1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
53 #include <linux/kvm_para.h>
55 #include "workqueue_internal.h"
61 * A bound pool is either associated or disassociated with its CPU.
62 * While associated (!DISASSOCIATED), all workers are bound to the
63 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 * While DISASSOCIATED, the cpu may be offline and all workers have
67 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 * be executing on any CPU. The pool behaves as an unbound one.
70 * Note that DISASSOCIATED should be flipped only while holding
71 * wq_pool_attach_mutex to avoid changing binding state while
72 * worker_attach_to_pool() is in progress.
74 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
75 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
78 WORKER_DIE = 1 << 1, /* die die die */
79 WORKER_IDLE = 1 << 2, /* is idle */
80 WORKER_PREP = 1 << 3, /* preparing to run works */
81 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
82 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
83 WORKER_REBOUND = 1 << 8, /* worker was rebound */
85 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
86 WORKER_UNBOUND | WORKER_REBOUND,
88 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
90 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
91 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
93 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
94 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
96 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
97 /* call for help after 10ms
99 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
100 CREATE_COOLDOWN = HZ, /* time to breath after fail */
103 * Rescue workers are used only on emergencies and shared by
104 * all cpus. Give MIN_NICE.
106 RESCUER_NICE_LEVEL = MIN_NICE,
107 HIGHPRI_NICE_LEVEL = MIN_NICE,
113 * Structure fields follow one of the following exclusion rules.
115 * I: Modifiable by initialization/destruction paths and read-only for
118 * P: Preemption protected. Disabling preemption is enough and should
119 * only be modified and accessed from the local cpu.
121 * L: pool->lock protected. Access with pool->lock held.
123 * X: During normal operation, modification requires pool->lock and should
124 * be done only from local cpu. Either disabling preemption on local
125 * cpu or grabbing pool->lock is enough for read access. If
126 * POOL_DISASSOCIATED is set, it's identical to L.
128 * A: wq_pool_attach_mutex protected.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
134 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
136 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
139 * WQ: wq->mutex protected.
141 * WR: wq->mutex protected for writes. RCU protected for reads.
143 * MD: wq_mayday_lock protected.
146 /* struct worker is defined in workqueue_internal.h */
149 spinlock_t lock; /* the pool lock */
150 int cpu; /* I: the associated cpu */
151 int node; /* I: the associated node ID */
152 int id; /* I: pool ID */
153 unsigned int flags; /* X: flags */
155 unsigned long watchdog_ts; /* L: watchdog timestamp */
157 struct list_head worklist; /* L: list of pending works */
159 int nr_workers; /* L: total number of workers */
160 int nr_idle; /* L: currently idle workers */
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
170 struct worker *manager; /* L: purely informational */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
174 struct ida worker_ida; /* worker IDs for task name */
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp;
188 * Destruction of pool is RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
225 * Structure used to wait for workqueue flush.
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
261 struct wq_device *wq_dev; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
265 struct lock_class_key key;
266 struct lockdep_map lockdep_map;
268 char name[WQ_NAME_LEN]; /* I: workqueue name */
271 * Destruction of workqueue_struct is RCU protected to allow walking
272 * the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
283 static struct kmem_cache *pwq_cache;
285 static cpumask_var_t *wq_numa_possible_cpumask;
286 /* possible CPUs of each node */
288 static bool wq_disable_numa;
289 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
291 /* see the comment above the definition of WQ_POWER_EFFICIENT */
292 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
293 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
295 static bool wq_online; /* can kworkers be created yet? */
297 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
299 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
300 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
302 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
303 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
304 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
305 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
307 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
308 static bool workqueue_freezing; /* PL: have wqs started freezing? */
310 /* PL: allowable cpus for unbound wqs and work items */
311 static cpumask_var_t wq_unbound_cpumask;
313 /* CPU where unbound work was last round robin scheduled from this CPU */
314 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
317 * Local execution of unbound work items is no longer guaranteed. The
318 * following always forces round-robin CPU selection on unbound work items
319 * to uncover usages which depend on it.
321 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
322 static bool wq_debug_force_rr_cpu = true;
324 static bool wq_debug_force_rr_cpu = false;
326 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
328 /* the per-cpu worker pools */
329 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
331 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
333 /* PL: hash of all unbound pools keyed by pool->attrs */
334 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
336 /* I: attributes used when instantiating standard unbound pools on demand */
337 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
339 /* I: attributes used when instantiating ordered pools on demand */
340 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
342 struct workqueue_struct *system_wq __read_mostly;
343 EXPORT_SYMBOL(system_wq);
344 struct workqueue_struct *system_highpri_wq __read_mostly;
345 EXPORT_SYMBOL_GPL(system_highpri_wq);
346 struct workqueue_struct *system_long_wq __read_mostly;
347 EXPORT_SYMBOL_GPL(system_long_wq);
348 struct workqueue_struct *system_unbound_wq __read_mostly;
349 EXPORT_SYMBOL_GPL(system_unbound_wq);
350 struct workqueue_struct *system_freezable_wq __read_mostly;
351 EXPORT_SYMBOL_GPL(system_freezable_wq);
352 struct workqueue_struct *system_power_efficient_wq __read_mostly;
353 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
354 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
355 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
357 static int worker_thread(void *__worker);
358 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
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)
860 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
861 * and the nr_running increment below, we may ruin the nr_running reset
862 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
863 * pool. Protect against such race.
866 if (!(worker->flags & WORKER_NOT_RUNNING))
867 atomic_inc(&worker->pool->nr_running);
869 worker->sleeping = 0;
873 * wq_worker_sleeping - a worker is going to sleep
874 * @task: task going to sleep
876 * This function is called from schedule() when a busy worker is
877 * going to sleep. Preemption needs to be disabled to protect ->sleeping
880 void wq_worker_sleeping(struct task_struct *task)
882 struct worker *next, *worker = kthread_data(task);
883 struct worker_pool *pool;
886 * Rescuers, which may not have all the fields set up like normal
887 * workers, also reach here, let's not access anything before
888 * checking NOT_RUNNING.
890 if (worker->flags & WORKER_NOT_RUNNING)
895 /* Return if preempted before wq_worker_running() was reached */
896 if (worker->sleeping)
899 worker->sleeping = 1;
900 spin_lock_irq(&pool->lock);
903 * The counterpart of the following dec_and_test, implied mb,
904 * worklist not empty test sequence is in insert_work().
905 * Please read comment there.
907 * NOT_RUNNING is clear. This means that we're bound to and
908 * running on the local cpu w/ rq lock held and preemption
909 * disabled, which in turn means that none else could be
910 * manipulating idle_list, so dereferencing idle_list without pool
913 if (atomic_dec_and_test(&pool->nr_running) &&
914 !list_empty(&pool->worklist)) {
915 next = first_idle_worker(pool);
917 wake_up_process(next->task);
919 spin_unlock_irq(&pool->lock);
923 * wq_worker_last_func - retrieve worker's last work function
924 * @task: Task to retrieve last work function of.
926 * Determine the last function a worker executed. This is called from
927 * the scheduler to get a worker's last known identity.
930 * spin_lock_irq(rq->lock)
932 * This function is called during schedule() when a kworker is going
933 * to sleep. It's used by psi to identify aggregation workers during
934 * dequeuing, to allow periodic aggregation to shut-off when that
935 * worker is the last task in the system or cgroup to go to sleep.
937 * As this function doesn't involve any workqueue-related locking, it
938 * only returns stable values when called from inside the scheduler's
939 * queuing and dequeuing paths, when @task, which must be a kworker,
940 * is guaranteed to not be processing any works.
943 * The last work function %current executed as a worker, NULL if it
944 * hasn't executed any work yet.
946 work_func_t wq_worker_last_func(struct task_struct *task)
948 struct worker *worker = kthread_data(task);
950 return worker->last_func;
954 * worker_set_flags - set worker flags and adjust nr_running accordingly
956 * @flags: flags to set
958 * Set @flags in @worker->flags and adjust nr_running accordingly.
961 * spin_lock_irq(pool->lock)
963 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
965 struct worker_pool *pool = worker->pool;
967 WARN_ON_ONCE(worker->task != current);
969 /* If transitioning into NOT_RUNNING, adjust nr_running. */
970 if ((flags & WORKER_NOT_RUNNING) &&
971 !(worker->flags & WORKER_NOT_RUNNING)) {
972 atomic_dec(&pool->nr_running);
975 worker->flags |= flags;
979 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
981 * @flags: flags to clear
983 * Clear @flags in @worker->flags and adjust nr_running accordingly.
986 * spin_lock_irq(pool->lock)
988 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
990 struct worker_pool *pool = worker->pool;
991 unsigned int oflags = worker->flags;
993 WARN_ON_ONCE(worker->task != current);
995 worker->flags &= ~flags;
998 * If transitioning out of NOT_RUNNING, increment nr_running. Note
999 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1000 * of multiple flags, not a single flag.
1002 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1003 if (!(worker->flags & WORKER_NOT_RUNNING))
1004 atomic_inc(&pool->nr_running);
1008 * find_worker_executing_work - find worker which is executing a work
1009 * @pool: pool of interest
1010 * @work: work to find worker for
1012 * Find a worker which is executing @work on @pool by searching
1013 * @pool->busy_hash which is keyed by the address of @work. For a worker
1014 * to match, its current execution should match the address of @work and
1015 * its work function. This is to avoid unwanted dependency between
1016 * unrelated work executions through a work item being recycled while still
1019 * This is a bit tricky. A work item may be freed once its execution
1020 * starts and nothing prevents the freed area from being recycled for
1021 * another work item. If the same work item address ends up being reused
1022 * before the original execution finishes, workqueue will identify the
1023 * recycled work item as currently executing and make it wait until the
1024 * current execution finishes, introducing an unwanted dependency.
1026 * This function checks the work item address and work function to avoid
1027 * false positives. Note that this isn't complete as one may construct a
1028 * work function which can introduce dependency onto itself through a
1029 * recycled work item. Well, if somebody wants to shoot oneself in the
1030 * foot that badly, there's only so much we can do, and if such deadlock
1031 * actually occurs, it should be easy to locate the culprit work function.
1034 * spin_lock_irq(pool->lock).
1037 * Pointer to worker which is executing @work if found, %NULL
1040 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1041 struct work_struct *work)
1043 struct worker *worker;
1045 hash_for_each_possible(pool->busy_hash, worker, hentry,
1046 (unsigned long)work)
1047 if (worker->current_work == work &&
1048 worker->current_func == work->func)
1055 * move_linked_works - move linked works to a list
1056 * @work: start of series of works to be scheduled
1057 * @head: target list to append @work to
1058 * @nextp: out parameter for nested worklist walking
1060 * Schedule linked works starting from @work to @head. Work series to
1061 * be scheduled starts at @work and includes any consecutive work with
1062 * WORK_STRUCT_LINKED set in its predecessor.
1064 * If @nextp is not NULL, it's updated to point to the next work of
1065 * the last scheduled work. This allows move_linked_works() to be
1066 * nested inside outer list_for_each_entry_safe().
1069 * spin_lock_irq(pool->lock).
1071 static void move_linked_works(struct work_struct *work, struct list_head *head,
1072 struct work_struct **nextp)
1074 struct work_struct *n;
1077 * Linked worklist will always end before the end of the list,
1078 * use NULL for list head.
1080 list_for_each_entry_safe_from(work, n, NULL, entry) {
1081 list_move_tail(&work->entry, head);
1082 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1087 * If we're already inside safe list traversal and have moved
1088 * multiple works to the scheduled queue, the next position
1089 * needs to be updated.
1096 * get_pwq - get an extra reference on the specified pool_workqueue
1097 * @pwq: pool_workqueue to get
1099 * Obtain an extra reference on @pwq. The caller should guarantee that
1100 * @pwq has positive refcnt and be holding the matching pool->lock.
1102 static void get_pwq(struct pool_workqueue *pwq)
1104 lockdep_assert_held(&pwq->pool->lock);
1105 WARN_ON_ONCE(pwq->refcnt <= 0);
1110 * put_pwq - put a pool_workqueue reference
1111 * @pwq: pool_workqueue to put
1113 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1114 * destruction. The caller should be holding the matching pool->lock.
1116 static void put_pwq(struct pool_workqueue *pwq)
1118 lockdep_assert_held(&pwq->pool->lock);
1119 if (likely(--pwq->refcnt))
1121 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1124 * @pwq can't be released under pool->lock, bounce to
1125 * pwq_unbound_release_workfn(). This never recurses on the same
1126 * pool->lock as this path is taken only for unbound workqueues and
1127 * the release work item is scheduled on a per-cpu workqueue. To
1128 * avoid lockdep warning, unbound pool->locks are given lockdep
1129 * subclass of 1 in get_unbound_pool().
1131 schedule_work(&pwq->unbound_release_work);
1135 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1136 * @pwq: pool_workqueue to put (can be %NULL)
1138 * put_pwq() with locking. This function also allows %NULL @pwq.
1140 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1144 * As both pwqs and pools are RCU protected, the
1145 * following lock operations are safe.
1147 spin_lock_irq(&pwq->pool->lock);
1149 spin_unlock_irq(&pwq->pool->lock);
1153 static void pwq_activate_delayed_work(struct work_struct *work)
1155 struct pool_workqueue *pwq = get_work_pwq(work);
1157 trace_workqueue_activate_work(work);
1158 if (list_empty(&pwq->pool->worklist))
1159 pwq->pool->watchdog_ts = jiffies;
1160 move_linked_works(work, &pwq->pool->worklist, NULL);
1161 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1165 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1167 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1168 struct work_struct, entry);
1170 pwq_activate_delayed_work(work);
1174 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1175 * @pwq: pwq of interest
1176 * @color: color of work which left the queue
1178 * A work either has completed or is removed from pending queue,
1179 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1182 * spin_lock_irq(pool->lock).
1184 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1186 /* uncolored work items don't participate in flushing or nr_active */
1187 if (color == WORK_NO_COLOR)
1190 pwq->nr_in_flight[color]--;
1193 if (!list_empty(&pwq->delayed_works)) {
1194 /* one down, submit a delayed one */
1195 if (pwq->nr_active < pwq->max_active)
1196 pwq_activate_first_delayed(pwq);
1199 /* is flush in progress and are we at the flushing tip? */
1200 if (likely(pwq->flush_color != color))
1203 /* are there still in-flight works? */
1204 if (pwq->nr_in_flight[color])
1207 /* this pwq is done, clear flush_color */
1208 pwq->flush_color = -1;
1211 * If this was the last pwq, wake up the first flusher. It
1212 * will handle the rest.
1214 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1215 complete(&pwq->wq->first_flusher->done);
1221 * try_to_grab_pending - steal work item from worklist and disable irq
1222 * @work: work item to steal
1223 * @is_dwork: @work is a delayed_work
1224 * @flags: place to store irq state
1226 * Try to grab PENDING bit of @work. This function can handle @work in any
1227 * stable state - idle, on timer or on worklist.
1230 * 1 if @work was pending and we successfully stole PENDING
1231 * 0 if @work was idle and we claimed PENDING
1232 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1233 * -ENOENT if someone else is canceling @work, this state may persist
1234 * for arbitrarily long
1237 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1238 * interrupted while holding PENDING and @work off queue, irq must be
1239 * disabled on entry. This, combined with delayed_work->timer being
1240 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1242 * On successful return, >= 0, irq is disabled and the caller is
1243 * responsible for releasing it using local_irq_restore(*@flags).
1245 * This function is safe to call from any context including IRQ handler.
1247 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1248 unsigned long *flags)
1250 struct worker_pool *pool;
1251 struct pool_workqueue *pwq;
1253 local_irq_save(*flags);
1255 /* try to steal the timer if it exists */
1257 struct delayed_work *dwork = to_delayed_work(work);
1260 * dwork->timer is irqsafe. If del_timer() fails, it's
1261 * guaranteed that the timer is not queued anywhere and not
1262 * running on the local CPU.
1264 if (likely(del_timer(&dwork->timer)))
1268 /* try to claim PENDING the normal way */
1269 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1274 * The queueing is in progress, or it is already queued. Try to
1275 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1277 pool = get_work_pool(work);
1281 spin_lock(&pool->lock);
1283 * work->data is guaranteed to point to pwq only while the work
1284 * item is queued on pwq->wq, and both updating work->data to point
1285 * to pwq on queueing and to pool on dequeueing are done under
1286 * pwq->pool->lock. This in turn guarantees that, if work->data
1287 * points to pwq which is associated with a locked pool, the work
1288 * item is currently queued on that pool.
1290 pwq = get_work_pwq(work);
1291 if (pwq && pwq->pool == pool) {
1292 debug_work_deactivate(work);
1295 * A delayed work item cannot be grabbed directly because
1296 * it might have linked NO_COLOR work items which, if left
1297 * on the delayed_list, will confuse pwq->nr_active
1298 * management later on and cause stall. Make sure the work
1299 * item is activated before grabbing.
1301 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1302 pwq_activate_delayed_work(work);
1304 list_del_init(&work->entry);
1305 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1307 /* work->data points to pwq iff queued, point to pool */
1308 set_work_pool_and_keep_pending(work, pool->id);
1310 spin_unlock(&pool->lock);
1314 spin_unlock(&pool->lock);
1317 local_irq_restore(*flags);
1318 if (work_is_canceling(work))
1325 * insert_work - insert a work into a pool
1326 * @pwq: pwq @work belongs to
1327 * @work: work to insert
1328 * @head: insertion point
1329 * @extra_flags: extra WORK_STRUCT_* flags to set
1331 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1332 * work_struct flags.
1335 * spin_lock_irq(pool->lock).
1337 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1338 struct list_head *head, unsigned int extra_flags)
1340 struct worker_pool *pool = pwq->pool;
1342 /* we own @work, set data and link */
1343 set_work_pwq(work, pwq, extra_flags);
1344 list_add_tail(&work->entry, head);
1348 * Ensure either wq_worker_sleeping() sees the above
1349 * list_add_tail() or we see zero nr_running to avoid workers lying
1350 * around lazily while there are works to be processed.
1354 if (__need_more_worker(pool))
1355 wake_up_worker(pool);
1359 * Test whether @work is being queued from another work executing on the
1362 static bool is_chained_work(struct workqueue_struct *wq)
1364 struct worker *worker;
1366 worker = current_wq_worker();
1368 * Return %true iff I'm a worker executing a work item on @wq. If
1369 * I'm @worker, it's safe to dereference it without locking.
1371 return worker && worker->current_pwq->wq == wq;
1375 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1376 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1377 * avoid perturbing sensitive tasks.
1379 static int wq_select_unbound_cpu(int cpu)
1381 static bool printed_dbg_warning;
1384 if (likely(!wq_debug_force_rr_cpu)) {
1385 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1387 } else if (!printed_dbg_warning) {
1388 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1389 printed_dbg_warning = true;
1392 if (cpumask_empty(wq_unbound_cpumask))
1395 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1396 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1397 if (unlikely(new_cpu >= nr_cpu_ids)) {
1398 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1399 if (unlikely(new_cpu >= nr_cpu_ids))
1402 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1407 static void __queue_work(int cpu, struct workqueue_struct *wq,
1408 struct work_struct *work)
1410 struct pool_workqueue *pwq;
1411 struct worker_pool *last_pool;
1412 struct list_head *worklist;
1413 unsigned int work_flags;
1414 unsigned int req_cpu = cpu;
1417 * While a work item is PENDING && off queue, a task trying to
1418 * steal the PENDING will busy-loop waiting for it to either get
1419 * queued or lose PENDING. Grabbing PENDING and queueing should
1420 * happen with IRQ disabled.
1422 lockdep_assert_irqs_disabled();
1425 /* if draining, only works from the same workqueue are allowed */
1426 if (unlikely(wq->flags & __WQ_DRAINING) &&
1427 WARN_ON_ONCE(!is_chained_work(wq)))
1431 /* pwq which will be used unless @work is executing elsewhere */
1432 if (wq->flags & WQ_UNBOUND) {
1433 if (req_cpu == WORK_CPU_UNBOUND)
1434 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1435 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1437 if (req_cpu == WORK_CPU_UNBOUND)
1438 cpu = raw_smp_processor_id();
1439 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1443 * If @work was previously on a different pool, it might still be
1444 * running there, in which case the work needs to be queued on that
1445 * pool to guarantee non-reentrancy.
1447 last_pool = get_work_pool(work);
1448 if (last_pool && last_pool != pwq->pool) {
1449 struct worker *worker;
1451 spin_lock(&last_pool->lock);
1453 worker = find_worker_executing_work(last_pool, work);
1455 if (worker && worker->current_pwq->wq == wq) {
1456 pwq = worker->current_pwq;
1458 /* meh... not running there, queue here */
1459 spin_unlock(&last_pool->lock);
1460 spin_lock(&pwq->pool->lock);
1463 spin_lock(&pwq->pool->lock);
1467 * pwq is determined and locked. For unbound pools, we could have
1468 * raced with pwq release and it could already be dead. If its
1469 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1470 * without another pwq replacing it in the numa_pwq_tbl or while
1471 * work items are executing on it, so the retrying is guaranteed to
1472 * make forward-progress.
1474 if (unlikely(!pwq->refcnt)) {
1475 if (wq->flags & WQ_UNBOUND) {
1476 spin_unlock(&pwq->pool->lock);
1481 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1485 /* pwq determined, queue */
1486 trace_workqueue_queue_work(req_cpu, pwq, work);
1488 if (WARN_ON(!list_empty(&work->entry)))
1491 pwq->nr_in_flight[pwq->work_color]++;
1492 work_flags = work_color_to_flags(pwq->work_color);
1494 if (likely(pwq->nr_active < pwq->max_active)) {
1495 trace_workqueue_activate_work(work);
1497 worklist = &pwq->pool->worklist;
1498 if (list_empty(worklist))
1499 pwq->pool->watchdog_ts = jiffies;
1501 work_flags |= WORK_STRUCT_DELAYED;
1502 worklist = &pwq->delayed_works;
1505 debug_work_activate(work);
1506 insert_work(pwq, work, worklist, work_flags);
1509 spin_unlock(&pwq->pool->lock);
1514 * queue_work_on - queue work on specific cpu
1515 * @cpu: CPU number to execute work on
1516 * @wq: workqueue to use
1517 * @work: work to queue
1519 * We queue the work to a specific CPU, the caller must ensure it
1522 * Return: %false if @work was already on a queue, %true otherwise.
1524 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1525 struct work_struct *work)
1528 unsigned long flags;
1530 local_irq_save(flags);
1532 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1533 __queue_work(cpu, wq, work);
1537 local_irq_restore(flags);
1540 EXPORT_SYMBOL(queue_work_on);
1543 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1544 * @node: NUMA node ID that we want to select a CPU from
1546 * This function will attempt to find a "random" cpu available on a given
1547 * node. If there are no CPUs available on the given node it will return
1548 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1549 * available CPU if we need to schedule this work.
1551 static int workqueue_select_cpu_near(int node)
1555 /* No point in doing this if NUMA isn't enabled for workqueues */
1556 if (!wq_numa_enabled)
1557 return WORK_CPU_UNBOUND;
1559 /* Delay binding to CPU if node is not valid or online */
1560 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1561 return WORK_CPU_UNBOUND;
1563 /* Use local node/cpu if we are already there */
1564 cpu = raw_smp_processor_id();
1565 if (node == cpu_to_node(cpu))
1568 /* Use "random" otherwise know as "first" online CPU of node */
1569 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1571 /* If CPU is valid return that, otherwise just defer */
1572 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1576 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1577 * @node: NUMA node that we are targeting the work for
1578 * @wq: workqueue to use
1579 * @work: work to queue
1581 * We queue the work to a "random" CPU within a given NUMA node. The basic
1582 * idea here is to provide a way to somehow associate work with a given
1585 * This function will only make a best effort attempt at getting this onto
1586 * the right NUMA node. If no node is requested or the requested node is
1587 * offline then we just fall back to standard queue_work behavior.
1589 * Currently the "random" CPU ends up being the first available CPU in the
1590 * intersection of cpu_online_mask and the cpumask of the node, unless we
1591 * are running on the node. In that case we just use the current CPU.
1593 * Return: %false if @work was already on a queue, %true otherwise.
1595 bool queue_work_node(int node, struct workqueue_struct *wq,
1596 struct work_struct *work)
1598 unsigned long flags;
1602 * This current implementation is specific to unbound workqueues.
1603 * Specifically we only return the first available CPU for a given
1604 * node instead of cycling through individual CPUs within the node.
1606 * If this is used with a per-cpu workqueue then the logic in
1607 * workqueue_select_cpu_near would need to be updated to allow for
1608 * some round robin type logic.
1610 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1612 local_irq_save(flags);
1614 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1615 int cpu = workqueue_select_cpu_near(node);
1617 __queue_work(cpu, wq, work);
1621 local_irq_restore(flags);
1624 EXPORT_SYMBOL_GPL(queue_work_node);
1626 void delayed_work_timer_fn(struct timer_list *t)
1628 struct delayed_work *dwork = from_timer(dwork, t, timer);
1630 /* should have been called from irqsafe timer with irq already off */
1631 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1633 EXPORT_SYMBOL(delayed_work_timer_fn);
1635 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1636 struct delayed_work *dwork, unsigned long delay)
1638 struct timer_list *timer = &dwork->timer;
1639 struct work_struct *work = &dwork->work;
1642 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1643 WARN_ON_ONCE(timer_pending(timer));
1644 WARN_ON_ONCE(!list_empty(&work->entry));
1647 * If @delay is 0, queue @dwork->work immediately. This is for
1648 * both optimization and correctness. The earliest @timer can
1649 * expire is on the closest next tick and delayed_work users depend
1650 * on that there's no such delay when @delay is 0.
1653 __queue_work(cpu, wq, &dwork->work);
1659 timer->expires = jiffies + delay;
1661 if (unlikely(cpu != WORK_CPU_UNBOUND))
1662 add_timer_on(timer, cpu);
1668 * queue_delayed_work_on - queue work on specific CPU after delay
1669 * @cpu: CPU number to execute work on
1670 * @wq: workqueue to use
1671 * @dwork: work to queue
1672 * @delay: number of jiffies to wait before queueing
1674 * Return: %false if @work was already on a queue, %true otherwise. If
1675 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1678 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1679 struct delayed_work *dwork, unsigned long delay)
1681 struct work_struct *work = &dwork->work;
1683 unsigned long flags;
1685 /* read the comment in __queue_work() */
1686 local_irq_save(flags);
1688 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1689 __queue_delayed_work(cpu, wq, dwork, delay);
1693 local_irq_restore(flags);
1696 EXPORT_SYMBOL(queue_delayed_work_on);
1699 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1700 * @cpu: CPU number to execute work on
1701 * @wq: workqueue to use
1702 * @dwork: work to queue
1703 * @delay: number of jiffies to wait before queueing
1705 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1706 * modify @dwork's timer so that it expires after @delay. If @delay is
1707 * zero, @work is guaranteed to be scheduled immediately regardless of its
1710 * Return: %false if @dwork was idle and queued, %true if @dwork was
1711 * pending and its timer was modified.
1713 * This function is safe to call from any context including IRQ handler.
1714 * See try_to_grab_pending() for details.
1716 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1717 struct delayed_work *dwork, unsigned long delay)
1719 unsigned long flags;
1723 ret = try_to_grab_pending(&dwork->work, true, &flags);
1724 } while (unlikely(ret == -EAGAIN));
1726 if (likely(ret >= 0)) {
1727 __queue_delayed_work(cpu, wq, dwork, delay);
1728 local_irq_restore(flags);
1731 /* -ENOENT from try_to_grab_pending() becomes %true */
1734 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1736 static void rcu_work_rcufn(struct rcu_head *rcu)
1738 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1740 /* read the comment in __queue_work() */
1741 local_irq_disable();
1742 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1747 * queue_rcu_work - queue work after a RCU grace period
1748 * @wq: workqueue to use
1749 * @rwork: work to queue
1751 * Return: %false if @rwork was already pending, %true otherwise. Note
1752 * that a full RCU grace period is guaranteed only after a %true return.
1753 * While @rwork is guaranteed to be executed after a %false return, the
1754 * execution may happen before a full RCU grace period has passed.
1756 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1758 struct work_struct *work = &rwork->work;
1760 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1762 call_rcu(&rwork->rcu, rcu_work_rcufn);
1768 EXPORT_SYMBOL(queue_rcu_work);
1771 * worker_enter_idle - enter idle state
1772 * @worker: worker which is entering idle state
1774 * @worker is entering idle state. Update stats and idle timer if
1778 * spin_lock_irq(pool->lock).
1780 static void worker_enter_idle(struct worker *worker)
1782 struct worker_pool *pool = worker->pool;
1784 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1785 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1786 (worker->hentry.next || worker->hentry.pprev)))
1789 /* can't use worker_set_flags(), also called from create_worker() */
1790 worker->flags |= WORKER_IDLE;
1792 worker->last_active = jiffies;
1794 /* idle_list is LIFO */
1795 list_add(&worker->entry, &pool->idle_list);
1797 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1798 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1801 * Sanity check nr_running. Because unbind_workers() releases
1802 * pool->lock between setting %WORKER_UNBOUND and zapping
1803 * nr_running, the warning may trigger spuriously. Check iff
1804 * unbind is not in progress.
1806 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1807 pool->nr_workers == pool->nr_idle &&
1808 atomic_read(&pool->nr_running));
1812 * worker_leave_idle - leave idle state
1813 * @worker: worker which is leaving idle state
1815 * @worker is leaving idle state. Update stats.
1818 * spin_lock_irq(pool->lock).
1820 static void worker_leave_idle(struct worker *worker)
1822 struct worker_pool *pool = worker->pool;
1824 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1826 worker_clr_flags(worker, WORKER_IDLE);
1828 list_del_init(&worker->entry);
1831 static struct worker *alloc_worker(int node)
1833 struct worker *worker;
1835 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1837 INIT_LIST_HEAD(&worker->entry);
1838 INIT_LIST_HEAD(&worker->scheduled);
1839 INIT_LIST_HEAD(&worker->node);
1840 /* on creation a worker is in !idle && prep state */
1841 worker->flags = WORKER_PREP;
1847 * worker_attach_to_pool() - attach a worker to a pool
1848 * @worker: worker to be attached
1849 * @pool: the target pool
1851 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1852 * cpu-binding of @worker are kept coordinated with the pool across
1855 static void worker_attach_to_pool(struct worker *worker,
1856 struct worker_pool *pool)
1858 mutex_lock(&wq_pool_attach_mutex);
1861 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1862 * stable across this function. See the comments above the flag
1863 * definition for details.
1865 if (pool->flags & POOL_DISASSOCIATED)
1866 worker->flags |= WORKER_UNBOUND;
1868 if (worker->rescue_wq)
1869 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1871 list_add_tail(&worker->node, &pool->workers);
1872 worker->pool = pool;
1874 mutex_unlock(&wq_pool_attach_mutex);
1878 * worker_detach_from_pool() - detach a worker from its pool
1879 * @worker: worker which is attached to its pool
1881 * Undo the attaching which had been done in worker_attach_to_pool(). The
1882 * caller worker shouldn't access to the pool after detached except it has
1883 * other reference to the pool.
1885 static void worker_detach_from_pool(struct worker *worker)
1887 struct worker_pool *pool = worker->pool;
1888 struct completion *detach_completion = NULL;
1890 mutex_lock(&wq_pool_attach_mutex);
1892 list_del(&worker->node);
1893 worker->pool = NULL;
1895 if (list_empty(&pool->workers))
1896 detach_completion = pool->detach_completion;
1897 mutex_unlock(&wq_pool_attach_mutex);
1899 /* clear leftover flags without pool->lock after it is detached */
1900 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1902 if (detach_completion)
1903 complete(detach_completion);
1907 * create_worker - create a new workqueue worker
1908 * @pool: pool the new worker will belong to
1910 * Create and start a new worker which is attached to @pool.
1913 * Might sleep. Does GFP_KERNEL allocations.
1916 * Pointer to the newly created worker.
1918 static struct worker *create_worker(struct worker_pool *pool)
1920 struct worker *worker = NULL;
1924 /* ID is needed to determine kthread name */
1925 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1929 worker = alloc_worker(pool->node);
1936 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1937 pool->attrs->nice < 0 ? "H" : "");
1939 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1941 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1942 "kworker/%s", id_buf);
1943 if (IS_ERR(worker->task))
1946 set_user_nice(worker->task, pool->attrs->nice);
1947 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1949 /* successful, attach the worker to the pool */
1950 worker_attach_to_pool(worker, pool);
1952 /* start the newly created worker */
1953 spin_lock_irq(&pool->lock);
1954 worker->pool->nr_workers++;
1955 worker_enter_idle(worker);
1956 wake_up_process(worker->task);
1957 spin_unlock_irq(&pool->lock);
1963 ida_simple_remove(&pool->worker_ida, id);
1969 * destroy_worker - destroy a workqueue worker
1970 * @worker: worker to be destroyed
1972 * Destroy @worker and adjust @pool stats accordingly. The worker should
1976 * spin_lock_irq(pool->lock).
1978 static void destroy_worker(struct worker *worker)
1980 struct worker_pool *pool = worker->pool;
1982 lockdep_assert_held(&pool->lock);
1984 /* sanity check frenzy */
1985 if (WARN_ON(worker->current_work) ||
1986 WARN_ON(!list_empty(&worker->scheduled)) ||
1987 WARN_ON(!(worker->flags & WORKER_IDLE)))
1993 list_del_init(&worker->entry);
1994 worker->flags |= WORKER_DIE;
1995 wake_up_process(worker->task);
1998 static void idle_worker_timeout(struct timer_list *t)
2000 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2002 spin_lock_irq(&pool->lock);
2004 while (too_many_workers(pool)) {
2005 struct worker *worker;
2006 unsigned long expires;
2008 /* idle_list is kept in LIFO order, check the last one */
2009 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2010 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2012 if (time_before(jiffies, expires)) {
2013 mod_timer(&pool->idle_timer, expires);
2017 destroy_worker(worker);
2020 spin_unlock_irq(&pool->lock);
2023 static void send_mayday(struct work_struct *work)
2025 struct pool_workqueue *pwq = get_work_pwq(work);
2026 struct workqueue_struct *wq = pwq->wq;
2028 lockdep_assert_held(&wq_mayday_lock);
2033 /* mayday mayday mayday */
2034 if (list_empty(&pwq->mayday_node)) {
2036 * If @pwq is for an unbound wq, its base ref may be put at
2037 * any time due to an attribute change. Pin @pwq until the
2038 * rescuer is done with it.
2041 list_add_tail(&pwq->mayday_node, &wq->maydays);
2042 wake_up_process(wq->rescuer->task);
2046 static void pool_mayday_timeout(struct timer_list *t)
2048 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2049 struct work_struct *work;
2051 spin_lock_irq(&pool->lock);
2052 spin_lock(&wq_mayday_lock); /* for wq->maydays */
2054 if (need_to_create_worker(pool)) {
2056 * We've been trying to create a new worker but
2057 * haven't been successful. We might be hitting an
2058 * allocation deadlock. Send distress signals to
2061 list_for_each_entry(work, &pool->worklist, entry)
2065 spin_unlock(&wq_mayday_lock);
2066 spin_unlock_irq(&pool->lock);
2068 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2072 * maybe_create_worker - create a new worker if necessary
2073 * @pool: pool to create a new worker for
2075 * Create a new worker for @pool if necessary. @pool is guaranteed to
2076 * have at least one idle worker on return from this function. If
2077 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2078 * sent to all rescuers with works scheduled on @pool to resolve
2079 * possible allocation deadlock.
2081 * On return, need_to_create_worker() is guaranteed to be %false and
2082 * may_start_working() %true.
2085 * spin_lock_irq(pool->lock) which may be released and regrabbed
2086 * multiple times. Does GFP_KERNEL allocations. Called only from
2089 static void maybe_create_worker(struct worker_pool *pool)
2090 __releases(&pool->lock)
2091 __acquires(&pool->lock)
2094 spin_unlock_irq(&pool->lock);
2096 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2097 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2100 if (create_worker(pool) || !need_to_create_worker(pool))
2103 schedule_timeout_interruptible(CREATE_COOLDOWN);
2105 if (!need_to_create_worker(pool))
2109 del_timer_sync(&pool->mayday_timer);
2110 spin_lock_irq(&pool->lock);
2112 * This is necessary even after a new worker was just successfully
2113 * created as @pool->lock was dropped and the new worker might have
2114 * already become busy.
2116 if (need_to_create_worker(pool))
2121 * manage_workers - manage worker pool
2124 * Assume the manager role and manage the worker pool @worker belongs
2125 * to. At any given time, there can be only zero or one manager per
2126 * pool. The exclusion is handled automatically by this function.
2128 * The caller can safely start processing works on false return. On
2129 * true return, it's guaranteed that need_to_create_worker() is false
2130 * and may_start_working() is true.
2133 * spin_lock_irq(pool->lock) which may be released and regrabbed
2134 * multiple times. Does GFP_KERNEL allocations.
2137 * %false if the pool doesn't need management and the caller can safely
2138 * start processing works, %true if management function was performed and
2139 * the conditions that the caller verified before calling the function may
2140 * no longer be true.
2142 static bool manage_workers(struct worker *worker)
2144 struct worker_pool *pool = worker->pool;
2146 if (pool->flags & POOL_MANAGER_ACTIVE)
2149 pool->flags |= POOL_MANAGER_ACTIVE;
2150 pool->manager = worker;
2152 maybe_create_worker(pool);
2154 pool->manager = NULL;
2155 pool->flags &= ~POOL_MANAGER_ACTIVE;
2156 wake_up(&wq_manager_wait);
2161 * process_one_work - process single work
2163 * @work: work to process
2165 * Process @work. This function contains all the logics necessary to
2166 * process a single work including synchronization against and
2167 * interaction with other workers on the same cpu, queueing and
2168 * flushing. As long as context requirement is met, any worker can
2169 * call this function to process a work.
2172 * spin_lock_irq(pool->lock) which is released and regrabbed.
2174 static void process_one_work(struct worker *worker, struct work_struct *work)
2175 __releases(&pool->lock)
2176 __acquires(&pool->lock)
2178 struct pool_workqueue *pwq = get_work_pwq(work);
2179 struct worker_pool *pool = worker->pool;
2180 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2182 struct worker *collision;
2183 #ifdef CONFIG_LOCKDEP
2185 * It is permissible to free the struct work_struct from
2186 * inside the function that is called from it, this we need to
2187 * take into account for lockdep too. To avoid bogus "held
2188 * lock freed" warnings as well as problems when looking into
2189 * work->lockdep_map, make a copy and use that here.
2191 struct lockdep_map lockdep_map;
2193 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2195 /* ensure we're on the correct CPU */
2196 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2197 raw_smp_processor_id() != pool->cpu);
2200 * A single work shouldn't be executed concurrently by
2201 * multiple workers on a single cpu. Check whether anyone is
2202 * already processing the work. If so, defer the work to the
2203 * currently executing one.
2205 collision = find_worker_executing_work(pool, work);
2206 if (unlikely(collision)) {
2207 move_linked_works(work, &collision->scheduled, NULL);
2211 /* claim and dequeue */
2212 debug_work_deactivate(work);
2213 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2214 worker->current_work = work;
2215 worker->current_func = work->func;
2216 worker->current_pwq = pwq;
2217 work_color = get_work_color(work);
2220 * Record wq name for cmdline and debug reporting, may get
2221 * overridden through set_worker_desc().
2223 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2225 list_del_init(&work->entry);
2228 * CPU intensive works don't participate in concurrency management.
2229 * They're the scheduler's responsibility. This takes @worker out
2230 * of concurrency management and the next code block will chain
2231 * execution of the pending work items.
2233 if (unlikely(cpu_intensive))
2234 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2237 * Wake up another worker if necessary. The condition is always
2238 * false for normal per-cpu workers since nr_running would always
2239 * be >= 1 at this point. This is used to chain execution of the
2240 * pending work items for WORKER_NOT_RUNNING workers such as the
2241 * UNBOUND and CPU_INTENSIVE ones.
2243 if (need_more_worker(pool))
2244 wake_up_worker(pool);
2247 * Record the last pool and clear PENDING which should be the last
2248 * update to @work. Also, do this inside @pool->lock so that
2249 * PENDING and queued state changes happen together while IRQ is
2252 set_work_pool_and_clear_pending(work, pool->id);
2254 spin_unlock_irq(&pool->lock);
2256 lock_map_acquire(&pwq->wq->lockdep_map);
2257 lock_map_acquire(&lockdep_map);
2259 * Strictly speaking we should mark the invariant state without holding
2260 * any locks, that is, before these two lock_map_acquire()'s.
2262 * However, that would result in:
2269 * Which would create W1->C->W1 dependencies, even though there is no
2270 * actual deadlock possible. There are two solutions, using a
2271 * read-recursive acquire on the work(queue) 'locks', but this will then
2272 * hit the lockdep limitation on recursive locks, or simply discard
2275 * AFAICT there is no possible deadlock scenario between the
2276 * flush_work() and complete() primitives (except for single-threaded
2277 * workqueues), so hiding them isn't a problem.
2279 lockdep_invariant_state(true);
2280 trace_workqueue_execute_start(work);
2281 worker->current_func(work);
2283 * While we must be careful to not use "work" after this, the trace
2284 * point will only record its address.
2286 trace_workqueue_execute_end(work);
2287 lock_map_release(&lockdep_map);
2288 lock_map_release(&pwq->wq->lockdep_map);
2290 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2291 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2292 " last function: %ps\n",
2293 current->comm, preempt_count(), task_pid_nr(current),
2294 worker->current_func);
2295 debug_show_held_locks(current);
2300 * The following prevents a kworker from hogging CPU on !PREEMPT
2301 * kernels, where a requeueing work item waiting for something to
2302 * happen could deadlock with stop_machine as such work item could
2303 * indefinitely requeue itself while all other CPUs are trapped in
2304 * stop_machine. At the same time, report a quiescent RCU state so
2305 * the same condition doesn't freeze RCU.
2309 spin_lock_irq(&pool->lock);
2311 /* clear cpu intensive status */
2312 if (unlikely(cpu_intensive))
2313 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2315 /* tag the worker for identification in schedule() */
2316 worker->last_func = worker->current_func;
2318 /* we're done with it, release */
2319 hash_del(&worker->hentry);
2320 worker->current_work = NULL;
2321 worker->current_func = NULL;
2322 worker->current_pwq = NULL;
2323 pwq_dec_nr_in_flight(pwq, work_color);
2327 * process_scheduled_works - process scheduled works
2330 * Process all scheduled works. Please note that the scheduled list
2331 * may change while processing a work, so this function repeatedly
2332 * fetches a work from the top and executes it.
2335 * spin_lock_irq(pool->lock) which may be released and regrabbed
2338 static void process_scheduled_works(struct worker *worker)
2340 while (!list_empty(&worker->scheduled)) {
2341 struct work_struct *work = list_first_entry(&worker->scheduled,
2342 struct work_struct, entry);
2343 process_one_work(worker, work);
2347 static void set_pf_worker(bool val)
2349 mutex_lock(&wq_pool_attach_mutex);
2351 current->flags |= PF_WQ_WORKER;
2353 current->flags &= ~PF_WQ_WORKER;
2354 mutex_unlock(&wq_pool_attach_mutex);
2358 * worker_thread - the worker thread function
2361 * The worker thread function. All workers belong to a worker_pool -
2362 * either a per-cpu one or dynamic unbound one. These workers process all
2363 * work items regardless of their specific target workqueue. The only
2364 * exception is work items which belong to workqueues with a rescuer which
2365 * will be explained in rescuer_thread().
2369 static int worker_thread(void *__worker)
2371 struct worker *worker = __worker;
2372 struct worker_pool *pool = worker->pool;
2374 /* tell the scheduler that this is a workqueue worker */
2375 set_pf_worker(true);
2377 spin_lock_irq(&pool->lock);
2379 /* am I supposed to die? */
2380 if (unlikely(worker->flags & WORKER_DIE)) {
2381 spin_unlock_irq(&pool->lock);
2382 WARN_ON_ONCE(!list_empty(&worker->entry));
2383 set_pf_worker(false);
2385 set_task_comm(worker->task, "kworker/dying");
2386 ida_simple_remove(&pool->worker_ida, worker->id);
2387 worker_detach_from_pool(worker);
2392 worker_leave_idle(worker);
2394 /* no more worker necessary? */
2395 if (!need_more_worker(pool))
2398 /* do we need to manage? */
2399 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2403 * ->scheduled list can only be filled while a worker is
2404 * preparing to process a work or actually processing it.
2405 * Make sure nobody diddled with it while I was sleeping.
2407 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2410 * Finish PREP stage. We're guaranteed to have at least one idle
2411 * worker or that someone else has already assumed the manager
2412 * role. This is where @worker starts participating in concurrency
2413 * management if applicable and concurrency management is restored
2414 * after being rebound. See rebind_workers() for details.
2416 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2419 struct work_struct *work =
2420 list_first_entry(&pool->worklist,
2421 struct work_struct, entry);
2423 pool->watchdog_ts = jiffies;
2425 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2426 /* optimization path, not strictly necessary */
2427 process_one_work(worker, work);
2428 if (unlikely(!list_empty(&worker->scheduled)))
2429 process_scheduled_works(worker);
2431 move_linked_works(work, &worker->scheduled, NULL);
2432 process_scheduled_works(worker);
2434 } while (keep_working(pool));
2436 worker_set_flags(worker, WORKER_PREP);
2439 * pool->lock is held and there's no work to process and no need to
2440 * manage, sleep. Workers are woken up only while holding
2441 * pool->lock or from local cpu, so setting the current state
2442 * before releasing pool->lock is enough to prevent losing any
2445 worker_enter_idle(worker);
2446 __set_current_state(TASK_IDLE);
2447 spin_unlock_irq(&pool->lock);
2453 * rescuer_thread - the rescuer thread function
2456 * Workqueue rescuer thread function. There's one rescuer for each
2457 * workqueue which has WQ_MEM_RECLAIM set.
2459 * Regular work processing on a pool may block trying to create a new
2460 * worker which uses GFP_KERNEL allocation which has slight chance of
2461 * developing into deadlock if some works currently on the same queue
2462 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2463 * the problem rescuer solves.
2465 * When such condition is possible, the pool summons rescuers of all
2466 * workqueues which have works queued on the pool and let them process
2467 * those works so that forward progress can be guaranteed.
2469 * This should happen rarely.
2473 static int rescuer_thread(void *__rescuer)
2475 struct worker *rescuer = __rescuer;
2476 struct workqueue_struct *wq = rescuer->rescue_wq;
2477 struct list_head *scheduled = &rescuer->scheduled;
2480 set_user_nice(current, RESCUER_NICE_LEVEL);
2483 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2484 * doesn't participate in concurrency management.
2486 set_pf_worker(true);
2488 set_current_state(TASK_IDLE);
2491 * By the time the rescuer is requested to stop, the workqueue
2492 * shouldn't have any work pending, but @wq->maydays may still have
2493 * pwq(s) queued. This can happen by non-rescuer workers consuming
2494 * all the work items before the rescuer got to them. Go through
2495 * @wq->maydays processing before acting on should_stop so that the
2496 * list is always empty on exit.
2498 should_stop = kthread_should_stop();
2500 /* see whether any pwq is asking for help */
2501 spin_lock_irq(&wq_mayday_lock);
2503 while (!list_empty(&wq->maydays)) {
2504 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2505 struct pool_workqueue, mayday_node);
2506 struct worker_pool *pool = pwq->pool;
2507 struct work_struct *work, *n;
2510 __set_current_state(TASK_RUNNING);
2511 list_del_init(&pwq->mayday_node);
2513 spin_unlock_irq(&wq_mayday_lock);
2515 worker_attach_to_pool(rescuer, pool);
2517 spin_lock_irq(&pool->lock);
2520 * Slurp in all works issued via this workqueue and
2523 WARN_ON_ONCE(!list_empty(scheduled));
2524 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2525 if (get_work_pwq(work) == pwq) {
2527 pool->watchdog_ts = jiffies;
2528 move_linked_works(work, scheduled, &n);
2533 if (!list_empty(scheduled)) {
2534 process_scheduled_works(rescuer);
2537 * The above execution of rescued work items could
2538 * have created more to rescue through
2539 * pwq_activate_first_delayed() or chained
2540 * queueing. Let's put @pwq back on mayday list so
2541 * that such back-to-back work items, which may be
2542 * being used to relieve memory pressure, don't
2543 * incur MAYDAY_INTERVAL delay inbetween.
2545 if (need_to_create_worker(pool)) {
2546 spin_lock(&wq_mayday_lock);
2548 * Queue iff we aren't racing destruction
2549 * and somebody else hasn't queued it already.
2551 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2553 list_add_tail(&pwq->mayday_node, &wq->maydays);
2555 spin_unlock(&wq_mayday_lock);
2560 * Put the reference grabbed by send_mayday(). @pool won't
2561 * go away while we're still attached to it.
2566 * Leave this pool. If need_more_worker() is %true, notify a
2567 * regular worker; otherwise, we end up with 0 concurrency
2568 * and stalling the execution.
2570 if (need_more_worker(pool))
2571 wake_up_worker(pool);
2573 spin_unlock_irq(&pool->lock);
2575 worker_detach_from_pool(rescuer);
2577 spin_lock_irq(&wq_mayday_lock);
2580 spin_unlock_irq(&wq_mayday_lock);
2583 __set_current_state(TASK_RUNNING);
2584 set_pf_worker(false);
2588 /* rescuers should never participate in concurrency management */
2589 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2595 * check_flush_dependency - check for flush dependency sanity
2596 * @target_wq: workqueue being flushed
2597 * @target_work: work item being flushed (NULL for workqueue flushes)
2599 * %current is trying to flush the whole @target_wq or @target_work on it.
2600 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2601 * reclaiming memory or running on a workqueue which doesn't have
2602 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2605 static void check_flush_dependency(struct workqueue_struct *target_wq,
2606 struct work_struct *target_work)
2608 work_func_t target_func = target_work ? target_work->func : NULL;
2609 struct worker *worker;
2611 if (target_wq->flags & WQ_MEM_RECLAIM)
2614 worker = current_wq_worker();
2616 WARN_ONCE(current->flags & PF_MEMALLOC,
2617 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2618 current->pid, current->comm, target_wq->name, target_func);
2619 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2620 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2621 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2622 worker->current_pwq->wq->name, worker->current_func,
2623 target_wq->name, target_func);
2627 struct work_struct work;
2628 struct completion done;
2629 struct task_struct *task; /* purely informational */
2632 static void wq_barrier_func(struct work_struct *work)
2634 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2635 complete(&barr->done);
2639 * insert_wq_barrier - insert a barrier work
2640 * @pwq: pwq to insert barrier into
2641 * @barr: wq_barrier to insert
2642 * @target: target work to attach @barr to
2643 * @worker: worker currently executing @target, NULL if @target is not executing
2645 * @barr is linked to @target such that @barr is completed only after
2646 * @target finishes execution. Please note that the ordering
2647 * guarantee is observed only with respect to @target and on the local
2650 * Currently, a queued barrier can't be canceled. This is because
2651 * try_to_grab_pending() can't determine whether the work to be
2652 * grabbed is at the head of the queue and thus can't clear LINKED
2653 * flag of the previous work while there must be a valid next work
2654 * after a work with LINKED flag set.
2656 * Note that when @worker is non-NULL, @target may be modified
2657 * underneath us, so we can't reliably determine pwq from @target.
2660 * spin_lock_irq(pool->lock).
2662 static void insert_wq_barrier(struct pool_workqueue *pwq,
2663 struct wq_barrier *barr,
2664 struct work_struct *target, struct worker *worker)
2666 struct list_head *head;
2667 unsigned int linked = 0;
2670 * debugobject calls are safe here even with pool->lock locked
2671 * as we know for sure that this will not trigger any of the
2672 * checks and call back into the fixup functions where we
2675 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2676 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2678 init_completion_map(&barr->done, &target->lockdep_map);
2680 barr->task = current;
2683 * If @target is currently being executed, schedule the
2684 * barrier to the worker; otherwise, put it after @target.
2687 head = worker->scheduled.next;
2689 unsigned long *bits = work_data_bits(target);
2691 head = target->entry.next;
2692 /* there can already be other linked works, inherit and set */
2693 linked = *bits & WORK_STRUCT_LINKED;
2694 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2697 debug_work_activate(&barr->work);
2698 insert_work(pwq, &barr->work, head,
2699 work_color_to_flags(WORK_NO_COLOR) | linked);
2703 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2704 * @wq: workqueue being flushed
2705 * @flush_color: new flush color, < 0 for no-op
2706 * @work_color: new work color, < 0 for no-op
2708 * Prepare pwqs for workqueue flushing.
2710 * If @flush_color is non-negative, flush_color on all pwqs should be
2711 * -1. If no pwq has in-flight commands at the specified color, all
2712 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2713 * has in flight commands, its pwq->flush_color is set to
2714 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2715 * wakeup logic is armed and %true is returned.
2717 * The caller should have initialized @wq->first_flusher prior to
2718 * calling this function with non-negative @flush_color. If
2719 * @flush_color is negative, no flush color update is done and %false
2722 * If @work_color is non-negative, all pwqs should have the same
2723 * work_color which is previous to @work_color and all will be
2724 * advanced to @work_color.
2727 * mutex_lock(wq->mutex).
2730 * %true if @flush_color >= 0 and there's something to flush. %false
2733 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2734 int flush_color, int work_color)
2737 struct pool_workqueue *pwq;
2739 if (flush_color >= 0) {
2740 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2741 atomic_set(&wq->nr_pwqs_to_flush, 1);
2744 for_each_pwq(pwq, wq) {
2745 struct worker_pool *pool = pwq->pool;
2747 spin_lock_irq(&pool->lock);
2749 if (flush_color >= 0) {
2750 WARN_ON_ONCE(pwq->flush_color != -1);
2752 if (pwq->nr_in_flight[flush_color]) {
2753 pwq->flush_color = flush_color;
2754 atomic_inc(&wq->nr_pwqs_to_flush);
2759 if (work_color >= 0) {
2760 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2761 pwq->work_color = work_color;
2764 spin_unlock_irq(&pool->lock);
2767 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2768 complete(&wq->first_flusher->done);
2774 * flush_workqueue - ensure that any scheduled work has run to completion.
2775 * @wq: workqueue to flush
2777 * This function sleeps until all work items which were queued on entry
2778 * have finished execution, but it is not livelocked by new incoming ones.
2780 void flush_workqueue(struct workqueue_struct *wq)
2782 struct wq_flusher this_flusher = {
2783 .list = LIST_HEAD_INIT(this_flusher.list),
2785 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2789 if (WARN_ON(!wq_online))
2792 lock_map_acquire(&wq->lockdep_map);
2793 lock_map_release(&wq->lockdep_map);
2795 mutex_lock(&wq->mutex);
2798 * Start-to-wait phase
2800 next_color = work_next_color(wq->work_color);
2802 if (next_color != wq->flush_color) {
2804 * Color space is not full. The current work_color
2805 * becomes our flush_color and work_color is advanced
2808 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2809 this_flusher.flush_color = wq->work_color;
2810 wq->work_color = next_color;
2812 if (!wq->first_flusher) {
2813 /* no flush in progress, become the first flusher */
2814 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2816 wq->first_flusher = &this_flusher;
2818 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2820 /* nothing to flush, done */
2821 wq->flush_color = next_color;
2822 wq->first_flusher = NULL;
2827 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2828 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2829 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2833 * Oops, color space is full, wait on overflow queue.
2834 * The next flush completion will assign us
2835 * flush_color and transfer to flusher_queue.
2837 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2840 check_flush_dependency(wq, NULL);
2842 mutex_unlock(&wq->mutex);
2844 wait_for_completion(&this_flusher.done);
2847 * Wake-up-and-cascade phase
2849 * First flushers are responsible for cascading flushes and
2850 * handling overflow. Non-first flushers can simply return.
2852 if (wq->first_flusher != &this_flusher)
2855 mutex_lock(&wq->mutex);
2857 /* we might have raced, check again with mutex held */
2858 if (wq->first_flusher != &this_flusher)
2861 wq->first_flusher = NULL;
2863 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2864 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2867 struct wq_flusher *next, *tmp;
2869 /* complete all the flushers sharing the current flush color */
2870 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2871 if (next->flush_color != wq->flush_color)
2873 list_del_init(&next->list);
2874 complete(&next->done);
2877 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2878 wq->flush_color != work_next_color(wq->work_color));
2880 /* this flush_color is finished, advance by one */
2881 wq->flush_color = work_next_color(wq->flush_color);
2883 /* one color has been freed, handle overflow queue */
2884 if (!list_empty(&wq->flusher_overflow)) {
2886 * Assign the same color to all overflowed
2887 * flushers, advance work_color and append to
2888 * flusher_queue. This is the start-to-wait
2889 * phase for these overflowed flushers.
2891 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2892 tmp->flush_color = wq->work_color;
2894 wq->work_color = work_next_color(wq->work_color);
2896 list_splice_tail_init(&wq->flusher_overflow,
2897 &wq->flusher_queue);
2898 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2901 if (list_empty(&wq->flusher_queue)) {
2902 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2907 * Need to flush more colors. Make the next flusher
2908 * the new first flusher and arm pwqs.
2910 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2911 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2913 list_del_init(&next->list);
2914 wq->first_flusher = next;
2916 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2920 * Meh... this color is already done, clear first
2921 * flusher and repeat cascading.
2923 wq->first_flusher = NULL;
2927 mutex_unlock(&wq->mutex);
2929 EXPORT_SYMBOL(flush_workqueue);
2932 * drain_workqueue - drain a workqueue
2933 * @wq: workqueue to drain
2935 * Wait until the workqueue becomes empty. While draining is in progress,
2936 * only chain queueing is allowed. IOW, only currently pending or running
2937 * work items on @wq can queue further work items on it. @wq is flushed
2938 * repeatedly until it becomes empty. The number of flushing is determined
2939 * by the depth of chaining and should be relatively short. Whine if it
2942 void drain_workqueue(struct workqueue_struct *wq)
2944 unsigned int flush_cnt = 0;
2945 struct pool_workqueue *pwq;
2948 * __queue_work() needs to test whether there are drainers, is much
2949 * hotter than drain_workqueue() and already looks at @wq->flags.
2950 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2952 mutex_lock(&wq->mutex);
2953 if (!wq->nr_drainers++)
2954 wq->flags |= __WQ_DRAINING;
2955 mutex_unlock(&wq->mutex);
2957 flush_workqueue(wq);
2959 mutex_lock(&wq->mutex);
2961 for_each_pwq(pwq, wq) {
2964 spin_lock_irq(&pwq->pool->lock);
2965 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2966 spin_unlock_irq(&pwq->pool->lock);
2971 if (++flush_cnt == 10 ||
2972 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2973 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2974 wq->name, flush_cnt);
2976 mutex_unlock(&wq->mutex);
2980 if (!--wq->nr_drainers)
2981 wq->flags &= ~__WQ_DRAINING;
2982 mutex_unlock(&wq->mutex);
2984 EXPORT_SYMBOL_GPL(drain_workqueue);
2986 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2989 struct worker *worker = NULL;
2990 struct worker_pool *pool;
2991 struct pool_workqueue *pwq;
2996 pool = get_work_pool(work);
3002 spin_lock_irq(&pool->lock);
3003 /* see the comment in try_to_grab_pending() with the same code */
3004 pwq = get_work_pwq(work);
3006 if (unlikely(pwq->pool != pool))
3009 worker = find_worker_executing_work(pool, work);
3012 pwq = worker->current_pwq;
3015 check_flush_dependency(pwq->wq, work);
3017 insert_wq_barrier(pwq, barr, work, worker);
3018 spin_unlock_irq(&pool->lock);
3021 * Force a lock recursion deadlock when using flush_work() inside a
3022 * single-threaded or rescuer equipped workqueue.
3024 * For single threaded workqueues the deadlock happens when the work
3025 * is after the work issuing the flush_work(). For rescuer equipped
3026 * workqueues the deadlock happens when the rescuer stalls, blocking
3030 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3031 lock_map_acquire(&pwq->wq->lockdep_map);
3032 lock_map_release(&pwq->wq->lockdep_map);
3037 spin_unlock_irq(&pool->lock);
3042 static bool __flush_work(struct work_struct *work, bool from_cancel)
3044 struct wq_barrier barr;
3046 if (WARN_ON(!wq_online))
3049 if (WARN_ON(!work->func))
3053 lock_map_acquire(&work->lockdep_map);
3054 lock_map_release(&work->lockdep_map);
3057 if (start_flush_work(work, &barr, from_cancel)) {
3058 wait_for_completion(&barr.done);
3059 destroy_work_on_stack(&barr.work);
3067 * flush_work - wait for a work to finish executing the last queueing instance
3068 * @work: the work to flush
3070 * Wait until @work has finished execution. @work is guaranteed to be idle
3071 * on return if it hasn't been requeued since flush started.
3074 * %true if flush_work() waited for the work to finish execution,
3075 * %false if it was already idle.
3077 bool flush_work(struct work_struct *work)
3079 return __flush_work(work, false);
3081 EXPORT_SYMBOL_GPL(flush_work);
3084 wait_queue_entry_t wait;
3085 struct work_struct *work;
3088 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3090 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3092 if (cwait->work != key)
3094 return autoremove_wake_function(wait, mode, sync, key);
3097 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3099 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3100 unsigned long flags;
3104 ret = try_to_grab_pending(work, is_dwork, &flags);
3106 * If someone else is already canceling, wait for it to
3107 * finish. flush_work() doesn't work for PREEMPT_NONE
3108 * because we may get scheduled between @work's completion
3109 * and the other canceling task resuming and clearing
3110 * CANCELING - flush_work() will return false immediately
3111 * as @work is no longer busy, try_to_grab_pending() will
3112 * return -ENOENT as @work is still being canceled and the
3113 * other canceling task won't be able to clear CANCELING as
3114 * we're hogging the CPU.
3116 * Let's wait for completion using a waitqueue. As this
3117 * may lead to the thundering herd problem, use a custom
3118 * wake function which matches @work along with exclusive
3121 if (unlikely(ret == -ENOENT)) {
3122 struct cwt_wait cwait;
3124 init_wait(&cwait.wait);
3125 cwait.wait.func = cwt_wakefn;
3128 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3129 TASK_UNINTERRUPTIBLE);
3130 if (work_is_canceling(work))
3132 finish_wait(&cancel_waitq, &cwait.wait);
3134 } while (unlikely(ret < 0));
3136 /* tell other tasks trying to grab @work to back off */
3137 mark_work_canceling(work);
3138 local_irq_restore(flags);
3141 * This allows canceling during early boot. We know that @work
3145 __flush_work(work, true);
3147 clear_work_data(work);
3150 * Paired with prepare_to_wait() above so that either
3151 * waitqueue_active() is visible here or !work_is_canceling() is
3155 if (waitqueue_active(&cancel_waitq))
3156 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3162 * cancel_work_sync - cancel a work and wait for it to finish
3163 * @work: the work to cancel
3165 * Cancel @work and wait for its execution to finish. This function
3166 * can be used even if the work re-queues itself or migrates to
3167 * another workqueue. On return from this function, @work is
3168 * guaranteed to be not pending or executing on any CPU.
3170 * cancel_work_sync(&delayed_work->work) must not be used for
3171 * delayed_work's. Use cancel_delayed_work_sync() instead.
3173 * The caller must ensure that the workqueue on which @work was last
3174 * queued can't be destroyed before this function returns.
3177 * %true if @work was pending, %false otherwise.
3179 bool cancel_work_sync(struct work_struct *work)
3181 return __cancel_work_timer(work, false);
3183 EXPORT_SYMBOL_GPL(cancel_work_sync);
3186 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3187 * @dwork: the delayed work to flush
3189 * Delayed timer is cancelled and the pending work is queued for
3190 * immediate execution. Like flush_work(), this function only
3191 * considers the last queueing instance of @dwork.
3194 * %true if flush_work() waited for the work to finish execution,
3195 * %false if it was already idle.
3197 bool flush_delayed_work(struct delayed_work *dwork)
3199 local_irq_disable();
3200 if (del_timer_sync(&dwork->timer))
3201 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3203 return flush_work(&dwork->work);
3205 EXPORT_SYMBOL(flush_delayed_work);
3208 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3209 * @rwork: the rcu work to flush
3212 * %true if flush_rcu_work() waited for the work to finish execution,
3213 * %false if it was already idle.
3215 bool flush_rcu_work(struct rcu_work *rwork)
3217 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3219 flush_work(&rwork->work);
3222 return flush_work(&rwork->work);
3225 EXPORT_SYMBOL(flush_rcu_work);
3227 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3229 unsigned long flags;
3233 ret = try_to_grab_pending(work, is_dwork, &flags);
3234 } while (unlikely(ret == -EAGAIN));
3236 if (unlikely(ret < 0))
3239 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3240 local_irq_restore(flags);
3245 * cancel_delayed_work - cancel a delayed work
3246 * @dwork: delayed_work to cancel
3248 * Kill off a pending delayed_work.
3250 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3254 * The work callback function may still be running on return, unless
3255 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3256 * use cancel_delayed_work_sync() to wait on it.
3258 * This function is safe to call from any context including IRQ handler.
3260 bool cancel_delayed_work(struct delayed_work *dwork)
3262 return __cancel_work(&dwork->work, true);
3264 EXPORT_SYMBOL(cancel_delayed_work);
3267 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3268 * @dwork: the delayed work cancel
3270 * This is cancel_work_sync() for delayed works.
3273 * %true if @dwork was pending, %false otherwise.
3275 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3277 return __cancel_work_timer(&dwork->work, true);
3279 EXPORT_SYMBOL(cancel_delayed_work_sync);
3282 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3283 * @func: the function to call
3285 * schedule_on_each_cpu() executes @func on each online CPU using the
3286 * system workqueue and blocks until all CPUs have completed.
3287 * schedule_on_each_cpu() is very slow.
3290 * 0 on success, -errno on failure.
3292 int schedule_on_each_cpu(work_func_t func)
3295 struct work_struct __percpu *works;
3297 works = alloc_percpu(struct work_struct);
3303 for_each_online_cpu(cpu) {
3304 struct work_struct *work = per_cpu_ptr(works, cpu);
3306 INIT_WORK(work, func);
3307 schedule_work_on(cpu, work);
3310 for_each_online_cpu(cpu)
3311 flush_work(per_cpu_ptr(works, cpu));
3319 * execute_in_process_context - reliably execute the routine with user context
3320 * @fn: the function to execute
3321 * @ew: guaranteed storage for the execute work structure (must
3322 * be available when the work executes)
3324 * Executes the function immediately if process context is available,
3325 * otherwise schedules the function for delayed execution.
3327 * Return: 0 - function was executed
3328 * 1 - function was scheduled for execution
3330 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3332 if (!in_interrupt()) {
3337 INIT_WORK(&ew->work, fn);
3338 schedule_work(&ew->work);
3342 EXPORT_SYMBOL_GPL(execute_in_process_context);
3345 * free_workqueue_attrs - free a workqueue_attrs
3346 * @attrs: workqueue_attrs to free
3348 * Undo alloc_workqueue_attrs().
3350 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3353 free_cpumask_var(attrs->cpumask);
3359 * alloc_workqueue_attrs - allocate a workqueue_attrs
3361 * Allocate a new workqueue_attrs, initialize with default settings and
3364 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3366 struct workqueue_attrs *alloc_workqueue_attrs(void)
3368 struct workqueue_attrs *attrs;
3370 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3373 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3376 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3379 free_workqueue_attrs(attrs);
3383 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3384 const struct workqueue_attrs *from)
3386 to->nice = from->nice;
3387 cpumask_copy(to->cpumask, from->cpumask);
3389 * Unlike hash and equality test, this function doesn't ignore
3390 * ->no_numa as it is used for both pool and wq attrs. Instead,
3391 * get_unbound_pool() explicitly clears ->no_numa after copying.
3393 to->no_numa = from->no_numa;
3396 /* hash value of the content of @attr */
3397 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3401 hash = jhash_1word(attrs->nice, hash);
3402 hash = jhash(cpumask_bits(attrs->cpumask),
3403 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3407 /* content equality test */
3408 static bool wqattrs_equal(const struct workqueue_attrs *a,
3409 const struct workqueue_attrs *b)
3411 if (a->nice != b->nice)
3413 if (!cpumask_equal(a->cpumask, b->cpumask))
3419 * init_worker_pool - initialize a newly zalloc'd worker_pool
3420 * @pool: worker_pool to initialize
3422 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3424 * Return: 0 on success, -errno on failure. Even on failure, all fields
3425 * inside @pool proper are initialized and put_unbound_pool() can be called
3426 * on @pool safely to release it.
3428 static int init_worker_pool(struct worker_pool *pool)
3430 spin_lock_init(&pool->lock);
3433 pool->node = NUMA_NO_NODE;
3434 pool->flags |= POOL_DISASSOCIATED;
3435 pool->watchdog_ts = jiffies;
3436 INIT_LIST_HEAD(&pool->worklist);
3437 INIT_LIST_HEAD(&pool->idle_list);
3438 hash_init(pool->busy_hash);
3440 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3442 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3444 INIT_LIST_HEAD(&pool->workers);
3446 ida_init(&pool->worker_ida);
3447 INIT_HLIST_NODE(&pool->hash_node);
3450 /* shouldn't fail above this point */
3451 pool->attrs = alloc_workqueue_attrs();
3457 #ifdef CONFIG_LOCKDEP
3458 static void wq_init_lockdep(struct workqueue_struct *wq)
3462 lockdep_register_key(&wq->key);
3463 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3465 lock_name = wq->name;
3467 wq->lock_name = lock_name;
3468 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3471 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3473 lockdep_unregister_key(&wq->key);
3476 static void wq_free_lockdep(struct workqueue_struct *wq)
3478 if (wq->lock_name != wq->name)
3479 kfree(wq->lock_name);
3482 static void wq_init_lockdep(struct workqueue_struct *wq)
3486 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3490 static void wq_free_lockdep(struct workqueue_struct *wq)
3495 static void rcu_free_wq(struct rcu_head *rcu)
3497 struct workqueue_struct *wq =
3498 container_of(rcu, struct workqueue_struct, rcu);
3500 wq_free_lockdep(wq);
3502 if (!(wq->flags & WQ_UNBOUND))
3503 free_percpu(wq->cpu_pwqs);
3505 free_workqueue_attrs(wq->unbound_attrs);
3511 static void rcu_free_pool(struct rcu_head *rcu)
3513 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3515 ida_destroy(&pool->worker_ida);
3516 free_workqueue_attrs(pool->attrs);
3521 * put_unbound_pool - put a worker_pool
3522 * @pool: worker_pool to put
3524 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3525 * safe manner. get_unbound_pool() calls this function on its failure path
3526 * and this function should be able to release pools which went through,
3527 * successfully or not, init_worker_pool().
3529 * Should be called with wq_pool_mutex held.
3531 static void put_unbound_pool(struct worker_pool *pool)
3533 DECLARE_COMPLETION_ONSTACK(detach_completion);
3534 struct worker *worker;
3536 lockdep_assert_held(&wq_pool_mutex);
3542 if (WARN_ON(!(pool->cpu < 0)) ||
3543 WARN_ON(!list_empty(&pool->worklist)))
3546 /* release id and unhash */
3548 idr_remove(&worker_pool_idr, pool->id);
3549 hash_del(&pool->hash_node);
3552 * Become the manager and destroy all workers. This prevents
3553 * @pool's workers from blocking on attach_mutex. We're the last
3554 * manager and @pool gets freed with the flag set.
3556 spin_lock_irq(&pool->lock);
3557 wait_event_lock_irq(wq_manager_wait,
3558 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3559 pool->flags |= POOL_MANAGER_ACTIVE;
3561 while ((worker = first_idle_worker(pool)))
3562 destroy_worker(worker);
3563 WARN_ON(pool->nr_workers || pool->nr_idle);
3564 spin_unlock_irq(&pool->lock);
3566 mutex_lock(&wq_pool_attach_mutex);
3567 if (!list_empty(&pool->workers))
3568 pool->detach_completion = &detach_completion;
3569 mutex_unlock(&wq_pool_attach_mutex);
3571 if (pool->detach_completion)
3572 wait_for_completion(pool->detach_completion);
3574 /* shut down the timers */
3575 del_timer_sync(&pool->idle_timer);
3576 del_timer_sync(&pool->mayday_timer);
3578 /* RCU protected to allow dereferences from get_work_pool() */
3579 call_rcu(&pool->rcu, rcu_free_pool);
3583 * get_unbound_pool - get a worker_pool with the specified attributes
3584 * @attrs: the attributes of the worker_pool to get
3586 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3587 * reference count and return it. If there already is a matching
3588 * worker_pool, it will be used; otherwise, this function attempts to
3591 * Should be called with wq_pool_mutex held.
3593 * Return: On success, a worker_pool with the same attributes as @attrs.
3594 * On failure, %NULL.
3596 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3598 u32 hash = wqattrs_hash(attrs);
3599 struct worker_pool *pool;
3601 int target_node = NUMA_NO_NODE;
3603 lockdep_assert_held(&wq_pool_mutex);
3605 /* do we already have a matching pool? */
3606 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3607 if (wqattrs_equal(pool->attrs, attrs)) {
3613 /* if cpumask is contained inside a NUMA node, we belong to that node */
3614 if (wq_numa_enabled) {
3615 for_each_node(node) {
3616 if (cpumask_subset(attrs->cpumask,
3617 wq_numa_possible_cpumask[node])) {
3624 /* nope, create a new one */
3625 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3626 if (!pool || init_worker_pool(pool) < 0)
3629 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3630 copy_workqueue_attrs(pool->attrs, attrs);
3631 pool->node = target_node;
3634 * no_numa isn't a worker_pool attribute, always clear it. See
3635 * 'struct workqueue_attrs' comments for detail.
3637 pool->attrs->no_numa = false;
3639 if (worker_pool_assign_id(pool) < 0)
3642 /* create and start the initial worker */
3643 if (wq_online && !create_worker(pool))
3647 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3652 put_unbound_pool(pool);
3656 static void rcu_free_pwq(struct rcu_head *rcu)
3658 kmem_cache_free(pwq_cache,
3659 container_of(rcu, struct pool_workqueue, rcu));
3663 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3664 * and needs to be destroyed.
3666 static void pwq_unbound_release_workfn(struct work_struct *work)
3668 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3669 unbound_release_work);
3670 struct workqueue_struct *wq = pwq->wq;
3671 struct worker_pool *pool = pwq->pool;
3672 bool is_last = false;
3675 * when @pwq is not linked, it doesn't hold any reference to the
3676 * @wq, and @wq is invalid to access.
3678 if (!list_empty(&pwq->pwqs_node)) {
3679 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3682 mutex_lock(&wq->mutex);
3683 list_del_rcu(&pwq->pwqs_node);
3684 is_last = list_empty(&wq->pwqs);
3685 mutex_unlock(&wq->mutex);
3688 mutex_lock(&wq_pool_mutex);
3689 put_unbound_pool(pool);
3690 mutex_unlock(&wq_pool_mutex);
3692 call_rcu(&pwq->rcu, rcu_free_pwq);
3695 * If we're the last pwq going away, @wq is already dead and no one
3696 * is gonna access it anymore. Schedule RCU free.
3699 wq_unregister_lockdep(wq);
3700 call_rcu(&wq->rcu, rcu_free_wq);
3705 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3706 * @pwq: target pool_workqueue
3708 * If @pwq isn't freezing, set @pwq->max_active to the associated
3709 * workqueue's saved_max_active and activate delayed work items
3710 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3712 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3714 struct workqueue_struct *wq = pwq->wq;
3715 bool freezable = wq->flags & WQ_FREEZABLE;
3716 unsigned long flags;
3718 /* for @wq->saved_max_active */
3719 lockdep_assert_held(&wq->mutex);
3721 /* fast exit for non-freezable wqs */
3722 if (!freezable && pwq->max_active == wq->saved_max_active)
3725 /* this function can be called during early boot w/ irq disabled */
3726 spin_lock_irqsave(&pwq->pool->lock, flags);
3729 * During [un]freezing, the caller is responsible for ensuring that
3730 * this function is called at least once after @workqueue_freezing
3731 * is updated and visible.
3733 if (!freezable || !workqueue_freezing) {
3736 pwq->max_active = wq->saved_max_active;
3738 while (!list_empty(&pwq->delayed_works) &&
3739 pwq->nr_active < pwq->max_active) {
3740 pwq_activate_first_delayed(pwq);
3745 * Need to kick a worker after thawed or an unbound wq's
3746 * max_active is bumped. In realtime scenarios, always kicking a
3747 * worker will cause interference on the isolated cpu cores, so
3748 * let's kick iff work items were activated.
3751 wake_up_worker(pwq->pool);
3753 pwq->max_active = 0;
3756 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3759 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3760 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3761 struct worker_pool *pool)
3763 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3765 memset(pwq, 0, sizeof(*pwq));
3769 pwq->flush_color = -1;
3771 INIT_LIST_HEAD(&pwq->delayed_works);
3772 INIT_LIST_HEAD(&pwq->pwqs_node);
3773 INIT_LIST_HEAD(&pwq->mayday_node);
3774 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3777 /* sync @pwq with the current state of its associated wq and link it */
3778 static void link_pwq(struct pool_workqueue *pwq)
3780 struct workqueue_struct *wq = pwq->wq;
3782 lockdep_assert_held(&wq->mutex);
3784 /* may be called multiple times, ignore if already linked */
3785 if (!list_empty(&pwq->pwqs_node))
3788 /* set the matching work_color */
3789 pwq->work_color = wq->work_color;
3791 /* sync max_active to the current setting */
3792 pwq_adjust_max_active(pwq);
3795 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3798 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3799 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3800 const struct workqueue_attrs *attrs)
3802 struct worker_pool *pool;
3803 struct pool_workqueue *pwq;
3805 lockdep_assert_held(&wq_pool_mutex);
3807 pool = get_unbound_pool(attrs);
3811 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3813 put_unbound_pool(pool);
3817 init_pwq(pwq, wq, pool);
3822 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3823 * @attrs: the wq_attrs of the default pwq of the target workqueue
3824 * @node: the target NUMA node
3825 * @cpu_going_down: if >= 0, the CPU to consider as offline
3826 * @cpumask: outarg, the resulting cpumask
3828 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3829 * @cpu_going_down is >= 0, that cpu is considered offline during
3830 * calculation. The result is stored in @cpumask.
3832 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3833 * enabled and @node has online CPUs requested by @attrs, the returned
3834 * cpumask is the intersection of the possible CPUs of @node and
3837 * The caller is responsible for ensuring that the cpumask of @node stays
3840 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3843 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3844 int cpu_going_down, cpumask_t *cpumask)
3846 if (!wq_numa_enabled || attrs->no_numa)
3849 /* does @node have any online CPUs @attrs wants? */
3850 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3851 if (cpu_going_down >= 0)
3852 cpumask_clear_cpu(cpu_going_down, cpumask);
3854 if (cpumask_empty(cpumask))
3857 /* yeap, return possible CPUs in @node that @attrs wants */
3858 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3860 if (cpumask_empty(cpumask)) {
3861 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3862 "possible intersect\n");
3866 return !cpumask_equal(cpumask, attrs->cpumask);
3869 cpumask_copy(cpumask, attrs->cpumask);
3873 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3874 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3876 struct pool_workqueue *pwq)
3878 struct pool_workqueue *old_pwq;
3880 lockdep_assert_held(&wq_pool_mutex);
3881 lockdep_assert_held(&wq->mutex);
3883 /* link_pwq() can handle duplicate calls */
3886 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3887 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3891 /* context to store the prepared attrs & pwqs before applying */
3892 struct apply_wqattrs_ctx {
3893 struct workqueue_struct *wq; /* target workqueue */
3894 struct workqueue_attrs *attrs; /* attrs to apply */
3895 struct list_head list; /* queued for batching commit */
3896 struct pool_workqueue *dfl_pwq;
3897 struct pool_workqueue *pwq_tbl[];
3900 /* free the resources after success or abort */
3901 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3907 put_pwq_unlocked(ctx->pwq_tbl[node]);
3908 put_pwq_unlocked(ctx->dfl_pwq);
3910 free_workqueue_attrs(ctx->attrs);
3916 /* allocate the attrs and pwqs for later installation */
3917 static struct apply_wqattrs_ctx *
3918 apply_wqattrs_prepare(struct workqueue_struct *wq,
3919 const struct workqueue_attrs *attrs)
3921 struct apply_wqattrs_ctx *ctx;
3922 struct workqueue_attrs *new_attrs, *tmp_attrs;
3925 lockdep_assert_held(&wq_pool_mutex);
3927 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3929 new_attrs = alloc_workqueue_attrs();
3930 tmp_attrs = alloc_workqueue_attrs();
3931 if (!ctx || !new_attrs || !tmp_attrs)
3935 * Calculate the attrs of the default pwq.
3936 * If the user configured cpumask doesn't overlap with the
3937 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3939 copy_workqueue_attrs(new_attrs, attrs);
3940 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3941 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3942 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3945 * We may create multiple pwqs with differing cpumasks. Make a
3946 * copy of @new_attrs which will be modified and used to obtain
3949 copy_workqueue_attrs(tmp_attrs, new_attrs);
3952 * If something goes wrong during CPU up/down, we'll fall back to
3953 * the default pwq covering whole @attrs->cpumask. Always create
3954 * it even if we don't use it immediately.
3956 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3960 for_each_node(node) {
3961 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3962 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3963 if (!ctx->pwq_tbl[node])
3966 ctx->dfl_pwq->refcnt++;
3967 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3971 /* save the user configured attrs and sanitize it. */
3972 copy_workqueue_attrs(new_attrs, attrs);
3973 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3974 ctx->attrs = new_attrs;
3977 free_workqueue_attrs(tmp_attrs);
3981 free_workqueue_attrs(tmp_attrs);
3982 free_workqueue_attrs(new_attrs);
3983 apply_wqattrs_cleanup(ctx);
3987 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3988 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3992 /* all pwqs have been created successfully, let's install'em */
3993 mutex_lock(&ctx->wq->mutex);
3995 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3997 /* save the previous pwq and install the new one */
3999 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
4000 ctx->pwq_tbl[node]);
4002 /* @dfl_pwq might not have been used, ensure it's linked */
4003 link_pwq(ctx->dfl_pwq);
4004 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4006 mutex_unlock(&ctx->wq->mutex);
4009 static void apply_wqattrs_lock(void)
4011 /* CPUs should stay stable across pwq creations and installations */
4013 mutex_lock(&wq_pool_mutex);
4016 static void apply_wqattrs_unlock(void)
4018 mutex_unlock(&wq_pool_mutex);
4022 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4023 const struct workqueue_attrs *attrs)
4025 struct apply_wqattrs_ctx *ctx;
4027 /* only unbound workqueues can change attributes */
4028 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4031 /* creating multiple pwqs breaks ordering guarantee */
4032 if (!list_empty(&wq->pwqs)) {
4033 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4036 wq->flags &= ~__WQ_ORDERED;
4039 ctx = apply_wqattrs_prepare(wq, attrs);
4043 /* the ctx has been prepared successfully, let's commit it */
4044 apply_wqattrs_commit(ctx);
4045 apply_wqattrs_cleanup(ctx);
4051 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4052 * @wq: the target workqueue
4053 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4055 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4056 * machines, this function maps a separate pwq to each NUMA node with
4057 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4058 * NUMA node it was issued on. Older pwqs are released as in-flight work
4059 * items finish. Note that a work item which repeatedly requeues itself
4060 * back-to-back will stay on its current pwq.
4062 * Performs GFP_KERNEL allocations.
4064 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4066 * Return: 0 on success and -errno on failure.
4068 int apply_workqueue_attrs(struct workqueue_struct *wq,
4069 const struct workqueue_attrs *attrs)
4073 lockdep_assert_cpus_held();
4075 mutex_lock(&wq_pool_mutex);
4076 ret = apply_workqueue_attrs_locked(wq, attrs);
4077 mutex_unlock(&wq_pool_mutex);
4083 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4084 * @wq: the target workqueue
4085 * @cpu: the CPU coming up or going down
4086 * @online: whether @cpu is coming up or going down
4088 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4089 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4092 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4093 * falls back to @wq->dfl_pwq which may not be optimal but is always
4096 * Note that when the last allowed CPU of a NUMA node goes offline for a
4097 * workqueue with a cpumask spanning multiple nodes, the workers which were
4098 * already executing the work items for the workqueue will lose their CPU
4099 * affinity and may execute on any CPU. This is similar to how per-cpu
4100 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4101 * affinity, it's the user's responsibility to flush the work item from
4104 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4107 int node = cpu_to_node(cpu);
4108 int cpu_off = online ? -1 : cpu;
4109 struct pool_workqueue *old_pwq = NULL, *pwq;
4110 struct workqueue_attrs *target_attrs;
4113 lockdep_assert_held(&wq_pool_mutex);
4115 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4116 wq->unbound_attrs->no_numa)
4120 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4121 * Let's use a preallocated one. The following buf is protected by
4122 * CPU hotplug exclusion.
4124 target_attrs = wq_update_unbound_numa_attrs_buf;
4125 cpumask = target_attrs->cpumask;
4127 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4128 pwq = unbound_pwq_by_node(wq, node);
4131 * Let's determine what needs to be done. If the target cpumask is
4132 * different from the default pwq's, we need to compare it to @pwq's
4133 * and create a new one if they don't match. If the target cpumask
4134 * equals the default pwq's, the default pwq should be used.
4136 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4137 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4143 /* create a new pwq */
4144 pwq = alloc_unbound_pwq(wq, target_attrs);
4146 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4151 /* Install the new pwq. */
4152 mutex_lock(&wq->mutex);
4153 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4157 mutex_lock(&wq->mutex);
4158 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4159 get_pwq(wq->dfl_pwq);
4160 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4161 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4163 mutex_unlock(&wq->mutex);
4164 put_pwq_unlocked(old_pwq);
4167 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4169 bool highpri = wq->flags & WQ_HIGHPRI;
4172 if (!(wq->flags & WQ_UNBOUND)) {
4173 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4177 for_each_possible_cpu(cpu) {
4178 struct pool_workqueue *pwq =
4179 per_cpu_ptr(wq->cpu_pwqs, cpu);
4180 struct worker_pool *cpu_pools =
4181 per_cpu(cpu_worker_pools, cpu);
4183 init_pwq(pwq, wq, &cpu_pools[highpri]);
4185 mutex_lock(&wq->mutex);
4187 mutex_unlock(&wq->mutex);
4193 if (wq->flags & __WQ_ORDERED) {
4194 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4195 /* there should only be single pwq for ordering guarantee */
4196 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4197 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4198 "ordering guarantee broken for workqueue %s\n", wq->name);
4200 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4207 static int wq_clamp_max_active(int max_active, unsigned int flags,
4210 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4212 if (max_active < 1 || max_active > lim)
4213 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4214 max_active, name, 1, lim);
4216 return clamp_val(max_active, 1, lim);
4220 * Workqueues which may be used during memory reclaim should have a rescuer
4221 * to guarantee forward progress.
4223 static int init_rescuer(struct workqueue_struct *wq)
4225 struct worker *rescuer;
4228 if (!(wq->flags & WQ_MEM_RECLAIM))
4231 rescuer = alloc_worker(NUMA_NO_NODE);
4235 rescuer->rescue_wq = wq;
4236 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4237 ret = PTR_ERR_OR_ZERO(rescuer->task);
4243 wq->rescuer = rescuer;
4244 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4245 wake_up_process(rescuer->task);
4251 struct workqueue_struct *alloc_workqueue(const char *fmt,
4253 int max_active, ...)
4255 size_t tbl_size = 0;
4257 struct workqueue_struct *wq;
4258 struct pool_workqueue *pwq;
4261 * Unbound && max_active == 1 used to imply ordered, which is no
4262 * longer the case on NUMA machines due to per-node pools. While
4263 * alloc_ordered_workqueue() is the right way to create an ordered
4264 * workqueue, keep the previous behavior to avoid subtle breakages
4267 if ((flags & WQ_UNBOUND) && max_active == 1)
4268 flags |= __WQ_ORDERED;
4270 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4271 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4272 flags |= WQ_UNBOUND;
4274 /* allocate wq and format name */
4275 if (flags & WQ_UNBOUND)
4276 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4278 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4282 if (flags & WQ_UNBOUND) {
4283 wq->unbound_attrs = alloc_workqueue_attrs();
4284 if (!wq->unbound_attrs)
4288 va_start(args, max_active);
4289 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4292 max_active = max_active ?: WQ_DFL_ACTIVE;
4293 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4297 wq->saved_max_active = max_active;
4298 mutex_init(&wq->mutex);
4299 atomic_set(&wq->nr_pwqs_to_flush, 0);
4300 INIT_LIST_HEAD(&wq->pwqs);
4301 INIT_LIST_HEAD(&wq->flusher_queue);
4302 INIT_LIST_HEAD(&wq->flusher_overflow);
4303 INIT_LIST_HEAD(&wq->maydays);
4305 wq_init_lockdep(wq);
4306 INIT_LIST_HEAD(&wq->list);
4308 if (alloc_and_link_pwqs(wq) < 0)
4309 goto err_unreg_lockdep;
4311 if (wq_online && init_rescuer(wq) < 0)
4314 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4318 * wq_pool_mutex protects global freeze state and workqueues list.
4319 * Grab it, adjust max_active and add the new @wq to workqueues
4322 mutex_lock(&wq_pool_mutex);
4324 mutex_lock(&wq->mutex);
4325 for_each_pwq(pwq, wq)
4326 pwq_adjust_max_active(pwq);
4327 mutex_unlock(&wq->mutex);
4329 list_add_tail_rcu(&wq->list, &workqueues);
4331 mutex_unlock(&wq_pool_mutex);
4336 wq_unregister_lockdep(wq);
4337 wq_free_lockdep(wq);
4339 free_workqueue_attrs(wq->unbound_attrs);
4343 destroy_workqueue(wq);
4346 EXPORT_SYMBOL_GPL(alloc_workqueue);
4349 * destroy_workqueue - safely terminate a workqueue
4350 * @wq: target workqueue
4352 * Safely destroy a workqueue. All work currently pending will be done first.
4354 void destroy_workqueue(struct workqueue_struct *wq)
4356 struct pool_workqueue *pwq;
4360 * Remove it from sysfs first so that sanity check failure doesn't
4361 * lead to sysfs name conflicts.
4363 workqueue_sysfs_unregister(wq);
4365 /* drain it before proceeding with destruction */
4366 drain_workqueue(wq);
4368 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4370 struct worker *rescuer = wq->rescuer;
4372 /* this prevents new queueing */
4373 spin_lock_irq(&wq_mayday_lock);
4375 spin_unlock_irq(&wq_mayday_lock);
4377 /* rescuer will empty maydays list before exiting */
4378 kthread_stop(rescuer->task);
4383 mutex_lock(&wq->mutex);
4384 for_each_pwq(pwq, wq) {
4387 for (i = 0; i < WORK_NR_COLORS; i++) {
4388 if (WARN_ON(pwq->nr_in_flight[i])) {
4389 mutex_unlock(&wq->mutex);
4390 show_workqueue_state();
4395 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4396 WARN_ON(pwq->nr_active) ||
4397 WARN_ON(!list_empty(&pwq->delayed_works))) {
4398 mutex_unlock(&wq->mutex);
4399 show_workqueue_state();
4403 mutex_unlock(&wq->mutex);
4406 * wq list is used to freeze wq, remove from list after
4407 * flushing is complete in case freeze races us.
4409 mutex_lock(&wq_pool_mutex);
4410 list_del_rcu(&wq->list);
4411 mutex_unlock(&wq_pool_mutex);
4413 if (!(wq->flags & WQ_UNBOUND)) {
4414 wq_unregister_lockdep(wq);
4416 * The base ref is never dropped on per-cpu pwqs. Directly
4417 * schedule RCU free.
4419 call_rcu(&wq->rcu, rcu_free_wq);
4422 * We're the sole accessor of @wq at this point. Directly
4423 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4424 * @wq will be freed when the last pwq is released.
4426 for_each_node(node) {
4427 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4428 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4429 put_pwq_unlocked(pwq);
4433 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4434 * put. Don't access it afterwards.
4438 put_pwq_unlocked(pwq);
4441 EXPORT_SYMBOL_GPL(destroy_workqueue);
4444 * workqueue_set_max_active - adjust max_active of a workqueue
4445 * @wq: target workqueue
4446 * @max_active: new max_active value.
4448 * Set max_active of @wq to @max_active.
4451 * Don't call from IRQ context.
4453 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4455 struct pool_workqueue *pwq;
4457 /* disallow meddling with max_active for ordered workqueues */
4458 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4461 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4463 mutex_lock(&wq->mutex);
4465 wq->flags &= ~__WQ_ORDERED;
4466 wq->saved_max_active = max_active;
4468 for_each_pwq(pwq, wq)
4469 pwq_adjust_max_active(pwq);
4471 mutex_unlock(&wq->mutex);
4473 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4476 * current_work - retrieve %current task's work struct
4478 * Determine if %current task is a workqueue worker and what it's working on.
4479 * Useful to find out the context that the %current task is running in.
4481 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4483 struct work_struct *current_work(void)
4485 struct worker *worker = current_wq_worker();
4487 return worker ? worker->current_work : NULL;
4489 EXPORT_SYMBOL(current_work);
4492 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4494 * Determine whether %current is a workqueue rescuer. Can be used from
4495 * work functions to determine whether it's being run off the rescuer task.
4497 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4499 bool current_is_workqueue_rescuer(void)
4501 struct worker *worker = current_wq_worker();
4503 return worker && worker->rescue_wq;
4507 * workqueue_congested - test whether a workqueue is congested
4508 * @cpu: CPU in question
4509 * @wq: target workqueue
4511 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4512 * no synchronization around this function and the test result is
4513 * unreliable and only useful as advisory hints or for debugging.
4515 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4516 * Note that both per-cpu and unbound workqueues may be associated with
4517 * multiple pool_workqueues which have separate congested states. A
4518 * workqueue being congested on one CPU doesn't mean the workqueue is also
4519 * contested on other CPUs / NUMA nodes.
4522 * %true if congested, %false otherwise.
4524 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4526 struct pool_workqueue *pwq;
4532 if (cpu == WORK_CPU_UNBOUND)
4533 cpu = smp_processor_id();
4535 if (!(wq->flags & WQ_UNBOUND))
4536 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4538 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4540 ret = !list_empty(&pwq->delayed_works);
4546 EXPORT_SYMBOL_GPL(workqueue_congested);
4549 * work_busy - test whether a work is currently pending or running
4550 * @work: the work to be tested
4552 * Test whether @work is currently pending or running. There is no
4553 * synchronization around this function and the test result is
4554 * unreliable and only useful as advisory hints or for debugging.
4557 * OR'd bitmask of WORK_BUSY_* bits.
4559 unsigned int work_busy(struct work_struct *work)
4561 struct worker_pool *pool;
4562 unsigned long flags;
4563 unsigned int ret = 0;
4565 if (work_pending(work))
4566 ret |= WORK_BUSY_PENDING;
4569 pool = get_work_pool(work);
4571 spin_lock_irqsave(&pool->lock, flags);
4572 if (find_worker_executing_work(pool, work))
4573 ret |= WORK_BUSY_RUNNING;
4574 spin_unlock_irqrestore(&pool->lock, flags);
4580 EXPORT_SYMBOL_GPL(work_busy);
4583 * set_worker_desc - set description for the current work item
4584 * @fmt: printf-style format string
4585 * @...: arguments for the format string
4587 * This function can be called by a running work function to describe what
4588 * the work item is about. If the worker task gets dumped, this
4589 * information will be printed out together to help debugging. The
4590 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4592 void set_worker_desc(const char *fmt, ...)
4594 struct worker *worker = current_wq_worker();
4598 va_start(args, fmt);
4599 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4603 EXPORT_SYMBOL_GPL(set_worker_desc);
4606 * print_worker_info - print out worker information and description
4607 * @log_lvl: the log level to use when printing
4608 * @task: target task
4610 * If @task is a worker and currently executing a work item, print out the
4611 * name of the workqueue being serviced and worker description set with
4612 * set_worker_desc() by the currently executing work item.
4614 * This function can be safely called on any task as long as the
4615 * task_struct itself is accessible. While safe, this function isn't
4616 * synchronized and may print out mixups or garbages of limited length.
4618 void print_worker_info(const char *log_lvl, struct task_struct *task)
4620 work_func_t *fn = NULL;
4621 char name[WQ_NAME_LEN] = { };
4622 char desc[WORKER_DESC_LEN] = { };
4623 struct pool_workqueue *pwq = NULL;
4624 struct workqueue_struct *wq = NULL;
4625 struct worker *worker;
4627 if (!(task->flags & PF_WQ_WORKER))
4631 * This function is called without any synchronization and @task
4632 * could be in any state. Be careful with dereferences.
4634 worker = kthread_probe_data(task);
4637 * Carefully copy the associated workqueue's workfn, name and desc.
4638 * Keep the original last '\0' in case the original is garbage.
4640 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4641 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4642 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4643 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4644 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4646 if (fn || name[0] || desc[0]) {
4647 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4648 if (strcmp(name, desc))
4649 pr_cont(" (%s)", desc);
4654 static void pr_cont_pool_info(struct worker_pool *pool)
4656 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4657 if (pool->node != NUMA_NO_NODE)
4658 pr_cont(" node=%d", pool->node);
4659 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4662 static void pr_cont_work(bool comma, struct work_struct *work)
4664 if (work->func == wq_barrier_func) {
4665 struct wq_barrier *barr;
4667 barr = container_of(work, struct wq_barrier, work);
4669 pr_cont("%s BAR(%d)", comma ? "," : "",
4670 task_pid_nr(barr->task));
4672 pr_cont("%s %ps", comma ? "," : "", work->func);
4676 static void show_pwq(struct pool_workqueue *pwq)
4678 struct worker_pool *pool = pwq->pool;
4679 struct work_struct *work;
4680 struct worker *worker;
4681 bool has_in_flight = false, has_pending = false;
4684 pr_info(" pwq %d:", pool->id);
4685 pr_cont_pool_info(pool);
4687 pr_cont(" active=%d/%d refcnt=%d%s\n",
4688 pwq->nr_active, pwq->max_active, pwq->refcnt,
4689 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4691 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4692 if (worker->current_pwq == pwq) {
4693 has_in_flight = true;
4697 if (has_in_flight) {
4700 pr_info(" in-flight:");
4701 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4702 if (worker->current_pwq != pwq)
4705 pr_cont("%s %d%s:%ps", comma ? "," : "",
4706 task_pid_nr(worker->task),
4707 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4708 worker->current_func);
4709 list_for_each_entry(work, &worker->scheduled, entry)
4710 pr_cont_work(false, work);
4716 list_for_each_entry(work, &pool->worklist, entry) {
4717 if (get_work_pwq(work) == pwq) {
4725 pr_info(" pending:");
4726 list_for_each_entry(work, &pool->worklist, entry) {
4727 if (get_work_pwq(work) != pwq)
4730 pr_cont_work(comma, work);
4731 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4736 if (!list_empty(&pwq->delayed_works)) {
4739 pr_info(" delayed:");
4740 list_for_each_entry(work, &pwq->delayed_works, entry) {
4741 pr_cont_work(comma, work);
4742 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4749 * show_workqueue_state - dump workqueue state
4751 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4752 * all busy workqueues and pools.
4754 void show_workqueue_state(void)
4756 struct workqueue_struct *wq;
4757 struct worker_pool *pool;
4758 unsigned long flags;
4763 pr_info("Showing busy workqueues and worker pools:\n");
4765 list_for_each_entry_rcu(wq, &workqueues, list) {
4766 struct pool_workqueue *pwq;
4769 for_each_pwq(pwq, wq) {
4770 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4778 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4780 for_each_pwq(pwq, wq) {
4781 spin_lock_irqsave(&pwq->pool->lock, flags);
4782 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4784 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4786 * We could be printing a lot from atomic context, e.g.
4787 * sysrq-t -> show_workqueue_state(). Avoid triggering
4790 touch_nmi_watchdog();
4794 for_each_pool(pool, pi) {
4795 struct worker *worker;
4798 spin_lock_irqsave(&pool->lock, flags);
4799 if (pool->nr_workers == pool->nr_idle)
4802 pr_info("pool %d:", pool->id);
4803 pr_cont_pool_info(pool);
4804 pr_cont(" hung=%us workers=%d",
4805 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4808 pr_cont(" manager: %d",
4809 task_pid_nr(pool->manager->task));
4810 list_for_each_entry(worker, &pool->idle_list, entry) {
4811 pr_cont(" %s%d", first ? "idle: " : "",
4812 task_pid_nr(worker->task));
4817 spin_unlock_irqrestore(&pool->lock, flags);
4819 * We could be printing a lot from atomic context, e.g.
4820 * sysrq-t -> show_workqueue_state(). Avoid triggering
4823 touch_nmi_watchdog();
4829 /* used to show worker information through /proc/PID/{comm,stat,status} */
4830 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4834 /* always show the actual comm */
4835 off = strscpy(buf, task->comm, size);
4839 /* stabilize PF_WQ_WORKER and worker pool association */
4840 mutex_lock(&wq_pool_attach_mutex);
4842 if (task->flags & PF_WQ_WORKER) {
4843 struct worker *worker = kthread_data(task);
4844 struct worker_pool *pool = worker->pool;
4847 spin_lock_irq(&pool->lock);
4849 * ->desc tracks information (wq name or
4850 * set_worker_desc()) for the latest execution. If
4851 * current, prepend '+', otherwise '-'.
4853 if (worker->desc[0] != '\0') {
4854 if (worker->current_work)
4855 scnprintf(buf + off, size - off, "+%s",
4858 scnprintf(buf + off, size - off, "-%s",
4861 spin_unlock_irq(&pool->lock);
4865 mutex_unlock(&wq_pool_attach_mutex);
4873 * There are two challenges in supporting CPU hotplug. Firstly, there
4874 * are a lot of assumptions on strong associations among work, pwq and
4875 * pool which make migrating pending and scheduled works very
4876 * difficult to implement without impacting hot paths. Secondly,
4877 * worker pools serve mix of short, long and very long running works making
4878 * blocked draining impractical.
4880 * This is solved by allowing the pools to be disassociated from the CPU
4881 * running as an unbound one and allowing it to be reattached later if the
4882 * cpu comes back online.
4885 static void unbind_workers(int cpu)
4887 struct worker_pool *pool;
4888 struct worker *worker;
4890 for_each_cpu_worker_pool(pool, cpu) {
4891 mutex_lock(&wq_pool_attach_mutex);
4892 spin_lock_irq(&pool->lock);
4895 * We've blocked all attach/detach operations. Make all workers
4896 * unbound and set DISASSOCIATED. Before this, all workers
4897 * except for the ones which are still executing works from
4898 * before the last CPU down must be on the cpu. After
4899 * this, they may become diasporas.
4901 for_each_pool_worker(worker, pool)
4902 worker->flags |= WORKER_UNBOUND;
4904 pool->flags |= POOL_DISASSOCIATED;
4906 spin_unlock_irq(&pool->lock);
4907 mutex_unlock(&wq_pool_attach_mutex);
4910 * Call schedule() so that we cross rq->lock and thus can
4911 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4912 * This is necessary as scheduler callbacks may be invoked
4918 * Sched callbacks are disabled now. Zap nr_running.
4919 * After this, nr_running stays zero and need_more_worker()
4920 * and keep_working() are always true as long as the
4921 * worklist is not empty. This pool now behaves as an
4922 * unbound (in terms of concurrency management) pool which
4923 * are served by workers tied to the pool.
4925 atomic_set(&pool->nr_running, 0);
4928 * With concurrency management just turned off, a busy
4929 * worker blocking could lead to lengthy stalls. Kick off
4930 * unbound chain execution of currently pending work items.
4932 spin_lock_irq(&pool->lock);
4933 wake_up_worker(pool);
4934 spin_unlock_irq(&pool->lock);
4939 * rebind_workers - rebind all workers of a pool to the associated CPU
4940 * @pool: pool of interest
4942 * @pool->cpu is coming online. Rebind all workers to the CPU.
4944 static void rebind_workers(struct worker_pool *pool)
4946 struct worker *worker;
4948 lockdep_assert_held(&wq_pool_attach_mutex);
4951 * Restore CPU affinity of all workers. As all idle workers should
4952 * be on the run-queue of the associated CPU before any local
4953 * wake-ups for concurrency management happen, restore CPU affinity
4954 * of all workers first and then clear UNBOUND. As we're called
4955 * from CPU_ONLINE, the following shouldn't fail.
4957 for_each_pool_worker(worker, pool)
4958 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4959 pool->attrs->cpumask) < 0);
4961 spin_lock_irq(&pool->lock);
4963 pool->flags &= ~POOL_DISASSOCIATED;
4965 for_each_pool_worker(worker, pool) {
4966 unsigned int worker_flags = worker->flags;
4969 * A bound idle worker should actually be on the runqueue
4970 * of the associated CPU for local wake-ups targeting it to
4971 * work. Kick all idle workers so that they migrate to the
4972 * associated CPU. Doing this in the same loop as
4973 * replacing UNBOUND with REBOUND is safe as no worker will
4974 * be bound before @pool->lock is released.
4976 if (worker_flags & WORKER_IDLE)
4977 wake_up_process(worker->task);
4980 * We want to clear UNBOUND but can't directly call
4981 * worker_clr_flags() or adjust nr_running. Atomically
4982 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4983 * @worker will clear REBOUND using worker_clr_flags() when
4984 * it initiates the next execution cycle thus restoring
4985 * concurrency management. Note that when or whether
4986 * @worker clears REBOUND doesn't affect correctness.
4988 * WRITE_ONCE() is necessary because @worker->flags may be
4989 * tested without holding any lock in
4990 * wq_worker_running(). Without it, NOT_RUNNING test may
4991 * fail incorrectly leading to premature concurrency
4992 * management operations.
4994 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4995 worker_flags |= WORKER_REBOUND;
4996 worker_flags &= ~WORKER_UNBOUND;
4997 WRITE_ONCE(worker->flags, worker_flags);
5000 spin_unlock_irq(&pool->lock);
5004 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5005 * @pool: unbound pool of interest
5006 * @cpu: the CPU which is coming up
5008 * An unbound pool may end up with a cpumask which doesn't have any online
5009 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5010 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5011 * online CPU before, cpus_allowed of all its workers should be restored.
5013 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5015 static cpumask_t cpumask;
5016 struct worker *worker;
5018 lockdep_assert_held(&wq_pool_attach_mutex);
5020 /* is @cpu allowed for @pool? */
5021 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5024 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5026 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5027 for_each_pool_worker(worker, pool)
5028 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5031 int workqueue_prepare_cpu(unsigned int cpu)
5033 struct worker_pool *pool;
5035 for_each_cpu_worker_pool(pool, cpu) {
5036 if (pool->nr_workers)
5038 if (!create_worker(pool))
5044 int workqueue_online_cpu(unsigned int cpu)
5046 struct worker_pool *pool;
5047 struct workqueue_struct *wq;
5050 mutex_lock(&wq_pool_mutex);
5052 for_each_pool(pool, pi) {
5053 mutex_lock(&wq_pool_attach_mutex);
5055 if (pool->cpu == cpu)
5056 rebind_workers(pool);
5057 else if (pool->cpu < 0)
5058 restore_unbound_workers_cpumask(pool, cpu);
5060 mutex_unlock(&wq_pool_attach_mutex);
5063 /* update NUMA affinity of unbound workqueues */
5064 list_for_each_entry(wq, &workqueues, list)
5065 wq_update_unbound_numa(wq, cpu, true);
5067 mutex_unlock(&wq_pool_mutex);
5071 int workqueue_offline_cpu(unsigned int cpu)
5073 struct workqueue_struct *wq;
5075 /* unbinding per-cpu workers should happen on the local CPU */
5076 if (WARN_ON(cpu != smp_processor_id()))
5079 unbind_workers(cpu);
5081 /* update NUMA affinity of unbound workqueues */
5082 mutex_lock(&wq_pool_mutex);
5083 list_for_each_entry(wq, &workqueues, list)
5084 wq_update_unbound_numa(wq, cpu, false);
5085 mutex_unlock(&wq_pool_mutex);
5090 struct work_for_cpu {
5091 struct work_struct work;
5097 static void work_for_cpu_fn(struct work_struct *work)
5099 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5101 wfc->ret = wfc->fn(wfc->arg);
5105 * work_on_cpu - run a function in thread context on a particular cpu
5106 * @cpu: the cpu to run on
5107 * @fn: the function to run
5108 * @arg: the function arg
5110 * It is up to the caller to ensure that the cpu doesn't go offline.
5111 * The caller must not hold any locks which would prevent @fn from completing.
5113 * Return: The value @fn returns.
5115 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5117 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5119 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5120 schedule_work_on(cpu, &wfc.work);
5121 flush_work(&wfc.work);
5122 destroy_work_on_stack(&wfc.work);
5125 EXPORT_SYMBOL_GPL(work_on_cpu);
5128 * work_on_cpu_safe - run a function in thread context on a particular cpu
5129 * @cpu: the cpu to run on
5130 * @fn: the function to run
5131 * @arg: the function argument
5133 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5134 * any locks which would prevent @fn from completing.
5136 * Return: The value @fn returns.
5138 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5143 if (cpu_online(cpu))
5144 ret = work_on_cpu(cpu, fn, arg);
5148 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5149 #endif /* CONFIG_SMP */
5151 #ifdef CONFIG_FREEZER
5154 * freeze_workqueues_begin - begin freezing workqueues
5156 * Start freezing workqueues. After this function returns, all freezable
5157 * workqueues will queue new works to their delayed_works list instead of
5161 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5163 void freeze_workqueues_begin(void)
5165 struct workqueue_struct *wq;
5166 struct pool_workqueue *pwq;
5168 mutex_lock(&wq_pool_mutex);
5170 WARN_ON_ONCE(workqueue_freezing);
5171 workqueue_freezing = true;
5173 list_for_each_entry(wq, &workqueues, list) {
5174 mutex_lock(&wq->mutex);
5175 for_each_pwq(pwq, wq)
5176 pwq_adjust_max_active(pwq);
5177 mutex_unlock(&wq->mutex);
5180 mutex_unlock(&wq_pool_mutex);
5184 * freeze_workqueues_busy - are freezable workqueues still busy?
5186 * Check whether freezing is complete. This function must be called
5187 * between freeze_workqueues_begin() and thaw_workqueues().
5190 * Grabs and releases wq_pool_mutex.
5193 * %true if some freezable workqueues are still busy. %false if freezing
5196 bool freeze_workqueues_busy(void)
5199 struct workqueue_struct *wq;
5200 struct pool_workqueue *pwq;
5202 mutex_lock(&wq_pool_mutex);
5204 WARN_ON_ONCE(!workqueue_freezing);
5206 list_for_each_entry(wq, &workqueues, list) {
5207 if (!(wq->flags & WQ_FREEZABLE))
5210 * nr_active is monotonically decreasing. It's safe
5211 * to peek without lock.
5214 for_each_pwq(pwq, wq) {
5215 WARN_ON_ONCE(pwq->nr_active < 0);
5216 if (pwq->nr_active) {
5225 mutex_unlock(&wq_pool_mutex);
5230 * thaw_workqueues - thaw workqueues
5232 * Thaw workqueues. Normal queueing is restored and all collected
5233 * frozen works are transferred to their respective pool worklists.
5236 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5238 void thaw_workqueues(void)
5240 struct workqueue_struct *wq;
5241 struct pool_workqueue *pwq;
5243 mutex_lock(&wq_pool_mutex);
5245 if (!workqueue_freezing)
5248 workqueue_freezing = false;
5250 /* restore max_active and repopulate worklist */
5251 list_for_each_entry(wq, &workqueues, list) {
5252 mutex_lock(&wq->mutex);
5253 for_each_pwq(pwq, wq)
5254 pwq_adjust_max_active(pwq);
5255 mutex_unlock(&wq->mutex);
5259 mutex_unlock(&wq_pool_mutex);
5261 #endif /* CONFIG_FREEZER */
5263 static int workqueue_apply_unbound_cpumask(void)
5267 struct workqueue_struct *wq;
5268 struct apply_wqattrs_ctx *ctx, *n;
5270 lockdep_assert_held(&wq_pool_mutex);
5272 list_for_each_entry(wq, &workqueues, list) {
5273 if (!(wq->flags & WQ_UNBOUND))
5275 /* creating multiple pwqs breaks ordering guarantee */
5276 if (wq->flags & __WQ_ORDERED)
5279 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5285 list_add_tail(&ctx->list, &ctxs);
5288 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5290 apply_wqattrs_commit(ctx);
5291 apply_wqattrs_cleanup(ctx);
5298 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5299 * @cpumask: the cpumask to set
5301 * The low-level workqueues cpumask is a global cpumask that limits
5302 * the affinity of all unbound workqueues. This function check the @cpumask
5303 * and apply it to all unbound workqueues and updates all pwqs of them.
5305 * Retun: 0 - Success
5306 * -EINVAL - Invalid @cpumask
5307 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5309 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5312 cpumask_var_t saved_cpumask;
5315 * Not excluding isolated cpus on purpose.
5316 * If the user wishes to include them, we allow that.
5318 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5319 if (!cpumask_empty(cpumask)) {
5320 apply_wqattrs_lock();
5321 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5326 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) {
5331 /* save the old wq_unbound_cpumask. */
5332 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5334 /* update wq_unbound_cpumask at first and apply it to wqs. */
5335 cpumask_copy(wq_unbound_cpumask, cpumask);
5336 ret = workqueue_apply_unbound_cpumask();
5338 /* restore the wq_unbound_cpumask when failed. */
5340 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5342 free_cpumask_var(saved_cpumask);
5344 apply_wqattrs_unlock();
5352 * Workqueues with WQ_SYSFS flag set is visible to userland via
5353 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5354 * following attributes.
5356 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5357 * max_active RW int : maximum number of in-flight work items
5359 * Unbound workqueues have the following extra attributes.
5361 * pool_ids RO int : the associated pool IDs for each node
5362 * nice RW int : nice value of the workers
5363 * cpumask RW mask : bitmask of allowed CPUs for the workers
5364 * numa RW bool : whether enable NUMA affinity
5367 struct workqueue_struct *wq;
5371 static struct workqueue_struct *dev_to_wq(struct device *dev)
5373 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5378 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5381 struct workqueue_struct *wq = dev_to_wq(dev);
5383 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5385 static DEVICE_ATTR_RO(per_cpu);
5387 static ssize_t max_active_show(struct device *dev,
5388 struct device_attribute *attr, char *buf)
5390 struct workqueue_struct *wq = dev_to_wq(dev);
5392 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5395 static ssize_t max_active_store(struct device *dev,
5396 struct device_attribute *attr, const char *buf,
5399 struct workqueue_struct *wq = dev_to_wq(dev);
5402 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5405 workqueue_set_max_active(wq, val);
5408 static DEVICE_ATTR_RW(max_active);
5410 static struct attribute *wq_sysfs_attrs[] = {
5411 &dev_attr_per_cpu.attr,
5412 &dev_attr_max_active.attr,
5415 ATTRIBUTE_GROUPS(wq_sysfs);
5417 static ssize_t wq_pool_ids_show(struct device *dev,
5418 struct device_attribute *attr, char *buf)
5420 struct workqueue_struct *wq = dev_to_wq(dev);
5421 const char *delim = "";
5422 int node, written = 0;
5426 for_each_node(node) {
5427 written += scnprintf(buf + written, PAGE_SIZE - written,
5428 "%s%d:%d", delim, node,
5429 unbound_pwq_by_node(wq, node)->pool->id);
5432 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5439 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5442 struct workqueue_struct *wq = dev_to_wq(dev);
5445 mutex_lock(&wq->mutex);
5446 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5447 mutex_unlock(&wq->mutex);
5452 /* prepare workqueue_attrs for sysfs store operations */
5453 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5455 struct workqueue_attrs *attrs;
5457 lockdep_assert_held(&wq_pool_mutex);
5459 attrs = alloc_workqueue_attrs();
5463 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5467 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5468 const char *buf, size_t count)
5470 struct workqueue_struct *wq = dev_to_wq(dev);
5471 struct workqueue_attrs *attrs;
5474 apply_wqattrs_lock();
5476 attrs = wq_sysfs_prep_attrs(wq);
5480 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5481 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5482 ret = apply_workqueue_attrs_locked(wq, attrs);
5487 apply_wqattrs_unlock();
5488 free_workqueue_attrs(attrs);
5489 return ret ?: count;
5492 static ssize_t wq_cpumask_show(struct device *dev,
5493 struct device_attribute *attr, char *buf)
5495 struct workqueue_struct *wq = dev_to_wq(dev);
5498 mutex_lock(&wq->mutex);
5499 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5500 cpumask_pr_args(wq->unbound_attrs->cpumask));
5501 mutex_unlock(&wq->mutex);
5505 static ssize_t wq_cpumask_store(struct device *dev,
5506 struct device_attribute *attr,
5507 const char *buf, size_t count)
5509 struct workqueue_struct *wq = dev_to_wq(dev);
5510 struct workqueue_attrs *attrs;
5513 apply_wqattrs_lock();
5515 attrs = wq_sysfs_prep_attrs(wq);
5519 ret = cpumask_parse(buf, attrs->cpumask);
5521 ret = apply_workqueue_attrs_locked(wq, attrs);
5524 apply_wqattrs_unlock();
5525 free_workqueue_attrs(attrs);
5526 return ret ?: count;
5529 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5532 struct workqueue_struct *wq = dev_to_wq(dev);
5535 mutex_lock(&wq->mutex);
5536 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5537 !wq->unbound_attrs->no_numa);
5538 mutex_unlock(&wq->mutex);
5543 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5544 const char *buf, size_t count)
5546 struct workqueue_struct *wq = dev_to_wq(dev);
5547 struct workqueue_attrs *attrs;
5548 int v, ret = -ENOMEM;
5550 apply_wqattrs_lock();
5552 attrs = wq_sysfs_prep_attrs(wq);
5557 if (sscanf(buf, "%d", &v) == 1) {
5558 attrs->no_numa = !v;
5559 ret = apply_workqueue_attrs_locked(wq, attrs);
5563 apply_wqattrs_unlock();
5564 free_workqueue_attrs(attrs);
5565 return ret ?: count;
5568 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5569 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5570 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5571 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5572 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5576 static struct bus_type wq_subsys = {
5577 .name = "workqueue",
5578 .dev_groups = wq_sysfs_groups,
5581 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5582 struct device_attribute *attr, char *buf)
5586 mutex_lock(&wq_pool_mutex);
5587 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5588 cpumask_pr_args(wq_unbound_cpumask));
5589 mutex_unlock(&wq_pool_mutex);
5594 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5595 struct device_attribute *attr, const char *buf, size_t count)
5597 cpumask_var_t cpumask;
5600 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5603 ret = cpumask_parse(buf, cpumask);
5605 ret = workqueue_set_unbound_cpumask(cpumask);
5607 free_cpumask_var(cpumask);
5608 return ret ? ret : count;
5611 static struct device_attribute wq_sysfs_cpumask_attr =
5612 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5613 wq_unbound_cpumask_store);
5615 static int __init wq_sysfs_init(void)
5619 err = subsys_virtual_register(&wq_subsys, NULL);
5623 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5625 core_initcall(wq_sysfs_init);
5627 static void wq_device_release(struct device *dev)
5629 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5635 * workqueue_sysfs_register - make a workqueue visible in sysfs
5636 * @wq: the workqueue to register
5638 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5639 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5640 * which is the preferred method.
5642 * Workqueue user should use this function directly iff it wants to apply
5643 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5644 * apply_workqueue_attrs() may race against userland updating the
5647 * Return: 0 on success, -errno on failure.
5649 int workqueue_sysfs_register(struct workqueue_struct *wq)
5651 struct wq_device *wq_dev;
5655 * Adjusting max_active or creating new pwqs by applying
5656 * attributes breaks ordering guarantee. Disallow exposing ordered
5659 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5662 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5667 wq_dev->dev.bus = &wq_subsys;
5668 wq_dev->dev.release = wq_device_release;
5669 dev_set_name(&wq_dev->dev, "%s", wq->name);
5672 * unbound_attrs are created separately. Suppress uevent until
5673 * everything is ready.
5675 dev_set_uevent_suppress(&wq_dev->dev, true);
5677 ret = device_register(&wq_dev->dev);
5679 put_device(&wq_dev->dev);
5684 if (wq->flags & WQ_UNBOUND) {
5685 struct device_attribute *attr;
5687 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5688 ret = device_create_file(&wq_dev->dev, attr);
5690 device_unregister(&wq_dev->dev);
5697 dev_set_uevent_suppress(&wq_dev->dev, false);
5698 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5703 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5704 * @wq: the workqueue to unregister
5706 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5708 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5710 struct wq_device *wq_dev = wq->wq_dev;
5716 device_unregister(&wq_dev->dev);
5718 #else /* CONFIG_SYSFS */
5719 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5720 #endif /* CONFIG_SYSFS */
5723 * Workqueue watchdog.
5725 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5726 * flush dependency, a concurrency managed work item which stays RUNNING
5727 * indefinitely. Workqueue stalls can be very difficult to debug as the
5728 * usual warning mechanisms don't trigger and internal workqueue state is
5731 * Workqueue watchdog monitors all worker pools periodically and dumps
5732 * state if some pools failed to make forward progress for a while where
5733 * forward progress is defined as the first item on ->worklist changing.
5735 * This mechanism is controlled through the kernel parameter
5736 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5737 * corresponding sysfs parameter file.
5739 #ifdef CONFIG_WQ_WATCHDOG
5741 static unsigned long wq_watchdog_thresh = 30;
5742 static struct timer_list wq_watchdog_timer;
5744 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5745 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5747 static void wq_watchdog_reset_touched(void)
5751 wq_watchdog_touched = jiffies;
5752 for_each_possible_cpu(cpu)
5753 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5756 static void wq_watchdog_timer_fn(struct timer_list *unused)
5758 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5759 bool lockup_detected = false;
5760 unsigned long now = jiffies;
5761 struct worker_pool *pool;
5769 for_each_pool(pool, pi) {
5770 unsigned long pool_ts, touched, ts;
5772 if (list_empty(&pool->worklist))
5776 * If a virtual machine is stopped by the host it can look to
5777 * the watchdog like a stall.
5779 kvm_check_and_clear_guest_paused();
5781 /* get the latest of pool and touched timestamps */
5782 pool_ts = READ_ONCE(pool->watchdog_ts);
5783 touched = READ_ONCE(wq_watchdog_touched);
5785 if (time_after(pool_ts, touched))
5790 if (pool->cpu >= 0) {
5791 unsigned long cpu_touched =
5792 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5794 if (time_after(cpu_touched, ts))
5799 if (time_after(now, ts + thresh)) {
5800 lockup_detected = true;
5801 pr_emerg("BUG: workqueue lockup - pool");
5802 pr_cont_pool_info(pool);
5803 pr_cont(" stuck for %us!\n",
5804 jiffies_to_msecs(now - pool_ts) / 1000);
5810 if (lockup_detected)
5811 show_workqueue_state();
5813 wq_watchdog_reset_touched();
5814 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5817 notrace void wq_watchdog_touch(int cpu)
5820 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5822 wq_watchdog_touched = jiffies;
5825 static void wq_watchdog_set_thresh(unsigned long thresh)
5827 wq_watchdog_thresh = 0;
5828 del_timer_sync(&wq_watchdog_timer);
5831 wq_watchdog_thresh = thresh;
5832 wq_watchdog_reset_touched();
5833 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5837 static int wq_watchdog_param_set_thresh(const char *val,
5838 const struct kernel_param *kp)
5840 unsigned long thresh;
5843 ret = kstrtoul(val, 0, &thresh);
5848 wq_watchdog_set_thresh(thresh);
5850 wq_watchdog_thresh = thresh;
5855 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5856 .set = wq_watchdog_param_set_thresh,
5857 .get = param_get_ulong,
5860 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5863 static void wq_watchdog_init(void)
5865 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5866 wq_watchdog_set_thresh(wq_watchdog_thresh);
5869 #else /* CONFIG_WQ_WATCHDOG */
5871 static inline void wq_watchdog_init(void) { }
5873 #endif /* CONFIG_WQ_WATCHDOG */
5875 static void __init wq_numa_init(void)
5880 if (num_possible_nodes() <= 1)
5883 if (wq_disable_numa) {
5884 pr_info("workqueue: NUMA affinity support disabled\n");
5888 for_each_possible_cpu(cpu) {
5889 if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) {
5890 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5895 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5896 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5899 * We want masks of possible CPUs of each node which isn't readily
5900 * available. Build one from cpu_to_node() which should have been
5901 * fully initialized by now.
5903 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5907 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5908 node_online(node) ? node : NUMA_NO_NODE));
5910 for_each_possible_cpu(cpu) {
5911 node = cpu_to_node(cpu);
5912 cpumask_set_cpu(cpu, tbl[node]);
5915 wq_numa_possible_cpumask = tbl;
5916 wq_numa_enabled = true;
5920 * workqueue_init_early - early init for workqueue subsystem
5922 * This is the first half of two-staged workqueue subsystem initialization
5923 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5924 * idr are up. It sets up all the data structures and system workqueues
5925 * and allows early boot code to create workqueues and queue/cancel work
5926 * items. Actual work item execution starts only after kthreads can be
5927 * created and scheduled right before early initcalls.
5929 int __init workqueue_init_early(void)
5931 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5932 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5935 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5937 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5938 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5940 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5942 /* initialize CPU pools */
5943 for_each_possible_cpu(cpu) {
5944 struct worker_pool *pool;
5947 for_each_cpu_worker_pool(pool, cpu) {
5948 BUG_ON(init_worker_pool(pool));
5950 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5951 pool->attrs->nice = std_nice[i++];
5952 pool->node = cpu_to_node(cpu);
5955 mutex_lock(&wq_pool_mutex);
5956 BUG_ON(worker_pool_assign_id(pool));
5957 mutex_unlock(&wq_pool_mutex);
5961 /* create default unbound and ordered wq attrs */
5962 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5963 struct workqueue_attrs *attrs;
5965 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5966 attrs->nice = std_nice[i];
5967 unbound_std_wq_attrs[i] = attrs;
5970 * An ordered wq should have only one pwq as ordering is
5971 * guaranteed by max_active which is enforced by pwqs.
5972 * Turn off NUMA so that dfl_pwq is used for all nodes.
5974 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5975 attrs->nice = std_nice[i];
5976 attrs->no_numa = true;
5977 ordered_wq_attrs[i] = attrs;
5980 system_wq = alloc_workqueue("events", 0, 0);
5981 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5982 system_long_wq = alloc_workqueue("events_long", 0, 0);
5983 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5984 WQ_UNBOUND_MAX_ACTIVE);
5985 system_freezable_wq = alloc_workqueue("events_freezable",
5987 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5988 WQ_POWER_EFFICIENT, 0);
5989 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5990 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5992 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5993 !system_unbound_wq || !system_freezable_wq ||
5994 !system_power_efficient_wq ||
5995 !system_freezable_power_efficient_wq);
6001 * workqueue_init - bring workqueue subsystem fully online
6003 * This is the latter half of two-staged workqueue subsystem initialization
6004 * and invoked as soon as kthreads can be created and scheduled.
6005 * Workqueues have been created and work items queued on them, but there
6006 * are no kworkers executing the work items yet. Populate the worker pools
6007 * with the initial workers and enable future kworker creations.
6009 int __init workqueue_init(void)
6011 struct workqueue_struct *wq;
6012 struct worker_pool *pool;
6016 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6017 * CPU to node mapping may not be available that early on some
6018 * archs such as power and arm64. As per-cpu pools created
6019 * previously could be missing node hint and unbound pools NUMA
6020 * affinity, fix them up.
6022 * Also, while iterating workqueues, create rescuers if requested.
6026 mutex_lock(&wq_pool_mutex);
6028 for_each_possible_cpu(cpu) {
6029 for_each_cpu_worker_pool(pool, cpu) {
6030 pool->node = cpu_to_node(cpu);
6034 list_for_each_entry(wq, &workqueues, list) {
6035 wq_update_unbound_numa(wq, smp_processor_id(), true);
6036 WARN(init_rescuer(wq),
6037 "workqueue: failed to create early rescuer for %s",
6041 mutex_unlock(&wq_pool_mutex);
6043 /* create the initial workers */
6044 for_each_online_cpu(cpu) {
6045 for_each_cpu_worker_pool(pool, cpu) {
6046 pool->flags &= ~POOL_DISASSOCIATED;
6047 BUG_ON(!create_worker(pool));
6051 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6052 BUG_ON(!create_worker(pool));