2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_MANAGER_ACTIVE
= 1 << 0, /* being managed */
72 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
75 WORKER_DIE
= 1 << 1, /* die die die */
76 WORKER_IDLE
= 1 << 2, /* is idle */
77 WORKER_PREP
= 1 << 3, /* preparing to run works */
78 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
79 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
80 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
82 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
83 WORKER_UNBOUND
| WORKER_REBOUND
,
85 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
87 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
88 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
90 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
91 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
93 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
94 /* call for help after 10ms
96 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
97 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
100 * Rescue workers are used only on emergencies and shared by
101 * all cpus. Give MIN_NICE.
103 RESCUER_NICE_LEVEL
= MIN_NICE
,
104 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
110 * Structure fields follow one of the following exclusion rules.
112 * I: Modifiable by initialization/destruction paths and read-only for
115 * P: Preemption protected. Disabling preemption is enough and should
116 * only be modified and accessed from the local cpu.
118 * L: pool->lock protected. Access with pool->lock held.
120 * X: During normal operation, modification requires pool->lock and should
121 * be done only from local cpu. Either disabling preemption on local
122 * cpu or grabbing pool->lock is enough for read access. If
123 * POOL_DISASSOCIATED is set, it's identical to L.
125 * A: pool->attach_mutex protected.
127 * PL: wq_pool_mutex protected.
129 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
133 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
134 * sched-RCU for reads.
136 * WQ: wq->mutex protected.
138 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
140 * MD: wq_mayday_lock protected.
143 /* struct worker is defined in workqueue_internal.h */
146 spinlock_t lock
; /* the pool lock */
147 int cpu
; /* I: the associated cpu */
148 int node
; /* I: the associated node ID */
149 int id
; /* I: pool ID */
150 unsigned int flags
; /* X: flags */
152 struct list_head worklist
; /* L: list of pending works */
153 int nr_workers
; /* L: total number of workers */
155 /* nr_idle includes the ones off idle_list for rebinding */
156 int nr_idle
; /* L: currently idle ones */
158 struct list_head idle_list
; /* X: list of idle workers */
159 struct timer_list idle_timer
; /* L: worker idle timeout */
160 struct timer_list mayday_timer
; /* L: SOS timer for workers */
162 /* a workers is either on busy_hash or idle_list, or the manager */
163 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
164 /* L: hash of busy workers */
166 /* see manage_workers() for details on the two manager mutexes */
167 struct worker
*manager
; /* L: purely informational */
168 struct mutex attach_mutex
; /* attach/detach exclusion */
169 struct list_head workers
; /* A: attached workers */
170 struct completion
*detach_completion
; /* all workers detached */
172 struct ida worker_ida
; /* worker IDs for task name */
174 struct workqueue_attrs
*attrs
; /* I: worker attributes */
175 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
176 int refcnt
; /* PL: refcnt for unbound pools */
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
183 atomic_t nr_running ____cacheline_aligned_in_smp
;
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
190 } ____cacheline_aligned_in_smp
;
193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
198 struct pool_workqueue
{
199 struct worker_pool
*pool
; /* I: the associated pool */
200 struct workqueue_struct
*wq
; /* I: the owning workqueue */
201 int work_color
; /* L: current color */
202 int flush_color
; /* L: flushing color */
203 int refcnt
; /* L: reference count */
204 int nr_in_flight
[WORK_NR_COLORS
];
205 /* L: nr of in_flight works */
206 int nr_active
; /* L: nr of active works */
207 int max_active
; /* L: max active works */
208 struct list_head delayed_works
; /* L: delayed works */
209 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
210 struct list_head mayday_node
; /* MD: node on wq->maydays */
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
216 * determined without grabbing wq->mutex.
218 struct work_struct unbound_release_work
;
220 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
223 * Structure used to wait for workqueue flush.
226 struct list_head list
; /* WQ: list of flushers */
227 int flush_color
; /* WQ: flush color waiting for */
228 struct completion done
; /* flush completion */
234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
237 struct workqueue_struct
{
238 struct list_head pwqs
; /* WR: all pwqs of this wq */
239 struct list_head list
; /* PR: list of all workqueues */
241 struct mutex mutex
; /* protects this wq */
242 int work_color
; /* WQ: current work color */
243 int flush_color
; /* WQ: current flush color */
244 atomic_t nr_pwqs_to_flush
; /* flush in progress */
245 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
246 struct list_head flusher_queue
; /* WQ: flush waiters */
247 struct list_head flusher_overflow
; /* WQ: flush overflow list */
249 struct list_head maydays
; /* MD: pwqs requesting rescue */
250 struct worker
*rescuer
; /* I: rescue worker */
252 int nr_drainers
; /* WQ: drain in progress */
253 int saved_max_active
; /* WQ: saved pwq max_active */
255 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
256 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
259 struct wq_device
*wq_dev
; /* I: for sysfs interface */
261 #ifdef CONFIG_LOCKDEP
262 struct lockdep_map lockdep_map
;
264 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
267 * Destruction of workqueue_struct is sched-RCU protected to allow
268 * walking the workqueues list without grabbing wq_pool_mutex.
269 * This is used to dump all workqueues from sysrq.
273 /* hot fields used during command issue, aligned to cacheline */
274 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
275 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
276 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
279 static struct kmem_cache
*pwq_cache
;
281 static cpumask_var_t
*wq_numa_possible_cpumask
;
282 /* possible CPUs of each node */
284 static bool wq_disable_numa
;
285 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
287 /* see the comment above the definition of WQ_POWER_EFFICIENT */
288 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
289 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
291 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
293 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
294 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
296 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
297 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
298 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait
); /* wait for manager to go away */
300 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
301 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
303 static cpumask_var_t wq_unbound_cpumask
; /* PL: low level cpumask for all unbound wqs */
305 /* the per-cpu worker pools */
306 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
309 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
311 /* PL: hash of all unbound pools keyed by pool->attrs */
312 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
314 /* I: attributes used when instantiating standard unbound pools on demand */
315 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
317 /* I: attributes used when instantiating ordered pools on demand */
318 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
320 struct workqueue_struct
*system_wq __read_mostly
;
321 EXPORT_SYMBOL(system_wq
);
322 struct workqueue_struct
*system_highpri_wq __read_mostly
;
323 EXPORT_SYMBOL_GPL(system_highpri_wq
);
324 struct workqueue_struct
*system_long_wq __read_mostly
;
325 EXPORT_SYMBOL_GPL(system_long_wq
);
326 struct workqueue_struct
*system_unbound_wq __read_mostly
;
327 EXPORT_SYMBOL_GPL(system_unbound_wq
);
328 struct workqueue_struct
*system_freezable_wq __read_mostly
;
329 EXPORT_SYMBOL_GPL(system_freezable_wq
);
330 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
331 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
332 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
333 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
335 static int worker_thread(void *__worker
);
336 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
338 #define CREATE_TRACE_POINTS
339 #include <trace/events/workqueue.h>
341 #define assert_rcu_or_pool_mutex() \
342 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
343 !lockdep_is_held(&wq_pool_mutex), \
344 "sched RCU or wq_pool_mutex should be held")
346 #define assert_rcu_or_wq_mutex(wq) \
347 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
348 !lockdep_is_held(&wq->mutex), \
349 "sched RCU or wq->mutex should be held")
351 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
352 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
353 !lockdep_is_held(&wq->mutex) && \
354 !lockdep_is_held(&wq_pool_mutex), \
355 "sched RCU, wq->mutex or wq_pool_mutex should be held")
357 #define for_each_cpu_worker_pool(pool, cpu) \
358 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
359 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
363 * for_each_pool - iterate through all worker_pools in the system
364 * @pool: iteration cursor
365 * @pi: integer used for iteration
367 * This must be called either with wq_pool_mutex held or sched RCU read
368 * locked. If the pool needs to be used beyond the locking in effect, the
369 * caller is responsible for guaranteeing that the pool stays online.
371 * The if/else clause exists only for the lockdep assertion and can be
374 #define for_each_pool(pool, pi) \
375 idr_for_each_entry(&worker_pool_idr, pool, pi) \
376 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
380 * for_each_pool_worker - iterate through all workers of a worker_pool
381 * @worker: iteration cursor
382 * @pool: worker_pool to iterate workers of
384 * This must be called with @pool->attach_mutex.
386 * The if/else clause exists only for the lockdep assertion and can be
389 #define for_each_pool_worker(worker, pool) \
390 list_for_each_entry((worker), &(pool)->workers, node) \
391 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
395 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
396 * @pwq: iteration cursor
397 * @wq: the target workqueue
399 * This must be called either with wq->mutex held or sched RCU read locked.
400 * If the pwq needs to be used beyond the locking in effect, the caller is
401 * responsible for guaranteeing that the pwq stays online.
403 * The if/else clause exists only for the lockdep assertion and can be
406 #define for_each_pwq(pwq, wq) \
407 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
408 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
411 #ifdef CONFIG_DEBUG_OBJECTS_WORK
413 static struct debug_obj_descr work_debug_descr
;
415 static void *work_debug_hint(void *addr
)
417 return ((struct work_struct
*) addr
)->func
;
421 * fixup_init is called when:
422 * - an active object is initialized
424 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
426 struct work_struct
*work
= addr
;
429 case ODEBUG_STATE_ACTIVE
:
430 cancel_work_sync(work
);
431 debug_object_init(work
, &work_debug_descr
);
439 * fixup_activate is called when:
440 * - an active object is activated
441 * - an unknown object is activated (might be a statically initialized object)
443 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
445 struct work_struct
*work
= addr
;
449 case ODEBUG_STATE_NOTAVAILABLE
:
451 * This is not really a fixup. The work struct was
452 * statically initialized. We just make sure that it
453 * is tracked in the object tracker.
455 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
456 debug_object_init(work
, &work_debug_descr
);
457 debug_object_activate(work
, &work_debug_descr
);
463 case ODEBUG_STATE_ACTIVE
:
472 * fixup_free is called when:
473 * - an active object is freed
475 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
477 struct work_struct
*work
= addr
;
480 case ODEBUG_STATE_ACTIVE
:
481 cancel_work_sync(work
);
482 debug_object_free(work
, &work_debug_descr
);
489 static struct debug_obj_descr work_debug_descr
= {
490 .name
= "work_struct",
491 .debug_hint
= work_debug_hint
,
492 .fixup_init
= work_fixup_init
,
493 .fixup_activate
= work_fixup_activate
,
494 .fixup_free
= work_fixup_free
,
497 static inline void debug_work_activate(struct work_struct
*work
)
499 debug_object_activate(work
, &work_debug_descr
);
502 static inline void debug_work_deactivate(struct work_struct
*work
)
504 debug_object_deactivate(work
, &work_debug_descr
);
507 void __init_work(struct work_struct
*work
, int onstack
)
510 debug_object_init_on_stack(work
, &work_debug_descr
);
512 debug_object_init(work
, &work_debug_descr
);
514 EXPORT_SYMBOL_GPL(__init_work
);
516 void destroy_work_on_stack(struct work_struct
*work
)
518 debug_object_free(work
, &work_debug_descr
);
520 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
522 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
524 destroy_timer_on_stack(&work
->timer
);
525 debug_object_free(&work
->work
, &work_debug_descr
);
527 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
530 static inline void debug_work_activate(struct work_struct
*work
) { }
531 static inline void debug_work_deactivate(struct work_struct
*work
) { }
535 * worker_pool_assign_id - allocate ID and assing it to @pool
536 * @pool: the pool pointer of interest
538 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
539 * successfully, -errno on failure.
541 static int worker_pool_assign_id(struct worker_pool
*pool
)
545 lockdep_assert_held(&wq_pool_mutex
);
547 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
557 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
558 * @wq: the target workqueue
561 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
563 * If the pwq needs to be used beyond the locking in effect, the caller is
564 * responsible for guaranteeing that the pwq stays online.
566 * Return: The unbound pool_workqueue for @node.
568 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
571 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
574 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
575 * delayed item is pending. The plan is to keep CPU -> NODE
576 * mapping valid and stable across CPU on/offlines. Once that
577 * happens, this workaround can be removed.
579 if (unlikely(node
== NUMA_NO_NODE
))
582 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
585 static unsigned int work_color_to_flags(int color
)
587 return color
<< WORK_STRUCT_COLOR_SHIFT
;
590 static int get_work_color(struct work_struct
*work
)
592 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
593 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
596 static int work_next_color(int color
)
598 return (color
+ 1) % WORK_NR_COLORS
;
602 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
603 * contain the pointer to the queued pwq. Once execution starts, the flag
604 * is cleared and the high bits contain OFFQ flags and pool ID.
606 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
607 * and clear_work_data() can be used to set the pwq, pool or clear
608 * work->data. These functions should only be called while the work is
609 * owned - ie. while the PENDING bit is set.
611 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
612 * corresponding to a work. Pool is available once the work has been
613 * queued anywhere after initialization until it is sync canceled. pwq is
614 * available only while the work item is queued.
616 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
617 * canceled. While being canceled, a work item may have its PENDING set
618 * but stay off timer and worklist for arbitrarily long and nobody should
619 * try to steal the PENDING bit.
621 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
624 WARN_ON_ONCE(!work_pending(work
));
625 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
628 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
629 unsigned long extra_flags
)
631 set_work_data(work
, (unsigned long)pwq
,
632 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
635 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
638 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
639 WORK_STRUCT_PENDING
);
642 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
646 * The following wmb is paired with the implied mb in
647 * test_and_set_bit(PENDING) and ensures all updates to @work made
648 * here are visible to and precede any updates by the next PENDING
652 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
654 * The following mb guarantees that previous clear of a PENDING bit
655 * will not be reordered with any speculative LOADS or STORES from
656 * work->current_func, which is executed afterwards. This possible
657 * reordering can lead to a missed execution on attempt to qeueue
658 * the same @work. E.g. consider this case:
661 * ---------------------------- --------------------------------
663 * 1 STORE event_indicated
664 * 2 queue_work_on() {
665 * 3 test_and_set_bit(PENDING)
666 * 4 } set_..._and_clear_pending() {
667 * 5 set_work_data() # clear bit
669 * 7 work->current_func() {
670 * 8 LOAD event_indicated
673 * Without an explicit full barrier speculative LOAD on line 8 can
674 * be executed before CPU#0 does STORE on line 1. If that happens,
675 * CPU#0 observes the PENDING bit is still set and new execution of
676 * a @work is not queued in a hope, that CPU#1 will eventually
677 * finish the queued @work. Meanwhile CPU#1 does not see
678 * event_indicated is set, because speculative LOAD was executed
679 * before actual STORE.
684 static void clear_work_data(struct work_struct
*work
)
686 smp_wmb(); /* see set_work_pool_and_clear_pending() */
687 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
690 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
692 unsigned long data
= atomic_long_read(&work
->data
);
694 if (data
& WORK_STRUCT_PWQ
)
695 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
701 * get_work_pool - return the worker_pool a given work was associated with
702 * @work: the work item of interest
704 * Pools are created and destroyed under wq_pool_mutex, and allows read
705 * access under sched-RCU read lock. As such, this function should be
706 * called under wq_pool_mutex or with preemption disabled.
708 * All fields of the returned pool are accessible as long as the above
709 * mentioned locking is in effect. If the returned pool needs to be used
710 * beyond the critical section, the caller is responsible for ensuring the
711 * returned pool is and stays online.
713 * Return: The worker_pool @work was last associated with. %NULL if none.
715 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
717 unsigned long data
= atomic_long_read(&work
->data
);
720 assert_rcu_or_pool_mutex();
722 if (data
& WORK_STRUCT_PWQ
)
723 return ((struct pool_workqueue
*)
724 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
726 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
727 if (pool_id
== WORK_OFFQ_POOL_NONE
)
730 return idr_find(&worker_pool_idr
, pool_id
);
734 * get_work_pool_id - return the worker pool ID a given work is associated with
735 * @work: the work item of interest
737 * Return: The worker_pool ID @work was last associated with.
738 * %WORK_OFFQ_POOL_NONE if none.
740 static int get_work_pool_id(struct work_struct
*work
)
742 unsigned long data
= atomic_long_read(&work
->data
);
744 if (data
& WORK_STRUCT_PWQ
)
745 return ((struct pool_workqueue
*)
746 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
748 return data
>> WORK_OFFQ_POOL_SHIFT
;
751 static void mark_work_canceling(struct work_struct
*work
)
753 unsigned long pool_id
= get_work_pool_id(work
);
755 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
756 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
759 static bool work_is_canceling(struct work_struct
*work
)
761 unsigned long data
= atomic_long_read(&work
->data
);
763 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
767 * Policy functions. These define the policies on how the global worker
768 * pools are managed. Unless noted otherwise, these functions assume that
769 * they're being called with pool->lock held.
772 static bool __need_more_worker(struct worker_pool
*pool
)
774 return !atomic_read(&pool
->nr_running
);
778 * Need to wake up a worker? Called from anything but currently
781 * Note that, because unbound workers never contribute to nr_running, this
782 * function will always return %true for unbound pools as long as the
783 * worklist isn't empty.
785 static bool need_more_worker(struct worker_pool
*pool
)
787 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
790 /* Can I start working? Called from busy but !running workers. */
791 static bool may_start_working(struct worker_pool
*pool
)
793 return pool
->nr_idle
;
796 /* Do I need to keep working? Called from currently running workers. */
797 static bool keep_working(struct worker_pool
*pool
)
799 return !list_empty(&pool
->worklist
) &&
800 atomic_read(&pool
->nr_running
) <= 1;
803 /* Do we need a new worker? Called from manager. */
804 static bool need_to_create_worker(struct worker_pool
*pool
)
806 return need_more_worker(pool
) && !may_start_working(pool
);
809 /* Do we have too many workers and should some go away? */
810 static bool too_many_workers(struct worker_pool
*pool
)
812 bool managing
= pool
->flags
& POOL_MANAGER_ACTIVE
;
813 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
814 int nr_busy
= pool
->nr_workers
- nr_idle
;
816 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
823 /* Return the first idle worker. Safe with preemption disabled */
824 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
826 if (unlikely(list_empty(&pool
->idle_list
)))
829 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
833 * wake_up_worker - wake up an idle worker
834 * @pool: worker pool to wake worker from
836 * Wake up the first idle worker of @pool.
839 * spin_lock_irq(pool->lock).
841 static void wake_up_worker(struct worker_pool
*pool
)
843 struct worker
*worker
= first_idle_worker(pool
);
846 wake_up_process(worker
->task
);
850 * wq_worker_waking_up - a worker is waking up
851 * @task: task waking up
852 * @cpu: CPU @task is waking up to
854 * This function is called during try_to_wake_up() when a worker is
858 * spin_lock_irq(rq->lock)
860 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
862 struct worker
*worker
= kthread_data(task
);
864 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
865 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
866 atomic_inc(&worker
->pool
->nr_running
);
871 * wq_worker_sleeping - a worker is going to sleep
872 * @task: task going to sleep
873 * @cpu: CPU in question, must be the current CPU number
875 * This function is called during schedule() when a busy worker is
876 * going to sleep. Worker on the same cpu can be woken up by
877 * returning pointer to its task.
880 * spin_lock_irq(rq->lock)
883 * Worker task on @cpu to wake up, %NULL if none.
885 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
887 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
888 struct worker_pool
*pool
;
891 * Rescuers, which may not have all the fields set up like normal
892 * workers, also reach here, let's not access anything before
893 * checking NOT_RUNNING.
895 if (worker
->flags
& WORKER_NOT_RUNNING
)
900 /* this can only happen on the local cpu */
901 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id() || pool
->cpu
!= cpu
))
905 * The counterpart of the following dec_and_test, implied mb,
906 * worklist not empty test sequence is in insert_work().
907 * Please read comment there.
909 * NOT_RUNNING is clear. This means that we're bound to and
910 * running on the local cpu w/ rq lock held and preemption
911 * disabled, which in turn means that none else could be
912 * manipulating idle_list, so dereferencing idle_list without pool
915 if (atomic_dec_and_test(&pool
->nr_running
) &&
916 !list_empty(&pool
->worklist
))
917 to_wakeup
= first_idle_worker(pool
);
918 return to_wakeup
? to_wakeup
->task
: NULL
;
922 * worker_set_flags - set worker flags and adjust nr_running accordingly
924 * @flags: flags to set
926 * Set @flags in @worker->flags and adjust nr_running accordingly.
929 * spin_lock_irq(pool->lock)
931 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
933 struct worker_pool
*pool
= worker
->pool
;
935 WARN_ON_ONCE(worker
->task
!= current
);
937 /* If transitioning into NOT_RUNNING, adjust nr_running. */
938 if ((flags
& WORKER_NOT_RUNNING
) &&
939 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
940 atomic_dec(&pool
->nr_running
);
943 worker
->flags
|= flags
;
947 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
949 * @flags: flags to clear
951 * Clear @flags in @worker->flags and adjust nr_running accordingly.
954 * spin_lock_irq(pool->lock)
956 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
958 struct worker_pool
*pool
= worker
->pool
;
959 unsigned int oflags
= worker
->flags
;
961 WARN_ON_ONCE(worker
->task
!= current
);
963 worker
->flags
&= ~flags
;
966 * If transitioning out of NOT_RUNNING, increment nr_running. Note
967 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
968 * of multiple flags, not a single flag.
970 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
971 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
972 atomic_inc(&pool
->nr_running
);
976 * find_worker_executing_work - find worker which is executing a work
977 * @pool: pool of interest
978 * @work: work to find worker for
980 * Find a worker which is executing @work on @pool by searching
981 * @pool->busy_hash which is keyed by the address of @work. For a worker
982 * to match, its current execution should match the address of @work and
983 * its work function. This is to avoid unwanted dependency between
984 * unrelated work executions through a work item being recycled while still
987 * This is a bit tricky. A work item may be freed once its execution
988 * starts and nothing prevents the freed area from being recycled for
989 * another work item. If the same work item address ends up being reused
990 * before the original execution finishes, workqueue will identify the
991 * recycled work item as currently executing and make it wait until the
992 * current execution finishes, introducing an unwanted dependency.
994 * This function checks the work item address and work function to avoid
995 * false positives. Note that this isn't complete as one may construct a
996 * work function which can introduce dependency onto itself through a
997 * recycled work item. Well, if somebody wants to shoot oneself in the
998 * foot that badly, there's only so much we can do, and if such deadlock
999 * actually occurs, it should be easy to locate the culprit work function.
1002 * spin_lock_irq(pool->lock).
1005 * Pointer to worker which is executing @work if found, %NULL
1008 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
1009 struct work_struct
*work
)
1011 struct worker
*worker
;
1013 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1014 (unsigned long)work
)
1015 if (worker
->current_work
== work
&&
1016 worker
->current_func
== work
->func
)
1023 * move_linked_works - move linked works to a list
1024 * @work: start of series of works to be scheduled
1025 * @head: target list to append @work to
1026 * @nextp: out parameter for nested worklist walking
1028 * Schedule linked works starting from @work to @head. Work series to
1029 * be scheduled starts at @work and includes any consecutive work with
1030 * WORK_STRUCT_LINKED set in its predecessor.
1032 * If @nextp is not NULL, it's updated to point to the next work of
1033 * the last scheduled work. This allows move_linked_works() to be
1034 * nested inside outer list_for_each_entry_safe().
1037 * spin_lock_irq(pool->lock).
1039 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1040 struct work_struct
**nextp
)
1042 struct work_struct
*n
;
1045 * Linked worklist will always end before the end of the list,
1046 * use NULL for list head.
1048 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1049 list_move_tail(&work
->entry
, head
);
1050 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1055 * If we're already inside safe list traversal and have moved
1056 * multiple works to the scheduled queue, the next position
1057 * needs to be updated.
1064 * get_pwq - get an extra reference on the specified pool_workqueue
1065 * @pwq: pool_workqueue to get
1067 * Obtain an extra reference on @pwq. The caller should guarantee that
1068 * @pwq has positive refcnt and be holding the matching pool->lock.
1070 static void get_pwq(struct pool_workqueue
*pwq
)
1072 lockdep_assert_held(&pwq
->pool
->lock
);
1073 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1078 * put_pwq - put a pool_workqueue reference
1079 * @pwq: pool_workqueue to put
1081 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1082 * destruction. The caller should be holding the matching pool->lock.
1084 static void put_pwq(struct pool_workqueue
*pwq
)
1086 lockdep_assert_held(&pwq
->pool
->lock
);
1087 if (likely(--pwq
->refcnt
))
1089 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1092 * @pwq can't be released under pool->lock, bounce to
1093 * pwq_unbound_release_workfn(). This never recurses on the same
1094 * pool->lock as this path is taken only for unbound workqueues and
1095 * the release work item is scheduled on a per-cpu workqueue. To
1096 * avoid lockdep warning, unbound pool->locks are given lockdep
1097 * subclass of 1 in get_unbound_pool().
1099 schedule_work(&pwq
->unbound_release_work
);
1103 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1104 * @pwq: pool_workqueue to put (can be %NULL)
1106 * put_pwq() with locking. This function also allows %NULL @pwq.
1108 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1112 * As both pwqs and pools are sched-RCU protected, the
1113 * following lock operations are safe.
1115 spin_lock_irq(&pwq
->pool
->lock
);
1117 spin_unlock_irq(&pwq
->pool
->lock
);
1121 static void pwq_activate_delayed_work(struct work_struct
*work
)
1123 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1125 trace_workqueue_activate_work(work
);
1126 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1127 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1131 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1133 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1134 struct work_struct
, entry
);
1136 pwq_activate_delayed_work(work
);
1140 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1141 * @pwq: pwq of interest
1142 * @color: color of work which left the queue
1144 * A work either has completed or is removed from pending queue,
1145 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1148 * spin_lock_irq(pool->lock).
1150 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1152 /* uncolored work items don't participate in flushing or nr_active */
1153 if (color
== WORK_NO_COLOR
)
1156 pwq
->nr_in_flight
[color
]--;
1159 if (!list_empty(&pwq
->delayed_works
)) {
1160 /* one down, submit a delayed one */
1161 if (pwq
->nr_active
< pwq
->max_active
)
1162 pwq_activate_first_delayed(pwq
);
1165 /* is flush in progress and are we at the flushing tip? */
1166 if (likely(pwq
->flush_color
!= color
))
1169 /* are there still in-flight works? */
1170 if (pwq
->nr_in_flight
[color
])
1173 /* this pwq is done, clear flush_color */
1174 pwq
->flush_color
= -1;
1177 * If this was the last pwq, wake up the first flusher. It
1178 * will handle the rest.
1180 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1181 complete(&pwq
->wq
->first_flusher
->done
);
1187 * try_to_grab_pending - steal work item from worklist and disable irq
1188 * @work: work item to steal
1189 * @is_dwork: @work is a delayed_work
1190 * @flags: place to store irq state
1192 * Try to grab PENDING bit of @work. This function can handle @work in any
1193 * stable state - idle, on timer or on worklist.
1196 * 1 if @work was pending and we successfully stole PENDING
1197 * 0 if @work was idle and we claimed PENDING
1198 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1199 * -ENOENT if someone else is canceling @work, this state may persist
1200 * for arbitrarily long
1203 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1204 * interrupted while holding PENDING and @work off queue, irq must be
1205 * disabled on entry. This, combined with delayed_work->timer being
1206 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1208 * On successful return, >= 0, irq is disabled and the caller is
1209 * responsible for releasing it using local_irq_restore(*@flags).
1211 * This function is safe to call from any context including IRQ handler.
1213 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1214 unsigned long *flags
)
1216 struct worker_pool
*pool
;
1217 struct pool_workqueue
*pwq
;
1219 local_irq_save(*flags
);
1221 /* try to steal the timer if it exists */
1223 struct delayed_work
*dwork
= to_delayed_work(work
);
1226 * dwork->timer is irqsafe. If del_timer() fails, it's
1227 * guaranteed that the timer is not queued anywhere and not
1228 * running on the local CPU.
1230 if (likely(del_timer(&dwork
->timer
)))
1234 /* try to claim PENDING the normal way */
1235 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1239 * The queueing is in progress, or it is already queued. Try to
1240 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1242 pool
= get_work_pool(work
);
1246 spin_lock(&pool
->lock
);
1248 * work->data is guaranteed to point to pwq only while the work
1249 * item is queued on pwq->wq, and both updating work->data to point
1250 * to pwq on queueing and to pool on dequeueing are done under
1251 * pwq->pool->lock. This in turn guarantees that, if work->data
1252 * points to pwq which is associated with a locked pool, the work
1253 * item is currently queued on that pool.
1255 pwq
= get_work_pwq(work
);
1256 if (pwq
&& pwq
->pool
== pool
) {
1257 debug_work_deactivate(work
);
1260 * A delayed work item cannot be grabbed directly because
1261 * it might have linked NO_COLOR work items which, if left
1262 * on the delayed_list, will confuse pwq->nr_active
1263 * management later on and cause stall. Make sure the work
1264 * item is activated before grabbing.
1266 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1267 pwq_activate_delayed_work(work
);
1269 list_del_init(&work
->entry
);
1270 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1272 /* work->data points to pwq iff queued, point to pool */
1273 set_work_pool_and_keep_pending(work
, pool
->id
);
1275 spin_unlock(&pool
->lock
);
1278 spin_unlock(&pool
->lock
);
1280 local_irq_restore(*flags
);
1281 if (work_is_canceling(work
))
1288 * insert_work - insert a work into a pool
1289 * @pwq: pwq @work belongs to
1290 * @work: work to insert
1291 * @head: insertion point
1292 * @extra_flags: extra WORK_STRUCT_* flags to set
1294 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1295 * work_struct flags.
1298 * spin_lock_irq(pool->lock).
1300 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1301 struct list_head
*head
, unsigned int extra_flags
)
1303 struct worker_pool
*pool
= pwq
->pool
;
1305 /* we own @work, set data and link */
1306 set_work_pwq(work
, pwq
, extra_flags
);
1307 list_add_tail(&work
->entry
, head
);
1311 * Ensure either wq_worker_sleeping() sees the above
1312 * list_add_tail() or we see zero nr_running to avoid workers lying
1313 * around lazily while there are works to be processed.
1317 if (__need_more_worker(pool
))
1318 wake_up_worker(pool
);
1322 * Test whether @work is being queued from another work executing on the
1325 static bool is_chained_work(struct workqueue_struct
*wq
)
1327 struct worker
*worker
;
1329 worker
= current_wq_worker();
1331 * Return %true iff I'm a worker execuing a work item on @wq. If
1332 * I'm @worker, it's safe to dereference it without locking.
1334 return worker
&& worker
->current_pwq
->wq
== wq
;
1337 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1338 struct work_struct
*work
)
1340 struct pool_workqueue
*pwq
;
1341 struct worker_pool
*last_pool
;
1342 struct list_head
*worklist
;
1343 unsigned int work_flags
;
1344 unsigned int req_cpu
= cpu
;
1347 * While a work item is PENDING && off queue, a task trying to
1348 * steal the PENDING will busy-loop waiting for it to either get
1349 * queued or lose PENDING. Grabbing PENDING and queueing should
1350 * happen with IRQ disabled.
1352 WARN_ON_ONCE(!irqs_disabled());
1354 debug_work_activate(work
);
1356 /* if draining, only works from the same workqueue are allowed */
1357 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1358 WARN_ON_ONCE(!is_chained_work(wq
)))
1361 if (req_cpu
== WORK_CPU_UNBOUND
)
1362 cpu
= raw_smp_processor_id();
1364 /* pwq which will be used unless @work is executing elsewhere */
1365 if (!(wq
->flags
& WQ_UNBOUND
))
1366 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1368 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1371 * If @work was previously on a different pool, it might still be
1372 * running there, in which case the work needs to be queued on that
1373 * pool to guarantee non-reentrancy.
1375 last_pool
= get_work_pool(work
);
1376 if (last_pool
&& last_pool
!= pwq
->pool
) {
1377 struct worker
*worker
;
1379 spin_lock(&last_pool
->lock
);
1381 worker
= find_worker_executing_work(last_pool
, work
);
1383 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1384 pwq
= worker
->current_pwq
;
1386 /* meh... not running there, queue here */
1387 spin_unlock(&last_pool
->lock
);
1388 spin_lock(&pwq
->pool
->lock
);
1391 spin_lock(&pwq
->pool
->lock
);
1395 * pwq is determined and locked. For unbound pools, we could have
1396 * raced with pwq release and it could already be dead. If its
1397 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1398 * without another pwq replacing it in the numa_pwq_tbl or while
1399 * work items are executing on it, so the retrying is guaranteed to
1400 * make forward-progress.
1402 if (unlikely(!pwq
->refcnt
)) {
1403 if (wq
->flags
& WQ_UNBOUND
) {
1404 spin_unlock(&pwq
->pool
->lock
);
1409 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1413 /* pwq determined, queue */
1414 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1416 if (WARN_ON(!list_empty(&work
->entry
))) {
1417 spin_unlock(&pwq
->pool
->lock
);
1421 pwq
->nr_in_flight
[pwq
->work_color
]++;
1422 work_flags
= work_color_to_flags(pwq
->work_color
);
1424 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1425 trace_workqueue_activate_work(work
);
1427 worklist
= &pwq
->pool
->worklist
;
1429 work_flags
|= WORK_STRUCT_DELAYED
;
1430 worklist
= &pwq
->delayed_works
;
1433 insert_work(pwq
, work
, worklist
, work_flags
);
1435 spin_unlock(&pwq
->pool
->lock
);
1439 * queue_work_on - queue work on specific cpu
1440 * @cpu: CPU number to execute work on
1441 * @wq: workqueue to use
1442 * @work: work to queue
1444 * We queue the work to a specific CPU, the caller must ensure it
1447 * Return: %false if @work was already on a queue, %true otherwise.
1449 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1450 struct work_struct
*work
)
1453 unsigned long flags
;
1455 local_irq_save(flags
);
1457 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1458 __queue_work(cpu
, wq
, work
);
1462 local_irq_restore(flags
);
1465 EXPORT_SYMBOL(queue_work_on
);
1467 void delayed_work_timer_fn(unsigned long __data
)
1469 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1471 /* should have been called from irqsafe timer with irq already off */
1472 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1474 EXPORT_SYMBOL(delayed_work_timer_fn
);
1476 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1477 struct delayed_work
*dwork
, unsigned long delay
)
1479 struct timer_list
*timer
= &dwork
->timer
;
1480 struct work_struct
*work
= &dwork
->work
;
1483 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1484 timer
->data
!= (unsigned long)dwork
);
1485 WARN_ON_ONCE(timer_pending(timer
));
1486 WARN_ON_ONCE(!list_empty(&work
->entry
));
1489 * If @delay is 0, queue @dwork->work immediately. This is for
1490 * both optimization and correctness. The earliest @timer can
1491 * expire is on the closest next tick and delayed_work users depend
1492 * on that there's no such delay when @delay is 0.
1495 __queue_work(cpu
, wq
, &dwork
->work
);
1499 timer_stats_timer_set_start_info(&dwork
->timer
);
1503 timer
->expires
= jiffies
+ delay
;
1505 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1506 add_timer_on(timer
, cpu
);
1512 * queue_delayed_work_on - queue work on specific CPU after delay
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 * Return: %false if @work was already on a queue, %true otherwise. If
1519 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1522 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1523 struct delayed_work
*dwork
, unsigned long delay
)
1525 struct work_struct
*work
= &dwork
->work
;
1527 unsigned long flags
;
1529 /* read the comment in __queue_work() */
1530 local_irq_save(flags
);
1532 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1533 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1537 local_irq_restore(flags
);
1540 EXPORT_SYMBOL(queue_delayed_work_on
);
1543 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1544 * @cpu: CPU number to execute work on
1545 * @wq: workqueue to use
1546 * @dwork: work to queue
1547 * @delay: number of jiffies to wait before queueing
1549 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1550 * modify @dwork's timer so that it expires after @delay. If @delay is
1551 * zero, @work is guaranteed to be scheduled immediately regardless of its
1554 * Return: %false if @dwork was idle and queued, %true if @dwork was
1555 * pending and its timer was modified.
1557 * This function is safe to call from any context including IRQ handler.
1558 * See try_to_grab_pending() for details.
1560 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1561 struct delayed_work
*dwork
, unsigned long delay
)
1563 unsigned long flags
;
1567 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1568 } while (unlikely(ret
== -EAGAIN
));
1570 if (likely(ret
>= 0)) {
1571 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1572 local_irq_restore(flags
);
1575 /* -ENOENT from try_to_grab_pending() becomes %true */
1578 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1581 * worker_enter_idle - enter idle state
1582 * @worker: worker which is entering idle state
1584 * @worker is entering idle state. Update stats and idle timer if
1588 * spin_lock_irq(pool->lock).
1590 static void worker_enter_idle(struct worker
*worker
)
1592 struct worker_pool
*pool
= worker
->pool
;
1594 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1595 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1596 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1599 /* can't use worker_set_flags(), also called from create_worker() */
1600 worker
->flags
|= WORKER_IDLE
;
1602 worker
->last_active
= jiffies
;
1604 /* idle_list is LIFO */
1605 list_add(&worker
->entry
, &pool
->idle_list
);
1607 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1608 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1611 * Sanity check nr_running. Because wq_unbind_fn() releases
1612 * pool->lock between setting %WORKER_UNBOUND and zapping
1613 * nr_running, the warning may trigger spuriously. Check iff
1614 * unbind is not in progress.
1616 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1617 pool
->nr_workers
== pool
->nr_idle
&&
1618 atomic_read(&pool
->nr_running
));
1622 * worker_leave_idle - leave idle state
1623 * @worker: worker which is leaving idle state
1625 * @worker is leaving idle state. Update stats.
1628 * spin_lock_irq(pool->lock).
1630 static void worker_leave_idle(struct worker
*worker
)
1632 struct worker_pool
*pool
= worker
->pool
;
1634 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1636 worker_clr_flags(worker
, WORKER_IDLE
);
1638 list_del_init(&worker
->entry
);
1641 static struct worker
*alloc_worker(int node
)
1643 struct worker
*worker
;
1645 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1647 INIT_LIST_HEAD(&worker
->entry
);
1648 INIT_LIST_HEAD(&worker
->scheduled
);
1649 INIT_LIST_HEAD(&worker
->node
);
1650 /* on creation a worker is in !idle && prep state */
1651 worker
->flags
= WORKER_PREP
;
1657 * worker_attach_to_pool() - attach a worker to a pool
1658 * @worker: worker to be attached
1659 * @pool: the target pool
1661 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1662 * cpu-binding of @worker are kept coordinated with the pool across
1665 static void worker_attach_to_pool(struct worker
*worker
,
1666 struct worker_pool
*pool
)
1668 mutex_lock(&pool
->attach_mutex
);
1671 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1672 * online CPUs. It'll be re-applied when any of the CPUs come up.
1674 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1677 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1678 * stable across this function. See the comments above the
1679 * flag definition for details.
1681 if (pool
->flags
& POOL_DISASSOCIATED
)
1682 worker
->flags
|= WORKER_UNBOUND
;
1684 list_add_tail(&worker
->node
, &pool
->workers
);
1686 mutex_unlock(&pool
->attach_mutex
);
1690 * worker_detach_from_pool() - detach a worker from its pool
1691 * @worker: worker which is attached to its pool
1692 * @pool: the pool @worker is attached to
1694 * Undo the attaching which had been done in worker_attach_to_pool(). The
1695 * caller worker shouldn't access to the pool after detached except it has
1696 * other reference to the pool.
1698 static void worker_detach_from_pool(struct worker
*worker
,
1699 struct worker_pool
*pool
)
1701 struct completion
*detach_completion
= NULL
;
1703 mutex_lock(&pool
->attach_mutex
);
1704 list_del(&worker
->node
);
1705 if (list_empty(&pool
->workers
))
1706 detach_completion
= pool
->detach_completion
;
1707 mutex_unlock(&pool
->attach_mutex
);
1709 /* clear leftover flags without pool->lock after it is detached */
1710 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1712 if (detach_completion
)
1713 complete(detach_completion
);
1717 * create_worker - create a new workqueue worker
1718 * @pool: pool the new worker will belong to
1720 * Create and start a new worker which is attached to @pool.
1723 * Might sleep. Does GFP_KERNEL allocations.
1726 * Pointer to the newly created worker.
1728 static struct worker
*create_worker(struct worker_pool
*pool
)
1730 struct worker
*worker
= NULL
;
1734 /* ID is needed to determine kthread name */
1735 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1739 worker
= alloc_worker(pool
->node
);
1743 worker
->pool
= pool
;
1747 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1748 pool
->attrs
->nice
< 0 ? "H" : "");
1750 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1752 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1753 "kworker/%s", id_buf
);
1754 if (IS_ERR(worker
->task
))
1757 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1758 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1760 /* successful, attach the worker to the pool */
1761 worker_attach_to_pool(worker
, pool
);
1763 /* start the newly created worker */
1764 spin_lock_irq(&pool
->lock
);
1765 worker
->pool
->nr_workers
++;
1766 worker_enter_idle(worker
);
1767 wake_up_process(worker
->task
);
1768 spin_unlock_irq(&pool
->lock
);
1774 ida_simple_remove(&pool
->worker_ida
, id
);
1780 * destroy_worker - destroy a workqueue worker
1781 * @worker: worker to be destroyed
1783 * Destroy @worker and adjust @pool stats accordingly. The worker should
1787 * spin_lock_irq(pool->lock).
1789 static void destroy_worker(struct worker
*worker
)
1791 struct worker_pool
*pool
= worker
->pool
;
1793 lockdep_assert_held(&pool
->lock
);
1795 /* sanity check frenzy */
1796 if (WARN_ON(worker
->current_work
) ||
1797 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1798 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1804 list_del_init(&worker
->entry
);
1805 worker
->flags
|= WORKER_DIE
;
1806 wake_up_process(worker
->task
);
1809 static void idle_worker_timeout(unsigned long __pool
)
1811 struct worker_pool
*pool
= (void *)__pool
;
1813 spin_lock_irq(&pool
->lock
);
1815 while (too_many_workers(pool
)) {
1816 struct worker
*worker
;
1817 unsigned long expires
;
1819 /* idle_list is kept in LIFO order, check the last one */
1820 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1821 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1823 if (time_before(jiffies
, expires
)) {
1824 mod_timer(&pool
->idle_timer
, expires
);
1828 destroy_worker(worker
);
1831 spin_unlock_irq(&pool
->lock
);
1834 static void send_mayday(struct work_struct
*work
)
1836 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1837 struct workqueue_struct
*wq
= pwq
->wq
;
1839 lockdep_assert_held(&wq_mayday_lock
);
1844 /* mayday mayday mayday */
1845 if (list_empty(&pwq
->mayday_node
)) {
1847 * If @pwq is for an unbound wq, its base ref may be put at
1848 * any time due to an attribute change. Pin @pwq until the
1849 * rescuer is done with it.
1852 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1853 wake_up_process(wq
->rescuer
->task
);
1857 static void pool_mayday_timeout(unsigned long __pool
)
1859 struct worker_pool
*pool
= (void *)__pool
;
1860 struct work_struct
*work
;
1862 spin_lock_irq(&pool
->lock
);
1863 spin_lock(&wq_mayday_lock
); /* for wq->maydays */
1865 if (need_to_create_worker(pool
)) {
1867 * We've been trying to create a new worker but
1868 * haven't been successful. We might be hitting an
1869 * allocation deadlock. Send distress signals to
1872 list_for_each_entry(work
, &pool
->worklist
, entry
)
1876 spin_unlock(&wq_mayday_lock
);
1877 spin_unlock_irq(&pool
->lock
);
1879 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1883 * maybe_create_worker - create a new worker if necessary
1884 * @pool: pool to create a new worker for
1886 * Create a new worker for @pool if necessary. @pool is guaranteed to
1887 * have at least one idle worker on return from this function. If
1888 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1889 * sent to all rescuers with works scheduled on @pool to resolve
1890 * possible allocation deadlock.
1892 * On return, need_to_create_worker() is guaranteed to be %false and
1893 * may_start_working() %true.
1896 * spin_lock_irq(pool->lock) which may be released and regrabbed
1897 * multiple times. Does GFP_KERNEL allocations. Called only from
1900 static void maybe_create_worker(struct worker_pool
*pool
)
1901 __releases(&pool
->lock
)
1902 __acquires(&pool
->lock
)
1905 spin_unlock_irq(&pool
->lock
);
1907 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1908 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1911 if (create_worker(pool
) || !need_to_create_worker(pool
))
1914 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1916 if (!need_to_create_worker(pool
))
1920 del_timer_sync(&pool
->mayday_timer
);
1921 spin_lock_irq(&pool
->lock
);
1923 * This is necessary even after a new worker was just successfully
1924 * created as @pool->lock was dropped and the new worker might have
1925 * already become busy.
1927 if (need_to_create_worker(pool
))
1932 * manage_workers - manage worker pool
1935 * Assume the manager role and manage the worker pool @worker belongs
1936 * to. At any given time, there can be only zero or one manager per
1937 * pool. The exclusion is handled automatically by this function.
1939 * The caller can safely start processing works on false return. On
1940 * true return, it's guaranteed that need_to_create_worker() is false
1941 * and may_start_working() is true.
1944 * spin_lock_irq(pool->lock) which may be released and regrabbed
1945 * multiple times. Does GFP_KERNEL allocations.
1948 * %false if the pool doesn't need management and the caller can safely
1949 * start processing works, %true if management function was performed and
1950 * the conditions that the caller verified before calling the function may
1951 * no longer be true.
1953 static bool manage_workers(struct worker
*worker
)
1955 struct worker_pool
*pool
= worker
->pool
;
1957 if (pool
->flags
& POOL_MANAGER_ACTIVE
)
1960 pool
->flags
|= POOL_MANAGER_ACTIVE
;
1961 pool
->manager
= worker
;
1963 maybe_create_worker(pool
);
1965 pool
->manager
= NULL
;
1966 pool
->flags
&= ~POOL_MANAGER_ACTIVE
;
1967 wake_up(&wq_manager_wait
);
1972 * process_one_work - process single work
1974 * @work: work to process
1976 * Process @work. This function contains all the logics necessary to
1977 * process a single work including synchronization against and
1978 * interaction with other workers on the same cpu, queueing and
1979 * flushing. As long as context requirement is met, any worker can
1980 * call this function to process a work.
1983 * spin_lock_irq(pool->lock) which is released and regrabbed.
1985 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1986 __releases(&pool
->lock
)
1987 __acquires(&pool
->lock
)
1989 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1990 struct worker_pool
*pool
= worker
->pool
;
1991 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1993 struct worker
*collision
;
1994 #ifdef CONFIG_LOCKDEP
1996 * It is permissible to free the struct work_struct from
1997 * inside the function that is called from it, this we need to
1998 * take into account for lockdep too. To avoid bogus "held
1999 * lock freed" warnings as well as problems when looking into
2000 * work->lockdep_map, make a copy and use that here.
2002 struct lockdep_map lockdep_map
;
2004 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2006 /* ensure we're on the correct CPU */
2007 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2008 raw_smp_processor_id() != pool
->cpu
);
2011 * A single work shouldn't be executed concurrently by
2012 * multiple workers on a single cpu. Check whether anyone is
2013 * already processing the work. If so, defer the work to the
2014 * currently executing one.
2016 collision
= find_worker_executing_work(pool
, work
);
2017 if (unlikely(collision
)) {
2018 move_linked_works(work
, &collision
->scheduled
, NULL
);
2022 /* claim and dequeue */
2023 debug_work_deactivate(work
);
2024 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2025 worker
->current_work
= work
;
2026 worker
->current_func
= work
->func
;
2027 worker
->current_pwq
= pwq
;
2028 work_color
= get_work_color(work
);
2030 list_del_init(&work
->entry
);
2033 * CPU intensive works don't participate in concurrency management.
2034 * They're the scheduler's responsibility. This takes @worker out
2035 * of concurrency management and the next code block will chain
2036 * execution of the pending work items.
2038 if (unlikely(cpu_intensive
))
2039 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2042 * Wake up another worker if necessary. The condition is always
2043 * false for normal per-cpu workers since nr_running would always
2044 * be >= 1 at this point. This is used to chain execution of the
2045 * pending work items for WORKER_NOT_RUNNING workers such as the
2046 * UNBOUND and CPU_INTENSIVE ones.
2048 if (need_more_worker(pool
))
2049 wake_up_worker(pool
);
2052 * Record the last pool and clear PENDING which should be the last
2053 * update to @work. Also, do this inside @pool->lock so that
2054 * PENDING and queued state changes happen together while IRQ is
2057 set_work_pool_and_clear_pending(work
, pool
->id
);
2059 spin_unlock_irq(&pool
->lock
);
2061 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2062 lock_map_acquire(&lockdep_map
);
2063 trace_workqueue_execute_start(work
);
2064 worker
->current_func(work
);
2066 * While we must be careful to not use "work" after this, the trace
2067 * point will only record its address.
2069 trace_workqueue_execute_end(work
);
2070 lock_map_release(&lockdep_map
);
2071 lock_map_release(&pwq
->wq
->lockdep_map
);
2073 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2074 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2075 " last function: %pf\n",
2076 current
->comm
, preempt_count(), task_pid_nr(current
),
2077 worker
->current_func
);
2078 debug_show_held_locks(current
);
2083 * The following prevents a kworker from hogging CPU on !PREEMPT
2084 * kernels, where a requeueing work item waiting for something to
2085 * happen could deadlock with stop_machine as such work item could
2086 * indefinitely requeue itself while all other CPUs are trapped in
2087 * stop_machine. At the same time, report a quiescent RCU state so
2088 * the same condition doesn't freeze RCU.
2090 cond_resched_rcu_qs();
2092 spin_lock_irq(&pool
->lock
);
2094 /* clear cpu intensive status */
2095 if (unlikely(cpu_intensive
))
2096 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2098 /* we're done with it, release */
2099 hash_del(&worker
->hentry
);
2100 worker
->current_work
= NULL
;
2101 worker
->current_func
= NULL
;
2102 worker
->current_pwq
= NULL
;
2103 worker
->desc_valid
= false;
2104 pwq_dec_nr_in_flight(pwq
, work_color
);
2108 * process_scheduled_works - process scheduled works
2111 * Process all scheduled works. Please note that the scheduled list
2112 * may change while processing a work, so this function repeatedly
2113 * fetches a work from the top and executes it.
2116 * spin_lock_irq(pool->lock) which may be released and regrabbed
2119 static void process_scheduled_works(struct worker
*worker
)
2121 while (!list_empty(&worker
->scheduled
)) {
2122 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2123 struct work_struct
, entry
);
2124 process_one_work(worker
, work
);
2129 * worker_thread - the worker thread function
2132 * The worker thread function. All workers belong to a worker_pool -
2133 * either a per-cpu one or dynamic unbound one. These workers process all
2134 * work items regardless of their specific target workqueue. The only
2135 * exception is work items which belong to workqueues with a rescuer which
2136 * will be explained in rescuer_thread().
2140 static int worker_thread(void *__worker
)
2142 struct worker
*worker
= __worker
;
2143 struct worker_pool
*pool
= worker
->pool
;
2145 /* tell the scheduler that this is a workqueue worker */
2146 worker
->task
->flags
|= PF_WQ_WORKER
;
2148 spin_lock_irq(&pool
->lock
);
2150 /* am I supposed to die? */
2151 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2152 spin_unlock_irq(&pool
->lock
);
2153 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2154 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2156 set_task_comm(worker
->task
, "kworker/dying");
2157 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2158 worker_detach_from_pool(worker
, pool
);
2163 worker_leave_idle(worker
);
2165 /* no more worker necessary? */
2166 if (!need_more_worker(pool
))
2169 /* do we need to manage? */
2170 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2174 * ->scheduled list can only be filled while a worker is
2175 * preparing to process a work or actually processing it.
2176 * Make sure nobody diddled with it while I was sleeping.
2178 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2181 * Finish PREP stage. We're guaranteed to have at least one idle
2182 * worker or that someone else has already assumed the manager
2183 * role. This is where @worker starts participating in concurrency
2184 * management if applicable and concurrency management is restored
2185 * after being rebound. See rebind_workers() for details.
2187 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2190 struct work_struct
*work
=
2191 list_first_entry(&pool
->worklist
,
2192 struct work_struct
, entry
);
2194 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2195 /* optimization path, not strictly necessary */
2196 process_one_work(worker
, work
);
2197 if (unlikely(!list_empty(&worker
->scheduled
)))
2198 process_scheduled_works(worker
);
2200 move_linked_works(work
, &worker
->scheduled
, NULL
);
2201 process_scheduled_works(worker
);
2203 } while (keep_working(pool
));
2205 worker_set_flags(worker
, WORKER_PREP
);
2208 * pool->lock is held and there's no work to process and no need to
2209 * manage, sleep. Workers are woken up only while holding
2210 * pool->lock or from local cpu, so setting the current state
2211 * before releasing pool->lock is enough to prevent losing any
2214 worker_enter_idle(worker
);
2215 __set_current_state(TASK_INTERRUPTIBLE
);
2216 spin_unlock_irq(&pool
->lock
);
2222 * rescuer_thread - the rescuer thread function
2225 * Workqueue rescuer thread function. There's one rescuer for each
2226 * workqueue which has WQ_MEM_RECLAIM set.
2228 * Regular work processing on a pool may block trying to create a new
2229 * worker which uses GFP_KERNEL allocation which has slight chance of
2230 * developing into deadlock if some works currently on the same queue
2231 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2232 * the problem rescuer solves.
2234 * When such condition is possible, the pool summons rescuers of all
2235 * workqueues which have works queued on the pool and let them process
2236 * those works so that forward progress can be guaranteed.
2238 * This should happen rarely.
2242 static int rescuer_thread(void *__rescuer
)
2244 struct worker
*rescuer
= __rescuer
;
2245 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2246 struct list_head
*scheduled
= &rescuer
->scheduled
;
2249 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2252 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2253 * doesn't participate in concurrency management.
2255 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2257 set_current_state(TASK_INTERRUPTIBLE
);
2260 * By the time the rescuer is requested to stop, the workqueue
2261 * shouldn't have any work pending, but @wq->maydays may still have
2262 * pwq(s) queued. This can happen by non-rescuer workers consuming
2263 * all the work items before the rescuer got to them. Go through
2264 * @wq->maydays processing before acting on should_stop so that the
2265 * list is always empty on exit.
2267 should_stop
= kthread_should_stop();
2269 /* see whether any pwq is asking for help */
2270 spin_lock_irq(&wq_mayday_lock
);
2272 while (!list_empty(&wq
->maydays
)) {
2273 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2274 struct pool_workqueue
, mayday_node
);
2275 struct worker_pool
*pool
= pwq
->pool
;
2276 struct work_struct
*work
, *n
;
2278 __set_current_state(TASK_RUNNING
);
2279 list_del_init(&pwq
->mayday_node
);
2281 spin_unlock_irq(&wq_mayday_lock
);
2283 worker_attach_to_pool(rescuer
, pool
);
2285 spin_lock_irq(&pool
->lock
);
2286 rescuer
->pool
= pool
;
2289 * Slurp in all works issued via this workqueue and
2292 WARN_ON_ONCE(!list_empty(scheduled
));
2293 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2294 if (get_work_pwq(work
) == pwq
)
2295 move_linked_works(work
, scheduled
, &n
);
2297 if (!list_empty(scheduled
)) {
2298 process_scheduled_works(rescuer
);
2301 * The above execution of rescued work items could
2302 * have created more to rescue through
2303 * pwq_activate_first_delayed() or chained
2304 * queueing. Let's put @pwq back on mayday list so
2305 * that such back-to-back work items, which may be
2306 * being used to relieve memory pressure, don't
2307 * incur MAYDAY_INTERVAL delay inbetween.
2309 if (need_to_create_worker(pool
)) {
2310 spin_lock(&wq_mayday_lock
);
2312 list_move_tail(&pwq
->mayday_node
, &wq
->maydays
);
2313 spin_unlock(&wq_mayday_lock
);
2318 * Put the reference grabbed by send_mayday(). @pool won't
2319 * go away while we're still attached to it.
2324 * Leave this pool. If need_more_worker() is %true, notify a
2325 * regular worker; otherwise, we end up with 0 concurrency
2326 * and stalling the execution.
2328 if (need_more_worker(pool
))
2329 wake_up_worker(pool
);
2331 rescuer
->pool
= NULL
;
2332 spin_unlock_irq(&pool
->lock
);
2334 worker_detach_from_pool(rescuer
, pool
);
2336 spin_lock_irq(&wq_mayday_lock
);
2339 spin_unlock_irq(&wq_mayday_lock
);
2342 __set_current_state(TASK_RUNNING
);
2343 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2347 /* rescuers should never participate in concurrency management */
2348 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2354 struct work_struct work
;
2355 struct completion done
;
2356 struct task_struct
*task
; /* purely informational */
2359 static void wq_barrier_func(struct work_struct
*work
)
2361 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2362 complete(&barr
->done
);
2366 * insert_wq_barrier - insert a barrier work
2367 * @pwq: pwq to insert barrier into
2368 * @barr: wq_barrier to insert
2369 * @target: target work to attach @barr to
2370 * @worker: worker currently executing @target, NULL if @target is not executing
2372 * @barr is linked to @target such that @barr is completed only after
2373 * @target finishes execution. Please note that the ordering
2374 * guarantee is observed only with respect to @target and on the local
2377 * Currently, a queued barrier can't be canceled. This is because
2378 * try_to_grab_pending() can't determine whether the work to be
2379 * grabbed is at the head of the queue and thus can't clear LINKED
2380 * flag of the previous work while there must be a valid next work
2381 * after a work with LINKED flag set.
2383 * Note that when @worker is non-NULL, @target may be modified
2384 * underneath us, so we can't reliably determine pwq from @target.
2387 * spin_lock_irq(pool->lock).
2389 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2390 struct wq_barrier
*barr
,
2391 struct work_struct
*target
, struct worker
*worker
)
2393 struct list_head
*head
;
2394 unsigned int linked
= 0;
2397 * debugobject calls are safe here even with pool->lock locked
2398 * as we know for sure that this will not trigger any of the
2399 * checks and call back into the fixup functions where we
2402 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2403 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2404 init_completion(&barr
->done
);
2405 barr
->task
= current
;
2408 * If @target is currently being executed, schedule the
2409 * barrier to the worker; otherwise, put it after @target.
2412 head
= worker
->scheduled
.next
;
2414 unsigned long *bits
= work_data_bits(target
);
2416 head
= target
->entry
.next
;
2417 /* there can already be other linked works, inherit and set */
2418 linked
= *bits
& WORK_STRUCT_LINKED
;
2419 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2422 debug_work_activate(&barr
->work
);
2423 insert_work(pwq
, &barr
->work
, head
,
2424 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2428 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2429 * @wq: workqueue being flushed
2430 * @flush_color: new flush color, < 0 for no-op
2431 * @work_color: new work color, < 0 for no-op
2433 * Prepare pwqs for workqueue flushing.
2435 * If @flush_color is non-negative, flush_color on all pwqs should be
2436 * -1. If no pwq has in-flight commands at the specified color, all
2437 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2438 * has in flight commands, its pwq->flush_color is set to
2439 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2440 * wakeup logic is armed and %true is returned.
2442 * The caller should have initialized @wq->first_flusher prior to
2443 * calling this function with non-negative @flush_color. If
2444 * @flush_color is negative, no flush color update is done and %false
2447 * If @work_color is non-negative, all pwqs should have the same
2448 * work_color which is previous to @work_color and all will be
2449 * advanced to @work_color.
2452 * mutex_lock(wq->mutex).
2455 * %true if @flush_color >= 0 and there's something to flush. %false
2458 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2459 int flush_color
, int work_color
)
2462 struct pool_workqueue
*pwq
;
2464 if (flush_color
>= 0) {
2465 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2466 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2469 for_each_pwq(pwq
, wq
) {
2470 struct worker_pool
*pool
= pwq
->pool
;
2472 spin_lock_irq(&pool
->lock
);
2474 if (flush_color
>= 0) {
2475 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2477 if (pwq
->nr_in_flight
[flush_color
]) {
2478 pwq
->flush_color
= flush_color
;
2479 atomic_inc(&wq
->nr_pwqs_to_flush
);
2484 if (work_color
>= 0) {
2485 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2486 pwq
->work_color
= work_color
;
2489 spin_unlock_irq(&pool
->lock
);
2492 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2493 complete(&wq
->first_flusher
->done
);
2499 * flush_workqueue - ensure that any scheduled work has run to completion.
2500 * @wq: workqueue to flush
2502 * This function sleeps until all work items which were queued on entry
2503 * have finished execution, but it is not livelocked by new incoming ones.
2505 void flush_workqueue(struct workqueue_struct
*wq
)
2507 struct wq_flusher this_flusher
= {
2508 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2510 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2514 lock_map_acquire(&wq
->lockdep_map
);
2515 lock_map_release(&wq
->lockdep_map
);
2517 mutex_lock(&wq
->mutex
);
2520 * Start-to-wait phase
2522 next_color
= work_next_color(wq
->work_color
);
2524 if (next_color
!= wq
->flush_color
) {
2526 * Color space is not full. The current work_color
2527 * becomes our flush_color and work_color is advanced
2530 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2531 this_flusher
.flush_color
= wq
->work_color
;
2532 wq
->work_color
= next_color
;
2534 if (!wq
->first_flusher
) {
2535 /* no flush in progress, become the first flusher */
2536 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2538 wq
->first_flusher
= &this_flusher
;
2540 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2542 /* nothing to flush, done */
2543 wq
->flush_color
= next_color
;
2544 wq
->first_flusher
= NULL
;
2549 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2550 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2551 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2555 * Oops, color space is full, wait on overflow queue.
2556 * The next flush completion will assign us
2557 * flush_color and transfer to flusher_queue.
2559 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2562 mutex_unlock(&wq
->mutex
);
2564 wait_for_completion(&this_flusher
.done
);
2567 * Wake-up-and-cascade phase
2569 * First flushers are responsible for cascading flushes and
2570 * handling overflow. Non-first flushers can simply return.
2572 if (wq
->first_flusher
!= &this_flusher
)
2575 mutex_lock(&wq
->mutex
);
2577 /* we might have raced, check again with mutex held */
2578 if (wq
->first_flusher
!= &this_flusher
)
2581 wq
->first_flusher
= NULL
;
2583 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2584 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2587 struct wq_flusher
*next
, *tmp
;
2589 /* complete all the flushers sharing the current flush color */
2590 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2591 if (next
->flush_color
!= wq
->flush_color
)
2593 list_del_init(&next
->list
);
2594 complete(&next
->done
);
2597 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2598 wq
->flush_color
!= work_next_color(wq
->work_color
));
2600 /* this flush_color is finished, advance by one */
2601 wq
->flush_color
= work_next_color(wq
->flush_color
);
2603 /* one color has been freed, handle overflow queue */
2604 if (!list_empty(&wq
->flusher_overflow
)) {
2606 * Assign the same color to all overflowed
2607 * flushers, advance work_color and append to
2608 * flusher_queue. This is the start-to-wait
2609 * phase for these overflowed flushers.
2611 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2612 tmp
->flush_color
= wq
->work_color
;
2614 wq
->work_color
= work_next_color(wq
->work_color
);
2616 list_splice_tail_init(&wq
->flusher_overflow
,
2617 &wq
->flusher_queue
);
2618 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2621 if (list_empty(&wq
->flusher_queue
)) {
2622 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2627 * Need to flush more colors. Make the next flusher
2628 * the new first flusher and arm pwqs.
2630 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2631 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2633 list_del_init(&next
->list
);
2634 wq
->first_flusher
= next
;
2636 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2640 * Meh... this color is already done, clear first
2641 * flusher and repeat cascading.
2643 wq
->first_flusher
= NULL
;
2647 mutex_unlock(&wq
->mutex
);
2649 EXPORT_SYMBOL(flush_workqueue
);
2652 * drain_workqueue - drain a workqueue
2653 * @wq: workqueue to drain
2655 * Wait until the workqueue becomes empty. While draining is in progress,
2656 * only chain queueing is allowed. IOW, only currently pending or running
2657 * work items on @wq can queue further work items on it. @wq is flushed
2658 * repeatedly until it becomes empty. The number of flushing is determined
2659 * by the depth of chaining and should be relatively short. Whine if it
2662 void drain_workqueue(struct workqueue_struct
*wq
)
2664 unsigned int flush_cnt
= 0;
2665 struct pool_workqueue
*pwq
;
2668 * __queue_work() needs to test whether there are drainers, is much
2669 * hotter than drain_workqueue() and already looks at @wq->flags.
2670 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2672 mutex_lock(&wq
->mutex
);
2673 if (!wq
->nr_drainers
++)
2674 wq
->flags
|= __WQ_DRAINING
;
2675 mutex_unlock(&wq
->mutex
);
2677 flush_workqueue(wq
);
2679 mutex_lock(&wq
->mutex
);
2681 for_each_pwq(pwq
, wq
) {
2684 spin_lock_irq(&pwq
->pool
->lock
);
2685 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2686 spin_unlock_irq(&pwq
->pool
->lock
);
2691 if (++flush_cnt
== 10 ||
2692 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2693 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2694 wq
->name
, flush_cnt
);
2696 mutex_unlock(&wq
->mutex
);
2700 if (!--wq
->nr_drainers
)
2701 wq
->flags
&= ~__WQ_DRAINING
;
2702 mutex_unlock(&wq
->mutex
);
2704 EXPORT_SYMBOL_GPL(drain_workqueue
);
2706 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2708 struct worker
*worker
= NULL
;
2709 struct worker_pool
*pool
;
2710 struct pool_workqueue
*pwq
;
2714 local_irq_disable();
2715 pool
= get_work_pool(work
);
2721 spin_lock(&pool
->lock
);
2722 /* see the comment in try_to_grab_pending() with the same code */
2723 pwq
= get_work_pwq(work
);
2725 if (unlikely(pwq
->pool
!= pool
))
2728 worker
= find_worker_executing_work(pool
, work
);
2731 pwq
= worker
->current_pwq
;
2734 insert_wq_barrier(pwq
, barr
, work
, worker
);
2735 spin_unlock_irq(&pool
->lock
);
2738 * If @max_active is 1 or rescuer is in use, flushing another work
2739 * item on the same workqueue may lead to deadlock. Make sure the
2740 * flusher is not running on the same workqueue by verifying write
2743 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2744 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2746 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2747 lock_map_release(&pwq
->wq
->lockdep_map
);
2751 spin_unlock_irq(&pool
->lock
);
2756 * flush_work - wait for a work to finish executing the last queueing instance
2757 * @work: the work to flush
2759 * Wait until @work has finished execution. @work is guaranteed to be idle
2760 * on return if it hasn't been requeued since flush started.
2763 * %true if flush_work() waited for the work to finish execution,
2764 * %false if it was already idle.
2766 bool flush_work(struct work_struct
*work
)
2768 struct wq_barrier barr
;
2770 lock_map_acquire(&work
->lockdep_map
);
2771 lock_map_release(&work
->lockdep_map
);
2773 if (start_flush_work(work
, &barr
)) {
2774 wait_for_completion(&barr
.done
);
2775 destroy_work_on_stack(&barr
.work
);
2781 EXPORT_SYMBOL_GPL(flush_work
);
2785 struct work_struct
*work
;
2788 static int cwt_wakefn(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
2790 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
2792 if (cwait
->work
!= key
)
2794 return autoremove_wake_function(wait
, mode
, sync
, key
);
2797 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2799 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
2800 unsigned long flags
;
2804 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2806 * If someone else is already canceling, wait for it to
2807 * finish. flush_work() doesn't work for PREEMPT_NONE
2808 * because we may get scheduled between @work's completion
2809 * and the other canceling task resuming and clearing
2810 * CANCELING - flush_work() will return false immediately
2811 * as @work is no longer busy, try_to_grab_pending() will
2812 * return -ENOENT as @work is still being canceled and the
2813 * other canceling task won't be able to clear CANCELING as
2814 * we're hogging the CPU.
2816 * Let's wait for completion using a waitqueue. As this
2817 * may lead to the thundering herd problem, use a custom
2818 * wake function which matches @work along with exclusive
2821 if (unlikely(ret
== -ENOENT
)) {
2822 struct cwt_wait cwait
;
2824 init_wait(&cwait
.wait
);
2825 cwait
.wait
.func
= cwt_wakefn
;
2828 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
2829 TASK_UNINTERRUPTIBLE
);
2830 if (work_is_canceling(work
))
2832 finish_wait(&cancel_waitq
, &cwait
.wait
);
2834 } while (unlikely(ret
< 0));
2836 /* tell other tasks trying to grab @work to back off */
2837 mark_work_canceling(work
);
2838 local_irq_restore(flags
);
2841 clear_work_data(work
);
2844 * Paired with prepare_to_wait() above so that either
2845 * waitqueue_active() is visible here or !work_is_canceling() is
2849 if (waitqueue_active(&cancel_waitq
))
2850 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
2856 * cancel_work_sync - cancel a work and wait for it to finish
2857 * @work: the work to cancel
2859 * Cancel @work and wait for its execution to finish. This function
2860 * can be used even if the work re-queues itself or migrates to
2861 * another workqueue. On return from this function, @work is
2862 * guaranteed to be not pending or executing on any CPU.
2864 * cancel_work_sync(&delayed_work->work) must not be used for
2865 * delayed_work's. Use cancel_delayed_work_sync() instead.
2867 * The caller must ensure that the workqueue on which @work was last
2868 * queued can't be destroyed before this function returns.
2871 * %true if @work was pending, %false otherwise.
2873 bool cancel_work_sync(struct work_struct
*work
)
2875 return __cancel_work_timer(work
, false);
2877 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2880 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2881 * @dwork: the delayed work to flush
2883 * Delayed timer is cancelled and the pending work is queued for
2884 * immediate execution. Like flush_work(), this function only
2885 * considers the last queueing instance of @dwork.
2888 * %true if flush_work() waited for the work to finish execution,
2889 * %false if it was already idle.
2891 bool flush_delayed_work(struct delayed_work
*dwork
)
2893 local_irq_disable();
2894 if (del_timer_sync(&dwork
->timer
))
2895 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2897 return flush_work(&dwork
->work
);
2899 EXPORT_SYMBOL(flush_delayed_work
);
2902 * cancel_delayed_work - cancel a delayed work
2903 * @dwork: delayed_work to cancel
2905 * Kill off a pending delayed_work.
2907 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2911 * The work callback function may still be running on return, unless
2912 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2913 * use cancel_delayed_work_sync() to wait on it.
2915 * This function is safe to call from any context including IRQ handler.
2917 bool cancel_delayed_work(struct delayed_work
*dwork
)
2919 unsigned long flags
;
2923 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2924 } while (unlikely(ret
== -EAGAIN
));
2926 if (unlikely(ret
< 0))
2929 set_work_pool_and_clear_pending(&dwork
->work
,
2930 get_work_pool_id(&dwork
->work
));
2931 local_irq_restore(flags
);
2934 EXPORT_SYMBOL(cancel_delayed_work
);
2937 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2938 * @dwork: the delayed work cancel
2940 * This is cancel_work_sync() for delayed works.
2943 * %true if @dwork was pending, %false otherwise.
2945 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2947 return __cancel_work_timer(&dwork
->work
, true);
2949 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2952 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2953 * @func: the function to call
2955 * schedule_on_each_cpu() executes @func on each online CPU using the
2956 * system workqueue and blocks until all CPUs have completed.
2957 * schedule_on_each_cpu() is very slow.
2960 * 0 on success, -errno on failure.
2962 int schedule_on_each_cpu(work_func_t func
)
2965 struct work_struct __percpu
*works
;
2967 works
= alloc_percpu(struct work_struct
);
2973 for_each_online_cpu(cpu
) {
2974 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2976 INIT_WORK(work
, func
);
2977 schedule_work_on(cpu
, work
);
2980 for_each_online_cpu(cpu
)
2981 flush_work(per_cpu_ptr(works
, cpu
));
2989 * execute_in_process_context - reliably execute the routine with user context
2990 * @fn: the function to execute
2991 * @ew: guaranteed storage for the execute work structure (must
2992 * be available when the work executes)
2994 * Executes the function immediately if process context is available,
2995 * otherwise schedules the function for delayed execution.
2997 * Return: 0 - function was executed
2998 * 1 - function was scheduled for execution
3000 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3002 if (!in_interrupt()) {
3007 INIT_WORK(&ew
->work
, fn
);
3008 schedule_work(&ew
->work
);
3012 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3015 * free_workqueue_attrs - free a workqueue_attrs
3016 * @attrs: workqueue_attrs to free
3018 * Undo alloc_workqueue_attrs().
3020 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3023 free_cpumask_var(attrs
->cpumask
);
3029 * alloc_workqueue_attrs - allocate a workqueue_attrs
3030 * @gfp_mask: allocation mask to use
3032 * Allocate a new workqueue_attrs, initialize with default settings and
3035 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3037 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3039 struct workqueue_attrs
*attrs
;
3041 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3044 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3047 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3050 free_workqueue_attrs(attrs
);
3054 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3055 const struct workqueue_attrs
*from
)
3057 to
->nice
= from
->nice
;
3058 cpumask_copy(to
->cpumask
, from
->cpumask
);
3060 * Unlike hash and equality test, this function doesn't ignore
3061 * ->no_numa as it is used for both pool and wq attrs. Instead,
3062 * get_unbound_pool() explicitly clears ->no_numa after copying.
3064 to
->no_numa
= from
->no_numa
;
3067 /* hash value of the content of @attr */
3068 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3072 hash
= jhash_1word(attrs
->nice
, hash
);
3073 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3074 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3078 /* content equality test */
3079 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3080 const struct workqueue_attrs
*b
)
3082 if (a
->nice
!= b
->nice
)
3084 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3090 * init_worker_pool - initialize a newly zalloc'd worker_pool
3091 * @pool: worker_pool to initialize
3093 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3095 * Return: 0 on success, -errno on failure. Even on failure, all fields
3096 * inside @pool proper are initialized and put_unbound_pool() can be called
3097 * on @pool safely to release it.
3099 static int init_worker_pool(struct worker_pool
*pool
)
3101 spin_lock_init(&pool
->lock
);
3104 pool
->node
= NUMA_NO_NODE
;
3105 pool
->flags
|= POOL_DISASSOCIATED
;
3106 INIT_LIST_HEAD(&pool
->worklist
);
3107 INIT_LIST_HEAD(&pool
->idle_list
);
3108 hash_init(pool
->busy_hash
);
3110 init_timer_deferrable(&pool
->idle_timer
);
3111 pool
->idle_timer
.function
= idle_worker_timeout
;
3112 pool
->idle_timer
.data
= (unsigned long)pool
;
3114 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3115 (unsigned long)pool
);
3117 mutex_init(&pool
->attach_mutex
);
3118 INIT_LIST_HEAD(&pool
->workers
);
3120 ida_init(&pool
->worker_ida
);
3121 INIT_HLIST_NODE(&pool
->hash_node
);
3124 /* shouldn't fail above this point */
3125 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3131 static void rcu_free_wq(struct rcu_head
*rcu
)
3133 struct workqueue_struct
*wq
=
3134 container_of(rcu
, struct workqueue_struct
, rcu
);
3136 if (!(wq
->flags
& WQ_UNBOUND
))
3137 free_percpu(wq
->cpu_pwqs
);
3139 free_workqueue_attrs(wq
->unbound_attrs
);
3145 static void rcu_free_pool(struct rcu_head
*rcu
)
3147 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3149 ida_destroy(&pool
->worker_ida
);
3150 free_workqueue_attrs(pool
->attrs
);
3155 * put_unbound_pool - put a worker_pool
3156 * @pool: worker_pool to put
3158 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3159 * safe manner. get_unbound_pool() calls this function on its failure path
3160 * and this function should be able to release pools which went through,
3161 * successfully or not, init_worker_pool().
3163 * Should be called with wq_pool_mutex held.
3165 static void put_unbound_pool(struct worker_pool
*pool
)
3167 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3168 struct worker
*worker
;
3170 lockdep_assert_held(&wq_pool_mutex
);
3176 if (WARN_ON(!(pool
->cpu
< 0)) ||
3177 WARN_ON(!list_empty(&pool
->worklist
)))
3180 /* release id and unhash */
3182 idr_remove(&worker_pool_idr
, pool
->id
);
3183 hash_del(&pool
->hash_node
);
3186 * Become the manager and destroy all workers. This prevents
3187 * @pool's workers from blocking on attach_mutex. We're the last
3188 * manager and @pool gets freed with the flag set.
3190 spin_lock_irq(&pool
->lock
);
3191 wait_event_lock_irq(wq_manager_wait
,
3192 !(pool
->flags
& POOL_MANAGER_ACTIVE
), pool
->lock
);
3193 pool
->flags
|= POOL_MANAGER_ACTIVE
;
3195 while ((worker
= first_idle_worker(pool
)))
3196 destroy_worker(worker
);
3197 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3198 spin_unlock_irq(&pool
->lock
);
3200 mutex_lock(&pool
->attach_mutex
);
3201 if (!list_empty(&pool
->workers
))
3202 pool
->detach_completion
= &detach_completion
;
3203 mutex_unlock(&pool
->attach_mutex
);
3205 if (pool
->detach_completion
)
3206 wait_for_completion(pool
->detach_completion
);
3208 /* shut down the timers */
3209 del_timer_sync(&pool
->idle_timer
);
3210 del_timer_sync(&pool
->mayday_timer
);
3212 /* sched-RCU protected to allow dereferences from get_work_pool() */
3213 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3217 * get_unbound_pool - get a worker_pool with the specified attributes
3218 * @attrs: the attributes of the worker_pool to get
3220 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3221 * reference count and return it. If there already is a matching
3222 * worker_pool, it will be used; otherwise, this function attempts to
3225 * Should be called with wq_pool_mutex held.
3227 * Return: On success, a worker_pool with the same attributes as @attrs.
3228 * On failure, %NULL.
3230 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3232 u32 hash
= wqattrs_hash(attrs
);
3233 struct worker_pool
*pool
;
3235 int target_node
= NUMA_NO_NODE
;
3237 lockdep_assert_held(&wq_pool_mutex
);
3239 /* do we already have a matching pool? */
3240 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3241 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3247 /* if cpumask is contained inside a NUMA node, we belong to that node */
3248 if (wq_numa_enabled
) {
3249 for_each_node(node
) {
3250 if (cpumask_subset(attrs
->cpumask
,
3251 wq_numa_possible_cpumask
[node
])) {
3258 /* nope, create a new one */
3259 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3260 if (!pool
|| init_worker_pool(pool
) < 0)
3263 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3264 copy_workqueue_attrs(pool
->attrs
, attrs
);
3265 pool
->node
= target_node
;
3268 * no_numa isn't a worker_pool attribute, always clear it. See
3269 * 'struct workqueue_attrs' comments for detail.
3271 pool
->attrs
->no_numa
= false;
3273 if (worker_pool_assign_id(pool
) < 0)
3276 /* create and start the initial worker */
3277 if (!create_worker(pool
))
3281 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3286 put_unbound_pool(pool
);
3290 static void rcu_free_pwq(struct rcu_head
*rcu
)
3292 kmem_cache_free(pwq_cache
,
3293 container_of(rcu
, struct pool_workqueue
, rcu
));
3297 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3298 * and needs to be destroyed.
3300 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3302 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3303 unbound_release_work
);
3304 struct workqueue_struct
*wq
= pwq
->wq
;
3305 struct worker_pool
*pool
= pwq
->pool
;
3308 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3311 mutex_lock(&wq
->mutex
);
3312 list_del_rcu(&pwq
->pwqs_node
);
3313 is_last
= list_empty(&wq
->pwqs
);
3314 mutex_unlock(&wq
->mutex
);
3316 mutex_lock(&wq_pool_mutex
);
3317 put_unbound_pool(pool
);
3318 mutex_unlock(&wq_pool_mutex
);
3320 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3323 * If we're the last pwq going away, @wq is already dead and no one
3324 * is gonna access it anymore. Schedule RCU free.
3327 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
3331 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3332 * @pwq: target pool_workqueue
3334 * If @pwq isn't freezing, set @pwq->max_active to the associated
3335 * workqueue's saved_max_active and activate delayed work items
3336 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3338 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3340 struct workqueue_struct
*wq
= pwq
->wq
;
3341 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3343 /* for @wq->saved_max_active */
3344 lockdep_assert_held(&wq
->mutex
);
3346 /* fast exit for non-freezable wqs */
3347 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3350 spin_lock_irq(&pwq
->pool
->lock
);
3353 * During [un]freezing, the caller is responsible for ensuring that
3354 * this function is called at least once after @workqueue_freezing
3355 * is updated and visible.
3357 if (!freezable
|| !workqueue_freezing
) {
3358 pwq
->max_active
= wq
->saved_max_active
;
3360 while (!list_empty(&pwq
->delayed_works
) &&
3361 pwq
->nr_active
< pwq
->max_active
)
3362 pwq_activate_first_delayed(pwq
);
3365 * Need to kick a worker after thawed or an unbound wq's
3366 * max_active is bumped. It's a slow path. Do it always.
3368 wake_up_worker(pwq
->pool
);
3370 pwq
->max_active
= 0;
3373 spin_unlock_irq(&pwq
->pool
->lock
);
3376 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3377 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3378 struct worker_pool
*pool
)
3380 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3382 memset(pwq
, 0, sizeof(*pwq
));
3386 pwq
->flush_color
= -1;
3388 INIT_LIST_HEAD(&pwq
->delayed_works
);
3389 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3390 INIT_LIST_HEAD(&pwq
->mayday_node
);
3391 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3394 /* sync @pwq with the current state of its associated wq and link it */
3395 static void link_pwq(struct pool_workqueue
*pwq
)
3397 struct workqueue_struct
*wq
= pwq
->wq
;
3399 lockdep_assert_held(&wq
->mutex
);
3401 /* may be called multiple times, ignore if already linked */
3402 if (!list_empty(&pwq
->pwqs_node
))
3405 /* set the matching work_color */
3406 pwq
->work_color
= wq
->work_color
;
3408 /* sync max_active to the current setting */
3409 pwq_adjust_max_active(pwq
);
3412 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3415 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3416 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3417 const struct workqueue_attrs
*attrs
)
3419 struct worker_pool
*pool
;
3420 struct pool_workqueue
*pwq
;
3422 lockdep_assert_held(&wq_pool_mutex
);
3424 pool
= get_unbound_pool(attrs
);
3428 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3430 put_unbound_pool(pool
);
3434 init_pwq(pwq
, wq
, pool
);
3439 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3440 * @attrs: the wq_attrs of the default pwq of the target workqueue
3441 * @node: the target NUMA node
3442 * @cpu_going_down: if >= 0, the CPU to consider as offline
3443 * @cpumask: outarg, the resulting cpumask
3445 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3446 * @cpu_going_down is >= 0, that cpu is considered offline during
3447 * calculation. The result is stored in @cpumask.
3449 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3450 * enabled and @node has online CPUs requested by @attrs, the returned
3451 * cpumask is the intersection of the possible CPUs of @node and
3454 * The caller is responsible for ensuring that the cpumask of @node stays
3457 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3460 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3461 int cpu_going_down
, cpumask_t
*cpumask
)
3463 if (!wq_numa_enabled
|| attrs
->no_numa
)
3466 /* does @node have any online CPUs @attrs wants? */
3467 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3468 if (cpu_going_down
>= 0)
3469 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3471 if (cpumask_empty(cpumask
))
3474 /* yeap, return possible CPUs in @node that @attrs wants */
3475 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3476 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3479 cpumask_copy(cpumask
, attrs
->cpumask
);
3483 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3484 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3486 struct pool_workqueue
*pwq
)
3488 struct pool_workqueue
*old_pwq
;
3490 lockdep_assert_held(&wq_pool_mutex
);
3491 lockdep_assert_held(&wq
->mutex
);
3493 /* link_pwq() can handle duplicate calls */
3496 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3497 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3501 /* context to store the prepared attrs & pwqs before applying */
3502 struct apply_wqattrs_ctx
{
3503 struct workqueue_struct
*wq
; /* target workqueue */
3504 struct workqueue_attrs
*attrs
; /* attrs to apply */
3505 struct list_head list
; /* queued for batching commit */
3506 struct pool_workqueue
*dfl_pwq
;
3507 struct pool_workqueue
*pwq_tbl
[];
3510 /* free the resources after success or abort */
3511 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3517 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3518 put_pwq_unlocked(ctx
->dfl_pwq
);
3520 free_workqueue_attrs(ctx
->attrs
);
3526 /* allocate the attrs and pwqs for later installation */
3527 static struct apply_wqattrs_ctx
*
3528 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3529 const struct workqueue_attrs
*attrs
)
3531 struct apply_wqattrs_ctx
*ctx
;
3532 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3535 lockdep_assert_held(&wq_pool_mutex
);
3537 ctx
= kzalloc(sizeof(*ctx
) + nr_node_ids
* sizeof(ctx
->pwq_tbl
[0]),
3540 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3541 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3542 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3546 * Calculate the attrs of the default pwq.
3547 * If the user configured cpumask doesn't overlap with the
3548 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3550 copy_workqueue_attrs(new_attrs
, attrs
);
3551 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3552 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3553 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3556 * We may create multiple pwqs with differing cpumasks. Make a
3557 * copy of @new_attrs which will be modified and used to obtain
3560 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3563 * If something goes wrong during CPU up/down, we'll fall back to
3564 * the default pwq covering whole @attrs->cpumask. Always create
3565 * it even if we don't use it immediately.
3567 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3571 for_each_node(node
) {
3572 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3573 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3574 if (!ctx
->pwq_tbl
[node
])
3577 ctx
->dfl_pwq
->refcnt
++;
3578 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3582 /* save the user configured attrs and sanitize it. */
3583 copy_workqueue_attrs(new_attrs
, attrs
);
3584 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3585 ctx
->attrs
= new_attrs
;
3588 free_workqueue_attrs(tmp_attrs
);
3592 free_workqueue_attrs(tmp_attrs
);
3593 free_workqueue_attrs(new_attrs
);
3594 apply_wqattrs_cleanup(ctx
);
3598 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3599 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3603 /* all pwqs have been created successfully, let's install'em */
3604 mutex_lock(&ctx
->wq
->mutex
);
3606 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3608 /* save the previous pwq and install the new one */
3610 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
3611 ctx
->pwq_tbl
[node
]);
3613 /* @dfl_pwq might not have been used, ensure it's linked */
3614 link_pwq(ctx
->dfl_pwq
);
3615 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
3617 mutex_unlock(&ctx
->wq
->mutex
);
3620 static void apply_wqattrs_lock(void)
3622 /* CPUs should stay stable across pwq creations and installations */
3624 mutex_lock(&wq_pool_mutex
);
3627 static void apply_wqattrs_unlock(void)
3629 mutex_unlock(&wq_pool_mutex
);
3633 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
3634 const struct workqueue_attrs
*attrs
)
3636 struct apply_wqattrs_ctx
*ctx
;
3639 /* only unbound workqueues can change attributes */
3640 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3643 /* creating multiple pwqs breaks ordering guarantee */
3644 if (!list_empty(&wq
->pwqs
)) {
3645 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
3648 wq
->flags
&= ~__WQ_ORDERED
;
3651 ctx
= apply_wqattrs_prepare(wq
, attrs
);
3653 /* the ctx has been prepared successfully, let's commit it */
3655 apply_wqattrs_commit(ctx
);
3659 apply_wqattrs_cleanup(ctx
);
3665 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3666 * @wq: the target workqueue
3667 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3669 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3670 * machines, this function maps a separate pwq to each NUMA node with
3671 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3672 * NUMA node it was issued on. Older pwqs are released as in-flight work
3673 * items finish. Note that a work item which repeatedly requeues itself
3674 * back-to-back will stay on its current pwq.
3676 * Performs GFP_KERNEL allocations.
3678 * Return: 0 on success and -errno on failure.
3680 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3681 const struct workqueue_attrs
*attrs
)
3685 apply_wqattrs_lock();
3686 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
3687 apply_wqattrs_unlock();
3693 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3694 * @wq: the target workqueue
3695 * @cpu: the CPU coming up or going down
3696 * @online: whether @cpu is coming up or going down
3698 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3699 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3702 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3703 * falls back to @wq->dfl_pwq which may not be optimal but is always
3706 * Note that when the last allowed CPU of a NUMA node goes offline for a
3707 * workqueue with a cpumask spanning multiple nodes, the workers which were
3708 * already executing the work items for the workqueue will lose their CPU
3709 * affinity and may execute on any CPU. This is similar to how per-cpu
3710 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3711 * affinity, it's the user's responsibility to flush the work item from
3714 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3717 int node
= cpu_to_node(cpu
);
3718 int cpu_off
= online
? -1 : cpu
;
3719 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3720 struct workqueue_attrs
*target_attrs
;
3723 lockdep_assert_held(&wq_pool_mutex
);
3725 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
3726 wq
->unbound_attrs
->no_numa
)
3730 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3731 * Let's use a preallocated one. The following buf is protected by
3732 * CPU hotplug exclusion.
3734 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3735 cpumask
= target_attrs
->cpumask
;
3737 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3738 pwq
= unbound_pwq_by_node(wq
, node
);
3741 * Let's determine what needs to be done. If the target cpumask is
3742 * different from the default pwq's, we need to compare it to @pwq's
3743 * and create a new one if they don't match. If the target cpumask
3744 * equals the default pwq's, the default pwq should be used.
3746 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
3747 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3753 /* create a new pwq */
3754 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3756 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3761 /* Install the new pwq. */
3762 mutex_lock(&wq
->mutex
);
3763 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3767 mutex_lock(&wq
->mutex
);
3768 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3769 get_pwq(wq
->dfl_pwq
);
3770 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3771 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
3773 mutex_unlock(&wq
->mutex
);
3774 put_pwq_unlocked(old_pwq
);
3777 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3779 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3782 if (!(wq
->flags
& WQ_UNBOUND
)) {
3783 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3787 for_each_possible_cpu(cpu
) {
3788 struct pool_workqueue
*pwq
=
3789 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3790 struct worker_pool
*cpu_pools
=
3791 per_cpu(cpu_worker_pools
, cpu
);
3793 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
3795 mutex_lock(&wq
->mutex
);
3797 mutex_unlock(&wq
->mutex
);
3800 } else if (wq
->flags
& __WQ_ORDERED
) {
3801 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
3802 /* there should only be single pwq for ordering guarantee */
3803 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
3804 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
3805 "ordering guarantee broken for workqueue %s\n", wq
->name
);
3808 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3812 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3815 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3817 if (max_active
< 1 || max_active
> lim
)
3818 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3819 max_active
, name
, 1, lim
);
3821 return clamp_val(max_active
, 1, lim
);
3824 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3827 struct lock_class_key
*key
,
3828 const char *lock_name
, ...)
3830 size_t tbl_size
= 0;
3832 struct workqueue_struct
*wq
;
3833 struct pool_workqueue
*pwq
;
3836 * Unbound && max_active == 1 used to imply ordered, which is no
3837 * longer the case on NUMA machines due to per-node pools. While
3838 * alloc_ordered_workqueue() is the right way to create an ordered
3839 * workqueue, keep the previous behavior to avoid subtle breakages
3842 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
3843 flags
|= __WQ_ORDERED
;
3845 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3846 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
3847 flags
|= WQ_UNBOUND
;
3849 /* allocate wq and format name */
3850 if (flags
& WQ_UNBOUND
)
3851 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
3853 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
3857 if (flags
& WQ_UNBOUND
) {
3858 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3859 if (!wq
->unbound_attrs
)
3863 va_start(args
, lock_name
);
3864 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
3867 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3868 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3872 wq
->saved_max_active
= max_active
;
3873 mutex_init(&wq
->mutex
);
3874 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3875 INIT_LIST_HEAD(&wq
->pwqs
);
3876 INIT_LIST_HEAD(&wq
->flusher_queue
);
3877 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3878 INIT_LIST_HEAD(&wq
->maydays
);
3880 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3881 INIT_LIST_HEAD(&wq
->list
);
3883 if (alloc_and_link_pwqs(wq
) < 0)
3887 * Workqueues which may be used during memory reclaim should
3888 * have a rescuer to guarantee forward progress.
3890 if (flags
& WQ_MEM_RECLAIM
) {
3891 struct worker
*rescuer
;
3893 rescuer
= alloc_worker(NUMA_NO_NODE
);
3897 rescuer
->rescue_wq
= wq
;
3898 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3900 if (IS_ERR(rescuer
->task
)) {
3905 wq
->rescuer
= rescuer
;
3906 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
3907 wake_up_process(rescuer
->task
);
3910 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
3914 * wq_pool_mutex protects global freeze state and workqueues list.
3915 * Grab it, adjust max_active and add the new @wq to workqueues
3918 mutex_lock(&wq_pool_mutex
);
3920 mutex_lock(&wq
->mutex
);
3921 for_each_pwq(pwq
, wq
)
3922 pwq_adjust_max_active(pwq
);
3923 mutex_unlock(&wq
->mutex
);
3925 list_add_tail_rcu(&wq
->list
, &workqueues
);
3927 mutex_unlock(&wq_pool_mutex
);
3932 free_workqueue_attrs(wq
->unbound_attrs
);
3936 destroy_workqueue(wq
);
3939 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3942 * destroy_workqueue - safely terminate a workqueue
3943 * @wq: target workqueue
3945 * Safely destroy a workqueue. All work currently pending will be done first.
3947 void destroy_workqueue(struct workqueue_struct
*wq
)
3949 struct pool_workqueue
*pwq
;
3952 /* drain it before proceeding with destruction */
3953 drain_workqueue(wq
);
3956 mutex_lock(&wq
->mutex
);
3957 for_each_pwq(pwq
, wq
) {
3960 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
3961 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
3962 mutex_unlock(&wq
->mutex
);
3967 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
3968 WARN_ON(pwq
->nr_active
) ||
3969 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
3970 mutex_unlock(&wq
->mutex
);
3974 mutex_unlock(&wq
->mutex
);
3977 * wq list is used to freeze wq, remove from list after
3978 * flushing is complete in case freeze races us.
3980 mutex_lock(&wq_pool_mutex
);
3981 list_del_rcu(&wq
->list
);
3982 mutex_unlock(&wq_pool_mutex
);
3984 workqueue_sysfs_unregister(wq
);
3987 kthread_stop(wq
->rescuer
->task
);
3989 if (!(wq
->flags
& WQ_UNBOUND
)) {
3991 * The base ref is never dropped on per-cpu pwqs. Directly
3992 * schedule RCU free.
3994 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
3997 * We're the sole accessor of @wq at this point. Directly
3998 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3999 * @wq will be freed when the last pwq is released.
4001 for_each_node(node
) {
4002 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4003 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4004 put_pwq_unlocked(pwq
);
4008 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4009 * put. Don't access it afterwards.
4013 put_pwq_unlocked(pwq
);
4016 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4019 * workqueue_set_max_active - adjust max_active of a workqueue
4020 * @wq: target workqueue
4021 * @max_active: new max_active value.
4023 * Set max_active of @wq to @max_active.
4026 * Don't call from IRQ context.
4028 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4030 struct pool_workqueue
*pwq
;
4032 /* disallow meddling with max_active for ordered workqueues */
4033 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4036 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4038 mutex_lock(&wq
->mutex
);
4040 wq
->flags
&= ~__WQ_ORDERED
;
4041 wq
->saved_max_active
= max_active
;
4043 for_each_pwq(pwq
, wq
)
4044 pwq_adjust_max_active(pwq
);
4046 mutex_unlock(&wq
->mutex
);
4048 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4051 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4053 * Determine whether %current is a workqueue rescuer. Can be used from
4054 * work functions to determine whether it's being run off the rescuer task.
4056 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4058 bool current_is_workqueue_rescuer(void)
4060 struct worker
*worker
= current_wq_worker();
4062 return worker
&& worker
->rescue_wq
;
4066 * workqueue_congested - test whether a workqueue is congested
4067 * @cpu: CPU in question
4068 * @wq: target workqueue
4070 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4071 * no synchronization around this function and the test result is
4072 * unreliable and only useful as advisory hints or for debugging.
4074 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4075 * Note that both per-cpu and unbound workqueues may be associated with
4076 * multiple pool_workqueues which have separate congested states. A
4077 * workqueue being congested on one CPU doesn't mean the workqueue is also
4078 * contested on other CPUs / NUMA nodes.
4081 * %true if congested, %false otherwise.
4083 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4085 struct pool_workqueue
*pwq
;
4088 rcu_read_lock_sched();
4090 if (cpu
== WORK_CPU_UNBOUND
)
4091 cpu
= smp_processor_id();
4093 if (!(wq
->flags
& WQ_UNBOUND
))
4094 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4096 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4098 ret
= !list_empty(&pwq
->delayed_works
);
4099 rcu_read_unlock_sched();
4103 EXPORT_SYMBOL_GPL(workqueue_congested
);
4106 * work_busy - test whether a work is currently pending or running
4107 * @work: the work to be tested
4109 * Test whether @work is currently pending or running. There is no
4110 * synchronization around this function and the test result is
4111 * unreliable and only useful as advisory hints or for debugging.
4114 * OR'd bitmask of WORK_BUSY_* bits.
4116 unsigned int work_busy(struct work_struct
*work
)
4118 struct worker_pool
*pool
;
4119 unsigned long flags
;
4120 unsigned int ret
= 0;
4122 if (work_pending(work
))
4123 ret
|= WORK_BUSY_PENDING
;
4125 local_irq_save(flags
);
4126 pool
= get_work_pool(work
);
4128 spin_lock(&pool
->lock
);
4129 if (find_worker_executing_work(pool
, work
))
4130 ret
|= WORK_BUSY_RUNNING
;
4131 spin_unlock(&pool
->lock
);
4133 local_irq_restore(flags
);
4137 EXPORT_SYMBOL_GPL(work_busy
);
4140 * set_worker_desc - set description for the current work item
4141 * @fmt: printf-style format string
4142 * @...: arguments for the format string
4144 * This function can be called by a running work function to describe what
4145 * the work item is about. If the worker task gets dumped, this
4146 * information will be printed out together to help debugging. The
4147 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4149 void set_worker_desc(const char *fmt
, ...)
4151 struct worker
*worker
= current_wq_worker();
4155 va_start(args
, fmt
);
4156 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4158 worker
->desc_valid
= true;
4163 * print_worker_info - print out worker information and description
4164 * @log_lvl: the log level to use when printing
4165 * @task: target task
4167 * If @task is a worker and currently executing a work item, print out the
4168 * name of the workqueue being serviced and worker description set with
4169 * set_worker_desc() by the currently executing work item.
4171 * This function can be safely called on any task as long as the
4172 * task_struct itself is accessible. While safe, this function isn't
4173 * synchronized and may print out mixups or garbages of limited length.
4175 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4177 work_func_t
*fn
= NULL
;
4178 char name
[WQ_NAME_LEN
] = { };
4179 char desc
[WORKER_DESC_LEN
] = { };
4180 struct pool_workqueue
*pwq
= NULL
;
4181 struct workqueue_struct
*wq
= NULL
;
4182 bool desc_valid
= false;
4183 struct worker
*worker
;
4185 if (!(task
->flags
& PF_WQ_WORKER
))
4189 * This function is called without any synchronization and @task
4190 * could be in any state. Be careful with dereferences.
4192 worker
= probe_kthread_data(task
);
4195 * Carefully copy the associated workqueue's workfn and name. Keep
4196 * the original last '\0' in case the original contains garbage.
4198 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4199 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4200 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4201 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4203 /* copy worker description */
4204 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4206 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4208 if (fn
|| name
[0] || desc
[0]) {
4209 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4211 pr_cont(" (%s)", desc
);
4216 static void pr_cont_pool_info(struct worker_pool
*pool
)
4218 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4219 if (pool
->node
!= NUMA_NO_NODE
)
4220 pr_cont(" node=%d", pool
->node
);
4221 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4224 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4226 if (work
->func
== wq_barrier_func
) {
4227 struct wq_barrier
*barr
;
4229 barr
= container_of(work
, struct wq_barrier
, work
);
4231 pr_cont("%s BAR(%d)", comma
? "," : "",
4232 task_pid_nr(barr
->task
));
4234 pr_cont("%s %pf", comma
? "," : "", work
->func
);
4238 static void show_pwq(struct pool_workqueue
*pwq
)
4240 struct worker_pool
*pool
= pwq
->pool
;
4241 struct work_struct
*work
;
4242 struct worker
*worker
;
4243 bool has_in_flight
= false, has_pending
= false;
4246 pr_info(" pwq %d:", pool
->id
);
4247 pr_cont_pool_info(pool
);
4249 pr_cont(" active=%d/%d%s\n", pwq
->nr_active
, pwq
->max_active
,
4250 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4252 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4253 if (worker
->current_pwq
== pwq
) {
4254 has_in_flight
= true;
4258 if (has_in_flight
) {
4261 pr_info(" in-flight:");
4262 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4263 if (worker
->current_pwq
!= pwq
)
4266 pr_cont("%s %d%s:%pf", comma
? "," : "",
4267 task_pid_nr(worker
->task
),
4268 worker
== pwq
->wq
->rescuer
? "(RESCUER)" : "",
4269 worker
->current_func
);
4270 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4271 pr_cont_work(false, work
);
4277 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4278 if (get_work_pwq(work
) == pwq
) {
4286 pr_info(" pending:");
4287 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4288 if (get_work_pwq(work
) != pwq
)
4291 pr_cont_work(comma
, work
);
4292 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4297 if (!list_empty(&pwq
->delayed_works
)) {
4300 pr_info(" delayed:");
4301 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4302 pr_cont_work(comma
, work
);
4303 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4310 * show_workqueue_state - dump workqueue state
4312 * Called from a sysrq handler and prints out all busy workqueues and
4315 void show_workqueue_state(void)
4317 struct workqueue_struct
*wq
;
4318 struct worker_pool
*pool
;
4319 unsigned long flags
;
4322 rcu_read_lock_sched();
4324 pr_info("Showing busy workqueues and worker pools:\n");
4326 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4327 struct pool_workqueue
*pwq
;
4330 for_each_pwq(pwq
, wq
) {
4331 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4339 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4341 for_each_pwq(pwq
, wq
) {
4342 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4343 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4345 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4349 for_each_pool(pool
, pi
) {
4350 struct worker
*worker
;
4353 spin_lock_irqsave(&pool
->lock
, flags
);
4354 if (pool
->nr_workers
== pool
->nr_idle
)
4357 pr_info("pool %d:", pool
->id
);
4358 pr_cont_pool_info(pool
);
4359 pr_cont(" workers=%d", pool
->nr_workers
);
4361 pr_cont(" manager: %d",
4362 task_pid_nr(pool
->manager
->task
));
4363 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4364 pr_cont(" %s%d", first
? "idle: " : "",
4365 task_pid_nr(worker
->task
));
4370 spin_unlock_irqrestore(&pool
->lock
, flags
);
4373 rcu_read_unlock_sched();
4379 * There are two challenges in supporting CPU hotplug. Firstly, there
4380 * are a lot of assumptions on strong associations among work, pwq and
4381 * pool which make migrating pending and scheduled works very
4382 * difficult to implement without impacting hot paths. Secondly,
4383 * worker pools serve mix of short, long and very long running works making
4384 * blocked draining impractical.
4386 * This is solved by allowing the pools to be disassociated from the CPU
4387 * running as an unbound one and allowing it to be reattached later if the
4388 * cpu comes back online.
4391 static void wq_unbind_fn(struct work_struct
*work
)
4393 int cpu
= smp_processor_id();
4394 struct worker_pool
*pool
;
4395 struct worker
*worker
;
4397 for_each_cpu_worker_pool(pool
, cpu
) {
4398 mutex_lock(&pool
->attach_mutex
);
4399 spin_lock_irq(&pool
->lock
);
4402 * We've blocked all attach/detach operations. Make all workers
4403 * unbound and set DISASSOCIATED. Before this, all workers
4404 * except for the ones which are still executing works from
4405 * before the last CPU down must be on the cpu. After
4406 * this, they may become diasporas.
4408 for_each_pool_worker(worker
, pool
)
4409 worker
->flags
|= WORKER_UNBOUND
;
4411 pool
->flags
|= POOL_DISASSOCIATED
;
4413 spin_unlock_irq(&pool
->lock
);
4414 mutex_unlock(&pool
->attach_mutex
);
4417 * Call schedule() so that we cross rq->lock and thus can
4418 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4419 * This is necessary as scheduler callbacks may be invoked
4425 * Sched callbacks are disabled now. Zap nr_running.
4426 * After this, nr_running stays zero and need_more_worker()
4427 * and keep_working() are always true as long as the
4428 * worklist is not empty. This pool now behaves as an
4429 * unbound (in terms of concurrency management) pool which
4430 * are served by workers tied to the pool.
4432 atomic_set(&pool
->nr_running
, 0);
4435 * With concurrency management just turned off, a busy
4436 * worker blocking could lead to lengthy stalls. Kick off
4437 * unbound chain execution of currently pending work items.
4439 spin_lock_irq(&pool
->lock
);
4440 wake_up_worker(pool
);
4441 spin_unlock_irq(&pool
->lock
);
4446 * rebind_workers - rebind all workers of a pool to the associated CPU
4447 * @pool: pool of interest
4449 * @pool->cpu is coming online. Rebind all workers to the CPU.
4451 static void rebind_workers(struct worker_pool
*pool
)
4453 struct worker
*worker
;
4455 lockdep_assert_held(&pool
->attach_mutex
);
4458 * Restore CPU affinity of all workers. As all idle workers should
4459 * be on the run-queue of the associated CPU before any local
4460 * wake-ups for concurrency management happen, restore CPU affinity
4461 * of all workers first and then clear UNBOUND. As we're called
4462 * from CPU_ONLINE, the following shouldn't fail.
4464 for_each_pool_worker(worker
, pool
)
4465 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4466 pool
->attrs
->cpumask
) < 0);
4468 spin_lock_irq(&pool
->lock
);
4471 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4472 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4473 * being reworked and this can go away in time.
4475 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
4476 spin_unlock_irq(&pool
->lock
);
4480 pool
->flags
&= ~POOL_DISASSOCIATED
;
4482 for_each_pool_worker(worker
, pool
) {
4483 unsigned int worker_flags
= worker
->flags
;
4486 * A bound idle worker should actually be on the runqueue
4487 * of the associated CPU for local wake-ups targeting it to
4488 * work. Kick all idle workers so that they migrate to the
4489 * associated CPU. Doing this in the same loop as
4490 * replacing UNBOUND with REBOUND is safe as no worker will
4491 * be bound before @pool->lock is released.
4493 if (worker_flags
& WORKER_IDLE
)
4494 wake_up_process(worker
->task
);
4497 * We want to clear UNBOUND but can't directly call
4498 * worker_clr_flags() or adjust nr_running. Atomically
4499 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4500 * @worker will clear REBOUND using worker_clr_flags() when
4501 * it initiates the next execution cycle thus restoring
4502 * concurrency management. Note that when or whether
4503 * @worker clears REBOUND doesn't affect correctness.
4505 * ACCESS_ONCE() is necessary because @worker->flags may be
4506 * tested without holding any lock in
4507 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4508 * fail incorrectly leading to premature concurrency
4509 * management operations.
4511 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4512 worker_flags
|= WORKER_REBOUND
;
4513 worker_flags
&= ~WORKER_UNBOUND
;
4514 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4517 spin_unlock_irq(&pool
->lock
);
4521 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4522 * @pool: unbound pool of interest
4523 * @cpu: the CPU which is coming up
4525 * An unbound pool may end up with a cpumask which doesn't have any online
4526 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4527 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4528 * online CPU before, cpus_allowed of all its workers should be restored.
4530 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4532 static cpumask_t cpumask
;
4533 struct worker
*worker
;
4535 lockdep_assert_held(&pool
->attach_mutex
);
4537 /* is @cpu allowed for @pool? */
4538 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4541 /* is @cpu the only online CPU? */
4542 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4543 if (cpumask_weight(&cpumask
) != 1)
4546 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4547 for_each_pool_worker(worker
, pool
)
4548 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4549 pool
->attrs
->cpumask
) < 0);
4553 * Workqueues should be brought up before normal priority CPU notifiers.
4554 * This will be registered high priority CPU notifier.
4556 static int workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4557 unsigned long action
,
4560 int cpu
= (unsigned long)hcpu
;
4561 struct worker_pool
*pool
;
4562 struct workqueue_struct
*wq
;
4565 switch (action
& ~CPU_TASKS_FROZEN
) {
4566 case CPU_UP_PREPARE
:
4567 for_each_cpu_worker_pool(pool
, cpu
) {
4568 if (pool
->nr_workers
)
4570 if (!create_worker(pool
))
4575 case CPU_DOWN_FAILED
:
4577 mutex_lock(&wq_pool_mutex
);
4579 for_each_pool(pool
, pi
) {
4580 mutex_lock(&pool
->attach_mutex
);
4582 if (pool
->cpu
== cpu
)
4583 rebind_workers(pool
);
4584 else if (pool
->cpu
< 0)
4585 restore_unbound_workers_cpumask(pool
, cpu
);
4587 mutex_unlock(&pool
->attach_mutex
);
4590 /* update NUMA affinity of unbound workqueues */
4591 list_for_each_entry(wq
, &workqueues
, list
)
4592 wq_update_unbound_numa(wq
, cpu
, true);
4594 mutex_unlock(&wq_pool_mutex
);
4601 * Workqueues should be brought down after normal priority CPU notifiers.
4602 * This will be registered as low priority CPU notifier.
4604 static int workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4605 unsigned long action
,
4608 int cpu
= (unsigned long)hcpu
;
4609 struct work_struct unbind_work
;
4610 struct workqueue_struct
*wq
;
4612 switch (action
& ~CPU_TASKS_FROZEN
) {
4613 case CPU_DOWN_PREPARE
:
4614 /* unbinding per-cpu workers should happen on the local CPU */
4615 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4616 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4618 /* update NUMA affinity of unbound workqueues */
4619 mutex_lock(&wq_pool_mutex
);
4620 list_for_each_entry(wq
, &workqueues
, list
)
4621 wq_update_unbound_numa(wq
, cpu
, false);
4622 mutex_unlock(&wq_pool_mutex
);
4624 /* wait for per-cpu unbinding to finish */
4625 flush_work(&unbind_work
);
4626 destroy_work_on_stack(&unbind_work
);
4634 struct work_for_cpu
{
4635 struct work_struct work
;
4641 static void work_for_cpu_fn(struct work_struct
*work
)
4643 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4645 wfc
->ret
= wfc
->fn(wfc
->arg
);
4649 * work_on_cpu - run a function in user context on a particular cpu
4650 * @cpu: the cpu to run on
4651 * @fn: the function to run
4652 * @arg: the function arg
4654 * It is up to the caller to ensure that the cpu doesn't go offline.
4655 * The caller must not hold any locks which would prevent @fn from completing.
4657 * Return: The value @fn returns.
4659 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4661 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4663 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4664 schedule_work_on(cpu
, &wfc
.work
);
4665 flush_work(&wfc
.work
);
4666 destroy_work_on_stack(&wfc
.work
);
4669 EXPORT_SYMBOL_GPL(work_on_cpu
);
4670 #endif /* CONFIG_SMP */
4672 #ifdef CONFIG_FREEZER
4675 * freeze_workqueues_begin - begin freezing workqueues
4677 * Start freezing workqueues. After this function returns, all freezable
4678 * workqueues will queue new works to their delayed_works list instead of
4682 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4684 void freeze_workqueues_begin(void)
4686 struct workqueue_struct
*wq
;
4687 struct pool_workqueue
*pwq
;
4689 mutex_lock(&wq_pool_mutex
);
4691 WARN_ON_ONCE(workqueue_freezing
);
4692 workqueue_freezing
= true;
4694 list_for_each_entry(wq
, &workqueues
, list
) {
4695 mutex_lock(&wq
->mutex
);
4696 for_each_pwq(pwq
, wq
)
4697 pwq_adjust_max_active(pwq
);
4698 mutex_unlock(&wq
->mutex
);
4701 mutex_unlock(&wq_pool_mutex
);
4705 * freeze_workqueues_busy - are freezable workqueues still busy?
4707 * Check whether freezing is complete. This function must be called
4708 * between freeze_workqueues_begin() and thaw_workqueues().
4711 * Grabs and releases wq_pool_mutex.
4714 * %true if some freezable workqueues are still busy. %false if freezing
4717 bool freeze_workqueues_busy(void)
4720 struct workqueue_struct
*wq
;
4721 struct pool_workqueue
*pwq
;
4723 mutex_lock(&wq_pool_mutex
);
4725 WARN_ON_ONCE(!workqueue_freezing
);
4727 list_for_each_entry(wq
, &workqueues
, list
) {
4728 if (!(wq
->flags
& WQ_FREEZABLE
))
4731 * nr_active is monotonically decreasing. It's safe
4732 * to peek without lock.
4734 rcu_read_lock_sched();
4735 for_each_pwq(pwq
, wq
) {
4736 WARN_ON_ONCE(pwq
->nr_active
< 0);
4737 if (pwq
->nr_active
) {
4739 rcu_read_unlock_sched();
4743 rcu_read_unlock_sched();
4746 mutex_unlock(&wq_pool_mutex
);
4751 * thaw_workqueues - thaw workqueues
4753 * Thaw workqueues. Normal queueing is restored and all collected
4754 * frozen works are transferred to their respective pool worklists.
4757 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4759 void thaw_workqueues(void)
4761 struct workqueue_struct
*wq
;
4762 struct pool_workqueue
*pwq
;
4764 mutex_lock(&wq_pool_mutex
);
4766 if (!workqueue_freezing
)
4769 workqueue_freezing
= false;
4771 /* restore max_active and repopulate worklist */
4772 list_for_each_entry(wq
, &workqueues
, list
) {
4773 mutex_lock(&wq
->mutex
);
4774 for_each_pwq(pwq
, wq
)
4775 pwq_adjust_max_active(pwq
);
4776 mutex_unlock(&wq
->mutex
);
4780 mutex_unlock(&wq_pool_mutex
);
4782 #endif /* CONFIG_FREEZER */
4784 static int workqueue_apply_unbound_cpumask(void)
4788 struct workqueue_struct
*wq
;
4789 struct apply_wqattrs_ctx
*ctx
, *n
;
4791 lockdep_assert_held(&wq_pool_mutex
);
4793 list_for_each_entry(wq
, &workqueues
, list
) {
4794 if (!(wq
->flags
& WQ_UNBOUND
))
4796 /* creating multiple pwqs breaks ordering guarantee */
4797 if (wq
->flags
& __WQ_ORDERED
)
4800 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
4806 list_add_tail(&ctx
->list
, &ctxs
);
4809 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
4811 apply_wqattrs_commit(ctx
);
4812 apply_wqattrs_cleanup(ctx
);
4819 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4820 * @cpumask: the cpumask to set
4822 * The low-level workqueues cpumask is a global cpumask that limits
4823 * the affinity of all unbound workqueues. This function check the @cpumask
4824 * and apply it to all unbound workqueues and updates all pwqs of them.
4826 * Retun: 0 - Success
4827 * -EINVAL - Invalid @cpumask
4828 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4830 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
4833 cpumask_var_t saved_cpumask
;
4835 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
4838 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
4839 if (!cpumask_empty(cpumask
)) {
4840 apply_wqattrs_lock();
4842 /* save the old wq_unbound_cpumask. */
4843 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
4845 /* update wq_unbound_cpumask at first and apply it to wqs. */
4846 cpumask_copy(wq_unbound_cpumask
, cpumask
);
4847 ret
= workqueue_apply_unbound_cpumask();
4849 /* restore the wq_unbound_cpumask when failed. */
4851 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
4853 apply_wqattrs_unlock();
4856 free_cpumask_var(saved_cpumask
);
4862 * Workqueues with WQ_SYSFS flag set is visible to userland via
4863 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4864 * following attributes.
4866 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4867 * max_active RW int : maximum number of in-flight work items
4869 * Unbound workqueues have the following extra attributes.
4871 * id RO int : the associated pool ID
4872 * nice RW int : nice value of the workers
4873 * cpumask RW mask : bitmask of allowed CPUs for the workers
4876 struct workqueue_struct
*wq
;
4880 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
4882 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
4887 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
4890 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4892 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
4894 static DEVICE_ATTR_RO(per_cpu
);
4896 static ssize_t
max_active_show(struct device
*dev
,
4897 struct device_attribute
*attr
, char *buf
)
4899 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4901 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
4904 static ssize_t
max_active_store(struct device
*dev
,
4905 struct device_attribute
*attr
, const char *buf
,
4908 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4911 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
4914 workqueue_set_max_active(wq
, val
);
4917 static DEVICE_ATTR_RW(max_active
);
4919 static struct attribute
*wq_sysfs_attrs
[] = {
4920 &dev_attr_per_cpu
.attr
,
4921 &dev_attr_max_active
.attr
,
4924 ATTRIBUTE_GROUPS(wq_sysfs
);
4926 static ssize_t
wq_pool_ids_show(struct device
*dev
,
4927 struct device_attribute
*attr
, char *buf
)
4929 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4930 const char *delim
= "";
4931 int node
, written
= 0;
4933 rcu_read_lock_sched();
4934 for_each_node(node
) {
4935 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
4936 "%s%d:%d", delim
, node
,
4937 unbound_pwq_by_node(wq
, node
)->pool
->id
);
4940 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
4941 rcu_read_unlock_sched();
4946 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
4949 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4952 mutex_lock(&wq
->mutex
);
4953 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
4954 mutex_unlock(&wq
->mutex
);
4959 /* prepare workqueue_attrs for sysfs store operations */
4960 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
4962 struct workqueue_attrs
*attrs
;
4964 lockdep_assert_held(&wq_pool_mutex
);
4966 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4970 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
4974 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
4975 const char *buf
, size_t count
)
4977 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4978 struct workqueue_attrs
*attrs
;
4981 apply_wqattrs_lock();
4983 attrs
= wq_sysfs_prep_attrs(wq
);
4987 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
4988 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
4989 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
4994 apply_wqattrs_unlock();
4995 free_workqueue_attrs(attrs
);
4996 return ret
?: count
;
4999 static ssize_t
wq_cpumask_show(struct device
*dev
,
5000 struct device_attribute
*attr
, char *buf
)
5002 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5005 mutex_lock(&wq
->mutex
);
5006 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5007 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5008 mutex_unlock(&wq
->mutex
);
5012 static ssize_t
wq_cpumask_store(struct device
*dev
,
5013 struct device_attribute
*attr
,
5014 const char *buf
, size_t count
)
5016 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5017 struct workqueue_attrs
*attrs
;
5020 apply_wqattrs_lock();
5022 attrs
= wq_sysfs_prep_attrs(wq
);
5026 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5028 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5031 apply_wqattrs_unlock();
5032 free_workqueue_attrs(attrs
);
5033 return ret
?: count
;
5036 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5039 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5042 mutex_lock(&wq
->mutex
);
5043 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5044 !wq
->unbound_attrs
->no_numa
);
5045 mutex_unlock(&wq
->mutex
);
5050 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5051 const char *buf
, size_t count
)
5053 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5054 struct workqueue_attrs
*attrs
;
5055 int v
, ret
= -ENOMEM
;
5057 apply_wqattrs_lock();
5059 attrs
= wq_sysfs_prep_attrs(wq
);
5064 if (sscanf(buf
, "%d", &v
) == 1) {
5065 attrs
->no_numa
= !v
;
5066 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5070 apply_wqattrs_unlock();
5071 free_workqueue_attrs(attrs
);
5072 return ret
?: count
;
5075 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5076 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5077 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5078 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5079 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5083 static struct bus_type wq_subsys
= {
5084 .name
= "workqueue",
5085 .dev_groups
= wq_sysfs_groups
,
5088 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5089 struct device_attribute
*attr
, char *buf
)
5093 mutex_lock(&wq_pool_mutex
);
5094 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5095 cpumask_pr_args(wq_unbound_cpumask
));
5096 mutex_unlock(&wq_pool_mutex
);
5101 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5102 struct device_attribute
*attr
, const char *buf
, size_t count
)
5104 cpumask_var_t cpumask
;
5107 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5110 ret
= cpumask_parse(buf
, cpumask
);
5112 ret
= workqueue_set_unbound_cpumask(cpumask
);
5114 free_cpumask_var(cpumask
);
5115 return ret
? ret
: count
;
5118 static struct device_attribute wq_sysfs_cpumask_attr
=
5119 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5120 wq_unbound_cpumask_store
);
5122 static int __init
wq_sysfs_init(void)
5126 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5130 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5132 core_initcall(wq_sysfs_init
);
5134 static void wq_device_release(struct device
*dev
)
5136 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5142 * workqueue_sysfs_register - make a workqueue visible in sysfs
5143 * @wq: the workqueue to register
5145 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5146 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5147 * which is the preferred method.
5149 * Workqueue user should use this function directly iff it wants to apply
5150 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5151 * apply_workqueue_attrs() may race against userland updating the
5154 * Return: 0 on success, -errno on failure.
5156 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5158 struct wq_device
*wq_dev
;
5162 * Adjusting max_active or creating new pwqs by applying
5163 * attributes breaks ordering guarantee. Disallow exposing ordered
5166 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
5169 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5174 wq_dev
->dev
.bus
= &wq_subsys
;
5175 wq_dev
->dev
.init_name
= wq
->name
;
5176 wq_dev
->dev
.release
= wq_device_release
;
5179 * unbound_attrs are created separately. Suppress uevent until
5180 * everything is ready.
5182 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5184 ret
= device_register(&wq_dev
->dev
);
5191 if (wq
->flags
& WQ_UNBOUND
) {
5192 struct device_attribute
*attr
;
5194 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5195 ret
= device_create_file(&wq_dev
->dev
, attr
);
5197 device_unregister(&wq_dev
->dev
);
5204 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5205 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5210 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5211 * @wq: the workqueue to unregister
5213 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5215 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5217 struct wq_device
*wq_dev
= wq
->wq_dev
;
5223 device_unregister(&wq_dev
->dev
);
5225 #else /* CONFIG_SYSFS */
5226 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5227 #endif /* CONFIG_SYSFS */
5229 static void __init
wq_numa_init(void)
5234 if (num_possible_nodes() <= 1)
5237 if (wq_disable_numa
) {
5238 pr_info("workqueue: NUMA affinity support disabled\n");
5242 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
5243 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5246 * We want masks of possible CPUs of each node which isn't readily
5247 * available. Build one from cpu_to_node() which should have been
5248 * fully initialized by now.
5250 tbl
= kzalloc(nr_node_ids
* sizeof(tbl
[0]), GFP_KERNEL
);
5254 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5255 node_online(node
) ? node
: NUMA_NO_NODE
));
5257 for_each_possible_cpu(cpu
) {
5258 node
= cpu_to_node(cpu
);
5259 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5260 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5261 /* happens iff arch is bonkers, let's just proceed */
5264 cpumask_set_cpu(cpu
, tbl
[node
]);
5267 wq_numa_possible_cpumask
= tbl
;
5268 wq_numa_enabled
= true;
5271 static int __init
init_workqueues(void)
5273 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5276 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5278 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5279 cpumask_copy(wq_unbound_cpumask
, cpu_possible_mask
);
5281 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5283 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
5284 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
5288 /* initialize CPU pools */
5289 for_each_possible_cpu(cpu
) {
5290 struct worker_pool
*pool
;
5293 for_each_cpu_worker_pool(pool
, cpu
) {
5294 BUG_ON(init_worker_pool(pool
));
5296 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5297 pool
->attrs
->nice
= std_nice
[i
++];
5298 pool
->node
= cpu_to_node(cpu
);
5301 mutex_lock(&wq_pool_mutex
);
5302 BUG_ON(worker_pool_assign_id(pool
));
5303 mutex_unlock(&wq_pool_mutex
);
5307 /* create the initial worker */
5308 for_each_online_cpu(cpu
) {
5309 struct worker_pool
*pool
;
5311 for_each_cpu_worker_pool(pool
, cpu
) {
5312 pool
->flags
&= ~POOL_DISASSOCIATED
;
5313 BUG_ON(!create_worker(pool
));
5317 /* create default unbound and ordered wq attrs */
5318 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5319 struct workqueue_attrs
*attrs
;
5321 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5322 attrs
->nice
= std_nice
[i
];
5323 unbound_std_wq_attrs
[i
] = attrs
;
5326 * An ordered wq should have only one pwq as ordering is
5327 * guaranteed by max_active which is enforced by pwqs.
5328 * Turn off NUMA so that dfl_pwq is used for all nodes.
5330 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5331 attrs
->nice
= std_nice
[i
];
5332 attrs
->no_numa
= true;
5333 ordered_wq_attrs
[i
] = attrs
;
5336 system_wq
= alloc_workqueue("events", 0, 0);
5337 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5338 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5339 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5340 WQ_UNBOUND_MAX_ACTIVE
);
5341 system_freezable_wq
= alloc_workqueue("events_freezable",
5343 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5344 WQ_POWER_EFFICIENT
, 0);
5345 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5346 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5348 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5349 !system_unbound_wq
|| !system_freezable_wq
||
5350 !system_power_efficient_wq
||
5351 !system_freezable_power_efficient_wq
);
5354 early_initcall(init_workqueues
);