2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/core-api/workqueue.rst for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/nmi.h>
52 #include <linux/debug-snapshot.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 * 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: pool->attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-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
136 * sched-RCU for reads.
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. Sched-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 */
157 int nr_workers
; /* L: total number of workers */
159 /* nr_idle includes the ones off idle_list for rebinding */
160 int nr_idle
; /* L: currently idle ones */
162 struct list_head idle_list
; /* X: list of idle workers */
163 struct timer_list idle_timer
; /* L: worker idle timeout */
164 struct timer_list mayday_timer
; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
168 /* L: hash of busy workers */
170 /* see manage_workers() for details on the two manager mutexes */
171 struct worker
*manager
; /* L: purely informational */
172 struct mutex attach_mutex
; /* attach/detach exclusion */
173 struct list_head workers
; /* A: attached workers */
174 struct completion
*detach_completion
; /* all workers detached */
176 struct ida worker_ida
; /* worker IDs for task name */
178 struct workqueue_attrs
*attrs
; /* I: worker attributes */
179 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
180 int refcnt
; /* PL: refcnt for unbound pools */
183 * The current concurrency level. As it's likely to be accessed
184 * from other CPUs during try_to_wake_up(), put it in a separate
187 atomic_t nr_running ____cacheline_aligned_in_smp
;
190 * Destruction of pool is sched-RCU protected to allow dereferences
191 * from get_work_pool().
194 } ____cacheline_aligned_in_smp
;
197 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
198 * of work_struct->data are used for flags and the remaining high bits
199 * point to the pwq; thus, pwqs need to be aligned at two's power of the
200 * number of flag bits.
202 struct pool_workqueue
{
203 struct worker_pool
*pool
; /* I: the associated pool */
204 struct workqueue_struct
*wq
; /* I: the owning workqueue */
205 int work_color
; /* L: current color */
206 int flush_color
; /* L: flushing color */
207 int refcnt
; /* L: reference count */
208 int nr_in_flight
[WORK_NR_COLORS
];
209 /* L: nr of in_flight works */
210 int nr_active
; /* L: nr of active works */
211 int max_active
; /* L: max active works */
212 struct list_head delayed_works
; /* L: delayed works */
213 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
214 struct list_head mayday_node
; /* MD: node on wq->maydays */
217 * Release of unbound pwq is punted to system_wq. See put_pwq()
218 * and pwq_unbound_release_workfn() for details. pool_workqueue
219 * itself is also sched-RCU protected so that the first pwq can be
220 * determined without grabbing wq->mutex.
222 struct work_struct unbound_release_work
;
224 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
227 * Structure used to wait for workqueue flush.
230 struct list_head list
; /* WQ: list of flushers */
231 int flush_color
; /* WQ: flush color waiting for */
232 struct completion done
; /* flush completion */
238 * The externally visible workqueue. It relays the issued work items to
239 * the appropriate worker_pool through its pool_workqueues.
241 struct workqueue_struct
{
242 struct list_head pwqs
; /* WR: all pwqs of this wq */
243 struct list_head list
; /* PR: list of all workqueues */
245 struct mutex mutex
; /* protects this wq */
246 int work_color
; /* WQ: current work color */
247 int flush_color
; /* WQ: current flush color */
248 atomic_t nr_pwqs_to_flush
; /* flush in progress */
249 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
250 struct list_head flusher_queue
; /* WQ: flush waiters */
251 struct list_head flusher_overflow
; /* WQ: flush overflow list */
253 struct list_head maydays
; /* MD: pwqs requesting rescue */
254 struct worker
*rescuer
; /* I: rescue worker */
256 int nr_drainers
; /* WQ: drain in progress */
257 int saved_max_active
; /* WQ: saved pwq max_active */
259 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
260 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
263 struct wq_device
*wq_dev
; /* I: for sysfs interface */
265 #ifdef CONFIG_LOCKDEP
266 struct lockdep_map lockdep_map
;
268 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
271 * Destruction of workqueue_struct is sched-RCU protected to allow
272 * walking the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
283 static struct kmem_cache
*pwq_cache
;
285 static cpumask_var_t
*wq_numa_possible_cpumask
;
286 /* possible CPUs of each node */
288 static bool wq_disable_numa
;
289 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
291 /* see the comment above the definition of WQ_POWER_EFFICIENT */
292 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
293 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
295 static bool wq_online
; /* can kworkers be created yet? */
297 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
299 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
300 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
302 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
303 static DEFINE_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
);
359 #define CREATE_TRACE_POINTS
360 #include <trace/events/workqueue.h>
362 #define assert_rcu_or_pool_mutex() \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq_pool_mutex), \
365 "sched RCU or wq_pool_mutex should be held")
367 #define assert_rcu_or_wq_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex), \
370 "sched RCU or wq->mutex should be held")
372 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
373 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
374 !lockdep_is_held(&wq->mutex) && \
375 !lockdep_is_held(&wq_pool_mutex), \
376 "sched RCU, wq->mutex or wq_pool_mutex should be held")
378 #define for_each_cpu_worker_pool(pool, cpu) \
379 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
380 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
384 * for_each_pool - iterate through all worker_pools in the system
385 * @pool: iteration cursor
386 * @pi: integer used for iteration
388 * This must be called either with wq_pool_mutex held or sched RCU read
389 * locked. If the pool needs to be used beyond the locking in effect, the
390 * caller is responsible for guaranteeing that the pool stays online.
392 * The if/else clause exists only for the lockdep assertion and can be
395 #define for_each_pool(pool, pi) \
396 idr_for_each_entry(&worker_pool_idr, pool, pi) \
397 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
401 * for_each_pool_worker - iterate through all workers of a worker_pool
402 * @worker: iteration cursor
403 * @pool: worker_pool to iterate workers of
405 * This must be called with @pool->attach_mutex.
407 * The if/else clause exists only for the lockdep assertion and can be
410 #define for_each_pool_worker(worker, pool) \
411 list_for_each_entry((worker), &(pool)->workers, node) \
412 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
416 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
417 * @pwq: iteration cursor
418 * @wq: the target workqueue
420 * This must be called either with wq->mutex held or sched RCU read locked.
421 * If the pwq needs to be used beyond the locking in effect, the caller is
422 * responsible for guaranteeing that the pwq stays online.
424 * The if/else clause exists only for the lockdep assertion and can be
427 #define for_each_pwq(pwq, wq) \
428 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
429 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
432 #ifdef CONFIG_DEBUG_OBJECTS_WORK
434 static struct debug_obj_descr work_debug_descr
;
436 static void *work_debug_hint(void *addr
)
438 return ((struct work_struct
*) addr
)->func
;
441 static bool work_is_static_object(void *addr
)
443 struct work_struct
*work
= addr
;
445 return test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
));
449 * fixup_init is called when:
450 * - an active object is initialized
452 static bool work_fixup_init(void *addr
, enum debug_obj_state state
)
454 struct work_struct
*work
= addr
;
457 case ODEBUG_STATE_ACTIVE
:
458 cancel_work_sync(work
);
459 debug_object_init(work
, &work_debug_descr
);
467 * fixup_free is called when:
468 * - an active object is freed
470 static bool work_fixup_free(void *addr
, enum debug_obj_state state
)
472 struct work_struct
*work
= addr
;
475 case ODEBUG_STATE_ACTIVE
:
476 cancel_work_sync(work
);
477 debug_object_free(work
, &work_debug_descr
);
484 static struct debug_obj_descr work_debug_descr
= {
485 .name
= "work_struct",
486 .debug_hint
= work_debug_hint
,
487 .is_static_object
= work_is_static_object
,
488 .fixup_init
= work_fixup_init
,
489 .fixup_free
= work_fixup_free
,
492 static inline void debug_work_activate(struct work_struct
*work
)
494 debug_object_activate(work
, &work_debug_descr
);
497 static inline void debug_work_deactivate(struct work_struct
*work
)
499 debug_object_deactivate(work
, &work_debug_descr
);
502 void __init_work(struct work_struct
*work
, int onstack
)
505 debug_object_init_on_stack(work
, &work_debug_descr
);
507 debug_object_init(work
, &work_debug_descr
);
509 EXPORT_SYMBOL_GPL(__init_work
);
511 void destroy_work_on_stack(struct work_struct
*work
)
513 debug_object_free(work
, &work_debug_descr
);
515 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
517 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
519 destroy_timer_on_stack(&work
->timer
);
520 debug_object_free(&work
->work
, &work_debug_descr
);
522 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
525 static inline void debug_work_activate(struct work_struct
*work
) { }
526 static inline void debug_work_deactivate(struct work_struct
*work
) { }
530 * worker_pool_assign_id - allocate ID and assing it to @pool
531 * @pool: the pool pointer of interest
533 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
534 * successfully, -errno on failure.
536 static int worker_pool_assign_id(struct worker_pool
*pool
)
540 lockdep_assert_held(&wq_pool_mutex
);
542 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
552 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
553 * @wq: the target workqueue
556 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
558 * If the pwq needs to be used beyond the locking in effect, the caller is
559 * responsible for guaranteeing that the pwq stays online.
561 * Return: The unbound pool_workqueue for @node.
563 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
566 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
569 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
570 * delayed item is pending. The plan is to keep CPU -> NODE
571 * mapping valid and stable across CPU on/offlines. Once that
572 * happens, this workaround can be removed.
574 if (unlikely(node
== NUMA_NO_NODE
))
577 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
580 static unsigned int work_color_to_flags(int color
)
582 return color
<< WORK_STRUCT_COLOR_SHIFT
;
585 static int get_work_color(struct work_struct
*work
)
587 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
588 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
591 static int work_next_color(int color
)
593 return (color
+ 1) % WORK_NR_COLORS
;
597 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
598 * contain the pointer to the queued pwq. Once execution starts, the flag
599 * is cleared and the high bits contain OFFQ flags and pool ID.
601 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
602 * and clear_work_data() can be used to set the pwq, pool or clear
603 * work->data. These functions should only be called while the work is
604 * owned - ie. while the PENDING bit is set.
606 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
607 * corresponding to a work. Pool is available once the work has been
608 * queued anywhere after initialization until it is sync canceled. pwq is
609 * available only while the work item is queued.
611 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
612 * canceled. While being canceled, a work item may have its PENDING set
613 * but stay off timer and worklist for arbitrarily long and nobody should
614 * try to steal the PENDING bit.
616 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
619 WARN_ON_ONCE(!work_pending(work
));
620 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
623 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
624 unsigned long extra_flags
)
626 set_work_data(work
, (unsigned long)pwq
,
627 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
630 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
633 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
634 WORK_STRUCT_PENDING
);
637 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
641 * The following wmb is paired with the implied mb in
642 * test_and_set_bit(PENDING) and ensures all updates to @work made
643 * here are visible to and precede any updates by the next PENDING
647 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
649 * The following mb guarantees that previous clear of a PENDING bit
650 * will not be reordered with any speculative LOADS or STORES from
651 * work->current_func, which is executed afterwards. This possible
652 * reordering can lead to a missed execution on attempt to qeueue
653 * the same @work. E.g. consider this case:
656 * ---------------------------- --------------------------------
658 * 1 STORE event_indicated
659 * 2 queue_work_on() {
660 * 3 test_and_set_bit(PENDING)
661 * 4 } set_..._and_clear_pending() {
662 * 5 set_work_data() # clear bit
664 * 7 work->current_func() {
665 * 8 LOAD event_indicated
668 * Without an explicit full barrier speculative LOAD on line 8 can
669 * be executed before CPU#0 does STORE on line 1. If that happens,
670 * CPU#0 observes the PENDING bit is still set and new execution of
671 * a @work is not queued in a hope, that CPU#1 will eventually
672 * finish the queued @work. Meanwhile CPU#1 does not see
673 * event_indicated is set, because speculative LOAD was executed
674 * before actual STORE.
679 static void clear_work_data(struct work_struct
*work
)
681 smp_wmb(); /* see set_work_pool_and_clear_pending() */
682 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
685 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
687 unsigned long data
= atomic_long_read(&work
->data
);
689 if (data
& WORK_STRUCT_PWQ
)
690 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
696 * get_work_pool - return the worker_pool a given work was associated with
697 * @work: the work item of interest
699 * Pools are created and destroyed under wq_pool_mutex, and allows read
700 * access under sched-RCU read lock. As such, this function should be
701 * called under wq_pool_mutex or with preemption disabled.
703 * All fields of the returned pool are accessible as long as the above
704 * mentioned locking is in effect. If the returned pool needs to be used
705 * beyond the critical section, the caller is responsible for ensuring the
706 * returned pool is and stays online.
708 * Return: The worker_pool @work was last associated with. %NULL if none.
710 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
712 unsigned long data
= atomic_long_read(&work
->data
);
715 assert_rcu_or_pool_mutex();
717 if (data
& WORK_STRUCT_PWQ
)
718 return ((struct pool_workqueue
*)
719 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
721 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
722 if (pool_id
== WORK_OFFQ_POOL_NONE
)
725 return idr_find(&worker_pool_idr
, pool_id
);
729 * get_work_pool_id - return the worker pool ID a given work is associated with
730 * @work: the work item of interest
732 * Return: The worker_pool ID @work was last associated with.
733 * %WORK_OFFQ_POOL_NONE if none.
735 static int get_work_pool_id(struct work_struct
*work
)
737 unsigned long data
= atomic_long_read(&work
->data
);
739 if (data
& WORK_STRUCT_PWQ
)
740 return ((struct pool_workqueue
*)
741 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
743 return data
>> WORK_OFFQ_POOL_SHIFT
;
746 static void mark_work_canceling(struct work_struct
*work
)
748 unsigned long pool_id
= get_work_pool_id(work
);
750 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
751 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
754 static bool work_is_canceling(struct work_struct
*work
)
756 unsigned long data
= atomic_long_read(&work
->data
);
758 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
762 * Policy functions. These define the policies on how the global worker
763 * pools are managed. Unless noted otherwise, these functions assume that
764 * they're being called with pool->lock held.
767 static bool __need_more_worker(struct worker_pool
*pool
)
769 return !atomic_read(&pool
->nr_running
);
773 * Need to wake up a worker? Called from anything but currently
776 * Note that, because unbound workers never contribute to nr_running, this
777 * function will always return %true for unbound pools as long as the
778 * worklist isn't empty.
780 static bool need_more_worker(struct worker_pool
*pool
)
782 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
785 /* Can I start working? Called from busy but !running workers. */
786 static bool may_start_working(struct worker_pool
*pool
)
788 return pool
->nr_idle
;
791 /* Do I need to keep working? Called from currently running workers. */
792 static bool keep_working(struct worker_pool
*pool
)
794 return !list_empty(&pool
->worklist
) &&
795 atomic_read(&pool
->nr_running
) <= 1;
798 /* Do we need a new worker? Called from manager. */
799 static bool need_to_create_worker(struct worker_pool
*pool
)
801 return need_more_worker(pool
) && !may_start_working(pool
);
804 /* Do we have too many workers and should some go away? */
805 static bool too_many_workers(struct worker_pool
*pool
)
807 bool managing
= pool
->flags
& POOL_MANAGER_ACTIVE
;
808 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
809 int nr_busy
= pool
->nr_workers
- nr_idle
;
811 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
818 /* Return the first idle worker. Safe with preemption disabled */
819 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
821 if (unlikely(list_empty(&pool
->idle_list
)))
824 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
828 * wake_up_worker - wake up an idle worker
829 * @pool: worker pool to wake worker from
831 * Wake up the first idle worker of @pool.
834 * spin_lock_irq(pool->lock).
836 static void wake_up_worker(struct worker_pool
*pool
)
838 struct worker
*worker
= first_idle_worker(pool
);
841 wake_up_process(worker
->task
);
845 * wq_worker_waking_up - a worker is waking up
846 * @task: task waking up
847 * @cpu: CPU @task is waking up to
849 * This function is called during try_to_wake_up() when a worker is
853 * spin_lock_irq(rq->lock)
855 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
857 struct worker
*worker
= kthread_data(task
);
859 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
860 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
861 atomic_inc(&worker
->pool
->nr_running
);
866 * wq_worker_sleeping - a worker is going to sleep
867 * @task: task going to sleep
869 * This function is called during schedule() when a busy worker is
870 * going to sleep. Worker on the same cpu can be woken up by
871 * returning pointer to its task.
874 * spin_lock_irq(rq->lock)
877 * Worker task on @cpu to wake up, %NULL if none.
879 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
)
881 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
882 struct worker_pool
*pool
;
885 * Rescuers, which may not have all the fields set up like normal
886 * workers, also reach here, let's not access anything before
887 * checking NOT_RUNNING.
889 if (worker
->flags
& WORKER_NOT_RUNNING
)
894 /* this can only happen on the local cpu */
895 if (WARN_ON_ONCE(pool
->cpu
!= raw_smp_processor_id()))
899 * The counterpart of the following dec_and_test, implied mb,
900 * worklist not empty test sequence is in insert_work().
901 * Please read comment there.
903 * NOT_RUNNING is clear. This means that we're bound to and
904 * running on the local cpu w/ rq lock held and preemption
905 * disabled, which in turn means that none else could be
906 * manipulating idle_list, so dereferencing idle_list without pool
909 if (atomic_dec_and_test(&pool
->nr_running
) &&
910 !list_empty(&pool
->worklist
))
911 to_wakeup
= first_idle_worker(pool
);
912 return to_wakeup
? to_wakeup
->task
: NULL
;
916 * wq_worker_last_func - retrieve worker's last work function
918 * Determine the last function a worker executed. This is called from
919 * the scheduler to get a worker's last known identity.
922 * spin_lock_irq(rq->lock)
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 sched-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
)))
1255 * The queueing is in progress, or it is already queued. Try to
1256 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1258 pool
= get_work_pool(work
);
1262 spin_lock(&pool
->lock
);
1264 * work->data is guaranteed to point to pwq only while the work
1265 * item is queued on pwq->wq, and both updating work->data to point
1266 * to pwq on queueing and to pool on dequeueing are done under
1267 * pwq->pool->lock. This in turn guarantees that, if work->data
1268 * points to pwq which is associated with a locked pool, the work
1269 * item is currently queued on that pool.
1271 pwq
= get_work_pwq(work
);
1272 if (pwq
&& pwq
->pool
== pool
) {
1273 debug_work_deactivate(work
);
1276 * A delayed work item cannot be grabbed directly because
1277 * it might have linked NO_COLOR work items which, if left
1278 * on the delayed_list, will confuse pwq->nr_active
1279 * management later on and cause stall. Make sure the work
1280 * item is activated before grabbing.
1282 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1283 pwq_activate_delayed_work(work
);
1285 list_del_init(&work
->entry
);
1286 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1288 /* work->data points to pwq iff queued, point to pool */
1289 set_work_pool_and_keep_pending(work
, pool
->id
);
1291 spin_unlock(&pool
->lock
);
1294 spin_unlock(&pool
->lock
);
1296 local_irq_restore(*flags
);
1297 if (work_is_canceling(work
))
1304 * insert_work - insert a work into a pool
1305 * @pwq: pwq @work belongs to
1306 * @work: work to insert
1307 * @head: insertion point
1308 * @extra_flags: extra WORK_STRUCT_* flags to set
1310 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1311 * work_struct flags.
1314 * spin_lock_irq(pool->lock).
1316 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1317 struct list_head
*head
, unsigned int extra_flags
)
1319 struct worker_pool
*pool
= pwq
->pool
;
1321 /* we own @work, set data and link */
1322 set_work_pwq(work
, pwq
, extra_flags
);
1323 list_add_tail(&work
->entry
, head
);
1327 * Ensure either wq_worker_sleeping() sees the above
1328 * list_add_tail() or we see zero nr_running to avoid workers lying
1329 * around lazily while there are works to be processed.
1333 if (__need_more_worker(pool
))
1334 wake_up_worker(pool
);
1338 * Test whether @work is being queued from another work executing on the
1341 static bool is_chained_work(struct workqueue_struct
*wq
)
1343 struct worker
*worker
;
1345 worker
= current_wq_worker();
1347 * Return %true iff I'm a worker execuing a work item on @wq. If
1348 * I'm @worker, it's safe to dereference it without locking.
1350 return worker
&& worker
->current_pwq
->wq
== wq
;
1354 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1355 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1356 * avoid perturbing sensitive tasks.
1358 static int wq_select_unbound_cpu(int cpu
)
1360 static bool printed_dbg_warning
;
1363 if (likely(!wq_debug_force_rr_cpu
)) {
1364 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1366 } else if (!printed_dbg_warning
) {
1367 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1368 printed_dbg_warning
= true;
1371 if (cpumask_empty(wq_unbound_cpumask
))
1374 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1375 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1376 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1377 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1378 if (unlikely(new_cpu
>= nr_cpu_ids
))
1381 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1386 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1387 struct work_struct
*work
)
1389 struct pool_workqueue
*pwq
;
1390 struct worker_pool
*last_pool
;
1391 struct list_head
*worklist
;
1392 unsigned int work_flags
;
1393 unsigned int req_cpu
= cpu
;
1396 * While a work item is PENDING && off queue, a task trying to
1397 * steal the PENDING will busy-loop waiting for it to either get
1398 * queued or lose PENDING. Grabbing PENDING and queueing should
1399 * happen with IRQ disabled.
1401 WARN_ON_ONCE(!irqs_disabled());
1403 debug_work_activate(work
);
1405 /* if draining, only works from the same workqueue are allowed */
1406 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1407 WARN_ON_ONCE(!is_chained_work(wq
)))
1410 if (req_cpu
== WORK_CPU_UNBOUND
)
1411 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1413 /* pwq which will be used unless @work is executing elsewhere */
1414 if (!(wq
->flags
& WQ_UNBOUND
))
1415 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1417 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1420 * If @work was previously on a different pool, it might still be
1421 * running there, in which case the work needs to be queued on that
1422 * pool to guarantee non-reentrancy.
1424 last_pool
= get_work_pool(work
);
1425 if (last_pool
&& last_pool
!= pwq
->pool
) {
1426 struct worker
*worker
;
1428 spin_lock(&last_pool
->lock
);
1430 worker
= find_worker_executing_work(last_pool
, work
);
1432 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1433 pwq
= worker
->current_pwq
;
1435 /* meh... not running there, queue here */
1436 spin_unlock(&last_pool
->lock
);
1437 spin_lock(&pwq
->pool
->lock
);
1440 spin_lock(&pwq
->pool
->lock
);
1444 * pwq is determined and locked. For unbound pools, we could have
1445 * raced with pwq release and it could already be dead. If its
1446 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1447 * without another pwq replacing it in the numa_pwq_tbl or while
1448 * work items are executing on it, so the retrying is guaranteed to
1449 * make forward-progress.
1451 if (unlikely(!pwq
->refcnt
)) {
1452 if (wq
->flags
& WQ_UNBOUND
) {
1453 spin_unlock(&pwq
->pool
->lock
);
1458 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1462 /* pwq determined, queue */
1463 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1465 if (WARN_ON(!list_empty(&work
->entry
))) {
1466 spin_unlock(&pwq
->pool
->lock
);
1470 pwq
->nr_in_flight
[pwq
->work_color
]++;
1471 work_flags
= work_color_to_flags(pwq
->work_color
);
1473 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1474 trace_workqueue_activate_work(work
);
1476 worklist
= &pwq
->pool
->worklist
;
1477 if (list_empty(worklist
))
1478 pwq
->pool
->watchdog_ts
= jiffies
;
1480 work_flags
|= WORK_STRUCT_DELAYED
;
1481 worklist
= &pwq
->delayed_works
;
1484 insert_work(pwq
, work
, worklist
, work_flags
);
1486 spin_unlock(&pwq
->pool
->lock
);
1490 * queue_work_on - queue work on specific cpu
1491 * @cpu: CPU number to execute work on
1492 * @wq: workqueue to use
1493 * @work: work to queue
1495 * We queue the work to a specific CPU, the caller must ensure it
1498 * Return: %false if @work was already on a queue, %true otherwise.
1500 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1501 struct work_struct
*work
)
1504 unsigned long flags
;
1506 local_irq_save(flags
);
1508 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1509 __queue_work(cpu
, wq
, work
);
1513 local_irq_restore(flags
);
1516 EXPORT_SYMBOL(queue_work_on
);
1518 void delayed_work_timer_fn(unsigned long __data
)
1520 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1522 /* should have been called from irqsafe timer with irq already off */
1523 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1525 EXPORT_SYMBOL(delayed_work_timer_fn
);
1527 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1528 struct delayed_work
*dwork
, unsigned long delay
)
1530 struct timer_list
*timer
= &dwork
->timer
;
1531 struct work_struct
*work
= &dwork
->work
;
1534 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1535 timer
->data
!= (unsigned long)dwork
);
1536 WARN_ON_ONCE(timer_pending(timer
));
1537 WARN_ON_ONCE(!list_empty(&work
->entry
));
1540 * If @delay is 0, queue @dwork->work immediately. This is for
1541 * both optimization and correctness. The earliest @timer can
1542 * expire is on the closest next tick and delayed_work users depend
1543 * on that there's no such delay when @delay is 0.
1546 __queue_work(cpu
, wq
, &dwork
->work
);
1552 timer
->expires
= jiffies
+ delay
;
1554 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1555 add_timer_on(timer
, cpu
);
1561 * queue_delayed_work_on - queue work on specific CPU after delay
1562 * @cpu: CPU number to execute work on
1563 * @wq: workqueue to use
1564 * @dwork: work to queue
1565 * @delay: number of jiffies to wait before queueing
1567 * Return: %false if @work was already on a queue, %true otherwise. If
1568 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1571 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1572 struct delayed_work
*dwork
, unsigned long delay
)
1574 struct work_struct
*work
= &dwork
->work
;
1576 unsigned long flags
;
1578 /* read the comment in __queue_work() */
1579 local_irq_save(flags
);
1581 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1582 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1586 local_irq_restore(flags
);
1589 EXPORT_SYMBOL(queue_delayed_work_on
);
1592 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1593 * @cpu: CPU number to execute work on
1594 * @wq: workqueue to use
1595 * @dwork: work to queue
1596 * @delay: number of jiffies to wait before queueing
1598 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1599 * modify @dwork's timer so that it expires after @delay. If @delay is
1600 * zero, @work is guaranteed to be scheduled immediately regardless of its
1603 * Return: %false if @dwork was idle and queued, %true if @dwork was
1604 * pending and its timer was modified.
1606 * This function is safe to call from any context including IRQ handler.
1607 * See try_to_grab_pending() for details.
1609 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1610 struct delayed_work
*dwork
, unsigned long delay
)
1612 unsigned long flags
;
1616 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1617 } while (unlikely(ret
== -EAGAIN
));
1619 if (likely(ret
>= 0)) {
1620 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1621 local_irq_restore(flags
);
1624 /* -ENOENT from try_to_grab_pending() becomes %true */
1627 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1630 * worker_enter_idle - enter idle state
1631 * @worker: worker which is entering idle state
1633 * @worker is entering idle state. Update stats and idle timer if
1637 * spin_lock_irq(pool->lock).
1639 static void worker_enter_idle(struct worker
*worker
)
1641 struct worker_pool
*pool
= worker
->pool
;
1643 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1644 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1645 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1648 /* can't use worker_set_flags(), also called from create_worker() */
1649 worker
->flags
|= WORKER_IDLE
;
1651 worker
->last_active
= jiffies
;
1653 /* idle_list is LIFO */
1654 list_add(&worker
->entry
, &pool
->idle_list
);
1656 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1657 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1660 * Sanity check nr_running. Because wq_unbind_fn() releases
1661 * pool->lock between setting %WORKER_UNBOUND and zapping
1662 * nr_running, the warning may trigger spuriously. Check iff
1663 * unbind is not in progress.
1665 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1666 pool
->nr_workers
== pool
->nr_idle
&&
1667 atomic_read(&pool
->nr_running
));
1671 * worker_leave_idle - leave idle state
1672 * @worker: worker which is leaving idle state
1674 * @worker is leaving idle state. Update stats.
1677 * spin_lock_irq(pool->lock).
1679 static void worker_leave_idle(struct worker
*worker
)
1681 struct worker_pool
*pool
= worker
->pool
;
1683 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1685 worker_clr_flags(worker
, WORKER_IDLE
);
1687 list_del_init(&worker
->entry
);
1690 static struct worker
*alloc_worker(int node
)
1692 struct worker
*worker
;
1694 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1696 INIT_LIST_HEAD(&worker
->entry
);
1697 INIT_LIST_HEAD(&worker
->scheduled
);
1698 INIT_LIST_HEAD(&worker
->node
);
1699 /* on creation a worker is in !idle && prep state */
1700 worker
->flags
= WORKER_PREP
;
1706 * worker_attach_to_pool() - attach a worker to a pool
1707 * @worker: worker to be attached
1708 * @pool: the target pool
1710 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1711 * cpu-binding of @worker are kept coordinated with the pool across
1714 static void worker_attach_to_pool(struct worker
*worker
,
1715 struct worker_pool
*pool
)
1717 mutex_lock(&pool
->attach_mutex
);
1720 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1721 * online CPUs. It'll be re-applied when any of the CPUs come up.
1723 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1726 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1727 * stable across this function. See the comments above the
1728 * flag definition for details.
1730 if (pool
->flags
& POOL_DISASSOCIATED
)
1731 worker
->flags
|= WORKER_UNBOUND
;
1733 list_add_tail(&worker
->node
, &pool
->workers
);
1735 mutex_unlock(&pool
->attach_mutex
);
1739 * worker_detach_from_pool() - detach a worker from its pool
1740 * @worker: worker which is attached to its pool
1741 * @pool: the pool @worker is attached to
1743 * Undo the attaching which had been done in worker_attach_to_pool(). The
1744 * caller worker shouldn't access to the pool after detached except it has
1745 * other reference to the pool.
1747 static void worker_detach_from_pool(struct worker
*worker
,
1748 struct worker_pool
*pool
)
1750 struct completion
*detach_completion
= NULL
;
1752 mutex_lock(&pool
->attach_mutex
);
1753 list_del(&worker
->node
);
1754 if (list_empty(&pool
->workers
))
1755 detach_completion
= pool
->detach_completion
;
1756 mutex_unlock(&pool
->attach_mutex
);
1758 /* clear leftover flags without pool->lock after it is detached */
1759 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1761 if (detach_completion
)
1762 complete(detach_completion
);
1766 * create_worker - create a new workqueue worker
1767 * @pool: pool the new worker will belong to
1769 * Create and start a new worker which is attached to @pool.
1772 * Might sleep. Does GFP_KERNEL allocations.
1775 * Pointer to the newly created worker.
1777 static struct worker
*create_worker(struct worker_pool
*pool
)
1779 struct worker
*worker
= NULL
;
1783 /* ID is needed to determine kthread name */
1784 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1788 worker
= alloc_worker(pool
->node
);
1792 worker
->pool
= pool
;
1796 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1797 pool
->attrs
->nice
< 0 ? "H" : "");
1799 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1801 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1802 "kworker/%s", id_buf
);
1803 if (IS_ERR(worker
->task
))
1806 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1807 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1809 /* successful, attach the worker to the pool */
1810 worker_attach_to_pool(worker
, pool
);
1812 /* start the newly created worker */
1813 spin_lock_irq(&pool
->lock
);
1814 worker
->pool
->nr_workers
++;
1815 worker_enter_idle(worker
);
1816 wake_up_process(worker
->task
);
1817 spin_unlock_irq(&pool
->lock
);
1823 ida_simple_remove(&pool
->worker_ida
, id
);
1829 * destroy_worker - destroy a workqueue worker
1830 * @worker: worker to be destroyed
1832 * Destroy @worker and adjust @pool stats accordingly. The worker should
1836 * spin_lock_irq(pool->lock).
1838 static void destroy_worker(struct worker
*worker
)
1840 struct worker_pool
*pool
= worker
->pool
;
1842 lockdep_assert_held(&pool
->lock
);
1844 /* sanity check frenzy */
1845 if (WARN_ON(worker
->current_work
) ||
1846 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1847 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1853 list_del_init(&worker
->entry
);
1854 worker
->flags
|= WORKER_DIE
;
1855 wake_up_process(worker
->task
);
1858 static void idle_worker_timeout(unsigned long __pool
)
1860 struct worker_pool
*pool
= (void *)__pool
;
1862 spin_lock_irq(&pool
->lock
);
1864 while (too_many_workers(pool
)) {
1865 struct worker
*worker
;
1866 unsigned long expires
;
1868 /* idle_list is kept in LIFO order, check the last one */
1869 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1870 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1872 if (time_before(jiffies
, expires
)) {
1873 mod_timer(&pool
->idle_timer
, expires
);
1877 destroy_worker(worker
);
1880 spin_unlock_irq(&pool
->lock
);
1883 static void send_mayday(struct work_struct
*work
)
1885 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1886 struct workqueue_struct
*wq
= pwq
->wq
;
1888 lockdep_assert_held(&wq_mayday_lock
);
1893 /* mayday mayday mayday */
1894 if (list_empty(&pwq
->mayday_node
)) {
1896 * If @pwq is for an unbound wq, its base ref may be put at
1897 * any time due to an attribute change. Pin @pwq until the
1898 * rescuer is done with it.
1901 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1902 wake_up_process(wq
->rescuer
->task
);
1906 static void pool_mayday_timeout(unsigned long __pool
)
1908 struct worker_pool
*pool
= (void *)__pool
;
1909 struct work_struct
*work
;
1911 spin_lock_irq(&pool
->lock
);
1912 spin_lock(&wq_mayday_lock
); /* for wq->maydays */
1914 if (need_to_create_worker(pool
)) {
1916 * We've been trying to create a new worker but
1917 * haven't been successful. We might be hitting an
1918 * allocation deadlock. Send distress signals to
1921 list_for_each_entry(work
, &pool
->worklist
, entry
)
1925 spin_unlock(&wq_mayday_lock
);
1926 spin_unlock_irq(&pool
->lock
);
1928 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1932 * maybe_create_worker - create a new worker if necessary
1933 * @pool: pool to create a new worker for
1935 * Create a new worker for @pool if necessary. @pool is guaranteed to
1936 * have at least one idle worker on return from this function. If
1937 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1938 * sent to all rescuers with works scheduled on @pool to resolve
1939 * possible allocation deadlock.
1941 * On return, need_to_create_worker() is guaranteed to be %false and
1942 * may_start_working() %true.
1945 * spin_lock_irq(pool->lock) which may be released and regrabbed
1946 * multiple times. Does GFP_KERNEL allocations. Called only from
1949 static void maybe_create_worker(struct worker_pool
*pool
)
1950 __releases(&pool
->lock
)
1951 __acquires(&pool
->lock
)
1954 spin_unlock_irq(&pool
->lock
);
1956 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1957 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1960 if (create_worker(pool
) || !need_to_create_worker(pool
))
1963 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1965 if (!need_to_create_worker(pool
))
1969 del_timer_sync(&pool
->mayday_timer
);
1970 spin_lock_irq(&pool
->lock
);
1972 * This is necessary even after a new worker was just successfully
1973 * created as @pool->lock was dropped and the new worker might have
1974 * already become busy.
1976 if (need_to_create_worker(pool
))
1981 * manage_workers - manage worker pool
1984 * Assume the manager role and manage the worker pool @worker belongs
1985 * to. At any given time, there can be only zero or one manager per
1986 * pool. The exclusion is handled automatically by this function.
1988 * The caller can safely start processing works on false return. On
1989 * true return, it's guaranteed that need_to_create_worker() is false
1990 * and may_start_working() is true.
1993 * spin_lock_irq(pool->lock) which may be released and regrabbed
1994 * multiple times. Does GFP_KERNEL allocations.
1997 * %false if the pool doesn't need management and the caller can safely
1998 * start processing works, %true if management function was performed and
1999 * the conditions that the caller verified before calling the function may
2000 * no longer be true.
2002 static bool manage_workers(struct worker
*worker
)
2004 struct worker_pool
*pool
= worker
->pool
;
2006 if (pool
->flags
& POOL_MANAGER_ACTIVE
)
2009 pool
->flags
|= POOL_MANAGER_ACTIVE
;
2010 pool
->manager
= worker
;
2012 maybe_create_worker(pool
);
2014 pool
->manager
= NULL
;
2015 pool
->flags
&= ~POOL_MANAGER_ACTIVE
;
2016 wake_up(&wq_manager_wait
);
2021 * process_one_work - process single work
2023 * @work: work to process
2025 * Process @work. This function contains all the logics necessary to
2026 * process a single work including synchronization against and
2027 * interaction with other workers on the same cpu, queueing and
2028 * flushing. As long as context requirement is met, any worker can
2029 * call this function to process a work.
2032 * spin_lock_irq(pool->lock) which is released and regrabbed.
2034 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2035 __releases(&pool
->lock
)
2036 __acquires(&pool
->lock
)
2038 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2039 struct worker_pool
*pool
= worker
->pool
;
2040 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2042 struct worker
*collision
;
2043 #ifdef CONFIG_LOCKDEP
2045 * It is permissible to free the struct work_struct from
2046 * inside the function that is called from it, this we need to
2047 * take into account for lockdep too. To avoid bogus "held
2048 * lock freed" warnings as well as problems when looking into
2049 * work->lockdep_map, make a copy and use that here.
2051 struct lockdep_map lockdep_map
;
2053 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2055 /* ensure we're on the correct CPU */
2056 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2057 raw_smp_processor_id() != pool
->cpu
);
2060 * A single work shouldn't be executed concurrently by
2061 * multiple workers on a single cpu. Check whether anyone is
2062 * already processing the work. If so, defer the work to the
2063 * currently executing one.
2065 collision
= find_worker_executing_work(pool
, work
);
2066 if (unlikely(collision
)) {
2067 move_linked_works(work
, &collision
->scheduled
, NULL
);
2071 /* claim and dequeue */
2072 debug_work_deactivate(work
);
2073 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2074 worker
->current_work
= work
;
2075 worker
->current_func
= work
->func
;
2076 worker
->current_pwq
= pwq
;
2077 work_color
= get_work_color(work
);
2079 list_del_init(&work
->entry
);
2082 * CPU intensive works don't participate in concurrency management.
2083 * They're the scheduler's responsibility. This takes @worker out
2084 * of concurrency management and the next code block will chain
2085 * execution of the pending work items.
2087 if (unlikely(cpu_intensive
))
2088 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2091 * Wake up another worker if necessary. The condition is always
2092 * false for normal per-cpu workers since nr_running would always
2093 * be >= 1 at this point. This is used to chain execution of the
2094 * pending work items for WORKER_NOT_RUNNING workers such as the
2095 * UNBOUND and CPU_INTENSIVE ones.
2097 if (need_more_worker(pool
))
2098 wake_up_worker(pool
);
2101 * Record the last pool and clear PENDING which should be the last
2102 * update to @work. Also, do this inside @pool->lock so that
2103 * PENDING and queued state changes happen together while IRQ is
2106 set_work_pool_and_clear_pending(work
, pool
->id
);
2108 spin_unlock_irq(&pool
->lock
);
2110 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2111 lock_map_acquire(&lockdep_map
);
2113 * Strictly speaking we should mark the invariant state without holding
2114 * any locks, that is, before these two lock_map_acquire()'s.
2116 * However, that would result in:
2123 * Which would create W1->C->W1 dependencies, even though there is no
2124 * actual deadlock possible. There are two solutions, using a
2125 * read-recursive acquire on the work(queue) 'locks', but this will then
2126 * hit the lockdep limitation on recursive locks, or simply discard
2129 * AFAICT there is no possible deadlock scenario between the
2130 * flush_work() and complete() primitives (except for single-threaded
2131 * workqueues), so hiding them isn't a problem.
2133 lockdep_invariant_state(true);
2134 trace_workqueue_execute_start(work
);
2135 dbg_snapshot_work(worker
, worker
->task
, worker
->current_func
, DSS_FLAG_IN
);
2136 worker
->current_func(work
);
2137 dbg_snapshot_work(worker
, worker
->task
, worker
->current_func
, DSS_FLAG_OUT
);
2139 * While we must be careful to not use "work" after this, the trace
2140 * point will only record its address.
2142 trace_workqueue_execute_end(work
);
2143 lock_map_release(&lockdep_map
);
2144 lock_map_release(&pwq
->wq
->lockdep_map
);
2146 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2147 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2148 " last function: %pf\n",
2149 current
->comm
, preempt_count(), task_pid_nr(current
),
2150 worker
->current_func
);
2151 debug_show_held_locks(current
);
2156 * The following prevents a kworker from hogging CPU on !PREEMPT
2157 * kernels, where a requeueing work item waiting for something to
2158 * happen could deadlock with stop_machine as such work item could
2159 * indefinitely requeue itself while all other CPUs are trapped in
2160 * stop_machine. At the same time, report a quiescent RCU state so
2161 * the same condition doesn't freeze RCU.
2163 cond_resched_rcu_qs();
2165 spin_lock_irq(&pool
->lock
);
2167 /* clear cpu intensive status */
2168 if (unlikely(cpu_intensive
))
2169 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2171 /* tag the worker for identification in schedule() */
2172 worker
->last_func
= worker
->current_func
;
2174 /* we're done with it, release */
2175 hash_del(&worker
->hentry
);
2176 worker
->current_work
= NULL
;
2177 worker
->current_func
= NULL
;
2178 worker
->current_pwq
= NULL
;
2179 worker
->desc_valid
= false;
2180 pwq_dec_nr_in_flight(pwq
, work_color
);
2184 * process_scheduled_works - process scheduled works
2187 * Process all scheduled works. Please note that the scheduled list
2188 * may change while processing a work, so this function repeatedly
2189 * fetches a work from the top and executes it.
2192 * spin_lock_irq(pool->lock) which may be released and regrabbed
2195 static void process_scheduled_works(struct worker
*worker
)
2197 while (!list_empty(&worker
->scheduled
)) {
2198 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2199 struct work_struct
, entry
);
2200 process_one_work(worker
, work
);
2205 * worker_thread - the worker thread function
2208 * The worker thread function. All workers belong to a worker_pool -
2209 * either a per-cpu one or dynamic unbound one. These workers process all
2210 * work items regardless of their specific target workqueue. The only
2211 * exception is work items which belong to workqueues with a rescuer which
2212 * will be explained in rescuer_thread().
2216 static int worker_thread(void *__worker
)
2218 struct worker
*worker
= __worker
;
2219 struct worker_pool
*pool
= worker
->pool
;
2221 /* tell the scheduler that this is a workqueue worker */
2222 worker
->task
->flags
|= PF_WQ_WORKER
;
2224 spin_lock_irq(&pool
->lock
);
2226 /* am I supposed to die? */
2227 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2228 spin_unlock_irq(&pool
->lock
);
2229 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2230 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2232 set_task_comm(worker
->task
, "kworker/dying");
2233 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2234 worker_detach_from_pool(worker
, pool
);
2239 worker_leave_idle(worker
);
2241 /* no more worker necessary? */
2242 if (!need_more_worker(pool
))
2245 /* do we need to manage? */
2246 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2250 * ->scheduled list can only be filled while a worker is
2251 * preparing to process a work or actually processing it.
2252 * Make sure nobody diddled with it while I was sleeping.
2254 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2257 * Finish PREP stage. We're guaranteed to have at least one idle
2258 * worker or that someone else has already assumed the manager
2259 * role. This is where @worker starts participating in concurrency
2260 * management if applicable and concurrency management is restored
2261 * after being rebound. See rebind_workers() for details.
2263 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2266 struct work_struct
*work
=
2267 list_first_entry(&pool
->worklist
,
2268 struct work_struct
, entry
);
2270 pool
->watchdog_ts
= jiffies
;
2272 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2273 /* optimization path, not strictly necessary */
2274 process_one_work(worker
, work
);
2275 if (unlikely(!list_empty(&worker
->scheduled
)))
2276 process_scheduled_works(worker
);
2278 move_linked_works(work
, &worker
->scheduled
, NULL
);
2279 process_scheduled_works(worker
);
2281 } while (keep_working(pool
));
2283 worker_set_flags(worker
, WORKER_PREP
);
2286 * pool->lock is held and there's no work to process and no need to
2287 * manage, sleep. Workers are woken up only while holding
2288 * pool->lock or from local cpu, so setting the current state
2289 * before releasing pool->lock is enough to prevent losing any
2292 worker_enter_idle(worker
);
2293 __set_current_state(TASK_IDLE
);
2294 spin_unlock_irq(&pool
->lock
);
2300 * rescuer_thread - the rescuer thread function
2303 * Workqueue rescuer thread function. There's one rescuer for each
2304 * workqueue which has WQ_MEM_RECLAIM set.
2306 * Regular work processing on a pool may block trying to create a new
2307 * worker which uses GFP_KERNEL allocation which has slight chance of
2308 * developing into deadlock if some works currently on the same queue
2309 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2310 * the problem rescuer solves.
2312 * When such condition is possible, the pool summons rescuers of all
2313 * workqueues which have works queued on the pool and let them process
2314 * those works so that forward progress can be guaranteed.
2316 * This should happen rarely.
2320 static int rescuer_thread(void *__rescuer
)
2322 struct worker
*rescuer
= __rescuer
;
2323 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2324 struct list_head
*scheduled
= &rescuer
->scheduled
;
2327 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2330 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2331 * doesn't participate in concurrency management.
2333 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2335 set_current_state(TASK_IDLE
);
2338 * By the time the rescuer is requested to stop, the workqueue
2339 * shouldn't have any work pending, but @wq->maydays may still have
2340 * pwq(s) queued. This can happen by non-rescuer workers consuming
2341 * all the work items before the rescuer got to them. Go through
2342 * @wq->maydays processing before acting on should_stop so that the
2343 * list is always empty on exit.
2345 should_stop
= kthread_should_stop();
2347 /* see whether any pwq is asking for help */
2348 spin_lock_irq(&wq_mayday_lock
);
2350 while (!list_empty(&wq
->maydays
)) {
2351 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2352 struct pool_workqueue
, mayday_node
);
2353 struct worker_pool
*pool
= pwq
->pool
;
2354 struct work_struct
*work
, *n
;
2357 __set_current_state(TASK_RUNNING
);
2358 list_del_init(&pwq
->mayday_node
);
2360 spin_unlock_irq(&wq_mayday_lock
);
2362 worker_attach_to_pool(rescuer
, pool
);
2364 spin_lock_irq(&pool
->lock
);
2365 rescuer
->pool
= pool
;
2368 * Slurp in all works issued via this workqueue and
2371 WARN_ON_ONCE(!list_empty(scheduled
));
2372 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2373 if (get_work_pwq(work
) == pwq
) {
2375 pool
->watchdog_ts
= jiffies
;
2376 move_linked_works(work
, scheduled
, &n
);
2381 if (!list_empty(scheduled
)) {
2382 process_scheduled_works(rescuer
);
2385 * The above execution of rescued work items could
2386 * have created more to rescue through
2387 * pwq_activate_first_delayed() or chained
2388 * queueing. Let's put @pwq back on mayday list so
2389 * that such back-to-back work items, which may be
2390 * being used to relieve memory pressure, don't
2391 * incur MAYDAY_INTERVAL delay inbetween.
2393 if (need_to_create_worker(pool
)) {
2394 spin_lock(&wq_mayday_lock
);
2396 list_move_tail(&pwq
->mayday_node
, &wq
->maydays
);
2397 spin_unlock(&wq_mayday_lock
);
2402 * Put the reference grabbed by send_mayday(). @pool won't
2403 * go away while we're still attached to it.
2408 * Leave this pool. If need_more_worker() is %true, notify a
2409 * regular worker; otherwise, we end up with 0 concurrency
2410 * and stalling the execution.
2412 if (need_more_worker(pool
))
2413 wake_up_worker(pool
);
2415 rescuer
->pool
= NULL
;
2416 spin_unlock_irq(&pool
->lock
);
2418 worker_detach_from_pool(rescuer
, pool
);
2420 spin_lock_irq(&wq_mayday_lock
);
2423 spin_unlock_irq(&wq_mayday_lock
);
2426 __set_current_state(TASK_RUNNING
);
2427 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2431 /* rescuers should never participate in concurrency management */
2432 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2438 * check_flush_dependency - check for flush dependency sanity
2439 * @target_wq: workqueue being flushed
2440 * @target_work: work item being flushed (NULL for workqueue flushes)
2442 * %current is trying to flush the whole @target_wq or @target_work on it.
2443 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2444 * reclaiming memory or running on a workqueue which doesn't have
2445 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2448 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2449 struct work_struct
*target_work
)
2451 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2452 struct worker
*worker
;
2454 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2457 worker
= current_wq_worker();
2459 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2460 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2461 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2462 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2463 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2464 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2465 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2466 target_wq
->name
, target_func
);
2470 struct work_struct work
;
2471 struct completion done
;
2472 struct task_struct
*task
; /* purely informational */
2475 static void wq_barrier_func(struct work_struct
*work
)
2477 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2478 complete(&barr
->done
);
2482 * insert_wq_barrier - insert a barrier work
2483 * @pwq: pwq to insert barrier into
2484 * @barr: wq_barrier to insert
2485 * @target: target work to attach @barr to
2486 * @worker: worker currently executing @target, NULL if @target is not executing
2488 * @barr is linked to @target such that @barr is completed only after
2489 * @target finishes execution. Please note that the ordering
2490 * guarantee is observed only with respect to @target and on the local
2493 * Currently, a queued barrier can't be canceled. This is because
2494 * try_to_grab_pending() can't determine whether the work to be
2495 * grabbed is at the head of the queue and thus can't clear LINKED
2496 * flag of the previous work while there must be a valid next work
2497 * after a work with LINKED flag set.
2499 * Note that when @worker is non-NULL, @target may be modified
2500 * underneath us, so we can't reliably determine pwq from @target.
2503 * spin_lock_irq(pool->lock).
2505 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2506 struct wq_barrier
*barr
,
2507 struct work_struct
*target
, struct worker
*worker
)
2509 struct list_head
*head
;
2510 unsigned int linked
= 0;
2513 * debugobject calls are safe here even with pool->lock locked
2514 * as we know for sure that this will not trigger any of the
2515 * checks and call back into the fixup functions where we
2518 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2519 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2522 * Explicitly init the crosslock for wq_barrier::done, make its lock
2523 * key a subkey of the corresponding work. As a result we won't
2524 * build a dependency between wq_barrier::done and unrelated work.
2526 lockdep_init_map_crosslock((struct lockdep_map
*)&barr
->done
.map
,
2527 "(complete)wq_barr::done",
2528 target
->lockdep_map
.key
, 1);
2529 __init_completion(&barr
->done
);
2530 barr
->task
= current
;
2533 * If @target is currently being executed, schedule the
2534 * barrier to the worker; otherwise, put it after @target.
2537 head
= worker
->scheduled
.next
;
2539 unsigned long *bits
= work_data_bits(target
);
2541 head
= target
->entry
.next
;
2542 /* there can already be other linked works, inherit and set */
2543 linked
= *bits
& WORK_STRUCT_LINKED
;
2544 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2547 debug_work_activate(&barr
->work
);
2548 insert_work(pwq
, &barr
->work
, head
,
2549 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2553 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2554 * @wq: workqueue being flushed
2555 * @flush_color: new flush color, < 0 for no-op
2556 * @work_color: new work color, < 0 for no-op
2558 * Prepare pwqs for workqueue flushing.
2560 * If @flush_color is non-negative, flush_color on all pwqs should be
2561 * -1. If no pwq has in-flight commands at the specified color, all
2562 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2563 * has in flight commands, its pwq->flush_color is set to
2564 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2565 * wakeup logic is armed and %true is returned.
2567 * The caller should have initialized @wq->first_flusher prior to
2568 * calling this function with non-negative @flush_color. If
2569 * @flush_color is negative, no flush color update is done and %false
2572 * If @work_color is non-negative, all pwqs should have the same
2573 * work_color which is previous to @work_color and all will be
2574 * advanced to @work_color.
2577 * mutex_lock(wq->mutex).
2580 * %true if @flush_color >= 0 and there's something to flush. %false
2583 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2584 int flush_color
, int work_color
)
2587 struct pool_workqueue
*pwq
;
2589 if (flush_color
>= 0) {
2590 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2591 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2594 for_each_pwq(pwq
, wq
) {
2595 struct worker_pool
*pool
= pwq
->pool
;
2597 spin_lock_irq(&pool
->lock
);
2599 if (flush_color
>= 0) {
2600 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2602 if (pwq
->nr_in_flight
[flush_color
]) {
2603 pwq
->flush_color
= flush_color
;
2604 atomic_inc(&wq
->nr_pwqs_to_flush
);
2609 if (work_color
>= 0) {
2610 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2611 pwq
->work_color
= work_color
;
2614 spin_unlock_irq(&pool
->lock
);
2617 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2618 complete(&wq
->first_flusher
->done
);
2624 * flush_workqueue - ensure that any scheduled work has run to completion.
2625 * @wq: workqueue to flush
2627 * This function sleeps until all work items which were queued on entry
2628 * have finished execution, but it is not livelocked by new incoming ones.
2630 void flush_workqueue(struct workqueue_struct
*wq
)
2632 struct wq_flusher this_flusher
= {
2633 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2635 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2639 if (WARN_ON(!wq_online
))
2642 lock_map_acquire(&wq
->lockdep_map
);
2643 lock_map_release(&wq
->lockdep_map
);
2645 mutex_lock(&wq
->mutex
);
2648 * Start-to-wait phase
2650 next_color
= work_next_color(wq
->work_color
);
2652 if (next_color
!= wq
->flush_color
) {
2654 * Color space is not full. The current work_color
2655 * becomes our flush_color and work_color is advanced
2658 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2659 this_flusher
.flush_color
= wq
->work_color
;
2660 wq
->work_color
= next_color
;
2662 if (!wq
->first_flusher
) {
2663 /* no flush in progress, become the first flusher */
2664 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2666 wq
->first_flusher
= &this_flusher
;
2668 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2670 /* nothing to flush, done */
2671 wq
->flush_color
= next_color
;
2672 wq
->first_flusher
= NULL
;
2677 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2678 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2679 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2683 * Oops, color space is full, wait on overflow queue.
2684 * The next flush completion will assign us
2685 * flush_color and transfer to flusher_queue.
2687 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2690 check_flush_dependency(wq
, NULL
);
2692 mutex_unlock(&wq
->mutex
);
2694 wait_for_completion(&this_flusher
.done
);
2697 * Wake-up-and-cascade phase
2699 * First flushers are responsible for cascading flushes and
2700 * handling overflow. Non-first flushers can simply return.
2702 if (wq
->first_flusher
!= &this_flusher
)
2705 mutex_lock(&wq
->mutex
);
2707 /* we might have raced, check again with mutex held */
2708 if (wq
->first_flusher
!= &this_flusher
)
2711 wq
->first_flusher
= NULL
;
2713 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2714 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2717 struct wq_flusher
*next
, *tmp
;
2719 /* complete all the flushers sharing the current flush color */
2720 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2721 if (next
->flush_color
!= wq
->flush_color
)
2723 list_del_init(&next
->list
);
2724 complete(&next
->done
);
2727 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2728 wq
->flush_color
!= work_next_color(wq
->work_color
));
2730 /* this flush_color is finished, advance by one */
2731 wq
->flush_color
= work_next_color(wq
->flush_color
);
2733 /* one color has been freed, handle overflow queue */
2734 if (!list_empty(&wq
->flusher_overflow
)) {
2736 * Assign the same color to all overflowed
2737 * flushers, advance work_color and append to
2738 * flusher_queue. This is the start-to-wait
2739 * phase for these overflowed flushers.
2741 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2742 tmp
->flush_color
= wq
->work_color
;
2744 wq
->work_color
= work_next_color(wq
->work_color
);
2746 list_splice_tail_init(&wq
->flusher_overflow
,
2747 &wq
->flusher_queue
);
2748 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2751 if (list_empty(&wq
->flusher_queue
)) {
2752 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2757 * Need to flush more colors. Make the next flusher
2758 * the new first flusher and arm pwqs.
2760 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2761 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2763 list_del_init(&next
->list
);
2764 wq
->first_flusher
= next
;
2766 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2770 * Meh... this color is already done, clear first
2771 * flusher and repeat cascading.
2773 wq
->first_flusher
= NULL
;
2777 mutex_unlock(&wq
->mutex
);
2779 EXPORT_SYMBOL(flush_workqueue
);
2782 * drain_workqueue - drain a workqueue
2783 * @wq: workqueue to drain
2785 * Wait until the workqueue becomes empty. While draining is in progress,
2786 * only chain queueing is allowed. IOW, only currently pending or running
2787 * work items on @wq can queue further work items on it. @wq is flushed
2788 * repeatedly until it becomes empty. The number of flushing is determined
2789 * by the depth of chaining and should be relatively short. Whine if it
2792 void drain_workqueue(struct workqueue_struct
*wq
)
2794 unsigned int flush_cnt
= 0;
2795 struct pool_workqueue
*pwq
;
2798 * __queue_work() needs to test whether there are drainers, is much
2799 * hotter than drain_workqueue() and already looks at @wq->flags.
2800 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2802 mutex_lock(&wq
->mutex
);
2803 if (!wq
->nr_drainers
++)
2804 wq
->flags
|= __WQ_DRAINING
;
2805 mutex_unlock(&wq
->mutex
);
2807 flush_workqueue(wq
);
2809 mutex_lock(&wq
->mutex
);
2811 for_each_pwq(pwq
, wq
) {
2814 spin_lock_irq(&pwq
->pool
->lock
);
2815 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2816 spin_unlock_irq(&pwq
->pool
->lock
);
2821 if (++flush_cnt
== 10 ||
2822 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2823 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2824 wq
->name
, flush_cnt
);
2826 mutex_unlock(&wq
->mutex
);
2830 if (!--wq
->nr_drainers
)
2831 wq
->flags
&= ~__WQ_DRAINING
;
2832 mutex_unlock(&wq
->mutex
);
2834 EXPORT_SYMBOL_GPL(drain_workqueue
);
2836 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2838 struct worker
*worker
= NULL
;
2839 struct worker_pool
*pool
;
2840 struct pool_workqueue
*pwq
;
2844 local_irq_disable();
2845 pool
= get_work_pool(work
);
2851 spin_lock(&pool
->lock
);
2852 /* see the comment in try_to_grab_pending() with the same code */
2853 pwq
= get_work_pwq(work
);
2855 if (unlikely(pwq
->pool
!= pool
))
2858 worker
= find_worker_executing_work(pool
, work
);
2861 pwq
= worker
->current_pwq
;
2864 check_flush_dependency(pwq
->wq
, work
);
2866 insert_wq_barrier(pwq
, barr
, work
, worker
);
2867 spin_unlock_irq(&pool
->lock
);
2870 * Force a lock recursion deadlock when using flush_work() inside a
2871 * single-threaded or rescuer equipped workqueue.
2873 * For single threaded workqueues the deadlock happens when the work
2874 * is after the work issuing the flush_work(). For rescuer equipped
2875 * workqueues the deadlock happens when the rescuer stalls, blocking
2878 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
) {
2879 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2880 lock_map_release(&pwq
->wq
->lockdep_map
);
2885 spin_unlock_irq(&pool
->lock
);
2890 * flush_work - wait for a work to finish executing the last queueing instance
2891 * @work: the work to flush
2893 * Wait until @work has finished execution. @work is guaranteed to be idle
2894 * on return if it hasn't been requeued since flush started.
2897 * %true if flush_work() waited for the work to finish execution,
2898 * %false if it was already idle.
2900 bool flush_work(struct work_struct
*work
)
2902 struct wq_barrier barr
;
2904 if (WARN_ON(!wq_online
))
2907 lock_map_acquire(&work
->lockdep_map
);
2908 lock_map_release(&work
->lockdep_map
);
2910 if (start_flush_work(work
, &barr
)) {
2911 wait_for_completion(&barr
.done
);
2912 destroy_work_on_stack(&barr
.work
);
2918 EXPORT_SYMBOL_GPL(flush_work
);
2921 wait_queue_entry_t wait
;
2922 struct work_struct
*work
;
2925 static int cwt_wakefn(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
2927 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
2929 if (cwait
->work
!= key
)
2931 return autoremove_wake_function(wait
, mode
, sync
, key
);
2934 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2936 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
2937 unsigned long flags
;
2941 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2943 * If someone else is already canceling, wait for it to
2944 * finish. flush_work() doesn't work for PREEMPT_NONE
2945 * because we may get scheduled between @work's completion
2946 * and the other canceling task resuming and clearing
2947 * CANCELING - flush_work() will return false immediately
2948 * as @work is no longer busy, try_to_grab_pending() will
2949 * return -ENOENT as @work is still being canceled and the
2950 * other canceling task won't be able to clear CANCELING as
2951 * we're hogging the CPU.
2953 * Let's wait for completion using a waitqueue. As this
2954 * may lead to the thundering herd problem, use a custom
2955 * wake function which matches @work along with exclusive
2958 if (unlikely(ret
== -ENOENT
)) {
2959 struct cwt_wait cwait
;
2961 init_wait(&cwait
.wait
);
2962 cwait
.wait
.func
= cwt_wakefn
;
2965 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
2966 TASK_UNINTERRUPTIBLE
);
2967 if (work_is_canceling(work
))
2969 finish_wait(&cancel_waitq
, &cwait
.wait
);
2971 } while (unlikely(ret
< 0));
2973 /* tell other tasks trying to grab @work to back off */
2974 mark_work_canceling(work
);
2975 local_irq_restore(flags
);
2978 * This allows canceling during early boot. We know that @work
2984 clear_work_data(work
);
2987 * Paired with prepare_to_wait() above so that either
2988 * waitqueue_active() is visible here or !work_is_canceling() is
2992 if (waitqueue_active(&cancel_waitq
))
2993 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
2999 * cancel_work_sync - cancel a work and wait for it to finish
3000 * @work: the work to cancel
3002 * Cancel @work and wait for its execution to finish. This function
3003 * can be used even if the work re-queues itself or migrates to
3004 * another workqueue. On return from this function, @work is
3005 * guaranteed to be not pending or executing on any CPU.
3007 * cancel_work_sync(&delayed_work->work) must not be used for
3008 * delayed_work's. Use cancel_delayed_work_sync() instead.
3010 * The caller must ensure that the workqueue on which @work was last
3011 * queued can't be destroyed before this function returns.
3014 * %true if @work was pending, %false otherwise.
3016 bool cancel_work_sync(struct work_struct
*work
)
3018 return __cancel_work_timer(work
, false);
3020 EXPORT_SYMBOL_GPL(cancel_work_sync
);
3023 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3024 * @dwork: the delayed work to flush
3026 * Delayed timer is cancelled and the pending work is queued for
3027 * immediate execution. Like flush_work(), this function only
3028 * considers the last queueing instance of @dwork.
3031 * %true if flush_work() waited for the work to finish execution,
3032 * %false if it was already idle.
3034 bool flush_delayed_work(struct delayed_work
*dwork
)
3036 local_irq_disable();
3037 if (del_timer_sync(&dwork
->timer
))
3038 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3040 return flush_work(&dwork
->work
);
3042 EXPORT_SYMBOL(flush_delayed_work
);
3044 static bool __cancel_work(struct work_struct
*work
, bool is_dwork
)
3046 unsigned long flags
;
3050 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3051 } while (unlikely(ret
== -EAGAIN
));
3053 if (unlikely(ret
< 0))
3056 set_work_pool_and_clear_pending(work
, get_work_pool_id(work
));
3057 local_irq_restore(flags
);
3062 * See cancel_delayed_work()
3064 bool cancel_work(struct work_struct
*work
)
3066 return __cancel_work(work
, false);
3070 * cancel_delayed_work - cancel a delayed work
3071 * @dwork: delayed_work to cancel
3073 * Kill off a pending delayed_work.
3075 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3079 * The work callback function may still be running on return, unless
3080 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3081 * use cancel_delayed_work_sync() to wait on it.
3083 * This function is safe to call from any context including IRQ handler.
3085 bool cancel_delayed_work(struct delayed_work
*dwork
)
3087 return __cancel_work(&dwork
->work
, true);
3089 EXPORT_SYMBOL(cancel_delayed_work
);
3092 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3093 * @dwork: the delayed work cancel
3095 * This is cancel_work_sync() for delayed works.
3098 * %true if @dwork was pending, %false otherwise.
3100 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3102 return __cancel_work_timer(&dwork
->work
, true);
3104 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3107 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3108 * @func: the function to call
3110 * schedule_on_each_cpu() executes @func on each online CPU using the
3111 * system workqueue and blocks until all CPUs have completed.
3112 * schedule_on_each_cpu() is very slow.
3115 * 0 on success, -errno on failure.
3117 int schedule_on_each_cpu(work_func_t func
)
3120 struct work_struct __percpu
*works
;
3122 works
= alloc_percpu(struct work_struct
);
3128 for_each_online_cpu(cpu
) {
3129 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3131 INIT_WORK(work
, func
);
3132 schedule_work_on(cpu
, work
);
3135 for_each_online_cpu(cpu
)
3136 flush_work(per_cpu_ptr(works
, cpu
));
3144 * execute_in_process_context - reliably execute the routine with user context
3145 * @fn: the function to execute
3146 * @ew: guaranteed storage for the execute work structure (must
3147 * be available when the work executes)
3149 * Executes the function immediately if process context is available,
3150 * otherwise schedules the function for delayed execution.
3152 * Return: 0 - function was executed
3153 * 1 - function was scheduled for execution
3155 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3157 if (!in_interrupt()) {
3162 INIT_WORK(&ew
->work
, fn
);
3163 schedule_work(&ew
->work
);
3167 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3170 * free_workqueue_attrs - free a workqueue_attrs
3171 * @attrs: workqueue_attrs to free
3173 * Undo alloc_workqueue_attrs().
3175 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3178 free_cpumask_var(attrs
->cpumask
);
3184 * alloc_workqueue_attrs - allocate a workqueue_attrs
3185 * @gfp_mask: allocation mask to use
3187 * Allocate a new workqueue_attrs, initialize with default settings and
3190 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3192 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3194 struct workqueue_attrs
*attrs
;
3196 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3199 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3202 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3205 free_workqueue_attrs(attrs
);
3209 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3210 const struct workqueue_attrs
*from
)
3212 to
->nice
= from
->nice
;
3213 cpumask_copy(to
->cpumask
, from
->cpumask
);
3215 * Unlike hash and equality test, this function doesn't ignore
3216 * ->no_numa as it is used for both pool and wq attrs. Instead,
3217 * get_unbound_pool() explicitly clears ->no_numa after copying.
3219 to
->no_numa
= from
->no_numa
;
3222 /* hash value of the content of @attr */
3223 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3227 hash
= jhash_1word(attrs
->nice
, hash
);
3228 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3229 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3233 /* content equality test */
3234 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3235 const struct workqueue_attrs
*b
)
3237 if (a
->nice
!= b
->nice
)
3239 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3245 * init_worker_pool - initialize a newly zalloc'd worker_pool
3246 * @pool: worker_pool to initialize
3248 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3250 * Return: 0 on success, -errno on failure. Even on failure, all fields
3251 * inside @pool proper are initialized and put_unbound_pool() can be called
3252 * on @pool safely to release it.
3254 static int init_worker_pool(struct worker_pool
*pool
)
3256 spin_lock_init(&pool
->lock
);
3259 pool
->node
= NUMA_NO_NODE
;
3260 pool
->flags
|= POOL_DISASSOCIATED
;
3261 pool
->watchdog_ts
= jiffies
;
3262 INIT_LIST_HEAD(&pool
->worklist
);
3263 INIT_LIST_HEAD(&pool
->idle_list
);
3264 hash_init(pool
->busy_hash
);
3266 setup_deferrable_timer(&pool
->idle_timer
, idle_worker_timeout
,
3267 (unsigned long)pool
);
3269 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3270 (unsigned long)pool
);
3272 mutex_init(&pool
->attach_mutex
);
3273 INIT_LIST_HEAD(&pool
->workers
);
3275 ida_init(&pool
->worker_ida
);
3276 INIT_HLIST_NODE(&pool
->hash_node
);
3279 /* shouldn't fail above this point */
3280 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3286 static void rcu_free_wq(struct rcu_head
*rcu
)
3288 struct workqueue_struct
*wq
=
3289 container_of(rcu
, struct workqueue_struct
, rcu
);
3291 if (!(wq
->flags
& WQ_UNBOUND
))
3292 free_percpu(wq
->cpu_pwqs
);
3294 free_workqueue_attrs(wq
->unbound_attrs
);
3300 static void rcu_free_pool(struct rcu_head
*rcu
)
3302 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3304 ida_destroy(&pool
->worker_ida
);
3305 free_workqueue_attrs(pool
->attrs
);
3310 * put_unbound_pool - put a worker_pool
3311 * @pool: worker_pool to put
3313 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3314 * safe manner. get_unbound_pool() calls this function on its failure path
3315 * and this function should be able to release pools which went through,
3316 * successfully or not, init_worker_pool().
3318 * Should be called with wq_pool_mutex held.
3320 static void put_unbound_pool(struct worker_pool
*pool
)
3322 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3323 struct worker
*worker
;
3325 lockdep_assert_held(&wq_pool_mutex
);
3331 if (WARN_ON(!(pool
->cpu
< 0)) ||
3332 WARN_ON(!list_empty(&pool
->worklist
)))
3335 /* release id and unhash */
3337 idr_remove(&worker_pool_idr
, pool
->id
);
3338 hash_del(&pool
->hash_node
);
3341 * Become the manager and destroy all workers. This prevents
3342 * @pool's workers from blocking on attach_mutex. We're the last
3343 * manager and @pool gets freed with the flag set.
3345 spin_lock_irq(&pool
->lock
);
3346 wait_event_lock_irq(wq_manager_wait
,
3347 !(pool
->flags
& POOL_MANAGER_ACTIVE
), pool
->lock
);
3348 pool
->flags
|= POOL_MANAGER_ACTIVE
;
3350 while ((worker
= first_idle_worker(pool
)))
3351 destroy_worker(worker
);
3352 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3353 spin_unlock_irq(&pool
->lock
);
3355 mutex_lock(&pool
->attach_mutex
);
3356 if (!list_empty(&pool
->workers
))
3357 pool
->detach_completion
= &detach_completion
;
3358 mutex_unlock(&pool
->attach_mutex
);
3360 if (pool
->detach_completion
)
3361 wait_for_completion(pool
->detach_completion
);
3363 /* shut down the timers */
3364 del_timer_sync(&pool
->idle_timer
);
3365 del_timer_sync(&pool
->mayday_timer
);
3367 /* sched-RCU protected to allow dereferences from get_work_pool() */
3368 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3372 * get_unbound_pool - get a worker_pool with the specified attributes
3373 * @attrs: the attributes of the worker_pool to get
3375 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3376 * reference count and return it. If there already is a matching
3377 * worker_pool, it will be used; otherwise, this function attempts to
3380 * Should be called with wq_pool_mutex held.
3382 * Return: On success, a worker_pool with the same attributes as @attrs.
3383 * On failure, %NULL.
3385 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3387 u32 hash
= wqattrs_hash(attrs
);
3388 struct worker_pool
*pool
;
3390 int target_node
= NUMA_NO_NODE
;
3392 lockdep_assert_held(&wq_pool_mutex
);
3394 /* do we already have a matching pool? */
3395 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3396 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3402 /* if cpumask is contained inside a NUMA node, we belong to that node */
3403 if (wq_numa_enabled
) {
3404 for_each_node(node
) {
3405 if (cpumask_subset(attrs
->cpumask
,
3406 wq_numa_possible_cpumask
[node
])) {
3413 /* nope, create a new one */
3414 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3415 if (!pool
|| init_worker_pool(pool
) < 0)
3418 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3419 copy_workqueue_attrs(pool
->attrs
, attrs
);
3420 pool
->node
= target_node
;
3423 * no_numa isn't a worker_pool attribute, always clear it. See
3424 * 'struct workqueue_attrs' comments for detail.
3426 pool
->attrs
->no_numa
= false;
3428 if (worker_pool_assign_id(pool
) < 0)
3431 /* create and start the initial worker */
3432 if (wq_online
&& !create_worker(pool
))
3436 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3441 put_unbound_pool(pool
);
3445 static void rcu_free_pwq(struct rcu_head
*rcu
)
3447 kmem_cache_free(pwq_cache
,
3448 container_of(rcu
, struct pool_workqueue
, rcu
));
3452 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3453 * and needs to be destroyed.
3455 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3457 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3458 unbound_release_work
);
3459 struct workqueue_struct
*wq
= pwq
->wq
;
3460 struct worker_pool
*pool
= pwq
->pool
;
3463 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3466 mutex_lock(&wq
->mutex
);
3467 list_del_rcu(&pwq
->pwqs_node
);
3468 is_last
= list_empty(&wq
->pwqs
);
3469 mutex_unlock(&wq
->mutex
);
3471 mutex_lock(&wq_pool_mutex
);
3472 put_unbound_pool(pool
);
3473 mutex_unlock(&wq_pool_mutex
);
3475 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3478 * If we're the last pwq going away, @wq is already dead and no one
3479 * is gonna access it anymore. Schedule RCU free.
3482 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
3486 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3487 * @pwq: target pool_workqueue
3489 * If @pwq isn't freezing, set @pwq->max_active to the associated
3490 * workqueue's saved_max_active and activate delayed work items
3491 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3493 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3495 struct workqueue_struct
*wq
= pwq
->wq
;
3496 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3497 unsigned long flags
;
3499 /* for @wq->saved_max_active */
3500 lockdep_assert_held(&wq
->mutex
);
3502 /* fast exit for non-freezable wqs */
3503 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3506 /* this function can be called during early boot w/ irq disabled */
3507 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
3510 * During [un]freezing, the caller is responsible for ensuring that
3511 * this function is called at least once after @workqueue_freezing
3512 * is updated and visible.
3514 if (!freezable
|| !workqueue_freezing
) {
3515 pwq
->max_active
= wq
->saved_max_active
;
3517 while (!list_empty(&pwq
->delayed_works
) &&
3518 pwq
->nr_active
< pwq
->max_active
)
3519 pwq_activate_first_delayed(pwq
);
3522 * Need to kick a worker after thawed or an unbound wq's
3523 * max_active is bumped. It's a slow path. Do it always.
3525 wake_up_worker(pwq
->pool
);
3527 pwq
->max_active
= 0;
3530 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
3533 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3534 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3535 struct worker_pool
*pool
)
3537 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3539 memset(pwq
, 0, sizeof(*pwq
));
3543 pwq
->flush_color
= -1;
3545 INIT_LIST_HEAD(&pwq
->delayed_works
);
3546 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3547 INIT_LIST_HEAD(&pwq
->mayday_node
);
3548 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3551 /* sync @pwq with the current state of its associated wq and link it */
3552 static void link_pwq(struct pool_workqueue
*pwq
)
3554 struct workqueue_struct
*wq
= pwq
->wq
;
3556 lockdep_assert_held(&wq
->mutex
);
3558 /* may be called multiple times, ignore if already linked */
3559 if (!list_empty(&pwq
->pwqs_node
))
3562 /* set the matching work_color */
3563 pwq
->work_color
= wq
->work_color
;
3565 /* sync max_active to the current setting */
3566 pwq_adjust_max_active(pwq
);
3569 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3572 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3573 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3574 const struct workqueue_attrs
*attrs
)
3576 struct worker_pool
*pool
;
3577 struct pool_workqueue
*pwq
;
3579 lockdep_assert_held(&wq_pool_mutex
);
3581 pool
= get_unbound_pool(attrs
);
3585 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3587 put_unbound_pool(pool
);
3591 init_pwq(pwq
, wq
, pool
);
3596 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3597 * @attrs: the wq_attrs of the default pwq of the target workqueue
3598 * @node: the target NUMA node
3599 * @cpu_going_down: if >= 0, the CPU to consider as offline
3600 * @cpumask: outarg, the resulting cpumask
3602 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3603 * @cpu_going_down is >= 0, that cpu is considered offline during
3604 * calculation. The result is stored in @cpumask.
3606 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3607 * enabled and @node has online CPUs requested by @attrs, the returned
3608 * cpumask is the intersection of the possible CPUs of @node and
3611 * The caller is responsible for ensuring that the cpumask of @node stays
3614 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3617 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3618 int cpu_going_down
, cpumask_t
*cpumask
)
3620 if (!wq_numa_enabled
|| attrs
->no_numa
)
3623 /* does @node have any online CPUs @attrs wants? */
3624 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3625 if (cpu_going_down
>= 0)
3626 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3628 if (cpumask_empty(cpumask
))
3631 /* yeap, return possible CPUs in @node that @attrs wants */
3632 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3634 if (cpumask_empty(cpumask
)) {
3635 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3636 "possible intersect\n");
3640 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3643 cpumask_copy(cpumask
, attrs
->cpumask
);
3647 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3648 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3650 struct pool_workqueue
*pwq
)
3652 struct pool_workqueue
*old_pwq
;
3654 lockdep_assert_held(&wq_pool_mutex
);
3655 lockdep_assert_held(&wq
->mutex
);
3657 /* link_pwq() can handle duplicate calls */
3660 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3661 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3665 /* context to store the prepared attrs & pwqs before applying */
3666 struct apply_wqattrs_ctx
{
3667 struct workqueue_struct
*wq
; /* target workqueue */
3668 struct workqueue_attrs
*attrs
; /* attrs to apply */
3669 struct list_head list
; /* queued for batching commit */
3670 struct pool_workqueue
*dfl_pwq
;
3671 struct pool_workqueue
*pwq_tbl
[];
3674 /* free the resources after success or abort */
3675 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3681 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3682 put_pwq_unlocked(ctx
->dfl_pwq
);
3684 free_workqueue_attrs(ctx
->attrs
);
3690 /* allocate the attrs and pwqs for later installation */
3691 static struct apply_wqattrs_ctx
*
3692 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3693 const struct workqueue_attrs
*attrs
)
3695 struct apply_wqattrs_ctx
*ctx
;
3696 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3699 lockdep_assert_held(&wq_pool_mutex
);
3701 ctx
= kzalloc(sizeof(*ctx
) + nr_node_ids
* sizeof(ctx
->pwq_tbl
[0]),
3704 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3705 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3706 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3710 * Calculate the attrs of the default pwq.
3711 * If the user configured cpumask doesn't overlap with the
3712 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3714 copy_workqueue_attrs(new_attrs
, attrs
);
3715 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3716 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3717 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3720 * We may create multiple pwqs with differing cpumasks. Make a
3721 * copy of @new_attrs which will be modified and used to obtain
3724 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3727 * If something goes wrong during CPU up/down, we'll fall back to
3728 * the default pwq covering whole @attrs->cpumask. Always create
3729 * it even if we don't use it immediately.
3731 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3735 for_each_node(node
) {
3736 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3737 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3738 if (!ctx
->pwq_tbl
[node
])
3741 ctx
->dfl_pwq
->refcnt
++;
3742 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3746 /* save the user configured attrs and sanitize it. */
3747 copy_workqueue_attrs(new_attrs
, attrs
);
3748 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3749 ctx
->attrs
= new_attrs
;
3752 free_workqueue_attrs(tmp_attrs
);
3756 free_workqueue_attrs(tmp_attrs
);
3757 free_workqueue_attrs(new_attrs
);
3758 apply_wqattrs_cleanup(ctx
);
3762 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3763 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3767 /* all pwqs have been created successfully, let's install'em */
3768 mutex_lock(&ctx
->wq
->mutex
);
3770 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3772 /* save the previous pwq and install the new one */
3774 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
3775 ctx
->pwq_tbl
[node
]);
3777 /* @dfl_pwq might not have been used, ensure it's linked */
3778 link_pwq(ctx
->dfl_pwq
);
3779 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
3781 mutex_unlock(&ctx
->wq
->mutex
);
3784 static void apply_wqattrs_lock(void)
3786 /* CPUs should stay stable across pwq creations and installations */
3788 mutex_lock(&wq_pool_mutex
);
3791 static void apply_wqattrs_unlock(void)
3793 mutex_unlock(&wq_pool_mutex
);
3797 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
3798 const struct workqueue_attrs
*attrs
)
3800 struct apply_wqattrs_ctx
*ctx
;
3802 /* only unbound workqueues can change attributes */
3803 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3806 /* creating multiple pwqs breaks ordering guarantee */
3807 if (!list_empty(&wq
->pwqs
)) {
3808 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
3811 wq
->flags
&= ~__WQ_ORDERED
;
3814 ctx
= apply_wqattrs_prepare(wq
, attrs
);
3818 /* the ctx has been prepared successfully, let's commit it */
3819 apply_wqattrs_commit(ctx
);
3820 apply_wqattrs_cleanup(ctx
);
3826 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3827 * @wq: the target workqueue
3828 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3830 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3831 * machines, this function maps a separate pwq to each NUMA node with
3832 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3833 * NUMA node it was issued on. Older pwqs are released as in-flight work
3834 * items finish. Note that a work item which repeatedly requeues itself
3835 * back-to-back will stay on its current pwq.
3837 * Performs GFP_KERNEL allocations.
3839 * Return: 0 on success and -errno on failure.
3841 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3842 const struct workqueue_attrs
*attrs
)
3846 apply_wqattrs_lock();
3847 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
3848 apply_wqattrs_unlock();
3854 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3855 * @wq: the target workqueue
3856 * @cpu: the CPU coming up or going down
3857 * @online: whether @cpu is coming up or going down
3859 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3860 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3863 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3864 * falls back to @wq->dfl_pwq which may not be optimal but is always
3867 * Note that when the last allowed CPU of a NUMA node goes offline for a
3868 * workqueue with a cpumask spanning multiple nodes, the workers which were
3869 * already executing the work items for the workqueue will lose their CPU
3870 * affinity and may execute on any CPU. This is similar to how per-cpu
3871 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3872 * affinity, it's the user's responsibility to flush the work item from
3875 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3878 int node
= cpu_to_node(cpu
);
3879 int cpu_off
= online
? -1 : cpu
;
3880 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3881 struct workqueue_attrs
*target_attrs
;
3884 lockdep_assert_held(&wq_pool_mutex
);
3886 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
3887 wq
->unbound_attrs
->no_numa
)
3891 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3892 * Let's use a preallocated one. The following buf is protected by
3893 * CPU hotplug exclusion.
3895 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3896 cpumask
= target_attrs
->cpumask
;
3898 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3899 pwq
= unbound_pwq_by_node(wq
, node
);
3902 * Let's determine what needs to be done. If the target cpumask is
3903 * different from the default pwq's, we need to compare it to @pwq's
3904 * and create a new one if they don't match. If the target cpumask
3905 * equals the default pwq's, the default pwq should be used.
3907 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
3908 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3914 /* create a new pwq */
3915 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3917 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3922 /* Install the new pwq. */
3923 mutex_lock(&wq
->mutex
);
3924 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3928 mutex_lock(&wq
->mutex
);
3929 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3930 get_pwq(wq
->dfl_pwq
);
3931 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3932 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
3934 mutex_unlock(&wq
->mutex
);
3935 put_pwq_unlocked(old_pwq
);
3938 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3940 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3943 if (!(wq
->flags
& WQ_UNBOUND
)) {
3944 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3948 for_each_possible_cpu(cpu
) {
3949 struct pool_workqueue
*pwq
=
3950 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3951 struct worker_pool
*cpu_pools
=
3952 per_cpu(cpu_worker_pools
, cpu
);
3954 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
3956 mutex_lock(&wq
->mutex
);
3958 mutex_unlock(&wq
->mutex
);
3961 } else if (wq
->flags
& __WQ_ORDERED
) {
3962 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
3963 /* there should only be single pwq for ordering guarantee */
3964 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
3965 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
3966 "ordering guarantee broken for workqueue %s\n", wq
->name
);
3969 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3973 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3976 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3978 if (max_active
< 1 || max_active
> lim
)
3979 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3980 max_active
, name
, 1, lim
);
3982 return clamp_val(max_active
, 1, lim
);
3985 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3988 struct lock_class_key
*key
,
3989 const char *lock_name
, ...)
3991 size_t tbl_size
= 0;
3993 struct workqueue_struct
*wq
;
3994 struct pool_workqueue
*pwq
;
3997 * Unbound && max_active == 1 used to imply ordered, which is no
3998 * longer the case on NUMA machines due to per-node pools. While
3999 * alloc_ordered_workqueue() is the right way to create an ordered
4000 * workqueue, keep the previous behavior to avoid subtle breakages
4003 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
4004 flags
|= __WQ_ORDERED
;
4006 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4007 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4008 flags
|= WQ_UNBOUND
;
4010 /* allocate wq and format name */
4011 if (flags
& WQ_UNBOUND
)
4012 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
4014 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4018 if (flags
& WQ_UNBOUND
) {
4019 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4020 if (!wq
->unbound_attrs
)
4024 va_start(args
, lock_name
);
4025 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4028 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4029 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4033 wq
->saved_max_active
= max_active
;
4034 mutex_init(&wq
->mutex
);
4035 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4036 INIT_LIST_HEAD(&wq
->pwqs
);
4037 INIT_LIST_HEAD(&wq
->flusher_queue
);
4038 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4039 INIT_LIST_HEAD(&wq
->maydays
);
4041 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4042 INIT_LIST_HEAD(&wq
->list
);
4044 if (alloc_and_link_pwqs(wq
) < 0)
4048 * Workqueues which may be used during memory reclaim should
4049 * have a rescuer to guarantee forward progress.
4051 if (flags
& WQ_MEM_RECLAIM
) {
4052 struct worker
*rescuer
;
4054 rescuer
= alloc_worker(NUMA_NO_NODE
);
4058 rescuer
->rescue_wq
= wq
;
4059 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4061 if (IS_ERR(rescuer
->task
)) {
4066 wq
->rescuer
= rescuer
;
4067 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
4068 wake_up_process(rescuer
->task
);
4071 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4075 * wq_pool_mutex protects global freeze state and workqueues list.
4076 * Grab it, adjust max_active and add the new @wq to workqueues
4079 mutex_lock(&wq_pool_mutex
);
4081 mutex_lock(&wq
->mutex
);
4082 for_each_pwq(pwq
, wq
)
4083 pwq_adjust_max_active(pwq
);
4084 mutex_unlock(&wq
->mutex
);
4086 list_add_tail_rcu(&wq
->list
, &workqueues
);
4088 mutex_unlock(&wq_pool_mutex
);
4093 free_workqueue_attrs(wq
->unbound_attrs
);
4097 destroy_workqueue(wq
);
4100 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4103 * destroy_workqueue - safely terminate a workqueue
4104 * @wq: target workqueue
4106 * Safely destroy a workqueue. All work currently pending will be done first.
4108 void destroy_workqueue(struct workqueue_struct
*wq
)
4110 struct pool_workqueue
*pwq
;
4113 /* drain it before proceeding with destruction */
4114 drain_workqueue(wq
);
4117 mutex_lock(&wq
->mutex
);
4118 for_each_pwq(pwq
, wq
) {
4121 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4122 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4123 mutex_unlock(&wq
->mutex
);
4124 show_workqueue_state();
4129 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4130 WARN_ON(pwq
->nr_active
) ||
4131 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4132 mutex_unlock(&wq
->mutex
);
4133 show_workqueue_state();
4137 mutex_unlock(&wq
->mutex
);
4140 * wq list is used to freeze wq, remove from list after
4141 * flushing is complete in case freeze races us.
4143 mutex_lock(&wq_pool_mutex
);
4144 list_del_rcu(&wq
->list
);
4145 mutex_unlock(&wq_pool_mutex
);
4147 workqueue_sysfs_unregister(wq
);
4150 kthread_stop(wq
->rescuer
->task
);
4152 if (!(wq
->flags
& WQ_UNBOUND
)) {
4154 * The base ref is never dropped on per-cpu pwqs. Directly
4155 * schedule RCU free.
4157 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
4160 * We're the sole accessor of @wq at this point. Directly
4161 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4162 * @wq will be freed when the last pwq is released.
4164 for_each_node(node
) {
4165 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4166 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4167 put_pwq_unlocked(pwq
);
4171 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4172 * put. Don't access it afterwards.
4176 put_pwq_unlocked(pwq
);
4179 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4182 * workqueue_set_max_active - adjust max_active of a workqueue
4183 * @wq: target workqueue
4184 * @max_active: new max_active value.
4186 * Set max_active of @wq to @max_active.
4189 * Don't call from IRQ context.
4191 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4193 struct pool_workqueue
*pwq
;
4195 /* disallow meddling with max_active for ordered workqueues */
4196 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4199 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4201 mutex_lock(&wq
->mutex
);
4203 wq
->flags
&= ~__WQ_ORDERED
;
4204 wq
->saved_max_active
= max_active
;
4206 for_each_pwq(pwq
, wq
)
4207 pwq_adjust_max_active(pwq
);
4209 mutex_unlock(&wq
->mutex
);
4211 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4214 * current_work - retrieve %current task's work struct
4216 * Determine if %current task is a workqueue worker and what it's working on.
4217 * Useful to find out the context that the %current task is running in.
4219 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4221 struct work_struct
*current_work(void)
4223 struct worker
*worker
= current_wq_worker();
4225 return worker
? worker
->current_work
: NULL
;
4227 EXPORT_SYMBOL(current_work
);
4230 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4232 * Determine whether %current is a workqueue rescuer. Can be used from
4233 * work functions to determine whether it's being run off the rescuer task.
4235 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4237 bool current_is_workqueue_rescuer(void)
4239 struct worker
*worker
= current_wq_worker();
4241 return worker
&& worker
->rescue_wq
;
4245 * workqueue_congested - test whether a workqueue is congested
4246 * @cpu: CPU in question
4247 * @wq: target workqueue
4249 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4250 * no synchronization around this function and the test result is
4251 * unreliable and only useful as advisory hints or for debugging.
4253 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4254 * Note that both per-cpu and unbound workqueues may be associated with
4255 * multiple pool_workqueues which have separate congested states. A
4256 * workqueue being congested on one CPU doesn't mean the workqueue is also
4257 * contested on other CPUs / NUMA nodes.
4260 * %true if congested, %false otherwise.
4262 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4264 struct pool_workqueue
*pwq
;
4267 rcu_read_lock_sched();
4269 if (cpu
== WORK_CPU_UNBOUND
)
4270 cpu
= smp_processor_id();
4272 if (!(wq
->flags
& WQ_UNBOUND
))
4273 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4275 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4277 ret
= !list_empty(&pwq
->delayed_works
);
4278 rcu_read_unlock_sched();
4282 EXPORT_SYMBOL_GPL(workqueue_congested
);
4285 * work_busy - test whether a work is currently pending or running
4286 * @work: the work to be tested
4288 * Test whether @work is currently pending or running. There is no
4289 * synchronization around this function and the test result is
4290 * unreliable and only useful as advisory hints or for debugging.
4293 * OR'd bitmask of WORK_BUSY_* bits.
4295 unsigned int work_busy(struct work_struct
*work
)
4297 struct worker_pool
*pool
;
4298 unsigned long flags
;
4299 unsigned int ret
= 0;
4301 if (work_pending(work
))
4302 ret
|= WORK_BUSY_PENDING
;
4304 local_irq_save(flags
);
4305 pool
= get_work_pool(work
);
4307 spin_lock(&pool
->lock
);
4308 if (find_worker_executing_work(pool
, work
))
4309 ret
|= WORK_BUSY_RUNNING
;
4310 spin_unlock(&pool
->lock
);
4312 local_irq_restore(flags
);
4316 EXPORT_SYMBOL_GPL(work_busy
);
4319 * set_worker_desc - set description for the current work item
4320 * @fmt: printf-style format string
4321 * @...: arguments for the format string
4323 * This function can be called by a running work function to describe what
4324 * the work item is about. If the worker task gets dumped, this
4325 * information will be printed out together to help debugging. The
4326 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4328 void set_worker_desc(const char *fmt
, ...)
4330 struct worker
*worker
= current_wq_worker();
4334 va_start(args
, fmt
);
4335 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4337 worker
->desc_valid
= true;
4342 * print_worker_info - print out worker information and description
4343 * @log_lvl: the log level to use when printing
4344 * @task: target task
4346 * If @task is a worker and currently executing a work item, print out the
4347 * name of the workqueue being serviced and worker description set with
4348 * set_worker_desc() by the currently executing work item.
4350 * This function can be safely called on any task as long as the
4351 * task_struct itself is accessible. While safe, this function isn't
4352 * synchronized and may print out mixups or garbages of limited length.
4354 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4356 work_func_t
*fn
= NULL
;
4357 char name
[WQ_NAME_LEN
] = { };
4358 char desc
[WORKER_DESC_LEN
] = { };
4359 struct pool_workqueue
*pwq
= NULL
;
4360 struct workqueue_struct
*wq
= NULL
;
4361 bool desc_valid
= false;
4362 struct worker
*worker
;
4364 if (!(task
->flags
& PF_WQ_WORKER
))
4368 * This function is called without any synchronization and @task
4369 * could be in any state. Be careful with dereferences.
4371 worker
= kthread_probe_data(task
);
4374 * Carefully copy the associated workqueue's workfn and name. Keep
4375 * the original last '\0' in case the original contains garbage.
4377 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4378 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4379 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4380 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4382 /* copy worker description */
4383 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4385 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4387 if (fn
|| name
[0] || desc
[0]) {
4388 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4390 pr_cont(" (%s)", desc
);
4395 static void pr_cont_pool_info(struct worker_pool
*pool
)
4397 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4398 if (pool
->node
!= NUMA_NO_NODE
)
4399 pr_cont(" node=%d", pool
->node
);
4400 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4403 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4405 if (work
->func
== wq_barrier_func
) {
4406 struct wq_barrier
*barr
;
4408 barr
= container_of(work
, struct wq_barrier
, work
);
4410 pr_cont("%s BAR(%d)", comma
? "," : "",
4411 task_pid_nr(barr
->task
));
4413 pr_cont("%s %pf", comma
? "," : "", work
->func
);
4417 static void show_pwq(struct pool_workqueue
*pwq
)
4419 struct worker_pool
*pool
= pwq
->pool
;
4420 struct work_struct
*work
;
4421 struct worker
*worker
;
4422 bool has_in_flight
= false, has_pending
= false;
4425 pr_info(" pwq %d:", pool
->id
);
4426 pr_cont_pool_info(pool
);
4428 pr_cont(" active=%d/%d%s\n", pwq
->nr_active
, pwq
->max_active
,
4429 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4431 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4432 if (worker
->current_pwq
== pwq
) {
4433 has_in_flight
= true;
4437 if (has_in_flight
) {
4440 pr_info(" in-flight:");
4441 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4442 if (worker
->current_pwq
!= pwq
)
4445 pr_cont("%s %d%s:%pf", comma
? "," : "",
4446 task_pid_nr(worker
->task
),
4447 worker
== pwq
->wq
->rescuer
? "(RESCUER)" : "",
4448 worker
->current_func
);
4449 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4450 pr_cont_work(false, work
);
4456 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4457 if (get_work_pwq(work
) == pwq
) {
4465 pr_info(" pending:");
4466 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4467 if (get_work_pwq(work
) != pwq
)
4470 pr_cont_work(comma
, work
);
4471 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4476 if (!list_empty(&pwq
->delayed_works
)) {
4479 pr_info(" delayed:");
4480 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4481 pr_cont_work(comma
, work
);
4482 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4489 * show_workqueue_state - dump workqueue state
4491 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4492 * all busy workqueues and pools.
4494 void show_workqueue_state(void)
4496 struct workqueue_struct
*wq
;
4497 struct worker_pool
*pool
;
4498 unsigned long flags
;
4501 rcu_read_lock_sched();
4503 pr_info("Showing busy workqueues and worker pools:\n");
4505 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4506 struct pool_workqueue
*pwq
;
4509 for_each_pwq(pwq
, wq
) {
4510 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4518 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4520 for_each_pwq(pwq
, wq
) {
4521 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4522 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4524 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4526 * We could be printing a lot from atomic context, e.g.
4527 * sysrq-t -> show_workqueue_state(). Avoid triggering
4530 touch_nmi_watchdog();
4534 for_each_pool(pool
, pi
) {
4535 struct worker
*worker
;
4538 spin_lock_irqsave(&pool
->lock
, flags
);
4539 if (pool
->nr_workers
== pool
->nr_idle
)
4542 pr_info("pool %d:", pool
->id
);
4543 pr_cont_pool_info(pool
);
4544 pr_cont(" hung=%us workers=%d",
4545 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4548 pr_cont(" manager: %d",
4549 task_pid_nr(pool
->manager
->task
));
4550 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4551 pr_cont(" %s%d", first
? "idle: " : "",
4552 task_pid_nr(worker
->task
));
4557 spin_unlock_irqrestore(&pool
->lock
, flags
);
4559 * We could be printing a lot from atomic context, e.g.
4560 * sysrq-t -> show_workqueue_state(). Avoid triggering
4563 touch_nmi_watchdog();
4566 rcu_read_unlock_sched();
4572 * There are two challenges in supporting CPU hotplug. Firstly, there
4573 * are a lot of assumptions on strong associations among work, pwq and
4574 * pool which make migrating pending and scheduled works very
4575 * difficult to implement without impacting hot paths. Secondly,
4576 * worker pools serve mix of short, long and very long running works making
4577 * blocked draining impractical.
4579 * This is solved by allowing the pools to be disassociated from the CPU
4580 * running as an unbound one and allowing it to be reattached later if the
4581 * cpu comes back online.
4584 static void wq_unbind_fn(struct work_struct
*work
)
4586 int cpu
= smp_processor_id();
4587 struct worker_pool
*pool
;
4588 struct worker
*worker
;
4590 for_each_cpu_worker_pool(pool
, cpu
) {
4591 mutex_lock(&pool
->attach_mutex
);
4592 spin_lock_irq(&pool
->lock
);
4595 * We've blocked all attach/detach operations. Make all workers
4596 * unbound and set DISASSOCIATED. Before this, all workers
4597 * except for the ones which are still executing works from
4598 * before the last CPU down must be on the cpu. After
4599 * this, they may become diasporas.
4601 for_each_pool_worker(worker
, pool
)
4602 worker
->flags
|= WORKER_UNBOUND
;
4604 pool
->flags
|= POOL_DISASSOCIATED
;
4606 spin_unlock_irq(&pool
->lock
);
4607 mutex_unlock(&pool
->attach_mutex
);
4610 * Call schedule() so that we cross rq->lock and thus can
4611 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4612 * This is necessary as scheduler callbacks may be invoked
4618 * Sched callbacks are disabled now. Zap nr_running.
4619 * After this, nr_running stays zero and need_more_worker()
4620 * and keep_working() are always true as long as the
4621 * worklist is not empty. This pool now behaves as an
4622 * unbound (in terms of concurrency management) pool which
4623 * are served by workers tied to the pool.
4625 atomic_set(&pool
->nr_running
, 0);
4628 * With concurrency management just turned off, a busy
4629 * worker blocking could lead to lengthy stalls. Kick off
4630 * unbound chain execution of currently pending work items.
4632 spin_lock_irq(&pool
->lock
);
4633 wake_up_worker(pool
);
4634 spin_unlock_irq(&pool
->lock
);
4639 * rebind_workers - rebind all workers of a pool to the associated CPU
4640 * @pool: pool of interest
4642 * @pool->cpu is coming online. Rebind all workers to the CPU.
4644 static void rebind_workers(struct worker_pool
*pool
)
4646 struct worker
*worker
;
4648 lockdep_assert_held(&pool
->attach_mutex
);
4651 * Restore CPU affinity of all workers. As all idle workers should
4652 * be on the run-queue of the associated CPU before any local
4653 * wake-ups for concurrency management happen, restore CPU affinity
4654 * of all workers first and then clear UNBOUND. As we're called
4655 * from CPU_ONLINE, the following shouldn't fail.
4657 for_each_pool_worker(worker
, pool
)
4658 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4659 pool
->attrs
->cpumask
) < 0);
4661 spin_lock_irq(&pool
->lock
);
4664 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4665 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4666 * being reworked and this can go away in time.
4668 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
4669 spin_unlock_irq(&pool
->lock
);
4673 pool
->flags
&= ~POOL_DISASSOCIATED
;
4675 for_each_pool_worker(worker
, pool
) {
4676 unsigned int worker_flags
= worker
->flags
;
4679 * A bound idle worker should actually be on the runqueue
4680 * of the associated CPU for local wake-ups targeting it to
4681 * work. Kick all idle workers so that they migrate to the
4682 * associated CPU. Doing this in the same loop as
4683 * replacing UNBOUND with REBOUND is safe as no worker will
4684 * be bound before @pool->lock is released.
4686 if (worker_flags
& WORKER_IDLE
)
4687 wake_up_process(worker
->task
);
4690 * We want to clear UNBOUND but can't directly call
4691 * worker_clr_flags() or adjust nr_running. Atomically
4692 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4693 * @worker will clear REBOUND using worker_clr_flags() when
4694 * it initiates the next execution cycle thus restoring
4695 * concurrency management. Note that when or whether
4696 * @worker clears REBOUND doesn't affect correctness.
4698 * ACCESS_ONCE() is necessary because @worker->flags may be
4699 * tested without holding any lock in
4700 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4701 * fail incorrectly leading to premature concurrency
4702 * management operations.
4704 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4705 worker_flags
|= WORKER_REBOUND
;
4706 worker_flags
&= ~WORKER_UNBOUND
;
4707 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4710 spin_unlock_irq(&pool
->lock
);
4714 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4715 * @pool: unbound pool of interest
4716 * @cpu: the CPU which is coming up
4718 * An unbound pool may end up with a cpumask which doesn't have any online
4719 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4720 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4721 * online CPU before, cpus_allowed of all its workers should be restored.
4723 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4725 static cpumask_t cpumask
;
4726 struct worker
*worker
;
4728 lockdep_assert_held(&pool
->attach_mutex
);
4730 /* is @cpu allowed for @pool? */
4731 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4734 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4736 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4737 for_each_pool_worker(worker
, pool
)
4738 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, &cpumask
) < 0);
4741 int workqueue_prepare_cpu(unsigned int cpu
)
4743 struct worker_pool
*pool
;
4745 for_each_cpu_worker_pool(pool
, cpu
) {
4746 if (pool
->nr_workers
)
4748 if (!create_worker(pool
))
4754 int workqueue_online_cpu(unsigned int cpu
)
4756 struct worker_pool
*pool
;
4757 struct workqueue_struct
*wq
;
4760 mutex_lock(&wq_pool_mutex
);
4762 for_each_pool(pool
, pi
) {
4763 mutex_lock(&pool
->attach_mutex
);
4765 if (pool
->cpu
== cpu
)
4766 rebind_workers(pool
);
4767 else if (pool
->cpu
< 0)
4768 restore_unbound_workers_cpumask(pool
, cpu
);
4770 mutex_unlock(&pool
->attach_mutex
);
4773 /* update NUMA affinity of unbound workqueues */
4774 list_for_each_entry(wq
, &workqueues
, list
)
4775 wq_update_unbound_numa(wq
, cpu
, true);
4777 mutex_unlock(&wq_pool_mutex
);
4781 int workqueue_offline_cpu(unsigned int cpu
)
4783 struct work_struct unbind_work
;
4784 struct workqueue_struct
*wq
;
4786 /* unbinding per-cpu workers should happen on the local CPU */
4787 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4788 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4790 /* update NUMA affinity of unbound workqueues */
4791 mutex_lock(&wq_pool_mutex
);
4792 list_for_each_entry(wq
, &workqueues
, list
)
4793 wq_update_unbound_numa(wq
, cpu
, false);
4794 mutex_unlock(&wq_pool_mutex
);
4796 /* wait for per-cpu unbinding to finish */
4797 flush_work(&unbind_work
);
4798 destroy_work_on_stack(&unbind_work
);
4804 struct work_for_cpu
{
4805 struct work_struct work
;
4811 static void work_for_cpu_fn(struct work_struct
*work
)
4813 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4815 wfc
->ret
= wfc
->fn(wfc
->arg
);
4819 * work_on_cpu - run a function in thread context on a particular cpu
4820 * @cpu: the cpu to run on
4821 * @fn: the function to run
4822 * @arg: the function arg
4824 * It is up to the caller to ensure that the cpu doesn't go offline.
4825 * The caller must not hold any locks which would prevent @fn from completing.
4827 * Return: The value @fn returns.
4829 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4831 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4833 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4834 schedule_work_on(cpu
, &wfc
.work
);
4835 flush_work(&wfc
.work
);
4836 destroy_work_on_stack(&wfc
.work
);
4839 EXPORT_SYMBOL_GPL(work_on_cpu
);
4842 * work_on_cpu_safe - run a function in thread context on a particular cpu
4843 * @cpu: the cpu to run on
4844 * @fn: the function to run
4845 * @arg: the function argument
4847 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4848 * any locks which would prevent @fn from completing.
4850 * Return: The value @fn returns.
4852 long work_on_cpu_safe(int cpu
, long (*fn
)(void *), void *arg
)
4857 if (cpu_online(cpu
))
4858 ret
= work_on_cpu(cpu
, fn
, arg
);
4862 EXPORT_SYMBOL_GPL(work_on_cpu_safe
);
4863 #endif /* CONFIG_SMP */
4865 #ifdef CONFIG_FREEZER
4868 * freeze_workqueues_begin - begin freezing workqueues
4870 * Start freezing workqueues. After this function returns, all freezable
4871 * workqueues will queue new works to their delayed_works list instead of
4875 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4877 void freeze_workqueues_begin(void)
4879 struct workqueue_struct
*wq
;
4880 struct pool_workqueue
*pwq
;
4882 mutex_lock(&wq_pool_mutex
);
4884 WARN_ON_ONCE(workqueue_freezing
);
4885 workqueue_freezing
= true;
4887 list_for_each_entry(wq
, &workqueues
, list
) {
4888 mutex_lock(&wq
->mutex
);
4889 for_each_pwq(pwq
, wq
)
4890 pwq_adjust_max_active(pwq
);
4891 mutex_unlock(&wq
->mutex
);
4894 mutex_unlock(&wq_pool_mutex
);
4898 * freeze_workqueues_busy - are freezable workqueues still busy?
4900 * Check whether freezing is complete. This function must be called
4901 * between freeze_workqueues_begin() and thaw_workqueues().
4904 * Grabs and releases wq_pool_mutex.
4907 * %true if some freezable workqueues are still busy. %false if freezing
4910 bool freeze_workqueues_busy(void)
4913 struct workqueue_struct
*wq
;
4914 struct pool_workqueue
*pwq
;
4916 mutex_lock(&wq_pool_mutex
);
4918 WARN_ON_ONCE(!workqueue_freezing
);
4920 list_for_each_entry(wq
, &workqueues
, list
) {
4921 if (!(wq
->flags
& WQ_FREEZABLE
))
4924 * nr_active is monotonically decreasing. It's safe
4925 * to peek without lock.
4927 rcu_read_lock_sched();
4928 for_each_pwq(pwq
, wq
) {
4929 WARN_ON_ONCE(pwq
->nr_active
< 0);
4930 if (pwq
->nr_active
) {
4932 rcu_read_unlock_sched();
4936 rcu_read_unlock_sched();
4939 mutex_unlock(&wq_pool_mutex
);
4944 * thaw_workqueues - thaw workqueues
4946 * Thaw workqueues. Normal queueing is restored and all collected
4947 * frozen works are transferred to their respective pool worklists.
4950 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4952 void thaw_workqueues(void)
4954 struct workqueue_struct
*wq
;
4955 struct pool_workqueue
*pwq
;
4957 mutex_lock(&wq_pool_mutex
);
4959 if (!workqueue_freezing
)
4962 workqueue_freezing
= false;
4964 /* restore max_active and repopulate worklist */
4965 list_for_each_entry(wq
, &workqueues
, list
) {
4966 mutex_lock(&wq
->mutex
);
4967 for_each_pwq(pwq
, wq
)
4968 pwq_adjust_max_active(pwq
);
4969 mutex_unlock(&wq
->mutex
);
4973 mutex_unlock(&wq_pool_mutex
);
4975 #endif /* CONFIG_FREEZER */
4977 static int workqueue_apply_unbound_cpumask(void)
4981 struct workqueue_struct
*wq
;
4982 struct apply_wqattrs_ctx
*ctx
, *n
;
4984 lockdep_assert_held(&wq_pool_mutex
);
4986 list_for_each_entry(wq
, &workqueues
, list
) {
4987 if (!(wq
->flags
& WQ_UNBOUND
))
4989 /* creating multiple pwqs breaks ordering guarantee */
4990 if (wq
->flags
& __WQ_ORDERED
)
4993 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
4999 list_add_tail(&ctx
->list
, &ctxs
);
5002 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
5004 apply_wqattrs_commit(ctx
);
5005 apply_wqattrs_cleanup(ctx
);
5012 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5013 * @cpumask: the cpumask to set
5015 * The low-level workqueues cpumask is a global cpumask that limits
5016 * the affinity of all unbound workqueues. This function check the @cpumask
5017 * and apply it to all unbound workqueues and updates all pwqs of them.
5019 * Retun: 0 - Success
5020 * -EINVAL - Invalid @cpumask
5021 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5023 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
5026 cpumask_var_t saved_cpumask
;
5028 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
5031 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
5032 if (!cpumask_empty(cpumask
)) {
5033 apply_wqattrs_lock();
5035 /* save the old wq_unbound_cpumask. */
5036 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
5038 /* update wq_unbound_cpumask at first and apply it to wqs. */
5039 cpumask_copy(wq_unbound_cpumask
, cpumask
);
5040 ret
= workqueue_apply_unbound_cpumask();
5042 /* restore the wq_unbound_cpumask when failed. */
5044 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
5046 apply_wqattrs_unlock();
5049 free_cpumask_var(saved_cpumask
);
5055 * Workqueues with WQ_SYSFS flag set is visible to userland via
5056 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5057 * following attributes.
5059 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5060 * max_active RW int : maximum number of in-flight work items
5062 * Unbound workqueues have the following extra attributes.
5064 * id RO int : the associated pool ID
5065 * nice RW int : nice value of the workers
5066 * cpumask RW mask : bitmask of allowed CPUs for the workers
5069 struct workqueue_struct
*wq
;
5073 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
5075 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5080 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
5083 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5085 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
5087 static DEVICE_ATTR_RO(per_cpu
);
5089 static ssize_t
max_active_show(struct device
*dev
,
5090 struct device_attribute
*attr
, char *buf
)
5092 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5094 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
5097 static ssize_t
max_active_store(struct device
*dev
,
5098 struct device_attribute
*attr
, const char *buf
,
5101 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5104 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
5107 workqueue_set_max_active(wq
, val
);
5110 static DEVICE_ATTR_RW(max_active
);
5112 static struct attribute
*wq_sysfs_attrs
[] = {
5113 &dev_attr_per_cpu
.attr
,
5114 &dev_attr_max_active
.attr
,
5117 ATTRIBUTE_GROUPS(wq_sysfs
);
5119 static ssize_t
wq_pool_ids_show(struct device
*dev
,
5120 struct device_attribute
*attr
, char *buf
)
5122 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5123 const char *delim
= "";
5124 int node
, written
= 0;
5126 rcu_read_lock_sched();
5127 for_each_node(node
) {
5128 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
5129 "%s%d:%d", delim
, node
,
5130 unbound_pwq_by_node(wq
, node
)->pool
->id
);
5133 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
5134 rcu_read_unlock_sched();
5139 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
5142 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5145 mutex_lock(&wq
->mutex
);
5146 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
5147 mutex_unlock(&wq
->mutex
);
5152 /* prepare workqueue_attrs for sysfs store operations */
5153 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
5155 struct workqueue_attrs
*attrs
;
5157 lockdep_assert_held(&wq_pool_mutex
);
5159 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
5163 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
5167 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
5168 const char *buf
, size_t count
)
5170 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5171 struct workqueue_attrs
*attrs
;
5174 apply_wqattrs_lock();
5176 attrs
= wq_sysfs_prep_attrs(wq
);
5180 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5181 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5182 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5187 apply_wqattrs_unlock();
5188 free_workqueue_attrs(attrs
);
5189 return ret
?: count
;
5192 static ssize_t
wq_cpumask_show(struct device
*dev
,
5193 struct device_attribute
*attr
, char *buf
)
5195 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5198 mutex_lock(&wq
->mutex
);
5199 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5200 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5201 mutex_unlock(&wq
->mutex
);
5205 static ssize_t
wq_cpumask_store(struct device
*dev
,
5206 struct device_attribute
*attr
,
5207 const char *buf
, size_t count
)
5209 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5210 struct workqueue_attrs
*attrs
;
5213 apply_wqattrs_lock();
5215 attrs
= wq_sysfs_prep_attrs(wq
);
5219 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5221 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5224 apply_wqattrs_unlock();
5225 free_workqueue_attrs(attrs
);
5226 return ret
?: count
;
5229 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5232 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5235 mutex_lock(&wq
->mutex
);
5236 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5237 !wq
->unbound_attrs
->no_numa
);
5238 mutex_unlock(&wq
->mutex
);
5243 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5244 const char *buf
, size_t count
)
5246 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5247 struct workqueue_attrs
*attrs
;
5248 int v
, ret
= -ENOMEM
;
5250 apply_wqattrs_lock();
5252 attrs
= wq_sysfs_prep_attrs(wq
);
5257 if (sscanf(buf
, "%d", &v
) == 1) {
5258 attrs
->no_numa
= !v
;
5259 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5263 apply_wqattrs_unlock();
5264 free_workqueue_attrs(attrs
);
5265 return ret
?: count
;
5268 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5269 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5270 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5271 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5272 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5276 static struct bus_type wq_subsys
= {
5277 .name
= "workqueue",
5278 .dev_groups
= wq_sysfs_groups
,
5281 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5282 struct device_attribute
*attr
, char *buf
)
5286 mutex_lock(&wq_pool_mutex
);
5287 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5288 cpumask_pr_args(wq_unbound_cpumask
));
5289 mutex_unlock(&wq_pool_mutex
);
5294 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5295 struct device_attribute
*attr
, const char *buf
, size_t count
)
5297 cpumask_var_t cpumask
;
5300 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5303 ret
= cpumask_parse(buf
, cpumask
);
5305 ret
= workqueue_set_unbound_cpumask(cpumask
);
5307 free_cpumask_var(cpumask
);
5308 return ret
? ret
: count
;
5311 static struct device_attribute wq_sysfs_cpumask_attr
=
5312 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5313 wq_unbound_cpumask_store
);
5315 static int __init
wq_sysfs_init(void)
5319 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5323 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5325 core_initcall(wq_sysfs_init
);
5327 static void wq_device_release(struct device
*dev
)
5329 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5335 * workqueue_sysfs_register - make a workqueue visible in sysfs
5336 * @wq: the workqueue to register
5338 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5339 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5340 * which is the preferred method.
5342 * Workqueue user should use this function directly iff it wants to apply
5343 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5344 * apply_workqueue_attrs() may race against userland updating the
5347 * Return: 0 on success, -errno on failure.
5349 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5351 struct wq_device
*wq_dev
;
5355 * Adjusting max_active or creating new pwqs by applying
5356 * attributes breaks ordering guarantee. Disallow exposing ordered
5359 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
5362 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5367 wq_dev
->dev
.bus
= &wq_subsys
;
5368 wq_dev
->dev
.release
= wq_device_release
;
5369 dev_set_name(&wq_dev
->dev
, "%s", wq
->name
);
5372 * unbound_attrs are created separately. Suppress uevent until
5373 * everything is ready.
5375 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5377 ret
= device_register(&wq_dev
->dev
);
5379 put_device(&wq_dev
->dev
);
5384 if (wq
->flags
& WQ_UNBOUND
) {
5385 struct device_attribute
*attr
;
5387 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5388 ret
= device_create_file(&wq_dev
->dev
, attr
);
5390 device_unregister(&wq_dev
->dev
);
5397 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5398 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5403 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5404 * @wq: the workqueue to unregister
5406 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5408 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5410 struct wq_device
*wq_dev
= wq
->wq_dev
;
5416 device_unregister(&wq_dev
->dev
);
5418 #else /* CONFIG_SYSFS */
5419 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5420 #endif /* CONFIG_SYSFS */
5423 * Workqueue watchdog.
5425 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5426 * flush dependency, a concurrency managed work item which stays RUNNING
5427 * indefinitely. Workqueue stalls can be very difficult to debug as the
5428 * usual warning mechanisms don't trigger and internal workqueue state is
5431 * Workqueue watchdog monitors all worker pools periodically and dumps
5432 * state if some pools failed to make forward progress for a while where
5433 * forward progress is defined as the first item on ->worklist changing.
5435 * This mechanism is controlled through the kernel parameter
5436 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5437 * corresponding sysfs parameter file.
5439 #ifdef CONFIG_WQ_WATCHDOG
5441 static void wq_watchdog_timer_fn(unsigned long data
);
5443 static unsigned long wq_watchdog_thresh
= 30;
5444 static struct timer_list wq_watchdog_timer
=
5445 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn
, 0, 0);
5447 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5448 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5450 static void wq_watchdog_reset_touched(void)
5454 wq_watchdog_touched
= jiffies
;
5455 for_each_possible_cpu(cpu
)
5456 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5459 static void wq_watchdog_timer_fn(unsigned long data
)
5461 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5462 bool lockup_detected
= false;
5463 struct worker_pool
*pool
;
5471 for_each_pool(pool
, pi
) {
5472 unsigned long pool_ts
, touched
, ts
;
5474 if (list_empty(&pool
->worklist
))
5477 /* get the latest of pool and touched timestamps */
5478 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5479 touched
= READ_ONCE(wq_watchdog_touched
);
5481 if (time_after(pool_ts
, touched
))
5486 if (pool
->cpu
>= 0) {
5487 unsigned long cpu_touched
=
5488 READ_ONCE(per_cpu(wq_watchdog_touched_cpu
,
5490 if (time_after(cpu_touched
, ts
))
5495 if (time_after(jiffies
, ts
+ thresh
)) {
5496 lockup_detected
= true;
5497 pr_emerg("BUG: workqueue lockup - pool");
5498 pr_cont_pool_info(pool
);
5499 pr_cont(" stuck for %us!\n",
5500 jiffies_to_msecs(jiffies
- pool_ts
) / 1000);
5506 if (lockup_detected
)
5507 show_workqueue_state();
5509 wq_watchdog_reset_touched();
5510 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5513 notrace
void wq_watchdog_touch(int cpu
)
5516 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5518 wq_watchdog_touched
= jiffies
;
5521 static void wq_watchdog_set_thresh(unsigned long thresh
)
5523 wq_watchdog_thresh
= 0;
5524 del_timer_sync(&wq_watchdog_timer
);
5527 wq_watchdog_thresh
= thresh
;
5528 wq_watchdog_reset_touched();
5529 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5533 static int wq_watchdog_param_set_thresh(const char *val
,
5534 const struct kernel_param
*kp
)
5536 unsigned long thresh
;
5539 ret
= kstrtoul(val
, 0, &thresh
);
5544 wq_watchdog_set_thresh(thresh
);
5546 wq_watchdog_thresh
= thresh
;
5551 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5552 .set
= wq_watchdog_param_set_thresh
,
5553 .get
= param_get_ulong
,
5556 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5559 static void wq_watchdog_init(void)
5561 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5564 #else /* CONFIG_WQ_WATCHDOG */
5566 static inline void wq_watchdog_init(void) { }
5568 #endif /* CONFIG_WQ_WATCHDOG */
5570 static void __init
wq_numa_init(void)
5575 if (num_possible_nodes() <= 1)
5578 if (wq_disable_numa
) {
5579 pr_info("workqueue: NUMA affinity support disabled\n");
5583 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
5584 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5587 * We want masks of possible CPUs of each node which isn't readily
5588 * available. Build one from cpu_to_node() which should have been
5589 * fully initialized by now.
5591 tbl
= kzalloc(nr_node_ids
* sizeof(tbl
[0]), GFP_KERNEL
);
5595 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5596 node_online(node
) ? node
: NUMA_NO_NODE
));
5598 for_each_possible_cpu(cpu
) {
5599 node
= cpu_to_node(cpu
);
5600 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5601 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5602 /* happens iff arch is bonkers, let's just proceed */
5605 cpumask_set_cpu(cpu
, tbl
[node
]);
5608 wq_numa_possible_cpumask
= tbl
;
5609 wq_numa_enabled
= true;
5613 * workqueue_init_early - early init for workqueue subsystem
5615 * This is the first half of two-staged workqueue subsystem initialization
5616 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5617 * idr are up. It sets up all the data structures and system workqueues
5618 * and allows early boot code to create workqueues and queue/cancel work
5619 * items. Actual work item execution starts only after kthreads can be
5620 * created and scheduled right before early initcalls.
5622 int __init
workqueue_init_early(void)
5624 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5627 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5629 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5630 cpumask_copy(wq_unbound_cpumask
, cpu_possible_mask
);
5632 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5634 /* initialize CPU pools */
5635 for_each_possible_cpu(cpu
) {
5636 struct worker_pool
*pool
;
5639 for_each_cpu_worker_pool(pool
, cpu
) {
5640 BUG_ON(init_worker_pool(pool
));
5642 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5643 pool
->attrs
->nice
= std_nice
[i
++];
5644 pool
->node
= cpu_to_node(cpu
);
5647 mutex_lock(&wq_pool_mutex
);
5648 BUG_ON(worker_pool_assign_id(pool
));
5649 mutex_unlock(&wq_pool_mutex
);
5653 /* create default unbound and ordered wq attrs */
5654 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5655 struct workqueue_attrs
*attrs
;
5657 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5658 attrs
->nice
= std_nice
[i
];
5659 unbound_std_wq_attrs
[i
] = attrs
;
5662 * An ordered wq should have only one pwq as ordering is
5663 * guaranteed by max_active which is enforced by pwqs.
5664 * Turn off NUMA so that dfl_pwq is used for all nodes.
5666 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5667 attrs
->nice
= std_nice
[i
];
5668 attrs
->no_numa
= true;
5669 ordered_wq_attrs
[i
] = attrs
;
5672 system_wq
= alloc_workqueue("events", 0, 0);
5673 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5674 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5675 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5676 WQ_UNBOUND_MAX_ACTIVE
);
5677 system_freezable_wq
= alloc_workqueue("events_freezable",
5679 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5680 WQ_POWER_EFFICIENT
, 0);
5681 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5682 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5684 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5685 !system_unbound_wq
|| !system_freezable_wq
||
5686 !system_power_efficient_wq
||
5687 !system_freezable_power_efficient_wq
);
5693 * workqueue_init - bring workqueue subsystem fully online
5695 * This is the latter half of two-staged workqueue subsystem initialization
5696 * and invoked as soon as kthreads can be created and scheduled.
5697 * Workqueues have been created and work items queued on them, but there
5698 * are no kworkers executing the work items yet. Populate the worker pools
5699 * with the initial workers and enable future kworker creations.
5701 int __init
workqueue_init(void)
5703 struct workqueue_struct
*wq
;
5704 struct worker_pool
*pool
;
5708 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5709 * CPU to node mapping may not be available that early on some
5710 * archs such as power and arm64. As per-cpu pools created
5711 * previously could be missing node hint and unbound pools NUMA
5712 * affinity, fix them up.
5716 mutex_lock(&wq_pool_mutex
);
5718 for_each_possible_cpu(cpu
) {
5719 for_each_cpu_worker_pool(pool
, cpu
) {
5720 pool
->node
= cpu_to_node(cpu
);
5724 list_for_each_entry(wq
, &workqueues
, list
)
5725 wq_update_unbound_numa(wq
, smp_processor_id(), true);
5727 mutex_unlock(&wq_pool_mutex
);
5729 /* create the initial workers */
5730 for_each_online_cpu(cpu
) {
5731 for_each_cpu_worker_pool(pool
, cpu
) {
5732 pool
->flags
&= ~POOL_DISASSOCIATED
;
5733 BUG_ON(!create_worker(pool
));
5737 hash_for_each(unbound_pool_hash
, bkt
, pool
, hash_node
)
5738 BUG_ON(!create_worker(pool
));