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>
54 #include "workqueue_internal.h"
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
69 * Note that DISASSOCIATED should be flipped only while holding
70 * wq_pool_attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
73 POOL_MANAGER_ACTIVE
= 1 << 0, /* being managed */
74 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
77 WORKER_DIE
= 1 << 1, /* die die die */
78 WORKER_IDLE
= 1 << 2, /* is idle */
79 WORKER_PREP
= 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
82 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
85 WORKER_UNBOUND
| WORKER_REBOUND
,
87 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
99 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
105 RESCUER_NICE_LEVEL
= MIN_NICE
,
106 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * A: wq_pool_attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. RCU protected for reads.
142 * MD: wq_mayday_lock protected.
145 /* struct worker is defined in workqueue_internal.h */
148 spinlock_t lock
; /* the pool lock */
149 int cpu
; /* I: the associated cpu */
150 int node
; /* I: the associated node ID */
151 int id
; /* I: pool ID */
152 unsigned int flags
; /* X: flags */
154 unsigned long watchdog_ts
; /* L: watchdog timestamp */
156 struct list_head worklist
; /* L: list of pending works */
158 int nr_workers
; /* L: total number of workers */
159 int nr_idle
; /* L: currently idle workers */
161 struct list_head idle_list
; /* X: list of idle workers */
162 struct timer_list idle_timer
; /* L: worker idle timeout */
163 struct timer_list mayday_timer
; /* L: SOS timer for workers */
165 /* a workers is either on busy_hash or idle_list, or the manager */
166 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
167 /* L: hash of busy workers */
169 struct worker
*manager
; /* L: purely informational */
170 struct list_head workers
; /* A: attached workers */
171 struct completion
*detach_completion
; /* all workers detached */
173 struct ida worker_ida
; /* worker IDs for task name */
175 struct workqueue_attrs
*attrs
; /* I: worker attributes */
176 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
177 int refcnt
; /* PL: refcnt for unbound pools */
180 * The current concurrency level. As it's likely to be accessed
181 * from other CPUs during try_to_wake_up(), put it in a separate
184 atomic_t nr_running ____cacheline_aligned_in_smp
;
187 * Destruction of pool is RCU protected to allow dereferences
188 * from get_work_pool().
191 } ____cacheline_aligned_in_smp
;
194 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
195 * of work_struct->data are used for flags and the remaining high bits
196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
197 * number of flag bits.
199 struct pool_workqueue
{
200 struct worker_pool
*pool
; /* I: the associated pool */
201 struct workqueue_struct
*wq
; /* I: the owning workqueue */
202 int work_color
; /* L: current color */
203 int flush_color
; /* L: flushing color */
204 int refcnt
; /* L: reference count */
205 int nr_in_flight
[WORK_NR_COLORS
];
206 /* L: nr of in_flight works */
207 int nr_active
; /* L: nr of active works */
208 int max_active
; /* L: max active works */
209 struct list_head delayed_works
; /* L: delayed works */
210 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
211 struct list_head mayday_node
; /* MD: node on wq->maydays */
214 * Release of unbound pwq is punted to system_wq. See put_pwq()
215 * and pwq_unbound_release_workfn() for details. pool_workqueue
216 * itself is also RCU protected so that the first pwq can be
217 * determined without grabbing wq->mutex.
219 struct work_struct unbound_release_work
;
221 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
224 * Structure used to wait for workqueue flush.
227 struct list_head list
; /* WQ: list of flushers */
228 int flush_color
; /* WQ: flush color waiting for */
229 struct completion done
; /* flush completion */
235 * The externally visible workqueue. It relays the issued work items to
236 * the appropriate worker_pool through its pool_workqueues.
238 struct workqueue_struct
{
239 struct list_head pwqs
; /* WR: all pwqs of this wq */
240 struct list_head list
; /* PR: list of all workqueues */
242 struct mutex mutex
; /* protects this wq */
243 int work_color
; /* WQ: current work color */
244 int flush_color
; /* WQ: current flush color */
245 atomic_t nr_pwqs_to_flush
; /* flush in progress */
246 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
247 struct list_head flusher_queue
; /* WQ: flush waiters */
248 struct list_head flusher_overflow
; /* WQ: flush overflow list */
250 struct list_head maydays
; /* MD: pwqs requesting rescue */
251 struct worker
*rescuer
; /* MD: rescue worker */
253 int nr_drainers
; /* WQ: drain in progress */
254 int saved_max_active
; /* WQ: saved pwq max_active */
256 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
257 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
260 struct wq_device
*wq_dev
; /* I: for sysfs interface */
262 #ifdef CONFIG_LOCKDEP
264 struct lock_class_key key
;
265 struct lockdep_map lockdep_map
;
267 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
270 * Destruction of workqueue_struct is RCU protected to allow walking
271 * the workqueues list without grabbing wq_pool_mutex.
272 * This is used to dump all workqueues from sysrq.
276 /* hot fields used during command issue, aligned to cacheline */
277 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
278 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
279 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
282 static struct kmem_cache
*pwq_cache
;
284 static cpumask_var_t
*wq_numa_possible_cpumask
;
285 /* possible CPUs of each node */
287 static bool wq_disable_numa
;
288 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
290 /* see the comment above the definition of WQ_POWER_EFFICIENT */
291 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
292 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
294 static bool wq_online
; /* can kworkers be created yet? */
296 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
298 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
301 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
302 static DEFINE_MUTEX(wq_pool_attach_mutex
); /* protects worker attach/detach */
303 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
304 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait
); /* wait for manager to go away */
306 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
307 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
309 /* PL: allowable cpus for unbound wqs and work items */
310 static cpumask_var_t wq_unbound_cpumask
;
312 /* CPU where unbound work was last round robin scheduled from this CPU */
313 static DEFINE_PER_CPU(int, wq_rr_cpu_last
);
316 * Local execution of unbound work items is no longer guaranteed. The
317 * following always forces round-robin CPU selection on unbound work items
318 * to uncover usages which depend on it.
320 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
321 static bool wq_debug_force_rr_cpu
= true;
323 static bool wq_debug_force_rr_cpu
= false;
325 module_param_named(debug_force_rr_cpu
, wq_debug_force_rr_cpu
, bool, 0644);
327 /* the per-cpu worker pools */
328 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
], cpu_worker_pools
);
330 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
332 /* PL: hash of all unbound pools keyed by pool->attrs */
333 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
335 /* I: attributes used when instantiating standard unbound pools on demand */
336 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
338 /* I: attributes used when instantiating ordered pools on demand */
339 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
341 struct workqueue_struct
*system_wq __read_mostly
;
342 EXPORT_SYMBOL(system_wq
);
343 struct workqueue_struct
*system_highpri_wq __read_mostly
;
344 EXPORT_SYMBOL_GPL(system_highpri_wq
);
345 struct workqueue_struct
*system_long_wq __read_mostly
;
346 EXPORT_SYMBOL_GPL(system_long_wq
);
347 struct workqueue_struct
*system_unbound_wq __read_mostly
;
348 EXPORT_SYMBOL_GPL(system_unbound_wq
);
349 struct workqueue_struct
*system_freezable_wq __read_mostly
;
350 EXPORT_SYMBOL_GPL(system_freezable_wq
);
351 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
352 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
353 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
354 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
356 static int worker_thread(void *__worker
);
357 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
358 static void show_pwq(struct pool_workqueue
*pwq
);
360 #define CREATE_TRACE_POINTS
361 #include <trace/events/workqueue.h>
363 #define assert_rcu_or_pool_mutex() \
364 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
365 !lockdep_is_held(&wq_pool_mutex), \
366 "RCU or wq_pool_mutex should be held")
368 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
369 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
370 !lockdep_is_held(&wq->mutex) && \
371 !lockdep_is_held(&wq_pool_mutex), \
372 "RCU, wq->mutex or wq_pool_mutex should be held")
374 #define for_each_cpu_worker_pool(pool, cpu) \
375 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
376 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
380 * for_each_pool - iterate through all worker_pools in the system
381 * @pool: iteration cursor
382 * @pi: integer used for iteration
384 * This must be called either with wq_pool_mutex held or RCU read
385 * locked. If the pool needs to be used beyond the locking in effect, the
386 * caller is responsible for guaranteeing that the pool stays online.
388 * The if/else clause exists only for the lockdep assertion and can be
391 #define for_each_pool(pool, pi) \
392 idr_for_each_entry(&worker_pool_idr, pool, pi) \
393 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
397 * for_each_pool_worker - iterate through all workers of a worker_pool
398 * @worker: iteration cursor
399 * @pool: worker_pool to iterate workers of
401 * This must be called with wq_pool_attach_mutex.
403 * The if/else clause exists only for the lockdep assertion and can be
406 #define for_each_pool_worker(worker, pool) \
407 list_for_each_entry((worker), &(pool)->workers, node) \
408 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
412 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
413 * @pwq: iteration cursor
414 * @wq: the target workqueue
416 * This must be called either with wq->mutex held or RCU read locked.
417 * If the pwq needs to be used beyond the locking in effect, the caller is
418 * responsible for guaranteeing that the pwq stays online.
420 * The if/else clause exists only for the lockdep assertion and can be
423 #define for_each_pwq(pwq, wq) \
424 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
425 lockdep_is_held(&(wq->mutex)))
427 #ifdef CONFIG_DEBUG_OBJECTS_WORK
429 static struct debug_obj_descr work_debug_descr
;
431 static void *work_debug_hint(void *addr
)
433 return ((struct work_struct
*) addr
)->func
;
436 static bool work_is_static_object(void *addr
)
438 struct work_struct
*work
= addr
;
440 return test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
));
444 * fixup_init is called when:
445 * - an active object is initialized
447 static bool work_fixup_init(void *addr
, enum debug_obj_state state
)
449 struct work_struct
*work
= addr
;
452 case ODEBUG_STATE_ACTIVE
:
453 cancel_work_sync(work
);
454 debug_object_init(work
, &work_debug_descr
);
462 * fixup_free is called when:
463 * - an active object is freed
465 static bool work_fixup_free(void *addr
, enum debug_obj_state state
)
467 struct work_struct
*work
= addr
;
470 case ODEBUG_STATE_ACTIVE
:
471 cancel_work_sync(work
);
472 debug_object_free(work
, &work_debug_descr
);
479 static struct debug_obj_descr work_debug_descr
= {
480 .name
= "work_struct",
481 .debug_hint
= work_debug_hint
,
482 .is_static_object
= work_is_static_object
,
483 .fixup_init
= work_fixup_init
,
484 .fixup_free
= work_fixup_free
,
487 static inline void debug_work_activate(struct work_struct
*work
)
489 debug_object_activate(work
, &work_debug_descr
);
492 static inline void debug_work_deactivate(struct work_struct
*work
)
494 debug_object_deactivate(work
, &work_debug_descr
);
497 void __init_work(struct work_struct
*work
, int onstack
)
500 debug_object_init_on_stack(work
, &work_debug_descr
);
502 debug_object_init(work
, &work_debug_descr
);
504 EXPORT_SYMBOL_GPL(__init_work
);
506 void destroy_work_on_stack(struct work_struct
*work
)
508 debug_object_free(work
, &work_debug_descr
);
510 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
512 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
514 destroy_timer_on_stack(&work
->timer
);
515 debug_object_free(&work
->work
, &work_debug_descr
);
517 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
520 static inline void debug_work_activate(struct work_struct
*work
) { }
521 static inline void debug_work_deactivate(struct work_struct
*work
) { }
525 * worker_pool_assign_id - allocate ID and assing it to @pool
526 * @pool: the pool pointer of interest
528 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
529 * successfully, -errno on failure.
531 static int worker_pool_assign_id(struct worker_pool
*pool
)
535 lockdep_assert_held(&wq_pool_mutex
);
537 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
547 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
548 * @wq: the target workqueue
551 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
553 * If the pwq needs to be used beyond the locking in effect, the caller is
554 * responsible for guaranteeing that the pwq stays online.
556 * Return: The unbound pool_workqueue for @node.
558 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
561 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
564 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
565 * delayed item is pending. The plan is to keep CPU -> NODE
566 * mapping valid and stable across CPU on/offlines. Once that
567 * happens, this workaround can be removed.
569 if (unlikely(node
== NUMA_NO_NODE
))
572 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
575 static unsigned int work_color_to_flags(int color
)
577 return color
<< WORK_STRUCT_COLOR_SHIFT
;
580 static int get_work_color(struct work_struct
*work
)
582 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
583 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
586 static int work_next_color(int color
)
588 return (color
+ 1) % WORK_NR_COLORS
;
592 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
593 * contain the pointer to the queued pwq. Once execution starts, the flag
594 * is cleared and the high bits contain OFFQ flags and pool ID.
596 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
597 * and clear_work_data() can be used to set the pwq, pool or clear
598 * work->data. These functions should only be called while the work is
599 * owned - ie. while the PENDING bit is set.
601 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
602 * corresponding to a work. Pool is available once the work has been
603 * queued anywhere after initialization until it is sync canceled. pwq is
604 * available only while the work item is queued.
606 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
607 * canceled. While being canceled, a work item may have its PENDING set
608 * but stay off timer and worklist for arbitrarily long and nobody should
609 * try to steal the PENDING bit.
611 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
614 WARN_ON_ONCE(!work_pending(work
));
615 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
618 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
619 unsigned long extra_flags
)
621 set_work_data(work
, (unsigned long)pwq
,
622 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
625 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
628 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
629 WORK_STRUCT_PENDING
);
632 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
636 * The following wmb is paired with the implied mb in
637 * test_and_set_bit(PENDING) and ensures all updates to @work made
638 * here are visible to and precede any updates by the next PENDING
642 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
644 * The following mb guarantees that previous clear of a PENDING bit
645 * will not be reordered with any speculative LOADS or STORES from
646 * work->current_func, which is executed afterwards. This possible
647 * reordering can lead to a missed execution on attempt to queue
648 * the same @work. E.g. consider this case:
651 * ---------------------------- --------------------------------
653 * 1 STORE event_indicated
654 * 2 queue_work_on() {
655 * 3 test_and_set_bit(PENDING)
656 * 4 } set_..._and_clear_pending() {
657 * 5 set_work_data() # clear bit
659 * 7 work->current_func() {
660 * 8 LOAD event_indicated
663 * Without an explicit full barrier speculative LOAD on line 8 can
664 * be executed before CPU#0 does STORE on line 1. If that happens,
665 * CPU#0 observes the PENDING bit is still set and new execution of
666 * a @work is not queued in a hope, that CPU#1 will eventually
667 * finish the queued @work. Meanwhile CPU#1 does not see
668 * event_indicated is set, because speculative LOAD was executed
669 * before actual STORE.
674 static void clear_work_data(struct work_struct
*work
)
676 smp_wmb(); /* see set_work_pool_and_clear_pending() */
677 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
680 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
682 unsigned long data
= atomic_long_read(&work
->data
);
684 if (data
& WORK_STRUCT_PWQ
)
685 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
691 * get_work_pool - return the worker_pool a given work was associated with
692 * @work: the work item of interest
694 * Pools are created and destroyed under wq_pool_mutex, and allows read
695 * access under RCU read lock. As such, this function should be
696 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
698 * All fields of the returned pool are accessible as long as the above
699 * mentioned locking is in effect. If the returned pool needs to be used
700 * beyond the critical section, the caller is responsible for ensuring the
701 * returned pool is and stays online.
703 * Return: The worker_pool @work was last associated with. %NULL if none.
705 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
707 unsigned long data
= atomic_long_read(&work
->data
);
710 assert_rcu_or_pool_mutex();
712 if (data
& WORK_STRUCT_PWQ
)
713 return ((struct pool_workqueue
*)
714 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
716 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
717 if (pool_id
== WORK_OFFQ_POOL_NONE
)
720 return idr_find(&worker_pool_idr
, pool_id
);
724 * get_work_pool_id - return the worker pool ID a given work is associated with
725 * @work: the work item of interest
727 * Return: The worker_pool ID @work was last associated with.
728 * %WORK_OFFQ_POOL_NONE if none.
730 static int get_work_pool_id(struct work_struct
*work
)
732 unsigned long data
= atomic_long_read(&work
->data
);
734 if (data
& WORK_STRUCT_PWQ
)
735 return ((struct pool_workqueue
*)
736 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
738 return data
>> WORK_OFFQ_POOL_SHIFT
;
741 static void mark_work_canceling(struct work_struct
*work
)
743 unsigned long pool_id
= get_work_pool_id(work
);
745 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
746 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
749 static bool work_is_canceling(struct work_struct
*work
)
751 unsigned long data
= atomic_long_read(&work
->data
);
753 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
757 * Policy functions. These define the policies on how the global worker
758 * pools are managed. Unless noted otherwise, these functions assume that
759 * they're being called with pool->lock held.
762 static bool __need_more_worker(struct worker_pool
*pool
)
764 return !atomic_read(&pool
->nr_running
);
768 * Need to wake up a worker? Called from anything but currently
771 * Note that, because unbound workers never contribute to nr_running, this
772 * function will always return %true for unbound pools as long as the
773 * worklist isn't empty.
775 static bool need_more_worker(struct worker_pool
*pool
)
777 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
780 /* Can I start working? Called from busy but !running workers. */
781 static bool may_start_working(struct worker_pool
*pool
)
783 return pool
->nr_idle
;
786 /* Do I need to keep working? Called from currently running workers. */
787 static bool keep_working(struct worker_pool
*pool
)
789 return !list_empty(&pool
->worklist
) &&
790 atomic_read(&pool
->nr_running
) <= 1;
793 /* Do we need a new worker? Called from manager. */
794 static bool need_to_create_worker(struct worker_pool
*pool
)
796 return need_more_worker(pool
) && !may_start_working(pool
);
799 /* Do we have too many workers and should some go away? */
800 static bool too_many_workers(struct worker_pool
*pool
)
802 bool managing
= pool
->flags
& POOL_MANAGER_ACTIVE
;
803 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
804 int nr_busy
= pool
->nr_workers
- nr_idle
;
806 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
813 /* Return the first idle worker. Safe with preemption disabled */
814 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
816 if (unlikely(list_empty(&pool
->idle_list
)))
819 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
823 * wake_up_worker - wake up an idle worker
824 * @pool: worker pool to wake worker from
826 * Wake up the first idle worker of @pool.
829 * spin_lock_irq(pool->lock).
831 static void wake_up_worker(struct worker_pool
*pool
)
833 struct worker
*worker
= first_idle_worker(pool
);
836 wake_up_process(worker
->task
);
840 * wq_worker_running - a worker is running again
841 * @task: task waking up
843 * This function is called when a worker returns from schedule()
845 void wq_worker_running(struct task_struct
*task
)
847 struct worker
*worker
= kthread_data(task
);
849 if (!worker
->sleeping
)
851 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
852 atomic_inc(&worker
->pool
->nr_running
);
853 worker
->sleeping
= 0;
857 * wq_worker_sleeping - a worker is going to sleep
858 * @task: task going to sleep
860 * This function is called from schedule() when a busy worker is
863 void wq_worker_sleeping(struct task_struct
*task
)
865 struct worker
*next
, *worker
= kthread_data(task
);
866 struct worker_pool
*pool
;
869 * Rescuers, which may not have all the fields set up like normal
870 * workers, also reach here, let's not access anything before
871 * checking NOT_RUNNING.
873 if (worker
->flags
& WORKER_NOT_RUNNING
)
878 if (WARN_ON_ONCE(worker
->sleeping
))
881 worker
->sleeping
= 1;
882 spin_lock_irq(&pool
->lock
);
885 * The counterpart of the following dec_and_test, implied mb,
886 * worklist not empty test sequence is in insert_work().
887 * Please read comment there.
889 * NOT_RUNNING is clear. This means that we're bound to and
890 * running on the local cpu w/ rq lock held and preemption
891 * disabled, which in turn means that none else could be
892 * manipulating idle_list, so dereferencing idle_list without pool
895 if (atomic_dec_and_test(&pool
->nr_running
) &&
896 !list_empty(&pool
->worklist
)) {
897 next
= first_idle_worker(pool
);
899 wake_up_process(next
->task
);
901 spin_unlock_irq(&pool
->lock
);
905 * wq_worker_last_func - retrieve worker's last work function
906 * @task: Task to retrieve last work function of.
908 * Determine the last function a worker executed. This is called from
909 * the scheduler to get a worker's last known identity.
912 * spin_lock_irq(rq->lock)
914 * This function is called during schedule() when a kworker is going
915 * to sleep. It's used by psi to identify aggregation workers during
916 * dequeuing, to allow periodic aggregation to shut-off when that
917 * worker is the last task in the system or cgroup to go to sleep.
919 * As this function doesn't involve any workqueue-related locking, it
920 * only returns stable values when called from inside the scheduler's
921 * queuing and dequeuing paths, when @task, which must be a kworker,
922 * is guaranteed to not be processing any works.
925 * The last work function %current executed as a worker, NULL if it
926 * hasn't executed any work yet.
928 work_func_t
wq_worker_last_func(struct task_struct
*task
)
930 struct worker
*worker
= kthread_data(task
);
932 return worker
->last_func
;
936 * worker_set_flags - set worker flags and adjust nr_running accordingly
938 * @flags: flags to set
940 * Set @flags in @worker->flags and adjust nr_running accordingly.
943 * spin_lock_irq(pool->lock)
945 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
947 struct worker_pool
*pool
= worker
->pool
;
949 WARN_ON_ONCE(worker
->task
!= current
);
951 /* If transitioning into NOT_RUNNING, adjust nr_running. */
952 if ((flags
& WORKER_NOT_RUNNING
) &&
953 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
954 atomic_dec(&pool
->nr_running
);
957 worker
->flags
|= flags
;
961 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
963 * @flags: flags to clear
965 * Clear @flags in @worker->flags and adjust nr_running accordingly.
968 * spin_lock_irq(pool->lock)
970 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
972 struct worker_pool
*pool
= worker
->pool
;
973 unsigned int oflags
= worker
->flags
;
975 WARN_ON_ONCE(worker
->task
!= current
);
977 worker
->flags
&= ~flags
;
980 * If transitioning out of NOT_RUNNING, increment nr_running. Note
981 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
982 * of multiple flags, not a single flag.
984 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
985 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
986 atomic_inc(&pool
->nr_running
);
990 * find_worker_executing_work - find worker which is executing a work
991 * @pool: pool of interest
992 * @work: work to find worker for
994 * Find a worker which is executing @work on @pool by searching
995 * @pool->busy_hash which is keyed by the address of @work. For a worker
996 * to match, its current execution should match the address of @work and
997 * its work function. This is to avoid unwanted dependency between
998 * unrelated work executions through a work item being recycled while still
1001 * This is a bit tricky. A work item may be freed once its execution
1002 * starts and nothing prevents the freed area from being recycled for
1003 * another work item. If the same work item address ends up being reused
1004 * before the original execution finishes, workqueue will identify the
1005 * recycled work item as currently executing and make it wait until the
1006 * current execution finishes, introducing an unwanted dependency.
1008 * This function checks the work item address and work function to avoid
1009 * false positives. Note that this isn't complete as one may construct a
1010 * work function which can introduce dependency onto itself through a
1011 * recycled work item. Well, if somebody wants to shoot oneself in the
1012 * foot that badly, there's only so much we can do, and if such deadlock
1013 * actually occurs, it should be easy to locate the culprit work function.
1016 * spin_lock_irq(pool->lock).
1019 * Pointer to worker which is executing @work if found, %NULL
1022 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
1023 struct work_struct
*work
)
1025 struct worker
*worker
;
1027 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1028 (unsigned long)work
)
1029 if (worker
->current_work
== work
&&
1030 worker
->current_func
== work
->func
)
1037 * move_linked_works - move linked works to a list
1038 * @work: start of series of works to be scheduled
1039 * @head: target list to append @work to
1040 * @nextp: out parameter for nested worklist walking
1042 * Schedule linked works starting from @work to @head. Work series to
1043 * be scheduled starts at @work and includes any consecutive work with
1044 * WORK_STRUCT_LINKED set in its predecessor.
1046 * If @nextp is not NULL, it's updated to point to the next work of
1047 * the last scheduled work. This allows move_linked_works() to be
1048 * nested inside outer list_for_each_entry_safe().
1051 * spin_lock_irq(pool->lock).
1053 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1054 struct work_struct
**nextp
)
1056 struct work_struct
*n
;
1059 * Linked worklist will always end before the end of the list,
1060 * use NULL for list head.
1062 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1063 list_move_tail(&work
->entry
, head
);
1064 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1069 * If we're already inside safe list traversal and have moved
1070 * multiple works to the scheduled queue, the next position
1071 * needs to be updated.
1078 * get_pwq - get an extra reference on the specified pool_workqueue
1079 * @pwq: pool_workqueue to get
1081 * Obtain an extra reference on @pwq. The caller should guarantee that
1082 * @pwq has positive refcnt and be holding the matching pool->lock.
1084 static void get_pwq(struct pool_workqueue
*pwq
)
1086 lockdep_assert_held(&pwq
->pool
->lock
);
1087 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1092 * put_pwq - put a pool_workqueue reference
1093 * @pwq: pool_workqueue to put
1095 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1096 * destruction. The caller should be holding the matching pool->lock.
1098 static void put_pwq(struct pool_workqueue
*pwq
)
1100 lockdep_assert_held(&pwq
->pool
->lock
);
1101 if (likely(--pwq
->refcnt
))
1103 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1106 * @pwq can't be released under pool->lock, bounce to
1107 * pwq_unbound_release_workfn(). This never recurses on the same
1108 * pool->lock as this path is taken only for unbound workqueues and
1109 * the release work item is scheduled on a per-cpu workqueue. To
1110 * avoid lockdep warning, unbound pool->locks are given lockdep
1111 * subclass of 1 in get_unbound_pool().
1113 schedule_work(&pwq
->unbound_release_work
);
1117 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1118 * @pwq: pool_workqueue to put (can be %NULL)
1120 * put_pwq() with locking. This function also allows %NULL @pwq.
1122 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1126 * As both pwqs and pools are RCU protected, the
1127 * following lock operations are safe.
1129 spin_lock_irq(&pwq
->pool
->lock
);
1131 spin_unlock_irq(&pwq
->pool
->lock
);
1135 static void pwq_activate_delayed_work(struct work_struct
*work
)
1137 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1139 trace_workqueue_activate_work(work
);
1140 if (list_empty(&pwq
->pool
->worklist
))
1141 pwq
->pool
->watchdog_ts
= jiffies
;
1142 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1143 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1147 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1149 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1150 struct work_struct
, entry
);
1152 pwq_activate_delayed_work(work
);
1156 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1157 * @pwq: pwq of interest
1158 * @color: color of work which left the queue
1160 * A work either has completed or is removed from pending queue,
1161 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1164 * spin_lock_irq(pool->lock).
1166 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1168 /* uncolored work items don't participate in flushing or nr_active */
1169 if (color
== WORK_NO_COLOR
)
1172 pwq
->nr_in_flight
[color
]--;
1175 if (!list_empty(&pwq
->delayed_works
)) {
1176 /* one down, submit a delayed one */
1177 if (pwq
->nr_active
< pwq
->max_active
)
1178 pwq_activate_first_delayed(pwq
);
1181 /* is flush in progress and are we at the flushing tip? */
1182 if (likely(pwq
->flush_color
!= color
))
1185 /* are there still in-flight works? */
1186 if (pwq
->nr_in_flight
[color
])
1189 /* this pwq is done, clear flush_color */
1190 pwq
->flush_color
= -1;
1193 * If this was the last pwq, wake up the first flusher. It
1194 * will handle the rest.
1196 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1197 complete(&pwq
->wq
->first_flusher
->done
);
1203 * try_to_grab_pending - steal work item from worklist and disable irq
1204 * @work: work item to steal
1205 * @is_dwork: @work is a delayed_work
1206 * @flags: place to store irq state
1208 * Try to grab PENDING bit of @work. This function can handle @work in any
1209 * stable state - idle, on timer or on worklist.
1212 * 1 if @work was pending and we successfully stole PENDING
1213 * 0 if @work was idle and we claimed PENDING
1214 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1215 * -ENOENT if someone else is canceling @work, this state may persist
1216 * for arbitrarily long
1219 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1220 * interrupted while holding PENDING and @work off queue, irq must be
1221 * disabled on entry. This, combined with delayed_work->timer being
1222 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1224 * On successful return, >= 0, irq is disabled and the caller is
1225 * responsible for releasing it using local_irq_restore(*@flags).
1227 * This function is safe to call from any context including IRQ handler.
1229 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1230 unsigned long *flags
)
1232 struct worker_pool
*pool
;
1233 struct pool_workqueue
*pwq
;
1235 local_irq_save(*flags
);
1237 /* try to steal the timer if it exists */
1239 struct delayed_work
*dwork
= to_delayed_work(work
);
1242 * dwork->timer is irqsafe. If del_timer() fails, it's
1243 * guaranteed that the timer is not queued anywhere and not
1244 * running on the local CPU.
1246 if (likely(del_timer(&dwork
->timer
)))
1250 /* try to claim PENDING the normal way */
1251 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1256 * The queueing is in progress, or it is already queued. Try to
1257 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1259 pool
= get_work_pool(work
);
1263 spin_lock(&pool
->lock
);
1265 * work->data is guaranteed to point to pwq only while the work
1266 * item is queued on pwq->wq, and both updating work->data to point
1267 * to pwq on queueing and to pool on dequeueing are done under
1268 * pwq->pool->lock. This in turn guarantees that, if work->data
1269 * points to pwq which is associated with a locked pool, the work
1270 * item is currently queued on that pool.
1272 pwq
= get_work_pwq(work
);
1273 if (pwq
&& pwq
->pool
== pool
) {
1274 debug_work_deactivate(work
);
1277 * A delayed work item cannot be grabbed directly because
1278 * it might have linked NO_COLOR work items which, if left
1279 * on the delayed_list, will confuse pwq->nr_active
1280 * management later on and cause stall. Make sure the work
1281 * item is activated before grabbing.
1283 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1284 pwq_activate_delayed_work(work
);
1286 list_del_init(&work
->entry
);
1287 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1289 /* work->data points to pwq iff queued, point to pool */
1290 set_work_pool_and_keep_pending(work
, pool
->id
);
1292 spin_unlock(&pool
->lock
);
1296 spin_unlock(&pool
->lock
);
1299 local_irq_restore(*flags
);
1300 if (work_is_canceling(work
))
1307 * insert_work - insert a work into a pool
1308 * @pwq: pwq @work belongs to
1309 * @work: work to insert
1310 * @head: insertion point
1311 * @extra_flags: extra WORK_STRUCT_* flags to set
1313 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1314 * work_struct flags.
1317 * spin_lock_irq(pool->lock).
1319 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1320 struct list_head
*head
, unsigned int extra_flags
)
1322 struct worker_pool
*pool
= pwq
->pool
;
1324 /* we own @work, set data and link */
1325 set_work_pwq(work
, pwq
, extra_flags
);
1326 list_add_tail(&work
->entry
, head
);
1330 * Ensure either wq_worker_sleeping() sees the above
1331 * list_add_tail() or we see zero nr_running to avoid workers lying
1332 * around lazily while there are works to be processed.
1336 if (__need_more_worker(pool
))
1337 wake_up_worker(pool
);
1341 * Test whether @work is being queued from another work executing on the
1344 static bool is_chained_work(struct workqueue_struct
*wq
)
1346 struct worker
*worker
;
1348 worker
= current_wq_worker();
1350 * Return %true iff I'm a worker executing a work item on @wq. If
1351 * I'm @worker, it's safe to dereference it without locking.
1353 return worker
&& worker
->current_pwq
->wq
== wq
;
1357 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1358 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1359 * avoid perturbing sensitive tasks.
1361 static int wq_select_unbound_cpu(int cpu
)
1363 static bool printed_dbg_warning
;
1366 if (likely(!wq_debug_force_rr_cpu
)) {
1367 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1369 } else if (!printed_dbg_warning
) {
1370 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1371 printed_dbg_warning
= true;
1374 if (cpumask_empty(wq_unbound_cpumask
))
1377 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1378 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1379 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1380 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1381 if (unlikely(new_cpu
>= nr_cpu_ids
))
1384 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1389 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1390 struct work_struct
*work
)
1392 struct pool_workqueue
*pwq
;
1393 struct worker_pool
*last_pool
;
1394 struct list_head
*worklist
;
1395 unsigned int work_flags
;
1396 unsigned int req_cpu
= cpu
;
1399 * While a work item is PENDING && off queue, a task trying to
1400 * steal the PENDING will busy-loop waiting for it to either get
1401 * queued or lose PENDING. Grabbing PENDING and queueing should
1402 * happen with IRQ disabled.
1404 lockdep_assert_irqs_disabled();
1406 debug_work_activate(work
);
1408 /* if draining, only works from the same workqueue are allowed */
1409 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1410 WARN_ON_ONCE(!is_chained_work(wq
)))
1414 /* pwq which will be used unless @work is executing elsewhere */
1415 if (wq
->flags
& WQ_UNBOUND
) {
1416 if (req_cpu
== WORK_CPU_UNBOUND
)
1417 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1418 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1420 if (req_cpu
== WORK_CPU_UNBOUND
)
1421 cpu
= raw_smp_processor_id();
1422 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1426 * If @work was previously on a different pool, it might still be
1427 * running there, in which case the work needs to be queued on that
1428 * pool to guarantee non-reentrancy.
1430 last_pool
= get_work_pool(work
);
1431 if (last_pool
&& last_pool
!= pwq
->pool
) {
1432 struct worker
*worker
;
1434 spin_lock(&last_pool
->lock
);
1436 worker
= find_worker_executing_work(last_pool
, work
);
1438 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1439 pwq
= worker
->current_pwq
;
1441 /* meh... not running there, queue here */
1442 spin_unlock(&last_pool
->lock
);
1443 spin_lock(&pwq
->pool
->lock
);
1446 spin_lock(&pwq
->pool
->lock
);
1450 * pwq is determined and locked. For unbound pools, we could have
1451 * raced with pwq release and it could already be dead. If its
1452 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1453 * without another pwq replacing it in the numa_pwq_tbl or while
1454 * work items are executing on it, so the retrying is guaranteed to
1455 * make forward-progress.
1457 if (unlikely(!pwq
->refcnt
)) {
1458 if (wq
->flags
& WQ_UNBOUND
) {
1459 spin_unlock(&pwq
->pool
->lock
);
1464 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1468 /* pwq determined, queue */
1469 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1471 if (WARN_ON(!list_empty(&work
->entry
)))
1474 pwq
->nr_in_flight
[pwq
->work_color
]++;
1475 work_flags
= work_color_to_flags(pwq
->work_color
);
1477 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1478 trace_workqueue_activate_work(work
);
1480 worklist
= &pwq
->pool
->worklist
;
1481 if (list_empty(worklist
))
1482 pwq
->pool
->watchdog_ts
= jiffies
;
1484 work_flags
|= WORK_STRUCT_DELAYED
;
1485 worklist
= &pwq
->delayed_works
;
1488 insert_work(pwq
, work
, worklist
, work_flags
);
1491 spin_unlock(&pwq
->pool
->lock
);
1496 * queue_work_on - queue work on specific cpu
1497 * @cpu: CPU number to execute work on
1498 * @wq: workqueue to use
1499 * @work: work to queue
1501 * We queue the work to a specific CPU, the caller must ensure it
1504 * Return: %false if @work was already on a queue, %true otherwise.
1506 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1507 struct work_struct
*work
)
1510 unsigned long flags
;
1512 local_irq_save(flags
);
1514 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1515 __queue_work(cpu
, wq
, work
);
1519 local_irq_restore(flags
);
1522 EXPORT_SYMBOL(queue_work_on
);
1525 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1526 * @node: NUMA node ID that we want to select a CPU from
1528 * This function will attempt to find a "random" cpu available on a given
1529 * node. If there are no CPUs available on the given node it will return
1530 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1531 * available CPU if we need to schedule this work.
1533 static int workqueue_select_cpu_near(int node
)
1537 /* No point in doing this if NUMA isn't enabled for workqueues */
1538 if (!wq_numa_enabled
)
1539 return WORK_CPU_UNBOUND
;
1541 /* Delay binding to CPU if node is not valid or online */
1542 if (node
< 0 || node
>= MAX_NUMNODES
|| !node_online(node
))
1543 return WORK_CPU_UNBOUND
;
1545 /* Use local node/cpu if we are already there */
1546 cpu
= raw_smp_processor_id();
1547 if (node
== cpu_to_node(cpu
))
1550 /* Use "random" otherwise know as "first" online CPU of node */
1551 cpu
= cpumask_any_and(cpumask_of_node(node
), cpu_online_mask
);
1553 /* If CPU is valid return that, otherwise just defer */
1554 return cpu
< nr_cpu_ids
? cpu
: WORK_CPU_UNBOUND
;
1558 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1559 * @node: NUMA node that we are targeting the work for
1560 * @wq: workqueue to use
1561 * @work: work to queue
1563 * We queue the work to a "random" CPU within a given NUMA node. The basic
1564 * idea here is to provide a way to somehow associate work with a given
1567 * This function will only make a best effort attempt at getting this onto
1568 * the right NUMA node. If no node is requested or the requested node is
1569 * offline then we just fall back to standard queue_work behavior.
1571 * Currently the "random" CPU ends up being the first available CPU in the
1572 * intersection of cpu_online_mask and the cpumask of the node, unless we
1573 * are running on the node. In that case we just use the current CPU.
1575 * Return: %false if @work was already on a queue, %true otherwise.
1577 bool queue_work_node(int node
, struct workqueue_struct
*wq
,
1578 struct work_struct
*work
)
1580 unsigned long flags
;
1584 * This current implementation is specific to unbound workqueues.
1585 * Specifically we only return the first available CPU for a given
1586 * node instead of cycling through individual CPUs within the node.
1588 * If this is used with a per-cpu workqueue then the logic in
1589 * workqueue_select_cpu_near would need to be updated to allow for
1590 * some round robin type logic.
1592 WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
));
1594 local_irq_save(flags
);
1596 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1597 int cpu
= workqueue_select_cpu_near(node
);
1599 __queue_work(cpu
, wq
, work
);
1603 local_irq_restore(flags
);
1606 EXPORT_SYMBOL_GPL(queue_work_node
);
1608 void delayed_work_timer_fn(struct timer_list
*t
)
1610 struct delayed_work
*dwork
= from_timer(dwork
, t
, timer
);
1612 /* should have been called from irqsafe timer with irq already off */
1613 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1615 EXPORT_SYMBOL(delayed_work_timer_fn
);
1617 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1618 struct delayed_work
*dwork
, unsigned long delay
)
1620 struct timer_list
*timer
= &dwork
->timer
;
1621 struct work_struct
*work
= &dwork
->work
;
1624 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
);
1625 WARN_ON_ONCE(timer_pending(timer
));
1626 WARN_ON_ONCE(!list_empty(&work
->entry
));
1629 * If @delay is 0, queue @dwork->work immediately. This is for
1630 * both optimization and correctness. The earliest @timer can
1631 * expire is on the closest next tick and delayed_work users depend
1632 * on that there's no such delay when @delay is 0.
1635 __queue_work(cpu
, wq
, &dwork
->work
);
1641 timer
->expires
= jiffies
+ delay
;
1643 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1644 add_timer_on(timer
, cpu
);
1650 * queue_delayed_work_on - queue work on specific CPU after delay
1651 * @cpu: CPU number to execute work on
1652 * @wq: workqueue to use
1653 * @dwork: work to queue
1654 * @delay: number of jiffies to wait before queueing
1656 * Return: %false if @work was already on a queue, %true otherwise. If
1657 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1660 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1661 struct delayed_work
*dwork
, unsigned long delay
)
1663 struct work_struct
*work
= &dwork
->work
;
1665 unsigned long flags
;
1667 /* read the comment in __queue_work() */
1668 local_irq_save(flags
);
1670 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1671 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1675 local_irq_restore(flags
);
1678 EXPORT_SYMBOL(queue_delayed_work_on
);
1681 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1682 * @cpu: CPU number to execute work on
1683 * @wq: workqueue to use
1684 * @dwork: work to queue
1685 * @delay: number of jiffies to wait before queueing
1687 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1688 * modify @dwork's timer so that it expires after @delay. If @delay is
1689 * zero, @work is guaranteed to be scheduled immediately regardless of its
1692 * Return: %false if @dwork was idle and queued, %true if @dwork was
1693 * pending and its timer was modified.
1695 * This function is safe to call from any context including IRQ handler.
1696 * See try_to_grab_pending() for details.
1698 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1699 struct delayed_work
*dwork
, unsigned long delay
)
1701 unsigned long flags
;
1705 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1706 } while (unlikely(ret
== -EAGAIN
));
1708 if (likely(ret
>= 0)) {
1709 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1710 local_irq_restore(flags
);
1713 /* -ENOENT from try_to_grab_pending() becomes %true */
1716 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1718 static void rcu_work_rcufn(struct rcu_head
*rcu
)
1720 struct rcu_work
*rwork
= container_of(rcu
, struct rcu_work
, rcu
);
1722 /* read the comment in __queue_work() */
1723 local_irq_disable();
1724 __queue_work(WORK_CPU_UNBOUND
, rwork
->wq
, &rwork
->work
);
1729 * queue_rcu_work - queue work after a RCU grace period
1730 * @wq: workqueue to use
1731 * @rwork: work to queue
1733 * Return: %false if @rwork was already pending, %true otherwise. Note
1734 * that a full RCU grace period is guaranteed only after a %true return.
1735 * While @rwork is guaranteed to be executed after a %false return, the
1736 * execution may happen before a full RCU grace period has passed.
1738 bool queue_rcu_work(struct workqueue_struct
*wq
, struct rcu_work
*rwork
)
1740 struct work_struct
*work
= &rwork
->work
;
1742 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1744 call_rcu(&rwork
->rcu
, rcu_work_rcufn
);
1750 EXPORT_SYMBOL(queue_rcu_work
);
1753 * worker_enter_idle - enter idle state
1754 * @worker: worker which is entering idle state
1756 * @worker is entering idle state. Update stats and idle timer if
1760 * spin_lock_irq(pool->lock).
1762 static void worker_enter_idle(struct worker
*worker
)
1764 struct worker_pool
*pool
= worker
->pool
;
1766 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1767 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1768 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1771 /* can't use worker_set_flags(), also called from create_worker() */
1772 worker
->flags
|= WORKER_IDLE
;
1774 worker
->last_active
= jiffies
;
1776 /* idle_list is LIFO */
1777 list_add(&worker
->entry
, &pool
->idle_list
);
1779 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1780 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1783 * Sanity check nr_running. Because unbind_workers() releases
1784 * pool->lock between setting %WORKER_UNBOUND and zapping
1785 * nr_running, the warning may trigger spuriously. Check iff
1786 * unbind is not in progress.
1788 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1789 pool
->nr_workers
== pool
->nr_idle
&&
1790 atomic_read(&pool
->nr_running
));
1794 * worker_leave_idle - leave idle state
1795 * @worker: worker which is leaving idle state
1797 * @worker is leaving idle state. Update stats.
1800 * spin_lock_irq(pool->lock).
1802 static void worker_leave_idle(struct worker
*worker
)
1804 struct worker_pool
*pool
= worker
->pool
;
1806 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1808 worker_clr_flags(worker
, WORKER_IDLE
);
1810 list_del_init(&worker
->entry
);
1813 static struct worker
*alloc_worker(int node
)
1815 struct worker
*worker
;
1817 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1819 INIT_LIST_HEAD(&worker
->entry
);
1820 INIT_LIST_HEAD(&worker
->scheduled
);
1821 INIT_LIST_HEAD(&worker
->node
);
1822 /* on creation a worker is in !idle && prep state */
1823 worker
->flags
= WORKER_PREP
;
1829 * worker_attach_to_pool() - attach a worker to a pool
1830 * @worker: worker to be attached
1831 * @pool: the target pool
1833 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1834 * cpu-binding of @worker are kept coordinated with the pool across
1837 static void worker_attach_to_pool(struct worker
*worker
,
1838 struct worker_pool
*pool
)
1840 mutex_lock(&wq_pool_attach_mutex
);
1843 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1844 * online CPUs. It'll be re-applied when any of the CPUs come up.
1846 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1849 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1850 * stable across this function. See the comments above the flag
1851 * definition for details.
1853 if (pool
->flags
& POOL_DISASSOCIATED
)
1854 worker
->flags
|= WORKER_UNBOUND
;
1856 list_add_tail(&worker
->node
, &pool
->workers
);
1857 worker
->pool
= pool
;
1859 mutex_unlock(&wq_pool_attach_mutex
);
1863 * worker_detach_from_pool() - detach a worker from its pool
1864 * @worker: worker which is attached to its pool
1866 * Undo the attaching which had been done in worker_attach_to_pool(). The
1867 * caller worker shouldn't access to the pool after detached except it has
1868 * other reference to the pool.
1870 static void worker_detach_from_pool(struct worker
*worker
)
1872 struct worker_pool
*pool
= worker
->pool
;
1873 struct completion
*detach_completion
= NULL
;
1875 mutex_lock(&wq_pool_attach_mutex
);
1877 list_del(&worker
->node
);
1878 worker
->pool
= NULL
;
1880 if (list_empty(&pool
->workers
))
1881 detach_completion
= pool
->detach_completion
;
1882 mutex_unlock(&wq_pool_attach_mutex
);
1884 /* clear leftover flags without pool->lock after it is detached */
1885 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1887 if (detach_completion
)
1888 complete(detach_completion
);
1892 * create_worker - create a new workqueue worker
1893 * @pool: pool the new worker will belong to
1895 * Create and start a new worker which is attached to @pool.
1898 * Might sleep. Does GFP_KERNEL allocations.
1901 * Pointer to the newly created worker.
1903 static struct worker
*create_worker(struct worker_pool
*pool
)
1905 struct worker
*worker
= NULL
;
1909 /* ID is needed to determine kthread name */
1910 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1914 worker
= alloc_worker(pool
->node
);
1921 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1922 pool
->attrs
->nice
< 0 ? "H" : "");
1924 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1926 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1927 "kworker/%s", id_buf
);
1928 if (IS_ERR(worker
->task
))
1931 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1932 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1934 /* successful, attach the worker to the pool */
1935 worker_attach_to_pool(worker
, pool
);
1937 /* start the newly created worker */
1938 spin_lock_irq(&pool
->lock
);
1939 worker
->pool
->nr_workers
++;
1940 worker_enter_idle(worker
);
1941 wake_up_process(worker
->task
);
1942 spin_unlock_irq(&pool
->lock
);
1948 ida_simple_remove(&pool
->worker_ida
, id
);
1954 * destroy_worker - destroy a workqueue worker
1955 * @worker: worker to be destroyed
1957 * Destroy @worker and adjust @pool stats accordingly. The worker should
1961 * spin_lock_irq(pool->lock).
1963 static void destroy_worker(struct worker
*worker
)
1965 struct worker_pool
*pool
= worker
->pool
;
1967 lockdep_assert_held(&pool
->lock
);
1969 /* sanity check frenzy */
1970 if (WARN_ON(worker
->current_work
) ||
1971 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1972 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1978 list_del_init(&worker
->entry
);
1979 worker
->flags
|= WORKER_DIE
;
1980 wake_up_process(worker
->task
);
1983 static void idle_worker_timeout(struct timer_list
*t
)
1985 struct worker_pool
*pool
= from_timer(pool
, t
, idle_timer
);
1987 spin_lock_irq(&pool
->lock
);
1989 while (too_many_workers(pool
)) {
1990 struct worker
*worker
;
1991 unsigned long expires
;
1993 /* idle_list is kept in LIFO order, check the last one */
1994 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1995 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1997 if (time_before(jiffies
, expires
)) {
1998 mod_timer(&pool
->idle_timer
, expires
);
2002 destroy_worker(worker
);
2005 spin_unlock_irq(&pool
->lock
);
2008 static void send_mayday(struct work_struct
*work
)
2010 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2011 struct workqueue_struct
*wq
= pwq
->wq
;
2013 lockdep_assert_held(&wq_mayday_lock
);
2018 /* mayday mayday mayday */
2019 if (list_empty(&pwq
->mayday_node
)) {
2021 * If @pwq is for an unbound wq, its base ref may be put at
2022 * any time due to an attribute change. Pin @pwq until the
2023 * rescuer is done with it.
2026 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2027 wake_up_process(wq
->rescuer
->task
);
2031 static void pool_mayday_timeout(struct timer_list
*t
)
2033 struct worker_pool
*pool
= from_timer(pool
, t
, mayday_timer
);
2034 struct work_struct
*work
;
2036 spin_lock_irq(&pool
->lock
);
2037 spin_lock(&wq_mayday_lock
); /* for wq->maydays */
2039 if (need_to_create_worker(pool
)) {
2041 * We've been trying to create a new worker but
2042 * haven't been successful. We might be hitting an
2043 * allocation deadlock. Send distress signals to
2046 list_for_each_entry(work
, &pool
->worklist
, entry
)
2050 spin_unlock(&wq_mayday_lock
);
2051 spin_unlock_irq(&pool
->lock
);
2053 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
2057 * maybe_create_worker - create a new worker if necessary
2058 * @pool: pool to create a new worker for
2060 * Create a new worker for @pool if necessary. @pool is guaranteed to
2061 * have at least one idle worker on return from this function. If
2062 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2063 * sent to all rescuers with works scheduled on @pool to resolve
2064 * possible allocation deadlock.
2066 * On return, need_to_create_worker() is guaranteed to be %false and
2067 * may_start_working() %true.
2070 * spin_lock_irq(pool->lock) which may be released and regrabbed
2071 * multiple times. Does GFP_KERNEL allocations. Called only from
2074 static void maybe_create_worker(struct worker_pool
*pool
)
2075 __releases(&pool
->lock
)
2076 __acquires(&pool
->lock
)
2079 spin_unlock_irq(&pool
->lock
);
2081 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2082 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
2085 if (create_worker(pool
) || !need_to_create_worker(pool
))
2088 schedule_timeout_interruptible(CREATE_COOLDOWN
);
2090 if (!need_to_create_worker(pool
))
2094 del_timer_sync(&pool
->mayday_timer
);
2095 spin_lock_irq(&pool
->lock
);
2097 * This is necessary even after a new worker was just successfully
2098 * created as @pool->lock was dropped and the new worker might have
2099 * already become busy.
2101 if (need_to_create_worker(pool
))
2106 * manage_workers - manage worker pool
2109 * Assume the manager role and manage the worker pool @worker belongs
2110 * to. At any given time, there can be only zero or one manager per
2111 * pool. The exclusion is handled automatically by this function.
2113 * The caller can safely start processing works on false return. On
2114 * true return, it's guaranteed that need_to_create_worker() is false
2115 * and may_start_working() is true.
2118 * spin_lock_irq(pool->lock) which may be released and regrabbed
2119 * multiple times. Does GFP_KERNEL allocations.
2122 * %false if the pool doesn't need management and the caller can safely
2123 * start processing works, %true if management function was performed and
2124 * the conditions that the caller verified before calling the function may
2125 * no longer be true.
2127 static bool manage_workers(struct worker
*worker
)
2129 struct worker_pool
*pool
= worker
->pool
;
2131 if (pool
->flags
& POOL_MANAGER_ACTIVE
)
2134 pool
->flags
|= POOL_MANAGER_ACTIVE
;
2135 pool
->manager
= worker
;
2137 maybe_create_worker(pool
);
2139 pool
->manager
= NULL
;
2140 pool
->flags
&= ~POOL_MANAGER_ACTIVE
;
2141 wake_up(&wq_manager_wait
);
2146 * process_one_work - process single work
2148 * @work: work to process
2150 * Process @work. This function contains all the logics necessary to
2151 * process a single work including synchronization against and
2152 * interaction with other workers on the same cpu, queueing and
2153 * flushing. As long as context requirement is met, any worker can
2154 * call this function to process a work.
2157 * spin_lock_irq(pool->lock) which is released and regrabbed.
2159 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2160 __releases(&pool
->lock
)
2161 __acquires(&pool
->lock
)
2163 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2164 struct worker_pool
*pool
= worker
->pool
;
2165 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2167 struct worker
*collision
;
2168 #ifdef CONFIG_LOCKDEP
2170 * It is permissible to free the struct work_struct from
2171 * inside the function that is called from it, this we need to
2172 * take into account for lockdep too. To avoid bogus "held
2173 * lock freed" warnings as well as problems when looking into
2174 * work->lockdep_map, make a copy and use that here.
2176 struct lockdep_map lockdep_map
;
2178 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2180 /* ensure we're on the correct CPU */
2181 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2182 raw_smp_processor_id() != pool
->cpu
);
2185 * A single work shouldn't be executed concurrently by
2186 * multiple workers on a single cpu. Check whether anyone is
2187 * already processing the work. If so, defer the work to the
2188 * currently executing one.
2190 collision
= find_worker_executing_work(pool
, work
);
2191 if (unlikely(collision
)) {
2192 move_linked_works(work
, &collision
->scheduled
, NULL
);
2196 /* claim and dequeue */
2197 debug_work_deactivate(work
);
2198 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2199 worker
->current_work
= work
;
2200 worker
->current_func
= work
->func
;
2201 worker
->current_pwq
= pwq
;
2202 work_color
= get_work_color(work
);
2205 * Record wq name for cmdline and debug reporting, may get
2206 * overridden through set_worker_desc().
2208 strscpy(worker
->desc
, pwq
->wq
->name
, WORKER_DESC_LEN
);
2210 list_del_init(&work
->entry
);
2213 * CPU intensive works don't participate in concurrency management.
2214 * They're the scheduler's responsibility. This takes @worker out
2215 * of concurrency management and the next code block will chain
2216 * execution of the pending work items.
2218 if (unlikely(cpu_intensive
))
2219 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2222 * Wake up another worker if necessary. The condition is always
2223 * false for normal per-cpu workers since nr_running would always
2224 * be >= 1 at this point. This is used to chain execution of the
2225 * pending work items for WORKER_NOT_RUNNING workers such as the
2226 * UNBOUND and CPU_INTENSIVE ones.
2228 if (need_more_worker(pool
))
2229 wake_up_worker(pool
);
2232 * Record the last pool and clear PENDING which should be the last
2233 * update to @work. Also, do this inside @pool->lock so that
2234 * PENDING and queued state changes happen together while IRQ is
2237 set_work_pool_and_clear_pending(work
, pool
->id
);
2239 spin_unlock_irq(&pool
->lock
);
2241 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2242 lock_map_acquire(&lockdep_map
);
2244 * Strictly speaking we should mark the invariant state without holding
2245 * any locks, that is, before these two lock_map_acquire()'s.
2247 * However, that would result in:
2254 * Which would create W1->C->W1 dependencies, even though there is no
2255 * actual deadlock possible. There are two solutions, using a
2256 * read-recursive acquire on the work(queue) 'locks', but this will then
2257 * hit the lockdep limitation on recursive locks, or simply discard
2260 * AFAICT there is no possible deadlock scenario between the
2261 * flush_work() and complete() primitives (except for single-threaded
2262 * workqueues), so hiding them isn't a problem.
2264 lockdep_invariant_state(true);
2265 trace_workqueue_execute_start(work
);
2266 worker
->current_func(work
);
2268 * While we must be careful to not use "work" after this, the trace
2269 * point will only record its address.
2271 trace_workqueue_execute_end(work
, worker
->current_func
);
2272 lock_map_release(&lockdep_map
);
2273 lock_map_release(&pwq
->wq
->lockdep_map
);
2275 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2276 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2277 " last function: %ps\n",
2278 current
->comm
, preempt_count(), task_pid_nr(current
),
2279 worker
->current_func
);
2280 debug_show_held_locks(current
);
2285 * The following prevents a kworker from hogging CPU on !PREEMPTION
2286 * kernels, where a requeueing work item waiting for something to
2287 * happen could deadlock with stop_machine as such work item could
2288 * indefinitely requeue itself while all other CPUs are trapped in
2289 * stop_machine. At the same time, report a quiescent RCU state so
2290 * the same condition doesn't freeze RCU.
2294 spin_lock_irq(&pool
->lock
);
2296 /* clear cpu intensive status */
2297 if (unlikely(cpu_intensive
))
2298 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2300 /* tag the worker for identification in schedule() */
2301 worker
->last_func
= worker
->current_func
;
2303 /* we're done with it, release */
2304 hash_del(&worker
->hentry
);
2305 worker
->current_work
= NULL
;
2306 worker
->current_func
= NULL
;
2307 worker
->current_pwq
= NULL
;
2308 pwq_dec_nr_in_flight(pwq
, work_color
);
2312 * process_scheduled_works - process scheduled works
2315 * Process all scheduled works. Please note that the scheduled list
2316 * may change while processing a work, so this function repeatedly
2317 * fetches a work from the top and executes it.
2320 * spin_lock_irq(pool->lock) which may be released and regrabbed
2323 static void process_scheduled_works(struct worker
*worker
)
2325 while (!list_empty(&worker
->scheduled
)) {
2326 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2327 struct work_struct
, entry
);
2328 process_one_work(worker
, work
);
2332 static void set_pf_worker(bool val
)
2334 mutex_lock(&wq_pool_attach_mutex
);
2336 current
->flags
|= PF_WQ_WORKER
;
2338 current
->flags
&= ~PF_WQ_WORKER
;
2339 mutex_unlock(&wq_pool_attach_mutex
);
2343 * worker_thread - the worker thread function
2346 * The worker thread function. All workers belong to a worker_pool -
2347 * either a per-cpu one or dynamic unbound one. These workers process all
2348 * work items regardless of their specific target workqueue. The only
2349 * exception is work items which belong to workqueues with a rescuer which
2350 * will be explained in rescuer_thread().
2354 static int worker_thread(void *__worker
)
2356 struct worker
*worker
= __worker
;
2357 struct worker_pool
*pool
= worker
->pool
;
2359 /* tell the scheduler that this is a workqueue worker */
2360 set_pf_worker(true);
2362 spin_lock_irq(&pool
->lock
);
2364 /* am I supposed to die? */
2365 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2366 spin_unlock_irq(&pool
->lock
);
2367 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2368 set_pf_worker(false);
2370 set_task_comm(worker
->task
, "kworker/dying");
2371 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2372 worker_detach_from_pool(worker
);
2377 worker_leave_idle(worker
);
2379 /* no more worker necessary? */
2380 if (!need_more_worker(pool
))
2383 /* do we need to manage? */
2384 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2388 * ->scheduled list can only be filled while a worker is
2389 * preparing to process a work or actually processing it.
2390 * Make sure nobody diddled with it while I was sleeping.
2392 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2395 * Finish PREP stage. We're guaranteed to have at least one idle
2396 * worker or that someone else has already assumed the manager
2397 * role. This is where @worker starts participating in concurrency
2398 * management if applicable and concurrency management is restored
2399 * after being rebound. See rebind_workers() for details.
2401 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2404 struct work_struct
*work
=
2405 list_first_entry(&pool
->worklist
,
2406 struct work_struct
, entry
);
2408 pool
->watchdog_ts
= jiffies
;
2410 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2411 /* optimization path, not strictly necessary */
2412 process_one_work(worker
, work
);
2413 if (unlikely(!list_empty(&worker
->scheduled
)))
2414 process_scheduled_works(worker
);
2416 move_linked_works(work
, &worker
->scheduled
, NULL
);
2417 process_scheduled_works(worker
);
2419 } while (keep_working(pool
));
2421 worker_set_flags(worker
, WORKER_PREP
);
2424 * pool->lock is held and there's no work to process and no need to
2425 * manage, sleep. Workers are woken up only while holding
2426 * pool->lock or from local cpu, so setting the current state
2427 * before releasing pool->lock is enough to prevent losing any
2430 worker_enter_idle(worker
);
2431 __set_current_state(TASK_IDLE
);
2432 spin_unlock_irq(&pool
->lock
);
2438 * rescuer_thread - the rescuer thread function
2441 * Workqueue rescuer thread function. There's one rescuer for each
2442 * workqueue which has WQ_MEM_RECLAIM set.
2444 * Regular work processing on a pool may block trying to create a new
2445 * worker which uses GFP_KERNEL allocation which has slight chance of
2446 * developing into deadlock if some works currently on the same queue
2447 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2448 * the problem rescuer solves.
2450 * When such condition is possible, the pool summons rescuers of all
2451 * workqueues which have works queued on the pool and let them process
2452 * those works so that forward progress can be guaranteed.
2454 * This should happen rarely.
2458 static int rescuer_thread(void *__rescuer
)
2460 struct worker
*rescuer
= __rescuer
;
2461 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2462 struct list_head
*scheduled
= &rescuer
->scheduled
;
2465 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2468 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2469 * doesn't participate in concurrency management.
2471 set_pf_worker(true);
2473 set_current_state(TASK_IDLE
);
2476 * By the time the rescuer is requested to stop, the workqueue
2477 * shouldn't have any work pending, but @wq->maydays may still have
2478 * pwq(s) queued. This can happen by non-rescuer workers consuming
2479 * all the work items before the rescuer got to them. Go through
2480 * @wq->maydays processing before acting on should_stop so that the
2481 * list is always empty on exit.
2483 should_stop
= kthread_should_stop();
2485 /* see whether any pwq is asking for help */
2486 spin_lock_irq(&wq_mayday_lock
);
2488 while (!list_empty(&wq
->maydays
)) {
2489 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2490 struct pool_workqueue
, mayday_node
);
2491 struct worker_pool
*pool
= pwq
->pool
;
2492 struct work_struct
*work
, *n
;
2495 __set_current_state(TASK_RUNNING
);
2496 list_del_init(&pwq
->mayday_node
);
2498 spin_unlock_irq(&wq_mayday_lock
);
2500 worker_attach_to_pool(rescuer
, pool
);
2502 spin_lock_irq(&pool
->lock
);
2505 * Slurp in all works issued via this workqueue and
2508 WARN_ON_ONCE(!list_empty(scheduled
));
2509 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2510 if (get_work_pwq(work
) == pwq
) {
2512 pool
->watchdog_ts
= jiffies
;
2513 move_linked_works(work
, scheduled
, &n
);
2518 if (!list_empty(scheduled
)) {
2519 process_scheduled_works(rescuer
);
2522 * The above execution of rescued work items could
2523 * have created more to rescue through
2524 * pwq_activate_first_delayed() or chained
2525 * queueing. Let's put @pwq back on mayday list so
2526 * that such back-to-back work items, which may be
2527 * being used to relieve memory pressure, don't
2528 * incur MAYDAY_INTERVAL delay inbetween.
2530 if (need_to_create_worker(pool
)) {
2531 spin_lock(&wq_mayday_lock
);
2533 * Queue iff we aren't racing destruction
2534 * and somebody else hasn't queued it already.
2536 if (wq
->rescuer
&& list_empty(&pwq
->mayday_node
)) {
2538 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2540 spin_unlock(&wq_mayday_lock
);
2545 * Put the reference grabbed by send_mayday(). @pool won't
2546 * go away while we're still attached to it.
2551 * Leave this pool. If need_more_worker() is %true, notify a
2552 * regular worker; otherwise, we end up with 0 concurrency
2553 * and stalling the execution.
2555 if (need_more_worker(pool
))
2556 wake_up_worker(pool
);
2558 spin_unlock_irq(&pool
->lock
);
2560 worker_detach_from_pool(rescuer
);
2562 spin_lock_irq(&wq_mayday_lock
);
2565 spin_unlock_irq(&wq_mayday_lock
);
2568 __set_current_state(TASK_RUNNING
);
2569 set_pf_worker(false);
2573 /* rescuers should never participate in concurrency management */
2574 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2580 * check_flush_dependency - check for flush dependency sanity
2581 * @target_wq: workqueue being flushed
2582 * @target_work: work item being flushed (NULL for workqueue flushes)
2584 * %current is trying to flush the whole @target_wq or @target_work on it.
2585 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2586 * reclaiming memory or running on a workqueue which doesn't have
2587 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2590 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2591 struct work_struct
*target_work
)
2593 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2594 struct worker
*worker
;
2596 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2599 worker
= current_wq_worker();
2601 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2602 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2603 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2604 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2605 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2606 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2607 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2608 target_wq
->name
, target_func
);
2612 struct work_struct work
;
2613 struct completion done
;
2614 struct task_struct
*task
; /* purely informational */
2617 static void wq_barrier_func(struct work_struct
*work
)
2619 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2620 complete(&barr
->done
);
2624 * insert_wq_barrier - insert a barrier work
2625 * @pwq: pwq to insert barrier into
2626 * @barr: wq_barrier to insert
2627 * @target: target work to attach @barr to
2628 * @worker: worker currently executing @target, NULL if @target is not executing
2630 * @barr is linked to @target such that @barr is completed only after
2631 * @target finishes execution. Please note that the ordering
2632 * guarantee is observed only with respect to @target and on the local
2635 * Currently, a queued barrier can't be canceled. This is because
2636 * try_to_grab_pending() can't determine whether the work to be
2637 * grabbed is at the head of the queue and thus can't clear LINKED
2638 * flag of the previous work while there must be a valid next work
2639 * after a work with LINKED flag set.
2641 * Note that when @worker is non-NULL, @target may be modified
2642 * underneath us, so we can't reliably determine pwq from @target.
2645 * spin_lock_irq(pool->lock).
2647 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2648 struct wq_barrier
*barr
,
2649 struct work_struct
*target
, struct worker
*worker
)
2651 struct list_head
*head
;
2652 unsigned int linked
= 0;
2655 * debugobject calls are safe here even with pool->lock locked
2656 * as we know for sure that this will not trigger any of the
2657 * checks and call back into the fixup functions where we
2660 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2661 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2663 init_completion_map(&barr
->done
, &target
->lockdep_map
);
2665 barr
->task
= current
;
2668 * If @target is currently being executed, schedule the
2669 * barrier to the worker; otherwise, put it after @target.
2672 head
= worker
->scheduled
.next
;
2674 unsigned long *bits
= work_data_bits(target
);
2676 head
= target
->entry
.next
;
2677 /* there can already be other linked works, inherit and set */
2678 linked
= *bits
& WORK_STRUCT_LINKED
;
2679 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2682 debug_work_activate(&barr
->work
);
2683 insert_work(pwq
, &barr
->work
, head
,
2684 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2688 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2689 * @wq: workqueue being flushed
2690 * @flush_color: new flush color, < 0 for no-op
2691 * @work_color: new work color, < 0 for no-op
2693 * Prepare pwqs for workqueue flushing.
2695 * If @flush_color is non-negative, flush_color on all pwqs should be
2696 * -1. If no pwq has in-flight commands at the specified color, all
2697 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2698 * has in flight commands, its pwq->flush_color is set to
2699 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2700 * wakeup logic is armed and %true is returned.
2702 * The caller should have initialized @wq->first_flusher prior to
2703 * calling this function with non-negative @flush_color. If
2704 * @flush_color is negative, no flush color update is done and %false
2707 * If @work_color is non-negative, all pwqs should have the same
2708 * work_color which is previous to @work_color and all will be
2709 * advanced to @work_color.
2712 * mutex_lock(wq->mutex).
2715 * %true if @flush_color >= 0 and there's something to flush. %false
2718 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2719 int flush_color
, int work_color
)
2722 struct pool_workqueue
*pwq
;
2724 if (flush_color
>= 0) {
2725 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2726 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2729 for_each_pwq(pwq
, wq
) {
2730 struct worker_pool
*pool
= pwq
->pool
;
2732 spin_lock_irq(&pool
->lock
);
2734 if (flush_color
>= 0) {
2735 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2737 if (pwq
->nr_in_flight
[flush_color
]) {
2738 pwq
->flush_color
= flush_color
;
2739 atomic_inc(&wq
->nr_pwqs_to_flush
);
2744 if (work_color
>= 0) {
2745 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2746 pwq
->work_color
= work_color
;
2749 spin_unlock_irq(&pool
->lock
);
2752 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2753 complete(&wq
->first_flusher
->done
);
2759 * flush_workqueue - ensure that any scheduled work has run to completion.
2760 * @wq: workqueue to flush
2762 * This function sleeps until all work items which were queued on entry
2763 * have finished execution, but it is not livelocked by new incoming ones.
2765 void flush_workqueue(struct workqueue_struct
*wq
)
2767 struct wq_flusher this_flusher
= {
2768 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2770 .done
= COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher
.done
, wq
->lockdep_map
),
2774 if (WARN_ON(!wq_online
))
2777 lock_map_acquire(&wq
->lockdep_map
);
2778 lock_map_release(&wq
->lockdep_map
);
2780 mutex_lock(&wq
->mutex
);
2783 * Start-to-wait phase
2785 next_color
= work_next_color(wq
->work_color
);
2787 if (next_color
!= wq
->flush_color
) {
2789 * Color space is not full. The current work_color
2790 * becomes our flush_color and work_color is advanced
2793 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2794 this_flusher
.flush_color
= wq
->work_color
;
2795 wq
->work_color
= next_color
;
2797 if (!wq
->first_flusher
) {
2798 /* no flush in progress, become the first flusher */
2799 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2801 wq
->first_flusher
= &this_flusher
;
2803 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2805 /* nothing to flush, done */
2806 wq
->flush_color
= next_color
;
2807 wq
->first_flusher
= NULL
;
2812 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2813 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2814 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2818 * Oops, color space is full, wait on overflow queue.
2819 * The next flush completion will assign us
2820 * flush_color and transfer to flusher_queue.
2822 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2825 check_flush_dependency(wq
, NULL
);
2827 mutex_unlock(&wq
->mutex
);
2829 wait_for_completion(&this_flusher
.done
);
2832 * Wake-up-and-cascade phase
2834 * First flushers are responsible for cascading flushes and
2835 * handling overflow. Non-first flushers can simply return.
2837 if (wq
->first_flusher
!= &this_flusher
)
2840 mutex_lock(&wq
->mutex
);
2842 /* we might have raced, check again with mutex held */
2843 if (wq
->first_flusher
!= &this_flusher
)
2846 wq
->first_flusher
= NULL
;
2848 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2849 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2852 struct wq_flusher
*next
, *tmp
;
2854 /* complete all the flushers sharing the current flush color */
2855 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2856 if (next
->flush_color
!= wq
->flush_color
)
2858 list_del_init(&next
->list
);
2859 complete(&next
->done
);
2862 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2863 wq
->flush_color
!= work_next_color(wq
->work_color
));
2865 /* this flush_color is finished, advance by one */
2866 wq
->flush_color
= work_next_color(wq
->flush_color
);
2868 /* one color has been freed, handle overflow queue */
2869 if (!list_empty(&wq
->flusher_overflow
)) {
2871 * Assign the same color to all overflowed
2872 * flushers, advance work_color and append to
2873 * flusher_queue. This is the start-to-wait
2874 * phase for these overflowed flushers.
2876 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2877 tmp
->flush_color
= wq
->work_color
;
2879 wq
->work_color
= work_next_color(wq
->work_color
);
2881 list_splice_tail_init(&wq
->flusher_overflow
,
2882 &wq
->flusher_queue
);
2883 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2886 if (list_empty(&wq
->flusher_queue
)) {
2887 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2892 * Need to flush more colors. Make the next flusher
2893 * the new first flusher and arm pwqs.
2895 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2896 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2898 list_del_init(&next
->list
);
2899 wq
->first_flusher
= next
;
2901 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2905 * Meh... this color is already done, clear first
2906 * flusher and repeat cascading.
2908 wq
->first_flusher
= NULL
;
2912 mutex_unlock(&wq
->mutex
);
2914 EXPORT_SYMBOL(flush_workqueue
);
2917 * drain_workqueue - drain a workqueue
2918 * @wq: workqueue to drain
2920 * Wait until the workqueue becomes empty. While draining is in progress,
2921 * only chain queueing is allowed. IOW, only currently pending or running
2922 * work items on @wq can queue further work items on it. @wq is flushed
2923 * repeatedly until it becomes empty. The number of flushing is determined
2924 * by the depth of chaining and should be relatively short. Whine if it
2927 void drain_workqueue(struct workqueue_struct
*wq
)
2929 unsigned int flush_cnt
= 0;
2930 struct pool_workqueue
*pwq
;
2933 * __queue_work() needs to test whether there are drainers, is much
2934 * hotter than drain_workqueue() and already looks at @wq->flags.
2935 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2937 mutex_lock(&wq
->mutex
);
2938 if (!wq
->nr_drainers
++)
2939 wq
->flags
|= __WQ_DRAINING
;
2940 mutex_unlock(&wq
->mutex
);
2942 flush_workqueue(wq
);
2944 mutex_lock(&wq
->mutex
);
2946 for_each_pwq(pwq
, wq
) {
2949 spin_lock_irq(&pwq
->pool
->lock
);
2950 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2951 spin_unlock_irq(&pwq
->pool
->lock
);
2956 if (++flush_cnt
== 10 ||
2957 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2958 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2959 wq
->name
, flush_cnt
);
2961 mutex_unlock(&wq
->mutex
);
2965 if (!--wq
->nr_drainers
)
2966 wq
->flags
&= ~__WQ_DRAINING
;
2967 mutex_unlock(&wq
->mutex
);
2969 EXPORT_SYMBOL_GPL(drain_workqueue
);
2971 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2974 struct worker
*worker
= NULL
;
2975 struct worker_pool
*pool
;
2976 struct pool_workqueue
*pwq
;
2981 pool
= get_work_pool(work
);
2987 spin_lock_irq(&pool
->lock
);
2988 /* see the comment in try_to_grab_pending() with the same code */
2989 pwq
= get_work_pwq(work
);
2991 if (unlikely(pwq
->pool
!= pool
))
2994 worker
= find_worker_executing_work(pool
, work
);
2997 pwq
= worker
->current_pwq
;
3000 check_flush_dependency(pwq
->wq
, work
);
3002 insert_wq_barrier(pwq
, barr
, work
, worker
);
3003 spin_unlock_irq(&pool
->lock
);
3006 * Force a lock recursion deadlock when using flush_work() inside a
3007 * single-threaded or rescuer equipped workqueue.
3009 * For single threaded workqueues the deadlock happens when the work
3010 * is after the work issuing the flush_work(). For rescuer equipped
3011 * workqueues the deadlock happens when the rescuer stalls, blocking
3015 (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)) {
3016 lock_map_acquire(&pwq
->wq
->lockdep_map
);
3017 lock_map_release(&pwq
->wq
->lockdep_map
);
3022 spin_unlock_irq(&pool
->lock
);
3027 static bool __flush_work(struct work_struct
*work
, bool from_cancel
)
3029 struct wq_barrier barr
;
3031 if (WARN_ON(!wq_online
))
3034 if (WARN_ON(!work
->func
))
3038 lock_map_acquire(&work
->lockdep_map
);
3039 lock_map_release(&work
->lockdep_map
);
3042 if (start_flush_work(work
, &barr
, from_cancel
)) {
3043 wait_for_completion(&barr
.done
);
3044 destroy_work_on_stack(&barr
.work
);
3052 * flush_work - wait for a work to finish executing the last queueing instance
3053 * @work: the work to flush
3055 * Wait until @work has finished execution. @work is guaranteed to be idle
3056 * on return if it hasn't been requeued since flush started.
3059 * %true if flush_work() waited for the work to finish execution,
3060 * %false if it was already idle.
3062 bool flush_work(struct work_struct
*work
)
3064 return __flush_work(work
, false);
3066 EXPORT_SYMBOL_GPL(flush_work
);
3069 wait_queue_entry_t wait
;
3070 struct work_struct
*work
;
3073 static int cwt_wakefn(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
3075 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
3077 if (cwait
->work
!= key
)
3079 return autoremove_wake_function(wait
, mode
, sync
, key
);
3082 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
3084 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
3085 unsigned long flags
;
3089 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3091 * If someone else is already canceling, wait for it to
3092 * finish. flush_work() doesn't work for PREEMPT_NONE
3093 * because we may get scheduled between @work's completion
3094 * and the other canceling task resuming and clearing
3095 * CANCELING - flush_work() will return false immediately
3096 * as @work is no longer busy, try_to_grab_pending() will
3097 * return -ENOENT as @work is still being canceled and the
3098 * other canceling task won't be able to clear CANCELING as
3099 * we're hogging the CPU.
3101 * Let's wait for completion using a waitqueue. As this
3102 * may lead to the thundering herd problem, use a custom
3103 * wake function which matches @work along with exclusive
3106 if (unlikely(ret
== -ENOENT
)) {
3107 struct cwt_wait cwait
;
3109 init_wait(&cwait
.wait
);
3110 cwait
.wait
.func
= cwt_wakefn
;
3113 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
3114 TASK_UNINTERRUPTIBLE
);
3115 if (work_is_canceling(work
))
3117 finish_wait(&cancel_waitq
, &cwait
.wait
);
3119 } while (unlikely(ret
< 0));
3121 /* tell other tasks trying to grab @work to back off */
3122 mark_work_canceling(work
);
3123 local_irq_restore(flags
);
3126 * This allows canceling during early boot. We know that @work
3130 __flush_work(work
, true);
3132 clear_work_data(work
);
3135 * Paired with prepare_to_wait() above so that either
3136 * waitqueue_active() is visible here or !work_is_canceling() is
3140 if (waitqueue_active(&cancel_waitq
))
3141 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
3147 * cancel_work_sync - cancel a work and wait for it to finish
3148 * @work: the work to cancel
3150 * Cancel @work and wait for its execution to finish. This function
3151 * can be used even if the work re-queues itself or migrates to
3152 * another workqueue. On return from this function, @work is
3153 * guaranteed to be not pending or executing on any CPU.
3155 * cancel_work_sync(&delayed_work->work) must not be used for
3156 * delayed_work's. Use cancel_delayed_work_sync() instead.
3158 * The caller must ensure that the workqueue on which @work was last
3159 * queued can't be destroyed before this function returns.
3162 * %true if @work was pending, %false otherwise.
3164 bool cancel_work_sync(struct work_struct
*work
)
3166 return __cancel_work_timer(work
, false);
3168 EXPORT_SYMBOL_GPL(cancel_work_sync
);
3171 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3172 * @dwork: the delayed work to flush
3174 * Delayed timer is cancelled and the pending work is queued for
3175 * immediate execution. Like flush_work(), this function only
3176 * considers the last queueing instance of @dwork.
3179 * %true if flush_work() waited for the work to finish execution,
3180 * %false if it was already idle.
3182 bool flush_delayed_work(struct delayed_work
*dwork
)
3184 local_irq_disable();
3185 if (del_timer_sync(&dwork
->timer
))
3186 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3188 return flush_work(&dwork
->work
);
3190 EXPORT_SYMBOL(flush_delayed_work
);
3193 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3194 * @rwork: the rcu work to flush
3197 * %true if flush_rcu_work() waited for the work to finish execution,
3198 * %false if it was already idle.
3200 bool flush_rcu_work(struct rcu_work
*rwork
)
3202 if (test_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&rwork
->work
))) {
3204 flush_work(&rwork
->work
);
3207 return flush_work(&rwork
->work
);
3210 EXPORT_SYMBOL(flush_rcu_work
);
3212 static bool __cancel_work(struct work_struct
*work
, bool is_dwork
)
3214 unsigned long flags
;
3218 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3219 } while (unlikely(ret
== -EAGAIN
));
3221 if (unlikely(ret
< 0))
3224 set_work_pool_and_clear_pending(work
, get_work_pool_id(work
));
3225 local_irq_restore(flags
);
3230 * cancel_delayed_work - cancel a delayed work
3231 * @dwork: delayed_work to cancel
3233 * Kill off a pending delayed_work.
3235 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3239 * The work callback function may still be running on return, unless
3240 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3241 * use cancel_delayed_work_sync() to wait on it.
3243 * This function is safe to call from any context including IRQ handler.
3245 bool cancel_delayed_work(struct delayed_work
*dwork
)
3247 return __cancel_work(&dwork
->work
, true);
3249 EXPORT_SYMBOL(cancel_delayed_work
);
3252 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3253 * @dwork: the delayed work cancel
3255 * This is cancel_work_sync() for delayed works.
3258 * %true if @dwork was pending, %false otherwise.
3260 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3262 return __cancel_work_timer(&dwork
->work
, true);
3264 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3267 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3268 * @func: the function to call
3270 * schedule_on_each_cpu() executes @func on each online CPU using the
3271 * system workqueue and blocks until all CPUs have completed.
3272 * schedule_on_each_cpu() is very slow.
3275 * 0 on success, -errno on failure.
3277 int schedule_on_each_cpu(work_func_t func
)
3280 struct work_struct __percpu
*works
;
3282 works
= alloc_percpu(struct work_struct
);
3288 for_each_online_cpu(cpu
) {
3289 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3291 INIT_WORK(work
, func
);
3292 schedule_work_on(cpu
, work
);
3295 for_each_online_cpu(cpu
)
3296 flush_work(per_cpu_ptr(works
, cpu
));
3304 * execute_in_process_context - reliably execute the routine with user context
3305 * @fn: the function to execute
3306 * @ew: guaranteed storage for the execute work structure (must
3307 * be available when the work executes)
3309 * Executes the function immediately if process context is available,
3310 * otherwise schedules the function for delayed execution.
3312 * Return: 0 - function was executed
3313 * 1 - function was scheduled for execution
3315 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3317 if (!in_interrupt()) {
3322 INIT_WORK(&ew
->work
, fn
);
3323 schedule_work(&ew
->work
);
3327 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3330 * free_workqueue_attrs - free a workqueue_attrs
3331 * @attrs: workqueue_attrs to free
3333 * Undo alloc_workqueue_attrs().
3335 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3338 free_cpumask_var(attrs
->cpumask
);
3344 * alloc_workqueue_attrs - allocate a workqueue_attrs
3346 * Allocate a new workqueue_attrs, initialize with default settings and
3349 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3351 struct workqueue_attrs
*alloc_workqueue_attrs(void)
3353 struct workqueue_attrs
*attrs
;
3355 attrs
= kzalloc(sizeof(*attrs
), GFP_KERNEL
);
3358 if (!alloc_cpumask_var(&attrs
->cpumask
, GFP_KERNEL
))
3361 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3364 free_workqueue_attrs(attrs
);
3368 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3369 const struct workqueue_attrs
*from
)
3371 to
->nice
= from
->nice
;
3372 cpumask_copy(to
->cpumask
, from
->cpumask
);
3374 * Unlike hash and equality test, this function doesn't ignore
3375 * ->no_numa as it is used for both pool and wq attrs. Instead,
3376 * get_unbound_pool() explicitly clears ->no_numa after copying.
3378 to
->no_numa
= from
->no_numa
;
3381 /* hash value of the content of @attr */
3382 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3386 hash
= jhash_1word(attrs
->nice
, hash
);
3387 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3388 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3392 /* content equality test */
3393 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3394 const struct workqueue_attrs
*b
)
3396 if (a
->nice
!= b
->nice
)
3398 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3404 * init_worker_pool - initialize a newly zalloc'd worker_pool
3405 * @pool: worker_pool to initialize
3407 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3409 * Return: 0 on success, -errno on failure. Even on failure, all fields
3410 * inside @pool proper are initialized and put_unbound_pool() can be called
3411 * on @pool safely to release it.
3413 static int init_worker_pool(struct worker_pool
*pool
)
3415 spin_lock_init(&pool
->lock
);
3418 pool
->node
= NUMA_NO_NODE
;
3419 pool
->flags
|= POOL_DISASSOCIATED
;
3420 pool
->watchdog_ts
= jiffies
;
3421 INIT_LIST_HEAD(&pool
->worklist
);
3422 INIT_LIST_HEAD(&pool
->idle_list
);
3423 hash_init(pool
->busy_hash
);
3425 timer_setup(&pool
->idle_timer
, idle_worker_timeout
, TIMER_DEFERRABLE
);
3427 timer_setup(&pool
->mayday_timer
, pool_mayday_timeout
, 0);
3429 INIT_LIST_HEAD(&pool
->workers
);
3431 ida_init(&pool
->worker_ida
);
3432 INIT_HLIST_NODE(&pool
->hash_node
);
3435 /* shouldn't fail above this point */
3436 pool
->attrs
= alloc_workqueue_attrs();
3442 #ifdef CONFIG_LOCKDEP
3443 static void wq_init_lockdep(struct workqueue_struct
*wq
)
3447 lockdep_register_key(&wq
->key
);
3448 lock_name
= kasprintf(GFP_KERNEL
, "%s%s", "(wq_completion)", wq
->name
);
3450 lock_name
= wq
->name
;
3452 wq
->lock_name
= lock_name
;
3453 lockdep_init_map(&wq
->lockdep_map
, lock_name
, &wq
->key
, 0);
3456 static void wq_unregister_lockdep(struct workqueue_struct
*wq
)
3458 lockdep_unregister_key(&wq
->key
);
3461 static void wq_free_lockdep(struct workqueue_struct
*wq
)
3463 if (wq
->lock_name
!= wq
->name
)
3464 kfree(wq
->lock_name
);
3467 static void wq_init_lockdep(struct workqueue_struct
*wq
)
3471 static void wq_unregister_lockdep(struct workqueue_struct
*wq
)
3475 static void wq_free_lockdep(struct workqueue_struct
*wq
)
3480 static void rcu_free_wq(struct rcu_head
*rcu
)
3482 struct workqueue_struct
*wq
=
3483 container_of(rcu
, struct workqueue_struct
, rcu
);
3485 wq_free_lockdep(wq
);
3487 if (!(wq
->flags
& WQ_UNBOUND
))
3488 free_percpu(wq
->cpu_pwqs
);
3490 free_workqueue_attrs(wq
->unbound_attrs
);
3496 static void rcu_free_pool(struct rcu_head
*rcu
)
3498 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3500 ida_destroy(&pool
->worker_ida
);
3501 free_workqueue_attrs(pool
->attrs
);
3506 * put_unbound_pool - put a worker_pool
3507 * @pool: worker_pool to put
3509 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3510 * safe manner. get_unbound_pool() calls this function on its failure path
3511 * and this function should be able to release pools which went through,
3512 * successfully or not, init_worker_pool().
3514 * Should be called with wq_pool_mutex held.
3516 static void put_unbound_pool(struct worker_pool
*pool
)
3518 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3519 struct worker
*worker
;
3521 lockdep_assert_held(&wq_pool_mutex
);
3527 if (WARN_ON(!(pool
->cpu
< 0)) ||
3528 WARN_ON(!list_empty(&pool
->worklist
)))
3531 /* release id and unhash */
3533 idr_remove(&worker_pool_idr
, pool
->id
);
3534 hash_del(&pool
->hash_node
);
3537 * Become the manager and destroy all workers. This prevents
3538 * @pool's workers from blocking on attach_mutex. We're the last
3539 * manager and @pool gets freed with the flag set.
3541 spin_lock_irq(&pool
->lock
);
3542 wait_event_lock_irq(wq_manager_wait
,
3543 !(pool
->flags
& POOL_MANAGER_ACTIVE
), pool
->lock
);
3544 pool
->flags
|= POOL_MANAGER_ACTIVE
;
3546 while ((worker
= first_idle_worker(pool
)))
3547 destroy_worker(worker
);
3548 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3549 spin_unlock_irq(&pool
->lock
);
3551 mutex_lock(&wq_pool_attach_mutex
);
3552 if (!list_empty(&pool
->workers
))
3553 pool
->detach_completion
= &detach_completion
;
3554 mutex_unlock(&wq_pool_attach_mutex
);
3556 if (pool
->detach_completion
)
3557 wait_for_completion(pool
->detach_completion
);
3559 /* shut down the timers */
3560 del_timer_sync(&pool
->idle_timer
);
3561 del_timer_sync(&pool
->mayday_timer
);
3563 /* RCU protected to allow dereferences from get_work_pool() */
3564 call_rcu(&pool
->rcu
, rcu_free_pool
);
3568 * get_unbound_pool - get a worker_pool with the specified attributes
3569 * @attrs: the attributes of the worker_pool to get
3571 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3572 * reference count and return it. If there already is a matching
3573 * worker_pool, it will be used; otherwise, this function attempts to
3576 * Should be called with wq_pool_mutex held.
3578 * Return: On success, a worker_pool with the same attributes as @attrs.
3579 * On failure, %NULL.
3581 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3583 u32 hash
= wqattrs_hash(attrs
);
3584 struct worker_pool
*pool
;
3586 int target_node
= NUMA_NO_NODE
;
3588 lockdep_assert_held(&wq_pool_mutex
);
3590 /* do we already have a matching pool? */
3591 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3592 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3598 /* if cpumask is contained inside a NUMA node, we belong to that node */
3599 if (wq_numa_enabled
) {
3600 for_each_node(node
) {
3601 if (cpumask_subset(attrs
->cpumask
,
3602 wq_numa_possible_cpumask
[node
])) {
3609 /* nope, create a new one */
3610 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3611 if (!pool
|| init_worker_pool(pool
) < 0)
3614 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3615 copy_workqueue_attrs(pool
->attrs
, attrs
);
3616 pool
->node
= target_node
;
3619 * no_numa isn't a worker_pool attribute, always clear it. See
3620 * 'struct workqueue_attrs' comments for detail.
3622 pool
->attrs
->no_numa
= false;
3624 if (worker_pool_assign_id(pool
) < 0)
3627 /* create and start the initial worker */
3628 if (wq_online
&& !create_worker(pool
))
3632 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3637 put_unbound_pool(pool
);
3641 static void rcu_free_pwq(struct rcu_head
*rcu
)
3643 kmem_cache_free(pwq_cache
,
3644 container_of(rcu
, struct pool_workqueue
, rcu
));
3648 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3649 * and needs to be destroyed.
3651 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3653 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3654 unbound_release_work
);
3655 struct workqueue_struct
*wq
= pwq
->wq
;
3656 struct worker_pool
*pool
= pwq
->pool
;
3659 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3662 mutex_lock(&wq
->mutex
);
3663 list_del_rcu(&pwq
->pwqs_node
);
3664 is_last
= list_empty(&wq
->pwqs
);
3665 mutex_unlock(&wq
->mutex
);
3667 mutex_lock(&wq_pool_mutex
);
3668 put_unbound_pool(pool
);
3669 mutex_unlock(&wq_pool_mutex
);
3671 call_rcu(&pwq
->rcu
, rcu_free_pwq
);
3674 * If we're the last pwq going away, @wq is already dead and no one
3675 * is gonna access it anymore. Schedule RCU free.
3678 wq_unregister_lockdep(wq
);
3679 call_rcu(&wq
->rcu
, rcu_free_wq
);
3684 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3685 * @pwq: target pool_workqueue
3687 * If @pwq isn't freezing, set @pwq->max_active to the associated
3688 * workqueue's saved_max_active and activate delayed work items
3689 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3691 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3693 struct workqueue_struct
*wq
= pwq
->wq
;
3694 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3695 unsigned long flags
;
3697 /* for @wq->saved_max_active */
3698 lockdep_assert_held(&wq
->mutex
);
3700 /* fast exit for non-freezable wqs */
3701 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3704 /* this function can be called during early boot w/ irq disabled */
3705 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
3708 * During [un]freezing, the caller is responsible for ensuring that
3709 * this function is called at least once after @workqueue_freezing
3710 * is updated and visible.
3712 if (!freezable
|| !workqueue_freezing
) {
3713 pwq
->max_active
= wq
->saved_max_active
;
3715 while (!list_empty(&pwq
->delayed_works
) &&
3716 pwq
->nr_active
< pwq
->max_active
)
3717 pwq_activate_first_delayed(pwq
);
3720 * Need to kick a worker after thawed or an unbound wq's
3721 * max_active is bumped. It's a slow path. Do it always.
3723 wake_up_worker(pwq
->pool
);
3725 pwq
->max_active
= 0;
3728 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
3731 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3732 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3733 struct worker_pool
*pool
)
3735 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3737 memset(pwq
, 0, sizeof(*pwq
));
3741 pwq
->flush_color
= -1;
3743 INIT_LIST_HEAD(&pwq
->delayed_works
);
3744 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3745 INIT_LIST_HEAD(&pwq
->mayday_node
);
3746 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3749 /* sync @pwq with the current state of its associated wq and link it */
3750 static void link_pwq(struct pool_workqueue
*pwq
)
3752 struct workqueue_struct
*wq
= pwq
->wq
;
3754 lockdep_assert_held(&wq
->mutex
);
3756 /* may be called multiple times, ignore if already linked */
3757 if (!list_empty(&pwq
->pwqs_node
))
3760 /* set the matching work_color */
3761 pwq
->work_color
= wq
->work_color
;
3763 /* sync max_active to the current setting */
3764 pwq_adjust_max_active(pwq
);
3767 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3770 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3771 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3772 const struct workqueue_attrs
*attrs
)
3774 struct worker_pool
*pool
;
3775 struct pool_workqueue
*pwq
;
3777 lockdep_assert_held(&wq_pool_mutex
);
3779 pool
= get_unbound_pool(attrs
);
3783 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3785 put_unbound_pool(pool
);
3789 init_pwq(pwq
, wq
, pool
);
3794 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3795 * @attrs: the wq_attrs of the default pwq of the target workqueue
3796 * @node: the target NUMA node
3797 * @cpu_going_down: if >= 0, the CPU to consider as offline
3798 * @cpumask: outarg, the resulting cpumask
3800 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3801 * @cpu_going_down is >= 0, that cpu is considered offline during
3802 * calculation. The result is stored in @cpumask.
3804 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3805 * enabled and @node has online CPUs requested by @attrs, the returned
3806 * cpumask is the intersection of the possible CPUs of @node and
3809 * The caller is responsible for ensuring that the cpumask of @node stays
3812 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3815 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3816 int cpu_going_down
, cpumask_t
*cpumask
)
3818 if (!wq_numa_enabled
|| attrs
->no_numa
)
3821 /* does @node have any online CPUs @attrs wants? */
3822 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3823 if (cpu_going_down
>= 0)
3824 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3826 if (cpumask_empty(cpumask
))
3829 /* yeap, return possible CPUs in @node that @attrs wants */
3830 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3832 if (cpumask_empty(cpumask
)) {
3833 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3834 "possible intersect\n");
3838 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3841 cpumask_copy(cpumask
, attrs
->cpumask
);
3845 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3846 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3848 struct pool_workqueue
*pwq
)
3850 struct pool_workqueue
*old_pwq
;
3852 lockdep_assert_held(&wq_pool_mutex
);
3853 lockdep_assert_held(&wq
->mutex
);
3855 /* link_pwq() can handle duplicate calls */
3858 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3859 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3863 /* context to store the prepared attrs & pwqs before applying */
3864 struct apply_wqattrs_ctx
{
3865 struct workqueue_struct
*wq
; /* target workqueue */
3866 struct workqueue_attrs
*attrs
; /* attrs to apply */
3867 struct list_head list
; /* queued for batching commit */
3868 struct pool_workqueue
*dfl_pwq
;
3869 struct pool_workqueue
*pwq_tbl
[];
3872 /* free the resources after success or abort */
3873 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3879 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3880 put_pwq_unlocked(ctx
->dfl_pwq
);
3882 free_workqueue_attrs(ctx
->attrs
);
3888 /* allocate the attrs and pwqs for later installation */
3889 static struct apply_wqattrs_ctx
*
3890 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3891 const struct workqueue_attrs
*attrs
)
3893 struct apply_wqattrs_ctx
*ctx
;
3894 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3897 lockdep_assert_held(&wq_pool_mutex
);
3899 ctx
= kzalloc(struct_size(ctx
, pwq_tbl
, nr_node_ids
), GFP_KERNEL
);
3901 new_attrs
= alloc_workqueue_attrs();
3902 tmp_attrs
= alloc_workqueue_attrs();
3903 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3907 * Calculate the attrs of the default pwq.
3908 * If the user configured cpumask doesn't overlap with the
3909 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3911 copy_workqueue_attrs(new_attrs
, attrs
);
3912 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3913 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3914 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3917 * We may create multiple pwqs with differing cpumasks. Make a
3918 * copy of @new_attrs which will be modified and used to obtain
3921 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3924 * If something goes wrong during CPU up/down, we'll fall back to
3925 * the default pwq covering whole @attrs->cpumask. Always create
3926 * it even if we don't use it immediately.
3928 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3932 for_each_node(node
) {
3933 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3934 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3935 if (!ctx
->pwq_tbl
[node
])
3938 ctx
->dfl_pwq
->refcnt
++;
3939 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3943 /* save the user configured attrs and sanitize it. */
3944 copy_workqueue_attrs(new_attrs
, attrs
);
3945 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3946 ctx
->attrs
= new_attrs
;
3949 free_workqueue_attrs(tmp_attrs
);
3953 free_workqueue_attrs(tmp_attrs
);
3954 free_workqueue_attrs(new_attrs
);
3955 apply_wqattrs_cleanup(ctx
);
3959 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3960 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3964 /* all pwqs have been created successfully, let's install'em */
3965 mutex_lock(&ctx
->wq
->mutex
);
3967 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3969 /* save the previous pwq and install the new one */
3971 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
3972 ctx
->pwq_tbl
[node
]);
3974 /* @dfl_pwq might not have been used, ensure it's linked */
3975 link_pwq(ctx
->dfl_pwq
);
3976 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
3978 mutex_unlock(&ctx
->wq
->mutex
);
3981 static void apply_wqattrs_lock(void)
3983 /* CPUs should stay stable across pwq creations and installations */
3985 mutex_lock(&wq_pool_mutex
);
3988 static void apply_wqattrs_unlock(void)
3990 mutex_unlock(&wq_pool_mutex
);
3994 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
3995 const struct workqueue_attrs
*attrs
)
3997 struct apply_wqattrs_ctx
*ctx
;
3999 /* only unbound workqueues can change attributes */
4000 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
4003 /* creating multiple pwqs breaks ordering guarantee */
4004 if (!list_empty(&wq
->pwqs
)) {
4005 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4008 wq
->flags
&= ~__WQ_ORDERED
;
4011 ctx
= apply_wqattrs_prepare(wq
, attrs
);
4015 /* the ctx has been prepared successfully, let's commit it */
4016 apply_wqattrs_commit(ctx
);
4017 apply_wqattrs_cleanup(ctx
);
4023 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4024 * @wq: the target workqueue
4025 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4027 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4028 * machines, this function maps a separate pwq to each NUMA node with
4029 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4030 * NUMA node it was issued on. Older pwqs are released as in-flight work
4031 * items finish. Note that a work item which repeatedly requeues itself
4032 * back-to-back will stay on its current pwq.
4034 * Performs GFP_KERNEL allocations.
4036 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4038 * Return: 0 on success and -errno on failure.
4040 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
4041 const struct workqueue_attrs
*attrs
)
4045 lockdep_assert_cpus_held();
4047 mutex_lock(&wq_pool_mutex
);
4048 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
4049 mutex_unlock(&wq_pool_mutex
);
4055 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4056 * @wq: the target workqueue
4057 * @cpu: the CPU coming up or going down
4058 * @online: whether @cpu is coming up or going down
4060 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4061 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4064 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4065 * falls back to @wq->dfl_pwq which may not be optimal but is always
4068 * Note that when the last allowed CPU of a NUMA node goes offline for a
4069 * workqueue with a cpumask spanning multiple nodes, the workers which were
4070 * already executing the work items for the workqueue will lose their CPU
4071 * affinity and may execute on any CPU. This is similar to how per-cpu
4072 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4073 * affinity, it's the user's responsibility to flush the work item from
4076 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
4079 int node
= cpu_to_node(cpu
);
4080 int cpu_off
= online
? -1 : cpu
;
4081 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
4082 struct workqueue_attrs
*target_attrs
;
4085 lockdep_assert_held(&wq_pool_mutex
);
4087 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
4088 wq
->unbound_attrs
->no_numa
)
4092 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4093 * Let's use a preallocated one. The following buf is protected by
4094 * CPU hotplug exclusion.
4096 target_attrs
= wq_update_unbound_numa_attrs_buf
;
4097 cpumask
= target_attrs
->cpumask
;
4099 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4100 pwq
= unbound_pwq_by_node(wq
, node
);
4103 * Let's determine what needs to be done. If the target cpumask is
4104 * different from the default pwq's, we need to compare it to @pwq's
4105 * and create a new one if they don't match. If the target cpumask
4106 * equals the default pwq's, the default pwq should be used.
4108 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
4109 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4115 /* create a new pwq */
4116 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4118 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4123 /* Install the new pwq. */
4124 mutex_lock(&wq
->mutex
);
4125 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4129 mutex_lock(&wq
->mutex
);
4130 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4131 get_pwq(wq
->dfl_pwq
);
4132 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4133 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4135 mutex_unlock(&wq
->mutex
);
4136 put_pwq_unlocked(old_pwq
);
4139 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4141 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4144 if (!(wq
->flags
& WQ_UNBOUND
)) {
4145 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4149 for_each_possible_cpu(cpu
) {
4150 struct pool_workqueue
*pwq
=
4151 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4152 struct worker_pool
*cpu_pools
=
4153 per_cpu(cpu_worker_pools
, cpu
);
4155 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4157 mutex_lock(&wq
->mutex
);
4159 mutex_unlock(&wq
->mutex
);
4165 if (wq
->flags
& __WQ_ORDERED
) {
4166 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4167 /* there should only be single pwq for ordering guarantee */
4168 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4169 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4170 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4172 ret
= apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4179 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4182 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4184 if (max_active
< 1 || max_active
> lim
)
4185 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4186 max_active
, name
, 1, lim
);
4188 return clamp_val(max_active
, 1, lim
);
4192 * Workqueues which may be used during memory reclaim should have a rescuer
4193 * to guarantee forward progress.
4195 static int init_rescuer(struct workqueue_struct
*wq
)
4197 struct worker
*rescuer
;
4200 if (!(wq
->flags
& WQ_MEM_RECLAIM
))
4203 rescuer
= alloc_worker(NUMA_NO_NODE
);
4207 rescuer
->rescue_wq
= wq
;
4208 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s", wq
->name
);
4209 ret
= PTR_ERR_OR_ZERO(rescuer
->task
);
4215 wq
->rescuer
= rescuer
;
4216 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
4217 wake_up_process(rescuer
->task
);
4223 struct workqueue_struct
*alloc_workqueue(const char *fmt
,
4225 int max_active
, ...)
4227 size_t tbl_size
= 0;
4229 struct workqueue_struct
*wq
;
4230 struct pool_workqueue
*pwq
;
4233 * Unbound && max_active == 1 used to imply ordered, which is no
4234 * longer the case on NUMA machines due to per-node pools. While
4235 * alloc_ordered_workqueue() is the right way to create an ordered
4236 * workqueue, keep the previous behavior to avoid subtle breakages
4239 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
4240 flags
|= __WQ_ORDERED
;
4242 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4243 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4244 flags
|= WQ_UNBOUND
;
4246 /* allocate wq and format name */
4247 if (flags
& WQ_UNBOUND
)
4248 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
4250 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4254 if (flags
& WQ_UNBOUND
) {
4255 wq
->unbound_attrs
= alloc_workqueue_attrs();
4256 if (!wq
->unbound_attrs
)
4260 va_start(args
, max_active
);
4261 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4264 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4265 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4269 wq
->saved_max_active
= max_active
;
4270 mutex_init(&wq
->mutex
);
4271 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4272 INIT_LIST_HEAD(&wq
->pwqs
);
4273 INIT_LIST_HEAD(&wq
->flusher_queue
);
4274 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4275 INIT_LIST_HEAD(&wq
->maydays
);
4277 wq_init_lockdep(wq
);
4278 INIT_LIST_HEAD(&wq
->list
);
4280 if (alloc_and_link_pwqs(wq
) < 0)
4281 goto err_unreg_lockdep
;
4283 if (wq_online
&& init_rescuer(wq
) < 0)
4286 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4290 * wq_pool_mutex protects global freeze state and workqueues list.
4291 * Grab it, adjust max_active and add the new @wq to workqueues
4294 mutex_lock(&wq_pool_mutex
);
4296 mutex_lock(&wq
->mutex
);
4297 for_each_pwq(pwq
, wq
)
4298 pwq_adjust_max_active(pwq
);
4299 mutex_unlock(&wq
->mutex
);
4301 list_add_tail_rcu(&wq
->list
, &workqueues
);
4303 mutex_unlock(&wq_pool_mutex
);
4308 wq_unregister_lockdep(wq
);
4309 wq_free_lockdep(wq
);
4311 free_workqueue_attrs(wq
->unbound_attrs
);
4315 destroy_workqueue(wq
);
4318 EXPORT_SYMBOL_GPL(alloc_workqueue
);
4320 static bool pwq_busy(struct pool_workqueue
*pwq
)
4324 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
4325 if (pwq
->nr_in_flight
[i
])
4328 if ((pwq
!= pwq
->wq
->dfl_pwq
) && (pwq
->refcnt
> 1))
4330 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4337 * destroy_workqueue - safely terminate a workqueue
4338 * @wq: target workqueue
4340 * Safely destroy a workqueue. All work currently pending will be done first.
4342 void destroy_workqueue(struct workqueue_struct
*wq
)
4344 struct pool_workqueue
*pwq
;
4348 * Remove it from sysfs first so that sanity check failure doesn't
4349 * lead to sysfs name conflicts.
4351 workqueue_sysfs_unregister(wq
);
4353 /* drain it before proceeding with destruction */
4354 drain_workqueue(wq
);
4356 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4358 struct worker
*rescuer
= wq
->rescuer
;
4360 /* this prevents new queueing */
4361 spin_lock_irq(&wq_mayday_lock
);
4363 spin_unlock_irq(&wq_mayday_lock
);
4365 /* rescuer will empty maydays list before exiting */
4366 kthread_stop(rescuer
->task
);
4371 * Sanity checks - grab all the locks so that we wait for all
4372 * in-flight operations which may do put_pwq().
4374 mutex_lock(&wq_pool_mutex
);
4375 mutex_lock(&wq
->mutex
);
4376 for_each_pwq(pwq
, wq
) {
4377 spin_lock_irq(&pwq
->pool
->lock
);
4378 if (WARN_ON(pwq_busy(pwq
))) {
4379 pr_warn("%s: %s has the following busy pwq\n",
4380 __func__
, wq
->name
);
4382 spin_unlock_irq(&pwq
->pool
->lock
);
4383 mutex_unlock(&wq
->mutex
);
4384 mutex_unlock(&wq_pool_mutex
);
4385 show_workqueue_state();
4388 spin_unlock_irq(&pwq
->pool
->lock
);
4390 mutex_unlock(&wq
->mutex
);
4391 mutex_unlock(&wq_pool_mutex
);
4394 * wq list is used to freeze wq, remove from list after
4395 * flushing is complete in case freeze races us.
4397 mutex_lock(&wq_pool_mutex
);
4398 list_del_rcu(&wq
->list
);
4399 mutex_unlock(&wq_pool_mutex
);
4401 if (!(wq
->flags
& WQ_UNBOUND
)) {
4402 wq_unregister_lockdep(wq
);
4404 * The base ref is never dropped on per-cpu pwqs. Directly
4405 * schedule RCU free.
4407 call_rcu(&wq
->rcu
, rcu_free_wq
);
4410 * We're the sole accessor of @wq at this point. Directly
4411 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4412 * @wq will be freed when the last pwq is released.
4414 for_each_node(node
) {
4415 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4416 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4417 put_pwq_unlocked(pwq
);
4421 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4422 * put. Don't access it afterwards.
4426 put_pwq_unlocked(pwq
);
4429 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4432 * workqueue_set_max_active - adjust max_active of a workqueue
4433 * @wq: target workqueue
4434 * @max_active: new max_active value.
4436 * Set max_active of @wq to @max_active.
4439 * Don't call from IRQ context.
4441 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4443 struct pool_workqueue
*pwq
;
4445 /* disallow meddling with max_active for ordered workqueues */
4446 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4449 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4451 mutex_lock(&wq
->mutex
);
4453 wq
->flags
&= ~__WQ_ORDERED
;
4454 wq
->saved_max_active
= max_active
;
4456 for_each_pwq(pwq
, wq
)
4457 pwq_adjust_max_active(pwq
);
4459 mutex_unlock(&wq
->mutex
);
4461 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4464 * current_work - retrieve %current task's work struct
4466 * Determine if %current task is a workqueue worker and what it's working on.
4467 * Useful to find out the context that the %current task is running in.
4469 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4471 struct work_struct
*current_work(void)
4473 struct worker
*worker
= current_wq_worker();
4475 return worker
? worker
->current_work
: NULL
;
4477 EXPORT_SYMBOL(current_work
);
4480 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4482 * Determine whether %current is a workqueue rescuer. Can be used from
4483 * work functions to determine whether it's being run off the rescuer task.
4485 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4487 bool current_is_workqueue_rescuer(void)
4489 struct worker
*worker
= current_wq_worker();
4491 return worker
&& worker
->rescue_wq
;
4495 * workqueue_congested - test whether a workqueue is congested
4496 * @cpu: CPU in question
4497 * @wq: target workqueue
4499 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4500 * no synchronization around this function and the test result is
4501 * unreliable and only useful as advisory hints or for debugging.
4503 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4504 * Note that both per-cpu and unbound workqueues may be associated with
4505 * multiple pool_workqueues which have separate congested states. A
4506 * workqueue being congested on one CPU doesn't mean the workqueue is also
4507 * contested on other CPUs / NUMA nodes.
4510 * %true if congested, %false otherwise.
4512 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4514 struct pool_workqueue
*pwq
;
4520 if (cpu
== WORK_CPU_UNBOUND
)
4521 cpu
= smp_processor_id();
4523 if (!(wq
->flags
& WQ_UNBOUND
))
4524 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4526 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4528 ret
= !list_empty(&pwq
->delayed_works
);
4534 EXPORT_SYMBOL_GPL(workqueue_congested
);
4537 * work_busy - test whether a work is currently pending or running
4538 * @work: the work to be tested
4540 * Test whether @work is currently pending or running. There is no
4541 * synchronization around this function and the test result is
4542 * unreliable and only useful as advisory hints or for debugging.
4545 * OR'd bitmask of WORK_BUSY_* bits.
4547 unsigned int work_busy(struct work_struct
*work
)
4549 struct worker_pool
*pool
;
4550 unsigned long flags
;
4551 unsigned int ret
= 0;
4553 if (work_pending(work
))
4554 ret
|= WORK_BUSY_PENDING
;
4557 pool
= get_work_pool(work
);
4559 spin_lock_irqsave(&pool
->lock
, flags
);
4560 if (find_worker_executing_work(pool
, work
))
4561 ret
|= WORK_BUSY_RUNNING
;
4562 spin_unlock_irqrestore(&pool
->lock
, flags
);
4568 EXPORT_SYMBOL_GPL(work_busy
);
4571 * set_worker_desc - set description for the current work item
4572 * @fmt: printf-style format string
4573 * @...: arguments for the format string
4575 * This function can be called by a running work function to describe what
4576 * the work item is about. If the worker task gets dumped, this
4577 * information will be printed out together to help debugging. The
4578 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4580 void set_worker_desc(const char *fmt
, ...)
4582 struct worker
*worker
= current_wq_worker();
4586 va_start(args
, fmt
);
4587 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4591 EXPORT_SYMBOL_GPL(set_worker_desc
);
4594 * print_worker_info - print out worker information and description
4595 * @log_lvl: the log level to use when printing
4596 * @task: target task
4598 * If @task is a worker and currently executing a work item, print out the
4599 * name of the workqueue being serviced and worker description set with
4600 * set_worker_desc() by the currently executing work item.
4602 * This function can be safely called on any task as long as the
4603 * task_struct itself is accessible. While safe, this function isn't
4604 * synchronized and may print out mixups or garbages of limited length.
4606 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4608 work_func_t
*fn
= NULL
;
4609 char name
[WQ_NAME_LEN
] = { };
4610 char desc
[WORKER_DESC_LEN
] = { };
4611 struct pool_workqueue
*pwq
= NULL
;
4612 struct workqueue_struct
*wq
= NULL
;
4613 struct worker
*worker
;
4615 if (!(task
->flags
& PF_WQ_WORKER
))
4619 * This function is called without any synchronization and @task
4620 * could be in any state. Be careful with dereferences.
4622 worker
= kthread_probe_data(task
);
4625 * Carefully copy the associated workqueue's workfn, name and desc.
4626 * Keep the original last '\0' in case the original is garbage.
4628 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4629 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4630 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4631 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4632 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4634 if (fn
|| name
[0] || desc
[0]) {
4635 printk("%sWorkqueue: %s %ps", log_lvl
, name
, fn
);
4636 if (strcmp(name
, desc
))
4637 pr_cont(" (%s)", desc
);
4642 static void pr_cont_pool_info(struct worker_pool
*pool
)
4644 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4645 if (pool
->node
!= NUMA_NO_NODE
)
4646 pr_cont(" node=%d", pool
->node
);
4647 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4650 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4652 if (work
->func
== wq_barrier_func
) {
4653 struct wq_barrier
*barr
;
4655 barr
= container_of(work
, struct wq_barrier
, work
);
4657 pr_cont("%s BAR(%d)", comma
? "," : "",
4658 task_pid_nr(barr
->task
));
4660 pr_cont("%s %ps", comma
? "," : "", work
->func
);
4664 static void show_pwq(struct pool_workqueue
*pwq
)
4666 struct worker_pool
*pool
= pwq
->pool
;
4667 struct work_struct
*work
;
4668 struct worker
*worker
;
4669 bool has_in_flight
= false, has_pending
= false;
4672 pr_info(" pwq %d:", pool
->id
);
4673 pr_cont_pool_info(pool
);
4675 pr_cont(" active=%d/%d refcnt=%d%s\n",
4676 pwq
->nr_active
, pwq
->max_active
, pwq
->refcnt
,
4677 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4679 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4680 if (worker
->current_pwq
== pwq
) {
4681 has_in_flight
= true;
4685 if (has_in_flight
) {
4688 pr_info(" in-flight:");
4689 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4690 if (worker
->current_pwq
!= pwq
)
4693 pr_cont("%s %d%s:%ps", comma
? "," : "",
4694 task_pid_nr(worker
->task
),
4695 worker
->rescue_wq
? "(RESCUER)" : "",
4696 worker
->current_func
);
4697 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4698 pr_cont_work(false, work
);
4704 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4705 if (get_work_pwq(work
) == pwq
) {
4713 pr_info(" pending:");
4714 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4715 if (get_work_pwq(work
) != pwq
)
4718 pr_cont_work(comma
, work
);
4719 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4724 if (!list_empty(&pwq
->delayed_works
)) {
4727 pr_info(" delayed:");
4728 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4729 pr_cont_work(comma
, work
);
4730 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4737 * show_workqueue_state - dump workqueue state
4739 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4740 * all busy workqueues and pools.
4742 void show_workqueue_state(void)
4744 struct workqueue_struct
*wq
;
4745 struct worker_pool
*pool
;
4746 unsigned long flags
;
4751 pr_info("Showing busy workqueues and worker pools:\n");
4753 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4754 struct pool_workqueue
*pwq
;
4757 for_each_pwq(pwq
, wq
) {
4758 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4766 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4768 for_each_pwq(pwq
, wq
) {
4769 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4770 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4772 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4774 * We could be printing a lot from atomic context, e.g.
4775 * sysrq-t -> show_workqueue_state(). Avoid triggering
4778 touch_nmi_watchdog();
4782 for_each_pool(pool
, pi
) {
4783 struct worker
*worker
;
4786 spin_lock_irqsave(&pool
->lock
, flags
);
4787 if (pool
->nr_workers
== pool
->nr_idle
)
4790 pr_info("pool %d:", pool
->id
);
4791 pr_cont_pool_info(pool
);
4792 pr_cont(" hung=%us workers=%d",
4793 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4796 pr_cont(" manager: %d",
4797 task_pid_nr(pool
->manager
->task
));
4798 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4799 pr_cont(" %s%d", first
? "idle: " : "",
4800 task_pid_nr(worker
->task
));
4805 spin_unlock_irqrestore(&pool
->lock
, flags
);
4807 * We could be printing a lot from atomic context, e.g.
4808 * sysrq-t -> show_workqueue_state(). Avoid triggering
4811 touch_nmi_watchdog();
4817 /* used to show worker information through /proc/PID/{comm,stat,status} */
4818 void wq_worker_comm(char *buf
, size_t size
, struct task_struct
*task
)
4822 /* always show the actual comm */
4823 off
= strscpy(buf
, task
->comm
, size
);
4827 /* stabilize PF_WQ_WORKER and worker pool association */
4828 mutex_lock(&wq_pool_attach_mutex
);
4830 if (task
->flags
& PF_WQ_WORKER
) {
4831 struct worker
*worker
= kthread_data(task
);
4832 struct worker_pool
*pool
= worker
->pool
;
4835 spin_lock_irq(&pool
->lock
);
4837 * ->desc tracks information (wq name or
4838 * set_worker_desc()) for the latest execution. If
4839 * current, prepend '+', otherwise '-'.
4841 if (worker
->desc
[0] != '\0') {
4842 if (worker
->current_work
)
4843 scnprintf(buf
+ off
, size
- off
, "+%s",
4846 scnprintf(buf
+ off
, size
- off
, "-%s",
4849 spin_unlock_irq(&pool
->lock
);
4853 mutex_unlock(&wq_pool_attach_mutex
);
4861 * There are two challenges in supporting CPU hotplug. Firstly, there
4862 * are a lot of assumptions on strong associations among work, pwq and
4863 * pool which make migrating pending and scheduled works very
4864 * difficult to implement without impacting hot paths. Secondly,
4865 * worker pools serve mix of short, long and very long running works making
4866 * blocked draining impractical.
4868 * This is solved by allowing the pools to be disassociated from the CPU
4869 * running as an unbound one and allowing it to be reattached later if the
4870 * cpu comes back online.
4873 static void unbind_workers(int cpu
)
4875 struct worker_pool
*pool
;
4876 struct worker
*worker
;
4878 for_each_cpu_worker_pool(pool
, cpu
) {
4879 mutex_lock(&wq_pool_attach_mutex
);
4880 spin_lock_irq(&pool
->lock
);
4883 * We've blocked all attach/detach operations. Make all workers
4884 * unbound and set DISASSOCIATED. Before this, all workers
4885 * except for the ones which are still executing works from
4886 * before the last CPU down must be on the cpu. After
4887 * this, they may become diasporas.
4889 for_each_pool_worker(worker
, pool
)
4890 worker
->flags
|= WORKER_UNBOUND
;
4892 pool
->flags
|= POOL_DISASSOCIATED
;
4894 spin_unlock_irq(&pool
->lock
);
4895 mutex_unlock(&wq_pool_attach_mutex
);
4898 * Call schedule() so that we cross rq->lock and thus can
4899 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4900 * This is necessary as scheduler callbacks may be invoked
4906 * Sched callbacks are disabled now. Zap nr_running.
4907 * After this, nr_running stays zero and need_more_worker()
4908 * and keep_working() are always true as long as the
4909 * worklist is not empty. This pool now behaves as an
4910 * unbound (in terms of concurrency management) pool which
4911 * are served by workers tied to the pool.
4913 atomic_set(&pool
->nr_running
, 0);
4916 * With concurrency management just turned off, a busy
4917 * worker blocking could lead to lengthy stalls. Kick off
4918 * unbound chain execution of currently pending work items.
4920 spin_lock_irq(&pool
->lock
);
4921 wake_up_worker(pool
);
4922 spin_unlock_irq(&pool
->lock
);
4927 * rebind_workers - rebind all workers of a pool to the associated CPU
4928 * @pool: pool of interest
4930 * @pool->cpu is coming online. Rebind all workers to the CPU.
4932 static void rebind_workers(struct worker_pool
*pool
)
4934 struct worker
*worker
;
4936 lockdep_assert_held(&wq_pool_attach_mutex
);
4939 * Restore CPU affinity of all workers. As all idle workers should
4940 * be on the run-queue of the associated CPU before any local
4941 * wake-ups for concurrency management happen, restore CPU affinity
4942 * of all workers first and then clear UNBOUND. As we're called
4943 * from CPU_ONLINE, the following shouldn't fail.
4945 for_each_pool_worker(worker
, pool
)
4946 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4947 pool
->attrs
->cpumask
) < 0);
4949 spin_lock_irq(&pool
->lock
);
4951 pool
->flags
&= ~POOL_DISASSOCIATED
;
4953 for_each_pool_worker(worker
, pool
) {
4954 unsigned int worker_flags
= worker
->flags
;
4957 * A bound idle worker should actually be on the runqueue
4958 * of the associated CPU for local wake-ups targeting it to
4959 * work. Kick all idle workers so that they migrate to the
4960 * associated CPU. Doing this in the same loop as
4961 * replacing UNBOUND with REBOUND is safe as no worker will
4962 * be bound before @pool->lock is released.
4964 if (worker_flags
& WORKER_IDLE
)
4965 wake_up_process(worker
->task
);
4968 * We want to clear UNBOUND but can't directly call
4969 * worker_clr_flags() or adjust nr_running. Atomically
4970 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4971 * @worker will clear REBOUND using worker_clr_flags() when
4972 * it initiates the next execution cycle thus restoring
4973 * concurrency management. Note that when or whether
4974 * @worker clears REBOUND doesn't affect correctness.
4976 * WRITE_ONCE() is necessary because @worker->flags may be
4977 * tested without holding any lock in
4978 * wq_worker_running(). Without it, NOT_RUNNING test may
4979 * fail incorrectly leading to premature concurrency
4980 * management operations.
4982 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4983 worker_flags
|= WORKER_REBOUND
;
4984 worker_flags
&= ~WORKER_UNBOUND
;
4985 WRITE_ONCE(worker
->flags
, worker_flags
);
4988 spin_unlock_irq(&pool
->lock
);
4992 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4993 * @pool: unbound pool of interest
4994 * @cpu: the CPU which is coming up
4996 * An unbound pool may end up with a cpumask which doesn't have any online
4997 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4998 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4999 * online CPU before, cpus_allowed of all its workers should be restored.
5001 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
5003 static cpumask_t cpumask
;
5004 struct worker
*worker
;
5006 lockdep_assert_held(&wq_pool_attach_mutex
);
5008 /* is @cpu allowed for @pool? */
5009 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
5012 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
5014 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5015 for_each_pool_worker(worker
, pool
)
5016 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, &cpumask
) < 0);
5019 int workqueue_prepare_cpu(unsigned int cpu
)
5021 struct worker_pool
*pool
;
5023 for_each_cpu_worker_pool(pool
, cpu
) {
5024 if (pool
->nr_workers
)
5026 if (!create_worker(pool
))
5032 int workqueue_online_cpu(unsigned int cpu
)
5034 struct worker_pool
*pool
;
5035 struct workqueue_struct
*wq
;
5038 mutex_lock(&wq_pool_mutex
);
5040 for_each_pool(pool
, pi
) {
5041 mutex_lock(&wq_pool_attach_mutex
);
5043 if (pool
->cpu
== cpu
)
5044 rebind_workers(pool
);
5045 else if (pool
->cpu
< 0)
5046 restore_unbound_workers_cpumask(pool
, cpu
);
5048 mutex_unlock(&wq_pool_attach_mutex
);
5051 /* update NUMA affinity of unbound workqueues */
5052 list_for_each_entry(wq
, &workqueues
, list
)
5053 wq_update_unbound_numa(wq
, cpu
, true);
5055 mutex_unlock(&wq_pool_mutex
);
5059 int workqueue_offline_cpu(unsigned int cpu
)
5061 struct workqueue_struct
*wq
;
5063 /* unbinding per-cpu workers should happen on the local CPU */
5064 if (WARN_ON(cpu
!= smp_processor_id()))
5067 unbind_workers(cpu
);
5069 /* update NUMA affinity of unbound workqueues */
5070 mutex_lock(&wq_pool_mutex
);
5071 list_for_each_entry(wq
, &workqueues
, list
)
5072 wq_update_unbound_numa(wq
, cpu
, false);
5073 mutex_unlock(&wq_pool_mutex
);
5078 struct work_for_cpu
{
5079 struct work_struct work
;
5085 static void work_for_cpu_fn(struct work_struct
*work
)
5087 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
5089 wfc
->ret
= wfc
->fn(wfc
->arg
);
5093 * work_on_cpu - run a function in thread context on a particular cpu
5094 * @cpu: the cpu to run on
5095 * @fn: the function to run
5096 * @arg: the function arg
5098 * It is up to the caller to ensure that the cpu doesn't go offline.
5099 * The caller must not hold any locks which would prevent @fn from completing.
5101 * Return: The value @fn returns.
5103 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
5105 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
5107 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
5108 schedule_work_on(cpu
, &wfc
.work
);
5109 flush_work(&wfc
.work
);
5110 destroy_work_on_stack(&wfc
.work
);
5113 EXPORT_SYMBOL_GPL(work_on_cpu
);
5116 * work_on_cpu_safe - run a function in thread context on a particular cpu
5117 * @cpu: the cpu to run on
5118 * @fn: the function to run
5119 * @arg: the function argument
5121 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5122 * any locks which would prevent @fn from completing.
5124 * Return: The value @fn returns.
5126 long work_on_cpu_safe(int cpu
, long (*fn
)(void *), void *arg
)
5131 if (cpu_online(cpu
))
5132 ret
= work_on_cpu(cpu
, fn
, arg
);
5136 EXPORT_SYMBOL_GPL(work_on_cpu_safe
);
5137 #endif /* CONFIG_SMP */
5139 #ifdef CONFIG_FREEZER
5142 * freeze_workqueues_begin - begin freezing workqueues
5144 * Start freezing workqueues. After this function returns, all freezable
5145 * workqueues will queue new works to their delayed_works list instead of
5149 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5151 void freeze_workqueues_begin(void)
5153 struct workqueue_struct
*wq
;
5154 struct pool_workqueue
*pwq
;
5156 mutex_lock(&wq_pool_mutex
);
5158 WARN_ON_ONCE(workqueue_freezing
);
5159 workqueue_freezing
= true;
5161 list_for_each_entry(wq
, &workqueues
, list
) {
5162 mutex_lock(&wq
->mutex
);
5163 for_each_pwq(pwq
, wq
)
5164 pwq_adjust_max_active(pwq
);
5165 mutex_unlock(&wq
->mutex
);
5168 mutex_unlock(&wq_pool_mutex
);
5172 * freeze_workqueues_busy - are freezable workqueues still busy?
5174 * Check whether freezing is complete. This function must be called
5175 * between freeze_workqueues_begin() and thaw_workqueues().
5178 * Grabs and releases wq_pool_mutex.
5181 * %true if some freezable workqueues are still busy. %false if freezing
5184 bool freeze_workqueues_busy(void)
5187 struct workqueue_struct
*wq
;
5188 struct pool_workqueue
*pwq
;
5190 mutex_lock(&wq_pool_mutex
);
5192 WARN_ON_ONCE(!workqueue_freezing
);
5194 list_for_each_entry(wq
, &workqueues
, list
) {
5195 if (!(wq
->flags
& WQ_FREEZABLE
))
5198 * nr_active is monotonically decreasing. It's safe
5199 * to peek without lock.
5202 for_each_pwq(pwq
, wq
) {
5203 WARN_ON_ONCE(pwq
->nr_active
< 0);
5204 if (pwq
->nr_active
) {
5213 mutex_unlock(&wq_pool_mutex
);
5218 * thaw_workqueues - thaw workqueues
5220 * Thaw workqueues. Normal queueing is restored and all collected
5221 * frozen works are transferred to their respective pool worklists.
5224 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5226 void thaw_workqueues(void)
5228 struct workqueue_struct
*wq
;
5229 struct pool_workqueue
*pwq
;
5231 mutex_lock(&wq_pool_mutex
);
5233 if (!workqueue_freezing
)
5236 workqueue_freezing
= false;
5238 /* restore max_active and repopulate worklist */
5239 list_for_each_entry(wq
, &workqueues
, list
) {
5240 mutex_lock(&wq
->mutex
);
5241 for_each_pwq(pwq
, wq
)
5242 pwq_adjust_max_active(pwq
);
5243 mutex_unlock(&wq
->mutex
);
5247 mutex_unlock(&wq_pool_mutex
);
5249 #endif /* CONFIG_FREEZER */
5251 static int workqueue_apply_unbound_cpumask(void)
5255 struct workqueue_struct
*wq
;
5256 struct apply_wqattrs_ctx
*ctx
, *n
;
5258 lockdep_assert_held(&wq_pool_mutex
);
5260 list_for_each_entry(wq
, &workqueues
, list
) {
5261 if (!(wq
->flags
& WQ_UNBOUND
))
5263 /* creating multiple pwqs breaks ordering guarantee */
5264 if (wq
->flags
& __WQ_ORDERED
)
5267 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
5273 list_add_tail(&ctx
->list
, &ctxs
);
5276 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
5278 apply_wqattrs_commit(ctx
);
5279 apply_wqattrs_cleanup(ctx
);
5286 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5287 * @cpumask: the cpumask to set
5289 * The low-level workqueues cpumask is a global cpumask that limits
5290 * the affinity of all unbound workqueues. This function check the @cpumask
5291 * and apply it to all unbound workqueues and updates all pwqs of them.
5293 * Retun: 0 - Success
5294 * -EINVAL - Invalid @cpumask
5295 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5297 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
5300 cpumask_var_t saved_cpumask
;
5302 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
5306 * Not excluding isolated cpus on purpose.
5307 * If the user wishes to include them, we allow that.
5309 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
5310 if (!cpumask_empty(cpumask
)) {
5311 apply_wqattrs_lock();
5313 /* save the old wq_unbound_cpumask. */
5314 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
5316 /* update wq_unbound_cpumask at first and apply it to wqs. */
5317 cpumask_copy(wq_unbound_cpumask
, cpumask
);
5318 ret
= workqueue_apply_unbound_cpumask();
5320 /* restore the wq_unbound_cpumask when failed. */
5322 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
5324 apply_wqattrs_unlock();
5327 free_cpumask_var(saved_cpumask
);
5333 * Workqueues with WQ_SYSFS flag set is visible to userland via
5334 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5335 * following attributes.
5337 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5338 * max_active RW int : maximum number of in-flight work items
5340 * Unbound workqueues have the following extra attributes.
5342 * pool_ids RO int : the associated pool IDs for each node
5343 * nice RW int : nice value of the workers
5344 * cpumask RW mask : bitmask of allowed CPUs for the workers
5345 * numa RW bool : whether enable NUMA affinity
5348 struct workqueue_struct
*wq
;
5352 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
5354 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5359 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
5362 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5364 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
5366 static DEVICE_ATTR_RO(per_cpu
);
5368 static ssize_t
max_active_show(struct device
*dev
,
5369 struct device_attribute
*attr
, char *buf
)
5371 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5373 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
5376 static ssize_t
max_active_store(struct device
*dev
,
5377 struct device_attribute
*attr
, const char *buf
,
5380 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5383 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
5386 workqueue_set_max_active(wq
, val
);
5389 static DEVICE_ATTR_RW(max_active
);
5391 static struct attribute
*wq_sysfs_attrs
[] = {
5392 &dev_attr_per_cpu
.attr
,
5393 &dev_attr_max_active
.attr
,
5396 ATTRIBUTE_GROUPS(wq_sysfs
);
5398 static ssize_t
wq_pool_ids_show(struct device
*dev
,
5399 struct device_attribute
*attr
, char *buf
)
5401 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5402 const char *delim
= "";
5403 int node
, written
= 0;
5407 for_each_node(node
) {
5408 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
5409 "%s%d:%d", delim
, node
,
5410 unbound_pwq_by_node(wq
, node
)->pool
->id
);
5413 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
5420 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
5423 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5426 mutex_lock(&wq
->mutex
);
5427 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
5428 mutex_unlock(&wq
->mutex
);
5433 /* prepare workqueue_attrs for sysfs store operations */
5434 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
5436 struct workqueue_attrs
*attrs
;
5438 lockdep_assert_held(&wq_pool_mutex
);
5440 attrs
= alloc_workqueue_attrs();
5444 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
5448 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
5449 const char *buf
, size_t count
)
5451 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5452 struct workqueue_attrs
*attrs
;
5455 apply_wqattrs_lock();
5457 attrs
= wq_sysfs_prep_attrs(wq
);
5461 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5462 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5463 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5468 apply_wqattrs_unlock();
5469 free_workqueue_attrs(attrs
);
5470 return ret
?: count
;
5473 static ssize_t
wq_cpumask_show(struct device
*dev
,
5474 struct device_attribute
*attr
, char *buf
)
5476 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5479 mutex_lock(&wq
->mutex
);
5480 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5481 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5482 mutex_unlock(&wq
->mutex
);
5486 static ssize_t
wq_cpumask_store(struct device
*dev
,
5487 struct device_attribute
*attr
,
5488 const char *buf
, size_t count
)
5490 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5491 struct workqueue_attrs
*attrs
;
5494 apply_wqattrs_lock();
5496 attrs
= wq_sysfs_prep_attrs(wq
);
5500 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5502 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5505 apply_wqattrs_unlock();
5506 free_workqueue_attrs(attrs
);
5507 return ret
?: count
;
5510 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5513 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5516 mutex_lock(&wq
->mutex
);
5517 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5518 !wq
->unbound_attrs
->no_numa
);
5519 mutex_unlock(&wq
->mutex
);
5524 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5525 const char *buf
, size_t count
)
5527 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5528 struct workqueue_attrs
*attrs
;
5529 int v
, ret
= -ENOMEM
;
5531 apply_wqattrs_lock();
5533 attrs
= wq_sysfs_prep_attrs(wq
);
5538 if (sscanf(buf
, "%d", &v
) == 1) {
5539 attrs
->no_numa
= !v
;
5540 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5544 apply_wqattrs_unlock();
5545 free_workqueue_attrs(attrs
);
5546 return ret
?: count
;
5549 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5550 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5551 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5552 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5553 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5557 static struct bus_type wq_subsys
= {
5558 .name
= "workqueue",
5559 .dev_groups
= wq_sysfs_groups
,
5562 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5563 struct device_attribute
*attr
, char *buf
)
5567 mutex_lock(&wq_pool_mutex
);
5568 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5569 cpumask_pr_args(wq_unbound_cpumask
));
5570 mutex_unlock(&wq_pool_mutex
);
5575 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5576 struct device_attribute
*attr
, const char *buf
, size_t count
)
5578 cpumask_var_t cpumask
;
5581 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5584 ret
= cpumask_parse(buf
, cpumask
);
5586 ret
= workqueue_set_unbound_cpumask(cpumask
);
5588 free_cpumask_var(cpumask
);
5589 return ret
? ret
: count
;
5592 static struct device_attribute wq_sysfs_cpumask_attr
=
5593 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5594 wq_unbound_cpumask_store
);
5596 static int __init
wq_sysfs_init(void)
5600 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5604 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5606 core_initcall(wq_sysfs_init
);
5608 static void wq_device_release(struct device
*dev
)
5610 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5616 * workqueue_sysfs_register - make a workqueue visible in sysfs
5617 * @wq: the workqueue to register
5619 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5620 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5621 * which is the preferred method.
5623 * Workqueue user should use this function directly iff it wants to apply
5624 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5625 * apply_workqueue_attrs() may race against userland updating the
5628 * Return: 0 on success, -errno on failure.
5630 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5632 struct wq_device
*wq_dev
;
5636 * Adjusting max_active or creating new pwqs by applying
5637 * attributes breaks ordering guarantee. Disallow exposing ordered
5640 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
5643 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5648 wq_dev
->dev
.bus
= &wq_subsys
;
5649 wq_dev
->dev
.release
= wq_device_release
;
5650 dev_set_name(&wq_dev
->dev
, "%s", wq
->name
);
5653 * unbound_attrs are created separately. Suppress uevent until
5654 * everything is ready.
5656 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5658 ret
= device_register(&wq_dev
->dev
);
5660 put_device(&wq_dev
->dev
);
5665 if (wq
->flags
& WQ_UNBOUND
) {
5666 struct device_attribute
*attr
;
5668 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5669 ret
= device_create_file(&wq_dev
->dev
, attr
);
5671 device_unregister(&wq_dev
->dev
);
5678 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5679 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5684 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5685 * @wq: the workqueue to unregister
5687 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5689 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5691 struct wq_device
*wq_dev
= wq
->wq_dev
;
5697 device_unregister(&wq_dev
->dev
);
5699 #else /* CONFIG_SYSFS */
5700 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5701 #endif /* CONFIG_SYSFS */
5704 * Workqueue watchdog.
5706 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5707 * flush dependency, a concurrency managed work item which stays RUNNING
5708 * indefinitely. Workqueue stalls can be very difficult to debug as the
5709 * usual warning mechanisms don't trigger and internal workqueue state is
5712 * Workqueue watchdog monitors all worker pools periodically and dumps
5713 * state if some pools failed to make forward progress for a while where
5714 * forward progress is defined as the first item on ->worklist changing.
5716 * This mechanism is controlled through the kernel parameter
5717 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5718 * corresponding sysfs parameter file.
5720 #ifdef CONFIG_WQ_WATCHDOG
5722 static unsigned long wq_watchdog_thresh
= 30;
5723 static struct timer_list wq_watchdog_timer
;
5725 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5726 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5728 static void wq_watchdog_reset_touched(void)
5732 wq_watchdog_touched
= jiffies
;
5733 for_each_possible_cpu(cpu
)
5734 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5737 static void wq_watchdog_timer_fn(struct timer_list
*unused
)
5739 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5740 bool lockup_detected
= false;
5741 struct worker_pool
*pool
;
5749 for_each_pool(pool
, pi
) {
5750 unsigned long pool_ts
, touched
, ts
;
5752 if (list_empty(&pool
->worklist
))
5755 /* get the latest of pool and touched timestamps */
5756 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5757 touched
= READ_ONCE(wq_watchdog_touched
);
5759 if (time_after(pool_ts
, touched
))
5764 if (pool
->cpu
>= 0) {
5765 unsigned long cpu_touched
=
5766 READ_ONCE(per_cpu(wq_watchdog_touched_cpu
,
5768 if (time_after(cpu_touched
, ts
))
5773 if (time_after(jiffies
, ts
+ thresh
)) {
5774 lockup_detected
= true;
5775 pr_emerg("BUG: workqueue lockup - pool");
5776 pr_cont_pool_info(pool
);
5777 pr_cont(" stuck for %us!\n",
5778 jiffies_to_msecs(jiffies
- pool_ts
) / 1000);
5784 if (lockup_detected
)
5785 show_workqueue_state();
5787 wq_watchdog_reset_touched();
5788 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5791 notrace
void wq_watchdog_touch(int cpu
)
5794 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5796 wq_watchdog_touched
= jiffies
;
5799 static void wq_watchdog_set_thresh(unsigned long thresh
)
5801 wq_watchdog_thresh
= 0;
5802 del_timer_sync(&wq_watchdog_timer
);
5805 wq_watchdog_thresh
= thresh
;
5806 wq_watchdog_reset_touched();
5807 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5811 static int wq_watchdog_param_set_thresh(const char *val
,
5812 const struct kernel_param
*kp
)
5814 unsigned long thresh
;
5817 ret
= kstrtoul(val
, 0, &thresh
);
5822 wq_watchdog_set_thresh(thresh
);
5824 wq_watchdog_thresh
= thresh
;
5829 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5830 .set
= wq_watchdog_param_set_thresh
,
5831 .get
= param_get_ulong
,
5834 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5837 static void wq_watchdog_init(void)
5839 timer_setup(&wq_watchdog_timer
, wq_watchdog_timer_fn
, TIMER_DEFERRABLE
);
5840 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5843 #else /* CONFIG_WQ_WATCHDOG */
5845 static inline void wq_watchdog_init(void) { }
5847 #endif /* CONFIG_WQ_WATCHDOG */
5849 static void __init
wq_numa_init(void)
5854 if (num_possible_nodes() <= 1)
5857 if (wq_disable_numa
) {
5858 pr_info("workqueue: NUMA affinity support disabled\n");
5862 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs();
5863 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5866 * We want masks of possible CPUs of each node which isn't readily
5867 * available. Build one from cpu_to_node() which should have been
5868 * fully initialized by now.
5870 tbl
= kcalloc(nr_node_ids
, sizeof(tbl
[0]), GFP_KERNEL
);
5874 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5875 node_online(node
) ? node
: NUMA_NO_NODE
));
5877 for_each_possible_cpu(cpu
) {
5878 node
= cpu_to_node(cpu
);
5879 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5880 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5881 /* happens iff arch is bonkers, let's just proceed */
5884 cpumask_set_cpu(cpu
, tbl
[node
]);
5887 wq_numa_possible_cpumask
= tbl
;
5888 wq_numa_enabled
= true;
5892 * workqueue_init_early - early init for workqueue subsystem
5894 * This is the first half of two-staged workqueue subsystem initialization
5895 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5896 * idr are up. It sets up all the data structures and system workqueues
5897 * and allows early boot code to create workqueues and queue/cancel work
5898 * items. Actual work item execution starts only after kthreads can be
5899 * created and scheduled right before early initcalls.
5901 int __init
workqueue_init_early(void)
5903 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5904 int hk_flags
= HK_FLAG_DOMAIN
| HK_FLAG_WQ
;
5907 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5909 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5910 cpumask_copy(wq_unbound_cpumask
, housekeeping_cpumask(hk_flags
));
5912 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5914 /* initialize CPU pools */
5915 for_each_possible_cpu(cpu
) {
5916 struct worker_pool
*pool
;
5919 for_each_cpu_worker_pool(pool
, cpu
) {
5920 BUG_ON(init_worker_pool(pool
));
5922 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5923 pool
->attrs
->nice
= std_nice
[i
++];
5924 pool
->node
= cpu_to_node(cpu
);
5927 mutex_lock(&wq_pool_mutex
);
5928 BUG_ON(worker_pool_assign_id(pool
));
5929 mutex_unlock(&wq_pool_mutex
);
5933 /* create default unbound and ordered wq attrs */
5934 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5935 struct workqueue_attrs
*attrs
;
5937 BUG_ON(!(attrs
= alloc_workqueue_attrs()));
5938 attrs
->nice
= std_nice
[i
];
5939 unbound_std_wq_attrs
[i
] = attrs
;
5942 * An ordered wq should have only one pwq as ordering is
5943 * guaranteed by max_active which is enforced by pwqs.
5944 * Turn off NUMA so that dfl_pwq is used for all nodes.
5946 BUG_ON(!(attrs
= alloc_workqueue_attrs()));
5947 attrs
->nice
= std_nice
[i
];
5948 attrs
->no_numa
= true;
5949 ordered_wq_attrs
[i
] = attrs
;
5952 system_wq
= alloc_workqueue("events", 0, 0);
5953 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5954 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5955 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5956 WQ_UNBOUND_MAX_ACTIVE
);
5957 system_freezable_wq
= alloc_workqueue("events_freezable",
5959 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5960 WQ_POWER_EFFICIENT
, 0);
5961 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5962 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5964 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5965 !system_unbound_wq
|| !system_freezable_wq
||
5966 !system_power_efficient_wq
||
5967 !system_freezable_power_efficient_wq
);
5973 * workqueue_init - bring workqueue subsystem fully online
5975 * This is the latter half of two-staged workqueue subsystem initialization
5976 * and invoked as soon as kthreads can be created and scheduled.
5977 * Workqueues have been created and work items queued on them, but there
5978 * are no kworkers executing the work items yet. Populate the worker pools
5979 * with the initial workers and enable future kworker creations.
5981 int __init
workqueue_init(void)
5983 struct workqueue_struct
*wq
;
5984 struct worker_pool
*pool
;
5988 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5989 * CPU to node mapping may not be available that early on some
5990 * archs such as power and arm64. As per-cpu pools created
5991 * previously could be missing node hint and unbound pools NUMA
5992 * affinity, fix them up.
5994 * Also, while iterating workqueues, create rescuers if requested.
5998 mutex_lock(&wq_pool_mutex
);
6000 for_each_possible_cpu(cpu
) {
6001 for_each_cpu_worker_pool(pool
, cpu
) {
6002 pool
->node
= cpu_to_node(cpu
);
6006 list_for_each_entry(wq
, &workqueues
, list
) {
6007 wq_update_unbound_numa(wq
, smp_processor_id(), true);
6008 WARN(init_rescuer(wq
),
6009 "workqueue: failed to create early rescuer for %s",
6013 mutex_unlock(&wq_pool_mutex
);
6015 /* create the initial workers */
6016 for_each_online_cpu(cpu
) {
6017 for_each_cpu_worker_pool(pool
, cpu
) {
6018 pool
->flags
&= ~POOL_DISASSOCIATED
;
6019 BUG_ON(!create_worker(pool
));
6023 hash_for_each(unbound_pool_hash
, bkt
, pool
, hash_node
)
6024 BUG_ON(!create_worker(pool
));