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 * manager_mutex to avoid changing binding state while
69 * create_worker() is in progress.
71 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
72 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
73 POOL_FREEZING
= 1 << 3, /* freeze in progress */
76 WORKER_STARTED
= 1 << 0, /* started */
77 WORKER_DIE
= 1 << 1, /* die die die */
78 WORKER_IDLE
= 1 << 2, /* is idle */
79 WORKER_PREP
= 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
82 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
85 WORKER_UNBOUND
| WORKER_REBOUND
,
87 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
99 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give -20.
105 RESCUER_NICE_LEVEL
= -20,
106 HIGHPRI_NICE_LEVEL
= -20,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * MG: pool->manager_mutex and pool->lock protected. Writes require both
128 * locks. Reads can happen under either lock.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
134 * WQ: wq->mutex protected.
136 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138 * MD: wq_mayday_lock protected.
141 /* struct worker is defined in workqueue_internal.h */
144 spinlock_t lock
; /* the pool lock */
145 int cpu
; /* I: the associated cpu */
146 int node
; /* I: the associated node ID */
147 int id
; /* I: pool ID */
148 unsigned int flags
; /* X: flags */
150 struct list_head worklist
; /* L: list of pending works */
151 int nr_workers
; /* L: total number of workers */
153 /* nr_idle includes the ones off idle_list for rebinding */
154 int nr_idle
; /* L: currently idle ones */
156 struct list_head idle_list
; /* X: list of idle workers */
157 struct timer_list idle_timer
; /* L: worker idle timeout */
158 struct timer_list mayday_timer
; /* L: SOS timer for workers */
160 /* a workers is either on busy_hash or idle_list, or the manager */
161 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
162 /* L: hash of busy workers */
164 /* see manage_workers() for details on the two manager mutexes */
165 struct mutex manager_arb
; /* manager arbitration */
166 struct mutex manager_mutex
; /* manager exclusion */
167 struct idr worker_idr
; /* MG: worker IDs and iteration */
169 struct workqueue_attrs
*attrs
; /* I: worker attributes */
170 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
171 int refcnt
; /* PL: refcnt for unbound pools */
174 * The current concurrency level. As it's likely to be accessed
175 * from other CPUs during try_to_wake_up(), put it in a separate
178 atomic_t nr_running ____cacheline_aligned_in_smp
;
181 * Destruction of pool is sched-RCU protected to allow dereferences
182 * from get_work_pool().
185 } ____cacheline_aligned_in_smp
;
188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
193 struct pool_workqueue
{
194 struct worker_pool
*pool
; /* I: the associated pool */
195 struct workqueue_struct
*wq
; /* I: the owning workqueue */
196 int work_color
; /* L: current color */
197 int flush_color
; /* L: flushing color */
198 int refcnt
; /* L: reference count */
199 int nr_in_flight
[WORK_NR_COLORS
];
200 /* L: nr of in_flight works */
201 int nr_active
; /* L: nr of active works */
202 int max_active
; /* L: max active works */
203 struct list_head delayed_works
; /* L: delayed works */
204 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
205 struct list_head mayday_node
; /* MD: node on wq->maydays */
208 * Release of unbound pwq is punted to system_wq. See put_pwq()
209 * and pwq_unbound_release_workfn() for details. pool_workqueue
210 * itself is also sched-RCU protected so that the first pwq can be
211 * determined without grabbing wq->mutex.
213 struct work_struct unbound_release_work
;
215 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
218 * Structure used to wait for workqueue flush.
221 struct list_head list
; /* WQ: list of flushers */
222 int flush_color
; /* WQ: flush color waiting for */
223 struct completion done
; /* flush completion */
229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
232 struct workqueue_struct
{
233 struct list_head pwqs
; /* WR: all pwqs of this wq */
234 struct list_head list
; /* PL: list of all workqueues */
236 struct mutex mutex
; /* protects this wq */
237 int work_color
; /* WQ: current work color */
238 int flush_color
; /* WQ: current flush color */
239 atomic_t nr_pwqs_to_flush
; /* flush in progress */
240 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
241 struct list_head flusher_queue
; /* WQ: flush waiters */
242 struct list_head flusher_overflow
; /* WQ: flush overflow list */
244 struct list_head maydays
; /* MD: pwqs requesting rescue */
245 struct worker
*rescuer
; /* I: rescue worker */
247 int nr_drainers
; /* WQ: drain in progress */
248 int saved_max_active
; /* WQ: saved pwq max_active */
250 struct workqueue_attrs
*unbound_attrs
; /* WQ: only for unbound wqs */
251 struct pool_workqueue
*dfl_pwq
; /* WQ: only for unbound wqs */
254 struct wq_device
*wq_dev
; /* I: for sysfs interface */
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map
;
259 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
261 /* hot fields used during command issue, aligned to cacheline */
262 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
263 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
264 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* FR: unbound pwqs indexed by node */
267 static struct kmem_cache
*pwq_cache
;
269 static int wq_numa_tbl_len
; /* highest possible NUMA node id + 1 */
270 static cpumask_var_t
*wq_numa_possible_cpumask
;
271 /* possible CPUs of each node */
273 static bool wq_disable_numa
;
274 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
276 /* see the comment above the definition of WQ_POWER_EFFICIENT */
277 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
278 static bool wq_power_efficient
= true;
280 static bool wq_power_efficient
;
283 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
285 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
287 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
288 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
290 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
291 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
293 static LIST_HEAD(workqueues
); /* PL: list of all workqueues */
294 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
296 /* the per-cpu worker pools */
297 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
300 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
302 /* PL: hash of all unbound pools keyed by pool->attrs */
303 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
305 /* I: attributes used when instantiating standard unbound pools on demand */
306 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
308 /* I: attributes used when instantiating ordered pools on demand */
309 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
311 struct workqueue_struct
*system_wq __read_mostly
;
312 EXPORT_SYMBOL(system_wq
);
313 struct workqueue_struct
*system_highpri_wq __read_mostly
;
314 EXPORT_SYMBOL_GPL(system_highpri_wq
);
315 struct workqueue_struct
*system_long_wq __read_mostly
;
316 EXPORT_SYMBOL_GPL(system_long_wq
);
317 struct workqueue_struct
*system_unbound_wq __read_mostly
;
318 EXPORT_SYMBOL_GPL(system_unbound_wq
);
319 struct workqueue_struct
*system_freezable_wq __read_mostly
;
320 EXPORT_SYMBOL_GPL(system_freezable_wq
);
321 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
322 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
323 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
324 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
326 static int worker_thread(void *__worker
);
327 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
328 const struct workqueue_attrs
*from
);
330 #define CREATE_TRACE_POINTS
331 #include <trace/events/workqueue.h>
333 #define assert_rcu_or_pool_mutex() \
334 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
335 lockdep_is_held(&wq_pool_mutex), \
336 "sched RCU or wq_pool_mutex should be held")
338 #define assert_rcu_or_wq_mutex(wq) \
339 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
340 lockdep_is_held(&wq->mutex), \
341 "sched RCU or wq->mutex should be held")
343 #ifdef CONFIG_LOCKDEP
344 #define assert_manager_or_pool_lock(pool) \
345 WARN_ONCE(debug_locks && \
346 !lockdep_is_held(&(pool)->manager_mutex) && \
347 !lockdep_is_held(&(pool)->lock), \
348 "pool->manager_mutex or ->lock should be held")
350 #define assert_manager_or_pool_lock(pool) do { } while (0)
353 #define for_each_cpu_worker_pool(pool, cpu) \
354 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
355 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
359 * for_each_pool - iterate through all worker_pools in the system
360 * @pool: iteration cursor
361 * @pi: integer used for iteration
363 * This must be called either with wq_pool_mutex held or sched RCU read
364 * locked. If the pool needs to be used beyond the locking in effect, the
365 * caller is responsible for guaranteeing that the pool stays online.
367 * The if/else clause exists only for the lockdep assertion and can be
370 #define for_each_pool(pool, pi) \
371 idr_for_each_entry(&worker_pool_idr, pool, pi) \
372 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
376 * for_each_pool_worker - iterate through all workers of a worker_pool
377 * @worker: iteration cursor
378 * @wi: integer used for iteration
379 * @pool: worker_pool to iterate workers of
381 * This must be called with either @pool->manager_mutex or ->lock held.
383 * The if/else clause exists only for the lockdep assertion and can be
386 #define for_each_pool_worker(worker, wi, pool) \
387 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
388 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
392 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
393 * @pwq: iteration cursor
394 * @wq: the target workqueue
396 * This must be called either with wq->mutex held or sched RCU read locked.
397 * If the pwq needs to be used beyond the locking in effect, the caller is
398 * responsible for guaranteeing that the pwq stays online.
400 * The if/else clause exists only for the lockdep assertion and can be
403 #define for_each_pwq(pwq, wq) \
404 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
405 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
408 #ifdef CONFIG_DEBUG_OBJECTS_WORK
410 static struct debug_obj_descr work_debug_descr
;
412 static void *work_debug_hint(void *addr
)
414 return ((struct work_struct
*) addr
)->func
;
418 * fixup_init is called when:
419 * - an active object is initialized
421 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
423 struct work_struct
*work
= addr
;
426 case ODEBUG_STATE_ACTIVE
:
427 cancel_work_sync(work
);
428 debug_object_init(work
, &work_debug_descr
);
436 * fixup_activate is called when:
437 * - an active object is activated
438 * - an unknown object is activated (might be a statically initialized object)
440 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
442 struct work_struct
*work
= addr
;
446 case ODEBUG_STATE_NOTAVAILABLE
:
448 * This is not really a fixup. The work struct was
449 * statically initialized. We just make sure that it
450 * is tracked in the object tracker.
452 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
453 debug_object_init(work
, &work_debug_descr
);
454 debug_object_activate(work
, &work_debug_descr
);
460 case ODEBUG_STATE_ACTIVE
:
469 * fixup_free is called when:
470 * - an active object is freed
472 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
474 struct work_struct
*work
= addr
;
477 case ODEBUG_STATE_ACTIVE
:
478 cancel_work_sync(work
);
479 debug_object_free(work
, &work_debug_descr
);
486 static struct debug_obj_descr work_debug_descr
= {
487 .name
= "work_struct",
488 .debug_hint
= work_debug_hint
,
489 .fixup_init
= work_fixup_init
,
490 .fixup_activate
= work_fixup_activate
,
491 .fixup_free
= work_fixup_free
,
494 static inline void debug_work_activate(struct work_struct
*work
)
496 debug_object_activate(work
, &work_debug_descr
);
499 static inline void debug_work_deactivate(struct work_struct
*work
)
501 debug_object_deactivate(work
, &work_debug_descr
);
504 void __init_work(struct work_struct
*work
, int onstack
)
507 debug_object_init_on_stack(work
, &work_debug_descr
);
509 debug_object_init(work
, &work_debug_descr
);
511 EXPORT_SYMBOL_GPL(__init_work
);
513 void destroy_work_on_stack(struct work_struct
*work
)
515 debug_object_free(work
, &work_debug_descr
);
517 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
520 static inline void debug_work_activate(struct work_struct
*work
) { }
521 static inline void debug_work_deactivate(struct work_struct
*work
) { }
524 /* allocate ID and assign it to @pool */
525 static int worker_pool_assign_id(struct worker_pool
*pool
)
529 lockdep_assert_held(&wq_pool_mutex
);
531 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, 0, GFP_KERNEL
);
540 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
541 * @wq: the target workqueue
544 * This must be called either with pwq_lock held or sched RCU read locked.
545 * If the pwq needs to be used beyond the locking in effect, the caller is
546 * responsible for guaranteeing that the pwq stays online.
548 * Return: The unbound pool_workqueue for @node.
550 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
553 assert_rcu_or_wq_mutex(wq
);
554 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
557 static unsigned int work_color_to_flags(int color
)
559 return color
<< WORK_STRUCT_COLOR_SHIFT
;
562 static int get_work_color(struct work_struct
*work
)
564 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
565 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
568 static int work_next_color(int color
)
570 return (color
+ 1) % WORK_NR_COLORS
;
574 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
575 * contain the pointer to the queued pwq. Once execution starts, the flag
576 * is cleared and the high bits contain OFFQ flags and pool ID.
578 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
579 * and clear_work_data() can be used to set the pwq, pool or clear
580 * work->data. These functions should only be called while the work is
581 * owned - ie. while the PENDING bit is set.
583 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
584 * corresponding to a work. Pool is available once the work has been
585 * queued anywhere after initialization until it is sync canceled. pwq is
586 * available only while the work item is queued.
588 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
589 * canceled. While being canceled, a work item may have its PENDING set
590 * but stay off timer and worklist for arbitrarily long and nobody should
591 * try to steal the PENDING bit.
593 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
596 WARN_ON_ONCE(!work_pending(work
));
597 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
600 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
601 unsigned long extra_flags
)
603 set_work_data(work
, (unsigned long)pwq
,
604 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
607 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
610 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
611 WORK_STRUCT_PENDING
);
614 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
618 * The following wmb is paired with the implied mb in
619 * test_and_set_bit(PENDING) and ensures all updates to @work made
620 * here are visible to and precede any updates by the next PENDING
624 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
627 static void clear_work_data(struct work_struct
*work
)
629 smp_wmb(); /* see set_work_pool_and_clear_pending() */
630 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
633 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
635 unsigned long data
= atomic_long_read(&work
->data
);
637 if (data
& WORK_STRUCT_PWQ
)
638 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
644 * get_work_pool - return the worker_pool a given work was associated with
645 * @work: the work item of interest
647 * Pools are created and destroyed under wq_pool_mutex, and allows read
648 * access under sched-RCU read lock. As such, this function should be
649 * called under wq_pool_mutex or with preemption disabled.
651 * All fields of the returned pool are accessible as long as the above
652 * mentioned locking is in effect. If the returned pool needs to be used
653 * beyond the critical section, the caller is responsible for ensuring the
654 * returned pool is and stays online.
656 * Return: The worker_pool @work was last associated with. %NULL if none.
658 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
660 unsigned long data
= atomic_long_read(&work
->data
);
663 assert_rcu_or_pool_mutex();
665 if (data
& WORK_STRUCT_PWQ
)
666 return ((struct pool_workqueue
*)
667 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
669 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
670 if (pool_id
== WORK_OFFQ_POOL_NONE
)
673 return idr_find(&worker_pool_idr
, pool_id
);
677 * get_work_pool_id - return the worker pool ID a given work is associated with
678 * @work: the work item of interest
680 * Return: The worker_pool ID @work was last associated with.
681 * %WORK_OFFQ_POOL_NONE if none.
683 static int get_work_pool_id(struct work_struct
*work
)
685 unsigned long data
= atomic_long_read(&work
->data
);
687 if (data
& WORK_STRUCT_PWQ
)
688 return ((struct pool_workqueue
*)
689 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
691 return data
>> WORK_OFFQ_POOL_SHIFT
;
694 static void mark_work_canceling(struct work_struct
*work
)
696 unsigned long pool_id
= get_work_pool_id(work
);
698 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
699 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
702 static bool work_is_canceling(struct work_struct
*work
)
704 unsigned long data
= atomic_long_read(&work
->data
);
706 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
710 * Policy functions. These define the policies on how the global worker
711 * pools are managed. Unless noted otherwise, these functions assume that
712 * they're being called with pool->lock held.
715 static bool __need_more_worker(struct worker_pool
*pool
)
717 return !atomic_read(&pool
->nr_running
);
721 * Need to wake up a worker? Called from anything but currently
724 * Note that, because unbound workers never contribute to nr_running, this
725 * function will always return %true for unbound pools as long as the
726 * worklist isn't empty.
728 static bool need_more_worker(struct worker_pool
*pool
)
730 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
733 /* Can I start working? Called from busy but !running workers. */
734 static bool may_start_working(struct worker_pool
*pool
)
736 return pool
->nr_idle
;
739 /* Do I need to keep working? Called from currently running workers. */
740 static bool keep_working(struct worker_pool
*pool
)
742 return !list_empty(&pool
->worklist
) &&
743 atomic_read(&pool
->nr_running
) <= 1;
746 /* Do we need a new worker? Called from manager. */
747 static bool need_to_create_worker(struct worker_pool
*pool
)
749 return need_more_worker(pool
) && !may_start_working(pool
);
752 /* Do I need to be the manager? */
753 static bool need_to_manage_workers(struct worker_pool
*pool
)
755 return need_to_create_worker(pool
) ||
756 (pool
->flags
& POOL_MANAGE_WORKERS
);
759 /* Do we have too many workers and should some go away? */
760 static bool too_many_workers(struct worker_pool
*pool
)
762 bool managing
= mutex_is_locked(&pool
->manager_arb
);
763 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
764 int nr_busy
= pool
->nr_workers
- nr_idle
;
767 * nr_idle and idle_list may disagree if idle rebinding is in
768 * progress. Never return %true if idle_list is empty.
770 if (list_empty(&pool
->idle_list
))
773 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
780 /* Return the first worker. Safe with preemption disabled */
781 static struct worker
*first_worker(struct worker_pool
*pool
)
783 if (unlikely(list_empty(&pool
->idle_list
)))
786 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
790 * wake_up_worker - wake up an idle worker
791 * @pool: worker pool to wake worker from
793 * Wake up the first idle worker of @pool.
796 * spin_lock_irq(pool->lock).
798 static void wake_up_worker(struct worker_pool
*pool
)
800 struct worker
*worker
= first_worker(pool
);
803 wake_up_process(worker
->task
);
807 * wq_worker_waking_up - a worker is waking up
808 * @task: task waking up
809 * @cpu: CPU @task is waking up to
811 * This function is called during try_to_wake_up() when a worker is
815 * spin_lock_irq(rq->lock)
817 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
819 struct worker
*worker
= kthread_data(task
);
821 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
822 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
823 atomic_inc(&worker
->pool
->nr_running
);
828 * wq_worker_sleeping - a worker is going to sleep
829 * @task: task going to sleep
830 * @cpu: CPU in question, must be the current CPU number
832 * This function is called during schedule() when a busy worker is
833 * going to sleep. Worker on the same cpu can be woken up by
834 * returning pointer to its task.
837 * spin_lock_irq(rq->lock)
840 * Worker task on @cpu to wake up, %NULL if none.
842 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
844 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
845 struct worker_pool
*pool
;
848 * Rescuers, which may not have all the fields set up like normal
849 * workers, also reach here, let's not access anything before
850 * checking NOT_RUNNING.
852 if (worker
->flags
& WORKER_NOT_RUNNING
)
857 /* this can only happen on the local cpu */
858 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
862 * The counterpart of the following dec_and_test, implied mb,
863 * worklist not empty test sequence is in insert_work().
864 * Please read comment there.
866 * NOT_RUNNING is clear. This means that we're bound to and
867 * running on the local cpu w/ rq lock held and preemption
868 * disabled, which in turn means that none else could be
869 * manipulating idle_list, so dereferencing idle_list without pool
872 if (atomic_dec_and_test(&pool
->nr_running
) &&
873 !list_empty(&pool
->worklist
))
874 to_wakeup
= first_worker(pool
);
875 return to_wakeup
? to_wakeup
->task
: NULL
;
879 * worker_set_flags - set worker flags and adjust nr_running accordingly
881 * @flags: flags to set
882 * @wakeup: wakeup an idle worker if necessary
884 * Set @flags in @worker->flags and adjust nr_running accordingly. If
885 * nr_running becomes zero and @wakeup is %true, an idle worker is
889 * spin_lock_irq(pool->lock)
891 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
894 struct worker_pool
*pool
= worker
->pool
;
896 WARN_ON_ONCE(worker
->task
!= current
);
899 * If transitioning into NOT_RUNNING, adjust nr_running and
900 * wake up an idle worker as necessary if requested by
903 if ((flags
& WORKER_NOT_RUNNING
) &&
904 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
906 if (atomic_dec_and_test(&pool
->nr_running
) &&
907 !list_empty(&pool
->worklist
))
908 wake_up_worker(pool
);
910 atomic_dec(&pool
->nr_running
);
913 worker
->flags
|= flags
;
917 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
919 * @flags: flags to clear
921 * Clear @flags in @worker->flags and adjust nr_running accordingly.
924 * spin_lock_irq(pool->lock)
926 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
928 struct worker_pool
*pool
= worker
->pool
;
929 unsigned int oflags
= worker
->flags
;
931 WARN_ON_ONCE(worker
->task
!= current
);
933 worker
->flags
&= ~flags
;
936 * If transitioning out of NOT_RUNNING, increment nr_running. Note
937 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
938 * of multiple flags, not a single flag.
940 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
941 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
942 atomic_inc(&pool
->nr_running
);
946 * find_worker_executing_work - find worker which is executing a work
947 * @pool: pool of interest
948 * @work: work to find worker for
950 * Find a worker which is executing @work on @pool by searching
951 * @pool->busy_hash which is keyed by the address of @work. For a worker
952 * to match, its current execution should match the address of @work and
953 * its work function. This is to avoid unwanted dependency between
954 * unrelated work executions through a work item being recycled while still
957 * This is a bit tricky. A work item may be freed once its execution
958 * starts and nothing prevents the freed area from being recycled for
959 * another work item. If the same work item address ends up being reused
960 * before the original execution finishes, workqueue will identify the
961 * recycled work item as currently executing and make it wait until the
962 * current execution finishes, introducing an unwanted dependency.
964 * This function checks the work item address and work function to avoid
965 * false positives. Note that this isn't complete as one may construct a
966 * work function which can introduce dependency onto itself through a
967 * recycled work item. Well, if somebody wants to shoot oneself in the
968 * foot that badly, there's only so much we can do, and if such deadlock
969 * actually occurs, it should be easy to locate the culprit work function.
972 * spin_lock_irq(pool->lock).
975 * Pointer to worker which is executing @work if found, %NULL
978 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
979 struct work_struct
*work
)
981 struct worker
*worker
;
983 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
985 if (worker
->current_work
== work
&&
986 worker
->current_func
== work
->func
)
993 * move_linked_works - move linked works to a list
994 * @work: start of series of works to be scheduled
995 * @head: target list to append @work to
996 * @nextp: out paramter for nested worklist walking
998 * Schedule linked works starting from @work to @head. Work series to
999 * be scheduled starts at @work and includes any consecutive work with
1000 * WORK_STRUCT_LINKED set in its predecessor.
1002 * If @nextp is not NULL, it's updated to point to the next work of
1003 * the last scheduled work. This allows move_linked_works() to be
1004 * nested inside outer list_for_each_entry_safe().
1007 * spin_lock_irq(pool->lock).
1009 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1010 struct work_struct
**nextp
)
1012 struct work_struct
*n
;
1015 * Linked worklist will always end before the end of the list,
1016 * use NULL for list head.
1018 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1019 list_move_tail(&work
->entry
, head
);
1020 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1025 * If we're already inside safe list traversal and have moved
1026 * multiple works to the scheduled queue, the next position
1027 * needs to be updated.
1034 * get_pwq - get an extra reference on the specified pool_workqueue
1035 * @pwq: pool_workqueue to get
1037 * Obtain an extra reference on @pwq. The caller should guarantee that
1038 * @pwq has positive refcnt and be holding the matching pool->lock.
1040 static void get_pwq(struct pool_workqueue
*pwq
)
1042 lockdep_assert_held(&pwq
->pool
->lock
);
1043 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1048 * put_pwq - put a pool_workqueue reference
1049 * @pwq: pool_workqueue to put
1051 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1052 * destruction. The caller should be holding the matching pool->lock.
1054 static void put_pwq(struct pool_workqueue
*pwq
)
1056 lockdep_assert_held(&pwq
->pool
->lock
);
1057 if (likely(--pwq
->refcnt
))
1059 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1062 * @pwq can't be released under pool->lock, bounce to
1063 * pwq_unbound_release_workfn(). This never recurses on the same
1064 * pool->lock as this path is taken only for unbound workqueues and
1065 * the release work item is scheduled on a per-cpu workqueue. To
1066 * avoid lockdep warning, unbound pool->locks are given lockdep
1067 * subclass of 1 in get_unbound_pool().
1069 schedule_work(&pwq
->unbound_release_work
);
1073 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1074 * @pwq: pool_workqueue to put (can be %NULL)
1076 * put_pwq() with locking. This function also allows %NULL @pwq.
1078 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1082 * As both pwqs and pools are sched-RCU protected, the
1083 * following lock operations are safe.
1085 spin_lock_irq(&pwq
->pool
->lock
);
1087 spin_unlock_irq(&pwq
->pool
->lock
);
1091 static void pwq_activate_delayed_work(struct work_struct
*work
)
1093 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1095 trace_workqueue_activate_work(work
);
1096 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1097 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1101 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1103 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1104 struct work_struct
, entry
);
1106 pwq_activate_delayed_work(work
);
1110 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1111 * @pwq: pwq of interest
1112 * @color: color of work which left the queue
1114 * A work either has completed or is removed from pending queue,
1115 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1118 * spin_lock_irq(pool->lock).
1120 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1122 /* uncolored work items don't participate in flushing or nr_active */
1123 if (color
== WORK_NO_COLOR
)
1126 pwq
->nr_in_flight
[color
]--;
1129 if (!list_empty(&pwq
->delayed_works
)) {
1130 /* one down, submit a delayed one */
1131 if (pwq
->nr_active
< pwq
->max_active
)
1132 pwq_activate_first_delayed(pwq
);
1135 /* is flush in progress and are we at the flushing tip? */
1136 if (likely(pwq
->flush_color
!= color
))
1139 /* are there still in-flight works? */
1140 if (pwq
->nr_in_flight
[color
])
1143 /* this pwq is done, clear flush_color */
1144 pwq
->flush_color
= -1;
1147 * If this was the last pwq, wake up the first flusher. It
1148 * will handle the rest.
1150 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1151 complete(&pwq
->wq
->first_flusher
->done
);
1157 * try_to_grab_pending - steal work item from worklist and disable irq
1158 * @work: work item to steal
1159 * @is_dwork: @work is a delayed_work
1160 * @flags: place to store irq state
1162 * Try to grab PENDING bit of @work. This function can handle @work in any
1163 * stable state - idle, on timer or on worklist.
1166 * 1 if @work was pending and we successfully stole PENDING
1167 * 0 if @work was idle and we claimed PENDING
1168 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1169 * -ENOENT if someone else is canceling @work, this state may persist
1170 * for arbitrarily long
1173 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1174 * interrupted while holding PENDING and @work off queue, irq must be
1175 * disabled on entry. This, combined with delayed_work->timer being
1176 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1178 * On successful return, >= 0, irq is disabled and the caller is
1179 * responsible for releasing it using local_irq_restore(*@flags).
1181 * This function is safe to call from any context including IRQ handler.
1183 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1184 unsigned long *flags
)
1186 struct worker_pool
*pool
;
1187 struct pool_workqueue
*pwq
;
1189 local_irq_save(*flags
);
1191 /* try to steal the timer if it exists */
1193 struct delayed_work
*dwork
= to_delayed_work(work
);
1196 * dwork->timer is irqsafe. If del_timer() fails, it's
1197 * guaranteed that the timer is not queued anywhere and not
1198 * running on the local CPU.
1200 if (likely(del_timer(&dwork
->timer
)))
1204 /* try to claim PENDING the normal way */
1205 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1209 * The queueing is in progress, or it is already queued. Try to
1210 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1212 pool
= get_work_pool(work
);
1216 spin_lock(&pool
->lock
);
1218 * work->data is guaranteed to point to pwq only while the work
1219 * item is queued on pwq->wq, and both updating work->data to point
1220 * to pwq on queueing and to pool on dequeueing are done under
1221 * pwq->pool->lock. This in turn guarantees that, if work->data
1222 * points to pwq which is associated with a locked pool, the work
1223 * item is currently queued on that pool.
1225 pwq
= get_work_pwq(work
);
1226 if (pwq
&& pwq
->pool
== pool
) {
1227 debug_work_deactivate(work
);
1230 * A delayed work item cannot be grabbed directly because
1231 * it might have linked NO_COLOR work items which, if left
1232 * on the delayed_list, will confuse pwq->nr_active
1233 * management later on and cause stall. Make sure the work
1234 * item is activated before grabbing.
1236 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1237 pwq_activate_delayed_work(work
);
1239 list_del_init(&work
->entry
);
1240 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1242 /* work->data points to pwq iff queued, point to pool */
1243 set_work_pool_and_keep_pending(work
, pool
->id
);
1245 spin_unlock(&pool
->lock
);
1248 spin_unlock(&pool
->lock
);
1250 local_irq_restore(*flags
);
1251 if (work_is_canceling(work
))
1258 * insert_work - insert a work into a pool
1259 * @pwq: pwq @work belongs to
1260 * @work: work to insert
1261 * @head: insertion point
1262 * @extra_flags: extra WORK_STRUCT_* flags to set
1264 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1265 * work_struct flags.
1268 * spin_lock_irq(pool->lock).
1270 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1271 struct list_head
*head
, unsigned int extra_flags
)
1273 struct worker_pool
*pool
= pwq
->pool
;
1275 /* we own @work, set data and link */
1276 set_work_pwq(work
, pwq
, extra_flags
);
1277 list_add_tail(&work
->entry
, head
);
1281 * Ensure either wq_worker_sleeping() sees the above
1282 * list_add_tail() or we see zero nr_running to avoid workers lying
1283 * around lazily while there are works to be processed.
1287 if (__need_more_worker(pool
))
1288 wake_up_worker(pool
);
1292 * Test whether @work is being queued from another work executing on the
1295 static bool is_chained_work(struct workqueue_struct
*wq
)
1297 struct worker
*worker
;
1299 worker
= current_wq_worker();
1301 * Return %true iff I'm a worker execuing a work item on @wq. If
1302 * I'm @worker, it's safe to dereference it without locking.
1304 return worker
&& worker
->current_pwq
->wq
== wq
;
1307 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1308 struct work_struct
*work
)
1310 struct pool_workqueue
*pwq
;
1311 struct worker_pool
*last_pool
;
1312 struct list_head
*worklist
;
1313 unsigned int work_flags
;
1314 unsigned int req_cpu
= cpu
;
1317 * While a work item is PENDING && off queue, a task trying to
1318 * steal the PENDING will busy-loop waiting for it to either get
1319 * queued or lose PENDING. Grabbing PENDING and queueing should
1320 * happen with IRQ disabled.
1322 WARN_ON_ONCE(!irqs_disabled());
1324 debug_work_activate(work
);
1326 /* if draining, only works from the same workqueue are allowed */
1327 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1328 WARN_ON_ONCE(!is_chained_work(wq
)))
1331 if (req_cpu
== WORK_CPU_UNBOUND
)
1332 cpu
= raw_smp_processor_id();
1334 /* pwq which will be used unless @work is executing elsewhere */
1335 if (!(wq
->flags
& WQ_UNBOUND
))
1336 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1338 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1341 * If @work was previously on a different pool, it might still be
1342 * running there, in which case the work needs to be queued on that
1343 * pool to guarantee non-reentrancy.
1345 last_pool
= get_work_pool(work
);
1346 if (last_pool
&& last_pool
!= pwq
->pool
) {
1347 struct worker
*worker
;
1349 spin_lock(&last_pool
->lock
);
1351 worker
= find_worker_executing_work(last_pool
, work
);
1353 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1354 pwq
= worker
->current_pwq
;
1356 /* meh... not running there, queue here */
1357 spin_unlock(&last_pool
->lock
);
1358 spin_lock(&pwq
->pool
->lock
);
1361 spin_lock(&pwq
->pool
->lock
);
1365 * pwq is determined and locked. For unbound pools, we could have
1366 * raced with pwq release and it could already be dead. If its
1367 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1368 * without another pwq replacing it in the numa_pwq_tbl or while
1369 * work items are executing on it, so the retrying is guaranteed to
1370 * make forward-progress.
1372 if (unlikely(!pwq
->refcnt
)) {
1373 if (wq
->flags
& WQ_UNBOUND
) {
1374 spin_unlock(&pwq
->pool
->lock
);
1379 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1383 /* pwq determined, queue */
1384 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1386 if (WARN_ON(!list_empty(&work
->entry
))) {
1387 spin_unlock(&pwq
->pool
->lock
);
1391 pwq
->nr_in_flight
[pwq
->work_color
]++;
1392 work_flags
= work_color_to_flags(pwq
->work_color
);
1394 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1395 trace_workqueue_activate_work(work
);
1397 worklist
= &pwq
->pool
->worklist
;
1399 work_flags
|= WORK_STRUCT_DELAYED
;
1400 worklist
= &pwq
->delayed_works
;
1403 insert_work(pwq
, work
, worklist
, work_flags
);
1405 spin_unlock(&pwq
->pool
->lock
);
1409 * queue_work_on - queue work on specific cpu
1410 * @cpu: CPU number to execute work on
1411 * @wq: workqueue to use
1412 * @work: work to queue
1414 * We queue the work to a specific CPU, the caller must ensure it
1417 * Return: %false if @work was already on a queue, %true otherwise.
1419 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1420 struct work_struct
*work
)
1423 unsigned long flags
;
1425 local_irq_save(flags
);
1427 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1428 __queue_work(cpu
, wq
, work
);
1432 local_irq_restore(flags
);
1435 EXPORT_SYMBOL(queue_work_on
);
1437 void delayed_work_timer_fn(unsigned long __data
)
1439 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1441 /* should have been called from irqsafe timer with irq already off */
1442 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1444 EXPORT_SYMBOL(delayed_work_timer_fn
);
1446 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1447 struct delayed_work
*dwork
, unsigned long delay
)
1449 struct timer_list
*timer
= &dwork
->timer
;
1450 struct work_struct
*work
= &dwork
->work
;
1452 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1453 timer
->data
!= (unsigned long)dwork
);
1454 WARN_ON_ONCE(timer_pending(timer
));
1455 WARN_ON_ONCE(!list_empty(&work
->entry
));
1458 * If @delay is 0, queue @dwork->work immediately. This is for
1459 * both optimization and correctness. The earliest @timer can
1460 * expire is on the closest next tick and delayed_work users depend
1461 * on that there's no such delay when @delay is 0.
1464 __queue_work(cpu
, wq
, &dwork
->work
);
1468 timer_stats_timer_set_start_info(&dwork
->timer
);
1472 timer
->expires
= jiffies
+ delay
;
1474 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1475 add_timer_on(timer
, cpu
);
1481 * queue_delayed_work_on - queue work on specific CPU after delay
1482 * @cpu: CPU number to execute work on
1483 * @wq: workqueue to use
1484 * @dwork: work to queue
1485 * @delay: number of jiffies to wait before queueing
1487 * Return: %false if @work was already on a queue, %true otherwise. If
1488 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1491 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1492 struct delayed_work
*dwork
, unsigned long delay
)
1494 struct work_struct
*work
= &dwork
->work
;
1496 unsigned long flags
;
1498 /* read the comment in __queue_work() */
1499 local_irq_save(flags
);
1501 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1502 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1506 local_irq_restore(flags
);
1509 EXPORT_SYMBOL(queue_delayed_work_on
);
1512 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1513 * @cpu: CPU number to execute work on
1514 * @wq: workqueue to use
1515 * @dwork: work to queue
1516 * @delay: number of jiffies to wait before queueing
1518 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1519 * modify @dwork's timer so that it expires after @delay. If @delay is
1520 * zero, @work is guaranteed to be scheduled immediately regardless of its
1523 * Return: %false if @dwork was idle and queued, %true if @dwork was
1524 * pending and its timer was modified.
1526 * This function is safe to call from any context including IRQ handler.
1527 * See try_to_grab_pending() for details.
1529 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1530 struct delayed_work
*dwork
, unsigned long delay
)
1532 unsigned long flags
;
1536 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1537 } while (unlikely(ret
== -EAGAIN
));
1539 if (likely(ret
>= 0)) {
1540 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1541 local_irq_restore(flags
);
1544 /* -ENOENT from try_to_grab_pending() becomes %true */
1547 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1550 * worker_enter_idle - enter idle state
1551 * @worker: worker which is entering idle state
1553 * @worker is entering idle state. Update stats and idle timer if
1557 * spin_lock_irq(pool->lock).
1559 static void worker_enter_idle(struct worker
*worker
)
1561 struct worker_pool
*pool
= worker
->pool
;
1563 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1564 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1565 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1568 /* can't use worker_set_flags(), also called from start_worker() */
1569 worker
->flags
|= WORKER_IDLE
;
1571 worker
->last_active
= jiffies
;
1573 /* idle_list is LIFO */
1574 list_add(&worker
->entry
, &pool
->idle_list
);
1576 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1577 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1580 * Sanity check nr_running. Because wq_unbind_fn() releases
1581 * pool->lock between setting %WORKER_UNBOUND and zapping
1582 * nr_running, the warning may trigger spuriously. Check iff
1583 * unbind is not in progress.
1585 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1586 pool
->nr_workers
== pool
->nr_idle
&&
1587 atomic_read(&pool
->nr_running
));
1591 * worker_leave_idle - leave idle state
1592 * @worker: worker which is leaving idle state
1594 * @worker is leaving idle state. Update stats.
1597 * spin_lock_irq(pool->lock).
1599 static void worker_leave_idle(struct worker
*worker
)
1601 struct worker_pool
*pool
= worker
->pool
;
1603 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1605 worker_clr_flags(worker
, WORKER_IDLE
);
1607 list_del_init(&worker
->entry
);
1611 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1612 * @pool: target worker_pool
1614 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1616 * Works which are scheduled while the cpu is online must at least be
1617 * scheduled to a worker which is bound to the cpu so that if they are
1618 * flushed from cpu callbacks while cpu is going down, they are
1619 * guaranteed to execute on the cpu.
1621 * This function is to be used by unbound workers and rescuers to bind
1622 * themselves to the target cpu and may race with cpu going down or
1623 * coming online. kthread_bind() can't be used because it may put the
1624 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1625 * verbatim as it's best effort and blocking and pool may be
1626 * [dis]associated in the meantime.
1628 * This function tries set_cpus_allowed() and locks pool and verifies the
1629 * binding against %POOL_DISASSOCIATED which is set during
1630 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1631 * enters idle state or fetches works without dropping lock, it can
1632 * guarantee the scheduling requirement described in the first paragraph.
1635 * Might sleep. Called without any lock but returns with pool->lock
1639 * %true if the associated pool is online (@worker is successfully
1640 * bound), %false if offline.
1642 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1643 __acquires(&pool
->lock
)
1647 * The following call may fail, succeed or succeed
1648 * without actually migrating the task to the cpu if
1649 * it races with cpu hotunplug operation. Verify
1650 * against POOL_DISASSOCIATED.
1652 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1653 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1655 spin_lock_irq(&pool
->lock
);
1656 if (pool
->flags
& POOL_DISASSOCIATED
)
1658 if (task_cpu(current
) == pool
->cpu
&&
1659 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1661 spin_unlock_irq(&pool
->lock
);
1664 * We've raced with CPU hot[un]plug. Give it a breather
1665 * and retry migration. cond_resched() is required here;
1666 * otherwise, we might deadlock against cpu_stop trying to
1667 * bring down the CPU on non-preemptive kernel.
1674 static struct worker
*alloc_worker(void)
1676 struct worker
*worker
;
1678 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1680 INIT_LIST_HEAD(&worker
->entry
);
1681 INIT_LIST_HEAD(&worker
->scheduled
);
1682 /* on creation a worker is in !idle && prep state */
1683 worker
->flags
= WORKER_PREP
;
1689 * create_worker - create a new workqueue worker
1690 * @pool: pool the new worker will belong to
1692 * Create a new worker which is bound to @pool. The returned worker
1693 * can be started by calling start_worker() or destroyed using
1697 * Might sleep. Does GFP_KERNEL allocations.
1700 * Pointer to the newly created worker.
1702 static struct worker
*create_worker(struct worker_pool
*pool
)
1704 struct worker
*worker
= NULL
;
1708 lockdep_assert_held(&pool
->manager_mutex
);
1711 * ID is needed to determine kthread name. Allocate ID first
1712 * without installing the pointer.
1714 idr_preload(GFP_KERNEL
);
1715 spin_lock_irq(&pool
->lock
);
1717 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1719 spin_unlock_irq(&pool
->lock
);
1724 worker
= alloc_worker();
1728 worker
->pool
= pool
;
1732 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1733 pool
->attrs
->nice
< 0 ? "H" : "");
1735 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1737 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1738 "kworker/%s", id_buf
);
1739 if (IS_ERR(worker
->task
))
1742 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1744 /* prevent userland from meddling with cpumask of workqueue workers */
1745 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1748 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1749 * online CPUs. It'll be re-applied when any of the CPUs come up.
1751 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1754 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1755 * remains stable across this function. See the comments above the
1756 * flag definition for details.
1758 if (pool
->flags
& POOL_DISASSOCIATED
)
1759 worker
->flags
|= WORKER_UNBOUND
;
1761 /* successful, commit the pointer to idr */
1762 spin_lock_irq(&pool
->lock
);
1763 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1764 spin_unlock_irq(&pool
->lock
);
1770 spin_lock_irq(&pool
->lock
);
1771 idr_remove(&pool
->worker_idr
, id
);
1772 spin_unlock_irq(&pool
->lock
);
1779 * start_worker - start a newly created worker
1780 * @worker: worker to start
1782 * Make the pool aware of @worker and start it.
1785 * spin_lock_irq(pool->lock).
1787 static void start_worker(struct worker
*worker
)
1789 worker
->flags
|= WORKER_STARTED
;
1790 worker
->pool
->nr_workers
++;
1791 worker_enter_idle(worker
);
1792 wake_up_process(worker
->task
);
1796 * create_and_start_worker - create and start a worker for a pool
1797 * @pool: the target pool
1799 * Grab the managership of @pool and create and start a new worker for it.
1801 * Return: 0 on success. A negative error code otherwise.
1803 static int create_and_start_worker(struct worker_pool
*pool
)
1805 struct worker
*worker
;
1807 mutex_lock(&pool
->manager_mutex
);
1809 worker
= create_worker(pool
);
1811 spin_lock_irq(&pool
->lock
);
1812 start_worker(worker
);
1813 spin_unlock_irq(&pool
->lock
);
1816 mutex_unlock(&pool
->manager_mutex
);
1818 return worker
? 0 : -ENOMEM
;
1822 * destroy_worker - destroy a workqueue worker
1823 * @worker: worker to be destroyed
1825 * Destroy @worker and adjust @pool stats accordingly.
1828 * spin_lock_irq(pool->lock) which is released and regrabbed.
1830 static void destroy_worker(struct worker
*worker
)
1832 struct worker_pool
*pool
= worker
->pool
;
1834 lockdep_assert_held(&pool
->manager_mutex
);
1835 lockdep_assert_held(&pool
->lock
);
1837 /* sanity check frenzy */
1838 if (WARN_ON(worker
->current_work
) ||
1839 WARN_ON(!list_empty(&worker
->scheduled
)))
1842 if (worker
->flags
& WORKER_STARTED
)
1844 if (worker
->flags
& WORKER_IDLE
)
1847 list_del_init(&worker
->entry
);
1848 worker
->flags
|= WORKER_DIE
;
1850 idr_remove(&pool
->worker_idr
, worker
->id
);
1852 spin_unlock_irq(&pool
->lock
);
1854 kthread_stop(worker
->task
);
1857 spin_lock_irq(&pool
->lock
);
1860 static void idle_worker_timeout(unsigned long __pool
)
1862 struct worker_pool
*pool
= (void *)__pool
;
1864 spin_lock_irq(&pool
->lock
);
1866 if (too_many_workers(pool
)) {
1867 struct worker
*worker
;
1868 unsigned long expires
;
1870 /* idle_list is kept in LIFO order, check the last one */
1871 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1872 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1874 if (time_before(jiffies
, expires
))
1875 mod_timer(&pool
->idle_timer
, expires
);
1877 /* it's been idle for too long, wake up manager */
1878 pool
->flags
|= POOL_MANAGE_WORKERS
;
1879 wake_up_worker(pool
);
1883 spin_unlock_irq(&pool
->lock
);
1886 static void send_mayday(struct work_struct
*work
)
1888 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1889 struct workqueue_struct
*wq
= pwq
->wq
;
1891 lockdep_assert_held(&wq_mayday_lock
);
1896 /* mayday mayday mayday */
1897 if (list_empty(&pwq
->mayday_node
)) {
1898 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1899 wake_up_process(wq
->rescuer
->task
);
1903 static void pool_mayday_timeout(unsigned long __pool
)
1905 struct worker_pool
*pool
= (void *)__pool
;
1906 struct work_struct
*work
;
1908 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1909 spin_lock(&pool
->lock
);
1911 if (need_to_create_worker(pool
)) {
1913 * We've been trying to create a new worker but
1914 * haven't been successful. We might be hitting an
1915 * allocation deadlock. Send distress signals to
1918 list_for_each_entry(work
, &pool
->worklist
, entry
)
1922 spin_unlock(&pool
->lock
);
1923 spin_unlock_irq(&wq_mayday_lock
);
1925 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1929 * maybe_create_worker - create a new worker if necessary
1930 * @pool: pool to create a new worker for
1932 * Create a new worker for @pool if necessary. @pool is guaranteed to
1933 * have at least one idle worker on return from this function. If
1934 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1935 * sent to all rescuers with works scheduled on @pool to resolve
1936 * possible allocation deadlock.
1938 * On return, need_to_create_worker() is guaranteed to be %false and
1939 * may_start_working() %true.
1942 * spin_lock_irq(pool->lock) which may be released and regrabbed
1943 * multiple times. Does GFP_KERNEL allocations. Called only from
1947 * %false if no action was taken and pool->lock stayed locked, %true
1950 static bool maybe_create_worker(struct worker_pool
*pool
)
1951 __releases(&pool
->lock
)
1952 __acquires(&pool
->lock
)
1954 if (!need_to_create_worker(pool
))
1957 spin_unlock_irq(&pool
->lock
);
1959 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1960 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1963 struct worker
*worker
;
1965 worker
= create_worker(pool
);
1967 del_timer_sync(&pool
->mayday_timer
);
1968 spin_lock_irq(&pool
->lock
);
1969 start_worker(worker
);
1970 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1975 if (!need_to_create_worker(pool
))
1978 __set_current_state(TASK_INTERRUPTIBLE
);
1979 schedule_timeout(CREATE_COOLDOWN
);
1981 if (!need_to_create_worker(pool
))
1985 del_timer_sync(&pool
->mayday_timer
);
1986 spin_lock_irq(&pool
->lock
);
1987 if (need_to_create_worker(pool
))
1993 * maybe_destroy_worker - destroy workers which have been idle for a while
1994 * @pool: pool to destroy workers for
1996 * Destroy @pool workers which have been idle for longer than
1997 * IDLE_WORKER_TIMEOUT.
2000 * spin_lock_irq(pool->lock) which may be released and regrabbed
2001 * multiple times. Called only from manager.
2004 * %false if no action was taken and pool->lock stayed locked, %true
2007 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2011 while (too_many_workers(pool
)) {
2012 struct worker
*worker
;
2013 unsigned long expires
;
2015 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2016 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2018 if (time_before(jiffies
, expires
)) {
2019 mod_timer(&pool
->idle_timer
, expires
);
2023 destroy_worker(worker
);
2031 * manage_workers - manage worker pool
2034 * Assume the manager role and manage the worker pool @worker belongs
2035 * to. At any given time, there can be only zero or one manager per
2036 * pool. The exclusion is handled automatically by this function.
2038 * The caller can safely start processing works on false return. On
2039 * true return, it's guaranteed that need_to_create_worker() is false
2040 * and may_start_working() is true.
2043 * spin_lock_irq(pool->lock) which may be released and regrabbed
2044 * multiple times. Does GFP_KERNEL allocations.
2047 * %false if the pool don't need management and the caller can safely start
2048 * processing works, %true indicates that the function released pool->lock
2049 * and reacquired it to perform some management function and that the
2050 * conditions that the caller verified while holding the lock before
2051 * calling the function might no longer be true.
2053 static bool manage_workers(struct worker
*worker
)
2055 struct worker_pool
*pool
= worker
->pool
;
2059 * Managership is governed by two mutexes - manager_arb and
2060 * manager_mutex. manager_arb handles arbitration of manager role.
2061 * Anyone who successfully grabs manager_arb wins the arbitration
2062 * and becomes the manager. mutex_trylock() on pool->manager_arb
2063 * failure while holding pool->lock reliably indicates that someone
2064 * else is managing the pool and the worker which failed trylock
2065 * can proceed to executing work items. This means that anyone
2066 * grabbing manager_arb is responsible for actually performing
2067 * manager duties. If manager_arb is grabbed and released without
2068 * actual management, the pool may stall indefinitely.
2070 * manager_mutex is used for exclusion of actual management
2071 * operations. The holder of manager_mutex can be sure that none
2072 * of management operations, including creation and destruction of
2073 * workers, won't take place until the mutex is released. Because
2074 * manager_mutex doesn't interfere with manager role arbitration,
2075 * it is guaranteed that the pool's management, while may be
2076 * delayed, won't be disturbed by someone else grabbing
2079 if (!mutex_trylock(&pool
->manager_arb
))
2083 * With manager arbitration won, manager_mutex would be free in
2084 * most cases. trylock first without dropping @pool->lock.
2086 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2087 spin_unlock_irq(&pool
->lock
);
2088 mutex_lock(&pool
->manager_mutex
);
2089 spin_lock_irq(&pool
->lock
);
2093 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2096 * Destroy and then create so that may_start_working() is true
2099 ret
|= maybe_destroy_workers(pool
);
2100 ret
|= maybe_create_worker(pool
);
2102 mutex_unlock(&pool
->manager_mutex
);
2103 mutex_unlock(&pool
->manager_arb
);
2108 * process_one_work - process single work
2110 * @work: work to process
2112 * Process @work. This function contains all the logics necessary to
2113 * process a single work including synchronization against and
2114 * interaction with other workers on the same cpu, queueing and
2115 * flushing. As long as context requirement is met, any worker can
2116 * call this function to process a work.
2119 * spin_lock_irq(pool->lock) which is released and regrabbed.
2121 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2122 __releases(&pool
->lock
)
2123 __acquires(&pool
->lock
)
2125 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2126 struct worker_pool
*pool
= worker
->pool
;
2127 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2129 struct worker
*collision
;
2130 #ifdef CONFIG_LOCKDEP
2132 * It is permissible to free the struct work_struct from
2133 * inside the function that is called from it, this we need to
2134 * take into account for lockdep too. To avoid bogus "held
2135 * lock freed" warnings as well as problems when looking into
2136 * work->lockdep_map, make a copy and use that here.
2138 struct lockdep_map lockdep_map
;
2140 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2143 * Ensure we're on the correct CPU. DISASSOCIATED test is
2144 * necessary to avoid spurious warnings from rescuers servicing the
2145 * unbound or a disassociated pool.
2147 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2148 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2149 raw_smp_processor_id() != pool
->cpu
);
2152 * A single work shouldn't be executed concurrently by
2153 * multiple workers on a single cpu. Check whether anyone is
2154 * already processing the work. If so, defer the work to the
2155 * currently executing one.
2157 collision
= find_worker_executing_work(pool
, work
);
2158 if (unlikely(collision
)) {
2159 move_linked_works(work
, &collision
->scheduled
, NULL
);
2163 /* claim and dequeue */
2164 debug_work_deactivate(work
);
2165 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2166 worker
->current_work
= work
;
2167 worker
->current_func
= work
->func
;
2168 worker
->current_pwq
= pwq
;
2169 work_color
= get_work_color(work
);
2171 list_del_init(&work
->entry
);
2174 * CPU intensive works don't participate in concurrency
2175 * management. They're the scheduler's responsibility.
2177 if (unlikely(cpu_intensive
))
2178 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2181 * Unbound pool isn't concurrency managed and work items should be
2182 * executed ASAP. Wake up another worker if necessary.
2184 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2185 wake_up_worker(pool
);
2188 * Record the last pool and clear PENDING which should be the last
2189 * update to @work. Also, do this inside @pool->lock so that
2190 * PENDING and queued state changes happen together while IRQ is
2193 set_work_pool_and_clear_pending(work
, pool
->id
);
2195 spin_unlock_irq(&pool
->lock
);
2197 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2198 lock_map_acquire(&lockdep_map
);
2199 trace_workqueue_execute_start(work
);
2200 worker
->current_func(work
);
2202 * While we must be careful to not use "work" after this, the trace
2203 * point will only record its address.
2205 trace_workqueue_execute_end(work
);
2206 lock_map_release(&lockdep_map
);
2207 lock_map_release(&pwq
->wq
->lockdep_map
);
2209 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2210 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2211 " last function: %pf\n",
2212 current
->comm
, preempt_count(), task_pid_nr(current
),
2213 worker
->current_func
);
2214 debug_show_held_locks(current
);
2219 * The following prevents a kworker from hogging CPU on !PREEMPT
2220 * kernels, where a requeueing work item waiting for something to
2221 * happen could deadlock with stop_machine as such work item could
2222 * indefinitely requeue itself while all other CPUs are trapped in
2227 spin_lock_irq(&pool
->lock
);
2229 /* clear cpu intensive status */
2230 if (unlikely(cpu_intensive
))
2231 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2233 /* we're done with it, release */
2234 hash_del(&worker
->hentry
);
2235 worker
->current_work
= NULL
;
2236 worker
->current_func
= NULL
;
2237 worker
->current_pwq
= NULL
;
2238 worker
->desc_valid
= false;
2239 pwq_dec_nr_in_flight(pwq
, work_color
);
2243 * process_scheduled_works - process scheduled works
2246 * Process all scheduled works. Please note that the scheduled list
2247 * may change while processing a work, so this function repeatedly
2248 * fetches a work from the top and executes it.
2251 * spin_lock_irq(pool->lock) which may be released and regrabbed
2254 static void process_scheduled_works(struct worker
*worker
)
2256 while (!list_empty(&worker
->scheduled
)) {
2257 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2258 struct work_struct
, entry
);
2259 process_one_work(worker
, work
);
2264 * worker_thread - the worker thread function
2267 * The worker thread function. All workers belong to a worker_pool -
2268 * either a per-cpu one or dynamic unbound one. These workers process all
2269 * work items regardless of their specific target workqueue. The only
2270 * exception is work items which belong to workqueues with a rescuer which
2271 * will be explained in rescuer_thread().
2275 static int worker_thread(void *__worker
)
2277 struct worker
*worker
= __worker
;
2278 struct worker_pool
*pool
= worker
->pool
;
2280 /* tell the scheduler that this is a workqueue worker */
2281 worker
->task
->flags
|= PF_WQ_WORKER
;
2283 spin_lock_irq(&pool
->lock
);
2285 /* am I supposed to die? */
2286 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2287 spin_unlock_irq(&pool
->lock
);
2288 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2289 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2293 worker_leave_idle(worker
);
2295 /* no more worker necessary? */
2296 if (!need_more_worker(pool
))
2299 /* do we need to manage? */
2300 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2304 * ->scheduled list can only be filled while a worker is
2305 * preparing to process a work or actually processing it.
2306 * Make sure nobody diddled with it while I was sleeping.
2308 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2311 * Finish PREP stage. We're guaranteed to have at least one idle
2312 * worker or that someone else has already assumed the manager
2313 * role. This is where @worker starts participating in concurrency
2314 * management if applicable and concurrency management is restored
2315 * after being rebound. See rebind_workers() for details.
2317 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2320 struct work_struct
*work
=
2321 list_first_entry(&pool
->worklist
,
2322 struct work_struct
, entry
);
2324 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2325 /* optimization path, not strictly necessary */
2326 process_one_work(worker
, work
);
2327 if (unlikely(!list_empty(&worker
->scheduled
)))
2328 process_scheduled_works(worker
);
2330 move_linked_works(work
, &worker
->scheduled
, NULL
);
2331 process_scheduled_works(worker
);
2333 } while (keep_working(pool
));
2335 worker_set_flags(worker
, WORKER_PREP
, false);
2337 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2341 * pool->lock is held and there's no work to process and no need to
2342 * manage, sleep. Workers are woken up only while holding
2343 * pool->lock or from local cpu, so setting the current state
2344 * before releasing pool->lock is enough to prevent losing any
2347 worker_enter_idle(worker
);
2348 __set_current_state(TASK_INTERRUPTIBLE
);
2349 spin_unlock_irq(&pool
->lock
);
2355 * rescuer_thread - the rescuer thread function
2358 * Workqueue rescuer thread function. There's one rescuer for each
2359 * workqueue which has WQ_MEM_RECLAIM set.
2361 * Regular work processing on a pool may block trying to create a new
2362 * worker which uses GFP_KERNEL allocation which has slight chance of
2363 * developing into deadlock if some works currently on the same queue
2364 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2365 * the problem rescuer solves.
2367 * When such condition is possible, the pool summons rescuers of all
2368 * workqueues which have works queued on the pool and let them process
2369 * those works so that forward progress can be guaranteed.
2371 * This should happen rarely.
2375 static int rescuer_thread(void *__rescuer
)
2377 struct worker
*rescuer
= __rescuer
;
2378 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2379 struct list_head
*scheduled
= &rescuer
->scheduled
;
2381 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2384 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2385 * doesn't participate in concurrency management.
2387 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2389 set_current_state(TASK_INTERRUPTIBLE
);
2391 if (kthread_should_stop()) {
2392 __set_current_state(TASK_RUNNING
);
2393 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2397 /* see whether any pwq is asking for help */
2398 spin_lock_irq(&wq_mayday_lock
);
2400 while (!list_empty(&wq
->maydays
)) {
2401 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2402 struct pool_workqueue
, mayday_node
);
2403 struct worker_pool
*pool
= pwq
->pool
;
2404 struct work_struct
*work
, *n
;
2406 __set_current_state(TASK_RUNNING
);
2407 list_del_init(&pwq
->mayday_node
);
2409 spin_unlock_irq(&wq_mayday_lock
);
2411 /* migrate to the target cpu if possible */
2412 worker_maybe_bind_and_lock(pool
);
2413 rescuer
->pool
= pool
;
2416 * Slurp in all works issued via this workqueue and
2419 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2420 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2421 if (get_work_pwq(work
) == pwq
)
2422 move_linked_works(work
, scheduled
, &n
);
2424 process_scheduled_works(rescuer
);
2427 * Leave this pool. If keep_working() is %true, notify a
2428 * regular worker; otherwise, we end up with 0 concurrency
2429 * and stalling the execution.
2431 if (keep_working(pool
))
2432 wake_up_worker(pool
);
2434 rescuer
->pool
= NULL
;
2435 spin_unlock(&pool
->lock
);
2436 spin_lock(&wq_mayday_lock
);
2439 spin_unlock_irq(&wq_mayday_lock
);
2441 /* rescuers should never participate in concurrency management */
2442 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2448 struct work_struct work
;
2449 struct completion done
;
2452 static void wq_barrier_func(struct work_struct
*work
)
2454 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2455 complete(&barr
->done
);
2459 * insert_wq_barrier - insert a barrier work
2460 * @pwq: pwq to insert barrier into
2461 * @barr: wq_barrier to insert
2462 * @target: target work to attach @barr to
2463 * @worker: worker currently executing @target, NULL if @target is not executing
2465 * @barr is linked to @target such that @barr is completed only after
2466 * @target finishes execution. Please note that the ordering
2467 * guarantee is observed only with respect to @target and on the local
2470 * Currently, a queued barrier can't be canceled. This is because
2471 * try_to_grab_pending() can't determine whether the work to be
2472 * grabbed is at the head of the queue and thus can't clear LINKED
2473 * flag of the previous work while there must be a valid next work
2474 * after a work with LINKED flag set.
2476 * Note that when @worker is non-NULL, @target may be modified
2477 * underneath us, so we can't reliably determine pwq from @target.
2480 * spin_lock_irq(pool->lock).
2482 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2483 struct wq_barrier
*barr
,
2484 struct work_struct
*target
, struct worker
*worker
)
2486 struct list_head
*head
;
2487 unsigned int linked
= 0;
2490 * debugobject calls are safe here even with pool->lock locked
2491 * as we know for sure that this will not trigger any of the
2492 * checks and call back into the fixup functions where we
2495 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2496 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2497 init_completion(&barr
->done
);
2500 * If @target is currently being executed, schedule the
2501 * barrier to the worker; otherwise, put it after @target.
2504 head
= worker
->scheduled
.next
;
2506 unsigned long *bits
= work_data_bits(target
);
2508 head
= target
->entry
.next
;
2509 /* there can already be other linked works, inherit and set */
2510 linked
= *bits
& WORK_STRUCT_LINKED
;
2511 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2514 debug_work_activate(&barr
->work
);
2515 insert_work(pwq
, &barr
->work
, head
,
2516 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2520 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2521 * @wq: workqueue being flushed
2522 * @flush_color: new flush color, < 0 for no-op
2523 * @work_color: new work color, < 0 for no-op
2525 * Prepare pwqs for workqueue flushing.
2527 * If @flush_color is non-negative, flush_color on all pwqs should be
2528 * -1. If no pwq has in-flight commands at the specified color, all
2529 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2530 * has in flight commands, its pwq->flush_color is set to
2531 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2532 * wakeup logic is armed and %true is returned.
2534 * The caller should have initialized @wq->first_flusher prior to
2535 * calling this function with non-negative @flush_color. If
2536 * @flush_color is negative, no flush color update is done and %false
2539 * If @work_color is non-negative, all pwqs should have the same
2540 * work_color which is previous to @work_color and all will be
2541 * advanced to @work_color.
2544 * mutex_lock(wq->mutex).
2547 * %true if @flush_color >= 0 and there's something to flush. %false
2550 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2551 int flush_color
, int work_color
)
2554 struct pool_workqueue
*pwq
;
2556 if (flush_color
>= 0) {
2557 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2558 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2561 for_each_pwq(pwq
, wq
) {
2562 struct worker_pool
*pool
= pwq
->pool
;
2564 spin_lock_irq(&pool
->lock
);
2566 if (flush_color
>= 0) {
2567 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2569 if (pwq
->nr_in_flight
[flush_color
]) {
2570 pwq
->flush_color
= flush_color
;
2571 atomic_inc(&wq
->nr_pwqs_to_flush
);
2576 if (work_color
>= 0) {
2577 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2578 pwq
->work_color
= work_color
;
2581 spin_unlock_irq(&pool
->lock
);
2584 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2585 complete(&wq
->first_flusher
->done
);
2591 * flush_workqueue - ensure that any scheduled work has run to completion.
2592 * @wq: workqueue to flush
2594 * This function sleeps until all work items which were queued on entry
2595 * have finished execution, but it is not livelocked by new incoming ones.
2597 void flush_workqueue(struct workqueue_struct
*wq
)
2599 struct wq_flusher this_flusher
= {
2600 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2602 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2606 lock_map_acquire(&wq
->lockdep_map
);
2607 lock_map_release(&wq
->lockdep_map
);
2609 mutex_lock(&wq
->mutex
);
2612 * Start-to-wait phase
2614 next_color
= work_next_color(wq
->work_color
);
2616 if (next_color
!= wq
->flush_color
) {
2618 * Color space is not full. The current work_color
2619 * becomes our flush_color and work_color is advanced
2622 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2623 this_flusher
.flush_color
= wq
->work_color
;
2624 wq
->work_color
= next_color
;
2626 if (!wq
->first_flusher
) {
2627 /* no flush in progress, become the first flusher */
2628 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2630 wq
->first_flusher
= &this_flusher
;
2632 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2634 /* nothing to flush, done */
2635 wq
->flush_color
= next_color
;
2636 wq
->first_flusher
= NULL
;
2641 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2642 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2643 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2647 * Oops, color space is full, wait on overflow queue.
2648 * The next flush completion will assign us
2649 * flush_color and transfer to flusher_queue.
2651 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2654 mutex_unlock(&wq
->mutex
);
2656 wait_for_completion(&this_flusher
.done
);
2659 * Wake-up-and-cascade phase
2661 * First flushers are responsible for cascading flushes and
2662 * handling overflow. Non-first flushers can simply return.
2664 if (wq
->first_flusher
!= &this_flusher
)
2667 mutex_lock(&wq
->mutex
);
2669 /* we might have raced, check again with mutex held */
2670 if (wq
->first_flusher
!= &this_flusher
)
2673 wq
->first_flusher
= NULL
;
2675 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2676 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2679 struct wq_flusher
*next
, *tmp
;
2681 /* complete all the flushers sharing the current flush color */
2682 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2683 if (next
->flush_color
!= wq
->flush_color
)
2685 list_del_init(&next
->list
);
2686 complete(&next
->done
);
2689 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2690 wq
->flush_color
!= work_next_color(wq
->work_color
));
2692 /* this flush_color is finished, advance by one */
2693 wq
->flush_color
= work_next_color(wq
->flush_color
);
2695 /* one color has been freed, handle overflow queue */
2696 if (!list_empty(&wq
->flusher_overflow
)) {
2698 * Assign the same color to all overflowed
2699 * flushers, advance work_color and append to
2700 * flusher_queue. This is the start-to-wait
2701 * phase for these overflowed flushers.
2703 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2704 tmp
->flush_color
= wq
->work_color
;
2706 wq
->work_color
= work_next_color(wq
->work_color
);
2708 list_splice_tail_init(&wq
->flusher_overflow
,
2709 &wq
->flusher_queue
);
2710 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2713 if (list_empty(&wq
->flusher_queue
)) {
2714 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2719 * Need to flush more colors. Make the next flusher
2720 * the new first flusher and arm pwqs.
2722 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2723 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2725 list_del_init(&next
->list
);
2726 wq
->first_flusher
= next
;
2728 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2732 * Meh... this color is already done, clear first
2733 * flusher and repeat cascading.
2735 wq
->first_flusher
= NULL
;
2739 mutex_unlock(&wq
->mutex
);
2741 EXPORT_SYMBOL_GPL(flush_workqueue
);
2744 * drain_workqueue - drain a workqueue
2745 * @wq: workqueue to drain
2747 * Wait until the workqueue becomes empty. While draining is in progress,
2748 * only chain queueing is allowed. IOW, only currently pending or running
2749 * work items on @wq can queue further work items on it. @wq is flushed
2750 * repeatedly until it becomes empty. The number of flushing is detemined
2751 * by the depth of chaining and should be relatively short. Whine if it
2754 void drain_workqueue(struct workqueue_struct
*wq
)
2756 unsigned int flush_cnt
= 0;
2757 struct pool_workqueue
*pwq
;
2760 * __queue_work() needs to test whether there are drainers, is much
2761 * hotter than drain_workqueue() and already looks at @wq->flags.
2762 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2764 mutex_lock(&wq
->mutex
);
2765 if (!wq
->nr_drainers
++)
2766 wq
->flags
|= __WQ_DRAINING
;
2767 mutex_unlock(&wq
->mutex
);
2769 flush_workqueue(wq
);
2771 mutex_lock(&wq
->mutex
);
2773 for_each_pwq(pwq
, wq
) {
2776 spin_lock_irq(&pwq
->pool
->lock
);
2777 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2778 spin_unlock_irq(&pwq
->pool
->lock
);
2783 if (++flush_cnt
== 10 ||
2784 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2785 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2786 wq
->name
, flush_cnt
);
2788 mutex_unlock(&wq
->mutex
);
2792 if (!--wq
->nr_drainers
)
2793 wq
->flags
&= ~__WQ_DRAINING
;
2794 mutex_unlock(&wq
->mutex
);
2796 EXPORT_SYMBOL_GPL(drain_workqueue
);
2798 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2800 struct worker
*worker
= NULL
;
2801 struct worker_pool
*pool
;
2802 struct pool_workqueue
*pwq
;
2806 local_irq_disable();
2807 pool
= get_work_pool(work
);
2813 spin_lock(&pool
->lock
);
2814 /* see the comment in try_to_grab_pending() with the same code */
2815 pwq
= get_work_pwq(work
);
2817 if (unlikely(pwq
->pool
!= pool
))
2820 worker
= find_worker_executing_work(pool
, work
);
2823 pwq
= worker
->current_pwq
;
2826 insert_wq_barrier(pwq
, barr
, work
, worker
);
2827 spin_unlock_irq(&pool
->lock
);
2830 * If @max_active is 1 or rescuer is in use, flushing another work
2831 * item on the same workqueue may lead to deadlock. Make sure the
2832 * flusher is not running on the same workqueue by verifying write
2835 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2836 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2838 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2839 lock_map_release(&pwq
->wq
->lockdep_map
);
2843 spin_unlock_irq(&pool
->lock
);
2847 static bool __flush_work(struct work_struct
*work
)
2849 struct wq_barrier barr
;
2851 if (start_flush_work(work
, &barr
)) {
2852 wait_for_completion(&barr
.done
);
2853 destroy_work_on_stack(&barr
.work
);
2861 * flush_work - wait for a work to finish executing the last queueing instance
2862 * @work: the work to flush
2864 * Wait until @work has finished execution. @work is guaranteed to be idle
2865 * on return if it hasn't been requeued since flush started.
2868 * %true if flush_work() waited for the work to finish execution,
2869 * %false if it was already idle.
2871 bool flush_work(struct work_struct
*work
)
2873 lock_map_acquire(&work
->lockdep_map
);
2874 lock_map_release(&work
->lockdep_map
);
2876 return __flush_work(work
);
2878 EXPORT_SYMBOL_GPL(flush_work
);
2880 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2882 unsigned long flags
;
2886 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2888 * If someone else is canceling, wait for the same event it
2889 * would be waiting for before retrying.
2891 if (unlikely(ret
== -ENOENT
))
2893 } while (unlikely(ret
< 0));
2895 /* tell other tasks trying to grab @work to back off */
2896 mark_work_canceling(work
);
2897 local_irq_restore(flags
);
2900 clear_work_data(work
);
2905 * cancel_work_sync - cancel a work and wait for it to finish
2906 * @work: the work to cancel
2908 * Cancel @work and wait for its execution to finish. This function
2909 * can be used even if the work re-queues itself or migrates to
2910 * another workqueue. On return from this function, @work is
2911 * guaranteed to be not pending or executing on any CPU.
2913 * cancel_work_sync(&delayed_work->work) must not be used for
2914 * delayed_work's. Use cancel_delayed_work_sync() instead.
2916 * The caller must ensure that the workqueue on which @work was last
2917 * queued can't be destroyed before this function returns.
2920 * %true if @work was pending, %false otherwise.
2922 bool cancel_work_sync(struct work_struct
*work
)
2924 return __cancel_work_timer(work
, false);
2926 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2929 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2930 * @dwork: the delayed work to flush
2932 * Delayed timer is cancelled and the pending work is queued for
2933 * immediate execution. Like flush_work(), this function only
2934 * considers the last queueing instance of @dwork.
2937 * %true if flush_work() waited for the work to finish execution,
2938 * %false if it was already idle.
2940 bool flush_delayed_work(struct delayed_work
*dwork
)
2942 local_irq_disable();
2943 if (del_timer_sync(&dwork
->timer
))
2944 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2946 return flush_work(&dwork
->work
);
2948 EXPORT_SYMBOL(flush_delayed_work
);
2951 * cancel_delayed_work - cancel a delayed work
2952 * @dwork: delayed_work to cancel
2954 * Kill off a pending delayed_work.
2956 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2960 * The work callback function may still be running on return, unless
2961 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2962 * use cancel_delayed_work_sync() to wait on it.
2964 * This function is safe to call from any context including IRQ handler.
2966 bool cancel_delayed_work(struct delayed_work
*dwork
)
2968 unsigned long flags
;
2972 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2973 } while (unlikely(ret
== -EAGAIN
));
2975 if (unlikely(ret
< 0))
2978 set_work_pool_and_clear_pending(&dwork
->work
,
2979 get_work_pool_id(&dwork
->work
));
2980 local_irq_restore(flags
);
2983 EXPORT_SYMBOL(cancel_delayed_work
);
2986 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2987 * @dwork: the delayed work cancel
2989 * This is cancel_work_sync() for delayed works.
2992 * %true if @dwork was pending, %false otherwise.
2994 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2996 return __cancel_work_timer(&dwork
->work
, true);
2998 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3001 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3002 * @func: the function to call
3004 * schedule_on_each_cpu() executes @func on each online CPU using the
3005 * system workqueue and blocks until all CPUs have completed.
3006 * schedule_on_each_cpu() is very slow.
3009 * 0 on success, -errno on failure.
3011 int schedule_on_each_cpu(work_func_t func
)
3014 struct work_struct __percpu
*works
;
3016 works
= alloc_percpu(struct work_struct
);
3022 for_each_online_cpu(cpu
) {
3023 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3025 INIT_WORK(work
, func
);
3026 schedule_work_on(cpu
, work
);
3029 for_each_online_cpu(cpu
)
3030 flush_work(per_cpu_ptr(works
, cpu
));
3038 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3040 * Forces execution of the kernel-global workqueue and blocks until its
3043 * Think twice before calling this function! It's very easy to get into
3044 * trouble if you don't take great care. Either of the following situations
3045 * will lead to deadlock:
3047 * One of the work items currently on the workqueue needs to acquire
3048 * a lock held by your code or its caller.
3050 * Your code is running in the context of a work routine.
3052 * They will be detected by lockdep when they occur, but the first might not
3053 * occur very often. It depends on what work items are on the workqueue and
3054 * what locks they need, which you have no control over.
3056 * In most situations flushing the entire workqueue is overkill; you merely
3057 * need to know that a particular work item isn't queued and isn't running.
3058 * In such cases you should use cancel_delayed_work_sync() or
3059 * cancel_work_sync() instead.
3061 void flush_scheduled_work(void)
3063 flush_workqueue(system_wq
);
3065 EXPORT_SYMBOL(flush_scheduled_work
);
3068 * execute_in_process_context - reliably execute the routine with user context
3069 * @fn: the function to execute
3070 * @ew: guaranteed storage for the execute work structure (must
3071 * be available when the work executes)
3073 * Executes the function immediately if process context is available,
3074 * otherwise schedules the function for delayed execution.
3076 * Return: 0 - function was executed
3077 * 1 - function was scheduled for execution
3079 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3081 if (!in_interrupt()) {
3086 INIT_WORK(&ew
->work
, fn
);
3087 schedule_work(&ew
->work
);
3091 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3095 * Workqueues with WQ_SYSFS flag set is visible to userland via
3096 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3097 * following attributes.
3099 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3100 * max_active RW int : maximum number of in-flight work items
3102 * Unbound workqueues have the following extra attributes.
3104 * id RO int : the associated pool ID
3105 * nice RW int : nice value of the workers
3106 * cpumask RW mask : bitmask of allowed CPUs for the workers
3109 struct workqueue_struct
*wq
;
3113 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3115 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3120 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
3123 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3125 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3127 static DEVICE_ATTR_RO(per_cpu
);
3129 static ssize_t
max_active_show(struct device
*dev
,
3130 struct device_attribute
*attr
, char *buf
)
3132 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3134 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3137 static ssize_t
max_active_store(struct device
*dev
,
3138 struct device_attribute
*attr
, const char *buf
,
3141 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3144 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3147 workqueue_set_max_active(wq
, val
);
3150 static DEVICE_ATTR_RW(max_active
);
3152 static struct attribute
*wq_sysfs_attrs
[] = {
3153 &dev_attr_per_cpu
.attr
,
3154 &dev_attr_max_active
.attr
,
3157 ATTRIBUTE_GROUPS(wq_sysfs
);
3159 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3160 struct device_attribute
*attr
, char *buf
)
3162 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3163 const char *delim
= "";
3164 int node
, written
= 0;
3166 rcu_read_lock_sched();
3167 for_each_node(node
) {
3168 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3169 "%s%d:%d", delim
, node
,
3170 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3173 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3174 rcu_read_unlock_sched();
3179 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3182 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3185 mutex_lock(&wq
->mutex
);
3186 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3187 mutex_unlock(&wq
->mutex
);
3192 /* prepare workqueue_attrs for sysfs store operations */
3193 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3195 struct workqueue_attrs
*attrs
;
3197 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3201 mutex_lock(&wq
->mutex
);
3202 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3203 mutex_unlock(&wq
->mutex
);
3207 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3208 const char *buf
, size_t count
)
3210 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3211 struct workqueue_attrs
*attrs
;
3214 attrs
= wq_sysfs_prep_attrs(wq
);
3218 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3219 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3220 ret
= apply_workqueue_attrs(wq
, attrs
);
3224 free_workqueue_attrs(attrs
);
3225 return ret
?: count
;
3228 static ssize_t
wq_cpumask_show(struct device
*dev
,
3229 struct device_attribute
*attr
, char *buf
)
3231 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3234 mutex_lock(&wq
->mutex
);
3235 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3236 mutex_unlock(&wq
->mutex
);
3238 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3242 static ssize_t
wq_cpumask_store(struct device
*dev
,
3243 struct device_attribute
*attr
,
3244 const char *buf
, size_t count
)
3246 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3247 struct workqueue_attrs
*attrs
;
3250 attrs
= wq_sysfs_prep_attrs(wq
);
3254 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3256 ret
= apply_workqueue_attrs(wq
, attrs
);
3258 free_workqueue_attrs(attrs
);
3259 return ret
?: count
;
3262 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3265 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3268 mutex_lock(&wq
->mutex
);
3269 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3270 !wq
->unbound_attrs
->no_numa
);
3271 mutex_unlock(&wq
->mutex
);
3276 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3277 const char *buf
, size_t count
)
3279 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3280 struct workqueue_attrs
*attrs
;
3283 attrs
= wq_sysfs_prep_attrs(wq
);
3288 if (sscanf(buf
, "%d", &v
) == 1) {
3289 attrs
->no_numa
= !v
;
3290 ret
= apply_workqueue_attrs(wq
, attrs
);
3293 free_workqueue_attrs(attrs
);
3294 return ret
?: count
;
3297 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3298 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3299 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3300 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3301 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3305 static struct bus_type wq_subsys
= {
3306 .name
= "workqueue",
3307 .dev_groups
= wq_sysfs_groups
,
3310 static int __init
wq_sysfs_init(void)
3312 return subsys_virtual_register(&wq_subsys
, NULL
);
3314 core_initcall(wq_sysfs_init
);
3316 static void wq_device_release(struct device
*dev
)
3318 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3324 * workqueue_sysfs_register - make a workqueue visible in sysfs
3325 * @wq: the workqueue to register
3327 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3328 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3329 * which is the preferred method.
3331 * Workqueue user should use this function directly iff it wants to apply
3332 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3333 * apply_workqueue_attrs() may race against userland updating the
3336 * Return: 0 on success, -errno on failure.
3338 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3340 struct wq_device
*wq_dev
;
3344 * Adjusting max_active or creating new pwqs by applyting
3345 * attributes breaks ordering guarantee. Disallow exposing ordered
3348 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3351 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3356 wq_dev
->dev
.bus
= &wq_subsys
;
3357 wq_dev
->dev
.init_name
= wq
->name
;
3358 wq_dev
->dev
.release
= wq_device_release
;
3361 * unbound_attrs are created separately. Suppress uevent until
3362 * everything is ready.
3364 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3366 ret
= device_register(&wq_dev
->dev
);
3373 if (wq
->flags
& WQ_UNBOUND
) {
3374 struct device_attribute
*attr
;
3376 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3377 ret
= device_create_file(&wq_dev
->dev
, attr
);
3379 device_unregister(&wq_dev
->dev
);
3386 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3391 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3392 * @wq: the workqueue to unregister
3394 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3396 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3398 struct wq_device
*wq_dev
= wq
->wq_dev
;
3404 device_unregister(&wq_dev
->dev
);
3406 #else /* CONFIG_SYSFS */
3407 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3408 #endif /* CONFIG_SYSFS */
3411 * free_workqueue_attrs - free a workqueue_attrs
3412 * @attrs: workqueue_attrs to free
3414 * Undo alloc_workqueue_attrs().
3416 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3419 free_cpumask_var(attrs
->cpumask
);
3425 * alloc_workqueue_attrs - allocate a workqueue_attrs
3426 * @gfp_mask: allocation mask to use
3428 * Allocate a new workqueue_attrs, initialize with default settings and
3431 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3433 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3435 struct workqueue_attrs
*attrs
;
3437 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3440 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3443 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3446 free_workqueue_attrs(attrs
);
3450 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3451 const struct workqueue_attrs
*from
)
3453 to
->nice
= from
->nice
;
3454 cpumask_copy(to
->cpumask
, from
->cpumask
);
3456 * Unlike hash and equality test, this function doesn't ignore
3457 * ->no_numa as it is used for both pool and wq attrs. Instead,
3458 * get_unbound_pool() explicitly clears ->no_numa after copying.
3460 to
->no_numa
= from
->no_numa
;
3463 /* hash value of the content of @attr */
3464 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3468 hash
= jhash_1word(attrs
->nice
, hash
);
3469 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3470 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3474 /* content equality test */
3475 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3476 const struct workqueue_attrs
*b
)
3478 if (a
->nice
!= b
->nice
)
3480 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3486 * init_worker_pool - initialize a newly zalloc'd worker_pool
3487 * @pool: worker_pool to initialize
3489 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3491 * Return: 0 on success, -errno on failure. Even on failure, all fields
3492 * inside @pool proper are initialized and put_unbound_pool() can be called
3493 * on @pool safely to release it.
3495 static int init_worker_pool(struct worker_pool
*pool
)
3497 spin_lock_init(&pool
->lock
);
3500 pool
->node
= NUMA_NO_NODE
;
3501 pool
->flags
|= POOL_DISASSOCIATED
;
3502 INIT_LIST_HEAD(&pool
->worklist
);
3503 INIT_LIST_HEAD(&pool
->idle_list
);
3504 hash_init(pool
->busy_hash
);
3506 init_timer_deferrable(&pool
->idle_timer
);
3507 pool
->idle_timer
.function
= idle_worker_timeout
;
3508 pool
->idle_timer
.data
= (unsigned long)pool
;
3510 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3511 (unsigned long)pool
);
3513 mutex_init(&pool
->manager_arb
);
3514 mutex_init(&pool
->manager_mutex
);
3515 idr_init(&pool
->worker_idr
);
3517 INIT_HLIST_NODE(&pool
->hash_node
);
3520 /* shouldn't fail above this point */
3521 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3527 static void rcu_free_pool(struct rcu_head
*rcu
)
3529 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3531 idr_destroy(&pool
->worker_idr
);
3532 free_workqueue_attrs(pool
->attrs
);
3537 * put_unbound_pool - put a worker_pool
3538 * @pool: worker_pool to put
3540 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3541 * safe manner. get_unbound_pool() calls this function on its failure path
3542 * and this function should be able to release pools which went through,
3543 * successfully or not, init_worker_pool().
3545 * Should be called with wq_pool_mutex held.
3547 static void put_unbound_pool(struct worker_pool
*pool
)
3549 struct worker
*worker
;
3551 lockdep_assert_held(&wq_pool_mutex
);
3557 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3558 WARN_ON(!list_empty(&pool
->worklist
)))
3561 /* release id and unhash */
3563 idr_remove(&worker_pool_idr
, pool
->id
);
3564 hash_del(&pool
->hash_node
);
3567 * Become the manager and destroy all workers. Grabbing
3568 * manager_arb prevents @pool's workers from blocking on
3571 mutex_lock(&pool
->manager_arb
);
3572 mutex_lock(&pool
->manager_mutex
);
3573 spin_lock_irq(&pool
->lock
);
3575 while ((worker
= first_worker(pool
)))
3576 destroy_worker(worker
);
3577 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3579 spin_unlock_irq(&pool
->lock
);
3580 mutex_unlock(&pool
->manager_mutex
);
3581 mutex_unlock(&pool
->manager_arb
);
3583 /* shut down the timers */
3584 del_timer_sync(&pool
->idle_timer
);
3585 del_timer_sync(&pool
->mayday_timer
);
3587 /* sched-RCU protected to allow dereferences from get_work_pool() */
3588 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3592 * get_unbound_pool - get a worker_pool with the specified attributes
3593 * @attrs: the attributes of the worker_pool to get
3595 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3596 * reference count and return it. If there already is a matching
3597 * worker_pool, it will be used; otherwise, this function attempts to
3600 * Should be called with wq_pool_mutex held.
3602 * Return: On success, a worker_pool with the same attributes as @attrs.
3603 * On failure, %NULL.
3605 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3607 u32 hash
= wqattrs_hash(attrs
);
3608 struct worker_pool
*pool
;
3611 lockdep_assert_held(&wq_pool_mutex
);
3613 /* do we already have a matching pool? */
3614 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3615 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3621 /* nope, create a new one */
3622 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3623 if (!pool
|| init_worker_pool(pool
) < 0)
3626 if (workqueue_freezing
)
3627 pool
->flags
|= POOL_FREEZING
;
3629 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3630 copy_workqueue_attrs(pool
->attrs
, attrs
);
3633 * no_numa isn't a worker_pool attribute, always clear it. See
3634 * 'struct workqueue_attrs' comments for detail.
3636 pool
->attrs
->no_numa
= false;
3638 /* if cpumask is contained inside a NUMA node, we belong to that node */
3639 if (wq_numa_enabled
) {
3640 for_each_node(node
) {
3641 if (cpumask_subset(pool
->attrs
->cpumask
,
3642 wq_numa_possible_cpumask
[node
])) {
3649 if (worker_pool_assign_id(pool
) < 0)
3652 /* create and start the initial worker */
3653 if (create_and_start_worker(pool
) < 0)
3657 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3662 put_unbound_pool(pool
);
3666 static void rcu_free_pwq(struct rcu_head
*rcu
)
3668 kmem_cache_free(pwq_cache
,
3669 container_of(rcu
, struct pool_workqueue
, rcu
));
3673 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3674 * and needs to be destroyed.
3676 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3678 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3679 unbound_release_work
);
3680 struct workqueue_struct
*wq
= pwq
->wq
;
3681 struct worker_pool
*pool
= pwq
->pool
;
3684 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3688 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3689 * necessary on release but do it anyway. It's easier to verify
3690 * and consistent with the linking path.
3692 mutex_lock(&wq
->mutex
);
3693 list_del_rcu(&pwq
->pwqs_node
);
3694 is_last
= list_empty(&wq
->pwqs
);
3695 mutex_unlock(&wq
->mutex
);
3697 mutex_lock(&wq_pool_mutex
);
3698 put_unbound_pool(pool
);
3699 mutex_unlock(&wq_pool_mutex
);
3701 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3704 * If we're the last pwq going away, @wq is already dead and no one
3705 * is gonna access it anymore. Free it.
3708 free_workqueue_attrs(wq
->unbound_attrs
);
3714 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3715 * @pwq: target pool_workqueue
3717 * If @pwq isn't freezing, set @pwq->max_active to the associated
3718 * workqueue's saved_max_active and activate delayed work items
3719 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3721 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3723 struct workqueue_struct
*wq
= pwq
->wq
;
3724 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3726 /* for @wq->saved_max_active */
3727 lockdep_assert_held(&wq
->mutex
);
3729 /* fast exit for non-freezable wqs */
3730 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3733 spin_lock_irq(&pwq
->pool
->lock
);
3735 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3736 pwq
->max_active
= wq
->saved_max_active
;
3738 while (!list_empty(&pwq
->delayed_works
) &&
3739 pwq
->nr_active
< pwq
->max_active
)
3740 pwq_activate_first_delayed(pwq
);
3743 * Need to kick a worker after thawed or an unbound wq's
3744 * max_active is bumped. It's a slow path. Do it always.
3746 wake_up_worker(pwq
->pool
);
3748 pwq
->max_active
= 0;
3751 spin_unlock_irq(&pwq
->pool
->lock
);
3754 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3755 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3756 struct worker_pool
*pool
)
3758 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3760 memset(pwq
, 0, sizeof(*pwq
));
3764 pwq
->flush_color
= -1;
3766 INIT_LIST_HEAD(&pwq
->delayed_works
);
3767 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3768 INIT_LIST_HEAD(&pwq
->mayday_node
);
3769 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3772 /* sync @pwq with the current state of its associated wq and link it */
3773 static void link_pwq(struct pool_workqueue
*pwq
)
3775 struct workqueue_struct
*wq
= pwq
->wq
;
3777 lockdep_assert_held(&wq
->mutex
);
3779 /* may be called multiple times, ignore if already linked */
3780 if (!list_empty(&pwq
->pwqs_node
))
3784 * Set the matching work_color. This is synchronized with
3785 * wq->mutex to avoid confusing flush_workqueue().
3787 pwq
->work_color
= wq
->work_color
;
3789 /* sync max_active to the current setting */
3790 pwq_adjust_max_active(pwq
);
3793 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3796 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3797 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3798 const struct workqueue_attrs
*attrs
)
3800 struct worker_pool
*pool
;
3801 struct pool_workqueue
*pwq
;
3803 lockdep_assert_held(&wq_pool_mutex
);
3805 pool
= get_unbound_pool(attrs
);
3809 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3811 put_unbound_pool(pool
);
3815 init_pwq(pwq
, wq
, pool
);
3819 /* undo alloc_unbound_pwq(), used only in the error path */
3820 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3822 lockdep_assert_held(&wq_pool_mutex
);
3825 put_unbound_pool(pwq
->pool
);
3826 kmem_cache_free(pwq_cache
, pwq
);
3831 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3832 * @attrs: the wq_attrs of interest
3833 * @node: the target NUMA node
3834 * @cpu_going_down: if >= 0, the CPU to consider as offline
3835 * @cpumask: outarg, the resulting cpumask
3837 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3838 * @cpu_going_down is >= 0, that cpu is considered offline during
3839 * calculation. The result is stored in @cpumask.
3841 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3842 * enabled and @node has online CPUs requested by @attrs, the returned
3843 * cpumask is the intersection of the possible CPUs of @node and
3846 * The caller is responsible for ensuring that the cpumask of @node stays
3849 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3852 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3853 int cpu_going_down
, cpumask_t
*cpumask
)
3855 if (!wq_numa_enabled
|| attrs
->no_numa
)
3858 /* does @node have any online CPUs @attrs wants? */
3859 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3860 if (cpu_going_down
>= 0)
3861 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3863 if (cpumask_empty(cpumask
))
3866 /* yeap, return possible CPUs in @node that @attrs wants */
3867 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3868 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3871 cpumask_copy(cpumask
, attrs
->cpumask
);
3875 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3876 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3878 struct pool_workqueue
*pwq
)
3880 struct pool_workqueue
*old_pwq
;
3882 lockdep_assert_held(&wq
->mutex
);
3884 /* link_pwq() can handle duplicate calls */
3887 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3888 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3893 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3894 * @wq: the target workqueue
3895 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3897 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3898 * machines, this function maps a separate pwq to each NUMA node with
3899 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3900 * NUMA node it was issued on. Older pwqs are released as in-flight work
3901 * items finish. Note that a work item which repeatedly requeues itself
3902 * back-to-back will stay on its current pwq.
3904 * Performs GFP_KERNEL allocations.
3906 * Return: 0 on success and -errno on failure.
3908 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3909 const struct workqueue_attrs
*attrs
)
3911 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3912 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3915 /* only unbound workqueues can change attributes */
3916 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3919 /* creating multiple pwqs breaks ordering guarantee */
3920 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3923 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3924 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3925 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3926 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3929 /* make a copy of @attrs and sanitize it */
3930 copy_workqueue_attrs(new_attrs
, attrs
);
3931 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3934 * We may create multiple pwqs with differing cpumasks. Make a
3935 * copy of @new_attrs which will be modified and used to obtain
3938 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3941 * CPUs should stay stable across pwq creations and installations.
3942 * Pin CPUs, determine the target cpumask for each node and create
3947 mutex_lock(&wq_pool_mutex
);
3950 * If something goes wrong during CPU up/down, we'll fall back to
3951 * the default pwq covering whole @attrs->cpumask. Always create
3952 * it even if we don't use it immediately.
3954 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3958 for_each_node(node
) {
3959 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3960 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3965 pwq_tbl
[node
] = dfl_pwq
;
3969 mutex_unlock(&wq_pool_mutex
);
3971 /* all pwqs have been created successfully, let's install'em */
3972 mutex_lock(&wq
->mutex
);
3974 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3976 /* save the previous pwq and install the new one */
3978 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3980 /* @dfl_pwq might not have been used, ensure it's linked */
3982 swap(wq
->dfl_pwq
, dfl_pwq
);
3984 mutex_unlock(&wq
->mutex
);
3986 /* put the old pwqs */
3988 put_pwq_unlocked(pwq_tbl
[node
]);
3989 put_pwq_unlocked(dfl_pwq
);
3995 free_workqueue_attrs(tmp_attrs
);
3996 free_workqueue_attrs(new_attrs
);
4001 free_unbound_pwq(dfl_pwq
);
4003 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
4004 free_unbound_pwq(pwq_tbl
[node
]);
4005 mutex_unlock(&wq_pool_mutex
);
4013 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4014 * @wq: the target workqueue
4015 * @cpu: the CPU coming up or going down
4016 * @online: whether @cpu is coming up or going down
4018 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4019 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4022 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4023 * falls back to @wq->dfl_pwq which may not be optimal but is always
4026 * Note that when the last allowed CPU of a NUMA node goes offline for a
4027 * workqueue with a cpumask spanning multiple nodes, the workers which were
4028 * already executing the work items for the workqueue will lose their CPU
4029 * affinity and may execute on any CPU. This is similar to how per-cpu
4030 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4031 * affinity, it's the user's responsibility to flush the work item from
4034 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
4037 int node
= cpu_to_node(cpu
);
4038 int cpu_off
= online
? -1 : cpu
;
4039 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
4040 struct workqueue_attrs
*target_attrs
;
4043 lockdep_assert_held(&wq_pool_mutex
);
4045 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
4049 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4050 * Let's use a preallocated one. The following buf is protected by
4051 * CPU hotplug exclusion.
4053 target_attrs
= wq_update_unbound_numa_attrs_buf
;
4054 cpumask
= target_attrs
->cpumask
;
4056 mutex_lock(&wq
->mutex
);
4057 if (wq
->unbound_attrs
->no_numa
)
4060 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4061 pwq
= unbound_pwq_by_node(wq
, node
);
4064 * Let's determine what needs to be done. If the target cpumask is
4065 * different from wq's, we need to compare it to @pwq's and create
4066 * a new one if they don't match. If the target cpumask equals
4067 * wq's, the default pwq should be used. If @pwq is already the
4068 * default one, nothing to do; otherwise, install the default one.
4070 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
4071 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4074 if (pwq
== wq
->dfl_pwq
)
4080 mutex_unlock(&wq
->mutex
);
4082 /* create a new pwq */
4083 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4085 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4091 * Install the new pwq. As this function is called only from CPU
4092 * hotplug callbacks and applying a new attrs is wrapped with
4093 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4096 mutex_lock(&wq
->mutex
);
4097 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4101 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4102 get_pwq(wq
->dfl_pwq
);
4103 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4104 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4106 mutex_unlock(&wq
->mutex
);
4107 put_pwq_unlocked(old_pwq
);
4110 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4112 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4115 if (!(wq
->flags
& WQ_UNBOUND
)) {
4116 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4120 for_each_possible_cpu(cpu
) {
4121 struct pool_workqueue
*pwq
=
4122 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4123 struct worker_pool
*cpu_pools
=
4124 per_cpu(cpu_worker_pools
, cpu
);
4126 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4128 mutex_lock(&wq
->mutex
);
4130 mutex_unlock(&wq
->mutex
);
4133 } else if (wq
->flags
& __WQ_ORDERED
) {
4134 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4135 /* there should only be single pwq for ordering guarantee */
4136 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4137 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4138 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4141 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4145 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4148 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4150 if (max_active
< 1 || max_active
> lim
)
4151 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4152 max_active
, name
, 1, lim
);
4154 return clamp_val(max_active
, 1, lim
);
4157 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4160 struct lock_class_key
*key
,
4161 const char *lock_name
, ...)
4163 size_t tbl_size
= 0;
4165 struct workqueue_struct
*wq
;
4166 struct pool_workqueue
*pwq
;
4168 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4169 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4170 flags
|= WQ_UNBOUND
;
4172 /* allocate wq and format name */
4173 if (flags
& WQ_UNBOUND
)
4174 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4176 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4180 if (flags
& WQ_UNBOUND
) {
4181 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4182 if (!wq
->unbound_attrs
)
4186 va_start(args
, lock_name
);
4187 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4190 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4191 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4195 wq
->saved_max_active
= max_active
;
4196 mutex_init(&wq
->mutex
);
4197 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4198 INIT_LIST_HEAD(&wq
->pwqs
);
4199 INIT_LIST_HEAD(&wq
->flusher_queue
);
4200 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4201 INIT_LIST_HEAD(&wq
->maydays
);
4203 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4204 INIT_LIST_HEAD(&wq
->list
);
4206 if (alloc_and_link_pwqs(wq
) < 0)
4210 * Workqueues which may be used during memory reclaim should
4211 * have a rescuer to guarantee forward progress.
4213 if (flags
& WQ_MEM_RECLAIM
) {
4214 struct worker
*rescuer
;
4216 rescuer
= alloc_worker();
4220 rescuer
->rescue_wq
= wq
;
4221 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4223 if (IS_ERR(rescuer
->task
)) {
4228 wq
->rescuer
= rescuer
;
4229 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4230 wake_up_process(rescuer
->task
);
4233 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4237 * wq_pool_mutex protects global freeze state and workqueues list.
4238 * Grab it, adjust max_active and add the new @wq to workqueues
4241 mutex_lock(&wq_pool_mutex
);
4243 mutex_lock(&wq
->mutex
);
4244 for_each_pwq(pwq
, wq
)
4245 pwq_adjust_max_active(pwq
);
4246 mutex_unlock(&wq
->mutex
);
4248 list_add(&wq
->list
, &workqueues
);
4250 mutex_unlock(&wq_pool_mutex
);
4255 free_workqueue_attrs(wq
->unbound_attrs
);
4259 destroy_workqueue(wq
);
4262 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4265 * destroy_workqueue - safely terminate a workqueue
4266 * @wq: target workqueue
4268 * Safely destroy a workqueue. All work currently pending will be done first.
4270 void destroy_workqueue(struct workqueue_struct
*wq
)
4272 struct pool_workqueue
*pwq
;
4275 /* drain it before proceeding with destruction */
4276 drain_workqueue(wq
);
4279 mutex_lock(&wq
->mutex
);
4280 for_each_pwq(pwq
, wq
) {
4283 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4284 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4285 mutex_unlock(&wq
->mutex
);
4290 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4291 WARN_ON(pwq
->nr_active
) ||
4292 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4293 mutex_unlock(&wq
->mutex
);
4297 mutex_unlock(&wq
->mutex
);
4300 * wq list is used to freeze wq, remove from list after
4301 * flushing is complete in case freeze races us.
4303 mutex_lock(&wq_pool_mutex
);
4304 list_del_init(&wq
->list
);
4305 mutex_unlock(&wq_pool_mutex
);
4307 workqueue_sysfs_unregister(wq
);
4310 kthread_stop(wq
->rescuer
->task
);
4315 if (!(wq
->flags
& WQ_UNBOUND
)) {
4317 * The base ref is never dropped on per-cpu pwqs. Directly
4318 * free the pwqs and wq.
4320 free_percpu(wq
->cpu_pwqs
);
4324 * We're the sole accessor of @wq at this point. Directly
4325 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4326 * @wq will be freed when the last pwq is released.
4328 for_each_node(node
) {
4329 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4330 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4331 put_pwq_unlocked(pwq
);
4335 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4336 * put. Don't access it afterwards.
4340 put_pwq_unlocked(pwq
);
4343 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4346 * workqueue_set_max_active - adjust max_active of a workqueue
4347 * @wq: target workqueue
4348 * @max_active: new max_active value.
4350 * Set max_active of @wq to @max_active.
4353 * Don't call from IRQ context.
4355 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4357 struct pool_workqueue
*pwq
;
4359 /* disallow meddling with max_active for ordered workqueues */
4360 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4363 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4365 mutex_lock(&wq
->mutex
);
4367 wq
->saved_max_active
= max_active
;
4369 for_each_pwq(pwq
, wq
)
4370 pwq_adjust_max_active(pwq
);
4372 mutex_unlock(&wq
->mutex
);
4374 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4377 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4379 * Determine whether %current is a workqueue rescuer. Can be used from
4380 * work functions to determine whether it's being run off the rescuer task.
4382 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4384 bool current_is_workqueue_rescuer(void)
4386 struct worker
*worker
= current_wq_worker();
4388 return worker
&& worker
->rescue_wq
;
4392 * workqueue_congested - test whether a workqueue is congested
4393 * @cpu: CPU in question
4394 * @wq: target workqueue
4396 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4397 * no synchronization around this function and the test result is
4398 * unreliable and only useful as advisory hints or for debugging.
4400 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4401 * Note that both per-cpu and unbound workqueues may be associated with
4402 * multiple pool_workqueues which have separate congested states. A
4403 * workqueue being congested on one CPU doesn't mean the workqueue is also
4404 * contested on other CPUs / NUMA nodes.
4407 * %true if congested, %false otherwise.
4409 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4411 struct pool_workqueue
*pwq
;
4414 rcu_read_lock_sched();
4416 if (cpu
== WORK_CPU_UNBOUND
)
4417 cpu
= smp_processor_id();
4419 if (!(wq
->flags
& WQ_UNBOUND
))
4420 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4422 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4424 ret
= !list_empty(&pwq
->delayed_works
);
4425 rcu_read_unlock_sched();
4429 EXPORT_SYMBOL_GPL(workqueue_congested
);
4432 * work_busy - test whether a work is currently pending or running
4433 * @work: the work to be tested
4435 * Test whether @work is currently pending or running. There is no
4436 * synchronization around this function and the test result is
4437 * unreliable and only useful as advisory hints or for debugging.
4440 * OR'd bitmask of WORK_BUSY_* bits.
4442 unsigned int work_busy(struct work_struct
*work
)
4444 struct worker_pool
*pool
;
4445 unsigned long flags
;
4446 unsigned int ret
= 0;
4448 if (work_pending(work
))
4449 ret
|= WORK_BUSY_PENDING
;
4451 local_irq_save(flags
);
4452 pool
= get_work_pool(work
);
4454 spin_lock(&pool
->lock
);
4455 if (find_worker_executing_work(pool
, work
))
4456 ret
|= WORK_BUSY_RUNNING
;
4457 spin_unlock(&pool
->lock
);
4459 local_irq_restore(flags
);
4463 EXPORT_SYMBOL_GPL(work_busy
);
4466 * set_worker_desc - set description for the current work item
4467 * @fmt: printf-style format string
4468 * @...: arguments for the format string
4470 * This function can be called by a running work function to describe what
4471 * the work item is about. If the worker task gets dumped, this
4472 * information will be printed out together to help debugging. The
4473 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4475 void set_worker_desc(const char *fmt
, ...)
4477 struct worker
*worker
= current_wq_worker();
4481 va_start(args
, fmt
);
4482 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4484 worker
->desc_valid
= true;
4489 * print_worker_info - print out worker information and description
4490 * @log_lvl: the log level to use when printing
4491 * @task: target task
4493 * If @task is a worker and currently executing a work item, print out the
4494 * name of the workqueue being serviced and worker description set with
4495 * set_worker_desc() by the currently executing work item.
4497 * This function can be safely called on any task as long as the
4498 * task_struct itself is accessible. While safe, this function isn't
4499 * synchronized and may print out mixups or garbages of limited length.
4501 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4503 work_func_t
*fn
= NULL
;
4504 char name
[WQ_NAME_LEN
] = { };
4505 char desc
[WORKER_DESC_LEN
] = { };
4506 struct pool_workqueue
*pwq
= NULL
;
4507 struct workqueue_struct
*wq
= NULL
;
4508 bool desc_valid
= false;
4509 struct worker
*worker
;
4511 if (!(task
->flags
& PF_WQ_WORKER
))
4515 * This function is called without any synchronization and @task
4516 * could be in any state. Be careful with dereferences.
4518 worker
= probe_kthread_data(task
);
4521 * Carefully copy the associated workqueue's workfn and name. Keep
4522 * the original last '\0' in case the original contains garbage.
4524 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4525 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4526 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4527 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4529 /* copy worker description */
4530 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4532 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4534 if (fn
|| name
[0] || desc
[0]) {
4535 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4537 pr_cont(" (%s)", desc
);
4545 * There are two challenges in supporting CPU hotplug. Firstly, there
4546 * are a lot of assumptions on strong associations among work, pwq and
4547 * pool which make migrating pending and scheduled works very
4548 * difficult to implement without impacting hot paths. Secondly,
4549 * worker pools serve mix of short, long and very long running works making
4550 * blocked draining impractical.
4552 * This is solved by allowing the pools to be disassociated from the CPU
4553 * running as an unbound one and allowing it to be reattached later if the
4554 * cpu comes back online.
4557 static void wq_unbind_fn(struct work_struct
*work
)
4559 int cpu
= smp_processor_id();
4560 struct worker_pool
*pool
;
4561 struct worker
*worker
;
4564 for_each_cpu_worker_pool(pool
, cpu
) {
4565 WARN_ON_ONCE(cpu
!= smp_processor_id());
4567 mutex_lock(&pool
->manager_mutex
);
4568 spin_lock_irq(&pool
->lock
);
4571 * We've blocked all manager operations. Make all workers
4572 * unbound and set DISASSOCIATED. Before this, all workers
4573 * except for the ones which are still executing works from
4574 * before the last CPU down must be on the cpu. After
4575 * this, they may become diasporas.
4577 for_each_pool_worker(worker
, wi
, pool
)
4578 worker
->flags
|= WORKER_UNBOUND
;
4580 pool
->flags
|= POOL_DISASSOCIATED
;
4582 spin_unlock_irq(&pool
->lock
);
4583 mutex_unlock(&pool
->manager_mutex
);
4586 * Call schedule() so that we cross rq->lock and thus can
4587 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4588 * This is necessary as scheduler callbacks may be invoked
4594 * Sched callbacks are disabled now. Zap nr_running.
4595 * After this, nr_running stays zero and need_more_worker()
4596 * and keep_working() are always true as long as the
4597 * worklist is not empty. This pool now behaves as an
4598 * unbound (in terms of concurrency management) pool which
4599 * are served by workers tied to the pool.
4601 atomic_set(&pool
->nr_running
, 0);
4604 * With concurrency management just turned off, a busy
4605 * worker blocking could lead to lengthy stalls. Kick off
4606 * unbound chain execution of currently pending work items.
4608 spin_lock_irq(&pool
->lock
);
4609 wake_up_worker(pool
);
4610 spin_unlock_irq(&pool
->lock
);
4615 * rebind_workers - rebind all workers of a pool to the associated CPU
4616 * @pool: pool of interest
4618 * @pool->cpu is coming online. Rebind all workers to the CPU.
4620 static void rebind_workers(struct worker_pool
*pool
)
4622 struct worker
*worker
;
4625 lockdep_assert_held(&pool
->manager_mutex
);
4628 * Restore CPU affinity of all workers. As all idle workers should
4629 * be on the run-queue of the associated CPU before any local
4630 * wake-ups for concurrency management happen, restore CPU affinty
4631 * of all workers first and then clear UNBOUND. As we're called
4632 * from CPU_ONLINE, the following shouldn't fail.
4634 for_each_pool_worker(worker
, wi
, pool
)
4635 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4636 pool
->attrs
->cpumask
) < 0);
4638 spin_lock_irq(&pool
->lock
);
4640 for_each_pool_worker(worker
, wi
, pool
) {
4641 unsigned int worker_flags
= worker
->flags
;
4644 * A bound idle worker should actually be on the runqueue
4645 * of the associated CPU for local wake-ups targeting it to
4646 * work. Kick all idle workers so that they migrate to the
4647 * associated CPU. Doing this in the same loop as
4648 * replacing UNBOUND with REBOUND is safe as no worker will
4649 * be bound before @pool->lock is released.
4651 if (worker_flags
& WORKER_IDLE
)
4652 wake_up_process(worker
->task
);
4655 * We want to clear UNBOUND but can't directly call
4656 * worker_clr_flags() or adjust nr_running. Atomically
4657 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4658 * @worker will clear REBOUND using worker_clr_flags() when
4659 * it initiates the next execution cycle thus restoring
4660 * concurrency management. Note that when or whether
4661 * @worker clears REBOUND doesn't affect correctness.
4663 * ACCESS_ONCE() is necessary because @worker->flags may be
4664 * tested without holding any lock in
4665 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4666 * fail incorrectly leading to premature concurrency
4667 * management operations.
4669 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4670 worker_flags
|= WORKER_REBOUND
;
4671 worker_flags
&= ~WORKER_UNBOUND
;
4672 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4675 spin_unlock_irq(&pool
->lock
);
4679 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4680 * @pool: unbound pool of interest
4681 * @cpu: the CPU which is coming up
4683 * An unbound pool may end up with a cpumask which doesn't have any online
4684 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4685 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4686 * online CPU before, cpus_allowed of all its workers should be restored.
4688 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4690 static cpumask_t cpumask
;
4691 struct worker
*worker
;
4694 lockdep_assert_held(&pool
->manager_mutex
);
4696 /* is @cpu allowed for @pool? */
4697 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4700 /* is @cpu the only online CPU? */
4701 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4702 if (cpumask_weight(&cpumask
) != 1)
4705 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4706 for_each_pool_worker(worker
, wi
, pool
)
4707 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4708 pool
->attrs
->cpumask
) < 0);
4712 * Workqueues should be brought up before normal priority CPU notifiers.
4713 * This will be registered high priority CPU notifier.
4715 static int workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4716 unsigned long action
,
4719 int cpu
= (unsigned long)hcpu
;
4720 struct worker_pool
*pool
;
4721 struct workqueue_struct
*wq
;
4724 switch (action
& ~CPU_TASKS_FROZEN
) {
4725 case CPU_UP_PREPARE
:
4726 for_each_cpu_worker_pool(pool
, cpu
) {
4727 if (pool
->nr_workers
)
4729 if (create_and_start_worker(pool
) < 0)
4734 case CPU_DOWN_FAILED
:
4736 mutex_lock(&wq_pool_mutex
);
4738 for_each_pool(pool
, pi
) {
4739 mutex_lock(&pool
->manager_mutex
);
4741 if (pool
->cpu
== cpu
) {
4742 spin_lock_irq(&pool
->lock
);
4743 pool
->flags
&= ~POOL_DISASSOCIATED
;
4744 spin_unlock_irq(&pool
->lock
);
4746 rebind_workers(pool
);
4747 } else if (pool
->cpu
< 0) {
4748 restore_unbound_workers_cpumask(pool
, cpu
);
4751 mutex_unlock(&pool
->manager_mutex
);
4754 /* update NUMA affinity of unbound workqueues */
4755 list_for_each_entry(wq
, &workqueues
, list
)
4756 wq_update_unbound_numa(wq
, cpu
, true);
4758 mutex_unlock(&wq_pool_mutex
);
4765 * Workqueues should be brought down after normal priority CPU notifiers.
4766 * This will be registered as low priority CPU notifier.
4768 static int workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4769 unsigned long action
,
4772 int cpu
= (unsigned long)hcpu
;
4773 struct work_struct unbind_work
;
4774 struct workqueue_struct
*wq
;
4776 switch (action
& ~CPU_TASKS_FROZEN
) {
4777 case CPU_DOWN_PREPARE
:
4778 /* unbinding per-cpu workers should happen on the local CPU */
4779 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4780 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4782 /* update NUMA affinity of unbound workqueues */
4783 mutex_lock(&wq_pool_mutex
);
4784 list_for_each_entry(wq
, &workqueues
, list
)
4785 wq_update_unbound_numa(wq
, cpu
, false);
4786 mutex_unlock(&wq_pool_mutex
);
4788 /* wait for per-cpu unbinding to finish */
4789 flush_work(&unbind_work
);
4797 struct work_for_cpu
{
4798 struct work_struct work
;
4804 static void work_for_cpu_fn(struct work_struct
*work
)
4806 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4808 wfc
->ret
= wfc
->fn(wfc
->arg
);
4812 * work_on_cpu - run a function in user context on a particular cpu
4813 * @cpu: the cpu to run on
4814 * @fn: the function to run
4815 * @arg: the function arg
4817 * It is up to the caller to ensure that the cpu doesn't go offline.
4818 * The caller must not hold any locks which would prevent @fn from completing.
4820 * Return: The value @fn returns.
4822 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4824 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4826 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4827 schedule_work_on(cpu
, &wfc
.work
);
4830 * The work item is on-stack and can't lead to deadlock through
4831 * flushing. Use __flush_work() to avoid spurious lockdep warnings
4832 * when work_on_cpu()s are nested.
4834 __flush_work(&wfc
.work
);
4838 EXPORT_SYMBOL_GPL(work_on_cpu
);
4839 #endif /* CONFIG_SMP */
4841 #ifdef CONFIG_FREEZER
4844 * freeze_workqueues_begin - begin freezing workqueues
4846 * Start freezing workqueues. After this function returns, all freezable
4847 * workqueues will queue new works to their delayed_works list instead of
4851 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4853 void freeze_workqueues_begin(void)
4855 struct worker_pool
*pool
;
4856 struct workqueue_struct
*wq
;
4857 struct pool_workqueue
*pwq
;
4860 mutex_lock(&wq_pool_mutex
);
4862 WARN_ON_ONCE(workqueue_freezing
);
4863 workqueue_freezing
= true;
4866 for_each_pool(pool
, pi
) {
4867 spin_lock_irq(&pool
->lock
);
4868 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4869 pool
->flags
|= POOL_FREEZING
;
4870 spin_unlock_irq(&pool
->lock
);
4873 list_for_each_entry(wq
, &workqueues
, list
) {
4874 mutex_lock(&wq
->mutex
);
4875 for_each_pwq(pwq
, wq
)
4876 pwq_adjust_max_active(pwq
);
4877 mutex_unlock(&wq
->mutex
);
4880 mutex_unlock(&wq_pool_mutex
);
4884 * freeze_workqueues_busy - are freezable workqueues still busy?
4886 * Check whether freezing is complete. This function must be called
4887 * between freeze_workqueues_begin() and thaw_workqueues().
4890 * Grabs and releases wq_pool_mutex.
4893 * %true if some freezable workqueues are still busy. %false if freezing
4896 bool freeze_workqueues_busy(void)
4899 struct workqueue_struct
*wq
;
4900 struct pool_workqueue
*pwq
;
4902 mutex_lock(&wq_pool_mutex
);
4904 WARN_ON_ONCE(!workqueue_freezing
);
4906 list_for_each_entry(wq
, &workqueues
, list
) {
4907 if (!(wq
->flags
& WQ_FREEZABLE
))
4910 * nr_active is monotonically decreasing. It's safe
4911 * to peek without lock.
4913 rcu_read_lock_sched();
4914 for_each_pwq(pwq
, wq
) {
4915 WARN_ON_ONCE(pwq
->nr_active
< 0);
4916 if (pwq
->nr_active
) {
4918 rcu_read_unlock_sched();
4922 rcu_read_unlock_sched();
4925 mutex_unlock(&wq_pool_mutex
);
4930 * thaw_workqueues - thaw workqueues
4932 * Thaw workqueues. Normal queueing is restored and all collected
4933 * frozen works are transferred to their respective pool worklists.
4936 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4938 void thaw_workqueues(void)
4940 struct workqueue_struct
*wq
;
4941 struct pool_workqueue
*pwq
;
4942 struct worker_pool
*pool
;
4945 mutex_lock(&wq_pool_mutex
);
4947 if (!workqueue_freezing
)
4950 /* clear FREEZING */
4951 for_each_pool(pool
, pi
) {
4952 spin_lock_irq(&pool
->lock
);
4953 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4954 pool
->flags
&= ~POOL_FREEZING
;
4955 spin_unlock_irq(&pool
->lock
);
4958 /* restore max_active and repopulate worklist */
4959 list_for_each_entry(wq
, &workqueues
, list
) {
4960 mutex_lock(&wq
->mutex
);
4961 for_each_pwq(pwq
, wq
)
4962 pwq_adjust_max_active(pwq
);
4963 mutex_unlock(&wq
->mutex
);
4966 workqueue_freezing
= false;
4968 mutex_unlock(&wq_pool_mutex
);
4970 #endif /* CONFIG_FREEZER */
4972 static void __init
wq_numa_init(void)
4977 /* determine NUMA pwq table len - highest node id + 1 */
4979 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4981 if (num_possible_nodes() <= 1)
4984 if (wq_disable_numa
) {
4985 pr_info("workqueue: NUMA affinity support disabled\n");
4989 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4990 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4993 * We want masks of possible CPUs of each node which isn't readily
4994 * available. Build one from cpu_to_node() which should have been
4995 * fully initialized by now.
4997 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
5001 BUG_ON(!alloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5002 node_online(node
) ? node
: NUMA_NO_NODE
));
5004 for_each_possible_cpu(cpu
) {
5005 node
= cpu_to_node(cpu
);
5006 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5007 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5008 /* happens iff arch is bonkers, let's just proceed */
5011 cpumask_set_cpu(cpu
, tbl
[node
]);
5014 wq_numa_possible_cpumask
= tbl
;
5015 wq_numa_enabled
= true;
5018 static int __init
init_workqueues(void)
5020 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5023 /* make sure we have enough bits for OFFQ pool ID */
5024 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
5025 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
5027 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5029 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5031 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
5032 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
5036 /* initialize CPU pools */
5037 for_each_possible_cpu(cpu
) {
5038 struct worker_pool
*pool
;
5041 for_each_cpu_worker_pool(pool
, cpu
) {
5042 BUG_ON(init_worker_pool(pool
));
5044 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5045 pool
->attrs
->nice
= std_nice
[i
++];
5046 pool
->node
= cpu_to_node(cpu
);
5049 mutex_lock(&wq_pool_mutex
);
5050 BUG_ON(worker_pool_assign_id(pool
));
5051 mutex_unlock(&wq_pool_mutex
);
5055 /* create the initial worker */
5056 for_each_online_cpu(cpu
) {
5057 struct worker_pool
*pool
;
5059 for_each_cpu_worker_pool(pool
, cpu
) {
5060 pool
->flags
&= ~POOL_DISASSOCIATED
;
5061 BUG_ON(create_and_start_worker(pool
) < 0);
5065 /* create default unbound and ordered wq attrs */
5066 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5067 struct workqueue_attrs
*attrs
;
5069 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5070 attrs
->nice
= std_nice
[i
];
5071 unbound_std_wq_attrs
[i
] = attrs
;
5074 * An ordered wq should have only one pwq as ordering is
5075 * guaranteed by max_active which is enforced by pwqs.
5076 * Turn off NUMA so that dfl_pwq is used for all nodes.
5078 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5079 attrs
->nice
= std_nice
[i
];
5080 attrs
->no_numa
= true;
5081 ordered_wq_attrs
[i
] = attrs
;
5084 system_wq
= alloc_workqueue("events", 0, 0);
5085 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5086 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5087 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5088 WQ_UNBOUND_MAX_ACTIVE
);
5089 system_freezable_wq
= alloc_workqueue("events_freezable",
5091 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5092 WQ_POWER_EFFICIENT
, 0);
5093 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5094 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5096 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5097 !system_unbound_wq
|| !system_freezable_wq
||
5098 !system_power_efficient_wq
||
5099 !system_freezable_power_efficient_wq
);
5102 early_initcall(init_workqueues
);