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;
762 * nr_idle and idle_list may disagree if idle rebinding is in
763 * progress. Never return %true if idle_list is empty.
765 if (list_empty(&pool->idle_list))
768 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
775 /* Return the first idle worker. Safe with preemption disabled */
776 static struct worker *first_idle_worker(struct worker_pool *pool)
778 if (unlikely(list_empty(&pool->idle_list)))
781 return list_first_entry(&pool->idle_list, struct worker, entry);
785 * wake_up_worker - wake up an idle worker
786 * @pool: worker pool to wake worker from
788 * Wake up the first idle worker of @pool.
791 * spin_lock_irq(pool->lock).
793 static void wake_up_worker(struct worker_pool *pool)
795 struct worker *worker = first_idle_worker(pool);
798 wake_up_process(worker->task);
802 * wq_worker_waking_up - a worker is waking up
803 * @task: task waking up
804 * @cpu: CPU @task is waking up to
806 * This function is called during try_to_wake_up() when a worker is
810 * spin_lock_irq(rq->lock)
812 void wq_worker_waking_up(struct task_struct *task, int cpu)
814 struct worker *worker = kthread_data(task);
816 if (!(worker->flags & WORKER_NOT_RUNNING)) {
817 WARN_ON_ONCE(worker->pool->cpu != cpu);
818 atomic_inc(&worker->pool->nr_running);
823 * wq_worker_sleeping - a worker is going to sleep
824 * @task: task going to sleep
825 * @cpu: CPU in question, must be the current CPU number
827 * This function is called during schedule() when a busy worker is
828 * going to sleep. Worker on the same cpu can be woken up by
829 * returning pointer to its task.
832 * spin_lock_irq(rq->lock)
835 * Worker task on @cpu to wake up, %NULL if none.
837 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
839 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
840 struct worker_pool *pool;
843 * Rescuers, which may not have all the fields set up like normal
844 * workers, also reach here, let's not access anything before
845 * checking NOT_RUNNING.
847 if (worker->flags & WORKER_NOT_RUNNING)
852 /* this can only happen on the local cpu */
853 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
857 * The counterpart of the following dec_and_test, implied mb,
858 * worklist not empty test sequence is in insert_work().
859 * Please read comment there.
861 * NOT_RUNNING is clear. This means that we're bound to and
862 * running on the local cpu w/ rq lock held and preemption
863 * disabled, which in turn means that none else could be
864 * manipulating idle_list, so dereferencing idle_list without pool
867 if (atomic_dec_and_test(&pool->nr_running) &&
868 !list_empty(&pool->worklist))
869 to_wakeup = first_idle_worker(pool);
870 return to_wakeup ? to_wakeup->task : NULL;
874 * worker_set_flags - set worker flags and adjust nr_running accordingly
876 * @flags: flags to set
877 * @wakeup: wakeup an idle worker if necessary
879 * Set @flags in @worker->flags and adjust nr_running accordingly. If
880 * nr_running becomes zero and @wakeup is %true, an idle worker is
884 * spin_lock_irq(pool->lock)
886 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
889 struct worker_pool *pool = worker->pool;
891 WARN_ON_ONCE(worker->task != current);
894 * If transitioning into NOT_RUNNING, adjust nr_running and
895 * wake up an idle worker as necessary if requested by
898 if ((flags & WORKER_NOT_RUNNING) &&
899 !(worker->flags & WORKER_NOT_RUNNING)) {
901 if (atomic_dec_and_test(&pool->nr_running) &&
902 !list_empty(&pool->worklist))
903 wake_up_worker(pool);
905 atomic_dec(&pool->nr_running);
908 worker->flags |= flags;
912 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
914 * @flags: flags to clear
916 * Clear @flags in @worker->flags and adjust nr_running accordingly.
919 * spin_lock_irq(pool->lock)
921 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
923 struct worker_pool *pool = worker->pool;
924 unsigned int oflags = worker->flags;
926 WARN_ON_ONCE(worker->task != current);
928 worker->flags &= ~flags;
931 * If transitioning out of NOT_RUNNING, increment nr_running. Note
932 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
933 * of multiple flags, not a single flag.
935 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
936 if (!(worker->flags & WORKER_NOT_RUNNING))
937 atomic_inc(&pool->nr_running);
941 * find_worker_executing_work - find worker which is executing a work
942 * @pool: pool of interest
943 * @work: work to find worker for
945 * Find a worker which is executing @work on @pool by searching
946 * @pool->busy_hash which is keyed by the address of @work. For a worker
947 * to match, its current execution should match the address of @work and
948 * its work function. This is to avoid unwanted dependency between
949 * unrelated work executions through a work item being recycled while still
952 * This is a bit tricky. A work item may be freed once its execution
953 * starts and nothing prevents the freed area from being recycled for
954 * another work item. If the same work item address ends up being reused
955 * before the original execution finishes, workqueue will identify the
956 * recycled work item as currently executing and make it wait until the
957 * current execution finishes, introducing an unwanted dependency.
959 * This function checks the work item address and work function to avoid
960 * false positives. Note that this isn't complete as one may construct a
961 * work function which can introduce dependency onto itself through a
962 * recycled work item. Well, if somebody wants to shoot oneself in the
963 * foot that badly, there's only so much we can do, and if such deadlock
964 * actually occurs, it should be easy to locate the culprit work function.
967 * spin_lock_irq(pool->lock).
970 * Pointer to worker which is executing @work if found, %NULL
973 static struct worker *find_worker_executing_work(struct worker_pool *pool,
974 struct work_struct *work)
976 struct worker *worker;
978 hash_for_each_possible(pool->busy_hash, worker, hentry,
980 if (worker->current_work == work &&
981 worker->current_func == work->func)
988 * move_linked_works - move linked works to a list
989 * @work: start of series of works to be scheduled
990 * @head: target list to append @work to
991 * @nextp: out paramter for nested worklist walking
993 * Schedule linked works starting from @work to @head. Work series to
994 * be scheduled starts at @work and includes any consecutive work with
995 * WORK_STRUCT_LINKED set in its predecessor.
997 * If @nextp is not NULL, it's updated to point to the next work of
998 * the last scheduled work. This allows move_linked_works() to be
999 * nested inside outer list_for_each_entry_safe().
1002 * spin_lock_irq(pool->lock).
1004 static void move_linked_works(struct work_struct *work, struct list_head *head,
1005 struct work_struct **nextp)
1007 struct work_struct *n;
1010 * Linked worklist will always end before the end of the list,
1011 * use NULL for list head.
1013 list_for_each_entry_safe_from(work, n, NULL, entry) {
1014 list_move_tail(&work->entry, head);
1015 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1020 * If we're already inside safe list traversal and have moved
1021 * multiple works to the scheduled queue, the next position
1022 * needs to be updated.
1029 * get_pwq - get an extra reference on the specified pool_workqueue
1030 * @pwq: pool_workqueue to get
1032 * Obtain an extra reference on @pwq. The caller should guarantee that
1033 * @pwq has positive refcnt and be holding the matching pool->lock.
1035 static void get_pwq(struct pool_workqueue *pwq)
1037 lockdep_assert_held(&pwq->pool->lock);
1038 WARN_ON_ONCE(pwq->refcnt <= 0);
1043 * put_pwq - put a pool_workqueue reference
1044 * @pwq: pool_workqueue to put
1046 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1047 * destruction. The caller should be holding the matching pool->lock.
1049 static void put_pwq(struct pool_workqueue *pwq)
1051 lockdep_assert_held(&pwq->pool->lock);
1052 if (likely(--pwq->refcnt))
1054 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1057 * @pwq can't be released under pool->lock, bounce to
1058 * pwq_unbound_release_workfn(). This never recurses on the same
1059 * pool->lock as this path is taken only for unbound workqueues and
1060 * the release work item is scheduled on a per-cpu workqueue. To
1061 * avoid lockdep warning, unbound pool->locks are given lockdep
1062 * subclass of 1 in get_unbound_pool().
1064 schedule_work(&pwq->unbound_release_work);
1068 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1069 * @pwq: pool_workqueue to put (can be %NULL)
1071 * put_pwq() with locking. This function also allows %NULL @pwq.
1073 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1077 * As both pwqs and pools are sched-RCU protected, the
1078 * following lock operations are safe.
1080 spin_lock_irq(&pwq->pool->lock);
1082 spin_unlock_irq(&pwq->pool->lock);
1086 static void pwq_activate_delayed_work(struct work_struct *work)
1088 struct pool_workqueue *pwq = get_work_pwq(work);
1090 trace_workqueue_activate_work(work);
1091 move_linked_works(work, &pwq->pool->worklist, NULL);
1092 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1096 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1098 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1099 struct work_struct, entry);
1101 pwq_activate_delayed_work(work);
1105 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1106 * @pwq: pwq of interest
1107 * @color: color of work which left the queue
1109 * A work either has completed or is removed from pending queue,
1110 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1113 * spin_lock_irq(pool->lock).
1115 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1117 /* uncolored work items don't participate in flushing or nr_active */
1118 if (color == WORK_NO_COLOR)
1121 pwq->nr_in_flight[color]--;
1124 if (!list_empty(&pwq->delayed_works)) {
1125 /* one down, submit a delayed one */
1126 if (pwq->nr_active < pwq->max_active)
1127 pwq_activate_first_delayed(pwq);
1130 /* is flush in progress and are we at the flushing tip? */
1131 if (likely(pwq->flush_color != color))
1134 /* are there still in-flight works? */
1135 if (pwq->nr_in_flight[color])
1138 /* this pwq is done, clear flush_color */
1139 pwq->flush_color = -1;
1142 * If this was the last pwq, wake up the first flusher. It
1143 * will handle the rest.
1145 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1146 complete(&pwq->wq->first_flusher->done);
1152 * try_to_grab_pending - steal work item from worklist and disable irq
1153 * @work: work item to steal
1154 * @is_dwork: @work is a delayed_work
1155 * @flags: place to store irq state
1157 * Try to grab PENDING bit of @work. This function can handle @work in any
1158 * stable state - idle, on timer or on worklist.
1161 * 1 if @work was pending and we successfully stole PENDING
1162 * 0 if @work was idle and we claimed PENDING
1163 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1164 * -ENOENT if someone else is canceling @work, this state may persist
1165 * for arbitrarily long
1168 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1169 * interrupted while holding PENDING and @work off queue, irq must be
1170 * disabled on entry. This, combined with delayed_work->timer being
1171 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1173 * On successful return, >= 0, irq is disabled and the caller is
1174 * responsible for releasing it using local_irq_restore(*@flags).
1176 * This function is safe to call from any context including IRQ handler.
1178 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1179 unsigned long *flags)
1181 struct worker_pool *pool;
1182 struct pool_workqueue *pwq;
1184 local_irq_save(*flags);
1186 /* try to steal the timer if it exists */
1188 struct delayed_work *dwork = to_delayed_work(work);
1191 * dwork->timer is irqsafe. If del_timer() fails, it's
1192 * guaranteed that the timer is not queued anywhere and not
1193 * running on the local CPU.
1195 if (likely(del_timer(&dwork->timer)))
1199 /* try to claim PENDING the normal way */
1200 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1204 * The queueing is in progress, or it is already queued. Try to
1205 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1207 pool = get_work_pool(work);
1211 spin_lock(&pool->lock);
1213 * work->data is guaranteed to point to pwq only while the work
1214 * item is queued on pwq->wq, and both updating work->data to point
1215 * to pwq on queueing and to pool on dequeueing are done under
1216 * pwq->pool->lock. This in turn guarantees that, if work->data
1217 * points to pwq which is associated with a locked pool, the work
1218 * item is currently queued on that pool.
1220 pwq = get_work_pwq(work);
1221 if (pwq && pwq->pool == pool) {
1222 debug_work_deactivate(work);
1225 * A delayed work item cannot be grabbed directly because
1226 * it might have linked NO_COLOR work items which, if left
1227 * on the delayed_list, will confuse pwq->nr_active
1228 * management later on and cause stall. Make sure the work
1229 * item is activated before grabbing.
1231 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1232 pwq_activate_delayed_work(work);
1234 list_del_init(&work->entry);
1235 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1237 /* work->data points to pwq iff queued, point to pool */
1238 set_work_pool_and_keep_pending(work, pool->id);
1240 spin_unlock(&pool->lock);
1243 spin_unlock(&pool->lock);
1245 local_irq_restore(*flags);
1246 if (work_is_canceling(work))
1253 * insert_work - insert a work into a pool
1254 * @pwq: pwq @work belongs to
1255 * @work: work to insert
1256 * @head: insertion point
1257 * @extra_flags: extra WORK_STRUCT_* flags to set
1259 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1260 * work_struct flags.
1263 * spin_lock_irq(pool->lock).
1265 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1266 struct list_head *head, unsigned int extra_flags)
1268 struct worker_pool *pool = pwq->pool;
1270 /* we own @work, set data and link */
1271 set_work_pwq(work, pwq, extra_flags);
1272 list_add_tail(&work->entry, head);
1276 * Ensure either wq_worker_sleeping() sees the above
1277 * list_add_tail() or we see zero nr_running to avoid workers lying
1278 * around lazily while there are works to be processed.
1282 if (__need_more_worker(pool))
1283 wake_up_worker(pool);
1287 * Test whether @work is being queued from another work executing on the
1290 static bool is_chained_work(struct workqueue_struct *wq)
1292 struct worker *worker;
1294 worker = current_wq_worker();
1296 * Return %true iff I'm a worker execuing a work item on @wq. If
1297 * I'm @worker, it's safe to dereference it without locking.
1299 return worker && worker->current_pwq->wq == wq;
1302 static void __queue_work(int cpu, struct workqueue_struct *wq,
1303 struct work_struct *work)
1305 struct pool_workqueue *pwq;
1306 struct worker_pool *last_pool;
1307 struct list_head *worklist;
1308 unsigned int work_flags;
1309 unsigned int req_cpu = cpu;
1312 * While a work item is PENDING && off queue, a task trying to
1313 * steal the PENDING will busy-loop waiting for it to either get
1314 * queued or lose PENDING. Grabbing PENDING and queueing should
1315 * happen with IRQ disabled.
1317 WARN_ON_ONCE(!irqs_disabled());
1319 debug_work_activate(work);
1321 /* if draining, only works from the same workqueue are allowed */
1322 if (unlikely(wq->flags & __WQ_DRAINING) &&
1323 WARN_ON_ONCE(!is_chained_work(wq)))
1326 if (req_cpu == WORK_CPU_UNBOUND)
1327 cpu = raw_smp_processor_id();
1329 /* pwq which will be used unless @work is executing elsewhere */
1330 if (!(wq->flags & WQ_UNBOUND))
1331 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1333 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1336 * If @work was previously on a different pool, it might still be
1337 * running there, in which case the work needs to be queued on that
1338 * pool to guarantee non-reentrancy.
1340 last_pool = get_work_pool(work);
1341 if (last_pool && last_pool != pwq->pool) {
1342 struct worker *worker;
1344 spin_lock(&last_pool->lock);
1346 worker = find_worker_executing_work(last_pool, work);
1348 if (worker && worker->current_pwq->wq == wq) {
1349 pwq = worker->current_pwq;
1351 /* meh... not running there, queue here */
1352 spin_unlock(&last_pool->lock);
1353 spin_lock(&pwq->pool->lock);
1356 spin_lock(&pwq->pool->lock);
1360 * pwq is determined and locked. For unbound pools, we could have
1361 * raced with pwq release and it could already be dead. If its
1362 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1363 * without another pwq replacing it in the numa_pwq_tbl or while
1364 * work items are executing on it, so the retrying is guaranteed to
1365 * make forward-progress.
1367 if (unlikely(!pwq->refcnt)) {
1368 if (wq->flags & WQ_UNBOUND) {
1369 spin_unlock(&pwq->pool->lock);
1374 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1378 /* pwq determined, queue */
1379 trace_workqueue_queue_work(req_cpu, pwq, work);
1381 if (WARN_ON(!list_empty(&work->entry))) {
1382 spin_unlock(&pwq->pool->lock);
1386 pwq->nr_in_flight[pwq->work_color]++;
1387 work_flags = work_color_to_flags(pwq->work_color);
1389 if (likely(pwq->nr_active < pwq->max_active)) {
1390 trace_workqueue_activate_work(work);
1392 worklist = &pwq->pool->worklist;
1394 work_flags |= WORK_STRUCT_DELAYED;
1395 worklist = &pwq->delayed_works;
1398 insert_work(pwq, work, worklist, work_flags);
1400 spin_unlock(&pwq->pool->lock);
1404 * queue_work_on - queue work on specific cpu
1405 * @cpu: CPU number to execute work on
1406 * @wq: workqueue to use
1407 * @work: work to queue
1409 * We queue the work to a specific CPU, the caller must ensure it
1412 * Return: %false if @work was already on a queue, %true otherwise.
1414 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1415 struct work_struct *work)
1418 unsigned long flags;
1420 local_irq_save(flags);
1422 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1423 __queue_work(cpu, wq, work);
1427 local_irq_restore(flags);
1430 EXPORT_SYMBOL(queue_work_on);
1432 void delayed_work_timer_fn(unsigned long __data)
1434 struct delayed_work *dwork = (struct delayed_work *)__data;
1436 /* should have been called from irqsafe timer with irq already off */
1437 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1439 EXPORT_SYMBOL(delayed_work_timer_fn);
1441 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1442 struct delayed_work *dwork, unsigned long delay)
1444 struct timer_list *timer = &dwork->timer;
1445 struct work_struct *work = &dwork->work;
1447 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1448 timer->data != (unsigned long)dwork);
1449 WARN_ON_ONCE(timer_pending(timer));
1450 WARN_ON_ONCE(!list_empty(&work->entry));
1453 * If @delay is 0, queue @dwork->work immediately. This is for
1454 * both optimization and correctness. The earliest @timer can
1455 * expire is on the closest next tick and delayed_work users depend
1456 * on that there's no such delay when @delay is 0.
1459 __queue_work(cpu, wq, &dwork->work);
1463 timer_stats_timer_set_start_info(&dwork->timer);
1467 timer->expires = jiffies + delay;
1469 if (unlikely(cpu != WORK_CPU_UNBOUND))
1470 add_timer_on(timer, cpu);
1476 * queue_delayed_work_on - queue work on specific CPU after delay
1477 * @cpu: CPU number to execute work on
1478 * @wq: workqueue to use
1479 * @dwork: work to queue
1480 * @delay: number of jiffies to wait before queueing
1482 * Return: %false if @work was already on a queue, %true otherwise. If
1483 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1486 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1487 struct delayed_work *dwork, unsigned long delay)
1489 struct work_struct *work = &dwork->work;
1491 unsigned long flags;
1493 /* read the comment in __queue_work() */
1494 local_irq_save(flags);
1496 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1497 __queue_delayed_work(cpu, wq, dwork, delay);
1501 local_irq_restore(flags);
1504 EXPORT_SYMBOL(queue_delayed_work_on);
1507 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1508 * @cpu: CPU number to execute work on
1509 * @wq: workqueue to use
1510 * @dwork: work to queue
1511 * @delay: number of jiffies to wait before queueing
1513 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1514 * modify @dwork's timer so that it expires after @delay. If @delay is
1515 * zero, @work is guaranteed to be scheduled immediately regardless of its
1518 * Return: %false if @dwork was idle and queued, %true if @dwork was
1519 * pending and its timer was modified.
1521 * This function is safe to call from any context including IRQ handler.
1522 * See try_to_grab_pending() for details.
1524 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1525 struct delayed_work *dwork, unsigned long delay)
1527 unsigned long flags;
1531 ret = try_to_grab_pending(&dwork->work, true, &flags);
1532 } while (unlikely(ret == -EAGAIN));
1534 if (likely(ret >= 0)) {
1535 __queue_delayed_work(cpu, wq, dwork, delay);
1536 local_irq_restore(flags);
1539 /* -ENOENT from try_to_grab_pending() becomes %true */
1542 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1545 * worker_enter_idle - enter idle state
1546 * @worker: worker which is entering idle state
1548 * @worker is entering idle state. Update stats and idle timer if
1552 * spin_lock_irq(pool->lock).
1554 static void worker_enter_idle(struct worker *worker)
1556 struct worker_pool *pool = worker->pool;
1558 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1559 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1560 (worker->hentry.next || worker->hentry.pprev)))
1563 /* can't use worker_set_flags(), also called from start_worker() */
1564 worker->flags |= WORKER_IDLE;
1566 worker->last_active = jiffies;
1568 /* idle_list is LIFO */
1569 list_add(&worker->entry, &pool->idle_list);
1571 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1572 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1575 * Sanity check nr_running. Because wq_unbind_fn() releases
1576 * pool->lock between setting %WORKER_UNBOUND and zapping
1577 * nr_running, the warning may trigger spuriously. Check iff
1578 * unbind is not in progress.
1580 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1581 pool->nr_workers == pool->nr_idle &&
1582 atomic_read(&pool->nr_running));
1586 * worker_leave_idle - leave idle state
1587 * @worker: worker which is leaving idle state
1589 * @worker is leaving idle state. Update stats.
1592 * spin_lock_irq(pool->lock).
1594 static void worker_leave_idle(struct worker *worker)
1596 struct worker_pool *pool = worker->pool;
1598 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1600 worker_clr_flags(worker, WORKER_IDLE);
1602 list_del_init(&worker->entry);
1605 static struct worker *alloc_worker(void)
1607 struct worker *worker;
1609 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1611 INIT_LIST_HEAD(&worker->entry);
1612 INIT_LIST_HEAD(&worker->scheduled);
1613 INIT_LIST_HEAD(&worker->node);
1614 /* on creation a worker is in !idle && prep state */
1615 worker->flags = WORKER_PREP;
1621 * worker_attach_to_pool() - attach a worker to a pool
1622 * @worker: worker to be attached
1623 * @pool: the target pool
1625 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1626 * cpu-binding of @worker are kept coordinated with the pool across
1629 static void worker_attach_to_pool(struct worker *worker,
1630 struct worker_pool *pool)
1632 mutex_lock(&pool->attach_mutex);
1635 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1636 * online CPUs. It'll be re-applied when any of the CPUs come up.
1638 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1641 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1642 * stable across this function. See the comments above the
1643 * flag definition for details.
1645 if (pool->flags & POOL_DISASSOCIATED)
1646 worker->flags |= WORKER_UNBOUND;
1648 list_add_tail(&worker->node, &pool->workers);
1650 mutex_unlock(&pool->attach_mutex);
1654 * worker_detach_from_pool() - detach a worker from its pool
1655 * @worker: worker which is attached to its pool
1656 * @pool: the pool @worker is attached to
1658 * Undo the attaching which had been done in worker_attach_to_pool(). The
1659 * caller worker shouldn't access to the pool after detached except it has
1660 * other reference to the pool.
1662 static void worker_detach_from_pool(struct worker *worker,
1663 struct worker_pool *pool)
1665 struct completion *detach_completion = NULL;
1667 mutex_lock(&pool->attach_mutex);
1668 list_del(&worker->node);
1669 if (list_empty(&pool->workers))
1670 detach_completion = pool->detach_completion;
1671 mutex_unlock(&pool->attach_mutex);
1673 if (detach_completion)
1674 complete(detach_completion);
1678 * create_worker - create a new workqueue worker
1679 * @pool: pool the new worker will belong to
1681 * Create a new worker which is attached to @pool. The new worker must be
1682 * started by start_worker().
1685 * Might sleep. Does GFP_KERNEL allocations.
1688 * Pointer to the newly created worker.
1690 static struct worker *create_worker(struct worker_pool *pool)
1692 struct worker *worker = NULL;
1696 /* ID is needed to determine kthread name */
1697 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1701 worker = alloc_worker();
1705 worker->pool = pool;
1709 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1710 pool->attrs->nice < 0 ? "H" : "");
1712 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1714 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1715 "kworker/%s", id_buf);
1716 if (IS_ERR(worker->task))
1719 set_user_nice(worker->task, pool->attrs->nice);
1721 /* prevent userland from meddling with cpumask of workqueue workers */
1722 worker->task->flags |= PF_NO_SETAFFINITY;
1724 /* successful, attach the worker to the pool */
1725 worker_attach_to_pool(worker, pool);
1731 ida_simple_remove(&pool->worker_ida, id);
1737 * start_worker - start a newly created worker
1738 * @worker: worker to start
1740 * Make the pool aware of @worker and start it.
1743 * spin_lock_irq(pool->lock).
1745 static void start_worker(struct worker *worker)
1747 worker->pool->nr_workers++;
1748 worker_enter_idle(worker);
1749 wake_up_process(worker->task);
1753 * create_and_start_worker - create and start a worker for a pool
1754 * @pool: the target pool
1756 * Grab the managership of @pool and create and start a new worker for it.
1758 * Return: 0 on success. A negative error code otherwise.
1760 static int create_and_start_worker(struct worker_pool *pool)
1762 struct worker *worker;
1764 worker = create_worker(pool);
1766 spin_lock_irq(&pool->lock);
1767 start_worker(worker);
1768 spin_unlock_irq(&pool->lock);
1771 return worker ? 0 : -ENOMEM;
1775 * destroy_worker - destroy a workqueue worker
1776 * @worker: worker to be destroyed
1778 * Destroy @worker and adjust @pool stats accordingly. The worker should
1782 * spin_lock_irq(pool->lock).
1784 static void destroy_worker(struct worker *worker)
1786 struct worker_pool *pool = worker->pool;
1788 lockdep_assert_held(&pool->lock);
1790 /* sanity check frenzy */
1791 if (WARN_ON(worker->current_work) ||
1792 WARN_ON(!list_empty(&worker->scheduled)) ||
1793 WARN_ON(!(worker->flags & WORKER_IDLE)))
1799 list_del_init(&worker->entry);
1800 worker->flags |= WORKER_DIE;
1801 wake_up_process(worker->task);
1804 static void idle_worker_timeout(unsigned long __pool)
1806 struct worker_pool *pool = (void *)__pool;
1808 spin_lock_irq(&pool->lock);
1810 while (too_many_workers(pool)) {
1811 struct worker *worker;
1812 unsigned long expires;
1814 /* idle_list is kept in LIFO order, check the last one */
1815 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1816 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1818 if (time_before(jiffies, expires)) {
1819 mod_timer(&pool->idle_timer, expires);
1823 destroy_worker(worker);
1826 spin_unlock_irq(&pool->lock);
1829 static void send_mayday(struct work_struct *work)
1831 struct pool_workqueue *pwq = get_work_pwq(work);
1832 struct workqueue_struct *wq = pwq->wq;
1834 lockdep_assert_held(&wq_mayday_lock);
1839 /* mayday mayday mayday */
1840 if (list_empty(&pwq->mayday_node)) {
1842 * If @pwq is for an unbound wq, its base ref may be put at
1843 * any time due to an attribute change. Pin @pwq until the
1844 * rescuer is done with it.
1847 list_add_tail(&pwq->mayday_node, &wq->maydays);
1848 wake_up_process(wq->rescuer->task);
1852 static void pool_mayday_timeout(unsigned long __pool)
1854 struct worker_pool *pool = (void *)__pool;
1855 struct work_struct *work;
1857 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1858 spin_lock(&pool->lock);
1860 if (need_to_create_worker(pool)) {
1862 * We've been trying to create a new worker but
1863 * haven't been successful. We might be hitting an
1864 * allocation deadlock. Send distress signals to
1867 list_for_each_entry(work, &pool->worklist, entry)
1871 spin_unlock(&pool->lock);
1872 spin_unlock_irq(&wq_mayday_lock);
1874 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1878 * maybe_create_worker - create a new worker if necessary
1879 * @pool: pool to create a new worker for
1881 * Create a new worker for @pool if necessary. @pool is guaranteed to
1882 * have at least one idle worker on return from this function. If
1883 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1884 * sent to all rescuers with works scheduled on @pool to resolve
1885 * possible allocation deadlock.
1887 * On return, need_to_create_worker() is guaranteed to be %false and
1888 * may_start_working() %true.
1891 * spin_lock_irq(pool->lock) which may be released and regrabbed
1892 * multiple times. Does GFP_KERNEL allocations. Called only from
1896 * %false if no action was taken and pool->lock stayed locked, %true
1899 static bool maybe_create_worker(struct worker_pool *pool)
1900 __releases(&pool->lock)
1901 __acquires(&pool->lock)
1903 if (!need_to_create_worker(pool))
1906 spin_unlock_irq(&pool->lock);
1908 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1909 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1912 struct worker *worker;
1914 worker = create_worker(pool);
1916 del_timer_sync(&pool->mayday_timer);
1917 spin_lock_irq(&pool->lock);
1918 start_worker(worker);
1919 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1924 if (!need_to_create_worker(pool))
1927 __set_current_state(TASK_INTERRUPTIBLE);
1928 schedule_timeout(CREATE_COOLDOWN);
1930 if (!need_to_create_worker(pool))
1934 del_timer_sync(&pool->mayday_timer);
1935 spin_lock_irq(&pool->lock);
1936 if (need_to_create_worker(pool))
1942 * manage_workers - manage worker pool
1945 * Assume the manager role and manage the worker pool @worker belongs
1946 * to. At any given time, there can be only zero or one manager per
1947 * pool. The exclusion is handled automatically by this function.
1949 * The caller can safely start processing works on false return. On
1950 * true return, it's guaranteed that need_to_create_worker() is false
1951 * and may_start_working() is true.
1954 * spin_lock_irq(pool->lock) which may be released and regrabbed
1955 * multiple times. Does GFP_KERNEL allocations.
1958 * %false if the pool don't need management and the caller can safely start
1959 * processing works, %true indicates that the function released pool->lock
1960 * and reacquired it to perform some management function and that the
1961 * conditions that the caller verified while holding the lock before
1962 * calling the function might no longer be true.
1964 static bool manage_workers(struct worker *worker)
1966 struct worker_pool *pool = worker->pool;
1970 * Anyone who successfully grabs manager_arb wins the arbitration
1971 * and becomes the manager. mutex_trylock() on pool->manager_arb
1972 * failure while holding pool->lock reliably indicates that someone
1973 * else is managing the pool and the worker which failed trylock
1974 * can proceed to executing work items. This means that anyone
1975 * grabbing manager_arb is responsible for actually performing
1976 * manager duties. If manager_arb is grabbed and released without
1977 * actual management, the pool may stall indefinitely.
1979 if (!mutex_trylock(&pool->manager_arb))
1982 ret |= maybe_create_worker(pool);
1984 mutex_unlock(&pool->manager_arb);
1989 * process_one_work - process single work
1991 * @work: work to process
1993 * Process @work. This function contains all the logics necessary to
1994 * process a single work including synchronization against and
1995 * interaction with other workers on the same cpu, queueing and
1996 * flushing. As long as context requirement is met, any worker can
1997 * call this function to process a work.
2000 * spin_lock_irq(pool->lock) which is released and regrabbed.
2002 static void process_one_work(struct worker *worker, struct work_struct *work)
2003 __releases(&pool->lock)
2004 __acquires(&pool->lock)
2006 struct pool_workqueue *pwq = get_work_pwq(work);
2007 struct worker_pool *pool = worker->pool;
2008 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2010 struct worker *collision;
2011 #ifdef CONFIG_LOCKDEP
2013 * It is permissible to free the struct work_struct from
2014 * inside the function that is called from it, this we need to
2015 * take into account for lockdep too. To avoid bogus "held
2016 * lock freed" warnings as well as problems when looking into
2017 * work->lockdep_map, make a copy and use that here.
2019 struct lockdep_map lockdep_map;
2021 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2024 * Ensure we're on the correct CPU. DISASSOCIATED test is
2025 * necessary to avoid spurious warnings from rescuers servicing the
2026 * unbound or a disassociated pool.
2028 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2029 !(pool->flags & POOL_DISASSOCIATED) &&
2030 raw_smp_processor_id() != pool->cpu);
2033 * A single work shouldn't be executed concurrently by
2034 * multiple workers on a single cpu. Check whether anyone is
2035 * already processing the work. If so, defer the work to the
2036 * currently executing one.
2038 collision = find_worker_executing_work(pool, work);
2039 if (unlikely(collision)) {
2040 move_linked_works(work, &collision->scheduled, NULL);
2044 /* claim and dequeue */
2045 debug_work_deactivate(work);
2046 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2047 worker->current_work = work;
2048 worker->current_func = work->func;
2049 worker->current_pwq = pwq;
2050 work_color = get_work_color(work);
2052 list_del_init(&work->entry);
2055 * CPU intensive works don't participate in concurrency
2056 * management. They're the scheduler's responsibility.
2058 if (unlikely(cpu_intensive))
2059 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2062 * Unbound pool isn't concurrency managed and work items should be
2063 * executed ASAP. Wake up another worker if necessary.
2065 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2066 wake_up_worker(pool);
2069 * Record the last pool and clear PENDING which should be the last
2070 * update to @work. Also, do this inside @pool->lock so that
2071 * PENDING and queued state changes happen together while IRQ is
2074 set_work_pool_and_clear_pending(work, pool->id);
2076 spin_unlock_irq(&pool->lock);
2078 lock_map_acquire_read(&pwq->wq->lockdep_map);
2079 lock_map_acquire(&lockdep_map);
2080 trace_workqueue_execute_start(work);
2081 worker->current_func(work);
2083 * While we must be careful to not use "work" after this, the trace
2084 * point will only record its address.
2086 trace_workqueue_execute_end(work);
2087 lock_map_release(&lockdep_map);
2088 lock_map_release(&pwq->wq->lockdep_map);
2090 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2091 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2092 " last function: %pf\n",
2093 current->comm, preempt_count(), task_pid_nr(current),
2094 worker->current_func);
2095 debug_show_held_locks(current);
2100 * The following prevents a kworker from hogging CPU on !PREEMPT
2101 * kernels, where a requeueing work item waiting for something to
2102 * happen could deadlock with stop_machine as such work item could
2103 * indefinitely requeue itself while all other CPUs are trapped in
2108 spin_lock_irq(&pool->lock);
2110 /* clear cpu intensive status */
2111 if (unlikely(cpu_intensive))
2112 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2114 /* we're done with it, release */
2115 hash_del(&worker->hentry);
2116 worker->current_work = NULL;
2117 worker->current_func = NULL;
2118 worker->current_pwq = NULL;
2119 worker->desc_valid = false;
2120 pwq_dec_nr_in_flight(pwq, work_color);
2124 * process_scheduled_works - process scheduled works
2127 * Process all scheduled works. Please note that the scheduled list
2128 * may change while processing a work, so this function repeatedly
2129 * fetches a work from the top and executes it.
2132 * spin_lock_irq(pool->lock) which may be released and regrabbed
2135 static void process_scheduled_works(struct worker *worker)
2137 while (!list_empty(&worker->scheduled)) {
2138 struct work_struct *work = list_first_entry(&worker->scheduled,
2139 struct work_struct, entry);
2140 process_one_work(worker, work);
2145 * worker_thread - the worker thread function
2148 * The worker thread function. All workers belong to a worker_pool -
2149 * either a per-cpu one or dynamic unbound one. These workers process all
2150 * work items regardless of their specific target workqueue. The only
2151 * exception is work items which belong to workqueues with a rescuer which
2152 * will be explained in rescuer_thread().
2156 static int worker_thread(void *__worker)
2158 struct worker *worker = __worker;
2159 struct worker_pool *pool = worker->pool;
2161 /* tell the scheduler that this is a workqueue worker */
2162 worker->task->flags |= PF_WQ_WORKER;
2164 spin_lock_irq(&pool->lock);
2166 /* am I supposed to die? */
2167 if (unlikely(worker->flags & WORKER_DIE)) {
2168 spin_unlock_irq(&pool->lock);
2169 WARN_ON_ONCE(!list_empty(&worker->entry));
2170 worker->task->flags &= ~PF_WQ_WORKER;
2172 set_task_comm(worker->task, "kworker/dying");
2173 ida_simple_remove(&pool->worker_ida, worker->id);
2174 worker_detach_from_pool(worker, pool);
2179 worker_leave_idle(worker);
2181 /* no more worker necessary? */
2182 if (!need_more_worker(pool))
2185 /* do we need to manage? */
2186 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2190 * ->scheduled list can only be filled while a worker is
2191 * preparing to process a work or actually processing it.
2192 * Make sure nobody diddled with it while I was sleeping.
2194 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2197 * Finish PREP stage. We're guaranteed to have at least one idle
2198 * worker or that someone else has already assumed the manager
2199 * role. This is where @worker starts participating in concurrency
2200 * management if applicable and concurrency management is restored
2201 * after being rebound. See rebind_workers() for details.
2203 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2206 struct work_struct *work =
2207 list_first_entry(&pool->worklist,
2208 struct work_struct, entry);
2210 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2211 /* optimization path, not strictly necessary */
2212 process_one_work(worker, work);
2213 if (unlikely(!list_empty(&worker->scheduled)))
2214 process_scheduled_works(worker);
2216 move_linked_works(work, &worker->scheduled, NULL);
2217 process_scheduled_works(worker);
2219 } while (keep_working(pool));
2221 worker_set_flags(worker, WORKER_PREP, false);
2224 * pool->lock is held and there's no work to process and no need to
2225 * manage, sleep. Workers are woken up only while holding
2226 * pool->lock or from local cpu, so setting the current state
2227 * before releasing pool->lock is enough to prevent losing any
2230 worker_enter_idle(worker);
2231 __set_current_state(TASK_INTERRUPTIBLE);
2232 spin_unlock_irq(&pool->lock);
2238 * rescuer_thread - the rescuer thread function
2241 * Workqueue rescuer thread function. There's one rescuer for each
2242 * workqueue which has WQ_MEM_RECLAIM set.
2244 * Regular work processing on a pool may block trying to create a new
2245 * worker which uses GFP_KERNEL allocation which has slight chance of
2246 * developing into deadlock if some works currently on the same queue
2247 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2248 * the problem rescuer solves.
2250 * When such condition is possible, the pool summons rescuers of all
2251 * workqueues which have works queued on the pool and let them process
2252 * those works so that forward progress can be guaranteed.
2254 * This should happen rarely.
2258 static int rescuer_thread(void *__rescuer)
2260 struct worker *rescuer = __rescuer;
2261 struct workqueue_struct *wq = rescuer->rescue_wq;
2262 struct list_head *scheduled = &rescuer->scheduled;
2265 set_user_nice(current, RESCUER_NICE_LEVEL);
2268 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2269 * doesn't participate in concurrency management.
2271 rescuer->task->flags |= PF_WQ_WORKER;
2273 set_current_state(TASK_INTERRUPTIBLE);
2276 * By the time the rescuer is requested to stop, the workqueue
2277 * shouldn't have any work pending, but @wq->maydays may still have
2278 * pwq(s) queued. This can happen by non-rescuer workers consuming
2279 * all the work items before the rescuer got to them. Go through
2280 * @wq->maydays processing before acting on should_stop so that the
2281 * list is always empty on exit.
2283 should_stop = kthread_should_stop();
2285 /* see whether any pwq is asking for help */
2286 spin_lock_irq(&wq_mayday_lock);
2288 while (!list_empty(&wq->maydays)) {
2289 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2290 struct pool_workqueue, mayday_node);
2291 struct worker_pool *pool = pwq->pool;
2292 struct work_struct *work, *n;
2294 __set_current_state(TASK_RUNNING);
2295 list_del_init(&pwq->mayday_node);
2297 spin_unlock_irq(&wq_mayday_lock);
2299 worker_attach_to_pool(rescuer, pool);
2301 spin_lock_irq(&pool->lock);
2302 rescuer->pool = pool;
2305 * Slurp in all works issued via this workqueue and
2308 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2309 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2310 if (get_work_pwq(work) == pwq)
2311 move_linked_works(work, scheduled, &n);
2313 process_scheduled_works(rescuer);
2314 spin_unlock_irq(&pool->lock);
2316 worker_detach_from_pool(rescuer, pool);
2318 spin_lock_irq(&pool->lock);
2321 * Put the reference grabbed by send_mayday(). @pool won't
2322 * go away while we're holding its lock.
2327 * Leave this pool. If keep_working() is %true, notify a
2328 * regular worker; otherwise, we end up with 0 concurrency
2329 * and stalling the execution.
2331 if (keep_working(pool))
2332 wake_up_worker(pool);
2334 rescuer->pool = NULL;
2335 spin_unlock(&pool->lock);
2336 spin_lock(&wq_mayday_lock);
2339 spin_unlock_irq(&wq_mayday_lock);
2342 __set_current_state(TASK_RUNNING);
2343 rescuer->task->flags &= ~PF_WQ_WORKER;
2347 /* rescuers should never participate in concurrency management */
2348 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2354 struct work_struct work;
2355 struct completion done;
2358 static void wq_barrier_func(struct work_struct *work)
2360 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2361 complete(&barr->done);
2365 * insert_wq_barrier - insert a barrier work
2366 * @pwq: pwq to insert barrier into
2367 * @barr: wq_barrier to insert
2368 * @target: target work to attach @barr to
2369 * @worker: worker currently executing @target, NULL if @target is not executing
2371 * @barr is linked to @target such that @barr is completed only after
2372 * @target finishes execution. Please note that the ordering
2373 * guarantee is observed only with respect to @target and on the local
2376 * Currently, a queued barrier can't be canceled. This is because
2377 * try_to_grab_pending() can't determine whether the work to be
2378 * grabbed is at the head of the queue and thus can't clear LINKED
2379 * flag of the previous work while there must be a valid next work
2380 * after a work with LINKED flag set.
2382 * Note that when @worker is non-NULL, @target may be modified
2383 * underneath us, so we can't reliably determine pwq from @target.
2386 * spin_lock_irq(pool->lock).
2388 static void insert_wq_barrier(struct pool_workqueue *pwq,
2389 struct wq_barrier *barr,
2390 struct work_struct *target, struct worker *worker)
2392 struct list_head *head;
2393 unsigned int linked = 0;
2396 * debugobject calls are safe here even with pool->lock locked
2397 * as we know for sure that this will not trigger any of the
2398 * checks and call back into the fixup functions where we
2401 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2402 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2403 init_completion(&barr->done);
2406 * If @target is currently being executed, schedule the
2407 * barrier to the worker; otherwise, put it after @target.
2410 head = worker->scheduled.next;
2412 unsigned long *bits = work_data_bits(target);
2414 head = target->entry.next;
2415 /* there can already be other linked works, inherit and set */
2416 linked = *bits & WORK_STRUCT_LINKED;
2417 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2420 debug_work_activate(&barr->work);
2421 insert_work(pwq, &barr->work, head,
2422 work_color_to_flags(WORK_NO_COLOR) | linked);
2426 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2427 * @wq: workqueue being flushed
2428 * @flush_color: new flush color, < 0 for no-op
2429 * @work_color: new work color, < 0 for no-op
2431 * Prepare pwqs for workqueue flushing.
2433 * If @flush_color is non-negative, flush_color on all pwqs should be
2434 * -1. If no pwq has in-flight commands at the specified color, all
2435 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2436 * has in flight commands, its pwq->flush_color is set to
2437 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2438 * wakeup logic is armed and %true is returned.
2440 * The caller should have initialized @wq->first_flusher prior to
2441 * calling this function with non-negative @flush_color. If
2442 * @flush_color is negative, no flush color update is done and %false
2445 * If @work_color is non-negative, all pwqs should have the same
2446 * work_color which is previous to @work_color and all will be
2447 * advanced to @work_color.
2450 * mutex_lock(wq->mutex).
2453 * %true if @flush_color >= 0 and there's something to flush. %false
2456 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2457 int flush_color, int work_color)
2460 struct pool_workqueue *pwq;
2462 if (flush_color >= 0) {
2463 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2464 atomic_set(&wq->nr_pwqs_to_flush, 1);
2467 for_each_pwq(pwq, wq) {
2468 struct worker_pool *pool = pwq->pool;
2470 spin_lock_irq(&pool->lock);
2472 if (flush_color >= 0) {
2473 WARN_ON_ONCE(pwq->flush_color != -1);
2475 if (pwq->nr_in_flight[flush_color]) {
2476 pwq->flush_color = flush_color;
2477 atomic_inc(&wq->nr_pwqs_to_flush);
2482 if (work_color >= 0) {
2483 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2484 pwq->work_color = work_color;
2487 spin_unlock_irq(&pool->lock);
2490 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2491 complete(&wq->first_flusher->done);
2497 * flush_workqueue - ensure that any scheduled work has run to completion.
2498 * @wq: workqueue to flush
2500 * This function sleeps until all work items which were queued on entry
2501 * have finished execution, but it is not livelocked by new incoming ones.
2503 void flush_workqueue(struct workqueue_struct *wq)
2505 struct wq_flusher this_flusher = {
2506 .list = LIST_HEAD_INIT(this_flusher.list),
2508 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2512 lock_map_acquire(&wq->lockdep_map);
2513 lock_map_release(&wq->lockdep_map);
2515 mutex_lock(&wq->mutex);
2518 * Start-to-wait phase
2520 next_color = work_next_color(wq->work_color);
2522 if (next_color != wq->flush_color) {
2524 * Color space is not full. The current work_color
2525 * becomes our flush_color and work_color is advanced
2528 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2529 this_flusher.flush_color = wq->work_color;
2530 wq->work_color = next_color;
2532 if (!wq->first_flusher) {
2533 /* no flush in progress, become the first flusher */
2534 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2536 wq->first_flusher = &this_flusher;
2538 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2540 /* nothing to flush, done */
2541 wq->flush_color = next_color;
2542 wq->first_flusher = NULL;
2547 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2548 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2549 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2553 * Oops, color space is full, wait on overflow queue.
2554 * The next flush completion will assign us
2555 * flush_color and transfer to flusher_queue.
2557 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2560 mutex_unlock(&wq->mutex);
2562 wait_for_completion(&this_flusher.done);
2565 * Wake-up-and-cascade phase
2567 * First flushers are responsible for cascading flushes and
2568 * handling overflow. Non-first flushers can simply return.
2570 if (wq->first_flusher != &this_flusher)
2573 mutex_lock(&wq->mutex);
2575 /* we might have raced, check again with mutex held */
2576 if (wq->first_flusher != &this_flusher)
2579 wq->first_flusher = NULL;
2581 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2582 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2585 struct wq_flusher *next, *tmp;
2587 /* complete all the flushers sharing the current flush color */
2588 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2589 if (next->flush_color != wq->flush_color)
2591 list_del_init(&next->list);
2592 complete(&next->done);
2595 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2596 wq->flush_color != work_next_color(wq->work_color));
2598 /* this flush_color is finished, advance by one */
2599 wq->flush_color = work_next_color(wq->flush_color);
2601 /* one color has been freed, handle overflow queue */
2602 if (!list_empty(&wq->flusher_overflow)) {
2604 * Assign the same color to all overflowed
2605 * flushers, advance work_color and append to
2606 * flusher_queue. This is the start-to-wait
2607 * phase for these overflowed flushers.
2609 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2610 tmp->flush_color = wq->work_color;
2612 wq->work_color = work_next_color(wq->work_color);
2614 list_splice_tail_init(&wq->flusher_overflow,
2615 &wq->flusher_queue);
2616 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2619 if (list_empty(&wq->flusher_queue)) {
2620 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2625 * Need to flush more colors. Make the next flusher
2626 * the new first flusher and arm pwqs.
2628 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2629 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2631 list_del_init(&next->list);
2632 wq->first_flusher = next;
2634 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2638 * Meh... this color is already done, clear first
2639 * flusher and repeat cascading.
2641 wq->first_flusher = NULL;
2645 mutex_unlock(&wq->mutex);
2647 EXPORT_SYMBOL_GPL(flush_workqueue);
2650 * drain_workqueue - drain a workqueue
2651 * @wq: workqueue to drain
2653 * Wait until the workqueue becomes empty. While draining is in progress,
2654 * only chain queueing is allowed. IOW, only currently pending or running
2655 * work items on @wq can queue further work items on it. @wq is flushed
2656 * repeatedly until it becomes empty. The number of flushing is detemined
2657 * by the depth of chaining and should be relatively short. Whine if it
2660 void drain_workqueue(struct workqueue_struct *wq)
2662 unsigned int flush_cnt = 0;
2663 struct pool_workqueue *pwq;
2666 * __queue_work() needs to test whether there are drainers, is much
2667 * hotter than drain_workqueue() and already looks at @wq->flags.
2668 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2670 mutex_lock(&wq->mutex);
2671 if (!wq->nr_drainers++)
2672 wq->flags |= __WQ_DRAINING;
2673 mutex_unlock(&wq->mutex);
2675 flush_workqueue(wq);
2677 mutex_lock(&wq->mutex);
2679 for_each_pwq(pwq, wq) {
2682 spin_lock_irq(&pwq->pool->lock);
2683 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2684 spin_unlock_irq(&pwq->pool->lock);
2689 if (++flush_cnt == 10 ||
2690 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2691 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2692 wq->name, flush_cnt);
2694 mutex_unlock(&wq->mutex);
2698 if (!--wq->nr_drainers)
2699 wq->flags &= ~__WQ_DRAINING;
2700 mutex_unlock(&wq->mutex);
2702 EXPORT_SYMBOL_GPL(drain_workqueue);
2704 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2706 struct worker *worker = NULL;
2707 struct worker_pool *pool;
2708 struct pool_workqueue *pwq;
2712 local_irq_disable();
2713 pool = get_work_pool(work);
2719 spin_lock(&pool->lock);
2720 /* see the comment in try_to_grab_pending() with the same code */
2721 pwq = get_work_pwq(work);
2723 if (unlikely(pwq->pool != pool))
2726 worker = find_worker_executing_work(pool, work);
2729 pwq = worker->current_pwq;
2732 insert_wq_barrier(pwq, barr, work, worker);
2733 spin_unlock_irq(&pool->lock);
2736 * If @max_active is 1 or rescuer is in use, flushing another work
2737 * item on the same workqueue may lead to deadlock. Make sure the
2738 * flusher is not running on the same workqueue by verifying write
2741 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2742 lock_map_acquire(&pwq->wq->lockdep_map);
2744 lock_map_acquire_read(&pwq->wq->lockdep_map);
2745 lock_map_release(&pwq->wq->lockdep_map);
2749 spin_unlock_irq(&pool->lock);
2754 * flush_work - wait for a work to finish executing the last queueing instance
2755 * @work: the work to flush
2757 * Wait until @work has finished execution. @work is guaranteed to be idle
2758 * on return if it hasn't been requeued since flush started.
2761 * %true if flush_work() waited for the work to finish execution,
2762 * %false if it was already idle.
2764 bool flush_work(struct work_struct *work)
2766 struct wq_barrier barr;
2768 lock_map_acquire(&work->lockdep_map);
2769 lock_map_release(&work->lockdep_map);
2771 if (start_flush_work(work, &barr)) {
2772 wait_for_completion(&barr.done);
2773 destroy_work_on_stack(&barr.work);
2779 EXPORT_SYMBOL_GPL(flush_work);
2781 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2783 unsigned long flags;
2787 ret = try_to_grab_pending(work, is_dwork, &flags);
2789 * If someone else is canceling, wait for the same event it
2790 * would be waiting for before retrying.
2792 if (unlikely(ret == -ENOENT))
2794 } while (unlikely(ret < 0));
2796 /* tell other tasks trying to grab @work to back off */
2797 mark_work_canceling(work);
2798 local_irq_restore(flags);
2801 clear_work_data(work);
2806 * cancel_work_sync - cancel a work and wait for it to finish
2807 * @work: the work to cancel
2809 * Cancel @work and wait for its execution to finish. This function
2810 * can be used even if the work re-queues itself or migrates to
2811 * another workqueue. On return from this function, @work is
2812 * guaranteed to be not pending or executing on any CPU.
2814 * cancel_work_sync(&delayed_work->work) must not be used for
2815 * delayed_work's. Use cancel_delayed_work_sync() instead.
2817 * The caller must ensure that the workqueue on which @work was last
2818 * queued can't be destroyed before this function returns.
2821 * %true if @work was pending, %false otherwise.
2823 bool cancel_work_sync(struct work_struct *work)
2825 return __cancel_work_timer(work, false);
2827 EXPORT_SYMBOL_GPL(cancel_work_sync);
2830 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2831 * @dwork: the delayed work to flush
2833 * Delayed timer is cancelled and the pending work is queued for
2834 * immediate execution. Like flush_work(), this function only
2835 * considers the last queueing instance of @dwork.
2838 * %true if flush_work() waited for the work to finish execution,
2839 * %false if it was already idle.
2841 bool flush_delayed_work(struct delayed_work *dwork)
2843 local_irq_disable();
2844 if (del_timer_sync(&dwork->timer))
2845 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2847 return flush_work(&dwork->work);
2849 EXPORT_SYMBOL(flush_delayed_work);
2852 * cancel_delayed_work - cancel a delayed work
2853 * @dwork: delayed_work to cancel
2855 * Kill off a pending delayed_work.
2857 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2861 * The work callback function may still be running on return, unless
2862 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2863 * use cancel_delayed_work_sync() to wait on it.
2865 * This function is safe to call from any context including IRQ handler.
2867 bool cancel_delayed_work(struct delayed_work *dwork)
2869 unsigned long flags;
2873 ret = try_to_grab_pending(&dwork->work, true, &flags);
2874 } while (unlikely(ret == -EAGAIN));
2876 if (unlikely(ret < 0))
2879 set_work_pool_and_clear_pending(&dwork->work,
2880 get_work_pool_id(&dwork->work));
2881 local_irq_restore(flags);
2884 EXPORT_SYMBOL(cancel_delayed_work);
2887 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2888 * @dwork: the delayed work cancel
2890 * This is cancel_work_sync() for delayed works.
2893 * %true if @dwork was pending, %false otherwise.
2895 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2897 return __cancel_work_timer(&dwork->work, true);
2899 EXPORT_SYMBOL(cancel_delayed_work_sync);
2902 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2903 * @func: the function to call
2905 * schedule_on_each_cpu() executes @func on each online CPU using the
2906 * system workqueue and blocks until all CPUs have completed.
2907 * schedule_on_each_cpu() is very slow.
2910 * 0 on success, -errno on failure.
2912 int schedule_on_each_cpu(work_func_t func)
2915 struct work_struct __percpu *works;
2917 works = alloc_percpu(struct work_struct);
2923 for_each_online_cpu(cpu) {
2924 struct work_struct *work = per_cpu_ptr(works, cpu);
2926 INIT_WORK(work, func);
2927 schedule_work_on(cpu, work);
2930 for_each_online_cpu(cpu)
2931 flush_work(per_cpu_ptr(works, cpu));
2939 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2941 * Forces execution of the kernel-global workqueue and blocks until its
2944 * Think twice before calling this function! It's very easy to get into
2945 * trouble if you don't take great care. Either of the following situations
2946 * will lead to deadlock:
2948 * One of the work items currently on the workqueue needs to acquire
2949 * a lock held by your code or its caller.
2951 * Your code is running in the context of a work routine.
2953 * They will be detected by lockdep when they occur, but the first might not
2954 * occur very often. It depends on what work items are on the workqueue and
2955 * what locks they need, which you have no control over.
2957 * In most situations flushing the entire workqueue is overkill; you merely
2958 * need to know that a particular work item isn't queued and isn't running.
2959 * In such cases you should use cancel_delayed_work_sync() or
2960 * cancel_work_sync() instead.
2962 void flush_scheduled_work(void)
2964 flush_workqueue(system_wq);
2966 EXPORT_SYMBOL(flush_scheduled_work);
2969 * execute_in_process_context - reliably execute the routine with user context
2970 * @fn: the function to execute
2971 * @ew: guaranteed storage for the execute work structure (must
2972 * be available when the work executes)
2974 * Executes the function immediately if process context is available,
2975 * otherwise schedules the function for delayed execution.
2977 * Return: 0 - function was executed
2978 * 1 - function was scheduled for execution
2980 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2982 if (!in_interrupt()) {
2987 INIT_WORK(&ew->work, fn);
2988 schedule_work(&ew->work);
2992 EXPORT_SYMBOL_GPL(execute_in_process_context);
2996 * Workqueues with WQ_SYSFS flag set is visible to userland via
2997 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
2998 * following attributes.
3000 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3001 * max_active RW int : maximum number of in-flight work items
3003 * Unbound workqueues have the following extra attributes.
3005 * id RO int : the associated pool ID
3006 * nice RW int : nice value of the workers
3007 * cpumask RW mask : bitmask of allowed CPUs for the workers
3010 struct workqueue_struct *wq;
3014 static struct workqueue_struct *dev_to_wq(struct device *dev)
3016 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3021 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3024 struct workqueue_struct *wq = dev_to_wq(dev);
3026 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3028 static DEVICE_ATTR_RO(per_cpu);
3030 static ssize_t max_active_show(struct device *dev,
3031 struct device_attribute *attr, char *buf)
3033 struct workqueue_struct *wq = dev_to_wq(dev);
3035 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3038 static ssize_t max_active_store(struct device *dev,
3039 struct device_attribute *attr, const char *buf,
3042 struct workqueue_struct *wq = dev_to_wq(dev);
3045 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3048 workqueue_set_max_active(wq, val);
3051 static DEVICE_ATTR_RW(max_active);
3053 static struct attribute *wq_sysfs_attrs[] = {
3054 &dev_attr_per_cpu.attr,
3055 &dev_attr_max_active.attr,
3058 ATTRIBUTE_GROUPS(wq_sysfs);
3060 static ssize_t wq_pool_ids_show(struct device *dev,
3061 struct device_attribute *attr, char *buf)
3063 struct workqueue_struct *wq = dev_to_wq(dev);
3064 const char *delim = "";
3065 int node, written = 0;
3067 rcu_read_lock_sched();
3068 for_each_node(node) {
3069 written += scnprintf(buf + written, PAGE_SIZE - written,
3070 "%s%d:%d", delim, node,
3071 unbound_pwq_by_node(wq, node)->pool->id);
3074 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3075 rcu_read_unlock_sched();
3080 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3083 struct workqueue_struct *wq = dev_to_wq(dev);
3086 mutex_lock(&wq->mutex);
3087 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3088 mutex_unlock(&wq->mutex);
3093 /* prepare workqueue_attrs for sysfs store operations */
3094 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3096 struct workqueue_attrs *attrs;
3098 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3102 mutex_lock(&wq->mutex);
3103 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3104 mutex_unlock(&wq->mutex);
3108 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3109 const char *buf, size_t count)
3111 struct workqueue_struct *wq = dev_to_wq(dev);
3112 struct workqueue_attrs *attrs;
3115 attrs = wq_sysfs_prep_attrs(wq);
3119 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3120 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
3121 ret = apply_workqueue_attrs(wq, attrs);
3125 free_workqueue_attrs(attrs);
3126 return ret ?: count;
3129 static ssize_t wq_cpumask_show(struct device *dev,
3130 struct device_attribute *attr, char *buf)
3132 struct workqueue_struct *wq = dev_to_wq(dev);
3135 mutex_lock(&wq->mutex);
3136 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3137 mutex_unlock(&wq->mutex);
3139 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3143 static ssize_t wq_cpumask_store(struct device *dev,
3144 struct device_attribute *attr,
3145 const char *buf, size_t count)
3147 struct workqueue_struct *wq = dev_to_wq(dev);
3148 struct workqueue_attrs *attrs;
3151 attrs = wq_sysfs_prep_attrs(wq);
3155 ret = cpumask_parse(buf, attrs->cpumask);
3157 ret = apply_workqueue_attrs(wq, attrs);
3159 free_workqueue_attrs(attrs);
3160 return ret ?: count;
3163 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3166 struct workqueue_struct *wq = dev_to_wq(dev);
3169 mutex_lock(&wq->mutex);
3170 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3171 !wq->unbound_attrs->no_numa);
3172 mutex_unlock(&wq->mutex);
3177 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3178 const char *buf, size_t count)
3180 struct workqueue_struct *wq = dev_to_wq(dev);
3181 struct workqueue_attrs *attrs;
3184 attrs = wq_sysfs_prep_attrs(wq);
3189 if (sscanf(buf, "%d", &v) == 1) {
3190 attrs->no_numa = !v;
3191 ret = apply_workqueue_attrs(wq, attrs);
3194 free_workqueue_attrs(attrs);
3195 return ret ?: count;
3198 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3199 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3200 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3201 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3202 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3206 static struct bus_type wq_subsys = {
3207 .name = "workqueue",
3208 .dev_groups = wq_sysfs_groups,
3211 static int __init wq_sysfs_init(void)
3213 return subsys_virtual_register(&wq_subsys, NULL);
3215 core_initcall(wq_sysfs_init);
3217 static void wq_device_release(struct device *dev)
3219 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3225 * workqueue_sysfs_register - make a workqueue visible in sysfs
3226 * @wq: the workqueue to register
3228 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3229 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3230 * which is the preferred method.
3232 * Workqueue user should use this function directly iff it wants to apply
3233 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3234 * apply_workqueue_attrs() may race against userland updating the
3237 * Return: 0 on success, -errno on failure.
3239 int workqueue_sysfs_register(struct workqueue_struct *wq)
3241 struct wq_device *wq_dev;
3245 * Adjusting max_active or creating new pwqs by applyting
3246 * attributes breaks ordering guarantee. Disallow exposing ordered
3249 if (WARN_ON(wq->flags & __WQ_ORDERED))
3252 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3257 wq_dev->dev.bus = &wq_subsys;
3258 wq_dev->dev.init_name = wq->name;
3259 wq_dev->dev.release = wq_device_release;
3262 * unbound_attrs are created separately. Suppress uevent until
3263 * everything is ready.
3265 dev_set_uevent_suppress(&wq_dev->dev, true);
3267 ret = device_register(&wq_dev->dev);
3274 if (wq->flags & WQ_UNBOUND) {
3275 struct device_attribute *attr;
3277 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3278 ret = device_create_file(&wq_dev->dev, attr);
3280 device_unregister(&wq_dev->dev);
3287 dev_set_uevent_suppress(&wq_dev->dev, false);
3288 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3293 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3294 * @wq: the workqueue to unregister
3296 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3298 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3300 struct wq_device *wq_dev = wq->wq_dev;
3306 device_unregister(&wq_dev->dev);
3308 #else /* CONFIG_SYSFS */
3309 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3310 #endif /* CONFIG_SYSFS */
3313 * free_workqueue_attrs - free a workqueue_attrs
3314 * @attrs: workqueue_attrs to free
3316 * Undo alloc_workqueue_attrs().
3318 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3321 free_cpumask_var(attrs->cpumask);
3327 * alloc_workqueue_attrs - allocate a workqueue_attrs
3328 * @gfp_mask: allocation mask to use
3330 * Allocate a new workqueue_attrs, initialize with default settings and
3333 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3335 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3337 struct workqueue_attrs *attrs;
3339 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3342 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3345 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3348 free_workqueue_attrs(attrs);
3352 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3353 const struct workqueue_attrs *from)
3355 to->nice = from->nice;
3356 cpumask_copy(to->cpumask, from->cpumask);
3358 * Unlike hash and equality test, this function doesn't ignore
3359 * ->no_numa as it is used for both pool and wq attrs. Instead,
3360 * get_unbound_pool() explicitly clears ->no_numa after copying.
3362 to->no_numa = from->no_numa;
3365 /* hash value of the content of @attr */
3366 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3370 hash = jhash_1word(attrs->nice, hash);
3371 hash = jhash(cpumask_bits(attrs->cpumask),
3372 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3376 /* content equality test */
3377 static bool wqattrs_equal(const struct workqueue_attrs *a,
3378 const struct workqueue_attrs *b)
3380 if (a->nice != b->nice)
3382 if (!cpumask_equal(a->cpumask, b->cpumask))
3388 * init_worker_pool - initialize a newly zalloc'd worker_pool
3389 * @pool: worker_pool to initialize
3391 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3393 * Return: 0 on success, -errno on failure. Even on failure, all fields
3394 * inside @pool proper are initialized and put_unbound_pool() can be called
3395 * on @pool safely to release it.
3397 static int init_worker_pool(struct worker_pool *pool)
3399 spin_lock_init(&pool->lock);
3402 pool->node = NUMA_NO_NODE;
3403 pool->flags |= POOL_DISASSOCIATED;
3404 INIT_LIST_HEAD(&pool->worklist);
3405 INIT_LIST_HEAD(&pool->idle_list);
3406 hash_init(pool->busy_hash);
3408 init_timer_deferrable(&pool->idle_timer);
3409 pool->idle_timer.function = idle_worker_timeout;
3410 pool->idle_timer.data = (unsigned long)pool;
3412 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3413 (unsigned long)pool);
3415 mutex_init(&pool->manager_arb);
3416 mutex_init(&pool->attach_mutex);
3417 INIT_LIST_HEAD(&pool->workers);
3419 ida_init(&pool->worker_ida);
3420 INIT_HLIST_NODE(&pool->hash_node);
3423 /* shouldn't fail above this point */
3424 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3430 static void rcu_free_pool(struct rcu_head *rcu)
3432 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3434 ida_destroy(&pool->worker_ida);
3435 free_workqueue_attrs(pool->attrs);
3440 * put_unbound_pool - put a worker_pool
3441 * @pool: worker_pool to put
3443 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3444 * safe manner. get_unbound_pool() calls this function on its failure path
3445 * and this function should be able to release pools which went through,
3446 * successfully or not, init_worker_pool().
3448 * Should be called with wq_pool_mutex held.
3450 static void put_unbound_pool(struct worker_pool *pool)
3452 DECLARE_COMPLETION_ONSTACK(detach_completion);
3453 struct worker *worker;
3455 lockdep_assert_held(&wq_pool_mutex);
3461 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3462 WARN_ON(!list_empty(&pool->worklist)))
3465 /* release id and unhash */
3467 idr_remove(&worker_pool_idr, pool->id);
3468 hash_del(&pool->hash_node);
3471 * Become the manager and destroy all workers. Grabbing
3472 * manager_arb prevents @pool's workers from blocking on
3475 mutex_lock(&pool->manager_arb);
3477 spin_lock_irq(&pool->lock);
3478 while ((worker = first_idle_worker(pool)))
3479 destroy_worker(worker);
3480 WARN_ON(pool->nr_workers || pool->nr_idle);
3481 spin_unlock_irq(&pool->lock);
3483 mutex_lock(&pool->attach_mutex);
3484 if (!list_empty(&pool->workers))
3485 pool->detach_completion = &detach_completion;
3486 mutex_unlock(&pool->attach_mutex);
3488 if (pool->detach_completion)
3489 wait_for_completion(pool->detach_completion);
3491 mutex_unlock(&pool->manager_arb);
3493 /* shut down the timers */
3494 del_timer_sync(&pool->idle_timer);
3495 del_timer_sync(&pool->mayday_timer);
3497 /* sched-RCU protected to allow dereferences from get_work_pool() */
3498 call_rcu_sched(&pool->rcu, rcu_free_pool);
3502 * get_unbound_pool - get a worker_pool with the specified attributes
3503 * @attrs: the attributes of the worker_pool to get
3505 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3506 * reference count and return it. If there already is a matching
3507 * worker_pool, it will be used; otherwise, this function attempts to
3510 * Should be called with wq_pool_mutex held.
3512 * Return: On success, a worker_pool with the same attributes as @attrs.
3513 * On failure, %NULL.
3515 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3517 u32 hash = wqattrs_hash(attrs);
3518 struct worker_pool *pool;
3521 lockdep_assert_held(&wq_pool_mutex);
3523 /* do we already have a matching pool? */
3524 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3525 if (wqattrs_equal(pool->attrs, attrs)) {
3531 /* nope, create a new one */
3532 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3533 if (!pool || init_worker_pool(pool) < 0)
3536 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3537 copy_workqueue_attrs(pool->attrs, attrs);
3540 * no_numa isn't a worker_pool attribute, always clear it. See
3541 * 'struct workqueue_attrs' comments for detail.
3543 pool->attrs->no_numa = false;
3545 /* if cpumask is contained inside a NUMA node, we belong to that node */
3546 if (wq_numa_enabled) {
3547 for_each_node(node) {
3548 if (cpumask_subset(pool->attrs->cpumask,
3549 wq_numa_possible_cpumask[node])) {
3556 if (worker_pool_assign_id(pool) < 0)
3559 /* create and start the initial worker */
3560 if (create_and_start_worker(pool) < 0)
3564 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3569 put_unbound_pool(pool);
3573 static void rcu_free_pwq(struct rcu_head *rcu)
3575 kmem_cache_free(pwq_cache,
3576 container_of(rcu, struct pool_workqueue, rcu));
3580 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3581 * and needs to be destroyed.
3583 static void pwq_unbound_release_workfn(struct work_struct *work)
3585 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3586 unbound_release_work);
3587 struct workqueue_struct *wq = pwq->wq;
3588 struct worker_pool *pool = pwq->pool;
3591 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3595 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3596 * necessary on release but do it anyway. It's easier to verify
3597 * and consistent with the linking path.
3599 mutex_lock(&wq->mutex);
3600 list_del_rcu(&pwq->pwqs_node);
3601 is_last = list_empty(&wq->pwqs);
3602 mutex_unlock(&wq->mutex);
3604 mutex_lock(&wq_pool_mutex);
3605 put_unbound_pool(pool);
3606 mutex_unlock(&wq_pool_mutex);
3608 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3611 * If we're the last pwq going away, @wq is already dead and no one
3612 * is gonna access it anymore. Free it.
3615 free_workqueue_attrs(wq->unbound_attrs);
3621 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3622 * @pwq: target pool_workqueue
3624 * If @pwq isn't freezing, set @pwq->max_active to the associated
3625 * workqueue's saved_max_active and activate delayed work items
3626 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3628 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3630 struct workqueue_struct *wq = pwq->wq;
3631 bool freezable = wq->flags & WQ_FREEZABLE;
3633 /* for @wq->saved_max_active */
3634 lockdep_assert_held(&wq->mutex);
3636 /* fast exit for non-freezable wqs */
3637 if (!freezable && pwq->max_active == wq->saved_max_active)
3640 spin_lock_irq(&pwq->pool->lock);
3643 * During [un]freezing, the caller is responsible for ensuring that
3644 * this function is called at least once after @workqueue_freezing
3645 * is updated and visible.
3647 if (!freezable || !workqueue_freezing) {
3648 pwq->max_active = wq->saved_max_active;
3650 while (!list_empty(&pwq->delayed_works) &&
3651 pwq->nr_active < pwq->max_active)
3652 pwq_activate_first_delayed(pwq);
3655 * Need to kick a worker after thawed or an unbound wq's
3656 * max_active is bumped. It's a slow path. Do it always.
3658 wake_up_worker(pwq->pool);
3660 pwq->max_active = 0;
3663 spin_unlock_irq(&pwq->pool->lock);
3666 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3667 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3668 struct worker_pool *pool)
3670 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3672 memset(pwq, 0, sizeof(*pwq));
3676 pwq->flush_color = -1;
3678 INIT_LIST_HEAD(&pwq->delayed_works);
3679 INIT_LIST_HEAD(&pwq->pwqs_node);
3680 INIT_LIST_HEAD(&pwq->mayday_node);
3681 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3684 /* sync @pwq with the current state of its associated wq and link it */
3685 static void link_pwq(struct pool_workqueue *pwq)
3687 struct workqueue_struct *wq = pwq->wq;
3689 lockdep_assert_held(&wq->mutex);
3691 /* may be called multiple times, ignore if already linked */
3692 if (!list_empty(&pwq->pwqs_node))
3696 * Set the matching work_color. This is synchronized with
3697 * wq->mutex to avoid confusing flush_workqueue().
3699 pwq->work_color = wq->work_color;
3701 /* sync max_active to the current setting */
3702 pwq_adjust_max_active(pwq);
3705 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3708 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3709 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3710 const struct workqueue_attrs *attrs)
3712 struct worker_pool *pool;
3713 struct pool_workqueue *pwq;
3715 lockdep_assert_held(&wq_pool_mutex);
3717 pool = get_unbound_pool(attrs);
3721 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3723 put_unbound_pool(pool);
3727 init_pwq(pwq, wq, pool);
3731 /* undo alloc_unbound_pwq(), used only in the error path */
3732 static void free_unbound_pwq(struct pool_workqueue *pwq)
3734 lockdep_assert_held(&wq_pool_mutex);
3737 put_unbound_pool(pwq->pool);
3738 kmem_cache_free(pwq_cache, pwq);
3743 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3744 * @attrs: the wq_attrs of interest
3745 * @node: the target NUMA node
3746 * @cpu_going_down: if >= 0, the CPU to consider as offline
3747 * @cpumask: outarg, the resulting cpumask
3749 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3750 * @cpu_going_down is >= 0, that cpu is considered offline during
3751 * calculation. The result is stored in @cpumask.
3753 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3754 * enabled and @node has online CPUs requested by @attrs, the returned
3755 * cpumask is the intersection of the possible CPUs of @node and
3758 * The caller is responsible for ensuring that the cpumask of @node stays
3761 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3764 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3765 int cpu_going_down, cpumask_t *cpumask)
3767 if (!wq_numa_enabled || attrs->no_numa)
3770 /* does @node have any online CPUs @attrs wants? */
3771 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3772 if (cpu_going_down >= 0)
3773 cpumask_clear_cpu(cpu_going_down, cpumask);
3775 if (cpumask_empty(cpumask))
3778 /* yeap, return possible CPUs in @node that @attrs wants */
3779 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3780 return !cpumask_equal(cpumask, attrs->cpumask);
3783 cpumask_copy(cpumask, attrs->cpumask);
3787 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3788 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3790 struct pool_workqueue *pwq)
3792 struct pool_workqueue *old_pwq;
3794 lockdep_assert_held(&wq->mutex);
3796 /* link_pwq() can handle duplicate calls */
3799 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3800 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3805 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3806 * @wq: the target workqueue
3807 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3809 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3810 * machines, this function maps a separate pwq to each NUMA node with
3811 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3812 * NUMA node it was issued on. Older pwqs are released as in-flight work
3813 * items finish. Note that a work item which repeatedly requeues itself
3814 * back-to-back will stay on its current pwq.
3816 * Performs GFP_KERNEL allocations.
3818 * Return: 0 on success and -errno on failure.
3820 int apply_workqueue_attrs(struct workqueue_struct *wq,
3821 const struct workqueue_attrs *attrs)
3823 struct workqueue_attrs *new_attrs, *tmp_attrs;
3824 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3827 /* only unbound workqueues can change attributes */
3828 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3831 /* creating multiple pwqs breaks ordering guarantee */
3832 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3835 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3836 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3837 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3838 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3841 /* make a copy of @attrs and sanitize it */
3842 copy_workqueue_attrs(new_attrs, attrs);
3843 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3846 * We may create multiple pwqs with differing cpumasks. Make a
3847 * copy of @new_attrs which will be modified and used to obtain
3850 copy_workqueue_attrs(tmp_attrs, new_attrs);
3853 * CPUs should stay stable across pwq creations and installations.
3854 * Pin CPUs, determine the target cpumask for each node and create
3859 mutex_lock(&wq_pool_mutex);
3862 * If something goes wrong during CPU up/down, we'll fall back to
3863 * the default pwq covering whole @attrs->cpumask. Always create
3864 * it even if we don't use it immediately.
3866 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3870 for_each_node(node) {
3871 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3872 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3877 pwq_tbl[node] = dfl_pwq;
3881 mutex_unlock(&wq_pool_mutex);
3883 /* all pwqs have been created successfully, let's install'em */
3884 mutex_lock(&wq->mutex);
3886 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3888 /* save the previous pwq and install the new one */
3890 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3892 /* @dfl_pwq might not have been used, ensure it's linked */
3894 swap(wq->dfl_pwq, dfl_pwq);
3896 mutex_unlock(&wq->mutex);
3898 /* put the old pwqs */
3900 put_pwq_unlocked(pwq_tbl[node]);
3901 put_pwq_unlocked(dfl_pwq);
3907 free_workqueue_attrs(tmp_attrs);
3908 free_workqueue_attrs(new_attrs);
3913 free_unbound_pwq(dfl_pwq);
3915 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3916 free_unbound_pwq(pwq_tbl[node]);
3917 mutex_unlock(&wq_pool_mutex);
3925 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3926 * @wq: the target workqueue
3927 * @cpu: the CPU coming up or going down
3928 * @online: whether @cpu is coming up or going down
3930 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3931 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3934 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3935 * falls back to @wq->dfl_pwq which may not be optimal but is always
3938 * Note that when the last allowed CPU of a NUMA node goes offline for a
3939 * workqueue with a cpumask spanning multiple nodes, the workers which were
3940 * already executing the work items for the workqueue will lose their CPU
3941 * affinity and may execute on any CPU. This is similar to how per-cpu
3942 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3943 * affinity, it's the user's responsibility to flush the work item from
3946 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3949 int node = cpu_to_node(cpu);
3950 int cpu_off = online ? -1 : cpu;
3951 struct pool_workqueue *old_pwq = NULL, *pwq;
3952 struct workqueue_attrs *target_attrs;
3955 lockdep_assert_held(&wq_pool_mutex);
3957 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3961 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3962 * Let's use a preallocated one. The following buf is protected by
3963 * CPU hotplug exclusion.
3965 target_attrs = wq_update_unbound_numa_attrs_buf;
3966 cpumask = target_attrs->cpumask;
3968 mutex_lock(&wq->mutex);
3969 if (wq->unbound_attrs->no_numa)
3972 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3973 pwq = unbound_pwq_by_node(wq, node);
3976 * Let's determine what needs to be done. If the target cpumask is
3977 * different from wq's, we need to compare it to @pwq's and create
3978 * a new one if they don't match. If the target cpumask equals
3979 * wq's, the default pwq should be used.
3981 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
3982 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3988 mutex_unlock(&wq->mutex);
3990 /* create a new pwq */
3991 pwq = alloc_unbound_pwq(wq, target_attrs);
3993 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3995 mutex_lock(&wq->mutex);
4000 * Install the new pwq. As this function is called only from CPU
4001 * hotplug callbacks and applying a new attrs is wrapped with
4002 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4005 mutex_lock(&wq->mutex);
4006 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4010 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4011 get_pwq(wq->dfl_pwq);
4012 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4013 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4015 mutex_unlock(&wq->mutex);
4016 put_pwq_unlocked(old_pwq);
4019 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4021 bool highpri = wq->flags & WQ_HIGHPRI;
4024 if (!(wq->flags & WQ_UNBOUND)) {
4025 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4029 for_each_possible_cpu(cpu) {
4030 struct pool_workqueue *pwq =
4031 per_cpu_ptr(wq->cpu_pwqs, cpu);
4032 struct worker_pool *cpu_pools =
4033 per_cpu(cpu_worker_pools, cpu);
4035 init_pwq(pwq, wq, &cpu_pools[highpri]);
4037 mutex_lock(&wq->mutex);
4039 mutex_unlock(&wq->mutex);
4042 } else if (wq->flags & __WQ_ORDERED) {
4043 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4044 /* there should only be single pwq for ordering guarantee */
4045 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4046 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4047 "ordering guarantee broken for workqueue %s\n", wq->name);
4050 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4054 static int wq_clamp_max_active(int max_active, unsigned int flags,
4057 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4059 if (max_active < 1 || max_active > lim)
4060 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4061 max_active, name, 1, lim);
4063 return clamp_val(max_active, 1, lim);
4066 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4069 struct lock_class_key *key,
4070 const char *lock_name, ...)
4072 size_t tbl_size = 0;
4074 struct workqueue_struct *wq;
4075 struct pool_workqueue *pwq;
4077 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4078 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4079 flags |= WQ_UNBOUND;
4081 /* allocate wq and format name */
4082 if (flags & WQ_UNBOUND)
4083 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4085 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4089 if (flags & WQ_UNBOUND) {
4090 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4091 if (!wq->unbound_attrs)
4095 va_start(args, lock_name);
4096 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4099 max_active = max_active ?: WQ_DFL_ACTIVE;
4100 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4104 wq->saved_max_active = max_active;
4105 mutex_init(&wq->mutex);
4106 atomic_set(&wq->nr_pwqs_to_flush, 0);
4107 INIT_LIST_HEAD(&wq->pwqs);
4108 INIT_LIST_HEAD(&wq->flusher_queue);
4109 INIT_LIST_HEAD(&wq->flusher_overflow);
4110 INIT_LIST_HEAD(&wq->maydays);
4112 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4113 INIT_LIST_HEAD(&wq->list);
4115 if (alloc_and_link_pwqs(wq) < 0)
4119 * Workqueues which may be used during memory reclaim should
4120 * have a rescuer to guarantee forward progress.
4122 if (flags & WQ_MEM_RECLAIM) {
4123 struct worker *rescuer;
4125 rescuer = alloc_worker();
4129 rescuer->rescue_wq = wq;
4130 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4132 if (IS_ERR(rescuer->task)) {
4137 wq->rescuer = rescuer;
4138 rescuer->task->flags |= PF_NO_SETAFFINITY;
4139 wake_up_process(rescuer->task);
4142 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4146 * wq_pool_mutex protects global freeze state and workqueues list.
4147 * Grab it, adjust max_active and add the new @wq to workqueues
4150 mutex_lock(&wq_pool_mutex);
4152 mutex_lock(&wq->mutex);
4153 for_each_pwq(pwq, wq)
4154 pwq_adjust_max_active(pwq);
4155 mutex_unlock(&wq->mutex);
4157 list_add(&wq->list, &workqueues);
4159 mutex_unlock(&wq_pool_mutex);
4164 free_workqueue_attrs(wq->unbound_attrs);
4168 destroy_workqueue(wq);
4171 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4174 * destroy_workqueue - safely terminate a workqueue
4175 * @wq: target workqueue
4177 * Safely destroy a workqueue. All work currently pending will be done first.
4179 void destroy_workqueue(struct workqueue_struct *wq)
4181 struct pool_workqueue *pwq;
4184 /* drain it before proceeding with destruction */
4185 drain_workqueue(wq);
4188 mutex_lock(&wq->mutex);
4189 for_each_pwq(pwq, wq) {
4192 for (i = 0; i < WORK_NR_COLORS; i++) {
4193 if (WARN_ON(pwq->nr_in_flight[i])) {
4194 mutex_unlock(&wq->mutex);
4199 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4200 WARN_ON(pwq->nr_active) ||
4201 WARN_ON(!list_empty(&pwq->delayed_works))) {
4202 mutex_unlock(&wq->mutex);
4206 mutex_unlock(&wq->mutex);
4209 * wq list is used to freeze wq, remove from list after
4210 * flushing is complete in case freeze races us.
4212 mutex_lock(&wq_pool_mutex);
4213 list_del_init(&wq->list);
4214 mutex_unlock(&wq_pool_mutex);
4216 workqueue_sysfs_unregister(wq);
4219 kthread_stop(wq->rescuer->task);
4224 if (!(wq->flags & WQ_UNBOUND)) {
4226 * The base ref is never dropped on per-cpu pwqs. Directly
4227 * free the pwqs and wq.
4229 free_percpu(wq->cpu_pwqs);
4233 * We're the sole accessor of @wq at this point. Directly
4234 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4235 * @wq will be freed when the last pwq is released.
4237 for_each_node(node) {
4238 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4239 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4240 put_pwq_unlocked(pwq);
4244 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4245 * put. Don't access it afterwards.
4249 put_pwq_unlocked(pwq);
4252 EXPORT_SYMBOL_GPL(destroy_workqueue);
4255 * workqueue_set_max_active - adjust max_active of a workqueue
4256 * @wq: target workqueue
4257 * @max_active: new max_active value.
4259 * Set max_active of @wq to @max_active.
4262 * Don't call from IRQ context.
4264 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4266 struct pool_workqueue *pwq;
4268 /* disallow meddling with max_active for ordered workqueues */
4269 if (WARN_ON(wq->flags & __WQ_ORDERED))
4272 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4274 mutex_lock(&wq->mutex);
4276 wq->saved_max_active = max_active;
4278 for_each_pwq(pwq, wq)
4279 pwq_adjust_max_active(pwq);
4281 mutex_unlock(&wq->mutex);
4283 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4286 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4288 * Determine whether %current is a workqueue rescuer. Can be used from
4289 * work functions to determine whether it's being run off the rescuer task.
4291 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4293 bool current_is_workqueue_rescuer(void)
4295 struct worker *worker = current_wq_worker();
4297 return worker && worker->rescue_wq;
4301 * workqueue_congested - test whether a workqueue is congested
4302 * @cpu: CPU in question
4303 * @wq: target workqueue
4305 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4306 * no synchronization around this function and the test result is
4307 * unreliable and only useful as advisory hints or for debugging.
4309 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4310 * Note that both per-cpu and unbound workqueues may be associated with
4311 * multiple pool_workqueues which have separate congested states. A
4312 * workqueue being congested on one CPU doesn't mean the workqueue is also
4313 * contested on other CPUs / NUMA nodes.
4316 * %true if congested, %false otherwise.
4318 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4320 struct pool_workqueue *pwq;
4323 rcu_read_lock_sched();
4325 if (cpu == WORK_CPU_UNBOUND)
4326 cpu = smp_processor_id();
4328 if (!(wq->flags & WQ_UNBOUND))
4329 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4331 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4333 ret = !list_empty(&pwq->delayed_works);
4334 rcu_read_unlock_sched();
4338 EXPORT_SYMBOL_GPL(workqueue_congested);
4341 * work_busy - test whether a work is currently pending or running
4342 * @work: the work to be tested
4344 * Test whether @work is currently pending or running. There is no
4345 * synchronization around this function and the test result is
4346 * unreliable and only useful as advisory hints or for debugging.
4349 * OR'd bitmask of WORK_BUSY_* bits.
4351 unsigned int work_busy(struct work_struct *work)
4353 struct worker_pool *pool;
4354 unsigned long flags;
4355 unsigned int ret = 0;
4357 if (work_pending(work))
4358 ret |= WORK_BUSY_PENDING;
4360 local_irq_save(flags);
4361 pool = get_work_pool(work);
4363 spin_lock(&pool->lock);
4364 if (find_worker_executing_work(pool, work))
4365 ret |= WORK_BUSY_RUNNING;
4366 spin_unlock(&pool->lock);
4368 local_irq_restore(flags);
4372 EXPORT_SYMBOL_GPL(work_busy);
4375 * set_worker_desc - set description for the current work item
4376 * @fmt: printf-style format string
4377 * @...: arguments for the format string
4379 * This function can be called by a running work function to describe what
4380 * the work item is about. If the worker task gets dumped, this
4381 * information will be printed out together to help debugging. The
4382 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4384 void set_worker_desc(const char *fmt, ...)
4386 struct worker *worker = current_wq_worker();
4390 va_start(args, fmt);
4391 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4393 worker->desc_valid = true;
4398 * print_worker_info - print out worker information and description
4399 * @log_lvl: the log level to use when printing
4400 * @task: target task
4402 * If @task is a worker and currently executing a work item, print out the
4403 * name of the workqueue being serviced and worker description set with
4404 * set_worker_desc() by the currently executing work item.
4406 * This function can be safely called on any task as long as the
4407 * task_struct itself is accessible. While safe, this function isn't
4408 * synchronized and may print out mixups or garbages of limited length.
4410 void print_worker_info(const char *log_lvl, struct task_struct *task)
4412 work_func_t *fn = NULL;
4413 char name[WQ_NAME_LEN] = { };
4414 char desc[WORKER_DESC_LEN] = { };
4415 struct pool_workqueue *pwq = NULL;
4416 struct workqueue_struct *wq = NULL;
4417 bool desc_valid = false;
4418 struct worker *worker;
4420 if (!(task->flags & PF_WQ_WORKER))
4424 * This function is called without any synchronization and @task
4425 * could be in any state. Be careful with dereferences.
4427 worker = probe_kthread_data(task);
4430 * Carefully copy the associated workqueue's workfn and name. Keep
4431 * the original last '\0' in case the original contains garbage.
4433 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4434 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4435 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4436 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4438 /* copy worker description */
4439 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4441 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4443 if (fn || name[0] || desc[0]) {
4444 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4446 pr_cont(" (%s)", desc);
4454 * There are two challenges in supporting CPU hotplug. Firstly, there
4455 * are a lot of assumptions on strong associations among work, pwq and
4456 * pool which make migrating pending and scheduled works very
4457 * difficult to implement without impacting hot paths. Secondly,
4458 * worker pools serve mix of short, long and very long running works making
4459 * blocked draining impractical.
4461 * This is solved by allowing the pools to be disassociated from the CPU
4462 * running as an unbound one and allowing it to be reattached later if the
4463 * cpu comes back online.
4466 static void wq_unbind_fn(struct work_struct *work)
4468 int cpu = smp_processor_id();
4469 struct worker_pool *pool;
4470 struct worker *worker;
4472 for_each_cpu_worker_pool(pool, cpu) {
4473 WARN_ON_ONCE(cpu != smp_processor_id());
4475 mutex_lock(&pool->attach_mutex);
4476 spin_lock_irq(&pool->lock);
4479 * We've blocked all attach/detach operations. Make all workers
4480 * unbound and set DISASSOCIATED. Before this, all workers
4481 * except for the ones which are still executing works from
4482 * before the last CPU down must be on the cpu. After
4483 * this, they may become diasporas.
4485 for_each_pool_worker(worker, pool)
4486 worker->flags |= WORKER_UNBOUND;
4488 pool->flags |= POOL_DISASSOCIATED;
4490 spin_unlock_irq(&pool->lock);
4491 mutex_unlock(&pool->attach_mutex);
4494 * Call schedule() so that we cross rq->lock and thus can
4495 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4496 * This is necessary as scheduler callbacks may be invoked
4502 * Sched callbacks are disabled now. Zap nr_running.
4503 * After this, nr_running stays zero and need_more_worker()
4504 * and keep_working() are always true as long as the
4505 * worklist is not empty. This pool now behaves as an
4506 * unbound (in terms of concurrency management) pool which
4507 * are served by workers tied to the pool.
4509 atomic_set(&pool->nr_running, 0);
4512 * With concurrency management just turned off, a busy
4513 * worker blocking could lead to lengthy stalls. Kick off
4514 * unbound chain execution of currently pending work items.
4516 spin_lock_irq(&pool->lock);
4517 wake_up_worker(pool);
4518 spin_unlock_irq(&pool->lock);
4523 * rebind_workers - rebind all workers of a pool to the associated CPU
4524 * @pool: pool of interest
4526 * @pool->cpu is coming online. Rebind all workers to the CPU.
4528 static void rebind_workers(struct worker_pool *pool)
4530 struct worker *worker;
4532 lockdep_assert_held(&pool->attach_mutex);
4535 * Restore CPU affinity of all workers. As all idle workers should
4536 * be on the run-queue of the associated CPU before any local
4537 * wake-ups for concurrency management happen, restore CPU affinty
4538 * of all workers first and then clear UNBOUND. As we're called
4539 * from CPU_ONLINE, the following shouldn't fail.
4541 for_each_pool_worker(worker, pool)
4542 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4543 pool->attrs->cpumask) < 0);
4545 spin_lock_irq(&pool->lock);
4547 for_each_pool_worker(worker, pool) {
4548 unsigned int worker_flags = worker->flags;
4551 * A bound idle worker should actually be on the runqueue
4552 * of the associated CPU for local wake-ups targeting it to
4553 * work. Kick all idle workers so that they migrate to the
4554 * associated CPU. Doing this in the same loop as
4555 * replacing UNBOUND with REBOUND is safe as no worker will
4556 * be bound before @pool->lock is released.
4558 if (worker_flags & WORKER_IDLE)
4559 wake_up_process(worker->task);
4562 * We want to clear UNBOUND but can't directly call
4563 * worker_clr_flags() or adjust nr_running. Atomically
4564 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4565 * @worker will clear REBOUND using worker_clr_flags() when
4566 * it initiates the next execution cycle thus restoring
4567 * concurrency management. Note that when or whether
4568 * @worker clears REBOUND doesn't affect correctness.
4570 * ACCESS_ONCE() is necessary because @worker->flags may be
4571 * tested without holding any lock in
4572 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4573 * fail incorrectly leading to premature concurrency
4574 * management operations.
4576 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4577 worker_flags |= WORKER_REBOUND;
4578 worker_flags &= ~WORKER_UNBOUND;
4579 ACCESS_ONCE(worker->flags) = worker_flags;
4582 spin_unlock_irq(&pool->lock);
4586 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4587 * @pool: unbound pool of interest
4588 * @cpu: the CPU which is coming up
4590 * An unbound pool may end up with a cpumask which doesn't have any online
4591 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4592 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4593 * online CPU before, cpus_allowed of all its workers should be restored.
4595 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4597 static cpumask_t cpumask;
4598 struct worker *worker;
4600 lockdep_assert_held(&pool->attach_mutex);
4602 /* is @cpu allowed for @pool? */
4603 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4606 /* is @cpu the only online CPU? */
4607 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4608 if (cpumask_weight(&cpumask) != 1)
4611 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4612 for_each_pool_worker(worker, pool)
4613 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4614 pool->attrs->cpumask) < 0);
4618 * Workqueues should be brought up before normal priority CPU notifiers.
4619 * This will be registered high priority CPU notifier.
4621 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4622 unsigned long action,
4625 int cpu = (unsigned long)hcpu;
4626 struct worker_pool *pool;
4627 struct workqueue_struct *wq;
4630 switch (action & ~CPU_TASKS_FROZEN) {
4631 case CPU_UP_PREPARE:
4632 for_each_cpu_worker_pool(pool, cpu) {
4633 if (pool->nr_workers)
4635 if (create_and_start_worker(pool) < 0)
4640 case CPU_DOWN_FAILED:
4642 mutex_lock(&wq_pool_mutex);
4644 for_each_pool(pool, pi) {
4645 mutex_lock(&pool->attach_mutex);
4647 if (pool->cpu == cpu) {
4648 spin_lock_irq(&pool->lock);
4649 pool->flags &= ~POOL_DISASSOCIATED;
4650 spin_unlock_irq(&pool->lock);
4652 rebind_workers(pool);
4653 } else if (pool->cpu < 0) {
4654 restore_unbound_workers_cpumask(pool, cpu);
4657 mutex_unlock(&pool->attach_mutex);
4660 /* update NUMA affinity of unbound workqueues */
4661 list_for_each_entry(wq, &workqueues, list)
4662 wq_update_unbound_numa(wq, cpu, true);
4664 mutex_unlock(&wq_pool_mutex);
4671 * Workqueues should be brought down after normal priority CPU notifiers.
4672 * This will be registered as low priority CPU notifier.
4674 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4675 unsigned long action,
4678 int cpu = (unsigned long)hcpu;
4679 struct work_struct unbind_work;
4680 struct workqueue_struct *wq;
4682 switch (action & ~CPU_TASKS_FROZEN) {
4683 case CPU_DOWN_PREPARE:
4684 /* unbinding per-cpu workers should happen on the local CPU */
4685 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4686 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4688 /* update NUMA affinity of unbound workqueues */
4689 mutex_lock(&wq_pool_mutex);
4690 list_for_each_entry(wq, &workqueues, list)
4691 wq_update_unbound_numa(wq, cpu, false);
4692 mutex_unlock(&wq_pool_mutex);
4694 /* wait for per-cpu unbinding to finish */
4695 flush_work(&unbind_work);
4696 destroy_work_on_stack(&unbind_work);
4704 struct work_for_cpu {
4705 struct work_struct work;
4711 static void work_for_cpu_fn(struct work_struct *work)
4713 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4715 wfc->ret = wfc->fn(wfc->arg);
4719 * work_on_cpu - run a function in user context on a particular cpu
4720 * @cpu: the cpu to run on
4721 * @fn: the function to run
4722 * @arg: the function arg
4724 * It is up to the caller to ensure that the cpu doesn't go offline.
4725 * The caller must not hold any locks which would prevent @fn from completing.
4727 * Return: The value @fn returns.
4729 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4731 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4733 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4734 schedule_work_on(cpu, &wfc.work);
4735 flush_work(&wfc.work);
4736 destroy_work_on_stack(&wfc.work);
4739 EXPORT_SYMBOL_GPL(work_on_cpu);
4740 #endif /* CONFIG_SMP */
4742 #ifdef CONFIG_FREEZER
4745 * freeze_workqueues_begin - begin freezing workqueues
4747 * Start freezing workqueues. After this function returns, all freezable
4748 * workqueues will queue new works to their delayed_works list instead of
4752 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4754 void freeze_workqueues_begin(void)
4756 struct workqueue_struct *wq;
4757 struct pool_workqueue *pwq;
4759 mutex_lock(&wq_pool_mutex);
4761 WARN_ON_ONCE(workqueue_freezing);
4762 workqueue_freezing = true;
4764 list_for_each_entry(wq, &workqueues, list) {
4765 mutex_lock(&wq->mutex);
4766 for_each_pwq(pwq, wq)
4767 pwq_adjust_max_active(pwq);
4768 mutex_unlock(&wq->mutex);
4771 mutex_unlock(&wq_pool_mutex);
4775 * freeze_workqueues_busy - are freezable workqueues still busy?
4777 * Check whether freezing is complete. This function must be called
4778 * between freeze_workqueues_begin() and thaw_workqueues().
4781 * Grabs and releases wq_pool_mutex.
4784 * %true if some freezable workqueues are still busy. %false if freezing
4787 bool freeze_workqueues_busy(void)
4790 struct workqueue_struct *wq;
4791 struct pool_workqueue *pwq;
4793 mutex_lock(&wq_pool_mutex);
4795 WARN_ON_ONCE(!workqueue_freezing);
4797 list_for_each_entry(wq, &workqueues, list) {
4798 if (!(wq->flags & WQ_FREEZABLE))
4801 * nr_active is monotonically decreasing. It's safe
4802 * to peek without lock.
4804 rcu_read_lock_sched();
4805 for_each_pwq(pwq, wq) {
4806 WARN_ON_ONCE(pwq->nr_active < 0);
4807 if (pwq->nr_active) {
4809 rcu_read_unlock_sched();
4813 rcu_read_unlock_sched();
4816 mutex_unlock(&wq_pool_mutex);
4821 * thaw_workqueues - thaw workqueues
4823 * Thaw workqueues. Normal queueing is restored and all collected
4824 * frozen works are transferred to their respective pool worklists.
4827 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4829 void thaw_workqueues(void)
4831 struct workqueue_struct *wq;
4832 struct pool_workqueue *pwq;
4834 mutex_lock(&wq_pool_mutex);
4836 if (!workqueue_freezing)
4839 workqueue_freezing = false;
4841 /* restore max_active and repopulate worklist */
4842 list_for_each_entry(wq, &workqueues, list) {
4843 mutex_lock(&wq->mutex);
4844 for_each_pwq(pwq, wq)
4845 pwq_adjust_max_active(pwq);
4846 mutex_unlock(&wq->mutex);
4850 mutex_unlock(&wq_pool_mutex);
4852 #endif /* CONFIG_FREEZER */
4854 static void __init wq_numa_init(void)
4859 /* determine NUMA pwq table len - highest node id + 1 */
4861 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4863 if (num_possible_nodes() <= 1)
4866 if (wq_disable_numa) {
4867 pr_info("workqueue: NUMA affinity support disabled\n");
4871 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4872 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4875 * We want masks of possible CPUs of each node which isn't readily
4876 * available. Build one from cpu_to_node() which should have been
4877 * fully initialized by now.
4879 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4883 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4884 node_online(node) ? node : NUMA_NO_NODE));
4886 for_each_possible_cpu(cpu) {
4887 node = cpu_to_node(cpu);
4888 if (WARN_ON(node == NUMA_NO_NODE)) {
4889 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4890 /* happens iff arch is bonkers, let's just proceed */
4893 cpumask_set_cpu(cpu, tbl[node]);
4896 wq_numa_possible_cpumask = tbl;
4897 wq_numa_enabled = true;
4900 static int __init init_workqueues(void)
4902 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4905 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4907 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4909 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4910 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4914 /* initialize CPU pools */
4915 for_each_possible_cpu(cpu) {
4916 struct worker_pool *pool;
4919 for_each_cpu_worker_pool(pool, cpu) {
4920 BUG_ON(init_worker_pool(pool));
4922 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4923 pool->attrs->nice = std_nice[i++];
4924 pool->node = cpu_to_node(cpu);
4927 mutex_lock(&wq_pool_mutex);
4928 BUG_ON(worker_pool_assign_id(pool));
4929 mutex_unlock(&wq_pool_mutex);
4933 /* create the initial worker */
4934 for_each_online_cpu(cpu) {
4935 struct worker_pool *pool;
4937 for_each_cpu_worker_pool(pool, cpu) {
4938 pool->flags &= ~POOL_DISASSOCIATED;
4939 BUG_ON(create_and_start_worker(pool) < 0);
4943 /* create default unbound and ordered wq attrs */
4944 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4945 struct workqueue_attrs *attrs;
4947 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4948 attrs->nice = std_nice[i];
4949 unbound_std_wq_attrs[i] = attrs;
4952 * An ordered wq should have only one pwq as ordering is
4953 * guaranteed by max_active which is enforced by pwqs.
4954 * Turn off NUMA so that dfl_pwq is used for all nodes.
4956 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4957 attrs->nice = std_nice[i];
4958 attrs->no_numa = true;
4959 ordered_wq_attrs[i] = attrs;
4962 system_wq = alloc_workqueue("events", 0, 0);
4963 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4964 system_long_wq = alloc_workqueue("events_long", 0, 0);
4965 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4966 WQ_UNBOUND_MAX_ACTIVE);
4967 system_freezable_wq = alloc_workqueue("events_freezable",
4969 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
4970 WQ_POWER_EFFICIENT, 0);
4971 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
4972 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
4974 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4975 !system_unbound_wq || !system_freezable_wq ||
4976 !system_power_efficient_wq ||
4977 !system_freezable_power_efficient_wq);
4980 early_initcall(init_workqueues);