1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
54 #include "workqueue_internal.h"
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
69 * Note that DISASSOCIATED should be flipped only while holding
70 * wq_pool_attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * A: wq_pool_attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. RCU protected for reads.
142 * MD: wq_mayday_lock protected.
145 /* struct worker is defined in workqueue_internal.h */
148 spinlock_t lock; /* the pool lock */
149 int cpu; /* I: the associated cpu */
150 int node; /* I: the associated node ID */
151 int id; /* I: pool ID */
152 unsigned int flags; /* X: flags */
154 unsigned long watchdog_ts; /* L: watchdog timestamp */
156 struct list_head worklist; /* L: list of pending works */
158 int nr_workers; /* L: total number of workers */
159 int nr_idle; /* L: currently idle workers */
161 struct list_head idle_list; /* X: list of idle workers */
162 struct timer_list idle_timer; /* L: worker idle timeout */
163 struct timer_list mayday_timer; /* L: SOS timer for workers */
165 /* a workers is either on busy_hash or idle_list, or the manager */
166 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
167 /* L: hash of busy workers */
169 struct worker *manager; /* L: purely informational */
170 struct list_head workers; /* A: attached workers */
171 struct completion *detach_completion; /* all workers detached */
173 struct ida worker_ida; /* worker IDs for task name */
175 struct workqueue_attrs *attrs; /* I: worker attributes */
176 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
177 int refcnt; /* PL: refcnt for unbound pools */
180 * The current concurrency level. As it's likely to be accessed
181 * from other CPUs during try_to_wake_up(), put it in a separate
184 atomic_t nr_running ____cacheline_aligned_in_smp;
187 * Destruction of pool is RCU protected to allow dereferences
188 * from get_work_pool().
191 } ____cacheline_aligned_in_smp;
194 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
195 * of work_struct->data are used for flags and the remaining high bits
196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
197 * number of flag bits.
199 struct pool_workqueue {
200 struct worker_pool *pool; /* I: the associated pool */
201 struct workqueue_struct *wq; /* I: the owning workqueue */
202 int work_color; /* L: current color */
203 int flush_color; /* L: flushing color */
204 int refcnt; /* L: reference count */
205 int nr_in_flight[WORK_NR_COLORS];
206 /* L: nr of in_flight works */
207 int nr_active; /* L: nr of active works */
208 int max_active; /* L: max active works */
209 struct list_head delayed_works; /* L: delayed works */
210 struct list_head pwqs_node; /* WR: node on wq->pwqs */
211 struct list_head mayday_node; /* MD: node on wq->maydays */
214 * Release of unbound pwq is punted to system_wq. See put_pwq()
215 * and pwq_unbound_release_workfn() for details. pool_workqueue
216 * itself is also RCU protected so that the first pwq can be
217 * determined without grabbing wq->mutex.
219 struct work_struct unbound_release_work;
221 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
224 * Structure used to wait for workqueue flush.
227 struct list_head list; /* WQ: list of flushers */
228 int flush_color; /* WQ: flush color waiting for */
229 struct completion done; /* flush completion */
235 * The externally visible workqueue. It relays the issued work items to
236 * the appropriate worker_pool through its pool_workqueues.
238 struct workqueue_struct {
239 struct list_head pwqs; /* WR: all pwqs of this wq */
240 struct list_head list; /* PR: list of all workqueues */
242 struct mutex mutex; /* protects this wq */
243 int work_color; /* WQ: current work color */
244 int flush_color; /* WQ: current flush color */
245 atomic_t nr_pwqs_to_flush; /* flush in progress */
246 struct wq_flusher *first_flusher; /* WQ: first flusher */
247 struct list_head flusher_queue; /* WQ: flush waiters */
248 struct list_head flusher_overflow; /* WQ: flush overflow list */
250 struct list_head maydays; /* MD: pwqs requesting rescue */
251 struct worker *rescuer; /* MD: rescue worker */
253 int nr_drainers; /* WQ: drain in progress */
254 int saved_max_active; /* WQ: saved pwq max_active */
256 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
257 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
260 struct wq_device *wq_dev; /* I: for sysfs interface */
262 #ifdef CONFIG_LOCKDEP
264 struct lock_class_key key;
265 struct lockdep_map lockdep_map;
267 char name[WQ_NAME_LEN]; /* I: workqueue name */
270 * Destruction of workqueue_struct is RCU protected to allow walking
271 * the workqueues list without grabbing wq_pool_mutex.
272 * This is used to dump all workqueues from sysrq.
276 /* hot fields used during command issue, aligned to cacheline */
277 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
278 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
279 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
282 static struct kmem_cache *pwq_cache;
284 static cpumask_var_t *wq_numa_possible_cpumask;
285 /* possible CPUs of each node */
287 static bool wq_disable_numa;
288 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
290 /* see the comment above the definition of WQ_POWER_EFFICIENT */
291 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
292 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
294 static bool wq_online; /* can kworkers be created yet? */
296 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
298 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
301 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
302 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
303 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
304 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
306 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
307 static bool workqueue_freezing; /* PL: have wqs started freezing? */
309 /* PL: allowable cpus for unbound wqs and work items */
310 static cpumask_var_t wq_unbound_cpumask;
312 /* CPU where unbound work was last round robin scheduled from this CPU */
313 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
316 * Local execution of unbound work items is no longer guaranteed. The
317 * following always forces round-robin CPU selection on unbound work items
318 * to uncover usages which depend on it.
320 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
321 static bool wq_debug_force_rr_cpu = true;
323 static bool wq_debug_force_rr_cpu = false;
325 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
327 /* the per-cpu worker pools */
328 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
330 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
332 /* PL: hash of all unbound pools keyed by pool->attrs */
333 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
335 /* I: attributes used when instantiating standard unbound pools on demand */
336 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
338 /* I: attributes used when instantiating ordered pools on demand */
339 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
341 struct workqueue_struct *system_wq __read_mostly;
342 EXPORT_SYMBOL(system_wq);
343 struct workqueue_struct *system_highpri_wq __read_mostly;
344 EXPORT_SYMBOL_GPL(system_highpri_wq);
345 struct workqueue_struct *system_long_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_long_wq);
347 struct workqueue_struct *system_unbound_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_unbound_wq);
349 struct workqueue_struct *system_freezable_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_freezable_wq);
351 struct workqueue_struct *system_power_efficient_wq __read_mostly;
352 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
353 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
354 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
356 static int worker_thread(void *__worker);
357 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
358 static void show_pwq(struct pool_workqueue *pwq);
360 #define CREATE_TRACE_POINTS
361 #include <trace/events/workqueue.h>
363 #define assert_rcu_or_pool_mutex() \
364 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
365 !lockdep_is_held(&wq_pool_mutex), \
366 "RCU or wq_pool_mutex should be held")
368 #define assert_rcu_or_wq_mutex(wq) \
369 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
370 !lockdep_is_held(&wq->mutex), \
371 "RCU or wq->mutex should be held")
373 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
374 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
375 !lockdep_is_held(&wq->mutex) && \
376 !lockdep_is_held(&wq_pool_mutex), \
377 "RCU, wq->mutex or wq_pool_mutex should be held")
379 #define for_each_cpu_worker_pool(pool, cpu) \
380 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
381 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
385 * for_each_pool - iterate through all worker_pools in the system
386 * @pool: iteration cursor
387 * @pi: integer used for iteration
389 * This must be called either with wq_pool_mutex held or RCU read
390 * locked. If the pool needs to be used beyond the locking in effect, the
391 * caller is responsible for guaranteeing that the pool stays online.
393 * The if/else clause exists only for the lockdep assertion and can be
396 #define for_each_pool(pool, pi) \
397 idr_for_each_entry(&worker_pool_idr, pool, pi) \
398 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
402 * for_each_pool_worker - iterate through all workers of a worker_pool
403 * @worker: iteration cursor
404 * @pool: worker_pool to iterate workers of
406 * This must be called with wq_pool_attach_mutex.
408 * The if/else clause exists only for the lockdep assertion and can be
411 #define for_each_pool_worker(worker, pool) \
412 list_for_each_entry((worker), &(pool)->workers, node) \
413 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
417 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
418 * @pwq: iteration cursor
419 * @wq: the target workqueue
421 * This must be called either with wq->mutex held or RCU read locked.
422 * If the pwq needs to be used beyond the locking in effect, the caller is
423 * responsible for guaranteeing that the pwq stays online.
425 * The if/else clause exists only for the lockdep assertion and can be
428 #define for_each_pwq(pwq, wq) \
429 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
430 lockdep_is_held(&wq->mutex)) \
431 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
434 #ifdef CONFIG_DEBUG_OBJECTS_WORK
436 static struct debug_obj_descr work_debug_descr;
438 static void *work_debug_hint(void *addr)
440 return ((struct work_struct *) addr)->func;
443 static bool work_is_static_object(void *addr)
445 struct work_struct *work = addr;
447 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
451 * fixup_init is called when:
452 * - an active object is initialized
454 static bool work_fixup_init(void *addr, enum debug_obj_state state)
456 struct work_struct *work = addr;
459 case ODEBUG_STATE_ACTIVE:
460 cancel_work_sync(work);
461 debug_object_init(work, &work_debug_descr);
469 * fixup_free is called when:
470 * - an active object is freed
472 static bool work_fixup_free(void *addr, enum debug_obj_state state)
474 struct work_struct *work = addr;
477 case ODEBUG_STATE_ACTIVE:
478 cancel_work_sync(work);
479 debug_object_free(work, &work_debug_descr);
486 static struct debug_obj_descr work_debug_descr = {
487 .name = "work_struct",
488 .debug_hint = work_debug_hint,
489 .is_static_object = work_is_static_object,
490 .fixup_init = work_fixup_init,
491 .fixup_free = work_fixup_free,
494 static inline void debug_work_activate(struct work_struct *work)
496 debug_object_activate(work, &work_debug_descr);
499 static inline void debug_work_deactivate(struct work_struct *work)
501 debug_object_deactivate(work, &work_debug_descr);
504 void __init_work(struct work_struct *work, int onstack)
507 debug_object_init_on_stack(work, &work_debug_descr);
509 debug_object_init(work, &work_debug_descr);
511 EXPORT_SYMBOL_GPL(__init_work);
513 void destroy_work_on_stack(struct work_struct *work)
515 debug_object_free(work, &work_debug_descr);
517 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
519 void destroy_delayed_work_on_stack(struct delayed_work *work)
521 destroy_timer_on_stack(&work->timer);
522 debug_object_free(&work->work, &work_debug_descr);
524 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
527 static inline void debug_work_activate(struct work_struct *work) { }
528 static inline void debug_work_deactivate(struct work_struct *work) { }
532 * worker_pool_assign_id - allocate ID and assing it to @pool
533 * @pool: the pool pointer of interest
535 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
536 * successfully, -errno on failure.
538 static int worker_pool_assign_id(struct worker_pool *pool)
542 lockdep_assert_held(&wq_pool_mutex);
544 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
554 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
555 * @wq: the target workqueue
558 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
560 * If the pwq needs to be used beyond the locking in effect, the caller is
561 * responsible for guaranteeing that the pwq stays online.
563 * Return: The unbound pool_workqueue for @node.
565 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
568 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
571 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
572 * delayed item is pending. The plan is to keep CPU -> NODE
573 * mapping valid and stable across CPU on/offlines. Once that
574 * happens, this workaround can be removed.
576 if (unlikely(node == NUMA_NO_NODE))
579 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
582 static unsigned int work_color_to_flags(int color)
584 return color << WORK_STRUCT_COLOR_SHIFT;
587 static int get_work_color(struct work_struct *work)
589 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
590 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
593 static int work_next_color(int color)
595 return (color + 1) % WORK_NR_COLORS;
599 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
600 * contain the pointer to the queued pwq. Once execution starts, the flag
601 * is cleared and the high bits contain OFFQ flags and pool ID.
603 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
604 * and clear_work_data() can be used to set the pwq, pool or clear
605 * work->data. These functions should only be called while the work is
606 * owned - ie. while the PENDING bit is set.
608 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
609 * corresponding to a work. Pool is available once the work has been
610 * queued anywhere after initialization until it is sync canceled. pwq is
611 * available only while the work item is queued.
613 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
614 * canceled. While being canceled, a work item may have its PENDING set
615 * but stay off timer and worklist for arbitrarily long and nobody should
616 * try to steal the PENDING bit.
618 static inline void set_work_data(struct work_struct *work, unsigned long data,
621 WARN_ON_ONCE(!work_pending(work));
622 atomic_long_set(&work->data, data | flags | work_static(work));
625 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
626 unsigned long extra_flags)
628 set_work_data(work, (unsigned long)pwq,
629 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
632 static void set_work_pool_and_keep_pending(struct work_struct *work,
635 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
636 WORK_STRUCT_PENDING);
639 static void set_work_pool_and_clear_pending(struct work_struct *work,
643 * The following wmb is paired with the implied mb in
644 * test_and_set_bit(PENDING) and ensures all updates to @work made
645 * here are visible to and precede any updates by the next PENDING
649 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
651 * The following mb guarantees that previous clear of a PENDING bit
652 * will not be reordered with any speculative LOADS or STORES from
653 * work->current_func, which is executed afterwards. This possible
654 * reordering can lead to a missed execution on attempt to queue
655 * the same @work. E.g. consider this case:
658 * ---------------------------- --------------------------------
660 * 1 STORE event_indicated
661 * 2 queue_work_on() {
662 * 3 test_and_set_bit(PENDING)
663 * 4 } set_..._and_clear_pending() {
664 * 5 set_work_data() # clear bit
666 * 7 work->current_func() {
667 * 8 LOAD event_indicated
670 * Without an explicit full barrier speculative LOAD on line 8 can
671 * be executed before CPU#0 does STORE on line 1. If that happens,
672 * CPU#0 observes the PENDING bit is still set and new execution of
673 * a @work is not queued in a hope, that CPU#1 will eventually
674 * finish the queued @work. Meanwhile CPU#1 does not see
675 * event_indicated is set, because speculative LOAD was executed
676 * before actual STORE.
681 static void clear_work_data(struct work_struct *work)
683 smp_wmb(); /* see set_work_pool_and_clear_pending() */
684 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
687 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
689 unsigned long data = atomic_long_read(&work->data);
691 if (data & WORK_STRUCT_PWQ)
692 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
698 * get_work_pool - return the worker_pool a given work was associated with
699 * @work: the work item of interest
701 * Pools are created and destroyed under wq_pool_mutex, and allows read
702 * access under RCU read lock. As such, this function should be
703 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
705 * All fields of the returned pool are accessible as long as the above
706 * mentioned locking is in effect. If the returned pool needs to be used
707 * beyond the critical section, the caller is responsible for ensuring the
708 * returned pool is and stays online.
710 * Return: The worker_pool @work was last associated with. %NULL if none.
712 static struct worker_pool *get_work_pool(struct work_struct *work)
714 unsigned long data = atomic_long_read(&work->data);
717 assert_rcu_or_pool_mutex();
719 if (data & WORK_STRUCT_PWQ)
720 return ((struct pool_workqueue *)
721 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
723 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
724 if (pool_id == WORK_OFFQ_POOL_NONE)
727 return idr_find(&worker_pool_idr, pool_id);
731 * get_work_pool_id - return the worker pool ID a given work is associated with
732 * @work: the work item of interest
734 * Return: The worker_pool ID @work was last associated with.
735 * %WORK_OFFQ_POOL_NONE if none.
737 static int get_work_pool_id(struct work_struct *work)
739 unsigned long data = atomic_long_read(&work->data);
741 if (data & WORK_STRUCT_PWQ)
742 return ((struct pool_workqueue *)
743 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
745 return data >> WORK_OFFQ_POOL_SHIFT;
748 static void mark_work_canceling(struct work_struct *work)
750 unsigned long pool_id = get_work_pool_id(work);
752 pool_id <<= WORK_OFFQ_POOL_SHIFT;
753 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
756 static bool work_is_canceling(struct work_struct *work)
758 unsigned long data = atomic_long_read(&work->data);
760 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
764 * Policy functions. These define the policies on how the global worker
765 * pools are managed. Unless noted otherwise, these functions assume that
766 * they're being called with pool->lock held.
769 static bool __need_more_worker(struct worker_pool *pool)
771 return !atomic_read(&pool->nr_running);
775 * Need to wake up a worker? Called from anything but currently
778 * Note that, because unbound workers never contribute to nr_running, this
779 * function will always return %true for unbound pools as long as the
780 * worklist isn't empty.
782 static bool need_more_worker(struct worker_pool *pool)
784 return !list_empty(&pool->worklist) && __need_more_worker(pool);
787 /* Can I start working? Called from busy but !running workers. */
788 static bool may_start_working(struct worker_pool *pool)
790 return pool->nr_idle;
793 /* Do I need to keep working? Called from currently running workers. */
794 static bool keep_working(struct worker_pool *pool)
796 return !list_empty(&pool->worklist) &&
797 atomic_read(&pool->nr_running) <= 1;
800 /* Do we need a new worker? Called from manager. */
801 static bool need_to_create_worker(struct worker_pool *pool)
803 return need_more_worker(pool) && !may_start_working(pool);
806 /* Do we have too many workers and should some go away? */
807 static bool too_many_workers(struct worker_pool *pool)
809 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
810 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
811 int nr_busy = pool->nr_workers - nr_idle;
813 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
820 /* Return the first idle worker. Safe with preemption disabled */
821 static struct worker *first_idle_worker(struct worker_pool *pool)
823 if (unlikely(list_empty(&pool->idle_list)))
826 return list_first_entry(&pool->idle_list, struct worker, entry);
830 * wake_up_worker - wake up an idle worker
831 * @pool: worker pool to wake worker from
833 * Wake up the first idle worker of @pool.
836 * spin_lock_irq(pool->lock).
838 static void wake_up_worker(struct worker_pool *pool)
840 struct worker *worker = first_idle_worker(pool);
843 wake_up_process(worker->task);
847 * wq_worker_running - a worker is running again
848 * @task: task waking up
850 * This function is called when a worker returns from schedule()
852 void wq_worker_running(struct task_struct *task)
854 struct worker *worker = kthread_data(task);
856 if (!worker->sleeping)
858 if (!(worker->flags & WORKER_NOT_RUNNING))
859 atomic_inc(&worker->pool->nr_running);
860 worker->sleeping = 0;
864 * wq_worker_sleeping - a worker is going to sleep
865 * @task: task going to sleep
867 * This function is called from schedule() when a busy worker is
870 void wq_worker_sleeping(struct task_struct *task)
872 struct worker *next, *worker = kthread_data(task);
873 struct worker_pool *pool;
876 * Rescuers, which may not have all the fields set up like normal
877 * workers, also reach here, let's not access anything before
878 * checking NOT_RUNNING.
880 if (worker->flags & WORKER_NOT_RUNNING)
885 if (WARN_ON_ONCE(worker->sleeping))
888 worker->sleeping = 1;
889 spin_lock_irq(&pool->lock);
892 * The counterpart of the following dec_and_test, implied mb,
893 * worklist not empty test sequence is in insert_work().
894 * Please read comment there.
896 * NOT_RUNNING is clear. This means that we're bound to and
897 * running on the local cpu w/ rq lock held and preemption
898 * disabled, which in turn means that none else could be
899 * manipulating idle_list, so dereferencing idle_list without pool
902 if (atomic_dec_and_test(&pool->nr_running) &&
903 !list_empty(&pool->worklist)) {
904 next = first_idle_worker(pool);
906 wake_up_process(next->task);
908 spin_unlock_irq(&pool->lock);
912 * wq_worker_last_func - retrieve worker's last work function
913 * @task: Task to retrieve last work function of.
915 * Determine the last function a worker executed. This is called from
916 * the scheduler to get a worker's last known identity.
919 * spin_lock_irq(rq->lock)
921 * This function is called during schedule() when a kworker is going
922 * to sleep. It's used by psi to identify aggregation workers during
923 * dequeuing, to allow periodic aggregation to shut-off when that
924 * worker is the last task in the system or cgroup to go to sleep.
926 * As this function doesn't involve any workqueue-related locking, it
927 * only returns stable values when called from inside the scheduler's
928 * queuing and dequeuing paths, when @task, which must be a kworker,
929 * is guaranteed to not be processing any works.
932 * The last work function %current executed as a worker, NULL if it
933 * hasn't executed any work yet.
935 work_func_t wq_worker_last_func(struct task_struct *task)
937 struct worker *worker = kthread_data(task);
939 return worker->last_func;
943 * worker_set_flags - set worker flags and adjust nr_running accordingly
945 * @flags: flags to set
947 * Set @flags in @worker->flags and adjust nr_running accordingly.
950 * spin_lock_irq(pool->lock)
952 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
954 struct worker_pool *pool = worker->pool;
956 WARN_ON_ONCE(worker->task != current);
958 /* If transitioning into NOT_RUNNING, adjust nr_running. */
959 if ((flags & WORKER_NOT_RUNNING) &&
960 !(worker->flags & WORKER_NOT_RUNNING)) {
961 atomic_dec(&pool->nr_running);
964 worker->flags |= flags;
968 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
970 * @flags: flags to clear
972 * Clear @flags in @worker->flags and adjust nr_running accordingly.
975 * spin_lock_irq(pool->lock)
977 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
979 struct worker_pool *pool = worker->pool;
980 unsigned int oflags = worker->flags;
982 WARN_ON_ONCE(worker->task != current);
984 worker->flags &= ~flags;
987 * If transitioning out of NOT_RUNNING, increment nr_running. Note
988 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
989 * of multiple flags, not a single flag.
991 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
992 if (!(worker->flags & WORKER_NOT_RUNNING))
993 atomic_inc(&pool->nr_running);
997 * find_worker_executing_work - find worker which is executing a work
998 * @pool: pool of interest
999 * @work: work to find worker for
1001 * Find a worker which is executing @work on @pool by searching
1002 * @pool->busy_hash which is keyed by the address of @work. For a worker
1003 * to match, its current execution should match the address of @work and
1004 * its work function. This is to avoid unwanted dependency between
1005 * unrelated work executions through a work item being recycled while still
1008 * This is a bit tricky. A work item may be freed once its execution
1009 * starts and nothing prevents the freed area from being recycled for
1010 * another work item. If the same work item address ends up being reused
1011 * before the original execution finishes, workqueue will identify the
1012 * recycled work item as currently executing and make it wait until the
1013 * current execution finishes, introducing an unwanted dependency.
1015 * This function checks the work item address and work function to avoid
1016 * false positives. Note that this isn't complete as one may construct a
1017 * work function which can introduce dependency onto itself through a
1018 * recycled work item. Well, if somebody wants to shoot oneself in the
1019 * foot that badly, there's only so much we can do, and if such deadlock
1020 * actually occurs, it should be easy to locate the culprit work function.
1023 * spin_lock_irq(pool->lock).
1026 * Pointer to worker which is executing @work if found, %NULL
1029 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1030 struct work_struct *work)
1032 struct worker *worker;
1034 hash_for_each_possible(pool->busy_hash, worker, hentry,
1035 (unsigned long)work)
1036 if (worker->current_work == work &&
1037 worker->current_func == work->func)
1044 * move_linked_works - move linked works to a list
1045 * @work: start of series of works to be scheduled
1046 * @head: target list to append @work to
1047 * @nextp: out parameter for nested worklist walking
1049 * Schedule linked works starting from @work to @head. Work series to
1050 * be scheduled starts at @work and includes any consecutive work with
1051 * WORK_STRUCT_LINKED set in its predecessor.
1053 * If @nextp is not NULL, it's updated to point to the next work of
1054 * the last scheduled work. This allows move_linked_works() to be
1055 * nested inside outer list_for_each_entry_safe().
1058 * spin_lock_irq(pool->lock).
1060 static void move_linked_works(struct work_struct *work, struct list_head *head,
1061 struct work_struct **nextp)
1063 struct work_struct *n;
1066 * Linked worklist will always end before the end of the list,
1067 * use NULL for list head.
1069 list_for_each_entry_safe_from(work, n, NULL, entry) {
1070 list_move_tail(&work->entry, head);
1071 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1076 * If we're already inside safe list traversal and have moved
1077 * multiple works to the scheduled queue, the next position
1078 * needs to be updated.
1085 * get_pwq - get an extra reference on the specified pool_workqueue
1086 * @pwq: pool_workqueue to get
1088 * Obtain an extra reference on @pwq. The caller should guarantee that
1089 * @pwq has positive refcnt and be holding the matching pool->lock.
1091 static void get_pwq(struct pool_workqueue *pwq)
1093 lockdep_assert_held(&pwq->pool->lock);
1094 WARN_ON_ONCE(pwq->refcnt <= 0);
1099 * put_pwq - put a pool_workqueue reference
1100 * @pwq: pool_workqueue to put
1102 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1103 * destruction. The caller should be holding the matching pool->lock.
1105 static void put_pwq(struct pool_workqueue *pwq)
1107 lockdep_assert_held(&pwq->pool->lock);
1108 if (likely(--pwq->refcnt))
1110 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1113 * @pwq can't be released under pool->lock, bounce to
1114 * pwq_unbound_release_workfn(). This never recurses on the same
1115 * pool->lock as this path is taken only for unbound workqueues and
1116 * the release work item is scheduled on a per-cpu workqueue. To
1117 * avoid lockdep warning, unbound pool->locks are given lockdep
1118 * subclass of 1 in get_unbound_pool().
1120 schedule_work(&pwq->unbound_release_work);
1124 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1125 * @pwq: pool_workqueue to put (can be %NULL)
1127 * put_pwq() with locking. This function also allows %NULL @pwq.
1129 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1133 * As both pwqs and pools are RCU protected, the
1134 * following lock operations are safe.
1136 spin_lock_irq(&pwq->pool->lock);
1138 spin_unlock_irq(&pwq->pool->lock);
1142 static void pwq_activate_delayed_work(struct work_struct *work)
1144 struct pool_workqueue *pwq = get_work_pwq(work);
1146 trace_workqueue_activate_work(work);
1147 if (list_empty(&pwq->pool->worklist))
1148 pwq->pool->watchdog_ts = jiffies;
1149 move_linked_works(work, &pwq->pool->worklist, NULL);
1150 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1154 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1156 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1157 struct work_struct, entry);
1159 pwq_activate_delayed_work(work);
1163 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1164 * @pwq: pwq of interest
1165 * @color: color of work which left the queue
1167 * A work either has completed or is removed from pending queue,
1168 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1171 * spin_lock_irq(pool->lock).
1173 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1175 /* uncolored work items don't participate in flushing or nr_active */
1176 if (color == WORK_NO_COLOR)
1179 pwq->nr_in_flight[color]--;
1182 if (!list_empty(&pwq->delayed_works)) {
1183 /* one down, submit a delayed one */
1184 if (pwq->nr_active < pwq->max_active)
1185 pwq_activate_first_delayed(pwq);
1188 /* is flush in progress and are we at the flushing tip? */
1189 if (likely(pwq->flush_color != color))
1192 /* are there still in-flight works? */
1193 if (pwq->nr_in_flight[color])
1196 /* this pwq is done, clear flush_color */
1197 pwq->flush_color = -1;
1200 * If this was the last pwq, wake up the first flusher. It
1201 * will handle the rest.
1203 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1204 complete(&pwq->wq->first_flusher->done);
1210 * try_to_grab_pending - steal work item from worklist and disable irq
1211 * @work: work item to steal
1212 * @is_dwork: @work is a delayed_work
1213 * @flags: place to store irq state
1215 * Try to grab PENDING bit of @work. This function can handle @work in any
1216 * stable state - idle, on timer or on worklist.
1219 * 1 if @work was pending and we successfully stole PENDING
1220 * 0 if @work was idle and we claimed PENDING
1221 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1222 * -ENOENT if someone else is canceling @work, this state may persist
1223 * for arbitrarily long
1226 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1227 * interrupted while holding PENDING and @work off queue, irq must be
1228 * disabled on entry. This, combined with delayed_work->timer being
1229 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1231 * On successful return, >= 0, irq is disabled and the caller is
1232 * responsible for releasing it using local_irq_restore(*@flags).
1234 * This function is safe to call from any context including IRQ handler.
1236 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1237 unsigned long *flags)
1239 struct worker_pool *pool;
1240 struct pool_workqueue *pwq;
1242 local_irq_save(*flags);
1244 /* try to steal the timer if it exists */
1246 struct delayed_work *dwork = to_delayed_work(work);
1249 * dwork->timer is irqsafe. If del_timer() fails, it's
1250 * guaranteed that the timer is not queued anywhere and not
1251 * running on the local CPU.
1253 if (likely(del_timer(&dwork->timer)))
1257 /* try to claim PENDING the normal way */
1258 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1263 * The queueing is in progress, or it is already queued. Try to
1264 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1266 pool = get_work_pool(work);
1270 spin_lock(&pool->lock);
1272 * work->data is guaranteed to point to pwq only while the work
1273 * item is queued on pwq->wq, and both updating work->data to point
1274 * to pwq on queueing and to pool on dequeueing are done under
1275 * pwq->pool->lock. This in turn guarantees that, if work->data
1276 * points to pwq which is associated with a locked pool, the work
1277 * item is currently queued on that pool.
1279 pwq = get_work_pwq(work);
1280 if (pwq && pwq->pool == pool) {
1281 debug_work_deactivate(work);
1284 * A delayed work item cannot be grabbed directly because
1285 * it might have linked NO_COLOR work items which, if left
1286 * on the delayed_list, will confuse pwq->nr_active
1287 * management later on and cause stall. Make sure the work
1288 * item is activated before grabbing.
1290 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1291 pwq_activate_delayed_work(work);
1293 list_del_init(&work->entry);
1294 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1296 /* work->data points to pwq iff queued, point to pool */
1297 set_work_pool_and_keep_pending(work, pool->id);
1299 spin_unlock(&pool->lock);
1303 spin_unlock(&pool->lock);
1306 local_irq_restore(*flags);
1307 if (work_is_canceling(work))
1314 * insert_work - insert a work into a pool
1315 * @pwq: pwq @work belongs to
1316 * @work: work to insert
1317 * @head: insertion point
1318 * @extra_flags: extra WORK_STRUCT_* flags to set
1320 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1321 * work_struct flags.
1324 * spin_lock_irq(pool->lock).
1326 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1327 struct list_head *head, unsigned int extra_flags)
1329 struct worker_pool *pool = pwq->pool;
1331 /* we own @work, set data and link */
1332 set_work_pwq(work, pwq, extra_flags);
1333 list_add_tail(&work->entry, head);
1337 * Ensure either wq_worker_sleeping() sees the above
1338 * list_add_tail() or we see zero nr_running to avoid workers lying
1339 * around lazily while there are works to be processed.
1343 if (__need_more_worker(pool))
1344 wake_up_worker(pool);
1348 * Test whether @work is being queued from another work executing on the
1351 static bool is_chained_work(struct workqueue_struct *wq)
1353 struct worker *worker;
1355 worker = current_wq_worker();
1357 * Return %true iff I'm a worker executing a work item on @wq. If
1358 * I'm @worker, it's safe to dereference it without locking.
1360 return worker && worker->current_pwq->wq == wq;
1364 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1365 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1366 * avoid perturbing sensitive tasks.
1368 static int wq_select_unbound_cpu(int cpu)
1370 static bool printed_dbg_warning;
1373 if (likely(!wq_debug_force_rr_cpu)) {
1374 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1376 } else if (!printed_dbg_warning) {
1377 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1378 printed_dbg_warning = true;
1381 if (cpumask_empty(wq_unbound_cpumask))
1384 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1385 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1386 if (unlikely(new_cpu >= nr_cpu_ids)) {
1387 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1388 if (unlikely(new_cpu >= nr_cpu_ids))
1391 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1396 static void __queue_work(int cpu, struct workqueue_struct *wq,
1397 struct work_struct *work)
1399 struct pool_workqueue *pwq;
1400 struct worker_pool *last_pool;
1401 struct list_head *worklist;
1402 unsigned int work_flags;
1403 unsigned int req_cpu = cpu;
1406 * While a work item is PENDING && off queue, a task trying to
1407 * steal the PENDING will busy-loop waiting for it to either get
1408 * queued or lose PENDING. Grabbing PENDING and queueing should
1409 * happen with IRQ disabled.
1411 lockdep_assert_irqs_disabled();
1413 debug_work_activate(work);
1415 /* if draining, only works from the same workqueue are allowed */
1416 if (unlikely(wq->flags & __WQ_DRAINING) &&
1417 WARN_ON_ONCE(!is_chained_work(wq)))
1421 if (req_cpu == WORK_CPU_UNBOUND)
1422 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1424 /* pwq which will be used unless @work is executing elsewhere */
1425 if (!(wq->flags & WQ_UNBOUND))
1426 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1428 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1431 * If @work was previously on a different pool, it might still be
1432 * running there, in which case the work needs to be queued on that
1433 * pool to guarantee non-reentrancy.
1435 last_pool = get_work_pool(work);
1436 if (last_pool && last_pool != pwq->pool) {
1437 struct worker *worker;
1439 spin_lock(&last_pool->lock);
1441 worker = find_worker_executing_work(last_pool, work);
1443 if (worker && worker->current_pwq->wq == wq) {
1444 pwq = worker->current_pwq;
1446 /* meh... not running there, queue here */
1447 spin_unlock(&last_pool->lock);
1448 spin_lock(&pwq->pool->lock);
1451 spin_lock(&pwq->pool->lock);
1455 * pwq is determined and locked. For unbound pools, we could have
1456 * raced with pwq release and it could already be dead. If its
1457 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1458 * without another pwq replacing it in the numa_pwq_tbl or while
1459 * work items are executing on it, so the retrying is guaranteed to
1460 * make forward-progress.
1462 if (unlikely(!pwq->refcnt)) {
1463 if (wq->flags & WQ_UNBOUND) {
1464 spin_unlock(&pwq->pool->lock);
1469 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1473 /* pwq determined, queue */
1474 trace_workqueue_queue_work(req_cpu, pwq, work);
1476 if (WARN_ON(!list_empty(&work->entry)))
1479 pwq->nr_in_flight[pwq->work_color]++;
1480 work_flags = work_color_to_flags(pwq->work_color);
1482 if (likely(pwq->nr_active < pwq->max_active)) {
1483 trace_workqueue_activate_work(work);
1485 worklist = &pwq->pool->worklist;
1486 if (list_empty(worklist))
1487 pwq->pool->watchdog_ts = jiffies;
1489 work_flags |= WORK_STRUCT_DELAYED;
1490 worklist = &pwq->delayed_works;
1493 insert_work(pwq, work, worklist, work_flags);
1496 spin_unlock(&pwq->pool->lock);
1501 * queue_work_on - queue work on specific cpu
1502 * @cpu: CPU number to execute work on
1503 * @wq: workqueue to use
1504 * @work: work to queue
1506 * We queue the work to a specific CPU, the caller must ensure it
1509 * Return: %false if @work was already on a queue, %true otherwise.
1511 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1512 struct work_struct *work)
1515 unsigned long flags;
1517 local_irq_save(flags);
1519 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1520 __queue_work(cpu, wq, work);
1524 local_irq_restore(flags);
1527 EXPORT_SYMBOL(queue_work_on);
1530 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1531 * @node: NUMA node ID that we want to select a CPU from
1533 * This function will attempt to find a "random" cpu available on a given
1534 * node. If there are no CPUs available on the given node it will return
1535 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1536 * available CPU if we need to schedule this work.
1538 static int workqueue_select_cpu_near(int node)
1542 /* No point in doing this if NUMA isn't enabled for workqueues */
1543 if (!wq_numa_enabled)
1544 return WORK_CPU_UNBOUND;
1546 /* Delay binding to CPU if node is not valid or online */
1547 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1548 return WORK_CPU_UNBOUND;
1550 /* Use local node/cpu if we are already there */
1551 cpu = raw_smp_processor_id();
1552 if (node == cpu_to_node(cpu))
1555 /* Use "random" otherwise know as "first" online CPU of node */
1556 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1558 /* If CPU is valid return that, otherwise just defer */
1559 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1563 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1564 * @node: NUMA node that we are targeting the work for
1565 * @wq: workqueue to use
1566 * @work: work to queue
1568 * We queue the work to a "random" CPU within a given NUMA node. The basic
1569 * idea here is to provide a way to somehow associate work with a given
1572 * This function will only make a best effort attempt at getting this onto
1573 * the right NUMA node. If no node is requested or the requested node is
1574 * offline then we just fall back to standard queue_work behavior.
1576 * Currently the "random" CPU ends up being the first available CPU in the
1577 * intersection of cpu_online_mask and the cpumask of the node, unless we
1578 * are running on the node. In that case we just use the current CPU.
1580 * Return: %false if @work was already on a queue, %true otherwise.
1582 bool queue_work_node(int node, struct workqueue_struct *wq,
1583 struct work_struct *work)
1585 unsigned long flags;
1589 * This current implementation is specific to unbound workqueues.
1590 * Specifically we only return the first available CPU for a given
1591 * node instead of cycling through individual CPUs within the node.
1593 * If this is used with a per-cpu workqueue then the logic in
1594 * workqueue_select_cpu_near would need to be updated to allow for
1595 * some round robin type logic.
1597 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1599 local_irq_save(flags);
1601 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1602 int cpu = workqueue_select_cpu_near(node);
1604 __queue_work(cpu, wq, work);
1608 local_irq_restore(flags);
1611 EXPORT_SYMBOL_GPL(queue_work_node);
1613 void delayed_work_timer_fn(struct timer_list *t)
1615 struct delayed_work *dwork = from_timer(dwork, t, timer);
1617 /* should have been called from irqsafe timer with irq already off */
1618 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1620 EXPORT_SYMBOL(delayed_work_timer_fn);
1622 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1623 struct delayed_work *dwork, unsigned long delay)
1625 struct timer_list *timer = &dwork->timer;
1626 struct work_struct *work = &dwork->work;
1629 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1630 WARN_ON_ONCE(timer_pending(timer));
1631 WARN_ON_ONCE(!list_empty(&work->entry));
1634 * If @delay is 0, queue @dwork->work immediately. This is for
1635 * both optimization and correctness. The earliest @timer can
1636 * expire is on the closest next tick and delayed_work users depend
1637 * on that there's no such delay when @delay is 0.
1640 __queue_work(cpu, wq, &dwork->work);
1646 timer->expires = jiffies + delay;
1648 if (unlikely(cpu != WORK_CPU_UNBOUND))
1649 add_timer_on(timer, cpu);
1655 * queue_delayed_work_on - queue work on specific CPU after delay
1656 * @cpu: CPU number to execute work on
1657 * @wq: workqueue to use
1658 * @dwork: work to queue
1659 * @delay: number of jiffies to wait before queueing
1661 * Return: %false if @work was already on a queue, %true otherwise. If
1662 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1665 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1666 struct delayed_work *dwork, unsigned long delay)
1668 struct work_struct *work = &dwork->work;
1670 unsigned long flags;
1672 /* read the comment in __queue_work() */
1673 local_irq_save(flags);
1675 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1676 __queue_delayed_work(cpu, wq, dwork, delay);
1680 local_irq_restore(flags);
1683 EXPORT_SYMBOL(queue_delayed_work_on);
1686 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1687 * @cpu: CPU number to execute work on
1688 * @wq: workqueue to use
1689 * @dwork: work to queue
1690 * @delay: number of jiffies to wait before queueing
1692 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1693 * modify @dwork's timer so that it expires after @delay. If @delay is
1694 * zero, @work is guaranteed to be scheduled immediately regardless of its
1697 * Return: %false if @dwork was idle and queued, %true if @dwork was
1698 * pending and its timer was modified.
1700 * This function is safe to call from any context including IRQ handler.
1701 * See try_to_grab_pending() for details.
1703 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1704 struct delayed_work *dwork, unsigned long delay)
1706 unsigned long flags;
1710 ret = try_to_grab_pending(&dwork->work, true, &flags);
1711 } while (unlikely(ret == -EAGAIN));
1713 if (likely(ret >= 0)) {
1714 __queue_delayed_work(cpu, wq, dwork, delay);
1715 local_irq_restore(flags);
1718 /* -ENOENT from try_to_grab_pending() becomes %true */
1721 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1723 static void rcu_work_rcufn(struct rcu_head *rcu)
1725 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1727 /* read the comment in __queue_work() */
1728 local_irq_disable();
1729 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1734 * queue_rcu_work - queue work after a RCU grace period
1735 * @wq: workqueue to use
1736 * @rwork: work to queue
1738 * Return: %false if @rwork was already pending, %true otherwise. Note
1739 * that a full RCU grace period is guaranteed only after a %true return.
1740 * While @rwork is guaranteed to be executed after a %false return, the
1741 * execution may happen before a full RCU grace period has passed.
1743 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1745 struct work_struct *work = &rwork->work;
1747 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1749 call_rcu(&rwork->rcu, rcu_work_rcufn);
1755 EXPORT_SYMBOL(queue_rcu_work);
1758 * worker_enter_idle - enter idle state
1759 * @worker: worker which is entering idle state
1761 * @worker is entering idle state. Update stats and idle timer if
1765 * spin_lock_irq(pool->lock).
1767 static void worker_enter_idle(struct worker *worker)
1769 struct worker_pool *pool = worker->pool;
1771 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1772 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1773 (worker->hentry.next || worker->hentry.pprev)))
1776 /* can't use worker_set_flags(), also called from create_worker() */
1777 worker->flags |= WORKER_IDLE;
1779 worker->last_active = jiffies;
1781 /* idle_list is LIFO */
1782 list_add(&worker->entry, &pool->idle_list);
1784 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1785 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1788 * Sanity check nr_running. Because unbind_workers() releases
1789 * pool->lock between setting %WORKER_UNBOUND and zapping
1790 * nr_running, the warning may trigger spuriously. Check iff
1791 * unbind is not in progress.
1793 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1794 pool->nr_workers == pool->nr_idle &&
1795 atomic_read(&pool->nr_running));
1799 * worker_leave_idle - leave idle state
1800 * @worker: worker which is leaving idle state
1802 * @worker is leaving idle state. Update stats.
1805 * spin_lock_irq(pool->lock).
1807 static void worker_leave_idle(struct worker *worker)
1809 struct worker_pool *pool = worker->pool;
1811 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1813 worker_clr_flags(worker, WORKER_IDLE);
1815 list_del_init(&worker->entry);
1818 static struct worker *alloc_worker(int node)
1820 struct worker *worker;
1822 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1824 INIT_LIST_HEAD(&worker->entry);
1825 INIT_LIST_HEAD(&worker->scheduled);
1826 INIT_LIST_HEAD(&worker->node);
1827 /* on creation a worker is in !idle && prep state */
1828 worker->flags = WORKER_PREP;
1834 * worker_attach_to_pool() - attach a worker to a pool
1835 * @worker: worker to be attached
1836 * @pool: the target pool
1838 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1839 * cpu-binding of @worker are kept coordinated with the pool across
1842 static void worker_attach_to_pool(struct worker *worker,
1843 struct worker_pool *pool)
1845 mutex_lock(&wq_pool_attach_mutex);
1848 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1849 * online CPUs. It'll be re-applied when any of the CPUs come up.
1851 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1854 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1855 * stable across this function. See the comments above the flag
1856 * definition for details.
1858 if (pool->flags & POOL_DISASSOCIATED)
1859 worker->flags |= WORKER_UNBOUND;
1861 list_add_tail(&worker->node, &pool->workers);
1862 worker->pool = pool;
1864 mutex_unlock(&wq_pool_attach_mutex);
1868 * worker_detach_from_pool() - detach a worker from its pool
1869 * @worker: worker which is attached to its pool
1871 * Undo the attaching which had been done in worker_attach_to_pool(). The
1872 * caller worker shouldn't access to the pool after detached except it has
1873 * other reference to the pool.
1875 static void worker_detach_from_pool(struct worker *worker)
1877 struct worker_pool *pool = worker->pool;
1878 struct completion *detach_completion = NULL;
1880 mutex_lock(&wq_pool_attach_mutex);
1882 list_del(&worker->node);
1883 worker->pool = NULL;
1885 if (list_empty(&pool->workers))
1886 detach_completion = pool->detach_completion;
1887 mutex_unlock(&wq_pool_attach_mutex);
1889 /* clear leftover flags without pool->lock after it is detached */
1890 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1892 if (detach_completion)
1893 complete(detach_completion);
1897 * create_worker - create a new workqueue worker
1898 * @pool: pool the new worker will belong to
1900 * Create and start a new worker which is attached to @pool.
1903 * Might sleep. Does GFP_KERNEL allocations.
1906 * Pointer to the newly created worker.
1908 static struct worker *create_worker(struct worker_pool *pool)
1910 struct worker *worker = NULL;
1914 /* ID is needed to determine kthread name */
1915 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1919 worker = alloc_worker(pool->node);
1926 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1927 pool->attrs->nice < 0 ? "H" : "");
1929 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1931 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1932 "kworker/%s", id_buf);
1933 if (IS_ERR(worker->task))
1936 set_user_nice(worker->task, pool->attrs->nice);
1937 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1939 /* successful, attach the worker to the pool */
1940 worker_attach_to_pool(worker, pool);
1942 /* start the newly created worker */
1943 spin_lock_irq(&pool->lock);
1944 worker->pool->nr_workers++;
1945 worker_enter_idle(worker);
1946 wake_up_process(worker->task);
1947 spin_unlock_irq(&pool->lock);
1953 ida_simple_remove(&pool->worker_ida, id);
1959 * destroy_worker - destroy a workqueue worker
1960 * @worker: worker to be destroyed
1962 * Destroy @worker and adjust @pool stats accordingly. The worker should
1966 * spin_lock_irq(pool->lock).
1968 static void destroy_worker(struct worker *worker)
1970 struct worker_pool *pool = worker->pool;
1972 lockdep_assert_held(&pool->lock);
1974 /* sanity check frenzy */
1975 if (WARN_ON(worker->current_work) ||
1976 WARN_ON(!list_empty(&worker->scheduled)) ||
1977 WARN_ON(!(worker->flags & WORKER_IDLE)))
1983 list_del_init(&worker->entry);
1984 worker->flags |= WORKER_DIE;
1985 wake_up_process(worker->task);
1988 static void idle_worker_timeout(struct timer_list *t)
1990 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1992 spin_lock_irq(&pool->lock);
1994 while (too_many_workers(pool)) {
1995 struct worker *worker;
1996 unsigned long expires;
1998 /* idle_list is kept in LIFO order, check the last one */
1999 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2000 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2002 if (time_before(jiffies, expires)) {
2003 mod_timer(&pool->idle_timer, expires);
2007 destroy_worker(worker);
2010 spin_unlock_irq(&pool->lock);
2013 static void send_mayday(struct work_struct *work)
2015 struct pool_workqueue *pwq = get_work_pwq(work);
2016 struct workqueue_struct *wq = pwq->wq;
2018 lockdep_assert_held(&wq_mayday_lock);
2023 /* mayday mayday mayday */
2024 if (list_empty(&pwq->mayday_node)) {
2026 * If @pwq is for an unbound wq, its base ref may be put at
2027 * any time due to an attribute change. Pin @pwq until the
2028 * rescuer is done with it.
2031 list_add_tail(&pwq->mayday_node, &wq->maydays);
2032 wake_up_process(wq->rescuer->task);
2036 static void pool_mayday_timeout(struct timer_list *t)
2038 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2039 struct work_struct *work;
2041 spin_lock_irq(&pool->lock);
2042 spin_lock(&wq_mayday_lock); /* for wq->maydays */
2044 if (need_to_create_worker(pool)) {
2046 * We've been trying to create a new worker but
2047 * haven't been successful. We might be hitting an
2048 * allocation deadlock. Send distress signals to
2051 list_for_each_entry(work, &pool->worklist, entry)
2055 spin_unlock(&wq_mayday_lock);
2056 spin_unlock_irq(&pool->lock);
2058 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2062 * maybe_create_worker - create a new worker if necessary
2063 * @pool: pool to create a new worker for
2065 * Create a new worker for @pool if necessary. @pool is guaranteed to
2066 * have at least one idle worker on return from this function. If
2067 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2068 * sent to all rescuers with works scheduled on @pool to resolve
2069 * possible allocation deadlock.
2071 * On return, need_to_create_worker() is guaranteed to be %false and
2072 * may_start_working() %true.
2075 * spin_lock_irq(pool->lock) which may be released and regrabbed
2076 * multiple times. Does GFP_KERNEL allocations. Called only from
2079 static void maybe_create_worker(struct worker_pool *pool)
2080 __releases(&pool->lock)
2081 __acquires(&pool->lock)
2084 spin_unlock_irq(&pool->lock);
2086 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2087 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2090 if (create_worker(pool) || !need_to_create_worker(pool))
2093 schedule_timeout_interruptible(CREATE_COOLDOWN);
2095 if (!need_to_create_worker(pool))
2099 del_timer_sync(&pool->mayday_timer);
2100 spin_lock_irq(&pool->lock);
2102 * This is necessary even after a new worker was just successfully
2103 * created as @pool->lock was dropped and the new worker might have
2104 * already become busy.
2106 if (need_to_create_worker(pool))
2111 * manage_workers - manage worker pool
2114 * Assume the manager role and manage the worker pool @worker belongs
2115 * to. At any given time, there can be only zero or one manager per
2116 * pool. The exclusion is handled automatically by this function.
2118 * The caller can safely start processing works on false return. On
2119 * true return, it's guaranteed that need_to_create_worker() is false
2120 * and may_start_working() is true.
2123 * spin_lock_irq(pool->lock) which may be released and regrabbed
2124 * multiple times. Does GFP_KERNEL allocations.
2127 * %false if the pool doesn't need management and the caller can safely
2128 * start processing works, %true if management function was performed and
2129 * the conditions that the caller verified before calling the function may
2130 * no longer be true.
2132 static bool manage_workers(struct worker *worker)
2134 struct worker_pool *pool = worker->pool;
2136 if (pool->flags & POOL_MANAGER_ACTIVE)
2139 pool->flags |= POOL_MANAGER_ACTIVE;
2140 pool->manager = worker;
2142 maybe_create_worker(pool);
2144 pool->manager = NULL;
2145 pool->flags &= ~POOL_MANAGER_ACTIVE;
2146 wake_up(&wq_manager_wait);
2151 * process_one_work - process single work
2153 * @work: work to process
2155 * Process @work. This function contains all the logics necessary to
2156 * process a single work including synchronization against and
2157 * interaction with other workers on the same cpu, queueing and
2158 * flushing. As long as context requirement is met, any worker can
2159 * call this function to process a work.
2162 * spin_lock_irq(pool->lock) which is released and regrabbed.
2164 static void process_one_work(struct worker *worker, struct work_struct *work)
2165 __releases(&pool->lock)
2166 __acquires(&pool->lock)
2168 struct pool_workqueue *pwq = get_work_pwq(work);
2169 struct worker_pool *pool = worker->pool;
2170 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2172 struct worker *collision;
2173 #ifdef CONFIG_LOCKDEP
2175 * It is permissible to free the struct work_struct from
2176 * inside the function that is called from it, this we need to
2177 * take into account for lockdep too. To avoid bogus "held
2178 * lock freed" warnings as well as problems when looking into
2179 * work->lockdep_map, make a copy and use that here.
2181 struct lockdep_map lockdep_map;
2183 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2185 /* ensure we're on the correct CPU */
2186 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2187 raw_smp_processor_id() != pool->cpu);
2190 * A single work shouldn't be executed concurrently by
2191 * multiple workers on a single cpu. Check whether anyone is
2192 * already processing the work. If so, defer the work to the
2193 * currently executing one.
2195 collision = find_worker_executing_work(pool, work);
2196 if (unlikely(collision)) {
2197 move_linked_works(work, &collision->scheduled, NULL);
2201 /* claim and dequeue */
2202 debug_work_deactivate(work);
2203 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2204 worker->current_work = work;
2205 worker->current_func = work->func;
2206 worker->current_pwq = pwq;
2207 work_color = get_work_color(work);
2210 * Record wq name for cmdline and debug reporting, may get
2211 * overridden through set_worker_desc().
2213 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2215 list_del_init(&work->entry);
2218 * CPU intensive works don't participate in concurrency management.
2219 * They're the scheduler's responsibility. This takes @worker out
2220 * of concurrency management and the next code block will chain
2221 * execution of the pending work items.
2223 if (unlikely(cpu_intensive))
2224 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2227 * Wake up another worker if necessary. The condition is always
2228 * false for normal per-cpu workers since nr_running would always
2229 * be >= 1 at this point. This is used to chain execution of the
2230 * pending work items for WORKER_NOT_RUNNING workers such as the
2231 * UNBOUND and CPU_INTENSIVE ones.
2233 if (need_more_worker(pool))
2234 wake_up_worker(pool);
2237 * Record the last pool and clear PENDING which should be the last
2238 * update to @work. Also, do this inside @pool->lock so that
2239 * PENDING and queued state changes happen together while IRQ is
2242 set_work_pool_and_clear_pending(work, pool->id);
2244 spin_unlock_irq(&pool->lock);
2246 lock_map_acquire(&pwq->wq->lockdep_map);
2247 lock_map_acquire(&lockdep_map);
2249 * Strictly speaking we should mark the invariant state without holding
2250 * any locks, that is, before these two lock_map_acquire()'s.
2252 * However, that would result in:
2259 * Which would create W1->C->W1 dependencies, even though there is no
2260 * actual deadlock possible. There are two solutions, using a
2261 * read-recursive acquire on the work(queue) 'locks', but this will then
2262 * hit the lockdep limitation on recursive locks, or simply discard
2265 * AFAICT there is no possible deadlock scenario between the
2266 * flush_work() and complete() primitives (except for single-threaded
2267 * workqueues), so hiding them isn't a problem.
2269 lockdep_invariant_state(true);
2270 trace_workqueue_execute_start(work);
2271 worker->current_func(work);
2273 * While we must be careful to not use "work" after this, the trace
2274 * point will only record its address.
2276 trace_workqueue_execute_end(work);
2277 lock_map_release(&lockdep_map);
2278 lock_map_release(&pwq->wq->lockdep_map);
2280 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2281 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2282 " last function: %ps\n",
2283 current->comm, preempt_count(), task_pid_nr(current),
2284 worker->current_func);
2285 debug_show_held_locks(current);
2290 * The following prevents a kworker from hogging CPU on !PREEMPT
2291 * kernels, where a requeueing work item waiting for something to
2292 * happen could deadlock with stop_machine as such work item could
2293 * indefinitely requeue itself while all other CPUs are trapped in
2294 * stop_machine. At the same time, report a quiescent RCU state so
2295 * the same condition doesn't freeze RCU.
2299 spin_lock_irq(&pool->lock);
2301 /* clear cpu intensive status */
2302 if (unlikely(cpu_intensive))
2303 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2305 /* tag the worker for identification in schedule() */
2306 worker->last_func = worker->current_func;
2308 /* we're done with it, release */
2309 hash_del(&worker->hentry);
2310 worker->current_work = NULL;
2311 worker->current_func = NULL;
2312 worker->current_pwq = NULL;
2313 pwq_dec_nr_in_flight(pwq, work_color);
2317 * process_scheduled_works - process scheduled works
2320 * Process all scheduled works. Please note that the scheduled list
2321 * may change while processing a work, so this function repeatedly
2322 * fetches a work from the top and executes it.
2325 * spin_lock_irq(pool->lock) which may be released and regrabbed
2328 static void process_scheduled_works(struct worker *worker)
2330 while (!list_empty(&worker->scheduled)) {
2331 struct work_struct *work = list_first_entry(&worker->scheduled,
2332 struct work_struct, entry);
2333 process_one_work(worker, work);
2337 static void set_pf_worker(bool val)
2339 mutex_lock(&wq_pool_attach_mutex);
2341 current->flags |= PF_WQ_WORKER;
2343 current->flags &= ~PF_WQ_WORKER;
2344 mutex_unlock(&wq_pool_attach_mutex);
2348 * worker_thread - the worker thread function
2351 * The worker thread function. All workers belong to a worker_pool -
2352 * either a per-cpu one or dynamic unbound one. These workers process all
2353 * work items regardless of their specific target workqueue. The only
2354 * exception is work items which belong to workqueues with a rescuer which
2355 * will be explained in rescuer_thread().
2359 static int worker_thread(void *__worker)
2361 struct worker *worker = __worker;
2362 struct worker_pool *pool = worker->pool;
2364 /* tell the scheduler that this is a workqueue worker */
2365 set_pf_worker(true);
2367 spin_lock_irq(&pool->lock);
2369 /* am I supposed to die? */
2370 if (unlikely(worker->flags & WORKER_DIE)) {
2371 spin_unlock_irq(&pool->lock);
2372 WARN_ON_ONCE(!list_empty(&worker->entry));
2373 set_pf_worker(false);
2375 set_task_comm(worker->task, "kworker/dying");
2376 ida_simple_remove(&pool->worker_ida, worker->id);
2377 worker_detach_from_pool(worker);
2382 worker_leave_idle(worker);
2384 /* no more worker necessary? */
2385 if (!need_more_worker(pool))
2388 /* do we need to manage? */
2389 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2393 * ->scheduled list can only be filled while a worker is
2394 * preparing to process a work or actually processing it.
2395 * Make sure nobody diddled with it while I was sleeping.
2397 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2400 * Finish PREP stage. We're guaranteed to have at least one idle
2401 * worker or that someone else has already assumed the manager
2402 * role. This is where @worker starts participating in concurrency
2403 * management if applicable and concurrency management is restored
2404 * after being rebound. See rebind_workers() for details.
2406 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2409 struct work_struct *work =
2410 list_first_entry(&pool->worklist,
2411 struct work_struct, entry);
2413 pool->watchdog_ts = jiffies;
2415 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2416 /* optimization path, not strictly necessary */
2417 process_one_work(worker, work);
2418 if (unlikely(!list_empty(&worker->scheduled)))
2419 process_scheduled_works(worker);
2421 move_linked_works(work, &worker->scheduled, NULL);
2422 process_scheduled_works(worker);
2424 } while (keep_working(pool));
2426 worker_set_flags(worker, WORKER_PREP);
2429 * pool->lock is held and there's no work to process and no need to
2430 * manage, sleep. Workers are woken up only while holding
2431 * pool->lock or from local cpu, so setting the current state
2432 * before releasing pool->lock is enough to prevent losing any
2435 worker_enter_idle(worker);
2436 __set_current_state(TASK_IDLE);
2437 spin_unlock_irq(&pool->lock);
2443 * rescuer_thread - the rescuer thread function
2446 * Workqueue rescuer thread function. There's one rescuer for each
2447 * workqueue which has WQ_MEM_RECLAIM set.
2449 * Regular work processing on a pool may block trying to create a new
2450 * worker which uses GFP_KERNEL allocation which has slight chance of
2451 * developing into deadlock if some works currently on the same queue
2452 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2453 * the problem rescuer solves.
2455 * When such condition is possible, the pool summons rescuers of all
2456 * workqueues which have works queued on the pool and let them process
2457 * those works so that forward progress can be guaranteed.
2459 * This should happen rarely.
2463 static int rescuer_thread(void *__rescuer)
2465 struct worker *rescuer = __rescuer;
2466 struct workqueue_struct *wq = rescuer->rescue_wq;
2467 struct list_head *scheduled = &rescuer->scheduled;
2470 set_user_nice(current, RESCUER_NICE_LEVEL);
2473 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2474 * doesn't participate in concurrency management.
2476 set_pf_worker(true);
2478 set_current_state(TASK_IDLE);
2481 * By the time the rescuer is requested to stop, the workqueue
2482 * shouldn't have any work pending, but @wq->maydays may still have
2483 * pwq(s) queued. This can happen by non-rescuer workers consuming
2484 * all the work items before the rescuer got to them. Go through
2485 * @wq->maydays processing before acting on should_stop so that the
2486 * list is always empty on exit.
2488 should_stop = kthread_should_stop();
2490 /* see whether any pwq is asking for help */
2491 spin_lock_irq(&wq_mayday_lock);
2493 while (!list_empty(&wq->maydays)) {
2494 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2495 struct pool_workqueue, mayday_node);
2496 struct worker_pool *pool = pwq->pool;
2497 struct work_struct *work, *n;
2500 __set_current_state(TASK_RUNNING);
2501 list_del_init(&pwq->mayday_node);
2503 spin_unlock_irq(&wq_mayday_lock);
2505 worker_attach_to_pool(rescuer, pool);
2507 spin_lock_irq(&pool->lock);
2510 * Slurp in all works issued via this workqueue and
2513 WARN_ON_ONCE(!list_empty(scheduled));
2514 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2515 if (get_work_pwq(work) == pwq) {
2517 pool->watchdog_ts = jiffies;
2518 move_linked_works(work, scheduled, &n);
2523 if (!list_empty(scheduled)) {
2524 process_scheduled_works(rescuer);
2527 * The above execution of rescued work items could
2528 * have created more to rescue through
2529 * pwq_activate_first_delayed() or chained
2530 * queueing. Let's put @pwq back on mayday list so
2531 * that such back-to-back work items, which may be
2532 * being used to relieve memory pressure, don't
2533 * incur MAYDAY_INTERVAL delay inbetween.
2535 if (need_to_create_worker(pool)) {
2536 spin_lock(&wq_mayday_lock);
2538 * Queue iff we aren't racing destruction
2539 * and somebody else hasn't queued it already.
2541 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2543 list_add_tail(&pwq->mayday_node, &wq->maydays);
2545 spin_unlock(&wq_mayday_lock);
2550 * Put the reference grabbed by send_mayday(). @pool won't
2551 * go away while we're still attached to it.
2556 * Leave this pool. If need_more_worker() is %true, notify a
2557 * regular worker; otherwise, we end up with 0 concurrency
2558 * and stalling the execution.
2560 if (need_more_worker(pool))
2561 wake_up_worker(pool);
2563 spin_unlock_irq(&pool->lock);
2565 worker_detach_from_pool(rescuer);
2567 spin_lock_irq(&wq_mayday_lock);
2570 spin_unlock_irq(&wq_mayday_lock);
2573 __set_current_state(TASK_RUNNING);
2574 set_pf_worker(false);
2578 /* rescuers should never participate in concurrency management */
2579 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2585 * check_flush_dependency - check for flush dependency sanity
2586 * @target_wq: workqueue being flushed
2587 * @target_work: work item being flushed (NULL for workqueue flushes)
2589 * %current is trying to flush the whole @target_wq or @target_work on it.
2590 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2591 * reclaiming memory or running on a workqueue which doesn't have
2592 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2595 static void check_flush_dependency(struct workqueue_struct *target_wq,
2596 struct work_struct *target_work)
2598 work_func_t target_func = target_work ? target_work->func : NULL;
2599 struct worker *worker;
2601 if (target_wq->flags & WQ_MEM_RECLAIM)
2604 worker = current_wq_worker();
2606 WARN_ONCE(current->flags & PF_MEMALLOC,
2607 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2608 current->pid, current->comm, target_wq->name, target_func);
2609 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2610 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2611 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2612 worker->current_pwq->wq->name, worker->current_func,
2613 target_wq->name, target_func);
2617 struct work_struct work;
2618 struct completion done;
2619 struct task_struct *task; /* purely informational */
2622 static void wq_barrier_func(struct work_struct *work)
2624 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2625 complete(&barr->done);
2629 * insert_wq_barrier - insert a barrier work
2630 * @pwq: pwq to insert barrier into
2631 * @barr: wq_barrier to insert
2632 * @target: target work to attach @barr to
2633 * @worker: worker currently executing @target, NULL if @target is not executing
2635 * @barr is linked to @target such that @barr is completed only after
2636 * @target finishes execution. Please note that the ordering
2637 * guarantee is observed only with respect to @target and on the local
2640 * Currently, a queued barrier can't be canceled. This is because
2641 * try_to_grab_pending() can't determine whether the work to be
2642 * grabbed is at the head of the queue and thus can't clear LINKED
2643 * flag of the previous work while there must be a valid next work
2644 * after a work with LINKED flag set.
2646 * Note that when @worker is non-NULL, @target may be modified
2647 * underneath us, so we can't reliably determine pwq from @target.
2650 * spin_lock_irq(pool->lock).
2652 static void insert_wq_barrier(struct pool_workqueue *pwq,
2653 struct wq_barrier *barr,
2654 struct work_struct *target, struct worker *worker)
2656 struct list_head *head;
2657 unsigned int linked = 0;
2660 * debugobject calls are safe here even with pool->lock locked
2661 * as we know for sure that this will not trigger any of the
2662 * checks and call back into the fixup functions where we
2665 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2666 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2668 init_completion_map(&barr->done, &target->lockdep_map);
2670 barr->task = current;
2673 * If @target is currently being executed, schedule the
2674 * barrier to the worker; otherwise, put it after @target.
2677 head = worker->scheduled.next;
2679 unsigned long *bits = work_data_bits(target);
2681 head = target->entry.next;
2682 /* there can already be other linked works, inherit and set */
2683 linked = *bits & WORK_STRUCT_LINKED;
2684 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2687 debug_work_activate(&barr->work);
2688 insert_work(pwq, &barr->work, head,
2689 work_color_to_flags(WORK_NO_COLOR) | linked);
2693 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2694 * @wq: workqueue being flushed
2695 * @flush_color: new flush color, < 0 for no-op
2696 * @work_color: new work color, < 0 for no-op
2698 * Prepare pwqs for workqueue flushing.
2700 * If @flush_color is non-negative, flush_color on all pwqs should be
2701 * -1. If no pwq has in-flight commands at the specified color, all
2702 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2703 * has in flight commands, its pwq->flush_color is set to
2704 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2705 * wakeup logic is armed and %true is returned.
2707 * The caller should have initialized @wq->first_flusher prior to
2708 * calling this function with non-negative @flush_color. If
2709 * @flush_color is negative, no flush color update is done and %false
2712 * If @work_color is non-negative, all pwqs should have the same
2713 * work_color which is previous to @work_color and all will be
2714 * advanced to @work_color.
2717 * mutex_lock(wq->mutex).
2720 * %true if @flush_color >= 0 and there's something to flush. %false
2723 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2724 int flush_color, int work_color)
2727 struct pool_workqueue *pwq;
2729 if (flush_color >= 0) {
2730 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2731 atomic_set(&wq->nr_pwqs_to_flush, 1);
2734 for_each_pwq(pwq, wq) {
2735 struct worker_pool *pool = pwq->pool;
2737 spin_lock_irq(&pool->lock);
2739 if (flush_color >= 0) {
2740 WARN_ON_ONCE(pwq->flush_color != -1);
2742 if (pwq->nr_in_flight[flush_color]) {
2743 pwq->flush_color = flush_color;
2744 atomic_inc(&wq->nr_pwqs_to_flush);
2749 if (work_color >= 0) {
2750 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2751 pwq->work_color = work_color;
2754 spin_unlock_irq(&pool->lock);
2757 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2758 complete(&wq->first_flusher->done);
2764 * flush_workqueue - ensure that any scheduled work has run to completion.
2765 * @wq: workqueue to flush
2767 * This function sleeps until all work items which were queued on entry
2768 * have finished execution, but it is not livelocked by new incoming ones.
2770 void flush_workqueue(struct workqueue_struct *wq)
2772 struct wq_flusher this_flusher = {
2773 .list = LIST_HEAD_INIT(this_flusher.list),
2775 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2779 if (WARN_ON(!wq_online))
2782 lock_map_acquire(&wq->lockdep_map);
2783 lock_map_release(&wq->lockdep_map);
2785 mutex_lock(&wq->mutex);
2788 * Start-to-wait phase
2790 next_color = work_next_color(wq->work_color);
2792 if (next_color != wq->flush_color) {
2794 * Color space is not full. The current work_color
2795 * becomes our flush_color and work_color is advanced
2798 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2799 this_flusher.flush_color = wq->work_color;
2800 wq->work_color = next_color;
2802 if (!wq->first_flusher) {
2803 /* no flush in progress, become the first flusher */
2804 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2806 wq->first_flusher = &this_flusher;
2808 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2810 /* nothing to flush, done */
2811 wq->flush_color = next_color;
2812 wq->first_flusher = NULL;
2817 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2818 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2819 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2823 * Oops, color space is full, wait on overflow queue.
2824 * The next flush completion will assign us
2825 * flush_color and transfer to flusher_queue.
2827 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2830 check_flush_dependency(wq, NULL);
2832 mutex_unlock(&wq->mutex);
2834 wait_for_completion(&this_flusher.done);
2837 * Wake-up-and-cascade phase
2839 * First flushers are responsible for cascading flushes and
2840 * handling overflow. Non-first flushers can simply return.
2842 if (wq->first_flusher != &this_flusher)
2845 mutex_lock(&wq->mutex);
2847 /* we might have raced, check again with mutex held */
2848 if (wq->first_flusher != &this_flusher)
2851 wq->first_flusher = NULL;
2853 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2854 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2857 struct wq_flusher *next, *tmp;
2859 /* complete all the flushers sharing the current flush color */
2860 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2861 if (next->flush_color != wq->flush_color)
2863 list_del_init(&next->list);
2864 complete(&next->done);
2867 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2868 wq->flush_color != work_next_color(wq->work_color));
2870 /* this flush_color is finished, advance by one */
2871 wq->flush_color = work_next_color(wq->flush_color);
2873 /* one color has been freed, handle overflow queue */
2874 if (!list_empty(&wq->flusher_overflow)) {
2876 * Assign the same color to all overflowed
2877 * flushers, advance work_color and append to
2878 * flusher_queue. This is the start-to-wait
2879 * phase for these overflowed flushers.
2881 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2882 tmp->flush_color = wq->work_color;
2884 wq->work_color = work_next_color(wq->work_color);
2886 list_splice_tail_init(&wq->flusher_overflow,
2887 &wq->flusher_queue);
2888 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2891 if (list_empty(&wq->flusher_queue)) {
2892 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2897 * Need to flush more colors. Make the next flusher
2898 * the new first flusher and arm pwqs.
2900 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2901 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2903 list_del_init(&next->list);
2904 wq->first_flusher = next;
2906 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2910 * Meh... this color is already done, clear first
2911 * flusher and repeat cascading.
2913 wq->first_flusher = NULL;
2917 mutex_unlock(&wq->mutex);
2919 EXPORT_SYMBOL(flush_workqueue);
2922 * drain_workqueue - drain a workqueue
2923 * @wq: workqueue to drain
2925 * Wait until the workqueue becomes empty. While draining is in progress,
2926 * only chain queueing is allowed. IOW, only currently pending or running
2927 * work items on @wq can queue further work items on it. @wq is flushed
2928 * repeatedly until it becomes empty. The number of flushing is determined
2929 * by the depth of chaining and should be relatively short. Whine if it
2932 void drain_workqueue(struct workqueue_struct *wq)
2934 unsigned int flush_cnt = 0;
2935 struct pool_workqueue *pwq;
2938 * __queue_work() needs to test whether there are drainers, is much
2939 * hotter than drain_workqueue() and already looks at @wq->flags.
2940 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2942 mutex_lock(&wq->mutex);
2943 if (!wq->nr_drainers++)
2944 wq->flags |= __WQ_DRAINING;
2945 mutex_unlock(&wq->mutex);
2947 flush_workqueue(wq);
2949 mutex_lock(&wq->mutex);
2951 for_each_pwq(pwq, wq) {
2954 spin_lock_irq(&pwq->pool->lock);
2955 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2956 spin_unlock_irq(&pwq->pool->lock);
2961 if (++flush_cnt == 10 ||
2962 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2963 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2964 wq->name, flush_cnt);
2966 mutex_unlock(&wq->mutex);
2970 if (!--wq->nr_drainers)
2971 wq->flags &= ~__WQ_DRAINING;
2972 mutex_unlock(&wq->mutex);
2974 EXPORT_SYMBOL_GPL(drain_workqueue);
2976 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2979 struct worker *worker = NULL;
2980 struct worker_pool *pool;
2981 struct pool_workqueue *pwq;
2986 pool = get_work_pool(work);
2992 spin_lock_irq(&pool->lock);
2993 /* see the comment in try_to_grab_pending() with the same code */
2994 pwq = get_work_pwq(work);
2996 if (unlikely(pwq->pool != pool))
2999 worker = find_worker_executing_work(pool, work);
3002 pwq = worker->current_pwq;
3005 check_flush_dependency(pwq->wq, work);
3007 insert_wq_barrier(pwq, barr, work, worker);
3008 spin_unlock_irq(&pool->lock);
3011 * Force a lock recursion deadlock when using flush_work() inside a
3012 * single-threaded or rescuer equipped workqueue.
3014 * For single threaded workqueues the deadlock happens when the work
3015 * is after the work issuing the flush_work(). For rescuer equipped
3016 * workqueues the deadlock happens when the rescuer stalls, blocking
3020 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3021 lock_map_acquire(&pwq->wq->lockdep_map);
3022 lock_map_release(&pwq->wq->lockdep_map);
3027 spin_unlock_irq(&pool->lock);
3032 static bool __flush_work(struct work_struct *work, bool from_cancel)
3034 struct wq_barrier barr;
3036 if (WARN_ON(!wq_online))
3039 if (WARN_ON(!work->func))
3043 lock_map_acquire(&work->lockdep_map);
3044 lock_map_release(&work->lockdep_map);
3047 if (start_flush_work(work, &barr, from_cancel)) {
3048 wait_for_completion(&barr.done);
3049 destroy_work_on_stack(&barr.work);
3057 * flush_work - wait for a work to finish executing the last queueing instance
3058 * @work: the work to flush
3060 * Wait until @work has finished execution. @work is guaranteed to be idle
3061 * on return if it hasn't been requeued since flush started.
3064 * %true if flush_work() waited for the work to finish execution,
3065 * %false if it was already idle.
3067 bool flush_work(struct work_struct *work)
3069 return __flush_work(work, false);
3071 EXPORT_SYMBOL_GPL(flush_work);
3074 wait_queue_entry_t wait;
3075 struct work_struct *work;
3078 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3080 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3082 if (cwait->work != key)
3084 return autoremove_wake_function(wait, mode, sync, key);
3087 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3089 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3090 unsigned long flags;
3094 ret = try_to_grab_pending(work, is_dwork, &flags);
3096 * If someone else is already canceling, wait for it to
3097 * finish. flush_work() doesn't work for PREEMPT_NONE
3098 * because we may get scheduled between @work's completion
3099 * and the other canceling task resuming and clearing
3100 * CANCELING - flush_work() will return false immediately
3101 * as @work is no longer busy, try_to_grab_pending() will
3102 * return -ENOENT as @work is still being canceled and the
3103 * other canceling task won't be able to clear CANCELING as
3104 * we're hogging the CPU.
3106 * Let's wait for completion using a waitqueue. As this
3107 * may lead to the thundering herd problem, use a custom
3108 * wake function which matches @work along with exclusive
3111 if (unlikely(ret == -ENOENT)) {
3112 struct cwt_wait cwait;
3114 init_wait(&cwait.wait);
3115 cwait.wait.func = cwt_wakefn;
3118 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3119 TASK_UNINTERRUPTIBLE);
3120 if (work_is_canceling(work))
3122 finish_wait(&cancel_waitq, &cwait.wait);
3124 } while (unlikely(ret < 0));
3126 /* tell other tasks trying to grab @work to back off */
3127 mark_work_canceling(work);
3128 local_irq_restore(flags);
3131 * This allows canceling during early boot. We know that @work
3135 __flush_work(work, true);
3137 clear_work_data(work);
3140 * Paired with prepare_to_wait() above so that either
3141 * waitqueue_active() is visible here or !work_is_canceling() is
3145 if (waitqueue_active(&cancel_waitq))
3146 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3152 * cancel_work_sync - cancel a work and wait for it to finish
3153 * @work: the work to cancel
3155 * Cancel @work and wait for its execution to finish. This function
3156 * can be used even if the work re-queues itself or migrates to
3157 * another workqueue. On return from this function, @work is
3158 * guaranteed to be not pending or executing on any CPU.
3160 * cancel_work_sync(&delayed_work->work) must not be used for
3161 * delayed_work's. Use cancel_delayed_work_sync() instead.
3163 * The caller must ensure that the workqueue on which @work was last
3164 * queued can't be destroyed before this function returns.
3167 * %true if @work was pending, %false otherwise.
3169 bool cancel_work_sync(struct work_struct *work)
3171 return __cancel_work_timer(work, false);
3173 EXPORT_SYMBOL_GPL(cancel_work_sync);
3176 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3177 * @dwork: the delayed work to flush
3179 * Delayed timer is cancelled and the pending work is queued for
3180 * immediate execution. Like flush_work(), this function only
3181 * considers the last queueing instance of @dwork.
3184 * %true if flush_work() waited for the work to finish execution,
3185 * %false if it was already idle.
3187 bool flush_delayed_work(struct delayed_work *dwork)
3189 local_irq_disable();
3190 if (del_timer_sync(&dwork->timer))
3191 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3193 return flush_work(&dwork->work);
3195 EXPORT_SYMBOL(flush_delayed_work);
3198 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3199 * @rwork: the rcu work to flush
3202 * %true if flush_rcu_work() waited for the work to finish execution,
3203 * %false if it was already idle.
3205 bool flush_rcu_work(struct rcu_work *rwork)
3207 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3209 flush_work(&rwork->work);
3212 return flush_work(&rwork->work);
3215 EXPORT_SYMBOL(flush_rcu_work);
3217 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3219 unsigned long flags;
3223 ret = try_to_grab_pending(work, is_dwork, &flags);
3224 } while (unlikely(ret == -EAGAIN));
3226 if (unlikely(ret < 0))
3229 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3230 local_irq_restore(flags);
3235 * cancel_delayed_work - cancel a delayed work
3236 * @dwork: delayed_work to cancel
3238 * Kill off a pending delayed_work.
3240 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3244 * The work callback function may still be running on return, unless
3245 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3246 * use cancel_delayed_work_sync() to wait on it.
3248 * This function is safe to call from any context including IRQ handler.
3250 bool cancel_delayed_work(struct delayed_work *dwork)
3252 return __cancel_work(&dwork->work, true);
3254 EXPORT_SYMBOL(cancel_delayed_work);
3257 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3258 * @dwork: the delayed work cancel
3260 * This is cancel_work_sync() for delayed works.
3263 * %true if @dwork was pending, %false otherwise.
3265 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3267 return __cancel_work_timer(&dwork->work, true);
3269 EXPORT_SYMBOL(cancel_delayed_work_sync);
3272 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3273 * @func: the function to call
3275 * schedule_on_each_cpu() executes @func on each online CPU using the
3276 * system workqueue and blocks until all CPUs have completed.
3277 * schedule_on_each_cpu() is very slow.
3280 * 0 on success, -errno on failure.
3282 int schedule_on_each_cpu(work_func_t func)
3285 struct work_struct __percpu *works;
3287 works = alloc_percpu(struct work_struct);
3293 for_each_online_cpu(cpu) {
3294 struct work_struct *work = per_cpu_ptr(works, cpu);
3296 INIT_WORK(work, func);
3297 schedule_work_on(cpu, work);
3300 for_each_online_cpu(cpu)
3301 flush_work(per_cpu_ptr(works, cpu));
3309 * execute_in_process_context - reliably execute the routine with user context
3310 * @fn: the function to execute
3311 * @ew: guaranteed storage for the execute work structure (must
3312 * be available when the work executes)
3314 * Executes the function immediately if process context is available,
3315 * otherwise schedules the function for delayed execution.
3317 * Return: 0 - function was executed
3318 * 1 - function was scheduled for execution
3320 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3322 if (!in_interrupt()) {
3327 INIT_WORK(&ew->work, fn);
3328 schedule_work(&ew->work);
3332 EXPORT_SYMBOL_GPL(execute_in_process_context);
3335 * free_workqueue_attrs - free a workqueue_attrs
3336 * @attrs: workqueue_attrs to free
3338 * Undo alloc_workqueue_attrs().
3340 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3343 free_cpumask_var(attrs->cpumask);
3349 * alloc_workqueue_attrs - allocate a workqueue_attrs
3351 * Allocate a new workqueue_attrs, initialize with default settings and
3354 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3356 struct workqueue_attrs *alloc_workqueue_attrs(void)
3358 struct workqueue_attrs *attrs;
3360 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3363 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3366 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3369 free_workqueue_attrs(attrs);
3373 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3374 const struct workqueue_attrs *from)
3376 to->nice = from->nice;
3377 cpumask_copy(to->cpumask, from->cpumask);
3379 * Unlike hash and equality test, this function doesn't ignore
3380 * ->no_numa as it is used for both pool and wq attrs. Instead,
3381 * get_unbound_pool() explicitly clears ->no_numa after copying.
3383 to->no_numa = from->no_numa;
3386 /* hash value of the content of @attr */
3387 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3391 hash = jhash_1word(attrs->nice, hash);
3392 hash = jhash(cpumask_bits(attrs->cpumask),
3393 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3397 /* content equality test */
3398 static bool wqattrs_equal(const struct workqueue_attrs *a,
3399 const struct workqueue_attrs *b)
3401 if (a->nice != b->nice)
3403 if (!cpumask_equal(a->cpumask, b->cpumask))
3409 * init_worker_pool - initialize a newly zalloc'd worker_pool
3410 * @pool: worker_pool to initialize
3412 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3414 * Return: 0 on success, -errno on failure. Even on failure, all fields
3415 * inside @pool proper are initialized and put_unbound_pool() can be called
3416 * on @pool safely to release it.
3418 static int init_worker_pool(struct worker_pool *pool)
3420 spin_lock_init(&pool->lock);
3423 pool->node = NUMA_NO_NODE;
3424 pool->flags |= POOL_DISASSOCIATED;
3425 pool->watchdog_ts = jiffies;
3426 INIT_LIST_HEAD(&pool->worklist);
3427 INIT_LIST_HEAD(&pool->idle_list);
3428 hash_init(pool->busy_hash);
3430 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3432 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3434 INIT_LIST_HEAD(&pool->workers);
3436 ida_init(&pool->worker_ida);
3437 INIT_HLIST_NODE(&pool->hash_node);
3440 /* shouldn't fail above this point */
3441 pool->attrs = alloc_workqueue_attrs();
3447 #ifdef CONFIG_LOCKDEP
3448 static void wq_init_lockdep(struct workqueue_struct *wq)
3452 lockdep_register_key(&wq->key);
3453 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3455 lock_name = wq->name;
3457 wq->lock_name = lock_name;
3458 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3461 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3463 lockdep_unregister_key(&wq->key);
3466 static void wq_free_lockdep(struct workqueue_struct *wq)
3468 if (wq->lock_name != wq->name)
3469 kfree(wq->lock_name);
3472 static void wq_init_lockdep(struct workqueue_struct *wq)
3476 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3480 static void wq_free_lockdep(struct workqueue_struct *wq)
3485 static void rcu_free_wq(struct rcu_head *rcu)
3487 struct workqueue_struct *wq =
3488 container_of(rcu, struct workqueue_struct, rcu);
3490 wq_free_lockdep(wq);
3492 if (!(wq->flags & WQ_UNBOUND))
3493 free_percpu(wq->cpu_pwqs);
3495 free_workqueue_attrs(wq->unbound_attrs);
3501 static void rcu_free_pool(struct rcu_head *rcu)
3503 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3505 ida_destroy(&pool->worker_ida);
3506 free_workqueue_attrs(pool->attrs);
3511 * put_unbound_pool - put a worker_pool
3512 * @pool: worker_pool to put
3514 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3515 * safe manner. get_unbound_pool() calls this function on its failure path
3516 * and this function should be able to release pools which went through,
3517 * successfully or not, init_worker_pool().
3519 * Should be called with wq_pool_mutex held.
3521 static void put_unbound_pool(struct worker_pool *pool)
3523 DECLARE_COMPLETION_ONSTACK(detach_completion);
3524 struct worker *worker;
3526 lockdep_assert_held(&wq_pool_mutex);
3532 if (WARN_ON(!(pool->cpu < 0)) ||
3533 WARN_ON(!list_empty(&pool->worklist)))
3536 /* release id and unhash */
3538 idr_remove(&worker_pool_idr, pool->id);
3539 hash_del(&pool->hash_node);
3542 * Become the manager and destroy all workers. This prevents
3543 * @pool's workers from blocking on attach_mutex. We're the last
3544 * manager and @pool gets freed with the flag set.
3546 spin_lock_irq(&pool->lock);
3547 wait_event_lock_irq(wq_manager_wait,
3548 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3549 pool->flags |= POOL_MANAGER_ACTIVE;
3551 while ((worker = first_idle_worker(pool)))
3552 destroy_worker(worker);
3553 WARN_ON(pool->nr_workers || pool->nr_idle);
3554 spin_unlock_irq(&pool->lock);
3556 mutex_lock(&wq_pool_attach_mutex);
3557 if (!list_empty(&pool->workers))
3558 pool->detach_completion = &detach_completion;
3559 mutex_unlock(&wq_pool_attach_mutex);
3561 if (pool->detach_completion)
3562 wait_for_completion(pool->detach_completion);
3564 /* shut down the timers */
3565 del_timer_sync(&pool->idle_timer);
3566 del_timer_sync(&pool->mayday_timer);
3568 /* RCU protected to allow dereferences from get_work_pool() */
3569 call_rcu(&pool->rcu, rcu_free_pool);
3573 * get_unbound_pool - get a worker_pool with the specified attributes
3574 * @attrs: the attributes of the worker_pool to get
3576 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3577 * reference count and return it. If there already is a matching
3578 * worker_pool, it will be used; otherwise, this function attempts to
3581 * Should be called with wq_pool_mutex held.
3583 * Return: On success, a worker_pool with the same attributes as @attrs.
3584 * On failure, %NULL.
3586 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3588 u32 hash = wqattrs_hash(attrs);
3589 struct worker_pool *pool;
3591 int target_node = NUMA_NO_NODE;
3593 lockdep_assert_held(&wq_pool_mutex);
3595 /* do we already have a matching pool? */
3596 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3597 if (wqattrs_equal(pool->attrs, attrs)) {
3603 /* if cpumask is contained inside a NUMA node, we belong to that node */
3604 if (wq_numa_enabled) {
3605 for_each_node(node) {
3606 if (cpumask_subset(attrs->cpumask,
3607 wq_numa_possible_cpumask[node])) {
3614 /* nope, create a new one */
3615 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3616 if (!pool || init_worker_pool(pool) < 0)
3619 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3620 copy_workqueue_attrs(pool->attrs, attrs);
3621 pool->node = target_node;
3624 * no_numa isn't a worker_pool attribute, always clear it. See
3625 * 'struct workqueue_attrs' comments for detail.
3627 pool->attrs->no_numa = false;
3629 if (worker_pool_assign_id(pool) < 0)
3632 /* create and start the initial worker */
3633 if (wq_online && !create_worker(pool))
3637 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3642 put_unbound_pool(pool);
3646 static void rcu_free_pwq(struct rcu_head *rcu)
3648 kmem_cache_free(pwq_cache,
3649 container_of(rcu, struct pool_workqueue, rcu));
3653 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3654 * and needs to be destroyed.
3656 static void pwq_unbound_release_workfn(struct work_struct *work)
3658 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3659 unbound_release_work);
3660 struct workqueue_struct *wq = pwq->wq;
3661 struct worker_pool *pool = pwq->pool;
3664 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3667 mutex_lock(&wq->mutex);
3668 list_del_rcu(&pwq->pwqs_node);
3669 is_last = list_empty(&wq->pwqs);
3670 mutex_unlock(&wq->mutex);
3672 mutex_lock(&wq_pool_mutex);
3673 put_unbound_pool(pool);
3674 mutex_unlock(&wq_pool_mutex);
3676 call_rcu(&pwq->rcu, rcu_free_pwq);
3679 * If we're the last pwq going away, @wq is already dead and no one
3680 * is gonna access it anymore. Schedule RCU free.
3683 wq_unregister_lockdep(wq);
3684 call_rcu(&wq->rcu, rcu_free_wq);
3689 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3690 * @pwq: target pool_workqueue
3692 * If @pwq isn't freezing, set @pwq->max_active to the associated
3693 * workqueue's saved_max_active and activate delayed work items
3694 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3696 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3698 struct workqueue_struct *wq = pwq->wq;
3699 bool freezable = wq->flags & WQ_FREEZABLE;
3700 unsigned long flags;
3702 /* for @wq->saved_max_active */
3703 lockdep_assert_held(&wq->mutex);
3705 /* fast exit for non-freezable wqs */
3706 if (!freezable && pwq->max_active == wq->saved_max_active)
3709 /* this function can be called during early boot w/ irq disabled */
3710 spin_lock_irqsave(&pwq->pool->lock, flags);
3713 * During [un]freezing, the caller is responsible for ensuring that
3714 * this function is called at least once after @workqueue_freezing
3715 * is updated and visible.
3717 if (!freezable || !workqueue_freezing) {
3718 pwq->max_active = wq->saved_max_active;
3720 while (!list_empty(&pwq->delayed_works) &&
3721 pwq->nr_active < pwq->max_active)
3722 pwq_activate_first_delayed(pwq);
3725 * Need to kick a worker after thawed or an unbound wq's
3726 * max_active is bumped. It's a slow path. Do it always.
3728 wake_up_worker(pwq->pool);
3730 pwq->max_active = 0;
3733 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3736 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3737 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3738 struct worker_pool *pool)
3740 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3742 memset(pwq, 0, sizeof(*pwq));
3746 pwq->flush_color = -1;
3748 INIT_LIST_HEAD(&pwq->delayed_works);
3749 INIT_LIST_HEAD(&pwq->pwqs_node);
3750 INIT_LIST_HEAD(&pwq->mayday_node);
3751 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3754 /* sync @pwq with the current state of its associated wq and link it */
3755 static void link_pwq(struct pool_workqueue *pwq)
3757 struct workqueue_struct *wq = pwq->wq;
3759 lockdep_assert_held(&wq->mutex);
3761 /* may be called multiple times, ignore if already linked */
3762 if (!list_empty(&pwq->pwqs_node))
3765 /* set the matching work_color */
3766 pwq->work_color = wq->work_color;
3768 /* sync max_active to the current setting */
3769 pwq_adjust_max_active(pwq);
3772 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3775 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3776 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3777 const struct workqueue_attrs *attrs)
3779 struct worker_pool *pool;
3780 struct pool_workqueue *pwq;
3782 lockdep_assert_held(&wq_pool_mutex);
3784 pool = get_unbound_pool(attrs);
3788 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3790 put_unbound_pool(pool);
3794 init_pwq(pwq, wq, pool);
3799 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3800 * @attrs: the wq_attrs of the default pwq of the target workqueue
3801 * @node: the target NUMA node
3802 * @cpu_going_down: if >= 0, the CPU to consider as offline
3803 * @cpumask: outarg, the resulting cpumask
3805 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3806 * @cpu_going_down is >= 0, that cpu is considered offline during
3807 * calculation. The result is stored in @cpumask.
3809 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3810 * enabled and @node has online CPUs requested by @attrs, the returned
3811 * cpumask is the intersection of the possible CPUs of @node and
3814 * The caller is responsible for ensuring that the cpumask of @node stays
3817 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3820 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3821 int cpu_going_down, cpumask_t *cpumask)
3823 if (!wq_numa_enabled || attrs->no_numa)
3826 /* does @node have any online CPUs @attrs wants? */
3827 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3828 if (cpu_going_down >= 0)
3829 cpumask_clear_cpu(cpu_going_down, cpumask);
3831 if (cpumask_empty(cpumask))
3834 /* yeap, return possible CPUs in @node that @attrs wants */
3835 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3837 if (cpumask_empty(cpumask)) {
3838 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3839 "possible intersect\n");
3843 return !cpumask_equal(cpumask, attrs->cpumask);
3846 cpumask_copy(cpumask, attrs->cpumask);
3850 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3851 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3853 struct pool_workqueue *pwq)
3855 struct pool_workqueue *old_pwq;
3857 lockdep_assert_held(&wq_pool_mutex);
3858 lockdep_assert_held(&wq->mutex);
3860 /* link_pwq() can handle duplicate calls */
3863 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3864 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3868 /* context to store the prepared attrs & pwqs before applying */
3869 struct apply_wqattrs_ctx {
3870 struct workqueue_struct *wq; /* target workqueue */
3871 struct workqueue_attrs *attrs; /* attrs to apply */
3872 struct list_head list; /* queued for batching commit */
3873 struct pool_workqueue *dfl_pwq;
3874 struct pool_workqueue *pwq_tbl[];
3877 /* free the resources after success or abort */
3878 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3884 put_pwq_unlocked(ctx->pwq_tbl[node]);
3885 put_pwq_unlocked(ctx->dfl_pwq);
3887 free_workqueue_attrs(ctx->attrs);
3893 /* allocate the attrs and pwqs for later installation */
3894 static struct apply_wqattrs_ctx *
3895 apply_wqattrs_prepare(struct workqueue_struct *wq,
3896 const struct workqueue_attrs *attrs)
3898 struct apply_wqattrs_ctx *ctx;
3899 struct workqueue_attrs *new_attrs, *tmp_attrs;
3902 lockdep_assert_held(&wq_pool_mutex);
3904 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3906 new_attrs = alloc_workqueue_attrs();
3907 tmp_attrs = alloc_workqueue_attrs();
3908 if (!ctx || !new_attrs || !tmp_attrs)
3912 * Calculate the attrs of the default pwq.
3913 * If the user configured cpumask doesn't overlap with the
3914 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3916 copy_workqueue_attrs(new_attrs, attrs);
3917 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3918 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3919 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3922 * We may create multiple pwqs with differing cpumasks. Make a
3923 * copy of @new_attrs which will be modified and used to obtain
3926 copy_workqueue_attrs(tmp_attrs, new_attrs);
3929 * If something goes wrong during CPU up/down, we'll fall back to
3930 * the default pwq covering whole @attrs->cpumask. Always create
3931 * it even if we don't use it immediately.
3933 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3937 for_each_node(node) {
3938 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3939 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3940 if (!ctx->pwq_tbl[node])
3943 ctx->dfl_pwq->refcnt++;
3944 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3948 /* save the user configured attrs and sanitize it. */
3949 copy_workqueue_attrs(new_attrs, attrs);
3950 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3951 ctx->attrs = new_attrs;
3954 free_workqueue_attrs(tmp_attrs);
3958 free_workqueue_attrs(tmp_attrs);
3959 free_workqueue_attrs(new_attrs);
3960 apply_wqattrs_cleanup(ctx);
3964 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3965 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3969 /* all pwqs have been created successfully, let's install'em */
3970 mutex_lock(&ctx->wq->mutex);
3972 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3974 /* save the previous pwq and install the new one */
3976 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3977 ctx->pwq_tbl[node]);
3979 /* @dfl_pwq might not have been used, ensure it's linked */
3980 link_pwq(ctx->dfl_pwq);
3981 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3983 mutex_unlock(&ctx->wq->mutex);
3986 static void apply_wqattrs_lock(void)
3988 /* CPUs should stay stable across pwq creations and installations */
3990 mutex_lock(&wq_pool_mutex);
3993 static void apply_wqattrs_unlock(void)
3995 mutex_unlock(&wq_pool_mutex);
3999 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4000 const struct workqueue_attrs *attrs)
4002 struct apply_wqattrs_ctx *ctx;
4004 /* only unbound workqueues can change attributes */
4005 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4008 /* creating multiple pwqs breaks ordering guarantee */
4009 if (!list_empty(&wq->pwqs)) {
4010 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4013 wq->flags &= ~__WQ_ORDERED;
4016 ctx = apply_wqattrs_prepare(wq, attrs);
4020 /* the ctx has been prepared successfully, let's commit it */
4021 apply_wqattrs_commit(ctx);
4022 apply_wqattrs_cleanup(ctx);
4028 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4029 * @wq: the target workqueue
4030 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4032 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4033 * machines, this function maps a separate pwq to each NUMA node with
4034 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4035 * NUMA node it was issued on. Older pwqs are released as in-flight work
4036 * items finish. Note that a work item which repeatedly requeues itself
4037 * back-to-back will stay on its current pwq.
4039 * Performs GFP_KERNEL allocations.
4041 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4043 * Return: 0 on success and -errno on failure.
4045 int apply_workqueue_attrs(struct workqueue_struct *wq,
4046 const struct workqueue_attrs *attrs)
4050 lockdep_assert_cpus_held();
4052 mutex_lock(&wq_pool_mutex);
4053 ret = apply_workqueue_attrs_locked(wq, attrs);
4054 mutex_unlock(&wq_pool_mutex);
4060 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4061 * @wq: the target workqueue
4062 * @cpu: the CPU coming up or going down
4063 * @online: whether @cpu is coming up or going down
4065 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4066 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4069 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4070 * falls back to @wq->dfl_pwq which may not be optimal but is always
4073 * Note that when the last allowed CPU of a NUMA node goes offline for a
4074 * workqueue with a cpumask spanning multiple nodes, the workers which were
4075 * already executing the work items for the workqueue will lose their CPU
4076 * affinity and may execute on any CPU. This is similar to how per-cpu
4077 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4078 * affinity, it's the user's responsibility to flush the work item from
4081 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4084 int node = cpu_to_node(cpu);
4085 int cpu_off = online ? -1 : cpu;
4086 struct pool_workqueue *old_pwq = NULL, *pwq;
4087 struct workqueue_attrs *target_attrs;
4090 lockdep_assert_held(&wq_pool_mutex);
4092 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4093 wq->unbound_attrs->no_numa)
4097 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4098 * Let's use a preallocated one. The following buf is protected by
4099 * CPU hotplug exclusion.
4101 target_attrs = wq_update_unbound_numa_attrs_buf;
4102 cpumask = target_attrs->cpumask;
4104 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4105 pwq = unbound_pwq_by_node(wq, node);
4108 * Let's determine what needs to be done. If the target cpumask is
4109 * different from the default pwq's, we need to compare it to @pwq's
4110 * and create a new one if they don't match. If the target cpumask
4111 * equals the default pwq's, the default pwq should be used.
4113 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4114 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4120 /* create a new pwq */
4121 pwq = alloc_unbound_pwq(wq, target_attrs);
4123 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4128 /* Install the new pwq. */
4129 mutex_lock(&wq->mutex);
4130 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4134 mutex_lock(&wq->mutex);
4135 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4136 get_pwq(wq->dfl_pwq);
4137 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4138 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4140 mutex_unlock(&wq->mutex);
4141 put_pwq_unlocked(old_pwq);
4144 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4146 bool highpri = wq->flags & WQ_HIGHPRI;
4149 if (!(wq->flags & WQ_UNBOUND)) {
4150 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4154 for_each_possible_cpu(cpu) {
4155 struct pool_workqueue *pwq =
4156 per_cpu_ptr(wq->cpu_pwqs, cpu);
4157 struct worker_pool *cpu_pools =
4158 per_cpu(cpu_worker_pools, cpu);
4160 init_pwq(pwq, wq, &cpu_pools[highpri]);
4162 mutex_lock(&wq->mutex);
4164 mutex_unlock(&wq->mutex);
4170 if (wq->flags & __WQ_ORDERED) {
4171 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4172 /* there should only be single pwq for ordering guarantee */
4173 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4174 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4175 "ordering guarantee broken for workqueue %s\n", wq->name);
4177 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4184 static int wq_clamp_max_active(int max_active, unsigned int flags,
4187 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4189 if (max_active < 1 || max_active > lim)
4190 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4191 max_active, name, 1, lim);
4193 return clamp_val(max_active, 1, lim);
4197 * Workqueues which may be used during memory reclaim should have a rescuer
4198 * to guarantee forward progress.
4200 static int init_rescuer(struct workqueue_struct *wq)
4202 struct worker *rescuer;
4205 if (!(wq->flags & WQ_MEM_RECLAIM))
4208 rescuer = alloc_worker(NUMA_NO_NODE);
4212 rescuer->rescue_wq = wq;
4213 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4214 ret = PTR_ERR_OR_ZERO(rescuer->task);
4220 wq->rescuer = rescuer;
4221 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4222 wake_up_process(rescuer->task);
4228 struct workqueue_struct *alloc_workqueue(const char *fmt,
4230 int max_active, ...)
4232 size_t tbl_size = 0;
4234 struct workqueue_struct *wq;
4235 struct pool_workqueue *pwq;
4238 * Unbound && max_active == 1 used to imply ordered, which is no
4239 * longer the case on NUMA machines due to per-node pools. While
4240 * alloc_ordered_workqueue() is the right way to create an ordered
4241 * workqueue, keep the previous behavior to avoid subtle breakages
4244 if ((flags & WQ_UNBOUND) && max_active == 1)
4245 flags |= __WQ_ORDERED;
4247 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4248 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4249 flags |= WQ_UNBOUND;
4251 /* allocate wq and format name */
4252 if (flags & WQ_UNBOUND)
4253 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4255 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4259 if (flags & WQ_UNBOUND) {
4260 wq->unbound_attrs = alloc_workqueue_attrs();
4261 if (!wq->unbound_attrs)
4265 va_start(args, max_active);
4266 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4269 max_active = max_active ?: WQ_DFL_ACTIVE;
4270 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4274 wq->saved_max_active = max_active;
4275 mutex_init(&wq->mutex);
4276 atomic_set(&wq->nr_pwqs_to_flush, 0);
4277 INIT_LIST_HEAD(&wq->pwqs);
4278 INIT_LIST_HEAD(&wq->flusher_queue);
4279 INIT_LIST_HEAD(&wq->flusher_overflow);
4280 INIT_LIST_HEAD(&wq->maydays);
4282 wq_init_lockdep(wq);
4283 INIT_LIST_HEAD(&wq->list);
4285 if (alloc_and_link_pwqs(wq) < 0)
4286 goto err_unreg_lockdep;
4288 if (wq_online && init_rescuer(wq) < 0)
4291 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4295 * wq_pool_mutex protects global freeze state and workqueues list.
4296 * Grab it, adjust max_active and add the new @wq to workqueues
4299 mutex_lock(&wq_pool_mutex);
4301 mutex_lock(&wq->mutex);
4302 for_each_pwq(pwq, wq)
4303 pwq_adjust_max_active(pwq);
4304 mutex_unlock(&wq->mutex);
4306 list_add_tail_rcu(&wq->list, &workqueues);
4308 mutex_unlock(&wq_pool_mutex);
4313 wq_unregister_lockdep(wq);
4314 wq_free_lockdep(wq);
4316 free_workqueue_attrs(wq->unbound_attrs);
4320 destroy_workqueue(wq);
4323 EXPORT_SYMBOL_GPL(alloc_workqueue);
4325 static bool pwq_busy(struct pool_workqueue *pwq)
4329 for (i = 0; i < WORK_NR_COLORS; i++)
4330 if (pwq->nr_in_flight[i])
4333 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4335 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4342 * destroy_workqueue - safely terminate a workqueue
4343 * @wq: target workqueue
4345 * Safely destroy a workqueue. All work currently pending will be done first.
4347 void destroy_workqueue(struct workqueue_struct *wq)
4349 struct pool_workqueue *pwq;
4353 * Remove it from sysfs first so that sanity check failure doesn't
4354 * lead to sysfs name conflicts.
4356 workqueue_sysfs_unregister(wq);
4358 /* drain it before proceeding with destruction */
4359 drain_workqueue(wq);
4361 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4363 struct worker *rescuer = wq->rescuer;
4365 /* this prevents new queueing */
4366 spin_lock_irq(&wq_mayday_lock);
4368 spin_unlock_irq(&wq_mayday_lock);
4370 /* rescuer will empty maydays list before exiting */
4371 kthread_stop(rescuer->task);
4376 * Sanity checks - grab all the locks so that we wait for all
4377 * in-flight operations which may do put_pwq().
4379 mutex_lock(&wq_pool_mutex);
4380 mutex_lock(&wq->mutex);
4381 for_each_pwq(pwq, wq) {
4382 spin_lock_irq(&pwq->pool->lock);
4383 if (WARN_ON(pwq_busy(pwq))) {
4384 pr_warning("%s: %s has the following busy pwq\n",
4385 __func__, wq->name);
4387 spin_unlock_irq(&pwq->pool->lock);
4388 mutex_unlock(&wq->mutex);
4389 mutex_unlock(&wq_pool_mutex);
4390 show_workqueue_state();
4393 spin_unlock_irq(&pwq->pool->lock);
4395 mutex_unlock(&wq->mutex);
4396 mutex_unlock(&wq_pool_mutex);
4399 * wq list is used to freeze wq, remove from list after
4400 * flushing is complete in case freeze races us.
4402 mutex_lock(&wq_pool_mutex);
4403 list_del_rcu(&wq->list);
4404 mutex_unlock(&wq_pool_mutex);
4406 if (!(wq->flags & WQ_UNBOUND)) {
4407 wq_unregister_lockdep(wq);
4409 * The base ref is never dropped on per-cpu pwqs. Directly
4410 * schedule RCU free.
4412 call_rcu(&wq->rcu, rcu_free_wq);
4415 * We're the sole accessor of @wq at this point. Directly
4416 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4417 * @wq will be freed when the last pwq is released.
4419 for_each_node(node) {
4420 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4421 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4422 put_pwq_unlocked(pwq);
4426 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4427 * put. Don't access it afterwards.
4431 put_pwq_unlocked(pwq);
4434 EXPORT_SYMBOL_GPL(destroy_workqueue);
4437 * workqueue_set_max_active - adjust max_active of a workqueue
4438 * @wq: target workqueue
4439 * @max_active: new max_active value.
4441 * Set max_active of @wq to @max_active.
4444 * Don't call from IRQ context.
4446 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4448 struct pool_workqueue *pwq;
4450 /* disallow meddling with max_active for ordered workqueues */
4451 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4454 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4456 mutex_lock(&wq->mutex);
4458 wq->flags &= ~__WQ_ORDERED;
4459 wq->saved_max_active = max_active;
4461 for_each_pwq(pwq, wq)
4462 pwq_adjust_max_active(pwq);
4464 mutex_unlock(&wq->mutex);
4466 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4469 * current_work - retrieve %current task's work struct
4471 * Determine if %current task is a workqueue worker and what it's working on.
4472 * Useful to find out the context that the %current task is running in.
4474 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4476 struct work_struct *current_work(void)
4478 struct worker *worker = current_wq_worker();
4480 return worker ? worker->current_work : NULL;
4482 EXPORT_SYMBOL(current_work);
4485 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4487 * Determine whether %current is a workqueue rescuer. Can be used from
4488 * work functions to determine whether it's being run off the rescuer task.
4490 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4492 bool current_is_workqueue_rescuer(void)
4494 struct worker *worker = current_wq_worker();
4496 return worker && worker->rescue_wq;
4500 * workqueue_congested - test whether a workqueue is congested
4501 * @cpu: CPU in question
4502 * @wq: target workqueue
4504 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4505 * no synchronization around this function and the test result is
4506 * unreliable and only useful as advisory hints or for debugging.
4508 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4509 * Note that both per-cpu and unbound workqueues may be associated with
4510 * multiple pool_workqueues which have separate congested states. A
4511 * workqueue being congested on one CPU doesn't mean the workqueue is also
4512 * contested on other CPUs / NUMA nodes.
4515 * %true if congested, %false otherwise.
4517 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4519 struct pool_workqueue *pwq;
4525 if (cpu == WORK_CPU_UNBOUND)
4526 cpu = smp_processor_id();
4528 if (!(wq->flags & WQ_UNBOUND))
4529 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4531 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4533 ret = !list_empty(&pwq->delayed_works);
4539 EXPORT_SYMBOL_GPL(workqueue_congested);
4542 * work_busy - test whether a work is currently pending or running
4543 * @work: the work to be tested
4545 * Test whether @work is currently pending or running. There is no
4546 * synchronization around this function and the test result is
4547 * unreliable and only useful as advisory hints or for debugging.
4550 * OR'd bitmask of WORK_BUSY_* bits.
4552 unsigned int work_busy(struct work_struct *work)
4554 struct worker_pool *pool;
4555 unsigned long flags;
4556 unsigned int ret = 0;
4558 if (work_pending(work))
4559 ret |= WORK_BUSY_PENDING;
4562 pool = get_work_pool(work);
4564 spin_lock_irqsave(&pool->lock, flags);
4565 if (find_worker_executing_work(pool, work))
4566 ret |= WORK_BUSY_RUNNING;
4567 spin_unlock_irqrestore(&pool->lock, flags);
4573 EXPORT_SYMBOL_GPL(work_busy);
4576 * set_worker_desc - set description for the current work item
4577 * @fmt: printf-style format string
4578 * @...: arguments for the format string
4580 * This function can be called by a running work function to describe what
4581 * the work item is about. If the worker task gets dumped, this
4582 * information will be printed out together to help debugging. The
4583 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4585 void set_worker_desc(const char *fmt, ...)
4587 struct worker *worker = current_wq_worker();
4591 va_start(args, fmt);
4592 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4596 EXPORT_SYMBOL_GPL(set_worker_desc);
4599 * print_worker_info - print out worker information and description
4600 * @log_lvl: the log level to use when printing
4601 * @task: target task
4603 * If @task is a worker and currently executing a work item, print out the
4604 * name of the workqueue being serviced and worker description set with
4605 * set_worker_desc() by the currently executing work item.
4607 * This function can be safely called on any task as long as the
4608 * task_struct itself is accessible. While safe, this function isn't
4609 * synchronized and may print out mixups or garbages of limited length.
4611 void print_worker_info(const char *log_lvl, struct task_struct *task)
4613 work_func_t *fn = NULL;
4614 char name[WQ_NAME_LEN] = { };
4615 char desc[WORKER_DESC_LEN] = { };
4616 struct pool_workqueue *pwq = NULL;
4617 struct workqueue_struct *wq = NULL;
4618 struct worker *worker;
4620 if (!(task->flags & PF_WQ_WORKER))
4624 * This function is called without any synchronization and @task
4625 * could be in any state. Be careful with dereferences.
4627 worker = kthread_probe_data(task);
4630 * Carefully copy the associated workqueue's workfn, name and desc.
4631 * Keep the original last '\0' in case the original is garbage.
4633 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4634 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4635 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4636 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4637 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4639 if (fn || name[0] || desc[0]) {
4640 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4641 if (strcmp(name, desc))
4642 pr_cont(" (%s)", desc);
4647 static void pr_cont_pool_info(struct worker_pool *pool)
4649 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4650 if (pool->node != NUMA_NO_NODE)
4651 pr_cont(" node=%d", pool->node);
4652 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4655 static void pr_cont_work(bool comma, struct work_struct *work)
4657 if (work->func == wq_barrier_func) {
4658 struct wq_barrier *barr;
4660 barr = container_of(work, struct wq_barrier, work);
4662 pr_cont("%s BAR(%d)", comma ? "," : "",
4663 task_pid_nr(barr->task));
4665 pr_cont("%s %ps", comma ? "," : "", work->func);
4669 static void show_pwq(struct pool_workqueue *pwq)
4671 struct worker_pool *pool = pwq->pool;
4672 struct work_struct *work;
4673 struct worker *worker;
4674 bool has_in_flight = false, has_pending = false;
4677 pr_info(" pwq %d:", pool->id);
4678 pr_cont_pool_info(pool);
4680 pr_cont(" active=%d/%d refcnt=%d%s\n",
4681 pwq->nr_active, pwq->max_active, pwq->refcnt,
4682 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4684 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4685 if (worker->current_pwq == pwq) {
4686 has_in_flight = true;
4690 if (has_in_flight) {
4693 pr_info(" in-flight:");
4694 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4695 if (worker->current_pwq != pwq)
4698 pr_cont("%s %d%s:%ps", comma ? "," : "",
4699 task_pid_nr(worker->task),
4700 worker->rescue_wq ? "(RESCUER)" : "",
4701 worker->current_func);
4702 list_for_each_entry(work, &worker->scheduled, entry)
4703 pr_cont_work(false, work);
4709 list_for_each_entry(work, &pool->worklist, entry) {
4710 if (get_work_pwq(work) == pwq) {
4718 pr_info(" pending:");
4719 list_for_each_entry(work, &pool->worklist, entry) {
4720 if (get_work_pwq(work) != pwq)
4723 pr_cont_work(comma, work);
4724 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4729 if (!list_empty(&pwq->delayed_works)) {
4732 pr_info(" delayed:");
4733 list_for_each_entry(work, &pwq->delayed_works, entry) {
4734 pr_cont_work(comma, work);
4735 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4742 * show_workqueue_state - dump workqueue state
4744 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4745 * all busy workqueues and pools.
4747 void show_workqueue_state(void)
4749 struct workqueue_struct *wq;
4750 struct worker_pool *pool;
4751 unsigned long flags;
4756 pr_info("Showing busy workqueues and worker pools:\n");
4758 list_for_each_entry_rcu(wq, &workqueues, list) {
4759 struct pool_workqueue *pwq;
4762 for_each_pwq(pwq, wq) {
4763 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4771 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4773 for_each_pwq(pwq, wq) {
4774 spin_lock_irqsave(&pwq->pool->lock, flags);
4775 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4777 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4779 * We could be printing a lot from atomic context, e.g.
4780 * sysrq-t -> show_workqueue_state(). Avoid triggering
4783 touch_nmi_watchdog();
4787 for_each_pool(pool, pi) {
4788 struct worker *worker;
4791 spin_lock_irqsave(&pool->lock, flags);
4792 if (pool->nr_workers == pool->nr_idle)
4795 pr_info("pool %d:", pool->id);
4796 pr_cont_pool_info(pool);
4797 pr_cont(" hung=%us workers=%d",
4798 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4801 pr_cont(" manager: %d",
4802 task_pid_nr(pool->manager->task));
4803 list_for_each_entry(worker, &pool->idle_list, entry) {
4804 pr_cont(" %s%d", first ? "idle: " : "",
4805 task_pid_nr(worker->task));
4810 spin_unlock_irqrestore(&pool->lock, flags);
4812 * We could be printing a lot from atomic context, e.g.
4813 * sysrq-t -> show_workqueue_state(). Avoid triggering
4816 touch_nmi_watchdog();
4822 /* used to show worker information through /proc/PID/{comm,stat,status} */
4823 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4827 /* always show the actual comm */
4828 off = strscpy(buf, task->comm, size);
4832 /* stabilize PF_WQ_WORKER and worker pool association */
4833 mutex_lock(&wq_pool_attach_mutex);
4835 if (task->flags & PF_WQ_WORKER) {
4836 struct worker *worker = kthread_data(task);
4837 struct worker_pool *pool = worker->pool;
4840 spin_lock_irq(&pool->lock);
4842 * ->desc tracks information (wq name or
4843 * set_worker_desc()) for the latest execution. If
4844 * current, prepend '+', otherwise '-'.
4846 if (worker->desc[0] != '\0') {
4847 if (worker->current_work)
4848 scnprintf(buf + off, size - off, "+%s",
4851 scnprintf(buf + off, size - off, "-%s",
4854 spin_unlock_irq(&pool->lock);
4858 mutex_unlock(&wq_pool_attach_mutex);
4866 * There are two challenges in supporting CPU hotplug. Firstly, there
4867 * are a lot of assumptions on strong associations among work, pwq and
4868 * pool which make migrating pending and scheduled works very
4869 * difficult to implement without impacting hot paths. Secondly,
4870 * worker pools serve mix of short, long and very long running works making
4871 * blocked draining impractical.
4873 * This is solved by allowing the pools to be disassociated from the CPU
4874 * running as an unbound one and allowing it to be reattached later if the
4875 * cpu comes back online.
4878 static void unbind_workers(int cpu)
4880 struct worker_pool *pool;
4881 struct worker *worker;
4883 for_each_cpu_worker_pool(pool, cpu) {
4884 mutex_lock(&wq_pool_attach_mutex);
4885 spin_lock_irq(&pool->lock);
4888 * We've blocked all attach/detach operations. Make all workers
4889 * unbound and set DISASSOCIATED. Before this, all workers
4890 * except for the ones which are still executing works from
4891 * before the last CPU down must be on the cpu. After
4892 * this, they may become diasporas.
4894 for_each_pool_worker(worker, pool)
4895 worker->flags |= WORKER_UNBOUND;
4897 pool->flags |= POOL_DISASSOCIATED;
4899 spin_unlock_irq(&pool->lock);
4900 mutex_unlock(&wq_pool_attach_mutex);
4903 * Call schedule() so that we cross rq->lock and thus can
4904 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4905 * This is necessary as scheduler callbacks may be invoked
4911 * Sched callbacks are disabled now. Zap nr_running.
4912 * After this, nr_running stays zero and need_more_worker()
4913 * and keep_working() are always true as long as the
4914 * worklist is not empty. This pool now behaves as an
4915 * unbound (in terms of concurrency management) pool which
4916 * are served by workers tied to the pool.
4918 atomic_set(&pool->nr_running, 0);
4921 * With concurrency management just turned off, a busy
4922 * worker blocking could lead to lengthy stalls. Kick off
4923 * unbound chain execution of currently pending work items.
4925 spin_lock_irq(&pool->lock);
4926 wake_up_worker(pool);
4927 spin_unlock_irq(&pool->lock);
4932 * rebind_workers - rebind all workers of a pool to the associated CPU
4933 * @pool: pool of interest
4935 * @pool->cpu is coming online. Rebind all workers to the CPU.
4937 static void rebind_workers(struct worker_pool *pool)
4939 struct worker *worker;
4941 lockdep_assert_held(&wq_pool_attach_mutex);
4944 * Restore CPU affinity of all workers. As all idle workers should
4945 * be on the run-queue of the associated CPU before any local
4946 * wake-ups for concurrency management happen, restore CPU affinity
4947 * of all workers first and then clear UNBOUND. As we're called
4948 * from CPU_ONLINE, the following shouldn't fail.
4950 for_each_pool_worker(worker, pool)
4951 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4952 pool->attrs->cpumask) < 0);
4954 spin_lock_irq(&pool->lock);
4956 pool->flags &= ~POOL_DISASSOCIATED;
4958 for_each_pool_worker(worker, pool) {
4959 unsigned int worker_flags = worker->flags;
4962 * A bound idle worker should actually be on the runqueue
4963 * of the associated CPU for local wake-ups targeting it to
4964 * work. Kick all idle workers so that they migrate to the
4965 * associated CPU. Doing this in the same loop as
4966 * replacing UNBOUND with REBOUND is safe as no worker will
4967 * be bound before @pool->lock is released.
4969 if (worker_flags & WORKER_IDLE)
4970 wake_up_process(worker->task);
4973 * We want to clear UNBOUND but can't directly call
4974 * worker_clr_flags() or adjust nr_running. Atomically
4975 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4976 * @worker will clear REBOUND using worker_clr_flags() when
4977 * it initiates the next execution cycle thus restoring
4978 * concurrency management. Note that when or whether
4979 * @worker clears REBOUND doesn't affect correctness.
4981 * WRITE_ONCE() is necessary because @worker->flags may be
4982 * tested without holding any lock in
4983 * wq_worker_running(). Without it, NOT_RUNNING test may
4984 * fail incorrectly leading to premature concurrency
4985 * management operations.
4987 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4988 worker_flags |= WORKER_REBOUND;
4989 worker_flags &= ~WORKER_UNBOUND;
4990 WRITE_ONCE(worker->flags, worker_flags);
4993 spin_unlock_irq(&pool->lock);
4997 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4998 * @pool: unbound pool of interest
4999 * @cpu: the CPU which is coming up
5001 * An unbound pool may end up with a cpumask which doesn't have any online
5002 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5003 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5004 * online CPU before, cpus_allowed of all its workers should be restored.
5006 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5008 static cpumask_t cpumask;
5009 struct worker *worker;
5011 lockdep_assert_held(&wq_pool_attach_mutex);
5013 /* is @cpu allowed for @pool? */
5014 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5017 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5019 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5020 for_each_pool_worker(worker, pool)
5021 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5024 int workqueue_prepare_cpu(unsigned int cpu)
5026 struct worker_pool *pool;
5028 for_each_cpu_worker_pool(pool, cpu) {
5029 if (pool->nr_workers)
5031 if (!create_worker(pool))
5037 int workqueue_online_cpu(unsigned int cpu)
5039 struct worker_pool *pool;
5040 struct workqueue_struct *wq;
5043 mutex_lock(&wq_pool_mutex);
5045 for_each_pool(pool, pi) {
5046 mutex_lock(&wq_pool_attach_mutex);
5048 if (pool->cpu == cpu)
5049 rebind_workers(pool);
5050 else if (pool->cpu < 0)
5051 restore_unbound_workers_cpumask(pool, cpu);
5053 mutex_unlock(&wq_pool_attach_mutex);
5056 /* update NUMA affinity of unbound workqueues */
5057 list_for_each_entry(wq, &workqueues, list)
5058 wq_update_unbound_numa(wq, cpu, true);
5060 mutex_unlock(&wq_pool_mutex);
5064 int workqueue_offline_cpu(unsigned int cpu)
5066 struct workqueue_struct *wq;
5068 /* unbinding per-cpu workers should happen on the local CPU */
5069 if (WARN_ON(cpu != smp_processor_id()))
5072 unbind_workers(cpu);
5074 /* update NUMA affinity of unbound workqueues */
5075 mutex_lock(&wq_pool_mutex);
5076 list_for_each_entry(wq, &workqueues, list)
5077 wq_update_unbound_numa(wq, cpu, false);
5078 mutex_unlock(&wq_pool_mutex);
5083 struct work_for_cpu {
5084 struct work_struct work;
5090 static void work_for_cpu_fn(struct work_struct *work)
5092 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5094 wfc->ret = wfc->fn(wfc->arg);
5098 * work_on_cpu - run a function in thread context on a particular cpu
5099 * @cpu: the cpu to run on
5100 * @fn: the function to run
5101 * @arg: the function arg
5103 * It is up to the caller to ensure that the cpu doesn't go offline.
5104 * The caller must not hold any locks which would prevent @fn from completing.
5106 * Return: The value @fn returns.
5108 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5110 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5112 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5113 schedule_work_on(cpu, &wfc.work);
5114 flush_work(&wfc.work);
5115 destroy_work_on_stack(&wfc.work);
5118 EXPORT_SYMBOL_GPL(work_on_cpu);
5121 * work_on_cpu_safe - run a function in thread context on a particular cpu
5122 * @cpu: the cpu to run on
5123 * @fn: the function to run
5124 * @arg: the function argument
5126 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5127 * any locks which would prevent @fn from completing.
5129 * Return: The value @fn returns.
5131 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5136 if (cpu_online(cpu))
5137 ret = work_on_cpu(cpu, fn, arg);
5141 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5142 #endif /* CONFIG_SMP */
5144 #ifdef CONFIG_FREEZER
5147 * freeze_workqueues_begin - begin freezing workqueues
5149 * Start freezing workqueues. After this function returns, all freezable
5150 * workqueues will queue new works to their delayed_works list instead of
5154 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5156 void freeze_workqueues_begin(void)
5158 struct workqueue_struct *wq;
5159 struct pool_workqueue *pwq;
5161 mutex_lock(&wq_pool_mutex);
5163 WARN_ON_ONCE(workqueue_freezing);
5164 workqueue_freezing = true;
5166 list_for_each_entry(wq, &workqueues, list) {
5167 mutex_lock(&wq->mutex);
5168 for_each_pwq(pwq, wq)
5169 pwq_adjust_max_active(pwq);
5170 mutex_unlock(&wq->mutex);
5173 mutex_unlock(&wq_pool_mutex);
5177 * freeze_workqueues_busy - are freezable workqueues still busy?
5179 * Check whether freezing is complete. This function must be called
5180 * between freeze_workqueues_begin() and thaw_workqueues().
5183 * Grabs and releases wq_pool_mutex.
5186 * %true if some freezable workqueues are still busy. %false if freezing
5189 bool freeze_workqueues_busy(void)
5192 struct workqueue_struct *wq;
5193 struct pool_workqueue *pwq;
5195 mutex_lock(&wq_pool_mutex);
5197 WARN_ON_ONCE(!workqueue_freezing);
5199 list_for_each_entry(wq, &workqueues, list) {
5200 if (!(wq->flags & WQ_FREEZABLE))
5203 * nr_active is monotonically decreasing. It's safe
5204 * to peek without lock.
5207 for_each_pwq(pwq, wq) {
5208 WARN_ON_ONCE(pwq->nr_active < 0);
5209 if (pwq->nr_active) {
5218 mutex_unlock(&wq_pool_mutex);
5223 * thaw_workqueues - thaw workqueues
5225 * Thaw workqueues. Normal queueing is restored and all collected
5226 * frozen works are transferred to their respective pool worklists.
5229 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5231 void thaw_workqueues(void)
5233 struct workqueue_struct *wq;
5234 struct pool_workqueue *pwq;
5236 mutex_lock(&wq_pool_mutex);
5238 if (!workqueue_freezing)
5241 workqueue_freezing = false;
5243 /* restore max_active and repopulate worklist */
5244 list_for_each_entry(wq, &workqueues, list) {
5245 mutex_lock(&wq->mutex);
5246 for_each_pwq(pwq, wq)
5247 pwq_adjust_max_active(pwq);
5248 mutex_unlock(&wq->mutex);
5252 mutex_unlock(&wq_pool_mutex);
5254 #endif /* CONFIG_FREEZER */
5256 static int workqueue_apply_unbound_cpumask(void)
5260 struct workqueue_struct *wq;
5261 struct apply_wqattrs_ctx *ctx, *n;
5263 lockdep_assert_held(&wq_pool_mutex);
5265 list_for_each_entry(wq, &workqueues, list) {
5266 if (!(wq->flags & WQ_UNBOUND))
5268 /* creating multiple pwqs breaks ordering guarantee */
5269 if (wq->flags & __WQ_ORDERED)
5272 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5278 list_add_tail(&ctx->list, &ctxs);
5281 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5283 apply_wqattrs_commit(ctx);
5284 apply_wqattrs_cleanup(ctx);
5291 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5292 * @cpumask: the cpumask to set
5294 * The low-level workqueues cpumask is a global cpumask that limits
5295 * the affinity of all unbound workqueues. This function check the @cpumask
5296 * and apply it to all unbound workqueues and updates all pwqs of them.
5298 * Retun: 0 - Success
5299 * -EINVAL - Invalid @cpumask
5300 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5302 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5305 cpumask_var_t saved_cpumask;
5307 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5311 * Not excluding isolated cpus on purpose.
5312 * If the user wishes to include them, we allow that.
5314 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5315 if (!cpumask_empty(cpumask)) {
5316 apply_wqattrs_lock();
5318 /* save the old wq_unbound_cpumask. */
5319 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5321 /* update wq_unbound_cpumask at first and apply it to wqs. */
5322 cpumask_copy(wq_unbound_cpumask, cpumask);
5323 ret = workqueue_apply_unbound_cpumask();
5325 /* restore the wq_unbound_cpumask when failed. */
5327 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5329 apply_wqattrs_unlock();
5332 free_cpumask_var(saved_cpumask);
5338 * Workqueues with WQ_SYSFS flag set is visible to userland via
5339 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5340 * following attributes.
5342 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5343 * max_active RW int : maximum number of in-flight work items
5345 * Unbound workqueues have the following extra attributes.
5347 * pool_ids RO int : the associated pool IDs for each node
5348 * nice RW int : nice value of the workers
5349 * cpumask RW mask : bitmask of allowed CPUs for the workers
5350 * numa RW bool : whether enable NUMA affinity
5353 struct workqueue_struct *wq;
5357 static struct workqueue_struct *dev_to_wq(struct device *dev)
5359 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5364 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5367 struct workqueue_struct *wq = dev_to_wq(dev);
5369 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5371 static DEVICE_ATTR_RO(per_cpu);
5373 static ssize_t max_active_show(struct device *dev,
5374 struct device_attribute *attr, char *buf)
5376 struct workqueue_struct *wq = dev_to_wq(dev);
5378 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5381 static ssize_t max_active_store(struct device *dev,
5382 struct device_attribute *attr, const char *buf,
5385 struct workqueue_struct *wq = dev_to_wq(dev);
5388 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5391 workqueue_set_max_active(wq, val);
5394 static DEVICE_ATTR_RW(max_active);
5396 static struct attribute *wq_sysfs_attrs[] = {
5397 &dev_attr_per_cpu.attr,
5398 &dev_attr_max_active.attr,
5401 ATTRIBUTE_GROUPS(wq_sysfs);
5403 static ssize_t wq_pool_ids_show(struct device *dev,
5404 struct device_attribute *attr, char *buf)
5406 struct workqueue_struct *wq = dev_to_wq(dev);
5407 const char *delim = "";
5408 int node, written = 0;
5412 for_each_node(node) {
5413 written += scnprintf(buf + written, PAGE_SIZE - written,
5414 "%s%d:%d", delim, node,
5415 unbound_pwq_by_node(wq, node)->pool->id);
5418 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5425 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5428 struct workqueue_struct *wq = dev_to_wq(dev);
5431 mutex_lock(&wq->mutex);
5432 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5433 mutex_unlock(&wq->mutex);
5438 /* prepare workqueue_attrs for sysfs store operations */
5439 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5441 struct workqueue_attrs *attrs;
5443 lockdep_assert_held(&wq_pool_mutex);
5445 attrs = alloc_workqueue_attrs();
5449 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5453 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5454 const char *buf, size_t count)
5456 struct workqueue_struct *wq = dev_to_wq(dev);
5457 struct workqueue_attrs *attrs;
5460 apply_wqattrs_lock();
5462 attrs = wq_sysfs_prep_attrs(wq);
5466 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5467 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5468 ret = apply_workqueue_attrs_locked(wq, attrs);
5473 apply_wqattrs_unlock();
5474 free_workqueue_attrs(attrs);
5475 return ret ?: count;
5478 static ssize_t wq_cpumask_show(struct device *dev,
5479 struct device_attribute *attr, char *buf)
5481 struct workqueue_struct *wq = dev_to_wq(dev);
5484 mutex_lock(&wq->mutex);
5485 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5486 cpumask_pr_args(wq->unbound_attrs->cpumask));
5487 mutex_unlock(&wq->mutex);
5491 static ssize_t wq_cpumask_store(struct device *dev,
5492 struct device_attribute *attr,
5493 const char *buf, size_t count)
5495 struct workqueue_struct *wq = dev_to_wq(dev);
5496 struct workqueue_attrs *attrs;
5499 apply_wqattrs_lock();
5501 attrs = wq_sysfs_prep_attrs(wq);
5505 ret = cpumask_parse(buf, attrs->cpumask);
5507 ret = apply_workqueue_attrs_locked(wq, attrs);
5510 apply_wqattrs_unlock();
5511 free_workqueue_attrs(attrs);
5512 return ret ?: count;
5515 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5518 struct workqueue_struct *wq = dev_to_wq(dev);
5521 mutex_lock(&wq->mutex);
5522 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5523 !wq->unbound_attrs->no_numa);
5524 mutex_unlock(&wq->mutex);
5529 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5530 const char *buf, size_t count)
5532 struct workqueue_struct *wq = dev_to_wq(dev);
5533 struct workqueue_attrs *attrs;
5534 int v, ret = -ENOMEM;
5536 apply_wqattrs_lock();
5538 attrs = wq_sysfs_prep_attrs(wq);
5543 if (sscanf(buf, "%d", &v) == 1) {
5544 attrs->no_numa = !v;
5545 ret = apply_workqueue_attrs_locked(wq, attrs);
5549 apply_wqattrs_unlock();
5550 free_workqueue_attrs(attrs);
5551 return ret ?: count;
5554 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5555 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5556 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5557 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5558 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5562 static struct bus_type wq_subsys = {
5563 .name = "workqueue",
5564 .dev_groups = wq_sysfs_groups,
5567 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5568 struct device_attribute *attr, char *buf)
5572 mutex_lock(&wq_pool_mutex);
5573 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5574 cpumask_pr_args(wq_unbound_cpumask));
5575 mutex_unlock(&wq_pool_mutex);
5580 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5581 struct device_attribute *attr, const char *buf, size_t count)
5583 cpumask_var_t cpumask;
5586 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5589 ret = cpumask_parse(buf, cpumask);
5591 ret = workqueue_set_unbound_cpumask(cpumask);
5593 free_cpumask_var(cpumask);
5594 return ret ? ret : count;
5597 static struct device_attribute wq_sysfs_cpumask_attr =
5598 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5599 wq_unbound_cpumask_store);
5601 static int __init wq_sysfs_init(void)
5605 err = subsys_virtual_register(&wq_subsys, NULL);
5609 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5611 core_initcall(wq_sysfs_init);
5613 static void wq_device_release(struct device *dev)
5615 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5621 * workqueue_sysfs_register - make a workqueue visible in sysfs
5622 * @wq: the workqueue to register
5624 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5625 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5626 * which is the preferred method.
5628 * Workqueue user should use this function directly iff it wants to apply
5629 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5630 * apply_workqueue_attrs() may race against userland updating the
5633 * Return: 0 on success, -errno on failure.
5635 int workqueue_sysfs_register(struct workqueue_struct *wq)
5637 struct wq_device *wq_dev;
5641 * Adjusting max_active or creating new pwqs by applying
5642 * attributes breaks ordering guarantee. Disallow exposing ordered
5645 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5648 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5653 wq_dev->dev.bus = &wq_subsys;
5654 wq_dev->dev.release = wq_device_release;
5655 dev_set_name(&wq_dev->dev, "%s", wq->name);
5658 * unbound_attrs are created separately. Suppress uevent until
5659 * everything is ready.
5661 dev_set_uevent_suppress(&wq_dev->dev, true);
5663 ret = device_register(&wq_dev->dev);
5665 put_device(&wq_dev->dev);
5670 if (wq->flags & WQ_UNBOUND) {
5671 struct device_attribute *attr;
5673 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5674 ret = device_create_file(&wq_dev->dev, attr);
5676 device_unregister(&wq_dev->dev);
5683 dev_set_uevent_suppress(&wq_dev->dev, false);
5684 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5689 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5690 * @wq: the workqueue to unregister
5692 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5694 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5696 struct wq_device *wq_dev = wq->wq_dev;
5702 device_unregister(&wq_dev->dev);
5704 #else /* CONFIG_SYSFS */
5705 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5706 #endif /* CONFIG_SYSFS */
5709 * Workqueue watchdog.
5711 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5712 * flush dependency, a concurrency managed work item which stays RUNNING
5713 * indefinitely. Workqueue stalls can be very difficult to debug as the
5714 * usual warning mechanisms don't trigger and internal workqueue state is
5717 * Workqueue watchdog monitors all worker pools periodically and dumps
5718 * state if some pools failed to make forward progress for a while where
5719 * forward progress is defined as the first item on ->worklist changing.
5721 * This mechanism is controlled through the kernel parameter
5722 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5723 * corresponding sysfs parameter file.
5725 #ifdef CONFIG_WQ_WATCHDOG
5727 static unsigned long wq_watchdog_thresh = 30;
5728 static struct timer_list wq_watchdog_timer;
5730 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5731 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5733 static void wq_watchdog_reset_touched(void)
5737 wq_watchdog_touched = jiffies;
5738 for_each_possible_cpu(cpu)
5739 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5742 static void wq_watchdog_timer_fn(struct timer_list *unused)
5744 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5745 bool lockup_detected = false;
5746 struct worker_pool *pool;
5754 for_each_pool(pool, pi) {
5755 unsigned long pool_ts, touched, ts;
5757 if (list_empty(&pool->worklist))
5760 /* get the latest of pool and touched timestamps */
5761 pool_ts = READ_ONCE(pool->watchdog_ts);
5762 touched = READ_ONCE(wq_watchdog_touched);
5764 if (time_after(pool_ts, touched))
5769 if (pool->cpu >= 0) {
5770 unsigned long cpu_touched =
5771 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5773 if (time_after(cpu_touched, ts))
5778 if (time_after(jiffies, ts + thresh)) {
5779 lockup_detected = true;
5780 pr_emerg("BUG: workqueue lockup - pool");
5781 pr_cont_pool_info(pool);
5782 pr_cont(" stuck for %us!\n",
5783 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5789 if (lockup_detected)
5790 show_workqueue_state();
5792 wq_watchdog_reset_touched();
5793 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5796 notrace void wq_watchdog_touch(int cpu)
5799 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5801 wq_watchdog_touched = jiffies;
5804 static void wq_watchdog_set_thresh(unsigned long thresh)
5806 wq_watchdog_thresh = 0;
5807 del_timer_sync(&wq_watchdog_timer);
5810 wq_watchdog_thresh = thresh;
5811 wq_watchdog_reset_touched();
5812 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5816 static int wq_watchdog_param_set_thresh(const char *val,
5817 const struct kernel_param *kp)
5819 unsigned long thresh;
5822 ret = kstrtoul(val, 0, &thresh);
5827 wq_watchdog_set_thresh(thresh);
5829 wq_watchdog_thresh = thresh;
5834 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5835 .set = wq_watchdog_param_set_thresh,
5836 .get = param_get_ulong,
5839 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5842 static void wq_watchdog_init(void)
5844 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5845 wq_watchdog_set_thresh(wq_watchdog_thresh);
5848 #else /* CONFIG_WQ_WATCHDOG */
5850 static inline void wq_watchdog_init(void) { }
5852 #endif /* CONFIG_WQ_WATCHDOG */
5854 static void __init wq_numa_init(void)
5859 if (num_possible_nodes() <= 1)
5862 if (wq_disable_numa) {
5863 pr_info("workqueue: NUMA affinity support disabled\n");
5867 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5868 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5871 * We want masks of possible CPUs of each node which isn't readily
5872 * available. Build one from cpu_to_node() which should have been
5873 * fully initialized by now.
5875 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5879 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5880 node_online(node) ? node : NUMA_NO_NODE));
5882 for_each_possible_cpu(cpu) {
5883 node = cpu_to_node(cpu);
5884 if (WARN_ON(node == NUMA_NO_NODE)) {
5885 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5886 /* happens iff arch is bonkers, let's just proceed */
5889 cpumask_set_cpu(cpu, tbl[node]);
5892 wq_numa_possible_cpumask = tbl;
5893 wq_numa_enabled = true;
5897 * workqueue_init_early - early init for workqueue subsystem
5899 * This is the first half of two-staged workqueue subsystem initialization
5900 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5901 * idr are up. It sets up all the data structures and system workqueues
5902 * and allows early boot code to create workqueues and queue/cancel work
5903 * items. Actual work item execution starts only after kthreads can be
5904 * created and scheduled right before early initcalls.
5906 int __init workqueue_init_early(void)
5908 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5909 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5912 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5914 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5915 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5917 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5919 /* initialize CPU pools */
5920 for_each_possible_cpu(cpu) {
5921 struct worker_pool *pool;
5924 for_each_cpu_worker_pool(pool, cpu) {
5925 BUG_ON(init_worker_pool(pool));
5927 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5928 pool->attrs->nice = std_nice[i++];
5929 pool->node = cpu_to_node(cpu);
5932 mutex_lock(&wq_pool_mutex);
5933 BUG_ON(worker_pool_assign_id(pool));
5934 mutex_unlock(&wq_pool_mutex);
5938 /* create default unbound and ordered wq attrs */
5939 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5940 struct workqueue_attrs *attrs;
5942 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5943 attrs->nice = std_nice[i];
5944 unbound_std_wq_attrs[i] = attrs;
5947 * An ordered wq should have only one pwq as ordering is
5948 * guaranteed by max_active which is enforced by pwqs.
5949 * Turn off NUMA so that dfl_pwq is used for all nodes.
5951 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5952 attrs->nice = std_nice[i];
5953 attrs->no_numa = true;
5954 ordered_wq_attrs[i] = attrs;
5957 system_wq = alloc_workqueue("events", 0, 0);
5958 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5959 system_long_wq = alloc_workqueue("events_long", 0, 0);
5960 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5961 WQ_UNBOUND_MAX_ACTIVE);
5962 system_freezable_wq = alloc_workqueue("events_freezable",
5964 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5965 WQ_POWER_EFFICIENT, 0);
5966 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5967 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5969 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5970 !system_unbound_wq || !system_freezable_wq ||
5971 !system_power_efficient_wq ||
5972 !system_freezable_power_efficient_wq);
5978 * workqueue_init - bring workqueue subsystem fully online
5980 * This is the latter half of two-staged workqueue subsystem initialization
5981 * and invoked as soon as kthreads can be created and scheduled.
5982 * Workqueues have been created and work items queued on them, but there
5983 * are no kworkers executing the work items yet. Populate the worker pools
5984 * with the initial workers and enable future kworker creations.
5986 int __init workqueue_init(void)
5988 struct workqueue_struct *wq;
5989 struct worker_pool *pool;
5993 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5994 * CPU to node mapping may not be available that early on some
5995 * archs such as power and arm64. As per-cpu pools created
5996 * previously could be missing node hint and unbound pools NUMA
5997 * affinity, fix them up.
5999 * Also, while iterating workqueues, create rescuers if requested.
6003 mutex_lock(&wq_pool_mutex);
6005 for_each_possible_cpu(cpu) {
6006 for_each_cpu_worker_pool(pool, cpu) {
6007 pool->node = cpu_to_node(cpu);
6011 list_for_each_entry(wq, &workqueues, list) {
6012 wq_update_unbound_numa(wq, smp_processor_id(), true);
6013 WARN(init_rescuer(wq),
6014 "workqueue: failed to create early rescuer for %s",
6018 mutex_unlock(&wq_pool_mutex);
6020 /* create the initial workers */
6021 for_each_online_cpu(cpu) {
6022 for_each_cpu_worker_pool(pool, cpu) {
6023 pool->flags &= ~POOL_DISASSOCIATED;
6024 BUG_ON(!create_worker(pool));
6028 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6029 BUG_ON(!create_worker(pool));