2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * WQ: wq->mutex protected.
132 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
134 * MD: wq_mayday_lock protected.
137 /* struct worker is defined in workqueue_internal.h */
140 spinlock_t lock; /* the pool lock */
141 int cpu; /* I: the associated cpu */
142 int node; /* I: the associated node ID */
143 int id; /* I: pool ID */
144 unsigned int flags; /* X: flags */
146 struct list_head worklist; /* L: list of pending works */
147 int nr_workers; /* L: total number of workers */
149 /* nr_idle includes the ones off idle_list for rebinding */
150 int nr_idle; /* L: currently idle ones */
152 struct list_head idle_list; /* X: list of idle workers */
153 struct timer_list idle_timer; /* L: worker idle timeout */
154 struct timer_list mayday_timer; /* L: SOS timer for workers */
156 /* a workers is either on busy_hash or idle_list, or the manager */
157 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
158 /* L: hash of busy workers */
160 /* see manage_workers() for details on the two manager mutexes */
161 struct mutex manager_arb; /* manager arbitration */
162 struct mutex attach_mutex; /* attach/detach exclusion */
163 struct list_head workers; /* A: attached workers */
164 struct completion *detach_completion; /* all workers detached */
166 struct ida worker_ida; /* worker IDs for task name */
168 struct workqueue_attrs *attrs; /* I: worker attributes */
169 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
170 int refcnt; /* PL: refcnt for unbound pools */
173 * The current concurrency level. As it's likely to be accessed
174 * from other CPUs during try_to_wake_up(), put it in a separate
177 atomic_t nr_running ____cacheline_aligned_in_smp;
180 * Destruction of pool is sched-RCU protected to allow dereferences
181 * from get_work_pool().
184 } ____cacheline_aligned_in_smp;
187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 * of work_struct->data are used for flags and the remaining high bits
189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 * number of flag bits.
192 struct pool_workqueue {
193 struct worker_pool *pool; /* I: the associated pool */
194 struct workqueue_struct *wq; /* I: the owning workqueue */
195 int work_color; /* L: current color */
196 int flush_color; /* L: flushing color */
197 int refcnt; /* L: reference count */
198 int nr_in_flight[WORK_NR_COLORS];
199 /* L: nr of in_flight works */
200 int nr_active; /* L: nr of active works */
201 int max_active; /* L: max active works */
202 struct list_head delayed_works; /* L: delayed works */
203 struct list_head pwqs_node; /* WR: node on wq->pwqs */
204 struct list_head mayday_node; /* MD: node on wq->maydays */
207 * Release of unbound pwq is punted to system_wq. See put_pwq()
208 * and pwq_unbound_release_workfn() for details. pool_workqueue
209 * itself is also sched-RCU protected so that the first pwq can be
210 * determined without grabbing wq->mutex.
212 struct work_struct unbound_release_work;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
217 * Structure used to wait for workqueue flush.
220 struct list_head list; /* WQ: list of flushers */
221 int flush_color; /* WQ: flush color waiting for */
222 struct completion done; /* flush completion */
228 * The externally visible workqueue. It relays the issued work items to
229 * the appropriate worker_pool through its pool_workqueues.
231 struct workqueue_struct {
232 struct list_head pwqs; /* WR: all pwqs of this wq */
233 struct list_head list; /* PL: list of all workqueues */
235 struct mutex mutex; /* protects this wq */
236 int work_color; /* WQ: current work color */
237 int flush_color; /* WQ: current flush color */
238 atomic_t nr_pwqs_to_flush; /* flush in progress */
239 struct wq_flusher *first_flusher; /* WQ: first flusher */
240 struct list_head flusher_queue; /* WQ: flush waiters */
241 struct list_head flusher_overflow; /* WQ: flush overflow list */
243 struct list_head maydays; /* MD: pwqs requesting rescue */
244 struct worker *rescuer; /* I: rescue worker */
246 int nr_drainers; /* WQ: drain in progress */
247 int saved_max_active; /* WQ: saved pwq max_active */
249 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
253 struct wq_device *wq_dev; /* I: for sysfs interface */
255 #ifdef CONFIG_LOCKDEP
256 struct lockdep_map lockdep_map;
258 char name[WQ_NAME_LEN]; /* I: workqueue name */
260 /* hot fields used during command issue, aligned to cacheline */
261 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
262 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
263 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
266 static struct kmem_cache *pwq_cache;
268 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
269 static cpumask_var_t *wq_numa_possible_cpumask;
270 /* possible CPUs of each node */
272 static bool wq_disable_numa;
273 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
275 /* see the comment above the definition of WQ_POWER_EFFICIENT */
276 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
277 static bool wq_power_efficient = true;
279 static bool wq_power_efficient;
282 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
284 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
286 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
287 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
289 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
290 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
292 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
293 static bool workqueue_freezing; /* PL: have wqs started freezing? */
295 /* the per-cpu worker pools */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
299 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
301 /* PL: hash of all unbound pools keyed by pool->attrs */
302 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
304 /* I: attributes used when instantiating standard unbound pools on demand */
305 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
307 /* I: attributes used when instantiating ordered pools on demand */
308 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
310 struct workqueue_struct *system_wq __read_mostly;
311 EXPORT_SYMBOL(system_wq);
312 struct workqueue_struct *system_highpri_wq __read_mostly;
313 EXPORT_SYMBOL_GPL(system_highpri_wq);
314 struct workqueue_struct *system_long_wq __read_mostly;
315 EXPORT_SYMBOL_GPL(system_long_wq);
316 struct workqueue_struct *system_unbound_wq __read_mostly;
317 EXPORT_SYMBOL_GPL(system_unbound_wq);
318 struct workqueue_struct *system_freezable_wq __read_mostly;
319 EXPORT_SYMBOL_GPL(system_freezable_wq);
320 struct workqueue_struct *system_power_efficient_wq __read_mostly;
321 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
322 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
323 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
325 static int worker_thread(void *__worker);
326 static void copy_workqueue_attrs(struct workqueue_attrs *to,
327 const struct workqueue_attrs *from);
329 #define CREATE_TRACE_POINTS
330 #include <trace/events/workqueue.h>
332 #define assert_rcu_or_pool_mutex() \
333 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
334 lockdep_is_held(&wq_pool_mutex), \
335 "sched RCU or wq_pool_mutex should be held")
337 #define assert_rcu_or_wq_mutex(wq) \
338 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
339 lockdep_is_held(&wq->mutex), \
340 "sched RCU or wq->mutex should be held")
342 #define for_each_cpu_worker_pool(pool, cpu) \
343 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
344 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
348 * for_each_pool - iterate through all worker_pools in the system
349 * @pool: iteration cursor
350 * @pi: integer used for iteration
352 * This must be called either with wq_pool_mutex held or sched RCU read
353 * locked. If the pool needs to be used beyond the locking in effect, the
354 * caller is responsible for guaranteeing that the pool stays online.
356 * The if/else clause exists only for the lockdep assertion and can be
359 #define for_each_pool(pool, pi) \
360 idr_for_each_entry(&worker_pool_idr, pool, pi) \
361 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
365 * for_each_pool_worker - iterate through all workers of a worker_pool
366 * @worker: iteration cursor
367 * @pool: worker_pool to iterate workers of
369 * This must be called with @pool->attach_mutex.
371 * The if/else clause exists only for the lockdep assertion and can be
374 #define for_each_pool_worker(worker, pool) \
375 list_for_each_entry((worker), &(pool)->workers, node) \
376 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
380 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
381 * @pwq: iteration cursor
382 * @wq: the target workqueue
384 * This must be called either with wq->mutex held or sched RCU read locked.
385 * If the pwq needs to be used beyond the locking in effect, the caller is
386 * responsible for guaranteeing that the pwq stays online.
388 * The if/else clause exists only for the lockdep assertion and can be
391 #define for_each_pwq(pwq, wq) \
392 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
393 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
396 #ifdef CONFIG_DEBUG_OBJECTS_WORK
398 static struct debug_obj_descr work_debug_descr;
400 static void *work_debug_hint(void *addr)
402 return ((struct work_struct *) addr)->func;
406 * fixup_init is called when:
407 * - an active object is initialized
409 static int work_fixup_init(void *addr, enum debug_obj_state state)
411 struct work_struct *work = addr;
414 case ODEBUG_STATE_ACTIVE:
415 cancel_work_sync(work);
416 debug_object_init(work, &work_debug_descr);
424 * fixup_activate is called when:
425 * - an active object is activated
426 * - an unknown object is activated (might be a statically initialized object)
428 static int work_fixup_activate(void *addr, enum debug_obj_state state)
430 struct work_struct *work = addr;
434 case ODEBUG_STATE_NOTAVAILABLE:
436 * This is not really a fixup. The work struct was
437 * statically initialized. We just make sure that it
438 * is tracked in the object tracker.
440 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
441 debug_object_init(work, &work_debug_descr);
442 debug_object_activate(work, &work_debug_descr);
448 case ODEBUG_STATE_ACTIVE:
457 * fixup_free is called when:
458 * - an active object is freed
460 static int work_fixup_free(void *addr, enum debug_obj_state state)
462 struct work_struct *work = addr;
465 case ODEBUG_STATE_ACTIVE:
466 cancel_work_sync(work);
467 debug_object_free(work, &work_debug_descr);
474 static struct debug_obj_descr work_debug_descr = {
475 .name = "work_struct",
476 .debug_hint = work_debug_hint,
477 .fixup_init = work_fixup_init,
478 .fixup_activate = work_fixup_activate,
479 .fixup_free = work_fixup_free,
482 static inline void debug_work_activate(struct work_struct *work)
484 debug_object_activate(work, &work_debug_descr);
487 static inline void debug_work_deactivate(struct work_struct *work)
489 debug_object_deactivate(work, &work_debug_descr);
492 void __init_work(struct work_struct *work, int onstack)
495 debug_object_init_on_stack(work, &work_debug_descr);
497 debug_object_init(work, &work_debug_descr);
499 EXPORT_SYMBOL_GPL(__init_work);
501 void destroy_work_on_stack(struct work_struct *work)
503 debug_object_free(work, &work_debug_descr);
505 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
507 void destroy_delayed_work_on_stack(struct delayed_work *work)
509 destroy_timer_on_stack(&work->timer);
510 debug_object_free(&work->work, &work_debug_descr);
512 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
515 static inline void debug_work_activate(struct work_struct *work) { }
516 static inline void debug_work_deactivate(struct work_struct *work) { }
520 * worker_pool_assign_id - allocate ID and assing it to @pool
521 * @pool: the pool pointer of interest
523 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
524 * successfully, -errno on failure.
526 static int worker_pool_assign_id(struct worker_pool *pool)
530 lockdep_assert_held(&wq_pool_mutex);
532 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
542 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
543 * @wq: the target workqueue
546 * This must be called either with pwq_lock held or sched RCU read locked.
547 * If the pwq needs to be used beyond the locking in effect, the caller is
548 * responsible for guaranteeing that the pwq stays online.
550 * Return: The unbound pool_workqueue for @node.
552 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
555 assert_rcu_or_wq_mutex(wq);
556 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
559 static unsigned int work_color_to_flags(int color)
561 return color << WORK_STRUCT_COLOR_SHIFT;
564 static int get_work_color(struct work_struct *work)
566 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
567 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
570 static int work_next_color(int color)
572 return (color + 1) % WORK_NR_COLORS;
576 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
577 * contain the pointer to the queued pwq. Once execution starts, the flag
578 * is cleared and the high bits contain OFFQ flags and pool ID.
580 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
581 * and clear_work_data() can be used to set the pwq, pool or clear
582 * work->data. These functions should only be called while the work is
583 * owned - ie. while the PENDING bit is set.
585 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
586 * corresponding to a work. Pool is available once the work has been
587 * queued anywhere after initialization until it is sync canceled. pwq is
588 * available only while the work item is queued.
590 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
591 * canceled. While being canceled, a work item may have its PENDING set
592 * but stay off timer and worklist for arbitrarily long and nobody should
593 * try to steal the PENDING bit.
595 static inline void set_work_data(struct work_struct *work, unsigned long data,
598 WARN_ON_ONCE(!work_pending(work));
599 atomic_long_set(&work->data, data | flags | work_static(work));
602 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
603 unsigned long extra_flags)
605 set_work_data(work, (unsigned long)pwq,
606 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
609 static void set_work_pool_and_keep_pending(struct work_struct *work,
612 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
613 WORK_STRUCT_PENDING);
616 static void set_work_pool_and_clear_pending(struct work_struct *work,
620 * The following wmb is paired with the implied mb in
621 * test_and_set_bit(PENDING) and ensures all updates to @work made
622 * here are visible to and precede any updates by the next PENDING
626 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
629 static void clear_work_data(struct work_struct *work)
631 smp_wmb(); /* see set_work_pool_and_clear_pending() */
632 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
635 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
637 unsigned long data = atomic_long_read(&work->data);
639 if (data & WORK_STRUCT_PWQ)
640 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
646 * get_work_pool - return the worker_pool a given work was associated with
647 * @work: the work item of interest
649 * Pools are created and destroyed under wq_pool_mutex, and allows read
650 * access under sched-RCU read lock. As such, this function should be
651 * called under wq_pool_mutex or with preemption disabled.
653 * All fields of the returned pool are accessible as long as the above
654 * mentioned locking is in effect. If the returned pool needs to be used
655 * beyond the critical section, the caller is responsible for ensuring the
656 * returned pool is and stays online.
658 * Return: The worker_pool @work was last associated with. %NULL if none.
660 static struct worker_pool *get_work_pool(struct work_struct *work)
662 unsigned long data = atomic_long_read(&work->data);
665 assert_rcu_or_pool_mutex();
667 if (data & WORK_STRUCT_PWQ)
668 return ((struct pool_workqueue *)
669 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
671 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
672 if (pool_id == WORK_OFFQ_POOL_NONE)
675 return idr_find(&worker_pool_idr, pool_id);
679 * get_work_pool_id - return the worker pool ID a given work is associated with
680 * @work: the work item of interest
682 * Return: The worker_pool ID @work was last associated with.
683 * %WORK_OFFQ_POOL_NONE if none.
685 static int get_work_pool_id(struct work_struct *work)
687 unsigned long data = atomic_long_read(&work->data);
689 if (data & WORK_STRUCT_PWQ)
690 return ((struct pool_workqueue *)
691 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
693 return data >> WORK_OFFQ_POOL_SHIFT;
696 static void mark_work_canceling(struct work_struct *work)
698 unsigned long pool_id = get_work_pool_id(work);
700 pool_id <<= WORK_OFFQ_POOL_SHIFT;
701 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
704 static bool work_is_canceling(struct work_struct *work)
706 unsigned long data = atomic_long_read(&work->data);
708 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
712 * Policy functions. These define the policies on how the global worker
713 * pools are managed. Unless noted otherwise, these functions assume that
714 * they're being called with pool->lock held.
717 static bool __need_more_worker(struct worker_pool *pool)
719 return !atomic_read(&pool->nr_running);
723 * Need to wake up a worker? Called from anything but currently
726 * Note that, because unbound workers never contribute to nr_running, this
727 * function will always return %true for unbound pools as long as the
728 * worklist isn't empty.
730 static bool need_more_worker(struct worker_pool *pool)
732 return !list_empty(&pool->worklist) && __need_more_worker(pool);
735 /* Can I start working? Called from busy but !running workers. */
736 static bool may_start_working(struct worker_pool *pool)
738 return pool->nr_idle;
741 /* Do I need to keep working? Called from currently running workers. */
742 static bool keep_working(struct worker_pool *pool)
744 return !list_empty(&pool->worklist) &&
745 atomic_read(&pool->nr_running) <= 1;
748 /* Do we need a new worker? Called from manager. */
749 static bool need_to_create_worker(struct worker_pool *pool)
751 return need_more_worker(pool) && !may_start_working(pool);
754 /* Do we have too many workers and should some go away? */
755 static bool too_many_workers(struct worker_pool *pool)
757 bool managing = mutex_is_locked(&pool->manager_arb);
758 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
759 int nr_busy = pool->nr_workers - nr_idle;
761 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
768 /* Return the first idle worker. Safe with preemption disabled */
769 static struct worker *first_idle_worker(struct worker_pool *pool)
771 if (unlikely(list_empty(&pool->idle_list)))
774 return list_first_entry(&pool->idle_list, struct worker, entry);
778 * wake_up_worker - wake up an idle worker
779 * @pool: worker pool to wake worker from
781 * Wake up the first idle worker of @pool.
784 * spin_lock_irq(pool->lock).
786 static void wake_up_worker(struct worker_pool *pool)
788 struct worker *worker = first_idle_worker(pool);
791 wake_up_process(worker->task);
795 * wq_worker_waking_up - a worker is waking up
796 * @task: task waking up
797 * @cpu: CPU @task is waking up to
799 * This function is called during try_to_wake_up() when a worker is
803 * spin_lock_irq(rq->lock)
805 void wq_worker_waking_up(struct task_struct *task, int cpu)
807 struct worker *worker = kthread_data(task);
809 if (!(worker->flags & WORKER_NOT_RUNNING)) {
810 WARN_ON_ONCE(worker->pool->cpu != cpu);
811 atomic_inc(&worker->pool->nr_running);
816 * wq_worker_sleeping - a worker is going to sleep
817 * @task: task going to sleep
818 * @cpu: CPU in question, must be the current CPU number
820 * This function is called during schedule() when a busy worker is
821 * going to sleep. Worker on the same cpu can be woken up by
822 * returning pointer to its task.
825 * spin_lock_irq(rq->lock)
828 * Worker task on @cpu to wake up, %NULL if none.
830 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
832 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
833 struct worker_pool *pool;
836 * Rescuers, which may not have all the fields set up like normal
837 * workers, also reach here, let's not access anything before
838 * checking NOT_RUNNING.
840 if (worker->flags & WORKER_NOT_RUNNING)
845 /* this can only happen on the local cpu */
846 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
850 * The counterpart of the following dec_and_test, implied mb,
851 * worklist not empty test sequence is in insert_work().
852 * Please read comment there.
854 * NOT_RUNNING is clear. This means that we're bound to and
855 * running on the local cpu w/ rq lock held and preemption
856 * disabled, which in turn means that none else could be
857 * manipulating idle_list, so dereferencing idle_list without pool
860 if (atomic_dec_and_test(&pool->nr_running) &&
861 !list_empty(&pool->worklist))
862 to_wakeup = first_idle_worker(pool);
863 return to_wakeup ? to_wakeup->task : NULL;
867 * worker_set_flags - set worker flags and adjust nr_running accordingly
869 * @flags: flags to set
871 * Set @flags in @worker->flags and adjust nr_running accordingly.
874 * spin_lock_irq(pool->lock)
876 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
878 struct worker_pool *pool = worker->pool;
880 WARN_ON_ONCE(worker->task != current);
882 /* If transitioning into NOT_RUNNING, adjust nr_running. */
883 if ((flags & WORKER_NOT_RUNNING) &&
884 !(worker->flags & WORKER_NOT_RUNNING)) {
885 atomic_dec(&pool->nr_running);
888 worker->flags |= flags;
892 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
894 * @flags: flags to clear
896 * Clear @flags in @worker->flags and adjust nr_running accordingly.
899 * spin_lock_irq(pool->lock)
901 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
903 struct worker_pool *pool = worker->pool;
904 unsigned int oflags = worker->flags;
906 WARN_ON_ONCE(worker->task != current);
908 worker->flags &= ~flags;
911 * If transitioning out of NOT_RUNNING, increment nr_running. Note
912 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
913 * of multiple flags, not a single flag.
915 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
916 if (!(worker->flags & WORKER_NOT_RUNNING))
917 atomic_inc(&pool->nr_running);
921 * find_worker_executing_work - find worker which is executing a work
922 * @pool: pool of interest
923 * @work: work to find worker for
925 * Find a worker which is executing @work on @pool by searching
926 * @pool->busy_hash which is keyed by the address of @work. For a worker
927 * to match, its current execution should match the address of @work and
928 * its work function. This is to avoid unwanted dependency between
929 * unrelated work executions through a work item being recycled while still
932 * This is a bit tricky. A work item may be freed once its execution
933 * starts and nothing prevents the freed area from being recycled for
934 * another work item. If the same work item address ends up being reused
935 * before the original execution finishes, workqueue will identify the
936 * recycled work item as currently executing and make it wait until the
937 * current execution finishes, introducing an unwanted dependency.
939 * This function checks the work item address and work function to avoid
940 * false positives. Note that this isn't complete as one may construct a
941 * work function which can introduce dependency onto itself through a
942 * recycled work item. Well, if somebody wants to shoot oneself in the
943 * foot that badly, there's only so much we can do, and if such deadlock
944 * actually occurs, it should be easy to locate the culprit work function.
947 * spin_lock_irq(pool->lock).
950 * Pointer to worker which is executing @work if found, %NULL
953 static struct worker *find_worker_executing_work(struct worker_pool *pool,
954 struct work_struct *work)
956 struct worker *worker;
958 hash_for_each_possible(pool->busy_hash, worker, hentry,
960 if (worker->current_work == work &&
961 worker->current_func == work->func)
968 * move_linked_works - move linked works to a list
969 * @work: start of series of works to be scheduled
970 * @head: target list to append @work to
971 * @nextp: out paramter for nested worklist walking
973 * Schedule linked works starting from @work to @head. Work series to
974 * be scheduled starts at @work and includes any consecutive work with
975 * WORK_STRUCT_LINKED set in its predecessor.
977 * If @nextp is not NULL, it's updated to point to the next work of
978 * the last scheduled work. This allows move_linked_works() to be
979 * nested inside outer list_for_each_entry_safe().
982 * spin_lock_irq(pool->lock).
984 static void move_linked_works(struct work_struct *work, struct list_head *head,
985 struct work_struct **nextp)
987 struct work_struct *n;
990 * Linked worklist will always end before the end of the list,
991 * use NULL for list head.
993 list_for_each_entry_safe_from(work, n, NULL, entry) {
994 list_move_tail(&work->entry, head);
995 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1000 * If we're already inside safe list traversal and have moved
1001 * multiple works to the scheduled queue, the next position
1002 * needs to be updated.
1009 * get_pwq - get an extra reference on the specified pool_workqueue
1010 * @pwq: pool_workqueue to get
1012 * Obtain an extra reference on @pwq. The caller should guarantee that
1013 * @pwq has positive refcnt and be holding the matching pool->lock.
1015 static void get_pwq(struct pool_workqueue *pwq)
1017 lockdep_assert_held(&pwq->pool->lock);
1018 WARN_ON_ONCE(pwq->refcnt <= 0);
1023 * put_pwq - put a pool_workqueue reference
1024 * @pwq: pool_workqueue to put
1026 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1027 * destruction. The caller should be holding the matching pool->lock.
1029 static void put_pwq(struct pool_workqueue *pwq)
1031 lockdep_assert_held(&pwq->pool->lock);
1032 if (likely(--pwq->refcnt))
1034 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1037 * @pwq can't be released under pool->lock, bounce to
1038 * pwq_unbound_release_workfn(). This never recurses on the same
1039 * pool->lock as this path is taken only for unbound workqueues and
1040 * the release work item is scheduled on a per-cpu workqueue. To
1041 * avoid lockdep warning, unbound pool->locks are given lockdep
1042 * subclass of 1 in get_unbound_pool().
1044 schedule_work(&pwq->unbound_release_work);
1048 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1049 * @pwq: pool_workqueue to put (can be %NULL)
1051 * put_pwq() with locking. This function also allows %NULL @pwq.
1053 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1057 * As both pwqs and pools are sched-RCU protected, the
1058 * following lock operations are safe.
1060 spin_lock_irq(&pwq->pool->lock);
1062 spin_unlock_irq(&pwq->pool->lock);
1066 static void pwq_activate_delayed_work(struct work_struct *work)
1068 struct pool_workqueue *pwq = get_work_pwq(work);
1070 trace_workqueue_activate_work(work);
1071 move_linked_works(work, &pwq->pool->worklist, NULL);
1072 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1076 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1078 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1079 struct work_struct, entry);
1081 pwq_activate_delayed_work(work);
1085 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1086 * @pwq: pwq of interest
1087 * @color: color of work which left the queue
1089 * A work either has completed or is removed from pending queue,
1090 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1093 * spin_lock_irq(pool->lock).
1095 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1097 /* uncolored work items don't participate in flushing or nr_active */
1098 if (color == WORK_NO_COLOR)
1101 pwq->nr_in_flight[color]--;
1104 if (!list_empty(&pwq->delayed_works)) {
1105 /* one down, submit a delayed one */
1106 if (pwq->nr_active < pwq->max_active)
1107 pwq_activate_first_delayed(pwq);
1110 /* is flush in progress and are we at the flushing tip? */
1111 if (likely(pwq->flush_color != color))
1114 /* are there still in-flight works? */
1115 if (pwq->nr_in_flight[color])
1118 /* this pwq is done, clear flush_color */
1119 pwq->flush_color = -1;
1122 * If this was the last pwq, wake up the first flusher. It
1123 * will handle the rest.
1125 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1126 complete(&pwq->wq->first_flusher->done);
1132 * try_to_grab_pending - steal work item from worklist and disable irq
1133 * @work: work item to steal
1134 * @is_dwork: @work is a delayed_work
1135 * @flags: place to store irq state
1137 * Try to grab PENDING bit of @work. This function can handle @work in any
1138 * stable state - idle, on timer or on worklist.
1141 * 1 if @work was pending and we successfully stole PENDING
1142 * 0 if @work was idle and we claimed PENDING
1143 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1144 * -ENOENT if someone else is canceling @work, this state may persist
1145 * for arbitrarily long
1148 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1149 * interrupted while holding PENDING and @work off queue, irq must be
1150 * disabled on entry. This, combined with delayed_work->timer being
1151 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1153 * On successful return, >= 0, irq is disabled and the caller is
1154 * responsible for releasing it using local_irq_restore(*@flags).
1156 * This function is safe to call from any context including IRQ handler.
1158 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1159 unsigned long *flags)
1161 struct worker_pool *pool;
1162 struct pool_workqueue *pwq;
1164 local_irq_save(*flags);
1166 /* try to steal the timer if it exists */
1168 struct delayed_work *dwork = to_delayed_work(work);
1171 * dwork->timer is irqsafe. If del_timer() fails, it's
1172 * guaranteed that the timer is not queued anywhere and not
1173 * running on the local CPU.
1175 if (likely(del_timer(&dwork->timer)))
1179 /* try to claim PENDING the normal way */
1180 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1184 * The queueing is in progress, or it is already queued. Try to
1185 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1187 pool = get_work_pool(work);
1191 spin_lock(&pool->lock);
1193 * work->data is guaranteed to point to pwq only while the work
1194 * item is queued on pwq->wq, and both updating work->data to point
1195 * to pwq on queueing and to pool on dequeueing are done under
1196 * pwq->pool->lock. This in turn guarantees that, if work->data
1197 * points to pwq which is associated with a locked pool, the work
1198 * item is currently queued on that pool.
1200 pwq = get_work_pwq(work);
1201 if (pwq && pwq->pool == pool) {
1202 debug_work_deactivate(work);
1205 * A delayed work item cannot be grabbed directly because
1206 * it might have linked NO_COLOR work items which, if left
1207 * on the delayed_list, will confuse pwq->nr_active
1208 * management later on and cause stall. Make sure the work
1209 * item is activated before grabbing.
1211 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1212 pwq_activate_delayed_work(work);
1214 list_del_init(&work->entry);
1215 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1217 /* work->data points to pwq iff queued, point to pool */
1218 set_work_pool_and_keep_pending(work, pool->id);
1220 spin_unlock(&pool->lock);
1223 spin_unlock(&pool->lock);
1225 local_irq_restore(*flags);
1226 if (work_is_canceling(work))
1233 * insert_work - insert a work into a pool
1234 * @pwq: pwq @work belongs to
1235 * @work: work to insert
1236 * @head: insertion point
1237 * @extra_flags: extra WORK_STRUCT_* flags to set
1239 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1240 * work_struct flags.
1243 * spin_lock_irq(pool->lock).
1245 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1246 struct list_head *head, unsigned int extra_flags)
1248 struct worker_pool *pool = pwq->pool;
1250 /* we own @work, set data and link */
1251 set_work_pwq(work, pwq, extra_flags);
1252 list_add_tail(&work->entry, head);
1256 * Ensure either wq_worker_sleeping() sees the above
1257 * list_add_tail() or we see zero nr_running to avoid workers lying
1258 * around lazily while there are works to be processed.
1262 if (__need_more_worker(pool))
1263 wake_up_worker(pool);
1267 * Test whether @work is being queued from another work executing on the
1270 static bool is_chained_work(struct workqueue_struct *wq)
1272 struct worker *worker;
1274 worker = current_wq_worker();
1276 * Return %true iff I'm a worker execuing a work item on @wq. If
1277 * I'm @worker, it's safe to dereference it without locking.
1279 return worker && worker->current_pwq->wq == wq;
1282 static void __queue_work(int cpu, struct workqueue_struct *wq,
1283 struct work_struct *work)
1285 struct pool_workqueue *pwq;
1286 struct worker_pool *last_pool;
1287 struct list_head *worklist;
1288 unsigned int work_flags;
1289 unsigned int req_cpu = cpu;
1292 * While a work item is PENDING && off queue, a task trying to
1293 * steal the PENDING will busy-loop waiting for it to either get
1294 * queued or lose PENDING. Grabbing PENDING and queueing should
1295 * happen with IRQ disabled.
1297 WARN_ON_ONCE(!irqs_disabled());
1299 debug_work_activate(work);
1301 /* if draining, only works from the same workqueue are allowed */
1302 if (unlikely(wq->flags & __WQ_DRAINING) &&
1303 WARN_ON_ONCE(!is_chained_work(wq)))
1306 if (req_cpu == WORK_CPU_UNBOUND)
1307 cpu = raw_smp_processor_id();
1309 /* pwq which will be used unless @work is executing elsewhere */
1310 if (!(wq->flags & WQ_UNBOUND))
1311 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1313 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1316 * If @work was previously on a different pool, it might still be
1317 * running there, in which case the work needs to be queued on that
1318 * pool to guarantee non-reentrancy.
1320 last_pool = get_work_pool(work);
1321 if (last_pool && last_pool != pwq->pool) {
1322 struct worker *worker;
1324 spin_lock(&last_pool->lock);
1326 worker = find_worker_executing_work(last_pool, work);
1328 if (worker && worker->current_pwq->wq == wq) {
1329 pwq = worker->current_pwq;
1331 /* meh... not running there, queue here */
1332 spin_unlock(&last_pool->lock);
1333 spin_lock(&pwq->pool->lock);
1336 spin_lock(&pwq->pool->lock);
1340 * pwq is determined and locked. For unbound pools, we could have
1341 * raced with pwq release and it could already be dead. If its
1342 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1343 * without another pwq replacing it in the numa_pwq_tbl or while
1344 * work items are executing on it, so the retrying is guaranteed to
1345 * make forward-progress.
1347 if (unlikely(!pwq->refcnt)) {
1348 if (wq->flags & WQ_UNBOUND) {
1349 spin_unlock(&pwq->pool->lock);
1354 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1358 /* pwq determined, queue */
1359 trace_workqueue_queue_work(req_cpu, pwq, work);
1361 if (WARN_ON(!list_empty(&work->entry))) {
1362 spin_unlock(&pwq->pool->lock);
1366 pwq->nr_in_flight[pwq->work_color]++;
1367 work_flags = work_color_to_flags(pwq->work_color);
1369 if (likely(pwq->nr_active < pwq->max_active)) {
1370 trace_workqueue_activate_work(work);
1372 worklist = &pwq->pool->worklist;
1374 work_flags |= WORK_STRUCT_DELAYED;
1375 worklist = &pwq->delayed_works;
1378 insert_work(pwq, work, worklist, work_flags);
1380 spin_unlock(&pwq->pool->lock);
1384 * queue_work_on - queue work on specific cpu
1385 * @cpu: CPU number to execute work on
1386 * @wq: workqueue to use
1387 * @work: work to queue
1389 * We queue the work to a specific CPU, the caller must ensure it
1392 * Return: %false if @work was already on a queue, %true otherwise.
1394 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1395 struct work_struct *work)
1398 unsigned long flags;
1400 local_irq_save(flags);
1402 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1403 __queue_work(cpu, wq, work);
1407 local_irq_restore(flags);
1410 EXPORT_SYMBOL(queue_work_on);
1412 void delayed_work_timer_fn(unsigned long __data)
1414 struct delayed_work *dwork = (struct delayed_work *)__data;
1416 /* should have been called from irqsafe timer with irq already off */
1417 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1419 EXPORT_SYMBOL(delayed_work_timer_fn);
1421 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1422 struct delayed_work *dwork, unsigned long delay)
1424 struct timer_list *timer = &dwork->timer;
1425 struct work_struct *work = &dwork->work;
1427 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1428 timer->data != (unsigned long)dwork);
1429 WARN_ON_ONCE(timer_pending(timer));
1430 WARN_ON_ONCE(!list_empty(&work->entry));
1433 * If @delay is 0, queue @dwork->work immediately. This is for
1434 * both optimization and correctness. The earliest @timer can
1435 * expire is on the closest next tick and delayed_work users depend
1436 * on that there's no such delay when @delay is 0.
1439 __queue_work(cpu, wq, &dwork->work);
1443 timer_stats_timer_set_start_info(&dwork->timer);
1447 timer->expires = jiffies + delay;
1449 if (unlikely(cpu != WORK_CPU_UNBOUND))
1450 add_timer_on(timer, cpu);
1456 * queue_delayed_work_on - queue work on specific CPU after delay
1457 * @cpu: CPU number to execute work on
1458 * @wq: workqueue to use
1459 * @dwork: work to queue
1460 * @delay: number of jiffies to wait before queueing
1462 * Return: %false if @work was already on a queue, %true otherwise. If
1463 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1466 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1467 struct delayed_work *dwork, unsigned long delay)
1469 struct work_struct *work = &dwork->work;
1471 unsigned long flags;
1473 /* read the comment in __queue_work() */
1474 local_irq_save(flags);
1476 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1477 __queue_delayed_work(cpu, wq, dwork, delay);
1481 local_irq_restore(flags);
1484 EXPORT_SYMBOL(queue_delayed_work_on);
1487 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1488 * @cpu: CPU number to execute work on
1489 * @wq: workqueue to use
1490 * @dwork: work to queue
1491 * @delay: number of jiffies to wait before queueing
1493 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1494 * modify @dwork's timer so that it expires after @delay. If @delay is
1495 * zero, @work is guaranteed to be scheduled immediately regardless of its
1498 * Return: %false if @dwork was idle and queued, %true if @dwork was
1499 * pending and its timer was modified.
1501 * This function is safe to call from any context including IRQ handler.
1502 * See try_to_grab_pending() for details.
1504 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1505 struct delayed_work *dwork, unsigned long delay)
1507 unsigned long flags;
1511 ret = try_to_grab_pending(&dwork->work, true, &flags);
1512 } while (unlikely(ret == -EAGAIN));
1514 if (likely(ret >= 0)) {
1515 __queue_delayed_work(cpu, wq, dwork, delay);
1516 local_irq_restore(flags);
1519 /* -ENOENT from try_to_grab_pending() becomes %true */
1522 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1525 * worker_enter_idle - enter idle state
1526 * @worker: worker which is entering idle state
1528 * @worker is entering idle state. Update stats and idle timer if
1532 * spin_lock_irq(pool->lock).
1534 static void worker_enter_idle(struct worker *worker)
1536 struct worker_pool *pool = worker->pool;
1538 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1539 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1540 (worker->hentry.next || worker->hentry.pprev)))
1543 /* can't use worker_set_flags(), also called from start_worker() */
1544 worker->flags |= WORKER_IDLE;
1546 worker->last_active = jiffies;
1548 /* idle_list is LIFO */
1549 list_add(&worker->entry, &pool->idle_list);
1551 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1552 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1555 * Sanity check nr_running. Because wq_unbind_fn() releases
1556 * pool->lock between setting %WORKER_UNBOUND and zapping
1557 * nr_running, the warning may trigger spuriously. Check iff
1558 * unbind is not in progress.
1560 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1561 pool->nr_workers == pool->nr_idle &&
1562 atomic_read(&pool->nr_running));
1566 * worker_leave_idle - leave idle state
1567 * @worker: worker which is leaving idle state
1569 * @worker is leaving idle state. Update stats.
1572 * spin_lock_irq(pool->lock).
1574 static void worker_leave_idle(struct worker *worker)
1576 struct worker_pool *pool = worker->pool;
1578 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1580 worker_clr_flags(worker, WORKER_IDLE);
1582 list_del_init(&worker->entry);
1585 static struct worker *alloc_worker(int node)
1587 struct worker *worker;
1589 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1591 INIT_LIST_HEAD(&worker->entry);
1592 INIT_LIST_HEAD(&worker->scheduled);
1593 INIT_LIST_HEAD(&worker->node);
1594 /* on creation a worker is in !idle && prep state */
1595 worker->flags = WORKER_PREP;
1601 * worker_attach_to_pool() - attach a worker to a pool
1602 * @worker: worker to be attached
1603 * @pool: the target pool
1605 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1606 * cpu-binding of @worker are kept coordinated with the pool across
1609 static void worker_attach_to_pool(struct worker *worker,
1610 struct worker_pool *pool)
1612 mutex_lock(&pool->attach_mutex);
1615 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1616 * online CPUs. It'll be re-applied when any of the CPUs come up.
1618 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1621 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1622 * stable across this function. See the comments above the
1623 * flag definition for details.
1625 if (pool->flags & POOL_DISASSOCIATED)
1626 worker->flags |= WORKER_UNBOUND;
1628 list_add_tail(&worker->node, &pool->workers);
1630 mutex_unlock(&pool->attach_mutex);
1634 * worker_detach_from_pool() - detach a worker from its pool
1635 * @worker: worker which is attached to its pool
1636 * @pool: the pool @worker is attached to
1638 * Undo the attaching which had been done in worker_attach_to_pool(). The
1639 * caller worker shouldn't access to the pool after detached except it has
1640 * other reference to the pool.
1642 static void worker_detach_from_pool(struct worker *worker,
1643 struct worker_pool *pool)
1645 struct completion *detach_completion = NULL;
1647 mutex_lock(&pool->attach_mutex);
1648 list_del(&worker->node);
1649 if (list_empty(&pool->workers))
1650 detach_completion = pool->detach_completion;
1651 mutex_unlock(&pool->attach_mutex);
1653 /* clear leftover flags without pool->lock after it is detached */
1654 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1656 if (detach_completion)
1657 complete(detach_completion);
1661 * create_worker - create a new workqueue worker
1662 * @pool: pool the new worker will belong to
1664 * Create a new worker which is attached to @pool. The new worker must be
1665 * started by start_worker().
1668 * Might sleep. Does GFP_KERNEL allocations.
1671 * Pointer to the newly created worker.
1673 static struct worker *create_worker(struct worker_pool *pool)
1675 struct worker *worker = NULL;
1679 /* ID is needed to determine kthread name */
1680 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1684 worker = alloc_worker(pool->node);
1688 worker->pool = pool;
1692 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1693 pool->attrs->nice < 0 ? "H" : "");
1695 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1697 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1698 "kworker/%s", id_buf);
1699 if (IS_ERR(worker->task))
1702 set_user_nice(worker->task, pool->attrs->nice);
1704 /* prevent userland from meddling with cpumask of workqueue workers */
1705 worker->task->flags |= PF_NO_SETAFFINITY;
1707 /* successful, attach the worker to the pool */
1708 worker_attach_to_pool(worker, pool);
1714 ida_simple_remove(&pool->worker_ida, id);
1720 * start_worker - start a newly created worker
1721 * @worker: worker to start
1723 * Make the pool aware of @worker and start it.
1726 * spin_lock_irq(pool->lock).
1728 static void start_worker(struct worker *worker)
1730 worker->pool->nr_workers++;
1731 worker_enter_idle(worker);
1732 wake_up_process(worker->task);
1736 * create_and_start_worker - create and start a worker for a pool
1737 * @pool: the target pool
1739 * Grab the managership of @pool and create and start a new worker for it.
1741 * Return: 0 on success. A negative error code otherwise.
1743 static int create_and_start_worker(struct worker_pool *pool)
1745 struct worker *worker;
1747 worker = create_worker(pool);
1749 spin_lock_irq(&pool->lock);
1750 start_worker(worker);
1751 spin_unlock_irq(&pool->lock);
1754 return worker ? 0 : -ENOMEM;
1758 * destroy_worker - destroy a workqueue worker
1759 * @worker: worker to be destroyed
1761 * Destroy @worker and adjust @pool stats accordingly. The worker should
1765 * spin_lock_irq(pool->lock).
1767 static void destroy_worker(struct worker *worker)
1769 struct worker_pool *pool = worker->pool;
1771 lockdep_assert_held(&pool->lock);
1773 /* sanity check frenzy */
1774 if (WARN_ON(worker->current_work) ||
1775 WARN_ON(!list_empty(&worker->scheduled)) ||
1776 WARN_ON(!(worker->flags & WORKER_IDLE)))
1782 list_del_init(&worker->entry);
1783 worker->flags |= WORKER_DIE;
1784 wake_up_process(worker->task);
1787 static void idle_worker_timeout(unsigned long __pool)
1789 struct worker_pool *pool = (void *)__pool;
1791 spin_lock_irq(&pool->lock);
1793 while (too_many_workers(pool)) {
1794 struct worker *worker;
1795 unsigned long expires;
1797 /* idle_list is kept in LIFO order, check the last one */
1798 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1799 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1801 if (time_before(jiffies, expires)) {
1802 mod_timer(&pool->idle_timer, expires);
1806 destroy_worker(worker);
1809 spin_unlock_irq(&pool->lock);
1812 static void send_mayday(struct work_struct *work)
1814 struct pool_workqueue *pwq = get_work_pwq(work);
1815 struct workqueue_struct *wq = pwq->wq;
1817 lockdep_assert_held(&wq_mayday_lock);
1822 /* mayday mayday mayday */
1823 if (list_empty(&pwq->mayday_node)) {
1825 * If @pwq is for an unbound wq, its base ref may be put at
1826 * any time due to an attribute change. Pin @pwq until the
1827 * rescuer is done with it.
1830 list_add_tail(&pwq->mayday_node, &wq->maydays);
1831 wake_up_process(wq->rescuer->task);
1835 static void pool_mayday_timeout(unsigned long __pool)
1837 struct worker_pool *pool = (void *)__pool;
1838 struct work_struct *work;
1840 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1841 spin_lock(&pool->lock);
1843 if (need_to_create_worker(pool)) {
1845 * We've been trying to create a new worker but
1846 * haven't been successful. We might be hitting an
1847 * allocation deadlock. Send distress signals to
1850 list_for_each_entry(work, &pool->worklist, entry)
1854 spin_unlock(&pool->lock);
1855 spin_unlock_irq(&wq_mayday_lock);
1857 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1861 * maybe_create_worker - create a new worker if necessary
1862 * @pool: pool to create a new worker for
1864 * Create a new worker for @pool if necessary. @pool is guaranteed to
1865 * have at least one idle worker on return from this function. If
1866 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1867 * sent to all rescuers with works scheduled on @pool to resolve
1868 * possible allocation deadlock.
1870 * On return, need_to_create_worker() is guaranteed to be %false and
1871 * may_start_working() %true.
1874 * spin_lock_irq(pool->lock) which may be released and regrabbed
1875 * multiple times. Does GFP_KERNEL allocations. Called only from
1879 * %false if no action was taken and pool->lock stayed locked, %true
1882 static bool maybe_create_worker(struct worker_pool *pool)
1883 __releases(&pool->lock)
1884 __acquires(&pool->lock)
1886 if (!need_to_create_worker(pool))
1889 spin_unlock_irq(&pool->lock);
1891 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1892 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1895 struct worker *worker;
1897 worker = create_worker(pool);
1899 del_timer_sync(&pool->mayday_timer);
1900 spin_lock_irq(&pool->lock);
1901 start_worker(worker);
1902 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1907 if (!need_to_create_worker(pool))
1910 schedule_timeout_interruptible(CREATE_COOLDOWN);
1912 if (!need_to_create_worker(pool))
1916 del_timer_sync(&pool->mayday_timer);
1917 spin_lock_irq(&pool->lock);
1918 if (need_to_create_worker(pool))
1924 * manage_workers - manage worker pool
1927 * Assume the manager role and manage the worker pool @worker belongs
1928 * to. At any given time, there can be only zero or one manager per
1929 * pool. The exclusion is handled automatically by this function.
1931 * The caller can safely start processing works on false return. On
1932 * true return, it's guaranteed that need_to_create_worker() is false
1933 * and may_start_working() is true.
1936 * spin_lock_irq(pool->lock) which may be released and regrabbed
1937 * multiple times. Does GFP_KERNEL allocations.
1940 * %false if the pool don't need management and the caller can safely start
1941 * processing works, %true indicates that the function released pool->lock
1942 * and reacquired it to perform some management function and that the
1943 * conditions that the caller verified while holding the lock before
1944 * calling the function might no longer be true.
1946 static bool manage_workers(struct worker *worker)
1948 struct worker_pool *pool = worker->pool;
1952 * Anyone who successfully grabs manager_arb wins the arbitration
1953 * and becomes the manager. mutex_trylock() on pool->manager_arb
1954 * failure while holding pool->lock reliably indicates that someone
1955 * else is managing the pool and the worker which failed trylock
1956 * can proceed to executing work items. This means that anyone
1957 * grabbing manager_arb is responsible for actually performing
1958 * manager duties. If manager_arb is grabbed and released without
1959 * actual management, the pool may stall indefinitely.
1961 if (!mutex_trylock(&pool->manager_arb))
1964 ret |= maybe_create_worker(pool);
1966 mutex_unlock(&pool->manager_arb);
1971 * process_one_work - process single work
1973 * @work: work to process
1975 * Process @work. This function contains all the logics necessary to
1976 * process a single work including synchronization against and
1977 * interaction with other workers on the same cpu, queueing and
1978 * flushing. As long as context requirement is met, any worker can
1979 * call this function to process a work.
1982 * spin_lock_irq(pool->lock) which is released and regrabbed.
1984 static void process_one_work(struct worker *worker, struct work_struct *work)
1985 __releases(&pool->lock)
1986 __acquires(&pool->lock)
1988 struct pool_workqueue *pwq = get_work_pwq(work);
1989 struct worker_pool *pool = worker->pool;
1990 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1992 struct worker *collision;
1993 #ifdef CONFIG_LOCKDEP
1995 * It is permissible to free the struct work_struct from
1996 * inside the function that is called from it, this we need to
1997 * take into account for lockdep too. To avoid bogus "held
1998 * lock freed" warnings as well as problems when looking into
1999 * work->lockdep_map, make a copy and use that here.
2001 struct lockdep_map lockdep_map;
2003 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2005 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2006 raw_smp_processor_id() != pool->cpu);
2009 * A single work shouldn't be executed concurrently by
2010 * multiple workers on a single cpu. Check whether anyone is
2011 * already processing the work. If so, defer the work to the
2012 * currently executing one.
2014 collision = find_worker_executing_work(pool, work);
2015 if (unlikely(collision)) {
2016 move_linked_works(work, &collision->scheduled, NULL);
2020 /* claim and dequeue */
2021 debug_work_deactivate(work);
2022 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2023 worker->current_work = work;
2024 worker->current_func = work->func;
2025 worker->current_pwq = pwq;
2026 work_color = get_work_color(work);
2028 list_del_init(&work->entry);
2031 * CPU intensive works don't participate in concurrency management.
2032 * They're the scheduler's responsibility. This takes @worker out
2033 * of concurrency management and the next code block will chain
2034 * execution of the pending work items.
2036 if (unlikely(cpu_intensive))
2037 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2040 * Wake up another worker if necessary. The condition is always
2041 * false for normal per-cpu workers since nr_running would always
2042 * be >= 1 at this point. This is used to chain execution of the
2043 * pending work items for WORKER_NOT_RUNNING workers such as the
2044 * UNBOUND and CPU_INTENSIVE ones.
2046 if (need_more_worker(pool))
2047 wake_up_worker(pool);
2050 * Record the last pool and clear PENDING which should be the last
2051 * update to @work. Also, do this inside @pool->lock so that
2052 * PENDING and queued state changes happen together while IRQ is
2055 set_work_pool_and_clear_pending(work, pool->id);
2057 spin_unlock_irq(&pool->lock);
2059 lock_map_acquire_read(&pwq->wq->lockdep_map);
2060 lock_map_acquire(&lockdep_map);
2061 trace_workqueue_execute_start(work);
2062 worker->current_func(work);
2064 * While we must be careful to not use "work" after this, the trace
2065 * point will only record its address.
2067 trace_workqueue_execute_end(work);
2068 lock_map_release(&lockdep_map);
2069 lock_map_release(&pwq->wq->lockdep_map);
2071 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2072 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2073 " last function: %pf\n",
2074 current->comm, preempt_count(), task_pid_nr(current),
2075 worker->current_func);
2076 debug_show_held_locks(current);
2081 * The following prevents a kworker from hogging CPU on !PREEMPT
2082 * kernels, where a requeueing work item waiting for something to
2083 * happen could deadlock with stop_machine as such work item could
2084 * indefinitely requeue itself while all other CPUs are trapped in
2089 spin_lock_irq(&pool->lock);
2091 /* clear cpu intensive status */
2092 if (unlikely(cpu_intensive))
2093 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2095 /* we're done with it, release */
2096 hash_del(&worker->hentry);
2097 worker->current_work = NULL;
2098 worker->current_func = NULL;
2099 worker->current_pwq = NULL;
2100 worker->desc_valid = false;
2101 pwq_dec_nr_in_flight(pwq, work_color);
2105 * process_scheduled_works - process scheduled works
2108 * Process all scheduled works. Please note that the scheduled list
2109 * may change while processing a work, so this function repeatedly
2110 * fetches a work from the top and executes it.
2113 * spin_lock_irq(pool->lock) which may be released and regrabbed
2116 static void process_scheduled_works(struct worker *worker)
2118 while (!list_empty(&worker->scheduled)) {
2119 struct work_struct *work = list_first_entry(&worker->scheduled,
2120 struct work_struct, entry);
2121 process_one_work(worker, work);
2126 * worker_thread - the worker thread function
2129 * The worker thread function. All workers belong to a worker_pool -
2130 * either a per-cpu one or dynamic unbound one. These workers process all
2131 * work items regardless of their specific target workqueue. The only
2132 * exception is work items which belong to workqueues with a rescuer which
2133 * will be explained in rescuer_thread().
2137 static int worker_thread(void *__worker)
2139 struct worker *worker = __worker;
2140 struct worker_pool *pool = worker->pool;
2142 /* tell the scheduler that this is a workqueue worker */
2143 worker->task->flags |= PF_WQ_WORKER;
2145 spin_lock_irq(&pool->lock);
2147 /* am I supposed to die? */
2148 if (unlikely(worker->flags & WORKER_DIE)) {
2149 spin_unlock_irq(&pool->lock);
2150 WARN_ON_ONCE(!list_empty(&worker->entry));
2151 worker->task->flags &= ~PF_WQ_WORKER;
2153 set_task_comm(worker->task, "kworker/dying");
2154 ida_simple_remove(&pool->worker_ida, worker->id);
2155 worker_detach_from_pool(worker, pool);
2160 worker_leave_idle(worker);
2162 /* no more worker necessary? */
2163 if (!need_more_worker(pool))
2166 /* do we need to manage? */
2167 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2171 * ->scheduled list can only be filled while a worker is
2172 * preparing to process a work or actually processing it.
2173 * Make sure nobody diddled with it while I was sleeping.
2175 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2178 * Finish PREP stage. We're guaranteed to have at least one idle
2179 * worker or that someone else has already assumed the manager
2180 * role. This is where @worker starts participating in concurrency
2181 * management if applicable and concurrency management is restored
2182 * after being rebound. See rebind_workers() for details.
2184 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2187 struct work_struct *work =
2188 list_first_entry(&pool->worklist,
2189 struct work_struct, entry);
2191 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2192 /* optimization path, not strictly necessary */
2193 process_one_work(worker, work);
2194 if (unlikely(!list_empty(&worker->scheduled)))
2195 process_scheduled_works(worker);
2197 move_linked_works(work, &worker->scheduled, NULL);
2198 process_scheduled_works(worker);
2200 } while (keep_working(pool));
2202 worker_set_flags(worker, WORKER_PREP);
2205 * pool->lock is held and there's no work to process and no need to
2206 * manage, sleep. Workers are woken up only while holding
2207 * pool->lock or from local cpu, so setting the current state
2208 * before releasing pool->lock is enough to prevent losing any
2211 worker_enter_idle(worker);
2212 __set_current_state(TASK_INTERRUPTIBLE);
2213 spin_unlock_irq(&pool->lock);
2219 * rescuer_thread - the rescuer thread function
2222 * Workqueue rescuer thread function. There's one rescuer for each
2223 * workqueue which has WQ_MEM_RECLAIM set.
2225 * Regular work processing on a pool may block trying to create a new
2226 * worker which uses GFP_KERNEL allocation which has slight chance of
2227 * developing into deadlock if some works currently on the same queue
2228 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2229 * the problem rescuer solves.
2231 * When such condition is possible, the pool summons rescuers of all
2232 * workqueues which have works queued on the pool and let them process
2233 * those works so that forward progress can be guaranteed.
2235 * This should happen rarely.
2239 static int rescuer_thread(void *__rescuer)
2241 struct worker *rescuer = __rescuer;
2242 struct workqueue_struct *wq = rescuer->rescue_wq;
2243 struct list_head *scheduled = &rescuer->scheduled;
2246 set_user_nice(current, RESCUER_NICE_LEVEL);
2249 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2250 * doesn't participate in concurrency management.
2252 rescuer->task->flags |= PF_WQ_WORKER;
2254 set_current_state(TASK_INTERRUPTIBLE);
2257 * By the time the rescuer is requested to stop, the workqueue
2258 * shouldn't have any work pending, but @wq->maydays may still have
2259 * pwq(s) queued. This can happen by non-rescuer workers consuming
2260 * all the work items before the rescuer got to them. Go through
2261 * @wq->maydays processing before acting on should_stop so that the
2262 * list is always empty on exit.
2264 should_stop = kthread_should_stop();
2266 /* see whether any pwq is asking for help */
2267 spin_lock_irq(&wq_mayday_lock);
2269 while (!list_empty(&wq->maydays)) {
2270 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2271 struct pool_workqueue, mayday_node);
2272 struct worker_pool *pool = pwq->pool;
2273 struct work_struct *work, *n;
2275 __set_current_state(TASK_RUNNING);
2276 list_del_init(&pwq->mayday_node);
2278 spin_unlock_irq(&wq_mayday_lock);
2280 worker_attach_to_pool(rescuer, pool);
2282 spin_lock_irq(&pool->lock);
2283 rescuer->pool = pool;
2286 * Slurp in all works issued via this workqueue and
2289 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2290 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2291 if (get_work_pwq(work) == pwq)
2292 move_linked_works(work, scheduled, &n);
2294 process_scheduled_works(rescuer);
2295 spin_unlock_irq(&pool->lock);
2297 worker_detach_from_pool(rescuer, pool);
2299 spin_lock_irq(&pool->lock);
2302 * Put the reference grabbed by send_mayday(). @pool won't
2303 * go away while we're holding its lock.
2308 * Leave this pool. If need_more_worker() is %true, notify a
2309 * regular worker; otherwise, we end up with 0 concurrency
2310 * and stalling the execution.
2312 if (need_more_worker(pool))
2313 wake_up_worker(pool);
2315 rescuer->pool = NULL;
2316 spin_unlock(&pool->lock);
2317 spin_lock(&wq_mayday_lock);
2320 spin_unlock_irq(&wq_mayday_lock);
2323 __set_current_state(TASK_RUNNING);
2324 rescuer->task->flags &= ~PF_WQ_WORKER;
2328 /* rescuers should never participate in concurrency management */
2329 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2335 struct work_struct work;
2336 struct completion done;
2339 static void wq_barrier_func(struct work_struct *work)
2341 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2342 complete(&barr->done);
2346 * insert_wq_barrier - insert a barrier work
2347 * @pwq: pwq to insert barrier into
2348 * @barr: wq_barrier to insert
2349 * @target: target work to attach @barr to
2350 * @worker: worker currently executing @target, NULL if @target is not executing
2352 * @barr is linked to @target such that @barr is completed only after
2353 * @target finishes execution. Please note that the ordering
2354 * guarantee is observed only with respect to @target and on the local
2357 * Currently, a queued barrier can't be canceled. This is because
2358 * try_to_grab_pending() can't determine whether the work to be
2359 * grabbed is at the head of the queue and thus can't clear LINKED
2360 * flag of the previous work while there must be a valid next work
2361 * after a work with LINKED flag set.
2363 * Note that when @worker is non-NULL, @target may be modified
2364 * underneath us, so we can't reliably determine pwq from @target.
2367 * spin_lock_irq(pool->lock).
2369 static void insert_wq_barrier(struct pool_workqueue *pwq,
2370 struct wq_barrier *barr,
2371 struct work_struct *target, struct worker *worker)
2373 struct list_head *head;
2374 unsigned int linked = 0;
2377 * debugobject calls are safe here even with pool->lock locked
2378 * as we know for sure that this will not trigger any of the
2379 * checks and call back into the fixup functions where we
2382 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2383 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2384 init_completion(&barr->done);
2387 * If @target is currently being executed, schedule the
2388 * barrier to the worker; otherwise, put it after @target.
2391 head = worker->scheduled.next;
2393 unsigned long *bits = work_data_bits(target);
2395 head = target->entry.next;
2396 /* there can already be other linked works, inherit and set */
2397 linked = *bits & WORK_STRUCT_LINKED;
2398 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2401 debug_work_activate(&barr->work);
2402 insert_work(pwq, &barr->work, head,
2403 work_color_to_flags(WORK_NO_COLOR) | linked);
2407 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2408 * @wq: workqueue being flushed
2409 * @flush_color: new flush color, < 0 for no-op
2410 * @work_color: new work color, < 0 for no-op
2412 * Prepare pwqs for workqueue flushing.
2414 * If @flush_color is non-negative, flush_color on all pwqs should be
2415 * -1. If no pwq has in-flight commands at the specified color, all
2416 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2417 * has in flight commands, its pwq->flush_color is set to
2418 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2419 * wakeup logic is armed and %true is returned.
2421 * The caller should have initialized @wq->first_flusher prior to
2422 * calling this function with non-negative @flush_color. If
2423 * @flush_color is negative, no flush color update is done and %false
2426 * If @work_color is non-negative, all pwqs should have the same
2427 * work_color which is previous to @work_color and all will be
2428 * advanced to @work_color.
2431 * mutex_lock(wq->mutex).
2434 * %true if @flush_color >= 0 and there's something to flush. %false
2437 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2438 int flush_color, int work_color)
2441 struct pool_workqueue *pwq;
2443 if (flush_color >= 0) {
2444 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2445 atomic_set(&wq->nr_pwqs_to_flush, 1);
2448 for_each_pwq(pwq, wq) {
2449 struct worker_pool *pool = pwq->pool;
2451 spin_lock_irq(&pool->lock);
2453 if (flush_color >= 0) {
2454 WARN_ON_ONCE(pwq->flush_color != -1);
2456 if (pwq->nr_in_flight[flush_color]) {
2457 pwq->flush_color = flush_color;
2458 atomic_inc(&wq->nr_pwqs_to_flush);
2463 if (work_color >= 0) {
2464 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2465 pwq->work_color = work_color;
2468 spin_unlock_irq(&pool->lock);
2471 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2472 complete(&wq->first_flusher->done);
2478 * flush_workqueue - ensure that any scheduled work has run to completion.
2479 * @wq: workqueue to flush
2481 * This function sleeps until all work items which were queued on entry
2482 * have finished execution, but it is not livelocked by new incoming ones.
2484 void flush_workqueue(struct workqueue_struct *wq)
2486 struct wq_flusher this_flusher = {
2487 .list = LIST_HEAD_INIT(this_flusher.list),
2489 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2493 lock_map_acquire(&wq->lockdep_map);
2494 lock_map_release(&wq->lockdep_map);
2496 mutex_lock(&wq->mutex);
2499 * Start-to-wait phase
2501 next_color = work_next_color(wq->work_color);
2503 if (next_color != wq->flush_color) {
2505 * Color space is not full. The current work_color
2506 * becomes our flush_color and work_color is advanced
2509 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2510 this_flusher.flush_color = wq->work_color;
2511 wq->work_color = next_color;
2513 if (!wq->first_flusher) {
2514 /* no flush in progress, become the first flusher */
2515 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2517 wq->first_flusher = &this_flusher;
2519 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2521 /* nothing to flush, done */
2522 wq->flush_color = next_color;
2523 wq->first_flusher = NULL;
2528 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2529 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2530 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2534 * Oops, color space is full, wait on overflow queue.
2535 * The next flush completion will assign us
2536 * flush_color and transfer to flusher_queue.
2538 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2541 mutex_unlock(&wq->mutex);
2543 wait_for_completion(&this_flusher.done);
2546 * Wake-up-and-cascade phase
2548 * First flushers are responsible for cascading flushes and
2549 * handling overflow. Non-first flushers can simply return.
2551 if (wq->first_flusher != &this_flusher)
2554 mutex_lock(&wq->mutex);
2556 /* we might have raced, check again with mutex held */
2557 if (wq->first_flusher != &this_flusher)
2560 wq->first_flusher = NULL;
2562 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2563 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2566 struct wq_flusher *next, *tmp;
2568 /* complete all the flushers sharing the current flush color */
2569 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2570 if (next->flush_color != wq->flush_color)
2572 list_del_init(&next->list);
2573 complete(&next->done);
2576 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2577 wq->flush_color != work_next_color(wq->work_color));
2579 /* this flush_color is finished, advance by one */
2580 wq->flush_color = work_next_color(wq->flush_color);
2582 /* one color has been freed, handle overflow queue */
2583 if (!list_empty(&wq->flusher_overflow)) {
2585 * Assign the same color to all overflowed
2586 * flushers, advance work_color and append to
2587 * flusher_queue. This is the start-to-wait
2588 * phase for these overflowed flushers.
2590 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2591 tmp->flush_color = wq->work_color;
2593 wq->work_color = work_next_color(wq->work_color);
2595 list_splice_tail_init(&wq->flusher_overflow,
2596 &wq->flusher_queue);
2597 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2600 if (list_empty(&wq->flusher_queue)) {
2601 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2606 * Need to flush more colors. Make the next flusher
2607 * the new first flusher and arm pwqs.
2609 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2610 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2612 list_del_init(&next->list);
2613 wq->first_flusher = next;
2615 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2619 * Meh... this color is already done, clear first
2620 * flusher and repeat cascading.
2622 wq->first_flusher = NULL;
2626 mutex_unlock(&wq->mutex);
2628 EXPORT_SYMBOL_GPL(flush_workqueue);
2631 * drain_workqueue - drain a workqueue
2632 * @wq: workqueue to drain
2634 * Wait until the workqueue becomes empty. While draining is in progress,
2635 * only chain queueing is allowed. IOW, only currently pending or running
2636 * work items on @wq can queue further work items on it. @wq is flushed
2637 * repeatedly until it becomes empty. The number of flushing is detemined
2638 * by the depth of chaining and should be relatively short. Whine if it
2641 void drain_workqueue(struct workqueue_struct *wq)
2643 unsigned int flush_cnt = 0;
2644 struct pool_workqueue *pwq;
2647 * __queue_work() needs to test whether there are drainers, is much
2648 * hotter than drain_workqueue() and already looks at @wq->flags.
2649 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2651 mutex_lock(&wq->mutex);
2652 if (!wq->nr_drainers++)
2653 wq->flags |= __WQ_DRAINING;
2654 mutex_unlock(&wq->mutex);
2656 flush_workqueue(wq);
2658 mutex_lock(&wq->mutex);
2660 for_each_pwq(pwq, wq) {
2663 spin_lock_irq(&pwq->pool->lock);
2664 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2665 spin_unlock_irq(&pwq->pool->lock);
2670 if (++flush_cnt == 10 ||
2671 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2672 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2673 wq->name, flush_cnt);
2675 mutex_unlock(&wq->mutex);
2679 if (!--wq->nr_drainers)
2680 wq->flags &= ~__WQ_DRAINING;
2681 mutex_unlock(&wq->mutex);
2683 EXPORT_SYMBOL_GPL(drain_workqueue);
2685 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2687 struct worker *worker = NULL;
2688 struct worker_pool *pool;
2689 struct pool_workqueue *pwq;
2693 local_irq_disable();
2694 pool = get_work_pool(work);
2700 spin_lock(&pool->lock);
2701 /* see the comment in try_to_grab_pending() with the same code */
2702 pwq = get_work_pwq(work);
2704 if (unlikely(pwq->pool != pool))
2707 worker = find_worker_executing_work(pool, work);
2710 pwq = worker->current_pwq;
2713 insert_wq_barrier(pwq, barr, work, worker);
2714 spin_unlock_irq(&pool->lock);
2717 * If @max_active is 1 or rescuer is in use, flushing another work
2718 * item on the same workqueue may lead to deadlock. Make sure the
2719 * flusher is not running on the same workqueue by verifying write
2722 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2723 lock_map_acquire(&pwq->wq->lockdep_map);
2725 lock_map_acquire_read(&pwq->wq->lockdep_map);
2726 lock_map_release(&pwq->wq->lockdep_map);
2730 spin_unlock_irq(&pool->lock);
2735 * flush_work - wait for a work to finish executing the last queueing instance
2736 * @work: the work to flush
2738 * Wait until @work has finished execution. @work is guaranteed to be idle
2739 * on return if it hasn't been requeued since flush started.
2742 * %true if flush_work() waited for the work to finish execution,
2743 * %false if it was already idle.
2745 bool flush_work(struct work_struct *work)
2747 struct wq_barrier barr;
2749 lock_map_acquire(&work->lockdep_map);
2750 lock_map_release(&work->lockdep_map);
2752 if (start_flush_work(work, &barr)) {
2753 wait_for_completion(&barr.done);
2754 destroy_work_on_stack(&barr.work);
2760 EXPORT_SYMBOL_GPL(flush_work);
2762 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2764 unsigned long flags;
2768 ret = try_to_grab_pending(work, is_dwork, &flags);
2770 * If someone else is canceling, wait for the same event it
2771 * would be waiting for before retrying.
2773 if (unlikely(ret == -ENOENT))
2775 } while (unlikely(ret < 0));
2777 /* tell other tasks trying to grab @work to back off */
2778 mark_work_canceling(work);
2779 local_irq_restore(flags);
2782 clear_work_data(work);
2787 * cancel_work_sync - cancel a work and wait for it to finish
2788 * @work: the work to cancel
2790 * Cancel @work and wait for its execution to finish. This function
2791 * can be used even if the work re-queues itself or migrates to
2792 * another workqueue. On return from this function, @work is
2793 * guaranteed to be not pending or executing on any CPU.
2795 * cancel_work_sync(&delayed_work->work) must not be used for
2796 * delayed_work's. Use cancel_delayed_work_sync() instead.
2798 * The caller must ensure that the workqueue on which @work was last
2799 * queued can't be destroyed before this function returns.
2802 * %true if @work was pending, %false otherwise.
2804 bool cancel_work_sync(struct work_struct *work)
2806 return __cancel_work_timer(work, false);
2808 EXPORT_SYMBOL_GPL(cancel_work_sync);
2811 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2812 * @dwork: the delayed work to flush
2814 * Delayed timer is cancelled and the pending work is queued for
2815 * immediate execution. Like flush_work(), this function only
2816 * considers the last queueing instance of @dwork.
2819 * %true if flush_work() waited for the work to finish execution,
2820 * %false if it was already idle.
2822 bool flush_delayed_work(struct delayed_work *dwork)
2824 local_irq_disable();
2825 if (del_timer_sync(&dwork->timer))
2826 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2828 return flush_work(&dwork->work);
2830 EXPORT_SYMBOL(flush_delayed_work);
2833 * cancel_delayed_work - cancel a delayed work
2834 * @dwork: delayed_work to cancel
2836 * Kill off a pending delayed_work.
2838 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2842 * The work callback function may still be running on return, unless
2843 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2844 * use cancel_delayed_work_sync() to wait on it.
2846 * This function is safe to call from any context including IRQ handler.
2848 bool cancel_delayed_work(struct delayed_work *dwork)
2850 unsigned long flags;
2854 ret = try_to_grab_pending(&dwork->work, true, &flags);
2855 } while (unlikely(ret == -EAGAIN));
2857 if (unlikely(ret < 0))
2860 set_work_pool_and_clear_pending(&dwork->work,
2861 get_work_pool_id(&dwork->work));
2862 local_irq_restore(flags);
2865 EXPORT_SYMBOL(cancel_delayed_work);
2868 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2869 * @dwork: the delayed work cancel
2871 * This is cancel_work_sync() for delayed works.
2874 * %true if @dwork was pending, %false otherwise.
2876 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2878 return __cancel_work_timer(&dwork->work, true);
2880 EXPORT_SYMBOL(cancel_delayed_work_sync);
2883 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2884 * @func: the function to call
2886 * schedule_on_each_cpu() executes @func on each online CPU using the
2887 * system workqueue and blocks until all CPUs have completed.
2888 * schedule_on_each_cpu() is very slow.
2891 * 0 on success, -errno on failure.
2893 int schedule_on_each_cpu(work_func_t func)
2896 struct work_struct __percpu *works;
2898 works = alloc_percpu(struct work_struct);
2904 for_each_online_cpu(cpu) {
2905 struct work_struct *work = per_cpu_ptr(works, cpu);
2907 INIT_WORK(work, func);
2908 schedule_work_on(cpu, work);
2911 for_each_online_cpu(cpu)
2912 flush_work(per_cpu_ptr(works, cpu));
2920 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2922 * Forces execution of the kernel-global workqueue and blocks until its
2925 * Think twice before calling this function! It's very easy to get into
2926 * trouble if you don't take great care. Either of the following situations
2927 * will lead to deadlock:
2929 * One of the work items currently on the workqueue needs to acquire
2930 * a lock held by your code or its caller.
2932 * Your code is running in the context of a work routine.
2934 * They will be detected by lockdep when they occur, but the first might not
2935 * occur very often. It depends on what work items are on the workqueue and
2936 * what locks they need, which you have no control over.
2938 * In most situations flushing the entire workqueue is overkill; you merely
2939 * need to know that a particular work item isn't queued and isn't running.
2940 * In such cases you should use cancel_delayed_work_sync() or
2941 * cancel_work_sync() instead.
2943 void flush_scheduled_work(void)
2945 flush_workqueue(system_wq);
2947 EXPORT_SYMBOL(flush_scheduled_work);
2950 * execute_in_process_context - reliably execute the routine with user context
2951 * @fn: the function to execute
2952 * @ew: guaranteed storage for the execute work structure (must
2953 * be available when the work executes)
2955 * Executes the function immediately if process context is available,
2956 * otherwise schedules the function for delayed execution.
2958 * Return: 0 - function was executed
2959 * 1 - function was scheduled for execution
2961 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2963 if (!in_interrupt()) {
2968 INIT_WORK(&ew->work, fn);
2969 schedule_work(&ew->work);
2973 EXPORT_SYMBOL_GPL(execute_in_process_context);
2977 * Workqueues with WQ_SYSFS flag set is visible to userland via
2978 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
2979 * following attributes.
2981 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
2982 * max_active RW int : maximum number of in-flight work items
2984 * Unbound workqueues have the following extra attributes.
2986 * id RO int : the associated pool ID
2987 * nice RW int : nice value of the workers
2988 * cpumask RW mask : bitmask of allowed CPUs for the workers
2991 struct workqueue_struct *wq;
2995 static struct workqueue_struct *dev_to_wq(struct device *dev)
2997 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3002 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3005 struct workqueue_struct *wq = dev_to_wq(dev);
3007 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3009 static DEVICE_ATTR_RO(per_cpu);
3011 static ssize_t max_active_show(struct device *dev,
3012 struct device_attribute *attr, char *buf)
3014 struct workqueue_struct *wq = dev_to_wq(dev);
3016 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3019 static ssize_t max_active_store(struct device *dev,
3020 struct device_attribute *attr, const char *buf,
3023 struct workqueue_struct *wq = dev_to_wq(dev);
3026 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3029 workqueue_set_max_active(wq, val);
3032 static DEVICE_ATTR_RW(max_active);
3034 static struct attribute *wq_sysfs_attrs[] = {
3035 &dev_attr_per_cpu.attr,
3036 &dev_attr_max_active.attr,
3039 ATTRIBUTE_GROUPS(wq_sysfs);
3041 static ssize_t wq_pool_ids_show(struct device *dev,
3042 struct device_attribute *attr, char *buf)
3044 struct workqueue_struct *wq = dev_to_wq(dev);
3045 const char *delim = "";
3046 int node, written = 0;
3048 rcu_read_lock_sched();
3049 for_each_node(node) {
3050 written += scnprintf(buf + written, PAGE_SIZE - written,
3051 "%s%d:%d", delim, node,
3052 unbound_pwq_by_node(wq, node)->pool->id);
3055 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3056 rcu_read_unlock_sched();
3061 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3064 struct workqueue_struct *wq = dev_to_wq(dev);
3067 mutex_lock(&wq->mutex);
3068 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3069 mutex_unlock(&wq->mutex);
3074 /* prepare workqueue_attrs for sysfs store operations */
3075 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3077 struct workqueue_attrs *attrs;
3079 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3083 mutex_lock(&wq->mutex);
3084 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3085 mutex_unlock(&wq->mutex);
3089 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3090 const char *buf, size_t count)
3092 struct workqueue_struct *wq = dev_to_wq(dev);
3093 struct workqueue_attrs *attrs;
3096 attrs = wq_sysfs_prep_attrs(wq);
3100 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3101 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
3102 ret = apply_workqueue_attrs(wq, attrs);
3106 free_workqueue_attrs(attrs);
3107 return ret ?: count;
3110 static ssize_t wq_cpumask_show(struct device *dev,
3111 struct device_attribute *attr, char *buf)
3113 struct workqueue_struct *wq = dev_to_wq(dev);
3116 mutex_lock(&wq->mutex);
3117 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3118 mutex_unlock(&wq->mutex);
3120 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3124 static ssize_t wq_cpumask_store(struct device *dev,
3125 struct device_attribute *attr,
3126 const char *buf, size_t count)
3128 struct workqueue_struct *wq = dev_to_wq(dev);
3129 struct workqueue_attrs *attrs;
3132 attrs = wq_sysfs_prep_attrs(wq);
3136 ret = cpumask_parse(buf, attrs->cpumask);
3138 ret = apply_workqueue_attrs(wq, attrs);
3140 free_workqueue_attrs(attrs);
3141 return ret ?: count;
3144 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3147 struct workqueue_struct *wq = dev_to_wq(dev);
3150 mutex_lock(&wq->mutex);
3151 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3152 !wq->unbound_attrs->no_numa);
3153 mutex_unlock(&wq->mutex);
3158 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3159 const char *buf, size_t count)
3161 struct workqueue_struct *wq = dev_to_wq(dev);
3162 struct workqueue_attrs *attrs;
3165 attrs = wq_sysfs_prep_attrs(wq);
3170 if (sscanf(buf, "%d", &v) == 1) {
3171 attrs->no_numa = !v;
3172 ret = apply_workqueue_attrs(wq, attrs);
3175 free_workqueue_attrs(attrs);
3176 return ret ?: count;
3179 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3180 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3181 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3182 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3183 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3187 static struct bus_type wq_subsys = {
3188 .name = "workqueue",
3189 .dev_groups = wq_sysfs_groups,
3192 static int __init wq_sysfs_init(void)
3194 return subsys_virtual_register(&wq_subsys, NULL);
3196 core_initcall(wq_sysfs_init);
3198 static void wq_device_release(struct device *dev)
3200 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3206 * workqueue_sysfs_register - make a workqueue visible in sysfs
3207 * @wq: the workqueue to register
3209 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3210 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3211 * which is the preferred method.
3213 * Workqueue user should use this function directly iff it wants to apply
3214 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3215 * apply_workqueue_attrs() may race against userland updating the
3218 * Return: 0 on success, -errno on failure.
3220 int workqueue_sysfs_register(struct workqueue_struct *wq)
3222 struct wq_device *wq_dev;
3226 * Adjusting max_active or creating new pwqs by applyting
3227 * attributes breaks ordering guarantee. Disallow exposing ordered
3230 if (WARN_ON(wq->flags & __WQ_ORDERED))
3233 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3238 wq_dev->dev.bus = &wq_subsys;
3239 wq_dev->dev.init_name = wq->name;
3240 wq_dev->dev.release = wq_device_release;
3243 * unbound_attrs are created separately. Suppress uevent until
3244 * everything is ready.
3246 dev_set_uevent_suppress(&wq_dev->dev, true);
3248 ret = device_register(&wq_dev->dev);
3255 if (wq->flags & WQ_UNBOUND) {
3256 struct device_attribute *attr;
3258 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3259 ret = device_create_file(&wq_dev->dev, attr);
3261 device_unregister(&wq_dev->dev);
3268 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3273 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3274 * @wq: the workqueue to unregister
3276 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3278 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3280 struct wq_device *wq_dev = wq->wq_dev;
3286 device_unregister(&wq_dev->dev);
3288 #else /* CONFIG_SYSFS */
3289 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3290 #endif /* CONFIG_SYSFS */
3293 * free_workqueue_attrs - free a workqueue_attrs
3294 * @attrs: workqueue_attrs to free
3296 * Undo alloc_workqueue_attrs().
3298 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3301 free_cpumask_var(attrs->cpumask);
3307 * alloc_workqueue_attrs - allocate a workqueue_attrs
3308 * @gfp_mask: allocation mask to use
3310 * Allocate a new workqueue_attrs, initialize with default settings and
3313 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3315 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3317 struct workqueue_attrs *attrs;
3319 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3322 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3325 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3328 free_workqueue_attrs(attrs);
3332 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3333 const struct workqueue_attrs *from)
3335 to->nice = from->nice;
3336 cpumask_copy(to->cpumask, from->cpumask);
3338 * Unlike hash and equality test, this function doesn't ignore
3339 * ->no_numa as it is used for both pool and wq attrs. Instead,
3340 * get_unbound_pool() explicitly clears ->no_numa after copying.
3342 to->no_numa = from->no_numa;
3345 /* hash value of the content of @attr */
3346 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3350 hash = jhash_1word(attrs->nice, hash);
3351 hash = jhash(cpumask_bits(attrs->cpumask),
3352 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3356 /* content equality test */
3357 static bool wqattrs_equal(const struct workqueue_attrs *a,
3358 const struct workqueue_attrs *b)
3360 if (a->nice != b->nice)
3362 if (!cpumask_equal(a->cpumask, b->cpumask))
3368 * init_worker_pool - initialize a newly zalloc'd worker_pool
3369 * @pool: worker_pool to initialize
3371 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3373 * Return: 0 on success, -errno on failure. Even on failure, all fields
3374 * inside @pool proper are initialized and put_unbound_pool() can be called
3375 * on @pool safely to release it.
3377 static int init_worker_pool(struct worker_pool *pool)
3379 spin_lock_init(&pool->lock);
3382 pool->node = NUMA_NO_NODE;
3383 pool->flags |= POOL_DISASSOCIATED;
3384 INIT_LIST_HEAD(&pool->worklist);
3385 INIT_LIST_HEAD(&pool->idle_list);
3386 hash_init(pool->busy_hash);
3388 init_timer_deferrable(&pool->idle_timer);
3389 pool->idle_timer.function = idle_worker_timeout;
3390 pool->idle_timer.data = (unsigned long)pool;
3392 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3393 (unsigned long)pool);
3395 mutex_init(&pool->manager_arb);
3396 mutex_init(&pool->attach_mutex);
3397 INIT_LIST_HEAD(&pool->workers);
3399 ida_init(&pool->worker_ida);
3400 INIT_HLIST_NODE(&pool->hash_node);
3403 /* shouldn't fail above this point */
3404 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3410 static void rcu_free_pool(struct rcu_head *rcu)
3412 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3414 ida_destroy(&pool->worker_ida);
3415 free_workqueue_attrs(pool->attrs);
3420 * put_unbound_pool - put a worker_pool
3421 * @pool: worker_pool to put
3423 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3424 * safe manner. get_unbound_pool() calls this function on its failure path
3425 * and this function should be able to release pools which went through,
3426 * successfully or not, init_worker_pool().
3428 * Should be called with wq_pool_mutex held.
3430 static void put_unbound_pool(struct worker_pool *pool)
3432 DECLARE_COMPLETION_ONSTACK(detach_completion);
3433 struct worker *worker;
3435 lockdep_assert_held(&wq_pool_mutex);
3441 if (WARN_ON(!(pool->cpu < 0)) ||
3442 WARN_ON(!list_empty(&pool->worklist)))
3445 /* release id and unhash */
3447 idr_remove(&worker_pool_idr, pool->id);
3448 hash_del(&pool->hash_node);
3451 * Become the manager and destroy all workers. Grabbing
3452 * manager_arb prevents @pool's workers from blocking on
3455 mutex_lock(&pool->manager_arb);
3457 spin_lock_irq(&pool->lock);
3458 while ((worker = first_idle_worker(pool)))
3459 destroy_worker(worker);
3460 WARN_ON(pool->nr_workers || pool->nr_idle);
3461 spin_unlock_irq(&pool->lock);
3463 mutex_lock(&pool->attach_mutex);
3464 if (!list_empty(&pool->workers))
3465 pool->detach_completion = &detach_completion;
3466 mutex_unlock(&pool->attach_mutex);
3468 if (pool->detach_completion)
3469 wait_for_completion(pool->detach_completion);
3471 mutex_unlock(&pool->manager_arb);
3473 /* shut down the timers */
3474 del_timer_sync(&pool->idle_timer);
3475 del_timer_sync(&pool->mayday_timer);
3477 /* sched-RCU protected to allow dereferences from get_work_pool() */
3478 call_rcu_sched(&pool->rcu, rcu_free_pool);
3482 * get_unbound_pool - get a worker_pool with the specified attributes
3483 * @attrs: the attributes of the worker_pool to get
3485 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3486 * reference count and return it. If there already is a matching
3487 * worker_pool, it will be used; otherwise, this function attempts to
3490 * Should be called with wq_pool_mutex held.
3492 * Return: On success, a worker_pool with the same attributes as @attrs.
3493 * On failure, %NULL.
3495 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3497 u32 hash = wqattrs_hash(attrs);
3498 struct worker_pool *pool;
3501 lockdep_assert_held(&wq_pool_mutex);
3503 /* do we already have a matching pool? */
3504 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3505 if (wqattrs_equal(pool->attrs, attrs)) {
3511 /* nope, create a new one */
3512 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3513 if (!pool || init_worker_pool(pool) < 0)
3516 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3517 copy_workqueue_attrs(pool->attrs, attrs);
3520 * no_numa isn't a worker_pool attribute, always clear it. See
3521 * 'struct workqueue_attrs' comments for detail.
3523 pool->attrs->no_numa = false;
3525 /* if cpumask is contained inside a NUMA node, we belong to that node */
3526 if (wq_numa_enabled) {
3527 for_each_node(node) {
3528 if (cpumask_subset(pool->attrs->cpumask,
3529 wq_numa_possible_cpumask[node])) {
3536 if (worker_pool_assign_id(pool) < 0)
3539 /* create and start the initial worker */
3540 if (create_and_start_worker(pool) < 0)
3544 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3549 put_unbound_pool(pool);
3553 static void rcu_free_pwq(struct rcu_head *rcu)
3555 kmem_cache_free(pwq_cache,
3556 container_of(rcu, struct pool_workqueue, rcu));
3560 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3561 * and needs to be destroyed.
3563 static void pwq_unbound_release_workfn(struct work_struct *work)
3565 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3566 unbound_release_work);
3567 struct workqueue_struct *wq = pwq->wq;
3568 struct worker_pool *pool = pwq->pool;
3571 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3575 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3576 * necessary on release but do it anyway. It's easier to verify
3577 * and consistent with the linking path.
3579 mutex_lock(&wq->mutex);
3580 list_del_rcu(&pwq->pwqs_node);
3581 is_last = list_empty(&wq->pwqs);
3582 mutex_unlock(&wq->mutex);
3584 mutex_lock(&wq_pool_mutex);
3585 put_unbound_pool(pool);
3586 mutex_unlock(&wq_pool_mutex);
3588 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3591 * If we're the last pwq going away, @wq is already dead and no one
3592 * is gonna access it anymore. Free it.
3595 free_workqueue_attrs(wq->unbound_attrs);
3601 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3602 * @pwq: target pool_workqueue
3604 * If @pwq isn't freezing, set @pwq->max_active to the associated
3605 * workqueue's saved_max_active and activate delayed work items
3606 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3608 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3610 struct workqueue_struct *wq = pwq->wq;
3611 bool freezable = wq->flags & WQ_FREEZABLE;
3613 /* for @wq->saved_max_active */
3614 lockdep_assert_held(&wq->mutex);
3616 /* fast exit for non-freezable wqs */
3617 if (!freezable && pwq->max_active == wq->saved_max_active)
3620 spin_lock_irq(&pwq->pool->lock);
3623 * During [un]freezing, the caller is responsible for ensuring that
3624 * this function is called at least once after @workqueue_freezing
3625 * is updated and visible.
3627 if (!freezable || !workqueue_freezing) {
3628 pwq->max_active = wq->saved_max_active;
3630 while (!list_empty(&pwq->delayed_works) &&
3631 pwq->nr_active < pwq->max_active)
3632 pwq_activate_first_delayed(pwq);
3635 * Need to kick a worker after thawed or an unbound wq's
3636 * max_active is bumped. It's a slow path. Do it always.
3638 wake_up_worker(pwq->pool);
3640 pwq->max_active = 0;
3643 spin_unlock_irq(&pwq->pool->lock);
3646 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3647 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3648 struct worker_pool *pool)
3650 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3652 memset(pwq, 0, sizeof(*pwq));
3656 pwq->flush_color = -1;
3658 INIT_LIST_HEAD(&pwq->delayed_works);
3659 INIT_LIST_HEAD(&pwq->pwqs_node);
3660 INIT_LIST_HEAD(&pwq->mayday_node);
3661 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3664 /* sync @pwq with the current state of its associated wq and link it */
3665 static void link_pwq(struct pool_workqueue *pwq)
3667 struct workqueue_struct *wq = pwq->wq;
3669 lockdep_assert_held(&wq->mutex);
3671 /* may be called multiple times, ignore if already linked */
3672 if (!list_empty(&pwq->pwqs_node))
3676 * Set the matching work_color. This is synchronized with
3677 * wq->mutex to avoid confusing flush_workqueue().
3679 pwq->work_color = wq->work_color;
3681 /* sync max_active to the current setting */
3682 pwq_adjust_max_active(pwq);
3685 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3688 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3689 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3690 const struct workqueue_attrs *attrs)
3692 struct worker_pool *pool;
3693 struct pool_workqueue *pwq;
3695 lockdep_assert_held(&wq_pool_mutex);
3697 pool = get_unbound_pool(attrs);
3701 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3703 put_unbound_pool(pool);
3707 init_pwq(pwq, wq, pool);
3711 /* undo alloc_unbound_pwq(), used only in the error path */
3712 static void free_unbound_pwq(struct pool_workqueue *pwq)
3714 lockdep_assert_held(&wq_pool_mutex);
3717 put_unbound_pool(pwq->pool);
3718 kmem_cache_free(pwq_cache, pwq);
3723 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3724 * @attrs: the wq_attrs of interest
3725 * @node: the target NUMA node
3726 * @cpu_going_down: if >= 0, the CPU to consider as offline
3727 * @cpumask: outarg, the resulting cpumask
3729 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3730 * @cpu_going_down is >= 0, that cpu is considered offline during
3731 * calculation. The result is stored in @cpumask.
3733 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3734 * enabled and @node has online CPUs requested by @attrs, the returned
3735 * cpumask is the intersection of the possible CPUs of @node and
3738 * The caller is responsible for ensuring that the cpumask of @node stays
3741 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3744 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3745 int cpu_going_down, cpumask_t *cpumask)
3747 if (!wq_numa_enabled || attrs->no_numa)
3750 /* does @node have any online CPUs @attrs wants? */
3751 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3752 if (cpu_going_down >= 0)
3753 cpumask_clear_cpu(cpu_going_down, cpumask);
3755 if (cpumask_empty(cpumask))
3758 /* yeap, return possible CPUs in @node that @attrs wants */
3759 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3760 return !cpumask_equal(cpumask, attrs->cpumask);
3763 cpumask_copy(cpumask, attrs->cpumask);
3767 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3768 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3770 struct pool_workqueue *pwq)
3772 struct pool_workqueue *old_pwq;
3774 lockdep_assert_held(&wq->mutex);
3776 /* link_pwq() can handle duplicate calls */
3779 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3780 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3785 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3786 * @wq: the target workqueue
3787 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3789 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3790 * machines, this function maps a separate pwq to each NUMA node with
3791 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3792 * NUMA node it was issued on. Older pwqs are released as in-flight work
3793 * items finish. Note that a work item which repeatedly requeues itself
3794 * back-to-back will stay on its current pwq.
3796 * Performs GFP_KERNEL allocations.
3798 * Return: 0 on success and -errno on failure.
3800 int apply_workqueue_attrs(struct workqueue_struct *wq,
3801 const struct workqueue_attrs *attrs)
3803 struct workqueue_attrs *new_attrs, *tmp_attrs;
3804 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3807 /* only unbound workqueues can change attributes */
3808 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3811 /* creating multiple pwqs breaks ordering guarantee */
3812 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3815 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3816 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3817 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3818 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3821 /* make a copy of @attrs and sanitize it */
3822 copy_workqueue_attrs(new_attrs, attrs);
3823 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3826 * We may create multiple pwqs with differing cpumasks. Make a
3827 * copy of @new_attrs which will be modified and used to obtain
3830 copy_workqueue_attrs(tmp_attrs, new_attrs);
3833 * CPUs should stay stable across pwq creations and installations.
3834 * Pin CPUs, determine the target cpumask for each node and create
3839 mutex_lock(&wq_pool_mutex);
3842 * If something goes wrong during CPU up/down, we'll fall back to
3843 * the default pwq covering whole @attrs->cpumask. Always create
3844 * it even if we don't use it immediately.
3846 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3850 for_each_node(node) {
3851 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3852 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3857 pwq_tbl[node] = dfl_pwq;
3861 mutex_unlock(&wq_pool_mutex);
3863 /* all pwqs have been created successfully, let's install'em */
3864 mutex_lock(&wq->mutex);
3866 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3868 /* save the previous pwq and install the new one */
3870 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3872 /* @dfl_pwq might not have been used, ensure it's linked */
3874 swap(wq->dfl_pwq, dfl_pwq);
3876 mutex_unlock(&wq->mutex);
3878 /* put the old pwqs */
3880 put_pwq_unlocked(pwq_tbl[node]);
3881 put_pwq_unlocked(dfl_pwq);
3887 free_workqueue_attrs(tmp_attrs);
3888 free_workqueue_attrs(new_attrs);
3893 free_unbound_pwq(dfl_pwq);
3895 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3896 free_unbound_pwq(pwq_tbl[node]);
3897 mutex_unlock(&wq_pool_mutex);
3905 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3906 * @wq: the target workqueue
3907 * @cpu: the CPU coming up or going down
3908 * @online: whether @cpu is coming up or going down
3910 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3911 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3914 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3915 * falls back to @wq->dfl_pwq which may not be optimal but is always
3918 * Note that when the last allowed CPU of a NUMA node goes offline for a
3919 * workqueue with a cpumask spanning multiple nodes, the workers which were
3920 * already executing the work items for the workqueue will lose their CPU
3921 * affinity and may execute on any CPU. This is similar to how per-cpu
3922 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3923 * affinity, it's the user's responsibility to flush the work item from
3926 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3929 int node = cpu_to_node(cpu);
3930 int cpu_off = online ? -1 : cpu;
3931 struct pool_workqueue *old_pwq = NULL, *pwq;
3932 struct workqueue_attrs *target_attrs;
3935 lockdep_assert_held(&wq_pool_mutex);
3937 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3941 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3942 * Let's use a preallocated one. The following buf is protected by
3943 * CPU hotplug exclusion.
3945 target_attrs = wq_update_unbound_numa_attrs_buf;
3946 cpumask = target_attrs->cpumask;
3948 mutex_lock(&wq->mutex);
3949 if (wq->unbound_attrs->no_numa)
3952 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3953 pwq = unbound_pwq_by_node(wq, node);
3956 * Let's determine what needs to be done. If the target cpumask is
3957 * different from wq's, we need to compare it to @pwq's and create
3958 * a new one if they don't match. If the target cpumask equals
3959 * wq's, the default pwq should be used.
3961 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
3962 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3968 mutex_unlock(&wq->mutex);
3970 /* create a new pwq */
3971 pwq = alloc_unbound_pwq(wq, target_attrs);
3973 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3975 mutex_lock(&wq->mutex);
3980 * Install the new pwq. As this function is called only from CPU
3981 * hotplug callbacks and applying a new attrs is wrapped with
3982 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3985 mutex_lock(&wq->mutex);
3986 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3990 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3991 get_pwq(wq->dfl_pwq);
3992 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3993 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3995 mutex_unlock(&wq->mutex);
3996 put_pwq_unlocked(old_pwq);
3999 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4001 bool highpri = wq->flags & WQ_HIGHPRI;
4004 if (!(wq->flags & WQ_UNBOUND)) {
4005 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4009 for_each_possible_cpu(cpu) {
4010 struct pool_workqueue *pwq =
4011 per_cpu_ptr(wq->cpu_pwqs, cpu);
4012 struct worker_pool *cpu_pools =
4013 per_cpu(cpu_worker_pools, cpu);
4015 init_pwq(pwq, wq, &cpu_pools[highpri]);
4017 mutex_lock(&wq->mutex);
4019 mutex_unlock(&wq->mutex);
4022 } else if (wq->flags & __WQ_ORDERED) {
4023 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4024 /* there should only be single pwq for ordering guarantee */
4025 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4026 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4027 "ordering guarantee broken for workqueue %s\n", wq->name);
4030 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4034 static int wq_clamp_max_active(int max_active, unsigned int flags,
4037 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4039 if (max_active < 1 || max_active > lim)
4040 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4041 max_active, name, 1, lim);
4043 return clamp_val(max_active, 1, lim);
4046 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4049 struct lock_class_key *key,
4050 const char *lock_name, ...)
4052 size_t tbl_size = 0;
4054 struct workqueue_struct *wq;
4055 struct pool_workqueue *pwq;
4057 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4058 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4059 flags |= WQ_UNBOUND;
4061 /* allocate wq and format name */
4062 if (flags & WQ_UNBOUND)
4063 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4065 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4069 if (flags & WQ_UNBOUND) {
4070 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4071 if (!wq->unbound_attrs)
4075 va_start(args, lock_name);
4076 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4079 max_active = max_active ?: WQ_DFL_ACTIVE;
4080 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4084 wq->saved_max_active = max_active;
4085 mutex_init(&wq->mutex);
4086 atomic_set(&wq->nr_pwqs_to_flush, 0);
4087 INIT_LIST_HEAD(&wq->pwqs);
4088 INIT_LIST_HEAD(&wq->flusher_queue);
4089 INIT_LIST_HEAD(&wq->flusher_overflow);
4090 INIT_LIST_HEAD(&wq->maydays);
4092 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4093 INIT_LIST_HEAD(&wq->list);
4095 if (alloc_and_link_pwqs(wq) < 0)
4099 * Workqueues which may be used during memory reclaim should
4100 * have a rescuer to guarantee forward progress.
4102 if (flags & WQ_MEM_RECLAIM) {
4103 struct worker *rescuer;
4105 rescuer = alloc_worker(NUMA_NO_NODE);
4109 rescuer->rescue_wq = wq;
4110 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4112 if (IS_ERR(rescuer->task)) {
4117 wq->rescuer = rescuer;
4118 rescuer->task->flags |= PF_NO_SETAFFINITY;
4119 wake_up_process(rescuer->task);
4122 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4126 * wq_pool_mutex protects global freeze state and workqueues list.
4127 * Grab it, adjust max_active and add the new @wq to workqueues
4130 mutex_lock(&wq_pool_mutex);
4132 mutex_lock(&wq->mutex);
4133 for_each_pwq(pwq, wq)
4134 pwq_adjust_max_active(pwq);
4135 mutex_unlock(&wq->mutex);
4137 list_add(&wq->list, &workqueues);
4139 mutex_unlock(&wq_pool_mutex);
4144 free_workqueue_attrs(wq->unbound_attrs);
4148 destroy_workqueue(wq);
4151 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4154 * destroy_workqueue - safely terminate a workqueue
4155 * @wq: target workqueue
4157 * Safely destroy a workqueue. All work currently pending will be done first.
4159 void destroy_workqueue(struct workqueue_struct *wq)
4161 struct pool_workqueue *pwq;
4164 /* drain it before proceeding with destruction */
4165 drain_workqueue(wq);
4168 mutex_lock(&wq->mutex);
4169 for_each_pwq(pwq, wq) {
4172 for (i = 0; i < WORK_NR_COLORS; i++) {
4173 if (WARN_ON(pwq->nr_in_flight[i])) {
4174 mutex_unlock(&wq->mutex);
4179 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4180 WARN_ON(pwq->nr_active) ||
4181 WARN_ON(!list_empty(&pwq->delayed_works))) {
4182 mutex_unlock(&wq->mutex);
4186 mutex_unlock(&wq->mutex);
4189 * wq list is used to freeze wq, remove from list after
4190 * flushing is complete in case freeze races us.
4192 mutex_lock(&wq_pool_mutex);
4193 list_del_init(&wq->list);
4194 mutex_unlock(&wq_pool_mutex);
4196 workqueue_sysfs_unregister(wq);
4199 kthread_stop(wq->rescuer->task);
4204 if (!(wq->flags & WQ_UNBOUND)) {
4206 * The base ref is never dropped on per-cpu pwqs. Directly
4207 * free the pwqs and wq.
4209 free_percpu(wq->cpu_pwqs);
4213 * We're the sole accessor of @wq at this point. Directly
4214 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4215 * @wq will be freed when the last pwq is released.
4217 for_each_node(node) {
4218 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4219 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4220 put_pwq_unlocked(pwq);
4224 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4225 * put. Don't access it afterwards.
4229 put_pwq_unlocked(pwq);
4232 EXPORT_SYMBOL_GPL(destroy_workqueue);
4235 * workqueue_set_max_active - adjust max_active of a workqueue
4236 * @wq: target workqueue
4237 * @max_active: new max_active value.
4239 * Set max_active of @wq to @max_active.
4242 * Don't call from IRQ context.
4244 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4246 struct pool_workqueue *pwq;
4248 /* disallow meddling with max_active for ordered workqueues */
4249 if (WARN_ON(wq->flags & __WQ_ORDERED))
4252 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4254 mutex_lock(&wq->mutex);
4256 wq->saved_max_active = max_active;
4258 for_each_pwq(pwq, wq)
4259 pwq_adjust_max_active(pwq);
4261 mutex_unlock(&wq->mutex);
4263 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4266 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4268 * Determine whether %current is a workqueue rescuer. Can be used from
4269 * work functions to determine whether it's being run off the rescuer task.
4271 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4273 bool current_is_workqueue_rescuer(void)
4275 struct worker *worker = current_wq_worker();
4277 return worker && worker->rescue_wq;
4281 * workqueue_congested - test whether a workqueue is congested
4282 * @cpu: CPU in question
4283 * @wq: target workqueue
4285 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4286 * no synchronization around this function and the test result is
4287 * unreliable and only useful as advisory hints or for debugging.
4289 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4290 * Note that both per-cpu and unbound workqueues may be associated with
4291 * multiple pool_workqueues which have separate congested states. A
4292 * workqueue being congested on one CPU doesn't mean the workqueue is also
4293 * contested on other CPUs / NUMA nodes.
4296 * %true if congested, %false otherwise.
4298 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4300 struct pool_workqueue *pwq;
4303 rcu_read_lock_sched();
4305 if (cpu == WORK_CPU_UNBOUND)
4306 cpu = smp_processor_id();
4308 if (!(wq->flags & WQ_UNBOUND))
4309 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4311 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4313 ret = !list_empty(&pwq->delayed_works);
4314 rcu_read_unlock_sched();
4318 EXPORT_SYMBOL_GPL(workqueue_congested);
4321 * work_busy - test whether a work is currently pending or running
4322 * @work: the work to be tested
4324 * Test whether @work is currently pending or running. There is no
4325 * synchronization around this function and the test result is
4326 * unreliable and only useful as advisory hints or for debugging.
4329 * OR'd bitmask of WORK_BUSY_* bits.
4331 unsigned int work_busy(struct work_struct *work)
4333 struct worker_pool *pool;
4334 unsigned long flags;
4335 unsigned int ret = 0;
4337 if (work_pending(work))
4338 ret |= WORK_BUSY_PENDING;
4340 local_irq_save(flags);
4341 pool = get_work_pool(work);
4343 spin_lock(&pool->lock);
4344 if (find_worker_executing_work(pool, work))
4345 ret |= WORK_BUSY_RUNNING;
4346 spin_unlock(&pool->lock);
4348 local_irq_restore(flags);
4352 EXPORT_SYMBOL_GPL(work_busy);
4355 * set_worker_desc - set description for the current work item
4356 * @fmt: printf-style format string
4357 * @...: arguments for the format string
4359 * This function can be called by a running work function to describe what
4360 * the work item is about. If the worker task gets dumped, this
4361 * information will be printed out together to help debugging. The
4362 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4364 void set_worker_desc(const char *fmt, ...)
4366 struct worker *worker = current_wq_worker();
4370 va_start(args, fmt);
4371 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4373 worker->desc_valid = true;
4378 * print_worker_info - print out worker information and description
4379 * @log_lvl: the log level to use when printing
4380 * @task: target task
4382 * If @task is a worker and currently executing a work item, print out the
4383 * name of the workqueue being serviced and worker description set with
4384 * set_worker_desc() by the currently executing work item.
4386 * This function can be safely called on any task as long as the
4387 * task_struct itself is accessible. While safe, this function isn't
4388 * synchronized and may print out mixups or garbages of limited length.
4390 void print_worker_info(const char *log_lvl, struct task_struct *task)
4392 work_func_t *fn = NULL;
4393 char name[WQ_NAME_LEN] = { };
4394 char desc[WORKER_DESC_LEN] = { };
4395 struct pool_workqueue *pwq = NULL;
4396 struct workqueue_struct *wq = NULL;
4397 bool desc_valid = false;
4398 struct worker *worker;
4400 if (!(task->flags & PF_WQ_WORKER))
4404 * This function is called without any synchronization and @task
4405 * could be in any state. Be careful with dereferences.
4407 worker = probe_kthread_data(task);
4410 * Carefully copy the associated workqueue's workfn and name. Keep
4411 * the original last '\0' in case the original contains garbage.
4413 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4414 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4415 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4416 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4418 /* copy worker description */
4419 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4421 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4423 if (fn || name[0] || desc[0]) {
4424 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4426 pr_cont(" (%s)", desc);
4434 * There are two challenges in supporting CPU hotplug. Firstly, there
4435 * are a lot of assumptions on strong associations among work, pwq and
4436 * pool which make migrating pending and scheduled works very
4437 * difficult to implement without impacting hot paths. Secondly,
4438 * worker pools serve mix of short, long and very long running works making
4439 * blocked draining impractical.
4441 * This is solved by allowing the pools to be disassociated from the CPU
4442 * running as an unbound one and allowing it to be reattached later if the
4443 * cpu comes back online.
4446 static void wq_unbind_fn(struct work_struct *work)
4448 int cpu = smp_processor_id();
4449 struct worker_pool *pool;
4450 struct worker *worker;
4452 for_each_cpu_worker_pool(pool, cpu) {
4453 mutex_lock(&pool->attach_mutex);
4454 spin_lock_irq(&pool->lock);
4457 * We've blocked all attach/detach operations. Make all workers
4458 * unbound and set DISASSOCIATED. Before this, all workers
4459 * except for the ones which are still executing works from
4460 * before the last CPU down must be on the cpu. After
4461 * this, they may become diasporas.
4463 for_each_pool_worker(worker, pool)
4464 worker->flags |= WORKER_UNBOUND;
4466 pool->flags |= POOL_DISASSOCIATED;
4468 spin_unlock_irq(&pool->lock);
4469 mutex_unlock(&pool->attach_mutex);
4472 * Call schedule() so that we cross rq->lock and thus can
4473 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4474 * This is necessary as scheduler callbacks may be invoked
4480 * Sched callbacks are disabled now. Zap nr_running.
4481 * After this, nr_running stays zero and need_more_worker()
4482 * and keep_working() are always true as long as the
4483 * worklist is not empty. This pool now behaves as an
4484 * unbound (in terms of concurrency management) pool which
4485 * are served by workers tied to the pool.
4487 atomic_set(&pool->nr_running, 0);
4490 * With concurrency management just turned off, a busy
4491 * worker blocking could lead to lengthy stalls. Kick off
4492 * unbound chain execution of currently pending work items.
4494 spin_lock_irq(&pool->lock);
4495 wake_up_worker(pool);
4496 spin_unlock_irq(&pool->lock);
4501 * rebind_workers - rebind all workers of a pool to the associated CPU
4502 * @pool: pool of interest
4504 * @pool->cpu is coming online. Rebind all workers to the CPU.
4506 static void rebind_workers(struct worker_pool *pool)
4508 struct worker *worker;
4510 lockdep_assert_held(&pool->attach_mutex);
4513 * Restore CPU affinity of all workers. As all idle workers should
4514 * be on the run-queue of the associated CPU before any local
4515 * wake-ups for concurrency management happen, restore CPU affinty
4516 * of all workers first and then clear UNBOUND. As we're called
4517 * from CPU_ONLINE, the following shouldn't fail.
4519 for_each_pool_worker(worker, pool)
4520 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4521 pool->attrs->cpumask) < 0);
4523 spin_lock_irq(&pool->lock);
4524 pool->flags &= ~POOL_DISASSOCIATED;
4526 for_each_pool_worker(worker, pool) {
4527 unsigned int worker_flags = worker->flags;
4530 * A bound idle worker should actually be on the runqueue
4531 * of the associated CPU for local wake-ups targeting it to
4532 * work. Kick all idle workers so that they migrate to the
4533 * associated CPU. Doing this in the same loop as
4534 * replacing UNBOUND with REBOUND is safe as no worker will
4535 * be bound before @pool->lock is released.
4537 if (worker_flags & WORKER_IDLE)
4538 wake_up_process(worker->task);
4541 * We want to clear UNBOUND but can't directly call
4542 * worker_clr_flags() or adjust nr_running. Atomically
4543 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4544 * @worker will clear REBOUND using worker_clr_flags() when
4545 * it initiates the next execution cycle thus restoring
4546 * concurrency management. Note that when or whether
4547 * @worker clears REBOUND doesn't affect correctness.
4549 * ACCESS_ONCE() is necessary because @worker->flags may be
4550 * tested without holding any lock in
4551 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4552 * fail incorrectly leading to premature concurrency
4553 * management operations.
4555 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4556 worker_flags |= WORKER_REBOUND;
4557 worker_flags &= ~WORKER_UNBOUND;
4558 ACCESS_ONCE(worker->flags) = worker_flags;
4561 spin_unlock_irq(&pool->lock);
4565 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4566 * @pool: unbound pool of interest
4567 * @cpu: the CPU which is coming up
4569 * An unbound pool may end up with a cpumask which doesn't have any online
4570 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4571 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4572 * online CPU before, cpus_allowed of all its workers should be restored.
4574 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4576 static cpumask_t cpumask;
4577 struct worker *worker;
4579 lockdep_assert_held(&pool->attach_mutex);
4581 /* is @cpu allowed for @pool? */
4582 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4585 /* is @cpu the only online CPU? */
4586 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4587 if (cpumask_weight(&cpumask) != 1)
4590 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4591 for_each_pool_worker(worker, pool)
4592 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4593 pool->attrs->cpumask) < 0);
4597 * Workqueues should be brought up before normal priority CPU notifiers.
4598 * This will be registered high priority CPU notifier.
4600 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4601 unsigned long action,
4604 int cpu = (unsigned long)hcpu;
4605 struct worker_pool *pool;
4606 struct workqueue_struct *wq;
4609 switch (action & ~CPU_TASKS_FROZEN) {
4610 case CPU_UP_PREPARE:
4611 for_each_cpu_worker_pool(pool, cpu) {
4612 if (pool->nr_workers)
4614 if (create_and_start_worker(pool) < 0)
4619 case CPU_DOWN_FAILED:
4621 mutex_lock(&wq_pool_mutex);
4623 for_each_pool(pool, pi) {
4624 mutex_lock(&pool->attach_mutex);
4626 if (pool->cpu == cpu) {
4627 rebind_workers(pool);
4628 } else if (pool->cpu < 0) {
4629 restore_unbound_workers_cpumask(pool, cpu);
4632 mutex_unlock(&pool->attach_mutex);
4635 /* update NUMA affinity of unbound workqueues */
4636 list_for_each_entry(wq, &workqueues, list)
4637 wq_update_unbound_numa(wq, cpu, true);
4639 mutex_unlock(&wq_pool_mutex);
4646 * Workqueues should be brought down after normal priority CPU notifiers.
4647 * This will be registered as low priority CPU notifier.
4649 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4650 unsigned long action,
4653 int cpu = (unsigned long)hcpu;
4654 struct work_struct unbind_work;
4655 struct workqueue_struct *wq;
4657 switch (action & ~CPU_TASKS_FROZEN) {
4658 case CPU_DOWN_PREPARE:
4659 /* unbinding per-cpu workers should happen on the local CPU */
4660 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4661 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4663 /* update NUMA affinity of unbound workqueues */
4664 mutex_lock(&wq_pool_mutex);
4665 list_for_each_entry(wq, &workqueues, list)
4666 wq_update_unbound_numa(wq, cpu, false);
4667 mutex_unlock(&wq_pool_mutex);
4669 /* wait for per-cpu unbinding to finish */
4670 flush_work(&unbind_work);
4671 destroy_work_on_stack(&unbind_work);
4679 struct work_for_cpu {
4680 struct work_struct work;
4686 static void work_for_cpu_fn(struct work_struct *work)
4688 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4690 wfc->ret = wfc->fn(wfc->arg);
4694 * work_on_cpu - run a function in user context on a particular cpu
4695 * @cpu: the cpu to run on
4696 * @fn: the function to run
4697 * @arg: the function arg
4699 * It is up to the caller to ensure that the cpu doesn't go offline.
4700 * The caller must not hold any locks which would prevent @fn from completing.
4702 * Return: The value @fn returns.
4704 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4706 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4708 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4709 schedule_work_on(cpu, &wfc.work);
4710 flush_work(&wfc.work);
4711 destroy_work_on_stack(&wfc.work);
4714 EXPORT_SYMBOL_GPL(work_on_cpu);
4715 #endif /* CONFIG_SMP */
4717 #ifdef CONFIG_FREEZER
4720 * freeze_workqueues_begin - begin freezing workqueues
4722 * Start freezing workqueues. After this function returns, all freezable
4723 * workqueues will queue new works to their delayed_works list instead of
4727 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4729 void freeze_workqueues_begin(void)
4731 struct workqueue_struct *wq;
4732 struct pool_workqueue *pwq;
4734 mutex_lock(&wq_pool_mutex);
4736 WARN_ON_ONCE(workqueue_freezing);
4737 workqueue_freezing = true;
4739 list_for_each_entry(wq, &workqueues, list) {
4740 mutex_lock(&wq->mutex);
4741 for_each_pwq(pwq, wq)
4742 pwq_adjust_max_active(pwq);
4743 mutex_unlock(&wq->mutex);
4746 mutex_unlock(&wq_pool_mutex);
4750 * freeze_workqueues_busy - are freezable workqueues still busy?
4752 * Check whether freezing is complete. This function must be called
4753 * between freeze_workqueues_begin() and thaw_workqueues().
4756 * Grabs and releases wq_pool_mutex.
4759 * %true if some freezable workqueues are still busy. %false if freezing
4762 bool freeze_workqueues_busy(void)
4765 struct workqueue_struct *wq;
4766 struct pool_workqueue *pwq;
4768 mutex_lock(&wq_pool_mutex);
4770 WARN_ON_ONCE(!workqueue_freezing);
4772 list_for_each_entry(wq, &workqueues, list) {
4773 if (!(wq->flags & WQ_FREEZABLE))
4776 * nr_active is monotonically decreasing. It's safe
4777 * to peek without lock.
4779 rcu_read_lock_sched();
4780 for_each_pwq(pwq, wq) {
4781 WARN_ON_ONCE(pwq->nr_active < 0);
4782 if (pwq->nr_active) {
4784 rcu_read_unlock_sched();
4788 rcu_read_unlock_sched();
4791 mutex_unlock(&wq_pool_mutex);
4796 * thaw_workqueues - thaw workqueues
4798 * Thaw workqueues. Normal queueing is restored and all collected
4799 * frozen works are transferred to their respective pool worklists.
4802 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4804 void thaw_workqueues(void)
4806 struct workqueue_struct *wq;
4807 struct pool_workqueue *pwq;
4809 mutex_lock(&wq_pool_mutex);
4811 if (!workqueue_freezing)
4814 workqueue_freezing = false;
4816 /* restore max_active and repopulate worklist */
4817 list_for_each_entry(wq, &workqueues, list) {
4818 mutex_lock(&wq->mutex);
4819 for_each_pwq(pwq, wq)
4820 pwq_adjust_max_active(pwq);
4821 mutex_unlock(&wq->mutex);
4825 mutex_unlock(&wq_pool_mutex);
4827 #endif /* CONFIG_FREEZER */
4829 static void __init wq_numa_init(void)
4834 /* determine NUMA pwq table len - highest node id + 1 */
4836 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4838 if (num_possible_nodes() <= 1)
4841 if (wq_disable_numa) {
4842 pr_info("workqueue: NUMA affinity support disabled\n");
4846 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4847 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4850 * We want masks of possible CPUs of each node which isn't readily
4851 * available. Build one from cpu_to_node() which should have been
4852 * fully initialized by now.
4854 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4858 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4859 node_online(node) ? node : NUMA_NO_NODE));
4861 for_each_possible_cpu(cpu) {
4862 node = cpu_to_node(cpu);
4863 if (WARN_ON(node == NUMA_NO_NODE)) {
4864 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4865 /* happens iff arch is bonkers, let's just proceed */
4868 cpumask_set_cpu(cpu, tbl[node]);
4871 wq_numa_possible_cpumask = tbl;
4872 wq_numa_enabled = true;
4875 static int __init init_workqueues(void)
4877 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4880 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4882 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4884 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4885 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4889 /* initialize CPU pools */
4890 for_each_possible_cpu(cpu) {
4891 struct worker_pool *pool;
4894 for_each_cpu_worker_pool(pool, cpu) {
4895 BUG_ON(init_worker_pool(pool));
4897 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4898 pool->attrs->nice = std_nice[i++];
4899 pool->node = cpu_to_node(cpu);
4902 mutex_lock(&wq_pool_mutex);
4903 BUG_ON(worker_pool_assign_id(pool));
4904 mutex_unlock(&wq_pool_mutex);
4908 /* create the initial worker */
4909 for_each_online_cpu(cpu) {
4910 struct worker_pool *pool;
4912 for_each_cpu_worker_pool(pool, cpu) {
4913 pool->flags &= ~POOL_DISASSOCIATED;
4914 BUG_ON(create_and_start_worker(pool) < 0);
4918 /* create default unbound and ordered wq attrs */
4919 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4920 struct workqueue_attrs *attrs;
4922 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4923 attrs->nice = std_nice[i];
4924 unbound_std_wq_attrs[i] = attrs;
4927 * An ordered wq should have only one pwq as ordering is
4928 * guaranteed by max_active which is enforced by pwqs.
4929 * Turn off NUMA so that dfl_pwq is used for all nodes.
4931 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4932 attrs->nice = std_nice[i];
4933 attrs->no_numa = true;
4934 ordered_wq_attrs[i] = attrs;
4937 system_wq = alloc_workqueue("events", 0, 0);
4938 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4939 system_long_wq = alloc_workqueue("events_long", 0, 0);
4940 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4941 WQ_UNBOUND_MAX_ACTIVE);
4942 system_freezable_wq = alloc_workqueue("events_freezable",
4944 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
4945 WQ_POWER_EFFICIENT, 0);
4946 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
4947 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
4949 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4950 !system_unbound_wq || !system_freezable_wq ||
4951 !system_power_efficient_wq ||
4952 !system_freezable_power_efficient_wq);
4955 early_initcall(init_workqueues);