2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/core-api/workqueue.rst for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
50 #include <linux/sched/isolation.h>
51 #include <linux/nmi.h>
53 #include "workqueue_internal.h"
59 * A bound pool is either associated or disassociated with its CPU.
60 * While associated (!DISASSOCIATED), all workers are bound to the
61 * CPU and none has %WORKER_UNBOUND set and concurrency management
64 * While DISASSOCIATED, the cpu may be offline and all workers have
65 * %WORKER_UNBOUND set and concurrency management disabled, and may
66 * be executing on any CPU. The pool behaves as an unbound one.
68 * Note that DISASSOCIATED should be flipped only while holding
69 * wq_pool_attach_mutex to avoid changing binding state while
70 * worker_attach_to_pool() is in progress.
72 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
73 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
76 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
78 WORKER_PREP = 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give MIN_NICE.
104 RESCUER_NICE_LEVEL = MIN_NICE,
105 HIGHPRI_NICE_LEVEL = MIN_NICE,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * A: wq_pool_attach_mutex protected.
128 * PL: wq_pool_mutex protected.
130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
135 * sched-RCU for reads.
137 * WQ: wq->mutex protected.
139 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
141 * MD: wq_mayday_lock protected.
144 /* struct worker is defined in workqueue_internal.h */
147 spinlock_t lock; /* the pool lock */
148 int cpu; /* I: the associated cpu */
149 int node; /* I: the associated node ID */
150 int id; /* I: pool ID */
151 unsigned int flags; /* X: flags */
153 unsigned long watchdog_ts; /* L: watchdog timestamp */
155 struct list_head worklist; /* L: list of pending works */
157 int nr_workers; /* L: total number of workers */
158 int nr_idle; /* L: currently idle workers */
160 struct list_head idle_list; /* X: list of idle workers */
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for workers */
164 /* a workers is either on busy_hash or idle_list, or the manager */
165 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
166 /* L: hash of busy workers */
168 struct worker *manager; /* L: purely informational */
169 struct list_head workers; /* A: attached workers */
170 struct completion *detach_completion; /* all workers detached */
172 struct ida worker_ida; /* worker IDs for task name */
174 struct workqueue_attrs *attrs; /* I: worker attributes */
175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
176 int refcnt; /* PL: refcnt for unbound pools */
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
183 atomic_t nr_running ____cacheline_aligned_in_smp;
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
190 } ____cacheline_aligned_in_smp;
193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
198 struct pool_workqueue {
199 struct worker_pool *pool; /* I: the associated pool */
200 struct workqueue_struct *wq; /* I: the owning workqueue */
201 int work_color; /* L: current color */
202 int flush_color; /* L: flushing color */
203 int refcnt; /* L: reference count */
204 int nr_in_flight[WORK_NR_COLORS];
205 /* L: nr of in_flight works */
206 int nr_active; /* L: nr of active works */
207 int max_active; /* L: max active works */
208 struct list_head delayed_works; /* L: delayed works */
209 struct list_head pwqs_node; /* WR: node on wq->pwqs */
210 struct list_head mayday_node; /* MD: node on wq->maydays */
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
216 * determined without grabbing wq->mutex.
218 struct work_struct unbound_release_work;
220 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
223 * Structure used to wait for workqueue flush.
226 struct list_head list; /* WQ: list of flushers */
227 int flush_color; /* WQ: flush color waiting for */
228 struct completion done; /* flush completion */
234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
237 struct workqueue_struct {
238 struct list_head pwqs; /* WR: all pwqs of this wq */
239 struct list_head list; /* PR: list of all workqueues */
241 struct mutex mutex; /* protects this wq */
242 int work_color; /* WQ: current work color */
243 int flush_color; /* WQ: current flush color */
244 atomic_t nr_pwqs_to_flush; /* flush in progress */
245 struct wq_flusher *first_flusher; /* WQ: first flusher */
246 struct list_head flusher_queue; /* WQ: flush waiters */
247 struct list_head flusher_overflow; /* WQ: flush overflow list */
249 struct list_head maydays; /* MD: pwqs requesting rescue */
250 struct worker *rescuer; /* I: rescue worker */
252 int nr_drainers; /* WQ: drain in progress */
253 int saved_max_active; /* WQ: saved pwq max_active */
255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259 struct wq_device *wq_dev; /* I: for sysfs interface */
261 #ifdef CONFIG_LOCKDEP
263 struct lock_class_key key;
264 struct lockdep_map lockdep_map;
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache *pwq_cache;
283 static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa;
287 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293 static bool wq_online; /* can kworkers be created yet? */
295 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
301 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
302 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
303 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
305 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
306 static bool workqueue_freezing; /* PL: have wqs started freezing? */
308 /* PL: allowable cpus for unbound wqs and work items */
309 static cpumask_var_t wq_unbound_cpumask;
311 /* CPU where unbound work was last round robin scheduled from this CPU */
312 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
315 * Local execution of unbound work items is no longer guaranteed. The
316 * following always forces round-robin CPU selection on unbound work items
317 * to uncover usages which depend on it.
319 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
320 static bool wq_debug_force_rr_cpu = true;
322 static bool wq_debug_force_rr_cpu = false;
324 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
326 /* the per-cpu worker pools */
327 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
329 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
331 /* PL: hash of all unbound pools keyed by pool->attrs */
332 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
334 /* I: attributes used when instantiating standard unbound pools on demand */
335 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
337 /* I: attributes used when instantiating ordered pools on demand */
338 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
340 struct workqueue_struct *system_wq __read_mostly;
341 EXPORT_SYMBOL(system_wq);
342 struct workqueue_struct *system_highpri_wq __read_mostly;
343 EXPORT_SYMBOL_GPL(system_highpri_wq);
344 struct workqueue_struct *system_long_wq __read_mostly;
345 EXPORT_SYMBOL_GPL(system_long_wq);
346 struct workqueue_struct *system_unbound_wq __read_mostly;
347 EXPORT_SYMBOL_GPL(system_unbound_wq);
348 struct workqueue_struct *system_freezable_wq __read_mostly;
349 EXPORT_SYMBOL_GPL(system_freezable_wq);
350 struct workqueue_struct *system_power_efficient_wq __read_mostly;
351 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
352 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
353 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
355 static int worker_thread(void *__worker);
356 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
358 #define CREATE_TRACE_POINTS
359 #include <trace/events/workqueue.h>
361 #define assert_rcu_or_pool_mutex() \
362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
363 !lockdep_is_held(&wq_pool_mutex), \
364 "sched RCU or wq_pool_mutex should be held")
366 #define assert_rcu_or_wq_mutex(wq) \
367 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
368 !lockdep_is_held(&wq->mutex), \
369 "sched RCU or wq->mutex should be held")
371 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
373 !lockdep_is_held(&wq->mutex) && \
374 !lockdep_is_held(&wq_pool_mutex), \
375 "sched RCU, wq->mutex or wq_pool_mutex should be held")
377 #define for_each_cpu_worker_pool(pool, cpu) \
378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
383 * for_each_pool - iterate through all worker_pools in the system
384 * @pool: iteration cursor
385 * @pi: integer used for iteration
387 * This must be called either with wq_pool_mutex held or sched RCU read
388 * locked. If the pool needs to be used beyond the locking in effect, the
389 * caller is responsible for guaranteeing that the pool stays online.
391 * The if/else clause exists only for the lockdep assertion and can be
394 #define for_each_pool(pool, pi) \
395 idr_for_each_entry(&worker_pool_idr, pool, pi) \
396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
400 * for_each_pool_worker - iterate through all workers of a worker_pool
401 * @worker: iteration cursor
402 * @pool: worker_pool to iterate workers of
404 * This must be called with wq_pool_attach_mutex.
406 * The if/else clause exists only for the lockdep assertion and can be
409 #define for_each_pool_worker(worker, pool) \
410 list_for_each_entry((worker), &(pool)->workers, node) \
411 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
416 * @pwq: iteration cursor
417 * @wq: the target workqueue
419 * This must be called either with wq->mutex held or sched RCU read locked.
420 * If the pwq needs to be used beyond the locking in effect, the caller is
421 * responsible for guaranteeing that the pwq stays online.
423 * The if/else clause exists only for the lockdep assertion and can be
426 #define for_each_pwq(pwq, wq) \
427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
431 #ifdef CONFIG_DEBUG_OBJECTS_WORK
433 static struct debug_obj_descr work_debug_descr;
435 static void *work_debug_hint(void *addr)
437 return ((struct work_struct *) addr)->func;
440 static bool work_is_static_object(void *addr)
442 struct work_struct *work = addr;
444 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
448 * fixup_init is called when:
449 * - an active object is initialized
451 static bool work_fixup_init(void *addr, enum debug_obj_state state)
453 struct work_struct *work = addr;
456 case ODEBUG_STATE_ACTIVE:
457 cancel_work_sync(work);
458 debug_object_init(work, &work_debug_descr);
466 * fixup_free is called when:
467 * - an active object is freed
469 static bool work_fixup_free(void *addr, enum debug_obj_state state)
471 struct work_struct *work = addr;
474 case ODEBUG_STATE_ACTIVE:
475 cancel_work_sync(work);
476 debug_object_free(work, &work_debug_descr);
483 static struct debug_obj_descr work_debug_descr = {
484 .name = "work_struct",
485 .debug_hint = work_debug_hint,
486 .is_static_object = work_is_static_object,
487 .fixup_init = work_fixup_init,
488 .fixup_free = work_fixup_free,
491 static inline void debug_work_activate(struct work_struct *work)
493 debug_object_activate(work, &work_debug_descr);
496 static inline void debug_work_deactivate(struct work_struct *work)
498 debug_object_deactivate(work, &work_debug_descr);
501 void __init_work(struct work_struct *work, int onstack)
504 debug_object_init_on_stack(work, &work_debug_descr);
506 debug_object_init(work, &work_debug_descr);
508 EXPORT_SYMBOL_GPL(__init_work);
510 void destroy_work_on_stack(struct work_struct *work)
512 debug_object_free(work, &work_debug_descr);
514 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
516 void destroy_delayed_work_on_stack(struct delayed_work *work)
518 destroy_timer_on_stack(&work->timer);
519 debug_object_free(&work->work, &work_debug_descr);
521 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
524 static inline void debug_work_activate(struct work_struct *work) { }
525 static inline void debug_work_deactivate(struct work_struct *work) { }
529 * worker_pool_assign_id - allocate ID and assing it to @pool
530 * @pool: the pool pointer of interest
532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
533 * successfully, -errno on failure.
535 static int worker_pool_assign_id(struct worker_pool *pool)
539 lockdep_assert_held(&wq_pool_mutex);
541 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
552 * @wq: the target workqueue
555 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
557 * If the pwq needs to be used beyond the locking in effect, the caller is
558 * responsible for guaranteeing that the pwq stays online.
560 * Return: The unbound pool_workqueue for @node.
562 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
565 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
569 * delayed item is pending. The plan is to keep CPU -> NODE
570 * mapping valid and stable across CPU on/offlines. Once that
571 * happens, this workaround can be removed.
573 if (unlikely(node == NUMA_NO_NODE))
576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
579 static unsigned int work_color_to_flags(int color)
581 return color << WORK_STRUCT_COLOR_SHIFT;
584 static int get_work_color(struct work_struct *work)
586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
587 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
590 static int work_next_color(int color)
592 return (color + 1) % WORK_NR_COLORS;
596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597 * contain the pointer to the queued pwq. Once execution starts, the flag
598 * is cleared and the high bits contain OFFQ flags and pool ID.
600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601 * and clear_work_data() can be used to set the pwq, pool or clear
602 * work->data. These functions should only be called while the work is
603 * owned - ie. while the PENDING bit is set.
605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
606 * corresponding to a work. Pool is available once the work has been
607 * queued anywhere after initialization until it is sync canceled. pwq is
608 * available only while the work item is queued.
610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
611 * canceled. While being canceled, a work item may have its PENDING set
612 * but stay off timer and worklist for arbitrarily long and nobody should
613 * try to steal the PENDING bit.
615 static inline void set_work_data(struct work_struct *work, unsigned long data,
618 WARN_ON_ONCE(!work_pending(work));
619 atomic_long_set(&work->data, data | flags | work_static(work));
622 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
623 unsigned long extra_flags)
625 set_work_data(work, (unsigned long)pwq,
626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
629 static void set_work_pool_and_keep_pending(struct work_struct *work,
632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
633 WORK_STRUCT_PENDING);
636 static void set_work_pool_and_clear_pending(struct work_struct *work,
640 * The following wmb is paired with the implied mb in
641 * test_and_set_bit(PENDING) and ensures all updates to @work made
642 * here are visible to and precede any updates by the next PENDING
646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
648 * The following mb guarantees that previous clear of a PENDING bit
649 * will not be reordered with any speculative LOADS or STORES from
650 * work->current_func, which is executed afterwards. This possible
651 * reordering can lead to a missed execution on attempt to qeueue
652 * the same @work. E.g. consider this case:
655 * ---------------------------- --------------------------------
657 * 1 STORE event_indicated
658 * 2 queue_work_on() {
659 * 3 test_and_set_bit(PENDING)
660 * 4 } set_..._and_clear_pending() {
661 * 5 set_work_data() # clear bit
663 * 7 work->current_func() {
664 * 8 LOAD event_indicated
667 * Without an explicit full barrier speculative LOAD on line 8 can
668 * be executed before CPU#0 does STORE on line 1. If that happens,
669 * CPU#0 observes the PENDING bit is still set and new execution of
670 * a @work is not queued in a hope, that CPU#1 will eventually
671 * finish the queued @work. Meanwhile CPU#1 does not see
672 * event_indicated is set, because speculative LOAD was executed
673 * before actual STORE.
678 static void clear_work_data(struct work_struct *work)
680 smp_wmb(); /* see set_work_pool_and_clear_pending() */
681 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
684 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
686 unsigned long data = atomic_long_read(&work->data);
688 if (data & WORK_STRUCT_PWQ)
689 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
695 * get_work_pool - return the worker_pool a given work was associated with
696 * @work: the work item of interest
698 * Pools are created and destroyed under wq_pool_mutex, and allows read
699 * access under sched-RCU read lock. As such, this function should be
700 * called under wq_pool_mutex or with preemption disabled.
702 * All fields of the returned pool are accessible as long as the above
703 * mentioned locking is in effect. If the returned pool needs to be used
704 * beyond the critical section, the caller is responsible for ensuring the
705 * returned pool is and stays online.
707 * Return: The worker_pool @work was last associated with. %NULL if none.
709 static struct worker_pool *get_work_pool(struct work_struct *work)
711 unsigned long data = atomic_long_read(&work->data);
714 assert_rcu_or_pool_mutex();
716 if (data & WORK_STRUCT_PWQ)
717 return ((struct pool_workqueue *)
718 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
720 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
721 if (pool_id == WORK_OFFQ_POOL_NONE)
724 return idr_find(&worker_pool_idr, pool_id);
728 * get_work_pool_id - return the worker pool ID a given work is associated with
729 * @work: the work item of interest
731 * Return: The worker_pool ID @work was last associated with.
732 * %WORK_OFFQ_POOL_NONE if none.
734 static int get_work_pool_id(struct work_struct *work)
736 unsigned long data = atomic_long_read(&work->data);
738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
742 return data >> WORK_OFFQ_POOL_SHIFT;
745 static void mark_work_canceling(struct work_struct *work)
747 unsigned long pool_id = get_work_pool_id(work);
749 pool_id <<= WORK_OFFQ_POOL_SHIFT;
750 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
753 static bool work_is_canceling(struct work_struct *work)
755 unsigned long data = atomic_long_read(&work->data);
757 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
761 * Policy functions. These define the policies on how the global worker
762 * pools are managed. Unless noted otherwise, these functions assume that
763 * they're being called with pool->lock held.
766 static bool __need_more_worker(struct worker_pool *pool)
768 return !atomic_read(&pool->nr_running);
772 * Need to wake up a worker? Called from anything but currently
775 * Note that, because unbound workers never contribute to nr_running, this
776 * function will always return %true for unbound pools as long as the
777 * worklist isn't empty.
779 static bool need_more_worker(struct worker_pool *pool)
781 return !list_empty(&pool->worklist) && __need_more_worker(pool);
784 /* Can I start working? Called from busy but !running workers. */
785 static bool may_start_working(struct worker_pool *pool)
787 return pool->nr_idle;
790 /* Do I need to keep working? Called from currently running workers. */
791 static bool keep_working(struct worker_pool *pool)
793 return !list_empty(&pool->worklist) &&
794 atomic_read(&pool->nr_running) <= 1;
797 /* Do we need a new worker? Called from manager. */
798 static bool need_to_create_worker(struct worker_pool *pool)
800 return need_more_worker(pool) && !may_start_working(pool);
803 /* Do we have too many workers and should some go away? */
804 static bool too_many_workers(struct worker_pool *pool)
806 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
807 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
808 int nr_busy = pool->nr_workers - nr_idle;
810 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
817 /* Return the first idle worker. Safe with preemption disabled */
818 static struct worker *first_idle_worker(struct worker_pool *pool)
820 if (unlikely(list_empty(&pool->idle_list)))
823 return list_first_entry(&pool->idle_list, struct worker, entry);
827 * wake_up_worker - wake up an idle worker
828 * @pool: worker pool to wake worker from
830 * Wake up the first idle worker of @pool.
833 * spin_lock_irq(pool->lock).
835 static void wake_up_worker(struct worker_pool *pool)
837 struct worker *worker = first_idle_worker(pool);
840 wake_up_process(worker->task);
844 * wq_worker_waking_up - a worker is waking up
845 * @task: task waking up
846 * @cpu: CPU @task is waking up to
848 * This function is called during try_to_wake_up() when a worker is
852 * spin_lock_irq(rq->lock)
854 void wq_worker_waking_up(struct task_struct *task, int cpu)
856 struct worker *worker = kthread_data(task);
858 if (!(worker->flags & WORKER_NOT_RUNNING)) {
859 WARN_ON_ONCE(worker->pool->cpu != cpu);
860 atomic_inc(&worker->pool->nr_running);
865 * wq_worker_sleeping - a worker is going to sleep
866 * @task: task going to sleep
868 * This function is called during schedule() when a busy worker is
869 * going to sleep. Worker on the same cpu can be woken up by
870 * returning pointer to its task.
873 * spin_lock_irq(rq->lock)
876 * Worker task on @cpu to wake up, %NULL if none.
878 struct task_struct *wq_worker_sleeping(struct task_struct *task)
880 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
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 /* this can only happen on the local cpu */
894 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
898 * The counterpart of the following dec_and_test, implied mb,
899 * worklist not empty test sequence is in insert_work().
900 * Please read comment there.
902 * NOT_RUNNING is clear. This means that we're bound to and
903 * running on the local cpu w/ rq lock held and preemption
904 * disabled, which in turn means that none else could be
905 * manipulating idle_list, so dereferencing idle_list without pool
908 if (atomic_dec_and_test(&pool->nr_running) &&
909 !list_empty(&pool->worklist))
910 to_wakeup = first_idle_worker(pool);
911 return to_wakeup ? to_wakeup->task : NULL;
915 * wq_worker_last_func - retrieve worker's last work function
917 * Determine the last function a worker executed. This is called from
918 * the scheduler to get a worker's last known identity.
921 * spin_lock_irq(rq->lock)
924 * The last work function %current executed as a worker, NULL if it
925 * hasn't executed any work yet.
927 work_func_t wq_worker_last_func(struct task_struct *task)
929 struct worker *worker = kthread_data(task);
931 return worker->last_func;
935 * worker_set_flags - set worker flags and adjust nr_running accordingly
937 * @flags: flags to set
939 * Set @flags in @worker->flags and adjust nr_running accordingly.
942 * spin_lock_irq(pool->lock)
944 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
946 struct worker_pool *pool = worker->pool;
948 WARN_ON_ONCE(worker->task != current);
950 /* If transitioning into NOT_RUNNING, adjust nr_running. */
951 if ((flags & WORKER_NOT_RUNNING) &&
952 !(worker->flags & WORKER_NOT_RUNNING)) {
953 atomic_dec(&pool->nr_running);
956 worker->flags |= flags;
960 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
962 * @flags: flags to clear
964 * Clear @flags in @worker->flags and adjust nr_running accordingly.
967 * spin_lock_irq(pool->lock)
969 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
971 struct worker_pool *pool = worker->pool;
972 unsigned int oflags = worker->flags;
974 WARN_ON_ONCE(worker->task != current);
976 worker->flags &= ~flags;
979 * If transitioning out of NOT_RUNNING, increment nr_running. Note
980 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
981 * of multiple flags, not a single flag.
983 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
984 if (!(worker->flags & WORKER_NOT_RUNNING))
985 atomic_inc(&pool->nr_running);
989 * find_worker_executing_work - find worker which is executing a work
990 * @pool: pool of interest
991 * @work: work to find worker for
993 * Find a worker which is executing @work on @pool by searching
994 * @pool->busy_hash which is keyed by the address of @work. For a worker
995 * to match, its current execution should match the address of @work and
996 * its work function. This is to avoid unwanted dependency between
997 * unrelated work executions through a work item being recycled while still
1000 * This is a bit tricky. A work item may be freed once its execution
1001 * starts and nothing prevents the freed area from being recycled for
1002 * another work item. If the same work item address ends up being reused
1003 * before the original execution finishes, workqueue will identify the
1004 * recycled work item as currently executing and make it wait until the
1005 * current execution finishes, introducing an unwanted dependency.
1007 * This function checks the work item address and work function to avoid
1008 * false positives. Note that this isn't complete as one may construct a
1009 * work function which can introduce dependency onto itself through a
1010 * recycled work item. Well, if somebody wants to shoot oneself in the
1011 * foot that badly, there's only so much we can do, and if such deadlock
1012 * actually occurs, it should be easy to locate the culprit work function.
1015 * spin_lock_irq(pool->lock).
1018 * Pointer to worker which is executing @work if found, %NULL
1021 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1022 struct work_struct *work)
1024 struct worker *worker;
1026 hash_for_each_possible(pool->busy_hash, worker, hentry,
1027 (unsigned long)work)
1028 if (worker->current_work == work &&
1029 worker->current_func == work->func)
1036 * move_linked_works - move linked works to a list
1037 * @work: start of series of works to be scheduled
1038 * @head: target list to append @work to
1039 * @nextp: out parameter for nested worklist walking
1041 * Schedule linked works starting from @work to @head. Work series to
1042 * be scheduled starts at @work and includes any consecutive work with
1043 * WORK_STRUCT_LINKED set in its predecessor.
1045 * If @nextp is not NULL, it's updated to point to the next work of
1046 * the last scheduled work. This allows move_linked_works() to be
1047 * nested inside outer list_for_each_entry_safe().
1050 * spin_lock_irq(pool->lock).
1052 static void move_linked_works(struct work_struct *work, struct list_head *head,
1053 struct work_struct **nextp)
1055 struct work_struct *n;
1058 * Linked worklist will always end before the end of the list,
1059 * use NULL for list head.
1061 list_for_each_entry_safe_from(work, n, NULL, entry) {
1062 list_move_tail(&work->entry, head);
1063 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1068 * If we're already inside safe list traversal and have moved
1069 * multiple works to the scheduled queue, the next position
1070 * needs to be updated.
1077 * get_pwq - get an extra reference on the specified pool_workqueue
1078 * @pwq: pool_workqueue to get
1080 * Obtain an extra reference on @pwq. The caller should guarantee that
1081 * @pwq has positive refcnt and be holding the matching pool->lock.
1083 static void get_pwq(struct pool_workqueue *pwq)
1085 lockdep_assert_held(&pwq->pool->lock);
1086 WARN_ON_ONCE(pwq->refcnt <= 0);
1091 * put_pwq - put a pool_workqueue reference
1092 * @pwq: pool_workqueue to put
1094 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1095 * destruction. The caller should be holding the matching pool->lock.
1097 static void put_pwq(struct pool_workqueue *pwq)
1099 lockdep_assert_held(&pwq->pool->lock);
1100 if (likely(--pwq->refcnt))
1102 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1105 * @pwq can't be released under pool->lock, bounce to
1106 * pwq_unbound_release_workfn(). This never recurses on the same
1107 * pool->lock as this path is taken only for unbound workqueues and
1108 * the release work item is scheduled on a per-cpu workqueue. To
1109 * avoid lockdep warning, unbound pool->locks are given lockdep
1110 * subclass of 1 in get_unbound_pool().
1112 schedule_work(&pwq->unbound_release_work);
1116 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1117 * @pwq: pool_workqueue to put (can be %NULL)
1119 * put_pwq() with locking. This function also allows %NULL @pwq.
1121 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1125 * As both pwqs and pools are sched-RCU protected, the
1126 * following lock operations are safe.
1128 spin_lock_irq(&pwq->pool->lock);
1130 spin_unlock_irq(&pwq->pool->lock);
1134 static void pwq_activate_delayed_work(struct work_struct *work)
1136 struct pool_workqueue *pwq = get_work_pwq(work);
1138 trace_workqueue_activate_work(work);
1139 if (list_empty(&pwq->pool->worklist))
1140 pwq->pool->watchdog_ts = jiffies;
1141 move_linked_works(work, &pwq->pool->worklist, NULL);
1142 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1146 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1148 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1149 struct work_struct, entry);
1151 pwq_activate_delayed_work(work);
1155 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1156 * @pwq: pwq of interest
1157 * @color: color of work which left the queue
1159 * A work either has completed or is removed from pending queue,
1160 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1163 * spin_lock_irq(pool->lock).
1165 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1167 /* uncolored work items don't participate in flushing or nr_active */
1168 if (color == WORK_NO_COLOR)
1171 pwq->nr_in_flight[color]--;
1174 if (!list_empty(&pwq->delayed_works)) {
1175 /* one down, submit a delayed one */
1176 if (pwq->nr_active < pwq->max_active)
1177 pwq_activate_first_delayed(pwq);
1180 /* is flush in progress and are we at the flushing tip? */
1181 if (likely(pwq->flush_color != color))
1184 /* are there still in-flight works? */
1185 if (pwq->nr_in_flight[color])
1188 /* this pwq is done, clear flush_color */
1189 pwq->flush_color = -1;
1192 * If this was the last pwq, wake up the first flusher. It
1193 * will handle the rest.
1195 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1196 complete(&pwq->wq->first_flusher->done);
1202 * try_to_grab_pending - steal work item from worklist and disable irq
1203 * @work: work item to steal
1204 * @is_dwork: @work is a delayed_work
1205 * @flags: place to store irq state
1207 * Try to grab PENDING bit of @work. This function can handle @work in any
1208 * stable state - idle, on timer or on worklist.
1211 * 1 if @work was pending and we successfully stole PENDING
1212 * 0 if @work was idle and we claimed PENDING
1213 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1214 * -ENOENT if someone else is canceling @work, this state may persist
1215 * for arbitrarily long
1218 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1219 * interrupted while holding PENDING and @work off queue, irq must be
1220 * disabled on entry. This, combined with delayed_work->timer being
1221 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1223 * On successful return, >= 0, irq is disabled and the caller is
1224 * responsible for releasing it using local_irq_restore(*@flags).
1226 * This function is safe to call from any context including IRQ handler.
1228 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1229 unsigned long *flags)
1231 struct worker_pool *pool;
1232 struct pool_workqueue *pwq;
1234 local_irq_save(*flags);
1236 /* try to steal the timer if it exists */
1238 struct delayed_work *dwork = to_delayed_work(work);
1241 * dwork->timer is irqsafe. If del_timer() fails, it's
1242 * guaranteed that the timer is not queued anywhere and not
1243 * running on the local CPU.
1245 if (likely(del_timer(&dwork->timer)))
1249 /* try to claim PENDING the normal way */
1250 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1254 * The queueing is in progress, or it is already queued. Try to
1255 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1257 pool = get_work_pool(work);
1261 spin_lock(&pool->lock);
1263 * work->data is guaranteed to point to pwq only while the work
1264 * item is queued on pwq->wq, and both updating work->data to point
1265 * to pwq on queueing and to pool on dequeueing are done under
1266 * pwq->pool->lock. This in turn guarantees that, if work->data
1267 * points to pwq which is associated with a locked pool, the work
1268 * item is currently queued on that pool.
1270 pwq = get_work_pwq(work);
1271 if (pwq && pwq->pool == pool) {
1272 debug_work_deactivate(work);
1275 * A delayed work item cannot be grabbed directly because
1276 * it might have linked NO_COLOR work items which, if left
1277 * on the delayed_list, will confuse pwq->nr_active
1278 * management later on and cause stall. Make sure the work
1279 * item is activated before grabbing.
1281 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1282 pwq_activate_delayed_work(work);
1284 list_del_init(&work->entry);
1285 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1287 /* work->data points to pwq iff queued, point to pool */
1288 set_work_pool_and_keep_pending(work, pool->id);
1290 spin_unlock(&pool->lock);
1293 spin_unlock(&pool->lock);
1295 local_irq_restore(*flags);
1296 if (work_is_canceling(work))
1303 * insert_work - insert a work into a pool
1304 * @pwq: pwq @work belongs to
1305 * @work: work to insert
1306 * @head: insertion point
1307 * @extra_flags: extra WORK_STRUCT_* flags to set
1309 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1310 * work_struct flags.
1313 * spin_lock_irq(pool->lock).
1315 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1316 struct list_head *head, unsigned int extra_flags)
1318 struct worker_pool *pool = pwq->pool;
1320 /* we own @work, set data and link */
1321 set_work_pwq(work, pwq, extra_flags);
1322 list_add_tail(&work->entry, head);
1326 * Ensure either wq_worker_sleeping() sees the above
1327 * list_add_tail() or we see zero nr_running to avoid workers lying
1328 * around lazily while there are works to be processed.
1332 if (__need_more_worker(pool))
1333 wake_up_worker(pool);
1337 * Test whether @work is being queued from another work executing on the
1340 static bool is_chained_work(struct workqueue_struct *wq)
1342 struct worker *worker;
1344 worker = current_wq_worker();
1346 * Return %true iff I'm a worker execuing a work item on @wq. If
1347 * I'm @worker, it's safe to dereference it without locking.
1349 return worker && worker->current_pwq->wq == wq;
1353 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1354 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1355 * avoid perturbing sensitive tasks.
1357 static int wq_select_unbound_cpu(int cpu)
1359 static bool printed_dbg_warning;
1362 if (likely(!wq_debug_force_rr_cpu)) {
1363 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1365 } else if (!printed_dbg_warning) {
1366 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1367 printed_dbg_warning = true;
1370 if (cpumask_empty(wq_unbound_cpumask))
1373 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1374 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1375 if (unlikely(new_cpu >= nr_cpu_ids)) {
1376 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1377 if (unlikely(new_cpu >= nr_cpu_ids))
1380 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1385 static void __queue_work(int cpu, struct workqueue_struct *wq,
1386 struct work_struct *work)
1388 struct pool_workqueue *pwq;
1389 struct worker_pool *last_pool;
1390 struct list_head *worklist;
1391 unsigned int work_flags;
1392 unsigned int req_cpu = cpu;
1395 * While a work item is PENDING && off queue, a task trying to
1396 * steal the PENDING will busy-loop waiting for it to either get
1397 * queued or lose PENDING. Grabbing PENDING and queueing should
1398 * happen with IRQ disabled.
1400 lockdep_assert_irqs_disabled();
1402 debug_work_activate(work);
1404 /* if draining, only works from the same workqueue are allowed */
1405 if (unlikely(wq->flags & __WQ_DRAINING) &&
1406 WARN_ON_ONCE(!is_chained_work(wq)))
1409 if (req_cpu == WORK_CPU_UNBOUND)
1410 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1412 /* pwq which will be used unless @work is executing elsewhere */
1413 if (!(wq->flags & WQ_UNBOUND))
1414 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1416 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1419 * If @work was previously on a different pool, it might still be
1420 * running there, in which case the work needs to be queued on that
1421 * pool to guarantee non-reentrancy.
1423 last_pool = get_work_pool(work);
1424 if (last_pool && last_pool != pwq->pool) {
1425 struct worker *worker;
1427 spin_lock(&last_pool->lock);
1429 worker = find_worker_executing_work(last_pool, work);
1431 if (worker && worker->current_pwq->wq == wq) {
1432 pwq = worker->current_pwq;
1434 /* meh... not running there, queue here */
1435 spin_unlock(&last_pool->lock);
1436 spin_lock(&pwq->pool->lock);
1439 spin_lock(&pwq->pool->lock);
1443 * pwq is determined and locked. For unbound pools, we could have
1444 * raced with pwq release and it could already be dead. If its
1445 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1446 * without another pwq replacing it in the numa_pwq_tbl or while
1447 * work items are executing on it, so the retrying is guaranteed to
1448 * make forward-progress.
1450 if (unlikely(!pwq->refcnt)) {
1451 if (wq->flags & WQ_UNBOUND) {
1452 spin_unlock(&pwq->pool->lock);
1457 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1461 /* pwq determined, queue */
1462 trace_workqueue_queue_work(req_cpu, pwq, work);
1464 if (WARN_ON(!list_empty(&work->entry))) {
1465 spin_unlock(&pwq->pool->lock);
1469 pwq->nr_in_flight[pwq->work_color]++;
1470 work_flags = work_color_to_flags(pwq->work_color);
1472 if (likely(pwq->nr_active < pwq->max_active)) {
1473 trace_workqueue_activate_work(work);
1475 worklist = &pwq->pool->worklist;
1476 if (list_empty(worklist))
1477 pwq->pool->watchdog_ts = jiffies;
1479 work_flags |= WORK_STRUCT_DELAYED;
1480 worklist = &pwq->delayed_works;
1483 insert_work(pwq, work, worklist, work_flags);
1485 spin_unlock(&pwq->pool->lock);
1489 * queue_work_on - queue work on specific cpu
1490 * @cpu: CPU number to execute work on
1491 * @wq: workqueue to use
1492 * @work: work to queue
1494 * We queue the work to a specific CPU, the caller must ensure it
1497 * Return: %false if @work was already on a queue, %true otherwise.
1499 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1500 struct work_struct *work)
1503 unsigned long flags;
1505 local_irq_save(flags);
1507 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1508 __queue_work(cpu, wq, work);
1512 local_irq_restore(flags);
1515 EXPORT_SYMBOL(queue_work_on);
1517 void delayed_work_timer_fn(struct timer_list *t)
1519 struct delayed_work *dwork = from_timer(dwork, t, timer);
1521 /* should have been called from irqsafe timer with irq already off */
1522 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1524 EXPORT_SYMBOL(delayed_work_timer_fn);
1526 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1527 struct delayed_work *dwork, unsigned long delay)
1529 struct timer_list *timer = &dwork->timer;
1530 struct work_struct *work = &dwork->work;
1533 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1534 WARN_ON_ONCE(timer_pending(timer));
1535 WARN_ON_ONCE(!list_empty(&work->entry));
1538 * If @delay is 0, queue @dwork->work immediately. This is for
1539 * both optimization and correctness. The earliest @timer can
1540 * expire is on the closest next tick and delayed_work users depend
1541 * on that there's no such delay when @delay is 0.
1544 __queue_work(cpu, wq, &dwork->work);
1550 timer->expires = jiffies + delay;
1552 if (unlikely(cpu != WORK_CPU_UNBOUND))
1553 add_timer_on(timer, cpu);
1559 * queue_delayed_work_on - queue work on specific CPU after delay
1560 * @cpu: CPU number to execute work on
1561 * @wq: workqueue to use
1562 * @dwork: work to queue
1563 * @delay: number of jiffies to wait before queueing
1565 * Return: %false if @work was already on a queue, %true otherwise. If
1566 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1569 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1570 struct delayed_work *dwork, unsigned long delay)
1572 struct work_struct *work = &dwork->work;
1574 unsigned long flags;
1576 /* read the comment in __queue_work() */
1577 local_irq_save(flags);
1579 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1580 __queue_delayed_work(cpu, wq, dwork, delay);
1584 local_irq_restore(flags);
1587 EXPORT_SYMBOL(queue_delayed_work_on);
1590 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1591 * @cpu: CPU number to execute work on
1592 * @wq: workqueue to use
1593 * @dwork: work to queue
1594 * @delay: number of jiffies to wait before queueing
1596 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1597 * modify @dwork's timer so that it expires after @delay. If @delay is
1598 * zero, @work is guaranteed to be scheduled immediately regardless of its
1601 * Return: %false if @dwork was idle and queued, %true if @dwork was
1602 * pending and its timer was modified.
1604 * This function is safe to call from any context including IRQ handler.
1605 * See try_to_grab_pending() for details.
1607 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1608 struct delayed_work *dwork, unsigned long delay)
1610 unsigned long flags;
1614 ret = try_to_grab_pending(&dwork->work, true, &flags);
1615 } while (unlikely(ret == -EAGAIN));
1617 if (likely(ret >= 0)) {
1618 __queue_delayed_work(cpu, wq, dwork, delay);
1619 local_irq_restore(flags);
1622 /* -ENOENT from try_to_grab_pending() becomes %true */
1625 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1627 static void rcu_work_rcufn(struct rcu_head *rcu)
1629 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1631 /* read the comment in __queue_work() */
1632 local_irq_disable();
1633 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1638 * queue_rcu_work - queue work after a RCU grace period
1639 * @wq: workqueue to use
1640 * @rwork: work to queue
1642 * Return: %false if @rwork was already pending, %true otherwise. Note
1643 * that a full RCU grace period is guaranteed only after a %true return.
1644 * While @rwork is guarnateed to be executed after a %false return, the
1645 * execution may happen before a full RCU grace period has passed.
1647 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1649 struct work_struct *work = &rwork->work;
1651 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1653 call_rcu(&rwork->rcu, rcu_work_rcufn);
1659 EXPORT_SYMBOL(queue_rcu_work);
1662 * worker_enter_idle - enter idle state
1663 * @worker: worker which is entering idle state
1665 * @worker is entering idle state. Update stats and idle timer if
1669 * spin_lock_irq(pool->lock).
1671 static void worker_enter_idle(struct worker *worker)
1673 struct worker_pool *pool = worker->pool;
1675 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1676 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1677 (worker->hentry.next || worker->hentry.pprev)))
1680 /* can't use worker_set_flags(), also called from create_worker() */
1681 worker->flags |= WORKER_IDLE;
1683 worker->last_active = jiffies;
1685 /* idle_list is LIFO */
1686 list_add(&worker->entry, &pool->idle_list);
1688 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1689 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1692 * Sanity check nr_running. Because unbind_workers() releases
1693 * pool->lock between setting %WORKER_UNBOUND and zapping
1694 * nr_running, the warning may trigger spuriously. Check iff
1695 * unbind is not in progress.
1697 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1698 pool->nr_workers == pool->nr_idle &&
1699 atomic_read(&pool->nr_running));
1703 * worker_leave_idle - leave idle state
1704 * @worker: worker which is leaving idle state
1706 * @worker is leaving idle state. Update stats.
1709 * spin_lock_irq(pool->lock).
1711 static void worker_leave_idle(struct worker *worker)
1713 struct worker_pool *pool = worker->pool;
1715 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1717 worker_clr_flags(worker, WORKER_IDLE);
1719 list_del_init(&worker->entry);
1722 static struct worker *alloc_worker(int node)
1724 struct worker *worker;
1726 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1728 INIT_LIST_HEAD(&worker->entry);
1729 INIT_LIST_HEAD(&worker->scheduled);
1730 INIT_LIST_HEAD(&worker->node);
1731 /* on creation a worker is in !idle && prep state */
1732 worker->flags = WORKER_PREP;
1738 * worker_attach_to_pool() - attach a worker to a pool
1739 * @worker: worker to be attached
1740 * @pool: the target pool
1742 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1743 * cpu-binding of @worker are kept coordinated with the pool across
1746 static void worker_attach_to_pool(struct worker *worker,
1747 struct worker_pool *pool)
1749 mutex_lock(&wq_pool_attach_mutex);
1752 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1753 * online CPUs. It'll be re-applied when any of the CPUs come up.
1755 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1758 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1759 * stable across this function. See the comments above the flag
1760 * definition for details.
1762 if (pool->flags & POOL_DISASSOCIATED)
1763 worker->flags |= WORKER_UNBOUND;
1765 list_add_tail(&worker->node, &pool->workers);
1766 worker->pool = pool;
1768 mutex_unlock(&wq_pool_attach_mutex);
1772 * worker_detach_from_pool() - detach a worker from its pool
1773 * @worker: worker which is attached to its pool
1775 * Undo the attaching which had been done in worker_attach_to_pool(). The
1776 * caller worker shouldn't access to the pool after detached except it has
1777 * other reference to the pool.
1779 static void worker_detach_from_pool(struct worker *worker)
1781 struct worker_pool *pool = worker->pool;
1782 struct completion *detach_completion = NULL;
1784 mutex_lock(&wq_pool_attach_mutex);
1786 list_del(&worker->node);
1787 worker->pool = NULL;
1789 if (list_empty(&pool->workers))
1790 detach_completion = pool->detach_completion;
1791 mutex_unlock(&wq_pool_attach_mutex);
1793 /* clear leftover flags without pool->lock after it is detached */
1794 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1796 if (detach_completion)
1797 complete(detach_completion);
1801 * create_worker - create a new workqueue worker
1802 * @pool: pool the new worker will belong to
1804 * Create and start a new worker which is attached to @pool.
1807 * Might sleep. Does GFP_KERNEL allocations.
1810 * Pointer to the newly created worker.
1812 static struct worker *create_worker(struct worker_pool *pool)
1814 struct worker *worker = NULL;
1818 /* ID is needed to determine kthread name */
1819 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1823 worker = alloc_worker(pool->node);
1830 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1831 pool->attrs->nice < 0 ? "H" : "");
1833 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1835 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1836 "kworker/%s", id_buf);
1837 if (IS_ERR(worker->task))
1840 set_user_nice(worker->task, pool->attrs->nice);
1841 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1843 /* successful, attach the worker to the pool */
1844 worker_attach_to_pool(worker, pool);
1846 /* start the newly created worker */
1847 spin_lock_irq(&pool->lock);
1848 worker->pool->nr_workers++;
1849 worker_enter_idle(worker);
1850 wake_up_process(worker->task);
1851 spin_unlock_irq(&pool->lock);
1857 ida_simple_remove(&pool->worker_ida, id);
1863 * destroy_worker - destroy a workqueue worker
1864 * @worker: worker to be destroyed
1866 * Destroy @worker and adjust @pool stats accordingly. The worker should
1870 * spin_lock_irq(pool->lock).
1872 static void destroy_worker(struct worker *worker)
1874 struct worker_pool *pool = worker->pool;
1876 lockdep_assert_held(&pool->lock);
1878 /* sanity check frenzy */
1879 if (WARN_ON(worker->current_work) ||
1880 WARN_ON(!list_empty(&worker->scheduled)) ||
1881 WARN_ON(!(worker->flags & WORKER_IDLE)))
1887 list_del_init(&worker->entry);
1888 worker->flags |= WORKER_DIE;
1889 wake_up_process(worker->task);
1892 static void idle_worker_timeout(struct timer_list *t)
1894 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1896 spin_lock_irq(&pool->lock);
1898 while (too_many_workers(pool)) {
1899 struct worker *worker;
1900 unsigned long expires;
1902 /* idle_list is kept in LIFO order, check the last one */
1903 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1904 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1906 if (time_before(jiffies, expires)) {
1907 mod_timer(&pool->idle_timer, expires);
1911 destroy_worker(worker);
1914 spin_unlock_irq(&pool->lock);
1917 static void send_mayday(struct work_struct *work)
1919 struct pool_workqueue *pwq = get_work_pwq(work);
1920 struct workqueue_struct *wq = pwq->wq;
1922 lockdep_assert_held(&wq_mayday_lock);
1927 /* mayday mayday mayday */
1928 if (list_empty(&pwq->mayday_node)) {
1930 * If @pwq is for an unbound wq, its base ref may be put at
1931 * any time due to an attribute change. Pin @pwq until the
1932 * rescuer is done with it.
1935 list_add_tail(&pwq->mayday_node, &wq->maydays);
1936 wake_up_process(wq->rescuer->task);
1940 static void pool_mayday_timeout(struct timer_list *t)
1942 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
1943 struct work_struct *work;
1945 spin_lock_irq(&pool->lock);
1946 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1948 if (need_to_create_worker(pool)) {
1950 * We've been trying to create a new worker but
1951 * haven't been successful. We might be hitting an
1952 * allocation deadlock. Send distress signals to
1955 list_for_each_entry(work, &pool->worklist, entry)
1959 spin_unlock(&wq_mayday_lock);
1960 spin_unlock_irq(&pool->lock);
1962 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1966 * maybe_create_worker - create a new worker if necessary
1967 * @pool: pool to create a new worker for
1969 * Create a new worker for @pool if necessary. @pool is guaranteed to
1970 * have at least one idle worker on return from this function. If
1971 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1972 * sent to all rescuers with works scheduled on @pool to resolve
1973 * possible allocation deadlock.
1975 * On return, need_to_create_worker() is guaranteed to be %false and
1976 * may_start_working() %true.
1979 * spin_lock_irq(pool->lock) which may be released and regrabbed
1980 * multiple times. Does GFP_KERNEL allocations. Called only from
1983 static void maybe_create_worker(struct worker_pool *pool)
1984 __releases(&pool->lock)
1985 __acquires(&pool->lock)
1988 spin_unlock_irq(&pool->lock);
1990 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1991 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1994 if (create_worker(pool) || !need_to_create_worker(pool))
1997 schedule_timeout_interruptible(CREATE_COOLDOWN);
1999 if (!need_to_create_worker(pool))
2003 del_timer_sync(&pool->mayday_timer);
2004 spin_lock_irq(&pool->lock);
2006 * This is necessary even after a new worker was just successfully
2007 * created as @pool->lock was dropped and the new worker might have
2008 * already become busy.
2010 if (need_to_create_worker(pool))
2015 * manage_workers - manage worker pool
2018 * Assume the manager role and manage the worker pool @worker belongs
2019 * to. At any given time, there can be only zero or one manager per
2020 * pool. The exclusion is handled automatically by this function.
2022 * The caller can safely start processing works on false return. On
2023 * true return, it's guaranteed that need_to_create_worker() is false
2024 * and may_start_working() is true.
2027 * spin_lock_irq(pool->lock) which may be released and regrabbed
2028 * multiple times. Does GFP_KERNEL allocations.
2031 * %false if the pool doesn't need management and the caller can safely
2032 * start processing works, %true if management function was performed and
2033 * the conditions that the caller verified before calling the function may
2034 * no longer be true.
2036 static bool manage_workers(struct worker *worker)
2038 struct worker_pool *pool = worker->pool;
2040 if (pool->flags & POOL_MANAGER_ACTIVE)
2043 pool->flags |= POOL_MANAGER_ACTIVE;
2044 pool->manager = worker;
2046 maybe_create_worker(pool);
2048 pool->manager = NULL;
2049 pool->flags &= ~POOL_MANAGER_ACTIVE;
2050 wake_up(&wq_manager_wait);
2055 * process_one_work - process single work
2057 * @work: work to process
2059 * Process @work. This function contains all the logics necessary to
2060 * process a single work including synchronization against and
2061 * interaction with other workers on the same cpu, queueing and
2062 * flushing. As long as context requirement is met, any worker can
2063 * call this function to process a work.
2066 * spin_lock_irq(pool->lock) which is released and regrabbed.
2068 static void process_one_work(struct worker *worker, struct work_struct *work)
2069 __releases(&pool->lock)
2070 __acquires(&pool->lock)
2072 struct pool_workqueue *pwq = get_work_pwq(work);
2073 struct worker_pool *pool = worker->pool;
2074 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2076 struct worker *collision;
2077 #ifdef CONFIG_LOCKDEP
2079 * It is permissible to free the struct work_struct from
2080 * inside the function that is called from it, this we need to
2081 * take into account for lockdep too. To avoid bogus "held
2082 * lock freed" warnings as well as problems when looking into
2083 * work->lockdep_map, make a copy and use that here.
2085 struct lockdep_map lockdep_map;
2087 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2089 /* ensure we're on the correct CPU */
2090 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2091 raw_smp_processor_id() != pool->cpu);
2094 * A single work shouldn't be executed concurrently by
2095 * multiple workers on a single cpu. Check whether anyone is
2096 * already processing the work. If so, defer the work to the
2097 * currently executing one.
2099 collision = find_worker_executing_work(pool, work);
2100 if (unlikely(collision)) {
2101 move_linked_works(work, &collision->scheduled, NULL);
2105 /* claim and dequeue */
2106 debug_work_deactivate(work);
2107 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2108 worker->current_work = work;
2109 worker->current_func = work->func;
2110 worker->current_pwq = pwq;
2111 work_color = get_work_color(work);
2114 * Record wq name for cmdline and debug reporting, may get
2115 * overridden through set_worker_desc().
2117 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2119 list_del_init(&work->entry);
2122 * CPU intensive works don't participate in concurrency management.
2123 * They're the scheduler's responsibility. This takes @worker out
2124 * of concurrency management and the next code block will chain
2125 * execution of the pending work items.
2127 if (unlikely(cpu_intensive))
2128 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2131 * Wake up another worker if necessary. The condition is always
2132 * false for normal per-cpu workers since nr_running would always
2133 * be >= 1 at this point. This is used to chain execution of the
2134 * pending work items for WORKER_NOT_RUNNING workers such as the
2135 * UNBOUND and CPU_INTENSIVE ones.
2137 if (need_more_worker(pool))
2138 wake_up_worker(pool);
2141 * Record the last pool and clear PENDING which should be the last
2142 * update to @work. Also, do this inside @pool->lock so that
2143 * PENDING and queued state changes happen together while IRQ is
2146 set_work_pool_and_clear_pending(work, pool->id);
2148 spin_unlock_irq(&pool->lock);
2150 lock_map_acquire(&pwq->wq->lockdep_map);
2151 lock_map_acquire(&lockdep_map);
2153 * Strictly speaking we should mark the invariant state without holding
2154 * any locks, that is, before these two lock_map_acquire()'s.
2156 * However, that would result in:
2163 * Which would create W1->C->W1 dependencies, even though there is no
2164 * actual deadlock possible. There are two solutions, using a
2165 * read-recursive acquire on the work(queue) 'locks', but this will then
2166 * hit the lockdep limitation on recursive locks, or simply discard
2169 * AFAICT there is no possible deadlock scenario between the
2170 * flush_work() and complete() primitives (except for single-threaded
2171 * workqueues), so hiding them isn't a problem.
2173 lockdep_invariant_state(true);
2174 trace_workqueue_execute_start(work);
2175 worker->current_func(work);
2177 * While we must be careful to not use "work" after this, the trace
2178 * point will only record its address.
2180 trace_workqueue_execute_end(work);
2181 lock_map_release(&lockdep_map);
2182 lock_map_release(&pwq->wq->lockdep_map);
2184 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2185 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2186 " last function: %pf\n",
2187 current->comm, preempt_count(), task_pid_nr(current),
2188 worker->current_func);
2189 debug_show_held_locks(current);
2194 * The following prevents a kworker from hogging CPU on !PREEMPT
2195 * kernels, where a requeueing work item waiting for something to
2196 * happen could deadlock with stop_machine as such work item could
2197 * indefinitely requeue itself while all other CPUs are trapped in
2198 * stop_machine. At the same time, report a quiescent RCU state so
2199 * the same condition doesn't freeze RCU.
2203 spin_lock_irq(&pool->lock);
2205 /* clear cpu intensive status */
2206 if (unlikely(cpu_intensive))
2207 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2209 /* tag the worker for identification in schedule() */
2210 worker->last_func = worker->current_func;
2212 /* we're done with it, release */
2213 hash_del(&worker->hentry);
2214 worker->current_work = NULL;
2215 worker->current_func = NULL;
2216 worker->current_pwq = NULL;
2217 pwq_dec_nr_in_flight(pwq, work_color);
2221 * process_scheduled_works - process scheduled works
2224 * Process all scheduled works. Please note that the scheduled list
2225 * may change while processing a work, so this function repeatedly
2226 * fetches a work from the top and executes it.
2229 * spin_lock_irq(pool->lock) which may be released and regrabbed
2232 static void process_scheduled_works(struct worker *worker)
2234 while (!list_empty(&worker->scheduled)) {
2235 struct work_struct *work = list_first_entry(&worker->scheduled,
2236 struct work_struct, entry);
2237 process_one_work(worker, work);
2241 static void set_pf_worker(bool val)
2243 mutex_lock(&wq_pool_attach_mutex);
2245 current->flags |= PF_WQ_WORKER;
2247 current->flags &= ~PF_WQ_WORKER;
2248 mutex_unlock(&wq_pool_attach_mutex);
2252 * worker_thread - the worker thread function
2255 * The worker thread function. All workers belong to a worker_pool -
2256 * either a per-cpu one or dynamic unbound one. These workers process all
2257 * work items regardless of their specific target workqueue. The only
2258 * exception is work items which belong to workqueues with a rescuer which
2259 * will be explained in rescuer_thread().
2263 static int worker_thread(void *__worker)
2265 struct worker *worker = __worker;
2266 struct worker_pool *pool = worker->pool;
2268 /* tell the scheduler that this is a workqueue worker */
2269 set_pf_worker(true);
2271 spin_lock_irq(&pool->lock);
2273 /* am I supposed to die? */
2274 if (unlikely(worker->flags & WORKER_DIE)) {
2275 spin_unlock_irq(&pool->lock);
2276 WARN_ON_ONCE(!list_empty(&worker->entry));
2277 set_pf_worker(false);
2279 set_task_comm(worker->task, "kworker/dying");
2280 ida_simple_remove(&pool->worker_ida, worker->id);
2281 worker_detach_from_pool(worker);
2286 worker_leave_idle(worker);
2288 /* no more worker necessary? */
2289 if (!need_more_worker(pool))
2292 /* do we need to manage? */
2293 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2297 * ->scheduled list can only be filled while a worker is
2298 * preparing to process a work or actually processing it.
2299 * Make sure nobody diddled with it while I was sleeping.
2301 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2304 * Finish PREP stage. We're guaranteed to have at least one idle
2305 * worker or that someone else has already assumed the manager
2306 * role. This is where @worker starts participating in concurrency
2307 * management if applicable and concurrency management is restored
2308 * after being rebound. See rebind_workers() for details.
2310 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2313 struct work_struct *work =
2314 list_first_entry(&pool->worklist,
2315 struct work_struct, entry);
2317 pool->watchdog_ts = jiffies;
2319 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2320 /* optimization path, not strictly necessary */
2321 process_one_work(worker, work);
2322 if (unlikely(!list_empty(&worker->scheduled)))
2323 process_scheduled_works(worker);
2325 move_linked_works(work, &worker->scheduled, NULL);
2326 process_scheduled_works(worker);
2328 } while (keep_working(pool));
2330 worker_set_flags(worker, WORKER_PREP);
2333 * pool->lock is held and there's no work to process and no need to
2334 * manage, sleep. Workers are woken up only while holding
2335 * pool->lock or from local cpu, so setting the current state
2336 * before releasing pool->lock is enough to prevent losing any
2339 worker_enter_idle(worker);
2340 __set_current_state(TASK_IDLE);
2341 spin_unlock_irq(&pool->lock);
2347 * rescuer_thread - the rescuer thread function
2350 * Workqueue rescuer thread function. There's one rescuer for each
2351 * workqueue which has WQ_MEM_RECLAIM set.
2353 * Regular work processing on a pool may block trying to create a new
2354 * worker which uses GFP_KERNEL allocation which has slight chance of
2355 * developing into deadlock if some works currently on the same queue
2356 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2357 * the problem rescuer solves.
2359 * When such condition is possible, the pool summons rescuers of all
2360 * workqueues which have works queued on the pool and let them process
2361 * those works so that forward progress can be guaranteed.
2363 * This should happen rarely.
2367 static int rescuer_thread(void *__rescuer)
2369 struct worker *rescuer = __rescuer;
2370 struct workqueue_struct *wq = rescuer->rescue_wq;
2371 struct list_head *scheduled = &rescuer->scheduled;
2374 set_user_nice(current, RESCUER_NICE_LEVEL);
2377 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2378 * doesn't participate in concurrency management.
2380 set_pf_worker(true);
2382 set_current_state(TASK_IDLE);
2385 * By the time the rescuer is requested to stop, the workqueue
2386 * shouldn't have any work pending, but @wq->maydays may still have
2387 * pwq(s) queued. This can happen by non-rescuer workers consuming
2388 * all the work items before the rescuer got to them. Go through
2389 * @wq->maydays processing before acting on should_stop so that the
2390 * list is always empty on exit.
2392 should_stop = kthread_should_stop();
2394 /* see whether any pwq is asking for help */
2395 spin_lock_irq(&wq_mayday_lock);
2397 while (!list_empty(&wq->maydays)) {
2398 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2399 struct pool_workqueue, mayday_node);
2400 struct worker_pool *pool = pwq->pool;
2401 struct work_struct *work, *n;
2404 __set_current_state(TASK_RUNNING);
2405 list_del_init(&pwq->mayday_node);
2407 spin_unlock_irq(&wq_mayday_lock);
2409 worker_attach_to_pool(rescuer, pool);
2411 spin_lock_irq(&pool->lock);
2414 * Slurp in all works issued via this workqueue and
2417 WARN_ON_ONCE(!list_empty(scheduled));
2418 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2419 if (get_work_pwq(work) == pwq) {
2421 pool->watchdog_ts = jiffies;
2422 move_linked_works(work, scheduled, &n);
2427 if (!list_empty(scheduled)) {
2428 process_scheduled_works(rescuer);
2431 * The above execution of rescued work items could
2432 * have created more to rescue through
2433 * pwq_activate_first_delayed() or chained
2434 * queueing. Let's put @pwq back on mayday list so
2435 * that such back-to-back work items, which may be
2436 * being used to relieve memory pressure, don't
2437 * incur MAYDAY_INTERVAL delay inbetween.
2439 if (need_to_create_worker(pool)) {
2440 spin_lock(&wq_mayday_lock);
2442 list_move_tail(&pwq->mayday_node, &wq->maydays);
2443 spin_unlock(&wq_mayday_lock);
2448 * Put the reference grabbed by send_mayday(). @pool won't
2449 * go away while we're still attached to it.
2454 * Leave this pool. If need_more_worker() is %true, notify a
2455 * regular worker; otherwise, we end up with 0 concurrency
2456 * and stalling the execution.
2458 if (need_more_worker(pool))
2459 wake_up_worker(pool);
2461 spin_unlock_irq(&pool->lock);
2463 worker_detach_from_pool(rescuer);
2465 spin_lock_irq(&wq_mayday_lock);
2468 spin_unlock_irq(&wq_mayday_lock);
2471 __set_current_state(TASK_RUNNING);
2472 set_pf_worker(false);
2476 /* rescuers should never participate in concurrency management */
2477 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2483 * check_flush_dependency - check for flush dependency sanity
2484 * @target_wq: workqueue being flushed
2485 * @target_work: work item being flushed (NULL for workqueue flushes)
2487 * %current is trying to flush the whole @target_wq or @target_work on it.
2488 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2489 * reclaiming memory or running on a workqueue which doesn't have
2490 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2493 static void check_flush_dependency(struct workqueue_struct *target_wq,
2494 struct work_struct *target_work)
2496 work_func_t target_func = target_work ? target_work->func : NULL;
2497 struct worker *worker;
2499 if (target_wq->flags & WQ_MEM_RECLAIM)
2502 worker = current_wq_worker();
2504 WARN_ONCE(current->flags & PF_MEMALLOC,
2505 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2506 current->pid, current->comm, target_wq->name, target_func);
2507 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2508 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2509 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2510 worker->current_pwq->wq->name, worker->current_func,
2511 target_wq->name, target_func);
2515 struct work_struct work;
2516 struct completion done;
2517 struct task_struct *task; /* purely informational */
2520 static void wq_barrier_func(struct work_struct *work)
2522 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2523 complete(&barr->done);
2527 * insert_wq_barrier - insert a barrier work
2528 * @pwq: pwq to insert barrier into
2529 * @barr: wq_barrier to insert
2530 * @target: target work to attach @barr to
2531 * @worker: worker currently executing @target, NULL if @target is not executing
2533 * @barr is linked to @target such that @barr is completed only after
2534 * @target finishes execution. Please note that the ordering
2535 * guarantee is observed only with respect to @target and on the local
2538 * Currently, a queued barrier can't be canceled. This is because
2539 * try_to_grab_pending() can't determine whether the work to be
2540 * grabbed is at the head of the queue and thus can't clear LINKED
2541 * flag of the previous work while there must be a valid next work
2542 * after a work with LINKED flag set.
2544 * Note that when @worker is non-NULL, @target may be modified
2545 * underneath us, so we can't reliably determine pwq from @target.
2548 * spin_lock_irq(pool->lock).
2550 static void insert_wq_barrier(struct pool_workqueue *pwq,
2551 struct wq_barrier *barr,
2552 struct work_struct *target, struct worker *worker)
2554 struct list_head *head;
2555 unsigned int linked = 0;
2558 * debugobject calls are safe here even with pool->lock locked
2559 * as we know for sure that this will not trigger any of the
2560 * checks and call back into the fixup functions where we
2563 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2564 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2566 init_completion_map(&barr->done, &target->lockdep_map);
2568 barr->task = current;
2571 * If @target is currently being executed, schedule the
2572 * barrier to the worker; otherwise, put it after @target.
2575 head = worker->scheduled.next;
2577 unsigned long *bits = work_data_bits(target);
2579 head = target->entry.next;
2580 /* there can already be other linked works, inherit and set */
2581 linked = *bits & WORK_STRUCT_LINKED;
2582 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2585 debug_work_activate(&barr->work);
2586 insert_work(pwq, &barr->work, head,
2587 work_color_to_flags(WORK_NO_COLOR) | linked);
2591 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2592 * @wq: workqueue being flushed
2593 * @flush_color: new flush color, < 0 for no-op
2594 * @work_color: new work color, < 0 for no-op
2596 * Prepare pwqs for workqueue flushing.
2598 * If @flush_color is non-negative, flush_color on all pwqs should be
2599 * -1. If no pwq has in-flight commands at the specified color, all
2600 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2601 * has in flight commands, its pwq->flush_color is set to
2602 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2603 * wakeup logic is armed and %true is returned.
2605 * The caller should have initialized @wq->first_flusher prior to
2606 * calling this function with non-negative @flush_color. If
2607 * @flush_color is negative, no flush color update is done and %false
2610 * If @work_color is non-negative, all pwqs should have the same
2611 * work_color which is previous to @work_color and all will be
2612 * advanced to @work_color.
2615 * mutex_lock(wq->mutex).
2618 * %true if @flush_color >= 0 and there's something to flush. %false
2621 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2622 int flush_color, int work_color)
2625 struct pool_workqueue *pwq;
2627 if (flush_color >= 0) {
2628 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2629 atomic_set(&wq->nr_pwqs_to_flush, 1);
2632 for_each_pwq(pwq, wq) {
2633 struct worker_pool *pool = pwq->pool;
2635 spin_lock_irq(&pool->lock);
2637 if (flush_color >= 0) {
2638 WARN_ON_ONCE(pwq->flush_color != -1);
2640 if (pwq->nr_in_flight[flush_color]) {
2641 pwq->flush_color = flush_color;
2642 atomic_inc(&wq->nr_pwqs_to_flush);
2647 if (work_color >= 0) {
2648 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2649 pwq->work_color = work_color;
2652 spin_unlock_irq(&pool->lock);
2655 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2656 complete(&wq->first_flusher->done);
2662 * flush_workqueue - ensure that any scheduled work has run to completion.
2663 * @wq: workqueue to flush
2665 * This function sleeps until all work items which were queued on entry
2666 * have finished execution, but it is not livelocked by new incoming ones.
2668 void flush_workqueue(struct workqueue_struct *wq)
2670 struct wq_flusher this_flusher = {
2671 .list = LIST_HEAD_INIT(this_flusher.list),
2673 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2677 if (WARN_ON(!wq_online))
2680 lock_map_acquire(&wq->lockdep_map);
2681 lock_map_release(&wq->lockdep_map);
2683 mutex_lock(&wq->mutex);
2686 * Start-to-wait phase
2688 next_color = work_next_color(wq->work_color);
2690 if (next_color != wq->flush_color) {
2692 * Color space is not full. The current work_color
2693 * becomes our flush_color and work_color is advanced
2696 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2697 this_flusher.flush_color = wq->work_color;
2698 wq->work_color = next_color;
2700 if (!wq->first_flusher) {
2701 /* no flush in progress, become the first flusher */
2702 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2704 wq->first_flusher = &this_flusher;
2706 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2708 /* nothing to flush, done */
2709 wq->flush_color = next_color;
2710 wq->first_flusher = NULL;
2715 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2716 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2717 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2721 * Oops, color space is full, wait on overflow queue.
2722 * The next flush completion will assign us
2723 * flush_color and transfer to flusher_queue.
2725 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2728 check_flush_dependency(wq, NULL);
2730 mutex_unlock(&wq->mutex);
2732 wait_for_completion(&this_flusher.done);
2735 * Wake-up-and-cascade phase
2737 * First flushers are responsible for cascading flushes and
2738 * handling overflow. Non-first flushers can simply return.
2740 if (wq->first_flusher != &this_flusher)
2743 mutex_lock(&wq->mutex);
2745 /* we might have raced, check again with mutex held */
2746 if (wq->first_flusher != &this_flusher)
2749 wq->first_flusher = NULL;
2751 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2752 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2755 struct wq_flusher *next, *tmp;
2757 /* complete all the flushers sharing the current flush color */
2758 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2759 if (next->flush_color != wq->flush_color)
2761 list_del_init(&next->list);
2762 complete(&next->done);
2765 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2766 wq->flush_color != work_next_color(wq->work_color));
2768 /* this flush_color is finished, advance by one */
2769 wq->flush_color = work_next_color(wq->flush_color);
2771 /* one color has been freed, handle overflow queue */
2772 if (!list_empty(&wq->flusher_overflow)) {
2774 * Assign the same color to all overflowed
2775 * flushers, advance work_color and append to
2776 * flusher_queue. This is the start-to-wait
2777 * phase for these overflowed flushers.
2779 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2780 tmp->flush_color = wq->work_color;
2782 wq->work_color = work_next_color(wq->work_color);
2784 list_splice_tail_init(&wq->flusher_overflow,
2785 &wq->flusher_queue);
2786 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2789 if (list_empty(&wq->flusher_queue)) {
2790 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2795 * Need to flush more colors. Make the next flusher
2796 * the new first flusher and arm pwqs.
2798 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2799 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2801 list_del_init(&next->list);
2802 wq->first_flusher = next;
2804 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2808 * Meh... this color is already done, clear first
2809 * flusher and repeat cascading.
2811 wq->first_flusher = NULL;
2815 mutex_unlock(&wq->mutex);
2817 EXPORT_SYMBOL(flush_workqueue);
2820 * drain_workqueue - drain a workqueue
2821 * @wq: workqueue to drain
2823 * Wait until the workqueue becomes empty. While draining is in progress,
2824 * only chain queueing is allowed. IOW, only currently pending or running
2825 * work items on @wq can queue further work items on it. @wq is flushed
2826 * repeatedly until it becomes empty. The number of flushing is determined
2827 * by the depth of chaining and should be relatively short. Whine if it
2830 void drain_workqueue(struct workqueue_struct *wq)
2832 unsigned int flush_cnt = 0;
2833 struct pool_workqueue *pwq;
2836 * __queue_work() needs to test whether there are drainers, is much
2837 * hotter than drain_workqueue() and already looks at @wq->flags.
2838 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2840 mutex_lock(&wq->mutex);
2841 if (!wq->nr_drainers++)
2842 wq->flags |= __WQ_DRAINING;
2843 mutex_unlock(&wq->mutex);
2845 flush_workqueue(wq);
2847 mutex_lock(&wq->mutex);
2849 for_each_pwq(pwq, wq) {
2852 spin_lock_irq(&pwq->pool->lock);
2853 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2854 spin_unlock_irq(&pwq->pool->lock);
2859 if (++flush_cnt == 10 ||
2860 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2861 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2862 wq->name, flush_cnt);
2864 mutex_unlock(&wq->mutex);
2868 if (!--wq->nr_drainers)
2869 wq->flags &= ~__WQ_DRAINING;
2870 mutex_unlock(&wq->mutex);
2872 EXPORT_SYMBOL_GPL(drain_workqueue);
2874 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2877 struct worker *worker = NULL;
2878 struct worker_pool *pool;
2879 struct pool_workqueue *pwq;
2883 local_irq_disable();
2884 pool = get_work_pool(work);
2890 spin_lock(&pool->lock);
2891 /* see the comment in try_to_grab_pending() with the same code */
2892 pwq = get_work_pwq(work);
2894 if (unlikely(pwq->pool != pool))
2897 worker = find_worker_executing_work(pool, work);
2900 pwq = worker->current_pwq;
2903 check_flush_dependency(pwq->wq, work);
2905 insert_wq_barrier(pwq, barr, work, worker);
2906 spin_unlock_irq(&pool->lock);
2909 * Force a lock recursion deadlock when using flush_work() inside a
2910 * single-threaded or rescuer equipped workqueue.
2912 * For single threaded workqueues the deadlock happens when the work
2913 * is after the work issuing the flush_work(). For rescuer equipped
2914 * workqueues the deadlock happens when the rescuer stalls, blocking
2918 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
2919 lock_map_acquire(&pwq->wq->lockdep_map);
2920 lock_map_release(&pwq->wq->lockdep_map);
2925 spin_unlock_irq(&pool->lock);
2929 static bool __flush_work(struct work_struct *work, bool from_cancel)
2931 struct wq_barrier barr;
2933 if (WARN_ON(!wq_online))
2937 lock_map_acquire(&work->lockdep_map);
2938 lock_map_release(&work->lockdep_map);
2941 if (start_flush_work(work, &barr, from_cancel)) {
2942 wait_for_completion(&barr.done);
2943 destroy_work_on_stack(&barr.work);
2951 * flush_work - wait for a work to finish executing the last queueing instance
2952 * @work: the work to flush
2954 * Wait until @work has finished execution. @work is guaranteed to be idle
2955 * on return if it hasn't been requeued since flush started.
2958 * %true if flush_work() waited for the work to finish execution,
2959 * %false if it was already idle.
2961 bool flush_work(struct work_struct *work)
2963 return __flush_work(work, false);
2965 EXPORT_SYMBOL_GPL(flush_work);
2968 wait_queue_entry_t wait;
2969 struct work_struct *work;
2972 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2974 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2976 if (cwait->work != key)
2978 return autoremove_wake_function(wait, mode, sync, key);
2981 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2983 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2984 unsigned long flags;
2988 ret = try_to_grab_pending(work, is_dwork, &flags);
2990 * If someone else is already canceling, wait for it to
2991 * finish. flush_work() doesn't work for PREEMPT_NONE
2992 * because we may get scheduled between @work's completion
2993 * and the other canceling task resuming and clearing
2994 * CANCELING - flush_work() will return false immediately
2995 * as @work is no longer busy, try_to_grab_pending() will
2996 * return -ENOENT as @work is still being canceled and the
2997 * other canceling task won't be able to clear CANCELING as
2998 * we're hogging the CPU.
3000 * Let's wait for completion using a waitqueue. As this
3001 * may lead to the thundering herd problem, use a custom
3002 * wake function which matches @work along with exclusive
3005 if (unlikely(ret == -ENOENT)) {
3006 struct cwt_wait cwait;
3008 init_wait(&cwait.wait);
3009 cwait.wait.func = cwt_wakefn;
3012 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3013 TASK_UNINTERRUPTIBLE);
3014 if (work_is_canceling(work))
3016 finish_wait(&cancel_waitq, &cwait.wait);
3018 } while (unlikely(ret < 0));
3020 /* tell other tasks trying to grab @work to back off */
3021 mark_work_canceling(work);
3022 local_irq_restore(flags);
3025 * This allows canceling during early boot. We know that @work
3029 __flush_work(work, true);
3031 clear_work_data(work);
3034 * Paired with prepare_to_wait() above so that either
3035 * waitqueue_active() is visible here or !work_is_canceling() is
3039 if (waitqueue_active(&cancel_waitq))
3040 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3046 * cancel_work_sync - cancel a work and wait for it to finish
3047 * @work: the work to cancel
3049 * Cancel @work and wait for its execution to finish. This function
3050 * can be used even if the work re-queues itself or migrates to
3051 * another workqueue. On return from this function, @work is
3052 * guaranteed to be not pending or executing on any CPU.
3054 * cancel_work_sync(&delayed_work->work) must not be used for
3055 * delayed_work's. Use cancel_delayed_work_sync() instead.
3057 * The caller must ensure that the workqueue on which @work was last
3058 * queued can't be destroyed before this function returns.
3061 * %true if @work was pending, %false otherwise.
3063 bool cancel_work_sync(struct work_struct *work)
3065 return __cancel_work_timer(work, false);
3067 EXPORT_SYMBOL_GPL(cancel_work_sync);
3070 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3071 * @dwork: the delayed work to flush
3073 * Delayed timer is cancelled and the pending work is queued for
3074 * immediate execution. Like flush_work(), this function only
3075 * considers the last queueing instance of @dwork.
3078 * %true if flush_work() waited for the work to finish execution,
3079 * %false if it was already idle.
3081 bool flush_delayed_work(struct delayed_work *dwork)
3083 local_irq_disable();
3084 if (del_timer_sync(&dwork->timer))
3085 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3087 return flush_work(&dwork->work);
3089 EXPORT_SYMBOL(flush_delayed_work);
3092 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3093 * @rwork: the rcu work to flush
3096 * %true if flush_rcu_work() waited for the work to finish execution,
3097 * %false if it was already idle.
3099 bool flush_rcu_work(struct rcu_work *rwork)
3101 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3103 flush_work(&rwork->work);
3106 return flush_work(&rwork->work);
3109 EXPORT_SYMBOL(flush_rcu_work);
3111 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3113 unsigned long flags;
3117 ret = try_to_grab_pending(work, is_dwork, &flags);
3118 } while (unlikely(ret == -EAGAIN));
3120 if (unlikely(ret < 0))
3123 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3124 local_irq_restore(flags);
3129 * cancel_delayed_work - cancel a delayed work
3130 * @dwork: delayed_work to cancel
3132 * Kill off a pending delayed_work.
3134 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3138 * The work callback function may still be running on return, unless
3139 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3140 * use cancel_delayed_work_sync() to wait on it.
3142 * This function is safe to call from any context including IRQ handler.
3144 bool cancel_delayed_work(struct delayed_work *dwork)
3146 return __cancel_work(&dwork->work, true);
3148 EXPORT_SYMBOL(cancel_delayed_work);
3151 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3152 * @dwork: the delayed work cancel
3154 * This is cancel_work_sync() for delayed works.
3157 * %true if @dwork was pending, %false otherwise.
3159 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3161 return __cancel_work_timer(&dwork->work, true);
3163 EXPORT_SYMBOL(cancel_delayed_work_sync);
3166 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3167 * @func: the function to call
3169 * schedule_on_each_cpu() executes @func on each online CPU using the
3170 * system workqueue and blocks until all CPUs have completed.
3171 * schedule_on_each_cpu() is very slow.
3174 * 0 on success, -errno on failure.
3176 int schedule_on_each_cpu(work_func_t func)
3179 struct work_struct __percpu *works;
3181 works = alloc_percpu(struct work_struct);
3187 for_each_online_cpu(cpu) {
3188 struct work_struct *work = per_cpu_ptr(works, cpu);
3190 INIT_WORK(work, func);
3191 schedule_work_on(cpu, work);
3194 for_each_online_cpu(cpu)
3195 flush_work(per_cpu_ptr(works, cpu));
3203 * execute_in_process_context - reliably execute the routine with user context
3204 * @fn: the function to execute
3205 * @ew: guaranteed storage for the execute work structure (must
3206 * be available when the work executes)
3208 * Executes the function immediately if process context is available,
3209 * otherwise schedules the function for delayed execution.
3211 * Return: 0 - function was executed
3212 * 1 - function was scheduled for execution
3214 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3216 if (!in_interrupt()) {
3221 INIT_WORK(&ew->work, fn);
3222 schedule_work(&ew->work);
3226 EXPORT_SYMBOL_GPL(execute_in_process_context);
3229 * free_workqueue_attrs - free a workqueue_attrs
3230 * @attrs: workqueue_attrs to free
3232 * Undo alloc_workqueue_attrs().
3234 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3237 free_cpumask_var(attrs->cpumask);
3243 * alloc_workqueue_attrs - allocate a workqueue_attrs
3244 * @gfp_mask: allocation mask to use
3246 * Allocate a new workqueue_attrs, initialize with default settings and
3249 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3251 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3253 struct workqueue_attrs *attrs;
3255 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3258 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3261 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3264 free_workqueue_attrs(attrs);
3268 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3269 const struct workqueue_attrs *from)
3271 to->nice = from->nice;
3272 cpumask_copy(to->cpumask, from->cpumask);
3274 * Unlike hash and equality test, this function doesn't ignore
3275 * ->no_numa as it is used for both pool and wq attrs. Instead,
3276 * get_unbound_pool() explicitly clears ->no_numa after copying.
3278 to->no_numa = from->no_numa;
3281 /* hash value of the content of @attr */
3282 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3286 hash = jhash_1word(attrs->nice, hash);
3287 hash = jhash(cpumask_bits(attrs->cpumask),
3288 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3292 /* content equality test */
3293 static bool wqattrs_equal(const struct workqueue_attrs *a,
3294 const struct workqueue_attrs *b)
3296 if (a->nice != b->nice)
3298 if (!cpumask_equal(a->cpumask, b->cpumask))
3304 * init_worker_pool - initialize a newly zalloc'd worker_pool
3305 * @pool: worker_pool to initialize
3307 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3309 * Return: 0 on success, -errno on failure. Even on failure, all fields
3310 * inside @pool proper are initialized and put_unbound_pool() can be called
3311 * on @pool safely to release it.
3313 static int init_worker_pool(struct worker_pool *pool)
3315 spin_lock_init(&pool->lock);
3318 pool->node = NUMA_NO_NODE;
3319 pool->flags |= POOL_DISASSOCIATED;
3320 pool->watchdog_ts = jiffies;
3321 INIT_LIST_HEAD(&pool->worklist);
3322 INIT_LIST_HEAD(&pool->idle_list);
3323 hash_init(pool->busy_hash);
3325 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3327 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3329 INIT_LIST_HEAD(&pool->workers);
3331 ida_init(&pool->worker_ida);
3332 INIT_HLIST_NODE(&pool->hash_node);
3335 /* shouldn't fail above this point */
3336 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3342 #ifdef CONFIG_LOCKDEP
3343 static void wq_init_lockdep(struct workqueue_struct *wq)
3347 lockdep_register_key(&wq->key);
3348 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3350 lock_name = wq->name;
3351 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3354 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3356 lockdep_unregister_key(&wq->key);
3359 static void wq_free_lockdep(struct workqueue_struct *wq)
3361 if (wq->lock_name != wq->name)
3362 kfree(wq->lock_name);
3365 static void wq_init_lockdep(struct workqueue_struct *wq)
3369 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3373 static void wq_free_lockdep(struct workqueue_struct *wq)
3378 static void rcu_free_wq(struct rcu_head *rcu)
3380 struct workqueue_struct *wq =
3381 container_of(rcu, struct workqueue_struct, rcu);
3383 wq_free_lockdep(wq);
3385 if (!(wq->flags & WQ_UNBOUND))
3386 free_percpu(wq->cpu_pwqs);
3388 free_workqueue_attrs(wq->unbound_attrs);
3394 static void rcu_free_pool(struct rcu_head *rcu)
3396 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3398 ida_destroy(&pool->worker_ida);
3399 free_workqueue_attrs(pool->attrs);
3404 * put_unbound_pool - put a worker_pool
3405 * @pool: worker_pool to put
3407 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3408 * safe manner. get_unbound_pool() calls this function on its failure path
3409 * and this function should be able to release pools which went through,
3410 * successfully or not, init_worker_pool().
3412 * Should be called with wq_pool_mutex held.
3414 static void put_unbound_pool(struct worker_pool *pool)
3416 DECLARE_COMPLETION_ONSTACK(detach_completion);
3417 struct worker *worker;
3419 lockdep_assert_held(&wq_pool_mutex);
3425 if (WARN_ON(!(pool->cpu < 0)) ||
3426 WARN_ON(!list_empty(&pool->worklist)))
3429 /* release id and unhash */
3431 idr_remove(&worker_pool_idr, pool->id);
3432 hash_del(&pool->hash_node);
3435 * Become the manager and destroy all workers. This prevents
3436 * @pool's workers from blocking on attach_mutex. We're the last
3437 * manager and @pool gets freed with the flag set.
3439 spin_lock_irq(&pool->lock);
3440 wait_event_lock_irq(wq_manager_wait,
3441 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3442 pool->flags |= POOL_MANAGER_ACTIVE;
3444 while ((worker = first_idle_worker(pool)))
3445 destroy_worker(worker);
3446 WARN_ON(pool->nr_workers || pool->nr_idle);
3447 spin_unlock_irq(&pool->lock);
3449 mutex_lock(&wq_pool_attach_mutex);
3450 if (!list_empty(&pool->workers))
3451 pool->detach_completion = &detach_completion;
3452 mutex_unlock(&wq_pool_attach_mutex);
3454 if (pool->detach_completion)
3455 wait_for_completion(pool->detach_completion);
3457 /* shut down the timers */
3458 del_timer_sync(&pool->idle_timer);
3459 del_timer_sync(&pool->mayday_timer);
3461 /* sched-RCU protected to allow dereferences from get_work_pool() */
3462 call_rcu(&pool->rcu, rcu_free_pool);
3466 * get_unbound_pool - get a worker_pool with the specified attributes
3467 * @attrs: the attributes of the worker_pool to get
3469 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3470 * reference count and return it. If there already is a matching
3471 * worker_pool, it will be used; otherwise, this function attempts to
3474 * Should be called with wq_pool_mutex held.
3476 * Return: On success, a worker_pool with the same attributes as @attrs.
3477 * On failure, %NULL.
3479 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3481 u32 hash = wqattrs_hash(attrs);
3482 struct worker_pool *pool;
3484 int target_node = NUMA_NO_NODE;
3486 lockdep_assert_held(&wq_pool_mutex);
3488 /* do we already have a matching pool? */
3489 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3490 if (wqattrs_equal(pool->attrs, attrs)) {
3496 /* if cpumask is contained inside a NUMA node, we belong to that node */
3497 if (wq_numa_enabled) {
3498 for_each_node(node) {
3499 if (cpumask_subset(attrs->cpumask,
3500 wq_numa_possible_cpumask[node])) {
3507 /* nope, create a new one */
3508 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3509 if (!pool || init_worker_pool(pool) < 0)
3512 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3513 copy_workqueue_attrs(pool->attrs, attrs);
3514 pool->node = target_node;
3517 * no_numa isn't a worker_pool attribute, always clear it. See
3518 * 'struct workqueue_attrs' comments for detail.
3520 pool->attrs->no_numa = false;
3522 if (worker_pool_assign_id(pool) < 0)
3525 /* create and start the initial worker */
3526 if (wq_online && !create_worker(pool))
3530 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3535 put_unbound_pool(pool);
3539 static void rcu_free_pwq(struct rcu_head *rcu)
3541 kmem_cache_free(pwq_cache,
3542 container_of(rcu, struct pool_workqueue, rcu));
3546 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3547 * and needs to be destroyed.
3549 static void pwq_unbound_release_workfn(struct work_struct *work)
3551 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3552 unbound_release_work);
3553 struct workqueue_struct *wq = pwq->wq;
3554 struct worker_pool *pool = pwq->pool;
3557 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3560 mutex_lock(&wq->mutex);
3561 list_del_rcu(&pwq->pwqs_node);
3562 is_last = list_empty(&wq->pwqs);
3563 mutex_unlock(&wq->mutex);
3565 mutex_lock(&wq_pool_mutex);
3566 put_unbound_pool(pool);
3567 mutex_unlock(&wq_pool_mutex);
3569 call_rcu(&pwq->rcu, rcu_free_pwq);
3572 * If we're the last pwq going away, @wq is already dead and no one
3573 * is gonna access it anymore. Schedule RCU free.
3576 wq_unregister_lockdep(wq);
3577 call_rcu(&wq->rcu, rcu_free_wq);
3582 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3583 * @pwq: target pool_workqueue
3585 * If @pwq isn't freezing, set @pwq->max_active to the associated
3586 * workqueue's saved_max_active and activate delayed work items
3587 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3589 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3591 struct workqueue_struct *wq = pwq->wq;
3592 bool freezable = wq->flags & WQ_FREEZABLE;
3593 unsigned long flags;
3595 /* for @wq->saved_max_active */
3596 lockdep_assert_held(&wq->mutex);
3598 /* fast exit for non-freezable wqs */
3599 if (!freezable && pwq->max_active == wq->saved_max_active)
3602 /* this function can be called during early boot w/ irq disabled */
3603 spin_lock_irqsave(&pwq->pool->lock, flags);
3606 * During [un]freezing, the caller is responsible for ensuring that
3607 * this function is called at least once after @workqueue_freezing
3608 * is updated and visible.
3610 if (!freezable || !workqueue_freezing) {
3611 pwq->max_active = wq->saved_max_active;
3613 while (!list_empty(&pwq->delayed_works) &&
3614 pwq->nr_active < pwq->max_active)
3615 pwq_activate_first_delayed(pwq);
3618 * Need to kick a worker after thawed or an unbound wq's
3619 * max_active is bumped. It's a slow path. Do it always.
3621 wake_up_worker(pwq->pool);
3623 pwq->max_active = 0;
3626 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3629 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3630 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3631 struct worker_pool *pool)
3633 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3635 memset(pwq, 0, sizeof(*pwq));
3639 pwq->flush_color = -1;
3641 INIT_LIST_HEAD(&pwq->delayed_works);
3642 INIT_LIST_HEAD(&pwq->pwqs_node);
3643 INIT_LIST_HEAD(&pwq->mayday_node);
3644 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3647 /* sync @pwq with the current state of its associated wq and link it */
3648 static void link_pwq(struct pool_workqueue *pwq)
3650 struct workqueue_struct *wq = pwq->wq;
3652 lockdep_assert_held(&wq->mutex);
3654 /* may be called multiple times, ignore if already linked */
3655 if (!list_empty(&pwq->pwqs_node))
3658 /* set the matching work_color */
3659 pwq->work_color = wq->work_color;
3661 /* sync max_active to the current setting */
3662 pwq_adjust_max_active(pwq);
3665 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3668 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3669 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3670 const struct workqueue_attrs *attrs)
3672 struct worker_pool *pool;
3673 struct pool_workqueue *pwq;
3675 lockdep_assert_held(&wq_pool_mutex);
3677 pool = get_unbound_pool(attrs);
3681 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3683 put_unbound_pool(pool);
3687 init_pwq(pwq, wq, pool);
3692 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3693 * @attrs: the wq_attrs of the default pwq of the target workqueue
3694 * @node: the target NUMA node
3695 * @cpu_going_down: if >= 0, the CPU to consider as offline
3696 * @cpumask: outarg, the resulting cpumask
3698 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3699 * @cpu_going_down is >= 0, that cpu is considered offline during
3700 * calculation. The result is stored in @cpumask.
3702 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3703 * enabled and @node has online CPUs requested by @attrs, the returned
3704 * cpumask is the intersection of the possible CPUs of @node and
3707 * The caller is responsible for ensuring that the cpumask of @node stays
3710 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3713 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3714 int cpu_going_down, cpumask_t *cpumask)
3716 if (!wq_numa_enabled || attrs->no_numa)
3719 /* does @node have any online CPUs @attrs wants? */
3720 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3721 if (cpu_going_down >= 0)
3722 cpumask_clear_cpu(cpu_going_down, cpumask);
3724 if (cpumask_empty(cpumask))
3727 /* yeap, return possible CPUs in @node that @attrs wants */
3728 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3730 if (cpumask_empty(cpumask)) {
3731 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3732 "possible intersect\n");
3736 return !cpumask_equal(cpumask, attrs->cpumask);
3739 cpumask_copy(cpumask, attrs->cpumask);
3743 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3744 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3746 struct pool_workqueue *pwq)
3748 struct pool_workqueue *old_pwq;
3750 lockdep_assert_held(&wq_pool_mutex);
3751 lockdep_assert_held(&wq->mutex);
3753 /* link_pwq() can handle duplicate calls */
3756 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3757 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3761 /* context to store the prepared attrs & pwqs before applying */
3762 struct apply_wqattrs_ctx {
3763 struct workqueue_struct *wq; /* target workqueue */
3764 struct workqueue_attrs *attrs; /* attrs to apply */
3765 struct list_head list; /* queued for batching commit */
3766 struct pool_workqueue *dfl_pwq;
3767 struct pool_workqueue *pwq_tbl[];
3770 /* free the resources after success or abort */
3771 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3777 put_pwq_unlocked(ctx->pwq_tbl[node]);
3778 put_pwq_unlocked(ctx->dfl_pwq);
3780 free_workqueue_attrs(ctx->attrs);
3786 /* allocate the attrs and pwqs for later installation */
3787 static struct apply_wqattrs_ctx *
3788 apply_wqattrs_prepare(struct workqueue_struct *wq,
3789 const struct workqueue_attrs *attrs)
3791 struct apply_wqattrs_ctx *ctx;
3792 struct workqueue_attrs *new_attrs, *tmp_attrs;
3795 lockdep_assert_held(&wq_pool_mutex);
3797 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3799 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3800 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3801 if (!ctx || !new_attrs || !tmp_attrs)
3805 * Calculate the attrs of the default pwq.
3806 * If the user configured cpumask doesn't overlap with the
3807 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3809 copy_workqueue_attrs(new_attrs, attrs);
3810 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3811 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3812 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3815 * We may create multiple pwqs with differing cpumasks. Make a
3816 * copy of @new_attrs which will be modified and used to obtain
3819 copy_workqueue_attrs(tmp_attrs, new_attrs);
3822 * If something goes wrong during CPU up/down, we'll fall back to
3823 * the default pwq covering whole @attrs->cpumask. Always create
3824 * it even if we don't use it immediately.
3826 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3830 for_each_node(node) {
3831 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3832 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3833 if (!ctx->pwq_tbl[node])
3836 ctx->dfl_pwq->refcnt++;
3837 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3841 /* save the user configured attrs and sanitize it. */
3842 copy_workqueue_attrs(new_attrs, attrs);
3843 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3844 ctx->attrs = new_attrs;
3847 free_workqueue_attrs(tmp_attrs);
3851 free_workqueue_attrs(tmp_attrs);
3852 free_workqueue_attrs(new_attrs);
3853 apply_wqattrs_cleanup(ctx);
3857 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3858 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3862 /* all pwqs have been created successfully, let's install'em */
3863 mutex_lock(&ctx->wq->mutex);
3865 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3867 /* save the previous pwq and install the new one */
3869 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3870 ctx->pwq_tbl[node]);
3872 /* @dfl_pwq might not have been used, ensure it's linked */
3873 link_pwq(ctx->dfl_pwq);
3874 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3876 mutex_unlock(&ctx->wq->mutex);
3879 static void apply_wqattrs_lock(void)
3881 /* CPUs should stay stable across pwq creations and installations */
3883 mutex_lock(&wq_pool_mutex);
3886 static void apply_wqattrs_unlock(void)
3888 mutex_unlock(&wq_pool_mutex);
3892 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3893 const struct workqueue_attrs *attrs)
3895 struct apply_wqattrs_ctx *ctx;
3897 /* only unbound workqueues can change attributes */
3898 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3901 /* creating multiple pwqs breaks ordering guarantee */
3902 if (!list_empty(&wq->pwqs)) {
3903 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3906 wq->flags &= ~__WQ_ORDERED;
3909 ctx = apply_wqattrs_prepare(wq, attrs);
3913 /* the ctx has been prepared successfully, let's commit it */
3914 apply_wqattrs_commit(ctx);
3915 apply_wqattrs_cleanup(ctx);
3921 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3922 * @wq: the target workqueue
3923 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3925 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3926 * machines, this function maps a separate pwq to each NUMA node with
3927 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3928 * NUMA node it was issued on. Older pwqs are released as in-flight work
3929 * items finish. Note that a work item which repeatedly requeues itself
3930 * back-to-back will stay on its current pwq.
3932 * Performs GFP_KERNEL allocations.
3934 * Return: 0 on success and -errno on failure.
3936 int apply_workqueue_attrs(struct workqueue_struct *wq,
3937 const struct workqueue_attrs *attrs)
3941 apply_wqattrs_lock();
3942 ret = apply_workqueue_attrs_locked(wq, attrs);
3943 apply_wqattrs_unlock();
3947 EXPORT_SYMBOL_GPL(apply_workqueue_attrs);
3950 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3951 * @wq: the target workqueue
3952 * @cpu: the CPU coming up or going down
3953 * @online: whether @cpu is coming up or going down
3955 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3956 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3959 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3960 * falls back to @wq->dfl_pwq which may not be optimal but is always
3963 * Note that when the last allowed CPU of a NUMA node goes offline for a
3964 * workqueue with a cpumask spanning multiple nodes, the workers which were
3965 * already executing the work items for the workqueue will lose their CPU
3966 * affinity and may execute on any CPU. This is similar to how per-cpu
3967 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3968 * affinity, it's the user's responsibility to flush the work item from
3971 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3974 int node = cpu_to_node(cpu);
3975 int cpu_off = online ? -1 : cpu;
3976 struct pool_workqueue *old_pwq = NULL, *pwq;
3977 struct workqueue_attrs *target_attrs;
3980 lockdep_assert_held(&wq_pool_mutex);
3982 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3983 wq->unbound_attrs->no_numa)
3987 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3988 * Let's use a preallocated one. The following buf is protected by
3989 * CPU hotplug exclusion.
3991 target_attrs = wq_update_unbound_numa_attrs_buf;
3992 cpumask = target_attrs->cpumask;
3994 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3995 pwq = unbound_pwq_by_node(wq, node);
3998 * Let's determine what needs to be done. If the target cpumask is
3999 * different from the default pwq's, we need to compare it to @pwq's
4000 * and create a new one if they don't match. If the target cpumask
4001 * equals the default pwq's, the default pwq should be used.
4003 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4004 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4010 /* create a new pwq */
4011 pwq = alloc_unbound_pwq(wq, target_attrs);
4013 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4018 /* Install the new pwq. */
4019 mutex_lock(&wq->mutex);
4020 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4024 mutex_lock(&wq->mutex);
4025 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4026 get_pwq(wq->dfl_pwq);
4027 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4028 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4030 mutex_unlock(&wq->mutex);
4031 put_pwq_unlocked(old_pwq);
4034 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4036 bool highpri = wq->flags & WQ_HIGHPRI;
4039 if (!(wq->flags & WQ_UNBOUND)) {
4040 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4044 for_each_possible_cpu(cpu) {
4045 struct pool_workqueue *pwq =
4046 per_cpu_ptr(wq->cpu_pwqs, cpu);
4047 struct worker_pool *cpu_pools =
4048 per_cpu(cpu_worker_pools, cpu);
4050 init_pwq(pwq, wq, &cpu_pools[highpri]);
4052 mutex_lock(&wq->mutex);
4054 mutex_unlock(&wq->mutex);
4057 } else if (wq->flags & __WQ_ORDERED) {
4058 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4059 /* there should only be single pwq for ordering guarantee */
4060 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4061 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4062 "ordering guarantee broken for workqueue %s\n", wq->name);
4065 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4069 static int wq_clamp_max_active(int max_active, unsigned int flags,
4072 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4074 if (max_active < 1 || max_active > lim)
4075 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4076 max_active, name, 1, lim);
4078 return clamp_val(max_active, 1, lim);
4082 * Workqueues which may be used during memory reclaim should have a rescuer
4083 * to guarantee forward progress.
4085 static int init_rescuer(struct workqueue_struct *wq)
4087 struct worker *rescuer;
4090 if (!(wq->flags & WQ_MEM_RECLAIM))
4093 rescuer = alloc_worker(NUMA_NO_NODE);
4097 rescuer->rescue_wq = wq;
4098 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4099 ret = PTR_ERR_OR_ZERO(rescuer->task);
4105 wq->rescuer = rescuer;
4106 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4107 wake_up_process(rescuer->task);
4112 struct workqueue_struct *alloc_workqueue(const char *fmt,
4114 int max_active, ...)
4116 size_t tbl_size = 0;
4118 struct workqueue_struct *wq;
4119 struct pool_workqueue *pwq;
4122 * Unbound && max_active == 1 used to imply ordered, which is no
4123 * longer the case on NUMA machines due to per-node pools. While
4124 * alloc_ordered_workqueue() is the right way to create an ordered
4125 * workqueue, keep the previous behavior to avoid subtle breakages
4128 if ((flags & WQ_UNBOUND) && max_active == 1)
4129 flags |= __WQ_ORDERED;
4131 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4132 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4133 flags |= WQ_UNBOUND;
4135 /* allocate wq and format name */
4136 if (flags & WQ_UNBOUND)
4137 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4139 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4143 if (flags & WQ_UNBOUND) {
4144 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4145 if (!wq->unbound_attrs)
4149 va_start(args, max_active);
4150 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4153 max_active = max_active ?: WQ_DFL_ACTIVE;
4154 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4158 wq->saved_max_active = max_active;
4159 mutex_init(&wq->mutex);
4160 atomic_set(&wq->nr_pwqs_to_flush, 0);
4161 INIT_LIST_HEAD(&wq->pwqs);
4162 INIT_LIST_HEAD(&wq->flusher_queue);
4163 INIT_LIST_HEAD(&wq->flusher_overflow);
4164 INIT_LIST_HEAD(&wq->maydays);
4166 wq_init_lockdep(wq);
4167 INIT_LIST_HEAD(&wq->list);
4169 if (alloc_and_link_pwqs(wq) < 0)
4172 if (wq_online && init_rescuer(wq) < 0)
4175 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4179 * wq_pool_mutex protects global freeze state and workqueues list.
4180 * Grab it, adjust max_active and add the new @wq to workqueues
4183 mutex_lock(&wq_pool_mutex);
4185 mutex_lock(&wq->mutex);
4186 for_each_pwq(pwq, wq)
4187 pwq_adjust_max_active(pwq);
4188 mutex_unlock(&wq->mutex);
4190 list_add_tail_rcu(&wq->list, &workqueues);
4192 mutex_unlock(&wq_pool_mutex);
4197 wq_unregister_lockdep(wq);
4198 wq_free_lockdep(wq);
4199 free_workqueue_attrs(wq->unbound_attrs);
4203 destroy_workqueue(wq);
4206 EXPORT_SYMBOL_GPL(alloc_workqueue);
4209 * destroy_workqueue - safely terminate a workqueue
4210 * @wq: target workqueue
4212 * Safely destroy a workqueue. All work currently pending will be done first.
4214 void destroy_workqueue(struct workqueue_struct *wq)
4216 struct pool_workqueue *pwq;
4219 /* drain it before proceeding with destruction */
4220 drain_workqueue(wq);
4223 mutex_lock(&wq->mutex);
4224 for_each_pwq(pwq, wq) {
4227 for (i = 0; i < WORK_NR_COLORS; i++) {
4228 if (WARN_ON(pwq->nr_in_flight[i])) {
4229 mutex_unlock(&wq->mutex);
4230 show_workqueue_state();
4235 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4236 WARN_ON(pwq->nr_active) ||
4237 WARN_ON(!list_empty(&pwq->delayed_works))) {
4238 mutex_unlock(&wq->mutex);
4239 show_workqueue_state();
4243 mutex_unlock(&wq->mutex);
4246 * wq list is used to freeze wq, remove from list after
4247 * flushing is complete in case freeze races us.
4249 mutex_lock(&wq_pool_mutex);
4250 list_del_rcu(&wq->list);
4251 mutex_unlock(&wq_pool_mutex);
4253 workqueue_sysfs_unregister(wq);
4256 kthread_stop(wq->rescuer->task);
4258 if (!(wq->flags & WQ_UNBOUND)) {
4259 wq_unregister_lockdep(wq);
4261 * The base ref is never dropped on per-cpu pwqs. Directly
4262 * schedule RCU free.
4264 call_rcu(&wq->rcu, rcu_free_wq);
4267 * We're the sole accessor of @wq at this point. Directly
4268 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4269 * @wq will be freed when the last pwq is released.
4271 for_each_node(node) {
4272 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4273 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4274 put_pwq_unlocked(pwq);
4278 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4279 * put. Don't access it afterwards.
4283 put_pwq_unlocked(pwq);
4286 EXPORT_SYMBOL_GPL(destroy_workqueue);
4289 * workqueue_set_max_active - adjust max_active of a workqueue
4290 * @wq: target workqueue
4291 * @max_active: new max_active value.
4293 * Set max_active of @wq to @max_active.
4296 * Don't call from IRQ context.
4298 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4300 struct pool_workqueue *pwq;
4302 /* disallow meddling with max_active for ordered workqueues */
4303 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4306 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4308 mutex_lock(&wq->mutex);
4310 wq->flags &= ~__WQ_ORDERED;
4311 wq->saved_max_active = max_active;
4313 for_each_pwq(pwq, wq)
4314 pwq_adjust_max_active(pwq);
4316 mutex_unlock(&wq->mutex);
4318 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4321 * current_work - retrieve %current task's work struct
4323 * Determine if %current task is a workqueue worker and what it's working on.
4324 * Useful to find out the context that the %current task is running in.
4326 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4328 struct work_struct *current_work(void)
4330 struct worker *worker = current_wq_worker();
4332 return worker ? worker->current_work : NULL;
4334 EXPORT_SYMBOL(current_work);
4337 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4339 * Determine whether %current is a workqueue rescuer. Can be used from
4340 * work functions to determine whether it's being run off the rescuer task.
4342 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4344 bool current_is_workqueue_rescuer(void)
4346 struct worker *worker = current_wq_worker();
4348 return worker && worker->rescue_wq;
4352 * workqueue_congested - test whether a workqueue is congested
4353 * @cpu: CPU in question
4354 * @wq: target workqueue
4356 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4357 * no synchronization around this function and the test result is
4358 * unreliable and only useful as advisory hints or for debugging.
4360 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4361 * Note that both per-cpu and unbound workqueues may be associated with
4362 * multiple pool_workqueues which have separate congested states. A
4363 * workqueue being congested on one CPU doesn't mean the workqueue is also
4364 * contested on other CPUs / NUMA nodes.
4367 * %true if congested, %false otherwise.
4369 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4371 struct pool_workqueue *pwq;
4374 rcu_read_lock_sched();
4376 if (cpu == WORK_CPU_UNBOUND)
4377 cpu = smp_processor_id();
4379 if (!(wq->flags & WQ_UNBOUND))
4380 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4382 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4384 ret = !list_empty(&pwq->delayed_works);
4385 rcu_read_unlock_sched();
4389 EXPORT_SYMBOL_GPL(workqueue_congested);
4392 * work_busy - test whether a work is currently pending or running
4393 * @work: the work to be tested
4395 * Test whether @work is currently pending or running. There is no
4396 * synchronization around this function and the test result is
4397 * unreliable and only useful as advisory hints or for debugging.
4400 * OR'd bitmask of WORK_BUSY_* bits.
4402 unsigned int work_busy(struct work_struct *work)
4404 struct worker_pool *pool;
4405 unsigned long flags;
4406 unsigned int ret = 0;
4408 if (work_pending(work))
4409 ret |= WORK_BUSY_PENDING;
4411 local_irq_save(flags);
4412 pool = get_work_pool(work);
4414 spin_lock(&pool->lock);
4415 if (find_worker_executing_work(pool, work))
4416 ret |= WORK_BUSY_RUNNING;
4417 spin_unlock(&pool->lock);
4419 local_irq_restore(flags);
4423 EXPORT_SYMBOL_GPL(work_busy);
4426 * set_worker_desc - set description for the current work item
4427 * @fmt: printf-style format string
4428 * @...: arguments for the format string
4430 * This function can be called by a running work function to describe what
4431 * the work item is about. If the worker task gets dumped, this
4432 * information will be printed out together to help debugging. The
4433 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4435 void set_worker_desc(const char *fmt, ...)
4437 struct worker *worker = current_wq_worker();
4441 va_start(args, fmt);
4442 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4446 EXPORT_SYMBOL_GPL(set_worker_desc);
4449 * print_worker_info - print out worker information and description
4450 * @log_lvl: the log level to use when printing
4451 * @task: target task
4453 * If @task is a worker and currently executing a work item, print out the
4454 * name of the workqueue being serviced and worker description set with
4455 * set_worker_desc() by the currently executing work item.
4457 * This function can be safely called on any task as long as the
4458 * task_struct itself is accessible. While safe, this function isn't
4459 * synchronized and may print out mixups or garbages of limited length.
4461 void print_worker_info(const char *log_lvl, struct task_struct *task)
4463 work_func_t *fn = NULL;
4464 char name[WQ_NAME_LEN] = { };
4465 char desc[WORKER_DESC_LEN] = { };
4466 struct pool_workqueue *pwq = NULL;
4467 struct workqueue_struct *wq = NULL;
4468 struct worker *worker;
4470 if (!(task->flags & PF_WQ_WORKER))
4474 * This function is called without any synchronization and @task
4475 * could be in any state. Be careful with dereferences.
4477 worker = kthread_probe_data(task);
4480 * Carefully copy the associated workqueue's workfn, name and desc.
4481 * Keep the original last '\0' in case the original is garbage.
4483 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4484 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4485 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4486 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4487 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4489 if (fn || name[0] || desc[0]) {
4490 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4491 if (strcmp(name, desc))
4492 pr_cont(" (%s)", desc);
4497 static void pr_cont_pool_info(struct worker_pool *pool)
4499 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4500 if (pool->node != NUMA_NO_NODE)
4501 pr_cont(" node=%d", pool->node);
4502 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4505 static void pr_cont_work(bool comma, struct work_struct *work)
4507 if (work->func == wq_barrier_func) {
4508 struct wq_barrier *barr;
4510 barr = container_of(work, struct wq_barrier, work);
4512 pr_cont("%s BAR(%d)", comma ? "," : "",
4513 task_pid_nr(barr->task));
4515 pr_cont("%s %pf", comma ? "," : "", work->func);
4519 static void show_pwq(struct pool_workqueue *pwq)
4521 struct worker_pool *pool = pwq->pool;
4522 struct work_struct *work;
4523 struct worker *worker;
4524 bool has_in_flight = false, has_pending = false;
4527 pr_info(" pwq %d:", pool->id);
4528 pr_cont_pool_info(pool);
4530 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4531 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4533 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4534 if (worker->current_pwq == pwq) {
4535 has_in_flight = true;
4539 if (has_in_flight) {
4542 pr_info(" in-flight:");
4543 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4544 if (worker->current_pwq != pwq)
4547 pr_cont("%s %d%s:%pf", comma ? "," : "",
4548 task_pid_nr(worker->task),
4549 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4550 worker->current_func);
4551 list_for_each_entry(work, &worker->scheduled, entry)
4552 pr_cont_work(false, work);
4558 list_for_each_entry(work, &pool->worklist, entry) {
4559 if (get_work_pwq(work) == pwq) {
4567 pr_info(" pending:");
4568 list_for_each_entry(work, &pool->worklist, entry) {
4569 if (get_work_pwq(work) != pwq)
4572 pr_cont_work(comma, work);
4573 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4578 if (!list_empty(&pwq->delayed_works)) {
4581 pr_info(" delayed:");
4582 list_for_each_entry(work, &pwq->delayed_works, entry) {
4583 pr_cont_work(comma, work);
4584 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4591 * show_workqueue_state - dump workqueue state
4593 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4594 * all busy workqueues and pools.
4596 void show_workqueue_state(void)
4598 struct workqueue_struct *wq;
4599 struct worker_pool *pool;
4600 unsigned long flags;
4603 rcu_read_lock_sched();
4605 pr_info("Showing busy workqueues and worker pools:\n");
4607 list_for_each_entry_rcu(wq, &workqueues, list) {
4608 struct pool_workqueue *pwq;
4611 for_each_pwq(pwq, wq) {
4612 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4620 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4622 for_each_pwq(pwq, wq) {
4623 spin_lock_irqsave(&pwq->pool->lock, flags);
4624 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4626 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4628 * We could be printing a lot from atomic context, e.g.
4629 * sysrq-t -> show_workqueue_state(). Avoid triggering
4632 touch_nmi_watchdog();
4636 for_each_pool(pool, pi) {
4637 struct worker *worker;
4640 spin_lock_irqsave(&pool->lock, flags);
4641 if (pool->nr_workers == pool->nr_idle)
4644 pr_info("pool %d:", pool->id);
4645 pr_cont_pool_info(pool);
4646 pr_cont(" hung=%us workers=%d",
4647 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4650 pr_cont(" manager: %d",
4651 task_pid_nr(pool->manager->task));
4652 list_for_each_entry(worker, &pool->idle_list, entry) {
4653 pr_cont(" %s%d", first ? "idle: " : "",
4654 task_pid_nr(worker->task));
4659 spin_unlock_irqrestore(&pool->lock, flags);
4661 * We could be printing a lot from atomic context, e.g.
4662 * sysrq-t -> show_workqueue_state(). Avoid triggering
4665 touch_nmi_watchdog();
4668 rcu_read_unlock_sched();
4671 /* used to show worker information through /proc/PID/{comm,stat,status} */
4672 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4676 /* always show the actual comm */
4677 off = strscpy(buf, task->comm, size);
4681 /* stabilize PF_WQ_WORKER and worker pool association */
4682 mutex_lock(&wq_pool_attach_mutex);
4684 if (task->flags & PF_WQ_WORKER) {
4685 struct worker *worker = kthread_data(task);
4686 struct worker_pool *pool = worker->pool;
4689 spin_lock_irq(&pool->lock);
4691 * ->desc tracks information (wq name or
4692 * set_worker_desc()) for the latest execution. If
4693 * current, prepend '+', otherwise '-'.
4695 if (worker->desc[0] != '\0') {
4696 if (worker->current_work)
4697 scnprintf(buf + off, size - off, "+%s",
4700 scnprintf(buf + off, size - off, "-%s",
4703 spin_unlock_irq(&pool->lock);
4707 mutex_unlock(&wq_pool_attach_mutex);
4715 * There are two challenges in supporting CPU hotplug. Firstly, there
4716 * are a lot of assumptions on strong associations among work, pwq and
4717 * pool which make migrating pending and scheduled works very
4718 * difficult to implement without impacting hot paths. Secondly,
4719 * worker pools serve mix of short, long and very long running works making
4720 * blocked draining impractical.
4722 * This is solved by allowing the pools to be disassociated from the CPU
4723 * running as an unbound one and allowing it to be reattached later if the
4724 * cpu comes back online.
4727 static void unbind_workers(int cpu)
4729 struct worker_pool *pool;
4730 struct worker *worker;
4732 for_each_cpu_worker_pool(pool, cpu) {
4733 mutex_lock(&wq_pool_attach_mutex);
4734 spin_lock_irq(&pool->lock);
4737 * We've blocked all attach/detach operations. Make all workers
4738 * unbound and set DISASSOCIATED. Before this, all workers
4739 * except for the ones which are still executing works from
4740 * before the last CPU down must be on the cpu. After
4741 * this, they may become diasporas.
4743 for_each_pool_worker(worker, pool)
4744 worker->flags |= WORKER_UNBOUND;
4746 pool->flags |= POOL_DISASSOCIATED;
4748 spin_unlock_irq(&pool->lock);
4749 mutex_unlock(&wq_pool_attach_mutex);
4752 * Call schedule() so that we cross rq->lock and thus can
4753 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4754 * This is necessary as scheduler callbacks may be invoked
4760 * Sched callbacks are disabled now. Zap nr_running.
4761 * After this, nr_running stays zero and need_more_worker()
4762 * and keep_working() are always true as long as the
4763 * worklist is not empty. This pool now behaves as an
4764 * unbound (in terms of concurrency management) pool which
4765 * are served by workers tied to the pool.
4767 atomic_set(&pool->nr_running, 0);
4770 * With concurrency management just turned off, a busy
4771 * worker blocking could lead to lengthy stalls. Kick off
4772 * unbound chain execution of currently pending work items.
4774 spin_lock_irq(&pool->lock);
4775 wake_up_worker(pool);
4776 spin_unlock_irq(&pool->lock);
4781 * rebind_workers - rebind all workers of a pool to the associated CPU
4782 * @pool: pool of interest
4784 * @pool->cpu is coming online. Rebind all workers to the CPU.
4786 static void rebind_workers(struct worker_pool *pool)
4788 struct worker *worker;
4790 lockdep_assert_held(&wq_pool_attach_mutex);
4793 * Restore CPU affinity of all workers. As all idle workers should
4794 * be on the run-queue of the associated CPU before any local
4795 * wake-ups for concurrency management happen, restore CPU affinity
4796 * of all workers first and then clear UNBOUND. As we're called
4797 * from CPU_ONLINE, the following shouldn't fail.
4799 for_each_pool_worker(worker, pool)
4800 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4801 pool->attrs->cpumask) < 0);
4803 spin_lock_irq(&pool->lock);
4805 pool->flags &= ~POOL_DISASSOCIATED;
4807 for_each_pool_worker(worker, pool) {
4808 unsigned int worker_flags = worker->flags;
4811 * A bound idle worker should actually be on the runqueue
4812 * of the associated CPU for local wake-ups targeting it to
4813 * work. Kick all idle workers so that they migrate to the
4814 * associated CPU. Doing this in the same loop as
4815 * replacing UNBOUND with REBOUND is safe as no worker will
4816 * be bound before @pool->lock is released.
4818 if (worker_flags & WORKER_IDLE)
4819 wake_up_process(worker->task);
4822 * We want to clear UNBOUND but can't directly call
4823 * worker_clr_flags() or adjust nr_running. Atomically
4824 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4825 * @worker will clear REBOUND using worker_clr_flags() when
4826 * it initiates the next execution cycle thus restoring
4827 * concurrency management. Note that when or whether
4828 * @worker clears REBOUND doesn't affect correctness.
4830 * WRITE_ONCE() is necessary because @worker->flags may be
4831 * tested without holding any lock in
4832 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4833 * fail incorrectly leading to premature concurrency
4834 * management operations.
4836 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4837 worker_flags |= WORKER_REBOUND;
4838 worker_flags &= ~WORKER_UNBOUND;
4839 WRITE_ONCE(worker->flags, worker_flags);
4842 spin_unlock_irq(&pool->lock);
4846 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4847 * @pool: unbound pool of interest
4848 * @cpu: the CPU which is coming up
4850 * An unbound pool may end up with a cpumask which doesn't have any online
4851 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4852 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4853 * online CPU before, cpus_allowed of all its workers should be restored.
4855 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4857 static cpumask_t cpumask;
4858 struct worker *worker;
4860 lockdep_assert_held(&wq_pool_attach_mutex);
4862 /* is @cpu allowed for @pool? */
4863 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4866 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4868 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4869 for_each_pool_worker(worker, pool)
4870 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4873 int workqueue_prepare_cpu(unsigned int cpu)
4875 struct worker_pool *pool;
4877 for_each_cpu_worker_pool(pool, cpu) {
4878 if (pool->nr_workers)
4880 if (!create_worker(pool))
4886 int workqueue_online_cpu(unsigned int cpu)
4888 struct worker_pool *pool;
4889 struct workqueue_struct *wq;
4892 mutex_lock(&wq_pool_mutex);
4894 for_each_pool(pool, pi) {
4895 mutex_lock(&wq_pool_attach_mutex);
4897 if (pool->cpu == cpu)
4898 rebind_workers(pool);
4899 else if (pool->cpu < 0)
4900 restore_unbound_workers_cpumask(pool, cpu);
4902 mutex_unlock(&wq_pool_attach_mutex);
4905 /* update NUMA affinity of unbound workqueues */
4906 list_for_each_entry(wq, &workqueues, list)
4907 wq_update_unbound_numa(wq, cpu, true);
4909 mutex_unlock(&wq_pool_mutex);
4913 int workqueue_offline_cpu(unsigned int cpu)
4915 struct workqueue_struct *wq;
4917 /* unbinding per-cpu workers should happen on the local CPU */
4918 if (WARN_ON(cpu != smp_processor_id()))
4921 unbind_workers(cpu);
4923 /* update NUMA affinity of unbound workqueues */
4924 mutex_lock(&wq_pool_mutex);
4925 list_for_each_entry(wq, &workqueues, list)
4926 wq_update_unbound_numa(wq, cpu, false);
4927 mutex_unlock(&wq_pool_mutex);
4932 struct work_for_cpu {
4933 struct work_struct work;
4939 static void work_for_cpu_fn(struct work_struct *work)
4941 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4943 wfc->ret = wfc->fn(wfc->arg);
4947 * work_on_cpu - run a function in thread context on a particular cpu
4948 * @cpu: the cpu to run on
4949 * @fn: the function to run
4950 * @arg: the function arg
4952 * It is up to the caller to ensure that the cpu doesn't go offline.
4953 * The caller must not hold any locks which would prevent @fn from completing.
4955 * Return: The value @fn returns.
4957 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4959 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4961 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4962 schedule_work_on(cpu, &wfc.work);
4963 flush_work(&wfc.work);
4964 destroy_work_on_stack(&wfc.work);
4967 EXPORT_SYMBOL_GPL(work_on_cpu);
4970 * work_on_cpu_safe - run a function in thread context on a particular cpu
4971 * @cpu: the cpu to run on
4972 * @fn: the function to run
4973 * @arg: the function argument
4975 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4976 * any locks which would prevent @fn from completing.
4978 * Return: The value @fn returns.
4980 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4985 if (cpu_online(cpu))
4986 ret = work_on_cpu(cpu, fn, arg);
4990 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
4991 #endif /* CONFIG_SMP */
4993 #ifdef CONFIG_FREEZER
4996 * freeze_workqueues_begin - begin freezing workqueues
4998 * Start freezing workqueues. After this function returns, all freezable
4999 * workqueues will queue new works to their delayed_works list instead of
5003 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5005 void freeze_workqueues_begin(void)
5007 struct workqueue_struct *wq;
5008 struct pool_workqueue *pwq;
5010 mutex_lock(&wq_pool_mutex);
5012 WARN_ON_ONCE(workqueue_freezing);
5013 workqueue_freezing = true;
5015 list_for_each_entry(wq, &workqueues, list) {
5016 mutex_lock(&wq->mutex);
5017 for_each_pwq(pwq, wq)
5018 pwq_adjust_max_active(pwq);
5019 mutex_unlock(&wq->mutex);
5022 mutex_unlock(&wq_pool_mutex);
5026 * freeze_workqueues_busy - are freezable workqueues still busy?
5028 * Check whether freezing is complete. This function must be called
5029 * between freeze_workqueues_begin() and thaw_workqueues().
5032 * Grabs and releases wq_pool_mutex.
5035 * %true if some freezable workqueues are still busy. %false if freezing
5038 bool freeze_workqueues_busy(void)
5041 struct workqueue_struct *wq;
5042 struct pool_workqueue *pwq;
5044 mutex_lock(&wq_pool_mutex);
5046 WARN_ON_ONCE(!workqueue_freezing);
5048 list_for_each_entry(wq, &workqueues, list) {
5049 if (!(wq->flags & WQ_FREEZABLE))
5052 * nr_active is monotonically decreasing. It's safe
5053 * to peek without lock.
5055 rcu_read_lock_sched();
5056 for_each_pwq(pwq, wq) {
5057 WARN_ON_ONCE(pwq->nr_active < 0);
5058 if (pwq->nr_active) {
5060 rcu_read_unlock_sched();
5064 rcu_read_unlock_sched();
5067 mutex_unlock(&wq_pool_mutex);
5072 * thaw_workqueues - thaw workqueues
5074 * Thaw workqueues. Normal queueing is restored and all collected
5075 * frozen works are transferred to their respective pool worklists.
5078 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5080 void thaw_workqueues(void)
5082 struct workqueue_struct *wq;
5083 struct pool_workqueue *pwq;
5085 mutex_lock(&wq_pool_mutex);
5087 if (!workqueue_freezing)
5090 workqueue_freezing = false;
5092 /* restore max_active and repopulate worklist */
5093 list_for_each_entry(wq, &workqueues, list) {
5094 mutex_lock(&wq->mutex);
5095 for_each_pwq(pwq, wq)
5096 pwq_adjust_max_active(pwq);
5097 mutex_unlock(&wq->mutex);
5101 mutex_unlock(&wq_pool_mutex);
5103 #endif /* CONFIG_FREEZER */
5105 static int workqueue_apply_unbound_cpumask(void)
5109 struct workqueue_struct *wq;
5110 struct apply_wqattrs_ctx *ctx, *n;
5112 lockdep_assert_held(&wq_pool_mutex);
5114 list_for_each_entry(wq, &workqueues, list) {
5115 if (!(wq->flags & WQ_UNBOUND))
5117 /* creating multiple pwqs breaks ordering guarantee */
5118 if (wq->flags & __WQ_ORDERED)
5121 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5127 list_add_tail(&ctx->list, &ctxs);
5130 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5132 apply_wqattrs_commit(ctx);
5133 apply_wqattrs_cleanup(ctx);
5140 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5141 * @cpumask: the cpumask to set
5143 * The low-level workqueues cpumask is a global cpumask that limits
5144 * the affinity of all unbound workqueues. This function check the @cpumask
5145 * and apply it to all unbound workqueues and updates all pwqs of them.
5147 * Retun: 0 - Success
5148 * -EINVAL - Invalid @cpumask
5149 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5151 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5154 cpumask_var_t saved_cpumask;
5156 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5160 * Not excluding isolated cpus on purpose.
5161 * If the user wishes to include them, we allow that.
5163 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5164 if (!cpumask_empty(cpumask)) {
5165 apply_wqattrs_lock();
5167 /* save the old wq_unbound_cpumask. */
5168 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5170 /* update wq_unbound_cpumask at first and apply it to wqs. */
5171 cpumask_copy(wq_unbound_cpumask, cpumask);
5172 ret = workqueue_apply_unbound_cpumask();
5174 /* restore the wq_unbound_cpumask when failed. */
5176 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5178 apply_wqattrs_unlock();
5181 free_cpumask_var(saved_cpumask);
5187 * Workqueues with WQ_SYSFS flag set is visible to userland via
5188 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5189 * following attributes.
5191 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5192 * max_active RW int : maximum number of in-flight work items
5194 * Unbound workqueues have the following extra attributes.
5196 * pool_ids RO int : the associated pool IDs for each node
5197 * nice RW int : nice value of the workers
5198 * cpumask RW mask : bitmask of allowed CPUs for the workers
5199 * numa RW bool : whether enable NUMA affinity
5202 struct workqueue_struct *wq;
5206 static struct workqueue_struct *dev_to_wq(struct device *dev)
5208 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5213 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5216 struct workqueue_struct *wq = dev_to_wq(dev);
5218 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5220 static DEVICE_ATTR_RO(per_cpu);
5222 static ssize_t max_active_show(struct device *dev,
5223 struct device_attribute *attr, char *buf)
5225 struct workqueue_struct *wq = dev_to_wq(dev);
5227 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5230 static ssize_t max_active_store(struct device *dev,
5231 struct device_attribute *attr, const char *buf,
5234 struct workqueue_struct *wq = dev_to_wq(dev);
5237 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5240 workqueue_set_max_active(wq, val);
5243 static DEVICE_ATTR_RW(max_active);
5245 static struct attribute *wq_sysfs_attrs[] = {
5246 &dev_attr_per_cpu.attr,
5247 &dev_attr_max_active.attr,
5250 ATTRIBUTE_GROUPS(wq_sysfs);
5252 static ssize_t wq_pool_ids_show(struct device *dev,
5253 struct device_attribute *attr, char *buf)
5255 struct workqueue_struct *wq = dev_to_wq(dev);
5256 const char *delim = "";
5257 int node, written = 0;
5259 rcu_read_lock_sched();
5260 for_each_node(node) {
5261 written += scnprintf(buf + written, PAGE_SIZE - written,
5262 "%s%d:%d", delim, node,
5263 unbound_pwq_by_node(wq, node)->pool->id);
5266 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5267 rcu_read_unlock_sched();
5272 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5275 struct workqueue_struct *wq = dev_to_wq(dev);
5278 mutex_lock(&wq->mutex);
5279 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5280 mutex_unlock(&wq->mutex);
5285 /* prepare workqueue_attrs for sysfs store operations */
5286 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5288 struct workqueue_attrs *attrs;
5290 lockdep_assert_held(&wq_pool_mutex);
5292 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5296 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5300 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5301 const char *buf, size_t count)
5303 struct workqueue_struct *wq = dev_to_wq(dev);
5304 struct workqueue_attrs *attrs;
5307 apply_wqattrs_lock();
5309 attrs = wq_sysfs_prep_attrs(wq);
5313 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5314 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5315 ret = apply_workqueue_attrs_locked(wq, attrs);
5320 apply_wqattrs_unlock();
5321 free_workqueue_attrs(attrs);
5322 return ret ?: count;
5325 static ssize_t wq_cpumask_show(struct device *dev,
5326 struct device_attribute *attr, char *buf)
5328 struct workqueue_struct *wq = dev_to_wq(dev);
5331 mutex_lock(&wq->mutex);
5332 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5333 cpumask_pr_args(wq->unbound_attrs->cpumask));
5334 mutex_unlock(&wq->mutex);
5338 static ssize_t wq_cpumask_store(struct device *dev,
5339 struct device_attribute *attr,
5340 const char *buf, size_t count)
5342 struct workqueue_struct *wq = dev_to_wq(dev);
5343 struct workqueue_attrs *attrs;
5346 apply_wqattrs_lock();
5348 attrs = wq_sysfs_prep_attrs(wq);
5352 ret = cpumask_parse(buf, attrs->cpumask);
5354 ret = apply_workqueue_attrs_locked(wq, attrs);
5357 apply_wqattrs_unlock();
5358 free_workqueue_attrs(attrs);
5359 return ret ?: count;
5362 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5365 struct workqueue_struct *wq = dev_to_wq(dev);
5368 mutex_lock(&wq->mutex);
5369 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5370 !wq->unbound_attrs->no_numa);
5371 mutex_unlock(&wq->mutex);
5376 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5377 const char *buf, size_t count)
5379 struct workqueue_struct *wq = dev_to_wq(dev);
5380 struct workqueue_attrs *attrs;
5381 int v, ret = -ENOMEM;
5383 apply_wqattrs_lock();
5385 attrs = wq_sysfs_prep_attrs(wq);
5390 if (sscanf(buf, "%d", &v) == 1) {
5391 attrs->no_numa = !v;
5392 ret = apply_workqueue_attrs_locked(wq, attrs);
5396 apply_wqattrs_unlock();
5397 free_workqueue_attrs(attrs);
5398 return ret ?: count;
5401 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5402 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5403 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5404 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5405 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5409 static struct bus_type wq_subsys = {
5410 .name = "workqueue",
5411 .dev_groups = wq_sysfs_groups,
5414 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5415 struct device_attribute *attr, char *buf)
5419 mutex_lock(&wq_pool_mutex);
5420 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5421 cpumask_pr_args(wq_unbound_cpumask));
5422 mutex_unlock(&wq_pool_mutex);
5427 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5428 struct device_attribute *attr, const char *buf, size_t count)
5430 cpumask_var_t cpumask;
5433 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5436 ret = cpumask_parse(buf, cpumask);
5438 ret = workqueue_set_unbound_cpumask(cpumask);
5440 free_cpumask_var(cpumask);
5441 return ret ? ret : count;
5444 static struct device_attribute wq_sysfs_cpumask_attr =
5445 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5446 wq_unbound_cpumask_store);
5448 static int __init wq_sysfs_init(void)
5452 err = subsys_virtual_register(&wq_subsys, NULL);
5456 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5458 core_initcall(wq_sysfs_init);
5460 static void wq_device_release(struct device *dev)
5462 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5468 * workqueue_sysfs_register - make a workqueue visible in sysfs
5469 * @wq: the workqueue to register
5471 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5472 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5473 * which is the preferred method.
5475 * Workqueue user should use this function directly iff it wants to apply
5476 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5477 * apply_workqueue_attrs() may race against userland updating the
5480 * Return: 0 on success, -errno on failure.
5482 int workqueue_sysfs_register(struct workqueue_struct *wq)
5484 struct wq_device *wq_dev;
5488 * Adjusting max_active or creating new pwqs by applying
5489 * attributes breaks ordering guarantee. Disallow exposing ordered
5492 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5495 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5500 wq_dev->dev.bus = &wq_subsys;
5501 wq_dev->dev.release = wq_device_release;
5502 dev_set_name(&wq_dev->dev, "%s", wq->name);
5505 * unbound_attrs are created separately. Suppress uevent until
5506 * everything is ready.
5508 dev_set_uevent_suppress(&wq_dev->dev, true);
5510 ret = device_register(&wq_dev->dev);
5512 put_device(&wq_dev->dev);
5517 if (wq->flags & WQ_UNBOUND) {
5518 struct device_attribute *attr;
5520 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5521 ret = device_create_file(&wq_dev->dev, attr);
5523 device_unregister(&wq_dev->dev);
5530 dev_set_uevent_suppress(&wq_dev->dev, false);
5531 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5536 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5537 * @wq: the workqueue to unregister
5539 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5541 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5543 struct wq_device *wq_dev = wq->wq_dev;
5549 device_unregister(&wq_dev->dev);
5551 #else /* CONFIG_SYSFS */
5552 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5553 #endif /* CONFIG_SYSFS */
5556 * Workqueue watchdog.
5558 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5559 * flush dependency, a concurrency managed work item which stays RUNNING
5560 * indefinitely. Workqueue stalls can be very difficult to debug as the
5561 * usual warning mechanisms don't trigger and internal workqueue state is
5564 * Workqueue watchdog monitors all worker pools periodically and dumps
5565 * state if some pools failed to make forward progress for a while where
5566 * forward progress is defined as the first item on ->worklist changing.
5568 * This mechanism is controlled through the kernel parameter
5569 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5570 * corresponding sysfs parameter file.
5572 #ifdef CONFIG_WQ_WATCHDOG
5574 static unsigned long wq_watchdog_thresh = 30;
5575 static struct timer_list wq_watchdog_timer;
5577 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5578 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5580 static void wq_watchdog_reset_touched(void)
5584 wq_watchdog_touched = jiffies;
5585 for_each_possible_cpu(cpu)
5586 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5589 static void wq_watchdog_timer_fn(struct timer_list *unused)
5591 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5592 bool lockup_detected = false;
5593 struct worker_pool *pool;
5601 for_each_pool(pool, pi) {
5602 unsigned long pool_ts, touched, ts;
5604 if (list_empty(&pool->worklist))
5607 /* get the latest of pool and touched timestamps */
5608 pool_ts = READ_ONCE(pool->watchdog_ts);
5609 touched = READ_ONCE(wq_watchdog_touched);
5611 if (time_after(pool_ts, touched))
5616 if (pool->cpu >= 0) {
5617 unsigned long cpu_touched =
5618 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5620 if (time_after(cpu_touched, ts))
5625 if (time_after(jiffies, ts + thresh)) {
5626 lockup_detected = true;
5627 pr_emerg("BUG: workqueue lockup - pool");
5628 pr_cont_pool_info(pool);
5629 pr_cont(" stuck for %us!\n",
5630 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5636 if (lockup_detected)
5637 show_workqueue_state();
5639 wq_watchdog_reset_touched();
5640 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5643 notrace void wq_watchdog_touch(int cpu)
5646 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5648 wq_watchdog_touched = jiffies;
5651 static void wq_watchdog_set_thresh(unsigned long thresh)
5653 wq_watchdog_thresh = 0;
5654 del_timer_sync(&wq_watchdog_timer);
5657 wq_watchdog_thresh = thresh;
5658 wq_watchdog_reset_touched();
5659 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5663 static int wq_watchdog_param_set_thresh(const char *val,
5664 const struct kernel_param *kp)
5666 unsigned long thresh;
5669 ret = kstrtoul(val, 0, &thresh);
5674 wq_watchdog_set_thresh(thresh);
5676 wq_watchdog_thresh = thresh;
5681 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5682 .set = wq_watchdog_param_set_thresh,
5683 .get = param_get_ulong,
5686 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5689 static void wq_watchdog_init(void)
5691 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5692 wq_watchdog_set_thresh(wq_watchdog_thresh);
5695 #else /* CONFIG_WQ_WATCHDOG */
5697 static inline void wq_watchdog_init(void) { }
5699 #endif /* CONFIG_WQ_WATCHDOG */
5701 static void __init wq_numa_init(void)
5706 if (num_possible_nodes() <= 1)
5709 if (wq_disable_numa) {
5710 pr_info("workqueue: NUMA affinity support disabled\n");
5714 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5715 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5718 * We want masks of possible CPUs of each node which isn't readily
5719 * available. Build one from cpu_to_node() which should have been
5720 * fully initialized by now.
5722 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5726 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5727 node_online(node) ? node : NUMA_NO_NODE));
5729 for_each_possible_cpu(cpu) {
5730 node = cpu_to_node(cpu);
5731 if (WARN_ON(node == NUMA_NO_NODE)) {
5732 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5733 /* happens iff arch is bonkers, let's just proceed */
5736 cpumask_set_cpu(cpu, tbl[node]);
5739 wq_numa_possible_cpumask = tbl;
5740 wq_numa_enabled = true;
5744 * workqueue_init_early - early init for workqueue subsystem
5746 * This is the first half of two-staged workqueue subsystem initialization
5747 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5748 * idr are up. It sets up all the data structures and system workqueues
5749 * and allows early boot code to create workqueues and queue/cancel work
5750 * items. Actual work item execution starts only after kthreads can be
5751 * created and scheduled right before early initcalls.
5753 int __init workqueue_init_early(void)
5755 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5756 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5759 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5761 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5762 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5764 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5766 /* initialize CPU pools */
5767 for_each_possible_cpu(cpu) {
5768 struct worker_pool *pool;
5771 for_each_cpu_worker_pool(pool, cpu) {
5772 BUG_ON(init_worker_pool(pool));
5774 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5775 pool->attrs->nice = std_nice[i++];
5776 pool->node = cpu_to_node(cpu);
5779 mutex_lock(&wq_pool_mutex);
5780 BUG_ON(worker_pool_assign_id(pool));
5781 mutex_unlock(&wq_pool_mutex);
5785 /* create default unbound and ordered wq attrs */
5786 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5787 struct workqueue_attrs *attrs;
5789 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5790 attrs->nice = std_nice[i];
5791 unbound_std_wq_attrs[i] = attrs;
5794 * An ordered wq should have only one pwq as ordering is
5795 * guaranteed by max_active which is enforced by pwqs.
5796 * Turn off NUMA so that dfl_pwq is used for all nodes.
5798 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5799 attrs->nice = std_nice[i];
5800 attrs->no_numa = true;
5801 ordered_wq_attrs[i] = attrs;
5804 system_wq = alloc_workqueue("events", 0, 0);
5805 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5806 system_long_wq = alloc_workqueue("events_long", 0, 0);
5807 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5808 WQ_UNBOUND_MAX_ACTIVE);
5809 system_freezable_wq = alloc_workqueue("events_freezable",
5811 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5812 WQ_POWER_EFFICIENT, 0);
5813 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5814 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5816 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5817 !system_unbound_wq || !system_freezable_wq ||
5818 !system_power_efficient_wq ||
5819 !system_freezable_power_efficient_wq);
5825 * workqueue_init - bring workqueue subsystem fully online
5827 * This is the latter half of two-staged workqueue subsystem initialization
5828 * and invoked as soon as kthreads can be created and scheduled.
5829 * Workqueues have been created and work items queued on them, but there
5830 * are no kworkers executing the work items yet. Populate the worker pools
5831 * with the initial workers and enable future kworker creations.
5833 int __init workqueue_init(void)
5835 struct workqueue_struct *wq;
5836 struct worker_pool *pool;
5840 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5841 * CPU to node mapping may not be available that early on some
5842 * archs such as power and arm64. As per-cpu pools created
5843 * previously could be missing node hint and unbound pools NUMA
5844 * affinity, fix them up.
5846 * Also, while iterating workqueues, create rescuers if requested.
5850 mutex_lock(&wq_pool_mutex);
5852 for_each_possible_cpu(cpu) {
5853 for_each_cpu_worker_pool(pool, cpu) {
5854 pool->node = cpu_to_node(cpu);
5858 list_for_each_entry(wq, &workqueues, list) {
5859 wq_update_unbound_numa(wq, smp_processor_id(), true);
5860 WARN(init_rescuer(wq),
5861 "workqueue: failed to create early rescuer for %s",
5865 mutex_unlock(&wq_pool_mutex);
5867 /* create the initial workers */
5868 for_each_online_cpu(cpu) {
5869 for_each_cpu_worker_pool(pool, cpu) {
5870 pool->flags &= ~POOL_DISASSOCIATED;
5871 BUG_ON(!create_worker(pool));
5875 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5876 BUG_ON(!create_worker(pool));