1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
53 #include <linux/kvm_para.h>
55 #include "workqueue_internal.h"
61 * A bound pool is either associated or disassociated with its CPU.
62 * While associated (!DISASSOCIATED), all workers are bound to the
63 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 * While DISASSOCIATED, the cpu may be offline and all workers have
67 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 * be executing on any CPU. The pool behaves as an unbound one.
70 * Note that DISASSOCIATED should be flipped only while holding
71 * wq_pool_attach_mutex to avoid changing binding state while
72 * worker_attach_to_pool() is in progress.
74 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
75 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
78 WORKER_DIE = 1 << 1, /* die die die */
79 WORKER_IDLE = 1 << 2, /* is idle */
80 WORKER_PREP = 1 << 3, /* preparing to run works */
81 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
82 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
83 WORKER_REBOUND = 1 << 8, /* worker was rebound */
85 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
86 WORKER_UNBOUND | WORKER_REBOUND,
88 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
90 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
91 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
93 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
94 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
96 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
97 /* call for help after 10ms
99 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
100 CREATE_COOLDOWN = HZ, /* time to breath after fail */
103 * Rescue workers are used only on emergencies and shared by
104 * all cpus. Give MIN_NICE.
106 RESCUER_NICE_LEVEL = MIN_NICE,
107 HIGHPRI_NICE_LEVEL = MIN_NICE,
113 * Structure fields follow one of the following exclusion rules.
115 * I: Modifiable by initialization/destruction paths and read-only for
118 * P: Preemption protected. Disabling preemption is enough and should
119 * only be modified and accessed from the local cpu.
121 * L: pool->lock protected. Access with pool->lock held.
123 * X: During normal operation, modification requires pool->lock and should
124 * be done only from local cpu. Either disabling preemption on local
125 * cpu or grabbing pool->lock is enough for read access. If
126 * POOL_DISASSOCIATED is set, it's identical to L.
128 * A: wq_pool_attach_mutex protected.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
134 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
136 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
139 * WQ: wq->mutex protected.
141 * WR: wq->mutex protected for writes. RCU protected for reads.
143 * MD: wq_mayday_lock protected.
146 /* struct worker is defined in workqueue_internal.h */
149 raw_spinlock_t lock; /* the pool lock */
150 int cpu; /* I: the associated cpu */
151 int node; /* I: the associated node ID */
152 int id; /* I: pool ID */
153 unsigned int flags; /* X: flags */
155 unsigned long watchdog_ts; /* L: watchdog timestamp */
157 struct list_head worklist; /* L: list of pending works */
159 int nr_workers; /* L: total number of workers */
160 int nr_idle; /* L: currently idle workers */
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
170 struct worker *manager; /* L: purely informational */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
174 struct ida worker_ida; /* worker IDs for task name */
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp;
188 * Destruction of pool is RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head inactive_works; /* L: inactive 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; /* MD: 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_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
305 /* wait for manager to go away */
306 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
308 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
309 static bool workqueue_freezing; /* PL: have wqs started freezing? */
311 /* PL: allowable cpus for unbound wqs and work items */
312 static cpumask_var_t wq_unbound_cpumask;
314 /* CPU where unbound work was last round robin scheduled from this CPU */
315 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
318 * Local execution of unbound work items is no longer guaranteed. The
319 * following always forces round-robin CPU selection on unbound work items
320 * to uncover usages which depend on it.
322 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
323 static bool wq_debug_force_rr_cpu = true;
325 static bool wq_debug_force_rr_cpu = false;
327 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
329 /* the per-cpu worker pools */
330 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
332 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
334 /* PL: hash of all unbound pools keyed by pool->attrs */
335 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
337 /* I: attributes used when instantiating standard unbound pools on demand */
338 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
340 /* I: attributes used when instantiating ordered pools on demand */
341 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
343 struct workqueue_struct *system_wq __read_mostly;
344 EXPORT_SYMBOL(system_wq);
345 struct workqueue_struct *system_highpri_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_highpri_wq);
347 struct workqueue_struct *system_long_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_long_wq);
349 struct workqueue_struct *system_unbound_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_unbound_wq);
351 struct workqueue_struct *system_freezable_wq __read_mostly;
352 EXPORT_SYMBOL_GPL(system_freezable_wq);
353 struct workqueue_struct *system_power_efficient_wq __read_mostly;
354 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
355 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
356 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
358 static int worker_thread(void *__worker);
359 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
360 static void show_pwq(struct pool_workqueue *pwq);
362 #define CREATE_TRACE_POINTS
363 #include <trace/events/workqueue.h>
365 #define assert_rcu_or_pool_mutex() \
366 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
367 !lockdep_is_held(&wq_pool_mutex), \
368 "RCU or wq_pool_mutex should be held")
370 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
371 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
372 !lockdep_is_held(&wq->mutex) && \
373 !lockdep_is_held(&wq_pool_mutex), \
374 "RCU, wq->mutex or wq_pool_mutex should be held")
376 #define for_each_cpu_worker_pool(pool, cpu) \
377 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
378 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
382 * for_each_pool - iterate through all worker_pools in the system
383 * @pool: iteration cursor
384 * @pi: integer used for iteration
386 * This must be called either with wq_pool_mutex held or RCU read
387 * locked. If the pool needs to be used beyond the locking in effect, the
388 * caller is responsible for guaranteeing that the pool stays online.
390 * The if/else clause exists only for the lockdep assertion and can be
393 #define for_each_pool(pool, pi) \
394 idr_for_each_entry(&worker_pool_idr, pool, pi) \
395 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
399 * for_each_pool_worker - iterate through all workers of a worker_pool
400 * @worker: iteration cursor
401 * @pool: worker_pool to iterate workers of
403 * This must be called with wq_pool_attach_mutex.
405 * The if/else clause exists only for the lockdep assertion and can be
408 #define for_each_pool_worker(worker, pool) \
409 list_for_each_entry((worker), &(pool)->workers, node) \
410 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
414 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
415 * @pwq: iteration cursor
416 * @wq: the target workqueue
418 * This must be called either with wq->mutex held or RCU read locked.
419 * If the pwq needs to be used beyond the locking in effect, the caller is
420 * responsible for guaranteeing that the pwq stays online.
422 * The if/else clause exists only for the lockdep assertion and can be
425 #define for_each_pwq(pwq, wq) \
426 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
427 lockdep_is_held(&(wq->mutex)))
429 #ifdef CONFIG_DEBUG_OBJECTS_WORK
431 static const struct debug_obj_descr work_debug_descr;
433 static void *work_debug_hint(void *addr)
435 return ((struct work_struct *) addr)->func;
438 static bool work_is_static_object(void *addr)
440 struct work_struct *work = addr;
442 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
446 * fixup_init is called when:
447 * - an active object is initialized
449 static bool work_fixup_init(void *addr, enum debug_obj_state state)
451 struct work_struct *work = addr;
454 case ODEBUG_STATE_ACTIVE:
455 cancel_work_sync(work);
456 debug_object_init(work, &work_debug_descr);
464 * fixup_free is called when:
465 * - an active object is freed
467 static bool work_fixup_free(void *addr, enum debug_obj_state state)
469 struct work_struct *work = addr;
472 case ODEBUG_STATE_ACTIVE:
473 cancel_work_sync(work);
474 debug_object_free(work, &work_debug_descr);
481 static const struct debug_obj_descr work_debug_descr = {
482 .name = "work_struct",
483 .debug_hint = work_debug_hint,
484 .is_static_object = work_is_static_object,
485 .fixup_init = work_fixup_init,
486 .fixup_free = work_fixup_free,
489 static inline void debug_work_activate(struct work_struct *work)
491 debug_object_activate(work, &work_debug_descr);
494 static inline void debug_work_deactivate(struct work_struct *work)
496 debug_object_deactivate(work, &work_debug_descr);
499 void __init_work(struct work_struct *work, int onstack)
502 debug_object_init_on_stack(work, &work_debug_descr);
504 debug_object_init(work, &work_debug_descr);
506 EXPORT_SYMBOL_GPL(__init_work);
508 void destroy_work_on_stack(struct work_struct *work)
510 debug_object_free(work, &work_debug_descr);
512 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
514 void destroy_delayed_work_on_stack(struct delayed_work *work)
516 destroy_timer_on_stack(&work->timer);
517 debug_object_free(&work->work, &work_debug_descr);
519 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
522 static inline void debug_work_activate(struct work_struct *work) { }
523 static inline void debug_work_deactivate(struct work_struct *work) { }
527 * worker_pool_assign_id - allocate ID and assing it to @pool
528 * @pool: the pool pointer of interest
530 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
531 * successfully, -errno on failure.
533 static int worker_pool_assign_id(struct worker_pool *pool)
537 lockdep_assert_held(&wq_pool_mutex);
539 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
549 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
550 * @wq: the target workqueue
553 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
555 * If the pwq needs to be used beyond the locking in effect, the caller is
556 * responsible for guaranteeing that the pwq stays online.
558 * Return: The unbound pool_workqueue for @node.
560 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
563 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
566 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
567 * delayed item is pending. The plan is to keep CPU -> NODE
568 * mapping valid and stable across CPU on/offlines. Once that
569 * happens, this workaround can be removed.
571 if (unlikely(node == NUMA_NO_NODE))
574 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
577 static unsigned int work_color_to_flags(int color)
579 return color << WORK_STRUCT_COLOR_SHIFT;
582 static int get_work_color(struct work_struct *work)
584 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
585 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
588 static int work_next_color(int color)
590 return (color + 1) % WORK_NR_COLORS;
594 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
595 * contain the pointer to the queued pwq. Once execution starts, the flag
596 * is cleared and the high bits contain OFFQ flags and pool ID.
598 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
599 * and clear_work_data() can be used to set the pwq, pool or clear
600 * work->data. These functions should only be called while the work is
601 * owned - ie. while the PENDING bit is set.
603 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
604 * corresponding to a work. Pool is available once the work has been
605 * queued anywhere after initialization until it is sync canceled. pwq is
606 * available only while the work item is queued.
608 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
609 * canceled. While being canceled, a work item may have its PENDING set
610 * but stay off timer and worklist for arbitrarily long and nobody should
611 * try to steal the PENDING bit.
613 static inline void set_work_data(struct work_struct *work, unsigned long data,
616 WARN_ON_ONCE(!work_pending(work));
617 atomic_long_set(&work->data, data | flags | work_static(work));
620 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
621 unsigned long extra_flags)
623 set_work_data(work, (unsigned long)pwq,
624 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
627 static void set_work_pool_and_keep_pending(struct work_struct *work,
630 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
631 WORK_STRUCT_PENDING);
634 static void set_work_pool_and_clear_pending(struct work_struct *work,
638 * The following wmb is paired with the implied mb in
639 * test_and_set_bit(PENDING) and ensures all updates to @work made
640 * here are visible to and precede any updates by the next PENDING
644 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
646 * The following mb guarantees that previous clear of a PENDING bit
647 * will not be reordered with any speculative LOADS or STORES from
648 * work->current_func, which is executed afterwards. This possible
649 * reordering can lead to a missed execution on attempt to queue
650 * the same @work. E.g. consider this case:
653 * ---------------------------- --------------------------------
655 * 1 STORE event_indicated
656 * 2 queue_work_on() {
657 * 3 test_and_set_bit(PENDING)
658 * 4 } set_..._and_clear_pending() {
659 * 5 set_work_data() # clear bit
661 * 7 work->current_func() {
662 * 8 LOAD event_indicated
665 * Without an explicit full barrier speculative LOAD on line 8 can
666 * be executed before CPU#0 does STORE on line 1. If that happens,
667 * CPU#0 observes the PENDING bit is still set and new execution of
668 * a @work is not queued in a hope, that CPU#1 will eventually
669 * finish the queued @work. Meanwhile CPU#1 does not see
670 * event_indicated is set, because speculative LOAD was executed
671 * before actual STORE.
676 static void clear_work_data(struct work_struct *work)
678 smp_wmb(); /* see set_work_pool_and_clear_pending() */
679 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
682 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
684 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
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 work_struct_pwq(data);
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 work_struct_pwq(data)->pool;
722 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
723 if (pool_id == WORK_OFFQ_POOL_NONE)
726 return idr_find(&worker_pool_idr, pool_id);
730 * get_work_pool_id - return the worker pool ID a given work is associated with
731 * @work: the work item of interest
733 * Return: The worker_pool ID @work was last associated with.
734 * %WORK_OFFQ_POOL_NONE if none.
736 static int get_work_pool_id(struct work_struct *work)
738 unsigned long data = atomic_long_read(&work->data);
740 if (data & WORK_STRUCT_PWQ)
741 return work_struct_pwq(data)->pool->id;
743 return data >> WORK_OFFQ_POOL_SHIFT;
746 static void mark_work_canceling(struct work_struct *work)
748 unsigned long pool_id = get_work_pool_id(work);
750 pool_id <<= WORK_OFFQ_POOL_SHIFT;
751 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
754 static bool work_is_canceling(struct work_struct *work)
756 unsigned long data = atomic_long_read(&work->data);
758 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
762 * Policy functions. These define the policies on how the global worker
763 * pools are managed. Unless noted otherwise, these functions assume that
764 * they're being called with pool->lock held.
767 static bool __need_more_worker(struct worker_pool *pool)
769 return !atomic_read(&pool->nr_running);
773 * Need to wake up a worker? Called from anything but currently
776 * Note that, because unbound workers never contribute to nr_running, this
777 * function will always return %true for unbound pools as long as the
778 * worklist isn't empty.
780 static bool need_more_worker(struct worker_pool *pool)
782 return !list_empty(&pool->worklist) && __need_more_worker(pool);
785 /* Can I start working? Called from busy but !running workers. */
786 static bool may_start_working(struct worker_pool *pool)
788 return pool->nr_idle;
791 /* Do I need to keep working? Called from currently running workers. */
792 static bool keep_working(struct worker_pool *pool)
794 return !list_empty(&pool->worklist) &&
795 atomic_read(&pool->nr_running) <= 1;
798 /* Do we need a new worker? Called from manager. */
799 static bool need_to_create_worker(struct worker_pool *pool)
801 return need_more_worker(pool) && !may_start_working(pool);
804 /* Do we have too many workers and should some go away? */
805 static bool too_many_workers(struct worker_pool *pool)
807 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
808 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
809 int nr_busy = pool->nr_workers - nr_idle;
811 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
818 /* Return the first idle worker. Safe with preemption disabled */
819 static struct worker *first_idle_worker(struct worker_pool *pool)
821 if (unlikely(list_empty(&pool->idle_list)))
824 return list_first_entry(&pool->idle_list, struct worker, entry);
828 * wake_up_worker - wake up an idle worker
829 * @pool: worker pool to wake worker from
831 * Wake up the first idle worker of @pool.
834 * raw_spin_lock_irq(pool->lock).
836 static void wake_up_worker(struct worker_pool *pool)
838 struct worker *worker = first_idle_worker(pool);
841 wake_up_process(worker->task);
845 * wq_worker_running - a worker is running again
846 * @task: task waking up
848 * This function is called when a worker returns from schedule()
850 void wq_worker_running(struct task_struct *task)
852 struct worker *worker = kthread_data(task);
854 if (!worker->sleeping)
858 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
859 * and the nr_running increment below, we may ruin the nr_running reset
860 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
861 * pool. Protect against such race.
864 if (!(worker->flags & WORKER_NOT_RUNNING))
865 atomic_inc(&worker->pool->nr_running);
867 worker->sleeping = 0;
871 * wq_worker_sleeping - a worker is going to sleep
872 * @task: task going to sleep
874 * This function is called from schedule() when a busy worker is
875 * going to sleep. Preemption needs to be disabled to protect ->sleeping
878 void wq_worker_sleeping(struct task_struct *task)
880 struct worker *next, *worker = kthread_data(task);
881 struct worker_pool *pool;
884 * Rescuers, which may not have all the fields set up like normal
885 * workers, also reach here, let's not access anything before
886 * checking NOT_RUNNING.
888 if (worker->flags & WORKER_NOT_RUNNING)
893 /* Return if preempted before wq_worker_running() was reached */
894 if (worker->sleeping)
897 worker->sleeping = 1;
898 raw_spin_lock_irq(&pool->lock);
901 * The counterpart of the following dec_and_test, implied mb,
902 * worklist not empty test sequence is in insert_work().
903 * Please read comment there.
905 * NOT_RUNNING is clear. This means that we're bound to and
906 * running on the local cpu w/ rq lock held and preemption
907 * disabled, which in turn means that none else could be
908 * manipulating idle_list, so dereferencing idle_list without pool
911 if (atomic_dec_and_test(&pool->nr_running) &&
912 !list_empty(&pool->worklist)) {
913 next = first_idle_worker(pool);
915 wake_up_process(next->task);
917 raw_spin_unlock_irq(&pool->lock);
921 * wq_worker_last_func - retrieve worker's last work function
922 * @task: Task to retrieve last work function of.
924 * Determine the last function a worker executed. This is called from
925 * the scheduler to get a worker's last known identity.
928 * raw_spin_lock_irq(rq->lock)
930 * This function is called during schedule() when a kworker is going
931 * to sleep. It's used by psi to identify aggregation workers during
932 * dequeuing, to allow periodic aggregation to shut-off when that
933 * worker is the last task in the system or cgroup to go to sleep.
935 * As this function doesn't involve any workqueue-related locking, it
936 * only returns stable values when called from inside the scheduler's
937 * queuing and dequeuing paths, when @task, which must be a kworker,
938 * is guaranteed to not be processing any works.
941 * The last work function %current executed as a worker, NULL if it
942 * hasn't executed any work yet.
944 work_func_t wq_worker_last_func(struct task_struct *task)
946 struct worker *worker = kthread_data(task);
948 return worker->last_func;
952 * worker_set_flags - set worker flags and adjust nr_running accordingly
954 * @flags: flags to set
956 * Set @flags in @worker->flags and adjust nr_running accordingly.
959 * raw_spin_lock_irq(pool->lock)
961 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
963 struct worker_pool *pool = worker->pool;
965 WARN_ON_ONCE(worker->task != current);
967 /* If transitioning into NOT_RUNNING, adjust nr_running. */
968 if ((flags & WORKER_NOT_RUNNING) &&
969 !(worker->flags & WORKER_NOT_RUNNING)) {
970 atomic_dec(&pool->nr_running);
973 worker->flags |= flags;
977 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
979 * @flags: flags to clear
981 * Clear @flags in @worker->flags and adjust nr_running accordingly.
984 * raw_spin_lock_irq(pool->lock)
986 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
988 struct worker_pool *pool = worker->pool;
989 unsigned int oflags = worker->flags;
991 WARN_ON_ONCE(worker->task != current);
993 worker->flags &= ~flags;
996 * If transitioning out of NOT_RUNNING, increment nr_running. Note
997 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
998 * of multiple flags, not a single flag.
1000 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1001 if (!(worker->flags & WORKER_NOT_RUNNING))
1002 atomic_inc(&pool->nr_running);
1006 * find_worker_executing_work - find worker which is executing a work
1007 * @pool: pool of interest
1008 * @work: work to find worker for
1010 * Find a worker which is executing @work on @pool by searching
1011 * @pool->busy_hash which is keyed by the address of @work. For a worker
1012 * to match, its current execution should match the address of @work and
1013 * its work function. This is to avoid unwanted dependency between
1014 * unrelated work executions through a work item being recycled while still
1017 * This is a bit tricky. A work item may be freed once its execution
1018 * starts and nothing prevents the freed area from being recycled for
1019 * another work item. If the same work item address ends up being reused
1020 * before the original execution finishes, workqueue will identify the
1021 * recycled work item as currently executing and make it wait until the
1022 * current execution finishes, introducing an unwanted dependency.
1024 * This function checks the work item address and work function to avoid
1025 * false positives. Note that this isn't complete as one may construct a
1026 * work function which can introduce dependency onto itself through a
1027 * recycled work item. Well, if somebody wants to shoot oneself in the
1028 * foot that badly, there's only so much we can do, and if such deadlock
1029 * actually occurs, it should be easy to locate the culprit work function.
1032 * raw_spin_lock_irq(pool->lock).
1035 * Pointer to worker which is executing @work if found, %NULL
1038 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1039 struct work_struct *work)
1041 struct worker *worker;
1043 hash_for_each_possible(pool->busy_hash, worker, hentry,
1044 (unsigned long)work)
1045 if (worker->current_work == work &&
1046 worker->current_func == work->func)
1053 * move_linked_works - move linked works to a list
1054 * @work: start of series of works to be scheduled
1055 * @head: target list to append @work to
1056 * @nextp: out parameter for nested worklist walking
1058 * Schedule linked works starting from @work to @head. Work series to
1059 * be scheduled starts at @work and includes any consecutive work with
1060 * WORK_STRUCT_LINKED set in its predecessor.
1062 * If @nextp is not NULL, it's updated to point to the next work of
1063 * the last scheduled work. This allows move_linked_works() to be
1064 * nested inside outer list_for_each_entry_safe().
1067 * raw_spin_lock_irq(pool->lock).
1069 static void move_linked_works(struct work_struct *work, struct list_head *head,
1070 struct work_struct **nextp)
1072 struct work_struct *n;
1075 * Linked worklist will always end before the end of the list,
1076 * use NULL for list head.
1078 list_for_each_entry_safe_from(work, n, NULL, entry) {
1079 list_move_tail(&work->entry, head);
1080 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1085 * If we're already inside safe list traversal and have moved
1086 * multiple works to the scheduled queue, the next position
1087 * needs to be updated.
1094 * get_pwq - get an extra reference on the specified pool_workqueue
1095 * @pwq: pool_workqueue to get
1097 * Obtain an extra reference on @pwq. The caller should guarantee that
1098 * @pwq has positive refcnt and be holding the matching pool->lock.
1100 static void get_pwq(struct pool_workqueue *pwq)
1102 lockdep_assert_held(&pwq->pool->lock);
1103 WARN_ON_ONCE(pwq->refcnt <= 0);
1108 * put_pwq - put a pool_workqueue reference
1109 * @pwq: pool_workqueue to put
1111 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1112 * destruction. The caller should be holding the matching pool->lock.
1114 static void put_pwq(struct pool_workqueue *pwq)
1116 lockdep_assert_held(&pwq->pool->lock);
1117 if (likely(--pwq->refcnt))
1119 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1122 * @pwq can't be released under pool->lock, bounce to
1123 * pwq_unbound_release_workfn(). This never recurses on the same
1124 * pool->lock as this path is taken only for unbound workqueues and
1125 * the release work item is scheduled on a per-cpu workqueue. To
1126 * avoid lockdep warning, unbound pool->locks are given lockdep
1127 * subclass of 1 in get_unbound_pool().
1129 schedule_work(&pwq->unbound_release_work);
1133 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1134 * @pwq: pool_workqueue to put (can be %NULL)
1136 * put_pwq() with locking. This function also allows %NULL @pwq.
1138 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1142 * As both pwqs and pools are RCU protected, the
1143 * following lock operations are safe.
1145 raw_spin_lock_irq(&pwq->pool->lock);
1147 raw_spin_unlock_irq(&pwq->pool->lock);
1151 static void pwq_activate_inactive_work(struct work_struct *work)
1153 struct pool_workqueue *pwq = get_work_pwq(work);
1155 trace_workqueue_activate_work(work);
1156 if (list_empty(&pwq->pool->worklist))
1157 pwq->pool->watchdog_ts = jiffies;
1158 move_linked_works(work, &pwq->pool->worklist, NULL);
1159 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1163 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1165 struct work_struct *work = list_first_entry(&pwq->inactive_works,
1166 struct work_struct, entry);
1168 pwq_activate_inactive_work(work);
1172 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1173 * @pwq: pwq of interest
1174 * @color: color of work which left the queue
1176 * A work either has completed or is removed from pending queue,
1177 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1180 * raw_spin_lock_irq(pool->lock).
1182 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1184 /* uncolored work items don't participate in flushing or nr_active */
1185 if (color == WORK_NO_COLOR)
1188 pwq->nr_in_flight[color]--;
1191 if (!list_empty(&pwq->inactive_works)) {
1192 /* one down, submit an inactive one */
1193 if (pwq->nr_active < pwq->max_active)
1194 pwq_activate_first_inactive(pwq);
1197 /* is flush in progress and are we at the flushing tip? */
1198 if (likely(pwq->flush_color != color))
1201 /* are there still in-flight works? */
1202 if (pwq->nr_in_flight[color])
1205 /* this pwq is done, clear flush_color */
1206 pwq->flush_color = -1;
1209 * If this was the last pwq, wake up the first flusher. It
1210 * will handle the rest.
1212 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1213 complete(&pwq->wq->first_flusher->done);
1219 * try_to_grab_pending - steal work item from worklist and disable irq
1220 * @work: work item to steal
1221 * @is_dwork: @work is a delayed_work
1222 * @flags: place to store irq state
1224 * Try to grab PENDING bit of @work. This function can handle @work in any
1225 * stable state - idle, on timer or on worklist.
1229 * ======== ================================================================
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
1235 * ======== ================================================================
1238 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1239 * interrupted while holding PENDING and @work off queue, irq must be
1240 * disabled on entry. This, combined with delayed_work->timer being
1241 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1243 * On successful return, >= 0, irq is disabled and the caller is
1244 * responsible for releasing it using local_irq_restore(*@flags).
1246 * This function is safe to call from any context including IRQ handler.
1248 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1249 unsigned long *flags)
1251 struct worker_pool *pool;
1252 struct pool_workqueue *pwq;
1254 local_irq_save(*flags);
1256 /* try to steal the timer if it exists */
1258 struct delayed_work *dwork = to_delayed_work(work);
1261 * dwork->timer is irqsafe. If del_timer() fails, it's
1262 * guaranteed that the timer is not queued anywhere and not
1263 * running on the local CPU.
1265 if (likely(del_timer(&dwork->timer)))
1269 /* try to claim PENDING the normal way */
1270 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1275 * The queueing is in progress, or it is already queued. Try to
1276 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1278 pool = get_work_pool(work);
1282 raw_spin_lock(&pool->lock);
1284 * work->data is guaranteed to point to pwq only while the work
1285 * item is queued on pwq->wq, and both updating work->data to point
1286 * to pwq on queueing and to pool on dequeueing are done under
1287 * pwq->pool->lock. This in turn guarantees that, if work->data
1288 * points to pwq which is associated with a locked pool, the work
1289 * item is currently queued on that pool.
1291 pwq = get_work_pwq(work);
1292 if (pwq && pwq->pool == pool) {
1293 debug_work_deactivate(work);
1296 * An inactive work item cannot be grabbed directly because
1297 * it might have linked NO_COLOR work items which, if left
1298 * on the inactive_works list, will confuse pwq->nr_active
1299 * management later on and cause stall. Make sure the work
1300 * item is activated before grabbing.
1302 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1303 pwq_activate_inactive_work(work);
1305 list_del_init(&work->entry);
1306 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1308 /* work->data points to pwq iff queued, point to pool */
1309 set_work_pool_and_keep_pending(work, pool->id);
1311 raw_spin_unlock(&pool->lock);
1315 raw_spin_unlock(&pool->lock);
1318 local_irq_restore(*flags);
1319 if (work_is_canceling(work))
1326 * insert_work - insert a work into a pool
1327 * @pwq: pwq @work belongs to
1328 * @work: work to insert
1329 * @head: insertion point
1330 * @extra_flags: extra WORK_STRUCT_* flags to set
1332 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1333 * work_struct flags.
1336 * raw_spin_lock_irq(pool->lock).
1338 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1339 struct list_head *head, unsigned int extra_flags)
1341 struct worker_pool *pool = pwq->pool;
1343 /* we own @work, set data and link */
1344 set_work_pwq(work, pwq, extra_flags);
1345 list_add_tail(&work->entry, head);
1349 * Ensure either wq_worker_sleeping() sees the above
1350 * list_add_tail() or we see zero nr_running to avoid workers lying
1351 * around lazily while there are works to be processed.
1355 if (__need_more_worker(pool))
1356 wake_up_worker(pool);
1360 * Test whether @work is being queued from another work executing on the
1363 static bool is_chained_work(struct workqueue_struct *wq)
1365 struct worker *worker;
1367 worker = current_wq_worker();
1369 * Return %true iff I'm a worker executing a work item on @wq. If
1370 * I'm @worker, it's safe to dereference it without locking.
1372 return worker && worker->current_pwq->wq == wq;
1376 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1377 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1378 * avoid perturbing sensitive tasks.
1380 static int wq_select_unbound_cpu(int cpu)
1382 static bool printed_dbg_warning;
1385 if (likely(!wq_debug_force_rr_cpu)) {
1386 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1388 } else if (!printed_dbg_warning) {
1389 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1390 printed_dbg_warning = true;
1393 if (cpumask_empty(wq_unbound_cpumask))
1396 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1397 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1398 if (unlikely(new_cpu >= nr_cpu_ids)) {
1399 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1400 if (unlikely(new_cpu >= nr_cpu_ids))
1403 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1408 static void __queue_work(int cpu, struct workqueue_struct *wq,
1409 struct work_struct *work)
1411 struct pool_workqueue *pwq;
1412 struct worker_pool *last_pool;
1413 struct list_head *worklist;
1414 unsigned int work_flags;
1415 unsigned int req_cpu = cpu;
1418 * While a work item is PENDING && off queue, a task trying to
1419 * steal the PENDING will busy-loop waiting for it to either get
1420 * queued or lose PENDING. Grabbing PENDING and queueing should
1421 * happen with IRQ disabled.
1423 lockdep_assert_irqs_disabled();
1426 /* if draining, only works from the same workqueue are allowed */
1427 if (unlikely(wq->flags & __WQ_DRAINING) &&
1428 WARN_ON_ONCE(!is_chained_work(wq)))
1432 /* pwq which will be used unless @work is executing elsewhere */
1433 if (wq->flags & WQ_UNBOUND) {
1434 if (req_cpu == WORK_CPU_UNBOUND)
1435 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1436 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1438 if (req_cpu == WORK_CPU_UNBOUND)
1439 cpu = raw_smp_processor_id();
1440 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1444 * If @work was previously on a different pool, it might still be
1445 * running there, in which case the work needs to be queued on that
1446 * pool to guarantee non-reentrancy.
1448 last_pool = get_work_pool(work);
1449 if (last_pool && last_pool != pwq->pool) {
1450 struct worker *worker;
1452 raw_spin_lock(&last_pool->lock);
1454 worker = find_worker_executing_work(last_pool, work);
1456 if (worker && worker->current_pwq->wq == wq) {
1457 pwq = worker->current_pwq;
1459 /* meh... not running there, queue here */
1460 raw_spin_unlock(&last_pool->lock);
1461 raw_spin_lock(&pwq->pool->lock);
1464 raw_spin_lock(&pwq->pool->lock);
1468 * pwq is determined and locked. For unbound pools, we could have
1469 * raced with pwq release and it could already be dead. If its
1470 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1471 * without another pwq replacing it in the numa_pwq_tbl or while
1472 * work items are executing on it, so the retrying is guaranteed to
1473 * make forward-progress.
1475 if (unlikely(!pwq->refcnt)) {
1476 if (wq->flags & WQ_UNBOUND) {
1477 raw_spin_unlock(&pwq->pool->lock);
1482 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1486 /* pwq determined, queue */
1487 trace_workqueue_queue_work(req_cpu, pwq, work);
1489 if (WARN_ON(!list_empty(&work->entry)))
1492 pwq->nr_in_flight[pwq->work_color]++;
1493 work_flags = work_color_to_flags(pwq->work_color);
1495 if (likely(pwq->nr_active < pwq->max_active)) {
1496 trace_workqueue_activate_work(work);
1498 worklist = &pwq->pool->worklist;
1499 if (list_empty(worklist))
1500 pwq->pool->watchdog_ts = jiffies;
1502 work_flags |= WORK_STRUCT_INACTIVE;
1503 worklist = &pwq->inactive_works;
1506 debug_work_activate(work);
1507 insert_work(pwq, work, worklist, work_flags);
1510 raw_spin_unlock(&pwq->pool->lock);
1515 * queue_work_on - queue work on specific cpu
1516 * @cpu: CPU number to execute work on
1517 * @wq: workqueue to use
1518 * @work: work to queue
1520 * We queue the work to a specific CPU, the caller must ensure it
1523 * Return: %false if @work was already on a queue, %true otherwise.
1525 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1526 struct work_struct *work)
1529 unsigned long flags;
1531 local_irq_save(flags);
1533 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1534 __queue_work(cpu, wq, work);
1538 local_irq_restore(flags);
1541 EXPORT_SYMBOL(queue_work_on);
1544 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1545 * @node: NUMA node ID that we want to select a CPU from
1547 * This function will attempt to find a "random" cpu available on a given
1548 * node. If there are no CPUs available on the given node it will return
1549 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1550 * available CPU if we need to schedule this work.
1552 static int workqueue_select_cpu_near(int node)
1556 /* No point in doing this if NUMA isn't enabled for workqueues */
1557 if (!wq_numa_enabled)
1558 return WORK_CPU_UNBOUND;
1560 /* Delay binding to CPU if node is not valid or online */
1561 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1562 return WORK_CPU_UNBOUND;
1564 /* Use local node/cpu if we are already there */
1565 cpu = raw_smp_processor_id();
1566 if (node == cpu_to_node(cpu))
1569 /* Use "random" otherwise know as "first" online CPU of node */
1570 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1572 /* If CPU is valid return that, otherwise just defer */
1573 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1577 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1578 * @node: NUMA node that we are targeting the work for
1579 * @wq: workqueue to use
1580 * @work: work to queue
1582 * We queue the work to a "random" CPU within a given NUMA node. The basic
1583 * idea here is to provide a way to somehow associate work with a given
1586 * This function will only make a best effort attempt at getting this onto
1587 * the right NUMA node. If no node is requested or the requested node is
1588 * offline then we just fall back to standard queue_work behavior.
1590 * Currently the "random" CPU ends up being the first available CPU in the
1591 * intersection of cpu_online_mask and the cpumask of the node, unless we
1592 * are running on the node. In that case we just use the current CPU.
1594 * Return: %false if @work was already on a queue, %true otherwise.
1596 bool queue_work_node(int node, struct workqueue_struct *wq,
1597 struct work_struct *work)
1599 unsigned long flags;
1603 * This current implementation is specific to unbound workqueues.
1604 * Specifically we only return the first available CPU for a given
1605 * node instead of cycling through individual CPUs within the node.
1607 * If this is used with a per-cpu workqueue then the logic in
1608 * workqueue_select_cpu_near would need to be updated to allow for
1609 * some round robin type logic.
1611 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1613 local_irq_save(flags);
1615 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1616 int cpu = workqueue_select_cpu_near(node);
1618 __queue_work(cpu, wq, work);
1622 local_irq_restore(flags);
1625 EXPORT_SYMBOL_GPL(queue_work_node);
1627 void delayed_work_timer_fn(struct timer_list *t)
1629 struct delayed_work *dwork = from_timer(dwork, t, timer);
1631 /* should have been called from irqsafe timer with irq already off */
1632 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1634 EXPORT_SYMBOL(delayed_work_timer_fn);
1636 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1637 struct delayed_work *dwork, unsigned long delay)
1639 struct timer_list *timer = &dwork->timer;
1640 struct work_struct *work = &dwork->work;
1643 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1644 WARN_ON_ONCE(timer_pending(timer));
1645 WARN_ON_ONCE(!list_empty(&work->entry));
1648 * If @delay is 0, queue @dwork->work immediately. This is for
1649 * both optimization and correctness. The earliest @timer can
1650 * expire is on the closest next tick and delayed_work users depend
1651 * on that there's no such delay when @delay is 0.
1654 __queue_work(cpu, wq, &dwork->work);
1660 timer->expires = jiffies + delay;
1662 if (unlikely(cpu != WORK_CPU_UNBOUND))
1663 add_timer_on(timer, cpu);
1669 * queue_delayed_work_on - queue work on specific CPU after delay
1670 * @cpu: CPU number to execute work on
1671 * @wq: workqueue to use
1672 * @dwork: work to queue
1673 * @delay: number of jiffies to wait before queueing
1675 * Return: %false if @work was already on a queue, %true otherwise. If
1676 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1679 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1680 struct delayed_work *dwork, unsigned long delay)
1682 struct work_struct *work = &dwork->work;
1684 unsigned long flags;
1686 /* read the comment in __queue_work() */
1687 local_irq_save(flags);
1689 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1690 __queue_delayed_work(cpu, wq, dwork, delay);
1694 local_irq_restore(flags);
1697 EXPORT_SYMBOL(queue_delayed_work_on);
1700 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1701 * @cpu: CPU number to execute work on
1702 * @wq: workqueue to use
1703 * @dwork: work to queue
1704 * @delay: number of jiffies to wait before queueing
1706 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1707 * modify @dwork's timer so that it expires after @delay. If @delay is
1708 * zero, @work is guaranteed to be scheduled immediately regardless of its
1711 * Return: %false if @dwork was idle and queued, %true if @dwork was
1712 * pending and its timer was modified.
1714 * This function is safe to call from any context including IRQ handler.
1715 * See try_to_grab_pending() for details.
1717 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1718 struct delayed_work *dwork, unsigned long delay)
1720 unsigned long flags;
1724 ret = try_to_grab_pending(&dwork->work, true, &flags);
1725 } while (unlikely(ret == -EAGAIN));
1727 if (likely(ret >= 0)) {
1728 __queue_delayed_work(cpu, wq, dwork, delay);
1729 local_irq_restore(flags);
1732 /* -ENOENT from try_to_grab_pending() becomes %true */
1735 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1737 static void rcu_work_rcufn(struct rcu_head *rcu)
1739 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1741 /* read the comment in __queue_work() */
1742 local_irq_disable();
1743 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1748 * queue_rcu_work - queue work after a RCU grace period
1749 * @wq: workqueue to use
1750 * @rwork: work to queue
1752 * Return: %false if @rwork was already pending, %true otherwise. Note
1753 * that a full RCU grace period is guaranteed only after a %true return.
1754 * While @rwork is guaranteed to be executed after a %false return, the
1755 * execution may happen before a full RCU grace period has passed.
1757 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1759 struct work_struct *work = &rwork->work;
1761 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1763 call_rcu(&rwork->rcu, rcu_work_rcufn);
1769 EXPORT_SYMBOL(queue_rcu_work);
1772 * worker_enter_idle - enter idle state
1773 * @worker: worker which is entering idle state
1775 * @worker is entering idle state. Update stats and idle timer if
1779 * raw_spin_lock_irq(pool->lock).
1781 static void worker_enter_idle(struct worker *worker)
1783 struct worker_pool *pool = worker->pool;
1785 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1786 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1787 (worker->hentry.next || worker->hentry.pprev)))
1790 /* can't use worker_set_flags(), also called from create_worker() */
1791 worker->flags |= WORKER_IDLE;
1793 worker->last_active = jiffies;
1795 /* idle_list is LIFO */
1796 list_add(&worker->entry, &pool->idle_list);
1798 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1799 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1802 * Sanity check nr_running. Because unbind_workers() releases
1803 * pool->lock between setting %WORKER_UNBOUND and zapping
1804 * nr_running, the warning may trigger spuriously. Check iff
1805 * unbind is not in progress.
1807 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1808 pool->nr_workers == pool->nr_idle &&
1809 atomic_read(&pool->nr_running));
1813 * worker_leave_idle - leave idle state
1814 * @worker: worker which is leaving idle state
1816 * @worker is leaving idle state. Update stats.
1819 * raw_spin_lock_irq(pool->lock).
1821 static void worker_leave_idle(struct worker *worker)
1823 struct worker_pool *pool = worker->pool;
1825 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1827 worker_clr_flags(worker, WORKER_IDLE);
1829 list_del_init(&worker->entry);
1832 static struct worker *alloc_worker(int node)
1834 struct worker *worker;
1836 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1838 INIT_LIST_HEAD(&worker->entry);
1839 INIT_LIST_HEAD(&worker->scheduled);
1840 INIT_LIST_HEAD(&worker->node);
1841 /* on creation a worker is in !idle && prep state */
1842 worker->flags = WORKER_PREP;
1848 * worker_attach_to_pool() - attach a worker to a pool
1849 * @worker: worker to be attached
1850 * @pool: the target pool
1852 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1853 * cpu-binding of @worker are kept coordinated with the pool across
1856 static void worker_attach_to_pool(struct worker *worker,
1857 struct worker_pool *pool)
1859 mutex_lock(&wq_pool_attach_mutex);
1862 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1863 * stable across this function. See the comments above the flag
1864 * definition for details.
1866 if (pool->flags & POOL_DISASSOCIATED)
1867 worker->flags |= WORKER_UNBOUND;
1869 if (worker->rescue_wq)
1870 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1872 list_add_tail(&worker->node, &pool->workers);
1873 worker->pool = pool;
1875 mutex_unlock(&wq_pool_attach_mutex);
1879 * worker_detach_from_pool() - detach a worker from its pool
1880 * @worker: worker which is attached to its pool
1882 * Undo the attaching which had been done in worker_attach_to_pool(). The
1883 * caller worker shouldn't access to the pool after detached except it has
1884 * other reference to the pool.
1886 static void worker_detach_from_pool(struct worker *worker)
1888 struct worker_pool *pool = worker->pool;
1889 struct completion *detach_completion = NULL;
1891 mutex_lock(&wq_pool_attach_mutex);
1893 list_del(&worker->node);
1894 worker->pool = NULL;
1896 if (list_empty(&pool->workers))
1897 detach_completion = pool->detach_completion;
1898 mutex_unlock(&wq_pool_attach_mutex);
1900 /* clear leftover flags without pool->lock after it is detached */
1901 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1903 if (detach_completion)
1904 complete(detach_completion);
1908 * create_worker - create a new workqueue worker
1909 * @pool: pool the new worker will belong to
1911 * Create and start a new worker which is attached to @pool.
1914 * Might sleep. Does GFP_KERNEL allocations.
1917 * Pointer to the newly created worker.
1919 static struct worker *create_worker(struct worker_pool *pool)
1921 struct worker *worker = NULL;
1925 /* ID is needed to determine kthread name */
1926 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1930 worker = alloc_worker(pool->node);
1937 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1938 pool->attrs->nice < 0 ? "H" : "");
1940 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1942 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1943 "kworker/%s", id_buf);
1944 if (IS_ERR(worker->task))
1947 set_user_nice(worker->task, pool->attrs->nice);
1948 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1950 /* successful, attach the worker to the pool */
1951 worker_attach_to_pool(worker, pool);
1953 /* start the newly created worker */
1954 raw_spin_lock_irq(&pool->lock);
1955 worker->pool->nr_workers++;
1956 worker_enter_idle(worker);
1957 wake_up_process(worker->task);
1958 raw_spin_unlock_irq(&pool->lock);
1964 ida_simple_remove(&pool->worker_ida, id);
1970 * destroy_worker - destroy a workqueue worker
1971 * @worker: worker to be destroyed
1973 * Destroy @worker and adjust @pool stats accordingly. The worker should
1977 * raw_spin_lock_irq(pool->lock).
1979 static void destroy_worker(struct worker *worker)
1981 struct worker_pool *pool = worker->pool;
1983 lockdep_assert_held(&pool->lock);
1985 /* sanity check frenzy */
1986 if (WARN_ON(worker->current_work) ||
1987 WARN_ON(!list_empty(&worker->scheduled)) ||
1988 WARN_ON(!(worker->flags & WORKER_IDLE)))
1994 list_del_init(&worker->entry);
1995 worker->flags |= WORKER_DIE;
1996 wake_up_process(worker->task);
1999 static void idle_worker_timeout(struct timer_list *t)
2001 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2003 raw_spin_lock_irq(&pool->lock);
2005 while (too_many_workers(pool)) {
2006 struct worker *worker;
2007 unsigned long expires;
2009 /* idle_list is kept in LIFO order, check the last one */
2010 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2011 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2013 if (time_before(jiffies, expires)) {
2014 mod_timer(&pool->idle_timer, expires);
2018 destroy_worker(worker);
2021 raw_spin_unlock_irq(&pool->lock);
2024 static void send_mayday(struct work_struct *work)
2026 struct pool_workqueue *pwq = get_work_pwq(work);
2027 struct workqueue_struct *wq = pwq->wq;
2029 lockdep_assert_held(&wq_mayday_lock);
2034 /* mayday mayday mayday */
2035 if (list_empty(&pwq->mayday_node)) {
2037 * If @pwq is for an unbound wq, its base ref may be put at
2038 * any time due to an attribute change. Pin @pwq until the
2039 * rescuer is done with it.
2042 list_add_tail(&pwq->mayday_node, &wq->maydays);
2043 wake_up_process(wq->rescuer->task);
2047 static void pool_mayday_timeout(struct timer_list *t)
2049 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2050 struct work_struct *work;
2052 raw_spin_lock_irq(&pool->lock);
2053 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2055 if (need_to_create_worker(pool)) {
2057 * We've been trying to create a new worker but
2058 * haven't been successful. We might be hitting an
2059 * allocation deadlock. Send distress signals to
2062 list_for_each_entry(work, &pool->worklist, entry)
2066 raw_spin_unlock(&wq_mayday_lock);
2067 raw_spin_unlock_irq(&pool->lock);
2069 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2073 * maybe_create_worker - create a new worker if necessary
2074 * @pool: pool to create a new worker for
2076 * Create a new worker for @pool if necessary. @pool is guaranteed to
2077 * have at least one idle worker on return from this function. If
2078 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2079 * sent to all rescuers with works scheduled on @pool to resolve
2080 * possible allocation deadlock.
2082 * On return, need_to_create_worker() is guaranteed to be %false and
2083 * may_start_working() %true.
2086 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2087 * multiple times. Does GFP_KERNEL allocations. Called only from
2090 static void maybe_create_worker(struct worker_pool *pool)
2091 __releases(&pool->lock)
2092 __acquires(&pool->lock)
2095 raw_spin_unlock_irq(&pool->lock);
2097 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2098 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2101 if (create_worker(pool) || !need_to_create_worker(pool))
2104 schedule_timeout_interruptible(CREATE_COOLDOWN);
2106 if (!need_to_create_worker(pool))
2110 del_timer_sync(&pool->mayday_timer);
2111 raw_spin_lock_irq(&pool->lock);
2113 * This is necessary even after a new worker was just successfully
2114 * created as @pool->lock was dropped and the new worker might have
2115 * already become busy.
2117 if (need_to_create_worker(pool))
2122 * manage_workers - manage worker pool
2125 * Assume the manager role and manage the worker pool @worker belongs
2126 * to. At any given time, there can be only zero or one manager per
2127 * pool. The exclusion is handled automatically by this function.
2129 * The caller can safely start processing works on false return. On
2130 * true return, it's guaranteed that need_to_create_worker() is false
2131 * and may_start_working() is true.
2134 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2135 * multiple times. Does GFP_KERNEL allocations.
2138 * %false if the pool doesn't need management and the caller can safely
2139 * start processing works, %true if management function was performed and
2140 * the conditions that the caller verified before calling the function may
2141 * no longer be true.
2143 static bool manage_workers(struct worker *worker)
2145 struct worker_pool *pool = worker->pool;
2147 if (pool->flags & POOL_MANAGER_ACTIVE)
2150 pool->flags |= POOL_MANAGER_ACTIVE;
2151 pool->manager = worker;
2153 maybe_create_worker(pool);
2155 pool->manager = NULL;
2156 pool->flags &= ~POOL_MANAGER_ACTIVE;
2157 rcuwait_wake_up(&manager_wait);
2162 * process_one_work - process single work
2164 * @work: work to process
2166 * Process @work. This function contains all the logics necessary to
2167 * process a single work including synchronization against and
2168 * interaction with other workers on the same cpu, queueing and
2169 * flushing. As long as context requirement is met, any worker can
2170 * call this function to process a work.
2173 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2175 static void process_one_work(struct worker *worker, struct work_struct *work)
2176 __releases(&pool->lock)
2177 __acquires(&pool->lock)
2179 struct pool_workqueue *pwq = get_work_pwq(work);
2180 struct worker_pool *pool = worker->pool;
2181 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2183 struct worker *collision;
2184 #ifdef CONFIG_LOCKDEP
2186 * It is permissible to free the struct work_struct from
2187 * inside the function that is called from it, this we need to
2188 * take into account for lockdep too. To avoid bogus "held
2189 * lock freed" warnings as well as problems when looking into
2190 * work->lockdep_map, make a copy and use that here.
2192 struct lockdep_map lockdep_map;
2194 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2196 /* ensure we're on the correct CPU */
2197 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2198 raw_smp_processor_id() != pool->cpu);
2201 * A single work shouldn't be executed concurrently by
2202 * multiple workers on a single cpu. Check whether anyone is
2203 * already processing the work. If so, defer the work to the
2204 * currently executing one.
2206 collision = find_worker_executing_work(pool, work);
2207 if (unlikely(collision)) {
2208 move_linked_works(work, &collision->scheduled, NULL);
2212 /* claim and dequeue */
2213 debug_work_deactivate(work);
2214 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2215 worker->current_work = work;
2216 worker->current_func = work->func;
2217 worker->current_pwq = pwq;
2218 work_color = get_work_color(work);
2221 * Record wq name for cmdline and debug reporting, may get
2222 * overridden through set_worker_desc().
2224 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2226 list_del_init(&work->entry);
2229 * CPU intensive works don't participate in concurrency management.
2230 * They're the scheduler's responsibility. This takes @worker out
2231 * of concurrency management and the next code block will chain
2232 * execution of the pending work items.
2234 if (unlikely(cpu_intensive))
2235 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2238 * Wake up another worker if necessary. The condition is always
2239 * false for normal per-cpu workers since nr_running would always
2240 * be >= 1 at this point. This is used to chain execution of the
2241 * pending work items for WORKER_NOT_RUNNING workers such as the
2242 * UNBOUND and CPU_INTENSIVE ones.
2244 if (need_more_worker(pool))
2245 wake_up_worker(pool);
2248 * Record the last pool and clear PENDING which should be the last
2249 * update to @work. Also, do this inside @pool->lock so that
2250 * PENDING and queued state changes happen together while IRQ is
2253 set_work_pool_and_clear_pending(work, pool->id);
2255 raw_spin_unlock_irq(&pool->lock);
2257 lock_map_acquire(&pwq->wq->lockdep_map);
2258 lock_map_acquire(&lockdep_map);
2260 * Strictly speaking we should mark the invariant state without holding
2261 * any locks, that is, before these two lock_map_acquire()'s.
2263 * However, that would result in:
2270 * Which would create W1->C->W1 dependencies, even though there is no
2271 * actual deadlock possible. There are two solutions, using a
2272 * read-recursive acquire on the work(queue) 'locks', but this will then
2273 * hit the lockdep limitation on recursive locks, or simply discard
2276 * AFAICT there is no possible deadlock scenario between the
2277 * flush_work() and complete() primitives (except for single-threaded
2278 * workqueues), so hiding them isn't a problem.
2280 lockdep_invariant_state(true);
2281 trace_workqueue_execute_start(work);
2282 worker->current_func(work);
2284 * While we must be careful to not use "work" after this, the trace
2285 * point will only record its address.
2287 trace_workqueue_execute_end(work, worker->current_func);
2288 lock_map_release(&lockdep_map);
2289 lock_map_release(&pwq->wq->lockdep_map);
2291 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2292 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2293 " last function: %ps\n",
2294 current->comm, preempt_count(), task_pid_nr(current),
2295 worker->current_func);
2296 debug_show_held_locks(current);
2301 * The following prevents a kworker from hogging CPU on !PREEMPTION
2302 * kernels, where a requeueing work item waiting for something to
2303 * happen could deadlock with stop_machine as such work item could
2304 * indefinitely requeue itself while all other CPUs are trapped in
2305 * stop_machine. At the same time, report a quiescent RCU state so
2306 * the same condition doesn't freeze RCU.
2310 raw_spin_lock_irq(&pool->lock);
2312 /* clear cpu intensive status */
2313 if (unlikely(cpu_intensive))
2314 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2316 /* tag the worker for identification in schedule() */
2317 worker->last_func = worker->current_func;
2319 /* we're done with it, release */
2320 hash_del(&worker->hentry);
2321 worker->current_work = NULL;
2322 worker->current_func = NULL;
2323 worker->current_pwq = NULL;
2324 pwq_dec_nr_in_flight(pwq, work_color);
2328 * process_scheduled_works - process scheduled works
2331 * Process all scheduled works. Please note that the scheduled list
2332 * may change while processing a work, so this function repeatedly
2333 * fetches a work from the top and executes it.
2336 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2339 static void process_scheduled_works(struct worker *worker)
2341 while (!list_empty(&worker->scheduled)) {
2342 struct work_struct *work = list_first_entry(&worker->scheduled,
2343 struct work_struct, entry);
2344 process_one_work(worker, work);
2348 static void set_pf_worker(bool val)
2350 mutex_lock(&wq_pool_attach_mutex);
2352 current->flags |= PF_WQ_WORKER;
2354 current->flags &= ~PF_WQ_WORKER;
2355 mutex_unlock(&wq_pool_attach_mutex);
2359 * worker_thread - the worker thread function
2362 * The worker thread function. All workers belong to a worker_pool -
2363 * either a per-cpu one or dynamic unbound one. These workers process all
2364 * work items regardless of their specific target workqueue. The only
2365 * exception is work items which belong to workqueues with a rescuer which
2366 * will be explained in rescuer_thread().
2370 static int worker_thread(void *__worker)
2372 struct worker *worker = __worker;
2373 struct worker_pool *pool = worker->pool;
2375 /* tell the scheduler that this is a workqueue worker */
2376 set_pf_worker(true);
2378 raw_spin_lock_irq(&pool->lock);
2380 /* am I supposed to die? */
2381 if (unlikely(worker->flags & WORKER_DIE)) {
2382 raw_spin_unlock_irq(&pool->lock);
2383 WARN_ON_ONCE(!list_empty(&worker->entry));
2384 set_pf_worker(false);
2386 set_task_comm(worker->task, "kworker/dying");
2387 ida_simple_remove(&pool->worker_ida, worker->id);
2388 worker_detach_from_pool(worker);
2393 worker_leave_idle(worker);
2395 /* no more worker necessary? */
2396 if (!need_more_worker(pool))
2399 /* do we need to manage? */
2400 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2404 * ->scheduled list can only be filled while a worker is
2405 * preparing to process a work or actually processing it.
2406 * Make sure nobody diddled with it while I was sleeping.
2408 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2411 * Finish PREP stage. We're guaranteed to have at least one idle
2412 * worker or that someone else has already assumed the manager
2413 * role. This is where @worker starts participating in concurrency
2414 * management if applicable and concurrency management is restored
2415 * after being rebound. See rebind_workers() for details.
2417 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2420 struct work_struct *work =
2421 list_first_entry(&pool->worklist,
2422 struct work_struct, entry);
2424 pool->watchdog_ts = jiffies;
2426 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2427 /* optimization path, not strictly necessary */
2428 process_one_work(worker, work);
2429 if (unlikely(!list_empty(&worker->scheduled)))
2430 process_scheduled_works(worker);
2432 move_linked_works(work, &worker->scheduled, NULL);
2433 process_scheduled_works(worker);
2435 } while (keep_working(pool));
2437 worker_set_flags(worker, WORKER_PREP);
2440 * pool->lock is held and there's no work to process and no need to
2441 * manage, sleep. Workers are woken up only while holding
2442 * pool->lock or from local cpu, so setting the current state
2443 * before releasing pool->lock is enough to prevent losing any
2446 worker_enter_idle(worker);
2447 __set_current_state(TASK_IDLE);
2448 raw_spin_unlock_irq(&pool->lock);
2454 * rescuer_thread - the rescuer thread function
2457 * Workqueue rescuer thread function. There's one rescuer for each
2458 * workqueue which has WQ_MEM_RECLAIM set.
2460 * Regular work processing on a pool may block trying to create a new
2461 * worker which uses GFP_KERNEL allocation which has slight chance of
2462 * developing into deadlock if some works currently on the same queue
2463 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2464 * the problem rescuer solves.
2466 * When such condition is possible, the pool summons rescuers of all
2467 * workqueues which have works queued on the pool and let them process
2468 * those works so that forward progress can be guaranteed.
2470 * This should happen rarely.
2474 static int rescuer_thread(void *__rescuer)
2476 struct worker *rescuer = __rescuer;
2477 struct workqueue_struct *wq = rescuer->rescue_wq;
2478 struct list_head *scheduled = &rescuer->scheduled;
2481 set_user_nice(current, RESCUER_NICE_LEVEL);
2484 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2485 * doesn't participate in concurrency management.
2487 set_pf_worker(true);
2489 set_current_state(TASK_IDLE);
2492 * By the time the rescuer is requested to stop, the workqueue
2493 * shouldn't have any work pending, but @wq->maydays may still have
2494 * pwq(s) queued. This can happen by non-rescuer workers consuming
2495 * all the work items before the rescuer got to them. Go through
2496 * @wq->maydays processing before acting on should_stop so that the
2497 * list is always empty on exit.
2499 should_stop = kthread_should_stop();
2501 /* see whether any pwq is asking for help */
2502 raw_spin_lock_irq(&wq_mayday_lock);
2504 while (!list_empty(&wq->maydays)) {
2505 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2506 struct pool_workqueue, mayday_node);
2507 struct worker_pool *pool = pwq->pool;
2508 struct work_struct *work, *n;
2511 __set_current_state(TASK_RUNNING);
2512 list_del_init(&pwq->mayday_node);
2514 raw_spin_unlock_irq(&wq_mayday_lock);
2516 worker_attach_to_pool(rescuer, pool);
2518 raw_spin_lock_irq(&pool->lock);
2521 * Slurp in all works issued via this workqueue and
2524 WARN_ON_ONCE(!list_empty(scheduled));
2525 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2526 if (get_work_pwq(work) == pwq) {
2528 pool->watchdog_ts = jiffies;
2529 move_linked_works(work, scheduled, &n);
2534 if (!list_empty(scheduled)) {
2535 process_scheduled_works(rescuer);
2538 * The above execution of rescued work items could
2539 * have created more to rescue through
2540 * pwq_activate_first_inactive() or chained
2541 * queueing. Let's put @pwq back on mayday list so
2542 * that such back-to-back work items, which may be
2543 * being used to relieve memory pressure, don't
2544 * incur MAYDAY_INTERVAL delay inbetween.
2546 if (pwq->nr_active && need_to_create_worker(pool)) {
2547 raw_spin_lock(&wq_mayday_lock);
2549 * Queue iff we aren't racing destruction
2550 * and somebody else hasn't queued it already.
2552 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2554 list_add_tail(&pwq->mayday_node, &wq->maydays);
2556 raw_spin_unlock(&wq_mayday_lock);
2561 * Put the reference grabbed by send_mayday(). @pool won't
2562 * go away while we're still attached to it.
2567 * Leave this pool. If need_more_worker() is %true, notify a
2568 * regular worker; otherwise, we end up with 0 concurrency
2569 * and stalling the execution.
2571 if (need_more_worker(pool))
2572 wake_up_worker(pool);
2574 raw_spin_unlock_irq(&pool->lock);
2576 worker_detach_from_pool(rescuer);
2578 raw_spin_lock_irq(&wq_mayday_lock);
2581 raw_spin_unlock_irq(&wq_mayday_lock);
2584 __set_current_state(TASK_RUNNING);
2585 set_pf_worker(false);
2589 /* rescuers should never participate in concurrency management */
2590 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2596 * check_flush_dependency - check for flush dependency sanity
2597 * @target_wq: workqueue being flushed
2598 * @target_work: work item being flushed (NULL for workqueue flushes)
2600 * %current is trying to flush the whole @target_wq or @target_work on it.
2601 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2602 * reclaiming memory or running on a workqueue which doesn't have
2603 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2606 static void check_flush_dependency(struct workqueue_struct *target_wq,
2607 struct work_struct *target_work)
2609 work_func_t target_func = target_work ? target_work->func : NULL;
2610 struct worker *worker;
2612 if (target_wq->flags & WQ_MEM_RECLAIM)
2615 worker = current_wq_worker();
2617 WARN_ONCE(current->flags & PF_MEMALLOC,
2618 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2619 current->pid, current->comm, target_wq->name, target_func);
2620 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2621 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2622 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2623 worker->current_pwq->wq->name, worker->current_func,
2624 target_wq->name, target_func);
2628 struct work_struct work;
2629 struct completion done;
2630 struct task_struct *task; /* purely informational */
2633 static void wq_barrier_func(struct work_struct *work)
2635 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2636 complete(&barr->done);
2640 * insert_wq_barrier - insert a barrier work
2641 * @pwq: pwq to insert barrier into
2642 * @barr: wq_barrier to insert
2643 * @target: target work to attach @barr to
2644 * @worker: worker currently executing @target, NULL if @target is not executing
2646 * @barr is linked to @target such that @barr is completed only after
2647 * @target finishes execution. Please note that the ordering
2648 * guarantee is observed only with respect to @target and on the local
2651 * Currently, a queued barrier can't be canceled. This is because
2652 * try_to_grab_pending() can't determine whether the work to be
2653 * grabbed is at the head of the queue and thus can't clear LINKED
2654 * flag of the previous work while there must be a valid next work
2655 * after a work with LINKED flag set.
2657 * Note that when @worker is non-NULL, @target may be modified
2658 * underneath us, so we can't reliably determine pwq from @target.
2661 * raw_spin_lock_irq(pool->lock).
2663 static void insert_wq_barrier(struct pool_workqueue *pwq,
2664 struct wq_barrier *barr,
2665 struct work_struct *target, struct worker *worker)
2667 struct list_head *head;
2668 unsigned int linked = 0;
2671 * debugobject calls are safe here even with pool->lock locked
2672 * as we know for sure that this will not trigger any of the
2673 * checks and call back into the fixup functions where we
2676 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2677 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2679 init_completion_map(&barr->done, &target->lockdep_map);
2681 barr->task = current;
2684 * If @target is currently being executed, schedule the
2685 * barrier to the worker; otherwise, put it after @target.
2688 head = worker->scheduled.next;
2690 unsigned long *bits = work_data_bits(target);
2692 head = target->entry.next;
2693 /* there can already be other linked works, inherit and set */
2694 linked = *bits & WORK_STRUCT_LINKED;
2695 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2698 debug_work_activate(&barr->work);
2699 insert_work(pwq, &barr->work, head,
2700 work_color_to_flags(WORK_NO_COLOR) | linked);
2704 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2705 * @wq: workqueue being flushed
2706 * @flush_color: new flush color, < 0 for no-op
2707 * @work_color: new work color, < 0 for no-op
2709 * Prepare pwqs for workqueue flushing.
2711 * If @flush_color is non-negative, flush_color on all pwqs should be
2712 * -1. If no pwq has in-flight commands at the specified color, all
2713 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2714 * has in flight commands, its pwq->flush_color is set to
2715 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2716 * wakeup logic is armed and %true is returned.
2718 * The caller should have initialized @wq->first_flusher prior to
2719 * calling this function with non-negative @flush_color. If
2720 * @flush_color is negative, no flush color update is done and %false
2723 * If @work_color is non-negative, all pwqs should have the same
2724 * work_color which is previous to @work_color and all will be
2725 * advanced to @work_color.
2728 * mutex_lock(wq->mutex).
2731 * %true if @flush_color >= 0 and there's something to flush. %false
2734 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2735 int flush_color, int work_color)
2738 struct pool_workqueue *pwq;
2740 if (flush_color >= 0) {
2741 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2742 atomic_set(&wq->nr_pwqs_to_flush, 1);
2745 for_each_pwq(pwq, wq) {
2746 struct worker_pool *pool = pwq->pool;
2748 raw_spin_lock_irq(&pool->lock);
2750 if (flush_color >= 0) {
2751 WARN_ON_ONCE(pwq->flush_color != -1);
2753 if (pwq->nr_in_flight[flush_color]) {
2754 pwq->flush_color = flush_color;
2755 atomic_inc(&wq->nr_pwqs_to_flush);
2760 if (work_color >= 0) {
2761 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2762 pwq->work_color = work_color;
2765 raw_spin_unlock_irq(&pool->lock);
2768 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2769 complete(&wq->first_flusher->done);
2775 * flush_workqueue - ensure that any scheduled work has run to completion.
2776 * @wq: workqueue to flush
2778 * This function sleeps until all work items which were queued on entry
2779 * have finished execution, but it is not livelocked by new incoming ones.
2781 void flush_workqueue(struct workqueue_struct *wq)
2783 struct wq_flusher this_flusher = {
2784 .list = LIST_HEAD_INIT(this_flusher.list),
2786 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2790 if (WARN_ON(!wq_online))
2793 lock_map_acquire(&wq->lockdep_map);
2794 lock_map_release(&wq->lockdep_map);
2796 mutex_lock(&wq->mutex);
2799 * Start-to-wait phase
2801 next_color = work_next_color(wq->work_color);
2803 if (next_color != wq->flush_color) {
2805 * Color space is not full. The current work_color
2806 * becomes our flush_color and work_color is advanced
2809 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2810 this_flusher.flush_color = wq->work_color;
2811 wq->work_color = next_color;
2813 if (!wq->first_flusher) {
2814 /* no flush in progress, become the first flusher */
2815 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2817 wq->first_flusher = &this_flusher;
2819 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2821 /* nothing to flush, done */
2822 wq->flush_color = next_color;
2823 wq->first_flusher = NULL;
2828 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2829 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2830 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2834 * Oops, color space is full, wait on overflow queue.
2835 * The next flush completion will assign us
2836 * flush_color and transfer to flusher_queue.
2838 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2841 check_flush_dependency(wq, NULL);
2843 mutex_unlock(&wq->mutex);
2845 wait_for_completion(&this_flusher.done);
2848 * Wake-up-and-cascade phase
2850 * First flushers are responsible for cascading flushes and
2851 * handling overflow. Non-first flushers can simply return.
2853 if (READ_ONCE(wq->first_flusher) != &this_flusher)
2856 mutex_lock(&wq->mutex);
2858 /* we might have raced, check again with mutex held */
2859 if (wq->first_flusher != &this_flusher)
2862 WRITE_ONCE(wq->first_flusher, NULL);
2864 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2865 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2868 struct wq_flusher *next, *tmp;
2870 /* complete all the flushers sharing the current flush color */
2871 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2872 if (next->flush_color != wq->flush_color)
2874 list_del_init(&next->list);
2875 complete(&next->done);
2878 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2879 wq->flush_color != work_next_color(wq->work_color));
2881 /* this flush_color is finished, advance by one */
2882 wq->flush_color = work_next_color(wq->flush_color);
2884 /* one color has been freed, handle overflow queue */
2885 if (!list_empty(&wq->flusher_overflow)) {
2887 * Assign the same color to all overflowed
2888 * flushers, advance work_color and append to
2889 * flusher_queue. This is the start-to-wait
2890 * phase for these overflowed flushers.
2892 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2893 tmp->flush_color = wq->work_color;
2895 wq->work_color = work_next_color(wq->work_color);
2897 list_splice_tail_init(&wq->flusher_overflow,
2898 &wq->flusher_queue);
2899 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2902 if (list_empty(&wq->flusher_queue)) {
2903 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2908 * Need to flush more colors. Make the next flusher
2909 * the new first flusher and arm pwqs.
2911 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2912 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2914 list_del_init(&next->list);
2915 wq->first_flusher = next;
2917 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2921 * Meh... this color is already done, clear first
2922 * flusher and repeat cascading.
2924 wq->first_flusher = NULL;
2928 mutex_unlock(&wq->mutex);
2930 EXPORT_SYMBOL(flush_workqueue);
2933 * drain_workqueue - drain a workqueue
2934 * @wq: workqueue to drain
2936 * Wait until the workqueue becomes empty. While draining is in progress,
2937 * only chain queueing is allowed. IOW, only currently pending or running
2938 * work items on @wq can queue further work items on it. @wq is flushed
2939 * repeatedly until it becomes empty. The number of flushing is determined
2940 * by the depth of chaining and should be relatively short. Whine if it
2943 void drain_workqueue(struct workqueue_struct *wq)
2945 unsigned int flush_cnt = 0;
2946 struct pool_workqueue *pwq;
2949 * __queue_work() needs to test whether there are drainers, is much
2950 * hotter than drain_workqueue() and already looks at @wq->flags.
2951 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2953 mutex_lock(&wq->mutex);
2954 if (!wq->nr_drainers++)
2955 wq->flags |= __WQ_DRAINING;
2956 mutex_unlock(&wq->mutex);
2958 flush_workqueue(wq);
2960 mutex_lock(&wq->mutex);
2962 for_each_pwq(pwq, wq) {
2965 raw_spin_lock_irq(&pwq->pool->lock);
2966 drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
2967 raw_spin_unlock_irq(&pwq->pool->lock);
2972 if (++flush_cnt == 10 ||
2973 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2974 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2975 wq->name, flush_cnt);
2977 mutex_unlock(&wq->mutex);
2981 if (!--wq->nr_drainers)
2982 wq->flags &= ~__WQ_DRAINING;
2983 mutex_unlock(&wq->mutex);
2985 EXPORT_SYMBOL_GPL(drain_workqueue);
2987 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2990 struct worker *worker = NULL;
2991 struct worker_pool *pool;
2992 struct pool_workqueue *pwq;
2997 pool = get_work_pool(work);
3003 raw_spin_lock_irq(&pool->lock);
3004 /* see the comment in try_to_grab_pending() with the same code */
3005 pwq = get_work_pwq(work);
3007 if (unlikely(pwq->pool != pool))
3010 worker = find_worker_executing_work(pool, work);
3013 pwq = worker->current_pwq;
3016 check_flush_dependency(pwq->wq, work);
3018 insert_wq_barrier(pwq, barr, work, worker);
3019 raw_spin_unlock_irq(&pool->lock);
3022 * Force a lock recursion deadlock when using flush_work() inside a
3023 * single-threaded or rescuer equipped workqueue.
3025 * For single threaded workqueues the deadlock happens when the work
3026 * is after the work issuing the flush_work(). For rescuer equipped
3027 * workqueues the deadlock happens when the rescuer stalls, blocking
3031 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3032 lock_map_acquire(&pwq->wq->lockdep_map);
3033 lock_map_release(&pwq->wq->lockdep_map);
3038 raw_spin_unlock_irq(&pool->lock);
3043 static bool __flush_work(struct work_struct *work, bool from_cancel)
3045 struct wq_barrier barr;
3047 if (WARN_ON(!wq_online))
3050 if (WARN_ON(!work->func))
3053 lock_map_acquire(&work->lockdep_map);
3054 lock_map_release(&work->lockdep_map);
3056 if (start_flush_work(work, &barr, from_cancel)) {
3057 wait_for_completion(&barr.done);
3058 destroy_work_on_stack(&barr.work);
3066 * flush_work - wait for a work to finish executing the last queueing instance
3067 * @work: the work to flush
3069 * Wait until @work has finished execution. @work is guaranteed to be idle
3070 * on return if it hasn't been requeued since flush started.
3073 * %true if flush_work() waited for the work to finish execution,
3074 * %false if it was already idle.
3076 bool flush_work(struct work_struct *work)
3078 return __flush_work(work, false);
3080 EXPORT_SYMBOL_GPL(flush_work);
3083 wait_queue_entry_t wait;
3084 struct work_struct *work;
3087 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3089 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3091 if (cwait->work != key)
3093 return autoremove_wake_function(wait, mode, sync, key);
3096 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3098 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3099 unsigned long flags;
3103 ret = try_to_grab_pending(work, is_dwork, &flags);
3105 * If someone else is already canceling, wait for it to
3106 * finish. flush_work() doesn't work for PREEMPT_NONE
3107 * because we may get scheduled between @work's completion
3108 * and the other canceling task resuming and clearing
3109 * CANCELING - flush_work() will return false immediately
3110 * as @work is no longer busy, try_to_grab_pending() will
3111 * return -ENOENT as @work is still being canceled and the
3112 * other canceling task won't be able to clear CANCELING as
3113 * we're hogging the CPU.
3115 * Let's wait for completion using a waitqueue. As this
3116 * may lead to the thundering herd problem, use a custom
3117 * wake function which matches @work along with exclusive
3120 if (unlikely(ret == -ENOENT)) {
3121 struct cwt_wait cwait;
3123 init_wait(&cwait.wait);
3124 cwait.wait.func = cwt_wakefn;
3127 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3128 TASK_UNINTERRUPTIBLE);
3129 if (work_is_canceling(work))
3131 finish_wait(&cancel_waitq, &cwait.wait);
3133 } while (unlikely(ret < 0));
3135 /* tell other tasks trying to grab @work to back off */
3136 mark_work_canceling(work);
3137 local_irq_restore(flags);
3140 * This allows canceling during early boot. We know that @work
3144 __flush_work(work, true);
3146 clear_work_data(work);
3149 * Paired with prepare_to_wait() above so that either
3150 * waitqueue_active() is visible here or !work_is_canceling() is
3154 if (waitqueue_active(&cancel_waitq))
3155 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3161 * cancel_work_sync - cancel a work and wait for it to finish
3162 * @work: the work to cancel
3164 * Cancel @work and wait for its execution to finish. This function
3165 * can be used even if the work re-queues itself or migrates to
3166 * another workqueue. On return from this function, @work is
3167 * guaranteed to be not pending or executing on any CPU.
3169 * cancel_work_sync(&delayed_work->work) must not be used for
3170 * delayed_work's. Use cancel_delayed_work_sync() instead.
3172 * The caller must ensure that the workqueue on which @work was last
3173 * queued can't be destroyed before this function returns.
3176 * %true if @work was pending, %false otherwise.
3178 bool cancel_work_sync(struct work_struct *work)
3180 return __cancel_work_timer(work, false);
3182 EXPORT_SYMBOL_GPL(cancel_work_sync);
3185 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3186 * @dwork: the delayed work to flush
3188 * Delayed timer is cancelled and the pending work is queued for
3189 * immediate execution. Like flush_work(), this function only
3190 * considers the last queueing instance of @dwork.
3193 * %true if flush_work() waited for the work to finish execution,
3194 * %false if it was already idle.
3196 bool flush_delayed_work(struct delayed_work *dwork)
3198 local_irq_disable();
3199 if (del_timer_sync(&dwork->timer))
3200 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3202 return flush_work(&dwork->work);
3204 EXPORT_SYMBOL(flush_delayed_work);
3207 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3208 * @rwork: the rcu work to flush
3211 * %true if flush_rcu_work() waited for the work to finish execution,
3212 * %false if it was already idle.
3214 bool flush_rcu_work(struct rcu_work *rwork)
3216 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3218 flush_work(&rwork->work);
3221 return flush_work(&rwork->work);
3224 EXPORT_SYMBOL(flush_rcu_work);
3226 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3228 unsigned long flags;
3232 ret = try_to_grab_pending(work, is_dwork, &flags);
3233 } while (unlikely(ret == -EAGAIN));
3235 if (unlikely(ret < 0))
3238 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3239 local_irq_restore(flags);
3244 * See cancel_delayed_work()
3246 bool cancel_work(struct work_struct *work)
3248 return __cancel_work(work, false);
3250 EXPORT_SYMBOL(cancel_work);
3253 * cancel_delayed_work - cancel a delayed work
3254 * @dwork: delayed_work to cancel
3256 * Kill off a pending delayed_work.
3258 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3262 * The work callback function may still be running on return, unless
3263 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3264 * use cancel_delayed_work_sync() to wait on it.
3266 * This function is safe to call from any context including IRQ handler.
3268 bool cancel_delayed_work(struct delayed_work *dwork)
3270 return __cancel_work(&dwork->work, true);
3272 EXPORT_SYMBOL(cancel_delayed_work);
3275 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3276 * @dwork: the delayed work cancel
3278 * This is cancel_work_sync() for delayed works.
3281 * %true if @dwork was pending, %false otherwise.
3283 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3285 return __cancel_work_timer(&dwork->work, true);
3287 EXPORT_SYMBOL(cancel_delayed_work_sync);
3290 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3291 * @func: the function to call
3293 * schedule_on_each_cpu() executes @func on each online CPU using the
3294 * system workqueue and blocks until all CPUs have completed.
3295 * schedule_on_each_cpu() is very slow.
3298 * 0 on success, -errno on failure.
3300 int schedule_on_each_cpu(work_func_t func)
3303 struct work_struct __percpu *works;
3305 works = alloc_percpu(struct work_struct);
3311 for_each_online_cpu(cpu) {
3312 struct work_struct *work = per_cpu_ptr(works, cpu);
3314 INIT_WORK(work, func);
3315 schedule_work_on(cpu, work);
3318 for_each_online_cpu(cpu)
3319 flush_work(per_cpu_ptr(works, cpu));
3327 * execute_in_process_context - reliably execute the routine with user context
3328 * @fn: the function to execute
3329 * @ew: guaranteed storage for the execute work structure (must
3330 * be available when the work executes)
3332 * Executes the function immediately if process context is available,
3333 * otherwise schedules the function for delayed execution.
3335 * Return: 0 - function was executed
3336 * 1 - function was scheduled for execution
3338 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3340 if (!in_interrupt()) {
3345 INIT_WORK(&ew->work, fn);
3346 schedule_work(&ew->work);
3350 EXPORT_SYMBOL_GPL(execute_in_process_context);
3353 * free_workqueue_attrs - free a workqueue_attrs
3354 * @attrs: workqueue_attrs to free
3356 * Undo alloc_workqueue_attrs().
3358 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3361 free_cpumask_var(attrs->cpumask);
3367 * alloc_workqueue_attrs - allocate a workqueue_attrs
3369 * Allocate a new workqueue_attrs, initialize with default settings and
3372 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3374 struct workqueue_attrs *alloc_workqueue_attrs(void)
3376 struct workqueue_attrs *attrs;
3378 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3381 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3384 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3387 free_workqueue_attrs(attrs);
3391 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3392 const struct workqueue_attrs *from)
3394 to->nice = from->nice;
3395 cpumask_copy(to->cpumask, from->cpumask);
3397 * Unlike hash and equality test, this function doesn't ignore
3398 * ->no_numa as it is used for both pool and wq attrs. Instead,
3399 * get_unbound_pool() explicitly clears ->no_numa after copying.
3401 to->no_numa = from->no_numa;
3404 /* hash value of the content of @attr */
3405 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3409 hash = jhash_1word(attrs->nice, hash);
3410 hash = jhash(cpumask_bits(attrs->cpumask),
3411 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3415 /* content equality test */
3416 static bool wqattrs_equal(const struct workqueue_attrs *a,
3417 const struct workqueue_attrs *b)
3419 if (a->nice != b->nice)
3421 if (!cpumask_equal(a->cpumask, b->cpumask))
3427 * init_worker_pool - initialize a newly zalloc'd worker_pool
3428 * @pool: worker_pool to initialize
3430 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3432 * Return: 0 on success, -errno on failure. Even on failure, all fields
3433 * inside @pool proper are initialized and put_unbound_pool() can be called
3434 * on @pool safely to release it.
3436 static int init_worker_pool(struct worker_pool *pool)
3438 raw_spin_lock_init(&pool->lock);
3441 pool->node = NUMA_NO_NODE;
3442 pool->flags |= POOL_DISASSOCIATED;
3443 pool->watchdog_ts = jiffies;
3444 INIT_LIST_HEAD(&pool->worklist);
3445 INIT_LIST_HEAD(&pool->idle_list);
3446 hash_init(pool->busy_hash);
3448 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3450 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3452 INIT_LIST_HEAD(&pool->workers);
3454 ida_init(&pool->worker_ida);
3455 INIT_HLIST_NODE(&pool->hash_node);
3458 /* shouldn't fail above this point */
3459 pool->attrs = alloc_workqueue_attrs();
3465 #ifdef CONFIG_LOCKDEP
3466 static void wq_init_lockdep(struct workqueue_struct *wq)
3470 lockdep_register_key(&wq->key);
3471 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3473 lock_name = wq->name;
3475 wq->lock_name = lock_name;
3476 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3479 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3481 lockdep_unregister_key(&wq->key);
3484 static void wq_free_lockdep(struct workqueue_struct *wq)
3486 if (wq->lock_name != wq->name)
3487 kfree(wq->lock_name);
3490 static void wq_init_lockdep(struct workqueue_struct *wq)
3494 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3498 static void wq_free_lockdep(struct workqueue_struct *wq)
3503 static void rcu_free_wq(struct rcu_head *rcu)
3505 struct workqueue_struct *wq =
3506 container_of(rcu, struct workqueue_struct, rcu);
3508 wq_free_lockdep(wq);
3510 if (!(wq->flags & WQ_UNBOUND))
3511 free_percpu(wq->cpu_pwqs);
3513 free_workqueue_attrs(wq->unbound_attrs);
3518 static void rcu_free_pool(struct rcu_head *rcu)
3520 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3522 ida_destroy(&pool->worker_ida);
3523 free_workqueue_attrs(pool->attrs);
3527 /* This returns with the lock held on success (pool manager is inactive). */
3528 static bool wq_manager_inactive(struct worker_pool *pool)
3530 raw_spin_lock_irq(&pool->lock);
3532 if (pool->flags & POOL_MANAGER_ACTIVE) {
3533 raw_spin_unlock_irq(&pool->lock);
3540 * put_unbound_pool - put a worker_pool
3541 * @pool: worker_pool to put
3543 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3544 * safe manner. get_unbound_pool() calls this function on its failure path
3545 * and this function should be able to release pools which went through,
3546 * successfully or not, init_worker_pool().
3548 * Should be called with wq_pool_mutex held.
3550 static void put_unbound_pool(struct worker_pool *pool)
3552 DECLARE_COMPLETION_ONSTACK(detach_completion);
3553 struct worker *worker;
3555 lockdep_assert_held(&wq_pool_mutex);
3561 if (WARN_ON(!(pool->cpu < 0)) ||
3562 WARN_ON(!list_empty(&pool->worklist)))
3565 /* release id and unhash */
3567 idr_remove(&worker_pool_idr, pool->id);
3568 hash_del(&pool->hash_node);
3571 * Become the manager and destroy all workers. This prevents
3572 * @pool's workers from blocking on attach_mutex. We're the last
3573 * manager and @pool gets freed with the flag set.
3574 * Because of how wq_manager_inactive() works, we will hold the
3575 * spinlock after a successful wait.
3577 rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
3578 TASK_UNINTERRUPTIBLE);
3579 pool->flags |= POOL_MANAGER_ACTIVE;
3581 while ((worker = first_idle_worker(pool)))
3582 destroy_worker(worker);
3583 WARN_ON(pool->nr_workers || pool->nr_idle);
3584 raw_spin_unlock_irq(&pool->lock);
3586 mutex_lock(&wq_pool_attach_mutex);
3587 if (!list_empty(&pool->workers))
3588 pool->detach_completion = &detach_completion;
3589 mutex_unlock(&wq_pool_attach_mutex);
3591 if (pool->detach_completion)
3592 wait_for_completion(pool->detach_completion);
3594 /* shut down the timers */
3595 del_timer_sync(&pool->idle_timer);
3596 del_timer_sync(&pool->mayday_timer);
3598 /* RCU protected to allow dereferences from get_work_pool() */
3599 call_rcu(&pool->rcu, rcu_free_pool);
3603 * get_unbound_pool - get a worker_pool with the specified attributes
3604 * @attrs: the attributes of the worker_pool to get
3606 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3607 * reference count and return it. If there already is a matching
3608 * worker_pool, it will be used; otherwise, this function attempts to
3611 * Should be called with wq_pool_mutex held.
3613 * Return: On success, a worker_pool with the same attributes as @attrs.
3614 * On failure, %NULL.
3616 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3618 u32 hash = wqattrs_hash(attrs);
3619 struct worker_pool *pool;
3621 int target_node = NUMA_NO_NODE;
3623 lockdep_assert_held(&wq_pool_mutex);
3625 /* do we already have a matching pool? */
3626 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3627 if (wqattrs_equal(pool->attrs, attrs)) {
3633 /* if cpumask is contained inside a NUMA node, we belong to that node */
3634 if (wq_numa_enabled) {
3635 for_each_node(node) {
3636 if (cpumask_subset(attrs->cpumask,
3637 wq_numa_possible_cpumask[node])) {
3644 /* nope, create a new one */
3645 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3646 if (!pool || init_worker_pool(pool) < 0)
3649 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3650 copy_workqueue_attrs(pool->attrs, attrs);
3651 pool->node = target_node;
3654 * no_numa isn't a worker_pool attribute, always clear it. See
3655 * 'struct workqueue_attrs' comments for detail.
3657 pool->attrs->no_numa = false;
3659 if (worker_pool_assign_id(pool) < 0)
3662 /* create and start the initial worker */
3663 if (wq_online && !create_worker(pool))
3667 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3672 put_unbound_pool(pool);
3676 static void rcu_free_pwq(struct rcu_head *rcu)
3678 kmem_cache_free(pwq_cache,
3679 container_of(rcu, struct pool_workqueue, rcu));
3683 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3684 * and needs to be destroyed.
3686 static void pwq_unbound_release_workfn(struct work_struct *work)
3688 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3689 unbound_release_work);
3690 struct workqueue_struct *wq = pwq->wq;
3691 struct worker_pool *pool = pwq->pool;
3692 bool is_last = false;
3695 * when @pwq is not linked, it doesn't hold any reference to the
3696 * @wq, and @wq is invalid to access.
3698 if (!list_empty(&pwq->pwqs_node)) {
3699 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3702 mutex_lock(&wq->mutex);
3703 list_del_rcu(&pwq->pwqs_node);
3704 is_last = list_empty(&wq->pwqs);
3705 mutex_unlock(&wq->mutex);
3708 mutex_lock(&wq_pool_mutex);
3709 put_unbound_pool(pool);
3710 mutex_unlock(&wq_pool_mutex);
3712 call_rcu(&pwq->rcu, rcu_free_pwq);
3715 * If we're the last pwq going away, @wq is already dead and no one
3716 * is gonna access it anymore. Schedule RCU free.
3719 wq_unregister_lockdep(wq);
3720 call_rcu(&wq->rcu, rcu_free_wq);
3725 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3726 * @pwq: target pool_workqueue
3728 * If @pwq isn't freezing, set @pwq->max_active to the associated
3729 * workqueue's saved_max_active and activate inactive work items
3730 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3732 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3734 struct workqueue_struct *wq = pwq->wq;
3735 bool freezable = wq->flags & WQ_FREEZABLE;
3736 unsigned long flags;
3738 /* for @wq->saved_max_active */
3739 lockdep_assert_held(&wq->mutex);
3741 /* fast exit for non-freezable wqs */
3742 if (!freezable && pwq->max_active == wq->saved_max_active)
3745 /* this function can be called during early boot w/ irq disabled */
3746 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
3749 * During [un]freezing, the caller is responsible for ensuring that
3750 * this function is called at least once after @workqueue_freezing
3751 * is updated and visible.
3753 if (!freezable || !workqueue_freezing) {
3756 pwq->max_active = wq->saved_max_active;
3758 while (!list_empty(&pwq->inactive_works) &&
3759 pwq->nr_active < pwq->max_active) {
3760 pwq_activate_first_inactive(pwq);
3765 * Need to kick a worker after thawed or an unbound wq's
3766 * max_active is bumped. In realtime scenarios, always kicking a
3767 * worker will cause interference on the isolated cpu cores, so
3768 * let's kick iff work items were activated.
3771 wake_up_worker(pwq->pool);
3773 pwq->max_active = 0;
3776 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
3779 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3780 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3781 struct worker_pool *pool)
3783 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3785 memset(pwq, 0, sizeof(*pwq));
3789 pwq->flush_color = -1;
3791 INIT_LIST_HEAD(&pwq->inactive_works);
3792 INIT_LIST_HEAD(&pwq->pwqs_node);
3793 INIT_LIST_HEAD(&pwq->mayday_node);
3794 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3797 /* sync @pwq with the current state of its associated wq and link it */
3798 static void link_pwq(struct pool_workqueue *pwq)
3800 struct workqueue_struct *wq = pwq->wq;
3802 lockdep_assert_held(&wq->mutex);
3804 /* may be called multiple times, ignore if already linked */
3805 if (!list_empty(&pwq->pwqs_node))
3808 /* set the matching work_color */
3809 pwq->work_color = wq->work_color;
3811 /* sync max_active to the current setting */
3812 pwq_adjust_max_active(pwq);
3815 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3818 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3819 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3820 const struct workqueue_attrs *attrs)
3822 struct worker_pool *pool;
3823 struct pool_workqueue *pwq;
3825 lockdep_assert_held(&wq_pool_mutex);
3827 pool = get_unbound_pool(attrs);
3831 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3833 put_unbound_pool(pool);
3837 init_pwq(pwq, wq, pool);
3842 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3843 * @attrs: the wq_attrs of the default pwq of the target workqueue
3844 * @node: the target NUMA node
3845 * @cpu_going_down: if >= 0, the CPU to consider as offline
3846 * @cpumask: outarg, the resulting cpumask
3848 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3849 * @cpu_going_down is >= 0, that cpu is considered offline during
3850 * calculation. The result is stored in @cpumask.
3852 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3853 * enabled and @node has online CPUs requested by @attrs, the returned
3854 * cpumask is the intersection of the possible CPUs of @node and
3857 * The caller is responsible for ensuring that the cpumask of @node stays
3860 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3863 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3864 int cpu_going_down, cpumask_t *cpumask)
3866 if (!wq_numa_enabled || attrs->no_numa)
3869 /* does @node have any online CPUs @attrs wants? */
3870 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3871 if (cpu_going_down >= 0)
3872 cpumask_clear_cpu(cpu_going_down, cpumask);
3874 if (cpumask_empty(cpumask))
3877 /* yeap, return possible CPUs in @node that @attrs wants */
3878 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3880 if (cpumask_empty(cpumask)) {
3881 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3882 "possible intersect\n");
3886 return !cpumask_equal(cpumask, attrs->cpumask);
3889 cpumask_copy(cpumask, attrs->cpumask);
3893 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3894 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3896 struct pool_workqueue *pwq)
3898 struct pool_workqueue *old_pwq;
3900 lockdep_assert_held(&wq_pool_mutex);
3901 lockdep_assert_held(&wq->mutex);
3903 /* link_pwq() can handle duplicate calls */
3906 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3907 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3911 /* context to store the prepared attrs & pwqs before applying */
3912 struct apply_wqattrs_ctx {
3913 struct workqueue_struct *wq; /* target workqueue */
3914 struct workqueue_attrs *attrs; /* attrs to apply */
3915 struct list_head list; /* queued for batching commit */
3916 struct pool_workqueue *dfl_pwq;
3917 struct pool_workqueue *pwq_tbl[];
3920 /* free the resources after success or abort */
3921 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3927 put_pwq_unlocked(ctx->pwq_tbl[node]);
3928 put_pwq_unlocked(ctx->dfl_pwq);
3930 free_workqueue_attrs(ctx->attrs);
3936 /* allocate the attrs and pwqs for later installation */
3937 static struct apply_wqattrs_ctx *
3938 apply_wqattrs_prepare(struct workqueue_struct *wq,
3939 const struct workqueue_attrs *attrs)
3941 struct apply_wqattrs_ctx *ctx;
3942 struct workqueue_attrs *new_attrs, *tmp_attrs;
3945 lockdep_assert_held(&wq_pool_mutex);
3947 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3949 new_attrs = alloc_workqueue_attrs();
3950 tmp_attrs = alloc_workqueue_attrs();
3951 if (!ctx || !new_attrs || !tmp_attrs)
3955 * Calculate the attrs of the default pwq.
3956 * If the user configured cpumask doesn't overlap with the
3957 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3959 copy_workqueue_attrs(new_attrs, attrs);
3960 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3961 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3962 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3965 * We may create multiple pwqs with differing cpumasks. Make a
3966 * copy of @new_attrs which will be modified and used to obtain
3969 copy_workqueue_attrs(tmp_attrs, new_attrs);
3972 * If something goes wrong during CPU up/down, we'll fall back to
3973 * the default pwq covering whole @attrs->cpumask. Always create
3974 * it even if we don't use it immediately.
3976 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3980 for_each_node(node) {
3981 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3982 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3983 if (!ctx->pwq_tbl[node])
3986 ctx->dfl_pwq->refcnt++;
3987 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3991 /* save the user configured attrs and sanitize it. */
3992 copy_workqueue_attrs(new_attrs, attrs);
3993 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3994 ctx->attrs = new_attrs;
3997 free_workqueue_attrs(tmp_attrs);
4001 free_workqueue_attrs(tmp_attrs);
4002 free_workqueue_attrs(new_attrs);
4003 apply_wqattrs_cleanup(ctx);
4007 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4008 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4012 /* all pwqs have been created successfully, let's install'em */
4013 mutex_lock(&ctx->wq->mutex);
4015 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4017 /* save the previous pwq and install the new one */
4019 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
4020 ctx->pwq_tbl[node]);
4022 /* @dfl_pwq might not have been used, ensure it's linked */
4023 link_pwq(ctx->dfl_pwq);
4024 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4026 mutex_unlock(&ctx->wq->mutex);
4029 static void apply_wqattrs_lock(void)
4031 /* CPUs should stay stable across pwq creations and installations */
4033 mutex_lock(&wq_pool_mutex);
4036 static void apply_wqattrs_unlock(void)
4038 mutex_unlock(&wq_pool_mutex);
4042 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4043 const struct workqueue_attrs *attrs)
4045 struct apply_wqattrs_ctx *ctx;
4047 /* only unbound workqueues can change attributes */
4048 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4051 /* creating multiple pwqs breaks ordering guarantee */
4052 if (!list_empty(&wq->pwqs)) {
4053 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4056 wq->flags &= ~__WQ_ORDERED;
4059 ctx = apply_wqattrs_prepare(wq, attrs);
4063 /* the ctx has been prepared successfully, let's commit it */
4064 apply_wqattrs_commit(ctx);
4065 apply_wqattrs_cleanup(ctx);
4071 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4072 * @wq: the target workqueue
4073 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4075 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4076 * machines, this function maps a separate pwq to each NUMA node with
4077 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4078 * NUMA node it was issued on. Older pwqs are released as in-flight work
4079 * items finish. Note that a work item which repeatedly requeues itself
4080 * back-to-back will stay on its current pwq.
4082 * Performs GFP_KERNEL allocations.
4084 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4086 * Return: 0 on success and -errno on failure.
4088 int apply_workqueue_attrs(struct workqueue_struct *wq,
4089 const struct workqueue_attrs *attrs)
4093 lockdep_assert_cpus_held();
4095 mutex_lock(&wq_pool_mutex);
4096 ret = apply_workqueue_attrs_locked(wq, attrs);
4097 mutex_unlock(&wq_pool_mutex);
4103 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4104 * @wq: the target workqueue
4105 * @cpu: the CPU coming up or going down
4106 * @online: whether @cpu is coming up or going down
4108 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4109 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4112 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4113 * falls back to @wq->dfl_pwq which may not be optimal but is always
4116 * Note that when the last allowed CPU of a NUMA node goes offline for a
4117 * workqueue with a cpumask spanning multiple nodes, the workers which were
4118 * already executing the work items for the workqueue will lose their CPU
4119 * affinity and may execute on any CPU. This is similar to how per-cpu
4120 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4121 * affinity, it's the user's responsibility to flush the work item from
4124 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4127 int node = cpu_to_node(cpu);
4128 int cpu_off = online ? -1 : cpu;
4129 struct pool_workqueue *old_pwq = NULL, *pwq;
4130 struct workqueue_attrs *target_attrs;
4133 lockdep_assert_held(&wq_pool_mutex);
4135 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4136 wq->unbound_attrs->no_numa)
4140 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4141 * Let's use a preallocated one. The following buf is protected by
4142 * CPU hotplug exclusion.
4144 target_attrs = wq_update_unbound_numa_attrs_buf;
4145 cpumask = target_attrs->cpumask;
4147 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4148 pwq = unbound_pwq_by_node(wq, node);
4151 * Let's determine what needs to be done. If the target cpumask is
4152 * different from the default pwq's, we need to compare it to @pwq's
4153 * and create a new one if they don't match. If the target cpumask
4154 * equals the default pwq's, the default pwq should be used.
4156 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4157 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4163 /* create a new pwq */
4164 pwq = alloc_unbound_pwq(wq, target_attrs);
4166 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4171 /* Install the new pwq. */
4172 mutex_lock(&wq->mutex);
4173 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4177 mutex_lock(&wq->mutex);
4178 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4179 get_pwq(wq->dfl_pwq);
4180 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4181 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4183 mutex_unlock(&wq->mutex);
4184 put_pwq_unlocked(old_pwq);
4187 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4189 bool highpri = wq->flags & WQ_HIGHPRI;
4192 if (!(wq->flags & WQ_UNBOUND)) {
4193 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4197 for_each_possible_cpu(cpu) {
4198 struct pool_workqueue *pwq =
4199 per_cpu_ptr(wq->cpu_pwqs, cpu);
4200 struct worker_pool *cpu_pools =
4201 per_cpu(cpu_worker_pools, cpu);
4203 init_pwq(pwq, wq, &cpu_pools[highpri]);
4205 mutex_lock(&wq->mutex);
4207 mutex_unlock(&wq->mutex);
4213 if (wq->flags & __WQ_ORDERED) {
4214 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4215 /* there should only be single pwq for ordering guarantee */
4216 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4217 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4218 "ordering guarantee broken for workqueue %s\n", wq->name);
4220 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4227 static int wq_clamp_max_active(int max_active, unsigned int flags,
4230 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4232 if (max_active < 1 || max_active > lim)
4233 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4234 max_active, name, 1, lim);
4236 return clamp_val(max_active, 1, lim);
4240 * Workqueues which may be used during memory reclaim should have a rescuer
4241 * to guarantee forward progress.
4243 static int init_rescuer(struct workqueue_struct *wq)
4245 struct worker *rescuer;
4248 if (!(wq->flags & WQ_MEM_RECLAIM))
4251 rescuer = alloc_worker(NUMA_NO_NODE);
4255 rescuer->rescue_wq = wq;
4256 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4257 if (IS_ERR(rescuer->task)) {
4258 ret = PTR_ERR(rescuer->task);
4263 wq->rescuer = rescuer;
4264 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4265 wake_up_process(rescuer->task);
4271 struct workqueue_struct *alloc_workqueue(const char *fmt,
4273 int max_active, ...)
4275 size_t tbl_size = 0;
4277 struct workqueue_struct *wq;
4278 struct pool_workqueue *pwq;
4281 * Unbound && max_active == 1 used to imply ordered, which is no
4282 * longer the case on NUMA machines due to per-node pools. While
4283 * alloc_ordered_workqueue() is the right way to create an ordered
4284 * workqueue, keep the previous behavior to avoid subtle breakages
4287 if ((flags & WQ_UNBOUND) && max_active == 1)
4288 flags |= __WQ_ORDERED;
4290 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4291 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4292 flags |= WQ_UNBOUND;
4294 /* allocate wq and format name */
4295 if (flags & WQ_UNBOUND)
4296 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4298 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4302 if (flags & WQ_UNBOUND) {
4303 wq->unbound_attrs = alloc_workqueue_attrs();
4304 if (!wq->unbound_attrs)
4308 va_start(args, max_active);
4309 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4312 max_active = max_active ?: WQ_DFL_ACTIVE;
4313 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4317 wq->saved_max_active = max_active;
4318 mutex_init(&wq->mutex);
4319 atomic_set(&wq->nr_pwqs_to_flush, 0);
4320 INIT_LIST_HEAD(&wq->pwqs);
4321 INIT_LIST_HEAD(&wq->flusher_queue);
4322 INIT_LIST_HEAD(&wq->flusher_overflow);
4323 INIT_LIST_HEAD(&wq->maydays);
4325 wq_init_lockdep(wq);
4326 INIT_LIST_HEAD(&wq->list);
4328 if (alloc_and_link_pwqs(wq) < 0)
4329 goto err_unreg_lockdep;
4331 if (wq_online && init_rescuer(wq) < 0)
4334 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4338 * wq_pool_mutex protects global freeze state and workqueues list.
4339 * Grab it, adjust max_active and add the new @wq to workqueues
4342 mutex_lock(&wq_pool_mutex);
4344 mutex_lock(&wq->mutex);
4345 for_each_pwq(pwq, wq)
4346 pwq_adjust_max_active(pwq);
4347 mutex_unlock(&wq->mutex);
4349 list_add_tail_rcu(&wq->list, &workqueues);
4351 mutex_unlock(&wq_pool_mutex);
4356 wq_unregister_lockdep(wq);
4357 wq_free_lockdep(wq);
4359 free_workqueue_attrs(wq->unbound_attrs);
4363 destroy_workqueue(wq);
4366 EXPORT_SYMBOL_GPL(alloc_workqueue);
4368 static bool pwq_busy(struct pool_workqueue *pwq)
4372 for (i = 0; i < WORK_NR_COLORS; i++)
4373 if (pwq->nr_in_flight[i])
4376 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4378 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4385 * destroy_workqueue - safely terminate a workqueue
4386 * @wq: target workqueue
4388 * Safely destroy a workqueue. All work currently pending will be done first.
4390 void destroy_workqueue(struct workqueue_struct *wq)
4392 struct pool_workqueue *pwq;
4396 * Remove it from sysfs first so that sanity check failure doesn't
4397 * lead to sysfs name conflicts.
4399 workqueue_sysfs_unregister(wq);
4401 /* drain it before proceeding with destruction */
4402 drain_workqueue(wq);
4404 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4406 struct worker *rescuer = wq->rescuer;
4408 /* this prevents new queueing */
4409 raw_spin_lock_irq(&wq_mayday_lock);
4411 raw_spin_unlock_irq(&wq_mayday_lock);
4413 /* rescuer will empty maydays list before exiting */
4414 kthread_stop(rescuer->task);
4419 * Sanity checks - grab all the locks so that we wait for all
4420 * in-flight operations which may do put_pwq().
4422 mutex_lock(&wq_pool_mutex);
4423 mutex_lock(&wq->mutex);
4424 for_each_pwq(pwq, wq) {
4425 raw_spin_lock_irq(&pwq->pool->lock);
4426 if (WARN_ON(pwq_busy(pwq))) {
4427 pr_warn("%s: %s has the following busy pwq\n",
4428 __func__, wq->name);
4430 raw_spin_unlock_irq(&pwq->pool->lock);
4431 mutex_unlock(&wq->mutex);
4432 mutex_unlock(&wq_pool_mutex);
4433 show_workqueue_state();
4436 raw_spin_unlock_irq(&pwq->pool->lock);
4438 mutex_unlock(&wq->mutex);
4441 * wq list is used to freeze wq, remove from list after
4442 * flushing is complete in case freeze races us.
4444 list_del_rcu(&wq->list);
4445 mutex_unlock(&wq_pool_mutex);
4447 if (!(wq->flags & WQ_UNBOUND)) {
4448 wq_unregister_lockdep(wq);
4450 * The base ref is never dropped on per-cpu pwqs. Directly
4451 * schedule RCU free.
4453 call_rcu(&wq->rcu, rcu_free_wq);
4456 * We're the sole accessor of @wq at this point. Directly
4457 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4458 * @wq will be freed when the last pwq is released.
4460 for_each_node(node) {
4461 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4462 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4463 put_pwq_unlocked(pwq);
4467 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4468 * put. Don't access it afterwards.
4472 put_pwq_unlocked(pwq);
4475 EXPORT_SYMBOL_GPL(destroy_workqueue);
4478 * workqueue_set_max_active - adjust max_active of a workqueue
4479 * @wq: target workqueue
4480 * @max_active: new max_active value.
4482 * Set max_active of @wq to @max_active.
4485 * Don't call from IRQ context.
4487 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4489 struct pool_workqueue *pwq;
4491 /* disallow meddling with max_active for ordered workqueues */
4492 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4495 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4497 mutex_lock(&wq->mutex);
4499 wq->flags &= ~__WQ_ORDERED;
4500 wq->saved_max_active = max_active;
4502 for_each_pwq(pwq, wq)
4503 pwq_adjust_max_active(pwq);
4505 mutex_unlock(&wq->mutex);
4507 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4510 * current_work - retrieve %current task's work struct
4512 * Determine if %current task is a workqueue worker and what it's working on.
4513 * Useful to find out the context that the %current task is running in.
4515 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4517 struct work_struct *current_work(void)
4519 struct worker *worker = current_wq_worker();
4521 return worker ? worker->current_work : NULL;
4523 EXPORT_SYMBOL(current_work);
4526 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4528 * Determine whether %current is a workqueue rescuer. Can be used from
4529 * work functions to determine whether it's being run off the rescuer task.
4531 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4533 bool current_is_workqueue_rescuer(void)
4535 struct worker *worker = current_wq_worker();
4537 return worker && worker->rescue_wq;
4541 * workqueue_congested - test whether a workqueue is congested
4542 * @cpu: CPU in question
4543 * @wq: target workqueue
4545 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4546 * no synchronization around this function and the test result is
4547 * unreliable and only useful as advisory hints or for debugging.
4549 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4550 * Note that both per-cpu and unbound workqueues may be associated with
4551 * multiple pool_workqueues which have separate congested states. A
4552 * workqueue being congested on one CPU doesn't mean the workqueue is also
4553 * contested on other CPUs / NUMA nodes.
4556 * %true if congested, %false otherwise.
4558 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4560 struct pool_workqueue *pwq;
4566 if (cpu == WORK_CPU_UNBOUND)
4567 cpu = smp_processor_id();
4569 if (!(wq->flags & WQ_UNBOUND))
4570 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4572 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4574 ret = !list_empty(&pwq->inactive_works);
4580 EXPORT_SYMBOL_GPL(workqueue_congested);
4583 * work_busy - test whether a work is currently pending or running
4584 * @work: the work to be tested
4586 * Test whether @work is currently pending or running. There is no
4587 * synchronization around this function and the test result is
4588 * unreliable and only useful as advisory hints or for debugging.
4591 * OR'd bitmask of WORK_BUSY_* bits.
4593 unsigned int work_busy(struct work_struct *work)
4595 struct worker_pool *pool;
4596 unsigned long flags;
4597 unsigned int ret = 0;
4599 if (work_pending(work))
4600 ret |= WORK_BUSY_PENDING;
4603 pool = get_work_pool(work);
4605 raw_spin_lock_irqsave(&pool->lock, flags);
4606 if (find_worker_executing_work(pool, work))
4607 ret |= WORK_BUSY_RUNNING;
4608 raw_spin_unlock_irqrestore(&pool->lock, flags);
4614 EXPORT_SYMBOL_GPL(work_busy);
4617 * set_worker_desc - set description for the current work item
4618 * @fmt: printf-style format string
4619 * @...: arguments for the format string
4621 * This function can be called by a running work function to describe what
4622 * the work item is about. If the worker task gets dumped, this
4623 * information will be printed out together to help debugging. The
4624 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4626 void set_worker_desc(const char *fmt, ...)
4628 struct worker *worker = current_wq_worker();
4632 va_start(args, fmt);
4633 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4637 EXPORT_SYMBOL_GPL(set_worker_desc);
4640 * print_worker_info - print out worker information and description
4641 * @log_lvl: the log level to use when printing
4642 * @task: target task
4644 * If @task is a worker and currently executing a work item, print out the
4645 * name of the workqueue being serviced and worker description set with
4646 * set_worker_desc() by the currently executing work item.
4648 * This function can be safely called on any task as long as the
4649 * task_struct itself is accessible. While safe, this function isn't
4650 * synchronized and may print out mixups or garbages of limited length.
4652 void print_worker_info(const char *log_lvl, struct task_struct *task)
4654 work_func_t *fn = NULL;
4655 char name[WQ_NAME_LEN] = { };
4656 char desc[WORKER_DESC_LEN] = { };
4657 struct pool_workqueue *pwq = NULL;
4658 struct workqueue_struct *wq = NULL;
4659 struct worker *worker;
4661 if (!(task->flags & PF_WQ_WORKER))
4665 * This function is called without any synchronization and @task
4666 * could be in any state. Be careful with dereferences.
4668 worker = kthread_probe_data(task);
4671 * Carefully copy the associated workqueue's workfn, name and desc.
4672 * Keep the original last '\0' in case the original is garbage.
4674 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
4675 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
4676 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
4677 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
4678 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
4680 if (fn || name[0] || desc[0]) {
4681 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4682 if (strcmp(name, desc))
4683 pr_cont(" (%s)", desc);
4688 static void pr_cont_pool_info(struct worker_pool *pool)
4690 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4691 if (pool->node != NUMA_NO_NODE)
4692 pr_cont(" node=%d", pool->node);
4693 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4696 static void pr_cont_work(bool comma, struct work_struct *work)
4698 if (work->func == wq_barrier_func) {
4699 struct wq_barrier *barr;
4701 barr = container_of(work, struct wq_barrier, work);
4703 pr_cont("%s BAR(%d)", comma ? "," : "",
4704 task_pid_nr(barr->task));
4706 pr_cont("%s %ps", comma ? "," : "", work->func);
4710 static void show_pwq(struct pool_workqueue *pwq)
4712 struct worker_pool *pool = pwq->pool;
4713 struct work_struct *work;
4714 struct worker *worker;
4715 bool has_in_flight = false, has_pending = false;
4718 pr_info(" pwq %d:", pool->id);
4719 pr_cont_pool_info(pool);
4721 pr_cont(" active=%d/%d refcnt=%d%s\n",
4722 pwq->nr_active, pwq->max_active, pwq->refcnt,
4723 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4725 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4726 if (worker->current_pwq == pwq) {
4727 has_in_flight = true;
4731 if (has_in_flight) {
4734 pr_info(" in-flight:");
4735 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4736 if (worker->current_pwq != pwq)
4739 pr_cont("%s %d%s:%ps", comma ? "," : "",
4740 task_pid_nr(worker->task),
4741 worker->rescue_wq ? "(RESCUER)" : "",
4742 worker->current_func);
4743 list_for_each_entry(work, &worker->scheduled, entry)
4744 pr_cont_work(false, work);
4750 list_for_each_entry(work, &pool->worklist, entry) {
4751 if (get_work_pwq(work) == pwq) {
4759 pr_info(" pending:");
4760 list_for_each_entry(work, &pool->worklist, entry) {
4761 if (get_work_pwq(work) != pwq)
4764 pr_cont_work(comma, work);
4765 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4770 if (!list_empty(&pwq->inactive_works)) {
4773 pr_info(" inactive:");
4774 list_for_each_entry(work, &pwq->inactive_works, entry) {
4775 pr_cont_work(comma, work);
4776 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4783 * show_workqueue_state - dump workqueue state
4785 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4786 * all busy workqueues and pools.
4788 void show_workqueue_state(void)
4790 struct workqueue_struct *wq;
4791 struct worker_pool *pool;
4792 unsigned long flags;
4797 pr_info("Showing busy workqueues and worker pools:\n");
4799 list_for_each_entry_rcu(wq, &workqueues, list) {
4800 struct pool_workqueue *pwq;
4803 for_each_pwq(pwq, wq) {
4804 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4812 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4814 for_each_pwq(pwq, wq) {
4815 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4816 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4818 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4820 * We could be printing a lot from atomic context, e.g.
4821 * sysrq-t -> show_workqueue_state(). Avoid triggering
4824 touch_nmi_watchdog();
4828 for_each_pool(pool, pi) {
4829 struct worker *worker;
4831 unsigned long hung = 0;
4833 raw_spin_lock_irqsave(&pool->lock, flags);
4834 if (pool->nr_workers == pool->nr_idle)
4837 /* How long the first pending work is waiting for a worker. */
4838 if (!list_empty(&pool->worklist))
4839 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
4841 pr_info("pool %d:", pool->id);
4842 pr_cont_pool_info(pool);
4843 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
4845 pr_cont(" manager: %d",
4846 task_pid_nr(pool->manager->task));
4847 list_for_each_entry(worker, &pool->idle_list, entry) {
4848 pr_cont(" %s%d", first ? "idle: " : "",
4849 task_pid_nr(worker->task));
4854 raw_spin_unlock_irqrestore(&pool->lock, flags);
4856 * We could be printing a lot from atomic context, e.g.
4857 * sysrq-t -> show_workqueue_state(). Avoid triggering
4860 touch_nmi_watchdog();
4866 /* used to show worker information through /proc/PID/{comm,stat,status} */
4867 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4871 /* always show the actual comm */
4872 off = strscpy(buf, task->comm, size);
4876 /* stabilize PF_WQ_WORKER and worker pool association */
4877 mutex_lock(&wq_pool_attach_mutex);
4879 if (task->flags & PF_WQ_WORKER) {
4880 struct worker *worker = kthread_data(task);
4881 struct worker_pool *pool = worker->pool;
4884 raw_spin_lock_irq(&pool->lock);
4886 * ->desc tracks information (wq name or
4887 * set_worker_desc()) for the latest execution. If
4888 * current, prepend '+', otherwise '-'.
4890 if (worker->desc[0] != '\0') {
4891 if (worker->current_work)
4892 scnprintf(buf + off, size - off, "+%s",
4895 scnprintf(buf + off, size - off, "-%s",
4898 raw_spin_unlock_irq(&pool->lock);
4902 mutex_unlock(&wq_pool_attach_mutex);
4910 * There are two challenges in supporting CPU hotplug. Firstly, there
4911 * are a lot of assumptions on strong associations among work, pwq and
4912 * pool which make migrating pending and scheduled works very
4913 * difficult to implement without impacting hot paths. Secondly,
4914 * worker pools serve mix of short, long and very long running works making
4915 * blocked draining impractical.
4917 * This is solved by allowing the pools to be disassociated from the CPU
4918 * running as an unbound one and allowing it to be reattached later if the
4919 * cpu comes back online.
4922 static void unbind_workers(int cpu)
4924 struct worker_pool *pool;
4925 struct worker *worker;
4927 for_each_cpu_worker_pool(pool, cpu) {
4928 mutex_lock(&wq_pool_attach_mutex);
4929 raw_spin_lock_irq(&pool->lock);
4932 * We've blocked all attach/detach operations. Make all workers
4933 * unbound and set DISASSOCIATED. Before this, all workers
4934 * except for the ones which are still executing works from
4935 * before the last CPU down must be on the cpu. After
4936 * this, they may become diasporas.
4938 for_each_pool_worker(worker, pool)
4939 worker->flags |= WORKER_UNBOUND;
4941 pool->flags |= POOL_DISASSOCIATED;
4943 raw_spin_unlock_irq(&pool->lock);
4944 mutex_unlock(&wq_pool_attach_mutex);
4947 * Call schedule() so that we cross rq->lock and thus can
4948 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4949 * This is necessary as scheduler callbacks may be invoked
4955 * Sched callbacks are disabled now. Zap nr_running.
4956 * After this, nr_running stays zero and need_more_worker()
4957 * and keep_working() are always true as long as the
4958 * worklist is not empty. This pool now behaves as an
4959 * unbound (in terms of concurrency management) pool which
4960 * are served by workers tied to the pool.
4962 atomic_set(&pool->nr_running, 0);
4965 * With concurrency management just turned off, a busy
4966 * worker blocking could lead to lengthy stalls. Kick off
4967 * unbound chain execution of currently pending work items.
4969 raw_spin_lock_irq(&pool->lock);
4970 wake_up_worker(pool);
4971 raw_spin_unlock_irq(&pool->lock);
4976 * rebind_workers - rebind all workers of a pool to the associated CPU
4977 * @pool: pool of interest
4979 * @pool->cpu is coming online. Rebind all workers to the CPU.
4981 static void rebind_workers(struct worker_pool *pool)
4983 struct worker *worker;
4985 lockdep_assert_held(&wq_pool_attach_mutex);
4988 * Restore CPU affinity of all workers. As all idle workers should
4989 * be on the run-queue of the associated CPU before any local
4990 * wake-ups for concurrency management happen, restore CPU affinity
4991 * of all workers first and then clear UNBOUND. As we're called
4992 * from CPU_ONLINE, the following shouldn't fail.
4994 for_each_pool_worker(worker, pool)
4995 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4996 pool->attrs->cpumask) < 0);
4998 raw_spin_lock_irq(&pool->lock);
5000 pool->flags &= ~POOL_DISASSOCIATED;
5002 for_each_pool_worker(worker, pool) {
5003 unsigned int worker_flags = worker->flags;
5006 * A bound idle worker should actually be on the runqueue
5007 * of the associated CPU for local wake-ups targeting it to
5008 * work. Kick all idle workers so that they migrate to the
5009 * associated CPU. Doing this in the same loop as
5010 * replacing UNBOUND with REBOUND is safe as no worker will
5011 * be bound before @pool->lock is released.
5013 if (worker_flags & WORKER_IDLE)
5014 wake_up_process(worker->task);
5017 * We want to clear UNBOUND but can't directly call
5018 * worker_clr_flags() or adjust nr_running. Atomically
5019 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5020 * @worker will clear REBOUND using worker_clr_flags() when
5021 * it initiates the next execution cycle thus restoring
5022 * concurrency management. Note that when or whether
5023 * @worker clears REBOUND doesn't affect correctness.
5025 * WRITE_ONCE() is necessary because @worker->flags may be
5026 * tested without holding any lock in
5027 * wq_worker_running(). Without it, NOT_RUNNING test may
5028 * fail incorrectly leading to premature concurrency
5029 * management operations.
5031 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5032 worker_flags |= WORKER_REBOUND;
5033 worker_flags &= ~WORKER_UNBOUND;
5034 WRITE_ONCE(worker->flags, worker_flags);
5037 raw_spin_unlock_irq(&pool->lock);
5041 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5042 * @pool: unbound pool of interest
5043 * @cpu: the CPU which is coming up
5045 * An unbound pool may end up with a cpumask which doesn't have any online
5046 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5047 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5048 * online CPU before, cpus_allowed of all its workers should be restored.
5050 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5052 static cpumask_t cpumask;
5053 struct worker *worker;
5055 lockdep_assert_held(&wq_pool_attach_mutex);
5057 /* is @cpu allowed for @pool? */
5058 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5061 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5063 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5064 for_each_pool_worker(worker, pool)
5065 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5068 int workqueue_prepare_cpu(unsigned int cpu)
5070 struct worker_pool *pool;
5072 for_each_cpu_worker_pool(pool, cpu) {
5073 if (pool->nr_workers)
5075 if (!create_worker(pool))
5081 int workqueue_online_cpu(unsigned int cpu)
5083 struct worker_pool *pool;
5084 struct workqueue_struct *wq;
5087 mutex_lock(&wq_pool_mutex);
5089 for_each_pool(pool, pi) {
5090 mutex_lock(&wq_pool_attach_mutex);
5092 if (pool->cpu == cpu)
5093 rebind_workers(pool);
5094 else if (pool->cpu < 0)
5095 restore_unbound_workers_cpumask(pool, cpu);
5097 mutex_unlock(&wq_pool_attach_mutex);
5100 /* update NUMA affinity of unbound workqueues */
5101 list_for_each_entry(wq, &workqueues, list)
5102 wq_update_unbound_numa(wq, cpu, true);
5104 mutex_unlock(&wq_pool_mutex);
5108 int workqueue_offline_cpu(unsigned int cpu)
5110 struct workqueue_struct *wq;
5112 /* unbinding per-cpu workers should happen on the local CPU */
5113 if (WARN_ON(cpu != smp_processor_id()))
5116 unbind_workers(cpu);
5118 /* update NUMA affinity of unbound workqueues */
5119 mutex_lock(&wq_pool_mutex);
5120 list_for_each_entry(wq, &workqueues, list)
5121 wq_update_unbound_numa(wq, cpu, false);
5122 mutex_unlock(&wq_pool_mutex);
5127 struct work_for_cpu {
5128 struct work_struct work;
5134 static void work_for_cpu_fn(struct work_struct *work)
5136 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5138 wfc->ret = wfc->fn(wfc->arg);
5142 * work_on_cpu - run a function in thread context on a particular cpu
5143 * @cpu: the cpu to run on
5144 * @fn: the function to run
5145 * @arg: the function arg
5147 * It is up to the caller to ensure that the cpu doesn't go offline.
5148 * The caller must not hold any locks which would prevent @fn from completing.
5150 * Return: The value @fn returns.
5152 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5154 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5156 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5157 schedule_work_on(cpu, &wfc.work);
5158 flush_work(&wfc.work);
5159 destroy_work_on_stack(&wfc.work);
5162 EXPORT_SYMBOL_GPL(work_on_cpu);
5165 * work_on_cpu_safe - run a function in thread context on a particular cpu
5166 * @cpu: the cpu to run on
5167 * @fn: the function to run
5168 * @arg: the function argument
5170 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5171 * any locks which would prevent @fn from completing.
5173 * Return: The value @fn returns.
5175 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5180 if (cpu_online(cpu))
5181 ret = work_on_cpu(cpu, fn, arg);
5185 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5186 #endif /* CONFIG_SMP */
5188 #ifdef CONFIG_FREEZER
5191 * freeze_workqueues_begin - begin freezing workqueues
5193 * Start freezing workqueues. After this function returns, all freezable
5194 * workqueues will queue new works to their inactive_works list instead of
5198 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5200 void freeze_workqueues_begin(void)
5202 struct workqueue_struct *wq;
5203 struct pool_workqueue *pwq;
5205 mutex_lock(&wq_pool_mutex);
5207 WARN_ON_ONCE(workqueue_freezing);
5208 workqueue_freezing = true;
5210 list_for_each_entry(wq, &workqueues, list) {
5211 mutex_lock(&wq->mutex);
5212 for_each_pwq(pwq, wq)
5213 pwq_adjust_max_active(pwq);
5214 mutex_unlock(&wq->mutex);
5217 mutex_unlock(&wq_pool_mutex);
5221 * freeze_workqueues_busy - are freezable workqueues still busy?
5223 * Check whether freezing is complete. This function must be called
5224 * between freeze_workqueues_begin() and thaw_workqueues().
5227 * Grabs and releases wq_pool_mutex.
5230 * %true if some freezable workqueues are still busy. %false if freezing
5233 bool freeze_workqueues_busy(void)
5236 struct workqueue_struct *wq;
5237 struct pool_workqueue *pwq;
5239 mutex_lock(&wq_pool_mutex);
5241 WARN_ON_ONCE(!workqueue_freezing);
5243 list_for_each_entry(wq, &workqueues, list) {
5244 if (!(wq->flags & WQ_FREEZABLE))
5247 * nr_active is monotonically decreasing. It's safe
5248 * to peek without lock.
5251 for_each_pwq(pwq, wq) {
5252 WARN_ON_ONCE(pwq->nr_active < 0);
5253 if (pwq->nr_active) {
5262 mutex_unlock(&wq_pool_mutex);
5267 * thaw_workqueues - thaw workqueues
5269 * Thaw workqueues. Normal queueing is restored and all collected
5270 * frozen works are transferred to their respective pool worklists.
5273 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5275 void thaw_workqueues(void)
5277 struct workqueue_struct *wq;
5278 struct pool_workqueue *pwq;
5280 mutex_lock(&wq_pool_mutex);
5282 if (!workqueue_freezing)
5285 workqueue_freezing = false;
5287 /* restore max_active and repopulate worklist */
5288 list_for_each_entry(wq, &workqueues, list) {
5289 mutex_lock(&wq->mutex);
5290 for_each_pwq(pwq, wq)
5291 pwq_adjust_max_active(pwq);
5292 mutex_unlock(&wq->mutex);
5296 mutex_unlock(&wq_pool_mutex);
5298 #endif /* CONFIG_FREEZER */
5300 static int workqueue_apply_unbound_cpumask(void)
5304 struct workqueue_struct *wq;
5305 struct apply_wqattrs_ctx *ctx, *n;
5307 lockdep_assert_held(&wq_pool_mutex);
5309 list_for_each_entry(wq, &workqueues, list) {
5310 if (!(wq->flags & WQ_UNBOUND))
5313 /* creating multiple pwqs breaks ordering guarantee */
5314 if (!list_empty(&wq->pwqs)) {
5315 if (wq->flags & __WQ_ORDERED_EXPLICIT)
5317 wq->flags &= ~__WQ_ORDERED;
5320 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5326 list_add_tail(&ctx->list, &ctxs);
5329 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5331 apply_wqattrs_commit(ctx);
5332 apply_wqattrs_cleanup(ctx);
5339 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5340 * @cpumask: the cpumask to set
5342 * The low-level workqueues cpumask is a global cpumask that limits
5343 * the affinity of all unbound workqueues. This function check the @cpumask
5344 * and apply it to all unbound workqueues and updates all pwqs of them.
5346 * Retun: 0 - Success
5347 * -EINVAL - Invalid @cpumask
5348 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5350 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5353 cpumask_var_t saved_cpumask;
5356 * Not excluding isolated cpus on purpose.
5357 * If the user wishes to include them, we allow that.
5359 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5360 if (!cpumask_empty(cpumask)) {
5361 apply_wqattrs_lock();
5362 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5367 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) {
5372 /* save the old wq_unbound_cpumask. */
5373 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5375 /* update wq_unbound_cpumask at first and apply it to wqs. */
5376 cpumask_copy(wq_unbound_cpumask, cpumask);
5377 ret = workqueue_apply_unbound_cpumask();
5379 /* restore the wq_unbound_cpumask when failed. */
5381 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5383 free_cpumask_var(saved_cpumask);
5385 apply_wqattrs_unlock();
5393 * Workqueues with WQ_SYSFS flag set is visible to userland via
5394 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5395 * following attributes.
5397 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5398 * max_active RW int : maximum number of in-flight work items
5400 * Unbound workqueues have the following extra attributes.
5402 * pool_ids RO int : the associated pool IDs for each node
5403 * nice RW int : nice value of the workers
5404 * cpumask RW mask : bitmask of allowed CPUs for the workers
5405 * numa RW bool : whether enable NUMA affinity
5408 struct workqueue_struct *wq;
5412 static struct workqueue_struct *dev_to_wq(struct device *dev)
5414 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5419 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5422 struct workqueue_struct *wq = dev_to_wq(dev);
5424 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5426 static DEVICE_ATTR_RO(per_cpu);
5428 static ssize_t max_active_show(struct device *dev,
5429 struct device_attribute *attr, char *buf)
5431 struct workqueue_struct *wq = dev_to_wq(dev);
5433 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5436 static ssize_t max_active_store(struct device *dev,
5437 struct device_attribute *attr, const char *buf,
5440 struct workqueue_struct *wq = dev_to_wq(dev);
5443 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5446 workqueue_set_max_active(wq, val);
5449 static DEVICE_ATTR_RW(max_active);
5451 static struct attribute *wq_sysfs_attrs[] = {
5452 &dev_attr_per_cpu.attr,
5453 &dev_attr_max_active.attr,
5456 ATTRIBUTE_GROUPS(wq_sysfs);
5458 static ssize_t wq_pool_ids_show(struct device *dev,
5459 struct device_attribute *attr, char *buf)
5461 struct workqueue_struct *wq = dev_to_wq(dev);
5462 const char *delim = "";
5463 int node, written = 0;
5467 for_each_node(node) {
5468 written += scnprintf(buf + written, PAGE_SIZE - written,
5469 "%s%d:%d", delim, node,
5470 unbound_pwq_by_node(wq, node)->pool->id);
5473 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5480 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5483 struct workqueue_struct *wq = dev_to_wq(dev);
5486 mutex_lock(&wq->mutex);
5487 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5488 mutex_unlock(&wq->mutex);
5493 /* prepare workqueue_attrs for sysfs store operations */
5494 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5496 struct workqueue_attrs *attrs;
5498 lockdep_assert_held(&wq_pool_mutex);
5500 attrs = alloc_workqueue_attrs();
5504 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5508 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5509 const char *buf, size_t count)
5511 struct workqueue_struct *wq = dev_to_wq(dev);
5512 struct workqueue_attrs *attrs;
5515 apply_wqattrs_lock();
5517 attrs = wq_sysfs_prep_attrs(wq);
5521 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5522 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5523 ret = apply_workqueue_attrs_locked(wq, attrs);
5528 apply_wqattrs_unlock();
5529 free_workqueue_attrs(attrs);
5530 return ret ?: count;
5533 static ssize_t wq_cpumask_show(struct device *dev,
5534 struct device_attribute *attr, char *buf)
5536 struct workqueue_struct *wq = dev_to_wq(dev);
5539 mutex_lock(&wq->mutex);
5540 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5541 cpumask_pr_args(wq->unbound_attrs->cpumask));
5542 mutex_unlock(&wq->mutex);
5546 static ssize_t wq_cpumask_store(struct device *dev,
5547 struct device_attribute *attr,
5548 const char *buf, size_t count)
5550 struct workqueue_struct *wq = dev_to_wq(dev);
5551 struct workqueue_attrs *attrs;
5554 apply_wqattrs_lock();
5556 attrs = wq_sysfs_prep_attrs(wq);
5560 ret = cpumask_parse(buf, attrs->cpumask);
5562 ret = apply_workqueue_attrs_locked(wq, attrs);
5565 apply_wqattrs_unlock();
5566 free_workqueue_attrs(attrs);
5567 return ret ?: count;
5570 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5573 struct workqueue_struct *wq = dev_to_wq(dev);
5576 mutex_lock(&wq->mutex);
5577 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5578 !wq->unbound_attrs->no_numa);
5579 mutex_unlock(&wq->mutex);
5584 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5585 const char *buf, size_t count)
5587 struct workqueue_struct *wq = dev_to_wq(dev);
5588 struct workqueue_attrs *attrs;
5589 int v, ret = -ENOMEM;
5591 apply_wqattrs_lock();
5593 attrs = wq_sysfs_prep_attrs(wq);
5598 if (sscanf(buf, "%d", &v) == 1) {
5599 attrs->no_numa = !v;
5600 ret = apply_workqueue_attrs_locked(wq, attrs);
5604 apply_wqattrs_unlock();
5605 free_workqueue_attrs(attrs);
5606 return ret ?: count;
5609 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5610 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5611 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5612 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5613 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5617 static struct bus_type wq_subsys = {
5618 .name = "workqueue",
5619 .dev_groups = wq_sysfs_groups,
5622 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5623 struct device_attribute *attr, char *buf)
5627 mutex_lock(&wq_pool_mutex);
5628 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5629 cpumask_pr_args(wq_unbound_cpumask));
5630 mutex_unlock(&wq_pool_mutex);
5635 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5636 struct device_attribute *attr, const char *buf, size_t count)
5638 cpumask_var_t cpumask;
5641 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5644 ret = cpumask_parse(buf, cpumask);
5646 ret = workqueue_set_unbound_cpumask(cpumask);
5648 free_cpumask_var(cpumask);
5649 return ret ? ret : count;
5652 static struct device_attribute wq_sysfs_cpumask_attr =
5653 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5654 wq_unbound_cpumask_store);
5656 static int __init wq_sysfs_init(void)
5660 err = subsys_virtual_register(&wq_subsys, NULL);
5664 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5666 core_initcall(wq_sysfs_init);
5668 static void wq_device_release(struct device *dev)
5670 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5676 * workqueue_sysfs_register - make a workqueue visible in sysfs
5677 * @wq: the workqueue to register
5679 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5680 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5681 * which is the preferred method.
5683 * Workqueue user should use this function directly iff it wants to apply
5684 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5685 * apply_workqueue_attrs() may race against userland updating the
5688 * Return: 0 on success, -errno on failure.
5690 int workqueue_sysfs_register(struct workqueue_struct *wq)
5692 struct wq_device *wq_dev;
5696 * Adjusting max_active or creating new pwqs by applying
5697 * attributes breaks ordering guarantee. Disallow exposing ordered
5700 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5703 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5708 wq_dev->dev.bus = &wq_subsys;
5709 wq_dev->dev.release = wq_device_release;
5710 dev_set_name(&wq_dev->dev, "%s", wq->name);
5713 * unbound_attrs are created separately. Suppress uevent until
5714 * everything is ready.
5716 dev_set_uevent_suppress(&wq_dev->dev, true);
5718 ret = device_register(&wq_dev->dev);
5720 put_device(&wq_dev->dev);
5725 if (wq->flags & WQ_UNBOUND) {
5726 struct device_attribute *attr;
5728 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5729 ret = device_create_file(&wq_dev->dev, attr);
5731 device_unregister(&wq_dev->dev);
5738 dev_set_uevent_suppress(&wq_dev->dev, false);
5739 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5744 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5745 * @wq: the workqueue to unregister
5747 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5749 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5751 struct wq_device *wq_dev = wq->wq_dev;
5757 device_unregister(&wq_dev->dev);
5759 #else /* CONFIG_SYSFS */
5760 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5761 #endif /* CONFIG_SYSFS */
5764 * Workqueue watchdog.
5766 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5767 * flush dependency, a concurrency managed work item which stays RUNNING
5768 * indefinitely. Workqueue stalls can be very difficult to debug as the
5769 * usual warning mechanisms don't trigger and internal workqueue state is
5772 * Workqueue watchdog monitors all worker pools periodically and dumps
5773 * state if some pools failed to make forward progress for a while where
5774 * forward progress is defined as the first item on ->worklist changing.
5776 * This mechanism is controlled through the kernel parameter
5777 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5778 * corresponding sysfs parameter file.
5780 #ifdef CONFIG_WQ_WATCHDOG
5782 static unsigned long wq_watchdog_thresh = 30;
5783 static struct timer_list wq_watchdog_timer;
5785 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5786 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5788 static void wq_watchdog_reset_touched(void)
5792 wq_watchdog_touched = jiffies;
5793 for_each_possible_cpu(cpu)
5794 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5797 static void wq_watchdog_timer_fn(struct timer_list *unused)
5799 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5800 bool lockup_detected = false;
5801 unsigned long now = jiffies;
5802 struct worker_pool *pool;
5810 for_each_pool(pool, pi) {
5811 unsigned long pool_ts, touched, ts;
5813 if (list_empty(&pool->worklist))
5817 * If a virtual machine is stopped by the host it can look to
5818 * the watchdog like a stall.
5820 kvm_check_and_clear_guest_paused();
5822 /* get the latest of pool and touched timestamps */
5823 pool_ts = READ_ONCE(pool->watchdog_ts);
5824 touched = READ_ONCE(wq_watchdog_touched);
5826 if (time_after(pool_ts, touched))
5831 if (pool->cpu >= 0) {
5832 unsigned long cpu_touched =
5833 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5835 if (time_after(cpu_touched, ts))
5840 if (time_after(now, ts + thresh)) {
5841 lockup_detected = true;
5842 pr_emerg("BUG: workqueue lockup - pool");
5843 pr_cont_pool_info(pool);
5844 pr_cont(" stuck for %us!\n",
5845 jiffies_to_msecs(now - pool_ts) / 1000);
5851 if (lockup_detected)
5852 show_workqueue_state();
5854 wq_watchdog_reset_touched();
5855 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5858 notrace void wq_watchdog_touch(int cpu)
5861 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5863 wq_watchdog_touched = jiffies;
5866 static void wq_watchdog_set_thresh(unsigned long thresh)
5868 wq_watchdog_thresh = 0;
5869 del_timer_sync(&wq_watchdog_timer);
5872 wq_watchdog_thresh = thresh;
5873 wq_watchdog_reset_touched();
5874 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5878 static int wq_watchdog_param_set_thresh(const char *val,
5879 const struct kernel_param *kp)
5881 unsigned long thresh;
5884 ret = kstrtoul(val, 0, &thresh);
5889 wq_watchdog_set_thresh(thresh);
5891 wq_watchdog_thresh = thresh;
5896 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5897 .set = wq_watchdog_param_set_thresh,
5898 .get = param_get_ulong,
5901 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5904 static void wq_watchdog_init(void)
5906 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5907 wq_watchdog_set_thresh(wq_watchdog_thresh);
5910 #else /* CONFIG_WQ_WATCHDOG */
5912 static inline void wq_watchdog_init(void) { }
5914 #endif /* CONFIG_WQ_WATCHDOG */
5916 static void __init wq_numa_init(void)
5921 if (num_possible_nodes() <= 1)
5924 if (wq_disable_numa) {
5925 pr_info("workqueue: NUMA affinity support disabled\n");
5929 for_each_possible_cpu(cpu) {
5930 if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) {
5931 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5936 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5937 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5940 * We want masks of possible CPUs of each node which isn't readily
5941 * available. Build one from cpu_to_node() which should have been
5942 * fully initialized by now.
5944 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5948 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5949 node_online(node) ? node : NUMA_NO_NODE));
5951 for_each_possible_cpu(cpu) {
5952 node = cpu_to_node(cpu);
5953 cpumask_set_cpu(cpu, tbl[node]);
5956 wq_numa_possible_cpumask = tbl;
5957 wq_numa_enabled = true;
5961 * workqueue_init_early - early init for workqueue subsystem
5963 * This is the first half of two-staged workqueue subsystem initialization
5964 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5965 * idr are up. It sets up all the data structures and system workqueues
5966 * and allows early boot code to create workqueues and queue/cancel work
5967 * items. Actual work item execution starts only after kthreads can be
5968 * created and scheduled right before early initcalls.
5970 void __init workqueue_init_early(void)
5972 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5973 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5976 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5978 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5979 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5981 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5983 /* initialize CPU pools */
5984 for_each_possible_cpu(cpu) {
5985 struct worker_pool *pool;
5988 for_each_cpu_worker_pool(pool, cpu) {
5989 BUG_ON(init_worker_pool(pool));
5991 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5992 pool->attrs->nice = std_nice[i++];
5993 pool->node = cpu_to_node(cpu);
5996 mutex_lock(&wq_pool_mutex);
5997 BUG_ON(worker_pool_assign_id(pool));
5998 mutex_unlock(&wq_pool_mutex);
6002 /* create default unbound and ordered wq attrs */
6003 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6004 struct workqueue_attrs *attrs;
6006 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6007 attrs->nice = std_nice[i];
6008 unbound_std_wq_attrs[i] = attrs;
6011 * An ordered wq should have only one pwq as ordering is
6012 * guaranteed by max_active which is enforced by pwqs.
6013 * Turn off NUMA so that dfl_pwq is used for all nodes.
6015 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6016 attrs->nice = std_nice[i];
6017 attrs->no_numa = true;
6018 ordered_wq_attrs[i] = attrs;
6021 system_wq = alloc_workqueue("events", 0, 0);
6022 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6023 system_long_wq = alloc_workqueue("events_long", 0, 0);
6024 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6025 WQ_UNBOUND_MAX_ACTIVE);
6026 system_freezable_wq = alloc_workqueue("events_freezable",
6028 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6029 WQ_POWER_EFFICIENT, 0);
6030 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6031 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6033 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6034 !system_unbound_wq || !system_freezable_wq ||
6035 !system_power_efficient_wq ||
6036 !system_freezable_power_efficient_wq);
6040 * workqueue_init - bring workqueue subsystem fully online
6042 * This is the latter half of two-staged workqueue subsystem initialization
6043 * and invoked as soon as kthreads can be created and scheduled.
6044 * Workqueues have been created and work items queued on them, but there
6045 * are no kworkers executing the work items yet. Populate the worker pools
6046 * with the initial workers and enable future kworker creations.
6048 void __init workqueue_init(void)
6050 struct workqueue_struct *wq;
6051 struct worker_pool *pool;
6055 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6056 * CPU to node mapping may not be available that early on some
6057 * archs such as power and arm64. As per-cpu pools created
6058 * previously could be missing node hint and unbound pools NUMA
6059 * affinity, fix them up.
6061 * Also, while iterating workqueues, create rescuers if requested.
6065 mutex_lock(&wq_pool_mutex);
6067 for_each_possible_cpu(cpu) {
6068 for_each_cpu_worker_pool(pool, cpu) {
6069 pool->node = cpu_to_node(cpu);
6073 list_for_each_entry(wq, &workqueues, list) {
6074 wq_update_unbound_numa(wq, smp_processor_id(), true);
6075 WARN(init_rescuer(wq),
6076 "workqueue: failed to create early rescuer for %s",
6080 mutex_unlock(&wq_pool_mutex);
6082 /* create the initial workers */
6083 for_each_online_cpu(cpu) {
6084 for_each_cpu_worker_pool(pool, cpu) {
6085 pool->flags &= ~POOL_DISASSOCIATED;
6086 BUG_ON(!create_worker(pool));
6090 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6091 BUG_ON(!create_worker(pool));