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 queue
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_running - a worker is running again
845 * @task: task waking up
847 * This function is called when a worker returns from schedule()
849 void wq_worker_running(struct task_struct *task)
851 struct worker *worker = kthread_data(task);
853 if (!worker->sleeping)
855 if (!(worker->flags & WORKER_NOT_RUNNING))
856 atomic_inc(&worker->pool->nr_running);
857 worker->sleeping = 0;
861 * wq_worker_sleeping - a worker is going to sleep
862 * @task: task going to sleep
864 * This function is called from schedule() when a busy worker is
867 void wq_worker_sleeping(struct task_struct *task)
869 struct worker *next, *worker = kthread_data(task);
870 struct worker_pool *pool;
873 * Rescuers, which may not have all the fields set up like normal
874 * workers, also reach here, let's not access anything before
875 * checking NOT_RUNNING.
877 if (worker->flags & WORKER_NOT_RUNNING)
882 if (WARN_ON_ONCE(worker->sleeping))
885 worker->sleeping = 1;
886 spin_lock_irq(&pool->lock);
889 * The counterpart of the following dec_and_test, implied mb,
890 * worklist not empty test sequence is in insert_work().
891 * Please read comment there.
893 * NOT_RUNNING is clear. This means that we're bound to and
894 * running on the local cpu w/ rq lock held and preemption
895 * disabled, which in turn means that none else could be
896 * manipulating idle_list, so dereferencing idle_list without pool
899 if (atomic_dec_and_test(&pool->nr_running) &&
900 !list_empty(&pool->worklist)) {
901 next = first_idle_worker(pool);
903 wake_up_process(next->task);
905 spin_unlock_irq(&pool->lock);
909 * wq_worker_last_func - retrieve worker's last work function
911 * Determine the last function a worker executed. This is called from
912 * the scheduler to get a worker's last known identity.
915 * spin_lock_irq(rq->lock)
917 * This function is called during schedule() when a kworker is going
918 * to sleep. It's used by psi to identify aggregation workers during
919 * dequeuing, to allow periodic aggregation to shut-off when that
920 * worker is the last task in the system or cgroup to go to sleep.
922 * As this function doesn't involve any workqueue-related locking, it
923 * only returns stable values when called from inside the scheduler's
924 * queuing and dequeuing paths, when @task, which must be a kworker,
925 * is guaranteed to not be processing any works.
928 * The last work function %current executed as a worker, NULL if it
929 * hasn't executed any work yet.
931 work_func_t wq_worker_last_func(struct task_struct *task)
933 struct worker *worker = kthread_data(task);
935 return worker->last_func;
939 * worker_set_flags - set worker flags and adjust nr_running accordingly
941 * @flags: flags to set
943 * Set @flags in @worker->flags and adjust nr_running accordingly.
946 * spin_lock_irq(pool->lock)
948 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
950 struct worker_pool *pool = worker->pool;
952 WARN_ON_ONCE(worker->task != current);
954 /* If transitioning into NOT_RUNNING, adjust nr_running. */
955 if ((flags & WORKER_NOT_RUNNING) &&
956 !(worker->flags & WORKER_NOT_RUNNING)) {
957 atomic_dec(&pool->nr_running);
960 worker->flags |= flags;
964 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
966 * @flags: flags to clear
968 * Clear @flags in @worker->flags and adjust nr_running accordingly.
971 * spin_lock_irq(pool->lock)
973 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
975 struct worker_pool *pool = worker->pool;
976 unsigned int oflags = worker->flags;
978 WARN_ON_ONCE(worker->task != current);
980 worker->flags &= ~flags;
983 * If transitioning out of NOT_RUNNING, increment nr_running. Note
984 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
985 * of multiple flags, not a single flag.
987 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
988 if (!(worker->flags & WORKER_NOT_RUNNING))
989 atomic_inc(&pool->nr_running);
993 * find_worker_executing_work - find worker which is executing a work
994 * @pool: pool of interest
995 * @work: work to find worker for
997 * Find a worker which is executing @work on @pool by searching
998 * @pool->busy_hash which is keyed by the address of @work. For a worker
999 * to match, its current execution should match the address of @work and
1000 * its work function. This is to avoid unwanted dependency between
1001 * unrelated work executions through a work item being recycled while still
1004 * This is a bit tricky. A work item may be freed once its execution
1005 * starts and nothing prevents the freed area from being recycled for
1006 * another work item. If the same work item address ends up being reused
1007 * before the original execution finishes, workqueue will identify the
1008 * recycled work item as currently executing and make it wait until the
1009 * current execution finishes, introducing an unwanted dependency.
1011 * This function checks the work item address and work function to avoid
1012 * false positives. Note that this isn't complete as one may construct a
1013 * work function which can introduce dependency onto itself through a
1014 * recycled work item. Well, if somebody wants to shoot oneself in the
1015 * foot that badly, there's only so much we can do, and if such deadlock
1016 * actually occurs, it should be easy to locate the culprit work function.
1019 * spin_lock_irq(pool->lock).
1022 * Pointer to worker which is executing @work if found, %NULL
1025 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1026 struct work_struct *work)
1028 struct worker *worker;
1030 hash_for_each_possible(pool->busy_hash, worker, hentry,
1031 (unsigned long)work)
1032 if (worker->current_work == work &&
1033 worker->current_func == work->func)
1040 * move_linked_works - move linked works to a list
1041 * @work: start of series of works to be scheduled
1042 * @head: target list to append @work to
1043 * @nextp: out parameter for nested worklist walking
1045 * Schedule linked works starting from @work to @head. Work series to
1046 * be scheduled starts at @work and includes any consecutive work with
1047 * WORK_STRUCT_LINKED set in its predecessor.
1049 * If @nextp is not NULL, it's updated to point to the next work of
1050 * the last scheduled work. This allows move_linked_works() to be
1051 * nested inside outer list_for_each_entry_safe().
1054 * spin_lock_irq(pool->lock).
1056 static void move_linked_works(struct work_struct *work, struct list_head *head,
1057 struct work_struct **nextp)
1059 struct work_struct *n;
1062 * Linked worklist will always end before the end of the list,
1063 * use NULL for list head.
1065 list_for_each_entry_safe_from(work, n, NULL, entry) {
1066 list_move_tail(&work->entry, head);
1067 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1072 * If we're already inside safe list traversal and have moved
1073 * multiple works to the scheduled queue, the next position
1074 * needs to be updated.
1081 * get_pwq - get an extra reference on the specified pool_workqueue
1082 * @pwq: pool_workqueue to get
1084 * Obtain an extra reference on @pwq. The caller should guarantee that
1085 * @pwq has positive refcnt and be holding the matching pool->lock.
1087 static void get_pwq(struct pool_workqueue *pwq)
1089 lockdep_assert_held(&pwq->pool->lock);
1090 WARN_ON_ONCE(pwq->refcnt <= 0);
1095 * put_pwq - put a pool_workqueue reference
1096 * @pwq: pool_workqueue to put
1098 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1099 * destruction. The caller should be holding the matching pool->lock.
1101 static void put_pwq(struct pool_workqueue *pwq)
1103 lockdep_assert_held(&pwq->pool->lock);
1104 if (likely(--pwq->refcnt))
1106 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1109 * @pwq can't be released under pool->lock, bounce to
1110 * pwq_unbound_release_workfn(). This never recurses on the same
1111 * pool->lock as this path is taken only for unbound workqueues and
1112 * the release work item is scheduled on a per-cpu workqueue. To
1113 * avoid lockdep warning, unbound pool->locks are given lockdep
1114 * subclass of 1 in get_unbound_pool().
1116 schedule_work(&pwq->unbound_release_work);
1120 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1121 * @pwq: pool_workqueue to put (can be %NULL)
1123 * put_pwq() with locking. This function also allows %NULL @pwq.
1125 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1129 * As both pwqs and pools are sched-RCU protected, the
1130 * following lock operations are safe.
1132 spin_lock_irq(&pwq->pool->lock);
1134 spin_unlock_irq(&pwq->pool->lock);
1138 static void pwq_activate_delayed_work(struct work_struct *work)
1140 struct pool_workqueue *pwq = get_work_pwq(work);
1142 trace_workqueue_activate_work(work);
1143 if (list_empty(&pwq->pool->worklist))
1144 pwq->pool->watchdog_ts = jiffies;
1145 move_linked_works(work, &pwq->pool->worklist, NULL);
1146 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1150 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1152 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1153 struct work_struct, entry);
1155 pwq_activate_delayed_work(work);
1159 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1160 * @pwq: pwq of interest
1161 * @color: color of work which left the queue
1163 * A work either has completed or is removed from pending queue,
1164 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1167 * spin_lock_irq(pool->lock).
1169 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1171 /* uncolored work items don't participate in flushing or nr_active */
1172 if (color == WORK_NO_COLOR)
1175 pwq->nr_in_flight[color]--;
1178 if (!list_empty(&pwq->delayed_works)) {
1179 /* one down, submit a delayed one */
1180 if (pwq->nr_active < pwq->max_active)
1181 pwq_activate_first_delayed(pwq);
1184 /* is flush in progress and are we at the flushing tip? */
1185 if (likely(pwq->flush_color != color))
1188 /* are there still in-flight works? */
1189 if (pwq->nr_in_flight[color])
1192 /* this pwq is done, clear flush_color */
1193 pwq->flush_color = -1;
1196 * If this was the last pwq, wake up the first flusher. It
1197 * will handle the rest.
1199 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1200 complete(&pwq->wq->first_flusher->done);
1206 * try_to_grab_pending - steal work item from worklist and disable irq
1207 * @work: work item to steal
1208 * @is_dwork: @work is a delayed_work
1209 * @flags: place to store irq state
1211 * Try to grab PENDING bit of @work. This function can handle @work in any
1212 * stable state - idle, on timer or on worklist.
1215 * 1 if @work was pending and we successfully stole PENDING
1216 * 0 if @work was idle and we claimed PENDING
1217 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1218 * -ENOENT if someone else is canceling @work, this state may persist
1219 * for arbitrarily long
1222 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1223 * interrupted while holding PENDING and @work off queue, irq must be
1224 * disabled on entry. This, combined with delayed_work->timer being
1225 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1227 * On successful return, >= 0, irq is disabled and the caller is
1228 * responsible for releasing it using local_irq_restore(*@flags).
1230 * This function is safe to call from any context including IRQ handler.
1232 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1233 unsigned long *flags)
1235 struct worker_pool *pool;
1236 struct pool_workqueue *pwq;
1238 local_irq_save(*flags);
1240 /* try to steal the timer if it exists */
1242 struct delayed_work *dwork = to_delayed_work(work);
1245 * dwork->timer is irqsafe. If del_timer() fails, it's
1246 * guaranteed that the timer is not queued anywhere and not
1247 * running on the local CPU.
1249 if (likely(del_timer(&dwork->timer)))
1253 /* try to claim PENDING the normal way */
1254 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1258 * The queueing is in progress, or it is already queued. Try to
1259 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1261 pool = get_work_pool(work);
1265 spin_lock(&pool->lock);
1267 * work->data is guaranteed to point to pwq only while the work
1268 * item is queued on pwq->wq, and both updating work->data to point
1269 * to pwq on queueing and to pool on dequeueing are done under
1270 * pwq->pool->lock. This in turn guarantees that, if work->data
1271 * points to pwq which is associated with a locked pool, the work
1272 * item is currently queued on that pool.
1274 pwq = get_work_pwq(work);
1275 if (pwq && pwq->pool == pool) {
1276 debug_work_deactivate(work);
1279 * A delayed work item cannot be grabbed directly because
1280 * it might have linked NO_COLOR work items which, if left
1281 * on the delayed_list, will confuse pwq->nr_active
1282 * management later on and cause stall. Make sure the work
1283 * item is activated before grabbing.
1285 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1286 pwq_activate_delayed_work(work);
1288 list_del_init(&work->entry);
1289 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1291 /* work->data points to pwq iff queued, point to pool */
1292 set_work_pool_and_keep_pending(work, pool->id);
1294 spin_unlock(&pool->lock);
1297 spin_unlock(&pool->lock);
1299 local_irq_restore(*flags);
1300 if (work_is_canceling(work))
1307 * insert_work - insert a work into a pool
1308 * @pwq: pwq @work belongs to
1309 * @work: work to insert
1310 * @head: insertion point
1311 * @extra_flags: extra WORK_STRUCT_* flags to set
1313 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1314 * work_struct flags.
1317 * spin_lock_irq(pool->lock).
1319 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1320 struct list_head *head, unsigned int extra_flags)
1322 struct worker_pool *pool = pwq->pool;
1324 /* we own @work, set data and link */
1325 set_work_pwq(work, pwq, extra_flags);
1326 list_add_tail(&work->entry, head);
1330 * Ensure either wq_worker_sleeping() sees the above
1331 * list_add_tail() or we see zero nr_running to avoid workers lying
1332 * around lazily while there are works to be processed.
1336 if (__need_more_worker(pool))
1337 wake_up_worker(pool);
1341 * Test whether @work is being queued from another work executing on the
1344 static bool is_chained_work(struct workqueue_struct *wq)
1346 struct worker *worker;
1348 worker = current_wq_worker();
1350 * Return %true iff I'm a worker executing a work item on @wq. If
1351 * I'm @worker, it's safe to dereference it without locking.
1353 return worker && worker->current_pwq->wq == wq;
1357 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1358 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1359 * avoid perturbing sensitive tasks.
1361 static int wq_select_unbound_cpu(int cpu)
1363 static bool printed_dbg_warning;
1366 if (likely(!wq_debug_force_rr_cpu)) {
1367 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1369 } else if (!printed_dbg_warning) {
1370 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1371 printed_dbg_warning = true;
1374 if (cpumask_empty(wq_unbound_cpumask))
1377 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1378 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1379 if (unlikely(new_cpu >= nr_cpu_ids)) {
1380 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1381 if (unlikely(new_cpu >= nr_cpu_ids))
1384 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1389 static void __queue_work(int cpu, struct workqueue_struct *wq,
1390 struct work_struct *work)
1392 struct pool_workqueue *pwq;
1393 struct worker_pool *last_pool;
1394 struct list_head *worklist;
1395 unsigned int work_flags;
1396 unsigned int req_cpu = cpu;
1399 * While a work item is PENDING && off queue, a task trying to
1400 * steal the PENDING will busy-loop waiting for it to either get
1401 * queued or lose PENDING. Grabbing PENDING and queueing should
1402 * happen with IRQ disabled.
1404 lockdep_assert_irqs_disabled();
1406 debug_work_activate(work);
1408 /* if draining, only works from the same workqueue are allowed */
1409 if (unlikely(wq->flags & __WQ_DRAINING) &&
1410 WARN_ON_ONCE(!is_chained_work(wq)))
1413 if (req_cpu == WORK_CPU_UNBOUND)
1414 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1416 /* pwq which will be used unless @work is executing elsewhere */
1417 if (!(wq->flags & WQ_UNBOUND))
1418 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1420 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1423 * If @work was previously on a different pool, it might still be
1424 * running there, in which case the work needs to be queued on that
1425 * pool to guarantee non-reentrancy.
1427 last_pool = get_work_pool(work);
1428 if (last_pool && last_pool != pwq->pool) {
1429 struct worker *worker;
1431 spin_lock(&last_pool->lock);
1433 worker = find_worker_executing_work(last_pool, work);
1435 if (worker && worker->current_pwq->wq == wq) {
1436 pwq = worker->current_pwq;
1438 /* meh... not running there, queue here */
1439 spin_unlock(&last_pool->lock);
1440 spin_lock(&pwq->pool->lock);
1443 spin_lock(&pwq->pool->lock);
1447 * pwq is determined and locked. For unbound pools, we could have
1448 * raced with pwq release and it could already be dead. If its
1449 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1450 * without another pwq replacing it in the numa_pwq_tbl or while
1451 * work items are executing on it, so the retrying is guaranteed to
1452 * make forward-progress.
1454 if (unlikely(!pwq->refcnt)) {
1455 if (wq->flags & WQ_UNBOUND) {
1456 spin_unlock(&pwq->pool->lock);
1461 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1465 /* pwq determined, queue */
1466 trace_workqueue_queue_work(req_cpu, pwq, work);
1468 if (WARN_ON(!list_empty(&work->entry))) {
1469 spin_unlock(&pwq->pool->lock);
1473 pwq->nr_in_flight[pwq->work_color]++;
1474 work_flags = work_color_to_flags(pwq->work_color);
1476 if (likely(pwq->nr_active < pwq->max_active)) {
1477 trace_workqueue_activate_work(work);
1479 worklist = &pwq->pool->worklist;
1480 if (list_empty(worklist))
1481 pwq->pool->watchdog_ts = jiffies;
1483 work_flags |= WORK_STRUCT_DELAYED;
1484 worklist = &pwq->delayed_works;
1487 insert_work(pwq, work, worklist, work_flags);
1489 spin_unlock(&pwq->pool->lock);
1493 * queue_work_on - queue work on specific cpu
1494 * @cpu: CPU number to execute work on
1495 * @wq: workqueue to use
1496 * @work: work to queue
1498 * We queue the work to a specific CPU, the caller must ensure it
1501 * Return: %false if @work was already on a queue, %true otherwise.
1503 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1504 struct work_struct *work)
1507 unsigned long flags;
1509 local_irq_save(flags);
1511 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1512 __queue_work(cpu, wq, work);
1516 local_irq_restore(flags);
1519 EXPORT_SYMBOL(queue_work_on);
1522 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1523 * @node: NUMA node ID that we want to select a CPU from
1525 * This function will attempt to find a "random" cpu available on a given
1526 * node. If there are no CPUs available on the given node it will return
1527 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1528 * available CPU if we need to schedule this work.
1530 static int workqueue_select_cpu_near(int node)
1534 /* No point in doing this if NUMA isn't enabled for workqueues */
1535 if (!wq_numa_enabled)
1536 return WORK_CPU_UNBOUND;
1538 /* Delay binding to CPU if node is not valid or online */
1539 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1540 return WORK_CPU_UNBOUND;
1542 /* Use local node/cpu if we are already there */
1543 cpu = raw_smp_processor_id();
1544 if (node == cpu_to_node(cpu))
1547 /* Use "random" otherwise know as "first" online CPU of node */
1548 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1550 /* If CPU is valid return that, otherwise just defer */
1551 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1555 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1556 * @node: NUMA node that we are targeting the work for
1557 * @wq: workqueue to use
1558 * @work: work to queue
1560 * We queue the work to a "random" CPU within a given NUMA node. The basic
1561 * idea here is to provide a way to somehow associate work with a given
1564 * This function will only make a best effort attempt at getting this onto
1565 * the right NUMA node. If no node is requested or the requested node is
1566 * offline then we just fall back to standard queue_work behavior.
1568 * Currently the "random" CPU ends up being the first available CPU in the
1569 * intersection of cpu_online_mask and the cpumask of the node, unless we
1570 * are running on the node. In that case we just use the current CPU.
1572 * Return: %false if @work was already on a queue, %true otherwise.
1574 bool queue_work_node(int node, struct workqueue_struct *wq,
1575 struct work_struct *work)
1577 unsigned long flags;
1581 * This current implementation is specific to unbound workqueues.
1582 * Specifically we only return the first available CPU for a given
1583 * node instead of cycling through individual CPUs within the node.
1585 * If this is used with a per-cpu workqueue then the logic in
1586 * workqueue_select_cpu_near would need to be updated to allow for
1587 * some round robin type logic.
1589 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1591 local_irq_save(flags);
1593 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1594 int cpu = workqueue_select_cpu_near(node);
1596 __queue_work(cpu, wq, work);
1600 local_irq_restore(flags);
1603 EXPORT_SYMBOL_GPL(queue_work_node);
1605 void delayed_work_timer_fn(struct timer_list *t)
1607 struct delayed_work *dwork = from_timer(dwork, t, timer);
1609 /* should have been called from irqsafe timer with irq already off */
1610 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1612 EXPORT_SYMBOL(delayed_work_timer_fn);
1614 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1615 struct delayed_work *dwork, unsigned long delay)
1617 struct timer_list *timer = &dwork->timer;
1618 struct work_struct *work = &dwork->work;
1621 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1622 WARN_ON_ONCE(timer_pending(timer));
1623 WARN_ON_ONCE(!list_empty(&work->entry));
1626 * If @delay is 0, queue @dwork->work immediately. This is for
1627 * both optimization and correctness. The earliest @timer can
1628 * expire is on the closest next tick and delayed_work users depend
1629 * on that there's no such delay when @delay is 0.
1632 __queue_work(cpu, wq, &dwork->work);
1638 timer->expires = jiffies + delay;
1640 if (unlikely(cpu != WORK_CPU_UNBOUND))
1641 add_timer_on(timer, cpu);
1647 * queue_delayed_work_on - queue work on specific CPU after delay
1648 * @cpu: CPU number to execute work on
1649 * @wq: workqueue to use
1650 * @dwork: work to queue
1651 * @delay: number of jiffies to wait before queueing
1653 * Return: %false if @work was already on a queue, %true otherwise. If
1654 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1657 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1658 struct delayed_work *dwork, unsigned long delay)
1660 struct work_struct *work = &dwork->work;
1662 unsigned long flags;
1664 /* read the comment in __queue_work() */
1665 local_irq_save(flags);
1667 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1668 __queue_delayed_work(cpu, wq, dwork, delay);
1672 local_irq_restore(flags);
1675 EXPORT_SYMBOL(queue_delayed_work_on);
1678 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1679 * @cpu: CPU number to execute work on
1680 * @wq: workqueue to use
1681 * @dwork: work to queue
1682 * @delay: number of jiffies to wait before queueing
1684 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1685 * modify @dwork's timer so that it expires after @delay. If @delay is
1686 * zero, @work is guaranteed to be scheduled immediately regardless of its
1689 * Return: %false if @dwork was idle and queued, %true if @dwork was
1690 * pending and its timer was modified.
1692 * This function is safe to call from any context including IRQ handler.
1693 * See try_to_grab_pending() for details.
1695 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1696 struct delayed_work *dwork, unsigned long delay)
1698 unsigned long flags;
1702 ret = try_to_grab_pending(&dwork->work, true, &flags);
1703 } while (unlikely(ret == -EAGAIN));
1705 if (likely(ret >= 0)) {
1706 __queue_delayed_work(cpu, wq, dwork, delay);
1707 local_irq_restore(flags);
1710 /* -ENOENT from try_to_grab_pending() becomes %true */
1713 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1715 static void rcu_work_rcufn(struct rcu_head *rcu)
1717 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1719 /* read the comment in __queue_work() */
1720 local_irq_disable();
1721 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1726 * queue_rcu_work - queue work after a RCU grace period
1727 * @wq: workqueue to use
1728 * @rwork: work to queue
1730 * Return: %false if @rwork was already pending, %true otherwise. Note
1731 * that a full RCU grace period is guaranteed only after a %true return.
1732 * While @rwork is guaranteed to be executed after a %false return, the
1733 * execution may happen before a full RCU grace period has passed.
1735 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1737 struct work_struct *work = &rwork->work;
1739 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1741 call_rcu(&rwork->rcu, rcu_work_rcufn);
1747 EXPORT_SYMBOL(queue_rcu_work);
1750 * worker_enter_idle - enter idle state
1751 * @worker: worker which is entering idle state
1753 * @worker is entering idle state. Update stats and idle timer if
1757 * spin_lock_irq(pool->lock).
1759 static void worker_enter_idle(struct worker *worker)
1761 struct worker_pool *pool = worker->pool;
1763 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1764 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1765 (worker->hentry.next || worker->hentry.pprev)))
1768 /* can't use worker_set_flags(), also called from create_worker() */
1769 worker->flags |= WORKER_IDLE;
1771 worker->last_active = jiffies;
1773 /* idle_list is LIFO */
1774 list_add(&worker->entry, &pool->idle_list);
1776 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1777 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1780 * Sanity check nr_running. Because unbind_workers() releases
1781 * pool->lock between setting %WORKER_UNBOUND and zapping
1782 * nr_running, the warning may trigger spuriously. Check iff
1783 * unbind is not in progress.
1785 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1786 pool->nr_workers == pool->nr_idle &&
1787 atomic_read(&pool->nr_running));
1791 * worker_leave_idle - leave idle state
1792 * @worker: worker which is leaving idle state
1794 * @worker is leaving idle state. Update stats.
1797 * spin_lock_irq(pool->lock).
1799 static void worker_leave_idle(struct worker *worker)
1801 struct worker_pool *pool = worker->pool;
1803 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1805 worker_clr_flags(worker, WORKER_IDLE);
1807 list_del_init(&worker->entry);
1810 static struct worker *alloc_worker(int node)
1812 struct worker *worker;
1814 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1816 INIT_LIST_HEAD(&worker->entry);
1817 INIT_LIST_HEAD(&worker->scheduled);
1818 INIT_LIST_HEAD(&worker->node);
1819 /* on creation a worker is in !idle && prep state */
1820 worker->flags = WORKER_PREP;
1826 * worker_attach_to_pool() - attach a worker to a pool
1827 * @worker: worker to be attached
1828 * @pool: the target pool
1830 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1831 * cpu-binding of @worker are kept coordinated with the pool across
1834 static void worker_attach_to_pool(struct worker *worker,
1835 struct worker_pool *pool)
1837 mutex_lock(&wq_pool_attach_mutex);
1840 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1841 * online CPUs. It'll be re-applied when any of the CPUs come up.
1843 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1846 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1847 * stable across this function. See the comments above the flag
1848 * definition for details.
1850 if (pool->flags & POOL_DISASSOCIATED)
1851 worker->flags |= WORKER_UNBOUND;
1853 list_add_tail(&worker->node, &pool->workers);
1854 worker->pool = pool;
1856 mutex_unlock(&wq_pool_attach_mutex);
1860 * worker_detach_from_pool() - detach a worker from its pool
1861 * @worker: worker which is attached to its pool
1863 * Undo the attaching which had been done in worker_attach_to_pool(). The
1864 * caller worker shouldn't access to the pool after detached except it has
1865 * other reference to the pool.
1867 static void worker_detach_from_pool(struct worker *worker)
1869 struct worker_pool *pool = worker->pool;
1870 struct completion *detach_completion = NULL;
1872 mutex_lock(&wq_pool_attach_mutex);
1874 list_del(&worker->node);
1875 worker->pool = NULL;
1877 if (list_empty(&pool->workers))
1878 detach_completion = pool->detach_completion;
1879 mutex_unlock(&wq_pool_attach_mutex);
1881 /* clear leftover flags without pool->lock after it is detached */
1882 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1884 if (detach_completion)
1885 complete(detach_completion);
1889 * create_worker - create a new workqueue worker
1890 * @pool: pool the new worker will belong to
1892 * Create and start a new worker which is attached to @pool.
1895 * Might sleep. Does GFP_KERNEL allocations.
1898 * Pointer to the newly created worker.
1900 static struct worker *create_worker(struct worker_pool *pool)
1902 struct worker *worker = NULL;
1906 /* ID is needed to determine kthread name */
1907 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1911 worker = alloc_worker(pool->node);
1918 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1919 pool->attrs->nice < 0 ? "H" : "");
1921 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1923 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1924 "kworker/%s", id_buf);
1925 if (IS_ERR(worker->task))
1928 set_user_nice(worker->task, pool->attrs->nice);
1929 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1931 /* successful, attach the worker to the pool */
1932 worker_attach_to_pool(worker, pool);
1934 /* start the newly created worker */
1935 spin_lock_irq(&pool->lock);
1936 worker->pool->nr_workers++;
1937 worker_enter_idle(worker);
1938 wake_up_process(worker->task);
1939 spin_unlock_irq(&pool->lock);
1945 ida_simple_remove(&pool->worker_ida, id);
1951 * destroy_worker - destroy a workqueue worker
1952 * @worker: worker to be destroyed
1954 * Destroy @worker and adjust @pool stats accordingly. The worker should
1958 * spin_lock_irq(pool->lock).
1960 static void destroy_worker(struct worker *worker)
1962 struct worker_pool *pool = worker->pool;
1964 lockdep_assert_held(&pool->lock);
1966 /* sanity check frenzy */
1967 if (WARN_ON(worker->current_work) ||
1968 WARN_ON(!list_empty(&worker->scheduled)) ||
1969 WARN_ON(!(worker->flags & WORKER_IDLE)))
1975 list_del_init(&worker->entry);
1976 worker->flags |= WORKER_DIE;
1977 wake_up_process(worker->task);
1980 static void idle_worker_timeout(struct timer_list *t)
1982 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1984 spin_lock_irq(&pool->lock);
1986 while (too_many_workers(pool)) {
1987 struct worker *worker;
1988 unsigned long expires;
1990 /* idle_list is kept in LIFO order, check the last one */
1991 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1992 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1994 if (time_before(jiffies, expires)) {
1995 mod_timer(&pool->idle_timer, expires);
1999 destroy_worker(worker);
2002 spin_unlock_irq(&pool->lock);
2005 static void send_mayday(struct work_struct *work)
2007 struct pool_workqueue *pwq = get_work_pwq(work);
2008 struct workqueue_struct *wq = pwq->wq;
2010 lockdep_assert_held(&wq_mayday_lock);
2015 /* mayday mayday mayday */
2016 if (list_empty(&pwq->mayday_node)) {
2018 * If @pwq is for an unbound wq, its base ref may be put at
2019 * any time due to an attribute change. Pin @pwq until the
2020 * rescuer is done with it.
2023 list_add_tail(&pwq->mayday_node, &wq->maydays);
2024 wake_up_process(wq->rescuer->task);
2028 static void pool_mayday_timeout(struct timer_list *t)
2030 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2031 struct work_struct *work;
2033 spin_lock_irq(&pool->lock);
2034 spin_lock(&wq_mayday_lock); /* for wq->maydays */
2036 if (need_to_create_worker(pool)) {
2038 * We've been trying to create a new worker but
2039 * haven't been successful. We might be hitting an
2040 * allocation deadlock. Send distress signals to
2043 list_for_each_entry(work, &pool->worklist, entry)
2047 spin_unlock(&wq_mayday_lock);
2048 spin_unlock_irq(&pool->lock);
2050 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2054 * maybe_create_worker - create a new worker if necessary
2055 * @pool: pool to create a new worker for
2057 * Create a new worker for @pool if necessary. @pool is guaranteed to
2058 * have at least one idle worker on return from this function. If
2059 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2060 * sent to all rescuers with works scheduled on @pool to resolve
2061 * possible allocation deadlock.
2063 * On return, need_to_create_worker() is guaranteed to be %false and
2064 * may_start_working() %true.
2067 * spin_lock_irq(pool->lock) which may be released and regrabbed
2068 * multiple times. Does GFP_KERNEL allocations. Called only from
2071 static void maybe_create_worker(struct worker_pool *pool)
2072 __releases(&pool->lock)
2073 __acquires(&pool->lock)
2076 spin_unlock_irq(&pool->lock);
2078 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2079 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2082 if (create_worker(pool) || !need_to_create_worker(pool))
2085 schedule_timeout_interruptible(CREATE_COOLDOWN);
2087 if (!need_to_create_worker(pool))
2091 del_timer_sync(&pool->mayday_timer);
2092 spin_lock_irq(&pool->lock);
2094 * This is necessary even after a new worker was just successfully
2095 * created as @pool->lock was dropped and the new worker might have
2096 * already become busy.
2098 if (need_to_create_worker(pool))
2103 * manage_workers - manage worker pool
2106 * Assume the manager role and manage the worker pool @worker belongs
2107 * to. At any given time, there can be only zero or one manager per
2108 * pool. The exclusion is handled automatically by this function.
2110 * The caller can safely start processing works on false return. On
2111 * true return, it's guaranteed that need_to_create_worker() is false
2112 * and may_start_working() is true.
2115 * spin_lock_irq(pool->lock) which may be released and regrabbed
2116 * multiple times. Does GFP_KERNEL allocations.
2119 * %false if the pool doesn't need management and the caller can safely
2120 * start processing works, %true if management function was performed and
2121 * the conditions that the caller verified before calling the function may
2122 * no longer be true.
2124 static bool manage_workers(struct worker *worker)
2126 struct worker_pool *pool = worker->pool;
2128 if (pool->flags & POOL_MANAGER_ACTIVE)
2131 pool->flags |= POOL_MANAGER_ACTIVE;
2132 pool->manager = worker;
2134 maybe_create_worker(pool);
2136 pool->manager = NULL;
2137 pool->flags &= ~POOL_MANAGER_ACTIVE;
2138 wake_up(&wq_manager_wait);
2143 * process_one_work - process single work
2145 * @work: work to process
2147 * Process @work. This function contains all the logics necessary to
2148 * process a single work including synchronization against and
2149 * interaction with other workers on the same cpu, queueing and
2150 * flushing. As long as context requirement is met, any worker can
2151 * call this function to process a work.
2154 * spin_lock_irq(pool->lock) which is released and regrabbed.
2156 static void process_one_work(struct worker *worker, struct work_struct *work)
2157 __releases(&pool->lock)
2158 __acquires(&pool->lock)
2160 struct pool_workqueue *pwq = get_work_pwq(work);
2161 struct worker_pool *pool = worker->pool;
2162 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2164 struct worker *collision;
2165 #ifdef CONFIG_LOCKDEP
2167 * It is permissible to free the struct work_struct from
2168 * inside the function that is called from it, this we need to
2169 * take into account for lockdep too. To avoid bogus "held
2170 * lock freed" warnings as well as problems when looking into
2171 * work->lockdep_map, make a copy and use that here.
2173 struct lockdep_map lockdep_map;
2175 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2177 /* ensure we're on the correct CPU */
2178 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2179 raw_smp_processor_id() != pool->cpu);
2182 * A single work shouldn't be executed concurrently by
2183 * multiple workers on a single cpu. Check whether anyone is
2184 * already processing the work. If so, defer the work to the
2185 * currently executing one.
2187 collision = find_worker_executing_work(pool, work);
2188 if (unlikely(collision)) {
2189 move_linked_works(work, &collision->scheduled, NULL);
2193 /* claim and dequeue */
2194 debug_work_deactivate(work);
2195 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2196 worker->current_work = work;
2197 worker->current_func = work->func;
2198 worker->current_pwq = pwq;
2199 work_color = get_work_color(work);
2202 * Record wq name for cmdline and debug reporting, may get
2203 * overridden through set_worker_desc().
2205 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2207 list_del_init(&work->entry);
2210 * CPU intensive works don't participate in concurrency management.
2211 * They're the scheduler's responsibility. This takes @worker out
2212 * of concurrency management and the next code block will chain
2213 * execution of the pending work items.
2215 if (unlikely(cpu_intensive))
2216 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2219 * Wake up another worker if necessary. The condition is always
2220 * false for normal per-cpu workers since nr_running would always
2221 * be >= 1 at this point. This is used to chain execution of the
2222 * pending work items for WORKER_NOT_RUNNING workers such as the
2223 * UNBOUND and CPU_INTENSIVE ones.
2225 if (need_more_worker(pool))
2226 wake_up_worker(pool);
2229 * Record the last pool and clear PENDING which should be the last
2230 * update to @work. Also, do this inside @pool->lock so that
2231 * PENDING and queued state changes happen together while IRQ is
2234 set_work_pool_and_clear_pending(work, pool->id);
2236 spin_unlock_irq(&pool->lock);
2238 lock_map_acquire(&pwq->wq->lockdep_map);
2239 lock_map_acquire(&lockdep_map);
2241 * Strictly speaking we should mark the invariant state without holding
2242 * any locks, that is, before these two lock_map_acquire()'s.
2244 * However, that would result in:
2251 * Which would create W1->C->W1 dependencies, even though there is no
2252 * actual deadlock possible. There are two solutions, using a
2253 * read-recursive acquire on the work(queue) 'locks', but this will then
2254 * hit the lockdep limitation on recursive locks, or simply discard
2257 * AFAICT there is no possible deadlock scenario between the
2258 * flush_work() and complete() primitives (except for single-threaded
2259 * workqueues), so hiding them isn't a problem.
2261 lockdep_invariant_state(true);
2262 trace_workqueue_execute_start(work);
2263 worker->current_func(work);
2265 * While we must be careful to not use "work" after this, the trace
2266 * point will only record its address.
2268 trace_workqueue_execute_end(work);
2269 lock_map_release(&lockdep_map);
2270 lock_map_release(&pwq->wq->lockdep_map);
2272 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2273 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2274 " last function: %ps\n",
2275 current->comm, preempt_count(), task_pid_nr(current),
2276 worker->current_func);
2277 debug_show_held_locks(current);
2282 * The following prevents a kworker from hogging CPU on !PREEMPT
2283 * kernels, where a requeueing work item waiting for something to
2284 * happen could deadlock with stop_machine as such work item could
2285 * indefinitely requeue itself while all other CPUs are trapped in
2286 * stop_machine. At the same time, report a quiescent RCU state so
2287 * the same condition doesn't freeze RCU.
2291 spin_lock_irq(&pool->lock);
2293 /* clear cpu intensive status */
2294 if (unlikely(cpu_intensive))
2295 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2297 /* tag the worker for identification in schedule() */
2298 worker->last_func = worker->current_func;
2300 /* we're done with it, release */
2301 hash_del(&worker->hentry);
2302 worker->current_work = NULL;
2303 worker->current_func = NULL;
2304 worker->current_pwq = NULL;
2305 pwq_dec_nr_in_flight(pwq, work_color);
2309 * process_scheduled_works - process scheduled works
2312 * Process all scheduled works. Please note that the scheduled list
2313 * may change while processing a work, so this function repeatedly
2314 * fetches a work from the top and executes it.
2317 * spin_lock_irq(pool->lock) which may be released and regrabbed
2320 static void process_scheduled_works(struct worker *worker)
2322 while (!list_empty(&worker->scheduled)) {
2323 struct work_struct *work = list_first_entry(&worker->scheduled,
2324 struct work_struct, entry);
2325 process_one_work(worker, work);
2329 static void set_pf_worker(bool val)
2331 mutex_lock(&wq_pool_attach_mutex);
2333 current->flags |= PF_WQ_WORKER;
2335 current->flags &= ~PF_WQ_WORKER;
2336 mutex_unlock(&wq_pool_attach_mutex);
2340 * worker_thread - the worker thread function
2343 * The worker thread function. All workers belong to a worker_pool -
2344 * either a per-cpu one or dynamic unbound one. These workers process all
2345 * work items regardless of their specific target workqueue. The only
2346 * exception is work items which belong to workqueues with a rescuer which
2347 * will be explained in rescuer_thread().
2351 static int worker_thread(void *__worker)
2353 struct worker *worker = __worker;
2354 struct worker_pool *pool = worker->pool;
2356 /* tell the scheduler that this is a workqueue worker */
2357 set_pf_worker(true);
2359 spin_lock_irq(&pool->lock);
2361 /* am I supposed to die? */
2362 if (unlikely(worker->flags & WORKER_DIE)) {
2363 spin_unlock_irq(&pool->lock);
2364 WARN_ON_ONCE(!list_empty(&worker->entry));
2365 set_pf_worker(false);
2367 set_task_comm(worker->task, "kworker/dying");
2368 ida_simple_remove(&pool->worker_ida, worker->id);
2369 worker_detach_from_pool(worker);
2374 worker_leave_idle(worker);
2376 /* no more worker necessary? */
2377 if (!need_more_worker(pool))
2380 /* do we need to manage? */
2381 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2385 * ->scheduled list can only be filled while a worker is
2386 * preparing to process a work or actually processing it.
2387 * Make sure nobody diddled with it while I was sleeping.
2389 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2392 * Finish PREP stage. We're guaranteed to have at least one idle
2393 * worker or that someone else has already assumed the manager
2394 * role. This is where @worker starts participating in concurrency
2395 * management if applicable and concurrency management is restored
2396 * after being rebound. See rebind_workers() for details.
2398 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2401 struct work_struct *work =
2402 list_first_entry(&pool->worklist,
2403 struct work_struct, entry);
2405 pool->watchdog_ts = jiffies;
2407 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2408 /* optimization path, not strictly necessary */
2409 process_one_work(worker, work);
2410 if (unlikely(!list_empty(&worker->scheduled)))
2411 process_scheduled_works(worker);
2413 move_linked_works(work, &worker->scheduled, NULL);
2414 process_scheduled_works(worker);
2416 } while (keep_working(pool));
2418 worker_set_flags(worker, WORKER_PREP);
2421 * pool->lock is held and there's no work to process and no need to
2422 * manage, sleep. Workers are woken up only while holding
2423 * pool->lock or from local cpu, so setting the current state
2424 * before releasing pool->lock is enough to prevent losing any
2427 worker_enter_idle(worker);
2428 __set_current_state(TASK_IDLE);
2429 spin_unlock_irq(&pool->lock);
2435 * rescuer_thread - the rescuer thread function
2438 * Workqueue rescuer thread function. There's one rescuer for each
2439 * workqueue which has WQ_MEM_RECLAIM set.
2441 * Regular work processing on a pool may block trying to create a new
2442 * worker which uses GFP_KERNEL allocation which has slight chance of
2443 * developing into deadlock if some works currently on the same queue
2444 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2445 * the problem rescuer solves.
2447 * When such condition is possible, the pool summons rescuers of all
2448 * workqueues which have works queued on the pool and let them process
2449 * those works so that forward progress can be guaranteed.
2451 * This should happen rarely.
2455 static int rescuer_thread(void *__rescuer)
2457 struct worker *rescuer = __rescuer;
2458 struct workqueue_struct *wq = rescuer->rescue_wq;
2459 struct list_head *scheduled = &rescuer->scheduled;
2462 set_user_nice(current, RESCUER_NICE_LEVEL);
2465 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2466 * doesn't participate in concurrency management.
2468 set_pf_worker(true);
2470 set_current_state(TASK_IDLE);
2473 * By the time the rescuer is requested to stop, the workqueue
2474 * shouldn't have any work pending, but @wq->maydays may still have
2475 * pwq(s) queued. This can happen by non-rescuer workers consuming
2476 * all the work items before the rescuer got to them. Go through
2477 * @wq->maydays processing before acting on should_stop so that the
2478 * list is always empty on exit.
2480 should_stop = kthread_should_stop();
2482 /* see whether any pwq is asking for help */
2483 spin_lock_irq(&wq_mayday_lock);
2485 while (!list_empty(&wq->maydays)) {
2486 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2487 struct pool_workqueue, mayday_node);
2488 struct worker_pool *pool = pwq->pool;
2489 struct work_struct *work, *n;
2492 __set_current_state(TASK_RUNNING);
2493 list_del_init(&pwq->mayday_node);
2495 spin_unlock_irq(&wq_mayday_lock);
2497 worker_attach_to_pool(rescuer, pool);
2499 spin_lock_irq(&pool->lock);
2502 * Slurp in all works issued via this workqueue and
2505 WARN_ON_ONCE(!list_empty(scheduled));
2506 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2507 if (get_work_pwq(work) == pwq) {
2509 pool->watchdog_ts = jiffies;
2510 move_linked_works(work, scheduled, &n);
2515 if (!list_empty(scheduled)) {
2516 process_scheduled_works(rescuer);
2519 * The above execution of rescued work items could
2520 * have created more to rescue through
2521 * pwq_activate_first_delayed() or chained
2522 * queueing. Let's put @pwq back on mayday list so
2523 * that such back-to-back work items, which may be
2524 * being used to relieve memory pressure, don't
2525 * incur MAYDAY_INTERVAL delay inbetween.
2527 if (need_to_create_worker(pool)) {
2528 spin_lock(&wq_mayday_lock);
2530 list_move_tail(&pwq->mayday_node, &wq->maydays);
2531 spin_unlock(&wq_mayday_lock);
2536 * Put the reference grabbed by send_mayday(). @pool won't
2537 * go away while we're still attached to it.
2542 * Leave this pool. If need_more_worker() is %true, notify a
2543 * regular worker; otherwise, we end up with 0 concurrency
2544 * and stalling the execution.
2546 if (need_more_worker(pool))
2547 wake_up_worker(pool);
2549 spin_unlock_irq(&pool->lock);
2551 worker_detach_from_pool(rescuer);
2553 spin_lock_irq(&wq_mayday_lock);
2556 spin_unlock_irq(&wq_mayday_lock);
2559 __set_current_state(TASK_RUNNING);
2560 set_pf_worker(false);
2564 /* rescuers should never participate in concurrency management */
2565 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2571 * check_flush_dependency - check for flush dependency sanity
2572 * @target_wq: workqueue being flushed
2573 * @target_work: work item being flushed (NULL for workqueue flushes)
2575 * %current is trying to flush the whole @target_wq or @target_work on it.
2576 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2577 * reclaiming memory or running on a workqueue which doesn't have
2578 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2581 static void check_flush_dependency(struct workqueue_struct *target_wq,
2582 struct work_struct *target_work)
2584 work_func_t target_func = target_work ? target_work->func : NULL;
2585 struct worker *worker;
2587 if (target_wq->flags & WQ_MEM_RECLAIM)
2590 worker = current_wq_worker();
2592 WARN_ONCE(current->flags & PF_MEMALLOC,
2593 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2594 current->pid, current->comm, target_wq->name, target_func);
2595 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2596 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2597 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2598 worker->current_pwq->wq->name, worker->current_func,
2599 target_wq->name, target_func);
2603 struct work_struct work;
2604 struct completion done;
2605 struct task_struct *task; /* purely informational */
2608 static void wq_barrier_func(struct work_struct *work)
2610 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2611 complete(&barr->done);
2615 * insert_wq_barrier - insert a barrier work
2616 * @pwq: pwq to insert barrier into
2617 * @barr: wq_barrier to insert
2618 * @target: target work to attach @barr to
2619 * @worker: worker currently executing @target, NULL if @target is not executing
2621 * @barr is linked to @target such that @barr is completed only after
2622 * @target finishes execution. Please note that the ordering
2623 * guarantee is observed only with respect to @target and on the local
2626 * Currently, a queued barrier can't be canceled. This is because
2627 * try_to_grab_pending() can't determine whether the work to be
2628 * grabbed is at the head of the queue and thus can't clear LINKED
2629 * flag of the previous work while there must be a valid next work
2630 * after a work with LINKED flag set.
2632 * Note that when @worker is non-NULL, @target may be modified
2633 * underneath us, so we can't reliably determine pwq from @target.
2636 * spin_lock_irq(pool->lock).
2638 static void insert_wq_barrier(struct pool_workqueue *pwq,
2639 struct wq_barrier *barr,
2640 struct work_struct *target, struct worker *worker)
2642 struct list_head *head;
2643 unsigned int linked = 0;
2646 * debugobject calls are safe here even with pool->lock locked
2647 * as we know for sure that this will not trigger any of the
2648 * checks and call back into the fixup functions where we
2651 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2652 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2654 init_completion_map(&barr->done, &target->lockdep_map);
2656 barr->task = current;
2659 * If @target is currently being executed, schedule the
2660 * barrier to the worker; otherwise, put it after @target.
2663 head = worker->scheduled.next;
2665 unsigned long *bits = work_data_bits(target);
2667 head = target->entry.next;
2668 /* there can already be other linked works, inherit and set */
2669 linked = *bits & WORK_STRUCT_LINKED;
2670 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2673 debug_work_activate(&barr->work);
2674 insert_work(pwq, &barr->work, head,
2675 work_color_to_flags(WORK_NO_COLOR) | linked);
2679 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2680 * @wq: workqueue being flushed
2681 * @flush_color: new flush color, < 0 for no-op
2682 * @work_color: new work color, < 0 for no-op
2684 * Prepare pwqs for workqueue flushing.
2686 * If @flush_color is non-negative, flush_color on all pwqs should be
2687 * -1. If no pwq has in-flight commands at the specified color, all
2688 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2689 * has in flight commands, its pwq->flush_color is set to
2690 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2691 * wakeup logic is armed and %true is returned.
2693 * The caller should have initialized @wq->first_flusher prior to
2694 * calling this function with non-negative @flush_color. If
2695 * @flush_color is negative, no flush color update is done and %false
2698 * If @work_color is non-negative, all pwqs should have the same
2699 * work_color which is previous to @work_color and all will be
2700 * advanced to @work_color.
2703 * mutex_lock(wq->mutex).
2706 * %true if @flush_color >= 0 and there's something to flush. %false
2709 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2710 int flush_color, int work_color)
2713 struct pool_workqueue *pwq;
2715 if (flush_color >= 0) {
2716 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2717 atomic_set(&wq->nr_pwqs_to_flush, 1);
2720 for_each_pwq(pwq, wq) {
2721 struct worker_pool *pool = pwq->pool;
2723 spin_lock_irq(&pool->lock);
2725 if (flush_color >= 0) {
2726 WARN_ON_ONCE(pwq->flush_color != -1);
2728 if (pwq->nr_in_flight[flush_color]) {
2729 pwq->flush_color = flush_color;
2730 atomic_inc(&wq->nr_pwqs_to_flush);
2735 if (work_color >= 0) {
2736 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2737 pwq->work_color = work_color;
2740 spin_unlock_irq(&pool->lock);
2743 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2744 complete(&wq->first_flusher->done);
2750 * flush_workqueue - ensure that any scheduled work has run to completion.
2751 * @wq: workqueue to flush
2753 * This function sleeps until all work items which were queued on entry
2754 * have finished execution, but it is not livelocked by new incoming ones.
2756 void flush_workqueue(struct workqueue_struct *wq)
2758 struct wq_flusher this_flusher = {
2759 .list = LIST_HEAD_INIT(this_flusher.list),
2761 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2765 if (WARN_ON(!wq_online))
2768 lock_map_acquire(&wq->lockdep_map);
2769 lock_map_release(&wq->lockdep_map);
2771 mutex_lock(&wq->mutex);
2774 * Start-to-wait phase
2776 next_color = work_next_color(wq->work_color);
2778 if (next_color != wq->flush_color) {
2780 * Color space is not full. The current work_color
2781 * becomes our flush_color and work_color is advanced
2784 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2785 this_flusher.flush_color = wq->work_color;
2786 wq->work_color = next_color;
2788 if (!wq->first_flusher) {
2789 /* no flush in progress, become the first flusher */
2790 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2792 wq->first_flusher = &this_flusher;
2794 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2796 /* nothing to flush, done */
2797 wq->flush_color = next_color;
2798 wq->first_flusher = NULL;
2803 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2804 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2805 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2809 * Oops, color space is full, wait on overflow queue.
2810 * The next flush completion will assign us
2811 * flush_color and transfer to flusher_queue.
2813 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2816 check_flush_dependency(wq, NULL);
2818 mutex_unlock(&wq->mutex);
2820 wait_for_completion(&this_flusher.done);
2823 * Wake-up-and-cascade phase
2825 * First flushers are responsible for cascading flushes and
2826 * handling overflow. Non-first flushers can simply return.
2828 if (wq->first_flusher != &this_flusher)
2831 mutex_lock(&wq->mutex);
2833 /* we might have raced, check again with mutex held */
2834 if (wq->first_flusher != &this_flusher)
2837 wq->first_flusher = NULL;
2839 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2840 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2843 struct wq_flusher *next, *tmp;
2845 /* complete all the flushers sharing the current flush color */
2846 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2847 if (next->flush_color != wq->flush_color)
2849 list_del_init(&next->list);
2850 complete(&next->done);
2853 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2854 wq->flush_color != work_next_color(wq->work_color));
2856 /* this flush_color is finished, advance by one */
2857 wq->flush_color = work_next_color(wq->flush_color);
2859 /* one color has been freed, handle overflow queue */
2860 if (!list_empty(&wq->flusher_overflow)) {
2862 * Assign the same color to all overflowed
2863 * flushers, advance work_color and append to
2864 * flusher_queue. This is the start-to-wait
2865 * phase for these overflowed flushers.
2867 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2868 tmp->flush_color = wq->work_color;
2870 wq->work_color = work_next_color(wq->work_color);
2872 list_splice_tail_init(&wq->flusher_overflow,
2873 &wq->flusher_queue);
2874 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2877 if (list_empty(&wq->flusher_queue)) {
2878 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2883 * Need to flush more colors. Make the next flusher
2884 * the new first flusher and arm pwqs.
2886 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2887 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2889 list_del_init(&next->list);
2890 wq->first_flusher = next;
2892 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2896 * Meh... this color is already done, clear first
2897 * flusher and repeat cascading.
2899 wq->first_flusher = NULL;
2903 mutex_unlock(&wq->mutex);
2905 EXPORT_SYMBOL(flush_workqueue);
2908 * drain_workqueue - drain a workqueue
2909 * @wq: workqueue to drain
2911 * Wait until the workqueue becomes empty. While draining is in progress,
2912 * only chain queueing is allowed. IOW, only currently pending or running
2913 * work items on @wq can queue further work items on it. @wq is flushed
2914 * repeatedly until it becomes empty. The number of flushing is determined
2915 * by the depth of chaining and should be relatively short. Whine if it
2918 void drain_workqueue(struct workqueue_struct *wq)
2920 unsigned int flush_cnt = 0;
2921 struct pool_workqueue *pwq;
2924 * __queue_work() needs to test whether there are drainers, is much
2925 * hotter than drain_workqueue() and already looks at @wq->flags.
2926 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2928 mutex_lock(&wq->mutex);
2929 if (!wq->nr_drainers++)
2930 wq->flags |= __WQ_DRAINING;
2931 mutex_unlock(&wq->mutex);
2933 flush_workqueue(wq);
2935 mutex_lock(&wq->mutex);
2937 for_each_pwq(pwq, wq) {
2940 spin_lock_irq(&pwq->pool->lock);
2941 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2942 spin_unlock_irq(&pwq->pool->lock);
2947 if (++flush_cnt == 10 ||
2948 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2949 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2950 wq->name, flush_cnt);
2952 mutex_unlock(&wq->mutex);
2956 if (!--wq->nr_drainers)
2957 wq->flags &= ~__WQ_DRAINING;
2958 mutex_unlock(&wq->mutex);
2960 EXPORT_SYMBOL_GPL(drain_workqueue);
2962 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2965 struct worker *worker = NULL;
2966 struct worker_pool *pool;
2967 struct pool_workqueue *pwq;
2971 local_irq_disable();
2972 pool = get_work_pool(work);
2978 spin_lock(&pool->lock);
2979 /* see the comment in try_to_grab_pending() with the same code */
2980 pwq = get_work_pwq(work);
2982 if (unlikely(pwq->pool != pool))
2985 worker = find_worker_executing_work(pool, work);
2988 pwq = worker->current_pwq;
2991 check_flush_dependency(pwq->wq, work);
2993 insert_wq_barrier(pwq, barr, work, worker);
2994 spin_unlock_irq(&pool->lock);
2997 * Force a lock recursion deadlock when using flush_work() inside a
2998 * single-threaded or rescuer equipped workqueue.
3000 * For single threaded workqueues the deadlock happens when the work
3001 * is after the work issuing the flush_work(). For rescuer equipped
3002 * workqueues the deadlock happens when the rescuer stalls, blocking
3006 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3007 lock_map_acquire(&pwq->wq->lockdep_map);
3008 lock_map_release(&pwq->wq->lockdep_map);
3013 spin_unlock_irq(&pool->lock);
3017 static bool __flush_work(struct work_struct *work, bool from_cancel)
3019 struct wq_barrier barr;
3021 if (WARN_ON(!wq_online))
3024 if (WARN_ON(!work->func))
3028 lock_map_acquire(&work->lockdep_map);
3029 lock_map_release(&work->lockdep_map);
3032 if (start_flush_work(work, &barr, from_cancel)) {
3033 wait_for_completion(&barr.done);
3034 destroy_work_on_stack(&barr.work);
3042 * flush_work - wait for a work to finish executing the last queueing instance
3043 * @work: the work to flush
3045 * Wait until @work has finished execution. @work is guaranteed to be idle
3046 * on return if it hasn't been requeued since flush started.
3049 * %true if flush_work() waited for the work to finish execution,
3050 * %false if it was already idle.
3052 bool flush_work(struct work_struct *work)
3054 return __flush_work(work, false);
3056 EXPORT_SYMBOL_GPL(flush_work);
3059 wait_queue_entry_t wait;
3060 struct work_struct *work;
3063 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3065 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3067 if (cwait->work != key)
3069 return autoremove_wake_function(wait, mode, sync, key);
3072 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3074 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3075 unsigned long flags;
3079 ret = try_to_grab_pending(work, is_dwork, &flags);
3081 * If someone else is already canceling, wait for it to
3082 * finish. flush_work() doesn't work for PREEMPT_NONE
3083 * because we may get scheduled between @work's completion
3084 * and the other canceling task resuming and clearing
3085 * CANCELING - flush_work() will return false immediately
3086 * as @work is no longer busy, try_to_grab_pending() will
3087 * return -ENOENT as @work is still being canceled and the
3088 * other canceling task won't be able to clear CANCELING as
3089 * we're hogging the CPU.
3091 * Let's wait for completion using a waitqueue. As this
3092 * may lead to the thundering herd problem, use a custom
3093 * wake function which matches @work along with exclusive
3096 if (unlikely(ret == -ENOENT)) {
3097 struct cwt_wait cwait;
3099 init_wait(&cwait.wait);
3100 cwait.wait.func = cwt_wakefn;
3103 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3104 TASK_UNINTERRUPTIBLE);
3105 if (work_is_canceling(work))
3107 finish_wait(&cancel_waitq, &cwait.wait);
3109 } while (unlikely(ret < 0));
3111 /* tell other tasks trying to grab @work to back off */
3112 mark_work_canceling(work);
3113 local_irq_restore(flags);
3116 * This allows canceling during early boot. We know that @work
3120 __flush_work(work, true);
3122 clear_work_data(work);
3125 * Paired with prepare_to_wait() above so that either
3126 * waitqueue_active() is visible here or !work_is_canceling() is
3130 if (waitqueue_active(&cancel_waitq))
3131 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3137 * cancel_work_sync - cancel a work and wait for it to finish
3138 * @work: the work to cancel
3140 * Cancel @work and wait for its execution to finish. This function
3141 * can be used even if the work re-queues itself or migrates to
3142 * another workqueue. On return from this function, @work is
3143 * guaranteed to be not pending or executing on any CPU.
3145 * cancel_work_sync(&delayed_work->work) must not be used for
3146 * delayed_work's. Use cancel_delayed_work_sync() instead.
3148 * The caller must ensure that the workqueue on which @work was last
3149 * queued can't be destroyed before this function returns.
3152 * %true if @work was pending, %false otherwise.
3154 bool cancel_work_sync(struct work_struct *work)
3156 return __cancel_work_timer(work, false);
3158 EXPORT_SYMBOL_GPL(cancel_work_sync);
3161 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3162 * @dwork: the delayed work to flush
3164 * Delayed timer is cancelled and the pending work is queued for
3165 * immediate execution. Like flush_work(), this function only
3166 * considers the last queueing instance of @dwork.
3169 * %true if flush_work() waited for the work to finish execution,
3170 * %false if it was already idle.
3172 bool flush_delayed_work(struct delayed_work *dwork)
3174 local_irq_disable();
3175 if (del_timer_sync(&dwork->timer))
3176 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3178 return flush_work(&dwork->work);
3180 EXPORT_SYMBOL(flush_delayed_work);
3183 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3184 * @rwork: the rcu work to flush
3187 * %true if flush_rcu_work() waited for the work to finish execution,
3188 * %false if it was already idle.
3190 bool flush_rcu_work(struct rcu_work *rwork)
3192 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3194 flush_work(&rwork->work);
3197 return flush_work(&rwork->work);
3200 EXPORT_SYMBOL(flush_rcu_work);
3202 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3204 unsigned long flags;
3208 ret = try_to_grab_pending(work, is_dwork, &flags);
3209 } while (unlikely(ret == -EAGAIN));
3211 if (unlikely(ret < 0))
3214 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3215 local_irq_restore(flags);
3220 * cancel_delayed_work - cancel a delayed work
3221 * @dwork: delayed_work to cancel
3223 * Kill off a pending delayed_work.
3225 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3229 * The work callback function may still be running on return, unless
3230 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3231 * use cancel_delayed_work_sync() to wait on it.
3233 * This function is safe to call from any context including IRQ handler.
3235 bool cancel_delayed_work(struct delayed_work *dwork)
3237 return __cancel_work(&dwork->work, true);
3239 EXPORT_SYMBOL(cancel_delayed_work);
3242 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3243 * @dwork: the delayed work cancel
3245 * This is cancel_work_sync() for delayed works.
3248 * %true if @dwork was pending, %false otherwise.
3250 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3252 return __cancel_work_timer(&dwork->work, true);
3254 EXPORT_SYMBOL(cancel_delayed_work_sync);
3257 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3258 * @func: the function to call
3260 * schedule_on_each_cpu() executes @func on each online CPU using the
3261 * system workqueue and blocks until all CPUs have completed.
3262 * schedule_on_each_cpu() is very slow.
3265 * 0 on success, -errno on failure.
3267 int schedule_on_each_cpu(work_func_t func)
3270 struct work_struct __percpu *works;
3272 works = alloc_percpu(struct work_struct);
3278 for_each_online_cpu(cpu) {
3279 struct work_struct *work = per_cpu_ptr(works, cpu);
3281 INIT_WORK(work, func);
3282 schedule_work_on(cpu, work);
3285 for_each_online_cpu(cpu)
3286 flush_work(per_cpu_ptr(works, cpu));
3294 * execute_in_process_context - reliably execute the routine with user context
3295 * @fn: the function to execute
3296 * @ew: guaranteed storage for the execute work structure (must
3297 * be available when the work executes)
3299 * Executes the function immediately if process context is available,
3300 * otherwise schedules the function for delayed execution.
3302 * Return: 0 - function was executed
3303 * 1 - function was scheduled for execution
3305 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3307 if (!in_interrupt()) {
3312 INIT_WORK(&ew->work, fn);
3313 schedule_work(&ew->work);
3317 EXPORT_SYMBOL_GPL(execute_in_process_context);
3320 * free_workqueue_attrs - free a workqueue_attrs
3321 * @attrs: workqueue_attrs to free
3323 * Undo alloc_workqueue_attrs().
3325 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3328 free_cpumask_var(attrs->cpumask);
3334 * alloc_workqueue_attrs - allocate a workqueue_attrs
3335 * @gfp_mask: allocation mask to use
3337 * Allocate a new workqueue_attrs, initialize with default settings and
3340 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3342 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3344 struct workqueue_attrs *attrs;
3346 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3349 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3352 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3355 free_workqueue_attrs(attrs);
3359 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3360 const struct workqueue_attrs *from)
3362 to->nice = from->nice;
3363 cpumask_copy(to->cpumask, from->cpumask);
3365 * Unlike hash and equality test, this function doesn't ignore
3366 * ->no_numa as it is used for both pool and wq attrs. Instead,
3367 * get_unbound_pool() explicitly clears ->no_numa after copying.
3369 to->no_numa = from->no_numa;
3372 /* hash value of the content of @attr */
3373 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3377 hash = jhash_1word(attrs->nice, hash);
3378 hash = jhash(cpumask_bits(attrs->cpumask),
3379 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3383 /* content equality test */
3384 static bool wqattrs_equal(const struct workqueue_attrs *a,
3385 const struct workqueue_attrs *b)
3387 if (a->nice != b->nice)
3389 if (!cpumask_equal(a->cpumask, b->cpumask))
3395 * init_worker_pool - initialize a newly zalloc'd worker_pool
3396 * @pool: worker_pool to initialize
3398 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3400 * Return: 0 on success, -errno on failure. Even on failure, all fields
3401 * inside @pool proper are initialized and put_unbound_pool() can be called
3402 * on @pool safely to release it.
3404 static int init_worker_pool(struct worker_pool *pool)
3406 spin_lock_init(&pool->lock);
3409 pool->node = NUMA_NO_NODE;
3410 pool->flags |= POOL_DISASSOCIATED;
3411 pool->watchdog_ts = jiffies;
3412 INIT_LIST_HEAD(&pool->worklist);
3413 INIT_LIST_HEAD(&pool->idle_list);
3414 hash_init(pool->busy_hash);
3416 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3418 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3420 INIT_LIST_HEAD(&pool->workers);
3422 ida_init(&pool->worker_ida);
3423 INIT_HLIST_NODE(&pool->hash_node);
3426 /* shouldn't fail above this point */
3427 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3433 #ifdef CONFIG_LOCKDEP
3434 static void wq_init_lockdep(struct workqueue_struct *wq)
3438 lockdep_register_key(&wq->key);
3439 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3441 lock_name = wq->name;
3443 wq->lock_name = lock_name;
3444 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3447 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3449 lockdep_unregister_key(&wq->key);
3452 static void wq_free_lockdep(struct workqueue_struct *wq)
3454 if (wq->lock_name != wq->name)
3455 kfree(wq->lock_name);
3458 static void wq_init_lockdep(struct workqueue_struct *wq)
3462 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3466 static void wq_free_lockdep(struct workqueue_struct *wq)
3471 static void rcu_free_wq(struct rcu_head *rcu)
3473 struct workqueue_struct *wq =
3474 container_of(rcu, struct workqueue_struct, rcu);
3476 wq_free_lockdep(wq);
3478 if (!(wq->flags & WQ_UNBOUND))
3479 free_percpu(wq->cpu_pwqs);
3481 free_workqueue_attrs(wq->unbound_attrs);
3487 static void rcu_free_pool(struct rcu_head *rcu)
3489 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3491 ida_destroy(&pool->worker_ida);
3492 free_workqueue_attrs(pool->attrs);
3497 * put_unbound_pool - put a worker_pool
3498 * @pool: worker_pool to put
3500 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3501 * safe manner. get_unbound_pool() calls this function on its failure path
3502 * and this function should be able to release pools which went through,
3503 * successfully or not, init_worker_pool().
3505 * Should be called with wq_pool_mutex held.
3507 static void put_unbound_pool(struct worker_pool *pool)
3509 DECLARE_COMPLETION_ONSTACK(detach_completion);
3510 struct worker *worker;
3512 lockdep_assert_held(&wq_pool_mutex);
3518 if (WARN_ON(!(pool->cpu < 0)) ||
3519 WARN_ON(!list_empty(&pool->worklist)))
3522 /* release id and unhash */
3524 idr_remove(&worker_pool_idr, pool->id);
3525 hash_del(&pool->hash_node);
3528 * Become the manager and destroy all workers. This prevents
3529 * @pool's workers from blocking on attach_mutex. We're the last
3530 * manager and @pool gets freed with the flag set.
3532 spin_lock_irq(&pool->lock);
3533 wait_event_lock_irq(wq_manager_wait,
3534 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3535 pool->flags |= POOL_MANAGER_ACTIVE;
3537 while ((worker = first_idle_worker(pool)))
3538 destroy_worker(worker);
3539 WARN_ON(pool->nr_workers || pool->nr_idle);
3540 spin_unlock_irq(&pool->lock);
3542 mutex_lock(&wq_pool_attach_mutex);
3543 if (!list_empty(&pool->workers))
3544 pool->detach_completion = &detach_completion;
3545 mutex_unlock(&wq_pool_attach_mutex);
3547 if (pool->detach_completion)
3548 wait_for_completion(pool->detach_completion);
3550 /* shut down the timers */
3551 del_timer_sync(&pool->idle_timer);
3552 del_timer_sync(&pool->mayday_timer);
3554 /* sched-RCU protected to allow dereferences from get_work_pool() */
3555 call_rcu(&pool->rcu, rcu_free_pool);
3559 * get_unbound_pool - get a worker_pool with the specified attributes
3560 * @attrs: the attributes of the worker_pool to get
3562 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3563 * reference count and return it. If there already is a matching
3564 * worker_pool, it will be used; otherwise, this function attempts to
3567 * Should be called with wq_pool_mutex held.
3569 * Return: On success, a worker_pool with the same attributes as @attrs.
3570 * On failure, %NULL.
3572 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3574 u32 hash = wqattrs_hash(attrs);
3575 struct worker_pool *pool;
3577 int target_node = NUMA_NO_NODE;
3579 lockdep_assert_held(&wq_pool_mutex);
3581 /* do we already have a matching pool? */
3582 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3583 if (wqattrs_equal(pool->attrs, attrs)) {
3589 /* if cpumask is contained inside a NUMA node, we belong to that node */
3590 if (wq_numa_enabled) {
3591 for_each_node(node) {
3592 if (cpumask_subset(attrs->cpumask,
3593 wq_numa_possible_cpumask[node])) {
3600 /* nope, create a new one */
3601 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3602 if (!pool || init_worker_pool(pool) < 0)
3605 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3606 copy_workqueue_attrs(pool->attrs, attrs);
3607 pool->node = target_node;
3610 * no_numa isn't a worker_pool attribute, always clear it. See
3611 * 'struct workqueue_attrs' comments for detail.
3613 pool->attrs->no_numa = false;
3615 if (worker_pool_assign_id(pool) < 0)
3618 /* create and start the initial worker */
3619 if (wq_online && !create_worker(pool))
3623 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3628 put_unbound_pool(pool);
3632 static void rcu_free_pwq(struct rcu_head *rcu)
3634 kmem_cache_free(pwq_cache,
3635 container_of(rcu, struct pool_workqueue, rcu));
3639 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3640 * and needs to be destroyed.
3642 static void pwq_unbound_release_workfn(struct work_struct *work)
3644 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3645 unbound_release_work);
3646 struct workqueue_struct *wq = pwq->wq;
3647 struct worker_pool *pool = pwq->pool;
3650 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3653 mutex_lock(&wq->mutex);
3654 list_del_rcu(&pwq->pwqs_node);
3655 is_last = list_empty(&wq->pwqs);
3656 mutex_unlock(&wq->mutex);
3658 mutex_lock(&wq_pool_mutex);
3659 put_unbound_pool(pool);
3660 mutex_unlock(&wq_pool_mutex);
3662 call_rcu(&pwq->rcu, rcu_free_pwq);
3665 * If we're the last pwq going away, @wq is already dead and no one
3666 * is gonna access it anymore. Schedule RCU free.
3669 wq_unregister_lockdep(wq);
3670 call_rcu(&wq->rcu, rcu_free_wq);
3675 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3676 * @pwq: target pool_workqueue
3678 * If @pwq isn't freezing, set @pwq->max_active to the associated
3679 * workqueue's saved_max_active and activate delayed work items
3680 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3682 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3684 struct workqueue_struct *wq = pwq->wq;
3685 bool freezable = wq->flags & WQ_FREEZABLE;
3686 unsigned long flags;
3688 /* for @wq->saved_max_active */
3689 lockdep_assert_held(&wq->mutex);
3691 /* fast exit for non-freezable wqs */
3692 if (!freezable && pwq->max_active == wq->saved_max_active)
3695 /* this function can be called during early boot w/ irq disabled */
3696 spin_lock_irqsave(&pwq->pool->lock, flags);
3699 * During [un]freezing, the caller is responsible for ensuring that
3700 * this function is called at least once after @workqueue_freezing
3701 * is updated and visible.
3703 if (!freezable || !workqueue_freezing) {
3704 pwq->max_active = wq->saved_max_active;
3706 while (!list_empty(&pwq->delayed_works) &&
3707 pwq->nr_active < pwq->max_active)
3708 pwq_activate_first_delayed(pwq);
3711 * Need to kick a worker after thawed or an unbound wq's
3712 * max_active is bumped. It's a slow path. Do it always.
3714 wake_up_worker(pwq->pool);
3716 pwq->max_active = 0;
3719 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3722 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3723 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3724 struct worker_pool *pool)
3726 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3728 memset(pwq, 0, sizeof(*pwq));
3732 pwq->flush_color = -1;
3734 INIT_LIST_HEAD(&pwq->delayed_works);
3735 INIT_LIST_HEAD(&pwq->pwqs_node);
3736 INIT_LIST_HEAD(&pwq->mayday_node);
3737 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3740 /* sync @pwq with the current state of its associated wq and link it */
3741 static void link_pwq(struct pool_workqueue *pwq)
3743 struct workqueue_struct *wq = pwq->wq;
3745 lockdep_assert_held(&wq->mutex);
3747 /* may be called multiple times, ignore if already linked */
3748 if (!list_empty(&pwq->pwqs_node))
3751 /* set the matching work_color */
3752 pwq->work_color = wq->work_color;
3754 /* sync max_active to the current setting */
3755 pwq_adjust_max_active(pwq);
3758 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3761 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3762 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3763 const struct workqueue_attrs *attrs)
3765 struct worker_pool *pool;
3766 struct pool_workqueue *pwq;
3768 lockdep_assert_held(&wq_pool_mutex);
3770 pool = get_unbound_pool(attrs);
3774 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3776 put_unbound_pool(pool);
3780 init_pwq(pwq, wq, pool);
3785 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3786 * @attrs: the wq_attrs of the default pwq of the target workqueue
3787 * @node: the target NUMA node
3788 * @cpu_going_down: if >= 0, the CPU to consider as offline
3789 * @cpumask: outarg, the resulting cpumask
3791 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3792 * @cpu_going_down is >= 0, that cpu is considered offline during
3793 * calculation. The result is stored in @cpumask.
3795 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3796 * enabled and @node has online CPUs requested by @attrs, the returned
3797 * cpumask is the intersection of the possible CPUs of @node and
3800 * The caller is responsible for ensuring that the cpumask of @node stays
3803 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3806 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3807 int cpu_going_down, cpumask_t *cpumask)
3809 if (!wq_numa_enabled || attrs->no_numa)
3812 /* does @node have any online CPUs @attrs wants? */
3813 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3814 if (cpu_going_down >= 0)
3815 cpumask_clear_cpu(cpu_going_down, cpumask);
3817 if (cpumask_empty(cpumask))
3820 /* yeap, return possible CPUs in @node that @attrs wants */
3821 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3823 if (cpumask_empty(cpumask)) {
3824 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3825 "possible intersect\n");
3829 return !cpumask_equal(cpumask, attrs->cpumask);
3832 cpumask_copy(cpumask, attrs->cpumask);
3836 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3837 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3839 struct pool_workqueue *pwq)
3841 struct pool_workqueue *old_pwq;
3843 lockdep_assert_held(&wq_pool_mutex);
3844 lockdep_assert_held(&wq->mutex);
3846 /* link_pwq() can handle duplicate calls */
3849 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3850 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3854 /* context to store the prepared attrs & pwqs before applying */
3855 struct apply_wqattrs_ctx {
3856 struct workqueue_struct *wq; /* target workqueue */
3857 struct workqueue_attrs *attrs; /* attrs to apply */
3858 struct list_head list; /* queued for batching commit */
3859 struct pool_workqueue *dfl_pwq;
3860 struct pool_workqueue *pwq_tbl[];
3863 /* free the resources after success or abort */
3864 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3870 put_pwq_unlocked(ctx->pwq_tbl[node]);
3871 put_pwq_unlocked(ctx->dfl_pwq);
3873 free_workqueue_attrs(ctx->attrs);
3879 /* allocate the attrs and pwqs for later installation */
3880 static struct apply_wqattrs_ctx *
3881 apply_wqattrs_prepare(struct workqueue_struct *wq,
3882 const struct workqueue_attrs *attrs)
3884 struct apply_wqattrs_ctx *ctx;
3885 struct workqueue_attrs *new_attrs, *tmp_attrs;
3888 lockdep_assert_held(&wq_pool_mutex);
3890 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3892 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3893 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3894 if (!ctx || !new_attrs || !tmp_attrs)
3898 * Calculate the attrs of the default pwq.
3899 * If the user configured cpumask doesn't overlap with the
3900 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3902 copy_workqueue_attrs(new_attrs, attrs);
3903 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3904 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3905 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3908 * We may create multiple pwqs with differing cpumasks. Make a
3909 * copy of @new_attrs which will be modified and used to obtain
3912 copy_workqueue_attrs(tmp_attrs, new_attrs);
3915 * If something goes wrong during CPU up/down, we'll fall back to
3916 * the default pwq covering whole @attrs->cpumask. Always create
3917 * it even if we don't use it immediately.
3919 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3923 for_each_node(node) {
3924 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3925 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3926 if (!ctx->pwq_tbl[node])
3929 ctx->dfl_pwq->refcnt++;
3930 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3934 /* save the user configured attrs and sanitize it. */
3935 copy_workqueue_attrs(new_attrs, attrs);
3936 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3937 ctx->attrs = new_attrs;
3940 free_workqueue_attrs(tmp_attrs);
3944 free_workqueue_attrs(tmp_attrs);
3945 free_workqueue_attrs(new_attrs);
3946 apply_wqattrs_cleanup(ctx);
3950 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3951 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3955 /* all pwqs have been created successfully, let's install'em */
3956 mutex_lock(&ctx->wq->mutex);
3958 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3960 /* save the previous pwq and install the new one */
3962 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3963 ctx->pwq_tbl[node]);
3965 /* @dfl_pwq might not have been used, ensure it's linked */
3966 link_pwq(ctx->dfl_pwq);
3967 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3969 mutex_unlock(&ctx->wq->mutex);
3972 static void apply_wqattrs_lock(void)
3974 /* CPUs should stay stable across pwq creations and installations */
3976 mutex_lock(&wq_pool_mutex);
3979 static void apply_wqattrs_unlock(void)
3981 mutex_unlock(&wq_pool_mutex);
3985 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3986 const struct workqueue_attrs *attrs)
3988 struct apply_wqattrs_ctx *ctx;
3990 /* only unbound workqueues can change attributes */
3991 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3994 /* creating multiple pwqs breaks ordering guarantee */
3995 if (!list_empty(&wq->pwqs)) {
3996 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3999 wq->flags &= ~__WQ_ORDERED;
4002 ctx = apply_wqattrs_prepare(wq, attrs);
4006 /* the ctx has been prepared successfully, let's commit it */
4007 apply_wqattrs_commit(ctx);
4008 apply_wqattrs_cleanup(ctx);
4014 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4015 * @wq: the target workqueue
4016 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4018 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4019 * machines, this function maps a separate pwq to each NUMA node with
4020 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4021 * NUMA node it was issued on. Older pwqs are released as in-flight work
4022 * items finish. Note that a work item which repeatedly requeues itself
4023 * back-to-back will stay on its current pwq.
4025 * Performs GFP_KERNEL allocations.
4027 * Return: 0 on success and -errno on failure.
4029 int apply_workqueue_attrs(struct workqueue_struct *wq,
4030 const struct workqueue_attrs *attrs)
4034 apply_wqattrs_lock();
4035 ret = apply_workqueue_attrs_locked(wq, attrs);
4036 apply_wqattrs_unlock();
4040 EXPORT_SYMBOL_GPL(apply_workqueue_attrs);
4043 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4044 * @wq: the target workqueue
4045 * @cpu: the CPU coming up or going down
4046 * @online: whether @cpu is coming up or going down
4048 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4049 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4052 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4053 * falls back to @wq->dfl_pwq which may not be optimal but is always
4056 * Note that when the last allowed CPU of a NUMA node goes offline for a
4057 * workqueue with a cpumask spanning multiple nodes, the workers which were
4058 * already executing the work items for the workqueue will lose their CPU
4059 * affinity and may execute on any CPU. This is similar to how per-cpu
4060 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4061 * affinity, it's the user's responsibility to flush the work item from
4064 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4067 int node = cpu_to_node(cpu);
4068 int cpu_off = online ? -1 : cpu;
4069 struct pool_workqueue *old_pwq = NULL, *pwq;
4070 struct workqueue_attrs *target_attrs;
4073 lockdep_assert_held(&wq_pool_mutex);
4075 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4076 wq->unbound_attrs->no_numa)
4080 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4081 * Let's use a preallocated one. The following buf is protected by
4082 * CPU hotplug exclusion.
4084 target_attrs = wq_update_unbound_numa_attrs_buf;
4085 cpumask = target_attrs->cpumask;
4087 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4088 pwq = unbound_pwq_by_node(wq, node);
4091 * Let's determine what needs to be done. If the target cpumask is
4092 * different from the default pwq's, we need to compare it to @pwq's
4093 * and create a new one if they don't match. If the target cpumask
4094 * equals the default pwq's, the default pwq should be used.
4096 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4097 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4103 /* create a new pwq */
4104 pwq = alloc_unbound_pwq(wq, target_attrs);
4106 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4111 /* Install the new pwq. */
4112 mutex_lock(&wq->mutex);
4113 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4117 mutex_lock(&wq->mutex);
4118 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4119 get_pwq(wq->dfl_pwq);
4120 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4121 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4123 mutex_unlock(&wq->mutex);
4124 put_pwq_unlocked(old_pwq);
4127 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4129 bool highpri = wq->flags & WQ_HIGHPRI;
4132 if (!(wq->flags & WQ_UNBOUND)) {
4133 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4137 for_each_possible_cpu(cpu) {
4138 struct pool_workqueue *pwq =
4139 per_cpu_ptr(wq->cpu_pwqs, cpu);
4140 struct worker_pool *cpu_pools =
4141 per_cpu(cpu_worker_pools, cpu);
4143 init_pwq(pwq, wq, &cpu_pools[highpri]);
4145 mutex_lock(&wq->mutex);
4147 mutex_unlock(&wq->mutex);
4150 } else if (wq->flags & __WQ_ORDERED) {
4151 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4152 /* there should only be single pwq for ordering guarantee */
4153 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4154 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4155 "ordering guarantee broken for workqueue %s\n", wq->name);
4158 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4162 static int wq_clamp_max_active(int max_active, unsigned int flags,
4165 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4167 if (max_active < 1 || max_active > lim)
4168 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4169 max_active, name, 1, lim);
4171 return clamp_val(max_active, 1, lim);
4175 * Workqueues which may be used during memory reclaim should have a rescuer
4176 * to guarantee forward progress.
4178 static int init_rescuer(struct workqueue_struct *wq)
4180 struct worker *rescuer;
4183 if (!(wq->flags & WQ_MEM_RECLAIM))
4186 rescuer = alloc_worker(NUMA_NO_NODE);
4190 rescuer->rescue_wq = wq;
4191 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4192 ret = PTR_ERR_OR_ZERO(rescuer->task);
4198 wq->rescuer = rescuer;
4199 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4200 wake_up_process(rescuer->task);
4205 struct workqueue_struct *alloc_workqueue(const char *fmt,
4207 int max_active, ...)
4209 size_t tbl_size = 0;
4211 struct workqueue_struct *wq;
4212 struct pool_workqueue *pwq;
4215 * Unbound && max_active == 1 used to imply ordered, which is no
4216 * longer the case on NUMA machines due to per-node pools. While
4217 * alloc_ordered_workqueue() is the right way to create an ordered
4218 * workqueue, keep the previous behavior to avoid subtle breakages
4221 if ((flags & WQ_UNBOUND) && max_active == 1)
4222 flags |= __WQ_ORDERED;
4224 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4225 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4226 flags |= WQ_UNBOUND;
4228 /* allocate wq and format name */
4229 if (flags & WQ_UNBOUND)
4230 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4232 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4236 if (flags & WQ_UNBOUND) {
4237 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4238 if (!wq->unbound_attrs)
4242 va_start(args, max_active);
4243 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4246 max_active = max_active ?: WQ_DFL_ACTIVE;
4247 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4251 wq->saved_max_active = max_active;
4252 mutex_init(&wq->mutex);
4253 atomic_set(&wq->nr_pwqs_to_flush, 0);
4254 INIT_LIST_HEAD(&wq->pwqs);
4255 INIT_LIST_HEAD(&wq->flusher_queue);
4256 INIT_LIST_HEAD(&wq->flusher_overflow);
4257 INIT_LIST_HEAD(&wq->maydays);
4259 wq_init_lockdep(wq);
4260 INIT_LIST_HEAD(&wq->list);
4262 if (alloc_and_link_pwqs(wq) < 0)
4263 goto err_unreg_lockdep;
4265 if (wq_online && init_rescuer(wq) < 0)
4268 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4272 * wq_pool_mutex protects global freeze state and workqueues list.
4273 * Grab it, adjust max_active and add the new @wq to workqueues
4276 mutex_lock(&wq_pool_mutex);
4278 mutex_lock(&wq->mutex);
4279 for_each_pwq(pwq, wq)
4280 pwq_adjust_max_active(pwq);
4281 mutex_unlock(&wq->mutex);
4283 list_add_tail_rcu(&wq->list, &workqueues);
4285 mutex_unlock(&wq_pool_mutex);
4290 wq_unregister_lockdep(wq);
4291 wq_free_lockdep(wq);
4293 free_workqueue_attrs(wq->unbound_attrs);
4297 destroy_workqueue(wq);
4300 EXPORT_SYMBOL_GPL(alloc_workqueue);
4303 * destroy_workqueue - safely terminate a workqueue
4304 * @wq: target workqueue
4306 * Safely destroy a workqueue. All work currently pending will be done first.
4308 void destroy_workqueue(struct workqueue_struct *wq)
4310 struct pool_workqueue *pwq;
4313 /* drain it before proceeding with destruction */
4314 drain_workqueue(wq);
4317 mutex_lock(&wq->mutex);
4318 for_each_pwq(pwq, wq) {
4321 for (i = 0; i < WORK_NR_COLORS; i++) {
4322 if (WARN_ON(pwq->nr_in_flight[i])) {
4323 mutex_unlock(&wq->mutex);
4324 show_workqueue_state();
4329 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4330 WARN_ON(pwq->nr_active) ||
4331 WARN_ON(!list_empty(&pwq->delayed_works))) {
4332 mutex_unlock(&wq->mutex);
4333 show_workqueue_state();
4337 mutex_unlock(&wq->mutex);
4340 * wq list is used to freeze wq, remove from list after
4341 * flushing is complete in case freeze races us.
4343 mutex_lock(&wq_pool_mutex);
4344 list_del_rcu(&wq->list);
4345 mutex_unlock(&wq_pool_mutex);
4347 workqueue_sysfs_unregister(wq);
4350 kthread_stop(wq->rescuer->task);
4352 if (!(wq->flags & WQ_UNBOUND)) {
4353 wq_unregister_lockdep(wq);
4355 * The base ref is never dropped on per-cpu pwqs. Directly
4356 * schedule RCU free.
4358 call_rcu(&wq->rcu, rcu_free_wq);
4361 * We're the sole accessor of @wq at this point. Directly
4362 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4363 * @wq will be freed when the last pwq is released.
4365 for_each_node(node) {
4366 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4367 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4368 put_pwq_unlocked(pwq);
4372 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4373 * put. Don't access it afterwards.
4377 put_pwq_unlocked(pwq);
4380 EXPORT_SYMBOL_GPL(destroy_workqueue);
4383 * workqueue_set_max_active - adjust max_active of a workqueue
4384 * @wq: target workqueue
4385 * @max_active: new max_active value.
4387 * Set max_active of @wq to @max_active.
4390 * Don't call from IRQ context.
4392 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4394 struct pool_workqueue *pwq;
4396 /* disallow meddling with max_active for ordered workqueues */
4397 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4400 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4402 mutex_lock(&wq->mutex);
4404 wq->flags &= ~__WQ_ORDERED;
4405 wq->saved_max_active = max_active;
4407 for_each_pwq(pwq, wq)
4408 pwq_adjust_max_active(pwq);
4410 mutex_unlock(&wq->mutex);
4412 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4415 * current_work - retrieve %current task's work struct
4417 * Determine if %current task is a workqueue worker and what it's working on.
4418 * Useful to find out the context that the %current task is running in.
4420 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4422 struct work_struct *current_work(void)
4424 struct worker *worker = current_wq_worker();
4426 return worker ? worker->current_work : NULL;
4428 EXPORT_SYMBOL(current_work);
4431 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4433 * Determine whether %current is a workqueue rescuer. Can be used from
4434 * work functions to determine whether it's being run off the rescuer task.
4436 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4438 bool current_is_workqueue_rescuer(void)
4440 struct worker *worker = current_wq_worker();
4442 return worker && worker->rescue_wq;
4446 * workqueue_congested - test whether a workqueue is congested
4447 * @cpu: CPU in question
4448 * @wq: target workqueue
4450 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4451 * no synchronization around this function and the test result is
4452 * unreliable and only useful as advisory hints or for debugging.
4454 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4455 * Note that both per-cpu and unbound workqueues may be associated with
4456 * multiple pool_workqueues which have separate congested states. A
4457 * workqueue being congested on one CPU doesn't mean the workqueue is also
4458 * contested on other CPUs / NUMA nodes.
4461 * %true if congested, %false otherwise.
4463 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4465 struct pool_workqueue *pwq;
4468 rcu_read_lock_sched();
4470 if (cpu == WORK_CPU_UNBOUND)
4471 cpu = smp_processor_id();
4473 if (!(wq->flags & WQ_UNBOUND))
4474 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4476 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4478 ret = !list_empty(&pwq->delayed_works);
4479 rcu_read_unlock_sched();
4483 EXPORT_SYMBOL_GPL(workqueue_congested);
4486 * work_busy - test whether a work is currently pending or running
4487 * @work: the work to be tested
4489 * Test whether @work is currently pending or running. There is no
4490 * synchronization around this function and the test result is
4491 * unreliable and only useful as advisory hints or for debugging.
4494 * OR'd bitmask of WORK_BUSY_* bits.
4496 unsigned int work_busy(struct work_struct *work)
4498 struct worker_pool *pool;
4499 unsigned long flags;
4500 unsigned int ret = 0;
4502 if (work_pending(work))
4503 ret |= WORK_BUSY_PENDING;
4505 local_irq_save(flags);
4506 pool = get_work_pool(work);
4508 spin_lock(&pool->lock);
4509 if (find_worker_executing_work(pool, work))
4510 ret |= WORK_BUSY_RUNNING;
4511 spin_unlock(&pool->lock);
4513 local_irq_restore(flags);
4517 EXPORT_SYMBOL_GPL(work_busy);
4520 * set_worker_desc - set description for the current work item
4521 * @fmt: printf-style format string
4522 * @...: arguments for the format string
4524 * This function can be called by a running work function to describe what
4525 * the work item is about. If the worker task gets dumped, this
4526 * information will be printed out together to help debugging. The
4527 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4529 void set_worker_desc(const char *fmt, ...)
4531 struct worker *worker = current_wq_worker();
4535 va_start(args, fmt);
4536 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4540 EXPORT_SYMBOL_GPL(set_worker_desc);
4543 * print_worker_info - print out worker information and description
4544 * @log_lvl: the log level to use when printing
4545 * @task: target task
4547 * If @task is a worker and currently executing a work item, print out the
4548 * name of the workqueue being serviced and worker description set with
4549 * set_worker_desc() by the currently executing work item.
4551 * This function can be safely called on any task as long as the
4552 * task_struct itself is accessible. While safe, this function isn't
4553 * synchronized and may print out mixups or garbages of limited length.
4555 void print_worker_info(const char *log_lvl, struct task_struct *task)
4557 work_func_t *fn = NULL;
4558 char name[WQ_NAME_LEN] = { };
4559 char desc[WORKER_DESC_LEN] = { };
4560 struct pool_workqueue *pwq = NULL;
4561 struct workqueue_struct *wq = NULL;
4562 struct worker *worker;
4564 if (!(task->flags & PF_WQ_WORKER))
4568 * This function is called without any synchronization and @task
4569 * could be in any state. Be careful with dereferences.
4571 worker = kthread_probe_data(task);
4574 * Carefully copy the associated workqueue's workfn, name and desc.
4575 * Keep the original last '\0' in case the original is garbage.
4577 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4578 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4579 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4580 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4581 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4583 if (fn || name[0] || desc[0]) {
4584 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4585 if (strcmp(name, desc))
4586 pr_cont(" (%s)", desc);
4591 static void pr_cont_pool_info(struct worker_pool *pool)
4593 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4594 if (pool->node != NUMA_NO_NODE)
4595 pr_cont(" node=%d", pool->node);
4596 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4599 static void pr_cont_work(bool comma, struct work_struct *work)
4601 if (work->func == wq_barrier_func) {
4602 struct wq_barrier *barr;
4604 barr = container_of(work, struct wq_barrier, work);
4606 pr_cont("%s BAR(%d)", comma ? "," : "",
4607 task_pid_nr(barr->task));
4609 pr_cont("%s %ps", comma ? "," : "", work->func);
4613 static void show_pwq(struct pool_workqueue *pwq)
4615 struct worker_pool *pool = pwq->pool;
4616 struct work_struct *work;
4617 struct worker *worker;
4618 bool has_in_flight = false, has_pending = false;
4621 pr_info(" pwq %d:", pool->id);
4622 pr_cont_pool_info(pool);
4624 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4625 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4627 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4628 if (worker->current_pwq == pwq) {
4629 has_in_flight = true;
4633 if (has_in_flight) {
4636 pr_info(" in-flight:");
4637 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4638 if (worker->current_pwq != pwq)
4641 pr_cont("%s %d%s:%ps", comma ? "," : "",
4642 task_pid_nr(worker->task),
4643 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4644 worker->current_func);
4645 list_for_each_entry(work, &worker->scheduled, entry)
4646 pr_cont_work(false, work);
4652 list_for_each_entry(work, &pool->worklist, entry) {
4653 if (get_work_pwq(work) == pwq) {
4661 pr_info(" pending:");
4662 list_for_each_entry(work, &pool->worklist, entry) {
4663 if (get_work_pwq(work) != pwq)
4666 pr_cont_work(comma, work);
4667 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4672 if (!list_empty(&pwq->delayed_works)) {
4675 pr_info(" delayed:");
4676 list_for_each_entry(work, &pwq->delayed_works, entry) {
4677 pr_cont_work(comma, work);
4678 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4685 * show_workqueue_state - dump workqueue state
4687 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4688 * all busy workqueues and pools.
4690 void show_workqueue_state(void)
4692 struct workqueue_struct *wq;
4693 struct worker_pool *pool;
4694 unsigned long flags;
4697 rcu_read_lock_sched();
4699 pr_info("Showing busy workqueues and worker pools:\n");
4701 list_for_each_entry_rcu(wq, &workqueues, list) {
4702 struct pool_workqueue *pwq;
4705 for_each_pwq(pwq, wq) {
4706 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4714 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4716 for_each_pwq(pwq, wq) {
4717 spin_lock_irqsave(&pwq->pool->lock, flags);
4718 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4720 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4722 * We could be printing a lot from atomic context, e.g.
4723 * sysrq-t -> show_workqueue_state(). Avoid triggering
4726 touch_nmi_watchdog();
4730 for_each_pool(pool, pi) {
4731 struct worker *worker;
4734 spin_lock_irqsave(&pool->lock, flags);
4735 if (pool->nr_workers == pool->nr_idle)
4738 pr_info("pool %d:", pool->id);
4739 pr_cont_pool_info(pool);
4740 pr_cont(" hung=%us workers=%d",
4741 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4744 pr_cont(" manager: %d",
4745 task_pid_nr(pool->manager->task));
4746 list_for_each_entry(worker, &pool->idle_list, entry) {
4747 pr_cont(" %s%d", first ? "idle: " : "",
4748 task_pid_nr(worker->task));
4753 spin_unlock_irqrestore(&pool->lock, flags);
4755 * We could be printing a lot from atomic context, e.g.
4756 * sysrq-t -> show_workqueue_state(). Avoid triggering
4759 touch_nmi_watchdog();
4762 rcu_read_unlock_sched();
4765 /* used to show worker information through /proc/PID/{comm,stat,status} */
4766 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4770 /* always show the actual comm */
4771 off = strscpy(buf, task->comm, size);
4775 /* stabilize PF_WQ_WORKER and worker pool association */
4776 mutex_lock(&wq_pool_attach_mutex);
4778 if (task->flags & PF_WQ_WORKER) {
4779 struct worker *worker = kthread_data(task);
4780 struct worker_pool *pool = worker->pool;
4783 spin_lock_irq(&pool->lock);
4785 * ->desc tracks information (wq name or
4786 * set_worker_desc()) for the latest execution. If
4787 * current, prepend '+', otherwise '-'.
4789 if (worker->desc[0] != '\0') {
4790 if (worker->current_work)
4791 scnprintf(buf + off, size - off, "+%s",
4794 scnprintf(buf + off, size - off, "-%s",
4797 spin_unlock_irq(&pool->lock);
4801 mutex_unlock(&wq_pool_attach_mutex);
4809 * There are two challenges in supporting CPU hotplug. Firstly, there
4810 * are a lot of assumptions on strong associations among work, pwq and
4811 * pool which make migrating pending and scheduled works very
4812 * difficult to implement without impacting hot paths. Secondly,
4813 * worker pools serve mix of short, long and very long running works making
4814 * blocked draining impractical.
4816 * This is solved by allowing the pools to be disassociated from the CPU
4817 * running as an unbound one and allowing it to be reattached later if the
4818 * cpu comes back online.
4821 static void unbind_workers(int cpu)
4823 struct worker_pool *pool;
4824 struct worker *worker;
4826 for_each_cpu_worker_pool(pool, cpu) {
4827 mutex_lock(&wq_pool_attach_mutex);
4828 spin_lock_irq(&pool->lock);
4831 * We've blocked all attach/detach operations. Make all workers
4832 * unbound and set DISASSOCIATED. Before this, all workers
4833 * except for the ones which are still executing works from
4834 * before the last CPU down must be on the cpu. After
4835 * this, they may become diasporas.
4837 for_each_pool_worker(worker, pool)
4838 worker->flags |= WORKER_UNBOUND;
4840 pool->flags |= POOL_DISASSOCIATED;
4842 spin_unlock_irq(&pool->lock);
4843 mutex_unlock(&wq_pool_attach_mutex);
4846 * Call schedule() so that we cross rq->lock and thus can
4847 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4848 * This is necessary as scheduler callbacks may be invoked
4854 * Sched callbacks are disabled now. Zap nr_running.
4855 * After this, nr_running stays zero and need_more_worker()
4856 * and keep_working() are always true as long as the
4857 * worklist is not empty. This pool now behaves as an
4858 * unbound (in terms of concurrency management) pool which
4859 * are served by workers tied to the pool.
4861 atomic_set(&pool->nr_running, 0);
4864 * With concurrency management just turned off, a busy
4865 * worker blocking could lead to lengthy stalls. Kick off
4866 * unbound chain execution of currently pending work items.
4868 spin_lock_irq(&pool->lock);
4869 wake_up_worker(pool);
4870 spin_unlock_irq(&pool->lock);
4875 * rebind_workers - rebind all workers of a pool to the associated CPU
4876 * @pool: pool of interest
4878 * @pool->cpu is coming online. Rebind all workers to the CPU.
4880 static void rebind_workers(struct worker_pool *pool)
4882 struct worker *worker;
4884 lockdep_assert_held(&wq_pool_attach_mutex);
4887 * Restore CPU affinity of all workers. As all idle workers should
4888 * be on the run-queue of the associated CPU before any local
4889 * wake-ups for concurrency management happen, restore CPU affinity
4890 * of all workers first and then clear UNBOUND. As we're called
4891 * from CPU_ONLINE, the following shouldn't fail.
4893 for_each_pool_worker(worker, pool)
4894 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4895 pool->attrs->cpumask) < 0);
4897 spin_lock_irq(&pool->lock);
4899 pool->flags &= ~POOL_DISASSOCIATED;
4901 for_each_pool_worker(worker, pool) {
4902 unsigned int worker_flags = worker->flags;
4905 * A bound idle worker should actually be on the runqueue
4906 * of the associated CPU for local wake-ups targeting it to
4907 * work. Kick all idle workers so that they migrate to the
4908 * associated CPU. Doing this in the same loop as
4909 * replacing UNBOUND with REBOUND is safe as no worker will
4910 * be bound before @pool->lock is released.
4912 if (worker_flags & WORKER_IDLE)
4913 wake_up_process(worker->task);
4916 * We want to clear UNBOUND but can't directly call
4917 * worker_clr_flags() or adjust nr_running. Atomically
4918 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4919 * @worker will clear REBOUND using worker_clr_flags() when
4920 * it initiates the next execution cycle thus restoring
4921 * concurrency management. Note that when or whether
4922 * @worker clears REBOUND doesn't affect correctness.
4924 * WRITE_ONCE() is necessary because @worker->flags may be
4925 * tested without holding any lock in
4926 * wq_worker_running(). Without it, NOT_RUNNING test may
4927 * fail incorrectly leading to premature concurrency
4928 * management operations.
4930 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4931 worker_flags |= WORKER_REBOUND;
4932 worker_flags &= ~WORKER_UNBOUND;
4933 WRITE_ONCE(worker->flags, worker_flags);
4936 spin_unlock_irq(&pool->lock);
4940 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4941 * @pool: unbound pool of interest
4942 * @cpu: the CPU which is coming up
4944 * An unbound pool may end up with a cpumask which doesn't have any online
4945 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4946 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4947 * online CPU before, cpus_allowed of all its workers should be restored.
4949 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4951 static cpumask_t cpumask;
4952 struct worker *worker;
4954 lockdep_assert_held(&wq_pool_attach_mutex);
4956 /* is @cpu allowed for @pool? */
4957 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4960 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4962 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4963 for_each_pool_worker(worker, pool)
4964 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4967 int workqueue_prepare_cpu(unsigned int cpu)
4969 struct worker_pool *pool;
4971 for_each_cpu_worker_pool(pool, cpu) {
4972 if (pool->nr_workers)
4974 if (!create_worker(pool))
4980 int workqueue_online_cpu(unsigned int cpu)
4982 struct worker_pool *pool;
4983 struct workqueue_struct *wq;
4986 mutex_lock(&wq_pool_mutex);
4988 for_each_pool(pool, pi) {
4989 mutex_lock(&wq_pool_attach_mutex);
4991 if (pool->cpu == cpu)
4992 rebind_workers(pool);
4993 else if (pool->cpu < 0)
4994 restore_unbound_workers_cpumask(pool, cpu);
4996 mutex_unlock(&wq_pool_attach_mutex);
4999 /* update NUMA affinity of unbound workqueues */
5000 list_for_each_entry(wq, &workqueues, list)
5001 wq_update_unbound_numa(wq, cpu, true);
5003 mutex_unlock(&wq_pool_mutex);
5007 int workqueue_offline_cpu(unsigned int cpu)
5009 struct workqueue_struct *wq;
5011 /* unbinding per-cpu workers should happen on the local CPU */
5012 if (WARN_ON(cpu != smp_processor_id()))
5015 unbind_workers(cpu);
5017 /* update NUMA affinity of unbound workqueues */
5018 mutex_lock(&wq_pool_mutex);
5019 list_for_each_entry(wq, &workqueues, list)
5020 wq_update_unbound_numa(wq, cpu, false);
5021 mutex_unlock(&wq_pool_mutex);
5026 struct work_for_cpu {
5027 struct work_struct work;
5033 static void work_for_cpu_fn(struct work_struct *work)
5035 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5037 wfc->ret = wfc->fn(wfc->arg);
5041 * work_on_cpu - run a function in thread context on a particular cpu
5042 * @cpu: the cpu to run on
5043 * @fn: the function to run
5044 * @arg: the function arg
5046 * It is up to the caller to ensure that the cpu doesn't go offline.
5047 * The caller must not hold any locks which would prevent @fn from completing.
5049 * Return: The value @fn returns.
5051 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5053 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5055 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5056 schedule_work_on(cpu, &wfc.work);
5057 flush_work(&wfc.work);
5058 destroy_work_on_stack(&wfc.work);
5061 EXPORT_SYMBOL_GPL(work_on_cpu);
5064 * work_on_cpu_safe - run a function in thread context on a particular cpu
5065 * @cpu: the cpu to run on
5066 * @fn: the function to run
5067 * @arg: the function argument
5069 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5070 * any locks which would prevent @fn from completing.
5072 * Return: The value @fn returns.
5074 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5079 if (cpu_online(cpu))
5080 ret = work_on_cpu(cpu, fn, arg);
5084 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5085 #endif /* CONFIG_SMP */
5087 #ifdef CONFIG_FREEZER
5090 * freeze_workqueues_begin - begin freezing workqueues
5092 * Start freezing workqueues. After this function returns, all freezable
5093 * workqueues will queue new works to their delayed_works list instead of
5097 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5099 void freeze_workqueues_begin(void)
5101 struct workqueue_struct *wq;
5102 struct pool_workqueue *pwq;
5104 mutex_lock(&wq_pool_mutex);
5106 WARN_ON_ONCE(workqueue_freezing);
5107 workqueue_freezing = true;
5109 list_for_each_entry(wq, &workqueues, list) {
5110 mutex_lock(&wq->mutex);
5111 for_each_pwq(pwq, wq)
5112 pwq_adjust_max_active(pwq);
5113 mutex_unlock(&wq->mutex);
5116 mutex_unlock(&wq_pool_mutex);
5120 * freeze_workqueues_busy - are freezable workqueues still busy?
5122 * Check whether freezing is complete. This function must be called
5123 * between freeze_workqueues_begin() and thaw_workqueues().
5126 * Grabs and releases wq_pool_mutex.
5129 * %true if some freezable workqueues are still busy. %false if freezing
5132 bool freeze_workqueues_busy(void)
5135 struct workqueue_struct *wq;
5136 struct pool_workqueue *pwq;
5138 mutex_lock(&wq_pool_mutex);
5140 WARN_ON_ONCE(!workqueue_freezing);
5142 list_for_each_entry(wq, &workqueues, list) {
5143 if (!(wq->flags & WQ_FREEZABLE))
5146 * nr_active is monotonically decreasing. It's safe
5147 * to peek without lock.
5149 rcu_read_lock_sched();
5150 for_each_pwq(pwq, wq) {
5151 WARN_ON_ONCE(pwq->nr_active < 0);
5152 if (pwq->nr_active) {
5154 rcu_read_unlock_sched();
5158 rcu_read_unlock_sched();
5161 mutex_unlock(&wq_pool_mutex);
5166 * thaw_workqueues - thaw workqueues
5168 * Thaw workqueues. Normal queueing is restored and all collected
5169 * frozen works are transferred to their respective pool worklists.
5172 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5174 void thaw_workqueues(void)
5176 struct workqueue_struct *wq;
5177 struct pool_workqueue *pwq;
5179 mutex_lock(&wq_pool_mutex);
5181 if (!workqueue_freezing)
5184 workqueue_freezing = false;
5186 /* restore max_active and repopulate worklist */
5187 list_for_each_entry(wq, &workqueues, list) {
5188 mutex_lock(&wq->mutex);
5189 for_each_pwq(pwq, wq)
5190 pwq_adjust_max_active(pwq);
5191 mutex_unlock(&wq->mutex);
5195 mutex_unlock(&wq_pool_mutex);
5197 #endif /* CONFIG_FREEZER */
5199 static int workqueue_apply_unbound_cpumask(void)
5203 struct workqueue_struct *wq;
5204 struct apply_wqattrs_ctx *ctx, *n;
5206 lockdep_assert_held(&wq_pool_mutex);
5208 list_for_each_entry(wq, &workqueues, list) {
5209 if (!(wq->flags & WQ_UNBOUND))
5211 /* creating multiple pwqs breaks ordering guarantee */
5212 if (wq->flags & __WQ_ORDERED)
5215 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5221 list_add_tail(&ctx->list, &ctxs);
5224 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5226 apply_wqattrs_commit(ctx);
5227 apply_wqattrs_cleanup(ctx);
5234 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5235 * @cpumask: the cpumask to set
5237 * The low-level workqueues cpumask is a global cpumask that limits
5238 * the affinity of all unbound workqueues. This function check the @cpumask
5239 * and apply it to all unbound workqueues and updates all pwqs of them.
5241 * Retun: 0 - Success
5242 * -EINVAL - Invalid @cpumask
5243 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5245 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5248 cpumask_var_t saved_cpumask;
5250 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5254 * Not excluding isolated cpus on purpose.
5255 * If the user wishes to include them, we allow that.
5257 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5258 if (!cpumask_empty(cpumask)) {
5259 apply_wqattrs_lock();
5261 /* save the old wq_unbound_cpumask. */
5262 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5264 /* update wq_unbound_cpumask at first and apply it to wqs. */
5265 cpumask_copy(wq_unbound_cpumask, cpumask);
5266 ret = workqueue_apply_unbound_cpumask();
5268 /* restore the wq_unbound_cpumask when failed. */
5270 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5272 apply_wqattrs_unlock();
5275 free_cpumask_var(saved_cpumask);
5281 * Workqueues with WQ_SYSFS flag set is visible to userland via
5282 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5283 * following attributes.
5285 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5286 * max_active RW int : maximum number of in-flight work items
5288 * Unbound workqueues have the following extra attributes.
5290 * pool_ids RO int : the associated pool IDs for each node
5291 * nice RW int : nice value of the workers
5292 * cpumask RW mask : bitmask of allowed CPUs for the workers
5293 * numa RW bool : whether enable NUMA affinity
5296 struct workqueue_struct *wq;
5300 static struct workqueue_struct *dev_to_wq(struct device *dev)
5302 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5307 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5310 struct workqueue_struct *wq = dev_to_wq(dev);
5312 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5314 static DEVICE_ATTR_RO(per_cpu);
5316 static ssize_t max_active_show(struct device *dev,
5317 struct device_attribute *attr, char *buf)
5319 struct workqueue_struct *wq = dev_to_wq(dev);
5321 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5324 static ssize_t max_active_store(struct device *dev,
5325 struct device_attribute *attr, const char *buf,
5328 struct workqueue_struct *wq = dev_to_wq(dev);
5331 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5334 workqueue_set_max_active(wq, val);
5337 static DEVICE_ATTR_RW(max_active);
5339 static struct attribute *wq_sysfs_attrs[] = {
5340 &dev_attr_per_cpu.attr,
5341 &dev_attr_max_active.attr,
5344 ATTRIBUTE_GROUPS(wq_sysfs);
5346 static ssize_t wq_pool_ids_show(struct device *dev,
5347 struct device_attribute *attr, char *buf)
5349 struct workqueue_struct *wq = dev_to_wq(dev);
5350 const char *delim = "";
5351 int node, written = 0;
5353 rcu_read_lock_sched();
5354 for_each_node(node) {
5355 written += scnprintf(buf + written, PAGE_SIZE - written,
5356 "%s%d:%d", delim, node,
5357 unbound_pwq_by_node(wq, node)->pool->id);
5360 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5361 rcu_read_unlock_sched();
5366 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5369 struct workqueue_struct *wq = dev_to_wq(dev);
5372 mutex_lock(&wq->mutex);
5373 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5374 mutex_unlock(&wq->mutex);
5379 /* prepare workqueue_attrs for sysfs store operations */
5380 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5382 struct workqueue_attrs *attrs;
5384 lockdep_assert_held(&wq_pool_mutex);
5386 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5390 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5394 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5395 const char *buf, size_t count)
5397 struct workqueue_struct *wq = dev_to_wq(dev);
5398 struct workqueue_attrs *attrs;
5401 apply_wqattrs_lock();
5403 attrs = wq_sysfs_prep_attrs(wq);
5407 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5408 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5409 ret = apply_workqueue_attrs_locked(wq, attrs);
5414 apply_wqattrs_unlock();
5415 free_workqueue_attrs(attrs);
5416 return ret ?: count;
5419 static ssize_t wq_cpumask_show(struct device *dev,
5420 struct device_attribute *attr, char *buf)
5422 struct workqueue_struct *wq = dev_to_wq(dev);
5425 mutex_lock(&wq->mutex);
5426 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5427 cpumask_pr_args(wq->unbound_attrs->cpumask));
5428 mutex_unlock(&wq->mutex);
5432 static ssize_t wq_cpumask_store(struct device *dev,
5433 struct device_attribute *attr,
5434 const char *buf, size_t count)
5436 struct workqueue_struct *wq = dev_to_wq(dev);
5437 struct workqueue_attrs *attrs;
5440 apply_wqattrs_lock();
5442 attrs = wq_sysfs_prep_attrs(wq);
5446 ret = cpumask_parse(buf, attrs->cpumask);
5448 ret = apply_workqueue_attrs_locked(wq, attrs);
5451 apply_wqattrs_unlock();
5452 free_workqueue_attrs(attrs);
5453 return ret ?: count;
5456 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5459 struct workqueue_struct *wq = dev_to_wq(dev);
5462 mutex_lock(&wq->mutex);
5463 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5464 !wq->unbound_attrs->no_numa);
5465 mutex_unlock(&wq->mutex);
5470 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5471 const char *buf, size_t count)
5473 struct workqueue_struct *wq = dev_to_wq(dev);
5474 struct workqueue_attrs *attrs;
5475 int v, ret = -ENOMEM;
5477 apply_wqattrs_lock();
5479 attrs = wq_sysfs_prep_attrs(wq);
5484 if (sscanf(buf, "%d", &v) == 1) {
5485 attrs->no_numa = !v;
5486 ret = apply_workqueue_attrs_locked(wq, attrs);
5490 apply_wqattrs_unlock();
5491 free_workqueue_attrs(attrs);
5492 return ret ?: count;
5495 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5496 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5497 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5498 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5499 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5503 static struct bus_type wq_subsys = {
5504 .name = "workqueue",
5505 .dev_groups = wq_sysfs_groups,
5508 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5509 struct device_attribute *attr, char *buf)
5513 mutex_lock(&wq_pool_mutex);
5514 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5515 cpumask_pr_args(wq_unbound_cpumask));
5516 mutex_unlock(&wq_pool_mutex);
5521 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5522 struct device_attribute *attr, const char *buf, size_t count)
5524 cpumask_var_t cpumask;
5527 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5530 ret = cpumask_parse(buf, cpumask);
5532 ret = workqueue_set_unbound_cpumask(cpumask);
5534 free_cpumask_var(cpumask);
5535 return ret ? ret : count;
5538 static struct device_attribute wq_sysfs_cpumask_attr =
5539 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5540 wq_unbound_cpumask_store);
5542 static int __init wq_sysfs_init(void)
5546 err = subsys_virtual_register(&wq_subsys, NULL);
5550 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5552 core_initcall(wq_sysfs_init);
5554 static void wq_device_release(struct device *dev)
5556 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5562 * workqueue_sysfs_register - make a workqueue visible in sysfs
5563 * @wq: the workqueue to register
5565 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5566 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5567 * which is the preferred method.
5569 * Workqueue user should use this function directly iff it wants to apply
5570 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5571 * apply_workqueue_attrs() may race against userland updating the
5574 * Return: 0 on success, -errno on failure.
5576 int workqueue_sysfs_register(struct workqueue_struct *wq)
5578 struct wq_device *wq_dev;
5582 * Adjusting max_active or creating new pwqs by applying
5583 * attributes breaks ordering guarantee. Disallow exposing ordered
5586 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5589 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5594 wq_dev->dev.bus = &wq_subsys;
5595 wq_dev->dev.release = wq_device_release;
5596 dev_set_name(&wq_dev->dev, "%s", wq->name);
5599 * unbound_attrs are created separately. Suppress uevent until
5600 * everything is ready.
5602 dev_set_uevent_suppress(&wq_dev->dev, true);
5604 ret = device_register(&wq_dev->dev);
5606 put_device(&wq_dev->dev);
5611 if (wq->flags & WQ_UNBOUND) {
5612 struct device_attribute *attr;
5614 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5615 ret = device_create_file(&wq_dev->dev, attr);
5617 device_unregister(&wq_dev->dev);
5624 dev_set_uevent_suppress(&wq_dev->dev, false);
5625 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5630 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5631 * @wq: the workqueue to unregister
5633 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5635 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5637 struct wq_device *wq_dev = wq->wq_dev;
5643 device_unregister(&wq_dev->dev);
5645 #else /* CONFIG_SYSFS */
5646 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5647 #endif /* CONFIG_SYSFS */
5650 * Workqueue watchdog.
5652 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5653 * flush dependency, a concurrency managed work item which stays RUNNING
5654 * indefinitely. Workqueue stalls can be very difficult to debug as the
5655 * usual warning mechanisms don't trigger and internal workqueue state is
5658 * Workqueue watchdog monitors all worker pools periodically and dumps
5659 * state if some pools failed to make forward progress for a while where
5660 * forward progress is defined as the first item on ->worklist changing.
5662 * This mechanism is controlled through the kernel parameter
5663 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5664 * corresponding sysfs parameter file.
5666 #ifdef CONFIG_WQ_WATCHDOG
5668 static unsigned long wq_watchdog_thresh = 30;
5669 static struct timer_list wq_watchdog_timer;
5671 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5672 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5674 static void wq_watchdog_reset_touched(void)
5678 wq_watchdog_touched = jiffies;
5679 for_each_possible_cpu(cpu)
5680 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5683 static void wq_watchdog_timer_fn(struct timer_list *unused)
5685 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5686 bool lockup_detected = false;
5687 struct worker_pool *pool;
5695 for_each_pool(pool, pi) {
5696 unsigned long pool_ts, touched, ts;
5698 if (list_empty(&pool->worklist))
5701 /* get the latest of pool and touched timestamps */
5702 pool_ts = READ_ONCE(pool->watchdog_ts);
5703 touched = READ_ONCE(wq_watchdog_touched);
5705 if (time_after(pool_ts, touched))
5710 if (pool->cpu >= 0) {
5711 unsigned long cpu_touched =
5712 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5714 if (time_after(cpu_touched, ts))
5719 if (time_after(jiffies, ts + thresh)) {
5720 lockup_detected = true;
5721 pr_emerg("BUG: workqueue lockup - pool");
5722 pr_cont_pool_info(pool);
5723 pr_cont(" stuck for %us!\n",
5724 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5730 if (lockup_detected)
5731 show_workqueue_state();
5733 wq_watchdog_reset_touched();
5734 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5737 notrace void wq_watchdog_touch(int cpu)
5740 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5742 wq_watchdog_touched = jiffies;
5745 static void wq_watchdog_set_thresh(unsigned long thresh)
5747 wq_watchdog_thresh = 0;
5748 del_timer_sync(&wq_watchdog_timer);
5751 wq_watchdog_thresh = thresh;
5752 wq_watchdog_reset_touched();
5753 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5757 static int wq_watchdog_param_set_thresh(const char *val,
5758 const struct kernel_param *kp)
5760 unsigned long thresh;
5763 ret = kstrtoul(val, 0, &thresh);
5768 wq_watchdog_set_thresh(thresh);
5770 wq_watchdog_thresh = thresh;
5775 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5776 .set = wq_watchdog_param_set_thresh,
5777 .get = param_get_ulong,
5780 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5783 static void wq_watchdog_init(void)
5785 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5786 wq_watchdog_set_thresh(wq_watchdog_thresh);
5789 #else /* CONFIG_WQ_WATCHDOG */
5791 static inline void wq_watchdog_init(void) { }
5793 #endif /* CONFIG_WQ_WATCHDOG */
5795 static void __init wq_numa_init(void)
5800 if (num_possible_nodes() <= 1)
5803 if (wq_disable_numa) {
5804 pr_info("workqueue: NUMA affinity support disabled\n");
5808 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5809 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5812 * We want masks of possible CPUs of each node which isn't readily
5813 * available. Build one from cpu_to_node() which should have been
5814 * fully initialized by now.
5816 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5820 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5821 node_online(node) ? node : NUMA_NO_NODE));
5823 for_each_possible_cpu(cpu) {
5824 node = cpu_to_node(cpu);
5825 if (WARN_ON(node == NUMA_NO_NODE)) {
5826 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5827 /* happens iff arch is bonkers, let's just proceed */
5830 cpumask_set_cpu(cpu, tbl[node]);
5833 wq_numa_possible_cpumask = tbl;
5834 wq_numa_enabled = true;
5838 * workqueue_init_early - early init for workqueue subsystem
5840 * This is the first half of two-staged workqueue subsystem initialization
5841 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5842 * idr are up. It sets up all the data structures and system workqueues
5843 * and allows early boot code to create workqueues and queue/cancel work
5844 * items. Actual work item execution starts only after kthreads can be
5845 * created and scheduled right before early initcalls.
5847 int __init workqueue_init_early(void)
5849 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5850 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5853 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5855 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5856 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5858 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5860 /* initialize CPU pools */
5861 for_each_possible_cpu(cpu) {
5862 struct worker_pool *pool;
5865 for_each_cpu_worker_pool(pool, cpu) {
5866 BUG_ON(init_worker_pool(pool));
5868 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5869 pool->attrs->nice = std_nice[i++];
5870 pool->node = cpu_to_node(cpu);
5873 mutex_lock(&wq_pool_mutex);
5874 BUG_ON(worker_pool_assign_id(pool));
5875 mutex_unlock(&wq_pool_mutex);
5879 /* create default unbound and ordered wq attrs */
5880 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5881 struct workqueue_attrs *attrs;
5883 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5884 attrs->nice = std_nice[i];
5885 unbound_std_wq_attrs[i] = attrs;
5888 * An ordered wq should have only one pwq as ordering is
5889 * guaranteed by max_active which is enforced by pwqs.
5890 * Turn off NUMA so that dfl_pwq is used for all nodes.
5892 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5893 attrs->nice = std_nice[i];
5894 attrs->no_numa = true;
5895 ordered_wq_attrs[i] = attrs;
5898 system_wq = alloc_workqueue("events", 0, 0);
5899 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5900 system_long_wq = alloc_workqueue("events_long", 0, 0);
5901 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5902 WQ_UNBOUND_MAX_ACTIVE);
5903 system_freezable_wq = alloc_workqueue("events_freezable",
5905 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5906 WQ_POWER_EFFICIENT, 0);
5907 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5908 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5910 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5911 !system_unbound_wq || !system_freezable_wq ||
5912 !system_power_efficient_wq ||
5913 !system_freezable_power_efficient_wq);
5919 * workqueue_init - bring workqueue subsystem fully online
5921 * This is the latter half of two-staged workqueue subsystem initialization
5922 * and invoked as soon as kthreads can be created and scheduled.
5923 * Workqueues have been created and work items queued on them, but there
5924 * are no kworkers executing the work items yet. Populate the worker pools
5925 * with the initial workers and enable future kworker creations.
5927 int __init workqueue_init(void)
5929 struct workqueue_struct *wq;
5930 struct worker_pool *pool;
5934 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5935 * CPU to node mapping may not be available that early on some
5936 * archs such as power and arm64. As per-cpu pools created
5937 * previously could be missing node hint and unbound pools NUMA
5938 * affinity, fix them up.
5940 * Also, while iterating workqueues, create rescuers if requested.
5944 mutex_lock(&wq_pool_mutex);
5946 for_each_possible_cpu(cpu) {
5947 for_each_cpu_worker_pool(pool, cpu) {
5948 pool->node = cpu_to_node(cpu);
5952 list_for_each_entry(wq, &workqueues, list) {
5953 wq_update_unbound_numa(wq, smp_processor_id(), true);
5954 WARN(init_rescuer(wq),
5955 "workqueue: failed to create early rescuer for %s",
5959 mutex_unlock(&wq_pool_mutex);
5961 /* create the initial workers */
5962 for_each_online_cpu(cpu) {
5963 for_each_cpu_worker_pool(pool, cpu) {
5964 pool->flags &= ~POOL_DISASSOCIATED;
5965 BUG_ON(!create_worker(pool));
5969 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5970 BUG_ON(!create_worker(pool));