1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
53 #include <linux/kvm_para.h>
55 #include "workqueue_internal.h"
61 * A bound pool is either associated or disassociated with its CPU.
62 * While associated (!DISASSOCIATED), all workers are bound to the
63 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 * While DISASSOCIATED, the cpu may be offline and all workers have
67 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 * be executing on any CPU. The pool behaves as an unbound one.
70 * Note that DISASSOCIATED should be flipped only while holding
71 * wq_pool_attach_mutex to avoid changing binding state while
72 * worker_attach_to_pool() is in progress.
74 POOL_MANAGER_ACTIVE
= 1 << 0, /* being managed */
75 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
78 WORKER_DIE
= 1 << 1, /* die die die */
79 WORKER_IDLE
= 1 << 2, /* is idle */
80 WORKER_PREP
= 1 << 3, /* preparing to run works */
81 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
82 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
83 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
85 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
86 WORKER_UNBOUND
| WORKER_REBOUND
,
88 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
90 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
91 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
93 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
94 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
96 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
97 /* call for help after 10ms
99 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
100 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
103 * Rescue workers are used only on emergencies and shared by
104 * all cpus. Give MIN_NICE.
106 RESCUER_NICE_LEVEL
= MIN_NICE
,
107 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
113 * Structure fields follow one of the following exclusion rules.
115 * I: Modifiable by initialization/destruction paths and read-only for
118 * P: Preemption protected. Disabling preemption is enough and should
119 * only be modified and accessed from the local cpu.
121 * L: pool->lock protected. Access with pool->lock held.
123 * X: During normal operation, modification requires pool->lock and should
124 * be done only from local cpu. Either disabling preemption on local
125 * cpu or grabbing pool->lock is enough for read access. If
126 * POOL_DISASSOCIATED is set, it's identical to L.
128 * A: wq_pool_attach_mutex protected.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
134 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
136 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
139 * WQ: wq->mutex protected.
141 * WR: wq->mutex protected for writes. RCU protected for reads.
143 * MD: wq_mayday_lock protected.
146 /* struct worker is defined in workqueue_internal.h */
149 raw_spinlock_t lock
; /* the pool lock */
150 int cpu
; /* I: the associated cpu */
151 int node
; /* I: the associated node ID */
152 int id
; /* I: pool ID */
153 unsigned int flags
; /* X: flags */
155 unsigned long watchdog_ts
; /* L: watchdog timestamp */
157 struct list_head worklist
; /* L: list of pending works */
159 int nr_workers
; /* L: total number of workers */
160 int nr_idle
; /* L: currently idle workers */
162 struct list_head idle_list
; /* X: list of idle workers */
163 struct timer_list idle_timer
; /* L: worker idle timeout */
164 struct timer_list mayday_timer
; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
168 /* L: hash of busy workers */
170 struct worker
*manager
; /* L: purely informational */
171 struct list_head workers
; /* A: attached workers */
172 struct completion
*detach_completion
; /* all workers detached */
174 struct ida worker_ida
; /* worker IDs for task name */
176 struct workqueue_attrs
*attrs
; /* I: worker attributes */
177 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
178 int refcnt
; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp
;
188 * Destruction of pool is RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp
;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue
{
201 struct worker_pool
*pool
; /* I: the associated pool */
202 struct workqueue_struct
*wq
; /* I: the owning workqueue */
203 int work_color
; /* L: current color */
204 int flush_color
; /* L: flushing color */
205 int refcnt
; /* L: reference count */
206 int nr_in_flight
[WORK_NR_COLORS
];
207 /* L: nr of in_flight works */
210 * nr_active management and WORK_STRUCT_INACTIVE:
212 * When pwq->nr_active >= max_active, new work item is queued to
213 * pwq->inactive_works instead of pool->worklist and marked with
214 * WORK_STRUCT_INACTIVE.
216 * All work items marked with WORK_STRUCT_INACTIVE do not participate
217 * in pwq->nr_active and all work items in pwq->inactive_works are
218 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
219 * work items are in pwq->inactive_works. Some of them are ready to
220 * run in pool->worklist or worker->scheduled. Those work itmes are
221 * only struct wq_barrier which is used for flush_work() and should
222 * not participate in pwq->nr_active. For non-barrier work item, it
223 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
225 int nr_active
; /* L: nr of active works */
226 int max_active
; /* L: max active works */
227 struct list_head inactive_works
; /* L: inactive works */
228 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
229 struct list_head mayday_node
; /* MD: node on wq->maydays */
232 * Release of unbound pwq is punted to system_wq. See put_pwq()
233 * and pwq_unbound_release_workfn() for details. pool_workqueue
234 * itself is also RCU protected so that the first pwq can be
235 * determined without grabbing wq->mutex.
237 struct work_struct unbound_release_work
;
239 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
242 * Structure used to wait for workqueue flush.
245 struct list_head list
; /* WQ: list of flushers */
246 int flush_color
; /* WQ: flush color waiting for */
247 struct completion done
; /* flush completion */
253 * The externally visible workqueue. It relays the issued work items to
254 * the appropriate worker_pool through its pool_workqueues.
256 struct workqueue_struct
{
257 struct list_head pwqs
; /* WR: all pwqs of this wq */
258 struct list_head list
; /* PR: list of all workqueues */
260 struct mutex mutex
; /* protects this wq */
261 int work_color
; /* WQ: current work color */
262 int flush_color
; /* WQ: current flush color */
263 atomic_t nr_pwqs_to_flush
; /* flush in progress */
264 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
265 struct list_head flusher_queue
; /* WQ: flush waiters */
266 struct list_head flusher_overflow
; /* WQ: flush overflow list */
268 struct list_head maydays
; /* MD: pwqs requesting rescue */
269 struct worker
*rescuer
; /* MD: rescue worker */
271 int nr_drainers
; /* WQ: drain in progress */
272 int saved_max_active
; /* WQ: saved pwq max_active */
274 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
275 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
278 struct wq_device
*wq_dev
; /* I: for sysfs interface */
280 #ifdef CONFIG_LOCKDEP
282 struct lock_class_key key
;
283 struct lockdep_map lockdep_map
;
285 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
288 * Destruction of workqueue_struct is RCU protected to allow walking
289 * the workqueues list without grabbing wq_pool_mutex.
290 * This is used to dump all workqueues from sysrq.
294 /* hot fields used during command issue, aligned to cacheline */
295 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
296 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
297 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
300 static struct kmem_cache
*pwq_cache
;
302 static cpumask_var_t
*wq_numa_possible_cpumask
;
303 /* possible CPUs of each node */
305 static bool wq_disable_numa
;
306 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
308 /* see the comment above the definition of WQ_POWER_EFFICIENT */
309 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
310 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
312 static bool wq_online
; /* can kworkers be created yet? */
314 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
316 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
317 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
319 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
320 static DEFINE_MUTEX(wq_pool_attach_mutex
); /* protects worker attach/detach */
321 static DEFINE_RAW_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
322 /* wait for manager to go away */
323 static struct rcuwait manager_wait
= __RCUWAIT_INITIALIZER(manager_wait
);
325 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
326 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
328 /* PL: allowable cpus for unbound wqs and work items */
329 static cpumask_var_t wq_unbound_cpumask
;
331 /* CPU where unbound work was last round robin scheduled from this CPU */
332 static DEFINE_PER_CPU(int, wq_rr_cpu_last
);
335 * Local execution of unbound work items is no longer guaranteed. The
336 * following always forces round-robin CPU selection on unbound work items
337 * to uncover usages which depend on it.
339 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
340 static bool wq_debug_force_rr_cpu
= true;
342 static bool wq_debug_force_rr_cpu
= false;
344 module_param_named(debug_force_rr_cpu
, wq_debug_force_rr_cpu
, bool, 0644);
346 /* the per-cpu worker pools */
347 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
], cpu_worker_pools
);
349 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
351 /* PL: hash of all unbound pools keyed by pool->attrs */
352 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
354 /* I: attributes used when instantiating standard unbound pools on demand */
355 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
357 /* I: attributes used when instantiating ordered pools on demand */
358 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
360 struct workqueue_struct
*system_wq __read_mostly
;
361 EXPORT_SYMBOL(system_wq
);
362 struct workqueue_struct
*system_highpri_wq __read_mostly
;
363 EXPORT_SYMBOL_GPL(system_highpri_wq
);
364 struct workqueue_struct
*system_long_wq __read_mostly
;
365 EXPORT_SYMBOL_GPL(system_long_wq
);
366 struct workqueue_struct
*system_unbound_wq __read_mostly
;
367 EXPORT_SYMBOL_GPL(system_unbound_wq
);
368 struct workqueue_struct
*system_freezable_wq __read_mostly
;
369 EXPORT_SYMBOL_GPL(system_freezable_wq
);
370 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
371 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
372 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
373 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
375 static int worker_thread(void *__worker
);
376 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
377 static void show_pwq(struct pool_workqueue
*pwq
);
378 static void show_one_worker_pool(struct worker_pool
*pool
);
380 #define CREATE_TRACE_POINTS
381 #include <trace/events/workqueue.h>
383 #define assert_rcu_or_pool_mutex() \
384 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
385 !lockdep_is_held(&wq_pool_mutex), \
386 "RCU or wq_pool_mutex should be held")
388 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
389 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
390 !lockdep_is_held(&wq->mutex) && \
391 !lockdep_is_held(&wq_pool_mutex), \
392 "RCU, wq->mutex or wq_pool_mutex should be held")
394 #define for_each_cpu_worker_pool(pool, cpu) \
395 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
396 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
400 * for_each_pool - iterate through all worker_pools in the system
401 * @pool: iteration cursor
402 * @pi: integer used for iteration
404 * This must be called either with wq_pool_mutex held or RCU read
405 * locked. If the pool needs to be used beyond the locking in effect, the
406 * caller is responsible for guaranteeing that the pool stays online.
408 * The if/else clause exists only for the lockdep assertion and can be
411 #define for_each_pool(pool, pi) \
412 idr_for_each_entry(&worker_pool_idr, pool, pi) \
413 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
417 * for_each_pool_worker - iterate through all workers of a worker_pool
418 * @worker: iteration cursor
419 * @pool: worker_pool to iterate workers of
421 * This must be called with wq_pool_attach_mutex.
423 * The if/else clause exists only for the lockdep assertion and can be
426 #define for_each_pool_worker(worker, pool) \
427 list_for_each_entry((worker), &(pool)->workers, node) \
428 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
432 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
433 * @pwq: iteration cursor
434 * @wq: the target workqueue
436 * This must be called either with wq->mutex held or RCU read locked.
437 * If the pwq needs to be used beyond the locking in effect, the caller is
438 * responsible for guaranteeing that the pwq stays online.
440 * The if/else clause exists only for the lockdep assertion and can be
443 #define for_each_pwq(pwq, wq) \
444 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
445 lockdep_is_held(&(wq->mutex)))
447 #ifdef CONFIG_DEBUG_OBJECTS_WORK
449 static const struct debug_obj_descr work_debug_descr
;
451 static void *work_debug_hint(void *addr
)
453 return ((struct work_struct
*) addr
)->func
;
456 static bool work_is_static_object(void *addr
)
458 struct work_struct
*work
= addr
;
460 return test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
));
464 * fixup_init is called when:
465 * - an active object is initialized
467 static bool work_fixup_init(void *addr
, enum debug_obj_state state
)
469 struct work_struct
*work
= addr
;
472 case ODEBUG_STATE_ACTIVE
:
473 cancel_work_sync(work
);
474 debug_object_init(work
, &work_debug_descr
);
482 * fixup_free is called when:
483 * - an active object is freed
485 static bool work_fixup_free(void *addr
, enum debug_obj_state state
)
487 struct work_struct
*work
= addr
;
490 case ODEBUG_STATE_ACTIVE
:
491 cancel_work_sync(work
);
492 debug_object_free(work
, &work_debug_descr
);
499 static const struct debug_obj_descr work_debug_descr
= {
500 .name
= "work_struct",
501 .debug_hint
= work_debug_hint
,
502 .is_static_object
= work_is_static_object
,
503 .fixup_init
= work_fixup_init
,
504 .fixup_free
= work_fixup_free
,
507 static inline void debug_work_activate(struct work_struct
*work
)
509 debug_object_activate(work
, &work_debug_descr
);
512 static inline void debug_work_deactivate(struct work_struct
*work
)
514 debug_object_deactivate(work
, &work_debug_descr
);
517 void __init_work(struct work_struct
*work
, int onstack
)
520 debug_object_init_on_stack(work
, &work_debug_descr
);
522 debug_object_init(work
, &work_debug_descr
);
524 EXPORT_SYMBOL_GPL(__init_work
);
526 void destroy_work_on_stack(struct work_struct
*work
)
528 debug_object_free(work
, &work_debug_descr
);
530 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
532 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
534 destroy_timer_on_stack(&work
->timer
);
535 debug_object_free(&work
->work
, &work_debug_descr
);
537 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
540 static inline void debug_work_activate(struct work_struct
*work
) { }
541 static inline void debug_work_deactivate(struct work_struct
*work
) { }
545 * worker_pool_assign_id - allocate ID and assign it to @pool
546 * @pool: the pool pointer of interest
548 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
549 * successfully, -errno on failure.
551 static int worker_pool_assign_id(struct worker_pool
*pool
)
555 lockdep_assert_held(&wq_pool_mutex
);
557 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
567 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
568 * @wq: the target workqueue
571 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
573 * If the pwq needs to be used beyond the locking in effect, the caller is
574 * responsible for guaranteeing that the pwq stays online.
576 * Return: The unbound pool_workqueue for @node.
578 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
581 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
584 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
585 * delayed item is pending. The plan is to keep CPU -> NODE
586 * mapping valid and stable across CPU on/offlines. Once that
587 * happens, this workaround can be removed.
589 if (unlikely(node
== NUMA_NO_NODE
))
592 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
595 static unsigned int work_color_to_flags(int color
)
597 return color
<< WORK_STRUCT_COLOR_SHIFT
;
600 static int get_work_color(unsigned long work_data
)
602 return (work_data
>> WORK_STRUCT_COLOR_SHIFT
) &
603 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
606 static int work_next_color(int color
)
608 return (color
+ 1) % WORK_NR_COLORS
;
612 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
613 * contain the pointer to the queued pwq. Once execution starts, the flag
614 * is cleared and the high bits contain OFFQ flags and pool ID.
616 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
617 * and clear_work_data() can be used to set the pwq, pool or clear
618 * work->data. These functions should only be called while the work is
619 * owned - ie. while the PENDING bit is set.
621 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
622 * corresponding to a work. Pool is available once the work has been
623 * queued anywhere after initialization until it is sync canceled. pwq is
624 * available only while the work item is queued.
626 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
627 * canceled. While being canceled, a work item may have its PENDING set
628 * but stay off timer and worklist for arbitrarily long and nobody should
629 * try to steal the PENDING bit.
631 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
634 WARN_ON_ONCE(!work_pending(work
));
635 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
638 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
639 unsigned long extra_flags
)
641 set_work_data(work
, (unsigned long)pwq
,
642 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
645 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
648 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
649 WORK_STRUCT_PENDING
);
652 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
656 * The following wmb is paired with the implied mb in
657 * test_and_set_bit(PENDING) and ensures all updates to @work made
658 * here are visible to and precede any updates by the next PENDING
662 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
664 * The following mb guarantees that previous clear of a PENDING bit
665 * will not be reordered with any speculative LOADS or STORES from
666 * work->current_func, which is executed afterwards. This possible
667 * reordering can lead to a missed execution on attempt to queue
668 * the same @work. E.g. consider this case:
671 * ---------------------------- --------------------------------
673 * 1 STORE event_indicated
674 * 2 queue_work_on() {
675 * 3 test_and_set_bit(PENDING)
676 * 4 } set_..._and_clear_pending() {
677 * 5 set_work_data() # clear bit
679 * 7 work->current_func() {
680 * 8 LOAD event_indicated
683 * Without an explicit full barrier speculative LOAD on line 8 can
684 * be executed before CPU#0 does STORE on line 1. If that happens,
685 * CPU#0 observes the PENDING bit is still set and new execution of
686 * a @work is not queued in a hope, that CPU#1 will eventually
687 * finish the queued @work. Meanwhile CPU#1 does not see
688 * event_indicated is set, because speculative LOAD was executed
689 * before actual STORE.
694 static void clear_work_data(struct work_struct
*work
)
696 smp_wmb(); /* see set_work_pool_and_clear_pending() */
697 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
700 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
702 unsigned long data
= atomic_long_read(&work
->data
);
704 if (data
& WORK_STRUCT_PWQ
)
705 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
711 * get_work_pool - return the worker_pool a given work was associated with
712 * @work: the work item of interest
714 * Pools are created and destroyed under wq_pool_mutex, and allows read
715 * access under RCU read lock. As such, this function should be
716 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
718 * All fields of the returned pool are accessible as long as the above
719 * mentioned locking is in effect. If the returned pool needs to be used
720 * beyond the critical section, the caller is responsible for ensuring the
721 * returned pool is and stays online.
723 * Return: The worker_pool @work was last associated with. %NULL if none.
725 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
727 unsigned long data
= atomic_long_read(&work
->data
);
730 assert_rcu_or_pool_mutex();
732 if (data
& WORK_STRUCT_PWQ
)
733 return ((struct pool_workqueue
*)
734 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
736 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
737 if (pool_id
== WORK_OFFQ_POOL_NONE
)
740 return idr_find(&worker_pool_idr
, pool_id
);
744 * get_work_pool_id - return the worker pool ID a given work is associated with
745 * @work: the work item of interest
747 * Return: The worker_pool ID @work was last associated with.
748 * %WORK_OFFQ_POOL_NONE if none.
750 static int get_work_pool_id(struct work_struct
*work
)
752 unsigned long data
= atomic_long_read(&work
->data
);
754 if (data
& WORK_STRUCT_PWQ
)
755 return ((struct pool_workqueue
*)
756 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
758 return data
>> WORK_OFFQ_POOL_SHIFT
;
761 static void mark_work_canceling(struct work_struct
*work
)
763 unsigned long pool_id
= get_work_pool_id(work
);
765 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
766 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
769 static bool work_is_canceling(struct work_struct
*work
)
771 unsigned long data
= atomic_long_read(&work
->data
);
773 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
777 * Policy functions. These define the policies on how the global worker
778 * pools are managed. Unless noted otherwise, these functions assume that
779 * they're being called with pool->lock held.
782 static bool __need_more_worker(struct worker_pool
*pool
)
784 return !atomic_read(&pool
->nr_running
);
788 * Need to wake up a worker? Called from anything but currently
791 * Note that, because unbound workers never contribute to nr_running, this
792 * function will always return %true for unbound pools as long as the
793 * worklist isn't empty.
795 static bool need_more_worker(struct worker_pool
*pool
)
797 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
800 /* Can I start working? Called from busy but !running workers. */
801 static bool may_start_working(struct worker_pool
*pool
)
803 return pool
->nr_idle
;
806 /* Do I need to keep working? Called from currently running workers. */
807 static bool keep_working(struct worker_pool
*pool
)
809 return !list_empty(&pool
->worklist
) &&
810 atomic_read(&pool
->nr_running
) <= 1;
813 /* Do we need a new worker? Called from manager. */
814 static bool need_to_create_worker(struct worker_pool
*pool
)
816 return need_more_worker(pool
) && !may_start_working(pool
);
819 /* Do we have too many workers and should some go away? */
820 static bool too_many_workers(struct worker_pool
*pool
)
822 bool managing
= pool
->flags
& POOL_MANAGER_ACTIVE
;
823 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
824 int nr_busy
= pool
->nr_workers
- nr_idle
;
826 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
833 /* Return the first idle worker. Safe with preemption disabled */
834 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
836 if (unlikely(list_empty(&pool
->idle_list
)))
839 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
843 * wake_up_worker - wake up an idle worker
844 * @pool: worker pool to wake worker from
846 * Wake up the first idle worker of @pool.
849 * raw_spin_lock_irq(pool->lock).
851 static void wake_up_worker(struct worker_pool
*pool
)
853 struct worker
*worker
= first_idle_worker(pool
);
856 wake_up_process(worker
->task
);
860 * wq_worker_running - a worker is running again
861 * @task: task waking up
863 * This function is called when a worker returns from schedule()
865 void wq_worker_running(struct task_struct
*task
)
867 struct worker
*worker
= kthread_data(task
);
869 if (!worker
->sleeping
)
871 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
872 atomic_inc(&worker
->pool
->nr_running
);
873 worker
->sleeping
= 0;
877 * wq_worker_sleeping - a worker is going to sleep
878 * @task: task going to sleep
880 * This function is called from schedule() when a busy worker is
881 * going to sleep. Preemption needs to be disabled to protect ->sleeping
884 void wq_worker_sleeping(struct task_struct
*task
)
886 struct worker
*next
, *worker
= kthread_data(task
);
887 struct worker_pool
*pool
;
890 * Rescuers, which may not have all the fields set up like normal
891 * workers, also reach here, let's not access anything before
892 * checking NOT_RUNNING.
894 if (worker
->flags
& WORKER_NOT_RUNNING
)
899 /* Return if preempted before wq_worker_running() was reached */
900 if (worker
->sleeping
)
903 worker
->sleeping
= 1;
904 raw_spin_lock_irq(&pool
->lock
);
907 * The counterpart of the following dec_and_test, implied mb,
908 * worklist not empty test sequence is in insert_work().
909 * Please read comment there.
911 * NOT_RUNNING is clear. This means that we're bound to and
912 * running on the local cpu w/ rq lock held and preemption
913 * disabled, which in turn means that none else could be
914 * manipulating idle_list, so dereferencing idle_list without pool
917 if (atomic_dec_and_test(&pool
->nr_running
) &&
918 !list_empty(&pool
->worklist
)) {
919 next
= first_idle_worker(pool
);
921 wake_up_process(next
->task
);
923 raw_spin_unlock_irq(&pool
->lock
);
927 * wq_worker_last_func - retrieve worker's last work function
928 * @task: Task to retrieve last work function of.
930 * Determine the last function a worker executed. This is called from
931 * the scheduler to get a worker's last known identity.
934 * raw_spin_lock_irq(rq->lock)
936 * This function is called during schedule() when a kworker is going
937 * to sleep. It's used by psi to identify aggregation workers during
938 * dequeuing, to allow periodic aggregation to shut-off when that
939 * worker is the last task in the system or cgroup to go to sleep.
941 * As this function doesn't involve any workqueue-related locking, it
942 * only returns stable values when called from inside the scheduler's
943 * queuing and dequeuing paths, when @task, which must be a kworker,
944 * is guaranteed to not be processing any works.
947 * The last work function %current executed as a worker, NULL if it
948 * hasn't executed any work yet.
950 work_func_t
wq_worker_last_func(struct task_struct
*task
)
952 struct worker
*worker
= kthread_data(task
);
954 return worker
->last_func
;
958 * worker_set_flags - set worker flags and adjust nr_running accordingly
960 * @flags: flags to set
962 * Set @flags in @worker->flags and adjust nr_running accordingly.
965 * raw_spin_lock_irq(pool->lock)
967 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
969 struct worker_pool
*pool
= worker
->pool
;
971 WARN_ON_ONCE(worker
->task
!= current
);
973 /* If transitioning into NOT_RUNNING, adjust nr_running. */
974 if ((flags
& WORKER_NOT_RUNNING
) &&
975 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
976 atomic_dec(&pool
->nr_running
);
979 worker
->flags
|= flags
;
983 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
985 * @flags: flags to clear
987 * Clear @flags in @worker->flags and adjust nr_running accordingly.
990 * raw_spin_lock_irq(pool->lock)
992 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
994 struct worker_pool
*pool
= worker
->pool
;
995 unsigned int oflags
= worker
->flags
;
997 WARN_ON_ONCE(worker
->task
!= current
);
999 worker
->flags
&= ~flags
;
1002 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1003 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1004 * of multiple flags, not a single flag.
1006 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
1007 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
1008 atomic_inc(&pool
->nr_running
);
1012 * find_worker_executing_work - find worker which is executing a work
1013 * @pool: pool of interest
1014 * @work: work to find worker for
1016 * Find a worker which is executing @work on @pool by searching
1017 * @pool->busy_hash which is keyed by the address of @work. For a worker
1018 * to match, its current execution should match the address of @work and
1019 * its work function. This is to avoid unwanted dependency between
1020 * unrelated work executions through a work item being recycled while still
1023 * This is a bit tricky. A work item may be freed once its execution
1024 * starts and nothing prevents the freed area from being recycled for
1025 * another work item. If the same work item address ends up being reused
1026 * before the original execution finishes, workqueue will identify the
1027 * recycled work item as currently executing and make it wait until the
1028 * current execution finishes, introducing an unwanted dependency.
1030 * This function checks the work item address and work function to avoid
1031 * false positives. Note that this isn't complete as one may construct a
1032 * work function which can introduce dependency onto itself through a
1033 * recycled work item. Well, if somebody wants to shoot oneself in the
1034 * foot that badly, there's only so much we can do, and if such deadlock
1035 * actually occurs, it should be easy to locate the culprit work function.
1038 * raw_spin_lock_irq(pool->lock).
1041 * Pointer to worker which is executing @work if found, %NULL
1044 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
1045 struct work_struct
*work
)
1047 struct worker
*worker
;
1049 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1050 (unsigned long)work
)
1051 if (worker
->current_work
== work
&&
1052 worker
->current_func
== work
->func
)
1059 * move_linked_works - move linked works to a list
1060 * @work: start of series of works to be scheduled
1061 * @head: target list to append @work to
1062 * @nextp: out parameter for nested worklist walking
1064 * Schedule linked works starting from @work to @head. Work series to
1065 * be scheduled starts at @work and includes any consecutive work with
1066 * WORK_STRUCT_LINKED set in its predecessor.
1068 * If @nextp is not NULL, it's updated to point to the next work of
1069 * the last scheduled work. This allows move_linked_works() to be
1070 * nested inside outer list_for_each_entry_safe().
1073 * raw_spin_lock_irq(pool->lock).
1075 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1076 struct work_struct
**nextp
)
1078 struct work_struct
*n
;
1081 * Linked worklist will always end before the end of the list,
1082 * use NULL for list head.
1084 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1085 list_move_tail(&work
->entry
, head
);
1086 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1091 * If we're already inside safe list traversal and have moved
1092 * multiple works to the scheduled queue, the next position
1093 * needs to be updated.
1100 * get_pwq - get an extra reference on the specified pool_workqueue
1101 * @pwq: pool_workqueue to get
1103 * Obtain an extra reference on @pwq. The caller should guarantee that
1104 * @pwq has positive refcnt and be holding the matching pool->lock.
1106 static void get_pwq(struct pool_workqueue
*pwq
)
1108 lockdep_assert_held(&pwq
->pool
->lock
);
1109 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1114 * put_pwq - put a pool_workqueue reference
1115 * @pwq: pool_workqueue to put
1117 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1118 * destruction. The caller should be holding the matching pool->lock.
1120 static void put_pwq(struct pool_workqueue
*pwq
)
1122 lockdep_assert_held(&pwq
->pool
->lock
);
1123 if (likely(--pwq
->refcnt
))
1125 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1128 * @pwq can't be released under pool->lock, bounce to
1129 * pwq_unbound_release_workfn(). This never recurses on the same
1130 * pool->lock as this path is taken only for unbound workqueues and
1131 * the release work item is scheduled on a per-cpu workqueue. To
1132 * avoid lockdep warning, unbound pool->locks are given lockdep
1133 * subclass of 1 in get_unbound_pool().
1135 schedule_work(&pwq
->unbound_release_work
);
1139 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1140 * @pwq: pool_workqueue to put (can be %NULL)
1142 * put_pwq() with locking. This function also allows %NULL @pwq.
1144 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1148 * As both pwqs and pools are RCU protected, the
1149 * following lock operations are safe.
1151 raw_spin_lock_irq(&pwq
->pool
->lock
);
1153 raw_spin_unlock_irq(&pwq
->pool
->lock
);
1157 static void pwq_activate_inactive_work(struct work_struct
*work
)
1159 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1161 trace_workqueue_activate_work(work
);
1162 if (list_empty(&pwq
->pool
->worklist
))
1163 pwq
->pool
->watchdog_ts
= jiffies
;
1164 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1165 __clear_bit(WORK_STRUCT_INACTIVE_BIT
, work_data_bits(work
));
1169 static void pwq_activate_first_inactive(struct pool_workqueue
*pwq
)
1171 struct work_struct
*work
= list_first_entry(&pwq
->inactive_works
,
1172 struct work_struct
, entry
);
1174 pwq_activate_inactive_work(work
);
1178 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1179 * @pwq: pwq of interest
1180 * @work_data: work_data of work which left the queue
1182 * A work either has completed or is removed from pending queue,
1183 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1186 * raw_spin_lock_irq(pool->lock).
1188 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, unsigned long work_data
)
1190 int color
= get_work_color(work_data
);
1192 if (!(work_data
& WORK_STRUCT_INACTIVE
)) {
1194 if (!list_empty(&pwq
->inactive_works
)) {
1195 /* one down, submit an inactive one */
1196 if (pwq
->nr_active
< pwq
->max_active
)
1197 pwq_activate_first_inactive(pwq
);
1201 pwq
->nr_in_flight
[color
]--;
1203 /* is flush in progress and are we at the flushing tip? */
1204 if (likely(pwq
->flush_color
!= color
))
1207 /* are there still in-flight works? */
1208 if (pwq
->nr_in_flight
[color
])
1211 /* this pwq is done, clear flush_color */
1212 pwq
->flush_color
= -1;
1215 * If this was the last pwq, wake up the first flusher. It
1216 * will handle the rest.
1218 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1219 complete(&pwq
->wq
->first_flusher
->done
);
1225 * try_to_grab_pending - steal work item from worklist and disable irq
1226 * @work: work item to steal
1227 * @is_dwork: @work is a delayed_work
1228 * @flags: place to store irq state
1230 * Try to grab PENDING bit of @work. This function can handle @work in any
1231 * stable state - idle, on timer or on worklist.
1235 * ======== ================================================================
1236 * 1 if @work was pending and we successfully stole PENDING
1237 * 0 if @work was idle and we claimed PENDING
1238 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1239 * -ENOENT if someone else is canceling @work, this state may persist
1240 * for arbitrarily long
1241 * ======== ================================================================
1244 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1245 * interrupted while holding PENDING and @work off queue, irq must be
1246 * disabled on entry. This, combined with delayed_work->timer being
1247 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1249 * On successful return, >= 0, irq is disabled and the caller is
1250 * responsible for releasing it using local_irq_restore(*@flags).
1252 * This function is safe to call from any context including IRQ handler.
1254 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1255 unsigned long *flags
)
1257 struct worker_pool
*pool
;
1258 struct pool_workqueue
*pwq
;
1260 local_irq_save(*flags
);
1262 /* try to steal the timer if it exists */
1264 struct delayed_work
*dwork
= to_delayed_work(work
);
1267 * dwork->timer is irqsafe. If del_timer() fails, it's
1268 * guaranteed that the timer is not queued anywhere and not
1269 * running on the local CPU.
1271 if (likely(del_timer(&dwork
->timer
)))
1275 /* try to claim PENDING the normal way */
1276 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1281 * The queueing is in progress, or it is already queued. Try to
1282 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1284 pool
= get_work_pool(work
);
1288 raw_spin_lock(&pool
->lock
);
1290 * work->data is guaranteed to point to pwq only while the work
1291 * item is queued on pwq->wq, and both updating work->data to point
1292 * to pwq on queueing and to pool on dequeueing are done under
1293 * pwq->pool->lock. This in turn guarantees that, if work->data
1294 * points to pwq which is associated with a locked pool, the work
1295 * item is currently queued on that pool.
1297 pwq
= get_work_pwq(work
);
1298 if (pwq
&& pwq
->pool
== pool
) {
1299 debug_work_deactivate(work
);
1302 * A cancelable inactive work item must be in the
1303 * pwq->inactive_works since a queued barrier can't be
1304 * canceled (see the comments in insert_wq_barrier()).
1306 * An inactive work item cannot be grabbed directly because
1307 * it might have linked barrier work items which, if left
1308 * on the inactive_works list, will confuse pwq->nr_active
1309 * management later on and cause stall. Make sure the work
1310 * item is activated before grabbing.
1312 if (*work_data_bits(work
) & WORK_STRUCT_INACTIVE
)
1313 pwq_activate_inactive_work(work
);
1315 list_del_init(&work
->entry
);
1316 pwq_dec_nr_in_flight(pwq
, *work_data_bits(work
));
1318 /* work->data points to pwq iff queued, point to pool */
1319 set_work_pool_and_keep_pending(work
, pool
->id
);
1321 raw_spin_unlock(&pool
->lock
);
1325 raw_spin_unlock(&pool
->lock
);
1328 local_irq_restore(*flags
);
1329 if (work_is_canceling(work
))
1336 * insert_work - insert a work into a pool
1337 * @pwq: pwq @work belongs to
1338 * @work: work to insert
1339 * @head: insertion point
1340 * @extra_flags: extra WORK_STRUCT_* flags to set
1342 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1343 * work_struct flags.
1346 * raw_spin_lock_irq(pool->lock).
1348 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1349 struct list_head
*head
, unsigned int extra_flags
)
1351 struct worker_pool
*pool
= pwq
->pool
;
1353 /* record the work call stack in order to print it in KASAN reports */
1354 kasan_record_aux_stack_noalloc(work
);
1356 /* we own @work, set data and link */
1357 set_work_pwq(work
, pwq
, extra_flags
);
1358 list_add_tail(&work
->entry
, head
);
1362 * Ensure either wq_worker_sleeping() sees the above
1363 * list_add_tail() or we see zero nr_running to avoid workers lying
1364 * around lazily while there are works to be processed.
1368 if (__need_more_worker(pool
))
1369 wake_up_worker(pool
);
1373 * Test whether @work is being queued from another work executing on the
1376 static bool is_chained_work(struct workqueue_struct
*wq
)
1378 struct worker
*worker
;
1380 worker
= current_wq_worker();
1382 * Return %true iff I'm a worker executing a work item on @wq. If
1383 * I'm @worker, it's safe to dereference it without locking.
1385 return worker
&& worker
->current_pwq
->wq
== wq
;
1389 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1390 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1391 * avoid perturbing sensitive tasks.
1393 static int wq_select_unbound_cpu(int cpu
)
1395 static bool printed_dbg_warning
;
1398 if (likely(!wq_debug_force_rr_cpu
)) {
1399 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1401 } else if (!printed_dbg_warning
) {
1402 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1403 printed_dbg_warning
= true;
1406 if (cpumask_empty(wq_unbound_cpumask
))
1409 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1410 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1411 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1412 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1413 if (unlikely(new_cpu
>= nr_cpu_ids
))
1416 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1421 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1422 struct work_struct
*work
)
1424 struct pool_workqueue
*pwq
;
1425 struct worker_pool
*last_pool
;
1426 struct list_head
*worklist
;
1427 unsigned int work_flags
;
1428 unsigned int req_cpu
= cpu
;
1431 * While a work item is PENDING && off queue, a task trying to
1432 * steal the PENDING will busy-loop waiting for it to either get
1433 * queued or lose PENDING. Grabbing PENDING and queueing should
1434 * happen with IRQ disabled.
1436 lockdep_assert_irqs_disabled();
1439 /* if draining, only works from the same workqueue are allowed */
1440 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1441 WARN_ON_ONCE(!is_chained_work(wq
)))
1445 /* pwq which will be used unless @work is executing elsewhere */
1446 if (wq
->flags
& WQ_UNBOUND
) {
1447 if (req_cpu
== WORK_CPU_UNBOUND
)
1448 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1449 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1451 if (req_cpu
== WORK_CPU_UNBOUND
)
1452 cpu
= raw_smp_processor_id();
1453 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1457 * If @work was previously on a different pool, it might still be
1458 * running there, in which case the work needs to be queued on that
1459 * pool to guarantee non-reentrancy.
1461 last_pool
= get_work_pool(work
);
1462 if (last_pool
&& last_pool
!= pwq
->pool
) {
1463 struct worker
*worker
;
1465 raw_spin_lock(&last_pool
->lock
);
1467 worker
= find_worker_executing_work(last_pool
, work
);
1469 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1470 pwq
= worker
->current_pwq
;
1472 /* meh... not running there, queue here */
1473 raw_spin_unlock(&last_pool
->lock
);
1474 raw_spin_lock(&pwq
->pool
->lock
);
1477 raw_spin_lock(&pwq
->pool
->lock
);
1481 * pwq is determined and locked. For unbound pools, we could have
1482 * raced with pwq release and it could already be dead. If its
1483 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1484 * without another pwq replacing it in the numa_pwq_tbl or while
1485 * work items are executing on it, so the retrying is guaranteed to
1486 * make forward-progress.
1488 if (unlikely(!pwq
->refcnt
)) {
1489 if (wq
->flags
& WQ_UNBOUND
) {
1490 raw_spin_unlock(&pwq
->pool
->lock
);
1495 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1499 /* pwq determined, queue */
1500 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1502 if (WARN_ON(!list_empty(&work
->entry
)))
1505 pwq
->nr_in_flight
[pwq
->work_color
]++;
1506 work_flags
= work_color_to_flags(pwq
->work_color
);
1508 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1509 trace_workqueue_activate_work(work
);
1511 worklist
= &pwq
->pool
->worklist
;
1512 if (list_empty(worklist
))
1513 pwq
->pool
->watchdog_ts
= jiffies
;
1515 work_flags
|= WORK_STRUCT_INACTIVE
;
1516 worklist
= &pwq
->inactive_works
;
1519 debug_work_activate(work
);
1520 insert_work(pwq
, work
, worklist
, work_flags
);
1523 raw_spin_unlock(&pwq
->pool
->lock
);
1528 * queue_work_on - queue work on specific cpu
1529 * @cpu: CPU number to execute work on
1530 * @wq: workqueue to use
1531 * @work: work to queue
1533 * We queue the work to a specific CPU, the caller must ensure it
1536 * Return: %false if @work was already on a queue, %true otherwise.
1538 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1539 struct work_struct
*work
)
1542 unsigned long flags
;
1544 local_irq_save(flags
);
1546 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1547 __queue_work(cpu
, wq
, work
);
1551 local_irq_restore(flags
);
1554 EXPORT_SYMBOL(queue_work_on
);
1557 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1558 * @node: NUMA node ID that we want to select a CPU from
1560 * This function will attempt to find a "random" cpu available on a given
1561 * node. If there are no CPUs available on the given node it will return
1562 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1563 * available CPU if we need to schedule this work.
1565 static int workqueue_select_cpu_near(int node
)
1569 /* No point in doing this if NUMA isn't enabled for workqueues */
1570 if (!wq_numa_enabled
)
1571 return WORK_CPU_UNBOUND
;
1573 /* Delay binding to CPU if node is not valid or online */
1574 if (node
< 0 || node
>= MAX_NUMNODES
|| !node_online(node
))
1575 return WORK_CPU_UNBOUND
;
1577 /* Use local node/cpu if we are already there */
1578 cpu
= raw_smp_processor_id();
1579 if (node
== cpu_to_node(cpu
))
1582 /* Use "random" otherwise know as "first" online CPU of node */
1583 cpu
= cpumask_any_and(cpumask_of_node(node
), cpu_online_mask
);
1585 /* If CPU is valid return that, otherwise just defer */
1586 return cpu
< nr_cpu_ids
? cpu
: WORK_CPU_UNBOUND
;
1590 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1591 * @node: NUMA node that we are targeting the work for
1592 * @wq: workqueue to use
1593 * @work: work to queue
1595 * We queue the work to a "random" CPU within a given NUMA node. The basic
1596 * idea here is to provide a way to somehow associate work with a given
1599 * This function will only make a best effort attempt at getting this onto
1600 * the right NUMA node. If no node is requested or the requested node is
1601 * offline then we just fall back to standard queue_work behavior.
1603 * Currently the "random" CPU ends up being the first available CPU in the
1604 * intersection of cpu_online_mask and the cpumask of the node, unless we
1605 * are running on the node. In that case we just use the current CPU.
1607 * Return: %false if @work was already on a queue, %true otherwise.
1609 bool queue_work_node(int node
, struct workqueue_struct
*wq
,
1610 struct work_struct
*work
)
1612 unsigned long flags
;
1616 * This current implementation is specific to unbound workqueues.
1617 * Specifically we only return the first available CPU for a given
1618 * node instead of cycling through individual CPUs within the node.
1620 * If this is used with a per-cpu workqueue then the logic in
1621 * workqueue_select_cpu_near would need to be updated to allow for
1622 * some round robin type logic.
1624 WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
));
1626 local_irq_save(flags
);
1628 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1629 int cpu
= workqueue_select_cpu_near(node
);
1631 __queue_work(cpu
, wq
, work
);
1635 local_irq_restore(flags
);
1638 EXPORT_SYMBOL_GPL(queue_work_node
);
1640 void delayed_work_timer_fn(struct timer_list
*t
)
1642 struct delayed_work
*dwork
= from_timer(dwork
, t
, timer
);
1644 /* should have been called from irqsafe timer with irq already off */
1645 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1647 EXPORT_SYMBOL(delayed_work_timer_fn
);
1649 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1650 struct delayed_work
*dwork
, unsigned long delay
)
1652 struct timer_list
*timer
= &dwork
->timer
;
1653 struct work_struct
*work
= &dwork
->work
;
1656 WARN_ON_FUNCTION_MISMATCH(timer
->function
, delayed_work_timer_fn
);
1657 WARN_ON_ONCE(timer_pending(timer
));
1658 WARN_ON_ONCE(!list_empty(&work
->entry
));
1661 * If @delay is 0, queue @dwork->work immediately. This is for
1662 * both optimization and correctness. The earliest @timer can
1663 * expire is on the closest next tick and delayed_work users depend
1664 * on that there's no such delay when @delay is 0.
1667 __queue_work(cpu
, wq
, &dwork
->work
);
1673 timer
->expires
= jiffies
+ delay
;
1675 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1676 add_timer_on(timer
, cpu
);
1682 * queue_delayed_work_on - queue work on specific CPU after delay
1683 * @cpu: CPU number to execute work on
1684 * @wq: workqueue to use
1685 * @dwork: work to queue
1686 * @delay: number of jiffies to wait before queueing
1688 * Return: %false if @work was already on a queue, %true otherwise. If
1689 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1692 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1693 struct delayed_work
*dwork
, unsigned long delay
)
1695 struct work_struct
*work
= &dwork
->work
;
1697 unsigned long flags
;
1699 /* read the comment in __queue_work() */
1700 local_irq_save(flags
);
1702 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1703 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1707 local_irq_restore(flags
);
1710 EXPORT_SYMBOL(queue_delayed_work_on
);
1713 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1714 * @cpu: CPU number to execute work on
1715 * @wq: workqueue to use
1716 * @dwork: work to queue
1717 * @delay: number of jiffies to wait before queueing
1719 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1720 * modify @dwork's timer so that it expires after @delay. If @delay is
1721 * zero, @work is guaranteed to be scheduled immediately regardless of its
1724 * Return: %false if @dwork was idle and queued, %true if @dwork was
1725 * pending and its timer was modified.
1727 * This function is safe to call from any context including IRQ handler.
1728 * See try_to_grab_pending() for details.
1730 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1731 struct delayed_work
*dwork
, unsigned long delay
)
1733 unsigned long flags
;
1737 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1738 } while (unlikely(ret
== -EAGAIN
));
1740 if (likely(ret
>= 0)) {
1741 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1742 local_irq_restore(flags
);
1745 /* -ENOENT from try_to_grab_pending() becomes %true */
1748 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1750 static void rcu_work_rcufn(struct rcu_head
*rcu
)
1752 struct rcu_work
*rwork
= container_of(rcu
, struct rcu_work
, rcu
);
1754 /* read the comment in __queue_work() */
1755 local_irq_disable();
1756 __queue_work(WORK_CPU_UNBOUND
, rwork
->wq
, &rwork
->work
);
1761 * queue_rcu_work - queue work after a RCU grace period
1762 * @wq: workqueue to use
1763 * @rwork: work to queue
1765 * Return: %false if @rwork was already pending, %true otherwise. Note
1766 * that a full RCU grace period is guaranteed only after a %true return.
1767 * While @rwork is guaranteed to be executed after a %false return, the
1768 * execution may happen before a full RCU grace period has passed.
1770 bool queue_rcu_work(struct workqueue_struct
*wq
, struct rcu_work
*rwork
)
1772 struct work_struct
*work
= &rwork
->work
;
1774 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1776 call_rcu(&rwork
->rcu
, rcu_work_rcufn
);
1782 EXPORT_SYMBOL(queue_rcu_work
);
1785 * worker_enter_idle - enter idle state
1786 * @worker: worker which is entering idle state
1788 * @worker is entering idle state. Update stats and idle timer if
1792 * raw_spin_lock_irq(pool->lock).
1794 static void worker_enter_idle(struct worker
*worker
)
1796 struct worker_pool
*pool
= worker
->pool
;
1798 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1799 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1800 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1803 /* can't use worker_set_flags(), also called from create_worker() */
1804 worker
->flags
|= WORKER_IDLE
;
1806 worker
->last_active
= jiffies
;
1808 /* idle_list is LIFO */
1809 list_add(&worker
->entry
, &pool
->idle_list
);
1811 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1812 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1815 * Sanity check nr_running. Because unbind_workers() releases
1816 * pool->lock between setting %WORKER_UNBOUND and zapping
1817 * nr_running, the warning may trigger spuriously. Check iff
1818 * unbind is not in progress.
1820 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1821 pool
->nr_workers
== pool
->nr_idle
&&
1822 atomic_read(&pool
->nr_running
));
1826 * worker_leave_idle - leave idle state
1827 * @worker: worker which is leaving idle state
1829 * @worker is leaving idle state. Update stats.
1832 * raw_spin_lock_irq(pool->lock).
1834 static void worker_leave_idle(struct worker
*worker
)
1836 struct worker_pool
*pool
= worker
->pool
;
1838 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1840 worker_clr_flags(worker
, WORKER_IDLE
);
1842 list_del_init(&worker
->entry
);
1845 static struct worker
*alloc_worker(int node
)
1847 struct worker
*worker
;
1849 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1851 INIT_LIST_HEAD(&worker
->entry
);
1852 INIT_LIST_HEAD(&worker
->scheduled
);
1853 INIT_LIST_HEAD(&worker
->node
);
1854 /* on creation a worker is in !idle && prep state */
1855 worker
->flags
= WORKER_PREP
;
1861 * worker_attach_to_pool() - attach a worker to a pool
1862 * @worker: worker to be attached
1863 * @pool: the target pool
1865 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1866 * cpu-binding of @worker are kept coordinated with the pool across
1869 static void worker_attach_to_pool(struct worker
*worker
,
1870 struct worker_pool
*pool
)
1872 mutex_lock(&wq_pool_attach_mutex
);
1875 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1876 * stable across this function. See the comments above the flag
1877 * definition for details.
1879 if (pool
->flags
& POOL_DISASSOCIATED
)
1880 worker
->flags
|= WORKER_UNBOUND
;
1882 kthread_set_per_cpu(worker
->task
, pool
->cpu
);
1884 if (worker
->rescue_wq
)
1885 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1887 list_add_tail(&worker
->node
, &pool
->workers
);
1888 worker
->pool
= pool
;
1890 mutex_unlock(&wq_pool_attach_mutex
);
1894 * worker_detach_from_pool() - detach a worker from its pool
1895 * @worker: worker which is attached to its pool
1897 * Undo the attaching which had been done in worker_attach_to_pool(). The
1898 * caller worker shouldn't access to the pool after detached except it has
1899 * other reference to the pool.
1901 static void worker_detach_from_pool(struct worker
*worker
)
1903 struct worker_pool
*pool
= worker
->pool
;
1904 struct completion
*detach_completion
= NULL
;
1906 mutex_lock(&wq_pool_attach_mutex
);
1908 kthread_set_per_cpu(worker
->task
, -1);
1909 list_del(&worker
->node
);
1910 worker
->pool
= NULL
;
1912 if (list_empty(&pool
->workers
))
1913 detach_completion
= pool
->detach_completion
;
1914 mutex_unlock(&wq_pool_attach_mutex
);
1916 /* clear leftover flags without pool->lock after it is detached */
1917 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1919 if (detach_completion
)
1920 complete(detach_completion
);
1924 * create_worker - create a new workqueue worker
1925 * @pool: pool the new worker will belong to
1927 * Create and start a new worker which is attached to @pool.
1930 * Might sleep. Does GFP_KERNEL allocations.
1933 * Pointer to the newly created worker.
1935 static struct worker
*create_worker(struct worker_pool
*pool
)
1937 struct worker
*worker
;
1941 /* ID is needed to determine kthread name */
1942 id
= ida_alloc(&pool
->worker_ida
, GFP_KERNEL
);
1946 worker
= alloc_worker(pool
->node
);
1953 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1954 pool
->attrs
->nice
< 0 ? "H" : "");
1956 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1958 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1959 "kworker/%s", id_buf
);
1960 if (IS_ERR(worker
->task
))
1963 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1964 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1966 /* successful, attach the worker to the pool */
1967 worker_attach_to_pool(worker
, pool
);
1969 /* start the newly created worker */
1970 raw_spin_lock_irq(&pool
->lock
);
1971 worker
->pool
->nr_workers
++;
1972 worker_enter_idle(worker
);
1973 wake_up_process(worker
->task
);
1974 raw_spin_unlock_irq(&pool
->lock
);
1979 ida_free(&pool
->worker_ida
, id
);
1985 * destroy_worker - destroy a workqueue worker
1986 * @worker: worker to be destroyed
1988 * Destroy @worker and adjust @pool stats accordingly. The worker should
1992 * raw_spin_lock_irq(pool->lock).
1994 static void destroy_worker(struct worker
*worker
)
1996 struct worker_pool
*pool
= worker
->pool
;
1998 lockdep_assert_held(&pool
->lock
);
2000 /* sanity check frenzy */
2001 if (WARN_ON(worker
->current_work
) ||
2002 WARN_ON(!list_empty(&worker
->scheduled
)) ||
2003 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
2009 list_del_init(&worker
->entry
);
2010 worker
->flags
|= WORKER_DIE
;
2011 wake_up_process(worker
->task
);
2014 static void idle_worker_timeout(struct timer_list
*t
)
2016 struct worker_pool
*pool
= from_timer(pool
, t
, idle_timer
);
2018 raw_spin_lock_irq(&pool
->lock
);
2020 while (too_many_workers(pool
)) {
2021 struct worker
*worker
;
2022 unsigned long expires
;
2024 /* idle_list is kept in LIFO order, check the last one */
2025 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2026 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2028 if (time_before(jiffies
, expires
)) {
2029 mod_timer(&pool
->idle_timer
, expires
);
2033 destroy_worker(worker
);
2036 raw_spin_unlock_irq(&pool
->lock
);
2039 static void send_mayday(struct work_struct
*work
)
2041 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2042 struct workqueue_struct
*wq
= pwq
->wq
;
2044 lockdep_assert_held(&wq_mayday_lock
);
2049 /* mayday mayday mayday */
2050 if (list_empty(&pwq
->mayday_node
)) {
2052 * If @pwq is for an unbound wq, its base ref may be put at
2053 * any time due to an attribute change. Pin @pwq until the
2054 * rescuer is done with it.
2057 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2058 wake_up_process(wq
->rescuer
->task
);
2062 static void pool_mayday_timeout(struct timer_list
*t
)
2064 struct worker_pool
*pool
= from_timer(pool
, t
, mayday_timer
);
2065 struct work_struct
*work
;
2067 raw_spin_lock_irq(&pool
->lock
);
2068 raw_spin_lock(&wq_mayday_lock
); /* for wq->maydays */
2070 if (need_to_create_worker(pool
)) {
2072 * We've been trying to create a new worker but
2073 * haven't been successful. We might be hitting an
2074 * allocation deadlock. Send distress signals to
2077 list_for_each_entry(work
, &pool
->worklist
, entry
)
2081 raw_spin_unlock(&wq_mayday_lock
);
2082 raw_spin_unlock_irq(&pool
->lock
);
2084 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
2088 * maybe_create_worker - create a new worker if necessary
2089 * @pool: pool to create a new worker for
2091 * Create a new worker for @pool if necessary. @pool is guaranteed to
2092 * have at least one idle worker on return from this function. If
2093 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2094 * sent to all rescuers with works scheduled on @pool to resolve
2095 * possible allocation deadlock.
2097 * On return, need_to_create_worker() is guaranteed to be %false and
2098 * may_start_working() %true.
2101 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2102 * multiple times. Does GFP_KERNEL allocations. Called only from
2105 static void maybe_create_worker(struct worker_pool
*pool
)
2106 __releases(&pool
->lock
)
2107 __acquires(&pool
->lock
)
2110 raw_spin_unlock_irq(&pool
->lock
);
2112 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2113 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
2116 if (create_worker(pool
) || !need_to_create_worker(pool
))
2119 schedule_timeout_interruptible(CREATE_COOLDOWN
);
2121 if (!need_to_create_worker(pool
))
2125 del_timer_sync(&pool
->mayday_timer
);
2126 raw_spin_lock_irq(&pool
->lock
);
2128 * This is necessary even after a new worker was just successfully
2129 * created as @pool->lock was dropped and the new worker might have
2130 * already become busy.
2132 if (need_to_create_worker(pool
))
2137 * manage_workers - manage worker pool
2140 * Assume the manager role and manage the worker pool @worker belongs
2141 * to. At any given time, there can be only zero or one manager per
2142 * pool. The exclusion is handled automatically by this function.
2144 * The caller can safely start processing works on false return. On
2145 * true return, it's guaranteed that need_to_create_worker() is false
2146 * and may_start_working() is true.
2149 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2150 * multiple times. Does GFP_KERNEL allocations.
2153 * %false if the pool doesn't need management and the caller can safely
2154 * start processing works, %true if management function was performed and
2155 * the conditions that the caller verified before calling the function may
2156 * no longer be true.
2158 static bool manage_workers(struct worker
*worker
)
2160 struct worker_pool
*pool
= worker
->pool
;
2162 if (pool
->flags
& POOL_MANAGER_ACTIVE
)
2165 pool
->flags
|= POOL_MANAGER_ACTIVE
;
2166 pool
->manager
= worker
;
2168 maybe_create_worker(pool
);
2170 pool
->manager
= NULL
;
2171 pool
->flags
&= ~POOL_MANAGER_ACTIVE
;
2172 rcuwait_wake_up(&manager_wait
);
2177 * process_one_work - process single work
2179 * @work: work to process
2181 * Process @work. This function contains all the logics necessary to
2182 * process a single work including synchronization against and
2183 * interaction with other workers on the same cpu, queueing and
2184 * flushing. As long as context requirement is met, any worker can
2185 * call this function to process a work.
2188 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2190 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2191 __releases(&pool
->lock
)
2192 __acquires(&pool
->lock
)
2194 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2195 struct worker_pool
*pool
= worker
->pool
;
2196 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2197 unsigned long work_data
;
2198 struct worker
*collision
;
2199 #ifdef CONFIG_LOCKDEP
2201 * It is permissible to free the struct work_struct from
2202 * inside the function that is called from it, this we need to
2203 * take into account for lockdep too. To avoid bogus "held
2204 * lock freed" warnings as well as problems when looking into
2205 * work->lockdep_map, make a copy and use that here.
2207 struct lockdep_map lockdep_map
;
2209 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2211 /* ensure we're on the correct CPU */
2212 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2213 raw_smp_processor_id() != pool
->cpu
);
2216 * A single work shouldn't be executed concurrently by
2217 * multiple workers on a single cpu. Check whether anyone is
2218 * already processing the work. If so, defer the work to the
2219 * currently executing one.
2221 collision
= find_worker_executing_work(pool
, work
);
2222 if (unlikely(collision
)) {
2223 move_linked_works(work
, &collision
->scheduled
, NULL
);
2227 /* claim and dequeue */
2228 debug_work_deactivate(work
);
2229 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2230 worker
->current_work
= work
;
2231 worker
->current_func
= work
->func
;
2232 worker
->current_pwq
= pwq
;
2233 work_data
= *work_data_bits(work
);
2234 worker
->current_color
= get_work_color(work_data
);
2237 * Record wq name for cmdline and debug reporting, may get
2238 * overridden through set_worker_desc().
2240 strscpy(worker
->desc
, pwq
->wq
->name
, WORKER_DESC_LEN
);
2242 list_del_init(&work
->entry
);
2245 * CPU intensive works don't participate in concurrency management.
2246 * They're the scheduler's responsibility. This takes @worker out
2247 * of concurrency management and the next code block will chain
2248 * execution of the pending work items.
2250 if (unlikely(cpu_intensive
))
2251 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2254 * Wake up another worker if necessary. The condition is always
2255 * false for normal per-cpu workers since nr_running would always
2256 * be >= 1 at this point. This is used to chain execution of the
2257 * pending work items for WORKER_NOT_RUNNING workers such as the
2258 * UNBOUND and CPU_INTENSIVE ones.
2260 if (need_more_worker(pool
))
2261 wake_up_worker(pool
);
2264 * Record the last pool and clear PENDING which should be the last
2265 * update to @work. Also, do this inside @pool->lock so that
2266 * PENDING and queued state changes happen together while IRQ is
2269 set_work_pool_and_clear_pending(work
, pool
->id
);
2271 raw_spin_unlock_irq(&pool
->lock
);
2273 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2274 lock_map_acquire(&lockdep_map
);
2276 * Strictly speaking we should mark the invariant state without holding
2277 * any locks, that is, before these two lock_map_acquire()'s.
2279 * However, that would result in:
2286 * Which would create W1->C->W1 dependencies, even though there is no
2287 * actual deadlock possible. There are two solutions, using a
2288 * read-recursive acquire on the work(queue) 'locks', but this will then
2289 * hit the lockdep limitation on recursive locks, or simply discard
2292 * AFAICT there is no possible deadlock scenario between the
2293 * flush_work() and complete() primitives (except for single-threaded
2294 * workqueues), so hiding them isn't a problem.
2296 lockdep_invariant_state(true);
2297 trace_workqueue_execute_start(work
);
2298 worker
->current_func(work
);
2300 * While we must be careful to not use "work" after this, the trace
2301 * point will only record its address.
2303 trace_workqueue_execute_end(work
, worker
->current_func
);
2304 lock_map_release(&lockdep_map
);
2305 lock_map_release(&pwq
->wq
->lockdep_map
);
2307 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2308 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2309 " last function: %ps\n",
2310 current
->comm
, preempt_count(), task_pid_nr(current
),
2311 worker
->current_func
);
2312 debug_show_held_locks(current
);
2317 * The following prevents a kworker from hogging CPU on !PREEMPTION
2318 * kernels, where a requeueing work item waiting for something to
2319 * happen could deadlock with stop_machine as such work item could
2320 * indefinitely requeue itself while all other CPUs are trapped in
2321 * stop_machine. At the same time, report a quiescent RCU state so
2322 * the same condition doesn't freeze RCU.
2326 raw_spin_lock_irq(&pool
->lock
);
2328 /* clear cpu intensive status */
2329 if (unlikely(cpu_intensive
))
2330 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2332 /* tag the worker for identification in schedule() */
2333 worker
->last_func
= worker
->current_func
;
2335 /* we're done with it, release */
2336 hash_del(&worker
->hentry
);
2337 worker
->current_work
= NULL
;
2338 worker
->current_func
= NULL
;
2339 worker
->current_pwq
= NULL
;
2340 worker
->current_color
= INT_MAX
;
2341 pwq_dec_nr_in_flight(pwq
, work_data
);
2345 * process_scheduled_works - process scheduled works
2348 * Process all scheduled works. Please note that the scheduled list
2349 * may change while processing a work, so this function repeatedly
2350 * fetches a work from the top and executes it.
2353 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2356 static void process_scheduled_works(struct worker
*worker
)
2358 while (!list_empty(&worker
->scheduled
)) {
2359 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2360 struct work_struct
, entry
);
2361 process_one_work(worker
, work
);
2365 static void set_pf_worker(bool val
)
2367 mutex_lock(&wq_pool_attach_mutex
);
2369 current
->flags
|= PF_WQ_WORKER
;
2371 current
->flags
&= ~PF_WQ_WORKER
;
2372 mutex_unlock(&wq_pool_attach_mutex
);
2376 * worker_thread - the worker thread function
2379 * The worker thread function. All workers belong to a worker_pool -
2380 * either a per-cpu one or dynamic unbound one. These workers process all
2381 * work items regardless of their specific target workqueue. The only
2382 * exception is work items which belong to workqueues with a rescuer which
2383 * will be explained in rescuer_thread().
2387 static int worker_thread(void *__worker
)
2389 struct worker
*worker
= __worker
;
2390 struct worker_pool
*pool
= worker
->pool
;
2392 /* tell the scheduler that this is a workqueue worker */
2393 set_pf_worker(true);
2395 raw_spin_lock_irq(&pool
->lock
);
2397 /* am I supposed to die? */
2398 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2399 raw_spin_unlock_irq(&pool
->lock
);
2400 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2401 set_pf_worker(false);
2403 set_task_comm(worker
->task
, "kworker/dying");
2404 ida_free(&pool
->worker_ida
, worker
->id
);
2405 worker_detach_from_pool(worker
);
2410 worker_leave_idle(worker
);
2412 /* no more worker necessary? */
2413 if (!need_more_worker(pool
))
2416 /* do we need to manage? */
2417 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2421 * ->scheduled list can only be filled while a worker is
2422 * preparing to process a work or actually processing it.
2423 * Make sure nobody diddled with it while I was sleeping.
2425 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2428 * Finish PREP stage. We're guaranteed to have at least one idle
2429 * worker or that someone else has already assumed the manager
2430 * role. This is where @worker starts participating in concurrency
2431 * management if applicable and concurrency management is restored
2432 * after being rebound. See rebind_workers() for details.
2434 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2437 struct work_struct
*work
=
2438 list_first_entry(&pool
->worklist
,
2439 struct work_struct
, entry
);
2441 pool
->watchdog_ts
= jiffies
;
2443 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2444 /* optimization path, not strictly necessary */
2445 process_one_work(worker
, work
);
2446 if (unlikely(!list_empty(&worker
->scheduled
)))
2447 process_scheduled_works(worker
);
2449 move_linked_works(work
, &worker
->scheduled
, NULL
);
2450 process_scheduled_works(worker
);
2452 } while (keep_working(pool
));
2454 worker_set_flags(worker
, WORKER_PREP
);
2457 * pool->lock is held and there's no work to process and no need to
2458 * manage, sleep. Workers are woken up only while holding
2459 * pool->lock or from local cpu, so setting the current state
2460 * before releasing pool->lock is enough to prevent losing any
2463 worker_enter_idle(worker
);
2464 __set_current_state(TASK_IDLE
);
2465 raw_spin_unlock_irq(&pool
->lock
);
2471 * rescuer_thread - the rescuer thread function
2474 * Workqueue rescuer thread function. There's one rescuer for each
2475 * workqueue which has WQ_MEM_RECLAIM set.
2477 * Regular work processing on a pool may block trying to create a new
2478 * worker which uses GFP_KERNEL allocation which has slight chance of
2479 * developing into deadlock if some works currently on the same queue
2480 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2481 * the problem rescuer solves.
2483 * When such condition is possible, the pool summons rescuers of all
2484 * workqueues which have works queued on the pool and let them process
2485 * those works so that forward progress can be guaranteed.
2487 * This should happen rarely.
2491 static int rescuer_thread(void *__rescuer
)
2493 struct worker
*rescuer
= __rescuer
;
2494 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2495 struct list_head
*scheduled
= &rescuer
->scheduled
;
2498 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2501 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2502 * doesn't participate in concurrency management.
2504 set_pf_worker(true);
2506 set_current_state(TASK_IDLE
);
2509 * By the time the rescuer is requested to stop, the workqueue
2510 * shouldn't have any work pending, but @wq->maydays may still have
2511 * pwq(s) queued. This can happen by non-rescuer workers consuming
2512 * all the work items before the rescuer got to them. Go through
2513 * @wq->maydays processing before acting on should_stop so that the
2514 * list is always empty on exit.
2516 should_stop
= kthread_should_stop();
2518 /* see whether any pwq is asking for help */
2519 raw_spin_lock_irq(&wq_mayday_lock
);
2521 while (!list_empty(&wq
->maydays
)) {
2522 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2523 struct pool_workqueue
, mayday_node
);
2524 struct worker_pool
*pool
= pwq
->pool
;
2525 struct work_struct
*work
, *n
;
2528 __set_current_state(TASK_RUNNING
);
2529 list_del_init(&pwq
->mayday_node
);
2531 raw_spin_unlock_irq(&wq_mayday_lock
);
2533 worker_attach_to_pool(rescuer
, pool
);
2535 raw_spin_lock_irq(&pool
->lock
);
2538 * Slurp in all works issued via this workqueue and
2541 WARN_ON_ONCE(!list_empty(scheduled
));
2542 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2543 if (get_work_pwq(work
) == pwq
) {
2545 pool
->watchdog_ts
= jiffies
;
2546 move_linked_works(work
, scheduled
, &n
);
2551 if (!list_empty(scheduled
)) {
2552 process_scheduled_works(rescuer
);
2555 * The above execution of rescued work items could
2556 * have created more to rescue through
2557 * pwq_activate_first_inactive() or chained
2558 * queueing. Let's put @pwq back on mayday list so
2559 * that such back-to-back work items, which may be
2560 * being used to relieve memory pressure, don't
2561 * incur MAYDAY_INTERVAL delay inbetween.
2563 if (pwq
->nr_active
&& need_to_create_worker(pool
)) {
2564 raw_spin_lock(&wq_mayday_lock
);
2566 * Queue iff we aren't racing destruction
2567 * and somebody else hasn't queued it already.
2569 if (wq
->rescuer
&& list_empty(&pwq
->mayday_node
)) {
2571 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2573 raw_spin_unlock(&wq_mayday_lock
);
2578 * Put the reference grabbed by send_mayday(). @pool won't
2579 * go away while we're still attached to it.
2584 * Leave this pool. If need_more_worker() is %true, notify a
2585 * regular worker; otherwise, we end up with 0 concurrency
2586 * and stalling the execution.
2588 if (need_more_worker(pool
))
2589 wake_up_worker(pool
);
2591 raw_spin_unlock_irq(&pool
->lock
);
2593 worker_detach_from_pool(rescuer
);
2595 raw_spin_lock_irq(&wq_mayday_lock
);
2598 raw_spin_unlock_irq(&wq_mayday_lock
);
2601 __set_current_state(TASK_RUNNING
);
2602 set_pf_worker(false);
2606 /* rescuers should never participate in concurrency management */
2607 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2613 * check_flush_dependency - check for flush dependency sanity
2614 * @target_wq: workqueue being flushed
2615 * @target_work: work item being flushed (NULL for workqueue flushes)
2617 * %current is trying to flush the whole @target_wq or @target_work on it.
2618 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2619 * reclaiming memory or running on a workqueue which doesn't have
2620 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2623 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2624 struct work_struct
*target_work
)
2626 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2627 struct worker
*worker
;
2629 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2632 worker
= current_wq_worker();
2634 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2635 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2636 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2637 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2638 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2639 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2640 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2641 target_wq
->name
, target_func
);
2645 struct work_struct work
;
2646 struct completion done
;
2647 struct task_struct
*task
; /* purely informational */
2650 static void wq_barrier_func(struct work_struct
*work
)
2652 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2653 complete(&barr
->done
);
2657 * insert_wq_barrier - insert a barrier work
2658 * @pwq: pwq to insert barrier into
2659 * @barr: wq_barrier to insert
2660 * @target: target work to attach @barr to
2661 * @worker: worker currently executing @target, NULL if @target is not executing
2663 * @barr is linked to @target such that @barr is completed only after
2664 * @target finishes execution. Please note that the ordering
2665 * guarantee is observed only with respect to @target and on the local
2668 * Currently, a queued barrier can't be canceled. This is because
2669 * try_to_grab_pending() can't determine whether the work to be
2670 * grabbed is at the head of the queue and thus can't clear LINKED
2671 * flag of the previous work while there must be a valid next work
2672 * after a work with LINKED flag set.
2674 * Note that when @worker is non-NULL, @target may be modified
2675 * underneath us, so we can't reliably determine pwq from @target.
2678 * raw_spin_lock_irq(pool->lock).
2680 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2681 struct wq_barrier
*barr
,
2682 struct work_struct
*target
, struct worker
*worker
)
2684 unsigned int work_flags
= 0;
2685 unsigned int work_color
;
2686 struct list_head
*head
;
2689 * debugobject calls are safe here even with pool->lock locked
2690 * as we know for sure that this will not trigger any of the
2691 * checks and call back into the fixup functions where we
2694 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2695 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2697 init_completion_map(&barr
->done
, &target
->lockdep_map
);
2699 barr
->task
= current
;
2701 /* The barrier work item does not participate in pwq->nr_active. */
2702 work_flags
|= WORK_STRUCT_INACTIVE
;
2705 * If @target is currently being executed, schedule the
2706 * barrier to the worker; otherwise, put it after @target.
2709 head
= worker
->scheduled
.next
;
2710 work_color
= worker
->current_color
;
2712 unsigned long *bits
= work_data_bits(target
);
2714 head
= target
->entry
.next
;
2715 /* there can already be other linked works, inherit and set */
2716 work_flags
|= *bits
& WORK_STRUCT_LINKED
;
2717 work_color
= get_work_color(*bits
);
2718 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2721 pwq
->nr_in_flight
[work_color
]++;
2722 work_flags
|= work_color_to_flags(work_color
);
2724 debug_work_activate(&barr
->work
);
2725 insert_work(pwq
, &barr
->work
, head
, work_flags
);
2729 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2730 * @wq: workqueue being flushed
2731 * @flush_color: new flush color, < 0 for no-op
2732 * @work_color: new work color, < 0 for no-op
2734 * Prepare pwqs for workqueue flushing.
2736 * If @flush_color is non-negative, flush_color on all pwqs should be
2737 * -1. If no pwq has in-flight commands at the specified color, all
2738 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2739 * has in flight commands, its pwq->flush_color is set to
2740 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2741 * wakeup logic is armed and %true is returned.
2743 * The caller should have initialized @wq->first_flusher prior to
2744 * calling this function with non-negative @flush_color. If
2745 * @flush_color is negative, no flush color update is done and %false
2748 * If @work_color is non-negative, all pwqs should have the same
2749 * work_color which is previous to @work_color and all will be
2750 * advanced to @work_color.
2753 * mutex_lock(wq->mutex).
2756 * %true if @flush_color >= 0 and there's something to flush. %false
2759 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2760 int flush_color
, int work_color
)
2763 struct pool_workqueue
*pwq
;
2765 if (flush_color
>= 0) {
2766 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2767 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2770 for_each_pwq(pwq
, wq
) {
2771 struct worker_pool
*pool
= pwq
->pool
;
2773 raw_spin_lock_irq(&pool
->lock
);
2775 if (flush_color
>= 0) {
2776 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2778 if (pwq
->nr_in_flight
[flush_color
]) {
2779 pwq
->flush_color
= flush_color
;
2780 atomic_inc(&wq
->nr_pwqs_to_flush
);
2785 if (work_color
>= 0) {
2786 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2787 pwq
->work_color
= work_color
;
2790 raw_spin_unlock_irq(&pool
->lock
);
2793 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2794 complete(&wq
->first_flusher
->done
);
2800 * flush_workqueue - ensure that any scheduled work has run to completion.
2801 * @wq: workqueue to flush
2803 * This function sleeps until all work items which were queued on entry
2804 * have finished execution, but it is not livelocked by new incoming ones.
2806 void flush_workqueue(struct workqueue_struct
*wq
)
2808 struct wq_flusher this_flusher
= {
2809 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2811 .done
= COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher
.done
, wq
->lockdep_map
),
2815 if (WARN_ON(!wq_online
))
2818 lock_map_acquire(&wq
->lockdep_map
);
2819 lock_map_release(&wq
->lockdep_map
);
2821 mutex_lock(&wq
->mutex
);
2824 * Start-to-wait phase
2826 next_color
= work_next_color(wq
->work_color
);
2828 if (next_color
!= wq
->flush_color
) {
2830 * Color space is not full. The current work_color
2831 * becomes our flush_color and work_color is advanced
2834 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2835 this_flusher
.flush_color
= wq
->work_color
;
2836 wq
->work_color
= next_color
;
2838 if (!wq
->first_flusher
) {
2839 /* no flush in progress, become the first flusher */
2840 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2842 wq
->first_flusher
= &this_flusher
;
2844 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2846 /* nothing to flush, done */
2847 wq
->flush_color
= next_color
;
2848 wq
->first_flusher
= NULL
;
2853 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2854 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2855 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2859 * Oops, color space is full, wait on overflow queue.
2860 * The next flush completion will assign us
2861 * flush_color and transfer to flusher_queue.
2863 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2866 check_flush_dependency(wq
, NULL
);
2868 mutex_unlock(&wq
->mutex
);
2870 wait_for_completion(&this_flusher
.done
);
2873 * Wake-up-and-cascade phase
2875 * First flushers are responsible for cascading flushes and
2876 * handling overflow. Non-first flushers can simply return.
2878 if (READ_ONCE(wq
->first_flusher
) != &this_flusher
)
2881 mutex_lock(&wq
->mutex
);
2883 /* we might have raced, check again with mutex held */
2884 if (wq
->first_flusher
!= &this_flusher
)
2887 WRITE_ONCE(wq
->first_flusher
, NULL
);
2889 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2890 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2893 struct wq_flusher
*next
, *tmp
;
2895 /* complete all the flushers sharing the current flush color */
2896 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2897 if (next
->flush_color
!= wq
->flush_color
)
2899 list_del_init(&next
->list
);
2900 complete(&next
->done
);
2903 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2904 wq
->flush_color
!= work_next_color(wq
->work_color
));
2906 /* this flush_color is finished, advance by one */
2907 wq
->flush_color
= work_next_color(wq
->flush_color
);
2909 /* one color has been freed, handle overflow queue */
2910 if (!list_empty(&wq
->flusher_overflow
)) {
2912 * Assign the same color to all overflowed
2913 * flushers, advance work_color and append to
2914 * flusher_queue. This is the start-to-wait
2915 * phase for these overflowed flushers.
2917 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2918 tmp
->flush_color
= wq
->work_color
;
2920 wq
->work_color
= work_next_color(wq
->work_color
);
2922 list_splice_tail_init(&wq
->flusher_overflow
,
2923 &wq
->flusher_queue
);
2924 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2927 if (list_empty(&wq
->flusher_queue
)) {
2928 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2933 * Need to flush more colors. Make the next flusher
2934 * the new first flusher and arm pwqs.
2936 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2937 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2939 list_del_init(&next
->list
);
2940 wq
->first_flusher
= next
;
2942 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2946 * Meh... this color is already done, clear first
2947 * flusher and repeat cascading.
2949 wq
->first_flusher
= NULL
;
2953 mutex_unlock(&wq
->mutex
);
2955 EXPORT_SYMBOL(flush_workqueue
);
2958 * drain_workqueue - drain a workqueue
2959 * @wq: workqueue to drain
2961 * Wait until the workqueue becomes empty. While draining is in progress,
2962 * only chain queueing is allowed. IOW, only currently pending or running
2963 * work items on @wq can queue further work items on it. @wq is flushed
2964 * repeatedly until it becomes empty. The number of flushing is determined
2965 * by the depth of chaining and should be relatively short. Whine if it
2968 void drain_workqueue(struct workqueue_struct
*wq
)
2970 unsigned int flush_cnt
= 0;
2971 struct pool_workqueue
*pwq
;
2974 * __queue_work() needs to test whether there are drainers, is much
2975 * hotter than drain_workqueue() and already looks at @wq->flags.
2976 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2978 mutex_lock(&wq
->mutex
);
2979 if (!wq
->nr_drainers
++)
2980 wq
->flags
|= __WQ_DRAINING
;
2981 mutex_unlock(&wq
->mutex
);
2983 flush_workqueue(wq
);
2985 mutex_lock(&wq
->mutex
);
2987 for_each_pwq(pwq
, wq
) {
2990 raw_spin_lock_irq(&pwq
->pool
->lock
);
2991 drained
= !pwq
->nr_active
&& list_empty(&pwq
->inactive_works
);
2992 raw_spin_unlock_irq(&pwq
->pool
->lock
);
2997 if (++flush_cnt
== 10 ||
2998 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2999 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3000 wq
->name
, __func__
, flush_cnt
);
3002 mutex_unlock(&wq
->mutex
);
3006 if (!--wq
->nr_drainers
)
3007 wq
->flags
&= ~__WQ_DRAINING
;
3008 mutex_unlock(&wq
->mutex
);
3010 EXPORT_SYMBOL_GPL(drain_workqueue
);
3012 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
3015 struct worker
*worker
= NULL
;
3016 struct worker_pool
*pool
;
3017 struct pool_workqueue
*pwq
;
3022 pool
= get_work_pool(work
);
3028 raw_spin_lock_irq(&pool
->lock
);
3029 /* see the comment in try_to_grab_pending() with the same code */
3030 pwq
= get_work_pwq(work
);
3032 if (unlikely(pwq
->pool
!= pool
))
3035 worker
= find_worker_executing_work(pool
, work
);
3038 pwq
= worker
->current_pwq
;
3041 check_flush_dependency(pwq
->wq
, work
);
3043 insert_wq_barrier(pwq
, barr
, work
, worker
);
3044 raw_spin_unlock_irq(&pool
->lock
);
3047 * Force a lock recursion deadlock when using flush_work() inside a
3048 * single-threaded or rescuer equipped workqueue.
3050 * For single threaded workqueues the deadlock happens when the work
3051 * is after the work issuing the flush_work(). For rescuer equipped
3052 * workqueues the deadlock happens when the rescuer stalls, blocking
3056 (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)) {
3057 lock_map_acquire(&pwq
->wq
->lockdep_map
);
3058 lock_map_release(&pwq
->wq
->lockdep_map
);
3063 raw_spin_unlock_irq(&pool
->lock
);
3068 static bool __flush_work(struct work_struct
*work
, bool from_cancel
)
3070 struct wq_barrier barr
;
3072 if (WARN_ON(!wq_online
))
3075 if (WARN_ON(!work
->func
))
3079 lock_map_acquire(&work
->lockdep_map
);
3080 lock_map_release(&work
->lockdep_map
);
3083 if (start_flush_work(work
, &barr
, from_cancel
)) {
3084 wait_for_completion(&barr
.done
);
3085 destroy_work_on_stack(&barr
.work
);
3093 * flush_work - wait for a work to finish executing the last queueing instance
3094 * @work: the work to flush
3096 * Wait until @work has finished execution. @work is guaranteed to be idle
3097 * on return if it hasn't been requeued since flush started.
3100 * %true if flush_work() waited for the work to finish execution,
3101 * %false if it was already idle.
3103 bool flush_work(struct work_struct
*work
)
3105 return __flush_work(work
, false);
3107 EXPORT_SYMBOL_GPL(flush_work
);
3110 wait_queue_entry_t wait
;
3111 struct work_struct
*work
;
3114 static int cwt_wakefn(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
3116 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
3118 if (cwait
->work
!= key
)
3120 return autoremove_wake_function(wait
, mode
, sync
, key
);
3123 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
3125 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
3126 unsigned long flags
;
3130 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3132 * If someone else is already canceling, wait for it to
3133 * finish. flush_work() doesn't work for PREEMPT_NONE
3134 * because we may get scheduled between @work's completion
3135 * and the other canceling task resuming and clearing
3136 * CANCELING - flush_work() will return false immediately
3137 * as @work is no longer busy, try_to_grab_pending() will
3138 * return -ENOENT as @work is still being canceled and the
3139 * other canceling task won't be able to clear CANCELING as
3140 * we're hogging the CPU.
3142 * Let's wait for completion using a waitqueue. As this
3143 * may lead to the thundering herd problem, use a custom
3144 * wake function which matches @work along with exclusive
3147 if (unlikely(ret
== -ENOENT
)) {
3148 struct cwt_wait cwait
;
3150 init_wait(&cwait
.wait
);
3151 cwait
.wait
.func
= cwt_wakefn
;
3154 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
3155 TASK_UNINTERRUPTIBLE
);
3156 if (work_is_canceling(work
))
3158 finish_wait(&cancel_waitq
, &cwait
.wait
);
3160 } while (unlikely(ret
< 0));
3162 /* tell other tasks trying to grab @work to back off */
3163 mark_work_canceling(work
);
3164 local_irq_restore(flags
);
3167 * This allows canceling during early boot. We know that @work
3171 __flush_work(work
, true);
3173 clear_work_data(work
);
3176 * Paired with prepare_to_wait() above so that either
3177 * waitqueue_active() is visible here or !work_is_canceling() is
3181 if (waitqueue_active(&cancel_waitq
))
3182 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
3188 * cancel_work_sync - cancel a work and wait for it to finish
3189 * @work: the work to cancel
3191 * Cancel @work and wait for its execution to finish. This function
3192 * can be used even if the work re-queues itself or migrates to
3193 * another workqueue. On return from this function, @work is
3194 * guaranteed to be not pending or executing on any CPU.
3196 * cancel_work_sync(&delayed_work->work) must not be used for
3197 * delayed_work's. Use cancel_delayed_work_sync() instead.
3199 * The caller must ensure that the workqueue on which @work was last
3200 * queued can't be destroyed before this function returns.
3203 * %true if @work was pending, %false otherwise.
3205 bool cancel_work_sync(struct work_struct
*work
)
3207 return __cancel_work_timer(work
, false);
3209 EXPORT_SYMBOL_GPL(cancel_work_sync
);
3212 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3213 * @dwork: the delayed work to flush
3215 * Delayed timer is cancelled and the pending work is queued for
3216 * immediate execution. Like flush_work(), this function only
3217 * considers the last queueing instance of @dwork.
3220 * %true if flush_work() waited for the work to finish execution,
3221 * %false if it was already idle.
3223 bool flush_delayed_work(struct delayed_work
*dwork
)
3225 local_irq_disable();
3226 if (del_timer_sync(&dwork
->timer
))
3227 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3229 return flush_work(&dwork
->work
);
3231 EXPORT_SYMBOL(flush_delayed_work
);
3234 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3235 * @rwork: the rcu work to flush
3238 * %true if flush_rcu_work() waited for the work to finish execution,
3239 * %false if it was already idle.
3241 bool flush_rcu_work(struct rcu_work
*rwork
)
3243 if (test_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&rwork
->work
))) {
3245 flush_work(&rwork
->work
);
3248 return flush_work(&rwork
->work
);
3251 EXPORT_SYMBOL(flush_rcu_work
);
3253 static bool __cancel_work(struct work_struct
*work
, bool is_dwork
)
3255 unsigned long flags
;
3259 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3260 } while (unlikely(ret
== -EAGAIN
));
3262 if (unlikely(ret
< 0))
3265 set_work_pool_and_clear_pending(work
, get_work_pool_id(work
));
3266 local_irq_restore(flags
);
3271 * cancel_delayed_work - cancel a delayed work
3272 * @dwork: delayed_work to cancel
3274 * Kill off a pending delayed_work.
3276 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3280 * The work callback function may still be running on return, unless
3281 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3282 * use cancel_delayed_work_sync() to wait on it.
3284 * This function is safe to call from any context including IRQ handler.
3286 bool cancel_delayed_work(struct delayed_work
*dwork
)
3288 return __cancel_work(&dwork
->work
, true);
3290 EXPORT_SYMBOL(cancel_delayed_work
);
3293 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3294 * @dwork: the delayed work cancel
3296 * This is cancel_work_sync() for delayed works.
3299 * %true if @dwork was pending, %false otherwise.
3301 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3303 return __cancel_work_timer(&dwork
->work
, true);
3305 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3308 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3309 * @func: the function to call
3311 * schedule_on_each_cpu() executes @func on each online CPU using the
3312 * system workqueue and blocks until all CPUs have completed.
3313 * schedule_on_each_cpu() is very slow.
3316 * 0 on success, -errno on failure.
3318 int schedule_on_each_cpu(work_func_t func
)
3321 struct work_struct __percpu
*works
;
3323 works
= alloc_percpu(struct work_struct
);
3329 for_each_online_cpu(cpu
) {
3330 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3332 INIT_WORK(work
, func
);
3333 schedule_work_on(cpu
, work
);
3336 for_each_online_cpu(cpu
)
3337 flush_work(per_cpu_ptr(works
, cpu
));
3345 * execute_in_process_context - reliably execute the routine with user context
3346 * @fn: the function to execute
3347 * @ew: guaranteed storage for the execute work structure (must
3348 * be available when the work executes)
3350 * Executes the function immediately if process context is available,
3351 * otherwise schedules the function for delayed execution.
3353 * Return: 0 - function was executed
3354 * 1 - function was scheduled for execution
3356 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3358 if (!in_interrupt()) {
3363 INIT_WORK(&ew
->work
, fn
);
3364 schedule_work(&ew
->work
);
3368 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3371 * free_workqueue_attrs - free a workqueue_attrs
3372 * @attrs: workqueue_attrs to free
3374 * Undo alloc_workqueue_attrs().
3376 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3379 free_cpumask_var(attrs
->cpumask
);
3385 * alloc_workqueue_attrs - allocate a workqueue_attrs
3387 * Allocate a new workqueue_attrs, initialize with default settings and
3390 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3392 struct workqueue_attrs
*alloc_workqueue_attrs(void)
3394 struct workqueue_attrs
*attrs
;
3396 attrs
= kzalloc(sizeof(*attrs
), GFP_KERNEL
);
3399 if (!alloc_cpumask_var(&attrs
->cpumask
, GFP_KERNEL
))
3402 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3405 free_workqueue_attrs(attrs
);
3409 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3410 const struct workqueue_attrs
*from
)
3412 to
->nice
= from
->nice
;
3413 cpumask_copy(to
->cpumask
, from
->cpumask
);
3415 * Unlike hash and equality test, this function doesn't ignore
3416 * ->no_numa as it is used for both pool and wq attrs. Instead,
3417 * get_unbound_pool() explicitly clears ->no_numa after copying.
3419 to
->no_numa
= from
->no_numa
;
3422 /* hash value of the content of @attr */
3423 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3427 hash
= jhash_1word(attrs
->nice
, hash
);
3428 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3429 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3433 /* content equality test */
3434 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3435 const struct workqueue_attrs
*b
)
3437 if (a
->nice
!= b
->nice
)
3439 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3445 * init_worker_pool - initialize a newly zalloc'd worker_pool
3446 * @pool: worker_pool to initialize
3448 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3450 * Return: 0 on success, -errno on failure. Even on failure, all fields
3451 * inside @pool proper are initialized and put_unbound_pool() can be called
3452 * on @pool safely to release it.
3454 static int init_worker_pool(struct worker_pool
*pool
)
3456 raw_spin_lock_init(&pool
->lock
);
3459 pool
->node
= NUMA_NO_NODE
;
3460 pool
->flags
|= POOL_DISASSOCIATED
;
3461 pool
->watchdog_ts
= jiffies
;
3462 INIT_LIST_HEAD(&pool
->worklist
);
3463 INIT_LIST_HEAD(&pool
->idle_list
);
3464 hash_init(pool
->busy_hash
);
3466 timer_setup(&pool
->idle_timer
, idle_worker_timeout
, TIMER_DEFERRABLE
);
3468 timer_setup(&pool
->mayday_timer
, pool_mayday_timeout
, 0);
3470 INIT_LIST_HEAD(&pool
->workers
);
3472 ida_init(&pool
->worker_ida
);
3473 INIT_HLIST_NODE(&pool
->hash_node
);
3476 /* shouldn't fail above this point */
3477 pool
->attrs
= alloc_workqueue_attrs();
3483 #ifdef CONFIG_LOCKDEP
3484 static void wq_init_lockdep(struct workqueue_struct
*wq
)
3488 lockdep_register_key(&wq
->key
);
3489 lock_name
= kasprintf(GFP_KERNEL
, "%s%s", "(wq_completion)", wq
->name
);
3491 lock_name
= wq
->name
;
3493 wq
->lock_name
= lock_name
;
3494 lockdep_init_map(&wq
->lockdep_map
, lock_name
, &wq
->key
, 0);
3497 static void wq_unregister_lockdep(struct workqueue_struct
*wq
)
3499 lockdep_unregister_key(&wq
->key
);
3502 static void wq_free_lockdep(struct workqueue_struct
*wq
)
3504 if (wq
->lock_name
!= wq
->name
)
3505 kfree(wq
->lock_name
);
3508 static void wq_init_lockdep(struct workqueue_struct
*wq
)
3512 static void wq_unregister_lockdep(struct workqueue_struct
*wq
)
3516 static void wq_free_lockdep(struct workqueue_struct
*wq
)
3521 static void rcu_free_wq(struct rcu_head
*rcu
)
3523 struct workqueue_struct
*wq
=
3524 container_of(rcu
, struct workqueue_struct
, rcu
);
3526 wq_free_lockdep(wq
);
3528 if (!(wq
->flags
& WQ_UNBOUND
))
3529 free_percpu(wq
->cpu_pwqs
);
3531 free_workqueue_attrs(wq
->unbound_attrs
);
3536 static void rcu_free_pool(struct rcu_head
*rcu
)
3538 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3540 ida_destroy(&pool
->worker_ida
);
3541 free_workqueue_attrs(pool
->attrs
);
3545 /* This returns with the lock held on success (pool manager is inactive). */
3546 static bool wq_manager_inactive(struct worker_pool
*pool
)
3548 raw_spin_lock_irq(&pool
->lock
);
3550 if (pool
->flags
& POOL_MANAGER_ACTIVE
) {
3551 raw_spin_unlock_irq(&pool
->lock
);
3558 * put_unbound_pool - put a worker_pool
3559 * @pool: worker_pool to put
3561 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3562 * safe manner. get_unbound_pool() calls this function on its failure path
3563 * and this function should be able to release pools which went through,
3564 * successfully or not, init_worker_pool().
3566 * Should be called with wq_pool_mutex held.
3568 static void put_unbound_pool(struct worker_pool
*pool
)
3570 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3571 struct worker
*worker
;
3573 lockdep_assert_held(&wq_pool_mutex
);
3579 if (WARN_ON(!(pool
->cpu
< 0)) ||
3580 WARN_ON(!list_empty(&pool
->worklist
)))
3583 /* release id and unhash */
3585 idr_remove(&worker_pool_idr
, pool
->id
);
3586 hash_del(&pool
->hash_node
);
3589 * Become the manager and destroy all workers. This prevents
3590 * @pool's workers from blocking on attach_mutex. We're the last
3591 * manager and @pool gets freed with the flag set.
3592 * Because of how wq_manager_inactive() works, we will hold the
3593 * spinlock after a successful wait.
3595 rcuwait_wait_event(&manager_wait
, wq_manager_inactive(pool
),
3596 TASK_UNINTERRUPTIBLE
);
3597 pool
->flags
|= POOL_MANAGER_ACTIVE
;
3599 while ((worker
= first_idle_worker(pool
)))
3600 destroy_worker(worker
);
3601 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3602 raw_spin_unlock_irq(&pool
->lock
);
3604 mutex_lock(&wq_pool_attach_mutex
);
3605 if (!list_empty(&pool
->workers
))
3606 pool
->detach_completion
= &detach_completion
;
3607 mutex_unlock(&wq_pool_attach_mutex
);
3609 if (pool
->detach_completion
)
3610 wait_for_completion(pool
->detach_completion
);
3612 /* shut down the timers */
3613 del_timer_sync(&pool
->idle_timer
);
3614 del_timer_sync(&pool
->mayday_timer
);
3616 /* RCU protected to allow dereferences from get_work_pool() */
3617 call_rcu(&pool
->rcu
, rcu_free_pool
);
3621 * get_unbound_pool - get a worker_pool with the specified attributes
3622 * @attrs: the attributes of the worker_pool to get
3624 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3625 * reference count and return it. If there already is a matching
3626 * worker_pool, it will be used; otherwise, this function attempts to
3629 * Should be called with wq_pool_mutex held.
3631 * Return: On success, a worker_pool with the same attributes as @attrs.
3632 * On failure, %NULL.
3634 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3636 u32 hash
= wqattrs_hash(attrs
);
3637 struct worker_pool
*pool
;
3639 int target_node
= NUMA_NO_NODE
;
3641 lockdep_assert_held(&wq_pool_mutex
);
3643 /* do we already have a matching pool? */
3644 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3645 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3651 /* if cpumask is contained inside a NUMA node, we belong to that node */
3652 if (wq_numa_enabled
) {
3653 for_each_node(node
) {
3654 if (cpumask_subset(attrs
->cpumask
,
3655 wq_numa_possible_cpumask
[node
])) {
3662 /* nope, create a new one */
3663 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3664 if (!pool
|| init_worker_pool(pool
) < 0)
3667 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3668 copy_workqueue_attrs(pool
->attrs
, attrs
);
3669 pool
->node
= target_node
;
3672 * no_numa isn't a worker_pool attribute, always clear it. See
3673 * 'struct workqueue_attrs' comments for detail.
3675 pool
->attrs
->no_numa
= false;
3677 if (worker_pool_assign_id(pool
) < 0)
3680 /* create and start the initial worker */
3681 if (wq_online
&& !create_worker(pool
))
3685 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3690 put_unbound_pool(pool
);
3694 static void rcu_free_pwq(struct rcu_head
*rcu
)
3696 kmem_cache_free(pwq_cache
,
3697 container_of(rcu
, struct pool_workqueue
, rcu
));
3701 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3702 * and needs to be destroyed.
3704 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3706 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3707 unbound_release_work
);
3708 struct workqueue_struct
*wq
= pwq
->wq
;
3709 struct worker_pool
*pool
= pwq
->pool
;
3710 bool is_last
= false;
3713 * when @pwq is not linked, it doesn't hold any reference to the
3714 * @wq, and @wq is invalid to access.
3716 if (!list_empty(&pwq
->pwqs_node
)) {
3717 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3720 mutex_lock(&wq
->mutex
);
3721 list_del_rcu(&pwq
->pwqs_node
);
3722 is_last
= list_empty(&wq
->pwqs
);
3723 mutex_unlock(&wq
->mutex
);
3726 mutex_lock(&wq_pool_mutex
);
3727 put_unbound_pool(pool
);
3728 mutex_unlock(&wq_pool_mutex
);
3730 call_rcu(&pwq
->rcu
, rcu_free_pwq
);
3733 * If we're the last pwq going away, @wq is already dead and no one
3734 * is gonna access it anymore. Schedule RCU free.
3737 wq_unregister_lockdep(wq
);
3738 call_rcu(&wq
->rcu
, rcu_free_wq
);
3743 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3744 * @pwq: target pool_workqueue
3746 * If @pwq isn't freezing, set @pwq->max_active to the associated
3747 * workqueue's saved_max_active and activate inactive work items
3748 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3750 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3752 struct workqueue_struct
*wq
= pwq
->wq
;
3753 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3754 unsigned long flags
;
3756 /* for @wq->saved_max_active */
3757 lockdep_assert_held(&wq
->mutex
);
3759 /* fast exit for non-freezable wqs */
3760 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3763 /* this function can be called during early boot w/ irq disabled */
3764 raw_spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
3767 * During [un]freezing, the caller is responsible for ensuring that
3768 * this function is called at least once after @workqueue_freezing
3769 * is updated and visible.
3771 if (!freezable
|| !workqueue_freezing
) {
3774 pwq
->max_active
= wq
->saved_max_active
;
3776 while (!list_empty(&pwq
->inactive_works
) &&
3777 pwq
->nr_active
< pwq
->max_active
) {
3778 pwq_activate_first_inactive(pwq
);
3783 * Need to kick a worker after thawed or an unbound wq's
3784 * max_active is bumped. In realtime scenarios, always kicking a
3785 * worker will cause interference on the isolated cpu cores, so
3786 * let's kick iff work items were activated.
3789 wake_up_worker(pwq
->pool
);
3791 pwq
->max_active
= 0;
3794 raw_spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
3797 /* initialize newly allocated @pwq which is associated with @wq and @pool */
3798 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3799 struct worker_pool
*pool
)
3801 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3803 memset(pwq
, 0, sizeof(*pwq
));
3807 pwq
->flush_color
= -1;
3809 INIT_LIST_HEAD(&pwq
->inactive_works
);
3810 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3811 INIT_LIST_HEAD(&pwq
->mayday_node
);
3812 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3815 /* sync @pwq with the current state of its associated wq and link it */
3816 static void link_pwq(struct pool_workqueue
*pwq
)
3818 struct workqueue_struct
*wq
= pwq
->wq
;
3820 lockdep_assert_held(&wq
->mutex
);
3822 /* may be called multiple times, ignore if already linked */
3823 if (!list_empty(&pwq
->pwqs_node
))
3826 /* set the matching work_color */
3827 pwq
->work_color
= wq
->work_color
;
3829 /* sync max_active to the current setting */
3830 pwq_adjust_max_active(pwq
);
3833 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3836 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3837 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3838 const struct workqueue_attrs
*attrs
)
3840 struct worker_pool
*pool
;
3841 struct pool_workqueue
*pwq
;
3843 lockdep_assert_held(&wq_pool_mutex
);
3845 pool
= get_unbound_pool(attrs
);
3849 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3851 put_unbound_pool(pool
);
3855 init_pwq(pwq
, wq
, pool
);
3860 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3861 * @attrs: the wq_attrs of the default pwq of the target workqueue
3862 * @node: the target NUMA node
3863 * @cpu_going_down: if >= 0, the CPU to consider as offline
3864 * @cpumask: outarg, the resulting cpumask
3866 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3867 * @cpu_going_down is >= 0, that cpu is considered offline during
3868 * calculation. The result is stored in @cpumask.
3870 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3871 * enabled and @node has online CPUs requested by @attrs, the returned
3872 * cpumask is the intersection of the possible CPUs of @node and
3875 * The caller is responsible for ensuring that the cpumask of @node stays
3878 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3881 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3882 int cpu_going_down
, cpumask_t
*cpumask
)
3884 if (!wq_numa_enabled
|| attrs
->no_numa
)
3887 /* does @node have any online CPUs @attrs wants? */
3888 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3889 if (cpu_going_down
>= 0)
3890 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3892 if (cpumask_empty(cpumask
))
3895 /* yeap, return possible CPUs in @node that @attrs wants */
3896 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3898 if (cpumask_empty(cpumask
)) {
3899 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3900 "possible intersect\n");
3904 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3907 cpumask_copy(cpumask
, attrs
->cpumask
);
3911 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3912 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3914 struct pool_workqueue
*pwq
)
3916 struct pool_workqueue
*old_pwq
;
3918 lockdep_assert_held(&wq_pool_mutex
);
3919 lockdep_assert_held(&wq
->mutex
);
3921 /* link_pwq() can handle duplicate calls */
3924 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3925 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3929 /* context to store the prepared attrs & pwqs before applying */
3930 struct apply_wqattrs_ctx
{
3931 struct workqueue_struct
*wq
; /* target workqueue */
3932 struct workqueue_attrs
*attrs
; /* attrs to apply */
3933 struct list_head list
; /* queued for batching commit */
3934 struct pool_workqueue
*dfl_pwq
;
3935 struct pool_workqueue
*pwq_tbl
[];
3938 /* free the resources after success or abort */
3939 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3945 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3946 put_pwq_unlocked(ctx
->dfl_pwq
);
3948 free_workqueue_attrs(ctx
->attrs
);
3954 /* allocate the attrs and pwqs for later installation */
3955 static struct apply_wqattrs_ctx
*
3956 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3957 const struct workqueue_attrs
*attrs
)
3959 struct apply_wqattrs_ctx
*ctx
;
3960 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3963 lockdep_assert_held(&wq_pool_mutex
);
3965 ctx
= kzalloc(struct_size(ctx
, pwq_tbl
, nr_node_ids
), GFP_KERNEL
);
3967 new_attrs
= alloc_workqueue_attrs();
3968 tmp_attrs
= alloc_workqueue_attrs();
3969 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3973 * Calculate the attrs of the default pwq.
3974 * If the user configured cpumask doesn't overlap with the
3975 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3977 copy_workqueue_attrs(new_attrs
, attrs
);
3978 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3979 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3980 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3983 * We may create multiple pwqs with differing cpumasks. Make a
3984 * copy of @new_attrs which will be modified and used to obtain
3987 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3990 * If something goes wrong during CPU up/down, we'll fall back to
3991 * the default pwq covering whole @attrs->cpumask. Always create
3992 * it even if we don't use it immediately.
3994 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3998 for_each_node(node
) {
3999 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
4000 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
4001 if (!ctx
->pwq_tbl
[node
])
4004 ctx
->dfl_pwq
->refcnt
++;
4005 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
4009 /* save the user configured attrs and sanitize it. */
4010 copy_workqueue_attrs(new_attrs
, attrs
);
4011 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
4012 ctx
->attrs
= new_attrs
;
4015 free_workqueue_attrs(tmp_attrs
);
4019 free_workqueue_attrs(tmp_attrs
);
4020 free_workqueue_attrs(new_attrs
);
4021 apply_wqattrs_cleanup(ctx
);
4025 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4026 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
4030 /* all pwqs have been created successfully, let's install'em */
4031 mutex_lock(&ctx
->wq
->mutex
);
4033 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
4035 /* save the previous pwq and install the new one */
4037 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
4038 ctx
->pwq_tbl
[node
]);
4040 /* @dfl_pwq might not have been used, ensure it's linked */
4041 link_pwq(ctx
->dfl_pwq
);
4042 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
4044 mutex_unlock(&ctx
->wq
->mutex
);
4047 static void apply_wqattrs_lock(void)
4049 /* CPUs should stay stable across pwq creations and installations */
4051 mutex_lock(&wq_pool_mutex
);
4054 static void apply_wqattrs_unlock(void)
4056 mutex_unlock(&wq_pool_mutex
);
4060 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
4061 const struct workqueue_attrs
*attrs
)
4063 struct apply_wqattrs_ctx
*ctx
;
4065 /* only unbound workqueues can change attributes */
4066 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
4069 /* creating multiple pwqs breaks ordering guarantee */
4070 if (!list_empty(&wq
->pwqs
)) {
4071 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4074 wq
->flags
&= ~__WQ_ORDERED
;
4077 ctx
= apply_wqattrs_prepare(wq
, attrs
);
4081 /* the ctx has been prepared successfully, let's commit it */
4082 apply_wqattrs_commit(ctx
);
4083 apply_wqattrs_cleanup(ctx
);
4089 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4090 * @wq: the target workqueue
4091 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4093 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4094 * machines, this function maps a separate pwq to each NUMA node with
4095 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4096 * NUMA node it was issued on. Older pwqs are released as in-flight work
4097 * items finish. Note that a work item which repeatedly requeues itself
4098 * back-to-back will stay on its current pwq.
4100 * Performs GFP_KERNEL allocations.
4102 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4104 * Return: 0 on success and -errno on failure.
4106 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
4107 const struct workqueue_attrs
*attrs
)
4111 lockdep_assert_cpus_held();
4113 mutex_lock(&wq_pool_mutex
);
4114 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
4115 mutex_unlock(&wq_pool_mutex
);
4121 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4122 * @wq: the target workqueue
4123 * @cpu: the CPU coming up or going down
4124 * @online: whether @cpu is coming up or going down
4126 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4127 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4130 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4131 * falls back to @wq->dfl_pwq which may not be optimal but is always
4134 * Note that when the last allowed CPU of a NUMA node goes offline for a
4135 * workqueue with a cpumask spanning multiple nodes, the workers which were
4136 * already executing the work items for the workqueue will lose their CPU
4137 * affinity and may execute on any CPU. This is similar to how per-cpu
4138 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4139 * affinity, it's the user's responsibility to flush the work item from
4142 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
4145 int node
= cpu_to_node(cpu
);
4146 int cpu_off
= online
? -1 : cpu
;
4147 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
4148 struct workqueue_attrs
*target_attrs
;
4151 lockdep_assert_held(&wq_pool_mutex
);
4153 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
4154 wq
->unbound_attrs
->no_numa
)
4158 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4159 * Let's use a preallocated one. The following buf is protected by
4160 * CPU hotplug exclusion.
4162 target_attrs
= wq_update_unbound_numa_attrs_buf
;
4163 cpumask
= target_attrs
->cpumask
;
4165 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4166 pwq
= unbound_pwq_by_node(wq
, node
);
4169 * Let's determine what needs to be done. If the target cpumask is
4170 * different from the default pwq's, we need to compare it to @pwq's
4171 * and create a new one if they don't match. If the target cpumask
4172 * equals the default pwq's, the default pwq should be used.
4174 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
4175 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4181 /* create a new pwq */
4182 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4184 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4189 /* Install the new pwq. */
4190 mutex_lock(&wq
->mutex
);
4191 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4195 mutex_lock(&wq
->mutex
);
4196 raw_spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4197 get_pwq(wq
->dfl_pwq
);
4198 raw_spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4199 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4201 mutex_unlock(&wq
->mutex
);
4202 put_pwq_unlocked(old_pwq
);
4205 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4207 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4210 if (!(wq
->flags
& WQ_UNBOUND
)) {
4211 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4215 for_each_possible_cpu(cpu
) {
4216 struct pool_workqueue
*pwq
=
4217 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4218 struct worker_pool
*cpu_pools
=
4219 per_cpu(cpu_worker_pools
, cpu
);
4221 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4223 mutex_lock(&wq
->mutex
);
4225 mutex_unlock(&wq
->mutex
);
4231 if (wq
->flags
& __WQ_ORDERED
) {
4232 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4233 /* there should only be single pwq for ordering guarantee */
4234 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4235 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4236 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4238 ret
= apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4245 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4248 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4250 if (max_active
< 1 || max_active
> lim
)
4251 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4252 max_active
, name
, 1, lim
);
4254 return clamp_val(max_active
, 1, lim
);
4258 * Workqueues which may be used during memory reclaim should have a rescuer
4259 * to guarantee forward progress.
4261 static int init_rescuer(struct workqueue_struct
*wq
)
4263 struct worker
*rescuer
;
4266 if (!(wq
->flags
& WQ_MEM_RECLAIM
))
4269 rescuer
= alloc_worker(NUMA_NO_NODE
);
4273 rescuer
->rescue_wq
= wq
;
4274 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s", wq
->name
);
4275 if (IS_ERR(rescuer
->task
)) {
4276 ret
= PTR_ERR(rescuer
->task
);
4281 wq
->rescuer
= rescuer
;
4282 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
4283 wake_up_process(rescuer
->task
);
4289 struct workqueue_struct
*alloc_workqueue(const char *fmt
,
4291 int max_active
, ...)
4293 size_t tbl_size
= 0;
4295 struct workqueue_struct
*wq
;
4296 struct pool_workqueue
*pwq
;
4299 * Unbound && max_active == 1 used to imply ordered, which is no
4300 * longer the case on NUMA machines due to per-node pools. While
4301 * alloc_ordered_workqueue() is the right way to create an ordered
4302 * workqueue, keep the previous behavior to avoid subtle breakages
4305 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
4306 flags
|= __WQ_ORDERED
;
4308 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4309 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4310 flags
|= WQ_UNBOUND
;
4312 /* allocate wq and format name */
4313 if (flags
& WQ_UNBOUND
)
4314 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
4316 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4320 if (flags
& WQ_UNBOUND
) {
4321 wq
->unbound_attrs
= alloc_workqueue_attrs();
4322 if (!wq
->unbound_attrs
)
4326 va_start(args
, max_active
);
4327 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4330 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4331 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4335 wq
->saved_max_active
= max_active
;
4336 mutex_init(&wq
->mutex
);
4337 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4338 INIT_LIST_HEAD(&wq
->pwqs
);
4339 INIT_LIST_HEAD(&wq
->flusher_queue
);
4340 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4341 INIT_LIST_HEAD(&wq
->maydays
);
4343 wq_init_lockdep(wq
);
4344 INIT_LIST_HEAD(&wq
->list
);
4346 if (alloc_and_link_pwqs(wq
) < 0)
4347 goto err_unreg_lockdep
;
4349 if (wq_online
&& init_rescuer(wq
) < 0)
4352 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4356 * wq_pool_mutex protects global freeze state and workqueues list.
4357 * Grab it, adjust max_active and add the new @wq to workqueues
4360 mutex_lock(&wq_pool_mutex
);
4362 mutex_lock(&wq
->mutex
);
4363 for_each_pwq(pwq
, wq
)
4364 pwq_adjust_max_active(pwq
);
4365 mutex_unlock(&wq
->mutex
);
4367 list_add_tail_rcu(&wq
->list
, &workqueues
);
4369 mutex_unlock(&wq_pool_mutex
);
4374 wq_unregister_lockdep(wq
);
4375 wq_free_lockdep(wq
);
4377 free_workqueue_attrs(wq
->unbound_attrs
);
4381 destroy_workqueue(wq
);
4384 EXPORT_SYMBOL_GPL(alloc_workqueue
);
4386 static bool pwq_busy(struct pool_workqueue
*pwq
)
4390 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
4391 if (pwq
->nr_in_flight
[i
])
4394 if ((pwq
!= pwq
->wq
->dfl_pwq
) && (pwq
->refcnt
> 1))
4396 if (pwq
->nr_active
|| !list_empty(&pwq
->inactive_works
))
4403 * destroy_workqueue - safely terminate a workqueue
4404 * @wq: target workqueue
4406 * Safely destroy a workqueue. All work currently pending will be done first.
4408 void destroy_workqueue(struct workqueue_struct
*wq
)
4410 struct pool_workqueue
*pwq
;
4414 * Remove it from sysfs first so that sanity check failure doesn't
4415 * lead to sysfs name conflicts.
4417 workqueue_sysfs_unregister(wq
);
4419 /* drain it before proceeding with destruction */
4420 drain_workqueue(wq
);
4422 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4424 struct worker
*rescuer
= wq
->rescuer
;
4426 /* this prevents new queueing */
4427 raw_spin_lock_irq(&wq_mayday_lock
);
4429 raw_spin_unlock_irq(&wq_mayday_lock
);
4431 /* rescuer will empty maydays list before exiting */
4432 kthread_stop(rescuer
->task
);
4437 * Sanity checks - grab all the locks so that we wait for all
4438 * in-flight operations which may do put_pwq().
4440 mutex_lock(&wq_pool_mutex
);
4441 mutex_lock(&wq
->mutex
);
4442 for_each_pwq(pwq
, wq
) {
4443 raw_spin_lock_irq(&pwq
->pool
->lock
);
4444 if (WARN_ON(pwq_busy(pwq
))) {
4445 pr_warn("%s: %s has the following busy pwq\n",
4446 __func__
, wq
->name
);
4448 raw_spin_unlock_irq(&pwq
->pool
->lock
);
4449 mutex_unlock(&wq
->mutex
);
4450 mutex_unlock(&wq_pool_mutex
);
4451 show_one_workqueue(wq
);
4454 raw_spin_unlock_irq(&pwq
->pool
->lock
);
4456 mutex_unlock(&wq
->mutex
);
4459 * wq list is used to freeze wq, remove from list after
4460 * flushing is complete in case freeze races us.
4462 list_del_rcu(&wq
->list
);
4463 mutex_unlock(&wq_pool_mutex
);
4465 if (!(wq
->flags
& WQ_UNBOUND
)) {
4466 wq_unregister_lockdep(wq
);
4468 * The base ref is never dropped on per-cpu pwqs. Directly
4469 * schedule RCU free.
4471 call_rcu(&wq
->rcu
, rcu_free_wq
);
4474 * We're the sole accessor of @wq at this point. Directly
4475 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4476 * @wq will be freed when the last pwq is released.
4478 for_each_node(node
) {
4479 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4480 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4481 put_pwq_unlocked(pwq
);
4485 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4486 * put. Don't access it afterwards.
4490 put_pwq_unlocked(pwq
);
4493 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4496 * workqueue_set_max_active - adjust max_active of a workqueue
4497 * @wq: target workqueue
4498 * @max_active: new max_active value.
4500 * Set max_active of @wq to @max_active.
4503 * Don't call from IRQ context.
4505 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4507 struct pool_workqueue
*pwq
;
4509 /* disallow meddling with max_active for ordered workqueues */
4510 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4513 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4515 mutex_lock(&wq
->mutex
);
4517 wq
->flags
&= ~__WQ_ORDERED
;
4518 wq
->saved_max_active
= max_active
;
4520 for_each_pwq(pwq
, wq
)
4521 pwq_adjust_max_active(pwq
);
4523 mutex_unlock(&wq
->mutex
);
4525 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4528 * current_work - retrieve %current task's work struct
4530 * Determine if %current task is a workqueue worker and what it's working on.
4531 * Useful to find out the context that the %current task is running in.
4533 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4535 struct work_struct
*current_work(void)
4537 struct worker
*worker
= current_wq_worker();
4539 return worker
? worker
->current_work
: NULL
;
4541 EXPORT_SYMBOL(current_work
);
4544 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4546 * Determine whether %current is a workqueue rescuer. Can be used from
4547 * work functions to determine whether it's being run off the rescuer task.
4549 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4551 bool current_is_workqueue_rescuer(void)
4553 struct worker
*worker
= current_wq_worker();
4555 return worker
&& worker
->rescue_wq
;
4559 * workqueue_congested - test whether a workqueue is congested
4560 * @cpu: CPU in question
4561 * @wq: target workqueue
4563 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4564 * no synchronization around this function and the test result is
4565 * unreliable and only useful as advisory hints or for debugging.
4567 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4568 * Note that both per-cpu and unbound workqueues may be associated with
4569 * multiple pool_workqueues which have separate congested states. A
4570 * workqueue being congested on one CPU doesn't mean the workqueue is also
4571 * contested on other CPUs / NUMA nodes.
4574 * %true if congested, %false otherwise.
4576 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4578 struct pool_workqueue
*pwq
;
4584 if (cpu
== WORK_CPU_UNBOUND
)
4585 cpu
= smp_processor_id();
4587 if (!(wq
->flags
& WQ_UNBOUND
))
4588 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4590 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4592 ret
= !list_empty(&pwq
->inactive_works
);
4598 EXPORT_SYMBOL_GPL(workqueue_congested
);
4601 * work_busy - test whether a work is currently pending or running
4602 * @work: the work to be tested
4604 * Test whether @work is currently pending or running. There is no
4605 * synchronization around this function and the test result is
4606 * unreliable and only useful as advisory hints or for debugging.
4609 * OR'd bitmask of WORK_BUSY_* bits.
4611 unsigned int work_busy(struct work_struct
*work
)
4613 struct worker_pool
*pool
;
4614 unsigned long flags
;
4615 unsigned int ret
= 0;
4617 if (work_pending(work
))
4618 ret
|= WORK_BUSY_PENDING
;
4621 pool
= get_work_pool(work
);
4623 raw_spin_lock_irqsave(&pool
->lock
, flags
);
4624 if (find_worker_executing_work(pool
, work
))
4625 ret
|= WORK_BUSY_RUNNING
;
4626 raw_spin_unlock_irqrestore(&pool
->lock
, flags
);
4632 EXPORT_SYMBOL_GPL(work_busy
);
4635 * set_worker_desc - set description for the current work item
4636 * @fmt: printf-style format string
4637 * @...: arguments for the format string
4639 * This function can be called by a running work function to describe what
4640 * the work item is about. If the worker task gets dumped, this
4641 * information will be printed out together to help debugging. The
4642 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4644 void set_worker_desc(const char *fmt
, ...)
4646 struct worker
*worker
= current_wq_worker();
4650 va_start(args
, fmt
);
4651 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4655 EXPORT_SYMBOL_GPL(set_worker_desc
);
4658 * print_worker_info - print out worker information and description
4659 * @log_lvl: the log level to use when printing
4660 * @task: target task
4662 * If @task is a worker and currently executing a work item, print out the
4663 * name of the workqueue being serviced and worker description set with
4664 * set_worker_desc() by the currently executing work item.
4666 * This function can be safely called on any task as long as the
4667 * task_struct itself is accessible. While safe, this function isn't
4668 * synchronized and may print out mixups or garbages of limited length.
4670 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4672 work_func_t
*fn
= NULL
;
4673 char name
[WQ_NAME_LEN
] = { };
4674 char desc
[WORKER_DESC_LEN
] = { };
4675 struct pool_workqueue
*pwq
= NULL
;
4676 struct workqueue_struct
*wq
= NULL
;
4677 struct worker
*worker
;
4679 if (!(task
->flags
& PF_WQ_WORKER
))
4683 * This function is called without any synchronization and @task
4684 * could be in any state. Be careful with dereferences.
4686 worker
= kthread_probe_data(task
);
4689 * Carefully copy the associated workqueue's workfn, name and desc.
4690 * Keep the original last '\0' in case the original is garbage.
4692 copy_from_kernel_nofault(&fn
, &worker
->current_func
, sizeof(fn
));
4693 copy_from_kernel_nofault(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4694 copy_from_kernel_nofault(&wq
, &pwq
->wq
, sizeof(wq
));
4695 copy_from_kernel_nofault(name
, wq
->name
, sizeof(name
) - 1);
4696 copy_from_kernel_nofault(desc
, worker
->desc
, sizeof(desc
) - 1);
4698 if (fn
|| name
[0] || desc
[0]) {
4699 printk("%sWorkqueue: %s %ps", log_lvl
, name
, fn
);
4700 if (strcmp(name
, desc
))
4701 pr_cont(" (%s)", desc
);
4706 static void pr_cont_pool_info(struct worker_pool
*pool
)
4708 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4709 if (pool
->node
!= NUMA_NO_NODE
)
4710 pr_cont(" node=%d", pool
->node
);
4711 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4714 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4716 if (work
->func
== wq_barrier_func
) {
4717 struct wq_barrier
*barr
;
4719 barr
= container_of(work
, struct wq_barrier
, work
);
4721 pr_cont("%s BAR(%d)", comma
? "," : "",
4722 task_pid_nr(barr
->task
));
4724 pr_cont("%s %ps", comma
? "," : "", work
->func
);
4728 static void show_pwq(struct pool_workqueue
*pwq
)
4730 struct worker_pool
*pool
= pwq
->pool
;
4731 struct work_struct
*work
;
4732 struct worker
*worker
;
4733 bool has_in_flight
= false, has_pending
= false;
4736 pr_info(" pwq %d:", pool
->id
);
4737 pr_cont_pool_info(pool
);
4739 pr_cont(" active=%d/%d refcnt=%d%s\n",
4740 pwq
->nr_active
, pwq
->max_active
, pwq
->refcnt
,
4741 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4743 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4744 if (worker
->current_pwq
== pwq
) {
4745 has_in_flight
= true;
4749 if (has_in_flight
) {
4752 pr_info(" in-flight:");
4753 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4754 if (worker
->current_pwq
!= pwq
)
4757 pr_cont("%s %d%s:%ps", comma
? "," : "",
4758 task_pid_nr(worker
->task
),
4759 worker
->rescue_wq
? "(RESCUER)" : "",
4760 worker
->current_func
);
4761 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4762 pr_cont_work(false, work
);
4768 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4769 if (get_work_pwq(work
) == pwq
) {
4777 pr_info(" pending:");
4778 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4779 if (get_work_pwq(work
) != pwq
)
4782 pr_cont_work(comma
, work
);
4783 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4788 if (!list_empty(&pwq
->inactive_works
)) {
4791 pr_info(" inactive:");
4792 list_for_each_entry(work
, &pwq
->inactive_works
, entry
) {
4793 pr_cont_work(comma
, work
);
4794 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4801 * show_one_workqueue - dump state of specified workqueue
4802 * @wq: workqueue whose state will be printed
4804 void show_one_workqueue(struct workqueue_struct
*wq
)
4806 struct pool_workqueue
*pwq
;
4808 unsigned long flags
;
4810 for_each_pwq(pwq
, wq
) {
4811 if (pwq
->nr_active
|| !list_empty(&pwq
->inactive_works
)) {
4816 if (idle
) /* Nothing to print for idle workqueue */
4819 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4821 for_each_pwq(pwq
, wq
) {
4822 raw_spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4823 if (pwq
->nr_active
|| !list_empty(&pwq
->inactive_works
)) {
4825 * Defer printing to avoid deadlocks in console
4826 * drivers that queue work while holding locks
4827 * also taken in their write paths.
4829 printk_deferred_enter();
4831 printk_deferred_exit();
4833 raw_spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4835 * We could be printing a lot from atomic context, e.g.
4836 * sysrq-t -> show_all_workqueues(). Avoid triggering
4839 touch_nmi_watchdog();
4845 * show_one_worker_pool - dump state of specified worker pool
4846 * @pool: worker pool whose state will be printed
4848 static void show_one_worker_pool(struct worker_pool
*pool
)
4850 struct worker
*worker
;
4852 unsigned long flags
;
4854 raw_spin_lock_irqsave(&pool
->lock
, flags
);
4855 if (pool
->nr_workers
== pool
->nr_idle
)
4858 * Defer printing to avoid deadlocks in console drivers that
4859 * queue work while holding locks also taken in their write
4862 printk_deferred_enter();
4863 pr_info("pool %d:", pool
->id
);
4864 pr_cont_pool_info(pool
);
4865 pr_cont(" hung=%us workers=%d",
4866 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4869 pr_cont(" manager: %d",
4870 task_pid_nr(pool
->manager
->task
));
4871 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4872 pr_cont(" %s%d", first
? "idle: " : "",
4873 task_pid_nr(worker
->task
));
4877 printk_deferred_exit();
4879 raw_spin_unlock_irqrestore(&pool
->lock
, flags
);
4881 * We could be printing a lot from atomic context, e.g.
4882 * sysrq-t -> show_all_workqueues(). Avoid triggering
4885 touch_nmi_watchdog();
4890 * show_all_workqueues - dump workqueue state
4892 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4893 * all busy workqueues and pools.
4895 void show_all_workqueues(void)
4897 struct workqueue_struct
*wq
;
4898 struct worker_pool
*pool
;
4903 pr_info("Showing busy workqueues and worker pools:\n");
4905 list_for_each_entry_rcu(wq
, &workqueues
, list
)
4906 show_one_workqueue(wq
);
4908 for_each_pool(pool
, pi
)
4909 show_one_worker_pool(pool
);
4914 /* used to show worker information through /proc/PID/{comm,stat,status} */
4915 void wq_worker_comm(char *buf
, size_t size
, struct task_struct
*task
)
4919 /* always show the actual comm */
4920 off
= strscpy(buf
, task
->comm
, size
);
4924 /* stabilize PF_WQ_WORKER and worker pool association */
4925 mutex_lock(&wq_pool_attach_mutex
);
4927 if (task
->flags
& PF_WQ_WORKER
) {
4928 struct worker
*worker
= kthread_data(task
);
4929 struct worker_pool
*pool
= worker
->pool
;
4932 raw_spin_lock_irq(&pool
->lock
);
4934 * ->desc tracks information (wq name or
4935 * set_worker_desc()) for the latest execution. If
4936 * current, prepend '+', otherwise '-'.
4938 if (worker
->desc
[0] != '\0') {
4939 if (worker
->current_work
)
4940 scnprintf(buf
+ off
, size
- off
, "+%s",
4943 scnprintf(buf
+ off
, size
- off
, "-%s",
4946 raw_spin_unlock_irq(&pool
->lock
);
4950 mutex_unlock(&wq_pool_attach_mutex
);
4958 * There are two challenges in supporting CPU hotplug. Firstly, there
4959 * are a lot of assumptions on strong associations among work, pwq and
4960 * pool which make migrating pending and scheduled works very
4961 * difficult to implement without impacting hot paths. Secondly,
4962 * worker pools serve mix of short, long and very long running works making
4963 * blocked draining impractical.
4965 * This is solved by allowing the pools to be disassociated from the CPU
4966 * running as an unbound one and allowing it to be reattached later if the
4967 * cpu comes back online.
4970 static void unbind_workers(int cpu
)
4972 struct worker_pool
*pool
;
4973 struct worker
*worker
;
4975 for_each_cpu_worker_pool(pool
, cpu
) {
4976 mutex_lock(&wq_pool_attach_mutex
);
4977 raw_spin_lock_irq(&pool
->lock
);
4980 * We've blocked all attach/detach operations. Make all workers
4981 * unbound and set DISASSOCIATED. Before this, all workers
4982 * except for the ones which are still executing works from
4983 * before the last CPU down must be on the cpu. After
4984 * this, they may become diasporas.
4986 for_each_pool_worker(worker
, pool
)
4987 worker
->flags
|= WORKER_UNBOUND
;
4989 pool
->flags
|= POOL_DISASSOCIATED
;
4991 raw_spin_unlock_irq(&pool
->lock
);
4993 for_each_pool_worker(worker
, pool
) {
4994 kthread_set_per_cpu(worker
->task
, -1);
4995 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, cpu_possible_mask
) < 0);
4998 mutex_unlock(&wq_pool_attach_mutex
);
5001 * Call schedule() so that we cross rq->lock and thus can
5002 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
5003 * This is necessary as scheduler callbacks may be invoked
5009 * Sched callbacks are disabled now. Zap nr_running.
5010 * After this, nr_running stays zero and need_more_worker()
5011 * and keep_working() are always true as long as the
5012 * worklist is not empty. This pool now behaves as an
5013 * unbound (in terms of concurrency management) pool which
5014 * are served by workers tied to the pool.
5016 atomic_set(&pool
->nr_running
, 0);
5019 * With concurrency management just turned off, a busy
5020 * worker blocking could lead to lengthy stalls. Kick off
5021 * unbound chain execution of currently pending work items.
5023 raw_spin_lock_irq(&pool
->lock
);
5024 wake_up_worker(pool
);
5025 raw_spin_unlock_irq(&pool
->lock
);
5030 * rebind_workers - rebind all workers of a pool to the associated CPU
5031 * @pool: pool of interest
5033 * @pool->cpu is coming online. Rebind all workers to the CPU.
5035 static void rebind_workers(struct worker_pool
*pool
)
5037 struct worker
*worker
;
5039 lockdep_assert_held(&wq_pool_attach_mutex
);
5042 * Restore CPU affinity of all workers. As all idle workers should
5043 * be on the run-queue of the associated CPU before any local
5044 * wake-ups for concurrency management happen, restore CPU affinity
5045 * of all workers first and then clear UNBOUND. As we're called
5046 * from CPU_ONLINE, the following shouldn't fail.
5048 for_each_pool_worker(worker
, pool
) {
5049 kthread_set_per_cpu(worker
->task
, pool
->cpu
);
5050 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
5051 pool
->attrs
->cpumask
) < 0);
5054 raw_spin_lock_irq(&pool
->lock
);
5056 pool
->flags
&= ~POOL_DISASSOCIATED
;
5058 for_each_pool_worker(worker
, pool
) {
5059 unsigned int worker_flags
= worker
->flags
;
5062 * A bound idle worker should actually be on the runqueue
5063 * of the associated CPU for local wake-ups targeting it to
5064 * work. Kick all idle workers so that they migrate to the
5065 * associated CPU. Doing this in the same loop as
5066 * replacing UNBOUND with REBOUND is safe as no worker will
5067 * be bound before @pool->lock is released.
5069 if (worker_flags
& WORKER_IDLE
)
5070 wake_up_process(worker
->task
);
5073 * We want to clear UNBOUND but can't directly call
5074 * worker_clr_flags() or adjust nr_running. Atomically
5075 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5076 * @worker will clear REBOUND using worker_clr_flags() when
5077 * it initiates the next execution cycle thus restoring
5078 * concurrency management. Note that when or whether
5079 * @worker clears REBOUND doesn't affect correctness.
5081 * WRITE_ONCE() is necessary because @worker->flags may be
5082 * tested without holding any lock in
5083 * wq_worker_running(). Without it, NOT_RUNNING test may
5084 * fail incorrectly leading to premature concurrency
5085 * management operations.
5087 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
5088 worker_flags
|= WORKER_REBOUND
;
5089 worker_flags
&= ~WORKER_UNBOUND
;
5090 WRITE_ONCE(worker
->flags
, worker_flags
);
5093 raw_spin_unlock_irq(&pool
->lock
);
5097 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5098 * @pool: unbound pool of interest
5099 * @cpu: the CPU which is coming up
5101 * An unbound pool may end up with a cpumask which doesn't have any online
5102 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5103 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5104 * online CPU before, cpus_allowed of all its workers should be restored.
5106 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
5108 static cpumask_t cpumask
;
5109 struct worker
*worker
;
5111 lockdep_assert_held(&wq_pool_attach_mutex
);
5113 /* is @cpu allowed for @pool? */
5114 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
5117 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
5119 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5120 for_each_pool_worker(worker
, pool
)
5121 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, &cpumask
) < 0);
5124 int workqueue_prepare_cpu(unsigned int cpu
)
5126 struct worker_pool
*pool
;
5128 for_each_cpu_worker_pool(pool
, cpu
) {
5129 if (pool
->nr_workers
)
5131 if (!create_worker(pool
))
5137 int workqueue_online_cpu(unsigned int cpu
)
5139 struct worker_pool
*pool
;
5140 struct workqueue_struct
*wq
;
5143 mutex_lock(&wq_pool_mutex
);
5145 for_each_pool(pool
, pi
) {
5146 mutex_lock(&wq_pool_attach_mutex
);
5148 if (pool
->cpu
== cpu
)
5149 rebind_workers(pool
);
5150 else if (pool
->cpu
< 0)
5151 restore_unbound_workers_cpumask(pool
, cpu
);
5153 mutex_unlock(&wq_pool_attach_mutex
);
5156 /* update NUMA affinity of unbound workqueues */
5157 list_for_each_entry(wq
, &workqueues
, list
)
5158 wq_update_unbound_numa(wq
, cpu
, true);
5160 mutex_unlock(&wq_pool_mutex
);
5164 int workqueue_offline_cpu(unsigned int cpu
)
5166 struct workqueue_struct
*wq
;
5168 /* unbinding per-cpu workers should happen on the local CPU */
5169 if (WARN_ON(cpu
!= smp_processor_id()))
5172 unbind_workers(cpu
);
5174 /* update NUMA affinity of unbound workqueues */
5175 mutex_lock(&wq_pool_mutex
);
5176 list_for_each_entry(wq
, &workqueues
, list
)
5177 wq_update_unbound_numa(wq
, cpu
, false);
5178 mutex_unlock(&wq_pool_mutex
);
5183 struct work_for_cpu
{
5184 struct work_struct work
;
5190 static void work_for_cpu_fn(struct work_struct
*work
)
5192 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
5194 wfc
->ret
= wfc
->fn(wfc
->arg
);
5198 * work_on_cpu - run a function in thread context on a particular cpu
5199 * @cpu: the cpu to run on
5200 * @fn: the function to run
5201 * @arg: the function arg
5203 * It is up to the caller to ensure that the cpu doesn't go offline.
5204 * The caller must not hold any locks which would prevent @fn from completing.
5206 * Return: The value @fn returns.
5208 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
5210 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
5212 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
5213 schedule_work_on(cpu
, &wfc
.work
);
5214 flush_work(&wfc
.work
);
5215 destroy_work_on_stack(&wfc
.work
);
5218 EXPORT_SYMBOL_GPL(work_on_cpu
);
5221 * work_on_cpu_safe - run a function in thread context on a particular cpu
5222 * @cpu: the cpu to run on
5223 * @fn: the function to run
5224 * @arg: the function argument
5226 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5227 * any locks which would prevent @fn from completing.
5229 * Return: The value @fn returns.
5231 long work_on_cpu_safe(int cpu
, long (*fn
)(void *), void *arg
)
5236 if (cpu_online(cpu
))
5237 ret
= work_on_cpu(cpu
, fn
, arg
);
5241 EXPORT_SYMBOL_GPL(work_on_cpu_safe
);
5242 #endif /* CONFIG_SMP */
5244 #ifdef CONFIG_FREEZER
5247 * freeze_workqueues_begin - begin freezing workqueues
5249 * Start freezing workqueues. After this function returns, all freezable
5250 * workqueues will queue new works to their inactive_works list instead of
5254 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5256 void freeze_workqueues_begin(void)
5258 struct workqueue_struct
*wq
;
5259 struct pool_workqueue
*pwq
;
5261 mutex_lock(&wq_pool_mutex
);
5263 WARN_ON_ONCE(workqueue_freezing
);
5264 workqueue_freezing
= true;
5266 list_for_each_entry(wq
, &workqueues
, list
) {
5267 mutex_lock(&wq
->mutex
);
5268 for_each_pwq(pwq
, wq
)
5269 pwq_adjust_max_active(pwq
);
5270 mutex_unlock(&wq
->mutex
);
5273 mutex_unlock(&wq_pool_mutex
);
5277 * freeze_workqueues_busy - are freezable workqueues still busy?
5279 * Check whether freezing is complete. This function must be called
5280 * between freeze_workqueues_begin() and thaw_workqueues().
5283 * Grabs and releases wq_pool_mutex.
5286 * %true if some freezable workqueues are still busy. %false if freezing
5289 bool freeze_workqueues_busy(void)
5292 struct workqueue_struct
*wq
;
5293 struct pool_workqueue
*pwq
;
5295 mutex_lock(&wq_pool_mutex
);
5297 WARN_ON_ONCE(!workqueue_freezing
);
5299 list_for_each_entry(wq
, &workqueues
, list
) {
5300 if (!(wq
->flags
& WQ_FREEZABLE
))
5303 * nr_active is monotonically decreasing. It's safe
5304 * to peek without lock.
5307 for_each_pwq(pwq
, wq
) {
5308 WARN_ON_ONCE(pwq
->nr_active
< 0);
5309 if (pwq
->nr_active
) {
5318 mutex_unlock(&wq_pool_mutex
);
5323 * thaw_workqueues - thaw workqueues
5325 * Thaw workqueues. Normal queueing is restored and all collected
5326 * frozen works are transferred to their respective pool worklists.
5329 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5331 void thaw_workqueues(void)
5333 struct workqueue_struct
*wq
;
5334 struct pool_workqueue
*pwq
;
5336 mutex_lock(&wq_pool_mutex
);
5338 if (!workqueue_freezing
)
5341 workqueue_freezing
= false;
5343 /* restore max_active and repopulate worklist */
5344 list_for_each_entry(wq
, &workqueues
, list
) {
5345 mutex_lock(&wq
->mutex
);
5346 for_each_pwq(pwq
, wq
)
5347 pwq_adjust_max_active(pwq
);
5348 mutex_unlock(&wq
->mutex
);
5352 mutex_unlock(&wq_pool_mutex
);
5354 #endif /* CONFIG_FREEZER */
5356 static int workqueue_apply_unbound_cpumask(void)
5360 struct workqueue_struct
*wq
;
5361 struct apply_wqattrs_ctx
*ctx
, *n
;
5363 lockdep_assert_held(&wq_pool_mutex
);
5365 list_for_each_entry(wq
, &workqueues
, list
) {
5366 if (!(wq
->flags
& WQ_UNBOUND
))
5368 /* creating multiple pwqs breaks ordering guarantee */
5369 if (wq
->flags
& __WQ_ORDERED
)
5372 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
5378 list_add_tail(&ctx
->list
, &ctxs
);
5381 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
5383 apply_wqattrs_commit(ctx
);
5384 apply_wqattrs_cleanup(ctx
);
5391 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5392 * @cpumask: the cpumask to set
5394 * The low-level workqueues cpumask is a global cpumask that limits
5395 * the affinity of all unbound workqueues. This function check the @cpumask
5396 * and apply it to all unbound workqueues and updates all pwqs of them.
5398 * Return: 0 - Success
5399 * -EINVAL - Invalid @cpumask
5400 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5402 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
5405 cpumask_var_t saved_cpumask
;
5408 * Not excluding isolated cpus on purpose.
5409 * If the user wishes to include them, we allow that.
5411 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
5412 if (!cpumask_empty(cpumask
)) {
5413 apply_wqattrs_lock();
5414 if (cpumask_equal(cpumask
, wq_unbound_cpumask
)) {
5419 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
)) {
5424 /* save the old wq_unbound_cpumask. */
5425 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
5427 /* update wq_unbound_cpumask at first and apply it to wqs. */
5428 cpumask_copy(wq_unbound_cpumask
, cpumask
);
5429 ret
= workqueue_apply_unbound_cpumask();
5431 /* restore the wq_unbound_cpumask when failed. */
5433 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
5435 free_cpumask_var(saved_cpumask
);
5437 apply_wqattrs_unlock();
5445 * Workqueues with WQ_SYSFS flag set is visible to userland via
5446 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5447 * following attributes.
5449 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5450 * max_active RW int : maximum number of in-flight work items
5452 * Unbound workqueues have the following extra attributes.
5454 * pool_ids RO int : the associated pool IDs for each node
5455 * nice RW int : nice value of the workers
5456 * cpumask RW mask : bitmask of allowed CPUs for the workers
5457 * numa RW bool : whether enable NUMA affinity
5460 struct workqueue_struct
*wq
;
5464 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
5466 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5471 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
5474 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5476 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
5478 static DEVICE_ATTR_RO(per_cpu
);
5480 static ssize_t
max_active_show(struct device
*dev
,
5481 struct device_attribute
*attr
, char *buf
)
5483 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5485 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
5488 static ssize_t
max_active_store(struct device
*dev
,
5489 struct device_attribute
*attr
, const char *buf
,
5492 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5495 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
5498 workqueue_set_max_active(wq
, val
);
5501 static DEVICE_ATTR_RW(max_active
);
5503 static struct attribute
*wq_sysfs_attrs
[] = {
5504 &dev_attr_per_cpu
.attr
,
5505 &dev_attr_max_active
.attr
,
5508 ATTRIBUTE_GROUPS(wq_sysfs
);
5510 static ssize_t
wq_pool_ids_show(struct device
*dev
,
5511 struct device_attribute
*attr
, char *buf
)
5513 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5514 const char *delim
= "";
5515 int node
, written
= 0;
5519 for_each_node(node
) {
5520 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
5521 "%s%d:%d", delim
, node
,
5522 unbound_pwq_by_node(wq
, node
)->pool
->id
);
5525 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
5532 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
5535 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5538 mutex_lock(&wq
->mutex
);
5539 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
5540 mutex_unlock(&wq
->mutex
);
5545 /* prepare workqueue_attrs for sysfs store operations */
5546 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
5548 struct workqueue_attrs
*attrs
;
5550 lockdep_assert_held(&wq_pool_mutex
);
5552 attrs
= alloc_workqueue_attrs();
5556 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
5560 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
5561 const char *buf
, size_t count
)
5563 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5564 struct workqueue_attrs
*attrs
;
5567 apply_wqattrs_lock();
5569 attrs
= wq_sysfs_prep_attrs(wq
);
5573 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5574 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5575 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5580 apply_wqattrs_unlock();
5581 free_workqueue_attrs(attrs
);
5582 return ret
?: count
;
5585 static ssize_t
wq_cpumask_show(struct device
*dev
,
5586 struct device_attribute
*attr
, char *buf
)
5588 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5591 mutex_lock(&wq
->mutex
);
5592 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5593 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5594 mutex_unlock(&wq
->mutex
);
5598 static ssize_t
wq_cpumask_store(struct device
*dev
,
5599 struct device_attribute
*attr
,
5600 const char *buf
, size_t count
)
5602 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5603 struct workqueue_attrs
*attrs
;
5606 apply_wqattrs_lock();
5608 attrs
= wq_sysfs_prep_attrs(wq
);
5612 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5614 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5617 apply_wqattrs_unlock();
5618 free_workqueue_attrs(attrs
);
5619 return ret
?: count
;
5622 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5625 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5628 mutex_lock(&wq
->mutex
);
5629 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5630 !wq
->unbound_attrs
->no_numa
);
5631 mutex_unlock(&wq
->mutex
);
5636 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5637 const char *buf
, size_t count
)
5639 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5640 struct workqueue_attrs
*attrs
;
5641 int v
, ret
= -ENOMEM
;
5643 apply_wqattrs_lock();
5645 attrs
= wq_sysfs_prep_attrs(wq
);
5650 if (sscanf(buf
, "%d", &v
) == 1) {
5651 attrs
->no_numa
= !v
;
5652 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5656 apply_wqattrs_unlock();
5657 free_workqueue_attrs(attrs
);
5658 return ret
?: count
;
5661 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5662 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5663 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5664 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5665 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5669 static struct bus_type wq_subsys
= {
5670 .name
= "workqueue",
5671 .dev_groups
= wq_sysfs_groups
,
5674 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5675 struct device_attribute
*attr
, char *buf
)
5679 mutex_lock(&wq_pool_mutex
);
5680 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5681 cpumask_pr_args(wq_unbound_cpumask
));
5682 mutex_unlock(&wq_pool_mutex
);
5687 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5688 struct device_attribute
*attr
, const char *buf
, size_t count
)
5690 cpumask_var_t cpumask
;
5693 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5696 ret
= cpumask_parse(buf
, cpumask
);
5698 ret
= workqueue_set_unbound_cpumask(cpumask
);
5700 free_cpumask_var(cpumask
);
5701 return ret
? ret
: count
;
5704 static struct device_attribute wq_sysfs_cpumask_attr
=
5705 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5706 wq_unbound_cpumask_store
);
5708 static int __init
wq_sysfs_init(void)
5712 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5716 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5718 core_initcall(wq_sysfs_init
);
5720 static void wq_device_release(struct device
*dev
)
5722 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5728 * workqueue_sysfs_register - make a workqueue visible in sysfs
5729 * @wq: the workqueue to register
5731 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5732 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5733 * which is the preferred method.
5735 * Workqueue user should use this function directly iff it wants to apply
5736 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5737 * apply_workqueue_attrs() may race against userland updating the
5740 * Return: 0 on success, -errno on failure.
5742 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5744 struct wq_device
*wq_dev
;
5748 * Adjusting max_active or creating new pwqs by applying
5749 * attributes breaks ordering guarantee. Disallow exposing ordered
5752 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
5755 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5760 wq_dev
->dev
.bus
= &wq_subsys
;
5761 wq_dev
->dev
.release
= wq_device_release
;
5762 dev_set_name(&wq_dev
->dev
, "%s", wq
->name
);
5765 * unbound_attrs are created separately. Suppress uevent until
5766 * everything is ready.
5768 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5770 ret
= device_register(&wq_dev
->dev
);
5772 put_device(&wq_dev
->dev
);
5777 if (wq
->flags
& WQ_UNBOUND
) {
5778 struct device_attribute
*attr
;
5780 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5781 ret
= device_create_file(&wq_dev
->dev
, attr
);
5783 device_unregister(&wq_dev
->dev
);
5790 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5791 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5796 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5797 * @wq: the workqueue to unregister
5799 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5801 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5803 struct wq_device
*wq_dev
= wq
->wq_dev
;
5809 device_unregister(&wq_dev
->dev
);
5811 #else /* CONFIG_SYSFS */
5812 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5813 #endif /* CONFIG_SYSFS */
5816 * Workqueue watchdog.
5818 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5819 * flush dependency, a concurrency managed work item which stays RUNNING
5820 * indefinitely. Workqueue stalls can be very difficult to debug as the
5821 * usual warning mechanisms don't trigger and internal workqueue state is
5824 * Workqueue watchdog monitors all worker pools periodically and dumps
5825 * state if some pools failed to make forward progress for a while where
5826 * forward progress is defined as the first item on ->worklist changing.
5828 * This mechanism is controlled through the kernel parameter
5829 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5830 * corresponding sysfs parameter file.
5832 #ifdef CONFIG_WQ_WATCHDOG
5834 static unsigned long wq_watchdog_thresh
= 30;
5835 static struct timer_list wq_watchdog_timer
;
5837 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5838 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5840 static void wq_watchdog_reset_touched(void)
5844 wq_watchdog_touched
= jiffies
;
5845 for_each_possible_cpu(cpu
)
5846 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5849 static void wq_watchdog_timer_fn(struct timer_list
*unused
)
5851 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5852 bool lockup_detected
= false;
5853 unsigned long now
= jiffies
;
5854 struct worker_pool
*pool
;
5862 for_each_pool(pool
, pi
) {
5863 unsigned long pool_ts
, touched
, ts
;
5865 if (list_empty(&pool
->worklist
))
5869 * If a virtual machine is stopped by the host it can look to
5870 * the watchdog like a stall.
5872 kvm_check_and_clear_guest_paused();
5874 /* get the latest of pool and touched timestamps */
5876 touched
= READ_ONCE(per_cpu(wq_watchdog_touched_cpu
, pool
->cpu
));
5878 touched
= READ_ONCE(wq_watchdog_touched
);
5879 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5881 if (time_after(pool_ts
, touched
))
5887 if (time_after(now
, ts
+ thresh
)) {
5888 lockup_detected
= true;
5889 pr_emerg("BUG: workqueue lockup - pool");
5890 pr_cont_pool_info(pool
);
5891 pr_cont(" stuck for %us!\n",
5892 jiffies_to_msecs(now
- pool_ts
) / 1000);
5898 if (lockup_detected
)
5899 show_all_workqueues();
5901 wq_watchdog_reset_touched();
5902 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5905 notrace
void wq_watchdog_touch(int cpu
)
5908 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5910 wq_watchdog_touched
= jiffies
;
5913 static void wq_watchdog_set_thresh(unsigned long thresh
)
5915 wq_watchdog_thresh
= 0;
5916 del_timer_sync(&wq_watchdog_timer
);
5919 wq_watchdog_thresh
= thresh
;
5920 wq_watchdog_reset_touched();
5921 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5925 static int wq_watchdog_param_set_thresh(const char *val
,
5926 const struct kernel_param
*kp
)
5928 unsigned long thresh
;
5931 ret
= kstrtoul(val
, 0, &thresh
);
5936 wq_watchdog_set_thresh(thresh
);
5938 wq_watchdog_thresh
= thresh
;
5943 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5944 .set
= wq_watchdog_param_set_thresh
,
5945 .get
= param_get_ulong
,
5948 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5951 static void wq_watchdog_init(void)
5953 timer_setup(&wq_watchdog_timer
, wq_watchdog_timer_fn
, TIMER_DEFERRABLE
);
5954 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5957 #else /* CONFIG_WQ_WATCHDOG */
5959 static inline void wq_watchdog_init(void) { }
5961 #endif /* CONFIG_WQ_WATCHDOG */
5963 static void __init
wq_numa_init(void)
5968 if (num_possible_nodes() <= 1)
5971 if (wq_disable_numa
) {
5972 pr_info("workqueue: NUMA affinity support disabled\n");
5976 for_each_possible_cpu(cpu
) {
5977 if (WARN_ON(cpu_to_node(cpu
) == NUMA_NO_NODE
)) {
5978 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5983 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs();
5984 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5987 * We want masks of possible CPUs of each node which isn't readily
5988 * available. Build one from cpu_to_node() which should have been
5989 * fully initialized by now.
5991 tbl
= kcalloc(nr_node_ids
, sizeof(tbl
[0]), GFP_KERNEL
);
5995 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5996 node_online(node
) ? node
: NUMA_NO_NODE
));
5998 for_each_possible_cpu(cpu
) {
5999 node
= cpu_to_node(cpu
);
6000 cpumask_set_cpu(cpu
, tbl
[node
]);
6003 wq_numa_possible_cpumask
= tbl
;
6004 wq_numa_enabled
= true;
6008 * workqueue_init_early - early init for workqueue subsystem
6010 * This is the first half of two-staged workqueue subsystem initialization
6011 * and invoked as soon as the bare basics - memory allocation, cpumasks and
6012 * idr are up. It sets up all the data structures and system workqueues
6013 * and allows early boot code to create workqueues and queue/cancel work
6014 * items. Actual work item execution starts only after kthreads can be
6015 * created and scheduled right before early initcalls.
6017 void __init
workqueue_init_early(void)
6019 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
6020 int hk_flags
= HK_FLAG_DOMAIN
| HK_FLAG_WQ
;
6023 BUILD_BUG_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
6025 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
6026 cpumask_copy(wq_unbound_cpumask
, housekeeping_cpumask(hk_flags
));
6028 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
6030 /* initialize CPU pools */
6031 for_each_possible_cpu(cpu
) {
6032 struct worker_pool
*pool
;
6035 for_each_cpu_worker_pool(pool
, cpu
) {
6036 BUG_ON(init_worker_pool(pool
));
6038 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
6039 pool
->attrs
->nice
= std_nice
[i
++];
6040 pool
->node
= cpu_to_node(cpu
);
6043 mutex_lock(&wq_pool_mutex
);
6044 BUG_ON(worker_pool_assign_id(pool
));
6045 mutex_unlock(&wq_pool_mutex
);
6049 /* create default unbound and ordered wq attrs */
6050 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
6051 struct workqueue_attrs
*attrs
;
6053 BUG_ON(!(attrs
= alloc_workqueue_attrs()));
6054 attrs
->nice
= std_nice
[i
];
6055 unbound_std_wq_attrs
[i
] = attrs
;
6058 * An ordered wq should have only one pwq as ordering is
6059 * guaranteed by max_active which is enforced by pwqs.
6060 * Turn off NUMA so that dfl_pwq is used for all nodes.
6062 BUG_ON(!(attrs
= alloc_workqueue_attrs()));
6063 attrs
->nice
= std_nice
[i
];
6064 attrs
->no_numa
= true;
6065 ordered_wq_attrs
[i
] = attrs
;
6068 system_wq
= alloc_workqueue("events", 0, 0);
6069 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
6070 system_long_wq
= alloc_workqueue("events_long", 0, 0);
6071 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
6072 WQ_UNBOUND_MAX_ACTIVE
);
6073 system_freezable_wq
= alloc_workqueue("events_freezable",
6075 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
6076 WQ_POWER_EFFICIENT
, 0);
6077 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
6078 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
6080 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
6081 !system_unbound_wq
|| !system_freezable_wq
||
6082 !system_power_efficient_wq
||
6083 !system_freezable_power_efficient_wq
);
6087 * workqueue_init - bring workqueue subsystem fully online
6089 * This is the latter half of two-staged workqueue subsystem initialization
6090 * and invoked as soon as kthreads can be created and scheduled.
6091 * Workqueues have been created and work items queued on them, but there
6092 * are no kworkers executing the work items yet. Populate the worker pools
6093 * with the initial workers and enable future kworker creations.
6095 void __init
workqueue_init(void)
6097 struct workqueue_struct
*wq
;
6098 struct worker_pool
*pool
;
6102 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6103 * CPU to node mapping may not be available that early on some
6104 * archs such as power and arm64. As per-cpu pools created
6105 * previously could be missing node hint and unbound pools NUMA
6106 * affinity, fix them up.
6108 * Also, while iterating workqueues, create rescuers if requested.
6112 mutex_lock(&wq_pool_mutex
);
6114 for_each_possible_cpu(cpu
) {
6115 for_each_cpu_worker_pool(pool
, cpu
) {
6116 pool
->node
= cpu_to_node(cpu
);
6120 list_for_each_entry(wq
, &workqueues
, list
) {
6121 wq_update_unbound_numa(wq
, smp_processor_id(), true);
6122 WARN(init_rescuer(wq
),
6123 "workqueue: failed to create early rescuer for %s",
6127 mutex_unlock(&wq_pool_mutex
);
6129 /* create the initial workers */
6130 for_each_online_cpu(cpu
) {
6131 for_each_cpu_worker_pool(pool
, cpu
) {
6132 pool
->flags
&= ~POOL_DISASSOCIATED
;
6133 BUG_ON(!create_worker(pool
));
6137 hash_for_each(unbound_pool_hash
, bkt
, pool
, hash_node
)
6138 BUG_ON(!create_worker(pool
));