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 */
72 POOL_FREEZING = 1 << 3, /* freeze in progress */
75 WORKER_DIE = 1 << 1, /* die die die */
76 WORKER_IDLE = 1 << 2, /* is idle */
77 WORKER_PREP = 1 << 3, /* preparing to run works */
78 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
79 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
80 WORKER_REBOUND = 1 << 8, /* worker was rebound */
82 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
83 WORKER_UNBOUND | WORKER_REBOUND,
85 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
87 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
88 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
90 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
91 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
93 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
94 /* call for help after 10ms
96 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
97 CREATE_COOLDOWN = HZ, /* time to breath after fail */
100 * Rescue workers are used only on emergencies and shared by
101 * all cpus. Give -20.
103 RESCUER_NICE_LEVEL = -20,
104 HIGHPRI_NICE_LEVEL = -20,
110 * Structure fields follow one of the following exclusion rules.
112 * I: Modifiable by initialization/destruction paths and read-only for
115 * P: Preemption protected. Disabling preemption is enough and should
116 * only be modified and accessed from the local cpu.
118 * L: pool->lock protected. Access with pool->lock held.
120 * X: During normal operation, modification requires pool->lock and should
121 * be done only from local cpu. Either disabling preemption on local
122 * cpu or grabbing pool->lock is enough for read access. If
123 * POOL_DISASSOCIATED is set, it's identical to L.
125 * A: pool->attach_mutex protected.
127 * PL: wq_pool_mutex protected.
129 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131 * WQ: wq->mutex protected.
133 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
135 * MD: wq_mayday_lock protected.
138 /* struct worker is defined in workqueue_internal.h */
141 spinlock_t lock; /* the pool lock */
142 int cpu; /* I: the associated cpu */
143 int node; /* I: the associated node ID */
144 int id; /* I: pool ID */
145 unsigned int flags; /* X: flags */
147 struct list_head worklist; /* L: list of pending works */
148 int nr_workers; /* L: total number of workers */
150 /* nr_idle includes the ones off idle_list for rebinding */
151 int nr_idle; /* L: currently idle ones */
153 struct list_head idle_list; /* X: list of idle workers */
154 struct timer_list idle_timer; /* L: worker idle timeout */
155 struct timer_list mayday_timer; /* L: SOS timer for workers */
157 /* a workers is either on busy_hash or idle_list, or the manager */
158 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
159 /* L: hash of busy workers */
161 /* see manage_workers() for details on the two manager mutexes */
162 struct mutex manager_arb; /* manager arbitration */
163 struct mutex attach_mutex; /* attach/detach exclusion */
164 struct list_head workers; /* A: attached workers */
165 struct completion *detach_completion; /* all workers detached */
167 struct ida worker_ida; /* worker IDs for task name */
169 struct workqueue_attrs *attrs; /* I: worker attributes */
170 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
171 int refcnt; /* PL: refcnt for unbound pools */
174 * The current concurrency level. As it's likely to be accessed
175 * from other CPUs during try_to_wake_up(), put it in a separate
178 atomic_t nr_running ____cacheline_aligned_in_smp;
181 * Destruction of pool is sched-RCU protected to allow dereferences
182 * from get_work_pool().
185 } ____cacheline_aligned_in_smp;
188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
193 struct pool_workqueue {
194 struct worker_pool *pool; /* I: the associated pool */
195 struct workqueue_struct *wq; /* I: the owning workqueue */
196 int work_color; /* L: current color */
197 int flush_color; /* L: flushing color */
198 int refcnt; /* L: reference count */
199 int nr_in_flight[WORK_NR_COLORS];
200 /* L: nr of in_flight works */
201 int nr_active; /* L: nr of active works */
202 int max_active; /* L: max active works */
203 struct list_head delayed_works; /* L: delayed works */
204 struct list_head pwqs_node; /* WR: node on wq->pwqs */
205 struct list_head mayday_node; /* MD: node on wq->maydays */
208 * Release of unbound pwq is punted to system_wq. See put_pwq()
209 * and pwq_unbound_release_workfn() for details. pool_workqueue
210 * itself is also sched-RCU protected so that the first pwq can be
211 * determined without grabbing wq->mutex.
213 struct work_struct unbound_release_work;
215 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
218 * Structure used to wait for workqueue flush.
221 struct list_head list; /* WQ: list of flushers */
222 int flush_color; /* WQ: flush color waiting for */
223 struct completion done; /* flush completion */
229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
232 struct workqueue_struct {
233 struct list_head pwqs; /* WR: all pwqs of this wq */
234 struct list_head list; /* PL: list of all workqueues */
236 struct mutex mutex; /* protects this wq */
237 int work_color; /* WQ: current work color */
238 int flush_color; /* WQ: current flush color */
239 atomic_t nr_pwqs_to_flush; /* flush in progress */
240 struct wq_flusher *first_flusher; /* WQ: first flusher */
241 struct list_head flusher_queue; /* WQ: flush waiters */
242 struct list_head flusher_overflow; /* WQ: flush overflow list */
244 struct list_head maydays; /* MD: pwqs requesting rescue */
245 struct worker *rescuer; /* I: rescue worker */
247 int nr_drainers; /* WQ: drain in progress */
248 int saved_max_active; /* WQ: saved pwq max_active */
250 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
251 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
254 struct wq_device *wq_dev; /* I: for sysfs interface */
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map;
259 char name[WQ_NAME_LEN]; /* I: workqueue name */
261 /* hot fields used during command issue, aligned to cacheline */
262 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
263 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
264 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
267 static struct kmem_cache *pwq_cache;
269 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
270 static cpumask_var_t *wq_numa_possible_cpumask;
271 /* possible CPUs of each node */
273 static bool wq_disable_numa;
274 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
276 /* see the comment above the definition of WQ_POWER_EFFICIENT */
277 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
278 static bool wq_power_efficient = true;
280 static bool wq_power_efficient;
283 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
285 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
287 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
288 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
290 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
291 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
293 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
294 static bool workqueue_freezing; /* PL: have wqs started freezing? */
296 /* the per-cpu worker pools */
297 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
300 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
302 /* PL: hash of all unbound pools keyed by pool->attrs */
303 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
305 /* I: attributes used when instantiating standard unbound pools on demand */
306 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
308 /* I: attributes used when instantiating ordered pools on demand */
309 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
311 struct workqueue_struct *system_wq __read_mostly;
312 EXPORT_SYMBOL(system_wq);
313 struct workqueue_struct *system_highpri_wq __read_mostly;
314 EXPORT_SYMBOL_GPL(system_highpri_wq);
315 struct workqueue_struct *system_long_wq __read_mostly;
316 EXPORT_SYMBOL_GPL(system_long_wq);
317 struct workqueue_struct *system_unbound_wq __read_mostly;
318 EXPORT_SYMBOL_GPL(system_unbound_wq);
319 struct workqueue_struct *system_freezable_wq __read_mostly;
320 EXPORT_SYMBOL_GPL(system_freezable_wq);
321 struct workqueue_struct *system_power_efficient_wq __read_mostly;
322 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
323 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
326 static int worker_thread(void *__worker);
327 static void copy_workqueue_attrs(struct workqueue_attrs *to,
328 const struct workqueue_attrs *from);
330 #define CREATE_TRACE_POINTS
331 #include <trace/events/workqueue.h>
333 #define assert_rcu_or_pool_mutex() \
334 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
335 lockdep_is_held(&wq_pool_mutex), \
336 "sched RCU or wq_pool_mutex should be held")
338 #define assert_rcu_or_wq_mutex(wq) \
339 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
340 lockdep_is_held(&wq->mutex), \
341 "sched RCU or wq->mutex should be held")
343 #define for_each_cpu_worker_pool(pool, cpu) \
344 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
345 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
349 * for_each_pool - iterate through all worker_pools in the system
350 * @pool: iteration cursor
351 * @pi: integer used for iteration
353 * This must be called either with wq_pool_mutex held or sched RCU read
354 * locked. If the pool needs to be used beyond the locking in effect, the
355 * caller is responsible for guaranteeing that the pool stays online.
357 * The if/else clause exists only for the lockdep assertion and can be
360 #define for_each_pool(pool, pi) \
361 idr_for_each_entry(&worker_pool_idr, pool, pi) \
362 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
366 * for_each_pool_worker - iterate through all workers of a worker_pool
367 * @worker: iteration cursor
368 * @pool: worker_pool to iterate workers of
370 * This must be called with @pool->attach_mutex.
372 * The if/else clause exists only for the lockdep assertion and can be
375 #define for_each_pool_worker(worker, pool) \
376 list_for_each_entry((worker), &(pool)->workers, node) \
377 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
381 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
382 * @pwq: iteration cursor
383 * @wq: the target workqueue
385 * This must be called either with wq->mutex held or sched RCU read locked.
386 * If the pwq needs to be used beyond the locking in effect, the caller is
387 * responsible for guaranteeing that the pwq stays online.
389 * The if/else clause exists only for the lockdep assertion and can be
392 #define for_each_pwq(pwq, wq) \
393 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
394 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
397 #ifdef CONFIG_DEBUG_OBJECTS_WORK
399 static struct debug_obj_descr work_debug_descr;
401 static void *work_debug_hint(void *addr)
403 return ((struct work_struct *) addr)->func;
407 * fixup_init is called when:
408 * - an active object is initialized
410 static int work_fixup_init(void *addr, enum debug_obj_state state)
412 struct work_struct *work = addr;
415 case ODEBUG_STATE_ACTIVE:
416 cancel_work_sync(work);
417 debug_object_init(work, &work_debug_descr);
425 * fixup_activate is called when:
426 * - an active object is activated
427 * - an unknown object is activated (might be a statically initialized object)
429 static int work_fixup_activate(void *addr, enum debug_obj_state state)
431 struct work_struct *work = addr;
435 case ODEBUG_STATE_NOTAVAILABLE:
437 * This is not really a fixup. The work struct was
438 * statically initialized. We just make sure that it
439 * is tracked in the object tracker.
441 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
442 debug_object_init(work, &work_debug_descr);
443 debug_object_activate(work, &work_debug_descr);
449 case ODEBUG_STATE_ACTIVE:
458 * fixup_free is called when:
459 * - an active object is freed
461 static int work_fixup_free(void *addr, enum debug_obj_state state)
463 struct work_struct *work = addr;
466 case ODEBUG_STATE_ACTIVE:
467 cancel_work_sync(work);
468 debug_object_free(work, &work_debug_descr);
475 static struct debug_obj_descr work_debug_descr = {
476 .name = "work_struct",
477 .debug_hint = work_debug_hint,
478 .fixup_init = work_fixup_init,
479 .fixup_activate = work_fixup_activate,
480 .fixup_free = work_fixup_free,
483 static inline void debug_work_activate(struct work_struct *work)
485 debug_object_activate(work, &work_debug_descr);
488 static inline void debug_work_deactivate(struct work_struct *work)
490 debug_object_deactivate(work, &work_debug_descr);
493 void __init_work(struct work_struct *work, int onstack)
496 debug_object_init_on_stack(work, &work_debug_descr);
498 debug_object_init(work, &work_debug_descr);
500 EXPORT_SYMBOL_GPL(__init_work);
502 void destroy_work_on_stack(struct work_struct *work)
504 debug_object_free(work, &work_debug_descr);
506 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
508 void destroy_delayed_work_on_stack(struct delayed_work *work)
510 destroy_timer_on_stack(&work->timer);
511 debug_object_free(&work->work, &work_debug_descr);
513 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
516 static inline void debug_work_activate(struct work_struct *work) { }
517 static inline void debug_work_deactivate(struct work_struct *work) { }
521 * worker_pool_assign_id - allocate ID and assing it to @pool
522 * @pool: the pool pointer of interest
524 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
525 * successfully, -errno on failure.
527 static int worker_pool_assign_id(struct worker_pool *pool)
531 lockdep_assert_held(&wq_pool_mutex);
533 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
543 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
544 * @wq: the target workqueue
547 * This must be called either with pwq_lock held or sched RCU read locked.
548 * If the pwq needs to be used beyond the locking in effect, the caller is
549 * responsible for guaranteeing that the pwq stays online.
551 * Return: The unbound pool_workqueue for @node.
553 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
556 assert_rcu_or_wq_mutex(wq);
557 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
560 static unsigned int work_color_to_flags(int color)
562 return color << WORK_STRUCT_COLOR_SHIFT;
565 static int get_work_color(struct work_struct *work)
567 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
568 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
571 static int work_next_color(int color)
573 return (color + 1) % WORK_NR_COLORS;
577 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
578 * contain the pointer to the queued pwq. Once execution starts, the flag
579 * is cleared and the high bits contain OFFQ flags and pool ID.
581 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
582 * and clear_work_data() can be used to set the pwq, pool or clear
583 * work->data. These functions should only be called while the work is
584 * owned - ie. while the PENDING bit is set.
586 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
587 * corresponding to a work. Pool is available once the work has been
588 * queued anywhere after initialization until it is sync canceled. pwq is
589 * available only while the work item is queued.
591 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
592 * canceled. While being canceled, a work item may have its PENDING set
593 * but stay off timer and worklist for arbitrarily long and nobody should
594 * try to steal the PENDING bit.
596 static inline void set_work_data(struct work_struct *work, unsigned long data,
599 WARN_ON_ONCE(!work_pending(work));
600 atomic_long_set(&work->data, data | flags | work_static(work));
603 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
604 unsigned long extra_flags)
606 set_work_data(work, (unsigned long)pwq,
607 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
610 static void set_work_pool_and_keep_pending(struct work_struct *work,
613 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
614 WORK_STRUCT_PENDING);
617 static void set_work_pool_and_clear_pending(struct work_struct *work,
621 * The following wmb is paired with the implied mb in
622 * test_and_set_bit(PENDING) and ensures all updates to @work made
623 * here are visible to and precede any updates by the next PENDING
627 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
630 static void clear_work_data(struct work_struct *work)
632 smp_wmb(); /* see set_work_pool_and_clear_pending() */
633 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
636 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
638 unsigned long data = atomic_long_read(&work->data);
640 if (data & WORK_STRUCT_PWQ)
641 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
647 * get_work_pool - return the worker_pool a given work was associated with
648 * @work: the work item of interest
650 * Pools are created and destroyed under wq_pool_mutex, and allows read
651 * access under sched-RCU read lock. As such, this function should be
652 * called under wq_pool_mutex or with preemption disabled.
654 * All fields of the returned pool are accessible as long as the above
655 * mentioned locking is in effect. If the returned pool needs to be used
656 * beyond the critical section, the caller is responsible for ensuring the
657 * returned pool is and stays online.
659 * Return: The worker_pool @work was last associated with. %NULL if none.
661 static struct worker_pool *get_work_pool(struct work_struct *work)
663 unsigned long data = atomic_long_read(&work->data);
666 assert_rcu_or_pool_mutex();
668 if (data & WORK_STRUCT_PWQ)
669 return ((struct pool_workqueue *)
670 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
672 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
673 if (pool_id == WORK_OFFQ_POOL_NONE)
676 return idr_find(&worker_pool_idr, pool_id);
680 * get_work_pool_id - return the worker pool ID a given work is associated with
681 * @work: the work item of interest
683 * Return: The worker_pool ID @work was last associated with.
684 * %WORK_OFFQ_POOL_NONE if none.
686 static int get_work_pool_id(struct work_struct *work)
688 unsigned long data = atomic_long_read(&work->data);
690 if (data & WORK_STRUCT_PWQ)
691 return ((struct pool_workqueue *)
692 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
694 return data >> WORK_OFFQ_POOL_SHIFT;
697 static void mark_work_canceling(struct work_struct *work)
699 unsigned long pool_id = get_work_pool_id(work);
701 pool_id <<= WORK_OFFQ_POOL_SHIFT;
702 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
705 static bool work_is_canceling(struct work_struct *work)
707 unsigned long data = atomic_long_read(&work->data);
709 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
713 * Policy functions. These define the policies on how the global worker
714 * pools are managed. Unless noted otherwise, these functions assume that
715 * they're being called with pool->lock held.
718 static bool __need_more_worker(struct worker_pool *pool)
720 return !atomic_read(&pool->nr_running);
724 * Need to wake up a worker? Called from anything but currently
727 * Note that, because unbound workers never contribute to nr_running, this
728 * function will always return %true for unbound pools as long as the
729 * worklist isn't empty.
731 static bool need_more_worker(struct worker_pool *pool)
733 return !list_empty(&pool->worklist) && __need_more_worker(pool);
736 /* Can I start working? Called from busy but !running workers. */
737 static bool may_start_working(struct worker_pool *pool)
739 return pool->nr_idle;
742 /* Do I need to keep working? Called from currently running workers. */
743 static bool keep_working(struct worker_pool *pool)
745 return !list_empty(&pool->worklist) &&
746 atomic_read(&pool->nr_running) <= 1;
749 /* Do we need a new worker? Called from manager. */
750 static bool need_to_create_worker(struct worker_pool *pool)
752 return need_more_worker(pool) && !may_start_working(pool);
755 /* Do we have too many workers and should some go away? */
756 static bool too_many_workers(struct worker_pool *pool)
758 bool managing = mutex_is_locked(&pool->manager_arb);
759 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
760 int nr_busy = pool->nr_workers - nr_idle;
763 * nr_idle and idle_list may disagree if idle rebinding is in
764 * progress. Never return %true if idle_list is empty.
766 if (list_empty(&pool->idle_list))
769 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
776 /* Return the first worker. Safe with preemption disabled */
777 static struct worker *first_worker(struct worker_pool *pool)
779 if (unlikely(list_empty(&pool->idle_list)))
782 return list_first_entry(&pool->idle_list, struct worker, entry);
786 * wake_up_worker - wake up an idle worker
787 * @pool: worker pool to wake worker from
789 * Wake up the first idle worker of @pool.
792 * spin_lock_irq(pool->lock).
794 static void wake_up_worker(struct worker_pool *pool)
796 struct worker *worker = first_worker(pool);
799 wake_up_process(worker->task);
803 * wq_worker_waking_up - a worker is waking up
804 * @task: task waking up
805 * @cpu: CPU @task is waking up to
807 * This function is called during try_to_wake_up() when a worker is
811 * spin_lock_irq(rq->lock)
813 void wq_worker_waking_up(struct task_struct *task, int cpu)
815 struct worker *worker = kthread_data(task);
817 if (!(worker->flags & WORKER_NOT_RUNNING)) {
818 WARN_ON_ONCE(worker->pool->cpu != cpu);
819 atomic_inc(&worker->pool->nr_running);
824 * wq_worker_sleeping - a worker is going to sleep
825 * @task: task going to sleep
826 * @cpu: CPU in question, must be the current CPU number
828 * This function is called during schedule() when a busy worker is
829 * going to sleep. Worker on the same cpu can be woken up by
830 * returning pointer to its task.
833 * spin_lock_irq(rq->lock)
836 * Worker task on @cpu to wake up, %NULL if none.
838 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
840 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
841 struct worker_pool *pool;
844 * Rescuers, which may not have all the fields set up like normal
845 * workers, also reach here, let's not access anything before
846 * checking NOT_RUNNING.
848 if (worker->flags & WORKER_NOT_RUNNING)
853 /* this can only happen on the local cpu */
854 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
858 * The counterpart of the following dec_and_test, implied mb,
859 * worklist not empty test sequence is in insert_work().
860 * Please read comment there.
862 * NOT_RUNNING is clear. This means that we're bound to and
863 * running on the local cpu w/ rq lock held and preemption
864 * disabled, which in turn means that none else could be
865 * manipulating idle_list, so dereferencing idle_list without pool
868 if (atomic_dec_and_test(&pool->nr_running) &&
869 !list_empty(&pool->worklist))
870 to_wakeup = first_worker(pool);
871 return to_wakeup ? to_wakeup->task : NULL;
875 * worker_set_flags - set worker flags and adjust nr_running accordingly
877 * @flags: flags to set
878 * @wakeup: wakeup an idle worker if necessary
880 * Set @flags in @worker->flags and adjust nr_running accordingly. If
881 * nr_running becomes zero and @wakeup is %true, an idle worker is
885 * spin_lock_irq(pool->lock)
887 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
890 struct worker_pool *pool = worker->pool;
892 WARN_ON_ONCE(worker->task != current);
895 * If transitioning into NOT_RUNNING, adjust nr_running and
896 * wake up an idle worker as necessary if requested by
899 if ((flags & WORKER_NOT_RUNNING) &&
900 !(worker->flags & WORKER_NOT_RUNNING)) {
902 if (atomic_dec_and_test(&pool->nr_running) &&
903 !list_empty(&pool->worklist))
904 wake_up_worker(pool);
906 atomic_dec(&pool->nr_running);
909 worker->flags |= flags;
913 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
915 * @flags: flags to clear
917 * Clear @flags in @worker->flags and adjust nr_running accordingly.
920 * spin_lock_irq(pool->lock)
922 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
924 struct worker_pool *pool = worker->pool;
925 unsigned int oflags = worker->flags;
927 WARN_ON_ONCE(worker->task != current);
929 worker->flags &= ~flags;
932 * If transitioning out of NOT_RUNNING, increment nr_running. Note
933 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
934 * of multiple flags, not a single flag.
936 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
937 if (!(worker->flags & WORKER_NOT_RUNNING))
938 atomic_inc(&pool->nr_running);
942 * find_worker_executing_work - find worker which is executing a work
943 * @pool: pool of interest
944 * @work: work to find worker for
946 * Find a worker which is executing @work on @pool by searching
947 * @pool->busy_hash which is keyed by the address of @work. For a worker
948 * to match, its current execution should match the address of @work and
949 * its work function. This is to avoid unwanted dependency between
950 * unrelated work executions through a work item being recycled while still
953 * This is a bit tricky. A work item may be freed once its execution
954 * starts and nothing prevents the freed area from being recycled for
955 * another work item. If the same work item address ends up being reused
956 * before the original execution finishes, workqueue will identify the
957 * recycled work item as currently executing and make it wait until the
958 * current execution finishes, introducing an unwanted dependency.
960 * This function checks the work item address and work function to avoid
961 * false positives. Note that this isn't complete as one may construct a
962 * work function which can introduce dependency onto itself through a
963 * recycled work item. Well, if somebody wants to shoot oneself in the
964 * foot that badly, there's only so much we can do, and if such deadlock
965 * actually occurs, it should be easy to locate the culprit work function.
968 * spin_lock_irq(pool->lock).
971 * Pointer to worker which is executing @work if found, %NULL
974 static struct worker *find_worker_executing_work(struct worker_pool *pool,
975 struct work_struct *work)
977 struct worker *worker;
979 hash_for_each_possible(pool->busy_hash, worker, hentry,
981 if (worker->current_work == work &&
982 worker->current_func == work->func)
989 * move_linked_works - move linked works to a list
990 * @work: start of series of works to be scheduled
991 * @head: target list to append @work to
992 * @nextp: out paramter for nested worklist walking
994 * Schedule linked works starting from @work to @head. Work series to
995 * be scheduled starts at @work and includes any consecutive work with
996 * WORK_STRUCT_LINKED set in its predecessor.
998 * If @nextp is not NULL, it's updated to point to the next work of
999 * the last scheduled work. This allows move_linked_works() to be
1000 * nested inside outer list_for_each_entry_safe().
1003 * spin_lock_irq(pool->lock).
1005 static void move_linked_works(struct work_struct *work, struct list_head *head,
1006 struct work_struct **nextp)
1008 struct work_struct *n;
1011 * Linked worklist will always end before the end of the list,
1012 * use NULL for list head.
1014 list_for_each_entry_safe_from(work, n, NULL, entry) {
1015 list_move_tail(&work->entry, head);
1016 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1021 * If we're already inside safe list traversal and have moved
1022 * multiple works to the scheduled queue, the next position
1023 * needs to be updated.
1030 * get_pwq - get an extra reference on the specified pool_workqueue
1031 * @pwq: pool_workqueue to get
1033 * Obtain an extra reference on @pwq. The caller should guarantee that
1034 * @pwq has positive refcnt and be holding the matching pool->lock.
1036 static void get_pwq(struct pool_workqueue *pwq)
1038 lockdep_assert_held(&pwq->pool->lock);
1039 WARN_ON_ONCE(pwq->refcnt <= 0);
1044 * put_pwq - put a pool_workqueue reference
1045 * @pwq: pool_workqueue to put
1047 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1048 * destruction. The caller should be holding the matching pool->lock.
1050 static void put_pwq(struct pool_workqueue *pwq)
1052 lockdep_assert_held(&pwq->pool->lock);
1053 if (likely(--pwq->refcnt))
1055 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1058 * @pwq can't be released under pool->lock, bounce to
1059 * pwq_unbound_release_workfn(). This never recurses on the same
1060 * pool->lock as this path is taken only for unbound workqueues and
1061 * the release work item is scheduled on a per-cpu workqueue. To
1062 * avoid lockdep warning, unbound pool->locks are given lockdep
1063 * subclass of 1 in get_unbound_pool().
1065 schedule_work(&pwq->unbound_release_work);
1069 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1070 * @pwq: pool_workqueue to put (can be %NULL)
1072 * put_pwq() with locking. This function also allows %NULL @pwq.
1074 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1078 * As both pwqs and pools are sched-RCU protected, the
1079 * following lock operations are safe.
1081 spin_lock_irq(&pwq->pool->lock);
1083 spin_unlock_irq(&pwq->pool->lock);
1087 static void pwq_activate_delayed_work(struct work_struct *work)
1089 struct pool_workqueue *pwq = get_work_pwq(work);
1091 trace_workqueue_activate_work(work);
1092 move_linked_works(work, &pwq->pool->worklist, NULL);
1093 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1097 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1099 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1100 struct work_struct, entry);
1102 pwq_activate_delayed_work(work);
1106 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1107 * @pwq: pwq of interest
1108 * @color: color of work which left the queue
1110 * A work either has completed or is removed from pending queue,
1111 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1114 * spin_lock_irq(pool->lock).
1116 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1118 /* uncolored work items don't participate in flushing or nr_active */
1119 if (color == WORK_NO_COLOR)
1122 pwq->nr_in_flight[color]--;
1125 if (!list_empty(&pwq->delayed_works)) {
1126 /* one down, submit a delayed one */
1127 if (pwq->nr_active < pwq->max_active)
1128 pwq_activate_first_delayed(pwq);
1131 /* is flush in progress and are we at the flushing tip? */
1132 if (likely(pwq->flush_color != color))
1135 /* are there still in-flight works? */
1136 if (pwq->nr_in_flight[color])
1139 /* this pwq is done, clear flush_color */
1140 pwq->flush_color = -1;
1143 * If this was the last pwq, wake up the first flusher. It
1144 * will handle the rest.
1146 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1147 complete(&pwq->wq->first_flusher->done);
1153 * try_to_grab_pending - steal work item from worklist and disable irq
1154 * @work: work item to steal
1155 * @is_dwork: @work is a delayed_work
1156 * @flags: place to store irq state
1158 * Try to grab PENDING bit of @work. This function can handle @work in any
1159 * stable state - idle, on timer or on worklist.
1162 * 1 if @work was pending and we successfully stole PENDING
1163 * 0 if @work was idle and we claimed PENDING
1164 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1165 * -ENOENT if someone else is canceling @work, this state may persist
1166 * for arbitrarily long
1169 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1170 * interrupted while holding PENDING and @work off queue, irq must be
1171 * disabled on entry. This, combined with delayed_work->timer being
1172 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1174 * On successful return, >= 0, irq is disabled and the caller is
1175 * responsible for releasing it using local_irq_restore(*@flags).
1177 * This function is safe to call from any context including IRQ handler.
1179 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1180 unsigned long *flags)
1182 struct worker_pool *pool;
1183 struct pool_workqueue *pwq;
1185 local_irq_save(*flags);
1187 /* try to steal the timer if it exists */
1189 struct delayed_work *dwork = to_delayed_work(work);
1192 * dwork->timer is irqsafe. If del_timer() fails, it's
1193 * guaranteed that the timer is not queued anywhere and not
1194 * running on the local CPU.
1196 if (likely(del_timer(&dwork->timer)))
1200 /* try to claim PENDING the normal way */
1201 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1205 * The queueing is in progress, or it is already queued. Try to
1206 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1208 pool = get_work_pool(work);
1212 spin_lock(&pool->lock);
1214 * work->data is guaranteed to point to pwq only while the work
1215 * item is queued on pwq->wq, and both updating work->data to point
1216 * to pwq on queueing and to pool on dequeueing are done under
1217 * pwq->pool->lock. This in turn guarantees that, if work->data
1218 * points to pwq which is associated with a locked pool, the work
1219 * item is currently queued on that pool.
1221 pwq = get_work_pwq(work);
1222 if (pwq && pwq->pool == pool) {
1223 debug_work_deactivate(work);
1226 * A delayed work item cannot be grabbed directly because
1227 * it might have linked NO_COLOR work items which, if left
1228 * on the delayed_list, will confuse pwq->nr_active
1229 * management later on and cause stall. Make sure the work
1230 * item is activated before grabbing.
1232 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1233 pwq_activate_delayed_work(work);
1235 list_del_init(&work->entry);
1236 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1238 /* work->data points to pwq iff queued, point to pool */
1239 set_work_pool_and_keep_pending(work, pool->id);
1241 spin_unlock(&pool->lock);
1244 spin_unlock(&pool->lock);
1246 local_irq_restore(*flags);
1247 if (work_is_canceling(work))
1254 * insert_work - insert a work into a pool
1255 * @pwq: pwq @work belongs to
1256 * @work: work to insert
1257 * @head: insertion point
1258 * @extra_flags: extra WORK_STRUCT_* flags to set
1260 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1261 * work_struct flags.
1264 * spin_lock_irq(pool->lock).
1266 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1267 struct list_head *head, unsigned int extra_flags)
1269 struct worker_pool *pool = pwq->pool;
1271 /* we own @work, set data and link */
1272 set_work_pwq(work, pwq, extra_flags);
1273 list_add_tail(&work->entry, head);
1277 * Ensure either wq_worker_sleeping() sees the above
1278 * list_add_tail() or we see zero nr_running to avoid workers lying
1279 * around lazily while there are works to be processed.
1283 if (__need_more_worker(pool))
1284 wake_up_worker(pool);
1288 * Test whether @work is being queued from another work executing on the
1291 static bool is_chained_work(struct workqueue_struct *wq)
1293 struct worker *worker;
1295 worker = current_wq_worker();
1297 * Return %true iff I'm a worker execuing a work item on @wq. If
1298 * I'm @worker, it's safe to dereference it without locking.
1300 return worker && worker->current_pwq->wq == wq;
1303 static void __queue_work(int cpu, struct workqueue_struct *wq,
1304 struct work_struct *work)
1306 struct pool_workqueue *pwq;
1307 struct worker_pool *last_pool;
1308 struct list_head *worklist;
1309 unsigned int work_flags;
1310 unsigned int req_cpu = cpu;
1313 * While a work item is PENDING && off queue, a task trying to
1314 * steal the PENDING will busy-loop waiting for it to either get
1315 * queued or lose PENDING. Grabbing PENDING and queueing should
1316 * happen with IRQ disabled.
1318 WARN_ON_ONCE(!irqs_disabled());
1320 debug_work_activate(work);
1322 /* if draining, only works from the same workqueue are allowed */
1323 if (unlikely(wq->flags & __WQ_DRAINING) &&
1324 WARN_ON_ONCE(!is_chained_work(wq)))
1327 if (req_cpu == WORK_CPU_UNBOUND)
1328 cpu = raw_smp_processor_id();
1330 /* pwq which will be used unless @work is executing elsewhere */
1331 if (!(wq->flags & WQ_UNBOUND))
1332 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1334 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1337 * If @work was previously on a different pool, it might still be
1338 * running there, in which case the work needs to be queued on that
1339 * pool to guarantee non-reentrancy.
1341 last_pool = get_work_pool(work);
1342 if (last_pool && last_pool != pwq->pool) {
1343 struct worker *worker;
1345 spin_lock(&last_pool->lock);
1347 worker = find_worker_executing_work(last_pool, work);
1349 if (worker && worker->current_pwq->wq == wq) {
1350 pwq = worker->current_pwq;
1352 /* meh... not running there, queue here */
1353 spin_unlock(&last_pool->lock);
1354 spin_lock(&pwq->pool->lock);
1357 spin_lock(&pwq->pool->lock);
1361 * pwq is determined and locked. For unbound pools, we could have
1362 * raced with pwq release and it could already be dead. If its
1363 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1364 * without another pwq replacing it in the numa_pwq_tbl or while
1365 * work items are executing on it, so the retrying is guaranteed to
1366 * make forward-progress.
1368 if (unlikely(!pwq->refcnt)) {
1369 if (wq->flags & WQ_UNBOUND) {
1370 spin_unlock(&pwq->pool->lock);
1375 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1379 /* pwq determined, queue */
1380 trace_workqueue_queue_work(req_cpu, pwq, work);
1382 if (WARN_ON(!list_empty(&work->entry))) {
1383 spin_unlock(&pwq->pool->lock);
1387 pwq->nr_in_flight[pwq->work_color]++;
1388 work_flags = work_color_to_flags(pwq->work_color);
1390 if (likely(pwq->nr_active < pwq->max_active)) {
1391 trace_workqueue_activate_work(work);
1393 worklist = &pwq->pool->worklist;
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 start_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);
1607 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1608 * @pool: target worker_pool
1610 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1612 * Works which are scheduled while the cpu is online must at least be
1613 * scheduled to a worker which is bound to the cpu so that if they are
1614 * flushed from cpu callbacks while cpu is going down, they are
1615 * guaranteed to execute on the cpu.
1617 * This function is to be used by unbound workers and rescuers to bind
1618 * themselves to the target cpu and may race with cpu going down or
1619 * coming online. kthread_bind() can't be used because it may put the
1620 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1621 * verbatim as it's best effort and blocking and pool may be
1622 * [dis]associated in the meantime.
1624 * This function tries set_cpus_allowed() and locks pool and verifies the
1625 * binding against %POOL_DISASSOCIATED which is set during
1626 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1627 * enters idle state or fetches works without dropping lock, it can
1628 * guarantee the scheduling requirement described in the first paragraph.
1631 * Might sleep. Called without any lock but returns with pool->lock
1635 * %true if the associated pool is online (@worker is successfully
1636 * bound), %false if offline.
1638 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1639 __acquires(&pool->lock)
1643 * The following call may fail, succeed or succeed
1644 * without actually migrating the task to the cpu if
1645 * it races with cpu hotunplug operation. Verify
1646 * against POOL_DISASSOCIATED.
1648 if (!(pool->flags & POOL_DISASSOCIATED))
1649 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1651 spin_lock_irq(&pool->lock);
1652 if (pool->flags & POOL_DISASSOCIATED)
1654 if (task_cpu(current) == pool->cpu &&
1655 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1657 spin_unlock_irq(&pool->lock);
1660 * We've raced with CPU hot[un]plug. Give it a breather
1661 * and retry migration. cond_resched() is required here;
1662 * otherwise, we might deadlock against cpu_stop trying to
1663 * bring down the CPU on non-preemptive kernel.
1670 static struct worker *alloc_worker(void)
1672 struct worker *worker;
1674 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1676 INIT_LIST_HEAD(&worker->entry);
1677 INIT_LIST_HEAD(&worker->scheduled);
1678 INIT_LIST_HEAD(&worker->node);
1679 /* on creation a worker is in !idle && prep state */
1680 worker->flags = WORKER_PREP;
1686 * worker_attach_to_pool() - attach a worker to a pool
1687 * @worker: worker to be attached
1688 * @pool: the target pool
1690 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1691 * cpu-binding of @worker are kept coordinated with the pool across
1694 static void worker_attach_to_pool(struct worker *worker,
1695 struct worker_pool *pool)
1697 mutex_lock(&pool->attach_mutex);
1700 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1701 * online CPUs. It'll be re-applied when any of the CPUs come up.
1703 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1706 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1707 * stable across this function. See the comments above the
1708 * flag definition for details.
1710 if (pool->flags & POOL_DISASSOCIATED)
1711 worker->flags |= WORKER_UNBOUND;
1713 list_add_tail(&worker->node, &pool->workers);
1715 mutex_unlock(&pool->attach_mutex);
1719 * worker_detach_from_pool() - detach a worker from its pool
1720 * @worker: worker which is attached to its pool
1721 * @pool: the pool @worker is attached to
1723 * Undo the attaching which had been done in worker_attach_to_pool(). The
1724 * caller worker shouldn't access to the pool after detached except it has
1725 * other reference to the pool.
1727 static void worker_detach_from_pool(struct worker *worker,
1728 struct worker_pool *pool)
1730 struct completion *detach_completion = NULL;
1732 mutex_lock(&pool->attach_mutex);
1733 list_del(&worker->node);
1734 if (list_empty(&pool->workers))
1735 detach_completion = pool->detach_completion;
1736 mutex_unlock(&pool->attach_mutex);
1738 if (detach_completion)
1739 complete(detach_completion);
1743 * create_worker - create a new workqueue worker
1744 * @pool: pool the new worker will belong to
1746 * Create a new worker which is attached to @pool. The new worker must be
1747 * started by start_worker().
1750 * Might sleep. Does GFP_KERNEL allocations.
1753 * Pointer to the newly created worker.
1755 static struct worker *create_worker(struct worker_pool *pool)
1757 struct worker *worker = NULL;
1761 /* ID is needed to determine kthread name */
1762 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1766 worker = alloc_worker();
1770 worker->pool = pool;
1774 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1775 pool->attrs->nice < 0 ? "H" : "");
1777 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1779 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1780 "kworker/%s", id_buf);
1781 if (IS_ERR(worker->task))
1784 set_user_nice(worker->task, pool->attrs->nice);
1786 /* prevent userland from meddling with cpumask of workqueue workers */
1787 worker->task->flags |= PF_NO_SETAFFINITY;
1789 /* successful, attach the worker to the pool */
1790 worker_attach_to_pool(worker, pool);
1796 ida_simple_remove(&pool->worker_ida, id);
1802 * start_worker - start a newly created worker
1803 * @worker: worker to start
1805 * Make the pool aware of @worker and start it.
1808 * spin_lock_irq(pool->lock).
1810 static void start_worker(struct worker *worker)
1812 worker->pool->nr_workers++;
1813 worker_enter_idle(worker);
1814 wake_up_process(worker->task);
1818 * create_and_start_worker - create and start a worker for a pool
1819 * @pool: the target pool
1821 * Grab the managership of @pool and create and start a new worker for it.
1823 * Return: 0 on success. A negative error code otherwise.
1825 static int create_and_start_worker(struct worker_pool *pool)
1827 struct worker *worker;
1829 worker = create_worker(pool);
1831 spin_lock_irq(&pool->lock);
1832 start_worker(worker);
1833 spin_unlock_irq(&pool->lock);
1836 return worker ? 0 : -ENOMEM;
1840 * destroy_worker - destroy a workqueue worker
1841 * @worker: worker to be destroyed
1843 * Destroy @worker and adjust @pool stats accordingly. The worker should
1847 * spin_lock_irq(pool->lock).
1849 static void destroy_worker(struct worker *worker)
1851 struct worker_pool *pool = worker->pool;
1853 lockdep_assert_held(&pool->lock);
1855 /* sanity check frenzy */
1856 if (WARN_ON(worker->current_work) ||
1857 WARN_ON(!list_empty(&worker->scheduled)) ||
1858 WARN_ON(!(worker->flags & WORKER_IDLE)))
1864 list_del_init(&worker->entry);
1865 worker->flags |= WORKER_DIE;
1866 wake_up_process(worker->task);
1869 static void idle_worker_timeout(unsigned long __pool)
1871 struct worker_pool *pool = (void *)__pool;
1873 spin_lock_irq(&pool->lock);
1875 while (too_many_workers(pool)) {
1876 struct worker *worker;
1877 unsigned long expires;
1879 /* idle_list is kept in LIFO order, check the last one */
1880 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1881 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1883 if (time_before(jiffies, expires)) {
1884 mod_timer(&pool->idle_timer, expires);
1888 destroy_worker(worker);
1891 spin_unlock_irq(&pool->lock);
1894 static void send_mayday(struct work_struct *work)
1896 struct pool_workqueue *pwq = get_work_pwq(work);
1897 struct workqueue_struct *wq = pwq->wq;
1899 lockdep_assert_held(&wq_mayday_lock);
1904 /* mayday mayday mayday */
1905 if (list_empty(&pwq->mayday_node)) {
1907 * If @pwq is for an unbound wq, its base ref may be put at
1908 * any time due to an attribute change. Pin @pwq until the
1909 * rescuer is done with it.
1912 list_add_tail(&pwq->mayday_node, &wq->maydays);
1913 wake_up_process(wq->rescuer->task);
1917 static void pool_mayday_timeout(unsigned long __pool)
1919 struct worker_pool *pool = (void *)__pool;
1920 struct work_struct *work;
1922 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1923 spin_lock(&pool->lock);
1925 if (need_to_create_worker(pool)) {
1927 * We've been trying to create a new worker but
1928 * haven't been successful. We might be hitting an
1929 * allocation deadlock. Send distress signals to
1932 list_for_each_entry(work, &pool->worklist, entry)
1936 spin_unlock(&pool->lock);
1937 spin_unlock_irq(&wq_mayday_lock);
1939 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1943 * maybe_create_worker - create a new worker if necessary
1944 * @pool: pool to create a new worker for
1946 * Create a new worker for @pool if necessary. @pool is guaranteed to
1947 * have at least one idle worker on return from this function. If
1948 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1949 * sent to all rescuers with works scheduled on @pool to resolve
1950 * possible allocation deadlock.
1952 * On return, need_to_create_worker() is guaranteed to be %false and
1953 * may_start_working() %true.
1956 * spin_lock_irq(pool->lock) which may be released and regrabbed
1957 * multiple times. Does GFP_KERNEL allocations. Called only from
1961 * %false if no action was taken and pool->lock stayed locked, %true
1964 static bool maybe_create_worker(struct worker_pool *pool)
1965 __releases(&pool->lock)
1966 __acquires(&pool->lock)
1968 if (!need_to_create_worker(pool))
1971 spin_unlock_irq(&pool->lock);
1973 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1974 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1977 struct worker *worker;
1979 worker = create_worker(pool);
1981 del_timer_sync(&pool->mayday_timer);
1982 spin_lock_irq(&pool->lock);
1983 start_worker(worker);
1984 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1989 if (!need_to_create_worker(pool))
1992 __set_current_state(TASK_INTERRUPTIBLE);
1993 schedule_timeout(CREATE_COOLDOWN);
1995 if (!need_to_create_worker(pool))
1999 del_timer_sync(&pool->mayday_timer);
2000 spin_lock_irq(&pool->lock);
2001 if (need_to_create_worker(pool))
2007 * manage_workers - manage worker pool
2010 * Assume the manager role and manage the worker pool @worker belongs
2011 * to. At any given time, there can be only zero or one manager per
2012 * pool. The exclusion is handled automatically by this function.
2014 * The caller can safely start processing works on false return. On
2015 * true return, it's guaranteed that need_to_create_worker() is false
2016 * and may_start_working() is true.
2019 * spin_lock_irq(pool->lock) which may be released and regrabbed
2020 * multiple times. Does GFP_KERNEL allocations.
2023 * %false if the pool don't need management and the caller can safely start
2024 * processing works, %true indicates that the function released pool->lock
2025 * and reacquired it to perform some management function and that the
2026 * conditions that the caller verified while holding the lock before
2027 * calling the function might no longer be true.
2029 static bool manage_workers(struct worker *worker)
2031 struct worker_pool *pool = worker->pool;
2035 * Anyone who successfully grabs manager_arb wins the arbitration
2036 * and becomes the manager. mutex_trylock() on pool->manager_arb
2037 * failure while holding pool->lock reliably indicates that someone
2038 * else is managing the pool and the worker which failed trylock
2039 * can proceed to executing work items. This means that anyone
2040 * grabbing manager_arb is responsible for actually performing
2041 * manager duties. If manager_arb is grabbed and released without
2042 * actual management, the pool may stall indefinitely.
2044 if (!mutex_trylock(&pool->manager_arb))
2047 ret |= maybe_create_worker(pool);
2049 mutex_unlock(&pool->manager_arb);
2054 * process_one_work - process single work
2056 * @work: work to process
2058 * Process @work. This function contains all the logics necessary to
2059 * process a single work including synchronization against and
2060 * interaction with other workers on the same cpu, queueing and
2061 * flushing. As long as context requirement is met, any worker can
2062 * call this function to process a work.
2065 * spin_lock_irq(pool->lock) which is released and regrabbed.
2067 static void process_one_work(struct worker *worker, struct work_struct *work)
2068 __releases(&pool->lock)
2069 __acquires(&pool->lock)
2071 struct pool_workqueue *pwq = get_work_pwq(work);
2072 struct worker_pool *pool = worker->pool;
2073 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2075 struct worker *collision;
2076 #ifdef CONFIG_LOCKDEP
2078 * It is permissible to free the struct work_struct from
2079 * inside the function that is called from it, this we need to
2080 * take into account for lockdep too. To avoid bogus "held
2081 * lock freed" warnings as well as problems when looking into
2082 * work->lockdep_map, make a copy and use that here.
2084 struct lockdep_map lockdep_map;
2086 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2089 * Ensure we're on the correct CPU. DISASSOCIATED test is
2090 * necessary to avoid spurious warnings from rescuers servicing the
2091 * unbound or a disassociated pool.
2093 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2094 !(pool->flags & POOL_DISASSOCIATED) &&
2095 raw_smp_processor_id() != pool->cpu);
2098 * A single work shouldn't be executed concurrently by
2099 * multiple workers on a single cpu. Check whether anyone is
2100 * already processing the work. If so, defer the work to the
2101 * currently executing one.
2103 collision = find_worker_executing_work(pool, work);
2104 if (unlikely(collision)) {
2105 move_linked_works(work, &collision->scheduled, NULL);
2109 /* claim and dequeue */
2110 debug_work_deactivate(work);
2111 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2112 worker->current_work = work;
2113 worker->current_func = work->func;
2114 worker->current_pwq = pwq;
2115 work_color = get_work_color(work);
2117 list_del_init(&work->entry);
2120 * CPU intensive works don't participate in concurrency
2121 * management. They're the scheduler's responsibility.
2123 if (unlikely(cpu_intensive))
2124 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2127 * Unbound pool isn't concurrency managed and work items should be
2128 * executed ASAP. Wake up another worker if necessary.
2130 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2131 wake_up_worker(pool);
2134 * Record the last pool and clear PENDING which should be the last
2135 * update to @work. Also, do this inside @pool->lock so that
2136 * PENDING and queued state changes happen together while IRQ is
2139 set_work_pool_and_clear_pending(work, pool->id);
2141 spin_unlock_irq(&pool->lock);
2143 lock_map_acquire_read(&pwq->wq->lockdep_map);
2144 lock_map_acquire(&lockdep_map);
2145 trace_workqueue_execute_start(work);
2146 worker->current_func(work);
2148 * While we must be careful to not use "work" after this, the trace
2149 * point will only record its address.
2151 trace_workqueue_execute_end(work);
2152 lock_map_release(&lockdep_map);
2153 lock_map_release(&pwq->wq->lockdep_map);
2155 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2156 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2157 " last function: %pf\n",
2158 current->comm, preempt_count(), task_pid_nr(current),
2159 worker->current_func);
2160 debug_show_held_locks(current);
2165 * The following prevents a kworker from hogging CPU on !PREEMPT
2166 * kernels, where a requeueing work item waiting for something to
2167 * happen could deadlock with stop_machine as such work item could
2168 * indefinitely requeue itself while all other CPUs are trapped in
2173 spin_lock_irq(&pool->lock);
2175 /* clear cpu intensive status */
2176 if (unlikely(cpu_intensive))
2177 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2179 /* we're done with it, release */
2180 hash_del(&worker->hentry);
2181 worker->current_work = NULL;
2182 worker->current_func = NULL;
2183 worker->current_pwq = NULL;
2184 worker->desc_valid = false;
2185 pwq_dec_nr_in_flight(pwq, work_color);
2189 * process_scheduled_works - process scheduled works
2192 * Process all scheduled works. Please note that the scheduled list
2193 * may change while processing a work, so this function repeatedly
2194 * fetches a work from the top and executes it.
2197 * spin_lock_irq(pool->lock) which may be released and regrabbed
2200 static void process_scheduled_works(struct worker *worker)
2202 while (!list_empty(&worker->scheduled)) {
2203 struct work_struct *work = list_first_entry(&worker->scheduled,
2204 struct work_struct, entry);
2205 process_one_work(worker, work);
2210 * worker_thread - the worker thread function
2213 * The worker thread function. All workers belong to a worker_pool -
2214 * either a per-cpu one or dynamic unbound one. These workers process all
2215 * work items regardless of their specific target workqueue. The only
2216 * exception is work items which belong to workqueues with a rescuer which
2217 * will be explained in rescuer_thread().
2221 static int worker_thread(void *__worker)
2223 struct worker *worker = __worker;
2224 struct worker_pool *pool = worker->pool;
2226 /* tell the scheduler that this is a workqueue worker */
2227 worker->task->flags |= PF_WQ_WORKER;
2229 spin_lock_irq(&pool->lock);
2231 /* am I supposed to die? */
2232 if (unlikely(worker->flags & WORKER_DIE)) {
2233 spin_unlock_irq(&pool->lock);
2234 WARN_ON_ONCE(!list_empty(&worker->entry));
2235 worker->task->flags &= ~PF_WQ_WORKER;
2237 set_task_comm(worker->task, "kworker/dying");
2238 ida_simple_remove(&pool->worker_ida, worker->id);
2239 worker_detach_from_pool(worker, pool);
2244 worker_leave_idle(worker);
2246 /* no more worker necessary? */
2247 if (!need_more_worker(pool))
2250 /* do we need to manage? */
2251 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2255 * ->scheduled list can only be filled while a worker is
2256 * preparing to process a work or actually processing it.
2257 * Make sure nobody diddled with it while I was sleeping.
2259 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2262 * Finish PREP stage. We're guaranteed to have at least one idle
2263 * worker or that someone else has already assumed the manager
2264 * role. This is where @worker starts participating in concurrency
2265 * management if applicable and concurrency management is restored
2266 * after being rebound. See rebind_workers() for details.
2268 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2271 struct work_struct *work =
2272 list_first_entry(&pool->worklist,
2273 struct work_struct, entry);
2275 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2276 /* optimization path, not strictly necessary */
2277 process_one_work(worker, work);
2278 if (unlikely(!list_empty(&worker->scheduled)))
2279 process_scheduled_works(worker);
2281 move_linked_works(work, &worker->scheduled, NULL);
2282 process_scheduled_works(worker);
2284 } while (keep_working(pool));
2286 worker_set_flags(worker, WORKER_PREP, false);
2289 * pool->lock is held and there's no work to process and no need to
2290 * manage, sleep. Workers are woken up only while holding
2291 * pool->lock or from local cpu, so setting the current state
2292 * before releasing pool->lock is enough to prevent losing any
2295 worker_enter_idle(worker);
2296 __set_current_state(TASK_INTERRUPTIBLE);
2297 spin_unlock_irq(&pool->lock);
2303 * rescuer_thread - the rescuer thread function
2306 * Workqueue rescuer thread function. There's one rescuer for each
2307 * workqueue which has WQ_MEM_RECLAIM set.
2309 * Regular work processing on a pool may block trying to create a new
2310 * worker which uses GFP_KERNEL allocation which has slight chance of
2311 * developing into deadlock if some works currently on the same queue
2312 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2313 * the problem rescuer solves.
2315 * When such condition is possible, the pool summons rescuers of all
2316 * workqueues which have works queued on the pool and let them process
2317 * those works so that forward progress can be guaranteed.
2319 * This should happen rarely.
2323 static int rescuer_thread(void *__rescuer)
2325 struct worker *rescuer = __rescuer;
2326 struct workqueue_struct *wq = rescuer->rescue_wq;
2327 struct list_head *scheduled = &rescuer->scheduled;
2330 set_user_nice(current, RESCUER_NICE_LEVEL);
2333 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2334 * doesn't participate in concurrency management.
2336 rescuer->task->flags |= PF_WQ_WORKER;
2338 set_current_state(TASK_INTERRUPTIBLE);
2341 * By the time the rescuer is requested to stop, the workqueue
2342 * shouldn't have any work pending, but @wq->maydays may still have
2343 * pwq(s) queued. This can happen by non-rescuer workers consuming
2344 * all the work items before the rescuer got to them. Go through
2345 * @wq->maydays processing before acting on should_stop so that the
2346 * list is always empty on exit.
2348 should_stop = kthread_should_stop();
2350 /* see whether any pwq is asking for help */
2351 spin_lock_irq(&wq_mayday_lock);
2353 while (!list_empty(&wq->maydays)) {
2354 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2355 struct pool_workqueue, mayday_node);
2356 struct worker_pool *pool = pwq->pool;
2357 struct work_struct *work, *n;
2359 __set_current_state(TASK_RUNNING);
2360 list_del_init(&pwq->mayday_node);
2362 spin_unlock_irq(&wq_mayday_lock);
2364 /* migrate to the target cpu if possible */
2365 worker_maybe_bind_and_lock(pool);
2366 rescuer->pool = pool;
2369 * Slurp in all works issued via this workqueue and
2372 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2373 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2374 if (get_work_pwq(work) == pwq)
2375 move_linked_works(work, scheduled, &n);
2377 process_scheduled_works(rescuer);
2380 * Put the reference grabbed by send_mayday(). @pool won't
2381 * go away while we're holding its lock.
2386 * Leave this pool. If keep_working() is %true, notify a
2387 * regular worker; otherwise, we end up with 0 concurrency
2388 * and stalling the execution.
2390 if (keep_working(pool))
2391 wake_up_worker(pool);
2393 rescuer->pool = NULL;
2394 spin_unlock(&pool->lock);
2395 spin_lock(&wq_mayday_lock);
2398 spin_unlock_irq(&wq_mayday_lock);
2401 __set_current_state(TASK_RUNNING);
2402 rescuer->task->flags &= ~PF_WQ_WORKER;
2406 /* rescuers should never participate in concurrency management */
2407 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2413 struct work_struct work;
2414 struct completion done;
2417 static void wq_barrier_func(struct work_struct *work)
2419 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2420 complete(&barr->done);
2424 * insert_wq_barrier - insert a barrier work
2425 * @pwq: pwq to insert barrier into
2426 * @barr: wq_barrier to insert
2427 * @target: target work to attach @barr to
2428 * @worker: worker currently executing @target, NULL if @target is not executing
2430 * @barr is linked to @target such that @barr is completed only after
2431 * @target finishes execution. Please note that the ordering
2432 * guarantee is observed only with respect to @target and on the local
2435 * Currently, a queued barrier can't be canceled. This is because
2436 * try_to_grab_pending() can't determine whether the work to be
2437 * grabbed is at the head of the queue and thus can't clear LINKED
2438 * flag of the previous work while there must be a valid next work
2439 * after a work with LINKED flag set.
2441 * Note that when @worker is non-NULL, @target may be modified
2442 * underneath us, so we can't reliably determine pwq from @target.
2445 * spin_lock_irq(pool->lock).
2447 static void insert_wq_barrier(struct pool_workqueue *pwq,
2448 struct wq_barrier *barr,
2449 struct work_struct *target, struct worker *worker)
2451 struct list_head *head;
2452 unsigned int linked = 0;
2455 * debugobject calls are safe here even with pool->lock locked
2456 * as we know for sure that this will not trigger any of the
2457 * checks and call back into the fixup functions where we
2460 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2461 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2462 init_completion(&barr->done);
2465 * If @target is currently being executed, schedule the
2466 * barrier to the worker; otherwise, put it after @target.
2469 head = worker->scheduled.next;
2471 unsigned long *bits = work_data_bits(target);
2473 head = target->entry.next;
2474 /* there can already be other linked works, inherit and set */
2475 linked = *bits & WORK_STRUCT_LINKED;
2476 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2479 debug_work_activate(&barr->work);
2480 insert_work(pwq, &barr->work, head,
2481 work_color_to_flags(WORK_NO_COLOR) | linked);
2485 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2486 * @wq: workqueue being flushed
2487 * @flush_color: new flush color, < 0 for no-op
2488 * @work_color: new work color, < 0 for no-op
2490 * Prepare pwqs for workqueue flushing.
2492 * If @flush_color is non-negative, flush_color on all pwqs should be
2493 * -1. If no pwq has in-flight commands at the specified color, all
2494 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2495 * has in flight commands, its pwq->flush_color is set to
2496 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2497 * wakeup logic is armed and %true is returned.
2499 * The caller should have initialized @wq->first_flusher prior to
2500 * calling this function with non-negative @flush_color. If
2501 * @flush_color is negative, no flush color update is done and %false
2504 * If @work_color is non-negative, all pwqs should have the same
2505 * work_color which is previous to @work_color and all will be
2506 * advanced to @work_color.
2509 * mutex_lock(wq->mutex).
2512 * %true if @flush_color >= 0 and there's something to flush. %false
2515 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2516 int flush_color, int work_color)
2519 struct pool_workqueue *pwq;
2521 if (flush_color >= 0) {
2522 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2523 atomic_set(&wq->nr_pwqs_to_flush, 1);
2526 for_each_pwq(pwq, wq) {
2527 struct worker_pool *pool = pwq->pool;
2529 spin_lock_irq(&pool->lock);
2531 if (flush_color >= 0) {
2532 WARN_ON_ONCE(pwq->flush_color != -1);
2534 if (pwq->nr_in_flight[flush_color]) {
2535 pwq->flush_color = flush_color;
2536 atomic_inc(&wq->nr_pwqs_to_flush);
2541 if (work_color >= 0) {
2542 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2543 pwq->work_color = work_color;
2546 spin_unlock_irq(&pool->lock);
2549 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2550 complete(&wq->first_flusher->done);
2556 * flush_workqueue - ensure that any scheduled work has run to completion.
2557 * @wq: workqueue to flush
2559 * This function sleeps until all work items which were queued on entry
2560 * have finished execution, but it is not livelocked by new incoming ones.
2562 void flush_workqueue(struct workqueue_struct *wq)
2564 struct wq_flusher this_flusher = {
2565 .list = LIST_HEAD_INIT(this_flusher.list),
2567 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2571 lock_map_acquire(&wq->lockdep_map);
2572 lock_map_release(&wq->lockdep_map);
2574 mutex_lock(&wq->mutex);
2577 * Start-to-wait phase
2579 next_color = work_next_color(wq->work_color);
2581 if (next_color != wq->flush_color) {
2583 * Color space is not full. The current work_color
2584 * becomes our flush_color and work_color is advanced
2587 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2588 this_flusher.flush_color = wq->work_color;
2589 wq->work_color = next_color;
2591 if (!wq->first_flusher) {
2592 /* no flush in progress, become the first flusher */
2593 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2595 wq->first_flusher = &this_flusher;
2597 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2599 /* nothing to flush, done */
2600 wq->flush_color = next_color;
2601 wq->first_flusher = NULL;
2606 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2607 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2608 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2612 * Oops, color space is full, wait on overflow queue.
2613 * The next flush completion will assign us
2614 * flush_color and transfer to flusher_queue.
2616 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2619 mutex_unlock(&wq->mutex);
2621 wait_for_completion(&this_flusher.done);
2624 * Wake-up-and-cascade phase
2626 * First flushers are responsible for cascading flushes and
2627 * handling overflow. Non-first flushers can simply return.
2629 if (wq->first_flusher != &this_flusher)
2632 mutex_lock(&wq->mutex);
2634 /* we might have raced, check again with mutex held */
2635 if (wq->first_flusher != &this_flusher)
2638 wq->first_flusher = NULL;
2640 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2641 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2644 struct wq_flusher *next, *tmp;
2646 /* complete all the flushers sharing the current flush color */
2647 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2648 if (next->flush_color != wq->flush_color)
2650 list_del_init(&next->list);
2651 complete(&next->done);
2654 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2655 wq->flush_color != work_next_color(wq->work_color));
2657 /* this flush_color is finished, advance by one */
2658 wq->flush_color = work_next_color(wq->flush_color);
2660 /* one color has been freed, handle overflow queue */
2661 if (!list_empty(&wq->flusher_overflow)) {
2663 * Assign the same color to all overflowed
2664 * flushers, advance work_color and append to
2665 * flusher_queue. This is the start-to-wait
2666 * phase for these overflowed flushers.
2668 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2669 tmp->flush_color = wq->work_color;
2671 wq->work_color = work_next_color(wq->work_color);
2673 list_splice_tail_init(&wq->flusher_overflow,
2674 &wq->flusher_queue);
2675 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2678 if (list_empty(&wq->flusher_queue)) {
2679 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2684 * Need to flush more colors. Make the next flusher
2685 * the new first flusher and arm pwqs.
2687 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2688 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2690 list_del_init(&next->list);
2691 wq->first_flusher = next;
2693 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2697 * Meh... this color is already done, clear first
2698 * flusher and repeat cascading.
2700 wq->first_flusher = NULL;
2704 mutex_unlock(&wq->mutex);
2706 EXPORT_SYMBOL_GPL(flush_workqueue);
2709 * drain_workqueue - drain a workqueue
2710 * @wq: workqueue to drain
2712 * Wait until the workqueue becomes empty. While draining is in progress,
2713 * only chain queueing is allowed. IOW, only currently pending or running
2714 * work items on @wq can queue further work items on it. @wq is flushed
2715 * repeatedly until it becomes empty. The number of flushing is detemined
2716 * by the depth of chaining and should be relatively short. Whine if it
2719 void drain_workqueue(struct workqueue_struct *wq)
2721 unsigned int flush_cnt = 0;
2722 struct pool_workqueue *pwq;
2725 * __queue_work() needs to test whether there are drainers, is much
2726 * hotter than drain_workqueue() and already looks at @wq->flags.
2727 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2729 mutex_lock(&wq->mutex);
2730 if (!wq->nr_drainers++)
2731 wq->flags |= __WQ_DRAINING;
2732 mutex_unlock(&wq->mutex);
2734 flush_workqueue(wq);
2736 mutex_lock(&wq->mutex);
2738 for_each_pwq(pwq, wq) {
2741 spin_lock_irq(&pwq->pool->lock);
2742 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2743 spin_unlock_irq(&pwq->pool->lock);
2748 if (++flush_cnt == 10 ||
2749 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2750 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2751 wq->name, flush_cnt);
2753 mutex_unlock(&wq->mutex);
2757 if (!--wq->nr_drainers)
2758 wq->flags &= ~__WQ_DRAINING;
2759 mutex_unlock(&wq->mutex);
2761 EXPORT_SYMBOL_GPL(drain_workqueue);
2763 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2765 struct worker *worker = NULL;
2766 struct worker_pool *pool;
2767 struct pool_workqueue *pwq;
2771 local_irq_disable();
2772 pool = get_work_pool(work);
2778 spin_lock(&pool->lock);
2779 /* see the comment in try_to_grab_pending() with the same code */
2780 pwq = get_work_pwq(work);
2782 if (unlikely(pwq->pool != pool))
2785 worker = find_worker_executing_work(pool, work);
2788 pwq = worker->current_pwq;
2791 insert_wq_barrier(pwq, barr, work, worker);
2792 spin_unlock_irq(&pool->lock);
2795 * If @max_active is 1 or rescuer is in use, flushing another work
2796 * item on the same workqueue may lead to deadlock. Make sure the
2797 * flusher is not running on the same workqueue by verifying write
2800 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2801 lock_map_acquire(&pwq->wq->lockdep_map);
2803 lock_map_acquire_read(&pwq->wq->lockdep_map);
2804 lock_map_release(&pwq->wq->lockdep_map);
2808 spin_unlock_irq(&pool->lock);
2813 * flush_work - wait for a work to finish executing the last queueing instance
2814 * @work: the work to flush
2816 * Wait until @work has finished execution. @work is guaranteed to be idle
2817 * on return if it hasn't been requeued since flush started.
2820 * %true if flush_work() waited for the work to finish execution,
2821 * %false if it was already idle.
2823 bool flush_work(struct work_struct *work)
2825 struct wq_barrier barr;
2827 lock_map_acquire(&work->lockdep_map);
2828 lock_map_release(&work->lockdep_map);
2830 if (start_flush_work(work, &barr)) {
2831 wait_for_completion(&barr.done);
2832 destroy_work_on_stack(&barr.work);
2838 EXPORT_SYMBOL_GPL(flush_work);
2840 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2842 unsigned long flags;
2846 ret = try_to_grab_pending(work, is_dwork, &flags);
2848 * If someone else is canceling, wait for the same event it
2849 * would be waiting for before retrying.
2851 if (unlikely(ret == -ENOENT))
2853 } while (unlikely(ret < 0));
2855 /* tell other tasks trying to grab @work to back off */
2856 mark_work_canceling(work);
2857 local_irq_restore(flags);
2860 clear_work_data(work);
2865 * cancel_work_sync - cancel a work and wait for it to finish
2866 * @work: the work to cancel
2868 * Cancel @work and wait for its execution to finish. This function
2869 * can be used even if the work re-queues itself or migrates to
2870 * another workqueue. On return from this function, @work is
2871 * guaranteed to be not pending or executing on any CPU.
2873 * cancel_work_sync(&delayed_work->work) must not be used for
2874 * delayed_work's. Use cancel_delayed_work_sync() instead.
2876 * The caller must ensure that the workqueue on which @work was last
2877 * queued can't be destroyed before this function returns.
2880 * %true if @work was pending, %false otherwise.
2882 bool cancel_work_sync(struct work_struct *work)
2884 return __cancel_work_timer(work, false);
2886 EXPORT_SYMBOL_GPL(cancel_work_sync);
2889 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2890 * @dwork: the delayed work to flush
2892 * Delayed timer is cancelled and the pending work is queued for
2893 * immediate execution. Like flush_work(), this function only
2894 * considers the last queueing instance of @dwork.
2897 * %true if flush_work() waited for the work to finish execution,
2898 * %false if it was already idle.
2900 bool flush_delayed_work(struct delayed_work *dwork)
2902 local_irq_disable();
2903 if (del_timer_sync(&dwork->timer))
2904 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2906 return flush_work(&dwork->work);
2908 EXPORT_SYMBOL(flush_delayed_work);
2911 * cancel_delayed_work - cancel a delayed work
2912 * @dwork: delayed_work to cancel
2914 * Kill off a pending delayed_work.
2916 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2920 * The work callback function may still be running on return, unless
2921 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2922 * use cancel_delayed_work_sync() to wait on it.
2924 * This function is safe to call from any context including IRQ handler.
2926 bool cancel_delayed_work(struct delayed_work *dwork)
2928 unsigned long flags;
2932 ret = try_to_grab_pending(&dwork->work, true, &flags);
2933 } while (unlikely(ret == -EAGAIN));
2935 if (unlikely(ret < 0))
2938 set_work_pool_and_clear_pending(&dwork->work,
2939 get_work_pool_id(&dwork->work));
2940 local_irq_restore(flags);
2943 EXPORT_SYMBOL(cancel_delayed_work);
2946 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2947 * @dwork: the delayed work cancel
2949 * This is cancel_work_sync() for delayed works.
2952 * %true if @dwork was pending, %false otherwise.
2954 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2956 return __cancel_work_timer(&dwork->work, true);
2958 EXPORT_SYMBOL(cancel_delayed_work_sync);
2961 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2962 * @func: the function to call
2964 * schedule_on_each_cpu() executes @func on each online CPU using the
2965 * system workqueue and blocks until all CPUs have completed.
2966 * schedule_on_each_cpu() is very slow.
2969 * 0 on success, -errno on failure.
2971 int schedule_on_each_cpu(work_func_t func)
2974 struct work_struct __percpu *works;
2976 works = alloc_percpu(struct work_struct);
2982 for_each_online_cpu(cpu) {
2983 struct work_struct *work = per_cpu_ptr(works, cpu);
2985 INIT_WORK(work, func);
2986 schedule_work_on(cpu, work);
2989 for_each_online_cpu(cpu)
2990 flush_work(per_cpu_ptr(works, cpu));
2998 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3000 * Forces execution of the kernel-global workqueue and blocks until its
3003 * Think twice before calling this function! It's very easy to get into
3004 * trouble if you don't take great care. Either of the following situations
3005 * will lead to deadlock:
3007 * One of the work items currently on the workqueue needs to acquire
3008 * a lock held by your code or its caller.
3010 * Your code is running in the context of a work routine.
3012 * They will be detected by lockdep when they occur, but the first might not
3013 * occur very often. It depends on what work items are on the workqueue and
3014 * what locks they need, which you have no control over.
3016 * In most situations flushing the entire workqueue is overkill; you merely
3017 * need to know that a particular work item isn't queued and isn't running.
3018 * In such cases you should use cancel_delayed_work_sync() or
3019 * cancel_work_sync() instead.
3021 void flush_scheduled_work(void)
3023 flush_workqueue(system_wq);
3025 EXPORT_SYMBOL(flush_scheduled_work);
3028 * execute_in_process_context - reliably execute the routine with user context
3029 * @fn: the function to execute
3030 * @ew: guaranteed storage for the execute work structure (must
3031 * be available when the work executes)
3033 * Executes the function immediately if process context is available,
3034 * otherwise schedules the function for delayed execution.
3036 * Return: 0 - function was executed
3037 * 1 - function was scheduled for execution
3039 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3041 if (!in_interrupt()) {
3046 INIT_WORK(&ew->work, fn);
3047 schedule_work(&ew->work);
3051 EXPORT_SYMBOL_GPL(execute_in_process_context);
3055 * Workqueues with WQ_SYSFS flag set is visible to userland via
3056 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3057 * following attributes.
3059 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3060 * max_active RW int : maximum number of in-flight work items
3062 * Unbound workqueues have the following extra attributes.
3064 * id RO int : the associated pool ID
3065 * nice RW int : nice value of the workers
3066 * cpumask RW mask : bitmask of allowed CPUs for the workers
3069 struct workqueue_struct *wq;
3073 static struct workqueue_struct *dev_to_wq(struct device *dev)
3075 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3080 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3083 struct workqueue_struct *wq = dev_to_wq(dev);
3085 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3087 static DEVICE_ATTR_RO(per_cpu);
3089 static ssize_t max_active_show(struct device *dev,
3090 struct device_attribute *attr, char *buf)
3092 struct workqueue_struct *wq = dev_to_wq(dev);
3094 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3097 static ssize_t max_active_store(struct device *dev,
3098 struct device_attribute *attr, const char *buf,
3101 struct workqueue_struct *wq = dev_to_wq(dev);
3104 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3107 workqueue_set_max_active(wq, val);
3110 static DEVICE_ATTR_RW(max_active);
3112 static struct attribute *wq_sysfs_attrs[] = {
3113 &dev_attr_per_cpu.attr,
3114 &dev_attr_max_active.attr,
3117 ATTRIBUTE_GROUPS(wq_sysfs);
3119 static ssize_t wq_pool_ids_show(struct device *dev,
3120 struct device_attribute *attr, char *buf)
3122 struct workqueue_struct *wq = dev_to_wq(dev);
3123 const char *delim = "";
3124 int node, written = 0;
3126 rcu_read_lock_sched();
3127 for_each_node(node) {
3128 written += scnprintf(buf + written, PAGE_SIZE - written,
3129 "%s%d:%d", delim, node,
3130 unbound_pwq_by_node(wq, node)->pool->id);
3133 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3134 rcu_read_unlock_sched();
3139 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3142 struct workqueue_struct *wq = dev_to_wq(dev);
3145 mutex_lock(&wq->mutex);
3146 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3147 mutex_unlock(&wq->mutex);
3152 /* prepare workqueue_attrs for sysfs store operations */
3153 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3155 struct workqueue_attrs *attrs;
3157 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3161 mutex_lock(&wq->mutex);
3162 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3163 mutex_unlock(&wq->mutex);
3167 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3168 const char *buf, size_t count)
3170 struct workqueue_struct *wq = dev_to_wq(dev);
3171 struct workqueue_attrs *attrs;
3174 attrs = wq_sysfs_prep_attrs(wq);
3178 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3179 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
3180 ret = apply_workqueue_attrs(wq, attrs);
3184 free_workqueue_attrs(attrs);
3185 return ret ?: count;
3188 static ssize_t wq_cpumask_show(struct device *dev,
3189 struct device_attribute *attr, char *buf)
3191 struct workqueue_struct *wq = dev_to_wq(dev);
3194 mutex_lock(&wq->mutex);
3195 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3196 mutex_unlock(&wq->mutex);
3198 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3202 static ssize_t wq_cpumask_store(struct device *dev,
3203 struct device_attribute *attr,
3204 const char *buf, size_t count)
3206 struct workqueue_struct *wq = dev_to_wq(dev);
3207 struct workqueue_attrs *attrs;
3210 attrs = wq_sysfs_prep_attrs(wq);
3214 ret = cpumask_parse(buf, attrs->cpumask);
3216 ret = apply_workqueue_attrs(wq, attrs);
3218 free_workqueue_attrs(attrs);
3219 return ret ?: count;
3222 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3225 struct workqueue_struct *wq = dev_to_wq(dev);
3228 mutex_lock(&wq->mutex);
3229 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3230 !wq->unbound_attrs->no_numa);
3231 mutex_unlock(&wq->mutex);
3236 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3237 const char *buf, size_t count)
3239 struct workqueue_struct *wq = dev_to_wq(dev);
3240 struct workqueue_attrs *attrs;
3243 attrs = wq_sysfs_prep_attrs(wq);
3248 if (sscanf(buf, "%d", &v) == 1) {
3249 attrs->no_numa = !v;
3250 ret = apply_workqueue_attrs(wq, attrs);
3253 free_workqueue_attrs(attrs);
3254 return ret ?: count;
3257 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3258 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3259 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3260 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3261 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3265 static struct bus_type wq_subsys = {
3266 .name = "workqueue",
3267 .dev_groups = wq_sysfs_groups,
3270 static int __init wq_sysfs_init(void)
3272 return subsys_virtual_register(&wq_subsys, NULL);
3274 core_initcall(wq_sysfs_init);
3276 static void wq_device_release(struct device *dev)
3278 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3284 * workqueue_sysfs_register - make a workqueue visible in sysfs
3285 * @wq: the workqueue to register
3287 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3288 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3289 * which is the preferred method.
3291 * Workqueue user should use this function directly iff it wants to apply
3292 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3293 * apply_workqueue_attrs() may race against userland updating the
3296 * Return: 0 on success, -errno on failure.
3298 int workqueue_sysfs_register(struct workqueue_struct *wq)
3300 struct wq_device *wq_dev;
3304 * Adjusting max_active or creating new pwqs by applyting
3305 * attributes breaks ordering guarantee. Disallow exposing ordered
3308 if (WARN_ON(wq->flags & __WQ_ORDERED))
3311 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3316 wq_dev->dev.bus = &wq_subsys;
3317 wq_dev->dev.init_name = wq->name;
3318 wq_dev->dev.release = wq_device_release;
3321 * unbound_attrs are created separately. Suppress uevent until
3322 * everything is ready.
3324 dev_set_uevent_suppress(&wq_dev->dev, true);
3326 ret = device_register(&wq_dev->dev);
3333 if (wq->flags & WQ_UNBOUND) {
3334 struct device_attribute *attr;
3336 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3337 ret = device_create_file(&wq_dev->dev, attr);
3339 device_unregister(&wq_dev->dev);
3346 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3351 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3352 * @wq: the workqueue to unregister
3354 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3356 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3358 struct wq_device *wq_dev = wq->wq_dev;
3364 device_unregister(&wq_dev->dev);
3366 #else /* CONFIG_SYSFS */
3367 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3368 #endif /* CONFIG_SYSFS */
3371 * free_workqueue_attrs - free a workqueue_attrs
3372 * @attrs: workqueue_attrs to free
3374 * Undo alloc_workqueue_attrs().
3376 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3379 free_cpumask_var(attrs->cpumask);
3385 * alloc_workqueue_attrs - allocate a workqueue_attrs
3386 * @gfp_mask: allocation mask to use
3388 * Allocate a new workqueue_attrs, initialize with default settings and
3391 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3393 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3395 struct workqueue_attrs *attrs;
3397 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3400 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3403 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3406 free_workqueue_attrs(attrs);
3410 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3411 const struct workqueue_attrs *from)
3413 to->nice = from->nice;
3414 cpumask_copy(to->cpumask, from->cpumask);
3416 * Unlike hash and equality test, this function doesn't ignore
3417 * ->no_numa as it is used for both pool and wq attrs. Instead,
3418 * get_unbound_pool() explicitly clears ->no_numa after copying.
3420 to->no_numa = from->no_numa;
3423 /* hash value of the content of @attr */
3424 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3428 hash = jhash_1word(attrs->nice, hash);
3429 hash = jhash(cpumask_bits(attrs->cpumask),
3430 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3434 /* content equality test */
3435 static bool wqattrs_equal(const struct workqueue_attrs *a,
3436 const struct workqueue_attrs *b)
3438 if (a->nice != b->nice)
3440 if (!cpumask_equal(a->cpumask, b->cpumask))
3446 * init_worker_pool - initialize a newly zalloc'd worker_pool
3447 * @pool: worker_pool to initialize
3449 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3451 * Return: 0 on success, -errno on failure. Even on failure, all fields
3452 * inside @pool proper are initialized and put_unbound_pool() can be called
3453 * on @pool safely to release it.
3455 static int init_worker_pool(struct worker_pool *pool)
3457 spin_lock_init(&pool->lock);
3460 pool->node = NUMA_NO_NODE;
3461 pool->flags |= POOL_DISASSOCIATED;
3462 INIT_LIST_HEAD(&pool->worklist);
3463 INIT_LIST_HEAD(&pool->idle_list);
3464 hash_init(pool->busy_hash);
3466 init_timer_deferrable(&pool->idle_timer);
3467 pool->idle_timer.function = idle_worker_timeout;
3468 pool->idle_timer.data = (unsigned long)pool;
3470 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3471 (unsigned long)pool);
3473 mutex_init(&pool->manager_arb);
3474 mutex_init(&pool->attach_mutex);
3475 INIT_LIST_HEAD(&pool->workers);
3477 ida_init(&pool->worker_ida);
3478 INIT_HLIST_NODE(&pool->hash_node);
3481 /* shouldn't fail above this point */
3482 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3488 static void rcu_free_pool(struct rcu_head *rcu)
3490 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3492 ida_destroy(&pool->worker_ida);
3493 free_workqueue_attrs(pool->attrs);
3498 * put_unbound_pool - put a worker_pool
3499 * @pool: worker_pool to put
3501 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3502 * safe manner. get_unbound_pool() calls this function on its failure path
3503 * and this function should be able to release pools which went through,
3504 * successfully or not, init_worker_pool().
3506 * Should be called with wq_pool_mutex held.
3508 static void put_unbound_pool(struct worker_pool *pool)
3510 DECLARE_COMPLETION_ONSTACK(detach_completion);
3511 struct worker *worker;
3513 lockdep_assert_held(&wq_pool_mutex);
3519 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3520 WARN_ON(!list_empty(&pool->worklist)))
3523 /* release id and unhash */
3525 idr_remove(&worker_pool_idr, pool->id);
3526 hash_del(&pool->hash_node);
3529 * Become the manager and destroy all workers. Grabbing
3530 * manager_arb prevents @pool's workers from blocking on
3533 mutex_lock(&pool->manager_arb);
3535 spin_lock_irq(&pool->lock);
3536 while ((worker = first_worker(pool)))
3537 destroy_worker(worker);
3538 WARN_ON(pool->nr_workers || pool->nr_idle);
3539 spin_unlock_irq(&pool->lock);
3541 mutex_lock(&pool->attach_mutex);
3542 if (!list_empty(&pool->workers))
3543 pool->detach_completion = &detach_completion;
3544 mutex_unlock(&pool->attach_mutex);
3546 if (pool->detach_completion)
3547 wait_for_completion(pool->detach_completion);
3549 mutex_unlock(&pool->manager_arb);
3551 /* shut down the timers */
3552 del_timer_sync(&pool->idle_timer);
3553 del_timer_sync(&pool->mayday_timer);
3555 /* sched-RCU protected to allow dereferences from get_work_pool() */
3556 call_rcu_sched(&pool->rcu, rcu_free_pool);
3560 * get_unbound_pool - get a worker_pool with the specified attributes
3561 * @attrs: the attributes of the worker_pool to get
3563 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3564 * reference count and return it. If there already is a matching
3565 * worker_pool, it will be used; otherwise, this function attempts to
3568 * Should be called with wq_pool_mutex held.
3570 * Return: On success, a worker_pool with the same attributes as @attrs.
3571 * On failure, %NULL.
3573 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3575 u32 hash = wqattrs_hash(attrs);
3576 struct worker_pool *pool;
3579 lockdep_assert_held(&wq_pool_mutex);
3581 /* do we already have a matching pool? */
3582 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3583 if (wqattrs_equal(pool->attrs, attrs)) {
3589 /* nope, create a new one */
3590 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3591 if (!pool || init_worker_pool(pool) < 0)
3594 if (workqueue_freezing)
3595 pool->flags |= POOL_FREEZING;
3597 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3598 copy_workqueue_attrs(pool->attrs, attrs);
3601 * no_numa isn't a worker_pool attribute, always clear it. See
3602 * 'struct workqueue_attrs' comments for detail.
3604 pool->attrs->no_numa = false;
3606 /* if cpumask is contained inside a NUMA node, we belong to that node */
3607 if (wq_numa_enabled) {
3608 for_each_node(node) {
3609 if (cpumask_subset(pool->attrs->cpumask,
3610 wq_numa_possible_cpumask[node])) {
3617 if (worker_pool_assign_id(pool) < 0)
3620 /* create and start the initial worker */
3621 if (create_and_start_worker(pool) < 0)
3625 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3630 put_unbound_pool(pool);
3634 static void rcu_free_pwq(struct rcu_head *rcu)
3636 kmem_cache_free(pwq_cache,
3637 container_of(rcu, struct pool_workqueue, rcu));
3641 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3642 * and needs to be destroyed.
3644 static void pwq_unbound_release_workfn(struct work_struct *work)
3646 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3647 unbound_release_work);
3648 struct workqueue_struct *wq = pwq->wq;
3649 struct worker_pool *pool = pwq->pool;
3652 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3656 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3657 * necessary on release but do it anyway. It's easier to verify
3658 * and consistent with the linking path.
3660 mutex_lock(&wq->mutex);
3661 list_del_rcu(&pwq->pwqs_node);
3662 is_last = list_empty(&wq->pwqs);
3663 mutex_unlock(&wq->mutex);
3665 mutex_lock(&wq_pool_mutex);
3666 put_unbound_pool(pool);
3667 mutex_unlock(&wq_pool_mutex);
3669 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3672 * If we're the last pwq going away, @wq is already dead and no one
3673 * is gonna access it anymore. Free it.
3676 free_workqueue_attrs(wq->unbound_attrs);
3682 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3683 * @pwq: target pool_workqueue
3685 * If @pwq isn't freezing, set @pwq->max_active to the associated
3686 * workqueue's saved_max_active and activate delayed work items
3687 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3689 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3691 struct workqueue_struct *wq = pwq->wq;
3692 bool freezable = wq->flags & WQ_FREEZABLE;
3694 /* for @wq->saved_max_active */
3695 lockdep_assert_held(&wq->mutex);
3697 /* fast exit for non-freezable wqs */
3698 if (!freezable && pwq->max_active == wq->saved_max_active)
3701 spin_lock_irq(&pwq->pool->lock);
3703 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3704 pwq->max_active = wq->saved_max_active;
3706 while (!list_empty(&pwq->delayed_works) &&
3707 pwq->nr_active < pwq->max_active)
3708 pwq_activate_first_delayed(pwq);
3711 * Need to kick a worker after thawed or an unbound wq's
3712 * max_active is bumped. It's a slow path. Do it always.
3714 wake_up_worker(pwq->pool);
3716 pwq->max_active = 0;
3719 spin_unlock_irq(&pwq->pool->lock);
3722 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3723 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3724 struct worker_pool *pool)
3726 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3728 memset(pwq, 0, sizeof(*pwq));
3732 pwq->flush_color = -1;
3734 INIT_LIST_HEAD(&pwq->delayed_works);
3735 INIT_LIST_HEAD(&pwq->pwqs_node);
3736 INIT_LIST_HEAD(&pwq->mayday_node);
3737 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3740 /* sync @pwq with the current state of its associated wq and link it */
3741 static void link_pwq(struct pool_workqueue *pwq)
3743 struct workqueue_struct *wq = pwq->wq;
3745 lockdep_assert_held(&wq->mutex);
3747 /* may be called multiple times, ignore if already linked */
3748 if (!list_empty(&pwq->pwqs_node))
3752 * Set the matching work_color. This is synchronized with
3753 * wq->mutex to avoid confusing flush_workqueue().
3755 pwq->work_color = wq->work_color;
3757 /* sync max_active to the current setting */
3758 pwq_adjust_max_active(pwq);
3761 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3764 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3765 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3766 const struct workqueue_attrs *attrs)
3768 struct worker_pool *pool;
3769 struct pool_workqueue *pwq;
3771 lockdep_assert_held(&wq_pool_mutex);
3773 pool = get_unbound_pool(attrs);
3777 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3779 put_unbound_pool(pool);
3783 init_pwq(pwq, wq, pool);
3787 /* undo alloc_unbound_pwq(), used only in the error path */
3788 static void free_unbound_pwq(struct pool_workqueue *pwq)
3790 lockdep_assert_held(&wq_pool_mutex);
3793 put_unbound_pool(pwq->pool);
3794 kmem_cache_free(pwq_cache, pwq);
3799 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3800 * @attrs: the wq_attrs of interest
3801 * @node: the target NUMA node
3802 * @cpu_going_down: if >= 0, the CPU to consider as offline
3803 * @cpumask: outarg, the resulting cpumask
3805 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3806 * @cpu_going_down is >= 0, that cpu is considered offline during
3807 * calculation. The result is stored in @cpumask.
3809 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3810 * enabled and @node has online CPUs requested by @attrs, the returned
3811 * cpumask is the intersection of the possible CPUs of @node and
3814 * The caller is responsible for ensuring that the cpumask of @node stays
3817 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3820 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3821 int cpu_going_down, cpumask_t *cpumask)
3823 if (!wq_numa_enabled || attrs->no_numa)
3826 /* does @node have any online CPUs @attrs wants? */
3827 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3828 if (cpu_going_down >= 0)
3829 cpumask_clear_cpu(cpu_going_down, cpumask);
3831 if (cpumask_empty(cpumask))
3834 /* yeap, return possible CPUs in @node that @attrs wants */
3835 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3836 return !cpumask_equal(cpumask, attrs->cpumask);
3839 cpumask_copy(cpumask, attrs->cpumask);
3843 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3844 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3846 struct pool_workqueue *pwq)
3848 struct pool_workqueue *old_pwq;
3850 lockdep_assert_held(&wq->mutex);
3852 /* link_pwq() can handle duplicate calls */
3855 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3856 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3861 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3862 * @wq: the target workqueue
3863 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3865 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3866 * machines, this function maps a separate pwq to each NUMA node with
3867 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3868 * NUMA node it was issued on. Older pwqs are released as in-flight work
3869 * items finish. Note that a work item which repeatedly requeues itself
3870 * back-to-back will stay on its current pwq.
3872 * Performs GFP_KERNEL allocations.
3874 * Return: 0 on success and -errno on failure.
3876 int apply_workqueue_attrs(struct workqueue_struct *wq,
3877 const struct workqueue_attrs *attrs)
3879 struct workqueue_attrs *new_attrs, *tmp_attrs;
3880 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3883 /* only unbound workqueues can change attributes */
3884 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3887 /* creating multiple pwqs breaks ordering guarantee */
3888 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3891 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3892 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3893 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3894 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3897 /* make a copy of @attrs and sanitize it */
3898 copy_workqueue_attrs(new_attrs, attrs);
3899 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3902 * We may create multiple pwqs with differing cpumasks. Make a
3903 * copy of @new_attrs which will be modified and used to obtain
3906 copy_workqueue_attrs(tmp_attrs, new_attrs);
3909 * CPUs should stay stable across pwq creations and installations.
3910 * Pin CPUs, determine the target cpumask for each node and create
3915 mutex_lock(&wq_pool_mutex);
3918 * If something goes wrong during CPU up/down, we'll fall back to
3919 * the default pwq covering whole @attrs->cpumask. Always create
3920 * it even if we don't use it immediately.
3922 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3926 for_each_node(node) {
3927 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3928 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3933 pwq_tbl[node] = dfl_pwq;
3937 mutex_unlock(&wq_pool_mutex);
3939 /* all pwqs have been created successfully, let's install'em */
3940 mutex_lock(&wq->mutex);
3942 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3944 /* save the previous pwq and install the new one */
3946 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3948 /* @dfl_pwq might not have been used, ensure it's linked */
3950 swap(wq->dfl_pwq, dfl_pwq);
3952 mutex_unlock(&wq->mutex);
3954 /* put the old pwqs */
3956 put_pwq_unlocked(pwq_tbl[node]);
3957 put_pwq_unlocked(dfl_pwq);
3963 free_workqueue_attrs(tmp_attrs);
3964 free_workqueue_attrs(new_attrs);
3969 free_unbound_pwq(dfl_pwq);
3971 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3972 free_unbound_pwq(pwq_tbl[node]);
3973 mutex_unlock(&wq_pool_mutex);
3981 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3982 * @wq: the target workqueue
3983 * @cpu: the CPU coming up or going down
3984 * @online: whether @cpu is coming up or going down
3986 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3987 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3990 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3991 * falls back to @wq->dfl_pwq which may not be optimal but is always
3994 * Note that when the last allowed CPU of a NUMA node goes offline for a
3995 * workqueue with a cpumask spanning multiple nodes, the workers which were
3996 * already executing the work items for the workqueue will lose their CPU
3997 * affinity and may execute on any CPU. This is similar to how per-cpu
3998 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3999 * affinity, it's the user's responsibility to flush the work item from
4002 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4005 int node = cpu_to_node(cpu);
4006 int cpu_off = online ? -1 : cpu;
4007 struct pool_workqueue *old_pwq = NULL, *pwq;
4008 struct workqueue_attrs *target_attrs;
4011 lockdep_assert_held(&wq_pool_mutex);
4013 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4017 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4018 * Let's use a preallocated one. The following buf is protected by
4019 * CPU hotplug exclusion.
4021 target_attrs = wq_update_unbound_numa_attrs_buf;
4022 cpumask = target_attrs->cpumask;
4024 mutex_lock(&wq->mutex);
4025 if (wq->unbound_attrs->no_numa)
4028 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4029 pwq = unbound_pwq_by_node(wq, node);
4032 * Let's determine what needs to be done. If the target cpumask is
4033 * different from wq's, we need to compare it to @pwq's and create
4034 * a new one if they don't match. If the target cpumask equals
4035 * wq's, the default pwq should be used.
4037 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4038 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4044 mutex_unlock(&wq->mutex);
4046 /* create a new pwq */
4047 pwq = alloc_unbound_pwq(wq, target_attrs);
4049 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4051 mutex_lock(&wq->mutex);
4056 * Install the new pwq. As this function is called only from CPU
4057 * hotplug callbacks and applying a new attrs is wrapped with
4058 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4061 mutex_lock(&wq->mutex);
4062 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4066 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4067 get_pwq(wq->dfl_pwq);
4068 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4069 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4071 mutex_unlock(&wq->mutex);
4072 put_pwq_unlocked(old_pwq);
4075 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4077 bool highpri = wq->flags & WQ_HIGHPRI;
4080 if (!(wq->flags & WQ_UNBOUND)) {
4081 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4085 for_each_possible_cpu(cpu) {
4086 struct pool_workqueue *pwq =
4087 per_cpu_ptr(wq->cpu_pwqs, cpu);
4088 struct worker_pool *cpu_pools =
4089 per_cpu(cpu_worker_pools, cpu);
4091 init_pwq(pwq, wq, &cpu_pools[highpri]);
4093 mutex_lock(&wq->mutex);
4095 mutex_unlock(&wq->mutex);
4098 } else if (wq->flags & __WQ_ORDERED) {
4099 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4100 /* there should only be single pwq for ordering guarantee */
4101 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4102 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4103 "ordering guarantee broken for workqueue %s\n", wq->name);
4106 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4110 static int wq_clamp_max_active(int max_active, unsigned int flags,
4113 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4115 if (max_active < 1 || max_active > lim)
4116 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4117 max_active, name, 1, lim);
4119 return clamp_val(max_active, 1, lim);
4122 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4125 struct lock_class_key *key,
4126 const char *lock_name, ...)
4128 size_t tbl_size = 0;
4130 struct workqueue_struct *wq;
4131 struct pool_workqueue *pwq;
4133 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4134 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4135 flags |= WQ_UNBOUND;
4137 /* allocate wq and format name */
4138 if (flags & WQ_UNBOUND)
4139 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4141 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4145 if (flags & WQ_UNBOUND) {
4146 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4147 if (!wq->unbound_attrs)
4151 va_start(args, lock_name);
4152 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4155 max_active = max_active ?: WQ_DFL_ACTIVE;
4156 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4160 wq->saved_max_active = max_active;
4161 mutex_init(&wq->mutex);
4162 atomic_set(&wq->nr_pwqs_to_flush, 0);
4163 INIT_LIST_HEAD(&wq->pwqs);
4164 INIT_LIST_HEAD(&wq->flusher_queue);
4165 INIT_LIST_HEAD(&wq->flusher_overflow);
4166 INIT_LIST_HEAD(&wq->maydays);
4168 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4169 INIT_LIST_HEAD(&wq->list);
4171 if (alloc_and_link_pwqs(wq) < 0)
4175 * Workqueues which may be used during memory reclaim should
4176 * have a rescuer to guarantee forward progress.
4178 if (flags & WQ_MEM_RECLAIM) {
4179 struct worker *rescuer;
4181 rescuer = alloc_worker();
4185 rescuer->rescue_wq = wq;
4186 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4188 if (IS_ERR(rescuer->task)) {
4193 wq->rescuer = rescuer;
4194 rescuer->task->flags |= PF_NO_SETAFFINITY;
4195 wake_up_process(rescuer->task);
4198 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4202 * wq_pool_mutex protects global freeze state and workqueues list.
4203 * Grab it, adjust max_active and add the new @wq to workqueues
4206 mutex_lock(&wq_pool_mutex);
4208 mutex_lock(&wq->mutex);
4209 for_each_pwq(pwq, wq)
4210 pwq_adjust_max_active(pwq);
4211 mutex_unlock(&wq->mutex);
4213 list_add(&wq->list, &workqueues);
4215 mutex_unlock(&wq_pool_mutex);
4220 free_workqueue_attrs(wq->unbound_attrs);
4224 destroy_workqueue(wq);
4227 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4230 * destroy_workqueue - safely terminate a workqueue
4231 * @wq: target workqueue
4233 * Safely destroy a workqueue. All work currently pending will be done first.
4235 void destroy_workqueue(struct workqueue_struct *wq)
4237 struct pool_workqueue *pwq;
4240 /* drain it before proceeding with destruction */
4241 drain_workqueue(wq);
4244 mutex_lock(&wq->mutex);
4245 for_each_pwq(pwq, wq) {
4248 for (i = 0; i < WORK_NR_COLORS; i++) {
4249 if (WARN_ON(pwq->nr_in_flight[i])) {
4250 mutex_unlock(&wq->mutex);
4255 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4256 WARN_ON(pwq->nr_active) ||
4257 WARN_ON(!list_empty(&pwq->delayed_works))) {
4258 mutex_unlock(&wq->mutex);
4262 mutex_unlock(&wq->mutex);
4265 * wq list is used to freeze wq, remove from list after
4266 * flushing is complete in case freeze races us.
4268 mutex_lock(&wq_pool_mutex);
4269 list_del_init(&wq->list);
4270 mutex_unlock(&wq_pool_mutex);
4272 workqueue_sysfs_unregister(wq);
4275 kthread_stop(wq->rescuer->task);
4280 if (!(wq->flags & WQ_UNBOUND)) {
4282 * The base ref is never dropped on per-cpu pwqs. Directly
4283 * free the pwqs and wq.
4285 free_percpu(wq->cpu_pwqs);
4289 * We're the sole accessor of @wq at this point. Directly
4290 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4291 * @wq will be freed when the last pwq is released.
4293 for_each_node(node) {
4294 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4295 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4296 put_pwq_unlocked(pwq);
4300 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4301 * put. Don't access it afterwards.
4305 put_pwq_unlocked(pwq);
4308 EXPORT_SYMBOL_GPL(destroy_workqueue);
4311 * workqueue_set_max_active - adjust max_active of a workqueue
4312 * @wq: target workqueue
4313 * @max_active: new max_active value.
4315 * Set max_active of @wq to @max_active.
4318 * Don't call from IRQ context.
4320 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4322 struct pool_workqueue *pwq;
4324 /* disallow meddling with max_active for ordered workqueues */
4325 if (WARN_ON(wq->flags & __WQ_ORDERED))
4328 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4330 mutex_lock(&wq->mutex);
4332 wq->saved_max_active = max_active;
4334 for_each_pwq(pwq, wq)
4335 pwq_adjust_max_active(pwq);
4337 mutex_unlock(&wq->mutex);
4339 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4342 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4344 * Determine whether %current is a workqueue rescuer. Can be used from
4345 * work functions to determine whether it's being run off the rescuer task.
4347 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4349 bool current_is_workqueue_rescuer(void)
4351 struct worker *worker = current_wq_worker();
4353 return worker && worker->rescue_wq;
4357 * workqueue_congested - test whether a workqueue is congested
4358 * @cpu: CPU in question
4359 * @wq: target workqueue
4361 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4362 * no synchronization around this function and the test result is
4363 * unreliable and only useful as advisory hints or for debugging.
4365 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4366 * Note that both per-cpu and unbound workqueues may be associated with
4367 * multiple pool_workqueues which have separate congested states. A
4368 * workqueue being congested on one CPU doesn't mean the workqueue is also
4369 * contested on other CPUs / NUMA nodes.
4372 * %true if congested, %false otherwise.
4374 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4376 struct pool_workqueue *pwq;
4379 rcu_read_lock_sched();
4381 if (cpu == WORK_CPU_UNBOUND)
4382 cpu = smp_processor_id();
4384 if (!(wq->flags & WQ_UNBOUND))
4385 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4387 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4389 ret = !list_empty(&pwq->delayed_works);
4390 rcu_read_unlock_sched();
4394 EXPORT_SYMBOL_GPL(workqueue_congested);
4397 * work_busy - test whether a work is currently pending or running
4398 * @work: the work to be tested
4400 * Test whether @work is currently pending or running. There is no
4401 * synchronization around this function and the test result is
4402 * unreliable and only useful as advisory hints or for debugging.
4405 * OR'd bitmask of WORK_BUSY_* bits.
4407 unsigned int work_busy(struct work_struct *work)
4409 struct worker_pool *pool;
4410 unsigned long flags;
4411 unsigned int ret = 0;
4413 if (work_pending(work))
4414 ret |= WORK_BUSY_PENDING;
4416 local_irq_save(flags);
4417 pool = get_work_pool(work);
4419 spin_lock(&pool->lock);
4420 if (find_worker_executing_work(pool, work))
4421 ret |= WORK_BUSY_RUNNING;
4422 spin_unlock(&pool->lock);
4424 local_irq_restore(flags);
4428 EXPORT_SYMBOL_GPL(work_busy);
4431 * set_worker_desc - set description for the current work item
4432 * @fmt: printf-style format string
4433 * @...: arguments for the format string
4435 * This function can be called by a running work function to describe what
4436 * the work item is about. If the worker task gets dumped, this
4437 * information will be printed out together to help debugging. The
4438 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4440 void set_worker_desc(const char *fmt, ...)
4442 struct worker *worker = current_wq_worker();
4446 va_start(args, fmt);
4447 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4449 worker->desc_valid = true;
4454 * print_worker_info - print out worker information and description
4455 * @log_lvl: the log level to use when printing
4456 * @task: target task
4458 * If @task is a worker and currently executing a work item, print out the
4459 * name of the workqueue being serviced and worker description set with
4460 * set_worker_desc() by the currently executing work item.
4462 * This function can be safely called on any task as long as the
4463 * task_struct itself is accessible. While safe, this function isn't
4464 * synchronized and may print out mixups or garbages of limited length.
4466 void print_worker_info(const char *log_lvl, struct task_struct *task)
4468 work_func_t *fn = NULL;
4469 char name[WQ_NAME_LEN] = { };
4470 char desc[WORKER_DESC_LEN] = { };
4471 struct pool_workqueue *pwq = NULL;
4472 struct workqueue_struct *wq = NULL;
4473 bool desc_valid = false;
4474 struct worker *worker;
4476 if (!(task->flags & PF_WQ_WORKER))
4480 * This function is called without any synchronization and @task
4481 * could be in any state. Be careful with dereferences.
4483 worker = probe_kthread_data(task);
4486 * Carefully copy the associated workqueue's workfn and name. Keep
4487 * the original last '\0' in case the original contains garbage.
4489 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4490 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4491 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4492 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4494 /* copy worker description */
4495 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4497 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4499 if (fn || name[0] || desc[0]) {
4500 pr_info("%sWorkqueue: %s %pf", log_lvl, name, fn);
4502 pr_cont(" (%s)", desc);
4510 * There are two challenges in supporting CPU hotplug. Firstly, there
4511 * are a lot of assumptions on strong associations among work, pwq and
4512 * pool which make migrating pending and scheduled works very
4513 * difficult to implement without impacting hot paths. Secondly,
4514 * worker pools serve mix of short, long and very long running works making
4515 * blocked draining impractical.
4517 * This is solved by allowing the pools to be disassociated from the CPU
4518 * running as an unbound one and allowing it to be reattached later if the
4519 * cpu comes back online.
4522 static void wq_unbind_fn(struct work_struct *work)
4524 int cpu = smp_processor_id();
4525 struct worker_pool *pool;
4526 struct worker *worker;
4528 for_each_cpu_worker_pool(pool, cpu) {
4529 WARN_ON_ONCE(cpu != smp_processor_id());
4531 mutex_lock(&pool->attach_mutex);
4532 spin_lock_irq(&pool->lock);
4535 * We've blocked all attach/detach operations. Make all workers
4536 * unbound and set DISASSOCIATED. Before this, all workers
4537 * except for the ones which are still executing works from
4538 * before the last CPU down must be on the cpu. After
4539 * this, they may become diasporas.
4541 for_each_pool_worker(worker, pool)
4542 worker->flags |= WORKER_UNBOUND;
4544 pool->flags |= POOL_DISASSOCIATED;
4546 spin_unlock_irq(&pool->lock);
4547 mutex_unlock(&pool->attach_mutex);
4550 * Call schedule() so that we cross rq->lock and thus can
4551 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4552 * This is necessary as scheduler callbacks may be invoked
4558 * Sched callbacks are disabled now. Zap nr_running.
4559 * After this, nr_running stays zero and need_more_worker()
4560 * and keep_working() are always true as long as the
4561 * worklist is not empty. This pool now behaves as an
4562 * unbound (in terms of concurrency management) pool which
4563 * are served by workers tied to the pool.
4565 atomic_set(&pool->nr_running, 0);
4568 * With concurrency management just turned off, a busy
4569 * worker blocking could lead to lengthy stalls. Kick off
4570 * unbound chain execution of currently pending work items.
4572 spin_lock_irq(&pool->lock);
4573 wake_up_worker(pool);
4574 spin_unlock_irq(&pool->lock);
4579 * rebind_workers - rebind all workers of a pool to the associated CPU
4580 * @pool: pool of interest
4582 * @pool->cpu is coming online. Rebind all workers to the CPU.
4584 static void rebind_workers(struct worker_pool *pool)
4586 struct worker *worker;
4588 lockdep_assert_held(&pool->attach_mutex);
4591 * Restore CPU affinity of all workers. As all idle workers should
4592 * be on the run-queue of the associated CPU before any local
4593 * wake-ups for concurrency management happen, restore CPU affinty
4594 * of all workers first and then clear UNBOUND. As we're called
4595 * from CPU_ONLINE, the following shouldn't fail.
4597 for_each_pool_worker(worker, pool)
4598 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4599 pool->attrs->cpumask) < 0);
4601 spin_lock_irq(&pool->lock);
4603 for_each_pool_worker(worker, pool) {
4604 unsigned int worker_flags = worker->flags;
4607 * A bound idle worker should actually be on the runqueue
4608 * of the associated CPU for local wake-ups targeting it to
4609 * work. Kick all idle workers so that they migrate to the
4610 * associated CPU. Doing this in the same loop as
4611 * replacing UNBOUND with REBOUND is safe as no worker will
4612 * be bound before @pool->lock is released.
4614 if (worker_flags & WORKER_IDLE)
4615 wake_up_process(worker->task);
4618 * We want to clear UNBOUND but can't directly call
4619 * worker_clr_flags() or adjust nr_running. Atomically
4620 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4621 * @worker will clear REBOUND using worker_clr_flags() when
4622 * it initiates the next execution cycle thus restoring
4623 * concurrency management. Note that when or whether
4624 * @worker clears REBOUND doesn't affect correctness.
4626 * ACCESS_ONCE() is necessary because @worker->flags may be
4627 * tested without holding any lock in
4628 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4629 * fail incorrectly leading to premature concurrency
4630 * management operations.
4632 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4633 worker_flags |= WORKER_REBOUND;
4634 worker_flags &= ~WORKER_UNBOUND;
4635 ACCESS_ONCE(worker->flags) = worker_flags;
4638 spin_unlock_irq(&pool->lock);
4642 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4643 * @pool: unbound pool of interest
4644 * @cpu: the CPU which is coming up
4646 * An unbound pool may end up with a cpumask which doesn't have any online
4647 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4648 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4649 * online CPU before, cpus_allowed of all its workers should be restored.
4651 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4653 static cpumask_t cpumask;
4654 struct worker *worker;
4656 lockdep_assert_held(&pool->attach_mutex);
4658 /* is @cpu allowed for @pool? */
4659 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4662 /* is @cpu the only online CPU? */
4663 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4664 if (cpumask_weight(&cpumask) != 1)
4667 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4668 for_each_pool_worker(worker, pool)
4669 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4670 pool->attrs->cpumask) < 0);
4674 * Workqueues should be brought up before normal priority CPU notifiers.
4675 * This will be registered high priority CPU notifier.
4677 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4678 unsigned long action,
4681 int cpu = (unsigned long)hcpu;
4682 struct worker_pool *pool;
4683 struct workqueue_struct *wq;
4686 switch (action & ~CPU_TASKS_FROZEN) {
4687 case CPU_UP_PREPARE:
4688 for_each_cpu_worker_pool(pool, cpu) {
4689 if (pool->nr_workers)
4691 if (create_and_start_worker(pool) < 0)
4696 case CPU_DOWN_FAILED:
4698 mutex_lock(&wq_pool_mutex);
4700 for_each_pool(pool, pi) {
4701 mutex_lock(&pool->attach_mutex);
4703 if (pool->cpu == cpu) {
4704 spin_lock_irq(&pool->lock);
4705 pool->flags &= ~POOL_DISASSOCIATED;
4706 spin_unlock_irq(&pool->lock);
4708 rebind_workers(pool);
4709 } else if (pool->cpu < 0) {
4710 restore_unbound_workers_cpumask(pool, cpu);
4713 mutex_unlock(&pool->attach_mutex);
4716 /* update NUMA affinity of unbound workqueues */
4717 list_for_each_entry(wq, &workqueues, list)
4718 wq_update_unbound_numa(wq, cpu, true);
4720 mutex_unlock(&wq_pool_mutex);
4727 * Workqueues should be brought down after normal priority CPU notifiers.
4728 * This will be registered as low priority CPU notifier.
4730 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4731 unsigned long action,
4734 int cpu = (unsigned long)hcpu;
4735 struct work_struct unbind_work;
4736 struct workqueue_struct *wq;
4738 switch (action & ~CPU_TASKS_FROZEN) {
4739 case CPU_DOWN_PREPARE:
4740 /* unbinding per-cpu workers should happen on the local CPU */
4741 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4742 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4744 /* update NUMA affinity of unbound workqueues */
4745 mutex_lock(&wq_pool_mutex);
4746 list_for_each_entry(wq, &workqueues, list)
4747 wq_update_unbound_numa(wq, cpu, false);
4748 mutex_unlock(&wq_pool_mutex);
4750 /* wait for per-cpu unbinding to finish */
4751 flush_work(&unbind_work);
4752 destroy_work_on_stack(&unbind_work);
4760 struct work_for_cpu {
4761 struct work_struct work;
4767 static void work_for_cpu_fn(struct work_struct *work)
4769 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4771 wfc->ret = wfc->fn(wfc->arg);
4775 * work_on_cpu - run a function in user context on a particular cpu
4776 * @cpu: the cpu to run on
4777 * @fn: the function to run
4778 * @arg: the function arg
4780 * It is up to the caller to ensure that the cpu doesn't go offline.
4781 * The caller must not hold any locks which would prevent @fn from completing.
4783 * Return: The value @fn returns.
4785 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4787 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4789 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4790 schedule_work_on(cpu, &wfc.work);
4791 flush_work(&wfc.work);
4792 destroy_work_on_stack(&wfc.work);
4795 EXPORT_SYMBOL_GPL(work_on_cpu);
4796 #endif /* CONFIG_SMP */
4798 #ifdef CONFIG_FREEZER
4801 * freeze_workqueues_begin - begin freezing workqueues
4803 * Start freezing workqueues. After this function returns, all freezable
4804 * workqueues will queue new works to their delayed_works list instead of
4808 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4810 void freeze_workqueues_begin(void)
4812 struct worker_pool *pool;
4813 struct workqueue_struct *wq;
4814 struct pool_workqueue *pwq;
4817 mutex_lock(&wq_pool_mutex);
4819 WARN_ON_ONCE(workqueue_freezing);
4820 workqueue_freezing = true;
4823 for_each_pool(pool, pi) {
4824 spin_lock_irq(&pool->lock);
4825 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4826 pool->flags |= POOL_FREEZING;
4827 spin_unlock_irq(&pool->lock);
4830 list_for_each_entry(wq, &workqueues, list) {
4831 mutex_lock(&wq->mutex);
4832 for_each_pwq(pwq, wq)
4833 pwq_adjust_max_active(pwq);
4834 mutex_unlock(&wq->mutex);
4837 mutex_unlock(&wq_pool_mutex);
4841 * freeze_workqueues_busy - are freezable workqueues still busy?
4843 * Check whether freezing is complete. This function must be called
4844 * between freeze_workqueues_begin() and thaw_workqueues().
4847 * Grabs and releases wq_pool_mutex.
4850 * %true if some freezable workqueues are still busy. %false if freezing
4853 bool freeze_workqueues_busy(void)
4856 struct workqueue_struct *wq;
4857 struct pool_workqueue *pwq;
4859 mutex_lock(&wq_pool_mutex);
4861 WARN_ON_ONCE(!workqueue_freezing);
4863 list_for_each_entry(wq, &workqueues, list) {
4864 if (!(wq->flags & WQ_FREEZABLE))
4867 * nr_active is monotonically decreasing. It's safe
4868 * to peek without lock.
4870 rcu_read_lock_sched();
4871 for_each_pwq(pwq, wq) {
4872 WARN_ON_ONCE(pwq->nr_active < 0);
4873 if (pwq->nr_active) {
4875 rcu_read_unlock_sched();
4879 rcu_read_unlock_sched();
4882 mutex_unlock(&wq_pool_mutex);
4887 * thaw_workqueues - thaw workqueues
4889 * Thaw workqueues. Normal queueing is restored and all collected
4890 * frozen works are transferred to their respective pool worklists.
4893 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4895 void thaw_workqueues(void)
4897 struct workqueue_struct *wq;
4898 struct pool_workqueue *pwq;
4899 struct worker_pool *pool;
4902 mutex_lock(&wq_pool_mutex);
4904 if (!workqueue_freezing)
4907 /* clear FREEZING */
4908 for_each_pool(pool, pi) {
4909 spin_lock_irq(&pool->lock);
4910 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4911 pool->flags &= ~POOL_FREEZING;
4912 spin_unlock_irq(&pool->lock);
4915 /* restore max_active and repopulate worklist */
4916 list_for_each_entry(wq, &workqueues, list) {
4917 mutex_lock(&wq->mutex);
4918 for_each_pwq(pwq, wq)
4919 pwq_adjust_max_active(pwq);
4920 mutex_unlock(&wq->mutex);
4923 workqueue_freezing = false;
4925 mutex_unlock(&wq_pool_mutex);
4927 #endif /* CONFIG_FREEZER */
4929 static void __init wq_numa_init(void)
4934 /* determine NUMA pwq table len - highest node id + 1 */
4936 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4938 if (num_possible_nodes() <= 1)
4941 if (wq_disable_numa) {
4942 pr_info("workqueue: NUMA affinity support disabled\n");
4946 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4947 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4950 * We want masks of possible CPUs of each node which isn't readily
4951 * available. Build one from cpu_to_node() which should have been
4952 * fully initialized by now.
4954 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4958 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4959 node_online(node) ? node : NUMA_NO_NODE));
4961 for_each_possible_cpu(cpu) {
4962 node = cpu_to_node(cpu);
4963 if (WARN_ON(node == NUMA_NO_NODE)) {
4964 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4965 /* happens iff arch is bonkers, let's just proceed */
4968 cpumask_set_cpu(cpu, tbl[node]);
4971 wq_numa_possible_cpumask = tbl;
4972 wq_numa_enabled = true;
4975 static int __init init_workqueues(void)
4977 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4980 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4982 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4984 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4985 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4989 /* initialize CPU pools */
4990 for_each_possible_cpu(cpu) {
4991 struct worker_pool *pool;
4994 for_each_cpu_worker_pool(pool, cpu) {
4995 BUG_ON(init_worker_pool(pool));
4997 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4998 pool->attrs->nice = std_nice[i++];
4999 pool->node = cpu_to_node(cpu);
5002 mutex_lock(&wq_pool_mutex);
5003 BUG_ON(worker_pool_assign_id(pool));
5004 mutex_unlock(&wq_pool_mutex);
5008 /* create the initial worker */
5009 for_each_online_cpu(cpu) {
5010 struct worker_pool *pool;
5012 for_each_cpu_worker_pool(pool, cpu) {
5013 pool->flags &= ~POOL_DISASSOCIATED;
5014 BUG_ON(create_and_start_worker(pool) < 0);
5018 /* create default unbound and ordered wq attrs */
5019 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5020 struct workqueue_attrs *attrs;
5022 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5023 attrs->nice = std_nice[i];
5024 unbound_std_wq_attrs[i] = attrs;
5027 * An ordered wq should have only one pwq as ordering is
5028 * guaranteed by max_active which is enforced by pwqs.
5029 * Turn off NUMA so that dfl_pwq is used for all nodes.
5031 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5032 attrs->nice = std_nice[i];
5033 attrs->no_numa = true;
5034 ordered_wq_attrs[i] = attrs;
5037 system_wq = alloc_workqueue("events", 0, 0);
5038 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5039 system_long_wq = alloc_workqueue("events_long", 0, 0);
5040 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5041 WQ_UNBOUND_MAX_ACTIVE);
5042 system_freezable_wq = alloc_workqueue("events_freezable",
5044 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5045 WQ_POWER_EFFICIENT, 0);
5046 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5047 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5049 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5050 !system_unbound_wq || !system_freezable_wq ||
5051 !system_power_efficient_wq ||
5052 !system_freezable_power_efficient_wq);
5055 early_initcall(init_workqueues);