1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.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>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
53 #include <linux/kvm_para.h>
55 #include "workqueue_internal.h"
61 * A bound pool is either associated or disassociated with its CPU.
62 * While associated (!DISASSOCIATED), all workers are bound to the
63 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 * While DISASSOCIATED, the cpu may be offline and all workers have
67 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 * be executing on any CPU. The pool behaves as an unbound one.
70 * Note that DISASSOCIATED should be flipped only while holding
71 * wq_pool_attach_mutex to avoid changing binding state while
72 * worker_attach_to_pool() is in progress.
74 POOL_MANAGER_ACTIVE
= 1 << 0, /* being managed */
75 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
78 WORKER_DIE
= 1 << 1, /* die die die */
79 WORKER_IDLE
= 1 << 2, /* is idle */
80 WORKER_PREP
= 1 << 3, /* preparing to run works */
81 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
82 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
83 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
85 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
86 WORKER_UNBOUND
| WORKER_REBOUND
,
88 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
90 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
91 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
93 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
94 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
96 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
97 /* call for help after 10ms
99 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
100 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
103 * Rescue workers are used only on emergencies and shared by
104 * all cpus. Give MIN_NICE.
106 RESCUER_NICE_LEVEL
= MIN_NICE
,
107 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
113 * Structure fields follow one of the following exclusion rules.
115 * I: Modifiable by initialization/destruction paths and read-only for
118 * P: Preemption protected. Disabling preemption is enough and should
119 * only be modified and accessed from the local cpu.
121 * L: pool->lock protected. Access with pool->lock held.
123 * X: During normal operation, modification requires pool->lock and should
124 * be done only from local cpu. Either disabling preemption on local
125 * cpu or grabbing pool->lock is enough for read access. If
126 * POOL_DISASSOCIATED is set, it's identical to L.
128 * A: wq_pool_attach_mutex protected.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
134 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
136 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
139 * WQ: wq->mutex protected.
141 * WR: wq->mutex protected for writes. RCU protected for reads.
143 * MD: wq_mayday_lock protected.
146 /* struct worker is defined in workqueue_internal.h */
149 raw_spinlock_t lock
; /* the pool lock */
150 int cpu
; /* I: the associated cpu */
151 int node
; /* I: the associated node ID */
152 int id
; /* I: pool ID */
153 unsigned int flags
; /* X: flags */
155 unsigned long watchdog_ts
; /* L: watchdog timestamp */
157 struct list_head worklist
; /* L: list of pending works */
159 int nr_workers
; /* L: total number of workers */
160 int nr_idle
; /* L: currently idle workers */
162 struct list_head idle_list
; /* X: list of idle workers */
163 struct timer_list idle_timer
; /* L: worker idle timeout */
164 struct timer_list mayday_timer
; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
168 /* L: hash of busy workers */
170 struct worker
*manager
; /* L: purely informational */
171 struct list_head workers
; /* A: attached workers */
172 struct completion
*detach_completion
; /* all workers detached */
174 struct ida worker_ida
; /* worker IDs for task name */
176 struct workqueue_attrs
*attrs
; /* I: worker attributes */
177 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
178 int refcnt
; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp
;
188 * Destruction of pool is RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp
;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue
{
201 struct worker_pool
*pool
; /* I: the associated pool */
202 struct workqueue_struct
*wq
; /* I: the owning workqueue */
203 int work_color
; /* L: current color */
204 int flush_color
; /* L: flushing color */
205 int refcnt
; /* L: reference count */
206 int nr_in_flight
[WORK_NR_COLORS
];
207 /* L: nr of in_flight works */
208 int nr_active
; /* L: nr of active works */
209 int max_active
; /* L: max active works */
210 struct list_head delayed_works
; /* L: delayed works */
211 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
212 struct list_head mayday_node
; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work
;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
225 * Structure used to wait for workqueue flush.
228 struct list_head list
; /* WQ: list of flushers */
229 int flush_color
; /* WQ: flush color waiting for */
230 struct completion done
; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct
{
240 struct list_head pwqs
; /* WR: all pwqs of this wq */
241 struct list_head list
; /* PR: list of all workqueues */
243 struct mutex mutex
; /* protects this wq */
244 int work_color
; /* WQ: current work color */
245 int flush_color
; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush
; /* flush in progress */
247 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
248 struct list_head flusher_queue
; /* WQ: flush waiters */
249 struct list_head flusher_overflow
; /* WQ: flush overflow list */
251 struct list_head maydays
; /* MD: pwqs requesting rescue */
252 struct worker
*rescuer
; /* MD: rescue worker */
254 int nr_drainers
; /* WQ: drain in progress */
255 int saved_max_active
; /* WQ: saved pwq max_active */
257 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
258 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
261 struct wq_device
*wq_dev
; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
265 struct lock_class_key key
;
266 struct lockdep_map lockdep_map
;
268 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
271 * Destruction of workqueue_struct is RCU protected to allow walking
272 * the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
283 static struct kmem_cache
*pwq_cache
;
285 static cpumask_var_t
*wq_numa_possible_cpumask
;
286 /* possible CPUs of each node */
288 static bool wq_disable_numa
;
289 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
291 /* see the comment above the definition of WQ_POWER_EFFICIENT */
292 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
293 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
295 static bool wq_online
; /* can kworkers be created yet? */
297 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
299 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
300 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
302 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
303 static DEFINE_MUTEX(wq_pool_attach_mutex
); /* protects worker attach/detach */
304 static DEFINE_RAW_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
305 /* wait for manager to go away */
306 static struct rcuwait manager_wait
= __RCUWAIT_INITIALIZER(manager_wait
);
308 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
309 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
311 /* PL: allowable cpus for unbound wqs and work items */
312 static cpumask_var_t wq_unbound_cpumask
;
314 /* CPU where unbound work was last round robin scheduled from this CPU */
315 static DEFINE_PER_CPU(int, wq_rr_cpu_last
);
318 * Local execution of unbound work items is no longer guaranteed. The
319 * following always forces round-robin CPU selection on unbound work items
320 * to uncover usages which depend on it.
322 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
323 static bool wq_debug_force_rr_cpu
= true;
325 static bool wq_debug_force_rr_cpu
= false;
327 module_param_named(debug_force_rr_cpu
, wq_debug_force_rr_cpu
, bool, 0644);
329 /* the per-cpu worker pools */
330 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
], cpu_worker_pools
);
332 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
334 /* PL: hash of all unbound pools keyed by pool->attrs */
335 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
337 /* I: attributes used when instantiating standard unbound pools on demand */
338 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
340 /* I: attributes used when instantiating ordered pools on demand */
341 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
343 struct workqueue_struct
*system_wq __read_mostly
;
344 EXPORT_SYMBOL(system_wq
);
345 struct workqueue_struct
*system_highpri_wq __read_mostly
;
346 EXPORT_SYMBOL_GPL(system_highpri_wq
);
347 struct workqueue_struct
*system_long_wq __read_mostly
;
348 EXPORT_SYMBOL_GPL(system_long_wq
);
349 struct workqueue_struct
*system_unbound_wq __read_mostly
;
350 EXPORT_SYMBOL_GPL(system_unbound_wq
);
351 struct workqueue_struct
*system_freezable_wq __read_mostly
;
352 EXPORT_SYMBOL_GPL(system_freezable_wq
);
353 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
354 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
355 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
356 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
358 static int worker_thread(void *__worker
);
359 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
360 static void show_pwq(struct pool_workqueue
*pwq
);
362 #define CREATE_TRACE_POINTS
363 #include <trace/events/workqueue.h>
365 #define assert_rcu_or_pool_mutex() \
366 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
367 !lockdep_is_held(&wq_pool_mutex), \
368 "RCU or wq_pool_mutex should be held")
370 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
371 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
372 !lockdep_is_held(&wq->mutex) && \
373 !lockdep_is_held(&wq_pool_mutex), \
374 "RCU, wq->mutex or wq_pool_mutex should be held")
376 #define for_each_cpu_worker_pool(pool, cpu) \
377 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
378 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
382 * for_each_pool - iterate through all worker_pools in the system
383 * @pool: iteration cursor
384 * @pi: integer used for iteration
386 * This must be called either with wq_pool_mutex held or RCU read
387 * locked. If the pool needs to be used beyond the locking in effect, the
388 * caller is responsible for guaranteeing that the pool stays online.
390 * The if/else clause exists only for the lockdep assertion and can be
393 #define for_each_pool(pool, pi) \
394 idr_for_each_entry(&worker_pool_idr, pool, pi) \
395 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
399 * for_each_pool_worker - iterate through all workers of a worker_pool
400 * @worker: iteration cursor
401 * @pool: worker_pool to iterate workers of
403 * This must be called with wq_pool_attach_mutex.
405 * The if/else clause exists only for the lockdep assertion and can be
408 #define for_each_pool_worker(worker, pool) \
409 list_for_each_entry((worker), &(pool)->workers, node) \
410 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
414 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
415 * @pwq: iteration cursor
416 * @wq: the target workqueue
418 * This must be called either with wq->mutex held or RCU read locked.
419 * If the pwq needs to be used beyond the locking in effect, the caller is
420 * responsible for guaranteeing that the pwq stays online.
422 * The if/else clause exists only for the lockdep assertion and can be
425 #define for_each_pwq(pwq, wq) \
426 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
427 lockdep_is_held(&(wq->mutex)))
429 #ifdef CONFIG_DEBUG_OBJECTS_WORK
431 static const struct debug_obj_descr work_debug_descr
;
433 static void *work_debug_hint(void *addr
)
435 return ((struct work_struct
*) addr
)->func
;
438 static bool work_is_static_object(void *addr
)
440 struct work_struct
*work
= addr
;
442 return test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
));
446 * fixup_init is called when:
447 * - an active object is initialized
449 static bool work_fixup_init(void *addr
, enum debug_obj_state state
)
451 struct work_struct
*work
= addr
;
454 case ODEBUG_STATE_ACTIVE
:
455 cancel_work_sync(work
);
456 debug_object_init(work
, &work_debug_descr
);
464 * fixup_free is called when:
465 * - an active object is freed
467 static bool work_fixup_free(void *addr
, enum debug_obj_state state
)
469 struct work_struct
*work
= addr
;
472 case ODEBUG_STATE_ACTIVE
:
473 cancel_work_sync(work
);
474 debug_object_free(work
, &work_debug_descr
);
481 static const struct debug_obj_descr work_debug_descr
= {
482 .name
= "work_struct",
483 .debug_hint
= work_debug_hint
,
484 .is_static_object
= work_is_static_object
,
485 .fixup_init
= work_fixup_init
,
486 .fixup_free
= work_fixup_free
,
489 static inline void debug_work_activate(struct work_struct
*work
)
491 debug_object_activate(work
, &work_debug_descr
);
494 static inline void debug_work_deactivate(struct work_struct
*work
)
496 debug_object_deactivate(work
, &work_debug_descr
);
499 void __init_work(struct work_struct
*work
, int onstack
)
502 debug_object_init_on_stack(work
, &work_debug_descr
);
504 debug_object_init(work
, &work_debug_descr
);
506 EXPORT_SYMBOL_GPL(__init_work
);
508 void destroy_work_on_stack(struct work_struct
*work
)
510 debug_object_free(work
, &work_debug_descr
);
512 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
514 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
516 destroy_timer_on_stack(&work
->timer
);
517 debug_object_free(&work
->work
, &work_debug_descr
);
519 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
522 static inline void debug_work_activate(struct work_struct
*work
) { }
523 static inline void debug_work_deactivate(struct work_struct
*work
) { }
527 * worker_pool_assign_id - allocate ID and assing it to @pool
528 * @pool: the pool pointer of interest
530 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
531 * successfully, -errno on failure.
533 static int worker_pool_assign_id(struct worker_pool
*pool
)
537 lockdep_assert_held(&wq_pool_mutex
);
539 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
549 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
550 * @wq: the target workqueue
553 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
555 * If the pwq needs to be used beyond the locking in effect, the caller is
556 * responsible for guaranteeing that the pwq stays online.
558 * Return: The unbound pool_workqueue for @node.
560 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
563 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
566 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
567 * delayed item is pending. The plan is to keep CPU -> NODE
568 * mapping valid and stable across CPU on/offlines. Once that
569 * happens, this workaround can be removed.
571 if (unlikely(node
== NUMA_NO_NODE
))
574 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
577 static unsigned int work_color_to_flags(int color
)
579 return color
<< WORK_STRUCT_COLOR_SHIFT
;
582 static int get_work_color(struct work_struct
*work
)
584 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
585 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
588 static int work_next_color(int color
)
590 return (color
+ 1) % WORK_NR_COLORS
;
594 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
595 * contain the pointer to the queued pwq. Once execution starts, the flag
596 * is cleared and the high bits contain OFFQ flags and pool ID.
598 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
599 * and clear_work_data() can be used to set the pwq, pool or clear
600 * work->data. These functions should only be called while the work is
601 * owned - ie. while the PENDING bit is set.
603 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
604 * corresponding to a work. Pool is available once the work has been
605 * queued anywhere after initialization until it is sync canceled. pwq is
606 * available only while the work item is queued.
608 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
609 * canceled. While being canceled, a work item may have its PENDING set
610 * but stay off timer and worklist for arbitrarily long and nobody should
611 * try to steal the PENDING bit.
613 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
616 WARN_ON_ONCE(!work_pending(work
));
617 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
620 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
621 unsigned long extra_flags
)
623 set_work_data(work
, (unsigned long)pwq
,
624 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
627 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
630 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
631 WORK_STRUCT_PENDING
);
634 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
638 * The following wmb is paired with the implied mb in
639 * test_and_set_bit(PENDING) and ensures all updates to @work made
640 * here are visible to and precede any updates by the next PENDING
644 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
646 * The following mb guarantees that previous clear of a PENDING bit
647 * will not be reordered with any speculative LOADS or STORES from
648 * work->current_func, which is executed afterwards. This possible
649 * reordering can lead to a missed execution on attempt to queue
650 * the same @work. E.g. consider this case:
653 * ---------------------------- --------------------------------
655 * 1 STORE event_indicated
656 * 2 queue_work_on() {
657 * 3 test_and_set_bit(PENDING)
658 * 4 } set_..._and_clear_pending() {
659 * 5 set_work_data() # clear bit
661 * 7 work->current_func() {
662 * 8 LOAD event_indicated
665 * Without an explicit full barrier speculative LOAD on line 8 can
666 * be executed before CPU#0 does STORE on line 1. If that happens,
667 * CPU#0 observes the PENDING bit is still set and new execution of
668 * a @work is not queued in a hope, that CPU#1 will eventually
669 * finish the queued @work. Meanwhile CPU#1 does not see
670 * event_indicated is set, because speculative LOAD was executed
671 * before actual STORE.
676 static void clear_work_data(struct work_struct
*work
)
678 smp_wmb(); /* see set_work_pool_and_clear_pending() */
679 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
682 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
684 unsigned long data
= atomic_long_read(&work
->data
);
686 if (data
& WORK_STRUCT_PWQ
)
687 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
693 * get_work_pool - return the worker_pool a given work was associated with
694 * @work: the work item of interest
696 * Pools are created and destroyed under wq_pool_mutex, and allows read
697 * access under RCU read lock. As such, this function should be
698 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
700 * All fields of the returned pool are accessible as long as the above
701 * mentioned locking is in effect. If the returned pool needs to be used
702 * beyond the critical section, the caller is responsible for ensuring the
703 * returned pool is and stays online.
705 * Return: The worker_pool @work was last associated with. %NULL if none.
707 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
709 unsigned long data
= atomic_long_read(&work
->data
);
712 assert_rcu_or_pool_mutex();
714 if (data
& WORK_STRUCT_PWQ
)
715 return ((struct pool_workqueue
*)
716 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
718 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
719 if (pool_id
== WORK_OFFQ_POOL_NONE
)
722 return idr_find(&worker_pool_idr
, pool_id
);
726 * get_work_pool_id - return the worker pool ID a given work is associated with
727 * @work: the work item of interest
729 * Return: The worker_pool ID @work was last associated with.
730 * %WORK_OFFQ_POOL_NONE if none.
732 static int get_work_pool_id(struct work_struct
*work
)
734 unsigned long data
= atomic_long_read(&work
->data
);
736 if (data
& WORK_STRUCT_PWQ
)
737 return ((struct pool_workqueue
*)
738 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
740 return data
>> WORK_OFFQ_POOL_SHIFT
;
743 static void mark_work_canceling(struct work_struct
*work
)
745 unsigned long pool_id
= get_work_pool_id(work
);
747 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
748 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
751 static bool work_is_canceling(struct work_struct
*work
)
753 unsigned long data
= atomic_long_read(&work
->data
);
755 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
759 * Policy functions. These define the policies on how the global worker
760 * pools are managed. Unless noted otherwise, these functions assume that
761 * they're being called with pool->lock held.
764 static bool __need_more_worker(struct worker_pool
*pool
)
766 return !atomic_read(&pool
->nr_running
);
770 * Need to wake up a worker? Called from anything but currently
773 * Note that, because unbound workers never contribute to nr_running, this
774 * function will always return %true for unbound pools as long as the
775 * worklist isn't empty.
777 static bool need_more_worker(struct worker_pool
*pool
)
779 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
782 /* Can I start working? Called from busy but !running workers. */
783 static bool may_start_working(struct worker_pool
*pool
)
785 return pool
->nr_idle
;
788 /* Do I need to keep working? Called from currently running workers. */
789 static bool keep_working(struct worker_pool
*pool
)
791 return !list_empty(&pool
->worklist
) &&
792 atomic_read(&pool
->nr_running
) <= 1;
795 /* Do we need a new worker? Called from manager. */
796 static bool need_to_create_worker(struct worker_pool
*pool
)
798 return need_more_worker(pool
) && !may_start_working(pool
);
801 /* Do we have too many workers and should some go away? */
802 static bool too_many_workers(struct worker_pool
*pool
)
804 bool managing
= pool
->flags
& POOL_MANAGER_ACTIVE
;
805 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
806 int nr_busy
= pool
->nr_workers
- nr_idle
;
808 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
815 /* Return the first idle worker. Safe with preemption disabled */
816 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
818 if (unlikely(list_empty(&pool
->idle_list
)))
821 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
825 * wake_up_worker - wake up an idle worker
826 * @pool: worker pool to wake worker from
828 * Wake up the first idle worker of @pool.
831 * raw_spin_lock_irq(pool->lock).
833 static void wake_up_worker(struct worker_pool
*pool
)
835 struct worker
*worker
= first_idle_worker(pool
);
838 wake_up_process(worker
->task
);
842 * wq_worker_running - a worker is running again
843 * @task: task waking up
845 * This function is called when a worker returns from schedule()
847 void wq_worker_running(struct task_struct
*task
)
849 struct worker
*worker
= kthread_data(task
);
851 if (!worker
->sleeping
)
853 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
854 atomic_inc(&worker
->pool
->nr_running
);
855 worker
->sleeping
= 0;
859 * wq_worker_sleeping - a worker is going to sleep
860 * @task: task going to sleep
862 * This function is called from schedule() when a busy worker is
863 * going to sleep. Preemption needs to be disabled to protect ->sleeping
866 void wq_worker_sleeping(struct task_struct
*task
)
868 struct worker
*next
, *worker
= kthread_data(task
);
869 struct worker_pool
*pool
;
872 * Rescuers, which may not have all the fields set up like normal
873 * workers, also reach here, let's not access anything before
874 * checking NOT_RUNNING.
876 if (worker
->flags
& WORKER_NOT_RUNNING
)
881 /* Return if preempted before wq_worker_running() was reached */
882 if (worker
->sleeping
)
885 worker
->sleeping
= 1;
886 raw_spin_lock_irq(&pool
->lock
);
889 * The counterpart of the following dec_and_test, implied mb,
890 * worklist not empty test sequence is in insert_work().
891 * Please read comment there.
893 * NOT_RUNNING is clear. This means that we're bound to and
894 * running on the local cpu w/ rq lock held and preemption
895 * disabled, which in turn means that none else could be
896 * manipulating idle_list, so dereferencing idle_list without pool
899 if (atomic_dec_and_test(&pool
->nr_running
) &&
900 !list_empty(&pool
->worklist
)) {
901 next
= first_idle_worker(pool
);
903 wake_up_process(next
->task
);
905 raw_spin_unlock_irq(&pool
->lock
);
909 * wq_worker_last_func - retrieve worker's last work function
910 * @task: Task to retrieve last work function of.
912 * Determine the last function a worker executed. This is called from
913 * the scheduler to get a worker's last known identity.
916 * raw_spin_lock_irq(rq->lock)
918 * This function is called during schedule() when a kworker is going
919 * to sleep. It's used by psi to identify aggregation workers during
920 * dequeuing, to allow periodic aggregation to shut-off when that
921 * worker is the last task in the system or cgroup to go to sleep.
923 * As this function doesn't involve any workqueue-related locking, it
924 * only returns stable values when called from inside the scheduler's
925 * queuing and dequeuing paths, when @task, which must be a kworker,
926 * is guaranteed to not be processing any works.
929 * The last work function %current executed as a worker, NULL if it
930 * hasn't executed any work yet.
932 work_func_t
wq_worker_last_func(struct task_struct
*task
)
934 struct worker
*worker
= kthread_data(task
);
936 return worker
->last_func
;
940 * worker_set_flags - set worker flags and adjust nr_running accordingly
942 * @flags: flags to set
944 * Set @flags in @worker->flags and adjust nr_running accordingly.
947 * raw_spin_lock_irq(pool->lock)
949 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
951 struct worker_pool
*pool
= worker
->pool
;
953 WARN_ON_ONCE(worker
->task
!= current
);
955 /* If transitioning into NOT_RUNNING, adjust nr_running. */
956 if ((flags
& WORKER_NOT_RUNNING
) &&
957 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
958 atomic_dec(&pool
->nr_running
);
961 worker
->flags
|= flags
;
965 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
967 * @flags: flags to clear
969 * Clear @flags in @worker->flags and adjust nr_running accordingly.
972 * raw_spin_lock_irq(pool->lock)
974 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
976 struct worker_pool
*pool
= worker
->pool
;
977 unsigned int oflags
= worker
->flags
;
979 WARN_ON_ONCE(worker
->task
!= current
);
981 worker
->flags
&= ~flags
;
984 * If transitioning out of NOT_RUNNING, increment nr_running. Note
985 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
986 * of multiple flags, not a single flag.
988 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
989 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
990 atomic_inc(&pool
->nr_running
);
994 * find_worker_executing_work - find worker which is executing a work
995 * @pool: pool of interest
996 * @work: work to find worker for
998 * Find a worker which is executing @work on @pool by searching
999 * @pool->busy_hash which is keyed by the address of @work. For a worker
1000 * to match, its current execution should match the address of @work and
1001 * its work function. This is to avoid unwanted dependency between
1002 * unrelated work executions through a work item being recycled while still
1005 * This is a bit tricky. A work item may be freed once its execution
1006 * starts and nothing prevents the freed area from being recycled for
1007 * another work item. If the same work item address ends up being reused
1008 * before the original execution finishes, workqueue will identify the
1009 * recycled work item as currently executing and make it wait until the
1010 * current execution finishes, introducing an unwanted dependency.
1012 * This function checks the work item address and work function to avoid
1013 * false positives. Note that this isn't complete as one may construct a
1014 * work function which can introduce dependency onto itself through a
1015 * recycled work item. Well, if somebody wants to shoot oneself in the
1016 * foot that badly, there's only so much we can do, and if such deadlock
1017 * actually occurs, it should be easy to locate the culprit work function.
1020 * raw_spin_lock_irq(pool->lock).
1023 * Pointer to worker which is executing @work if found, %NULL
1026 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
1027 struct work_struct
*work
)
1029 struct worker
*worker
;
1031 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1032 (unsigned long)work
)
1033 if (worker
->current_work
== work
&&
1034 worker
->current_func
== work
->func
)
1041 * move_linked_works - move linked works to a list
1042 * @work: start of series of works to be scheduled
1043 * @head: target list to append @work to
1044 * @nextp: out parameter for nested worklist walking
1046 * Schedule linked works starting from @work to @head. Work series to
1047 * be scheduled starts at @work and includes any consecutive work with
1048 * WORK_STRUCT_LINKED set in its predecessor.
1050 * If @nextp is not NULL, it's updated to point to the next work of
1051 * the last scheduled work. This allows move_linked_works() to be
1052 * nested inside outer list_for_each_entry_safe().
1055 * raw_spin_lock_irq(pool->lock).
1057 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1058 struct work_struct
**nextp
)
1060 struct work_struct
*n
;
1063 * Linked worklist will always end before the end of the list,
1064 * use NULL for list head.
1066 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1067 list_move_tail(&work
->entry
, head
);
1068 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1073 * If we're already inside safe list traversal and have moved
1074 * multiple works to the scheduled queue, the next position
1075 * needs to be updated.
1082 * get_pwq - get an extra reference on the specified pool_workqueue
1083 * @pwq: pool_workqueue to get
1085 * Obtain an extra reference on @pwq. The caller should guarantee that
1086 * @pwq has positive refcnt and be holding the matching pool->lock.
1088 static void get_pwq(struct pool_workqueue
*pwq
)
1090 lockdep_assert_held(&pwq
->pool
->lock
);
1091 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1096 * put_pwq - put a pool_workqueue reference
1097 * @pwq: pool_workqueue to put
1099 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1100 * destruction. The caller should be holding the matching pool->lock.
1102 static void put_pwq(struct pool_workqueue
*pwq
)
1104 lockdep_assert_held(&pwq
->pool
->lock
);
1105 if (likely(--pwq
->refcnt
))
1107 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1110 * @pwq can't be released under pool->lock, bounce to
1111 * pwq_unbound_release_workfn(). This never recurses on the same
1112 * pool->lock as this path is taken only for unbound workqueues and
1113 * the release work item is scheduled on a per-cpu workqueue. To
1114 * avoid lockdep warning, unbound pool->locks are given lockdep
1115 * subclass of 1 in get_unbound_pool().
1117 schedule_work(&pwq
->unbound_release_work
);
1121 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1122 * @pwq: pool_workqueue to put (can be %NULL)
1124 * put_pwq() with locking. This function also allows %NULL @pwq.
1126 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1130 * As both pwqs and pools are RCU protected, the
1131 * following lock operations are safe.
1133 raw_spin_lock_irq(&pwq
->pool
->lock
);
1135 raw_spin_unlock_irq(&pwq
->pool
->lock
);
1139 static void pwq_activate_delayed_work(struct work_struct
*work
)
1141 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1143 trace_workqueue_activate_work(work
);
1144 if (list_empty(&pwq
->pool
->worklist
))
1145 pwq
->pool
->watchdog_ts
= jiffies
;
1146 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1147 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1151 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1153 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1154 struct work_struct
, entry
);
1156 pwq_activate_delayed_work(work
);
1160 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1161 * @pwq: pwq of interest
1162 * @color: color of work which left the queue
1164 * A work either has completed or is removed from pending queue,
1165 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1168 * raw_spin_lock_irq(pool->lock).
1170 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1172 /* uncolored work items don't participate in flushing or nr_active */
1173 if (color
== WORK_NO_COLOR
)
1176 pwq
->nr_in_flight
[color
]--;
1179 if (!list_empty(&pwq
->delayed_works
)) {
1180 /* one down, submit a delayed one */
1181 if (pwq
->nr_active
< pwq
->max_active
)
1182 pwq_activate_first_delayed(pwq
);
1185 /* is flush in progress and are we at the flushing tip? */
1186 if (likely(pwq
->flush_color
!= color
))
1189 /* are there still in-flight works? */
1190 if (pwq
->nr_in_flight
[color
])
1193 /* this pwq is done, clear flush_color */
1194 pwq
->flush_color
= -1;
1197 * If this was the last pwq, wake up the first flusher. It
1198 * will handle the rest.
1200 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1201 complete(&pwq
->wq
->first_flusher
->done
);
1207 * try_to_grab_pending - steal work item from worklist and disable irq
1208 * @work: work item to steal
1209 * @is_dwork: @work is a delayed_work
1210 * @flags: place to store irq state
1212 * Try to grab PENDING bit of @work. This function can handle @work in any
1213 * stable state - idle, on timer or on worklist.
1217 * ======== ================================================================
1218 * 1 if @work was pending and we successfully stole PENDING
1219 * 0 if @work was idle and we claimed PENDING
1220 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1221 * -ENOENT if someone else is canceling @work, this state may persist
1222 * for arbitrarily long
1223 * ======== ================================================================
1226 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1227 * interrupted while holding PENDING and @work off queue, irq must be
1228 * disabled on entry. This, combined with delayed_work->timer being
1229 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1231 * On successful return, >= 0, irq is disabled and the caller is
1232 * responsible for releasing it using local_irq_restore(*@flags).
1234 * This function is safe to call from any context including IRQ handler.
1236 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1237 unsigned long *flags
)
1239 struct worker_pool
*pool
;
1240 struct pool_workqueue
*pwq
;
1242 local_irq_save(*flags
);
1244 /* try to steal the timer if it exists */
1246 struct delayed_work
*dwork
= to_delayed_work(work
);
1249 * dwork->timer is irqsafe. If del_timer() fails, it's
1250 * guaranteed that the timer is not queued anywhere and not
1251 * running on the local CPU.
1253 if (likely(del_timer(&dwork
->timer
)))
1257 /* try to claim PENDING the normal way */
1258 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1263 * The queueing is in progress, or it is already queued. Try to
1264 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1266 pool
= get_work_pool(work
);
1270 raw_spin_lock(&pool
->lock
);
1272 * work->data is guaranteed to point to pwq only while the work
1273 * item is queued on pwq->wq, and both updating work->data to point
1274 * to pwq on queueing and to pool on dequeueing are done under
1275 * pwq->pool->lock. This in turn guarantees that, if work->data
1276 * points to pwq which is associated with a locked pool, the work
1277 * item is currently queued on that pool.
1279 pwq
= get_work_pwq(work
);
1280 if (pwq
&& pwq
->pool
== pool
) {
1281 debug_work_deactivate(work
);
1284 * A delayed work item cannot be grabbed directly because
1285 * it might have linked NO_COLOR work items which, if left
1286 * on the delayed_list, will confuse pwq->nr_active
1287 * management later on and cause stall. Make sure the work
1288 * item is activated before grabbing.
1290 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1291 pwq_activate_delayed_work(work
);
1293 list_del_init(&work
->entry
);
1294 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1296 /* work->data points to pwq iff queued, point to pool */
1297 set_work_pool_and_keep_pending(work
, pool
->id
);
1299 raw_spin_unlock(&pool
->lock
);
1303 raw_spin_unlock(&pool
->lock
);
1306 local_irq_restore(*flags
);
1307 if (work_is_canceling(work
))
1314 * insert_work - insert a work into a pool
1315 * @pwq: pwq @work belongs to
1316 * @work: work to insert
1317 * @head: insertion point
1318 * @extra_flags: extra WORK_STRUCT_* flags to set
1320 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1321 * work_struct flags.
1324 * raw_spin_lock_irq(pool->lock).
1326 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1327 struct list_head
*head
, unsigned int extra_flags
)
1329 struct worker_pool
*pool
= pwq
->pool
;
1331 /* record the work call stack in order to print it in KASAN reports */
1332 kasan_record_aux_stack(work
);
1334 /* we own @work, set data and link */
1335 set_work_pwq(work
, pwq
, extra_flags
);
1336 list_add_tail(&work
->entry
, head
);
1340 * Ensure either wq_worker_sleeping() sees the above
1341 * list_add_tail() or we see zero nr_running to avoid workers lying
1342 * around lazily while there are works to be processed.
1346 if (__need_more_worker(pool
))
1347 wake_up_worker(pool
);
1351 * Test whether @work is being queued from another work executing on the
1354 static bool is_chained_work(struct workqueue_struct
*wq
)
1356 struct worker
*worker
;
1358 worker
= current_wq_worker();
1360 * Return %true iff I'm a worker executing a work item on @wq. If
1361 * I'm @worker, it's safe to dereference it without locking.
1363 return worker
&& worker
->current_pwq
->wq
== wq
;
1367 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1368 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1369 * avoid perturbing sensitive tasks.
1371 static int wq_select_unbound_cpu(int cpu
)
1373 static bool printed_dbg_warning
;
1376 if (likely(!wq_debug_force_rr_cpu
)) {
1377 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1379 } else if (!printed_dbg_warning
) {
1380 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1381 printed_dbg_warning
= true;
1384 if (cpumask_empty(wq_unbound_cpumask
))
1387 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1388 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1389 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1390 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1391 if (unlikely(new_cpu
>= nr_cpu_ids
))
1394 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1399 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1400 struct work_struct
*work
)
1402 struct pool_workqueue
*pwq
;
1403 struct worker_pool
*last_pool
;
1404 struct list_head
*worklist
;
1405 unsigned int work_flags
;
1406 unsigned int req_cpu
= cpu
;
1409 * While a work item is PENDING && off queue, a task trying to
1410 * steal the PENDING will busy-loop waiting for it to either get
1411 * queued or lose PENDING. Grabbing PENDING and queueing should
1412 * happen with IRQ disabled.
1414 lockdep_assert_irqs_disabled();
1417 /* if draining, only works from the same workqueue are allowed */
1418 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1419 WARN_ON_ONCE(!is_chained_work(wq
)))
1423 /* pwq which will be used unless @work is executing elsewhere */
1424 if (wq
->flags
& WQ_UNBOUND
) {
1425 if (req_cpu
== WORK_CPU_UNBOUND
)
1426 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1427 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1429 if (req_cpu
== WORK_CPU_UNBOUND
)
1430 cpu
= raw_smp_processor_id();
1431 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1435 * If @work was previously on a different pool, it might still be
1436 * running there, in which case the work needs to be queued on that
1437 * pool to guarantee non-reentrancy.
1439 last_pool
= get_work_pool(work
);
1440 if (last_pool
&& last_pool
!= pwq
->pool
) {
1441 struct worker
*worker
;
1443 raw_spin_lock(&last_pool
->lock
);
1445 worker
= find_worker_executing_work(last_pool
, work
);
1447 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1448 pwq
= worker
->current_pwq
;
1450 /* meh... not running there, queue here */
1451 raw_spin_unlock(&last_pool
->lock
);
1452 raw_spin_lock(&pwq
->pool
->lock
);
1455 raw_spin_lock(&pwq
->pool
->lock
);
1459 * pwq is determined and locked. For unbound pools, we could have
1460 * raced with pwq release and it could already be dead. If its
1461 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1462 * without another pwq replacing it in the numa_pwq_tbl or while
1463 * work items are executing on it, so the retrying is guaranteed to
1464 * make forward-progress.
1466 if (unlikely(!pwq
->refcnt
)) {
1467 if (wq
->flags
& WQ_UNBOUND
) {
1468 raw_spin_unlock(&pwq
->pool
->lock
);
1473 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1477 /* pwq determined, queue */
1478 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1480 if (WARN_ON(!list_empty(&work
->entry
)))
1483 pwq
->nr_in_flight
[pwq
->work_color
]++;
1484 work_flags
= work_color_to_flags(pwq
->work_color
);
1486 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1487 trace_workqueue_activate_work(work
);
1489 worklist
= &pwq
->pool
->worklist
;
1490 if (list_empty(worklist
))
1491 pwq
->pool
->watchdog_ts
= jiffies
;
1493 work_flags
|= WORK_STRUCT_DELAYED
;
1494 worklist
= &pwq
->delayed_works
;
1497 debug_work_activate(work
);
1498 insert_work(pwq
, work
, worklist
, work_flags
);
1501 raw_spin_unlock(&pwq
->pool
->lock
);
1506 * queue_work_on - queue work on specific cpu
1507 * @cpu: CPU number to execute work on
1508 * @wq: workqueue to use
1509 * @work: work to queue
1511 * We queue the work to a specific CPU, the caller must ensure it
1514 * Return: %false if @work was already on a queue, %true otherwise.
1516 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1517 struct work_struct
*work
)
1520 unsigned long flags
;
1522 local_irq_save(flags
);
1524 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1525 __queue_work(cpu
, wq
, work
);
1529 local_irq_restore(flags
);
1532 EXPORT_SYMBOL(queue_work_on
);
1535 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1536 * @node: NUMA node ID that we want to select a CPU from
1538 * This function will attempt to find a "random" cpu available on a given
1539 * node. If there are no CPUs available on the given node it will return
1540 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1541 * available CPU if we need to schedule this work.
1543 static int workqueue_select_cpu_near(int node
)
1547 /* No point in doing this if NUMA isn't enabled for workqueues */
1548 if (!wq_numa_enabled
)
1549 return WORK_CPU_UNBOUND
;
1551 /* Delay binding to CPU if node is not valid or online */
1552 if (node
< 0 || node
>= MAX_NUMNODES
|| !node_online(node
))
1553 return WORK_CPU_UNBOUND
;
1555 /* Use local node/cpu if we are already there */
1556 cpu
= raw_smp_processor_id();
1557 if (node
== cpu_to_node(cpu
))
1560 /* Use "random" otherwise know as "first" online CPU of node */
1561 cpu
= cpumask_any_and(cpumask_of_node(node
), cpu_online_mask
);
1563 /* If CPU is valid return that, otherwise just defer */
1564 return cpu
< nr_cpu_ids
? cpu
: WORK_CPU_UNBOUND
;
1568 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1569 * @node: NUMA node that we are targeting the work for
1570 * @wq: workqueue to use
1571 * @work: work to queue
1573 * We queue the work to a "random" CPU within a given NUMA node. The basic
1574 * idea here is to provide a way to somehow associate work with a given
1577 * This function will only make a best effort attempt at getting this onto
1578 * the right NUMA node. If no node is requested or the requested node is
1579 * offline then we just fall back to standard queue_work behavior.
1581 * Currently the "random" CPU ends up being the first available CPU in the
1582 * intersection of cpu_online_mask and the cpumask of the node, unless we
1583 * are running on the node. In that case we just use the current CPU.
1585 * Return: %false if @work was already on a queue, %true otherwise.
1587 bool queue_work_node(int node
, struct workqueue_struct
*wq
,
1588 struct work_struct
*work
)
1590 unsigned long flags
;
1594 * This current implementation is specific to unbound workqueues.
1595 * Specifically we only return the first available CPU for a given
1596 * node instead of cycling through individual CPUs within the node.
1598 * If this is used with a per-cpu workqueue then the logic in
1599 * workqueue_select_cpu_near would need to be updated to allow for
1600 * some round robin type logic.
1602 WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
));
1604 local_irq_save(flags
);
1606 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1607 int cpu
= workqueue_select_cpu_near(node
);
1609 __queue_work(cpu
, wq
, work
);
1613 local_irq_restore(flags
);
1616 EXPORT_SYMBOL_GPL(queue_work_node
);
1618 void delayed_work_timer_fn(struct timer_list
*t
)
1620 struct delayed_work
*dwork
= from_timer(dwork
, t
, timer
);
1622 /* should have been called from irqsafe timer with irq already off */
1623 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1625 EXPORT_SYMBOL(delayed_work_timer_fn
);
1627 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1628 struct delayed_work
*dwork
, unsigned long delay
)
1630 struct timer_list
*timer
= &dwork
->timer
;
1631 struct work_struct
*work
= &dwork
->work
;
1634 WARN_ON_FUNCTION_MISMATCH(timer
->function
, delayed_work_timer_fn
);
1635 WARN_ON_ONCE(timer_pending(timer
));
1636 WARN_ON_ONCE(!list_empty(&work
->entry
));
1639 * If @delay is 0, queue @dwork->work immediately. This is for
1640 * both optimization and correctness. The earliest @timer can
1641 * expire is on the closest next tick and delayed_work users depend
1642 * on that there's no such delay when @delay is 0.
1645 __queue_work(cpu
, wq
, &dwork
->work
);
1651 timer
->expires
= jiffies
+ delay
;
1653 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1654 add_timer_on(timer
, cpu
);
1660 * queue_delayed_work_on - queue work on specific CPU after delay
1661 * @cpu: CPU number to execute work on
1662 * @wq: workqueue to use
1663 * @dwork: work to queue
1664 * @delay: number of jiffies to wait before queueing
1666 * Return: %false if @work was already on a queue, %true otherwise. If
1667 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1670 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1671 struct delayed_work
*dwork
, unsigned long delay
)
1673 struct work_struct
*work
= &dwork
->work
;
1675 unsigned long flags
;
1677 /* read the comment in __queue_work() */
1678 local_irq_save(flags
);
1680 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1681 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1685 local_irq_restore(flags
);
1688 EXPORT_SYMBOL(queue_delayed_work_on
);
1691 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1692 * @cpu: CPU number to execute work on
1693 * @wq: workqueue to use
1694 * @dwork: work to queue
1695 * @delay: number of jiffies to wait before queueing
1697 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1698 * modify @dwork's timer so that it expires after @delay. If @delay is
1699 * zero, @work is guaranteed to be scheduled immediately regardless of its
1702 * Return: %false if @dwork was idle and queued, %true if @dwork was
1703 * pending and its timer was modified.
1705 * This function is safe to call from any context including IRQ handler.
1706 * See try_to_grab_pending() for details.
1708 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1709 struct delayed_work
*dwork
, unsigned long delay
)
1711 unsigned long flags
;
1715 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1716 } while (unlikely(ret
== -EAGAIN
));
1718 if (likely(ret
>= 0)) {
1719 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1720 local_irq_restore(flags
);
1723 /* -ENOENT from try_to_grab_pending() becomes %true */
1726 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1728 static void rcu_work_rcufn(struct rcu_head
*rcu
)
1730 struct rcu_work
*rwork
= container_of(rcu
, struct rcu_work
, rcu
);
1732 /* read the comment in __queue_work() */
1733 local_irq_disable();
1734 __queue_work(WORK_CPU_UNBOUND
, rwork
->wq
, &rwork
->work
);
1739 * queue_rcu_work - queue work after a RCU grace period
1740 * @wq: workqueue to use
1741 * @rwork: work to queue
1743 * Return: %false if @rwork was already pending, %true otherwise. Note
1744 * that a full RCU grace period is guaranteed only after a %true return.
1745 * While @rwork is guaranteed to be executed after a %false return, the
1746 * execution may happen before a full RCU grace period has passed.
1748 bool queue_rcu_work(struct workqueue_struct
*wq
, struct rcu_work
*rwork
)
1750 struct work_struct
*work
= &rwork
->work
;
1752 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1754 call_rcu(&rwork
->rcu
, rcu_work_rcufn
);
1760 EXPORT_SYMBOL(queue_rcu_work
);
1763 * worker_enter_idle - enter idle state
1764 * @worker: worker which is entering idle state
1766 * @worker is entering idle state. Update stats and idle timer if
1770 * raw_spin_lock_irq(pool->lock).
1772 static void worker_enter_idle(struct worker
*worker
)
1774 struct worker_pool
*pool
= worker
->pool
;
1776 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1777 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1778 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1781 /* can't use worker_set_flags(), also called from create_worker() */
1782 worker
->flags
|= WORKER_IDLE
;
1784 worker
->last_active
= jiffies
;
1786 /* idle_list is LIFO */
1787 list_add(&worker
->entry
, &pool
->idle_list
);
1789 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1790 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1793 * Sanity check nr_running. Because unbind_workers() releases
1794 * pool->lock between setting %WORKER_UNBOUND and zapping
1795 * nr_running, the warning may trigger spuriously. Check iff
1796 * unbind is not in progress.
1798 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1799 pool
->nr_workers
== pool
->nr_idle
&&
1800 atomic_read(&pool
->nr_running
));
1804 * worker_leave_idle - leave idle state
1805 * @worker: worker which is leaving idle state
1807 * @worker is leaving idle state. Update stats.
1810 * raw_spin_lock_irq(pool->lock).
1812 static void worker_leave_idle(struct worker
*worker
)
1814 struct worker_pool
*pool
= worker
->pool
;
1816 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1818 worker_clr_flags(worker
, WORKER_IDLE
);
1820 list_del_init(&worker
->entry
);
1823 static struct worker
*alloc_worker(int node
)
1825 struct worker
*worker
;
1827 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1829 INIT_LIST_HEAD(&worker
->entry
);
1830 INIT_LIST_HEAD(&worker
->scheduled
);
1831 INIT_LIST_HEAD(&worker
->node
);
1832 /* on creation a worker is in !idle && prep state */
1833 worker
->flags
= WORKER_PREP
;
1839 * worker_attach_to_pool() - attach a worker to a pool
1840 * @worker: worker to be attached
1841 * @pool: the target pool
1843 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1844 * cpu-binding of @worker are kept coordinated with the pool across
1847 static void worker_attach_to_pool(struct worker
*worker
,
1848 struct worker_pool
*pool
)
1850 mutex_lock(&wq_pool_attach_mutex
);
1853 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1854 * stable across this function. See the comments above the flag
1855 * definition for details.
1857 if (pool
->flags
& POOL_DISASSOCIATED
)
1858 worker
->flags
|= WORKER_UNBOUND
;
1860 kthread_set_per_cpu(worker
->task
, pool
->cpu
);
1862 if (worker
->rescue_wq
)
1863 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1865 list_add_tail(&worker
->node
, &pool
->workers
);
1866 worker
->pool
= pool
;
1868 mutex_unlock(&wq_pool_attach_mutex
);
1872 * worker_detach_from_pool() - detach a worker from its pool
1873 * @worker: worker which is attached to its pool
1875 * Undo the attaching which had been done in worker_attach_to_pool(). The
1876 * caller worker shouldn't access to the pool after detached except it has
1877 * other reference to the pool.
1879 static void worker_detach_from_pool(struct worker
*worker
)
1881 struct worker_pool
*pool
= worker
->pool
;
1882 struct completion
*detach_completion
= NULL
;
1884 mutex_lock(&wq_pool_attach_mutex
);
1886 kthread_set_per_cpu(worker
->task
, -1);
1887 list_del(&worker
->node
);
1888 worker
->pool
= NULL
;
1890 if (list_empty(&pool
->workers
))
1891 detach_completion
= pool
->detach_completion
;
1892 mutex_unlock(&wq_pool_attach_mutex
);
1894 /* clear leftover flags without pool->lock after it is detached */
1895 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1897 if (detach_completion
)
1898 complete(detach_completion
);
1902 * create_worker - create a new workqueue worker
1903 * @pool: pool the new worker will belong to
1905 * Create and start a new worker which is attached to @pool.
1908 * Might sleep. Does GFP_KERNEL allocations.
1911 * Pointer to the newly created worker.
1913 static struct worker
*create_worker(struct worker_pool
*pool
)
1915 struct worker
*worker
= NULL
;
1919 /* ID is needed to determine kthread name */
1920 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1924 worker
= alloc_worker(pool
->node
);
1931 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1932 pool
->attrs
->nice
< 0 ? "H" : "");
1934 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1936 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1937 "kworker/%s", id_buf
);
1938 if (IS_ERR(worker
->task
))
1941 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1942 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1944 /* successful, attach the worker to the pool */
1945 worker_attach_to_pool(worker
, pool
);
1947 /* start the newly created worker */
1948 raw_spin_lock_irq(&pool
->lock
);
1949 worker
->pool
->nr_workers
++;
1950 worker_enter_idle(worker
);
1951 wake_up_process(worker
->task
);
1952 raw_spin_unlock_irq(&pool
->lock
);
1958 ida_simple_remove(&pool
->worker_ida
, id
);
1964 * destroy_worker - destroy a workqueue worker
1965 * @worker: worker to be destroyed
1967 * Destroy @worker and adjust @pool stats accordingly. The worker should
1971 * raw_spin_lock_irq(pool->lock).
1973 static void destroy_worker(struct worker
*worker
)
1975 struct worker_pool
*pool
= worker
->pool
;
1977 lockdep_assert_held(&pool
->lock
);
1979 /* sanity check frenzy */
1980 if (WARN_ON(worker
->current_work
) ||
1981 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1982 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1988 list_del_init(&worker
->entry
);
1989 worker
->flags
|= WORKER_DIE
;
1990 wake_up_process(worker
->task
);
1993 static void idle_worker_timeout(struct timer_list
*t
)
1995 struct worker_pool
*pool
= from_timer(pool
, t
, idle_timer
);
1997 raw_spin_lock_irq(&pool
->lock
);
1999 while (too_many_workers(pool
)) {
2000 struct worker
*worker
;
2001 unsigned long expires
;
2003 /* idle_list is kept in LIFO order, check the last one */
2004 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2005 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2007 if (time_before(jiffies
, expires
)) {
2008 mod_timer(&pool
->idle_timer
, expires
);
2012 destroy_worker(worker
);
2015 raw_spin_unlock_irq(&pool
->lock
);
2018 static void send_mayday(struct work_struct
*work
)
2020 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2021 struct workqueue_struct
*wq
= pwq
->wq
;
2023 lockdep_assert_held(&wq_mayday_lock
);
2028 /* mayday mayday mayday */
2029 if (list_empty(&pwq
->mayday_node
)) {
2031 * If @pwq is for an unbound wq, its base ref may be put at
2032 * any time due to an attribute change. Pin @pwq until the
2033 * rescuer is done with it.
2036 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2037 wake_up_process(wq
->rescuer
->task
);
2041 static void pool_mayday_timeout(struct timer_list
*t
)
2043 struct worker_pool
*pool
= from_timer(pool
, t
, mayday_timer
);
2044 struct work_struct
*work
;
2046 raw_spin_lock_irq(&pool
->lock
);
2047 raw_spin_lock(&wq_mayday_lock
); /* for wq->maydays */
2049 if (need_to_create_worker(pool
)) {
2051 * We've been trying to create a new worker but
2052 * haven't been successful. We might be hitting an
2053 * allocation deadlock. Send distress signals to
2056 list_for_each_entry(work
, &pool
->worklist
, entry
)
2060 raw_spin_unlock(&wq_mayday_lock
);
2061 raw_spin_unlock_irq(&pool
->lock
);
2063 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
2067 * maybe_create_worker - create a new worker if necessary
2068 * @pool: pool to create a new worker for
2070 * Create a new worker for @pool if necessary. @pool is guaranteed to
2071 * have at least one idle worker on return from this function. If
2072 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2073 * sent to all rescuers with works scheduled on @pool to resolve
2074 * possible allocation deadlock.
2076 * On return, need_to_create_worker() is guaranteed to be %false and
2077 * may_start_working() %true.
2080 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2081 * multiple times. Does GFP_KERNEL allocations. Called only from
2084 static void maybe_create_worker(struct worker_pool
*pool
)
2085 __releases(&pool
->lock
)
2086 __acquires(&pool
->lock
)
2089 raw_spin_unlock_irq(&pool
->lock
);
2091 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2092 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
2095 if (create_worker(pool
) || !need_to_create_worker(pool
))
2098 schedule_timeout_interruptible(CREATE_COOLDOWN
);
2100 if (!need_to_create_worker(pool
))
2104 del_timer_sync(&pool
->mayday_timer
);
2105 raw_spin_lock_irq(&pool
->lock
);
2107 * This is necessary even after a new worker was just successfully
2108 * created as @pool->lock was dropped and the new worker might have
2109 * already become busy.
2111 if (need_to_create_worker(pool
))
2116 * manage_workers - manage worker pool
2119 * Assume the manager role and manage the worker pool @worker belongs
2120 * to. At any given time, there can be only zero or one manager per
2121 * pool. The exclusion is handled automatically by this function.
2123 * The caller can safely start processing works on false return. On
2124 * true return, it's guaranteed that need_to_create_worker() is false
2125 * and may_start_working() is true.
2128 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2129 * multiple times. Does GFP_KERNEL allocations.
2132 * %false if the pool doesn't need management and the caller can safely
2133 * start processing works, %true if management function was performed and
2134 * the conditions that the caller verified before calling the function may
2135 * no longer be true.
2137 static bool manage_workers(struct worker
*worker
)
2139 struct worker_pool
*pool
= worker
->pool
;
2141 if (pool
->flags
& POOL_MANAGER_ACTIVE
)
2144 pool
->flags
|= POOL_MANAGER_ACTIVE
;
2145 pool
->manager
= worker
;
2147 maybe_create_worker(pool
);
2149 pool
->manager
= NULL
;
2150 pool
->flags
&= ~POOL_MANAGER_ACTIVE
;
2151 rcuwait_wake_up(&manager_wait
);
2156 * process_one_work - process single work
2158 * @work: work to process
2160 * Process @work. This function contains all the logics necessary to
2161 * process a single work including synchronization against and
2162 * interaction with other workers on the same cpu, queueing and
2163 * flushing. As long as context requirement is met, any worker can
2164 * call this function to process a work.
2167 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2169 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2170 __releases(&pool
->lock
)
2171 __acquires(&pool
->lock
)
2173 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2174 struct worker_pool
*pool
= worker
->pool
;
2175 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2177 struct worker
*collision
;
2178 #ifdef CONFIG_LOCKDEP
2180 * It is permissible to free the struct work_struct from
2181 * inside the function that is called from it, this we need to
2182 * take into account for lockdep too. To avoid bogus "held
2183 * lock freed" warnings as well as problems when looking into
2184 * work->lockdep_map, make a copy and use that here.
2186 struct lockdep_map lockdep_map
;
2188 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2190 /* ensure we're on the correct CPU */
2191 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2192 raw_smp_processor_id() != pool
->cpu
);
2195 * A single work shouldn't be executed concurrently by
2196 * multiple workers on a single cpu. Check whether anyone is
2197 * already processing the work. If so, defer the work to the
2198 * currently executing one.
2200 collision
= find_worker_executing_work(pool
, work
);
2201 if (unlikely(collision
)) {
2202 move_linked_works(work
, &collision
->scheduled
, NULL
);
2206 /* claim and dequeue */
2207 debug_work_deactivate(work
);
2208 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2209 worker
->current_work
= work
;
2210 worker
->current_func
= work
->func
;
2211 worker
->current_pwq
= pwq
;
2212 work_color
= get_work_color(work
);
2215 * Record wq name for cmdline and debug reporting, may get
2216 * overridden through set_worker_desc().
2218 strscpy(worker
->desc
, pwq
->wq
->name
, WORKER_DESC_LEN
);
2220 list_del_init(&work
->entry
);
2223 * CPU intensive works don't participate in concurrency management.
2224 * They're the scheduler's responsibility. This takes @worker out
2225 * of concurrency management and the next code block will chain
2226 * execution of the pending work items.
2228 if (unlikely(cpu_intensive
))
2229 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2232 * Wake up another worker if necessary. The condition is always
2233 * false for normal per-cpu workers since nr_running would always
2234 * be >= 1 at this point. This is used to chain execution of the
2235 * pending work items for WORKER_NOT_RUNNING workers such as the
2236 * UNBOUND and CPU_INTENSIVE ones.
2238 if (need_more_worker(pool
))
2239 wake_up_worker(pool
);
2242 * Record the last pool and clear PENDING which should be the last
2243 * update to @work. Also, do this inside @pool->lock so that
2244 * PENDING and queued state changes happen together while IRQ is
2247 set_work_pool_and_clear_pending(work
, pool
->id
);
2249 raw_spin_unlock_irq(&pool
->lock
);
2251 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2252 lock_map_acquire(&lockdep_map
);
2254 * Strictly speaking we should mark the invariant state without holding
2255 * any locks, that is, before these two lock_map_acquire()'s.
2257 * However, that would result in:
2264 * Which would create W1->C->W1 dependencies, even though there is no
2265 * actual deadlock possible. There are two solutions, using a
2266 * read-recursive acquire on the work(queue) 'locks', but this will then
2267 * hit the lockdep limitation on recursive locks, or simply discard
2270 * AFAICT there is no possible deadlock scenario between the
2271 * flush_work() and complete() primitives (except for single-threaded
2272 * workqueues), so hiding them isn't a problem.
2274 lockdep_invariant_state(true);
2275 trace_workqueue_execute_start(work
);
2276 worker
->current_func(work
);
2278 * While we must be careful to not use "work" after this, the trace
2279 * point will only record its address.
2281 trace_workqueue_execute_end(work
, worker
->current_func
);
2282 lock_map_release(&lockdep_map
);
2283 lock_map_release(&pwq
->wq
->lockdep_map
);
2285 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2286 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2287 " last function: %ps\n",
2288 current
->comm
, preempt_count(), task_pid_nr(current
),
2289 worker
->current_func
);
2290 debug_show_held_locks(current
);
2295 * The following prevents a kworker from hogging CPU on !PREEMPTION
2296 * kernels, where a requeueing work item waiting for something to
2297 * happen could deadlock with stop_machine as such work item could
2298 * indefinitely requeue itself while all other CPUs are trapped in
2299 * stop_machine. At the same time, report a quiescent RCU state so
2300 * the same condition doesn't freeze RCU.
2304 raw_spin_lock_irq(&pool
->lock
);
2306 /* clear cpu intensive status */
2307 if (unlikely(cpu_intensive
))
2308 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2310 /* tag the worker for identification in schedule() */
2311 worker
->last_func
= worker
->current_func
;
2313 /* we're done with it, release */
2314 hash_del(&worker
->hentry
);
2315 worker
->current_work
= NULL
;
2316 worker
->current_func
= NULL
;
2317 worker
->current_pwq
= NULL
;
2318 pwq_dec_nr_in_flight(pwq
, work_color
);
2322 * process_scheduled_works - process scheduled works
2325 * Process all scheduled works. Please note that the scheduled list
2326 * may change while processing a work, so this function repeatedly
2327 * fetches a work from the top and executes it.
2330 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2333 static void process_scheduled_works(struct worker
*worker
)
2335 while (!list_empty(&worker
->scheduled
)) {
2336 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2337 struct work_struct
, entry
);
2338 process_one_work(worker
, work
);
2342 static void set_pf_worker(bool val
)
2344 mutex_lock(&wq_pool_attach_mutex
);
2346 current
->flags
|= PF_WQ_WORKER
;
2348 current
->flags
&= ~PF_WQ_WORKER
;
2349 mutex_unlock(&wq_pool_attach_mutex
);
2353 * worker_thread - the worker thread function
2356 * The worker thread function. All workers belong to a worker_pool -
2357 * either a per-cpu one or dynamic unbound one. These workers process all
2358 * work items regardless of their specific target workqueue. The only
2359 * exception is work items which belong to workqueues with a rescuer which
2360 * will be explained in rescuer_thread().
2364 static int worker_thread(void *__worker
)
2366 struct worker
*worker
= __worker
;
2367 struct worker_pool
*pool
= worker
->pool
;
2369 /* tell the scheduler that this is a workqueue worker */
2370 set_pf_worker(true);
2372 raw_spin_lock_irq(&pool
->lock
);
2374 /* am I supposed to die? */
2375 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2376 raw_spin_unlock_irq(&pool
->lock
);
2377 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2378 set_pf_worker(false);
2380 set_task_comm(worker
->task
, "kworker/dying");
2381 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2382 worker_detach_from_pool(worker
);
2387 worker_leave_idle(worker
);
2389 /* no more worker necessary? */
2390 if (!need_more_worker(pool
))
2393 /* do we need to manage? */
2394 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2398 * ->scheduled list can only be filled while a worker is
2399 * preparing to process a work or actually processing it.
2400 * Make sure nobody diddled with it while I was sleeping.
2402 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2405 * Finish PREP stage. We're guaranteed to have at least one idle
2406 * worker or that someone else has already assumed the manager
2407 * role. This is where @worker starts participating in concurrency
2408 * management if applicable and concurrency management is restored
2409 * after being rebound. See rebind_workers() for details.
2411 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2414 struct work_struct
*work
=
2415 list_first_entry(&pool
->worklist
,
2416 struct work_struct
, entry
);
2418 pool
->watchdog_ts
= jiffies
;
2420 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2421 /* optimization path, not strictly necessary */
2422 process_one_work(worker
, work
);
2423 if (unlikely(!list_empty(&worker
->scheduled
)))
2424 process_scheduled_works(worker
);
2426 move_linked_works(work
, &worker
->scheduled
, NULL
);
2427 process_scheduled_works(worker
);
2429 } while (keep_working(pool
));
2431 worker_set_flags(worker
, WORKER_PREP
);
2434 * pool->lock is held and there's no work to process and no need to
2435 * manage, sleep. Workers are woken up only while holding
2436 * pool->lock or from local cpu, so setting the current state
2437 * before releasing pool->lock is enough to prevent losing any
2440 worker_enter_idle(worker
);
2441 __set_current_state(TASK_IDLE
);
2442 raw_spin_unlock_irq(&pool
->lock
);
2448 * rescuer_thread - the rescuer thread function
2451 * Workqueue rescuer thread function. There's one rescuer for each
2452 * workqueue which has WQ_MEM_RECLAIM set.
2454 * Regular work processing on a pool may block trying to create a new
2455 * worker which uses GFP_KERNEL allocation which has slight chance of
2456 * developing into deadlock if some works currently on the same queue
2457 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2458 * the problem rescuer solves.
2460 * When such condition is possible, the pool summons rescuers of all
2461 * workqueues which have works queued on the pool and let them process
2462 * those works so that forward progress can be guaranteed.
2464 * This should happen rarely.
2468 static int rescuer_thread(void *__rescuer
)
2470 struct worker
*rescuer
= __rescuer
;
2471 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2472 struct list_head
*scheduled
= &rescuer
->scheduled
;
2475 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2478 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2479 * doesn't participate in concurrency management.
2481 set_pf_worker(true);
2483 set_current_state(TASK_IDLE
);
2486 * By the time the rescuer is requested to stop, the workqueue
2487 * shouldn't have any work pending, but @wq->maydays may still have
2488 * pwq(s) queued. This can happen by non-rescuer workers consuming
2489 * all the work items before the rescuer got to them. Go through
2490 * @wq->maydays processing before acting on should_stop so that the
2491 * list is always empty on exit.
2493 should_stop
= kthread_should_stop();
2495 /* see whether any pwq is asking for help */
2496 raw_spin_lock_irq(&wq_mayday_lock
);
2498 while (!list_empty(&wq
->maydays
)) {
2499 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2500 struct pool_workqueue
, mayday_node
);
2501 struct worker_pool
*pool
= pwq
->pool
;
2502 struct work_struct
*work
, *n
;
2505 __set_current_state(TASK_RUNNING
);
2506 list_del_init(&pwq
->mayday_node
);
2508 raw_spin_unlock_irq(&wq_mayday_lock
);
2510 worker_attach_to_pool(rescuer
, pool
);
2512 raw_spin_lock_irq(&pool
->lock
);
2515 * Slurp in all works issued via this workqueue and
2518 WARN_ON_ONCE(!list_empty(scheduled
));
2519 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2520 if (get_work_pwq(work
) == pwq
) {
2522 pool
->watchdog_ts
= jiffies
;
2523 move_linked_works(work
, scheduled
, &n
);
2528 if (!list_empty(scheduled
)) {
2529 process_scheduled_works(rescuer
);
2532 * The above execution of rescued work items could
2533 * have created more to rescue through
2534 * pwq_activate_first_delayed() or chained
2535 * queueing. Let's put @pwq back on mayday list so
2536 * that such back-to-back work items, which may be
2537 * being used to relieve memory pressure, don't
2538 * incur MAYDAY_INTERVAL delay inbetween.
2540 if (pwq
->nr_active
&& need_to_create_worker(pool
)) {
2541 raw_spin_lock(&wq_mayday_lock
);
2543 * Queue iff we aren't racing destruction
2544 * and somebody else hasn't queued it already.
2546 if (wq
->rescuer
&& list_empty(&pwq
->mayday_node
)) {
2548 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2550 raw_spin_unlock(&wq_mayday_lock
);
2555 * Put the reference grabbed by send_mayday(). @pool won't
2556 * go away while we're still attached to it.
2561 * Leave this pool. If need_more_worker() is %true, notify a
2562 * regular worker; otherwise, we end up with 0 concurrency
2563 * and stalling the execution.
2565 if (need_more_worker(pool
))
2566 wake_up_worker(pool
);
2568 raw_spin_unlock_irq(&pool
->lock
);
2570 worker_detach_from_pool(rescuer
);
2572 raw_spin_lock_irq(&wq_mayday_lock
);
2575 raw_spin_unlock_irq(&wq_mayday_lock
);
2578 __set_current_state(TASK_RUNNING
);
2579 set_pf_worker(false);
2583 /* rescuers should never participate in concurrency management */
2584 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2590 * check_flush_dependency - check for flush dependency sanity
2591 * @target_wq: workqueue being flushed
2592 * @target_work: work item being flushed (NULL for workqueue flushes)
2594 * %current is trying to flush the whole @target_wq or @target_work on it.
2595 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2596 * reclaiming memory or running on a workqueue which doesn't have
2597 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2600 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2601 struct work_struct
*target_work
)
2603 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2604 struct worker
*worker
;
2606 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2609 worker
= current_wq_worker();
2611 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2612 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2613 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2614 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2615 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2616 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2617 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2618 target_wq
->name
, target_func
);
2622 struct work_struct work
;
2623 struct completion done
;
2624 struct task_struct
*task
; /* purely informational */
2627 static void wq_barrier_func(struct work_struct
*work
)
2629 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2630 complete(&barr
->done
);
2634 * insert_wq_barrier - insert a barrier work
2635 * @pwq: pwq to insert barrier into
2636 * @barr: wq_barrier to insert
2637 * @target: target work to attach @barr to
2638 * @worker: worker currently executing @target, NULL if @target is not executing
2640 * @barr is linked to @target such that @barr is completed only after
2641 * @target finishes execution. Please note that the ordering
2642 * guarantee is observed only with respect to @target and on the local
2645 * Currently, a queued barrier can't be canceled. This is because
2646 * try_to_grab_pending() can't determine whether the work to be
2647 * grabbed is at the head of the queue and thus can't clear LINKED
2648 * flag of the previous work while there must be a valid next work
2649 * after a work with LINKED flag set.
2651 * Note that when @worker is non-NULL, @target may be modified
2652 * underneath us, so we can't reliably determine pwq from @target.
2655 * raw_spin_lock_irq(pool->lock).
2657 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2658 struct wq_barrier
*barr
,
2659 struct work_struct
*target
, struct worker
*worker
)
2661 struct list_head
*head
;
2662 unsigned int linked
= 0;
2665 * debugobject calls are safe here even with pool->lock locked
2666 * as we know for sure that this will not trigger any of the
2667 * checks and call back into the fixup functions where we
2670 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2671 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2673 init_completion_map(&barr
->done
, &target
->lockdep_map
);
2675 barr
->task
= current
;
2678 * If @target is currently being executed, schedule the
2679 * barrier to the worker; otherwise, put it after @target.
2682 head
= worker
->scheduled
.next
;
2684 unsigned long *bits
= work_data_bits(target
);
2686 head
= target
->entry
.next
;
2687 /* there can already be other linked works, inherit and set */
2688 linked
= *bits
& WORK_STRUCT_LINKED
;
2689 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2692 debug_work_activate(&barr
->work
);
2693 insert_work(pwq
, &barr
->work
, head
,
2694 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2698 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2699 * @wq: workqueue being flushed
2700 * @flush_color: new flush color, < 0 for no-op
2701 * @work_color: new work color, < 0 for no-op
2703 * Prepare pwqs for workqueue flushing.
2705 * If @flush_color is non-negative, flush_color on all pwqs should be
2706 * -1. If no pwq has in-flight commands at the specified color, all
2707 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2708 * has in flight commands, its pwq->flush_color is set to
2709 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2710 * wakeup logic is armed and %true is returned.
2712 * The caller should have initialized @wq->first_flusher prior to
2713 * calling this function with non-negative @flush_color. If
2714 * @flush_color is negative, no flush color update is done and %false
2717 * If @work_color is non-negative, all pwqs should have the same
2718 * work_color which is previous to @work_color and all will be
2719 * advanced to @work_color.
2722 * mutex_lock(wq->mutex).
2725 * %true if @flush_color >= 0 and there's something to flush. %false
2728 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2729 int flush_color
, int work_color
)
2732 struct pool_workqueue
*pwq
;
2734 if (flush_color
>= 0) {
2735 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2736 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2739 for_each_pwq(pwq
, wq
) {
2740 struct worker_pool
*pool
= pwq
->pool
;
2742 raw_spin_lock_irq(&pool
->lock
);
2744 if (flush_color
>= 0) {
2745 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2747 if (pwq
->nr_in_flight
[flush_color
]) {
2748 pwq
->flush_color
= flush_color
;
2749 atomic_inc(&wq
->nr_pwqs_to_flush
);
2754 if (work_color
>= 0) {
2755 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2756 pwq
->work_color
= work_color
;
2759 raw_spin_unlock_irq(&pool
->lock
);
2762 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2763 complete(&wq
->first_flusher
->done
);
2769 * flush_workqueue - ensure that any scheduled work has run to completion.
2770 * @wq: workqueue to flush
2772 * This function sleeps until all work items which were queued on entry
2773 * have finished execution, but it is not livelocked by new incoming ones.
2775 void flush_workqueue(struct workqueue_struct
*wq
)
2777 struct wq_flusher this_flusher
= {
2778 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2780 .done
= COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher
.done
, wq
->lockdep_map
),
2784 if (WARN_ON(!wq_online
))
2787 lock_map_acquire(&wq
->lockdep_map
);
2788 lock_map_release(&wq
->lockdep_map
);
2790 mutex_lock(&wq
->mutex
);
2793 * Start-to-wait phase
2795 next_color
= work_next_color(wq
->work_color
);
2797 if (next_color
!= wq
->flush_color
) {
2799 * Color space is not full. The current work_color
2800 * becomes our flush_color and work_color is advanced
2803 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2804 this_flusher
.flush_color
= wq
->work_color
;
2805 wq
->work_color
= next_color
;
2807 if (!wq
->first_flusher
) {
2808 /* no flush in progress, become the first flusher */
2809 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2811 wq
->first_flusher
= &this_flusher
;
2813 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2815 /* nothing to flush, done */
2816 wq
->flush_color
= next_color
;
2817 wq
->first_flusher
= NULL
;
2822 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2823 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2824 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2828 * Oops, color space is full, wait on overflow queue.
2829 * The next flush completion will assign us
2830 * flush_color and transfer to flusher_queue.
2832 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2835 check_flush_dependency(wq
, NULL
);
2837 mutex_unlock(&wq
->mutex
);
2839 wait_for_completion(&this_flusher
.done
);
2842 * Wake-up-and-cascade phase
2844 * First flushers are responsible for cascading flushes and
2845 * handling overflow. Non-first flushers can simply return.
2847 if (READ_ONCE(wq
->first_flusher
) != &this_flusher
)
2850 mutex_lock(&wq
->mutex
);
2852 /* we might have raced, check again with mutex held */
2853 if (wq
->first_flusher
!= &this_flusher
)
2856 WRITE_ONCE(wq
->first_flusher
, NULL
);
2858 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2859 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2862 struct wq_flusher
*next
, *tmp
;
2864 /* complete all the flushers sharing the current flush color */
2865 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2866 if (next
->flush_color
!= wq
->flush_color
)
2868 list_del_init(&next
->list
);
2869 complete(&next
->done
);
2872 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2873 wq
->flush_color
!= work_next_color(wq
->work_color
));
2875 /* this flush_color is finished, advance by one */
2876 wq
->flush_color
= work_next_color(wq
->flush_color
);
2878 /* one color has been freed, handle overflow queue */
2879 if (!list_empty(&wq
->flusher_overflow
)) {
2881 * Assign the same color to all overflowed
2882 * flushers, advance work_color and append to
2883 * flusher_queue. This is the start-to-wait
2884 * phase for these overflowed flushers.
2886 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2887 tmp
->flush_color
= wq
->work_color
;
2889 wq
->work_color
= work_next_color(wq
->work_color
);
2891 list_splice_tail_init(&wq
->flusher_overflow
,
2892 &wq
->flusher_queue
);
2893 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2896 if (list_empty(&wq
->flusher_queue
)) {
2897 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2902 * Need to flush more colors. Make the next flusher
2903 * the new first flusher and arm pwqs.
2905 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2906 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2908 list_del_init(&next
->list
);
2909 wq
->first_flusher
= next
;
2911 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2915 * Meh... this color is already done, clear first
2916 * flusher and repeat cascading.
2918 wq
->first_flusher
= NULL
;
2922 mutex_unlock(&wq
->mutex
);
2924 EXPORT_SYMBOL(flush_workqueue
);
2927 * drain_workqueue - drain a workqueue
2928 * @wq: workqueue to drain
2930 * Wait until the workqueue becomes empty. While draining is in progress,
2931 * only chain queueing is allowed. IOW, only currently pending or running
2932 * work items on @wq can queue further work items on it. @wq is flushed
2933 * repeatedly until it becomes empty. The number of flushing is determined
2934 * by the depth of chaining and should be relatively short. Whine if it
2937 void drain_workqueue(struct workqueue_struct
*wq
)
2939 unsigned int flush_cnt
= 0;
2940 struct pool_workqueue
*pwq
;
2943 * __queue_work() needs to test whether there are drainers, is much
2944 * hotter than drain_workqueue() and already looks at @wq->flags.
2945 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2947 mutex_lock(&wq
->mutex
);
2948 if (!wq
->nr_drainers
++)
2949 wq
->flags
|= __WQ_DRAINING
;
2950 mutex_unlock(&wq
->mutex
);
2952 flush_workqueue(wq
);
2954 mutex_lock(&wq
->mutex
);
2956 for_each_pwq(pwq
, wq
) {
2959 raw_spin_lock_irq(&pwq
->pool
->lock
);
2960 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2961 raw_spin_unlock_irq(&pwq
->pool
->lock
);
2966 if (++flush_cnt
== 10 ||
2967 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2968 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
2969 wq
->name
, __func__
, flush_cnt
);
2971 mutex_unlock(&wq
->mutex
);
2975 if (!--wq
->nr_drainers
)
2976 wq
->flags
&= ~__WQ_DRAINING
;
2977 mutex_unlock(&wq
->mutex
);
2979 EXPORT_SYMBOL_GPL(drain_workqueue
);
2981 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2984 struct worker
*worker
= NULL
;
2985 struct worker_pool
*pool
;
2986 struct pool_workqueue
*pwq
;
2991 pool
= get_work_pool(work
);
2997 raw_spin_lock_irq(&pool
->lock
);
2998 /* see the comment in try_to_grab_pending() with the same code */
2999 pwq
= get_work_pwq(work
);
3001 if (unlikely(pwq
->pool
!= pool
))
3004 worker
= find_worker_executing_work(pool
, work
);
3007 pwq
= worker
->current_pwq
;
3010 check_flush_dependency(pwq
->wq
, work
);
3012 insert_wq_barrier(pwq
, barr
, work
, worker
);
3013 raw_spin_unlock_irq(&pool
->lock
);
3016 * Force a lock recursion deadlock when using flush_work() inside a
3017 * single-threaded or rescuer equipped workqueue.
3019 * For single threaded workqueues the deadlock happens when the work
3020 * is after the work issuing the flush_work(). For rescuer equipped
3021 * workqueues the deadlock happens when the rescuer stalls, blocking
3025 (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)) {
3026 lock_map_acquire(&pwq
->wq
->lockdep_map
);
3027 lock_map_release(&pwq
->wq
->lockdep_map
);
3032 raw_spin_unlock_irq(&pool
->lock
);
3037 static bool __flush_work(struct work_struct
*work
, bool from_cancel
)
3039 struct wq_barrier barr
;
3041 if (WARN_ON(!wq_online
))
3044 if (WARN_ON(!work
->func
))
3048 lock_map_acquire(&work
->lockdep_map
);
3049 lock_map_release(&work
->lockdep_map
);
3052 if (start_flush_work(work
, &barr
, from_cancel
)) {
3053 wait_for_completion(&barr
.done
);
3054 destroy_work_on_stack(&barr
.work
);
3062 * flush_work - wait for a work to finish executing the last queueing instance
3063 * @work: the work to flush
3065 * Wait until @work has finished execution. @work is guaranteed to be idle
3066 * on return if it hasn't been requeued since flush started.
3069 * %true if flush_work() waited for the work to finish execution,
3070 * %false if it was already idle.
3072 bool flush_work(struct work_struct
*work
)
3074 return __flush_work(work
, false);
3076 EXPORT_SYMBOL_GPL(flush_work
);
3079 wait_queue_entry_t wait
;
3080 struct work_struct
*work
;
3083 static int cwt_wakefn(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
3085 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
3087 if (cwait
->work
!= key
)
3089 return autoremove_wake_function(wait
, mode
, sync
, key
);
3092 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
3094 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
3095 unsigned long flags
;
3099 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3101 * If someone else is already canceling, wait for it to
3102 * finish. flush_work() doesn't work for PREEMPT_NONE
3103 * because we may get scheduled between @work's completion
3104 * and the other canceling task resuming and clearing
3105 * CANCELING - flush_work() will return false immediately
3106 * as @work is no longer busy, try_to_grab_pending() will
3107 * return -ENOENT as @work is still being canceled and the
3108 * other canceling task won't be able to clear CANCELING as
3109 * we're hogging the CPU.
3111 * Let's wait for completion using a waitqueue. As this
3112 * may lead to the thundering herd problem, use a custom
3113 * wake function which matches @work along with exclusive
3116 if (unlikely(ret
== -ENOENT
)) {
3117 struct cwt_wait cwait
;
3119 init_wait(&cwait
.wait
);
3120 cwait
.wait
.func
= cwt_wakefn
;
3123 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
3124 TASK_UNINTERRUPTIBLE
);
3125 if (work_is_canceling(work
))
3127 finish_wait(&cancel_waitq
, &cwait
.wait
);
3129 } while (unlikely(ret
< 0));
3131 /* tell other tasks trying to grab @work to back off */
3132 mark_work_canceling(work
);
3133 local_irq_restore(flags
);
3136 * This allows canceling during early boot. We know that @work
3140 __flush_work(work
, true);
3142 clear_work_data(work
);
3145 * Paired with prepare_to_wait() above so that either
3146 * waitqueue_active() is visible here or !work_is_canceling() is
3150 if (waitqueue_active(&cancel_waitq
))
3151 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
3157 * cancel_work_sync - cancel a work and wait for it to finish
3158 * @work: the work to cancel
3160 * Cancel @work and wait for its execution to finish. This function
3161 * can be used even if the work re-queues itself or migrates to
3162 * another workqueue. On return from this function, @work is
3163 * guaranteed to be not pending or executing on any CPU.
3165 * cancel_work_sync(&delayed_work->work) must not be used for
3166 * delayed_work's. Use cancel_delayed_work_sync() instead.
3168 * The caller must ensure that the workqueue on which @work was last
3169 * queued can't be destroyed before this function returns.
3172 * %true if @work was pending, %false otherwise.
3174 bool cancel_work_sync(struct work_struct
*work
)
3176 return __cancel_work_timer(work
, false);
3178 EXPORT_SYMBOL_GPL(cancel_work_sync
);
3181 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3182 * @dwork: the delayed work to flush
3184 * Delayed timer is cancelled and the pending work is queued for
3185 * immediate execution. Like flush_work(), this function only
3186 * considers the last queueing instance of @dwork.
3189 * %true if flush_work() waited for the work to finish execution,
3190 * %false if it was already idle.
3192 bool flush_delayed_work(struct delayed_work
*dwork
)
3194 local_irq_disable();
3195 if (del_timer_sync(&dwork
->timer
))
3196 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3198 return flush_work(&dwork
->work
);
3200 EXPORT_SYMBOL(flush_delayed_work
);
3203 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3204 * @rwork: the rcu work to flush
3207 * %true if flush_rcu_work() waited for the work to finish execution,
3208 * %false if it was already idle.
3210 bool flush_rcu_work(struct rcu_work
*rwork
)
3212 if (test_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&rwork
->work
))) {
3214 flush_work(&rwork
->work
);
3217 return flush_work(&rwork
->work
);
3220 EXPORT_SYMBOL(flush_rcu_work
);
3222 static bool __cancel_work(struct work_struct
*work
, bool is_dwork
)
3224 unsigned long flags
;
3228 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3229 } while (unlikely(ret
== -EAGAIN
));
3231 if (unlikely(ret
< 0))
3234 set_work_pool_and_clear_pending(work
, get_work_pool_id(work
));
3235 local_irq_restore(flags
);
3240 * cancel_delayed_work - cancel a delayed work
3241 * @dwork: delayed_work to cancel
3243 * Kill off a pending delayed_work.
3245 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3249 * The work callback function may still be running on return, unless
3250 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3251 * use cancel_delayed_work_sync() to wait on it.
3253 * This function is safe to call from any context including IRQ handler.
3255 bool cancel_delayed_work(struct delayed_work
*dwork
)
3257 return __cancel_work(&dwork
->work
, true);
3259 EXPORT_SYMBOL(cancel_delayed_work
);
3262 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3263 * @dwork: the delayed work cancel
3265 * This is cancel_work_sync() for delayed works.
3268 * %true if @dwork was pending, %false otherwise.
3270 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3272 return __cancel_work_timer(&dwork
->work
, true);
3274 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3277 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3278 * @func: the function to call
3280 * schedule_on_each_cpu() executes @func on each online CPU using the
3281 * system workqueue and blocks until all CPUs have completed.
3282 * schedule_on_each_cpu() is very slow.
3285 * 0 on success, -errno on failure.
3287 int schedule_on_each_cpu(work_func_t func
)
3290 struct work_struct __percpu
*works
;
3292 works
= alloc_percpu(struct work_struct
);
3298 for_each_online_cpu(cpu
) {
3299 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3301 INIT_WORK(work
, func
);
3302 schedule_work_on(cpu
, work
);
3305 for_each_online_cpu(cpu
)
3306 flush_work(per_cpu_ptr(works
, cpu
));
3314 * execute_in_process_context - reliably execute the routine with user context
3315 * @fn: the function to execute
3316 * @ew: guaranteed storage for the execute work structure (must
3317 * be available when the work executes)
3319 * Executes the function immediately if process context is available,
3320 * otherwise schedules the function for delayed execution.
3322 * Return: 0 - function was executed
3323 * 1 - function was scheduled for execution
3325 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3327 if (!in_interrupt()) {
3332 INIT_WORK(&ew
->work
, fn
);
3333 schedule_work(&ew
->work
);
3337 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3340 * free_workqueue_attrs - free a workqueue_attrs
3341 * @attrs: workqueue_attrs to free
3343 * Undo alloc_workqueue_attrs().
3345 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3348 free_cpumask_var(attrs
->cpumask
);
3354 * alloc_workqueue_attrs - allocate a workqueue_attrs
3356 * Allocate a new workqueue_attrs, initialize with default settings and
3359 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3361 struct workqueue_attrs
*alloc_workqueue_attrs(void)
3363 struct workqueue_attrs
*attrs
;
3365 attrs
= kzalloc(sizeof(*attrs
), GFP_KERNEL
);
3368 if (!alloc_cpumask_var(&attrs
->cpumask
, GFP_KERNEL
))
3371 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3374 free_workqueue_attrs(attrs
);
3378 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3379 const struct workqueue_attrs
*from
)
3381 to
->nice
= from
->nice
;
3382 cpumask_copy(to
->cpumask
, from
->cpumask
);
3384 * Unlike hash and equality test, this function doesn't ignore
3385 * ->no_numa as it is used for both pool and wq attrs. Instead,
3386 * get_unbound_pool() explicitly clears ->no_numa after copying.
3388 to
->no_numa
= from
->no_numa
;
3391 /* hash value of the content of @attr */
3392 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3396 hash
= jhash_1word(attrs
->nice
, hash
);
3397 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3398 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3402 /* content equality test */
3403 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3404 const struct workqueue_attrs
*b
)
3406 if (a
->nice
!= b
->nice
)
3408 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3414 * init_worker_pool - initialize a newly zalloc'd worker_pool
3415 * @pool: worker_pool to initialize
3417 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3419 * Return: 0 on success, -errno on failure. Even on failure, all fields
3420 * inside @pool proper are initialized and put_unbound_pool() can be called
3421 * on @pool safely to release it.
3423 static int init_worker_pool(struct worker_pool
*pool
)
3425 raw_spin_lock_init(&pool
->lock
);
3428 pool
->node
= NUMA_NO_NODE
;
3429 pool
->flags
|= POOL_DISASSOCIATED
;
3430 pool
->watchdog_ts
= jiffies
;
3431 INIT_LIST_HEAD(&pool
->worklist
);
3432 INIT_LIST_HEAD(&pool
->idle_list
);
3433 hash_init(pool
->busy_hash
);
3435 timer_setup(&pool
->idle_timer
, idle_worker_timeout
, TIMER_DEFERRABLE
);
3437 timer_setup(&pool
->mayday_timer
, pool_mayday_timeout
, 0);
3439 INIT_LIST_HEAD(&pool
->workers
);
3441 ida_init(&pool
->worker_ida
);
3442 INIT_HLIST_NODE(&pool
->hash_node
);
3445 /* shouldn't fail above this point */
3446 pool
->attrs
= alloc_workqueue_attrs();
3452 #ifdef CONFIG_LOCKDEP
3453 static void wq_init_lockdep(struct workqueue_struct
*wq
)
3457 lockdep_register_key(&wq
->key
);
3458 lock_name
= kasprintf(GFP_KERNEL
, "%s%s", "(wq_completion)", wq
->name
);
3460 lock_name
= wq
->name
;
3462 wq
->lock_name
= lock_name
;
3463 lockdep_init_map(&wq
->lockdep_map
, lock_name
, &wq
->key
, 0);
3466 static void wq_unregister_lockdep(struct workqueue_struct
*wq
)
3468 lockdep_unregister_key(&wq
->key
);
3471 static void wq_free_lockdep(struct workqueue_struct
*wq
)
3473 if (wq
->lock_name
!= wq
->name
)
3474 kfree(wq
->lock_name
);
3477 static void wq_init_lockdep(struct workqueue_struct
*wq
)
3481 static void wq_unregister_lockdep(struct workqueue_struct
*wq
)
3485 static void wq_free_lockdep(struct workqueue_struct
*wq
)
3490 static void rcu_free_wq(struct rcu_head
*rcu
)
3492 struct workqueue_struct
*wq
=
3493 container_of(rcu
, struct workqueue_struct
, rcu
);
3495 wq_free_lockdep(wq
);
3497 if (!(wq
->flags
& WQ_UNBOUND
))
3498 free_percpu(wq
->cpu_pwqs
);
3500 free_workqueue_attrs(wq
->unbound_attrs
);
3505 static void rcu_free_pool(struct rcu_head
*rcu
)
3507 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3509 ida_destroy(&pool
->worker_ida
);
3510 free_workqueue_attrs(pool
->attrs
);
3514 /* This returns with the lock held on success (pool manager is inactive). */
3515 static bool wq_manager_inactive(struct worker_pool
*pool
)
3517 raw_spin_lock_irq(&pool
->lock
);
3519 if (pool
->flags
& POOL_MANAGER_ACTIVE
) {
3520 raw_spin_unlock_irq(&pool
->lock
);
3527 * put_unbound_pool - put a worker_pool
3528 * @pool: worker_pool to put
3530 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3531 * safe manner. get_unbound_pool() calls this function on its failure path
3532 * and this function should be able to release pools which went through,
3533 * successfully or not, init_worker_pool().
3535 * Should be called with wq_pool_mutex held.
3537 static void put_unbound_pool(struct worker_pool
*pool
)
3539 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3540 struct worker
*worker
;
3542 lockdep_assert_held(&wq_pool_mutex
);
3548 if (WARN_ON(!(pool
->cpu
< 0)) ||
3549 WARN_ON(!list_empty(&pool
->worklist
)))
3552 /* release id and unhash */
3554 idr_remove(&worker_pool_idr
, pool
->id
);
3555 hash_del(&pool
->hash_node
);
3558 * Become the manager and destroy all workers. This prevents
3559 * @pool's workers from blocking on attach_mutex. We're the last
3560 * manager and @pool gets freed with the flag set.
3561 * Because of how wq_manager_inactive() works, we will hold the
3562 * spinlock after a successful wait.
3564 rcuwait_wait_event(&manager_wait
, wq_manager_inactive(pool
),
3565 TASK_UNINTERRUPTIBLE
);
3566 pool
->flags
|= POOL_MANAGER_ACTIVE
;
3568 while ((worker
= first_idle_worker(pool
)))
3569 destroy_worker(worker
);
3570 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3571 raw_spin_unlock_irq(&pool
->lock
);
3573 mutex_lock(&wq_pool_attach_mutex
);
3574 if (!list_empty(&pool
->workers
))
3575 pool
->detach_completion
= &detach_completion
;
3576 mutex_unlock(&wq_pool_attach_mutex
);
3578 if (pool
->detach_completion
)
3579 wait_for_completion(pool
->detach_completion
);
3581 /* shut down the timers */
3582 del_timer_sync(&pool
->idle_timer
);
3583 del_timer_sync(&pool
->mayday_timer
);
3585 /* RCU protected to allow dereferences from get_work_pool() */
3586 call_rcu(&pool
->rcu
, rcu_free_pool
);
3590 * get_unbound_pool - get a worker_pool with the specified attributes
3591 * @attrs: the attributes of the worker_pool to get
3593 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3594 * reference count and return it. If there already is a matching
3595 * worker_pool, it will be used; otherwise, this function attempts to
3598 * Should be called with wq_pool_mutex held.
3600 * Return: On success, a worker_pool with the same attributes as @attrs.
3601 * On failure, %NULL.
3603 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3605 u32 hash
= wqattrs_hash(attrs
);
3606 struct worker_pool
*pool
;
3608 int target_node
= NUMA_NO_NODE
;
3610 lockdep_assert_held(&wq_pool_mutex
);
3612 /* do we already have a matching pool? */
3613 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3614 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3620 /* if cpumask is contained inside a NUMA node, we belong to that node */
3621 if (wq_numa_enabled
) {
3622 for_each_node(node
) {
3623 if (cpumask_subset(attrs
->cpumask
,
3624 wq_numa_possible_cpumask
[node
])) {
3631 /* nope, create a new one */
3632 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3633 if (!pool
|| init_worker_pool(pool
) < 0)
3636 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3637 copy_workqueue_attrs(pool
->attrs
, attrs
);
3638 pool
->node
= target_node
;
3641 * no_numa isn't a worker_pool attribute, always clear it. See
3642 * 'struct workqueue_attrs' comments for detail.
3644 pool
->attrs
->no_numa
= false;
3646 if (worker_pool_assign_id(pool
) < 0)
3649 /* create and start the initial worker */
3650 if (wq_online
&& !create_worker(pool
))
3654 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3659 put_unbound_pool(pool
);
3663 static void rcu_free_pwq(struct rcu_head
*rcu
)
3665 kmem_cache_free(pwq_cache
,
3666 container_of(rcu
, struct pool_workqueue
, rcu
));
3670 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3671 * and needs to be destroyed.
3673 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3675 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3676 unbound_release_work
);
3677 struct workqueue_struct
*wq
= pwq
->wq
;
3678 struct worker_pool
*pool
= pwq
->pool
;
3681 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3684 mutex_lock(&wq
->mutex
);
3685 list_del_rcu(&pwq
->pwqs_node
);
3686 is_last
= list_empty(&wq
->pwqs
);
3687 mutex_unlock(&wq
->mutex
);
3689 mutex_lock(&wq_pool_mutex
);
3690 put_unbound_pool(pool
);
3691 mutex_unlock(&wq_pool_mutex
);
3693 call_rcu(&pwq
->rcu
, rcu_free_pwq
);
3696 * If we're the last pwq going away, @wq is already dead and no one
3697 * is gonna access it anymore. Schedule RCU free.
3700 wq_unregister_lockdep(wq
);
3701 call_rcu(&wq
->rcu
, rcu_free_wq
);
3706 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3707 * @pwq: target pool_workqueue
3709 * If @pwq isn't freezing, set @pwq->max_active to the associated
3710 * workqueue's saved_max_active and activate delayed work items
3711 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3713 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3715 struct workqueue_struct
*wq
= pwq
->wq
;
3716 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3717 unsigned long flags
;
3719 /* for @wq->saved_max_active */
3720 lockdep_assert_held(&wq
->mutex
);
3722 /* fast exit for non-freezable wqs */
3723 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3726 /* this function can be called during early boot w/ irq disabled */
3727 raw_spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
3730 * During [un]freezing, the caller is responsible for ensuring that
3731 * this function is called at least once after @workqueue_freezing
3732 * is updated and visible.
3734 if (!freezable
|| !workqueue_freezing
) {
3737 pwq
->max_active
= wq
->saved_max_active
;
3739 while (!list_empty(&pwq
->delayed_works
) &&
3740 pwq
->nr_active
< pwq
->max_active
) {
3741 pwq_activate_first_delayed(pwq
);
3746 * Need to kick a worker after thawed or an unbound wq's
3747 * max_active is bumped. In realtime scenarios, always kicking a
3748 * worker will cause interference on the isolated cpu cores, so
3749 * let's kick iff work items were activated.
3752 wake_up_worker(pwq
->pool
);
3754 pwq
->max_active
= 0;
3757 raw_spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
3760 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3761 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3762 struct worker_pool
*pool
)
3764 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3766 memset(pwq
, 0, sizeof(*pwq
));
3770 pwq
->flush_color
= -1;
3772 INIT_LIST_HEAD(&pwq
->delayed_works
);
3773 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3774 INIT_LIST_HEAD(&pwq
->mayday_node
);
3775 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3778 /* sync @pwq with the current state of its associated wq and link it */
3779 static void link_pwq(struct pool_workqueue
*pwq
)
3781 struct workqueue_struct
*wq
= pwq
->wq
;
3783 lockdep_assert_held(&wq
->mutex
);
3785 /* may be called multiple times, ignore if already linked */
3786 if (!list_empty(&pwq
->pwqs_node
))
3789 /* set the matching work_color */
3790 pwq
->work_color
= wq
->work_color
;
3792 /* sync max_active to the current setting */
3793 pwq_adjust_max_active(pwq
);
3796 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3799 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3800 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3801 const struct workqueue_attrs
*attrs
)
3803 struct worker_pool
*pool
;
3804 struct pool_workqueue
*pwq
;
3806 lockdep_assert_held(&wq_pool_mutex
);
3808 pool
= get_unbound_pool(attrs
);
3812 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3814 put_unbound_pool(pool
);
3818 init_pwq(pwq
, wq
, pool
);
3823 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3824 * @attrs: the wq_attrs of the default pwq of the target workqueue
3825 * @node: the target NUMA node
3826 * @cpu_going_down: if >= 0, the CPU to consider as offline
3827 * @cpumask: outarg, the resulting cpumask
3829 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3830 * @cpu_going_down is >= 0, that cpu is considered offline during
3831 * calculation. The result is stored in @cpumask.
3833 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3834 * enabled and @node has online CPUs requested by @attrs, the returned
3835 * cpumask is the intersection of the possible CPUs of @node and
3838 * The caller is responsible for ensuring that the cpumask of @node stays
3841 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3844 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3845 int cpu_going_down
, cpumask_t
*cpumask
)
3847 if (!wq_numa_enabled
|| attrs
->no_numa
)
3850 /* does @node have any online CPUs @attrs wants? */
3851 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3852 if (cpu_going_down
>= 0)
3853 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3855 if (cpumask_empty(cpumask
))
3858 /* yeap, return possible CPUs in @node that @attrs wants */
3859 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3861 if (cpumask_empty(cpumask
)) {
3862 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3863 "possible intersect\n");
3867 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3870 cpumask_copy(cpumask
, attrs
->cpumask
);
3874 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3875 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3877 struct pool_workqueue
*pwq
)
3879 struct pool_workqueue
*old_pwq
;
3881 lockdep_assert_held(&wq_pool_mutex
);
3882 lockdep_assert_held(&wq
->mutex
);
3884 /* link_pwq() can handle duplicate calls */
3887 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3888 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3892 /* context to store the prepared attrs & pwqs before applying */
3893 struct apply_wqattrs_ctx
{
3894 struct workqueue_struct
*wq
; /* target workqueue */
3895 struct workqueue_attrs
*attrs
; /* attrs to apply */
3896 struct list_head list
; /* queued for batching commit */
3897 struct pool_workqueue
*dfl_pwq
;
3898 struct pool_workqueue
*pwq_tbl
[];
3901 /* free the resources after success or abort */
3902 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3908 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3909 put_pwq_unlocked(ctx
->dfl_pwq
);
3911 free_workqueue_attrs(ctx
->attrs
);
3917 /* allocate the attrs and pwqs for later installation */
3918 static struct apply_wqattrs_ctx
*
3919 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3920 const struct workqueue_attrs
*attrs
)
3922 struct apply_wqattrs_ctx
*ctx
;
3923 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3926 lockdep_assert_held(&wq_pool_mutex
);
3928 ctx
= kzalloc(struct_size(ctx
, pwq_tbl
, nr_node_ids
), GFP_KERNEL
);
3930 new_attrs
= alloc_workqueue_attrs();
3931 tmp_attrs
= alloc_workqueue_attrs();
3932 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3936 * Calculate the attrs of the default pwq.
3937 * If the user configured cpumask doesn't overlap with the
3938 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3940 copy_workqueue_attrs(new_attrs
, attrs
);
3941 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3942 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3943 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3946 * We may create multiple pwqs with differing cpumasks. Make a
3947 * copy of @new_attrs which will be modified and used to obtain
3950 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3953 * If something goes wrong during CPU up/down, we'll fall back to
3954 * the default pwq covering whole @attrs->cpumask. Always create
3955 * it even if we don't use it immediately.
3957 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3961 for_each_node(node
) {
3962 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3963 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3964 if (!ctx
->pwq_tbl
[node
])
3967 ctx
->dfl_pwq
->refcnt
++;
3968 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3972 /* save the user configured attrs and sanitize it. */
3973 copy_workqueue_attrs(new_attrs
, attrs
);
3974 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3975 ctx
->attrs
= new_attrs
;
3978 free_workqueue_attrs(tmp_attrs
);
3982 free_workqueue_attrs(tmp_attrs
);
3983 free_workqueue_attrs(new_attrs
);
3984 apply_wqattrs_cleanup(ctx
);
3988 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3989 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3993 /* all pwqs have been created successfully, let's install'em */
3994 mutex_lock(&ctx
->wq
->mutex
);
3996 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3998 /* save the previous pwq and install the new one */
4000 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
4001 ctx
->pwq_tbl
[node
]);
4003 /* @dfl_pwq might not have been used, ensure it's linked */
4004 link_pwq(ctx
->dfl_pwq
);
4005 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
4007 mutex_unlock(&ctx
->wq
->mutex
);
4010 static void apply_wqattrs_lock(void)
4012 /* CPUs should stay stable across pwq creations and installations */
4014 mutex_lock(&wq_pool_mutex
);
4017 static void apply_wqattrs_unlock(void)
4019 mutex_unlock(&wq_pool_mutex
);
4023 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
4024 const struct workqueue_attrs
*attrs
)
4026 struct apply_wqattrs_ctx
*ctx
;
4028 /* only unbound workqueues can change attributes */
4029 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
4032 /* creating multiple pwqs breaks ordering guarantee */
4033 if (!list_empty(&wq
->pwqs
)) {
4034 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4037 wq
->flags
&= ~__WQ_ORDERED
;
4040 ctx
= apply_wqattrs_prepare(wq
, attrs
);
4044 /* the ctx has been prepared successfully, let's commit it */
4045 apply_wqattrs_commit(ctx
);
4046 apply_wqattrs_cleanup(ctx
);
4052 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4053 * @wq: the target workqueue
4054 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4056 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4057 * machines, this function maps a separate pwq to each NUMA node with
4058 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4059 * NUMA node it was issued on. Older pwqs are released as in-flight work
4060 * items finish. Note that a work item which repeatedly requeues itself
4061 * back-to-back will stay on its current pwq.
4063 * Performs GFP_KERNEL allocations.
4065 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4067 * Return: 0 on success and -errno on failure.
4069 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
4070 const struct workqueue_attrs
*attrs
)
4074 lockdep_assert_cpus_held();
4076 mutex_lock(&wq_pool_mutex
);
4077 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
4078 mutex_unlock(&wq_pool_mutex
);
4084 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4085 * @wq: the target workqueue
4086 * @cpu: the CPU coming up or going down
4087 * @online: whether @cpu is coming up or going down
4089 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4090 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4093 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4094 * falls back to @wq->dfl_pwq which may not be optimal but is always
4097 * Note that when the last allowed CPU of a NUMA node goes offline for a
4098 * workqueue with a cpumask spanning multiple nodes, the workers which were
4099 * already executing the work items for the workqueue will lose their CPU
4100 * affinity and may execute on any CPU. This is similar to how per-cpu
4101 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4102 * affinity, it's the user's responsibility to flush the work item from
4105 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
4108 int node
= cpu_to_node(cpu
);
4109 int cpu_off
= online
? -1 : cpu
;
4110 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
4111 struct workqueue_attrs
*target_attrs
;
4114 lockdep_assert_held(&wq_pool_mutex
);
4116 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
4117 wq
->unbound_attrs
->no_numa
)
4121 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4122 * Let's use a preallocated one. The following buf is protected by
4123 * CPU hotplug exclusion.
4125 target_attrs
= wq_update_unbound_numa_attrs_buf
;
4126 cpumask
= target_attrs
->cpumask
;
4128 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4129 pwq
= unbound_pwq_by_node(wq
, node
);
4132 * Let's determine what needs to be done. If the target cpumask is
4133 * different from the default pwq's, we need to compare it to @pwq's
4134 * and create a new one if they don't match. If the target cpumask
4135 * equals the default pwq's, the default pwq should be used.
4137 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
4138 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4144 /* create a new pwq */
4145 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4147 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4152 /* Install the new pwq. */
4153 mutex_lock(&wq
->mutex
);
4154 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4158 mutex_lock(&wq
->mutex
);
4159 raw_spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4160 get_pwq(wq
->dfl_pwq
);
4161 raw_spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4162 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4164 mutex_unlock(&wq
->mutex
);
4165 put_pwq_unlocked(old_pwq
);
4168 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4170 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4173 if (!(wq
->flags
& WQ_UNBOUND
)) {
4174 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4178 for_each_possible_cpu(cpu
) {
4179 struct pool_workqueue
*pwq
=
4180 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4181 struct worker_pool
*cpu_pools
=
4182 per_cpu(cpu_worker_pools
, cpu
);
4184 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4186 mutex_lock(&wq
->mutex
);
4188 mutex_unlock(&wq
->mutex
);
4194 if (wq
->flags
& __WQ_ORDERED
) {
4195 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4196 /* there should only be single pwq for ordering guarantee */
4197 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4198 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4199 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4201 ret
= apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4208 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4211 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4213 if (max_active
< 1 || max_active
> lim
)
4214 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4215 max_active
, name
, 1, lim
);
4217 return clamp_val(max_active
, 1, lim
);
4221 * Workqueues which may be used during memory reclaim should have a rescuer
4222 * to guarantee forward progress.
4224 static int init_rescuer(struct workqueue_struct
*wq
)
4226 struct worker
*rescuer
;
4229 if (!(wq
->flags
& WQ_MEM_RECLAIM
))
4232 rescuer
= alloc_worker(NUMA_NO_NODE
);
4236 rescuer
->rescue_wq
= wq
;
4237 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s", wq
->name
);
4238 if (IS_ERR(rescuer
->task
)) {
4239 ret
= PTR_ERR(rescuer
->task
);
4244 wq
->rescuer
= rescuer
;
4245 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
4246 wake_up_process(rescuer
->task
);
4252 struct workqueue_struct
*alloc_workqueue(const char *fmt
,
4254 int max_active
, ...)
4256 size_t tbl_size
= 0;
4258 struct workqueue_struct
*wq
;
4259 struct pool_workqueue
*pwq
;
4262 * Unbound && max_active == 1 used to imply ordered, which is no
4263 * longer the case on NUMA machines due to per-node pools. While
4264 * alloc_ordered_workqueue() is the right way to create an ordered
4265 * workqueue, keep the previous behavior to avoid subtle breakages
4268 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
4269 flags
|= __WQ_ORDERED
;
4271 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4272 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4273 flags
|= WQ_UNBOUND
;
4275 /* allocate wq and format name */
4276 if (flags
& WQ_UNBOUND
)
4277 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
4279 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4283 if (flags
& WQ_UNBOUND
) {
4284 wq
->unbound_attrs
= alloc_workqueue_attrs();
4285 if (!wq
->unbound_attrs
)
4289 va_start(args
, max_active
);
4290 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4293 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4294 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4298 wq
->saved_max_active
= max_active
;
4299 mutex_init(&wq
->mutex
);
4300 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4301 INIT_LIST_HEAD(&wq
->pwqs
);
4302 INIT_LIST_HEAD(&wq
->flusher_queue
);
4303 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4304 INIT_LIST_HEAD(&wq
->maydays
);
4306 wq_init_lockdep(wq
);
4307 INIT_LIST_HEAD(&wq
->list
);
4309 if (alloc_and_link_pwqs(wq
) < 0)
4310 goto err_unreg_lockdep
;
4312 if (wq_online
&& init_rescuer(wq
) < 0)
4315 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4319 * wq_pool_mutex protects global freeze state and workqueues list.
4320 * Grab it, adjust max_active and add the new @wq to workqueues
4323 mutex_lock(&wq_pool_mutex
);
4325 mutex_lock(&wq
->mutex
);
4326 for_each_pwq(pwq
, wq
)
4327 pwq_adjust_max_active(pwq
);
4328 mutex_unlock(&wq
->mutex
);
4330 list_add_tail_rcu(&wq
->list
, &workqueues
);
4332 mutex_unlock(&wq_pool_mutex
);
4337 wq_unregister_lockdep(wq
);
4338 wq_free_lockdep(wq
);
4340 free_workqueue_attrs(wq
->unbound_attrs
);
4344 destroy_workqueue(wq
);
4347 EXPORT_SYMBOL_GPL(alloc_workqueue
);
4349 static bool pwq_busy(struct pool_workqueue
*pwq
)
4353 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
4354 if (pwq
->nr_in_flight
[i
])
4357 if ((pwq
!= pwq
->wq
->dfl_pwq
) && (pwq
->refcnt
> 1))
4359 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4366 * destroy_workqueue - safely terminate a workqueue
4367 * @wq: target workqueue
4369 * Safely destroy a workqueue. All work currently pending will be done first.
4371 void destroy_workqueue(struct workqueue_struct
*wq
)
4373 struct pool_workqueue
*pwq
;
4377 * Remove it from sysfs first so that sanity check failure doesn't
4378 * lead to sysfs name conflicts.
4380 workqueue_sysfs_unregister(wq
);
4382 /* drain it before proceeding with destruction */
4383 drain_workqueue(wq
);
4385 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4387 struct worker
*rescuer
= wq
->rescuer
;
4389 /* this prevents new queueing */
4390 raw_spin_lock_irq(&wq_mayday_lock
);
4392 raw_spin_unlock_irq(&wq_mayday_lock
);
4394 /* rescuer will empty maydays list before exiting */
4395 kthread_stop(rescuer
->task
);
4400 * Sanity checks - grab all the locks so that we wait for all
4401 * in-flight operations which may do put_pwq().
4403 mutex_lock(&wq_pool_mutex
);
4404 mutex_lock(&wq
->mutex
);
4405 for_each_pwq(pwq
, wq
) {
4406 raw_spin_lock_irq(&pwq
->pool
->lock
);
4407 if (WARN_ON(pwq_busy(pwq
))) {
4408 pr_warn("%s: %s has the following busy pwq\n",
4409 __func__
, wq
->name
);
4411 raw_spin_unlock_irq(&pwq
->pool
->lock
);
4412 mutex_unlock(&wq
->mutex
);
4413 mutex_unlock(&wq_pool_mutex
);
4414 show_workqueue_state();
4417 raw_spin_unlock_irq(&pwq
->pool
->lock
);
4419 mutex_unlock(&wq
->mutex
);
4422 * wq list is used to freeze wq, remove from list after
4423 * flushing is complete in case freeze races us.
4425 list_del_rcu(&wq
->list
);
4426 mutex_unlock(&wq_pool_mutex
);
4428 if (!(wq
->flags
& WQ_UNBOUND
)) {
4429 wq_unregister_lockdep(wq
);
4431 * The base ref is never dropped on per-cpu pwqs. Directly
4432 * schedule RCU free.
4434 call_rcu(&wq
->rcu
, rcu_free_wq
);
4437 * We're the sole accessor of @wq at this point. Directly
4438 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4439 * @wq will be freed when the last pwq is released.
4441 for_each_node(node
) {
4442 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4443 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4444 put_pwq_unlocked(pwq
);
4448 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4449 * put. Don't access it afterwards.
4453 put_pwq_unlocked(pwq
);
4456 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4459 * workqueue_set_max_active - adjust max_active of a workqueue
4460 * @wq: target workqueue
4461 * @max_active: new max_active value.
4463 * Set max_active of @wq to @max_active.
4466 * Don't call from IRQ context.
4468 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4470 struct pool_workqueue
*pwq
;
4472 /* disallow meddling with max_active for ordered workqueues */
4473 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4476 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4478 mutex_lock(&wq
->mutex
);
4480 wq
->flags
&= ~__WQ_ORDERED
;
4481 wq
->saved_max_active
= max_active
;
4483 for_each_pwq(pwq
, wq
)
4484 pwq_adjust_max_active(pwq
);
4486 mutex_unlock(&wq
->mutex
);
4488 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4491 * current_work - retrieve %current task's work struct
4493 * Determine if %current task is a workqueue worker and what it's working on.
4494 * Useful to find out the context that the %current task is running in.
4496 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4498 struct work_struct
*current_work(void)
4500 struct worker
*worker
= current_wq_worker();
4502 return worker
? worker
->current_work
: NULL
;
4504 EXPORT_SYMBOL(current_work
);
4507 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4509 * Determine whether %current is a workqueue rescuer. Can be used from
4510 * work functions to determine whether it's being run off the rescuer task.
4512 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4514 bool current_is_workqueue_rescuer(void)
4516 struct worker
*worker
= current_wq_worker();
4518 return worker
&& worker
->rescue_wq
;
4522 * workqueue_congested - test whether a workqueue is congested
4523 * @cpu: CPU in question
4524 * @wq: target workqueue
4526 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4527 * no synchronization around this function and the test result is
4528 * unreliable and only useful as advisory hints or for debugging.
4530 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4531 * Note that both per-cpu and unbound workqueues may be associated with
4532 * multiple pool_workqueues which have separate congested states. A
4533 * workqueue being congested on one CPU doesn't mean the workqueue is also
4534 * contested on other CPUs / NUMA nodes.
4537 * %true if congested, %false otherwise.
4539 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4541 struct pool_workqueue
*pwq
;
4547 if (cpu
== WORK_CPU_UNBOUND
)
4548 cpu
= smp_processor_id();
4550 if (!(wq
->flags
& WQ_UNBOUND
))
4551 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4553 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4555 ret
= !list_empty(&pwq
->delayed_works
);
4561 EXPORT_SYMBOL_GPL(workqueue_congested
);
4564 * work_busy - test whether a work is currently pending or running
4565 * @work: the work to be tested
4567 * Test whether @work is currently pending or running. There is no
4568 * synchronization around this function and the test result is
4569 * unreliable and only useful as advisory hints or for debugging.
4572 * OR'd bitmask of WORK_BUSY_* bits.
4574 unsigned int work_busy(struct work_struct
*work
)
4576 struct worker_pool
*pool
;
4577 unsigned long flags
;
4578 unsigned int ret
= 0;
4580 if (work_pending(work
))
4581 ret
|= WORK_BUSY_PENDING
;
4584 pool
= get_work_pool(work
);
4586 raw_spin_lock_irqsave(&pool
->lock
, flags
);
4587 if (find_worker_executing_work(pool
, work
))
4588 ret
|= WORK_BUSY_RUNNING
;
4589 raw_spin_unlock_irqrestore(&pool
->lock
, flags
);
4595 EXPORT_SYMBOL_GPL(work_busy
);
4598 * set_worker_desc - set description for the current work item
4599 * @fmt: printf-style format string
4600 * @...: arguments for the format string
4602 * This function can be called by a running work function to describe what
4603 * the work item is about. If the worker task gets dumped, this
4604 * information will be printed out together to help debugging. The
4605 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4607 void set_worker_desc(const char *fmt
, ...)
4609 struct worker
*worker
= current_wq_worker();
4613 va_start(args
, fmt
);
4614 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4618 EXPORT_SYMBOL_GPL(set_worker_desc
);
4621 * print_worker_info - print out worker information and description
4622 * @log_lvl: the log level to use when printing
4623 * @task: target task
4625 * If @task is a worker and currently executing a work item, print out the
4626 * name of the workqueue being serviced and worker description set with
4627 * set_worker_desc() by the currently executing work item.
4629 * This function can be safely called on any task as long as the
4630 * task_struct itself is accessible. While safe, this function isn't
4631 * synchronized and may print out mixups or garbages of limited length.
4633 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4635 work_func_t
*fn
= NULL
;
4636 char name
[WQ_NAME_LEN
] = { };
4637 char desc
[WORKER_DESC_LEN
] = { };
4638 struct pool_workqueue
*pwq
= NULL
;
4639 struct workqueue_struct
*wq
= NULL
;
4640 struct worker
*worker
;
4642 if (!(task
->flags
& PF_WQ_WORKER
))
4646 * This function is called without any synchronization and @task
4647 * could be in any state. Be careful with dereferences.
4649 worker
= kthread_probe_data(task
);
4652 * Carefully copy the associated workqueue's workfn, name and desc.
4653 * Keep the original last '\0' in case the original is garbage.
4655 copy_from_kernel_nofault(&fn
, &worker
->current_func
, sizeof(fn
));
4656 copy_from_kernel_nofault(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4657 copy_from_kernel_nofault(&wq
, &pwq
->wq
, sizeof(wq
));
4658 copy_from_kernel_nofault(name
, wq
->name
, sizeof(name
) - 1);
4659 copy_from_kernel_nofault(desc
, worker
->desc
, sizeof(desc
) - 1);
4661 if (fn
|| name
[0] || desc
[0]) {
4662 printk("%sWorkqueue: %s %ps", log_lvl
, name
, fn
);
4663 if (strcmp(name
, desc
))
4664 pr_cont(" (%s)", desc
);
4669 static void pr_cont_pool_info(struct worker_pool
*pool
)
4671 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4672 if (pool
->node
!= NUMA_NO_NODE
)
4673 pr_cont(" node=%d", pool
->node
);
4674 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4677 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4679 if (work
->func
== wq_barrier_func
) {
4680 struct wq_barrier
*barr
;
4682 barr
= container_of(work
, struct wq_barrier
, work
);
4684 pr_cont("%s BAR(%d)", comma
? "," : "",
4685 task_pid_nr(barr
->task
));
4687 pr_cont("%s %ps", comma
? "," : "", work
->func
);
4691 static void show_pwq(struct pool_workqueue
*pwq
)
4693 struct worker_pool
*pool
= pwq
->pool
;
4694 struct work_struct
*work
;
4695 struct worker
*worker
;
4696 bool has_in_flight
= false, has_pending
= false;
4699 pr_info(" pwq %d:", pool
->id
);
4700 pr_cont_pool_info(pool
);
4702 pr_cont(" active=%d/%d refcnt=%d%s\n",
4703 pwq
->nr_active
, pwq
->max_active
, pwq
->refcnt
,
4704 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4706 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4707 if (worker
->current_pwq
== pwq
) {
4708 has_in_flight
= true;
4712 if (has_in_flight
) {
4715 pr_info(" in-flight:");
4716 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4717 if (worker
->current_pwq
!= pwq
)
4720 pr_cont("%s %d%s:%ps", comma
? "," : "",
4721 task_pid_nr(worker
->task
),
4722 worker
->rescue_wq
? "(RESCUER)" : "",
4723 worker
->current_func
);
4724 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4725 pr_cont_work(false, work
);
4731 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4732 if (get_work_pwq(work
) == pwq
) {
4740 pr_info(" pending:");
4741 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4742 if (get_work_pwq(work
) != pwq
)
4745 pr_cont_work(comma
, work
);
4746 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4751 if (!list_empty(&pwq
->delayed_works
)) {
4754 pr_info(" delayed:");
4755 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4756 pr_cont_work(comma
, work
);
4757 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4764 * show_workqueue_state - dump workqueue state
4766 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4767 * all busy workqueues and pools.
4769 void show_workqueue_state(void)
4771 struct workqueue_struct
*wq
;
4772 struct worker_pool
*pool
;
4773 unsigned long flags
;
4778 pr_info("Showing busy workqueues and worker pools:\n");
4780 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4781 struct pool_workqueue
*pwq
;
4784 for_each_pwq(pwq
, wq
) {
4785 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4793 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4795 for_each_pwq(pwq
, wq
) {
4796 raw_spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4797 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4799 raw_spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4801 * We could be printing a lot from atomic context, e.g.
4802 * sysrq-t -> show_workqueue_state(). Avoid triggering
4805 touch_nmi_watchdog();
4809 for_each_pool(pool
, pi
) {
4810 struct worker
*worker
;
4813 raw_spin_lock_irqsave(&pool
->lock
, flags
);
4814 if (pool
->nr_workers
== pool
->nr_idle
)
4817 pr_info("pool %d:", pool
->id
);
4818 pr_cont_pool_info(pool
);
4819 pr_cont(" hung=%us workers=%d",
4820 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4823 pr_cont(" manager: %d",
4824 task_pid_nr(pool
->manager
->task
));
4825 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4826 pr_cont(" %s%d", first
? "idle: " : "",
4827 task_pid_nr(worker
->task
));
4832 raw_spin_unlock_irqrestore(&pool
->lock
, flags
);
4834 * We could be printing a lot from atomic context, e.g.
4835 * sysrq-t -> show_workqueue_state(). Avoid triggering
4838 touch_nmi_watchdog();
4844 /* used to show worker information through /proc/PID/{comm,stat,status} */
4845 void wq_worker_comm(char *buf
, size_t size
, struct task_struct
*task
)
4849 /* always show the actual comm */
4850 off
= strscpy(buf
, task
->comm
, size
);
4854 /* stabilize PF_WQ_WORKER and worker pool association */
4855 mutex_lock(&wq_pool_attach_mutex
);
4857 if (task
->flags
& PF_WQ_WORKER
) {
4858 struct worker
*worker
= kthread_data(task
);
4859 struct worker_pool
*pool
= worker
->pool
;
4862 raw_spin_lock_irq(&pool
->lock
);
4864 * ->desc tracks information (wq name or
4865 * set_worker_desc()) for the latest execution. If
4866 * current, prepend '+', otherwise '-'.
4868 if (worker
->desc
[0] != '\0') {
4869 if (worker
->current_work
)
4870 scnprintf(buf
+ off
, size
- off
, "+%s",
4873 scnprintf(buf
+ off
, size
- off
, "-%s",
4876 raw_spin_unlock_irq(&pool
->lock
);
4880 mutex_unlock(&wq_pool_attach_mutex
);
4888 * There are two challenges in supporting CPU hotplug. Firstly, there
4889 * are a lot of assumptions on strong associations among work, pwq and
4890 * pool which make migrating pending and scheduled works very
4891 * difficult to implement without impacting hot paths. Secondly,
4892 * worker pools serve mix of short, long and very long running works making
4893 * blocked draining impractical.
4895 * This is solved by allowing the pools to be disassociated from the CPU
4896 * running as an unbound one and allowing it to be reattached later if the
4897 * cpu comes back online.
4900 static void unbind_workers(int cpu
)
4902 struct worker_pool
*pool
;
4903 struct worker
*worker
;
4905 for_each_cpu_worker_pool(pool
, cpu
) {
4906 mutex_lock(&wq_pool_attach_mutex
);
4907 raw_spin_lock_irq(&pool
->lock
);
4910 * We've blocked all attach/detach operations. Make all workers
4911 * unbound and set DISASSOCIATED. Before this, all workers
4912 * except for the ones which are still executing works from
4913 * before the last CPU down must be on the cpu. After
4914 * this, they may become diasporas.
4916 for_each_pool_worker(worker
, pool
)
4917 worker
->flags
|= WORKER_UNBOUND
;
4919 pool
->flags
|= POOL_DISASSOCIATED
;
4921 raw_spin_unlock_irq(&pool
->lock
);
4923 for_each_pool_worker(worker
, pool
) {
4924 kthread_set_per_cpu(worker
->task
, -1);
4925 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, cpu_possible_mask
) < 0);
4928 mutex_unlock(&wq_pool_attach_mutex
);
4931 * Call schedule() so that we cross rq->lock and thus can
4932 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4933 * This is necessary as scheduler callbacks may be invoked
4939 * Sched callbacks are disabled now. Zap nr_running.
4940 * After this, nr_running stays zero and need_more_worker()
4941 * and keep_working() are always true as long as the
4942 * worklist is not empty. This pool now behaves as an
4943 * unbound (in terms of concurrency management) pool which
4944 * are served by workers tied to the pool.
4946 atomic_set(&pool
->nr_running
, 0);
4949 * With concurrency management just turned off, a busy
4950 * worker blocking could lead to lengthy stalls. Kick off
4951 * unbound chain execution of currently pending work items.
4953 raw_spin_lock_irq(&pool
->lock
);
4954 wake_up_worker(pool
);
4955 raw_spin_unlock_irq(&pool
->lock
);
4960 * rebind_workers - rebind all workers of a pool to the associated CPU
4961 * @pool: pool of interest
4963 * @pool->cpu is coming online. Rebind all workers to the CPU.
4965 static void rebind_workers(struct worker_pool
*pool
)
4967 struct worker
*worker
;
4969 lockdep_assert_held(&wq_pool_attach_mutex
);
4972 * Restore CPU affinity of all workers. As all idle workers should
4973 * be on the run-queue of the associated CPU before any local
4974 * wake-ups for concurrency management happen, restore CPU affinity
4975 * of all workers first and then clear UNBOUND. As we're called
4976 * from CPU_ONLINE, the following shouldn't fail.
4978 for_each_pool_worker(worker
, pool
) {
4979 kthread_set_per_cpu(worker
->task
, pool
->cpu
);
4980 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4981 pool
->attrs
->cpumask
) < 0);
4984 raw_spin_lock_irq(&pool
->lock
);
4986 pool
->flags
&= ~POOL_DISASSOCIATED
;
4988 for_each_pool_worker(worker
, pool
) {
4989 unsigned int worker_flags
= worker
->flags
;
4992 * A bound idle worker should actually be on the runqueue
4993 * of the associated CPU for local wake-ups targeting it to
4994 * work. Kick all idle workers so that they migrate to the
4995 * associated CPU. Doing this in the same loop as
4996 * replacing UNBOUND with REBOUND is safe as no worker will
4997 * be bound before @pool->lock is released.
4999 if (worker_flags
& WORKER_IDLE
)
5000 wake_up_process(worker
->task
);
5003 * We want to clear UNBOUND but can't directly call
5004 * worker_clr_flags() or adjust nr_running. Atomically
5005 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5006 * @worker will clear REBOUND using worker_clr_flags() when
5007 * it initiates the next execution cycle thus restoring
5008 * concurrency management. Note that when or whether
5009 * @worker clears REBOUND doesn't affect correctness.
5011 * WRITE_ONCE() is necessary because @worker->flags may be
5012 * tested without holding any lock in
5013 * wq_worker_running(). Without it, NOT_RUNNING test may
5014 * fail incorrectly leading to premature concurrency
5015 * management operations.
5017 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
5018 worker_flags
|= WORKER_REBOUND
;
5019 worker_flags
&= ~WORKER_UNBOUND
;
5020 WRITE_ONCE(worker
->flags
, worker_flags
);
5023 raw_spin_unlock_irq(&pool
->lock
);
5027 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5028 * @pool: unbound pool of interest
5029 * @cpu: the CPU which is coming up
5031 * An unbound pool may end up with a cpumask which doesn't have any online
5032 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5033 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5034 * online CPU before, cpus_allowed of all its workers should be restored.
5036 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
5038 static cpumask_t cpumask
;
5039 struct worker
*worker
;
5041 lockdep_assert_held(&wq_pool_attach_mutex
);
5043 /* is @cpu allowed for @pool? */
5044 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
5047 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
5049 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5050 for_each_pool_worker(worker
, pool
)
5051 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, &cpumask
) < 0);
5054 int workqueue_prepare_cpu(unsigned int cpu
)
5056 struct worker_pool
*pool
;
5058 for_each_cpu_worker_pool(pool
, cpu
) {
5059 if (pool
->nr_workers
)
5061 if (!create_worker(pool
))
5067 int workqueue_online_cpu(unsigned int cpu
)
5069 struct worker_pool
*pool
;
5070 struct workqueue_struct
*wq
;
5073 mutex_lock(&wq_pool_mutex
);
5075 for_each_pool(pool
, pi
) {
5076 mutex_lock(&wq_pool_attach_mutex
);
5078 if (pool
->cpu
== cpu
)
5079 rebind_workers(pool
);
5080 else if (pool
->cpu
< 0)
5081 restore_unbound_workers_cpumask(pool
, cpu
);
5083 mutex_unlock(&wq_pool_attach_mutex
);
5086 /* update NUMA affinity of unbound workqueues */
5087 list_for_each_entry(wq
, &workqueues
, list
)
5088 wq_update_unbound_numa(wq
, cpu
, true);
5090 mutex_unlock(&wq_pool_mutex
);
5094 int workqueue_offline_cpu(unsigned int cpu
)
5096 struct workqueue_struct
*wq
;
5098 /* unbinding per-cpu workers should happen on the local CPU */
5099 if (WARN_ON(cpu
!= smp_processor_id()))
5102 unbind_workers(cpu
);
5104 /* update NUMA affinity of unbound workqueues */
5105 mutex_lock(&wq_pool_mutex
);
5106 list_for_each_entry(wq
, &workqueues
, list
)
5107 wq_update_unbound_numa(wq
, cpu
, false);
5108 mutex_unlock(&wq_pool_mutex
);
5113 struct work_for_cpu
{
5114 struct work_struct work
;
5120 static void work_for_cpu_fn(struct work_struct
*work
)
5122 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
5124 wfc
->ret
= wfc
->fn(wfc
->arg
);
5128 * work_on_cpu - run a function in thread context on a particular cpu
5129 * @cpu: the cpu to run on
5130 * @fn: the function to run
5131 * @arg: the function arg
5133 * It is up to the caller to ensure that the cpu doesn't go offline.
5134 * The caller must not hold any locks which would prevent @fn from completing.
5136 * Return: The value @fn returns.
5138 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
5140 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
5142 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
5143 schedule_work_on(cpu
, &wfc
.work
);
5144 flush_work(&wfc
.work
);
5145 destroy_work_on_stack(&wfc
.work
);
5148 EXPORT_SYMBOL_GPL(work_on_cpu
);
5151 * work_on_cpu_safe - run a function in thread context on a particular cpu
5152 * @cpu: the cpu to run on
5153 * @fn: the function to run
5154 * @arg: the function argument
5156 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5157 * any locks which would prevent @fn from completing.
5159 * Return: The value @fn returns.
5161 long work_on_cpu_safe(int cpu
, long (*fn
)(void *), void *arg
)
5166 if (cpu_online(cpu
))
5167 ret
= work_on_cpu(cpu
, fn
, arg
);
5171 EXPORT_SYMBOL_GPL(work_on_cpu_safe
);
5172 #endif /* CONFIG_SMP */
5174 #ifdef CONFIG_FREEZER
5177 * freeze_workqueues_begin - begin freezing workqueues
5179 * Start freezing workqueues. After this function returns, all freezable
5180 * workqueues will queue new works to their delayed_works list instead of
5184 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5186 void freeze_workqueues_begin(void)
5188 struct workqueue_struct
*wq
;
5189 struct pool_workqueue
*pwq
;
5191 mutex_lock(&wq_pool_mutex
);
5193 WARN_ON_ONCE(workqueue_freezing
);
5194 workqueue_freezing
= true;
5196 list_for_each_entry(wq
, &workqueues
, list
) {
5197 mutex_lock(&wq
->mutex
);
5198 for_each_pwq(pwq
, wq
)
5199 pwq_adjust_max_active(pwq
);
5200 mutex_unlock(&wq
->mutex
);
5203 mutex_unlock(&wq_pool_mutex
);
5207 * freeze_workqueues_busy - are freezable workqueues still busy?
5209 * Check whether freezing is complete. This function must be called
5210 * between freeze_workqueues_begin() and thaw_workqueues().
5213 * Grabs and releases wq_pool_mutex.
5216 * %true if some freezable workqueues are still busy. %false if freezing
5219 bool freeze_workqueues_busy(void)
5222 struct workqueue_struct
*wq
;
5223 struct pool_workqueue
*pwq
;
5225 mutex_lock(&wq_pool_mutex
);
5227 WARN_ON_ONCE(!workqueue_freezing
);
5229 list_for_each_entry(wq
, &workqueues
, list
) {
5230 if (!(wq
->flags
& WQ_FREEZABLE
))
5233 * nr_active is monotonically decreasing. It's safe
5234 * to peek without lock.
5237 for_each_pwq(pwq
, wq
) {
5238 WARN_ON_ONCE(pwq
->nr_active
< 0);
5239 if (pwq
->nr_active
) {
5248 mutex_unlock(&wq_pool_mutex
);
5253 * thaw_workqueues - thaw workqueues
5255 * Thaw workqueues. Normal queueing is restored and all collected
5256 * frozen works are transferred to their respective pool worklists.
5259 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5261 void thaw_workqueues(void)
5263 struct workqueue_struct
*wq
;
5264 struct pool_workqueue
*pwq
;
5266 mutex_lock(&wq_pool_mutex
);
5268 if (!workqueue_freezing
)
5271 workqueue_freezing
= false;
5273 /* restore max_active and repopulate worklist */
5274 list_for_each_entry(wq
, &workqueues
, list
) {
5275 mutex_lock(&wq
->mutex
);
5276 for_each_pwq(pwq
, wq
)
5277 pwq_adjust_max_active(pwq
);
5278 mutex_unlock(&wq
->mutex
);
5282 mutex_unlock(&wq_pool_mutex
);
5284 #endif /* CONFIG_FREEZER */
5286 static int workqueue_apply_unbound_cpumask(void)
5290 struct workqueue_struct
*wq
;
5291 struct apply_wqattrs_ctx
*ctx
, *n
;
5293 lockdep_assert_held(&wq_pool_mutex
);
5295 list_for_each_entry(wq
, &workqueues
, list
) {
5296 if (!(wq
->flags
& WQ_UNBOUND
))
5298 /* creating multiple pwqs breaks ordering guarantee */
5299 if (wq
->flags
& __WQ_ORDERED
)
5302 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
5308 list_add_tail(&ctx
->list
, &ctxs
);
5311 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
5313 apply_wqattrs_commit(ctx
);
5314 apply_wqattrs_cleanup(ctx
);
5321 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5322 * @cpumask: the cpumask to set
5324 * The low-level workqueues cpumask is a global cpumask that limits
5325 * the affinity of all unbound workqueues. This function check the @cpumask
5326 * and apply it to all unbound workqueues and updates all pwqs of them.
5328 * Retun: 0 - Success
5329 * -EINVAL - Invalid @cpumask
5330 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5332 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
5335 cpumask_var_t saved_cpumask
;
5337 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
5341 * Not excluding isolated cpus on purpose.
5342 * If the user wishes to include them, we allow that.
5344 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
5345 if (!cpumask_empty(cpumask
)) {
5346 apply_wqattrs_lock();
5348 /* save the old wq_unbound_cpumask. */
5349 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
5351 /* update wq_unbound_cpumask at first and apply it to wqs. */
5352 cpumask_copy(wq_unbound_cpumask
, cpumask
);
5353 ret
= workqueue_apply_unbound_cpumask();
5355 /* restore the wq_unbound_cpumask when failed. */
5357 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
5359 apply_wqattrs_unlock();
5362 free_cpumask_var(saved_cpumask
);
5368 * Workqueues with WQ_SYSFS flag set is visible to userland via
5369 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5370 * following attributes.
5372 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5373 * max_active RW int : maximum number of in-flight work items
5375 * Unbound workqueues have the following extra attributes.
5377 * pool_ids RO int : the associated pool IDs for each node
5378 * nice RW int : nice value of the workers
5379 * cpumask RW mask : bitmask of allowed CPUs for the workers
5380 * numa RW bool : whether enable NUMA affinity
5383 struct workqueue_struct
*wq
;
5387 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
5389 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5394 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
5397 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5399 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
5401 static DEVICE_ATTR_RO(per_cpu
);
5403 static ssize_t
max_active_show(struct device
*dev
,
5404 struct device_attribute
*attr
, char *buf
)
5406 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5408 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
5411 static ssize_t
max_active_store(struct device
*dev
,
5412 struct device_attribute
*attr
, const char *buf
,
5415 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5418 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
5421 workqueue_set_max_active(wq
, val
);
5424 static DEVICE_ATTR_RW(max_active
);
5426 static struct attribute
*wq_sysfs_attrs
[] = {
5427 &dev_attr_per_cpu
.attr
,
5428 &dev_attr_max_active
.attr
,
5431 ATTRIBUTE_GROUPS(wq_sysfs
);
5433 static ssize_t
wq_pool_ids_show(struct device
*dev
,
5434 struct device_attribute
*attr
, char *buf
)
5436 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5437 const char *delim
= "";
5438 int node
, written
= 0;
5442 for_each_node(node
) {
5443 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
5444 "%s%d:%d", delim
, node
,
5445 unbound_pwq_by_node(wq
, node
)->pool
->id
);
5448 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
5455 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
5458 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5461 mutex_lock(&wq
->mutex
);
5462 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
5463 mutex_unlock(&wq
->mutex
);
5468 /* prepare workqueue_attrs for sysfs store operations */
5469 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
5471 struct workqueue_attrs
*attrs
;
5473 lockdep_assert_held(&wq_pool_mutex
);
5475 attrs
= alloc_workqueue_attrs();
5479 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
5483 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
5484 const char *buf
, size_t count
)
5486 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5487 struct workqueue_attrs
*attrs
;
5490 apply_wqattrs_lock();
5492 attrs
= wq_sysfs_prep_attrs(wq
);
5496 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5497 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5498 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5503 apply_wqattrs_unlock();
5504 free_workqueue_attrs(attrs
);
5505 return ret
?: count
;
5508 static ssize_t
wq_cpumask_show(struct device
*dev
,
5509 struct device_attribute
*attr
, char *buf
)
5511 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5514 mutex_lock(&wq
->mutex
);
5515 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5516 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5517 mutex_unlock(&wq
->mutex
);
5521 static ssize_t
wq_cpumask_store(struct device
*dev
,
5522 struct device_attribute
*attr
,
5523 const char *buf
, size_t count
)
5525 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5526 struct workqueue_attrs
*attrs
;
5529 apply_wqattrs_lock();
5531 attrs
= wq_sysfs_prep_attrs(wq
);
5535 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5537 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5540 apply_wqattrs_unlock();
5541 free_workqueue_attrs(attrs
);
5542 return ret
?: count
;
5545 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5548 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5551 mutex_lock(&wq
->mutex
);
5552 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5553 !wq
->unbound_attrs
->no_numa
);
5554 mutex_unlock(&wq
->mutex
);
5559 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5560 const char *buf
, size_t count
)
5562 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5563 struct workqueue_attrs
*attrs
;
5564 int v
, ret
= -ENOMEM
;
5566 apply_wqattrs_lock();
5568 attrs
= wq_sysfs_prep_attrs(wq
);
5573 if (sscanf(buf
, "%d", &v
) == 1) {
5574 attrs
->no_numa
= !v
;
5575 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5579 apply_wqattrs_unlock();
5580 free_workqueue_attrs(attrs
);
5581 return ret
?: count
;
5584 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5585 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5586 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5587 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5588 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5592 static struct bus_type wq_subsys
= {
5593 .name
= "workqueue",
5594 .dev_groups
= wq_sysfs_groups
,
5597 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5598 struct device_attribute
*attr
, char *buf
)
5602 mutex_lock(&wq_pool_mutex
);
5603 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5604 cpumask_pr_args(wq_unbound_cpumask
));
5605 mutex_unlock(&wq_pool_mutex
);
5610 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5611 struct device_attribute
*attr
, const char *buf
, size_t count
)
5613 cpumask_var_t cpumask
;
5616 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5619 ret
= cpumask_parse(buf
, cpumask
);
5621 ret
= workqueue_set_unbound_cpumask(cpumask
);
5623 free_cpumask_var(cpumask
);
5624 return ret
? ret
: count
;
5627 static struct device_attribute wq_sysfs_cpumask_attr
=
5628 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5629 wq_unbound_cpumask_store
);
5631 static int __init
wq_sysfs_init(void)
5635 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5639 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5641 core_initcall(wq_sysfs_init
);
5643 static void wq_device_release(struct device
*dev
)
5645 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5651 * workqueue_sysfs_register - make a workqueue visible in sysfs
5652 * @wq: the workqueue to register
5654 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5655 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5656 * which is the preferred method.
5658 * Workqueue user should use this function directly iff it wants to apply
5659 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5660 * apply_workqueue_attrs() may race against userland updating the
5663 * Return: 0 on success, -errno on failure.
5665 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5667 struct wq_device
*wq_dev
;
5671 * Adjusting max_active or creating new pwqs by applying
5672 * attributes breaks ordering guarantee. Disallow exposing ordered
5675 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
5678 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5683 wq_dev
->dev
.bus
= &wq_subsys
;
5684 wq_dev
->dev
.release
= wq_device_release
;
5685 dev_set_name(&wq_dev
->dev
, "%s", wq
->name
);
5688 * unbound_attrs are created separately. Suppress uevent until
5689 * everything is ready.
5691 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5693 ret
= device_register(&wq_dev
->dev
);
5695 put_device(&wq_dev
->dev
);
5700 if (wq
->flags
& WQ_UNBOUND
) {
5701 struct device_attribute
*attr
;
5703 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5704 ret
= device_create_file(&wq_dev
->dev
, attr
);
5706 device_unregister(&wq_dev
->dev
);
5713 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5714 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5719 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5720 * @wq: the workqueue to unregister
5722 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5724 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5726 struct wq_device
*wq_dev
= wq
->wq_dev
;
5732 device_unregister(&wq_dev
->dev
);
5734 #else /* CONFIG_SYSFS */
5735 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5736 #endif /* CONFIG_SYSFS */
5739 * Workqueue watchdog.
5741 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5742 * flush dependency, a concurrency managed work item which stays RUNNING
5743 * indefinitely. Workqueue stalls can be very difficult to debug as the
5744 * usual warning mechanisms don't trigger and internal workqueue state is
5747 * Workqueue watchdog monitors all worker pools periodically and dumps
5748 * state if some pools failed to make forward progress for a while where
5749 * forward progress is defined as the first item on ->worklist changing.
5751 * This mechanism is controlled through the kernel parameter
5752 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5753 * corresponding sysfs parameter file.
5755 #ifdef CONFIG_WQ_WATCHDOG
5757 static unsigned long wq_watchdog_thresh
= 30;
5758 static struct timer_list wq_watchdog_timer
;
5760 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5761 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5763 static void wq_watchdog_reset_touched(void)
5767 wq_watchdog_touched
= jiffies
;
5768 for_each_possible_cpu(cpu
)
5769 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5772 static void wq_watchdog_timer_fn(struct timer_list
*unused
)
5774 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5775 bool lockup_detected
= false;
5776 unsigned long now
= jiffies
;
5777 struct worker_pool
*pool
;
5785 for_each_pool(pool
, pi
) {
5786 unsigned long pool_ts
, touched
, ts
;
5788 if (list_empty(&pool
->worklist
))
5792 * If a virtual machine is stopped by the host it can look to
5793 * the watchdog like a stall.
5795 kvm_check_and_clear_guest_paused();
5797 /* get the latest of pool and touched timestamps */
5799 touched
= READ_ONCE(per_cpu(wq_watchdog_touched_cpu
, pool
->cpu
));
5801 touched
= READ_ONCE(wq_watchdog_touched
);
5802 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5804 if (time_after(pool_ts
, touched
))
5810 if (time_after(now
, ts
+ thresh
)) {
5811 lockup_detected
= true;
5812 pr_emerg("BUG: workqueue lockup - pool");
5813 pr_cont_pool_info(pool
);
5814 pr_cont(" stuck for %us!\n",
5815 jiffies_to_msecs(now
- pool_ts
) / 1000);
5821 if (lockup_detected
)
5822 show_workqueue_state();
5824 wq_watchdog_reset_touched();
5825 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5828 notrace
void wq_watchdog_touch(int cpu
)
5831 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5833 wq_watchdog_touched
= jiffies
;
5836 static void wq_watchdog_set_thresh(unsigned long thresh
)
5838 wq_watchdog_thresh
= 0;
5839 del_timer_sync(&wq_watchdog_timer
);
5842 wq_watchdog_thresh
= thresh
;
5843 wq_watchdog_reset_touched();
5844 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5848 static int wq_watchdog_param_set_thresh(const char *val
,
5849 const struct kernel_param
*kp
)
5851 unsigned long thresh
;
5854 ret
= kstrtoul(val
, 0, &thresh
);
5859 wq_watchdog_set_thresh(thresh
);
5861 wq_watchdog_thresh
= thresh
;
5866 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5867 .set
= wq_watchdog_param_set_thresh
,
5868 .get
= param_get_ulong
,
5871 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5874 static void wq_watchdog_init(void)
5876 timer_setup(&wq_watchdog_timer
, wq_watchdog_timer_fn
, TIMER_DEFERRABLE
);
5877 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5880 #else /* CONFIG_WQ_WATCHDOG */
5882 static inline void wq_watchdog_init(void) { }
5884 #endif /* CONFIG_WQ_WATCHDOG */
5886 static void __init
wq_numa_init(void)
5891 if (num_possible_nodes() <= 1)
5894 if (wq_disable_numa
) {
5895 pr_info("workqueue: NUMA affinity support disabled\n");
5899 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs();
5900 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5903 * We want masks of possible CPUs of each node which isn't readily
5904 * available. Build one from cpu_to_node() which should have been
5905 * fully initialized by now.
5907 tbl
= kcalloc(nr_node_ids
, sizeof(tbl
[0]), GFP_KERNEL
);
5911 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5912 node_online(node
) ? node
: NUMA_NO_NODE
));
5914 for_each_possible_cpu(cpu
) {
5915 node
= cpu_to_node(cpu
);
5916 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5917 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5918 /* happens iff arch is bonkers, let's just proceed */
5921 cpumask_set_cpu(cpu
, tbl
[node
]);
5924 wq_numa_possible_cpumask
= tbl
;
5925 wq_numa_enabled
= true;
5929 * workqueue_init_early - early init for workqueue subsystem
5931 * This is the first half of two-staged workqueue subsystem initialization
5932 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5933 * idr are up. It sets up all the data structures and system workqueues
5934 * and allows early boot code to create workqueues and queue/cancel work
5935 * items. Actual work item execution starts only after kthreads can be
5936 * created and scheduled right before early initcalls.
5938 void __init
workqueue_init_early(void)
5940 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5941 int hk_flags
= HK_FLAG_DOMAIN
| HK_FLAG_WQ
;
5944 BUILD_BUG_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5946 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5947 cpumask_copy(wq_unbound_cpumask
, housekeeping_cpumask(hk_flags
));
5949 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5951 /* initialize CPU pools */
5952 for_each_possible_cpu(cpu
) {
5953 struct worker_pool
*pool
;
5956 for_each_cpu_worker_pool(pool
, cpu
) {
5957 BUG_ON(init_worker_pool(pool
));
5959 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5960 pool
->attrs
->nice
= std_nice
[i
++];
5961 pool
->node
= cpu_to_node(cpu
);
5964 mutex_lock(&wq_pool_mutex
);
5965 BUG_ON(worker_pool_assign_id(pool
));
5966 mutex_unlock(&wq_pool_mutex
);
5970 /* create default unbound and ordered wq attrs */
5971 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5972 struct workqueue_attrs
*attrs
;
5974 BUG_ON(!(attrs
= alloc_workqueue_attrs()));
5975 attrs
->nice
= std_nice
[i
];
5976 unbound_std_wq_attrs
[i
] = attrs
;
5979 * An ordered wq should have only one pwq as ordering is
5980 * guaranteed by max_active which is enforced by pwqs.
5981 * Turn off NUMA so that dfl_pwq is used for all nodes.
5983 BUG_ON(!(attrs
= alloc_workqueue_attrs()));
5984 attrs
->nice
= std_nice
[i
];
5985 attrs
->no_numa
= true;
5986 ordered_wq_attrs
[i
] = attrs
;
5989 system_wq
= alloc_workqueue("events", 0, 0);
5990 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5991 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5992 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5993 WQ_UNBOUND_MAX_ACTIVE
);
5994 system_freezable_wq
= alloc_workqueue("events_freezable",
5996 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5997 WQ_POWER_EFFICIENT
, 0);
5998 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5999 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
6001 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
6002 !system_unbound_wq
|| !system_freezable_wq
||
6003 !system_power_efficient_wq
||
6004 !system_freezable_power_efficient_wq
);
6008 * workqueue_init - bring workqueue subsystem fully online
6010 * This is the latter half of two-staged workqueue subsystem initialization
6011 * and invoked as soon as kthreads can be created and scheduled.
6012 * Workqueues have been created and work items queued on them, but there
6013 * are no kworkers executing the work items yet. Populate the worker pools
6014 * with the initial workers and enable future kworker creations.
6016 void __init
workqueue_init(void)
6018 struct workqueue_struct
*wq
;
6019 struct worker_pool
*pool
;
6023 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6024 * CPU to node mapping may not be available that early on some
6025 * archs such as power and arm64. As per-cpu pools created
6026 * previously could be missing node hint and unbound pools NUMA
6027 * affinity, fix them up.
6029 * Also, while iterating workqueues, create rescuers if requested.
6033 mutex_lock(&wq_pool_mutex
);
6035 for_each_possible_cpu(cpu
) {
6036 for_each_cpu_worker_pool(pool
, cpu
) {
6037 pool
->node
= cpu_to_node(cpu
);
6041 list_for_each_entry(wq
, &workqueues
, list
) {
6042 wq_update_unbound_numa(wq
, smp_processor_id(), true);
6043 WARN(init_rescuer(wq
),
6044 "workqueue: failed to create early rescuer for %s",
6048 mutex_unlock(&wq_pool_mutex
);
6050 /* create the initial workers */
6051 for_each_online_cpu(cpu
) {
6052 for_each_cpu_worker_pool(pool
, cpu
) {
6053 pool
->flags
&= ~POOL_DISASSOCIATED
;
6054 BUG_ON(!create_worker(pool
));
6058 hash_for_each(unbound_pool_hash
, bkt
, pool
, hash_node
)
6059 BUG_ON(!create_worker(pool
));