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/workqueue.txt 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/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
135 * WQ: wq->mutex protected.
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
139 * MD: wq_mayday_lock protected.
142 /* struct worker is defined in workqueue_internal.h */
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
151 unsigned long watchdog_ts; /* L: watchdog timestamp */
153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle; /* L: currently idle ones */
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb; /* manager arbitration */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
174 struct ida worker_ida; /* worker IDs for task name */
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp;
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
225 * Structure used to wait for workqueue flush.
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
261 struct wq_device *wq_dev; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
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_numa_enabled; /* unbound NUMA affinity enabled */
295 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
298 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
301 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302 static bool workqueue_freezing; /* PL: have wqs started freezing? */
304 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
306 /* the per-cpu worker pools */
307 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
310 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
312 /* PL: hash of all unbound pools keyed by pool->attrs */
313 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
315 /* I: attributes used when instantiating standard unbound pools on demand */
316 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
318 /* I: attributes used when instantiating ordered pools on demand */
319 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
321 struct workqueue_struct *system_wq __read_mostly;
322 EXPORT_SYMBOL(system_wq);
323 struct workqueue_struct *system_highpri_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_highpri_wq);
325 struct workqueue_struct *system_long_wq __read_mostly;
326 EXPORT_SYMBOL_GPL(system_long_wq);
327 struct workqueue_struct *system_unbound_wq __read_mostly;
328 EXPORT_SYMBOL_GPL(system_unbound_wq);
329 struct workqueue_struct *system_freezable_wq __read_mostly;
330 EXPORT_SYMBOL_GPL(system_freezable_wq);
331 struct workqueue_struct *system_power_efficient_wq __read_mostly;
332 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
333 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
334 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
336 static int worker_thread(void *__worker);
337 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
339 #define CREATE_TRACE_POINTS
340 #include <trace/events/workqueue.h>
342 #define assert_rcu_or_pool_mutex() \
343 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
344 !lockdep_is_held(&wq_pool_mutex), \
345 "sched RCU or wq_pool_mutex should be held")
347 #define assert_rcu_or_wq_mutex(wq) \
348 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
349 !lockdep_is_held(&wq->mutex), \
350 "sched RCU or wq->mutex should be held")
352 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
353 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
354 !lockdep_is_held(&wq->mutex) && \
355 !lockdep_is_held(&wq_pool_mutex), \
356 "sched RCU, wq->mutex or wq_pool_mutex should be held")
358 #define for_each_cpu_worker_pool(pool, cpu) \
359 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
360 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
364 * for_each_pool - iterate through all worker_pools in the system
365 * @pool: iteration cursor
366 * @pi: integer used for iteration
368 * This must be called either with wq_pool_mutex held or sched RCU read
369 * locked. If the pool needs to be used beyond the locking in effect, the
370 * caller is responsible for guaranteeing that the pool stays online.
372 * The if/else clause exists only for the lockdep assertion and can be
375 #define for_each_pool(pool, pi) \
376 idr_for_each_entry(&worker_pool_idr, pool, pi) \
377 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
381 * for_each_pool_worker - iterate through all workers of a worker_pool
382 * @worker: iteration cursor
383 * @pool: worker_pool to iterate workers of
385 * This must be called with @pool->attach_mutex.
387 * The if/else clause exists only for the lockdep assertion and can be
390 #define for_each_pool_worker(worker, pool) \
391 list_for_each_entry((worker), &(pool)->workers, node) \
392 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
396 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
397 * @pwq: iteration cursor
398 * @wq: the target workqueue
400 * This must be called either with wq->mutex held or sched RCU read locked.
401 * If the pwq needs to be used beyond the locking in effect, the caller is
402 * responsible for guaranteeing that the pwq stays online.
404 * The if/else clause exists only for the lockdep assertion and can be
407 #define for_each_pwq(pwq, wq) \
408 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
409 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
412 #ifdef CONFIG_DEBUG_OBJECTS_WORK
414 static struct debug_obj_descr work_debug_descr;
416 static void *work_debug_hint(void *addr)
418 return ((struct work_struct *) addr)->func;
422 * fixup_init is called when:
423 * - an active object is initialized
425 static int work_fixup_init(void *addr, enum debug_obj_state state)
427 struct work_struct *work = addr;
430 case ODEBUG_STATE_ACTIVE:
431 cancel_work_sync(work);
432 debug_object_init(work, &work_debug_descr);
440 * fixup_activate is called when:
441 * - an active object is activated
442 * - an unknown object is activated (might be a statically initialized object)
444 static int work_fixup_activate(void *addr, enum debug_obj_state state)
446 struct work_struct *work = addr;
450 case ODEBUG_STATE_NOTAVAILABLE:
452 * This is not really a fixup. The work struct was
453 * statically initialized. We just make sure that it
454 * is tracked in the object tracker.
456 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
457 debug_object_init(work, &work_debug_descr);
458 debug_object_activate(work, &work_debug_descr);
464 case ODEBUG_STATE_ACTIVE:
473 * fixup_free is called when:
474 * - an active object is freed
476 static int work_fixup_free(void *addr, enum debug_obj_state state)
478 struct work_struct *work = addr;
481 case ODEBUG_STATE_ACTIVE:
482 cancel_work_sync(work);
483 debug_object_free(work, &work_debug_descr);
490 static struct debug_obj_descr work_debug_descr = {
491 .name = "work_struct",
492 .debug_hint = work_debug_hint,
493 .fixup_init = work_fixup_init,
494 .fixup_activate = work_fixup_activate,
495 .fixup_free = work_fixup_free,
498 static inline void debug_work_activate(struct work_struct *work)
500 debug_object_activate(work, &work_debug_descr);
503 static inline void debug_work_deactivate(struct work_struct *work)
505 debug_object_deactivate(work, &work_debug_descr);
508 void __init_work(struct work_struct *work, int onstack)
511 debug_object_init_on_stack(work, &work_debug_descr);
513 debug_object_init(work, &work_debug_descr);
515 EXPORT_SYMBOL_GPL(__init_work);
517 void destroy_work_on_stack(struct work_struct *work)
519 debug_object_free(work, &work_debug_descr);
521 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
523 void destroy_delayed_work_on_stack(struct delayed_work *work)
525 destroy_timer_on_stack(&work->timer);
526 debug_object_free(&work->work, &work_debug_descr);
528 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
531 static inline void debug_work_activate(struct work_struct *work) { }
532 static inline void debug_work_deactivate(struct work_struct *work) { }
536 * worker_pool_assign_id - allocate ID and assing it to @pool
537 * @pool: the pool pointer of interest
539 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
540 * successfully, -errno on failure.
542 static int worker_pool_assign_id(struct worker_pool *pool)
546 lockdep_assert_held(&wq_pool_mutex);
548 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
558 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
559 * @wq: the target workqueue
562 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
564 * If the pwq needs to be used beyond the locking in effect, the caller is
565 * responsible for guaranteeing that the pwq stays online.
567 * Return: The unbound pool_workqueue for @node.
569 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
572 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
573 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
576 static unsigned int work_color_to_flags(int color)
578 return color << WORK_STRUCT_COLOR_SHIFT;
581 static int get_work_color(struct work_struct *work)
583 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
584 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
587 static int work_next_color(int color)
589 return (color + 1) % WORK_NR_COLORS;
593 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
594 * contain the pointer to the queued pwq. Once execution starts, the flag
595 * is cleared and the high bits contain OFFQ flags and pool ID.
597 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
598 * and clear_work_data() can be used to set the pwq, pool or clear
599 * work->data. These functions should only be called while the work is
600 * owned - ie. while the PENDING bit is set.
602 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
603 * corresponding to a work. Pool is available once the work has been
604 * queued anywhere after initialization until it is sync canceled. pwq is
605 * available only while the work item is queued.
607 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
608 * canceled. While being canceled, a work item may have its PENDING set
609 * but stay off timer and worklist for arbitrarily long and nobody should
610 * try to steal the PENDING bit.
612 static inline void set_work_data(struct work_struct *work, unsigned long data,
615 WARN_ON_ONCE(!work_pending(work));
616 atomic_long_set(&work->data, data | flags | work_static(work));
619 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
620 unsigned long extra_flags)
622 set_work_data(work, (unsigned long)pwq,
623 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
626 static void set_work_pool_and_keep_pending(struct work_struct *work,
629 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
630 WORK_STRUCT_PENDING);
633 static void set_work_pool_and_clear_pending(struct work_struct *work,
637 * The following wmb is paired with the implied mb in
638 * test_and_set_bit(PENDING) and ensures all updates to @work made
639 * here are visible to and precede any updates by the next PENDING
643 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
646 static void clear_work_data(struct work_struct *work)
648 smp_wmb(); /* see set_work_pool_and_clear_pending() */
649 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
652 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
654 unsigned long data = atomic_long_read(&work->data);
656 if (data & WORK_STRUCT_PWQ)
657 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
663 * get_work_pool - return the worker_pool a given work was associated with
664 * @work: the work item of interest
666 * Pools are created and destroyed under wq_pool_mutex, and allows read
667 * access under sched-RCU read lock. As such, this function should be
668 * called under wq_pool_mutex or with preemption disabled.
670 * All fields of the returned pool are accessible as long as the above
671 * mentioned locking is in effect. If the returned pool needs to be used
672 * beyond the critical section, the caller is responsible for ensuring the
673 * returned pool is and stays online.
675 * Return: The worker_pool @work was last associated with. %NULL if none.
677 static struct worker_pool *get_work_pool(struct work_struct *work)
679 unsigned long data = atomic_long_read(&work->data);
682 assert_rcu_or_pool_mutex();
684 if (data & WORK_STRUCT_PWQ)
685 return ((struct pool_workqueue *)
686 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
688 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
689 if (pool_id == WORK_OFFQ_POOL_NONE)
692 return idr_find(&worker_pool_idr, pool_id);
696 * get_work_pool_id - return the worker pool ID a given work is associated with
697 * @work: the work item of interest
699 * Return: The worker_pool ID @work was last associated with.
700 * %WORK_OFFQ_POOL_NONE if none.
702 static int get_work_pool_id(struct work_struct *work)
704 unsigned long data = atomic_long_read(&work->data);
706 if (data & WORK_STRUCT_PWQ)
707 return ((struct pool_workqueue *)
708 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
710 return data >> WORK_OFFQ_POOL_SHIFT;
713 static void mark_work_canceling(struct work_struct *work)
715 unsigned long pool_id = get_work_pool_id(work);
717 pool_id <<= WORK_OFFQ_POOL_SHIFT;
718 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
721 static bool work_is_canceling(struct work_struct *work)
723 unsigned long data = atomic_long_read(&work->data);
725 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
729 * Policy functions. These define the policies on how the global worker
730 * pools are managed. Unless noted otherwise, these functions assume that
731 * they're being called with pool->lock held.
734 static bool __need_more_worker(struct worker_pool *pool)
736 return !atomic_read(&pool->nr_running);
740 * Need to wake up a worker? Called from anything but currently
743 * Note that, because unbound workers never contribute to nr_running, this
744 * function will always return %true for unbound pools as long as the
745 * worklist isn't empty.
747 static bool need_more_worker(struct worker_pool *pool)
749 return !list_empty(&pool->worklist) && __need_more_worker(pool);
752 /* Can I start working? Called from busy but !running workers. */
753 static bool may_start_working(struct worker_pool *pool)
755 return pool->nr_idle;
758 /* Do I need to keep working? Called from currently running workers. */
759 static bool keep_working(struct worker_pool *pool)
761 return !list_empty(&pool->worklist) &&
762 atomic_read(&pool->nr_running) <= 1;
765 /* Do we need a new worker? Called from manager. */
766 static bool need_to_create_worker(struct worker_pool *pool)
768 return need_more_worker(pool) && !may_start_working(pool);
771 /* Do we have too many workers and should some go away? */
772 static bool too_many_workers(struct worker_pool *pool)
774 bool managing = mutex_is_locked(&pool->manager_arb);
775 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
776 int nr_busy = pool->nr_workers - nr_idle;
778 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
785 /* Return the first idle worker. Safe with preemption disabled */
786 static struct worker *first_idle_worker(struct worker_pool *pool)
788 if (unlikely(list_empty(&pool->idle_list)))
791 return list_first_entry(&pool->idle_list, struct worker, entry);
795 * wake_up_worker - wake up an idle worker
796 * @pool: worker pool to wake worker from
798 * Wake up the first idle worker of @pool.
801 * spin_lock_irq(pool->lock).
803 static void wake_up_worker(struct worker_pool *pool)
805 struct worker *worker = first_idle_worker(pool);
808 wake_up_process(worker->task);
812 * wq_worker_waking_up - a worker is waking up
813 * @task: task waking up
814 * @cpu: CPU @task is waking up to
816 * This function is called during try_to_wake_up() when a worker is
820 * spin_lock_irq(rq->lock)
822 void wq_worker_waking_up(struct task_struct *task, int cpu)
824 struct worker *worker = kthread_data(task);
826 if (!(worker->flags & WORKER_NOT_RUNNING)) {
827 WARN_ON_ONCE(worker->pool->cpu != cpu);
828 atomic_inc(&worker->pool->nr_running);
833 * wq_worker_sleeping - a worker is going to sleep
834 * @task: task going to sleep
835 * @cpu: CPU in question, must be the current CPU number
837 * This function is called during schedule() when a busy worker is
838 * going to sleep. Worker on the same cpu can be woken up by
839 * returning pointer to its task.
842 * spin_lock_irq(rq->lock)
845 * Worker task on @cpu to wake up, %NULL if none.
847 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
849 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
850 struct worker_pool *pool;
853 * Rescuers, which may not have all the fields set up like normal
854 * workers, also reach here, let's not access anything before
855 * checking NOT_RUNNING.
857 if (worker->flags & WORKER_NOT_RUNNING)
862 /* this can only happen on the local cpu */
863 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
867 * The counterpart of the following dec_and_test, implied mb,
868 * worklist not empty test sequence is in insert_work().
869 * Please read comment there.
871 * NOT_RUNNING is clear. This means that we're bound to and
872 * running on the local cpu w/ rq lock held and preemption
873 * disabled, which in turn means that none else could be
874 * manipulating idle_list, so dereferencing idle_list without pool
877 if (atomic_dec_and_test(&pool->nr_running) &&
878 !list_empty(&pool->worklist))
879 to_wakeup = first_idle_worker(pool);
880 return to_wakeup ? to_wakeup->task : NULL;
884 * worker_set_flags - set worker flags and adjust nr_running accordingly
886 * @flags: flags to set
888 * Set @flags in @worker->flags and adjust nr_running accordingly.
891 * spin_lock_irq(pool->lock)
893 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
895 struct worker_pool *pool = worker->pool;
897 WARN_ON_ONCE(worker->task != current);
899 /* If transitioning into NOT_RUNNING, adjust nr_running. */
900 if ((flags & WORKER_NOT_RUNNING) &&
901 !(worker->flags & WORKER_NOT_RUNNING)) {
902 atomic_dec(&pool->nr_running);
905 worker->flags |= flags;
909 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
911 * @flags: flags to clear
913 * Clear @flags in @worker->flags and adjust nr_running accordingly.
916 * spin_lock_irq(pool->lock)
918 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
920 struct worker_pool *pool = worker->pool;
921 unsigned int oflags = worker->flags;
923 WARN_ON_ONCE(worker->task != current);
925 worker->flags &= ~flags;
928 * If transitioning out of NOT_RUNNING, increment nr_running. Note
929 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
930 * of multiple flags, not a single flag.
932 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
933 if (!(worker->flags & WORKER_NOT_RUNNING))
934 atomic_inc(&pool->nr_running);
938 * find_worker_executing_work - find worker which is executing a work
939 * @pool: pool of interest
940 * @work: work to find worker for
942 * Find a worker which is executing @work on @pool by searching
943 * @pool->busy_hash which is keyed by the address of @work. For a worker
944 * to match, its current execution should match the address of @work and
945 * its work function. This is to avoid unwanted dependency between
946 * unrelated work executions through a work item being recycled while still
949 * This is a bit tricky. A work item may be freed once its execution
950 * starts and nothing prevents the freed area from being recycled for
951 * another work item. If the same work item address ends up being reused
952 * before the original execution finishes, workqueue will identify the
953 * recycled work item as currently executing and make it wait until the
954 * current execution finishes, introducing an unwanted dependency.
956 * This function checks the work item address and work function to avoid
957 * false positives. Note that this isn't complete as one may construct a
958 * work function which can introduce dependency onto itself through a
959 * recycled work item. Well, if somebody wants to shoot oneself in the
960 * foot that badly, there's only so much we can do, and if such deadlock
961 * actually occurs, it should be easy to locate the culprit work function.
964 * spin_lock_irq(pool->lock).
967 * Pointer to worker which is executing @work if found, %NULL
970 static struct worker *find_worker_executing_work(struct worker_pool *pool,
971 struct work_struct *work)
973 struct worker *worker;
975 hash_for_each_possible(pool->busy_hash, worker, hentry,
977 if (worker->current_work == work &&
978 worker->current_func == work->func)
985 * move_linked_works - move linked works to a list
986 * @work: start of series of works to be scheduled
987 * @head: target list to append @work to
988 * @nextp: out parameter for nested worklist walking
990 * Schedule linked works starting from @work to @head. Work series to
991 * be scheduled starts at @work and includes any consecutive work with
992 * WORK_STRUCT_LINKED set in its predecessor.
994 * If @nextp is not NULL, it's updated to point to the next work of
995 * the last scheduled work. This allows move_linked_works() to be
996 * nested inside outer list_for_each_entry_safe().
999 * spin_lock_irq(pool->lock).
1001 static void move_linked_works(struct work_struct *work, struct list_head *head,
1002 struct work_struct **nextp)
1004 struct work_struct *n;
1007 * Linked worklist will always end before the end of the list,
1008 * use NULL for list head.
1010 list_for_each_entry_safe_from(work, n, NULL, entry) {
1011 list_move_tail(&work->entry, head);
1012 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1017 * If we're already inside safe list traversal and have moved
1018 * multiple works to the scheduled queue, the next position
1019 * needs to be updated.
1026 * get_pwq - get an extra reference on the specified pool_workqueue
1027 * @pwq: pool_workqueue to get
1029 * Obtain an extra reference on @pwq. The caller should guarantee that
1030 * @pwq has positive refcnt and be holding the matching pool->lock.
1032 static void get_pwq(struct pool_workqueue *pwq)
1034 lockdep_assert_held(&pwq->pool->lock);
1035 WARN_ON_ONCE(pwq->refcnt <= 0);
1040 * put_pwq - put a pool_workqueue reference
1041 * @pwq: pool_workqueue to put
1043 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1044 * destruction. The caller should be holding the matching pool->lock.
1046 static void put_pwq(struct pool_workqueue *pwq)
1048 lockdep_assert_held(&pwq->pool->lock);
1049 if (likely(--pwq->refcnt))
1051 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1054 * @pwq can't be released under pool->lock, bounce to
1055 * pwq_unbound_release_workfn(). This never recurses on the same
1056 * pool->lock as this path is taken only for unbound workqueues and
1057 * the release work item is scheduled on a per-cpu workqueue. To
1058 * avoid lockdep warning, unbound pool->locks are given lockdep
1059 * subclass of 1 in get_unbound_pool().
1061 schedule_work(&pwq->unbound_release_work);
1065 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1066 * @pwq: pool_workqueue to put (can be %NULL)
1068 * put_pwq() with locking. This function also allows %NULL @pwq.
1070 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1074 * As both pwqs and pools are sched-RCU protected, the
1075 * following lock operations are safe.
1077 spin_lock_irq(&pwq->pool->lock);
1079 spin_unlock_irq(&pwq->pool->lock);
1083 static void pwq_activate_delayed_work(struct work_struct *work)
1085 struct pool_workqueue *pwq = get_work_pwq(work);
1087 trace_workqueue_activate_work(work);
1088 if (list_empty(&pwq->pool->worklist))
1089 pwq->pool->watchdog_ts = jiffies;
1090 move_linked_works(work, &pwq->pool->worklist, NULL);
1091 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1095 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1097 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1098 struct work_struct, entry);
1100 pwq_activate_delayed_work(work);
1104 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1105 * @pwq: pwq of interest
1106 * @color: color of work which left the queue
1108 * A work either has completed or is removed from pending queue,
1109 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1112 * spin_lock_irq(pool->lock).
1114 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1116 /* uncolored work items don't participate in flushing or nr_active */
1117 if (color == WORK_NO_COLOR)
1120 pwq->nr_in_flight[color]--;
1123 if (!list_empty(&pwq->delayed_works)) {
1124 /* one down, submit a delayed one */
1125 if (pwq->nr_active < pwq->max_active)
1126 pwq_activate_first_delayed(pwq);
1129 /* is flush in progress and are we at the flushing tip? */
1130 if (likely(pwq->flush_color != color))
1133 /* are there still in-flight works? */
1134 if (pwq->nr_in_flight[color])
1137 /* this pwq is done, clear flush_color */
1138 pwq->flush_color = -1;
1141 * If this was the last pwq, wake up the first flusher. It
1142 * will handle the rest.
1144 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1145 complete(&pwq->wq->first_flusher->done);
1151 * try_to_grab_pending - steal work item from worklist and disable irq
1152 * @work: work item to steal
1153 * @is_dwork: @work is a delayed_work
1154 * @flags: place to store irq state
1156 * Try to grab PENDING bit of @work. This function can handle @work in any
1157 * stable state - idle, on timer or on worklist.
1160 * 1 if @work was pending and we successfully stole PENDING
1161 * 0 if @work was idle and we claimed PENDING
1162 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1163 * -ENOENT if someone else is canceling @work, this state may persist
1164 * for arbitrarily long
1167 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1168 * interrupted while holding PENDING and @work off queue, irq must be
1169 * disabled on entry. This, combined with delayed_work->timer being
1170 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1172 * On successful return, >= 0, irq is disabled and the caller is
1173 * responsible for releasing it using local_irq_restore(*@flags).
1175 * This function is safe to call from any context including IRQ handler.
1177 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1178 unsigned long *flags)
1180 struct worker_pool *pool;
1181 struct pool_workqueue *pwq;
1183 local_irq_save(*flags);
1185 /* try to steal the timer if it exists */
1187 struct delayed_work *dwork = to_delayed_work(work);
1190 * dwork->timer is irqsafe. If del_timer() fails, it's
1191 * guaranteed that the timer is not queued anywhere and not
1192 * running on the local CPU.
1194 if (likely(del_timer(&dwork->timer)))
1198 /* try to claim PENDING the normal way */
1199 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1203 * The queueing is in progress, or it is already queued. Try to
1204 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1206 pool = get_work_pool(work);
1210 spin_lock(&pool->lock);
1212 * work->data is guaranteed to point to pwq only while the work
1213 * item is queued on pwq->wq, and both updating work->data to point
1214 * to pwq on queueing and to pool on dequeueing are done under
1215 * pwq->pool->lock. This in turn guarantees that, if work->data
1216 * points to pwq which is associated with a locked pool, the work
1217 * item is currently queued on that pool.
1219 pwq = get_work_pwq(work);
1220 if (pwq && pwq->pool == pool) {
1221 debug_work_deactivate(work);
1224 * A delayed work item cannot be grabbed directly because
1225 * it might have linked NO_COLOR work items which, if left
1226 * on the delayed_list, will confuse pwq->nr_active
1227 * management later on and cause stall. Make sure the work
1228 * item is activated before grabbing.
1230 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1231 pwq_activate_delayed_work(work);
1233 list_del_init(&work->entry);
1234 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1236 /* work->data points to pwq iff queued, point to pool */
1237 set_work_pool_and_keep_pending(work, pool->id);
1239 spin_unlock(&pool->lock);
1242 spin_unlock(&pool->lock);
1244 local_irq_restore(*flags);
1245 if (work_is_canceling(work))
1252 * insert_work - insert a work into a pool
1253 * @pwq: pwq @work belongs to
1254 * @work: work to insert
1255 * @head: insertion point
1256 * @extra_flags: extra WORK_STRUCT_* flags to set
1258 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1259 * work_struct flags.
1262 * spin_lock_irq(pool->lock).
1264 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1265 struct list_head *head, unsigned int extra_flags)
1267 struct worker_pool *pool = pwq->pool;
1269 /* we own @work, set data and link */
1270 set_work_pwq(work, pwq, extra_flags);
1271 list_add_tail(&work->entry, head);
1275 * Ensure either wq_worker_sleeping() sees the above
1276 * list_add_tail() or we see zero nr_running to avoid workers lying
1277 * around lazily while there are works to be processed.
1281 if (__need_more_worker(pool))
1282 wake_up_worker(pool);
1286 * Test whether @work is being queued from another work executing on the
1289 static bool is_chained_work(struct workqueue_struct *wq)
1291 struct worker *worker;
1293 worker = current_wq_worker();
1295 * Return %true iff I'm a worker execuing a work item on @wq. If
1296 * I'm @worker, it's safe to dereference it without locking.
1298 return worker && worker->current_pwq->wq == wq;
1301 static void __queue_work(int cpu, struct workqueue_struct *wq,
1302 struct work_struct *work)
1304 struct pool_workqueue *pwq;
1305 struct worker_pool *last_pool;
1306 struct list_head *worklist;
1307 unsigned int work_flags;
1308 unsigned int req_cpu = cpu;
1311 * While a work item is PENDING && off queue, a task trying to
1312 * steal the PENDING will busy-loop waiting for it to either get
1313 * queued or lose PENDING. Grabbing PENDING and queueing should
1314 * happen with IRQ disabled.
1316 WARN_ON_ONCE(!irqs_disabled());
1318 debug_work_activate(work);
1320 /* if draining, only works from the same workqueue are allowed */
1321 if (unlikely(wq->flags & __WQ_DRAINING) &&
1322 WARN_ON_ONCE(!is_chained_work(wq)))
1325 if (req_cpu == WORK_CPU_UNBOUND)
1326 cpu = raw_smp_processor_id();
1328 /* pwq which will be used unless @work is executing elsewhere */
1329 if (!(wq->flags & WQ_UNBOUND))
1330 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1332 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1335 * If @work was previously on a different pool, it might still be
1336 * running there, in which case the work needs to be queued on that
1337 * pool to guarantee non-reentrancy.
1339 last_pool = get_work_pool(work);
1340 if (last_pool && last_pool != pwq->pool) {
1341 struct worker *worker;
1343 spin_lock(&last_pool->lock);
1345 worker = find_worker_executing_work(last_pool, work);
1347 if (worker && worker->current_pwq->wq == wq) {
1348 pwq = worker->current_pwq;
1350 /* meh... not running there, queue here */
1351 spin_unlock(&last_pool->lock);
1352 spin_lock(&pwq->pool->lock);
1355 spin_lock(&pwq->pool->lock);
1359 * pwq is determined and locked. For unbound pools, we could have
1360 * raced with pwq release and it could already be dead. If its
1361 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1362 * without another pwq replacing it in the numa_pwq_tbl or while
1363 * work items are executing on it, so the retrying is guaranteed to
1364 * make forward-progress.
1366 if (unlikely(!pwq->refcnt)) {
1367 if (wq->flags & WQ_UNBOUND) {
1368 spin_unlock(&pwq->pool->lock);
1373 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1377 /* pwq determined, queue */
1378 trace_workqueue_queue_work(req_cpu, pwq, work);
1380 if (WARN_ON(!list_empty(&work->entry))) {
1381 spin_unlock(&pwq->pool->lock);
1385 pwq->nr_in_flight[pwq->work_color]++;
1386 work_flags = work_color_to_flags(pwq->work_color);
1388 if (likely(pwq->nr_active < pwq->max_active)) {
1389 trace_workqueue_activate_work(work);
1391 worklist = &pwq->pool->worklist;
1392 if (list_empty(worklist))
1393 pwq->pool->watchdog_ts = jiffies;
1395 work_flags |= WORK_STRUCT_DELAYED;
1396 worklist = &pwq->delayed_works;
1399 insert_work(pwq, work, worklist, work_flags);
1401 spin_unlock(&pwq->pool->lock);
1405 * queue_work_on - queue work on specific cpu
1406 * @cpu: CPU number to execute work on
1407 * @wq: workqueue to use
1408 * @work: work to queue
1410 * We queue the work to a specific CPU, the caller must ensure it
1413 * Return: %false if @work was already on a queue, %true otherwise.
1415 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1416 struct work_struct *work)
1419 unsigned long flags;
1421 local_irq_save(flags);
1423 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1424 __queue_work(cpu, wq, work);
1428 local_irq_restore(flags);
1431 EXPORT_SYMBOL(queue_work_on);
1433 void delayed_work_timer_fn(unsigned long __data)
1435 struct delayed_work *dwork = (struct delayed_work *)__data;
1437 /* should have been called from irqsafe timer with irq already off */
1438 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1440 EXPORT_SYMBOL(delayed_work_timer_fn);
1442 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1443 struct delayed_work *dwork, unsigned long delay)
1445 struct timer_list *timer = &dwork->timer;
1446 struct work_struct *work = &dwork->work;
1448 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1449 timer->data != (unsigned long)dwork);
1450 WARN_ON_ONCE(timer_pending(timer));
1451 WARN_ON_ONCE(!list_empty(&work->entry));
1454 * If @delay is 0, queue @dwork->work immediately. This is for
1455 * both optimization and correctness. The earliest @timer can
1456 * expire is on the closest next tick and delayed_work users depend
1457 * on that there's no such delay when @delay is 0.
1460 __queue_work(cpu, wq, &dwork->work);
1464 timer_stats_timer_set_start_info(&dwork->timer);
1468 timer->expires = jiffies + delay;
1470 if (unlikely(cpu != WORK_CPU_UNBOUND))
1471 add_timer_on(timer, cpu);
1477 * queue_delayed_work_on - queue work on specific CPU after delay
1478 * @cpu: CPU number to execute work on
1479 * @wq: workqueue to use
1480 * @dwork: work to queue
1481 * @delay: number of jiffies to wait before queueing
1483 * Return: %false if @work was already on a queue, %true otherwise. If
1484 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1487 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1488 struct delayed_work *dwork, unsigned long delay)
1490 struct work_struct *work = &dwork->work;
1492 unsigned long flags;
1494 /* read the comment in __queue_work() */
1495 local_irq_save(flags);
1497 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1498 __queue_delayed_work(cpu, wq, dwork, delay);
1502 local_irq_restore(flags);
1505 EXPORT_SYMBOL(queue_delayed_work_on);
1508 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1509 * @cpu: CPU number to execute work on
1510 * @wq: workqueue to use
1511 * @dwork: work to queue
1512 * @delay: number of jiffies to wait before queueing
1514 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1515 * modify @dwork's timer so that it expires after @delay. If @delay is
1516 * zero, @work is guaranteed to be scheduled immediately regardless of its
1519 * Return: %false if @dwork was idle and queued, %true if @dwork was
1520 * pending and its timer was modified.
1522 * This function is safe to call from any context including IRQ handler.
1523 * See try_to_grab_pending() for details.
1525 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1526 struct delayed_work *dwork, unsigned long delay)
1528 unsigned long flags;
1532 ret = try_to_grab_pending(&dwork->work, true, &flags);
1533 } while (unlikely(ret == -EAGAIN));
1535 if (likely(ret >= 0)) {
1536 __queue_delayed_work(cpu, wq, dwork, delay);
1537 local_irq_restore(flags);
1540 /* -ENOENT from try_to_grab_pending() becomes %true */
1543 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1546 * worker_enter_idle - enter idle state
1547 * @worker: worker which is entering idle state
1549 * @worker is entering idle state. Update stats and idle timer if
1553 * spin_lock_irq(pool->lock).
1555 static void worker_enter_idle(struct worker *worker)
1557 struct worker_pool *pool = worker->pool;
1559 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1560 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1561 (worker->hentry.next || worker->hentry.pprev)))
1564 /* can't use worker_set_flags(), also called from create_worker() */
1565 worker->flags |= WORKER_IDLE;
1567 worker->last_active = jiffies;
1569 /* idle_list is LIFO */
1570 list_add(&worker->entry, &pool->idle_list);
1572 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1573 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1576 * Sanity check nr_running. Because wq_unbind_fn() releases
1577 * pool->lock between setting %WORKER_UNBOUND and zapping
1578 * nr_running, the warning may trigger spuriously. Check iff
1579 * unbind is not in progress.
1581 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1582 pool->nr_workers == pool->nr_idle &&
1583 atomic_read(&pool->nr_running));
1587 * worker_leave_idle - leave idle state
1588 * @worker: worker which is leaving idle state
1590 * @worker is leaving idle state. Update stats.
1593 * spin_lock_irq(pool->lock).
1595 static void worker_leave_idle(struct worker *worker)
1597 struct worker_pool *pool = worker->pool;
1599 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1601 worker_clr_flags(worker, WORKER_IDLE);
1603 list_del_init(&worker->entry);
1606 static struct worker *alloc_worker(int node)
1608 struct worker *worker;
1610 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1612 INIT_LIST_HEAD(&worker->entry);
1613 INIT_LIST_HEAD(&worker->scheduled);
1614 INIT_LIST_HEAD(&worker->node);
1615 /* on creation a worker is in !idle && prep state */
1616 worker->flags = WORKER_PREP;
1622 * worker_attach_to_pool() - attach a worker to a pool
1623 * @worker: worker to be attached
1624 * @pool: the target pool
1626 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1627 * cpu-binding of @worker are kept coordinated with the pool across
1630 static void worker_attach_to_pool(struct worker *worker,
1631 struct worker_pool *pool)
1633 mutex_lock(&pool->attach_mutex);
1636 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1637 * online CPUs. It'll be re-applied when any of the CPUs come up.
1639 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1642 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1643 * stable across this function. See the comments above the
1644 * flag definition for details.
1646 if (pool->flags & POOL_DISASSOCIATED)
1647 worker->flags |= WORKER_UNBOUND;
1649 list_add_tail(&worker->node, &pool->workers);
1651 mutex_unlock(&pool->attach_mutex);
1655 * worker_detach_from_pool() - detach a worker from its pool
1656 * @worker: worker which is attached to its pool
1657 * @pool: the pool @worker is attached to
1659 * Undo the attaching which had been done in worker_attach_to_pool(). The
1660 * caller worker shouldn't access to the pool after detached except it has
1661 * other reference to the pool.
1663 static void worker_detach_from_pool(struct worker *worker,
1664 struct worker_pool *pool)
1666 struct completion *detach_completion = NULL;
1668 mutex_lock(&pool->attach_mutex);
1669 list_del(&worker->node);
1670 if (list_empty(&pool->workers))
1671 detach_completion = pool->detach_completion;
1672 mutex_unlock(&pool->attach_mutex);
1674 /* clear leftover flags without pool->lock after it is detached */
1675 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1677 if (detach_completion)
1678 complete(detach_completion);
1682 * create_worker - create a new workqueue worker
1683 * @pool: pool the new worker will belong to
1685 * Create and start a new worker which is attached to @pool.
1688 * Might sleep. Does GFP_KERNEL allocations.
1691 * Pointer to the newly created worker.
1693 static struct worker *create_worker(struct worker_pool *pool)
1695 struct worker *worker = NULL;
1699 /* ID is needed to determine kthread name */
1700 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1704 worker = alloc_worker(pool->node);
1708 worker->pool = pool;
1712 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1713 pool->attrs->nice < 0 ? "H" : "");
1715 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1717 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1718 "kworker/%s", id_buf);
1719 if (IS_ERR(worker->task))
1722 set_user_nice(worker->task, pool->attrs->nice);
1723 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1725 /* successful, attach the worker to the pool */
1726 worker_attach_to_pool(worker, pool);
1728 /* start the newly created worker */
1729 spin_lock_irq(&pool->lock);
1730 worker->pool->nr_workers++;
1731 worker_enter_idle(worker);
1732 wake_up_process(worker->task);
1733 spin_unlock_irq(&pool->lock);
1739 ida_simple_remove(&pool->worker_ida, id);
1745 * destroy_worker - destroy a workqueue worker
1746 * @worker: worker to be destroyed
1748 * Destroy @worker and adjust @pool stats accordingly. The worker should
1752 * spin_lock_irq(pool->lock).
1754 static void destroy_worker(struct worker *worker)
1756 struct worker_pool *pool = worker->pool;
1758 lockdep_assert_held(&pool->lock);
1760 /* sanity check frenzy */
1761 if (WARN_ON(worker->current_work) ||
1762 WARN_ON(!list_empty(&worker->scheduled)) ||
1763 WARN_ON(!(worker->flags & WORKER_IDLE)))
1769 list_del_init(&worker->entry);
1770 worker->flags |= WORKER_DIE;
1771 wake_up_process(worker->task);
1774 static void idle_worker_timeout(unsigned long __pool)
1776 struct worker_pool *pool = (void *)__pool;
1778 spin_lock_irq(&pool->lock);
1780 while (too_many_workers(pool)) {
1781 struct worker *worker;
1782 unsigned long expires;
1784 /* idle_list is kept in LIFO order, check the last one */
1785 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1786 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1788 if (time_before(jiffies, expires)) {
1789 mod_timer(&pool->idle_timer, expires);
1793 destroy_worker(worker);
1796 spin_unlock_irq(&pool->lock);
1799 static void send_mayday(struct work_struct *work)
1801 struct pool_workqueue *pwq = get_work_pwq(work);
1802 struct workqueue_struct *wq = pwq->wq;
1804 lockdep_assert_held(&wq_mayday_lock);
1809 /* mayday mayday mayday */
1810 if (list_empty(&pwq->mayday_node)) {
1812 * If @pwq is for an unbound wq, its base ref may be put at
1813 * any time due to an attribute change. Pin @pwq until the
1814 * rescuer is done with it.
1817 list_add_tail(&pwq->mayday_node, &wq->maydays);
1818 wake_up_process(wq->rescuer->task);
1822 static void pool_mayday_timeout(unsigned long __pool)
1824 struct worker_pool *pool = (void *)__pool;
1825 struct work_struct *work;
1827 spin_lock_irq(&pool->lock);
1828 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1830 if (need_to_create_worker(pool)) {
1832 * We've been trying to create a new worker but
1833 * haven't been successful. We might be hitting an
1834 * allocation deadlock. Send distress signals to
1837 list_for_each_entry(work, &pool->worklist, entry)
1841 spin_unlock(&wq_mayday_lock);
1842 spin_unlock_irq(&pool->lock);
1844 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1848 * maybe_create_worker - create a new worker if necessary
1849 * @pool: pool to create a new worker for
1851 * Create a new worker for @pool if necessary. @pool is guaranteed to
1852 * have at least one idle worker on return from this function. If
1853 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1854 * sent to all rescuers with works scheduled on @pool to resolve
1855 * possible allocation deadlock.
1857 * On return, need_to_create_worker() is guaranteed to be %false and
1858 * may_start_working() %true.
1861 * spin_lock_irq(pool->lock) which may be released and regrabbed
1862 * multiple times. Does GFP_KERNEL allocations. Called only from
1865 static void maybe_create_worker(struct worker_pool *pool)
1866 __releases(&pool->lock)
1867 __acquires(&pool->lock)
1870 spin_unlock_irq(&pool->lock);
1872 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1873 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1876 if (create_worker(pool) || !need_to_create_worker(pool))
1879 schedule_timeout_interruptible(CREATE_COOLDOWN);
1881 if (!need_to_create_worker(pool))
1885 del_timer_sync(&pool->mayday_timer);
1886 spin_lock_irq(&pool->lock);
1888 * This is necessary even after a new worker was just successfully
1889 * created as @pool->lock was dropped and the new worker might have
1890 * already become busy.
1892 if (need_to_create_worker(pool))
1897 * manage_workers - manage worker pool
1900 * Assume the manager role and manage the worker pool @worker belongs
1901 * to. At any given time, there can be only zero or one manager per
1902 * pool. The exclusion is handled automatically by this function.
1904 * The caller can safely start processing works on false return. On
1905 * true return, it's guaranteed that need_to_create_worker() is false
1906 * and may_start_working() is true.
1909 * spin_lock_irq(pool->lock) which may be released and regrabbed
1910 * multiple times. Does GFP_KERNEL allocations.
1913 * %false if the pool doesn't need management and the caller can safely
1914 * start processing works, %true if management function was performed and
1915 * the conditions that the caller verified before calling the function may
1916 * no longer be true.
1918 static bool manage_workers(struct worker *worker)
1920 struct worker_pool *pool = worker->pool;
1923 * Anyone who successfully grabs manager_arb wins the arbitration
1924 * and becomes the manager. mutex_trylock() on pool->manager_arb
1925 * failure while holding pool->lock reliably indicates that someone
1926 * else is managing the pool and the worker which failed trylock
1927 * can proceed to executing work items. This means that anyone
1928 * grabbing manager_arb is responsible for actually performing
1929 * manager duties. If manager_arb is grabbed and released without
1930 * actual management, the pool may stall indefinitely.
1932 if (!mutex_trylock(&pool->manager_arb))
1934 pool->manager = worker;
1936 maybe_create_worker(pool);
1938 pool->manager = NULL;
1939 mutex_unlock(&pool->manager_arb);
1944 * process_one_work - process single work
1946 * @work: work to process
1948 * Process @work. This function contains all the logics necessary to
1949 * process a single work including synchronization against and
1950 * interaction with other workers on the same cpu, queueing and
1951 * flushing. As long as context requirement is met, any worker can
1952 * call this function to process a work.
1955 * spin_lock_irq(pool->lock) which is released and regrabbed.
1957 static void process_one_work(struct worker *worker, struct work_struct *work)
1958 __releases(&pool->lock)
1959 __acquires(&pool->lock)
1961 struct pool_workqueue *pwq = get_work_pwq(work);
1962 struct worker_pool *pool = worker->pool;
1963 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1965 struct worker *collision;
1966 #ifdef CONFIG_LOCKDEP
1968 * It is permissible to free the struct work_struct from
1969 * inside the function that is called from it, this we need to
1970 * take into account for lockdep too. To avoid bogus "held
1971 * lock freed" warnings as well as problems when looking into
1972 * work->lockdep_map, make a copy and use that here.
1974 struct lockdep_map lockdep_map;
1976 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1978 /* ensure we're on the correct CPU */
1979 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1980 raw_smp_processor_id() != pool->cpu);
1983 * A single work shouldn't be executed concurrently by
1984 * multiple workers on a single cpu. Check whether anyone is
1985 * already processing the work. If so, defer the work to the
1986 * currently executing one.
1988 collision = find_worker_executing_work(pool, work);
1989 if (unlikely(collision)) {
1990 move_linked_works(work, &collision->scheduled, NULL);
1994 /* claim and dequeue */
1995 debug_work_deactivate(work);
1996 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
1997 worker->current_work = work;
1998 worker->current_func = work->func;
1999 worker->current_pwq = pwq;
2000 work_color = get_work_color(work);
2002 list_del_init(&work->entry);
2005 * CPU intensive works don't participate in concurrency management.
2006 * They're the scheduler's responsibility. This takes @worker out
2007 * of concurrency management and the next code block will chain
2008 * execution of the pending work items.
2010 if (unlikely(cpu_intensive))
2011 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2014 * Wake up another worker if necessary. The condition is always
2015 * false for normal per-cpu workers since nr_running would always
2016 * be >= 1 at this point. This is used to chain execution of the
2017 * pending work items for WORKER_NOT_RUNNING workers such as the
2018 * UNBOUND and CPU_INTENSIVE ones.
2020 if (need_more_worker(pool))
2021 wake_up_worker(pool);
2024 * Record the last pool and clear PENDING which should be the last
2025 * update to @work. Also, do this inside @pool->lock so that
2026 * PENDING and queued state changes happen together while IRQ is
2029 set_work_pool_and_clear_pending(work, pool->id);
2031 spin_unlock_irq(&pool->lock);
2033 lock_map_acquire_read(&pwq->wq->lockdep_map);
2034 lock_map_acquire(&lockdep_map);
2035 trace_workqueue_execute_start(work);
2036 worker->current_func(work);
2038 * While we must be careful to not use "work" after this, the trace
2039 * point will only record its address.
2041 trace_workqueue_execute_end(work);
2042 lock_map_release(&lockdep_map);
2043 lock_map_release(&pwq->wq->lockdep_map);
2045 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2046 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2047 " last function: %pf\n",
2048 current->comm, preempt_count(), task_pid_nr(current),
2049 worker->current_func);
2050 debug_show_held_locks(current);
2055 * The following prevents a kworker from hogging CPU on !PREEMPT
2056 * kernels, where a requeueing work item waiting for something to
2057 * happen could deadlock with stop_machine as such work item could
2058 * indefinitely requeue itself while all other CPUs are trapped in
2059 * stop_machine. At the same time, report a quiescent RCU state so
2060 * the same condition doesn't freeze RCU.
2062 cond_resched_rcu_qs();
2064 spin_lock_irq(&pool->lock);
2066 /* clear cpu intensive status */
2067 if (unlikely(cpu_intensive))
2068 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2070 /* we're done with it, release */
2071 hash_del(&worker->hentry);
2072 worker->current_work = NULL;
2073 worker->current_func = NULL;
2074 worker->current_pwq = NULL;
2075 worker->desc_valid = false;
2076 pwq_dec_nr_in_flight(pwq, work_color);
2080 * process_scheduled_works - process scheduled works
2083 * Process all scheduled works. Please note that the scheduled list
2084 * may change while processing a work, so this function repeatedly
2085 * fetches a work from the top and executes it.
2088 * spin_lock_irq(pool->lock) which may be released and regrabbed
2091 static void process_scheduled_works(struct worker *worker)
2093 while (!list_empty(&worker->scheduled)) {
2094 struct work_struct *work = list_first_entry(&worker->scheduled,
2095 struct work_struct, entry);
2096 process_one_work(worker, work);
2101 * worker_thread - the worker thread function
2104 * The worker thread function. All workers belong to a worker_pool -
2105 * either a per-cpu one or dynamic unbound one. These workers process all
2106 * work items regardless of their specific target workqueue. The only
2107 * exception is work items which belong to workqueues with a rescuer which
2108 * will be explained in rescuer_thread().
2112 static int worker_thread(void *__worker)
2114 struct worker *worker = __worker;
2115 struct worker_pool *pool = worker->pool;
2117 /* tell the scheduler that this is a workqueue worker */
2118 worker->task->flags |= PF_WQ_WORKER;
2120 spin_lock_irq(&pool->lock);
2122 /* am I supposed to die? */
2123 if (unlikely(worker->flags & WORKER_DIE)) {
2124 spin_unlock_irq(&pool->lock);
2125 WARN_ON_ONCE(!list_empty(&worker->entry));
2126 worker->task->flags &= ~PF_WQ_WORKER;
2128 set_task_comm(worker->task, "kworker/dying");
2129 ida_simple_remove(&pool->worker_ida, worker->id);
2130 worker_detach_from_pool(worker, pool);
2135 worker_leave_idle(worker);
2137 /* no more worker necessary? */
2138 if (!need_more_worker(pool))
2141 /* do we need to manage? */
2142 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2146 * ->scheduled list can only be filled while a worker is
2147 * preparing to process a work or actually processing it.
2148 * Make sure nobody diddled with it while I was sleeping.
2150 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2153 * Finish PREP stage. We're guaranteed to have at least one idle
2154 * worker or that someone else has already assumed the manager
2155 * role. This is where @worker starts participating in concurrency
2156 * management if applicable and concurrency management is restored
2157 * after being rebound. See rebind_workers() for details.
2159 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2162 struct work_struct *work =
2163 list_first_entry(&pool->worklist,
2164 struct work_struct, entry);
2166 pool->watchdog_ts = jiffies;
2168 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2169 /* optimization path, not strictly necessary */
2170 process_one_work(worker, work);
2171 if (unlikely(!list_empty(&worker->scheduled)))
2172 process_scheduled_works(worker);
2174 move_linked_works(work, &worker->scheduled, NULL);
2175 process_scheduled_works(worker);
2177 } while (keep_working(pool));
2179 worker_set_flags(worker, WORKER_PREP);
2182 * pool->lock is held and there's no work to process and no need to
2183 * manage, sleep. Workers are woken up only while holding
2184 * pool->lock or from local cpu, so setting the current state
2185 * before releasing pool->lock is enough to prevent losing any
2188 worker_enter_idle(worker);
2189 __set_current_state(TASK_INTERRUPTIBLE);
2190 spin_unlock_irq(&pool->lock);
2196 * rescuer_thread - the rescuer thread function
2199 * Workqueue rescuer thread function. There's one rescuer for each
2200 * workqueue which has WQ_MEM_RECLAIM set.
2202 * Regular work processing on a pool may block trying to create a new
2203 * worker which uses GFP_KERNEL allocation which has slight chance of
2204 * developing into deadlock if some works currently on the same queue
2205 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2206 * the problem rescuer solves.
2208 * When such condition is possible, the pool summons rescuers of all
2209 * workqueues which have works queued on the pool and let them process
2210 * those works so that forward progress can be guaranteed.
2212 * This should happen rarely.
2216 static int rescuer_thread(void *__rescuer)
2218 struct worker *rescuer = __rescuer;
2219 struct workqueue_struct *wq = rescuer->rescue_wq;
2220 struct list_head *scheduled = &rescuer->scheduled;
2223 set_user_nice(current, RESCUER_NICE_LEVEL);
2226 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2227 * doesn't participate in concurrency management.
2229 rescuer->task->flags |= PF_WQ_WORKER;
2231 set_current_state(TASK_INTERRUPTIBLE);
2234 * By the time the rescuer is requested to stop, the workqueue
2235 * shouldn't have any work pending, but @wq->maydays may still have
2236 * pwq(s) queued. This can happen by non-rescuer workers consuming
2237 * all the work items before the rescuer got to them. Go through
2238 * @wq->maydays processing before acting on should_stop so that the
2239 * list is always empty on exit.
2241 should_stop = kthread_should_stop();
2243 /* see whether any pwq is asking for help */
2244 spin_lock_irq(&wq_mayday_lock);
2246 while (!list_empty(&wq->maydays)) {
2247 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2248 struct pool_workqueue, mayday_node);
2249 struct worker_pool *pool = pwq->pool;
2250 struct work_struct *work, *n;
2253 __set_current_state(TASK_RUNNING);
2254 list_del_init(&pwq->mayday_node);
2256 spin_unlock_irq(&wq_mayday_lock);
2258 worker_attach_to_pool(rescuer, pool);
2260 spin_lock_irq(&pool->lock);
2261 rescuer->pool = pool;
2264 * Slurp in all works issued via this workqueue and
2267 WARN_ON_ONCE(!list_empty(scheduled));
2268 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2269 if (get_work_pwq(work) == pwq) {
2271 pool->watchdog_ts = jiffies;
2272 move_linked_works(work, scheduled, &n);
2277 if (!list_empty(scheduled)) {
2278 process_scheduled_works(rescuer);
2281 * The above execution of rescued work items could
2282 * have created more to rescue through
2283 * pwq_activate_first_delayed() or chained
2284 * queueing. Let's put @pwq back on mayday list so
2285 * that such back-to-back work items, which may be
2286 * being used to relieve memory pressure, don't
2287 * incur MAYDAY_INTERVAL delay inbetween.
2289 if (need_to_create_worker(pool)) {
2290 spin_lock(&wq_mayday_lock);
2292 list_move_tail(&pwq->mayday_node, &wq->maydays);
2293 spin_unlock(&wq_mayday_lock);
2298 * Put the reference grabbed by send_mayday(). @pool won't
2299 * go away while we're still attached to it.
2304 * Leave this pool. If need_more_worker() is %true, notify a
2305 * regular worker; otherwise, we end up with 0 concurrency
2306 * and stalling the execution.
2308 if (need_more_worker(pool))
2309 wake_up_worker(pool);
2311 rescuer->pool = NULL;
2312 spin_unlock_irq(&pool->lock);
2314 worker_detach_from_pool(rescuer, pool);
2316 spin_lock_irq(&wq_mayday_lock);
2319 spin_unlock_irq(&wq_mayday_lock);
2322 __set_current_state(TASK_RUNNING);
2323 rescuer->task->flags &= ~PF_WQ_WORKER;
2327 /* rescuers should never participate in concurrency management */
2328 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2334 * check_flush_dependency - check for flush dependency sanity
2335 * @target_wq: workqueue being flushed
2336 * @target_work: work item being flushed (NULL for workqueue flushes)
2338 * %current is trying to flush the whole @target_wq or @target_work on it.
2339 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2340 * reclaiming memory or running on a workqueue which doesn't have
2341 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2344 static void check_flush_dependency(struct workqueue_struct *target_wq,
2345 struct work_struct *target_work)
2347 work_func_t target_func = target_work ? target_work->func : NULL;
2348 struct worker *worker;
2350 if (target_wq->flags & WQ_MEM_RECLAIM)
2353 worker = current_wq_worker();
2355 WARN_ONCE(current->flags & PF_MEMALLOC,
2356 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2357 current->pid, current->comm, target_wq->name, target_func);
2358 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2359 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2360 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2361 worker->current_pwq->wq->name, worker->current_func,
2362 target_wq->name, target_func);
2366 struct work_struct work;
2367 struct completion done;
2368 struct task_struct *task; /* purely informational */
2371 static void wq_barrier_func(struct work_struct *work)
2373 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2374 complete(&barr->done);
2378 * insert_wq_barrier - insert a barrier work
2379 * @pwq: pwq to insert barrier into
2380 * @barr: wq_barrier to insert
2381 * @target: target work to attach @barr to
2382 * @worker: worker currently executing @target, NULL if @target is not executing
2384 * @barr is linked to @target such that @barr is completed only after
2385 * @target finishes execution. Please note that the ordering
2386 * guarantee is observed only with respect to @target and on the local
2389 * Currently, a queued barrier can't be canceled. This is because
2390 * try_to_grab_pending() can't determine whether the work to be
2391 * grabbed is at the head of the queue and thus can't clear LINKED
2392 * flag of the previous work while there must be a valid next work
2393 * after a work with LINKED flag set.
2395 * Note that when @worker is non-NULL, @target may be modified
2396 * underneath us, so we can't reliably determine pwq from @target.
2399 * spin_lock_irq(pool->lock).
2401 static void insert_wq_barrier(struct pool_workqueue *pwq,
2402 struct wq_barrier *barr,
2403 struct work_struct *target, struct worker *worker)
2405 struct list_head *head;
2406 unsigned int linked = 0;
2409 * debugobject calls are safe here even with pool->lock locked
2410 * as we know for sure that this will not trigger any of the
2411 * checks and call back into the fixup functions where we
2414 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2415 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2416 init_completion(&barr->done);
2417 barr->task = current;
2420 * If @target is currently being executed, schedule the
2421 * barrier to the worker; otherwise, put it after @target.
2424 head = worker->scheduled.next;
2426 unsigned long *bits = work_data_bits(target);
2428 head = target->entry.next;
2429 /* there can already be other linked works, inherit and set */
2430 linked = *bits & WORK_STRUCT_LINKED;
2431 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2434 debug_work_activate(&barr->work);
2435 insert_work(pwq, &barr->work, head,
2436 work_color_to_flags(WORK_NO_COLOR) | linked);
2440 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2441 * @wq: workqueue being flushed
2442 * @flush_color: new flush color, < 0 for no-op
2443 * @work_color: new work color, < 0 for no-op
2445 * Prepare pwqs for workqueue flushing.
2447 * If @flush_color is non-negative, flush_color on all pwqs should be
2448 * -1. If no pwq has in-flight commands at the specified color, all
2449 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2450 * has in flight commands, its pwq->flush_color is set to
2451 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2452 * wakeup logic is armed and %true is returned.
2454 * The caller should have initialized @wq->first_flusher prior to
2455 * calling this function with non-negative @flush_color. If
2456 * @flush_color is negative, no flush color update is done and %false
2459 * If @work_color is non-negative, all pwqs should have the same
2460 * work_color which is previous to @work_color and all will be
2461 * advanced to @work_color.
2464 * mutex_lock(wq->mutex).
2467 * %true if @flush_color >= 0 and there's something to flush. %false
2470 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2471 int flush_color, int work_color)
2474 struct pool_workqueue *pwq;
2476 if (flush_color >= 0) {
2477 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2478 atomic_set(&wq->nr_pwqs_to_flush, 1);
2481 for_each_pwq(pwq, wq) {
2482 struct worker_pool *pool = pwq->pool;
2484 spin_lock_irq(&pool->lock);
2486 if (flush_color >= 0) {
2487 WARN_ON_ONCE(pwq->flush_color != -1);
2489 if (pwq->nr_in_flight[flush_color]) {
2490 pwq->flush_color = flush_color;
2491 atomic_inc(&wq->nr_pwqs_to_flush);
2496 if (work_color >= 0) {
2497 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2498 pwq->work_color = work_color;
2501 spin_unlock_irq(&pool->lock);
2504 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2505 complete(&wq->first_flusher->done);
2511 * flush_workqueue - ensure that any scheduled work has run to completion.
2512 * @wq: workqueue to flush
2514 * This function sleeps until all work items which were queued on entry
2515 * have finished execution, but it is not livelocked by new incoming ones.
2517 void flush_workqueue(struct workqueue_struct *wq)
2519 struct wq_flusher this_flusher = {
2520 .list = LIST_HEAD_INIT(this_flusher.list),
2522 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2526 lock_map_acquire(&wq->lockdep_map);
2527 lock_map_release(&wq->lockdep_map);
2529 mutex_lock(&wq->mutex);
2532 * Start-to-wait phase
2534 next_color = work_next_color(wq->work_color);
2536 if (next_color != wq->flush_color) {
2538 * Color space is not full. The current work_color
2539 * becomes our flush_color and work_color is advanced
2542 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2543 this_flusher.flush_color = wq->work_color;
2544 wq->work_color = next_color;
2546 if (!wq->first_flusher) {
2547 /* no flush in progress, become the first flusher */
2548 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2550 wq->first_flusher = &this_flusher;
2552 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2554 /* nothing to flush, done */
2555 wq->flush_color = next_color;
2556 wq->first_flusher = NULL;
2561 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2562 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2563 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2567 * Oops, color space is full, wait on overflow queue.
2568 * The next flush completion will assign us
2569 * flush_color and transfer to flusher_queue.
2571 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2574 check_flush_dependency(wq, NULL);
2576 mutex_unlock(&wq->mutex);
2578 wait_for_completion(&this_flusher.done);
2581 * Wake-up-and-cascade phase
2583 * First flushers are responsible for cascading flushes and
2584 * handling overflow. Non-first flushers can simply return.
2586 if (wq->first_flusher != &this_flusher)
2589 mutex_lock(&wq->mutex);
2591 /* we might have raced, check again with mutex held */
2592 if (wq->first_flusher != &this_flusher)
2595 wq->first_flusher = NULL;
2597 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2598 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2601 struct wq_flusher *next, *tmp;
2603 /* complete all the flushers sharing the current flush color */
2604 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2605 if (next->flush_color != wq->flush_color)
2607 list_del_init(&next->list);
2608 complete(&next->done);
2611 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2612 wq->flush_color != work_next_color(wq->work_color));
2614 /* this flush_color is finished, advance by one */
2615 wq->flush_color = work_next_color(wq->flush_color);
2617 /* one color has been freed, handle overflow queue */
2618 if (!list_empty(&wq->flusher_overflow)) {
2620 * Assign the same color to all overflowed
2621 * flushers, advance work_color and append to
2622 * flusher_queue. This is the start-to-wait
2623 * phase for these overflowed flushers.
2625 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2626 tmp->flush_color = wq->work_color;
2628 wq->work_color = work_next_color(wq->work_color);
2630 list_splice_tail_init(&wq->flusher_overflow,
2631 &wq->flusher_queue);
2632 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2635 if (list_empty(&wq->flusher_queue)) {
2636 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2641 * Need to flush more colors. Make the next flusher
2642 * the new first flusher and arm pwqs.
2644 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2645 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2647 list_del_init(&next->list);
2648 wq->first_flusher = next;
2650 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2654 * Meh... this color is already done, clear first
2655 * flusher and repeat cascading.
2657 wq->first_flusher = NULL;
2661 mutex_unlock(&wq->mutex);
2663 EXPORT_SYMBOL(flush_workqueue);
2666 * drain_workqueue - drain a workqueue
2667 * @wq: workqueue to drain
2669 * Wait until the workqueue becomes empty. While draining is in progress,
2670 * only chain queueing is allowed. IOW, only currently pending or running
2671 * work items on @wq can queue further work items on it. @wq is flushed
2672 * repeatedly until it becomes empty. The number of flushing is determined
2673 * by the depth of chaining and should be relatively short. Whine if it
2676 void drain_workqueue(struct workqueue_struct *wq)
2678 unsigned int flush_cnt = 0;
2679 struct pool_workqueue *pwq;
2682 * __queue_work() needs to test whether there are drainers, is much
2683 * hotter than drain_workqueue() and already looks at @wq->flags.
2684 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2686 mutex_lock(&wq->mutex);
2687 if (!wq->nr_drainers++)
2688 wq->flags |= __WQ_DRAINING;
2689 mutex_unlock(&wq->mutex);
2691 flush_workqueue(wq);
2693 mutex_lock(&wq->mutex);
2695 for_each_pwq(pwq, wq) {
2698 spin_lock_irq(&pwq->pool->lock);
2699 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2700 spin_unlock_irq(&pwq->pool->lock);
2705 if (++flush_cnt == 10 ||
2706 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2707 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2708 wq->name, flush_cnt);
2710 mutex_unlock(&wq->mutex);
2714 if (!--wq->nr_drainers)
2715 wq->flags &= ~__WQ_DRAINING;
2716 mutex_unlock(&wq->mutex);
2718 EXPORT_SYMBOL_GPL(drain_workqueue);
2720 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2722 struct worker *worker = NULL;
2723 struct worker_pool *pool;
2724 struct pool_workqueue *pwq;
2728 local_irq_disable();
2729 pool = get_work_pool(work);
2735 spin_lock(&pool->lock);
2736 /* see the comment in try_to_grab_pending() with the same code */
2737 pwq = get_work_pwq(work);
2739 if (unlikely(pwq->pool != pool))
2742 worker = find_worker_executing_work(pool, work);
2745 pwq = worker->current_pwq;
2748 check_flush_dependency(pwq->wq, work);
2750 insert_wq_barrier(pwq, barr, work, worker);
2751 spin_unlock_irq(&pool->lock);
2754 * If @max_active is 1 or rescuer is in use, flushing another work
2755 * item on the same workqueue may lead to deadlock. Make sure the
2756 * flusher is not running on the same workqueue by verifying write
2759 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2760 lock_map_acquire(&pwq->wq->lockdep_map);
2762 lock_map_acquire_read(&pwq->wq->lockdep_map);
2763 lock_map_release(&pwq->wq->lockdep_map);
2767 spin_unlock_irq(&pool->lock);
2772 * flush_work - wait for a work to finish executing the last queueing instance
2773 * @work: the work to flush
2775 * Wait until @work has finished execution. @work is guaranteed to be idle
2776 * on return if it hasn't been requeued since flush started.
2779 * %true if flush_work() waited for the work to finish execution,
2780 * %false if it was already idle.
2782 bool flush_work(struct work_struct *work)
2784 struct wq_barrier barr;
2786 lock_map_acquire(&work->lockdep_map);
2787 lock_map_release(&work->lockdep_map);
2789 if (start_flush_work(work, &barr)) {
2790 wait_for_completion(&barr.done);
2791 destroy_work_on_stack(&barr.work);
2797 EXPORT_SYMBOL_GPL(flush_work);
2801 struct work_struct *work;
2804 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2806 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2808 if (cwait->work != key)
2810 return autoremove_wake_function(wait, mode, sync, key);
2813 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2815 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2816 unsigned long flags;
2820 ret = try_to_grab_pending(work, is_dwork, &flags);
2822 * If someone else is already canceling, wait for it to
2823 * finish. flush_work() doesn't work for PREEMPT_NONE
2824 * because we may get scheduled between @work's completion
2825 * and the other canceling task resuming and clearing
2826 * CANCELING - flush_work() will return false immediately
2827 * as @work is no longer busy, try_to_grab_pending() will
2828 * return -ENOENT as @work is still being canceled and the
2829 * other canceling task won't be able to clear CANCELING as
2830 * we're hogging the CPU.
2832 * Let's wait for completion using a waitqueue. As this
2833 * may lead to the thundering herd problem, use a custom
2834 * wake function which matches @work along with exclusive
2837 if (unlikely(ret == -ENOENT)) {
2838 struct cwt_wait cwait;
2840 init_wait(&cwait.wait);
2841 cwait.wait.func = cwt_wakefn;
2844 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2845 TASK_UNINTERRUPTIBLE);
2846 if (work_is_canceling(work))
2848 finish_wait(&cancel_waitq, &cwait.wait);
2850 } while (unlikely(ret < 0));
2852 /* tell other tasks trying to grab @work to back off */
2853 mark_work_canceling(work);
2854 local_irq_restore(flags);
2857 clear_work_data(work);
2860 * Paired with prepare_to_wait() above so that either
2861 * waitqueue_active() is visible here or !work_is_canceling() is
2865 if (waitqueue_active(&cancel_waitq))
2866 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2872 * cancel_work_sync - cancel a work and wait for it to finish
2873 * @work: the work to cancel
2875 * Cancel @work and wait for its execution to finish. This function
2876 * can be used even if the work re-queues itself or migrates to
2877 * another workqueue. On return from this function, @work is
2878 * guaranteed to be not pending or executing on any CPU.
2880 * cancel_work_sync(&delayed_work->work) must not be used for
2881 * delayed_work's. Use cancel_delayed_work_sync() instead.
2883 * The caller must ensure that the workqueue on which @work was last
2884 * queued can't be destroyed before this function returns.
2887 * %true if @work was pending, %false otherwise.
2889 bool cancel_work_sync(struct work_struct *work)
2891 return __cancel_work_timer(work, false);
2893 EXPORT_SYMBOL_GPL(cancel_work_sync);
2896 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2897 * @dwork: the delayed work to flush
2899 * Delayed timer is cancelled and the pending work is queued for
2900 * immediate execution. Like flush_work(), this function only
2901 * considers the last queueing instance of @dwork.
2904 * %true if flush_work() waited for the work to finish execution,
2905 * %false if it was already idle.
2907 bool flush_delayed_work(struct delayed_work *dwork)
2909 local_irq_disable();
2910 if (del_timer_sync(&dwork->timer))
2911 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2913 return flush_work(&dwork->work);
2915 EXPORT_SYMBOL(flush_delayed_work);
2918 * cancel_delayed_work - cancel a delayed work
2919 * @dwork: delayed_work to cancel
2921 * Kill off a pending delayed_work.
2923 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2927 * The work callback function may still be running on return, unless
2928 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2929 * use cancel_delayed_work_sync() to wait on it.
2931 * This function is safe to call from any context including IRQ handler.
2933 bool cancel_delayed_work(struct delayed_work *dwork)
2935 unsigned long flags;
2939 ret = try_to_grab_pending(&dwork->work, true, &flags);
2940 } while (unlikely(ret == -EAGAIN));
2942 if (unlikely(ret < 0))
2945 set_work_pool_and_clear_pending(&dwork->work,
2946 get_work_pool_id(&dwork->work));
2947 local_irq_restore(flags);
2950 EXPORT_SYMBOL(cancel_delayed_work);
2953 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2954 * @dwork: the delayed work cancel
2956 * This is cancel_work_sync() for delayed works.
2959 * %true if @dwork was pending, %false otherwise.
2961 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2963 return __cancel_work_timer(&dwork->work, true);
2965 EXPORT_SYMBOL(cancel_delayed_work_sync);
2968 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2969 * @func: the function to call
2971 * schedule_on_each_cpu() executes @func on each online CPU using the
2972 * system workqueue and blocks until all CPUs have completed.
2973 * schedule_on_each_cpu() is very slow.
2976 * 0 on success, -errno on failure.
2978 int schedule_on_each_cpu(work_func_t func)
2981 struct work_struct __percpu *works;
2983 works = alloc_percpu(struct work_struct);
2989 for_each_online_cpu(cpu) {
2990 struct work_struct *work = per_cpu_ptr(works, cpu);
2992 INIT_WORK(work, func);
2993 schedule_work_on(cpu, work);
2996 for_each_online_cpu(cpu)
2997 flush_work(per_cpu_ptr(works, cpu));
3005 * execute_in_process_context - reliably execute the routine with user context
3006 * @fn: the function to execute
3007 * @ew: guaranteed storage for the execute work structure (must
3008 * be available when the work executes)
3010 * Executes the function immediately if process context is available,
3011 * otherwise schedules the function for delayed execution.
3013 * Return: 0 - function was executed
3014 * 1 - function was scheduled for execution
3016 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3018 if (!in_interrupt()) {
3023 INIT_WORK(&ew->work, fn);
3024 schedule_work(&ew->work);
3028 EXPORT_SYMBOL_GPL(execute_in_process_context);
3031 * free_workqueue_attrs - free a workqueue_attrs
3032 * @attrs: workqueue_attrs to free
3034 * Undo alloc_workqueue_attrs().
3036 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3039 free_cpumask_var(attrs->cpumask);
3045 * alloc_workqueue_attrs - allocate a workqueue_attrs
3046 * @gfp_mask: allocation mask to use
3048 * Allocate a new workqueue_attrs, initialize with default settings and
3051 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3053 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3055 struct workqueue_attrs *attrs;
3057 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3060 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3063 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3066 free_workqueue_attrs(attrs);
3070 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3071 const struct workqueue_attrs *from)
3073 to->nice = from->nice;
3074 cpumask_copy(to->cpumask, from->cpumask);
3076 * Unlike hash and equality test, this function doesn't ignore
3077 * ->no_numa as it is used for both pool and wq attrs. Instead,
3078 * get_unbound_pool() explicitly clears ->no_numa after copying.
3080 to->no_numa = from->no_numa;
3083 /* hash value of the content of @attr */
3084 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3088 hash = jhash_1word(attrs->nice, hash);
3089 hash = jhash(cpumask_bits(attrs->cpumask),
3090 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3094 /* content equality test */
3095 static bool wqattrs_equal(const struct workqueue_attrs *a,
3096 const struct workqueue_attrs *b)
3098 if (a->nice != b->nice)
3100 if (!cpumask_equal(a->cpumask, b->cpumask))
3106 * init_worker_pool - initialize a newly zalloc'd worker_pool
3107 * @pool: worker_pool to initialize
3109 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3111 * Return: 0 on success, -errno on failure. Even on failure, all fields
3112 * inside @pool proper are initialized and put_unbound_pool() can be called
3113 * on @pool safely to release it.
3115 static int init_worker_pool(struct worker_pool *pool)
3117 spin_lock_init(&pool->lock);
3120 pool->node = NUMA_NO_NODE;
3121 pool->flags |= POOL_DISASSOCIATED;
3122 pool->watchdog_ts = jiffies;
3123 INIT_LIST_HEAD(&pool->worklist);
3124 INIT_LIST_HEAD(&pool->idle_list);
3125 hash_init(pool->busy_hash);
3127 init_timer_deferrable(&pool->idle_timer);
3128 pool->idle_timer.function = idle_worker_timeout;
3129 pool->idle_timer.data = (unsigned long)pool;
3131 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3132 (unsigned long)pool);
3134 mutex_init(&pool->manager_arb);
3135 mutex_init(&pool->attach_mutex);
3136 INIT_LIST_HEAD(&pool->workers);
3138 ida_init(&pool->worker_ida);
3139 INIT_HLIST_NODE(&pool->hash_node);
3142 /* shouldn't fail above this point */
3143 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3149 static void rcu_free_wq(struct rcu_head *rcu)
3151 struct workqueue_struct *wq =
3152 container_of(rcu, struct workqueue_struct, rcu);
3154 if (!(wq->flags & WQ_UNBOUND))
3155 free_percpu(wq->cpu_pwqs);
3157 free_workqueue_attrs(wq->unbound_attrs);
3163 static void rcu_free_pool(struct rcu_head *rcu)
3165 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3167 ida_destroy(&pool->worker_ida);
3168 free_workqueue_attrs(pool->attrs);
3173 * put_unbound_pool - put a worker_pool
3174 * @pool: worker_pool to put
3176 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3177 * safe manner. get_unbound_pool() calls this function on its failure path
3178 * and this function should be able to release pools which went through,
3179 * successfully or not, init_worker_pool().
3181 * Should be called with wq_pool_mutex held.
3183 static void put_unbound_pool(struct worker_pool *pool)
3185 DECLARE_COMPLETION_ONSTACK(detach_completion);
3186 struct worker *worker;
3188 lockdep_assert_held(&wq_pool_mutex);
3194 if (WARN_ON(!(pool->cpu < 0)) ||
3195 WARN_ON(!list_empty(&pool->worklist)))
3198 /* release id and unhash */
3200 idr_remove(&worker_pool_idr, pool->id);
3201 hash_del(&pool->hash_node);
3204 * Become the manager and destroy all workers. Grabbing
3205 * manager_arb prevents @pool's workers from blocking on
3208 mutex_lock(&pool->manager_arb);
3210 spin_lock_irq(&pool->lock);
3211 while ((worker = first_idle_worker(pool)))
3212 destroy_worker(worker);
3213 WARN_ON(pool->nr_workers || pool->nr_idle);
3214 spin_unlock_irq(&pool->lock);
3216 mutex_lock(&pool->attach_mutex);
3217 if (!list_empty(&pool->workers))
3218 pool->detach_completion = &detach_completion;
3219 mutex_unlock(&pool->attach_mutex);
3221 if (pool->detach_completion)
3222 wait_for_completion(pool->detach_completion);
3224 mutex_unlock(&pool->manager_arb);
3226 /* shut down the timers */
3227 del_timer_sync(&pool->idle_timer);
3228 del_timer_sync(&pool->mayday_timer);
3230 /* sched-RCU protected to allow dereferences from get_work_pool() */
3231 call_rcu_sched(&pool->rcu, rcu_free_pool);
3235 * get_unbound_pool - get a worker_pool with the specified attributes
3236 * @attrs: the attributes of the worker_pool to get
3238 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3239 * reference count and return it. If there already is a matching
3240 * worker_pool, it will be used; otherwise, this function attempts to
3243 * Should be called with wq_pool_mutex held.
3245 * Return: On success, a worker_pool with the same attributes as @attrs.
3246 * On failure, %NULL.
3248 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3250 u32 hash = wqattrs_hash(attrs);
3251 struct worker_pool *pool;
3253 int target_node = NUMA_NO_NODE;
3255 lockdep_assert_held(&wq_pool_mutex);
3257 /* do we already have a matching pool? */
3258 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3259 if (wqattrs_equal(pool->attrs, attrs)) {
3265 /* if cpumask is contained inside a NUMA node, we belong to that node */
3266 if (wq_numa_enabled) {
3267 for_each_node(node) {
3268 if (cpumask_subset(attrs->cpumask,
3269 wq_numa_possible_cpumask[node])) {
3276 /* nope, create a new one */
3277 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3278 if (!pool || init_worker_pool(pool) < 0)
3281 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3282 copy_workqueue_attrs(pool->attrs, attrs);
3283 pool->node = target_node;
3286 * no_numa isn't a worker_pool attribute, always clear it. See
3287 * 'struct workqueue_attrs' comments for detail.
3289 pool->attrs->no_numa = false;
3291 if (worker_pool_assign_id(pool) < 0)
3294 /* create and start the initial worker */
3295 if (!create_worker(pool))
3299 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3304 put_unbound_pool(pool);
3308 static void rcu_free_pwq(struct rcu_head *rcu)
3310 kmem_cache_free(pwq_cache,
3311 container_of(rcu, struct pool_workqueue, rcu));
3315 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3316 * and needs to be destroyed.
3318 static void pwq_unbound_release_workfn(struct work_struct *work)
3320 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3321 unbound_release_work);
3322 struct workqueue_struct *wq = pwq->wq;
3323 struct worker_pool *pool = pwq->pool;
3326 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3329 mutex_lock(&wq->mutex);
3330 list_del_rcu(&pwq->pwqs_node);
3331 is_last = list_empty(&wq->pwqs);
3332 mutex_unlock(&wq->mutex);
3334 mutex_lock(&wq_pool_mutex);
3335 put_unbound_pool(pool);
3336 mutex_unlock(&wq_pool_mutex);
3338 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3341 * If we're the last pwq going away, @wq is already dead and no one
3342 * is gonna access it anymore. Schedule RCU free.
3345 call_rcu_sched(&wq->rcu, rcu_free_wq);
3349 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3350 * @pwq: target pool_workqueue
3352 * If @pwq isn't freezing, set @pwq->max_active to the associated
3353 * workqueue's saved_max_active and activate delayed work items
3354 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3356 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3358 struct workqueue_struct *wq = pwq->wq;
3359 bool freezable = wq->flags & WQ_FREEZABLE;
3361 /* for @wq->saved_max_active */
3362 lockdep_assert_held(&wq->mutex);
3364 /* fast exit for non-freezable wqs */
3365 if (!freezable && pwq->max_active == wq->saved_max_active)
3368 spin_lock_irq(&pwq->pool->lock);
3371 * During [un]freezing, the caller is responsible for ensuring that
3372 * this function is called at least once after @workqueue_freezing
3373 * is updated and visible.
3375 if (!freezable || !workqueue_freezing) {
3376 pwq->max_active = wq->saved_max_active;
3378 while (!list_empty(&pwq->delayed_works) &&
3379 pwq->nr_active < pwq->max_active)
3380 pwq_activate_first_delayed(pwq);
3383 * Need to kick a worker after thawed or an unbound wq's
3384 * max_active is bumped. It's a slow path. Do it always.
3386 wake_up_worker(pwq->pool);
3388 pwq->max_active = 0;
3391 spin_unlock_irq(&pwq->pool->lock);
3394 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3395 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3396 struct worker_pool *pool)
3398 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3400 memset(pwq, 0, sizeof(*pwq));
3404 pwq->flush_color = -1;
3406 INIT_LIST_HEAD(&pwq->delayed_works);
3407 INIT_LIST_HEAD(&pwq->pwqs_node);
3408 INIT_LIST_HEAD(&pwq->mayday_node);
3409 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3412 /* sync @pwq with the current state of its associated wq and link it */
3413 static void link_pwq(struct pool_workqueue *pwq)
3415 struct workqueue_struct *wq = pwq->wq;
3417 lockdep_assert_held(&wq->mutex);
3419 /* may be called multiple times, ignore if already linked */
3420 if (!list_empty(&pwq->pwqs_node))
3423 /* set the matching work_color */
3424 pwq->work_color = wq->work_color;
3426 /* sync max_active to the current setting */
3427 pwq_adjust_max_active(pwq);
3430 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3433 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3434 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3435 const struct workqueue_attrs *attrs)
3437 struct worker_pool *pool;
3438 struct pool_workqueue *pwq;
3440 lockdep_assert_held(&wq_pool_mutex);
3442 pool = get_unbound_pool(attrs);
3446 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3448 put_unbound_pool(pool);
3452 init_pwq(pwq, wq, pool);
3457 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3458 * @attrs: the wq_attrs of the default pwq of the target workqueue
3459 * @node: the target NUMA node
3460 * @cpu_going_down: if >= 0, the CPU to consider as offline
3461 * @cpumask: outarg, the resulting cpumask
3463 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3464 * @cpu_going_down is >= 0, that cpu is considered offline during
3465 * calculation. The result is stored in @cpumask.
3467 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3468 * enabled and @node has online CPUs requested by @attrs, the returned
3469 * cpumask is the intersection of the possible CPUs of @node and
3472 * The caller is responsible for ensuring that the cpumask of @node stays
3475 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3478 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3479 int cpu_going_down, cpumask_t *cpumask)
3481 if (!wq_numa_enabled || attrs->no_numa)
3484 /* does @node have any online CPUs @attrs wants? */
3485 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3486 if (cpu_going_down >= 0)
3487 cpumask_clear_cpu(cpu_going_down, cpumask);
3489 if (cpumask_empty(cpumask))
3492 /* yeap, return possible CPUs in @node that @attrs wants */
3493 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3494 return !cpumask_equal(cpumask, attrs->cpumask);
3497 cpumask_copy(cpumask, attrs->cpumask);
3501 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3502 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3504 struct pool_workqueue *pwq)
3506 struct pool_workqueue *old_pwq;
3508 lockdep_assert_held(&wq_pool_mutex);
3509 lockdep_assert_held(&wq->mutex);
3511 /* link_pwq() can handle duplicate calls */
3514 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3515 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3519 /* context to store the prepared attrs & pwqs before applying */
3520 struct apply_wqattrs_ctx {
3521 struct workqueue_struct *wq; /* target workqueue */
3522 struct workqueue_attrs *attrs; /* attrs to apply */
3523 struct list_head list; /* queued for batching commit */
3524 struct pool_workqueue *dfl_pwq;
3525 struct pool_workqueue *pwq_tbl[];
3528 /* free the resources after success or abort */
3529 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3535 put_pwq_unlocked(ctx->pwq_tbl[node]);
3536 put_pwq_unlocked(ctx->dfl_pwq);
3538 free_workqueue_attrs(ctx->attrs);
3544 /* allocate the attrs and pwqs for later installation */
3545 static struct apply_wqattrs_ctx *
3546 apply_wqattrs_prepare(struct workqueue_struct *wq,
3547 const struct workqueue_attrs *attrs)
3549 struct apply_wqattrs_ctx *ctx;
3550 struct workqueue_attrs *new_attrs, *tmp_attrs;
3553 lockdep_assert_held(&wq_pool_mutex);
3555 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3558 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3559 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3560 if (!ctx || !new_attrs || !tmp_attrs)
3564 * Calculate the attrs of the default pwq.
3565 * If the user configured cpumask doesn't overlap with the
3566 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3568 copy_workqueue_attrs(new_attrs, attrs);
3569 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3570 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3571 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3574 * We may create multiple pwqs with differing cpumasks. Make a
3575 * copy of @new_attrs which will be modified and used to obtain
3578 copy_workqueue_attrs(tmp_attrs, new_attrs);
3581 * If something goes wrong during CPU up/down, we'll fall back to
3582 * the default pwq covering whole @attrs->cpumask. Always create
3583 * it even if we don't use it immediately.
3585 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3589 for_each_node(node) {
3590 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3591 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3592 if (!ctx->pwq_tbl[node])
3595 ctx->dfl_pwq->refcnt++;
3596 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3600 /* save the user configured attrs and sanitize it. */
3601 copy_workqueue_attrs(new_attrs, attrs);
3602 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3603 ctx->attrs = new_attrs;
3606 free_workqueue_attrs(tmp_attrs);
3610 free_workqueue_attrs(tmp_attrs);
3611 free_workqueue_attrs(new_attrs);
3612 apply_wqattrs_cleanup(ctx);
3616 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3617 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3621 /* all pwqs have been created successfully, let's install'em */
3622 mutex_lock(&ctx->wq->mutex);
3624 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3626 /* save the previous pwq and install the new one */
3628 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3629 ctx->pwq_tbl[node]);
3631 /* @dfl_pwq might not have been used, ensure it's linked */
3632 link_pwq(ctx->dfl_pwq);
3633 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3635 mutex_unlock(&ctx->wq->mutex);
3638 static void apply_wqattrs_lock(void)
3640 /* CPUs should stay stable across pwq creations and installations */
3642 mutex_lock(&wq_pool_mutex);
3645 static void apply_wqattrs_unlock(void)
3647 mutex_unlock(&wq_pool_mutex);
3651 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3652 const struct workqueue_attrs *attrs)
3654 struct apply_wqattrs_ctx *ctx;
3656 /* only unbound workqueues can change attributes */
3657 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3660 /* creating multiple pwqs breaks ordering guarantee */
3661 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3664 ctx = apply_wqattrs_prepare(wq, attrs);
3668 /* the ctx has been prepared successfully, let's commit it */
3669 apply_wqattrs_commit(ctx);
3670 apply_wqattrs_cleanup(ctx);
3676 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3677 * @wq: the target workqueue
3678 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3680 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3681 * machines, this function maps a separate pwq to each NUMA node with
3682 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3683 * NUMA node it was issued on. Older pwqs are released as in-flight work
3684 * items finish. Note that a work item which repeatedly requeues itself
3685 * back-to-back will stay on its current pwq.
3687 * Performs GFP_KERNEL allocations.
3689 * Return: 0 on success and -errno on failure.
3691 int apply_workqueue_attrs(struct workqueue_struct *wq,
3692 const struct workqueue_attrs *attrs)
3696 apply_wqattrs_lock();
3697 ret = apply_workqueue_attrs_locked(wq, attrs);
3698 apply_wqattrs_unlock();
3704 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3705 * @wq: the target workqueue
3706 * @cpu: the CPU coming up or going down
3707 * @online: whether @cpu is coming up or going down
3709 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3710 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3713 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3714 * falls back to @wq->dfl_pwq which may not be optimal but is always
3717 * Note that when the last allowed CPU of a NUMA node goes offline for a
3718 * workqueue with a cpumask spanning multiple nodes, the workers which were
3719 * already executing the work items for the workqueue will lose their CPU
3720 * affinity and may execute on any CPU. This is similar to how per-cpu
3721 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3722 * affinity, it's the user's responsibility to flush the work item from
3725 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3728 int node = cpu_to_node(cpu);
3729 int cpu_off = online ? -1 : cpu;
3730 struct pool_workqueue *old_pwq = NULL, *pwq;
3731 struct workqueue_attrs *target_attrs;
3734 lockdep_assert_held(&wq_pool_mutex);
3736 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3737 wq->unbound_attrs->no_numa)
3741 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3742 * Let's use a preallocated one. The following buf is protected by
3743 * CPU hotplug exclusion.
3745 target_attrs = wq_update_unbound_numa_attrs_buf;
3746 cpumask = target_attrs->cpumask;
3748 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3749 pwq = unbound_pwq_by_node(wq, node);
3752 * Let's determine what needs to be done. If the target cpumask is
3753 * different from the default pwq's, we need to compare it to @pwq's
3754 * and create a new one if they don't match. If the target cpumask
3755 * equals the default pwq's, the default pwq should be used.
3757 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3758 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3764 /* create a new pwq */
3765 pwq = alloc_unbound_pwq(wq, target_attrs);
3767 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3772 /* Install the new pwq. */
3773 mutex_lock(&wq->mutex);
3774 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3778 mutex_lock(&wq->mutex);
3779 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3780 get_pwq(wq->dfl_pwq);
3781 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3782 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3784 mutex_unlock(&wq->mutex);
3785 put_pwq_unlocked(old_pwq);
3788 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3790 bool highpri = wq->flags & WQ_HIGHPRI;
3793 if (!(wq->flags & WQ_UNBOUND)) {
3794 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3798 for_each_possible_cpu(cpu) {
3799 struct pool_workqueue *pwq =
3800 per_cpu_ptr(wq->cpu_pwqs, cpu);
3801 struct worker_pool *cpu_pools =
3802 per_cpu(cpu_worker_pools, cpu);
3804 init_pwq(pwq, wq, &cpu_pools[highpri]);
3806 mutex_lock(&wq->mutex);
3808 mutex_unlock(&wq->mutex);
3811 } else if (wq->flags & __WQ_ORDERED) {
3812 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3813 /* there should only be single pwq for ordering guarantee */
3814 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3815 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3816 "ordering guarantee broken for workqueue %s\n", wq->name);
3819 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3823 static int wq_clamp_max_active(int max_active, unsigned int flags,
3826 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3828 if (max_active < 1 || max_active > lim)
3829 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3830 max_active, name, 1, lim);
3832 return clamp_val(max_active, 1, lim);
3835 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3838 struct lock_class_key *key,
3839 const char *lock_name, ...)
3841 size_t tbl_size = 0;
3843 struct workqueue_struct *wq;
3844 struct pool_workqueue *pwq;
3846 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3847 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3848 flags |= WQ_UNBOUND;
3850 /* allocate wq and format name */
3851 if (flags & WQ_UNBOUND)
3852 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3854 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3858 if (flags & WQ_UNBOUND) {
3859 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3860 if (!wq->unbound_attrs)
3864 va_start(args, lock_name);
3865 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3868 max_active = max_active ?: WQ_DFL_ACTIVE;
3869 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3873 wq->saved_max_active = max_active;
3874 mutex_init(&wq->mutex);
3875 atomic_set(&wq->nr_pwqs_to_flush, 0);
3876 INIT_LIST_HEAD(&wq->pwqs);
3877 INIT_LIST_HEAD(&wq->flusher_queue);
3878 INIT_LIST_HEAD(&wq->flusher_overflow);
3879 INIT_LIST_HEAD(&wq->maydays);
3881 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3882 INIT_LIST_HEAD(&wq->list);
3884 if (alloc_and_link_pwqs(wq) < 0)
3888 * Workqueues which may be used during memory reclaim should
3889 * have a rescuer to guarantee forward progress.
3891 if (flags & WQ_MEM_RECLAIM) {
3892 struct worker *rescuer;
3894 rescuer = alloc_worker(NUMA_NO_NODE);
3898 rescuer->rescue_wq = wq;
3899 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3901 if (IS_ERR(rescuer->task)) {
3906 wq->rescuer = rescuer;
3907 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3908 wake_up_process(rescuer->task);
3911 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3915 * wq_pool_mutex protects global freeze state and workqueues list.
3916 * Grab it, adjust max_active and add the new @wq to workqueues
3919 mutex_lock(&wq_pool_mutex);
3921 mutex_lock(&wq->mutex);
3922 for_each_pwq(pwq, wq)
3923 pwq_adjust_max_active(pwq);
3924 mutex_unlock(&wq->mutex);
3926 list_add_tail_rcu(&wq->list, &workqueues);
3928 mutex_unlock(&wq_pool_mutex);
3933 free_workqueue_attrs(wq->unbound_attrs);
3937 destroy_workqueue(wq);
3940 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3943 * destroy_workqueue - safely terminate a workqueue
3944 * @wq: target workqueue
3946 * Safely destroy a workqueue. All work currently pending will be done first.
3948 void destroy_workqueue(struct workqueue_struct *wq)
3950 struct pool_workqueue *pwq;
3953 /* drain it before proceeding with destruction */
3954 drain_workqueue(wq);
3957 mutex_lock(&wq->mutex);
3958 for_each_pwq(pwq, wq) {
3961 for (i = 0; i < WORK_NR_COLORS; i++) {
3962 if (WARN_ON(pwq->nr_in_flight[i])) {
3963 mutex_unlock(&wq->mutex);
3968 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3969 WARN_ON(pwq->nr_active) ||
3970 WARN_ON(!list_empty(&pwq->delayed_works))) {
3971 mutex_unlock(&wq->mutex);
3975 mutex_unlock(&wq->mutex);
3978 * wq list is used to freeze wq, remove from list after
3979 * flushing is complete in case freeze races us.
3981 mutex_lock(&wq_pool_mutex);
3982 list_del_rcu(&wq->list);
3983 mutex_unlock(&wq_pool_mutex);
3985 workqueue_sysfs_unregister(wq);
3988 kthread_stop(wq->rescuer->task);
3990 if (!(wq->flags & WQ_UNBOUND)) {
3992 * The base ref is never dropped on per-cpu pwqs. Directly
3993 * schedule RCU free.
3995 call_rcu_sched(&wq->rcu, rcu_free_wq);
3998 * We're the sole accessor of @wq at this point. Directly
3999 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4000 * @wq will be freed when the last pwq is released.
4002 for_each_node(node) {
4003 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4004 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4005 put_pwq_unlocked(pwq);
4009 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4010 * put. Don't access it afterwards.
4014 put_pwq_unlocked(pwq);
4017 EXPORT_SYMBOL_GPL(destroy_workqueue);
4020 * workqueue_set_max_active - adjust max_active of a workqueue
4021 * @wq: target workqueue
4022 * @max_active: new max_active value.
4024 * Set max_active of @wq to @max_active.
4027 * Don't call from IRQ context.
4029 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4031 struct pool_workqueue *pwq;
4033 /* disallow meddling with max_active for ordered workqueues */
4034 if (WARN_ON(wq->flags & __WQ_ORDERED))
4037 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4039 mutex_lock(&wq->mutex);
4041 wq->saved_max_active = max_active;
4043 for_each_pwq(pwq, wq)
4044 pwq_adjust_max_active(pwq);
4046 mutex_unlock(&wq->mutex);
4048 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4051 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4053 * Determine whether %current is a workqueue rescuer. Can be used from
4054 * work functions to determine whether it's being run off the rescuer task.
4056 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4058 bool current_is_workqueue_rescuer(void)
4060 struct worker *worker = current_wq_worker();
4062 return worker && worker->rescue_wq;
4066 * workqueue_congested - test whether a workqueue is congested
4067 * @cpu: CPU in question
4068 * @wq: target workqueue
4070 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4071 * no synchronization around this function and the test result is
4072 * unreliable and only useful as advisory hints or for debugging.
4074 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4075 * Note that both per-cpu and unbound workqueues may be associated with
4076 * multiple pool_workqueues which have separate congested states. A
4077 * workqueue being congested on one CPU doesn't mean the workqueue is also
4078 * contested on other CPUs / NUMA nodes.
4081 * %true if congested, %false otherwise.
4083 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4085 struct pool_workqueue *pwq;
4088 rcu_read_lock_sched();
4090 if (cpu == WORK_CPU_UNBOUND)
4091 cpu = smp_processor_id();
4093 if (!(wq->flags & WQ_UNBOUND))
4094 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4096 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4098 ret = !list_empty(&pwq->delayed_works);
4099 rcu_read_unlock_sched();
4103 EXPORT_SYMBOL_GPL(workqueue_congested);
4106 * work_busy - test whether a work is currently pending or running
4107 * @work: the work to be tested
4109 * Test whether @work is currently pending or running. There is no
4110 * synchronization around this function and the test result is
4111 * unreliable and only useful as advisory hints or for debugging.
4114 * OR'd bitmask of WORK_BUSY_* bits.
4116 unsigned int work_busy(struct work_struct *work)
4118 struct worker_pool *pool;
4119 unsigned long flags;
4120 unsigned int ret = 0;
4122 if (work_pending(work))
4123 ret |= WORK_BUSY_PENDING;
4125 local_irq_save(flags);
4126 pool = get_work_pool(work);
4128 spin_lock(&pool->lock);
4129 if (find_worker_executing_work(pool, work))
4130 ret |= WORK_BUSY_RUNNING;
4131 spin_unlock(&pool->lock);
4133 local_irq_restore(flags);
4137 EXPORT_SYMBOL_GPL(work_busy);
4140 * set_worker_desc - set description for the current work item
4141 * @fmt: printf-style format string
4142 * @...: arguments for the format string
4144 * This function can be called by a running work function to describe what
4145 * the work item is about. If the worker task gets dumped, this
4146 * information will be printed out together to help debugging. The
4147 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4149 void set_worker_desc(const char *fmt, ...)
4151 struct worker *worker = current_wq_worker();
4155 va_start(args, fmt);
4156 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4158 worker->desc_valid = true;
4163 * print_worker_info - print out worker information and description
4164 * @log_lvl: the log level to use when printing
4165 * @task: target task
4167 * If @task is a worker and currently executing a work item, print out the
4168 * name of the workqueue being serviced and worker description set with
4169 * set_worker_desc() by the currently executing work item.
4171 * This function can be safely called on any task as long as the
4172 * task_struct itself is accessible. While safe, this function isn't
4173 * synchronized and may print out mixups or garbages of limited length.
4175 void print_worker_info(const char *log_lvl, struct task_struct *task)
4177 work_func_t *fn = NULL;
4178 char name[WQ_NAME_LEN] = { };
4179 char desc[WORKER_DESC_LEN] = { };
4180 struct pool_workqueue *pwq = NULL;
4181 struct workqueue_struct *wq = NULL;
4182 bool desc_valid = false;
4183 struct worker *worker;
4185 if (!(task->flags & PF_WQ_WORKER))
4189 * This function is called without any synchronization and @task
4190 * could be in any state. Be careful with dereferences.
4192 worker = probe_kthread_data(task);
4195 * Carefully copy the associated workqueue's workfn and name. Keep
4196 * the original last '\0' in case the original contains garbage.
4198 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4199 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4200 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4201 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4203 /* copy worker description */
4204 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4206 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4208 if (fn || name[0] || desc[0]) {
4209 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4211 pr_cont(" (%s)", desc);
4216 static void pr_cont_pool_info(struct worker_pool *pool)
4218 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4219 if (pool->node != NUMA_NO_NODE)
4220 pr_cont(" node=%d", pool->node);
4221 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4224 static void pr_cont_work(bool comma, struct work_struct *work)
4226 if (work->func == wq_barrier_func) {
4227 struct wq_barrier *barr;
4229 barr = container_of(work, struct wq_barrier, work);
4231 pr_cont("%s BAR(%d)", comma ? "," : "",
4232 task_pid_nr(barr->task));
4234 pr_cont("%s %pf", comma ? "," : "", work->func);
4238 static void show_pwq(struct pool_workqueue *pwq)
4240 struct worker_pool *pool = pwq->pool;
4241 struct work_struct *work;
4242 struct worker *worker;
4243 bool has_in_flight = false, has_pending = false;
4246 pr_info(" pwq %d:", pool->id);
4247 pr_cont_pool_info(pool);
4249 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4250 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4252 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4253 if (worker->current_pwq == pwq) {
4254 has_in_flight = true;
4258 if (has_in_flight) {
4261 pr_info(" in-flight:");
4262 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4263 if (worker->current_pwq != pwq)
4266 pr_cont("%s %d%s:%pf", comma ? "," : "",
4267 task_pid_nr(worker->task),
4268 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4269 worker->current_func);
4270 list_for_each_entry(work, &worker->scheduled, entry)
4271 pr_cont_work(false, work);
4277 list_for_each_entry(work, &pool->worklist, entry) {
4278 if (get_work_pwq(work) == pwq) {
4286 pr_info(" pending:");
4287 list_for_each_entry(work, &pool->worklist, entry) {
4288 if (get_work_pwq(work) != pwq)
4291 pr_cont_work(comma, work);
4292 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4297 if (!list_empty(&pwq->delayed_works)) {
4300 pr_info(" delayed:");
4301 list_for_each_entry(work, &pwq->delayed_works, entry) {
4302 pr_cont_work(comma, work);
4303 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4310 * show_workqueue_state - dump workqueue state
4312 * Called from a sysrq handler and prints out all busy workqueues and
4315 void show_workqueue_state(void)
4317 struct workqueue_struct *wq;
4318 struct worker_pool *pool;
4319 unsigned long flags;
4322 rcu_read_lock_sched();
4324 pr_info("Showing busy workqueues and worker pools:\n");
4326 list_for_each_entry_rcu(wq, &workqueues, list) {
4327 struct pool_workqueue *pwq;
4330 for_each_pwq(pwq, wq) {
4331 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4339 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4341 for_each_pwq(pwq, wq) {
4342 spin_lock_irqsave(&pwq->pool->lock, flags);
4343 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4345 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4349 for_each_pool(pool, pi) {
4350 struct worker *worker;
4353 spin_lock_irqsave(&pool->lock, flags);
4354 if (pool->nr_workers == pool->nr_idle)
4357 pr_info("pool %d:", pool->id);
4358 pr_cont_pool_info(pool);
4359 pr_cont(" hung=%us workers=%d",
4360 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4363 pr_cont(" manager: %d",
4364 task_pid_nr(pool->manager->task));
4365 list_for_each_entry(worker, &pool->idle_list, entry) {
4366 pr_cont(" %s%d", first ? "idle: " : "",
4367 task_pid_nr(worker->task));
4372 spin_unlock_irqrestore(&pool->lock, flags);
4375 rcu_read_unlock_sched();
4381 * There are two challenges in supporting CPU hotplug. Firstly, there
4382 * are a lot of assumptions on strong associations among work, pwq and
4383 * pool which make migrating pending and scheduled works very
4384 * difficult to implement without impacting hot paths. Secondly,
4385 * worker pools serve mix of short, long and very long running works making
4386 * blocked draining impractical.
4388 * This is solved by allowing the pools to be disassociated from the CPU
4389 * running as an unbound one and allowing it to be reattached later if the
4390 * cpu comes back online.
4393 static void wq_unbind_fn(struct work_struct *work)
4395 int cpu = smp_processor_id();
4396 struct worker_pool *pool;
4397 struct worker *worker;
4399 for_each_cpu_worker_pool(pool, cpu) {
4400 mutex_lock(&pool->attach_mutex);
4401 spin_lock_irq(&pool->lock);
4404 * We've blocked all attach/detach operations. Make all workers
4405 * unbound and set DISASSOCIATED. Before this, all workers
4406 * except for the ones which are still executing works from
4407 * before the last CPU down must be on the cpu. After
4408 * this, they may become diasporas.
4410 for_each_pool_worker(worker, pool)
4411 worker->flags |= WORKER_UNBOUND;
4413 pool->flags |= POOL_DISASSOCIATED;
4415 spin_unlock_irq(&pool->lock);
4416 mutex_unlock(&pool->attach_mutex);
4419 * Call schedule() so that we cross rq->lock and thus can
4420 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4421 * This is necessary as scheduler callbacks may be invoked
4427 * Sched callbacks are disabled now. Zap nr_running.
4428 * After this, nr_running stays zero and need_more_worker()
4429 * and keep_working() are always true as long as the
4430 * worklist is not empty. This pool now behaves as an
4431 * unbound (in terms of concurrency management) pool which
4432 * are served by workers tied to the pool.
4434 atomic_set(&pool->nr_running, 0);
4437 * With concurrency management just turned off, a busy
4438 * worker blocking could lead to lengthy stalls. Kick off
4439 * unbound chain execution of currently pending work items.
4441 spin_lock_irq(&pool->lock);
4442 wake_up_worker(pool);
4443 spin_unlock_irq(&pool->lock);
4448 * rebind_workers - rebind all workers of a pool to the associated CPU
4449 * @pool: pool of interest
4451 * @pool->cpu is coming online. Rebind all workers to the CPU.
4453 static void rebind_workers(struct worker_pool *pool)
4455 struct worker *worker;
4457 lockdep_assert_held(&pool->attach_mutex);
4460 * Restore CPU affinity of all workers. As all idle workers should
4461 * be on the run-queue of the associated CPU before any local
4462 * wake-ups for concurrency management happen, restore CPU affinity
4463 * of all workers first and then clear UNBOUND. As we're called
4464 * from CPU_ONLINE, the following shouldn't fail.
4466 for_each_pool_worker(worker, pool)
4467 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4468 pool->attrs->cpumask) < 0);
4470 spin_lock_irq(&pool->lock);
4471 pool->flags &= ~POOL_DISASSOCIATED;
4473 for_each_pool_worker(worker, pool) {
4474 unsigned int worker_flags = worker->flags;
4477 * A bound idle worker should actually be on the runqueue
4478 * of the associated CPU for local wake-ups targeting it to
4479 * work. Kick all idle workers so that they migrate to the
4480 * associated CPU. Doing this in the same loop as
4481 * replacing UNBOUND with REBOUND is safe as no worker will
4482 * be bound before @pool->lock is released.
4484 if (worker_flags & WORKER_IDLE)
4485 wake_up_process(worker->task);
4488 * We want to clear UNBOUND but can't directly call
4489 * worker_clr_flags() or adjust nr_running. Atomically
4490 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4491 * @worker will clear REBOUND using worker_clr_flags() when
4492 * it initiates the next execution cycle thus restoring
4493 * concurrency management. Note that when or whether
4494 * @worker clears REBOUND doesn't affect correctness.
4496 * ACCESS_ONCE() is necessary because @worker->flags may be
4497 * tested without holding any lock in
4498 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4499 * fail incorrectly leading to premature concurrency
4500 * management operations.
4502 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4503 worker_flags |= WORKER_REBOUND;
4504 worker_flags &= ~WORKER_UNBOUND;
4505 ACCESS_ONCE(worker->flags) = worker_flags;
4508 spin_unlock_irq(&pool->lock);
4512 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4513 * @pool: unbound pool of interest
4514 * @cpu: the CPU which is coming up
4516 * An unbound pool may end up with a cpumask which doesn't have any online
4517 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4518 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4519 * online CPU before, cpus_allowed of all its workers should be restored.
4521 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4523 static cpumask_t cpumask;
4524 struct worker *worker;
4526 lockdep_assert_held(&pool->attach_mutex);
4528 /* is @cpu allowed for @pool? */
4529 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4532 /* is @cpu the only online CPU? */
4533 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4534 if (cpumask_weight(&cpumask) != 1)
4537 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4538 for_each_pool_worker(worker, pool)
4539 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4540 pool->attrs->cpumask) < 0);
4544 * Workqueues should be brought up before normal priority CPU notifiers.
4545 * This will be registered high priority CPU notifier.
4547 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4548 unsigned long action,
4551 int cpu = (unsigned long)hcpu;
4552 struct worker_pool *pool;
4553 struct workqueue_struct *wq;
4556 switch (action & ~CPU_TASKS_FROZEN) {
4557 case CPU_UP_PREPARE:
4558 for_each_cpu_worker_pool(pool, cpu) {
4559 if (pool->nr_workers)
4561 if (!create_worker(pool))
4566 case CPU_DOWN_FAILED:
4568 mutex_lock(&wq_pool_mutex);
4570 for_each_pool(pool, pi) {
4571 mutex_lock(&pool->attach_mutex);
4573 if (pool->cpu == cpu)
4574 rebind_workers(pool);
4575 else if (pool->cpu < 0)
4576 restore_unbound_workers_cpumask(pool, cpu);
4578 mutex_unlock(&pool->attach_mutex);
4581 /* update NUMA affinity of unbound workqueues */
4582 list_for_each_entry(wq, &workqueues, list)
4583 wq_update_unbound_numa(wq, cpu, true);
4585 mutex_unlock(&wq_pool_mutex);
4592 * Workqueues should be brought down after normal priority CPU notifiers.
4593 * This will be registered as low priority CPU notifier.
4595 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4596 unsigned long action,
4599 int cpu = (unsigned long)hcpu;
4600 struct work_struct unbind_work;
4601 struct workqueue_struct *wq;
4603 switch (action & ~CPU_TASKS_FROZEN) {
4604 case CPU_DOWN_PREPARE:
4605 /* unbinding per-cpu workers should happen on the local CPU */
4606 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4607 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4609 /* update NUMA affinity of unbound workqueues */
4610 mutex_lock(&wq_pool_mutex);
4611 list_for_each_entry(wq, &workqueues, list)
4612 wq_update_unbound_numa(wq, cpu, false);
4613 mutex_unlock(&wq_pool_mutex);
4615 /* wait for per-cpu unbinding to finish */
4616 flush_work(&unbind_work);
4617 destroy_work_on_stack(&unbind_work);
4625 struct work_for_cpu {
4626 struct work_struct work;
4632 static void work_for_cpu_fn(struct work_struct *work)
4634 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4636 wfc->ret = wfc->fn(wfc->arg);
4640 * work_on_cpu - run a function in user context on a particular cpu
4641 * @cpu: the cpu to run on
4642 * @fn: the function to run
4643 * @arg: the function arg
4645 * It is up to the caller to ensure that the cpu doesn't go offline.
4646 * The caller must not hold any locks which would prevent @fn from completing.
4648 * Return: The value @fn returns.
4650 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4652 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4654 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4655 schedule_work_on(cpu, &wfc.work);
4656 flush_work(&wfc.work);
4657 destroy_work_on_stack(&wfc.work);
4660 EXPORT_SYMBOL_GPL(work_on_cpu);
4661 #endif /* CONFIG_SMP */
4663 #ifdef CONFIG_FREEZER
4666 * freeze_workqueues_begin - begin freezing workqueues
4668 * Start freezing workqueues. After this function returns, all freezable
4669 * workqueues will queue new works to their delayed_works list instead of
4673 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4675 void freeze_workqueues_begin(void)
4677 struct workqueue_struct *wq;
4678 struct pool_workqueue *pwq;
4680 mutex_lock(&wq_pool_mutex);
4682 WARN_ON_ONCE(workqueue_freezing);
4683 workqueue_freezing = true;
4685 list_for_each_entry(wq, &workqueues, list) {
4686 mutex_lock(&wq->mutex);
4687 for_each_pwq(pwq, wq)
4688 pwq_adjust_max_active(pwq);
4689 mutex_unlock(&wq->mutex);
4692 mutex_unlock(&wq_pool_mutex);
4696 * freeze_workqueues_busy - are freezable workqueues still busy?
4698 * Check whether freezing is complete. This function must be called
4699 * between freeze_workqueues_begin() and thaw_workqueues().
4702 * Grabs and releases wq_pool_mutex.
4705 * %true if some freezable workqueues are still busy. %false if freezing
4708 bool freeze_workqueues_busy(void)
4711 struct workqueue_struct *wq;
4712 struct pool_workqueue *pwq;
4714 mutex_lock(&wq_pool_mutex);
4716 WARN_ON_ONCE(!workqueue_freezing);
4718 list_for_each_entry(wq, &workqueues, list) {
4719 if (!(wq->flags & WQ_FREEZABLE))
4722 * nr_active is monotonically decreasing. It's safe
4723 * to peek without lock.
4725 rcu_read_lock_sched();
4726 for_each_pwq(pwq, wq) {
4727 WARN_ON_ONCE(pwq->nr_active < 0);
4728 if (pwq->nr_active) {
4730 rcu_read_unlock_sched();
4734 rcu_read_unlock_sched();
4737 mutex_unlock(&wq_pool_mutex);
4742 * thaw_workqueues - thaw workqueues
4744 * Thaw workqueues. Normal queueing is restored and all collected
4745 * frozen works are transferred to their respective pool worklists.
4748 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4750 void thaw_workqueues(void)
4752 struct workqueue_struct *wq;
4753 struct pool_workqueue *pwq;
4755 mutex_lock(&wq_pool_mutex);
4757 if (!workqueue_freezing)
4760 workqueue_freezing = false;
4762 /* restore max_active and repopulate worklist */
4763 list_for_each_entry(wq, &workqueues, list) {
4764 mutex_lock(&wq->mutex);
4765 for_each_pwq(pwq, wq)
4766 pwq_adjust_max_active(pwq);
4767 mutex_unlock(&wq->mutex);
4771 mutex_unlock(&wq_pool_mutex);
4773 #endif /* CONFIG_FREEZER */
4775 static int workqueue_apply_unbound_cpumask(void)
4779 struct workqueue_struct *wq;
4780 struct apply_wqattrs_ctx *ctx, *n;
4782 lockdep_assert_held(&wq_pool_mutex);
4784 list_for_each_entry(wq, &workqueues, list) {
4785 if (!(wq->flags & WQ_UNBOUND))
4787 /* creating multiple pwqs breaks ordering guarantee */
4788 if (wq->flags & __WQ_ORDERED)
4791 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4797 list_add_tail(&ctx->list, &ctxs);
4800 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4802 apply_wqattrs_commit(ctx);
4803 apply_wqattrs_cleanup(ctx);
4810 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4811 * @cpumask: the cpumask to set
4813 * The low-level workqueues cpumask is a global cpumask that limits
4814 * the affinity of all unbound workqueues. This function check the @cpumask
4815 * and apply it to all unbound workqueues and updates all pwqs of them.
4817 * Retun: 0 - Success
4818 * -EINVAL - Invalid @cpumask
4819 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4821 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4824 cpumask_var_t saved_cpumask;
4826 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4829 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4830 if (!cpumask_empty(cpumask)) {
4831 apply_wqattrs_lock();
4833 /* save the old wq_unbound_cpumask. */
4834 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4836 /* update wq_unbound_cpumask at first and apply it to wqs. */
4837 cpumask_copy(wq_unbound_cpumask, cpumask);
4838 ret = workqueue_apply_unbound_cpumask();
4840 /* restore the wq_unbound_cpumask when failed. */
4842 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4844 apply_wqattrs_unlock();
4847 free_cpumask_var(saved_cpumask);
4853 * Workqueues with WQ_SYSFS flag set is visible to userland via
4854 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4855 * following attributes.
4857 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4858 * max_active RW int : maximum number of in-flight work items
4860 * Unbound workqueues have the following extra attributes.
4862 * id RO int : the associated pool ID
4863 * nice RW int : nice value of the workers
4864 * cpumask RW mask : bitmask of allowed CPUs for the workers
4867 struct workqueue_struct *wq;
4871 static struct workqueue_struct *dev_to_wq(struct device *dev)
4873 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4878 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4881 struct workqueue_struct *wq = dev_to_wq(dev);
4883 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4885 static DEVICE_ATTR_RO(per_cpu);
4887 static ssize_t max_active_show(struct device *dev,
4888 struct device_attribute *attr, char *buf)
4890 struct workqueue_struct *wq = dev_to_wq(dev);
4892 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4895 static ssize_t max_active_store(struct device *dev,
4896 struct device_attribute *attr, const char *buf,
4899 struct workqueue_struct *wq = dev_to_wq(dev);
4902 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4905 workqueue_set_max_active(wq, val);
4908 static DEVICE_ATTR_RW(max_active);
4910 static struct attribute *wq_sysfs_attrs[] = {
4911 &dev_attr_per_cpu.attr,
4912 &dev_attr_max_active.attr,
4915 ATTRIBUTE_GROUPS(wq_sysfs);
4917 static ssize_t wq_pool_ids_show(struct device *dev,
4918 struct device_attribute *attr, char *buf)
4920 struct workqueue_struct *wq = dev_to_wq(dev);
4921 const char *delim = "";
4922 int node, written = 0;
4924 rcu_read_lock_sched();
4925 for_each_node(node) {
4926 written += scnprintf(buf + written, PAGE_SIZE - written,
4927 "%s%d:%d", delim, node,
4928 unbound_pwq_by_node(wq, node)->pool->id);
4931 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4932 rcu_read_unlock_sched();
4937 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4940 struct workqueue_struct *wq = dev_to_wq(dev);
4943 mutex_lock(&wq->mutex);
4944 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4945 mutex_unlock(&wq->mutex);
4950 /* prepare workqueue_attrs for sysfs store operations */
4951 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4953 struct workqueue_attrs *attrs;
4955 lockdep_assert_held(&wq_pool_mutex);
4957 attrs = alloc_workqueue_attrs(GFP_KERNEL);
4961 copy_workqueue_attrs(attrs, wq->unbound_attrs);
4965 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4966 const char *buf, size_t count)
4968 struct workqueue_struct *wq = dev_to_wq(dev);
4969 struct workqueue_attrs *attrs;
4972 apply_wqattrs_lock();
4974 attrs = wq_sysfs_prep_attrs(wq);
4978 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4979 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4980 ret = apply_workqueue_attrs_locked(wq, attrs);
4985 apply_wqattrs_unlock();
4986 free_workqueue_attrs(attrs);
4987 return ret ?: count;
4990 static ssize_t wq_cpumask_show(struct device *dev,
4991 struct device_attribute *attr, char *buf)
4993 struct workqueue_struct *wq = dev_to_wq(dev);
4996 mutex_lock(&wq->mutex);
4997 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4998 cpumask_pr_args(wq->unbound_attrs->cpumask));
4999 mutex_unlock(&wq->mutex);
5003 static ssize_t wq_cpumask_store(struct device *dev,
5004 struct device_attribute *attr,
5005 const char *buf, size_t count)
5007 struct workqueue_struct *wq = dev_to_wq(dev);
5008 struct workqueue_attrs *attrs;
5011 apply_wqattrs_lock();
5013 attrs = wq_sysfs_prep_attrs(wq);
5017 ret = cpumask_parse(buf, attrs->cpumask);
5019 ret = apply_workqueue_attrs_locked(wq, attrs);
5022 apply_wqattrs_unlock();
5023 free_workqueue_attrs(attrs);
5024 return ret ?: count;
5027 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5030 struct workqueue_struct *wq = dev_to_wq(dev);
5033 mutex_lock(&wq->mutex);
5034 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5035 !wq->unbound_attrs->no_numa);
5036 mutex_unlock(&wq->mutex);
5041 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5042 const char *buf, size_t count)
5044 struct workqueue_struct *wq = dev_to_wq(dev);
5045 struct workqueue_attrs *attrs;
5046 int v, ret = -ENOMEM;
5048 apply_wqattrs_lock();
5050 attrs = wq_sysfs_prep_attrs(wq);
5055 if (sscanf(buf, "%d", &v) == 1) {
5056 attrs->no_numa = !v;
5057 ret = apply_workqueue_attrs_locked(wq, attrs);
5061 apply_wqattrs_unlock();
5062 free_workqueue_attrs(attrs);
5063 return ret ?: count;
5066 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5067 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5068 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5069 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5070 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5074 static struct bus_type wq_subsys = {
5075 .name = "workqueue",
5076 .dev_groups = wq_sysfs_groups,
5079 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5080 struct device_attribute *attr, char *buf)
5084 mutex_lock(&wq_pool_mutex);
5085 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5086 cpumask_pr_args(wq_unbound_cpumask));
5087 mutex_unlock(&wq_pool_mutex);
5092 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5093 struct device_attribute *attr, const char *buf, size_t count)
5095 cpumask_var_t cpumask;
5098 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5101 ret = cpumask_parse(buf, cpumask);
5103 ret = workqueue_set_unbound_cpumask(cpumask);
5105 free_cpumask_var(cpumask);
5106 return ret ? ret : count;
5109 static struct device_attribute wq_sysfs_cpumask_attr =
5110 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5111 wq_unbound_cpumask_store);
5113 static int __init wq_sysfs_init(void)
5117 err = subsys_virtual_register(&wq_subsys, NULL);
5121 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5123 core_initcall(wq_sysfs_init);
5125 static void wq_device_release(struct device *dev)
5127 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5133 * workqueue_sysfs_register - make a workqueue visible in sysfs
5134 * @wq: the workqueue to register
5136 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5137 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5138 * which is the preferred method.
5140 * Workqueue user should use this function directly iff it wants to apply
5141 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5142 * apply_workqueue_attrs() may race against userland updating the
5145 * Return: 0 on success, -errno on failure.
5147 int workqueue_sysfs_register(struct workqueue_struct *wq)
5149 struct wq_device *wq_dev;
5153 * Adjusting max_active or creating new pwqs by applying
5154 * attributes breaks ordering guarantee. Disallow exposing ordered
5157 if (WARN_ON(wq->flags & __WQ_ORDERED))
5160 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5165 wq_dev->dev.bus = &wq_subsys;
5166 wq_dev->dev.init_name = wq->name;
5167 wq_dev->dev.release = wq_device_release;
5170 * unbound_attrs are created separately. Suppress uevent until
5171 * everything is ready.
5173 dev_set_uevent_suppress(&wq_dev->dev, true);
5175 ret = device_register(&wq_dev->dev);
5182 if (wq->flags & WQ_UNBOUND) {
5183 struct device_attribute *attr;
5185 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5186 ret = device_create_file(&wq_dev->dev, attr);
5188 device_unregister(&wq_dev->dev);
5195 dev_set_uevent_suppress(&wq_dev->dev, false);
5196 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5201 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5202 * @wq: the workqueue to unregister
5204 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5206 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5208 struct wq_device *wq_dev = wq->wq_dev;
5214 device_unregister(&wq_dev->dev);
5216 #else /* CONFIG_SYSFS */
5217 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5218 #endif /* CONFIG_SYSFS */
5221 * Workqueue watchdog.
5223 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5224 * flush dependency, a concurrency managed work item which stays RUNNING
5225 * indefinitely. Workqueue stalls can be very difficult to debug as the
5226 * usual warning mechanisms don't trigger and internal workqueue state is
5229 * Workqueue watchdog monitors all worker pools periodically and dumps
5230 * state if some pools failed to make forward progress for a while where
5231 * forward progress is defined as the first item on ->worklist changing.
5233 * This mechanism is controlled through the kernel parameter
5234 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5235 * corresponding sysfs parameter file.
5237 #ifdef CONFIG_WQ_WATCHDOG
5239 static void wq_watchdog_timer_fn(unsigned long data);
5241 static unsigned long wq_watchdog_thresh = 30;
5242 static struct timer_list wq_watchdog_timer =
5243 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5245 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5246 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5248 static void wq_watchdog_reset_touched(void)
5252 wq_watchdog_touched = jiffies;
5253 for_each_possible_cpu(cpu)
5254 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5257 static void wq_watchdog_timer_fn(unsigned long data)
5259 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5260 bool lockup_detected = false;
5261 struct worker_pool *pool;
5269 for_each_pool(pool, pi) {
5270 unsigned long pool_ts, touched, ts;
5272 if (list_empty(&pool->worklist))
5275 /* get the latest of pool and touched timestamps */
5276 pool_ts = READ_ONCE(pool->watchdog_ts);
5277 touched = READ_ONCE(wq_watchdog_touched);
5279 if (time_after(pool_ts, touched))
5284 if (pool->cpu >= 0) {
5285 unsigned long cpu_touched =
5286 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5288 if (time_after(cpu_touched, ts))
5293 if (time_after(jiffies, ts + thresh)) {
5294 lockup_detected = true;
5295 pr_emerg("BUG: workqueue lockup - pool");
5296 pr_cont_pool_info(pool);
5297 pr_cont(" stuck for %us!\n",
5298 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5304 if (lockup_detected)
5305 show_workqueue_state();
5307 wq_watchdog_reset_touched();
5308 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5311 void wq_watchdog_touch(int cpu)
5314 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5316 wq_watchdog_touched = jiffies;
5319 static void wq_watchdog_set_thresh(unsigned long thresh)
5321 wq_watchdog_thresh = 0;
5322 del_timer_sync(&wq_watchdog_timer);
5325 wq_watchdog_thresh = thresh;
5326 wq_watchdog_reset_touched();
5327 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5331 static int wq_watchdog_param_set_thresh(const char *val,
5332 const struct kernel_param *kp)
5334 unsigned long thresh;
5337 ret = kstrtoul(val, 0, &thresh);
5342 wq_watchdog_set_thresh(thresh);
5344 wq_watchdog_thresh = thresh;
5349 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5350 .set = wq_watchdog_param_set_thresh,
5351 .get = param_get_ulong,
5354 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5357 static void wq_watchdog_init(void)
5359 wq_watchdog_set_thresh(wq_watchdog_thresh);
5362 #else /* CONFIG_WQ_WATCHDOG */
5364 static inline void wq_watchdog_init(void) { }
5366 #endif /* CONFIG_WQ_WATCHDOG */
5368 static void __init wq_numa_init(void)
5373 if (num_possible_nodes() <= 1)
5376 if (wq_disable_numa) {
5377 pr_info("workqueue: NUMA affinity support disabled\n");
5381 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5382 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5385 * We want masks of possible CPUs of each node which isn't readily
5386 * available. Build one from cpu_to_node() which should have been
5387 * fully initialized by now.
5389 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5393 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5394 node_online(node) ? node : NUMA_NO_NODE));
5396 for_each_possible_cpu(cpu) {
5397 node = cpu_to_node(cpu);
5398 if (WARN_ON(node == NUMA_NO_NODE)) {
5399 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5400 /* happens iff arch is bonkers, let's just proceed */
5403 cpumask_set_cpu(cpu, tbl[node]);
5406 wq_numa_possible_cpumask = tbl;
5407 wq_numa_enabled = true;
5410 static int __init init_workqueues(void)
5412 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5415 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5417 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5418 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5420 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5422 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5423 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5427 /* initialize CPU pools */
5428 for_each_possible_cpu(cpu) {
5429 struct worker_pool *pool;
5432 for_each_cpu_worker_pool(pool, cpu) {
5433 BUG_ON(init_worker_pool(pool));
5435 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5436 pool->attrs->nice = std_nice[i++];
5437 pool->node = cpu_to_node(cpu);
5440 mutex_lock(&wq_pool_mutex);
5441 BUG_ON(worker_pool_assign_id(pool));
5442 mutex_unlock(&wq_pool_mutex);
5446 /* create the initial worker */
5447 for_each_online_cpu(cpu) {
5448 struct worker_pool *pool;
5450 for_each_cpu_worker_pool(pool, cpu) {
5451 pool->flags &= ~POOL_DISASSOCIATED;
5452 BUG_ON(!create_worker(pool));
5456 /* create default unbound and ordered wq attrs */
5457 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5458 struct workqueue_attrs *attrs;
5460 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5461 attrs->nice = std_nice[i];
5462 unbound_std_wq_attrs[i] = attrs;
5465 * An ordered wq should have only one pwq as ordering is
5466 * guaranteed by max_active which is enforced by pwqs.
5467 * Turn off NUMA so that dfl_pwq is used for all nodes.
5469 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5470 attrs->nice = std_nice[i];
5471 attrs->no_numa = true;
5472 ordered_wq_attrs[i] = attrs;
5475 system_wq = alloc_workqueue("events", 0, 0);
5476 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5477 system_long_wq = alloc_workqueue("events_long", 0, 0);
5478 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5479 WQ_UNBOUND_MAX_ACTIVE);
5480 system_freezable_wq = alloc_workqueue("events_freezable",
5482 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5483 WQ_POWER_EFFICIENT, 0);
5484 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5485 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5487 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5488 !system_unbound_wq || !system_freezable_wq ||
5489 !system_power_efficient_wq ||
5490 !system_freezable_power_efficient_wq);
5496 early_initcall(init_workqueues);