbed directly because
1219 * it might have linked NO_COLOR work items which, if left
1220 * on the delayed_list, will confuse pwq->nr_active
1221 * management later on and cause stall. Make sure the work
1222 * item is activated before grabbing.
1224 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1225 pwq_activate_delayed_work(work
);
1227 list_del_init(&work
->entry
);
1228 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1230 /* work->data points to pwq iff queued, point to pool */
1231 set_work_pool_and_keep_pending(work
, pool
->id
);
1233 spin_unlock(&pool
->lock
);
1236 spin_unlock(&pool
->lock
);
1238 local_irq_restore(*flags
);
1239 if (work_is_canceling(work
))
1246 * insert_work - insert a work into a pool
1247 * @pwq: pwq @work belongs to
1248 * @work: work to insert
1249 * @head: insertion point
1250 * @extra_flags: extra WORK_STRUCT_* flags to set
1252 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1253 * work_struct flags.
1256 * spin_lock_irq(pool->lock).
1258 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1259 struct list_head
*head
, unsigned int extra_flags
)
1261 struct worker_pool
*pool
= pwq
->pool
;
1263 /* we own @work, set data and link */
1264 set_work_pwq(work
, pwq
, extra_flags
);
1265 list_add_tail(&work
->entry
, head
);
1269 * Ensure either wq_worker_sleeping() sees the above
1270 * list_add_tail() or we see zero nr_running to avoid workers lying
1271 * around lazily while there are works to be processed.
1275 if (__need_more_worker(pool
))
1276 wake_up_worker(pool
);
1280 * Test whether @work is being queued from another work executing on the
1283 static bool is_chained_work(struct workqueue_struct
*wq
)
1285 struct worker
*worker
;
1287 worker
= current_wq_worker();
1289 * Return %true iff I'm a worker execuing a work item on @wq. If
1290 * I'm @worker, it's safe to dereference it without locking.
1292 return worker
&& worker
->current_pwq
->wq
== wq
;
1295 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1296 struct work_struct
*work
)
1298 struct pool_workqueue
*pwq
;
1299 struct worker_pool
*last_pool
;
1300 struct list_head
*worklist
;
1301 unsigned int work_flags
;
1302 unsigned int req_cpu
= cpu
;
1305 * While a work item is PENDING && off queue, a task trying to
1306 * steal the PENDING will busy-loop waiting for it to either get
1307 * queued or lose PENDING. Grabbing PENDING and queueing should
1308 * happen with IRQ disabled.
1310 WARN_ON_ONCE(!irqs_disabled());
1312 debug_work_activate(work
);
1314 /* if draining, only works from the same workqueue are allowed */
1315 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1316 WARN_ON_ONCE(!is_chained_work(wq
)))
1319 if (req_cpu
== WORK_CPU_UNBOUND
)
1320 cpu
= raw_smp_processor_id();
1322 /* pwq which will be used unless @work is executing elsewhere */
1323 if (!(wq
->flags
& WQ_UNBOUND
))
1324 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1326 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1329 * If @work was previously on a different pool, it might still be
1330 * running there, in which case the work needs to be queued on that
1331 * pool to guarantee non-reentrancy.
1333 last_pool
= get_work_pool(work
);
1334 if (last_pool
&& last_pool
!= pwq
->pool
) {
1335 struct worker
*worker
;
1337 spin_lock(&last_pool
->lock
);
1339 worker
= find_worker_executing_work(last_pool
, work
);
1341 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1342 pwq
= worker
->current_pwq
;
1344 /* meh... not running there, queue here */
1345 spin_unlock(&last_pool
->lock
);
1346 spin_lock(&pwq
->pool
->lock
);
1349 spin_lock(&pwq
->pool
->lock
);
1353 * pwq is determined and locked. For unbound pools, we could have
1354 * raced with pwq release and it could already be dead. If its
1355 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1356 * without another pwq replacing it in the numa_pwq_tbl or while
1357 * work items are executing on it, so the retrying is guaranteed to
1358 * make forward-progress.
1360 if (unlikely(!pwq
->refcnt
)) {
1361 if (wq
->flags
& WQ_UNBOUND
) {
1362 spin_unlock(&pwq
->pool
->lock
);
1367 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1371 /* pwq determined, queue */
1372 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1374 if (WARN_ON(!list_empty(&work
->entry
))) {
1375 spin_unlock(&pwq
->pool
->lock
);
1379 pwq
->nr_in_flight
[pwq
->work_color
]++;
1380 work_flags
= work_color_to_flags(pwq
->work_color
);
1382 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1383 trace_workqueue_activate_work(work
);
1385 worklist
= &pwq
->pool
->worklist
;
1387 work_flags
|= WORK_STRUCT_DELAYED
;
1388 worklist
= &pwq
->delayed_works
;
1391 insert_work(pwq
, work
, worklist
, work_flags
);
1393 spin_unlock(&pwq
->pool
->lock
);
1397 * queue_work_on - queue work on specific cpu
1398 * @cpu: CPU number to execute work on
1399 * @wq: workqueue to use
1400 * @work: work to queue
1402 * We queue the work to a specific CPU, the caller must ensure it
1405 * Return: %false if @work was already on a queue, %true otherwise.
1407 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1408 struct work_struct
*work
)
1411 unsigned long flags
;
1413 local_irq_save(flags
);
1415 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1416 __queue_work(cpu
, wq
, work
);
1420 local_irq_restore(flags
);
1423 EXPORT_SYMBOL(queue_work_on
);
1425 void delayed_work_timer_fn(unsigned long __data
)
1427 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1429 /* should have been called from irqsafe timer with irq already off */
1430 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1432 EXPORT_SYMBOL(delayed_work_timer_fn
);
1434 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1435 struct delayed_work
*dwork
, unsigned long delay
)
1437 struct timer_list
*timer
= &dwork
->timer
;
1438 struct work_struct
*work
= &dwork
->work
;
1440 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1441 timer
->data
!= (unsigned long)dwork
);
1442 WARN_ON_ONCE(timer_pending(timer
));
1443 WARN_ON_ONCE(!list_empty(&work
->entry
));
1446 * If @delay is 0, queue @dwork->work immediately. This is for
1447 * both optimization and correctness. The earliest @timer can
1448 * expire is on the closest next tick and delayed_work users depend
1449 * on that there's no such delay when @delay is 0.
1452 __queue_work(cpu
, wq
, &dwork
->work
);
1456 timer_stats_timer_set_start_info(&dwork
->timer
);
1460 timer
->expires
= jiffies
+ delay
;
1462 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1463 add_timer_on(timer
, cpu
);
1469 * queue_delayed_work_on - queue work on specific CPU after delay
1470 * @cpu: CPU number to execute work on
1471 * @wq: workqueue to use
1472 * @dwork: work to queue
1473 * @delay: number of jiffies to wait before queueing
1475 * Return: %false if @work was already on a queue, %true otherwise. If
1476 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1479 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1480 struct delayed_work
*dwork
, unsigned long delay
)
1482 struct work_struct
*work
= &dwork
->work
;
1484 unsigned long flags
;
1486 /* read the comment in __queue_work() */
1487 local_irq_save(flags
);
1489 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1490 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1494 local_irq_restore(flags
);
1497 EXPORT_SYMBOL(queue_delayed_work_on
);
1500 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1501 * @cpu: CPU number to execute work on
1502 * @wq: workqueue to use
1503 * @dwork: work to queue
1504 * @delay: number of jiffies to wait before queueing
1506 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1507 * modify @dwork's timer so that it expires after @delay. If @delay is
1508 * zero, @work is guaranteed to be scheduled immediately regardless of its
1511 * Return: %false if @dwork was idle and queued, %true if @dwork was
1512 * pending and its timer was modified.
1514 * This function is safe to call from any context including IRQ handler.
1515 * See try_to_grab_pending() for details.
1517 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1518 struct delayed_work
*dwork
, unsigned long delay
)
1520 unsigned long flags
;
1524 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1525 } while (unlikely(ret
== -EAGAIN
));
1527 if (likely(ret
>= 0)) {
1528 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1529 local_irq_restore(flags
);
1532 /* -ENOENT from try_to_grab_pending() becomes %true */
1535 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1538 * worker_enter_idle - enter idle state
1539 * @worker: worker which is entering idle state
1541 * @worker is entering idle state. Update stats and idle timer if
1545 * spin_lock_irq(pool->lock).
1547 static void worker_enter_idle(struct worker
*worker
)
1549 struct worker_pool
*pool
= worker
->pool
;
1551 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1552 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1553 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1556 /* can't use worker_set_flags(), also called from start_worker() */
1557 worker
->flags
|= WORKER_IDLE
;
1559 worker
->last_active
= jiffies
;
1561 /* idle_list is LIFO */
1562 list_add(&worker
->entry
, &pool
->idle_list
);
1564 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1565 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1568 * Sanity check nr_running. Because wq_unbind_fn() releases
1569 * pool->lock between setting %WORKER_UNBOUND and zapping
1570 * nr_running, the warning may trigger spuriously. Check iff
1571 * unbind is not in progress.
1573 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1574 pool
->nr_workers
== pool
->nr_idle
&&
1575 atomic_read(&pool
->nr_running
));
1579 * worker_leave_idle - leave idle state
1580 * @worker: worker which is leaving idle state
1582 * @worker is leaving idle state. Update stats.
1585 * spin_lock_irq(pool->lock).
1587 static void worker_leave_idle(struct worker
*worker
)
1589 struct worker_pool
*pool
= worker
->pool
;
1591 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1593 worker_clr_flags(worker
, WORKER_IDLE
);
1595 list_del_init(&worker
->entry
);
1598 static struct worker
*alloc_worker(void)
1600 struct worker
*worker
;
1602 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1604 INIT_LIST_HEAD(&worker
->entry
);
1605 INIT_LIST_HEAD(&worker
->scheduled
);
1606 INIT_LIST_HEAD(&worker
->node
);
1607 /* on creation a worker is in !idle && prep state */
1608 worker
->flags
= WORKER_PREP
;
1614 * worker_attach_to_pool() - attach a worker to a pool
1615 * @worker: worker to be attached
1616 * @pool: the target pool
1618 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1619 * cpu-binding of @worker are kept coordinated with the pool across
1622 static void worker_attach_to_pool(struct worker
*worker
,
1623 struct worker_pool
*pool
)
1625 mutex_lock(&pool
->attach_mutex
);
1628 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1629 * online CPUs. It'll be re-applied when any of the CPUs come up.
1631 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1634 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1635 * stable across this function. See the comments above the
1636 * flag definition for details.
1638 if (pool
->flags
& POOL_DISASSOCIATED
)
1639 worker
->flags
|= WORKER_UNBOUND
;
1641 list_add_tail(&worker
->node
, &pool
->workers
);
1643 mutex_unlock(&pool
->attach_mutex
);
1647 * worker_detach_from_pool() - detach a worker from its pool
1648 * @worker: worker which is attached to its pool
1649 * @pool: the pool @worker is attached to
1651 * Undo the attaching which had been done in worker_attach_to_pool(). The
1652 * caller worker shouldn't access to the pool after detached except it has
1653 * other reference to the pool.
1655 static void worker_detach_from_pool(struct worker
*worker
,
1656 struct worker_pool
*pool
)
1658 struct completion
*detach_completion
= NULL
;
1660 mutex_lock(&pool
->attach_mutex
);
1661 list_del(&worker
->node
);
1662 if (list_empty(&pool
->workers
))
1663 detach_completion
= pool
->detach_completion
;
1664 mutex_unlock(&pool
->attach_mutex
);
1666 /* clear leftover flags without pool->lock after it is detached */
1667 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1669 if (detach_completion
)
1670 complete(detach_completion
);
1674 * create_worker - create a new workqueue worker
1675 * @pool: pool the new worker will belong to
1677 * Create a new worker which is attached to @pool. The new worker must be
1678 * started by start_worker().
1681 * Might sleep. Does GFP_KERNEL allocations.
1684 * Pointer to the newly created worker.
1686 static struct worker
*create_worker(struct worker_pool
*pool
)
1688 struct worker
*worker
= NULL
;
1692 /* ID is needed to determine kthread name */
1693 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1697 worker
= alloc_worker();
1701 worker
->pool
= pool
;
1705 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1706 pool
->attrs
->nice
< 0 ? "H" : "");
1708 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1710 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1711 "kworker/%s", id_buf
);
1712 if (IS_ERR(worker
->task
))
1715 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1717 /* prevent userland from meddling with cpumask of workqueue workers */
1718 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1720 /* successful, attach the worker to the pool */
1721 worker_attach_to_pool(worker
, pool
);
1727 ida_simple_remove(&pool
->worker_ida
, id
);
1733 * start_worker - start a newly created worker
1734 * @worker: worker to start
1736 * Make the pool aware of @worker and start it.
1739 * spin_lock_irq(pool->lock).
1741 static void start_worker(struct worker
*worker
)
1743 worker
->pool
->nr_workers
++;
1744 worker_enter_idle(worker
);
1745 wake_up_process(worker
->task
);
1749 * create_and_start_worker - create and start a worker for a pool
1750 * @pool: the target pool
1752 * Grab the managership of @pool and create and start a new worker for it.
1754 * Return: 0 on success. A negative error code otherwise.
1756 static int create_and_start_worker(struct worker_pool
*pool
)
1758 struct worker
*worker
;
1760 worker
= create_worker(pool
);
1762 spin_lock_irq(&pool
->lock
);
1763 start_worker(worker
);
1764 spin_unlock_irq(&pool
->lock
);
1767 return worker
? 0 : -ENOMEM
;
1771 * destroy_worker - destroy a workqueue worker
1772 * @worker: worker to be destroyed
1774 * Destroy @worker and adjust @pool stats accordingly. The worker should
1778 * spin_lock_irq(pool->lock).
1780 static void destroy_worker(struct worker
*worker
)
1782 struct worker_pool
*pool
= worker
->pool
;
1784 lockdep_assert_held(&pool
->lock
);
1786 /* sanity check frenzy */
1787 if (WARN_ON(worker
->current_work
) ||
1788 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1789 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1795 list_del_init(&worker
->entry
);
1796 worker
->flags
|= WORKER_DIE
;
1797 wake_up_process(worker
->task
);
1800 static void idle_worker_timeout(unsigned long __pool
)
1802 struct worker_pool
*pool
= (void *)__pool
;
1804 spin_lock_irq(&pool
->lock
);
1806 while (too_many_workers(pool
)) {
1807 struct worker
*worker
;
1808 unsigned long expires
;
1810 /* idle_list is kept in LIFO order, check the last one */
1811 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1812 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1814 if (time_before(jiffies
, expires
)) {
1815 mod_timer(&pool
->idle_timer
, expires
);
1819 destroy_worker(worker
);
1822 spin_unlock_irq(&pool
->lock
);
1825 static void send_mayday(struct work_struct
*work
)
1827 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1828 struct workqueue_struct
*wq
= pwq
->wq
;
1830 lockdep_assert_held(&wq_mayday_lock
);
1835 /* mayday mayday mayday */
1836 if (list_empty(&pwq
->mayday_node
)) {
1838 * If @pwq is for an unbound wq, its base ref may be put at
1839 * any time due to an attribute change. Pin @pwq until the
1840 * rescuer is done with it.
1843 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1844 wake_up_process(wq
->rescuer
->task
);
1848 static void pool_mayday_timeout(unsigned long __pool
)
1850 struct worker_pool
*pool
= (void *)__pool
;
1851 struct work_struct
*work
;
1853 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1854 spin_lock(&pool
->lock
);
1856 if (need_to_create_worker(pool
)) {
1858 * We've been trying to create a new worker but
1859 * haven't been successful. We might be hitting an
1860 * allocation deadlock. Send distress signals to
1863 list_for_each_entry(work
, &pool
->worklist
, entry
)
1867 spin_unlock(&pool
->lock
);
1868 spin_unlock_irq(&wq_mayday_lock
);
1870 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1874 * maybe_create_worker - create a new worker if necessary
1875 * @pool: pool to create a new worker for
1877 * Create a new worker for @pool if necessary. @pool is guaranteed to
1878 * have at least one idle worker on return from this function. If
1879 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1880 * sent to all rescuers with works scheduled on @pool to resolve
1881 * possible allocation deadlock.
1883 * On return, need_to_create_worker() is guaranteed to be %false and
1884 * may_start_working() %true.
1887 * spin_lock_irq(pool->lock) which may be released and regrabbed
1888 * multiple times. Does GFP_KERNEL allocations. Called only from
1892 * %false if no action was taken and pool->lock stayed locked, %true
1895 static bool maybe_create_worker(struct worker_pool
*pool
)
1896 __releases(&pool
->lock
)
1897 __acquires(&pool
->lock
)
1899 if (!need_to_create_worker(pool
))
1902 spin_unlock_irq(&pool
->lock
);
1904 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1905 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1908 struct worker
*worker
;
1910 worker
= create_worker(pool
);
1912 del_timer_sync(&pool
->mayday_timer
);
1913 spin_lock_irq(&pool
->lock
);
1914 start_worker(worker
);
1915 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1920 if (!need_to_create_worker(pool
))
1923 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1925 if (!need_to_create_worker(pool
))
1929 del_timer_sync(&pool
->mayday_timer
);
1930 spin_lock_irq(&pool
->lock
);
1931 if (need_to_create_worker(pool
))
1937 * manage_workers - manage worker pool
1940 * Assume the manager role and manage the worker pool @worker belongs
1941 * to. At any given time, there can be only zero or one manager per
1942 * pool. The exclusion is handled automatically by this function.
1944 * The caller can safely start processing works on false return. On
1945 * true return, it's guaranteed that need_to_create_worker() is false
1946 * and may_start_working() is true.
1949 * spin_lock_irq(pool->lock) which may be released and regrabbed
1950 * multiple times. Does GFP_KERNEL allocations.
1953 * %false if the pool don't need management and the caller can safely start
1954 * processing works, %true indicates that the function released pool->lock
1955 * and reacquired it to perform some management function and that the
1956 * conditions that the caller verified while holding the lock before
1957 * calling the function might no longer be true.
1959 static bool manage_workers(struct worker
*worker
)
1961 struct worker_pool
*pool
= worker
->pool
;
1965 * Anyone who successfully grabs manager_arb wins the arbitration
1966 * and becomes the manager. mutex_trylock() on pool->manager_arb
1967 * failure while holding pool->lock reliably indicates that someone
1968 * else is managing the pool and the worker which failed trylock
1969 * can proceed to executing work items. This means that anyone
1970 * grabbing manager_arb is responsible for actually performing
1971 * manager duties. If manager_arb is grabbed and released without
1972 * actual management, the pool may stall indefinitely.
1974 if (!mutex_trylock(&pool
->manager_arb
))
1977 ret
|= maybe_create_worker(pool
);
1979 mutex_unlock(&pool
->manager_arb
);
1984 * process_one_work - process single work
1986 * @work: work to process
1988 * Process @work. This function contains all the logics necessary to
1989 * process a single work including synchronization against and
1990 * interaction with other workers on the same cpu, queueing and
1991 * flushing. As long as context requirement is met, any worker can
1992 * call this function to process a work.
1995 * spin_lock_irq(pool->lock) which is released and regrabbed.
1997 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1998 __releases(&pool
->lock
)
1999 __acquires(&pool
->lock
)
2001 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2002 struct worker_pool
*pool
= worker
->pool
;
2003 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2005 struct worker
*collision
;
2006 #ifdef CONFIG_LOCKDEP
2008 * It is permissible to free the struct work_struct from
2009 * inside the function that is called from it, this we need to
2010 * take into account for lockdep too. To avoid bogus "held
2011 * lock freed" warnings as well as problems when looking into
2012 * work->lockdep_map, make a copy and use that here.
2014 struct lockdep_map lockdep_map
;
2016 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2019 * Ensure we're on the correct CPU. DISASSOCIATED test is
2020 * necessary to avoid spurious warnings from rescuers servicing the
2021 * unbound or a disassociated pool.
2023 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2024 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2025 raw_smp_processor_id() != pool
->cpu
);
2028 * A single work shouldn't be executed concurrently by
2029 * multiple workers on a single cpu. Check whether anyone is
2030 * already processing the work. If so, defer the work to the
2031 * currently executing one.
2033 collision
= find_worker_executing_work(pool
, work
);
2034 if (unlikely(collision
)) {
2035 move_linked_works(work
, &collision
->scheduled
, NULL
);
2039 /* claim and dequeue */
2040 debug_work_deactivate(work
);
2041 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2042 worker
->current_work
= work
;
2043 worker
->current_func
= work
->func
;
2044 worker
->current_pwq
= pwq
;
2045 work_color
= get_work_color(work
);
2047 list_del_init(&work
->entry
);
2050 * CPU intensive works don't participate in concurrency
2051 * management. They're the scheduler's responsibility.
2053 if (unlikely(cpu_intensive
))
2054 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2057 * Unbound pool isn't concurrency managed and work items should be
2058 * executed ASAP. Wake up another worker if necessary.
2060 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2061 wake_up_worker(pool
);
2064 * Record the last pool and clear PENDING which should be the last
2065 * update to @work. Also, do this inside @pool->lock so that
2066 * PENDING and queued state changes happen together while IRQ is
2069 set_work_pool_and_clear_pending(work
, pool
->id
);
2071 spin_unlock_irq(&pool
->lock
);
2073 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2074 lock_map_acquire(&lockdep_map
);
2075 trace_workqueue_execute_start(work
);
2076 worker
->current_func(work
);
2078 * While we must be careful to not use "work" after this, the trace
2079 * point will only record its address.
2081 trace_workqueue_execute_end(work
);
2082 lock_map_release(&lockdep_map
);
2083 lock_map_release(&pwq
->wq
->lockdep_map
);
2085 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2086 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2087 " last function: %pf\n",
2088 current
->comm
, preempt_count(), task_pid_nr(current
),
2089 worker
->current_func
);
2090 debug_show_held_locks(current
);
2095 * The following prevents a kworker from hogging CPU on !PREEMPT
2096 * kernels, where a requeueing work item waiting for something to
2097 * happen could deadlock with stop_machine as such work item could
2098 * indefinitely requeue itself while all other CPUs are trapped in
2103 spin_lock_irq(&pool
->lock
);
2105 /* clear cpu intensive status */
2106 if (unlikely(cpu_intensive
))
2107 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2109 /* we're done with it, release */
2110 hash_del(&worker
->hentry
);
2111 worker
->current_work
= NULL
;
2112 worker
->current_func
= NULL
;
2113 worker
->current_pwq
= NULL
;
2114 worker
->desc_valid
= false;
2115 pwq_dec_nr_in_flight(pwq
, work_color
);
2119 * process_scheduled_works - process scheduled works
2122 * Process all scheduled works. Please note that the scheduled list
2123 * may change while processing a work, so this function repeatedly
2124 * fetches a work from the top and executes it.
2127 * spin_lock_irq(pool->lock) which may be released and regrabbed
2130 static void process_scheduled_works(struct worker
*worker
)
2132 while (!list_empty(&worker
->scheduled
)) {
2133 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2134 struct work_struct
, entry
);
2135 process_one_work(worker
, work
);
2140 * worker_thread - the worker thread function
2143 * The worker thread function. All workers belong to a worker_pool -
2144 * either a per-cpu one or dynamic unbound one. These workers process all
2145 * work items regardless of their specific target workqueue. The only
2146 * exception is work items which belong to workqueues with a rescuer which
2147 * will be explained in rescuer_thread().
2151 static int worker_thread(void *__worker
)
2153 struct worker
*worker
= __worker
;
2154 struct worker_pool
*pool
= worker
->pool
;
2156 /* tell the scheduler that this is a workqueue worker */
2157 worker
->task
->flags
|= PF_WQ_WORKER
;
2159 spin_lock_irq(&pool
->lock
);
2161 /* am I supposed to die? */
2162 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2163 spin_unlock_irq(&pool
->lock
);
2164 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2165 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2167 set_task_comm(worker
->task
, "kworker/dying");
2168 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2169 worker_detach_from_pool(worker
, pool
);
2174 worker_leave_idle(worker
);
2176 /* no more worker necessary? */
2177 if (!need_more_worker(pool
))
2180 /* do we need to manage? */
2181 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2185 * ->scheduled list can only be filled while a worker is
2186 * preparing to process a work or actually processing it.
2187 * Make sure nobody diddled with it while I was sleeping.
2189 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2192 * Finish PREP stage. We're guaranteed to have at least one idle
2193 * worker or that someone else has already assumed the manager
2194 * role. This is where @worker starts participating in concurrency
2195 * management if applicable and concurrency management is restored
2196 * after being rebound. See rebind_workers() for details.
2198 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2201 struct work_struct
*work
=
2202 list_first_entry(&pool
->worklist
,
2203 struct work_struct
, entry
);
2205 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2206 /* optimization path, not strictly necessary */
2207 process_one_work(worker
, work
);
2208 if (unlikely(!list_empty(&worker
->scheduled
)))
2209 process_scheduled_works(worker
);
2211 move_linked_works(work
, &worker
->scheduled
, NULL
);
2212 process_scheduled_works(worker
);
2214 } while (keep_working(pool
));
2216 worker_set_flags(worker
, WORKER_PREP
, false);
2219 * pool->lock is held and there's no work to process and no need to
2220 * manage, sleep. Workers are woken up only while holding
2221 * pool->lock or from local cpu, so setting the current state
2222 * before releasing pool->lock is enough to prevent losing any
2225 worker_enter_idle(worker
);
2226 __set_current_state(TASK_INTERRUPTIBLE
);
2227 spin_unlock_irq(&pool
->lock
);
2233 * rescuer_thread - the rescuer thread function
2236 * Workqueue rescuer thread function. There's one rescuer for each
2237 * workqueue which has WQ_MEM_RECLAIM set.
2239 * Regular work processing on a pool may block trying to create a new
2240 * worker which uses GFP_KERNEL allocation which has slight chance of
2241 * developing into deadlock if some works currently on the same queue
2242 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2243 * the problem rescuer solves.
2245 * When such condition is possible, the pool summons rescuers of all
2246 * workqueues which have works queued on the pool and let them process
2247 * those works so that forward progress can be guaranteed.
2249 * This should happen rarely.
2253 static int rescuer_thread(void *__rescuer
)
2255 struct worker
*rescuer
= __rescuer
;
2256 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2257 struct list_head
*scheduled
= &rescuer
->scheduled
;
2260 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2263 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2264 * doesn't participate in concurrency management.
2266 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2268 set_current_state(TASK_INTERRUPTIBLE
);
2271 * By the time the rescuer is requested to stop, the workqueue
2272 * shouldn't have any work pending, but @wq->maydays may still have
2273 * pwq(s) queued. This can happen by non-rescuer workers consuming
2274 * all the work items before the rescuer got to them. Go through
2275 * @wq->maydays processing before acting on should_stop so that the
2276 * list is always empty on exit.
2278 should_stop
= kthread_should_stop();
2280 /* see whether any pwq is asking for help */
2281 spin_lock_irq(&wq_mayday_lock
);
2283 while (!list_empty(&wq
->maydays
)) {
2284 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2285 struct pool_workqueue
, mayday_node
);
2286 struct worker_pool
*pool
= pwq
->pool
;
2287 struct work_struct
*work
, *n
;
2289 __set_current_state(TASK_RUNNING
);
2290 list_del_init(&pwq
->mayday_node
);
2292 spin_unlock_irq(&wq_mayday_lock
);
2294 worker_attach_to_pool(rescuer
, pool
);
2296 spin_lock_irq(&pool
->lock
);
2297 rescuer
->pool
= pool
;
2300 * Slurp in all works issued via this workqueue and
2303 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2304 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2305 if (get_work_pwq(work
) == pwq
)
2306 move_linked_works(work
, scheduled
, &n
);
2308 process_scheduled_works(rescuer
);
2309 spin_unlock_irq(&pool
->lock
);
2311 worker_detach_from_pool(rescuer
, pool
);
2313 spin_lock_irq(&pool
->lock
);
2316 * Put the reference grabbed by send_mayday(). @pool won't
2317 * go away while we're holding its lock.
2322 * Leave this pool. If keep_working() is %true, notify a
2323 * regular worker; otherwise, we end up with 0 concurrency
2324 * and stalling the execution.
2326 if (keep_working(pool
))
2327 wake_up_worker(pool
);
2329 rescuer
->pool
= NULL
;
2330 spin_unlock(&pool
->lock
);
2331 spin_lock(&wq_mayday_lock
);
2334 spin_unlock_irq(&wq_mayday_lock
);
2337 __set_current_state(TASK_RUNNING
);
2338 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2342 /* rescuers should never participate in concurrency management */
2343 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2349 struct work_struct work
;
2350 struct completion done
;
2353 static void wq_barrier_func(struct work_struct
*work
)
2355 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2356 complete(&barr
->done
);
2360 * insert_wq_barrier - insert a barrier work
2361 * @pwq: pwq to insert barrier into
2362 * @barr: wq_barrier to insert
2363 * @target: target work to attach @barr to
2364 * @worker: worker currently executing @target, NULL if @target is not executing
2366 * @barr is linked to @target such that @barr is completed only after
2367 * @target finishes execution. Please note that the ordering
2368 * guarantee is observed only with respect to @target and on the local
2371 * Currently, a queued barrier can't be canceled. This is because
2372 * try_to_grab_pending() can't determine whether the work to be
2373 * grabbed is at the head of the queue and thus can't clear LINKED
2374 * flag of the previous work while there must be a valid next work
2375 * after a work with LINKED flag set.
2377 * Note that when @worker is non-NULL, @target may be modified
2378 * underneath us, so we can't reliably determine pwq from @target.
2381 * spin_lock_irq(pool->lock).
2383 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2384 struct wq_barrier
*barr
,
2385 struct work_struct
*target
, struct worker
*worker
)
2387 struct list_head
*head
;
2388 unsigned int linked
= 0;
2391 * debugobject calls are safe here even with pool->lock locked
2392 * as we know for sure that this will not trigger any of the
2393 * checks and call back into the fixup functions where we
2396 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2397 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2398 init_completion(&barr
->done
);
2401 * If @target is currently being executed, schedule the
2402 * barrier to the worker; otherwise, put it after @target.
2405 head
= worker
->scheduled
.next
;
2407 unsigned long *bits
= work_data_bits(target
);
2409 head
= target
->entry
.next
;
2410 /* there can already be other linked works, inherit and set */
2411 linked
= *bits
& WORK_STRUCT_LINKED
;
2412 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2415 debug_work_activate(&barr
->work
);
2416 insert_work(pwq
, &barr
->work
, head
,
2417 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2421 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2422 * @wq: workqueue being flushed
2423 * @flush_color: new flush color, < 0 for no-op
2424 * @work_color: new work color, < 0 for no-op
2426 * Prepare pwqs for workqueue flushing.
2428 * If @flush_color is non-negative, flush_color on all pwqs should be
2429 * -1. If no pwq has in-flight commands at the specified color, all
2430 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2431 * has in flight commands, its pwq->flush_color is set to
2432 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2433 * wakeup logic is armed and %true is returned.
2435 * The caller should have initialized @wq->first_flusher prior to
2436 * calling this function with non-negative @flush_color. If
2437 * @flush_color is negative, no flush color update is done and %false
2440 * If @work_color is non-negative, all pwqs should have the same
2441 * work_color which is previous to @work_color and all will be
2442 * advanced to @work_color.
2445 * mutex_lock(wq->mutex).
2448 * %true if @flush_color >= 0 and there's something to flush. %false
2451 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2452 int flush_color
, int work_color
)
2455 struct pool_workqueue
*pwq
;
2457 if (flush_color
>= 0) {
2458 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2459 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2462 for_each_pwq(pwq
, wq
) {
2463 struct worker_pool
*pool
= pwq
->pool
;
2465 spin_lock_irq(&pool
->lock
);
2467 if (flush_color
>= 0) {
2468 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2470 if (pwq
->nr_in_flight
[flush_color
]) {
2471 pwq
->flush_color
= flush_color
;
2472 atomic_inc(&wq
->nr_pwqs_to_flush
);
2477 if (work_color
>= 0) {
2478 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2479 pwq
->work_color
= work_color
;
2482 spin_unlock_irq(&pool
->lock
);
2485 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2486 complete(&wq
->first_flusher
->done
);
2492 * flush_workqueue - ensure that any scheduled work has run to completion.
2493 * @wq: workqueue to flush
2495 * This function sleeps until all work items which were queued on entry
2496 * have finished execution, but it is not livelocked by new incoming ones.
2498 void flush_workqueue(struct workqueue_struct
*wq
)
2500 struct wq_flusher this_flusher
= {
2501 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2503 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2507 lock_map_acquire(&wq
->lockdep_map
);
2508 lock_map_release(&wq
->lockdep_map
);
2510 mutex_lock(&wq
->mutex
);
2513 * Start-to-wait phase
2515 next_color
= work_next_color(wq
->work_color
);
2517 if (next_color
!= wq
->flush_color
) {
2519 * Color space is not full. The current work_color
2520 * becomes our flush_color and work_color is advanced
2523 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2524 this_flusher
.flush_color
= wq
->work_color
;
2525 wq
->work_color
= next_color
;
2527 if (!wq
->first_flusher
) {
2528 /* no flush in progress, become the first flusher */
2529 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2531 wq
->first_flusher
= &this_flusher
;
2533 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2535 /* nothing to flush, done */
2536 wq
->flush_color
= next_color
;
2537 wq
->first_flusher
= NULL
;
2542 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2543 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2544 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2548 * Oops, color space is full, wait on overflow queue.
2549 * The next flush completion will assign us
2550 * flush_color and transfer to flusher_queue.
2552 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2555 mutex_unlock(&wq
->mutex
);
2557 wait_for_completion(&this_flusher
.done
);
2560 * Wake-up-and-cascade phase
2562 * First flushers are responsible for cascading flushes and
2563 * handling overflow. Non-first flushers can simply return.
2565 if (wq
->first_flusher
!= &this_flusher
)
2568 mutex_lock(&wq
->mutex
);
2570 /* we might have raced, check again with mutex held */
2571 if (wq
->first_flusher
!= &this_flusher
)
2574 wq
->first_flusher
= NULL
;
2576 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2577 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2580 struct wq_flusher
*next
, *tmp
;
2582 /* complete all the flushers sharing the current flush color */
2583 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2584 if (next
->flush_color
!= wq
->flush_color
)
2586 list_del_init(&next
->list
);
2587 complete(&next
->done
);
2590 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2591 wq
->flush_color
!= work_next_color(wq
->work_color
));
2593 /* this flush_color is finished, advance by one */
2594 wq
->flush_color
= work_next_color(wq
->flush_color
);
2596 /* one color has been freed, handle overflow queue */
2597 if (!list_empty(&wq
->flusher_overflow
)) {
2599 * Assign the same color to all overflowed
2600 * flushers, advance work_color and append to
2601 * flusher_queue. This is the start-to-wait
2602 * phase for these overflowed flushers.
2604 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2605 tmp
->flush_color
= wq
->work_color
;
2607 wq
->work_color
= work_next_color(wq
->work_color
);
2609 list_splice_tail_init(&wq
->flusher_overflow
,
2610 &wq
->flusher_queue
);
2611 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2614 if (list_empty(&wq
->flusher_queue
)) {
2615 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2620 * Need to flush more colors. Make the next flusher
2621 * the new first flusher and arm pwqs.
2623 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2624 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2626 list_del_init(&next
->list
);
2627 wq
->first_flusher
= next
;
2629 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2633 * Meh... this color is already done, clear first
2634 * flusher and repeat cascading.
2636 wq
->first_flusher
= NULL
;
2640 mutex_unlock(&wq
->mutex
);
2642 EXPORT_SYMBOL_GPL(flush_workqueue
);
2645 * drain_workqueue - drain a workqueue
2646 * @wq: workqueue to drain
2648 * Wait until the workqueue becomes empty. While draining is in progress,
2649 * only chain queueing is allowed. IOW, only currently pending or running
2650 * work items on @wq can queue further work items on it. @wq is flushed
2651 * repeatedly until it becomes empty. The number of flushing is detemined
2652 * by the depth of chaining and should be relatively short. Whine if it
2655 void drain_workqueue(struct workqueue_struct
*wq
)
2657 unsigned int flush_cnt
= 0;
2658 struct pool_workqueue
*pwq
;
2661 * __queue_work() needs to test whether there are drainers, is much
2662 * hotter than drain_workqueue() and already looks at @wq->flags.
2663 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2665 mutex_lock(&wq
->mutex
);
2666 if (!wq
->nr_drainers
++)
2667 wq
->flags
|= __WQ_DRAINING
;
2668 mutex_unlock(&wq
->mutex
);
2670 flush_workqueue(wq
);
2672 mutex_lock(&wq
->mutex
);
2674 for_each_pwq(pwq
, wq
) {
2677 spin_lock_irq(&pwq
->pool
->lock
);
2678 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2679 spin_unlock_irq(&pwq
->pool
->lock
);
2684 if (++flush_cnt
== 10 ||
2685 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2686 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2687 wq
->name
, flush_cnt
);
2689 mutex_unlock(&wq
->mutex
);
2693 if (!--wq
->nr_drainers
)
2694 wq
->flags
&= ~__WQ_DRAINING
;
2695 mutex_unlock(&wq
->mutex
);
2697 EXPORT_SYMBOL_GPL(drain_workqueue
);
2699 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2701 struct worker
*worker
= NULL
;
2702 struct worker_pool
*pool
;
2703 struct pool_workqueue
*pwq
;
2707 local_irq_disable();
2708 pool
= get_work_pool(work
);
2714 spin_lock(&pool
->lock
);
2715 /* see the comment in try_to_grab_pending() with the same code */
2716 pwq
= get_work_pwq(work
);
2718 if (unlikely(pwq
->pool
!= pool
))
2721 worker
= find_worker_executing_work(pool
, work
);
2724 pwq
= worker
->current_pwq
;
2727 insert_wq_barrier(pwq
, barr
, work
, worker
);
2728 spin_unlock_irq(&pool
->lock
);
2731 * If @max_active is 1 or rescuer is in use, flushing another work
2732 * item on the same workqueue may lead to deadlock. Make sure the
2733 * flusher is not running on the same workqueue by verifying write
2736 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2737 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2739 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2740 lock_map_release(&pwq
->wq
->lockdep_map
);
2744 spin_unlock_irq(&pool
->lock
);
2749 * flush_work - wait for a work to finish executing the last queueing instance
2750 * @work: the work to flush
2752 * Wait until @work has finished execution. @work is guaranteed to be idle
2753 * on return if it hasn't been requeued since flush started.
2756 * %true if flush_work() waited for the work to finish execution,
2757 * %false if it was already idle.
2759 bool flush_work(struct work_struct
*work
)
2761 struct wq_barrier barr
;
2763 lock_map_acquire(&work
->lockdep_map
);
2764 lock_map_release(&work
->lockdep_map
);
2766 if (start_flush_work(work
, &barr
)) {
2767 wait_for_completion(&barr
.done
);
2768 destroy_work_on_stack(&barr
.work
);
2774 EXPORT_SYMBOL_GPL(flush_work
);
2776 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2778 unsigned long flags
;
2782 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2784 * If someone else is canceling, wait for the same event it
2785 * would be waiting for before retrying.
2787 if (unlikely(ret
== -ENOENT
))
2789 } while (unlikely(ret
< 0));
2791 /* tell other tasks trying to grab @work to back off */
2792 mark_work_canceling(work
);
2793 local_irq_restore(flags
);
2796 clear_work_data(work
);
2801 * cancel_work_sync - cancel a work and wait for it to finish
2802 * @work: the work to cancel
2804 * Cancel @work and wait for its execution to finish. This function
2805 * can be used even if the work re-queues itself or migrates to
2806 * another workqueue. On return from this function, @work is
2807 * guaranteed to be not pending or executing on any CPU.
2809 * cancel_work_sync(&delayed_work->work) must not be used for
2810 * delayed_work's. Use cancel_delayed_work_sync() instead.
2812 * The caller must ensure that the workqueue on which @work was last
2813 * queued can't be destroyed before this function returns.
2816 * %true if @work was pending, %false otherwise.
2818 bool cancel_work_sync(struct work_struct
*work
)
2820 return __cancel_work_timer(work
, false);
2822 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2825 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2826 * @dwork: the delayed work to flush
2828 * Delayed timer is cancelled and the pending work is queued for
2829 * immediate execution. Like flush_work(), this function only
2830 * considers the last queueing instance of @dwork.
2833 * %true if flush_work() waited for the work to finish execution,
2834 * %false if it was already idle.
2836 bool flush_delayed_work(struct delayed_work
*dwork
)
2838 local_irq_disable();
2839 if (del_timer_sync(&dwork
->timer
))
2840 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2842 return flush_work(&dwork
->work
);
2844 EXPORT_SYMBOL(flush_delayed_work
);
2847 * cancel_delayed_work - cancel a delayed work
2848 * @dwork: delayed_work to cancel
2850 * Kill off a pending delayed_work.
2852 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2856 * The work callback function may still be running on return, unless
2857 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2858 * use cancel_delayed_work_sync() to wait on it.
2860 * This function is safe to call from any context including IRQ handler.
2862 bool cancel_delayed_work(struct delayed_work
*dwork
)
2864 unsigned long flags
;
2868 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2869 } while (unlikely(ret
== -EAGAIN
));
2871 if (unlikely(ret
< 0))
2874 set_work_pool_and_clear_pending(&dwork
->work
,
2875 get_work_pool_id(&dwork
->work
));
2876 local_irq_restore(flags
);
2879 EXPORT_SYMBOL(cancel_delayed_work
);
2882 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2883 * @dwork: the delayed work cancel
2885 * This is cancel_work_sync() for delayed works.
2888 * %true if @dwork was pending, %false otherwise.
2890 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2892 return __cancel_work_timer(&dwork
->work
, true);
2894 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2897 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2898 * @func: the function to call
2900 * schedule_on_each_cpu() executes @func on each online CPU using the
2901 * system workqueue and blocks until all CPUs have completed.
2902 * schedule_on_each_cpu() is very slow.
2905 * 0 on success, -errno on failure.
2907 int schedule_on_each_cpu(work_func_t func
)
2910 struct work_struct __percpu
*works
;
2912 works
= alloc_percpu(struct work_struct
);
2918 for_each_online_cpu(cpu
) {
2919 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2921 INIT_WORK(work
, func
);
2922 schedule_work_on(cpu
, work
);
2925 for_each_online_cpu(cpu
)
2926 flush_work(per_cpu_ptr(works
, cpu
));
2934 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2936 * Forces execution of the kernel-global workqueue and blocks until its
2939 * Think twice before calling this function! It's very easy to get into
2940 * trouble if you don't take great care. Either of the following situations
2941 * will lead to deadlock:
2943 * One of the work items currently on the workqueue needs to acquire
2944 * a lock held by your code or its caller.
2946 * Your code is running in the context of a work routine.
2948 * They will be detected by lockdep when they occur, but the first might not
2949 * occur very often. It depends on what work items are on the workqueue and
2950 * what locks they need, which you have no control over.
2952 * In most situations flushing the entire workqueue is overkill; you merely
2953 * need to know that a particular work item isn't queued and isn't running.
2954 * In such cases you should use cancel_delayed_work_sync() or
2955 * cancel_work_sync() instead.
2957 void flush_scheduled_work(void)
2959 flush_workqueue(system_wq
);
2961 EXPORT_SYMBOL(flush_scheduled_work
);
2964 * execute_in_process_context - reliably execute the routine with user context
2965 * @fn: the function to execute
2966 * @ew: guaranteed storage for the execute work structure (must
2967 * be available when the work executes)
2969 * Executes the function immediately if process context is available,
2970 * otherwise schedules the function for delayed execution.
2972 * Return: 0 - function was executed
2973 * 1 - function was scheduled for execution
2975 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
2977 if (!in_interrupt()) {
2982 INIT_WORK(&ew
->work
, fn
);
2983 schedule_work(&ew
->work
);
2987 EXPORT_SYMBOL_GPL(execute_in_process_context
);
2991 * Workqueues with WQ_SYSFS flag set is visible to userland via
2992 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
2993 * following attributes.
2995 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
2996 * max_active RW int : maximum number of in-flight work items
2998 * Unbound workqueues have the following extra attributes.
3000 * id RO int : the associated pool ID
3001 * nice RW int : nice value of the workers
3002 * cpumask RW mask : bitmask of allowed CPUs for the workers
3005 struct workqueue_struct
*wq
;
3009 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3011 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3016 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
3019 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3021 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3023 static DEVICE_ATTR_RO(per_cpu
);
3025 static ssize_t
max_active_show(struct device
*dev
,
3026 struct device_attribute
*attr
, char *buf
)
3028 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3030 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3033 static ssize_t
max_active_store(struct device
*dev
,
3034 struct device_attribute
*attr
, const char *buf
,
3037 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3040 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3043 workqueue_set_max_active(wq
, val
);
3046 static DEVICE_ATTR_RW(max_active
);
3048 static struct attribute
*wq_sysfs_attrs
[] = {
3049 &dev_attr_per_cpu
.attr
,
3050 &dev_attr_max_active
.attr
,
3053 ATTRIBUTE_GROUPS(wq_sysfs
);
3055 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3056 struct device_attribute
*attr
, char *buf
)
3058 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3059 const char *delim
= "";
3060 int node
, written
= 0;
3062 rcu_read_lock_sched();
3063 for_each_node(node
) {
3064 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3065 "%s%d:%d", delim
, node
,
3066 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3069 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3070 rcu_read_unlock_sched();
3075 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3078 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3081 mutex_lock(&wq
->mutex
);
3082 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3083 mutex_unlock(&wq
->mutex
);
3088 /* prepare workqueue_attrs for sysfs store operations */
3089 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3091 struct workqueue_attrs
*attrs
;
3093 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3097 mutex_lock(&wq
->mutex
);
3098 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3099 mutex_unlock(&wq
->mutex
);
3103 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3104 const char *buf
, size_t count
)
3106 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3107 struct workqueue_attrs
*attrs
;
3110 attrs
= wq_sysfs_prep_attrs(wq
);
3114 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3115 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
3116 ret
= apply_workqueue_attrs(wq
, attrs
);
3120 free_workqueue_attrs(attrs
);
3121 return ret
?: count
;
3124 static ssize_t
wq_cpumask_show(struct device
*dev
,
3125 struct device_attribute
*attr
, char *buf
)
3127 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3130 mutex_lock(&wq
->mutex
);
3131 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3132 mutex_unlock(&wq
->mutex
);
3134 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3138 static ssize_t
wq_cpumask_store(struct device
*dev
,
3139 struct device_attribute
*attr
,
3140 const char *buf
, size_t count
)
3142 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3143 struct workqueue_attrs
*attrs
;
3146 attrs
= wq_sysfs_prep_attrs(wq
);
3150 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3152 ret
= apply_workqueue_attrs(wq
, attrs
);
3154 free_workqueue_attrs(attrs
);
3155 return ret
?: count
;
3158 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3161 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3164 mutex_lock(&wq
->mutex
);
3165 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3166 !wq
->unbound_attrs
->no_numa
);
3167 mutex_unlock(&wq
->mutex
);
3172 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3173 const char *buf
, size_t count
)
3175 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3176 struct workqueue_attrs
*attrs
;
3179 attrs
= wq_sysfs_prep_attrs(wq
);
3184 if (sscanf(buf
, "%d", &v
) == 1) {
3185 attrs
->no_numa
= !v
;
3186 ret
= apply_workqueue_attrs(wq
, attrs
);
3189 free_workqueue_attrs(attrs
);
3190 return ret
?: count
;
3193 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3194 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3195 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3196 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3197 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3201 static struct bus_type wq_subsys
= {
3202 .name
= "workqueue",
3203 .dev_groups
= wq_sysfs_groups
,
3206 static int __init
wq_sysfs_init(void)
3208 return subsys_virtual_register(&wq_subsys
, NULL
);
3210 core_initcall(wq_sysfs_init
);
3212 static void wq_device_release(struct device
*dev
)
3214 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3220 * workqueue_sysfs_register - make a workqueue visible in sysfs
3221 * @wq: the workqueue to register
3223 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3224 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3225 * which is the preferred method.
3227 * Workqueue user should use this function directly iff it wants to apply
3228 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3229 * apply_workqueue_attrs() may race against userland updating the
3232 * Return: 0 on success, -errno on failure.
3234 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3236 struct wq_device
*wq_dev
;
3240 * Adjusting max_active or creating new pwqs by applyting
3241 * attributes breaks ordering guarantee. Disallow exposing ordered
3244 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3247 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3252 wq_dev
->dev
.bus
= &wq_subsys
;
3253 wq_dev
->dev
.init_name
= wq
->name
;
3254 wq_dev
->dev
.release
= wq_device_release
;
3257 * unbound_attrs are created separately. Suppress uevent until
3258 * everything is ready.
3260 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3262 ret
= device_register(&wq_dev
->dev
);
3269 if (wq
->flags
& WQ_UNBOUND
) {
3270 struct device_attribute
*attr
;
3272 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3273 ret
= device_create_file(&wq_dev
->dev
, attr
);
3275 device_unregister(&wq_dev
->dev
);
3282 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3287 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3288 * @wq: the workqueue to unregister
3290 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3292 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3294 struct wq_device
*wq_dev
= wq
->wq_dev
;
3300 device_unregister(&wq_dev
->dev
);
3302 #else /* CONFIG_SYSFS */
3303 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3304 #endif /* CONFIG_SYSFS */
3307 * free_workqueue_attrs - free a workqueue_attrs
3308 * @attrs: workqueue_attrs to free
3310 * Undo alloc_workqueue_attrs().
3312 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3315 free_cpumask_var(attrs
->cpumask
);
3321 * alloc_workqueue_attrs - allocate a workqueue_attrs
3322 * @gfp_mask: allocation mask to use
3324 * Allocate a new workqueue_attrs, initialize with default settings and
3327 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3329 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3331 struct workqueue_attrs
*attrs
;
3333 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3336 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3339 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3342 free_workqueue_attrs(attrs
);
3346 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3347 const struct workqueue_attrs
*from
)
3349 to
->nice
= from
->nice
;
3350 cpumask_copy(to
->cpumask
, from
->cpumask
);
3352 * Unlike hash and equality test, this function doesn't ignore
3353 * ->no_numa as it is used for both pool and wq attrs. Instead,
3354 * get_unbound_pool() explicitly clears ->no_numa after copying.
3356 to
->no_numa
= from
->no_numa
;
3359 /* hash value of the content of @attr */
3360 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3364 hash
= jhash_1word(attrs
->nice
, hash
);
3365 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3366 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3370 /* content equality test */
3371 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3372 const struct workqueue_attrs
*b
)
3374 if (a
->nice
!= b
->nice
)
3376 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3382 * init_worker_pool - initialize a newly zalloc'd worker_pool
3383 * @pool: worker_pool to initialize
3385 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3387 * Return: 0 on success, -errno on failure. Even on failure, all fields
3388 * inside @pool proper are initialized and put_unbound_pool() can be called
3389 * on @pool safely to release it.
3391 static int init_worker_pool(struct worker_pool
*pool
)
3393 spin_lock_init(&pool
->lock
);
3396 pool
->node
= NUMA_NO_NODE
;
3397 pool
->flags
|= POOL_DISASSOCIATED
;
3398 INIT_LIST_HEAD(&pool
->worklist
);
3399 INIT_LIST_HEAD(&pool
->idle_list
);
3400 hash_init(pool
->busy_hash
);
3402 init_timer_deferrable(&pool
->idle_timer
);
3403 pool
->idle_timer
.function
= idle_worker_timeout
;
3404 pool
->idle_timer
.data
= (unsigned long)pool
;
3406 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3407 (unsigned long)pool
);
3409 mutex_init(&pool
->manager_arb
);
3410 mutex_init(&pool
->attach_mutex
);
3411 INIT_LIST_HEAD(&pool
->workers
);
3413 ida_init(&pool
->worker_ida
);
3414 INIT_HLIST_NODE(&pool
->hash_node
);
3417 /* shouldn't fail above this point */
3418 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3424 static void rcu_free_pool(struct rcu_head
*rcu
)
3426 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3428 ida_destroy(&pool
->worker_ida
);
3429 free_workqueue_attrs(pool
->attrs
);
3434 * put_unbound_pool - put a worker_pool
3435 * @pool: worker_pool to put
3437 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3438 * safe manner. get_unbound_pool() calls this function on its failure path
3439 * and this function should be able to release pools which went through,
3440 * successfully or not, init_worker_pool().
3442 * Should be called with wq_pool_mutex held.
3444 static void put_unbound_pool(struct worker_pool
*pool
)
3446 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3447 struct worker
*worker
;
3449 lockdep_assert_held(&wq_pool_mutex
);
3455 if (WARN_ON(!(pool
->cpu
< 0)) ||
3456 WARN_ON(!list_empty(&pool
->worklist
)))
3459 /* release id and unhash */
3461 idr_remove(&worker_pool_idr
, pool
->id
);
3462 hash_del(&pool
->hash_node
);
3465 * Become the manager and destroy all workers. Grabbing
3466 * manager_arb prevents @pool's workers from blocking on
3469 mutex_lock(&pool
->manager_arb
);
3471 spin_lock_irq(&pool
->lock
);
3472 while ((worker
= first_idle_worker(pool
)))
3473 destroy_worker(worker
);
3474 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3475 spin_unlock_irq(&pool
->lock
);
3477 mutex_lock(&pool
->attach_mutex
);
3478 if (!list_empty(&pool
->workers
))
3479 pool
->detach_completion
= &detach_completion
;
3480 mutex_unlock(&pool
->attach_mutex
);
3482 if (pool
->detach_completion
)
3483 wait_for_completion(pool
->detach_completion
);
3485 mutex_unlock(&pool
->manager_arb
);
3487 /* shut down the timers */
3488 del_timer_sync(&pool
->idle_timer
);
3489 del_timer_sync(&pool
->mayday_timer
);
3491 /* sched-RCU protected to allow dereferences from get_work_pool() */
3492 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3496 * get_unbound_pool - get a worker_pool with the specified attributes
3497 * @attrs: the attributes of the worker_pool to get
3499 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3500 * reference count and return it. If there already is a matching
3501 * worker_pool, it will be used; otherwise, this function attempts to
3504 * Should be called with wq_pool_mutex held.
3506 * Return: On success, a worker_pool with the same attributes as @attrs.
3507 * On failure, %NULL.
3509 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3511 u32 hash
= wqattrs_hash(attrs
);
3512 struct worker_pool
*pool
;
3515 lockdep_assert_held(&wq_pool_mutex
);
3517 /* do we already have a matching pool? */
3518 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3519 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3525 /* nope, create a new one */
3526 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3527 if (!pool
|| init_worker_pool(pool
) < 0)
3530 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3531 copy_workqueue_attrs(pool
->attrs
, attrs
);
3534 * no_numa isn't a worker_pool attribute, always clear it. See
3535 * 'struct workqueue_attrs' comments for detail.
3537 pool
->attrs
->no_numa
= false;
3539 /* if cpumask is contained inside a NUMA node, we belong to that node */
3540 if (wq_numa_enabled
) {
3541 for_each_node(node
) {
3542 if (cpumask_subset(pool
->attrs
->cpumask
,
3543 wq_numa_possible_cpumask
[node
])) {
3550 if (worker_pool_assign_id(pool
) < 0)
3553 /* create and start the initial worker */
3554 if (create_and_start_worker(pool
) < 0)
3558 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3563 put_unbound_pool(pool
);
3567 static void rcu_free_pwq(struct rcu_head
*rcu
)
3569 kmem_cache_free(pwq_cache
,
3570 container_of(rcu
, struct pool_workqueue
, rcu
));
3574 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3575 * and needs to be destroyed.
3577 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3579 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3580 unbound_release_work
);
3581 struct workqueue_struct
*wq
= pwq
->wq
;
3582 struct worker_pool
*pool
= pwq
->pool
;
3585 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3589 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3590 * necessary on release but do it anyway. It's easier to verify
3591 * and consistent with the linking path.
3593 mutex_lock(&wq
->mutex
);
3594 list_del_rcu(&pwq
->pwqs_node
);
3595 is_last
= list_empty(&wq
->pwqs
);
3596 mutex_unlock(&wq
->mutex
);
3598 mutex_lock(&wq_pool_mutex
);
3599 put_unbound_pool(pool
);
3600 mutex_unlock(&wq_pool_mutex
);
3602 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3605 * If we're the last pwq going away, @wq is already dead and no one
3606 * is gonna access it anymore. Free it.
3609 free_workqueue_attrs(wq
->unbound_attrs
);
3615 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3616 * @pwq: target pool_workqueue
3618 * If @pwq isn't freezing, set @pwq->max_active to the associated
3619 * workqueue's saved_max_active and activate delayed work items
3620 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3622 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3624 struct workqueue_struct
*wq
= pwq
->wq
;
3625 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3627 /* for @wq->saved_max_active */
3628 lockdep_assert_held(&wq
->mutex
);
3630 /* fast exit for non-freezable wqs */
3631 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3634 spin_lock_irq(&pwq
->pool
->lock
);
3637 * During [un]freezing, the caller is responsible for ensuring that
3638 * this function is called at least once after @workqueue_freezing
3639 * is updated and visible.
3641 if (!freezable
|| !workqueue_freezing
) {
3642 pwq
->max_active
= wq
->saved_max_active
;
3644 while (!list_empty(&pwq
->delayed_works
) &&
3645 pwq
->nr_active
< pwq
->max_active
)
3646 pwq_activate_first_delayed(pwq
);
3649 * Need to kick a worker after thawed or an unbound wq's
3650 * max_active is bumped. It's a slow path. Do it always.
3652 wake_up_worker(pwq
->pool
);
3654 pwq
->max_active
= 0;
3657 spin_unlock_irq(&pwq
->pool
->lock
);
3660 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3661 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3662 struct worker_pool
*pool
)
3664 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3666 memset(pwq
, 0, sizeof(*pwq
));
3670 pwq
->flush_color
= -1;
3672 INIT_LIST_HEAD(&pwq
->delayed_works
);
3673 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3674 INIT_LIST_HEAD(&pwq
->mayday_node
);
3675 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3678 /* sync @pwq with the current state of its associated wq and link it */
3679 static void link_pwq(struct pool_workqueue
*pwq
)
3681 struct workqueue_struct
*wq
= pwq
->wq
;
3683 lockdep_assert_held(&wq
->mutex
);
3685 /* may be called multiple times, ignore if already linked */
3686 if (!list_empty(&pwq
->pwqs_node
))
3690 * Set the matching work_color. This is synchronized with
3691 * wq->mutex to avoid confusing flush_workqueue().
3693 pwq
->work_color
= wq
->work_color
;
3695 /* sync max_active to the current setting */
3696 pwq_adjust_max_active(pwq
);
3699 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3702 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3703 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3704 const struct workqueue_attrs
*attrs
)
3706 struct worker_pool
*pool
;
3707 struct pool_workqueue
*pwq
;
3709 lockdep_assert_held(&wq_pool_mutex
);
3711 pool
= get_unbound_pool(attrs
);
3715 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3717 put_unbound_pool(pool
);
3721 init_pwq(pwq
, wq
, pool
);
3725 /* undo alloc_unbound_pwq(), used only in the error path */
3726 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3728 lockdep_assert_held(&wq_pool_mutex
);
3731 put_unbound_pool(pwq
->pool
);
3732 kmem_cache_free(pwq_cache
, pwq
);
3737 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3738 * @attrs: the wq_attrs of interest
3739 * @node: the target NUMA node
3740 * @cpu_going_down: if >= 0, the CPU to consider as offline
3741 * @cpumask: outarg, the resulting cpumask
3743 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3744 * @cpu_going_down is >= 0, that cpu is considered offline during
3745 * calculation. The result is stored in @cpumask.
3747 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3748 * enabled and @node has online CPUs requested by @attrs, the returned
3749 * cpumask is the intersection of the possible CPUs of @node and
3752 * The caller is responsible for ensuring that the cpumask of @node stays
3755 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3758 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3759 int cpu_going_down
, cpumask_t
*cpumask
)
3761 if (!wq_numa_enabled
|| attrs
->no_numa
)
3764 /* does @node have any online CPUs @attrs wants? */
3765 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3766 if (cpu_going_down
>= 0)
3767 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3769 if (cpumask_empty(cpumask
))
3772 /* yeap, return possible CPUs in @node that @attrs wants */
3773 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3774 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3777 cpumask_copy(cpumask
, attrs
->cpumask
);
3781 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3782 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3784 struct pool_workqueue
*pwq
)
3786 struct pool_workqueue
*old_pwq
;
3788 lockdep_assert_held(&wq
->mutex
);
3790 /* link_pwq() can handle duplicate calls */
3793 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3794 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3799 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3800 * @wq: the target workqueue
3801 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3803 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3804 * machines, this function maps a separate pwq to each NUMA node with
3805 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3806 * NUMA node it was issued on. Older pwqs are released as in-flight work
3807 * items finish. Note that a work item which repeatedly requeues itself
3808 * back-to-back will stay on its current pwq.
3810 * Performs GFP_KERNEL allocations.
3812 * Return: 0 on success and -errno on failure.
3814 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3815 const struct workqueue_attrs
*attrs
)
3817 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3818 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3821 /* only unbound workqueues can change attributes */
3822 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3825 /* creating multiple pwqs breaks ordering guarantee */
3826 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3829 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3830 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3831 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3832 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3835 /* make a copy of @attrs and sanitize it */
3836 copy_workqueue_attrs(new_attrs
, attrs
);
3837 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3840 * We may create multiple pwqs with differing cpumasks. Make a
3841 * copy of @new_attrs which will be modified and used to obtain
3844 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3847 * CPUs should stay stable across pwq creations and installations.
3848 * Pin CPUs, determine the target cpumask for each node and create
3853 mutex_lock(&wq_pool_mutex
);
3856 * If something goes wrong during CPU up/down, we'll fall back to
3857 * the default pwq covering whole @attrs->cpumask. Always create
3858 * it even if we don't use it immediately.
3860 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3864 for_each_node(node
) {
3865 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3866 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3871 pwq_tbl
[node
] = dfl_pwq
;
3875 mutex_unlock(&wq_pool_mutex
);
3877 /* all pwqs have been created successfully, let's install'em */
3878 mutex_lock(&wq
->mutex
);
3880 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3882 /* save the previous pwq and install the new one */
3884 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3886 /* @dfl_pwq might not have been used, ensure it's linked */
3888 swap(wq
->dfl_pwq
, dfl_pwq
);
3890 mutex_unlock(&wq
->mutex
);
3892 /* put the old pwqs */
3894 put_pwq_unlocked(pwq_tbl
[node
]);
3895 put_pwq_unlocked(dfl_pwq
);
3901 free_workqueue_attrs(tmp_attrs
);
3902 free_workqueue_attrs(new_attrs
);
3907 free_unbound_pwq(dfl_pwq
);
3909 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
3910 free_unbound_pwq(pwq_tbl
[node
]);
3911 mutex_unlock(&wq_pool_mutex
);
3919 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3920 * @wq: the target workqueue
3921 * @cpu: the CPU coming up or going down
3922 * @online: whether @cpu is coming up or going down
3924 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3925 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3928 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3929 * falls back to @wq->dfl_pwq which may not be optimal but is always
3932 * Note that when the last allowed CPU of a NUMA node goes offline for a
3933 * workqueue with a cpumask spanning multiple nodes, the workers which were
3934 * already executing the work items for the workqueue will lose their CPU
3935 * affinity and may execute on any CPU. This is similar to how per-cpu
3936 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3937 * affinity, it's the user's responsibility to flush the work item from
3940 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3943 int node
= cpu_to_node(cpu
);
3944 int cpu_off
= online
? -1 : cpu
;
3945 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3946 struct workqueue_attrs
*target_attrs
;
3949 lockdep_assert_held(&wq_pool_mutex
);
3951 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
3955 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3956 * Let's use a preallocated one. The following buf is protected by
3957 * CPU hotplug exclusion.
3959 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3960 cpumask
= target_attrs
->cpumask
;
3962 mutex_lock(&wq
->mutex
);
3963 if (wq
->unbound_attrs
->no_numa
)
3966 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3967 pwq
= unbound_pwq_by_node(wq
, node
);
3970 * Let's determine what needs to be done. If the target cpumask is
3971 * different from wq's, we need to compare it to @pwq's and create
3972 * a new one if they don't match. If the target cpumask equals
3973 * wq's, the default pwq should be used.
3975 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
3976 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3982 mutex_unlock(&wq
->mutex
);
3984 /* create a new pwq */
3985 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3987 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3989 mutex_lock(&wq
->mutex
);
3994 * Install the new pwq. As this function is called only from CPU
3995 * hotplug callbacks and applying a new attrs is wrapped with
3996 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3999 mutex_lock(&wq
->mutex
);
4000 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4004 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4005 get_pwq(wq
->dfl_pwq
);
4006 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4007 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4009 mutex_unlock(&wq
->mutex
);
4010 put_pwq_unlocked(old_pwq
);
4013 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4015 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4018 if (!(wq
->flags
& WQ_UNBOUND
)) {
4019 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4023 for_each_possible_cpu(cpu
) {
4024 struct pool_workqueue
*pwq
=
4025 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4026 struct worker_pool
*cpu_pools
=
4027 per_cpu(cpu_worker_pools
, cpu
);
4029 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4031 mutex_lock(&wq
->mutex
);
4033 mutex_unlock(&wq
->mutex
);
4036 } else if (wq
->flags
& __WQ_ORDERED
) {
4037 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4038 /* there should only be single pwq for ordering guarantee */
4039 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4040 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4041 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4044 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4048 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4051 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4053 if (max_active
< 1 || max_active
> lim
)
4054 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4055 max_active
, name
, 1, lim
);
4057 return clamp_val(max_active
, 1, lim
);
4060 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4063 struct lock_class_key
*key
,
4064 const char *lock_name
, ...)
4066 size_t tbl_size
= 0;
4068 struct workqueue_struct
*wq
;
4069 struct pool_workqueue
*pwq
;
4071 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4072 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4073 flags
|= WQ_UNBOUND
;
4075 /* allocate wq and format name */
4076 if (flags
& WQ_UNBOUND
)
4077 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4079 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4083 if (flags
& WQ_UNBOUND
) {
4084 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4085 if (!wq
->unbound_attrs
)
4089 va_start(args
, lock_name
);
4090 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4093 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4094 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4098 wq
->saved_max_active
= max_active
;
4099 mutex_init(&wq
->mutex
);
4100 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4101 INIT_LIST_HEAD(&wq
->pwqs
);
4102 INIT_LIST_HEAD(&wq
->flusher_queue
);
4103 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4104 INIT_LIST_HEAD(&wq
->maydays
);
4106 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4107 INIT_LIST_HEAD(&wq
->list
);
4109 if (alloc_and_link_pwqs(wq
) < 0)
4113 * Workqueues which may be used during memory reclaim should
4114 * have a rescuer to guarantee forward progress.
4116 if (flags
& WQ_MEM_RECLAIM
) {
4117 struct worker
*rescuer
;
4119 rescuer
= alloc_worker();
4123 rescuer
->rescue_wq
= wq
;
4124 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4126 if (IS_ERR(rescuer
->task
)) {
4131 wq
->rescuer
= rescuer
;
4132 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4133 wake_up_process(rescuer
->task
);
4136 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4140 * wq_pool_mutex protects global freeze state and workqueues list.
4141 * Grab it, adjust max_active and add the new @wq to workqueues
4144 mutex_lock(&wq_pool_mutex
);
4146 mutex_lock(&wq
->mutex
);
4147 for_each_pwq(pwq
, wq
)
4148 pwq_adjust_max_active(pwq
);
4149 mutex_unlock(&wq
->mutex
);
4151 list_add(&wq
->list
, &workqueues
);
4153 mutex_unlock(&wq_pool_mutex
);
4158 free_workqueue_attrs(wq
->unbound_attrs
);
4162 destroy_workqueue(wq
);
4165 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4168 * destroy_workqueue - safely terminate a workqueue
4169 * @wq: target workqueue
4171 * Safely destroy a workqueue. All work currently pending will be done first.
4173 void destroy_workqueue(struct workqueue_struct
*wq
)
4175 struct pool_workqueue
*pwq
;
4178 /* drain it before proceeding with destruction */
4179 drain_workqueue(wq
);
4182 mutex_lock(&wq
->mutex
);
4183 for_each_pwq(pwq
, wq
) {
4186 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4187 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4188 mutex_unlock(&wq
->mutex
);
4193 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4194 WARN_ON(pwq
->nr_active
) ||
4195 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4196 mutex_unlock(&wq
->mutex
);
4200 mutex_unlock(&wq
->mutex
);
4203 * wq list is used to freeze wq, remove from list after
4204 * flushing is complete in case freeze races us.
4206 mutex_lock(&wq_pool_mutex
);
4207 list_del_init(&wq
->list
);
4208 mutex_unlock(&wq_pool_mutex
);
4210 workqueue_sysfs_unregister(wq
);
4213 kthread_stop(wq
->rescuer
->task
);
4218 if (!(wq
->flags
& WQ_UNBOUND
)) {
4220 * The base ref is never dropped on per-cpu pwqs. Directly
4221 * free the pwqs and wq.
4223 free_percpu(wq
->cpu_pwqs
);
4227 * We're the sole accessor of @wq at this point. Directly
4228 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4229 * @wq will be freed when the last pwq is released.
4231 for_each_node(node
) {
4232 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4233 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4234 put_pwq_unlocked(pwq
);
4238 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4239 * put. Don't access it afterwards.
4243 put_pwq_unlocked(pwq
);
4246 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4249 * workqueue_set_max_active - adjust max_active of a workqueue
4250 * @wq: target workqueue
4251 * @max_active: new max_active value.
4253 * Set max_active of @wq to @max_active.
4256 * Don't call from IRQ context.
4258 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4260 struct pool_workqueue
*pwq
;
4262 /* disallow meddling with max_active for ordered workqueues */
4263 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4266 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4268 mutex_lock(&wq
->mutex
);
4270 wq
->saved_max_active
= max_active
;
4272 for_each_pwq(pwq
, wq
)
4273 pwq_adjust_max_active(pwq
);
4275 mutex_unlock(&wq
->mutex
);
4277 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4280 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4282 * Determine whether %current is a workqueue rescuer. Can be used from
4283 * work functions to determine whether it's being run off the rescuer task.
4285 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4287 bool current_is_workqueue_rescuer(void)
4289 struct worker
*worker
= current_wq_worker();
4291 return worker
&& worker
->rescue_wq
;
4295 * workqueue_congested - test whether a workqueue is congested
4296 * @cpu: CPU in question
4297 * @wq: target workqueue
4299 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4300 * no synchronization around this function and the test result is
4301 * unreliable and only useful as advisory hints or for debugging.
4303 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4304 * Note that both per-cpu and unbound workqueues may be associated with
4305 * multiple pool_workqueues which have separate congested states. A
4306 * workqueue being congested on one CPU doesn't mean the workqueue is also
4307 * contested on other CPUs / NUMA nodes.
4310 * %true if congested, %false otherwise.
4312 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4314 struct pool_workqueue
*pwq
;
4317 rcu_read_lock_sched();
4319 if (cpu
== WORK_CPU_UNBOUND
)
4320 cpu
= smp_processor_id();
4322 if (!(wq
->flags
& WQ_UNBOUND
))
4323 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4325 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4327 ret
= !list_empty(&pwq
->delayed_works
);
4328 rcu_read_unlock_sched();
4332 EXPORT_SYMBOL_GPL(workqueue_congested
);
4335 * work_busy - test whether a work is currently pending or running
4336 * @work: the work to be tested
4338 * Test whether @work is currently pending or running. There is no
4339 * synchronization around this function and the test result is
4340 * unreliable and only useful as advisory hints or for debugging.
4343 * OR'd bitmask of WORK_BUSY_* bits.
4345 unsigned int work_busy(struct work_struct
*work
)
4347 struct worker_pool
*pool
;
4348 unsigned long flags
;
4349 unsigned int ret
= 0;
4351 if (work_pending(work
))
4352 ret
|= WORK_BUSY_PENDING
;
4354 local_irq_save(flags
);
4355 pool
= get_work_pool(work
);
4357 spin_lock(&pool
->lock
);
4358 if (find_worker_executing_work(pool
, work
))
4359 ret
|= WORK_BUSY_RUNNING
;
4360 spin_unlock(&pool
->lock
);
4362 local_irq_restore(flags
);
4366 EXPORT_SYMBOL_GPL(work_busy
);
4369 * set_worker_desc - set description for the current work item
4370 * @fmt: printf-style format string
4371 * @...: arguments for the format string
4373 * This function can be called by a running work function to describe what
4374 * the work item is about. If the worker task gets dumped, this
4375 * information will be printed out together to help debugging. The
4376 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4378 void set_worker_desc(const char *fmt
, ...)
4380 struct worker
*worker
= current_wq_worker();
4384 va_start(args
, fmt
);
4385 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4387 worker
->desc_valid
= true;
4392 * print_worker_info - print out worker information and description
4393 * @log_lvl: the log level to use when printing
4394 * @task: target task
4396 * If @task is a worker and currently executing a work item, print out the
4397 * name of the workqueue being serviced and worker description set with
4398 * set_worker_desc() by the currently executing work item.
4400 * This function can be safely called on any task as long as the
4401 * task_struct itself is accessible. While safe, this function isn't
4402 * synchronized and may print out mixups or garbages of limited length.
4404 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4406 work_func_t
*fn
= NULL
;
4407 char name
[WQ_NAME_LEN
] = { };
4408 char desc
[WORKER_DESC_LEN
] = { };
4409 struct pool_workqueue
*pwq
= NULL
;
4410 struct workqueue_struct
*wq
= NULL
;
4411 bool desc_valid
= false;
4412 struct worker
*worker
;
4414 if (!(task
->flags
& PF_WQ_WORKER
))
4418 * This function is called without any synchronization and @task
4419 * could be in any state. Be careful with dereferences.
4421 worker
= probe_kthread_data(task
);
4424 * Carefully copy the associated workqueue's workfn and name. Keep
4425 * the original last '\0' in case the original contains garbage.
4427 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4428 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4429 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4430 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4432 /* copy worker description */
4433 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4435 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4437 if (fn
|| name
[0] || desc
[0]) {
4438 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4440 pr_cont(" (%s)", desc
);
4448 * There are two challenges in supporting CPU hotplug. Firstly, there
4449 * are a lot of assumptions on strong associations among work, pwq and
4450 * pool which make migrating pending and scheduled works very
4451 * difficult to implement without impacting hot paths. Secondly,
4452 * worker pools serve mix of short, long and very long running works making
4453 * blocked draining impractical.
4455 * This is solved by allowing the pools to be disassociated from the CPU
4456 * running as an unbound one and allowing it to be reattached later if the
4457 * cpu comes back online.
4460 static void wq_unbind_fn(struct work_struct
*work
)
4462 int cpu
= smp_processor_id();
4463 struct worker_pool
*pool
;
4464 struct worker
*worker
;
4466 for_each_cpu_worker_pool(pool
, cpu
) {
4467 mutex_lock(&pool
->attach_mutex
);
4468 spin_lock_irq(&pool
->lock
);
4471 * We've blocked all attach/detach operations. Make all workers
4472 * unbound and set DISASSOCIATED. Before this, all workers
4473 * except for the ones which are still executing works from
4474 * before the last CPU down must be on the cpu. After
4475 * this, they may become diasporas.
4477 for_each_pool_worker(worker
, pool
)
4478 worker
->flags
|= WORKER_UNBOUND
;
4480 pool
->flags
|= POOL_DISASSOCIATED
;
4482 spin_unlock_irq(&pool
->lock
);
4483 mutex_unlock(&pool
->attach_mutex
);
4486 * Call schedule() so that we cross rq->lock and thus can
4487 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4488 * This is necessary as scheduler callbacks may be invoked
4494 * Sched callbacks are disabled now. Zap nr_running.
4495 * After this, nr_running stays zero and need_more_worker()
4496 * and keep_working() are always true as long as the
4497 * worklist is not empty. This pool now behaves as an
4498 * unbound (in terms of concurrency management) pool which
4499 * are served by workers tied to the pool.
4501 atomic_set(&pool
->nr_running
, 0);
4504 * With concurrency management just turned off, a busy
4505 * worker blocking could lead to lengthy stalls. Kick off
4506 * unbound chain execution of currently pending work items.
4508 spin_lock_irq(&pool
->lock
);
4509 wake_up_worker(pool
);
4510 spin_unlock_irq(&pool
->lock
);
4515 * rebind_workers - rebind all workers of a pool to the associated CPU
4516 * @pool: pool of interest
4518 * @pool->cpu is coming online. Rebind all workers to the CPU.
4520 static void rebind_workers(struct worker_pool
*pool
)
4522 struct worker
*worker
;
4524 lockdep_assert_held(&pool
->attach_mutex
);
4527 * Restore CPU affinity of all workers. As all idle workers should
4528 * be on the run-queue of the associated CPU before any local
4529 * wake-ups for concurrency management happen, restore CPU affinty
4530 * of all workers first and then clear UNBOUND. As we're called
4531 * from CPU_ONLINE, the following shouldn't fail.
4533 for_each_pool_worker(worker
, pool
)
4534 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4535 pool
->attrs
->cpumask
) < 0);
4537 spin_lock_irq(&pool
->lock
);
4539 for_each_pool_worker(worker
, pool
) {
4540 unsigned int worker_flags
= worker
->flags
;
4543 * A bound idle worker should actually be on the runqueue
4544 * of the associated CPU for local wake-ups targeting it to
4545 * work. Kick all idle workers so that they migrate to the
4546 * associated CPU. Doing this in the same loop as
4547 * replacing UNBOUND with REBOUND is safe as no worker will
4548 * be bound before @pool->lock is released.
4550 if (worker_flags
& WORKER_IDLE
)
4551 wake_up_process(worker
->task
);
4554 * We want to clear UNBOUND but can't directly call
4555 * worker_clr_flags() or adjust nr_running. Atomically
4556 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4557 * @worker will clear REBOUND using worker_clr_flags() when
4558 * it initiates the next execution cycle thus restoring
4559 * concurrency management. Note that when or whether
4560 * @worker clears REBOUND doesn't affect correctness.
4562 * ACCESS_ONCE() is necessary because @worker->flags may be
4563 * tested without holding any lock in
4564 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4565 * fail incorrectly leading to premature concurrency
4566 * management operations.
4568 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4569 worker_flags
|= WORKER_REBOUND
;
4570 worker_flags
&= ~WORKER_UNBOUND
;
4571 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4574 spin_unlock_irq(&pool
->lock
);
4578 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4579 * @pool: unbound pool of interest
4580 * @cpu: the CPU which is coming up
4582 * An unbound pool may end up with a cpumask which doesn't have any online
4583 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4584 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4585 * online CPU before, cpus_allowed of all its workers should be restored.
4587 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4589 static cpumask_t cpumask
;
4590 struct worker
*worker
;
4592 lockdep_assert_held(&pool
->attach_mutex
);
4594 /* is @cpu allowed for @pool? */
4595 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4598 /* is @cpu the only online CPU? */
4599 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4600 if (cpumask_weight(&cpumask
) != 1)
4603 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4604 for_each_pool_worker(worker
, pool
)
4605 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4606 pool
->attrs
->cpumask
) < 0);
4610 * Workqueues should be brought up before normal priority CPU notifiers.
4611 * This will be registered high priority CPU notifier.
4613 static int workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4614 unsigned long action
,
4617 int cpu
= (unsigned long)hcpu
;
4618 struct worker_pool
*pool
;
4619 struct workqueue_struct
*wq
;
4622 switch (action
& ~CPU_TASKS_FROZEN
) {
4623 case CPU_UP_PREPARE
:
4624 for_each_cpu_worker_pool(pool
, cpu
) {
4625 if (pool
->nr_workers
)
4627 if (create_and_start_worker(pool
) < 0)
4632 case CPU_DOWN_FAILED
:
4634 mutex_lock(&wq_pool_mutex
);
4636 for_each_pool(pool
, pi
) {
4637 mutex_lock(&pool
->attach_mutex
);
4639 if (pool
->cpu
== cpu
) {
4640 spin_lock_irq(&pool
->lock
);
4641 pool
->flags
&= ~POOL_DISASSOCIATED
;
4642 spin_unlock_irq(&pool
->lock
);
4644 rebind_workers(pool
);
4645 } else if (pool
->cpu
< 0) {
4646 restore_unbound_workers_cpumask(pool
, cpu
);
4649 mutex_unlock(&pool
->attach_mutex
);
4652 /* update NUMA affinity of unbound workqueues */
4653 list_for_each_entry(wq
, &workqueues
, list
)
4654 wq_update_unbound_numa(wq
, cpu
, true);
4656 mutex_unlock(&wq_pool_mutex
);
4663 * Workqueues should be brought down after normal priority CPU notifiers.
4664 * This will be registered as low priority CPU notifier.
4666 static int workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4667 unsigned long action
,
4670 int cpu
= (unsigned long)hcpu
;
4671 struct work_struct unbind_work
;
4672 struct workqueue_struct
*wq
;
4674 switch (action
& ~CPU_TASKS_FROZEN
) {
4675 case CPU_DOWN_PREPARE
:
4676 /* unbinding per-cpu workers should happen on the local CPU */
4677 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4678 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4680 /* update NUMA affinity of unbound workqueues */
4681 mutex_lock(&wq_pool_mutex
);
4682 list_for_each_entry(wq
, &workqueues
, list
)
4683 wq_update_unbound_numa(wq
, cpu
, false);
4684 mutex_unlock(&wq_pool_mutex
);
4686 /* wait for per-cpu unbinding to finish */
4687 flush_work(&unbind_work
);
4688 destroy_work_on_stack(&unbind_work
);
4696 struct work_for_cpu
{
4697 struct work_struct work
;
4703 static void work_for_cpu_fn(struct work_struct
*work
)
4705 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4707 wfc
->ret
= wfc
->fn(wfc
->arg
);
4711 * work_on_cpu - run a function in user context on a particular cpu
4712 * @cpu: the cpu to run on
4713 * @fn: the function to run
4714 * @arg: the function arg
4716 * It is up to the caller to ensure that the cpu doesn't go offline.
4717 * The caller must not hold any locks which would prevent @fn from completing.
4719 * Return: The value @fn returns.
4721 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4723 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4725 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4726 schedule_work_on(cpu
, &wfc
.work
);
4727 flush_work(&wfc
.work
);
4728 destroy_work_on_stack(&wfc
.work
);
4731 EXPORT_SYMBOL_GPL(work_on_cpu
);
4732 #endif /* CONFIG_SMP */
4734 #ifdef CONFIG_FREEZER
4737 * freeze_workqueues_begin - begin freezing workqueues
4739 * Start freezing workqueues. After this function returns, all freezable
4740 * workqueues will queue new works to their delayed_works list instead of
4744 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4746 void freeze_workqueues_begin(void)
4748 struct workqueue_struct
*wq
;
4749 struct pool_workqueue
*pwq
;
4751 mutex_lock(&wq_pool_mutex
);
4753 WARN_ON_ONCE(workqueue_freezing
);
4754 workqueue_freezing
= true;
4756 list_for_each_entry(wq
, &workqueues
, list
) {
4757 mutex_lock(&wq
->mutex
);
4758 for_each_pwq(pwq
, wq
)
4759 pwq_adjust_max_active(pwq
);
4760 mutex_unlock(&wq
->mutex
);
4763 mutex_unlock(&wq_pool_mutex
);
4767 * freeze_workqueues_busy - are freezable workqueues still busy?
4769 * Check whether freezing is complete. This function must be called
4770 * between freeze_workqueues_begin() and thaw_workqueues().
4773 * Grabs and releases wq_pool_mutex.
4776 * %true if some freezable workqueues are still busy. %false if freezing
4779 bool freeze_workqueues_busy(void)
4782 struct workqueue_struct
*wq
;
4783 struct pool_workqueue
*pwq
;
4785 mutex_lock(&wq_pool_mutex
);
4787 WARN_ON_ONCE(!workqueue_freezing
);
4789 list_for_each_entry(wq
, &workqueues
, list
) {
4790 if (!(wq
->flags
& WQ_FREEZABLE
))
4793 * nr_active is monotonically decreasing. It's safe
4794 * to peek without lock.
4796 rcu_read_lock_sched();
4797 for_each_pwq(pwq
, wq
) {
4798 WARN_ON_ONCE(pwq
->nr_active
< 0);
4799 if (pwq
->nr_active
) {
4801 rcu_read_unlock_sched();
4805 rcu_read_unlock_sched();
4808 mutex_unlock(&wq_pool_mutex
);
4813 * thaw_workqueues - thaw workqueues
4815 * Thaw workqueues. Normal queueing is restored and all collected
4816 * frozen works are transferred to their respective pool worklists.
4819 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4821 void thaw_workqueues(void)
4823 struct workqueue_struct
*wq
;
4824 struct pool_workqueue
*pwq
;
4826 mutex_lock(&wq_pool_mutex
);
4828 if (!workqueue_freezing
)
4831 workqueue_freezing
= false;
4833 /* restore max_active and repopulate worklist */
4834 list_for_each_entry(wq
, &workqueues
, list
) {
4835 mutex_lock(&wq
->mutex
);
4836 for_each_pwq(pwq
, wq
)
4837 pwq_adjust_max_active(pwq
);
4838 mutex_unlock(&wq
->mutex
);
4842 mutex_unlock(&wq_pool_mutex
);
4844 #endif /* CONFIG_FREEZER */
4846 static void __init
wq_numa_init(void)
4851 /* determine NUMA pwq table len - highest node id + 1 */
4853 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4855 if (num_possible_nodes() <= 1)
4858 if (wq_disable_numa
) {
4859 pr_info("workqueue: NUMA affinity support disabled\n");
4863 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4864 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4867 * We want masks of possible CPUs of each node which isn't readily
4868 * available. Build one from cpu_to_node() which should have been
4869 * fully initialized by now.
4871 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
4875 BUG_ON(!alloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
4876 node_online(node
) ? node
: NUMA_NO_NODE
));
4878 for_each_possible_cpu(cpu
) {
4879 node
= cpu_to_node(cpu
);
4880 if (WARN_ON(node
== NUMA_NO_NODE
)) {
4881 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
4882 /* happens iff arch is bonkers, let's just proceed */
4885 cpumask_set_cpu(cpu
, tbl
[node
]);
4888 wq_numa_possible_cpumask
= tbl
;
4889 wq_numa_enabled
= true;
4892 static int __init
init_workqueues(void)
4894 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4897 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4899 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4901 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4902 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4906 /* initialize CPU pools */
4907 for_each_possible_cpu(cpu
) {
4908 struct worker_pool
*pool
;
4911 for_each_cpu_worker_pool(pool
, cpu
) {
4912 BUG_ON(init_worker_pool(pool
));
4914 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4915 pool
->attrs
->nice
= std_nice
[i
++];
4916 pool
->node
= cpu_to_node(cpu
);
4919 mutex_lock(&wq_pool_mutex
);
4920 BUG_ON(worker_pool_assign_id(pool
));
4921 mutex_unlock(&wq_pool_mutex
);
4925 /* create the initial worker */
4926 for_each_online_cpu(cpu
) {
4927 struct worker_pool
*pool
;
4929 for_each_cpu_worker_pool(pool
, cpu
) {
4930 pool
->flags
&= ~POOL_DISASSOCIATED
;
4931 BUG_ON(create_and_start_worker(pool
) < 0);
4935 /* create default unbound and ordered wq attrs */
4936 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4937 struct workqueue_attrs
*attrs
;
4939 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4940 attrs
->nice
= std_nice
[i
];
4941 unbound_std_wq_attrs
[i
] = attrs
;
4944 * An ordered wq should have only one pwq as ordering is
4945 * guaranteed by max_active which is enforced by pwqs.
4946 * Turn off NUMA so that dfl_pwq is used for all nodes.
4948 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4949 attrs
->nice
= std_nice
[i
];
4950 attrs
->no_numa
= true;
4951 ordered_wq_attrs
[i
] = attrs
;
4954 system_wq
= alloc_workqueue("events", 0, 0);
4955 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4956 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4957 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4958 WQ_UNBOUND_MAX_ACTIVE
);
4959 system_freezable_wq
= alloc_workqueue("events_freezable",
4961 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
4962 WQ_POWER_EFFICIENT
, 0);
4963 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
4964 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
4966 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
4967 !system_unbound_wq
|| !system_freezable_wq
||
4968 !system_power_efficient_wq
||
4969 !system_freezable_power_efficient_wq
);
4972 early_initcall(init_workqueues
);