2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
135 * WQ: wq->mutex protected.
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
139 * MD: wq_mayday_lock protected.
142 /* struct worker is defined in workqueue_internal.h */
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
151 struct list_head worklist; /* L: list of pending works */
152 int nr_workers; /* L: total number of workers */
154 /* nr_idle includes the ones off idle_list for rebinding */
155 int nr_idle; /* L: currently idle ones */
157 struct list_head idle_list; /* X: list of idle workers */
158 struct timer_list idle_timer; /* L: worker idle timeout */
159 struct timer_list mayday_timer; /* L: SOS timer for workers */
161 /* a workers is either on busy_hash or idle_list, or the manager */
162 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
163 /* L: hash of busy workers */
165 /* see manage_workers() for details on the two manager mutexes */
166 struct mutex manager_arb; /* manager arbitration */
167 struct worker *manager; /* L: purely informational */
168 struct mutex attach_mutex; /* attach/detach exclusion */
169 struct list_head workers; /* A: attached workers */
170 struct completion *detach_completion; /* all workers detached */
172 struct ida worker_ida; /* worker IDs for task name */
174 struct workqueue_attrs *attrs; /* I: worker attributes */
175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
176 int refcnt; /* PL: refcnt for unbound pools */
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
183 atomic_t nr_running ____cacheline_aligned_in_smp;
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
190 } ____cacheline_aligned_in_smp;
193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
198 struct pool_workqueue {
199 struct worker_pool *pool; /* I: the associated pool */
200 struct workqueue_struct *wq; /* I: the owning workqueue */
201 int work_color; /* L: current color */
202 int flush_color; /* L: flushing color */
203 int refcnt; /* L: reference count */
204 int nr_in_flight[WORK_NR_COLORS];
205 /* L: nr of in_flight works */
206 int nr_active; /* L: nr of active works */
207 int max_active; /* L: max active works */
208 struct list_head delayed_works; /* L: delayed works */
209 struct list_head pwqs_node; /* WR: node on wq->pwqs */
210 struct list_head mayday_node; /* MD: node on wq->maydays */
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
216 * determined without grabbing wq->mutex.
218 struct work_struct unbound_release_work;
220 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
223 * Structure used to wait for workqueue flush.
226 struct list_head list; /* WQ: list of flushers */
227 int flush_color; /* WQ: flush color waiting for */
228 struct completion done; /* flush completion */
234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
237 struct workqueue_struct {
238 struct list_head pwqs; /* WR: all pwqs of this wq */
239 struct list_head list; /* PR: list of all workqueues */
241 struct mutex mutex; /* protects this wq */
242 int work_color; /* WQ: current work color */
243 int flush_color; /* WQ: current flush color */
244 atomic_t nr_pwqs_to_flush; /* flush in progress */
245 struct wq_flusher *first_flusher; /* WQ: first flusher */
246 struct list_head flusher_queue; /* WQ: flush waiters */
247 struct list_head flusher_overflow; /* WQ: flush overflow list */
249 struct list_head maydays; /* MD: pwqs requesting rescue */
250 struct worker *rescuer; /* I: rescue worker */
252 int nr_drainers; /* WQ: drain in progress */
253 int saved_max_active; /* WQ: saved pwq max_active */
255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259 struct wq_device *wq_dev; /* I: for sysfs interface */
261 #ifdef CONFIG_LOCKDEP
262 struct lockdep_map lockdep_map;
264 char name[WQ_NAME_LEN]; /* I: workqueue name */
267 * Destruction of workqueue_struct is sched-RCU protected to allow
268 * walking the workqueues list without grabbing wq_pool_mutex.
269 * This is used to dump all workqueues from sysrq.
273 /* hot fields used during command issue, aligned to cacheline */
274 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
275 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
276 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279 static struct kmem_cache *pwq_cache;
281 static cpumask_var_t *wq_numa_possible_cpumask;
282 /* possible CPUs of each node */
284 static bool wq_disable_numa;
285 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
287 /* see the comment above the definition of WQ_POWER_EFFICIENT */
288 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
289 static bool wq_power_efficient = true;
291 static bool wq_power_efficient;
294 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
296 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
298 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
301 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
302 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
304 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
305 static bool workqueue_freezing; /* PL: have wqs started freezing? */
307 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
309 /* the per-cpu worker pools */
310 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
313 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
315 /* PL: hash of all unbound pools keyed by pool->attrs */
316 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
318 /* I: attributes used when instantiating standard unbound pools on demand */
319 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
321 /* I: attributes used when instantiating ordered pools on demand */
322 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
324 struct workqueue_struct *system_wq __read_mostly;
325 EXPORT_SYMBOL(system_wq);
326 struct workqueue_struct *system_highpri_wq __read_mostly;
327 EXPORT_SYMBOL_GPL(system_highpri_wq);
328 struct workqueue_struct *system_long_wq __read_mostly;
329 EXPORT_SYMBOL_GPL(system_long_wq);
330 struct workqueue_struct *system_unbound_wq __read_mostly;
331 EXPORT_SYMBOL_GPL(system_unbound_wq);
332 struct workqueue_struct *system_freezable_wq __read_mostly;
333 EXPORT_SYMBOL_GPL(system_freezable_wq);
334 struct workqueue_struct *system_power_efficient_wq __read_mostly;
335 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
336 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
337 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
339 static int worker_thread(void *__worker);
340 static void copy_workqueue_attrs(struct workqueue_attrs *to,
341 const struct workqueue_attrs *from);
342 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
344 #define CREATE_TRACE_POINTS
345 #include <trace/events/workqueue.h>
347 #define assert_rcu_or_pool_mutex() \
348 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
349 lockdep_is_held(&wq_pool_mutex), \
350 "sched RCU or wq_pool_mutex should be held")
352 #define assert_rcu_or_wq_mutex(wq) \
353 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
354 lockdep_is_held(&wq->mutex), \
355 "sched RCU or wq->mutex should be held")
357 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
358 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
359 lockdep_is_held(&wq->mutex) || \
360 lockdep_is_held(&wq_pool_mutex), \
361 "sched RCU, wq->mutex or wq_pool_mutex should be held")
363 #define for_each_cpu_worker_pool(pool, cpu) \
364 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
365 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
369 * for_each_pool - iterate through all worker_pools in the system
370 * @pool: iteration cursor
371 * @pi: integer used for iteration
373 * This must be called either with wq_pool_mutex held or sched RCU read
374 * locked. If the pool needs to be used beyond the locking in effect, the
375 * caller is responsible for guaranteeing that the pool stays online.
377 * The if/else clause exists only for the lockdep assertion and can be
380 #define for_each_pool(pool, pi) \
381 idr_for_each_entry(&worker_pool_idr, pool, pi) \
382 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
386 * for_each_pool_worker - iterate through all workers of a worker_pool
387 * @worker: iteration cursor
388 * @pool: worker_pool to iterate workers of
390 * This must be called with @pool->attach_mutex.
392 * The if/else clause exists only for the lockdep assertion and can be
395 #define for_each_pool_worker(worker, pool) \
396 list_for_each_entry((worker), &(pool)->workers, node) \
397 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
401 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
402 * @pwq: iteration cursor
403 * @wq: the target workqueue
405 * This must be called either with wq->mutex held or sched RCU read locked.
406 * If the pwq needs to be used beyond the locking in effect, the caller is
407 * responsible for guaranteeing that the pwq stays online.
409 * The if/else clause exists only for the lockdep assertion and can be
412 #define for_each_pwq(pwq, wq) \
413 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
414 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
417 #ifdef CONFIG_DEBUG_OBJECTS_WORK
419 static struct debug_obj_descr work_debug_descr;
421 static void *work_debug_hint(void *addr)
423 return ((struct work_struct *) addr)->func;
427 * fixup_init is called when:
428 * - an active object is initialized
430 static int work_fixup_init(void *addr, enum debug_obj_state state)
432 struct work_struct *work = addr;
435 case ODEBUG_STATE_ACTIVE:
436 cancel_work_sync(work);
437 debug_object_init(work, &work_debug_descr);
445 * fixup_activate is called when:
446 * - an active object is activated
447 * - an unknown object is activated (might be a statically initialized object)
449 static int work_fixup_activate(void *addr, enum debug_obj_state state)
451 struct work_struct *work = addr;
455 case ODEBUG_STATE_NOTAVAILABLE:
457 * This is not really a fixup. The work struct was
458 * statically initialized. We just make sure that it
459 * is tracked in the object tracker.
461 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
462 debug_object_init(work, &work_debug_descr);
463 debug_object_activate(work, &work_debug_descr);
469 case ODEBUG_STATE_ACTIVE:
478 * fixup_free is called when:
479 * - an active object is freed
481 static int work_fixup_free(void *addr, enum debug_obj_state state)
483 struct work_struct *work = addr;
486 case ODEBUG_STATE_ACTIVE:
487 cancel_work_sync(work);
488 debug_object_free(work, &work_debug_descr);
495 static struct debug_obj_descr work_debug_descr = {
496 .name = "work_struct",
497 .debug_hint = work_debug_hint,
498 .fixup_init = work_fixup_init,
499 .fixup_activate = work_fixup_activate,
500 .fixup_free = work_fixup_free,
503 static inline void debug_work_activate(struct work_struct *work)
505 debug_object_activate(work, &work_debug_descr);
508 static inline void debug_work_deactivate(struct work_struct *work)
510 debug_object_deactivate(work, &work_debug_descr);
513 void __init_work(struct work_struct *work, int onstack)
516 debug_object_init_on_stack(work, &work_debug_descr);
518 debug_object_init(work, &work_debug_descr);
520 EXPORT_SYMBOL_GPL(__init_work);
522 void destroy_work_on_stack(struct work_struct *work)
524 debug_object_free(work, &work_debug_descr);
526 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
528 void destroy_delayed_work_on_stack(struct delayed_work *work)
530 destroy_timer_on_stack(&work->timer);
531 debug_object_free(&work->work, &work_debug_descr);
533 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
536 static inline void debug_work_activate(struct work_struct *work) { }
537 static inline void debug_work_deactivate(struct work_struct *work) { }
541 * worker_pool_assign_id - allocate ID and assing it to @pool
542 * @pool: the pool pointer of interest
544 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
545 * successfully, -errno on failure.
547 static int worker_pool_assign_id(struct worker_pool *pool)
551 lockdep_assert_held(&wq_pool_mutex);
553 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
563 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
564 * @wq: the target workqueue
567 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
569 * If the pwq needs to be used beyond the locking in effect, the caller is
570 * responsible for guaranteeing that the pwq stays online.
572 * Return: The unbound pool_workqueue for @node.
574 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
577 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
578 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
581 static unsigned int work_color_to_flags(int color)
583 return color << WORK_STRUCT_COLOR_SHIFT;
586 static int get_work_color(struct work_struct *work)
588 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
589 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
592 static int work_next_color(int color)
594 return (color + 1) % WORK_NR_COLORS;
598 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
599 * contain the pointer to the queued pwq. Once execution starts, the flag
600 * is cleared and the high bits contain OFFQ flags and pool ID.
602 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
603 * and clear_work_data() can be used to set the pwq, pool or clear
604 * work->data. These functions should only be called while the work is
605 * owned - ie. while the PENDING bit is set.
607 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
608 * corresponding to a work. Pool is available once the work has been
609 * queued anywhere after initialization until it is sync canceled. pwq is
610 * available only while the work item is queued.
612 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
613 * canceled. While being canceled, a work item may have its PENDING set
614 * but stay off timer and worklist for arbitrarily long and nobody should
615 * try to steal the PENDING bit.
617 static inline void set_work_data(struct work_struct *work, unsigned long data,
620 WARN_ON_ONCE(!work_pending(work));
621 atomic_long_set(&work->data, data | flags | work_static(work));
624 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
625 unsigned long extra_flags)
627 set_work_data(work, (unsigned long)pwq,
628 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
631 static void set_work_pool_and_keep_pending(struct work_struct *work,
634 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
635 WORK_STRUCT_PENDING);
638 static void set_work_pool_and_clear_pending(struct work_struct *work,
642 * The following wmb is paired with the implied mb in
643 * test_and_set_bit(PENDING) and ensures all updates to @work made
644 * here are visible to and precede any updates by the next PENDING
648 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
651 static void clear_work_data(struct work_struct *work)
653 smp_wmb(); /* see set_work_pool_and_clear_pending() */
654 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
657 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
659 unsigned long data = atomic_long_read(&work->data);
661 if (data & WORK_STRUCT_PWQ)
662 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
668 * get_work_pool - return the worker_pool a given work was associated with
669 * @work: the work item of interest
671 * Pools are created and destroyed under wq_pool_mutex, and allows read
672 * access under sched-RCU read lock. As such, this function should be
673 * called under wq_pool_mutex or with preemption disabled.
675 * All fields of the returned pool are accessible as long as the above
676 * mentioned locking is in effect. If the returned pool needs to be used
677 * beyond the critical section, the caller is responsible for ensuring the
678 * returned pool is and stays online.
680 * Return: The worker_pool @work was last associated with. %NULL if none.
682 static struct worker_pool *get_work_pool(struct work_struct *work)
684 unsigned long data = atomic_long_read(&work->data);
687 assert_rcu_or_pool_mutex();
689 if (data & WORK_STRUCT_PWQ)
690 return ((struct pool_workqueue *)
691 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
693 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
694 if (pool_id == WORK_OFFQ_POOL_NONE)
697 return idr_find(&worker_pool_idr, pool_id);
701 * get_work_pool_id - return the worker pool ID a given work is associated with
702 * @work: the work item of interest
704 * Return: The worker_pool ID @work was last associated with.
705 * %WORK_OFFQ_POOL_NONE if none.
707 static int get_work_pool_id(struct work_struct *work)
709 unsigned long data = atomic_long_read(&work->data);
711 if (data & WORK_STRUCT_PWQ)
712 return ((struct pool_workqueue *)
713 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
715 return data >> WORK_OFFQ_POOL_SHIFT;
718 static void mark_work_canceling(struct work_struct *work)
720 unsigned long pool_id = get_work_pool_id(work);
722 pool_id <<= WORK_OFFQ_POOL_SHIFT;
723 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
726 static bool work_is_canceling(struct work_struct *work)
728 unsigned long data = atomic_long_read(&work->data);
730 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
734 * Policy functions. These define the policies on how the global worker
735 * pools are managed. Unless noted otherwise, these functions assume that
736 * they're being called with pool->lock held.
739 static bool __need_more_worker(struct worker_pool *pool)
741 return !atomic_read(&pool->nr_running);
745 * Need to wake up a worker? Called from anything but currently
748 * Note that, because unbound workers never contribute to nr_running, this
749 * function will always return %true for unbound pools as long as the
750 * worklist isn't empty.
752 static bool need_more_worker(struct worker_pool *pool)
754 return !list_empty(&pool->worklist) && __need_more_worker(pool);
757 /* Can I start working? Called from busy but !running workers. */
758 static bool may_start_working(struct worker_pool *pool)
760 return pool->nr_idle;
763 /* Do I need to keep working? Called from currently running workers. */
764 static bool keep_working(struct worker_pool *pool)
766 return !list_empty(&pool->worklist) &&
767 atomic_read(&pool->nr_running) <= 1;
770 /* Do we need a new worker? Called from manager. */
771 static bool need_to_create_worker(struct worker_pool *pool)
773 return need_more_worker(pool) && !may_start_working(pool);
776 /* Do we have too many workers and should some go away? */
777 static bool too_many_workers(struct worker_pool *pool)
779 bool managing = mutex_is_locked(&pool->manager_arb);
780 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
781 int nr_busy = pool->nr_workers - nr_idle;
783 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
790 /* Return the first idle worker. Safe with preemption disabled */
791 static struct worker *first_idle_worker(struct worker_pool *pool)
793 if (unlikely(list_empty(&pool->idle_list)))
796 return list_first_entry(&pool->idle_list, struct worker, entry);
800 * wake_up_worker - wake up an idle worker
801 * @pool: worker pool to wake worker from
803 * Wake up the first idle worker of @pool.
806 * spin_lock_irq(pool->lock).
808 static void wake_up_worker(struct worker_pool *pool)
810 struct worker *worker = first_idle_worker(pool);
813 wake_up_process(worker->task);
817 * wq_worker_waking_up - a worker is waking up
818 * @task: task waking up
819 * @cpu: CPU @task is waking up to
821 * This function is called during try_to_wake_up() when a worker is
825 * spin_lock_irq(rq->lock)
827 void wq_worker_waking_up(struct task_struct *task, int cpu)
829 struct worker *worker = kthread_data(task);
831 if (!(worker->flags & WORKER_NOT_RUNNING)) {
832 WARN_ON_ONCE(worker->pool->cpu != cpu);
833 atomic_inc(&worker->pool->nr_running);
838 * wq_worker_sleeping - a worker is going to sleep
839 * @task: task going to sleep
840 * @cpu: CPU in question, must be the current CPU number
842 * This function is called during schedule() when a busy worker is
843 * going to sleep. Worker on the same cpu can be woken up by
844 * returning pointer to its task.
847 * spin_lock_irq(rq->lock)
850 * Worker task on @cpu to wake up, %NULL if none.
852 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
854 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
855 struct worker_pool *pool;
858 * Rescuers, which may not have all the fields set up like normal
859 * workers, also reach here, let's not access anything before
860 * checking NOT_RUNNING.
862 if (worker->flags & WORKER_NOT_RUNNING)
867 /* this can only happen on the local cpu */
868 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
872 * The counterpart of the following dec_and_test, implied mb,
873 * worklist not empty test sequence is in insert_work().
874 * Please read comment there.
876 * NOT_RUNNING is clear. This means that we're bound to and
877 * running on the local cpu w/ rq lock held and preemption
878 * disabled, which in turn means that none else could be
879 * manipulating idle_list, so dereferencing idle_list without pool
882 if (atomic_dec_and_test(&pool->nr_running) &&
883 !list_empty(&pool->worklist))
884 to_wakeup = first_idle_worker(pool);
885 return to_wakeup ? to_wakeup->task : NULL;
889 * worker_set_flags - set worker flags and adjust nr_running accordingly
891 * @flags: flags to set
893 * Set @flags in @worker->flags and adjust nr_running accordingly.
896 * spin_lock_irq(pool->lock)
898 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
900 struct worker_pool *pool = worker->pool;
902 WARN_ON_ONCE(worker->task != current);
904 /* If transitioning into NOT_RUNNING, adjust nr_running. */
905 if ((flags & WORKER_NOT_RUNNING) &&
906 !(worker->flags & WORKER_NOT_RUNNING)) {
907 atomic_dec(&pool->nr_running);
910 worker->flags |= flags;
914 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
916 * @flags: flags to clear
918 * Clear @flags in @worker->flags and adjust nr_running accordingly.
921 * spin_lock_irq(pool->lock)
923 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
925 struct worker_pool *pool = worker->pool;
926 unsigned int oflags = worker->flags;
928 WARN_ON_ONCE(worker->task != current);
930 worker->flags &= ~flags;
933 * If transitioning out of NOT_RUNNING, increment nr_running. Note
934 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
935 * of multiple flags, not a single flag.
937 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
938 if (!(worker->flags & WORKER_NOT_RUNNING))
939 atomic_inc(&pool->nr_running);
943 * find_worker_executing_work - find worker which is executing a work
944 * @pool: pool of interest
945 * @work: work to find worker for
947 * Find a worker which is executing @work on @pool by searching
948 * @pool->busy_hash which is keyed by the address of @work. For a worker
949 * to match, its current execution should match the address of @work and
950 * its work function. This is to avoid unwanted dependency between
951 * unrelated work executions through a work item being recycled while still
954 * This is a bit tricky. A work item may be freed once its execution
955 * starts and nothing prevents the freed area from being recycled for
956 * another work item. If the same work item address ends up being reused
957 * before the original execution finishes, workqueue will identify the
958 * recycled work item as currently executing and make it wait until the
959 * current execution finishes, introducing an unwanted dependency.
961 * This function checks the work item address and work function to avoid
962 * false positives. Note that this isn't complete as one may construct a
963 * work function which can introduce dependency onto itself through a
964 * recycled work item. Well, if somebody wants to shoot oneself in the
965 * foot that badly, there's only so much we can do, and if such deadlock
966 * actually occurs, it should be easy to locate the culprit work function.
969 * spin_lock_irq(pool->lock).
972 * Pointer to worker which is executing @work if found, %NULL
975 static struct worker *find_worker_executing_work(struct worker_pool *pool,
976 struct work_struct *work)
978 struct worker *worker;
980 hash_for_each_possible(pool->busy_hash, worker, hentry,
982 if (worker->current_work == work &&
983 worker->current_func == work->func)
990 * move_linked_works - move linked works to a list
991 * @work: start of series of works to be scheduled
992 * @head: target list to append @work to
993 * @nextp: out paramter for nested worklist walking
995 * Schedule linked works starting from @work to @head. Work series to
996 * be scheduled starts at @work and includes any consecutive work with
997 * WORK_STRUCT_LINKED set in its predecessor.
999 * If @nextp is not NULL, it's updated to point to the next work of
1000 * the last scheduled work. This allows move_linked_works() to be
1001 * nested inside outer list_for_each_entry_safe().
1004 * spin_lock_irq(pool->lock).
1006 static void move_linked_works(struct work_struct *work, struct list_head *head,
1007 struct work_struct **nextp)
1009 struct work_struct *n;
1012 * Linked worklist will always end before the end of the list,
1013 * use NULL for list head.
1015 list_for_each_entry_safe_from(work, n, NULL, entry) {
1016 list_move_tail(&work->entry, head);
1017 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1022 * If we're already inside safe list traversal and have moved
1023 * multiple works to the scheduled queue, the next position
1024 * needs to be updated.
1031 * get_pwq - get an extra reference on the specified pool_workqueue
1032 * @pwq: pool_workqueue to get
1034 * Obtain an extra reference on @pwq. The caller should guarantee that
1035 * @pwq has positive refcnt and be holding the matching pool->lock.
1037 static void get_pwq(struct pool_workqueue *pwq)
1039 lockdep_assert_held(&pwq->pool->lock);
1040 WARN_ON_ONCE(pwq->refcnt <= 0);
1045 * put_pwq - put a pool_workqueue reference
1046 * @pwq: pool_workqueue to put
1048 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1049 * destruction. The caller should be holding the matching pool->lock.
1051 static void put_pwq(struct pool_workqueue *pwq)
1053 lockdep_assert_held(&pwq->pool->lock);
1054 if (likely(--pwq->refcnt))
1056 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1059 * @pwq can't be released under pool->lock, bounce to
1060 * pwq_unbound_release_workfn(). This never recurses on the same
1061 * pool->lock as this path is taken only for unbound workqueues and
1062 * the release work item is scheduled on a per-cpu workqueue. To
1063 * avoid lockdep warning, unbound pool->locks are given lockdep
1064 * subclass of 1 in get_unbound_pool().
1066 schedule_work(&pwq->unbound_release_work);
1070 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1071 * @pwq: pool_workqueue to put (can be %NULL)
1073 * put_pwq() with locking. This function also allows %NULL @pwq.
1075 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1079 * As both pwqs and pools are sched-RCU protected, the
1080 * following lock operations are safe.
1082 spin_lock_irq(&pwq->pool->lock);
1084 spin_unlock_irq(&pwq->pool->lock);
1088 static void pwq_activate_delayed_work(struct work_struct *work)
1090 struct pool_workqueue *pwq = get_work_pwq(work);
1092 trace_workqueue_activate_work(work);
1093 move_linked_works(work, &pwq->pool->worklist, NULL);
1094 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1098 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1100 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1101 struct work_struct, entry);
1103 pwq_activate_delayed_work(work);
1107 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1108 * @pwq: pwq of interest
1109 * @color: color of work which left the queue
1111 * A work either has completed or is removed from pending queue,
1112 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1115 * spin_lock_irq(pool->lock).
1117 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1119 /* uncolored work items don't participate in flushing or nr_active */
1120 if (color == WORK_NO_COLOR)
1123 pwq->nr_in_flight[color]--;
1126 if (!list_empty(&pwq->delayed_works)) {
1127 /* one down, submit a delayed one */
1128 if (pwq->nr_active < pwq->max_active)
1129 pwq_activate_first_delayed(pwq);
1132 /* is flush in progress and are we at the flushing tip? */
1133 if (likely(pwq->flush_color != color))
1136 /* are there still in-flight works? */
1137 if (pwq->nr_in_flight[color])
1140 /* this pwq is done, clear flush_color */
1141 pwq->flush_color = -1;
1144 * If this was the last pwq, wake up the first flusher. It
1145 * will handle the rest.
1147 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1148 complete(&pwq->wq->first_flusher->done);
1154 * try_to_grab_pending - steal work item from worklist and disable irq
1155 * @work: work item to steal
1156 * @is_dwork: @work is a delayed_work
1157 * @flags: place to store irq state
1159 * Try to grab PENDING bit of @work. This function can handle @work in any
1160 * stable state - idle, on timer or on worklist.
1163 * 1 if @work was pending and we successfully stole PENDING
1164 * 0 if @work was idle and we claimed PENDING
1165 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1166 * -ENOENT if someone else is canceling @work, this state may persist
1167 * for arbitrarily long
1170 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1171 * interrupted while holding PENDING and @work off queue, irq must be
1172 * disabled on entry. This, combined with delayed_work->timer being
1173 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1175 * On successful return, >= 0, irq is disabled and the caller is
1176 * responsible for releasing it using local_irq_restore(*@flags).
1178 * This function is safe to call from any context including IRQ handler.
1180 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1181 unsigned long *flags)
1183 struct worker_pool *pool;
1184 struct pool_workqueue *pwq;
1186 local_irq_save(*flags);
1188 /* try to steal the timer if it exists */
1190 struct delayed_work *dwork = to_delayed_work(work);
1193 * dwork->timer is irqsafe. If del_timer() fails, it's
1194 * guaranteed that the timer is not queued anywhere and not
1195 * running on the local CPU.
1197 if (likely(del_timer(&dwork->timer)))
1201 /* try to claim PENDING the normal way */
1202 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1206 * The queueing is in progress, or it is already queued. Try to
1207 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1209 pool = get_work_pool(work);
1213 spin_lock(&pool->lock);
1215 * work->data is guaranteed to point to pwq only while the work
1216 * item is queued on pwq->wq, and both updating work->data to point
1217 * to pwq on queueing and to pool on dequeueing are done under
1218 * pwq->pool->lock. This in turn guarantees that, if work->data
1219 * points to pwq which is associated with a locked pool, the work
1220 * item is currently queued on that pool.
1222 pwq = get_work_pwq(work);
1223 if (pwq && pwq->pool == pool) {
1224 debug_work_deactivate(work);
1227 * A delayed work item cannot be grabbed directly because
1228 * it might have linked NO_COLOR work items which, if left
1229 * on the delayed_list, will confuse pwq->nr_active
1230 * management later on and cause stall. Make sure the work
1231 * item is activated before grabbing.
1233 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1234 pwq_activate_delayed_work(work);
1236 list_del_init(&work->entry);
1237 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1239 /* work->data points to pwq iff queued, point to pool */
1240 set_work_pool_and_keep_pending(work, pool->id);
1242 spin_unlock(&pool->lock);
1245 spin_unlock(&pool->lock);
1247 local_irq_restore(*flags);
1248 if (work_is_canceling(work))
1255 * insert_work - insert a work into a pool
1256 * @pwq: pwq @work belongs to
1257 * @work: work to insert
1258 * @head: insertion point
1259 * @extra_flags: extra WORK_STRUCT_* flags to set
1261 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1262 * work_struct flags.
1265 * spin_lock_irq(pool->lock).
1267 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1268 struct list_head *head, unsigned int extra_flags)
1270 struct worker_pool *pool = pwq->pool;
1272 /* we own @work, set data and link */
1273 set_work_pwq(work, pwq, extra_flags);
1274 list_add_tail(&work->entry, head);
1278 * Ensure either wq_worker_sleeping() sees the above
1279 * list_add_tail() or we see zero nr_running to avoid workers lying
1280 * around lazily while there are works to be processed.
1284 if (__need_more_worker(pool))
1285 wake_up_worker(pool);
1289 * Test whether @work is being queued from another work executing on the
1292 static bool is_chained_work(struct workqueue_struct *wq)
1294 struct worker *worker;
1296 worker = current_wq_worker();
1298 * Return %true iff I'm a worker execuing a work item on @wq. If
1299 * I'm @worker, it's safe to dereference it without locking.
1301 return worker && worker->current_pwq->wq == wq;
1304 static void __queue_work(int cpu, struct workqueue_struct *wq,
1305 struct work_struct *work)
1307 struct pool_workqueue *pwq;
1308 struct worker_pool *last_pool;
1309 struct list_head *worklist;
1310 unsigned int work_flags;
1311 unsigned int req_cpu = cpu;
1314 * While a work item is PENDING && off queue, a task trying to
1315 * steal the PENDING will busy-loop waiting for it to either get
1316 * queued or lose PENDING. Grabbing PENDING and queueing should
1317 * happen with IRQ disabled.
1319 WARN_ON_ONCE(!irqs_disabled());
1321 debug_work_activate(work);
1323 /* if draining, only works from the same workqueue are allowed */
1324 if (unlikely(wq->flags & __WQ_DRAINING) &&
1325 WARN_ON_ONCE(!is_chained_work(wq)))
1328 if (req_cpu == WORK_CPU_UNBOUND)
1329 cpu = raw_smp_processor_id();
1331 /* pwq which will be used unless @work is executing elsewhere */
1332 if (!(wq->flags & WQ_UNBOUND))
1333 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1335 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1338 * If @work was previously on a different pool, it might still be
1339 * running there, in which case the work needs to be queued on that
1340 * pool to guarantee non-reentrancy.
1342 last_pool = get_work_pool(work);
1343 if (last_pool && last_pool != pwq->pool) {
1344 struct worker *worker;
1346 spin_lock(&last_pool->lock);
1348 worker = find_worker_executing_work(last_pool, work);
1350 if (worker && worker->current_pwq->wq == wq) {
1351 pwq = worker->current_pwq;
1353 /* meh... not running there, queue here */
1354 spin_unlock(&last_pool->lock);
1355 spin_lock(&pwq->pool->lock);
1358 spin_lock(&pwq->pool->lock);
1362 * pwq is determined and locked. For unbound pools, we could have
1363 * raced with pwq release and it could already be dead. If its
1364 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1365 * without another pwq replacing it in the numa_pwq_tbl or while
1366 * work items are executing on it, so the retrying is guaranteed to
1367 * make forward-progress.
1369 if (unlikely(!pwq->refcnt)) {
1370 if (wq->flags & WQ_UNBOUND) {
1371 spin_unlock(&pwq->pool->lock);
1376 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1380 /* pwq determined, queue */
1381 trace_workqueue_queue_work(req_cpu, pwq, work);
1383 if (WARN_ON(!list_empty(&work->entry))) {
1384 spin_unlock(&pwq->pool->lock);
1388 pwq->nr_in_flight[pwq->work_color]++;
1389 work_flags = work_color_to_flags(pwq->work_color);
1391 if (likely(pwq->nr_active < pwq->max_active)) {
1392 trace_workqueue_activate_work(work);
1394 worklist = &pwq->pool->worklist;
1396 work_flags |= WORK_STRUCT_DELAYED;
1397 worklist = &pwq->delayed_works;
1400 insert_work(pwq, work, worklist, work_flags);
1402 spin_unlock(&pwq->pool->lock);
1406 * queue_work_on - queue work on specific cpu
1407 * @cpu: CPU number to execute work on
1408 * @wq: workqueue to use
1409 * @work: work to queue
1411 * We queue the work to a specific CPU, the caller must ensure it
1414 * Return: %false if @work was already on a queue, %true otherwise.
1416 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1417 struct work_struct *work)
1420 unsigned long flags;
1422 local_irq_save(flags);
1424 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1425 __queue_work(cpu, wq, work);
1429 local_irq_restore(flags);
1432 EXPORT_SYMBOL(queue_work_on);
1434 void delayed_work_timer_fn(unsigned long __data)
1436 struct delayed_work *dwork = (struct delayed_work *)__data;
1438 /* should have been called from irqsafe timer with irq already off */
1439 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1441 EXPORT_SYMBOL(delayed_work_timer_fn);
1443 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1444 struct delayed_work *dwork, unsigned long delay)
1446 struct timer_list *timer = &dwork->timer;
1447 struct work_struct *work = &dwork->work;
1449 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1450 timer->data != (unsigned long)dwork);
1451 WARN_ON_ONCE(timer_pending(timer));
1452 WARN_ON_ONCE(!list_empty(&work->entry));
1455 * If @delay is 0, queue @dwork->work immediately. This is for
1456 * both optimization and correctness. The earliest @timer can
1457 * expire is on the closest next tick and delayed_work users depend
1458 * on that there's no such delay when @delay is 0.
1461 __queue_work(cpu, wq, &dwork->work);
1465 timer_stats_timer_set_start_info(&dwork->timer);
1469 timer->expires = jiffies + delay;
1471 if (unlikely(cpu != WORK_CPU_UNBOUND))
1472 add_timer_on(timer, cpu);
1478 * queue_delayed_work_on - queue work on specific CPU after delay
1479 * @cpu: CPU number to execute work on
1480 * @wq: workqueue to use
1481 * @dwork: work to queue
1482 * @delay: number of jiffies to wait before queueing
1484 * Return: %false if @work was already on a queue, %true otherwise. If
1485 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1488 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1489 struct delayed_work *dwork, unsigned long delay)
1491 struct work_struct *work = &dwork->work;
1493 unsigned long flags;
1495 /* read the comment in __queue_work() */
1496 local_irq_save(flags);
1498 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1499 __queue_delayed_work(cpu, wq, dwork, delay);
1503 local_irq_restore(flags);
1506 EXPORT_SYMBOL(queue_delayed_work_on);
1509 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1510 * @cpu: CPU number to execute work on
1511 * @wq: workqueue to use
1512 * @dwork: work to queue
1513 * @delay: number of jiffies to wait before queueing
1515 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1516 * modify @dwork's timer so that it expires after @delay. If @delay is
1517 * zero, @work is guaranteed to be scheduled immediately regardless of its
1520 * Return: %false if @dwork was idle and queued, %true if @dwork was
1521 * pending and its timer was modified.
1523 * This function is safe to call from any context including IRQ handler.
1524 * See try_to_grab_pending() for details.
1526 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1527 struct delayed_work *dwork, unsigned long delay)
1529 unsigned long flags;
1533 ret = try_to_grab_pending(&dwork->work, true, &flags);
1534 } while (unlikely(ret == -EAGAIN));
1536 if (likely(ret >= 0)) {
1537 __queue_delayed_work(cpu, wq, dwork, delay);
1538 local_irq_restore(flags);
1541 /* -ENOENT from try_to_grab_pending() becomes %true */
1544 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1547 * worker_enter_idle - enter idle state
1548 * @worker: worker which is entering idle state
1550 * @worker is entering idle state. Update stats and idle timer if
1554 * spin_lock_irq(pool->lock).
1556 static void worker_enter_idle(struct worker *worker)
1558 struct worker_pool *pool = worker->pool;
1560 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1561 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1562 (worker->hentry.next || worker->hentry.pprev)))
1565 /* can't use worker_set_flags(), also called from create_worker() */
1566 worker->flags |= WORKER_IDLE;
1568 worker->last_active = jiffies;
1570 /* idle_list is LIFO */
1571 list_add(&worker->entry, &pool->idle_list);
1573 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1574 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1577 * Sanity check nr_running. Because wq_unbind_fn() releases
1578 * pool->lock between setting %WORKER_UNBOUND and zapping
1579 * nr_running, the warning may trigger spuriously. Check iff
1580 * unbind is not in progress.
1582 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1583 pool->nr_workers == pool->nr_idle &&
1584 atomic_read(&pool->nr_running));
1588 * worker_leave_idle - leave idle state
1589 * @worker: worker which is leaving idle state
1591 * @worker is leaving idle state. Update stats.
1594 * spin_lock_irq(pool->lock).
1596 static void worker_leave_idle(struct worker *worker)
1598 struct worker_pool *pool = worker->pool;
1600 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1602 worker_clr_flags(worker, WORKER_IDLE);
1604 list_del_init(&worker->entry);
1607 static struct worker *alloc_worker(int node)
1609 struct worker *worker;
1611 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1613 INIT_LIST_HEAD(&worker->entry);
1614 INIT_LIST_HEAD(&worker->scheduled);
1615 INIT_LIST_HEAD(&worker->node);
1616 /* on creation a worker is in !idle && prep state */
1617 worker->flags = WORKER_PREP;
1623 * worker_attach_to_pool() - attach a worker to a pool
1624 * @worker: worker to be attached
1625 * @pool: the target pool
1627 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1628 * cpu-binding of @worker are kept coordinated with the pool across
1631 static void worker_attach_to_pool(struct worker *worker,
1632 struct worker_pool *pool)
1634 mutex_lock(&pool->attach_mutex);
1637 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1638 * online CPUs. It'll be re-applied when any of the CPUs come up.
1640 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1643 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1644 * stable across this function. See the comments above the
1645 * flag definition for details.
1647 if (pool->flags & POOL_DISASSOCIATED)
1648 worker->flags |= WORKER_UNBOUND;
1650 list_add_tail(&worker->node, &pool->workers);
1652 mutex_unlock(&pool->attach_mutex);
1656 * worker_detach_from_pool() - detach a worker from its pool
1657 * @worker: worker which is attached to its pool
1658 * @pool: the pool @worker is attached to
1660 * Undo the attaching which had been done in worker_attach_to_pool(). The
1661 * caller worker shouldn't access to the pool after detached except it has
1662 * other reference to the pool.
1664 static void worker_detach_from_pool(struct worker *worker,
1665 struct worker_pool *pool)
1667 struct completion *detach_completion = NULL;
1669 mutex_lock(&pool->attach_mutex);
1670 list_del(&worker->node);
1671 if (list_empty(&pool->workers))
1672 detach_completion = pool->detach_completion;
1673 mutex_unlock(&pool->attach_mutex);
1675 /* clear leftover flags without pool->lock after it is detached */
1676 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1678 if (detach_completion)
1679 complete(detach_completion);
1683 * create_worker - create a new workqueue worker
1684 * @pool: pool the new worker will belong to
1686 * Create and start a new worker which is attached to @pool.
1689 * Might sleep. Does GFP_KERNEL allocations.
1692 * Pointer to the newly created worker.
1694 static struct worker *create_worker(struct worker_pool *pool)
1696 struct worker *worker = NULL;
1700 /* ID is needed to determine kthread name */
1701 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1705 worker = alloc_worker(pool->node);
1709 worker->pool = pool;
1713 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1714 pool->attrs->nice < 0 ? "H" : "");
1716 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1718 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1719 "kworker/%s", id_buf);
1720 if (IS_ERR(worker->task))
1723 set_user_nice(worker->task, pool->attrs->nice);
1725 /* prevent userland from meddling with cpumask of workqueue workers */
1726 worker->task->flags |= PF_NO_SETAFFINITY;
1728 /* successful, attach the worker to the pool */
1729 worker_attach_to_pool(worker, pool);
1731 /* start the newly created worker */
1732 spin_lock_irq(&pool->lock);
1733 worker->pool->nr_workers++;
1734 worker_enter_idle(worker);
1735 wake_up_process(worker->task);
1736 spin_unlock_irq(&pool->lock);
1742 ida_simple_remove(&pool->worker_ida, id);
1748 * destroy_worker - destroy a workqueue worker
1749 * @worker: worker to be destroyed
1751 * Destroy @worker and adjust @pool stats accordingly. The worker should
1755 * spin_lock_irq(pool->lock).
1757 static void destroy_worker(struct worker *worker)
1759 struct worker_pool *pool = worker->pool;
1761 lockdep_assert_held(&pool->lock);
1763 /* sanity check frenzy */
1764 if (WARN_ON(worker->current_work) ||
1765 WARN_ON(!list_empty(&worker->scheduled)) ||
1766 WARN_ON(!(worker->flags & WORKER_IDLE)))
1772 list_del_init(&worker->entry);
1773 worker->flags |= WORKER_DIE;
1774 wake_up_process(worker->task);
1777 static void idle_worker_timeout(unsigned long __pool)
1779 struct worker_pool *pool = (void *)__pool;
1781 spin_lock_irq(&pool->lock);
1783 while (too_many_workers(pool)) {
1784 struct worker *worker;
1785 unsigned long expires;
1787 /* idle_list is kept in LIFO order, check the last one */
1788 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1789 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1791 if (time_before(jiffies, expires)) {
1792 mod_timer(&pool->idle_timer, expires);
1796 destroy_worker(worker);
1799 spin_unlock_irq(&pool->lock);
1802 static void send_mayday(struct work_struct *work)
1804 struct pool_workqueue *pwq = get_work_pwq(work);
1805 struct workqueue_struct *wq = pwq->wq;
1807 lockdep_assert_held(&wq_mayday_lock);
1812 /* mayday mayday mayday */
1813 if (list_empty(&pwq->mayday_node)) {
1815 * If @pwq is for an unbound wq, its base ref may be put at
1816 * any time due to an attribute change. Pin @pwq until the
1817 * rescuer is done with it.
1820 list_add_tail(&pwq->mayday_node, &wq->maydays);
1821 wake_up_process(wq->rescuer->task);
1825 static void pool_mayday_timeout(unsigned long __pool)
1827 struct worker_pool *pool = (void *)__pool;
1828 struct work_struct *work;
1830 spin_lock_irq(&pool->lock);
1831 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1833 if (need_to_create_worker(pool)) {
1835 * We've been trying to create a new worker but
1836 * haven't been successful. We might be hitting an
1837 * allocation deadlock. Send distress signals to
1840 list_for_each_entry(work, &pool->worklist, entry)
1844 spin_unlock(&wq_mayday_lock);
1845 spin_unlock_irq(&pool->lock);
1847 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1851 * maybe_create_worker - create a new worker if necessary
1852 * @pool: pool to create a new worker for
1854 * Create a new worker for @pool if necessary. @pool is guaranteed to
1855 * have at least one idle worker on return from this function. If
1856 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1857 * sent to all rescuers with works scheduled on @pool to resolve
1858 * possible allocation deadlock.
1860 * On return, need_to_create_worker() is guaranteed to be %false and
1861 * may_start_working() %true.
1864 * spin_lock_irq(pool->lock) which may be released and regrabbed
1865 * multiple times. Does GFP_KERNEL allocations. Called only from
1868 static void maybe_create_worker(struct worker_pool *pool)
1869 __releases(&pool->lock)
1870 __acquires(&pool->lock)
1873 spin_unlock_irq(&pool->lock);
1875 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1876 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1879 if (create_worker(pool) || !need_to_create_worker(pool))
1882 schedule_timeout_interruptible(CREATE_COOLDOWN);
1884 if (!need_to_create_worker(pool))
1888 del_timer_sync(&pool->mayday_timer);
1889 spin_lock_irq(&pool->lock);
1891 * This is necessary even after a new worker was just successfully
1892 * created as @pool->lock was dropped and the new worker might have
1893 * already become busy.
1895 if (need_to_create_worker(pool))
1900 * manage_workers - manage worker pool
1903 * Assume the manager role and manage the worker pool @worker belongs
1904 * to. At any given time, there can be only zero or one manager per
1905 * pool. The exclusion is handled automatically by this function.
1907 * The caller can safely start processing works on false return. On
1908 * true return, it's guaranteed that need_to_create_worker() is false
1909 * and may_start_working() is true.
1912 * spin_lock_irq(pool->lock) which may be released and regrabbed
1913 * multiple times. Does GFP_KERNEL allocations.
1916 * %false if the pool doesn't need management and the caller can safely
1917 * start processing works, %true if management function was performed and
1918 * the conditions that the caller verified before calling the function may
1919 * no longer be true.
1921 static bool manage_workers(struct worker *worker)
1923 struct worker_pool *pool = worker->pool;
1926 * Anyone who successfully grabs manager_arb wins the arbitration
1927 * and becomes the manager. mutex_trylock() on pool->manager_arb
1928 * failure while holding pool->lock reliably indicates that someone
1929 * else is managing the pool and the worker which failed trylock
1930 * can proceed to executing work items. This means that anyone
1931 * grabbing manager_arb is responsible for actually performing
1932 * manager duties. If manager_arb is grabbed and released without
1933 * actual management, the pool may stall indefinitely.
1935 if (!mutex_trylock(&pool->manager_arb))
1937 pool->manager = worker;
1939 maybe_create_worker(pool);
1941 pool->manager = NULL;
1942 mutex_unlock(&pool->manager_arb);
1947 * process_one_work - process single work
1949 * @work: work to process
1951 * Process @work. This function contains all the logics necessary to
1952 * process a single work including synchronization against and
1953 * interaction with other workers on the same cpu, queueing and
1954 * flushing. As long as context requirement is met, any worker can
1955 * call this function to process a work.
1958 * spin_lock_irq(pool->lock) which is released and regrabbed.
1960 static void process_one_work(struct worker *worker, struct work_struct *work)
1961 __releases(&pool->lock)
1962 __acquires(&pool->lock)
1964 struct pool_workqueue *pwq = get_work_pwq(work);
1965 struct worker_pool *pool = worker->pool;
1966 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1968 struct worker *collision;
1969 #ifdef CONFIG_LOCKDEP
1971 * It is permissible to free the struct work_struct from
1972 * inside the function that is called from it, this we need to
1973 * take into account for lockdep too. To avoid bogus "held
1974 * lock freed" warnings as well as problems when looking into
1975 * work->lockdep_map, make a copy and use that here.
1977 struct lockdep_map lockdep_map;
1979 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1981 /* ensure we're on the correct CPU */
1982 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1983 raw_smp_processor_id() != pool->cpu);
1986 * A single work shouldn't be executed concurrently by
1987 * multiple workers on a single cpu. Check whether anyone is
1988 * already processing the work. If so, defer the work to the
1989 * currently executing one.
1991 collision = find_worker_executing_work(pool, work);
1992 if (unlikely(collision)) {
1993 move_linked_works(work, &collision->scheduled, NULL);
1997 /* claim and dequeue */
1998 debug_work_deactivate(work);
1999 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2000 worker->current_work = work;
2001 worker->current_func = work->func;
2002 worker->current_pwq = pwq;
2003 work_color = get_work_color(work);
2005 list_del_init(&work->entry);
2008 * CPU intensive works don't participate in concurrency management.
2009 * They're the scheduler's responsibility. This takes @worker out
2010 * of concurrency management and the next code block will chain
2011 * execution of the pending work items.
2013 if (unlikely(cpu_intensive))
2014 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2017 * Wake up another worker if necessary. The condition is always
2018 * false for normal per-cpu workers since nr_running would always
2019 * be >= 1 at this point. This is used to chain execution of the
2020 * pending work items for WORKER_NOT_RUNNING workers such as the
2021 * UNBOUND and CPU_INTENSIVE ones.
2023 if (need_more_worker(pool))
2024 wake_up_worker(pool);
2027 * Record the last pool and clear PENDING which should be the last
2028 * update to @work. Also, do this inside @pool->lock so that
2029 * PENDING and queued state changes happen together while IRQ is
2032 set_work_pool_and_clear_pending(work, pool->id);
2034 spin_unlock_irq(&pool->lock);
2036 lock_map_acquire_read(&pwq->wq->lockdep_map);
2037 lock_map_acquire(&lockdep_map);
2038 trace_workqueue_execute_start(work);
2039 worker->current_func(work);
2041 * While we must be careful to not use "work" after this, the trace
2042 * point will only record its address.
2044 trace_workqueue_execute_end(work);
2045 lock_map_release(&lockdep_map);
2046 lock_map_release(&pwq->wq->lockdep_map);
2048 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2049 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2050 " last function: %pf\n",
2051 current->comm, preempt_count(), task_pid_nr(current),
2052 worker->current_func);
2053 debug_show_held_locks(current);
2058 * The following prevents a kworker from hogging CPU on !PREEMPT
2059 * kernels, where a requeueing work item waiting for something to
2060 * happen could deadlock with stop_machine as such work item could
2061 * indefinitely requeue itself while all other CPUs are trapped in
2062 * stop_machine. At the same time, report a quiescent RCU state so
2063 * the same condition doesn't freeze RCU.
2065 cond_resched_rcu_qs();
2067 spin_lock_irq(&pool->lock);
2069 /* clear cpu intensive status */
2070 if (unlikely(cpu_intensive))
2071 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2073 /* we're done with it, release */
2074 hash_del(&worker->hentry);
2075 worker->current_work = NULL;
2076 worker->current_func = NULL;
2077 worker->current_pwq = NULL;
2078 worker->desc_valid = false;
2079 pwq_dec_nr_in_flight(pwq, work_color);
2083 * process_scheduled_works - process scheduled works
2086 * Process all scheduled works. Please note that the scheduled list
2087 * may change while processing a work, so this function repeatedly
2088 * fetches a work from the top and executes it.
2091 * spin_lock_irq(pool->lock) which may be released and regrabbed
2094 static void process_scheduled_works(struct worker *worker)
2096 while (!list_empty(&worker->scheduled)) {
2097 struct work_struct *work = list_first_entry(&worker->scheduled,
2098 struct work_struct, entry);
2099 process_one_work(worker, work);
2104 * worker_thread - the worker thread function
2107 * The worker thread function. All workers belong to a worker_pool -
2108 * either a per-cpu one or dynamic unbound one. These workers process all
2109 * work items regardless of their specific target workqueue. The only
2110 * exception is work items which belong to workqueues with a rescuer which
2111 * will be explained in rescuer_thread().
2115 static int worker_thread(void *__worker)
2117 struct worker *worker = __worker;
2118 struct worker_pool *pool = worker->pool;
2120 /* tell the scheduler that this is a workqueue worker */
2121 worker->task->flags |= PF_WQ_WORKER;
2123 spin_lock_irq(&pool->lock);
2125 /* am I supposed to die? */
2126 if (unlikely(worker->flags & WORKER_DIE)) {
2127 spin_unlock_irq(&pool->lock);
2128 WARN_ON_ONCE(!list_empty(&worker->entry));
2129 worker->task->flags &= ~PF_WQ_WORKER;
2131 set_task_comm(worker->task, "kworker/dying");
2132 ida_simple_remove(&pool->worker_ida, worker->id);
2133 worker_detach_from_pool(worker, pool);
2138 worker_leave_idle(worker);
2140 /* no more worker necessary? */
2141 if (!need_more_worker(pool))
2144 /* do we need to manage? */
2145 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2149 * ->scheduled list can only be filled while a worker is
2150 * preparing to process a work or actually processing it.
2151 * Make sure nobody diddled with it while I was sleeping.
2153 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2156 * Finish PREP stage. We're guaranteed to have at least one idle
2157 * worker or that someone else has already assumed the manager
2158 * role. This is where @worker starts participating in concurrency
2159 * management if applicable and concurrency management is restored
2160 * after being rebound. See rebind_workers() for details.
2162 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2165 struct work_struct *work =
2166 list_first_entry(&pool->worklist,
2167 struct work_struct, entry);
2169 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2170 /* optimization path, not strictly necessary */
2171 process_one_work(worker, work);
2172 if (unlikely(!list_empty(&worker->scheduled)))
2173 process_scheduled_works(worker);
2175 move_linked_works(work, &worker->scheduled, NULL);
2176 process_scheduled_works(worker);
2178 } while (keep_working(pool));
2180 worker_set_flags(worker, WORKER_PREP);
2183 * pool->lock is held and there's no work to process and no need to
2184 * manage, sleep. Workers are woken up only while holding
2185 * pool->lock or from local cpu, so setting the current state
2186 * before releasing pool->lock is enough to prevent losing any
2189 worker_enter_idle(worker);
2190 __set_current_state(TASK_INTERRUPTIBLE);
2191 spin_unlock_irq(&pool->lock);
2197 * rescuer_thread - the rescuer thread function
2200 * Workqueue rescuer thread function. There's one rescuer for each
2201 * workqueue which has WQ_MEM_RECLAIM set.
2203 * Regular work processing on a pool may block trying to create a new
2204 * worker which uses GFP_KERNEL allocation which has slight chance of
2205 * developing into deadlock if some works currently on the same queue
2206 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2207 * the problem rescuer solves.
2209 * When such condition is possible, the pool summons rescuers of all
2210 * workqueues which have works queued on the pool and let them process
2211 * those works so that forward progress can be guaranteed.
2213 * This should happen rarely.
2217 static int rescuer_thread(void *__rescuer)
2219 struct worker *rescuer = __rescuer;
2220 struct workqueue_struct *wq = rescuer->rescue_wq;
2221 struct list_head *scheduled = &rescuer->scheduled;
2224 set_user_nice(current, RESCUER_NICE_LEVEL);
2227 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2228 * doesn't participate in concurrency management.
2230 rescuer->task->flags |= PF_WQ_WORKER;
2232 set_current_state(TASK_INTERRUPTIBLE);
2235 * By the time the rescuer is requested to stop, the workqueue
2236 * shouldn't have any work pending, but @wq->maydays may still have
2237 * pwq(s) queued. This can happen by non-rescuer workers consuming
2238 * all the work items before the rescuer got to them. Go through
2239 * @wq->maydays processing before acting on should_stop so that the
2240 * list is always empty on exit.
2242 should_stop = kthread_should_stop();
2244 /* see whether any pwq is asking for help */
2245 spin_lock_irq(&wq_mayday_lock);
2247 while (!list_empty(&wq->maydays)) {
2248 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2249 struct pool_workqueue, mayday_node);
2250 struct worker_pool *pool = pwq->pool;
2251 struct work_struct *work, *n;
2253 __set_current_state(TASK_RUNNING);
2254 list_del_init(&pwq->mayday_node);
2256 spin_unlock_irq(&wq_mayday_lock);
2258 worker_attach_to_pool(rescuer, pool);
2260 spin_lock_irq(&pool->lock);
2261 rescuer->pool = pool;
2264 * Slurp in all works issued via this workqueue and
2267 WARN_ON_ONCE(!list_empty(scheduled));
2268 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2269 if (get_work_pwq(work) == pwq)
2270 move_linked_works(work, scheduled, &n);
2272 if (!list_empty(scheduled)) {
2273 process_scheduled_works(rescuer);
2276 * The above execution of rescued work items could
2277 * have created more to rescue through
2278 * pwq_activate_first_delayed() or chained
2279 * queueing. Let's put @pwq back on mayday list so
2280 * that such back-to-back work items, which may be
2281 * being used to relieve memory pressure, don't
2282 * incur MAYDAY_INTERVAL delay inbetween.
2284 if (need_to_create_worker(pool)) {
2285 spin_lock(&wq_mayday_lock);
2287 list_move_tail(&pwq->mayday_node, &wq->maydays);
2288 spin_unlock(&wq_mayday_lock);
2293 * Put the reference grabbed by send_mayday(). @pool won't
2294 * go away while we're still attached to it.
2299 * Leave this pool. If need_more_worker() is %true, notify a
2300 * regular worker; otherwise, we end up with 0 concurrency
2301 * and stalling the execution.
2303 if (need_more_worker(pool))
2304 wake_up_worker(pool);
2306 rescuer->pool = NULL;
2307 spin_unlock_irq(&pool->lock);
2309 worker_detach_from_pool(rescuer, pool);
2311 spin_lock_irq(&wq_mayday_lock);
2314 spin_unlock_irq(&wq_mayday_lock);
2317 __set_current_state(TASK_RUNNING);
2318 rescuer->task->flags &= ~PF_WQ_WORKER;
2322 /* rescuers should never participate in concurrency management */
2323 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2329 struct work_struct work;
2330 struct completion done;
2331 struct task_struct *task; /* purely informational */
2334 static void wq_barrier_func(struct work_struct *work)
2336 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2337 complete(&barr->done);
2341 * insert_wq_barrier - insert a barrier work
2342 * @pwq: pwq to insert barrier into
2343 * @barr: wq_barrier to insert
2344 * @target: target work to attach @barr to
2345 * @worker: worker currently executing @target, NULL if @target is not executing
2347 * @barr is linked to @target such that @barr is completed only after
2348 * @target finishes execution. Please note that the ordering
2349 * guarantee is observed only with respect to @target and on the local
2352 * Currently, a queued barrier can't be canceled. This is because
2353 * try_to_grab_pending() can't determine whether the work to be
2354 * grabbed is at the head of the queue and thus can't clear LINKED
2355 * flag of the previous work while there must be a valid next work
2356 * after a work with LINKED flag set.
2358 * Note that when @worker is non-NULL, @target may be modified
2359 * underneath us, so we can't reliably determine pwq from @target.
2362 * spin_lock_irq(pool->lock).
2364 static void insert_wq_barrier(struct pool_workqueue *pwq,
2365 struct wq_barrier *barr,
2366 struct work_struct *target, struct worker *worker)
2368 struct list_head *head;
2369 unsigned int linked = 0;
2372 * debugobject calls are safe here even with pool->lock locked
2373 * as we know for sure that this will not trigger any of the
2374 * checks and call back into the fixup functions where we
2377 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2378 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2379 init_completion(&barr->done);
2380 barr->task = current;
2383 * If @target is currently being executed, schedule the
2384 * barrier to the worker; otherwise, put it after @target.
2387 head = worker->scheduled.next;
2389 unsigned long *bits = work_data_bits(target);
2391 head = target->entry.next;
2392 /* there can already be other linked works, inherit and set */
2393 linked = *bits & WORK_STRUCT_LINKED;
2394 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2397 debug_work_activate(&barr->work);
2398 insert_work(pwq, &barr->work, head,
2399 work_color_to_flags(WORK_NO_COLOR) | linked);
2403 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2404 * @wq: workqueue being flushed
2405 * @flush_color: new flush color, < 0 for no-op
2406 * @work_color: new work color, < 0 for no-op
2408 * Prepare pwqs for workqueue flushing.
2410 * If @flush_color is non-negative, flush_color on all pwqs should be
2411 * -1. If no pwq has in-flight commands at the specified color, all
2412 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2413 * has in flight commands, its pwq->flush_color is set to
2414 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2415 * wakeup logic is armed and %true is returned.
2417 * The caller should have initialized @wq->first_flusher prior to
2418 * calling this function with non-negative @flush_color. If
2419 * @flush_color is negative, no flush color update is done and %false
2422 * If @work_color is non-negative, all pwqs should have the same
2423 * work_color which is previous to @work_color and all will be
2424 * advanced to @work_color.
2427 * mutex_lock(wq->mutex).
2430 * %true if @flush_color >= 0 and there's something to flush. %false
2433 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2434 int flush_color, int work_color)
2437 struct pool_workqueue *pwq;
2439 if (flush_color >= 0) {
2440 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2441 atomic_set(&wq->nr_pwqs_to_flush, 1);
2444 for_each_pwq(pwq, wq) {
2445 struct worker_pool *pool = pwq->pool;
2447 spin_lock_irq(&pool->lock);
2449 if (flush_color >= 0) {
2450 WARN_ON_ONCE(pwq->flush_color != -1);
2452 if (pwq->nr_in_flight[flush_color]) {
2453 pwq->flush_color = flush_color;
2454 atomic_inc(&wq->nr_pwqs_to_flush);
2459 if (work_color >= 0) {
2460 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2461 pwq->work_color = work_color;
2464 spin_unlock_irq(&pool->lock);
2467 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2468 complete(&wq->first_flusher->done);
2474 * flush_workqueue - ensure that any scheduled work has run to completion.
2475 * @wq: workqueue to flush
2477 * This function sleeps until all work items which were queued on entry
2478 * have finished execution, but it is not livelocked by new incoming ones.
2480 void flush_workqueue(struct workqueue_struct *wq)
2482 struct wq_flusher this_flusher = {
2483 .list = LIST_HEAD_INIT(this_flusher.list),
2485 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2489 lock_map_acquire(&wq->lockdep_map);
2490 lock_map_release(&wq->lockdep_map);
2492 mutex_lock(&wq->mutex);
2495 * Start-to-wait phase
2497 next_color = work_next_color(wq->work_color);
2499 if (next_color != wq->flush_color) {
2501 * Color space is not full. The current work_color
2502 * becomes our flush_color and work_color is advanced
2505 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2506 this_flusher.flush_color = wq->work_color;
2507 wq->work_color = next_color;
2509 if (!wq->first_flusher) {
2510 /* no flush in progress, become the first flusher */
2511 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2513 wq->first_flusher = &this_flusher;
2515 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2517 /* nothing to flush, done */
2518 wq->flush_color = next_color;
2519 wq->first_flusher = NULL;
2524 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2525 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2526 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2530 * Oops, color space is full, wait on overflow queue.
2531 * The next flush completion will assign us
2532 * flush_color and transfer to flusher_queue.
2534 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2537 mutex_unlock(&wq->mutex);
2539 wait_for_completion(&this_flusher.done);
2542 * Wake-up-and-cascade phase
2544 * First flushers are responsible for cascading flushes and
2545 * handling overflow. Non-first flushers can simply return.
2547 if (wq->first_flusher != &this_flusher)
2550 mutex_lock(&wq->mutex);
2552 /* we might have raced, check again with mutex held */
2553 if (wq->first_flusher != &this_flusher)
2556 wq->first_flusher = NULL;
2558 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2559 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2562 struct wq_flusher *next, *tmp;
2564 /* complete all the flushers sharing the current flush color */
2565 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2566 if (next->flush_color != wq->flush_color)
2568 list_del_init(&next->list);
2569 complete(&next->done);
2572 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2573 wq->flush_color != work_next_color(wq->work_color));
2575 /* this flush_color is finished, advance by one */
2576 wq->flush_color = work_next_color(wq->flush_color);
2578 /* one color has been freed, handle overflow queue */
2579 if (!list_empty(&wq->flusher_overflow)) {
2581 * Assign the same color to all overflowed
2582 * flushers, advance work_color and append to
2583 * flusher_queue. This is the start-to-wait
2584 * phase for these overflowed flushers.
2586 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2587 tmp->flush_color = wq->work_color;
2589 wq->work_color = work_next_color(wq->work_color);
2591 list_splice_tail_init(&wq->flusher_overflow,
2592 &wq->flusher_queue);
2593 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2596 if (list_empty(&wq->flusher_queue)) {
2597 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2602 * Need to flush more colors. Make the next flusher
2603 * the new first flusher and arm pwqs.
2605 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2606 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2608 list_del_init(&next->list);
2609 wq->first_flusher = next;
2611 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2615 * Meh... this color is already done, clear first
2616 * flusher and repeat cascading.
2618 wq->first_flusher = NULL;
2622 mutex_unlock(&wq->mutex);
2624 EXPORT_SYMBOL_GPL(flush_workqueue);
2627 * drain_workqueue - drain a workqueue
2628 * @wq: workqueue to drain
2630 * Wait until the workqueue becomes empty. While draining is in progress,
2631 * only chain queueing is allowed. IOW, only currently pending or running
2632 * work items on @wq can queue further work items on it. @wq is flushed
2633 * repeatedly until it becomes empty. The number of flushing is determined
2634 * by the depth of chaining and should be relatively short. Whine if it
2637 void drain_workqueue(struct workqueue_struct *wq)
2639 unsigned int flush_cnt = 0;
2640 struct pool_workqueue *pwq;
2643 * __queue_work() needs to test whether there are drainers, is much
2644 * hotter than drain_workqueue() and already looks at @wq->flags.
2645 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2647 mutex_lock(&wq->mutex);
2648 if (!wq->nr_drainers++)
2649 wq->flags |= __WQ_DRAINING;
2650 mutex_unlock(&wq->mutex);
2652 flush_workqueue(wq);
2654 mutex_lock(&wq->mutex);
2656 for_each_pwq(pwq, wq) {
2659 spin_lock_irq(&pwq->pool->lock);
2660 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2661 spin_unlock_irq(&pwq->pool->lock);
2666 if (++flush_cnt == 10 ||
2667 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2668 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2669 wq->name, flush_cnt);
2671 mutex_unlock(&wq->mutex);
2675 if (!--wq->nr_drainers)
2676 wq->flags &= ~__WQ_DRAINING;
2677 mutex_unlock(&wq->mutex);
2679 EXPORT_SYMBOL_GPL(drain_workqueue);
2681 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2683 struct worker *worker = NULL;
2684 struct worker_pool *pool;
2685 struct pool_workqueue *pwq;
2689 local_irq_disable();
2690 pool = get_work_pool(work);
2696 spin_lock(&pool->lock);
2697 /* see the comment in try_to_grab_pending() with the same code */
2698 pwq = get_work_pwq(work);
2700 if (unlikely(pwq->pool != pool))
2703 worker = find_worker_executing_work(pool, work);
2706 pwq = worker->current_pwq;
2709 insert_wq_barrier(pwq, barr, work, worker);
2710 spin_unlock_irq(&pool->lock);
2713 * If @max_active is 1 or rescuer is in use, flushing another work
2714 * item on the same workqueue may lead to deadlock. Make sure the
2715 * flusher is not running on the same workqueue by verifying write
2718 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2719 lock_map_acquire(&pwq->wq->lockdep_map);
2721 lock_map_acquire_read(&pwq->wq->lockdep_map);
2722 lock_map_release(&pwq->wq->lockdep_map);
2726 spin_unlock_irq(&pool->lock);
2731 * flush_work - wait for a work to finish executing the last queueing instance
2732 * @work: the work to flush
2734 * Wait until @work has finished execution. @work is guaranteed to be idle
2735 * on return if it hasn't been requeued since flush started.
2738 * %true if flush_work() waited for the work to finish execution,
2739 * %false if it was already idle.
2741 bool flush_work(struct work_struct *work)
2743 struct wq_barrier barr;
2745 lock_map_acquire(&work->lockdep_map);
2746 lock_map_release(&work->lockdep_map);
2748 if (start_flush_work(work, &barr)) {
2749 wait_for_completion(&barr.done);
2750 destroy_work_on_stack(&barr.work);
2756 EXPORT_SYMBOL_GPL(flush_work);
2760 struct work_struct *work;
2763 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2765 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2767 if (cwait->work != key)
2769 return autoremove_wake_function(wait, mode, sync, key);
2772 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2774 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2775 unsigned long flags;
2779 ret = try_to_grab_pending(work, is_dwork, &flags);
2781 * If someone else is already canceling, wait for it to
2782 * finish. flush_work() doesn't work for PREEMPT_NONE
2783 * because we may get scheduled between @work's completion
2784 * and the other canceling task resuming and clearing
2785 * CANCELING - flush_work() will return false immediately
2786 * as @work is no longer busy, try_to_grab_pending() will
2787 * return -ENOENT as @work is still being canceled and the
2788 * other canceling task won't be able to clear CANCELING as
2789 * we're hogging the CPU.
2791 * Let's wait for completion using a waitqueue. As this
2792 * may lead to the thundering herd problem, use a custom
2793 * wake function which matches @work along with exclusive
2796 if (unlikely(ret == -ENOENT)) {
2797 struct cwt_wait cwait;
2799 init_wait(&cwait.wait);
2800 cwait.wait.func = cwt_wakefn;
2803 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2804 TASK_UNINTERRUPTIBLE);
2805 if (work_is_canceling(work))
2807 finish_wait(&cancel_waitq, &cwait.wait);
2809 } while (unlikely(ret < 0));
2811 /* tell other tasks trying to grab @work to back off */
2812 mark_work_canceling(work);
2813 local_irq_restore(flags);
2816 clear_work_data(work);
2819 * Paired with prepare_to_wait() above so that either
2820 * waitqueue_active() is visible here or !work_is_canceling() is
2824 if (waitqueue_active(&cancel_waitq))
2825 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2831 * cancel_work_sync - cancel a work and wait for it to finish
2832 * @work: the work to cancel
2834 * Cancel @work and wait for its execution to finish. This function
2835 * can be used even if the work re-queues itself or migrates to
2836 * another workqueue. On return from this function, @work is
2837 * guaranteed to be not pending or executing on any CPU.
2839 * cancel_work_sync(&delayed_work->work) must not be used for
2840 * delayed_work's. Use cancel_delayed_work_sync() instead.
2842 * The caller must ensure that the workqueue on which @work was last
2843 * queued can't be destroyed before this function returns.
2846 * %true if @work was pending, %false otherwise.
2848 bool cancel_work_sync(struct work_struct *work)
2850 return __cancel_work_timer(work, false);
2852 EXPORT_SYMBOL_GPL(cancel_work_sync);
2855 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2856 * @dwork: the delayed work to flush
2858 * Delayed timer is cancelled and the pending work is queued for
2859 * immediate execution. Like flush_work(), this function only
2860 * considers the last queueing instance of @dwork.
2863 * %true if flush_work() waited for the work to finish execution,
2864 * %false if it was already idle.
2866 bool flush_delayed_work(struct delayed_work *dwork)
2868 local_irq_disable();
2869 if (del_timer_sync(&dwork->timer))
2870 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2872 return flush_work(&dwork->work);
2874 EXPORT_SYMBOL(flush_delayed_work);
2877 * cancel_delayed_work - cancel a delayed work
2878 * @dwork: delayed_work to cancel
2880 * Kill off a pending delayed_work.
2882 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2886 * The work callback function may still be running on return, unless
2887 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2888 * use cancel_delayed_work_sync() to wait on it.
2890 * This function is safe to call from any context including IRQ handler.
2892 bool cancel_delayed_work(struct delayed_work *dwork)
2894 unsigned long flags;
2898 ret = try_to_grab_pending(&dwork->work, true, &flags);
2899 } while (unlikely(ret == -EAGAIN));
2901 if (unlikely(ret < 0))
2904 set_work_pool_and_clear_pending(&dwork->work,
2905 get_work_pool_id(&dwork->work));
2906 local_irq_restore(flags);
2909 EXPORT_SYMBOL(cancel_delayed_work);
2912 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2913 * @dwork: the delayed work cancel
2915 * This is cancel_work_sync() for delayed works.
2918 * %true if @dwork was pending, %false otherwise.
2920 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2922 return __cancel_work_timer(&dwork->work, true);
2924 EXPORT_SYMBOL(cancel_delayed_work_sync);
2927 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2928 * @func: the function to call
2930 * schedule_on_each_cpu() executes @func on each online CPU using the
2931 * system workqueue and blocks until all CPUs have completed.
2932 * schedule_on_each_cpu() is very slow.
2935 * 0 on success, -errno on failure.
2937 int schedule_on_each_cpu(work_func_t func)
2940 struct work_struct __percpu *works;
2942 works = alloc_percpu(struct work_struct);
2948 for_each_online_cpu(cpu) {
2949 struct work_struct *work = per_cpu_ptr(works, cpu);
2951 INIT_WORK(work, func);
2952 schedule_work_on(cpu, work);
2955 for_each_online_cpu(cpu)
2956 flush_work(per_cpu_ptr(works, cpu));
2964 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2966 * Forces execution of the kernel-global workqueue and blocks until its
2969 * Think twice before calling this function! It's very easy to get into
2970 * trouble if you don't take great care. Either of the following situations
2971 * will lead to deadlock:
2973 * One of the work items currently on the workqueue needs to acquire
2974 * a lock held by your code or its caller.
2976 * Your code is running in the context of a work routine.
2978 * They will be detected by lockdep when they occur, but the first might not
2979 * occur very often. It depends on what work items are on the workqueue and
2980 * what locks they need, which you have no control over.
2982 * In most situations flushing the entire workqueue is overkill; you merely
2983 * need to know that a particular work item isn't queued and isn't running.
2984 * In such cases you should use cancel_delayed_work_sync() or
2985 * cancel_work_sync() instead.
2987 void flush_scheduled_work(void)
2989 flush_workqueue(system_wq);
2991 EXPORT_SYMBOL(flush_scheduled_work);
2994 * execute_in_process_context - reliably execute the routine with user context
2995 * @fn: the function to execute
2996 * @ew: guaranteed storage for the execute work structure (must
2997 * be available when the work executes)
2999 * Executes the function immediately if process context is available,
3000 * otherwise schedules the function for delayed execution.
3002 * Return: 0 - function was executed
3003 * 1 - function was scheduled for execution
3005 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3007 if (!in_interrupt()) {
3012 INIT_WORK(&ew->work, fn);
3013 schedule_work(&ew->work);
3017 EXPORT_SYMBOL_GPL(execute_in_process_context);
3020 * free_workqueue_attrs - free a workqueue_attrs
3021 * @attrs: workqueue_attrs to free
3023 * Undo alloc_workqueue_attrs().
3025 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3028 free_cpumask_var(attrs->cpumask);
3034 * alloc_workqueue_attrs - allocate a workqueue_attrs
3035 * @gfp_mask: allocation mask to use
3037 * Allocate a new workqueue_attrs, initialize with default settings and
3040 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3042 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3044 struct workqueue_attrs *attrs;
3046 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3049 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3052 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3055 free_workqueue_attrs(attrs);
3059 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3060 const struct workqueue_attrs *from)
3062 to->nice = from->nice;
3063 cpumask_copy(to->cpumask, from->cpumask);
3065 * Unlike hash and equality test, this function doesn't ignore
3066 * ->no_numa as it is used for both pool and wq attrs. Instead,
3067 * get_unbound_pool() explicitly clears ->no_numa after copying.
3069 to->no_numa = from->no_numa;
3072 /* hash value of the content of @attr */
3073 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3077 hash = jhash_1word(attrs->nice, hash);
3078 hash = jhash(cpumask_bits(attrs->cpumask),
3079 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3083 /* content equality test */
3084 static bool wqattrs_equal(const struct workqueue_attrs *a,
3085 const struct workqueue_attrs *b)
3087 if (a->nice != b->nice)
3089 if (!cpumask_equal(a->cpumask, b->cpumask))
3095 * init_worker_pool - initialize a newly zalloc'd worker_pool
3096 * @pool: worker_pool to initialize
3098 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3100 * Return: 0 on success, -errno on failure. Even on failure, all fields
3101 * inside @pool proper are initialized and put_unbound_pool() can be called
3102 * on @pool safely to release it.
3104 static int init_worker_pool(struct worker_pool *pool)
3106 spin_lock_init(&pool->lock);
3109 pool->node = NUMA_NO_NODE;
3110 pool->flags |= POOL_DISASSOCIATED;
3111 INIT_LIST_HEAD(&pool->worklist);
3112 INIT_LIST_HEAD(&pool->idle_list);
3113 hash_init(pool->busy_hash);
3115 init_timer_deferrable(&pool->idle_timer);
3116 pool->idle_timer.function = idle_worker_timeout;
3117 pool->idle_timer.data = (unsigned long)pool;
3119 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3120 (unsigned long)pool);
3122 mutex_init(&pool->manager_arb);
3123 mutex_init(&pool->attach_mutex);
3124 INIT_LIST_HEAD(&pool->workers);
3126 ida_init(&pool->worker_ida);
3127 INIT_HLIST_NODE(&pool->hash_node);
3130 /* shouldn't fail above this point */
3131 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3137 static void rcu_free_wq(struct rcu_head *rcu)
3139 struct workqueue_struct *wq =
3140 container_of(rcu, struct workqueue_struct, rcu);
3142 if (!(wq->flags & WQ_UNBOUND))
3143 free_percpu(wq->cpu_pwqs);
3145 free_workqueue_attrs(wq->unbound_attrs);
3151 static void rcu_free_pool(struct rcu_head *rcu)
3153 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3155 ida_destroy(&pool->worker_ida);
3156 free_workqueue_attrs(pool->attrs);
3161 * put_unbound_pool - put a worker_pool
3162 * @pool: worker_pool to put
3164 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3165 * safe manner. get_unbound_pool() calls this function on its failure path
3166 * and this function should be able to release pools which went through,
3167 * successfully or not, init_worker_pool().
3169 * Should be called with wq_pool_mutex held.
3171 static void put_unbound_pool(struct worker_pool *pool)
3173 DECLARE_COMPLETION_ONSTACK(detach_completion);
3174 struct worker *worker;
3176 lockdep_assert_held(&wq_pool_mutex);
3182 if (WARN_ON(!(pool->cpu < 0)) ||
3183 WARN_ON(!list_empty(&pool->worklist)))
3186 /* release id and unhash */
3188 idr_remove(&worker_pool_idr, pool->id);
3189 hash_del(&pool->hash_node);
3192 * Become the manager and destroy all workers. Grabbing
3193 * manager_arb prevents @pool's workers from blocking on
3196 mutex_lock(&pool->manager_arb);
3198 spin_lock_irq(&pool->lock);
3199 while ((worker = first_idle_worker(pool)))
3200 destroy_worker(worker);
3201 WARN_ON(pool->nr_workers || pool->nr_idle);
3202 spin_unlock_irq(&pool->lock);
3204 mutex_lock(&pool->attach_mutex);
3205 if (!list_empty(&pool->workers))
3206 pool->detach_completion = &detach_completion;
3207 mutex_unlock(&pool->attach_mutex);
3209 if (pool->detach_completion)
3210 wait_for_completion(pool->detach_completion);
3212 mutex_unlock(&pool->manager_arb);
3214 /* shut down the timers */
3215 del_timer_sync(&pool->idle_timer);
3216 del_timer_sync(&pool->mayday_timer);
3218 /* sched-RCU protected to allow dereferences from get_work_pool() */
3219 call_rcu_sched(&pool->rcu, rcu_free_pool);
3223 * get_unbound_pool - get a worker_pool with the specified attributes
3224 * @attrs: the attributes of the worker_pool to get
3226 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3227 * reference count and return it. If there already is a matching
3228 * worker_pool, it will be used; otherwise, this function attempts to
3231 * Should be called with wq_pool_mutex held.
3233 * Return: On success, a worker_pool with the same attributes as @attrs.
3234 * On failure, %NULL.
3236 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3238 u32 hash = wqattrs_hash(attrs);
3239 struct worker_pool *pool;
3242 lockdep_assert_held(&wq_pool_mutex);
3244 /* do we already have a matching pool? */
3245 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3246 if (wqattrs_equal(pool->attrs, attrs)) {
3252 /* nope, create a new one */
3253 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3254 if (!pool || init_worker_pool(pool) < 0)
3257 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3258 copy_workqueue_attrs(pool->attrs, attrs);
3261 * no_numa isn't a worker_pool attribute, always clear it. See
3262 * 'struct workqueue_attrs' comments for detail.
3264 pool->attrs->no_numa = false;
3266 /* if cpumask is contained inside a NUMA node, we belong to that node */
3267 if (wq_numa_enabled) {
3268 for_each_node(node) {
3269 if (cpumask_subset(pool->attrs->cpumask,
3270 wq_numa_possible_cpumask[node])) {
3277 if (worker_pool_assign_id(pool) < 0)
3280 /* create and start the initial worker */
3281 if (!create_worker(pool))
3285 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3290 put_unbound_pool(pool);
3294 static void rcu_free_pwq(struct rcu_head *rcu)
3296 kmem_cache_free(pwq_cache,
3297 container_of(rcu, struct pool_workqueue, rcu));
3301 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3302 * and needs to be destroyed.
3304 static void pwq_unbound_release_workfn(struct work_struct *work)
3306 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3307 unbound_release_work);
3308 struct workqueue_struct *wq = pwq->wq;
3309 struct worker_pool *pool = pwq->pool;
3312 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3315 mutex_lock(&wq->mutex);
3316 list_del_rcu(&pwq->pwqs_node);
3317 is_last = list_empty(&wq->pwqs);
3318 mutex_unlock(&wq->mutex);
3320 mutex_lock(&wq_pool_mutex);
3321 put_unbound_pool(pool);
3322 mutex_unlock(&wq_pool_mutex);
3324 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3327 * If we're the last pwq going away, @wq is already dead and no one
3328 * is gonna access it anymore. Schedule RCU free.
3331 call_rcu_sched(&wq->rcu, rcu_free_wq);
3335 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3336 * @pwq: target pool_workqueue
3338 * If @pwq isn't freezing, set @pwq->max_active to the associated
3339 * workqueue's saved_max_active and activate delayed work items
3340 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3342 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3344 struct workqueue_struct *wq = pwq->wq;
3345 bool freezable = wq->flags & WQ_FREEZABLE;
3347 /* for @wq->saved_max_active */
3348 lockdep_assert_held(&wq->mutex);
3350 /* fast exit for non-freezable wqs */
3351 if (!freezable && pwq->max_active == wq->saved_max_active)
3354 spin_lock_irq(&pwq->pool->lock);
3357 * During [un]freezing, the caller is responsible for ensuring that
3358 * this function is called at least once after @workqueue_freezing
3359 * is updated and visible.
3361 if (!freezable || !workqueue_freezing) {
3362 pwq->max_active = wq->saved_max_active;
3364 while (!list_empty(&pwq->delayed_works) &&
3365 pwq->nr_active < pwq->max_active)
3366 pwq_activate_first_delayed(pwq);
3369 * Need to kick a worker after thawed or an unbound wq's
3370 * max_active is bumped. It's a slow path. Do it always.
3372 wake_up_worker(pwq->pool);
3374 pwq->max_active = 0;
3377 spin_unlock_irq(&pwq->pool->lock);
3380 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3381 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3382 struct worker_pool *pool)
3384 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3386 memset(pwq, 0, sizeof(*pwq));
3390 pwq->flush_color = -1;
3392 INIT_LIST_HEAD(&pwq->delayed_works);
3393 INIT_LIST_HEAD(&pwq->pwqs_node);
3394 INIT_LIST_HEAD(&pwq->mayday_node);
3395 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3398 /* sync @pwq with the current state of its associated wq and link it */
3399 static void link_pwq(struct pool_workqueue *pwq)
3401 struct workqueue_struct *wq = pwq->wq;
3403 lockdep_assert_held(&wq->mutex);
3405 /* may be called multiple times, ignore if already linked */
3406 if (!list_empty(&pwq->pwqs_node))
3409 /* set the matching work_color */
3410 pwq->work_color = wq->work_color;
3412 /* sync max_active to the current setting */
3413 pwq_adjust_max_active(pwq);
3416 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3419 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3420 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3421 const struct workqueue_attrs *attrs)
3423 struct worker_pool *pool;
3424 struct pool_workqueue *pwq;
3426 lockdep_assert_held(&wq_pool_mutex);
3428 pool = get_unbound_pool(attrs);
3432 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3434 put_unbound_pool(pool);
3438 init_pwq(pwq, wq, pool);
3443 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3444 * @attrs: the wq_attrs of the default pwq of the target workqueue
3445 * @node: the target NUMA node
3446 * @cpu_going_down: if >= 0, the CPU to consider as offline
3447 * @cpumask: outarg, the resulting cpumask
3449 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3450 * @cpu_going_down is >= 0, that cpu is considered offline during
3451 * calculation. The result is stored in @cpumask.
3453 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3454 * enabled and @node has online CPUs requested by @attrs, the returned
3455 * cpumask is the intersection of the possible CPUs of @node and
3458 * The caller is responsible for ensuring that the cpumask of @node stays
3461 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3464 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3465 int cpu_going_down, cpumask_t *cpumask)
3467 if (!wq_numa_enabled || attrs->no_numa)
3470 /* does @node have any online CPUs @attrs wants? */
3471 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3472 if (cpu_going_down >= 0)
3473 cpumask_clear_cpu(cpu_going_down, cpumask);
3475 if (cpumask_empty(cpumask))
3478 /* yeap, return possible CPUs in @node that @attrs wants */
3479 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3480 return !cpumask_equal(cpumask, attrs->cpumask);
3483 cpumask_copy(cpumask, attrs->cpumask);
3487 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3488 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3490 struct pool_workqueue *pwq)
3492 struct pool_workqueue *old_pwq;
3494 lockdep_assert_held(&wq_pool_mutex);
3495 lockdep_assert_held(&wq->mutex);
3497 /* link_pwq() can handle duplicate calls */
3500 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3501 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3505 /* context to store the prepared attrs & pwqs before applying */
3506 struct apply_wqattrs_ctx {
3507 struct workqueue_struct *wq; /* target workqueue */
3508 struct workqueue_attrs *attrs; /* attrs to apply */
3509 struct list_head list; /* queued for batching commit */
3510 struct pool_workqueue *dfl_pwq;
3511 struct pool_workqueue *pwq_tbl[];
3514 /* free the resources after success or abort */
3515 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3521 put_pwq_unlocked(ctx->pwq_tbl[node]);
3522 put_pwq_unlocked(ctx->dfl_pwq);
3524 free_workqueue_attrs(ctx->attrs);
3530 /* allocate the attrs and pwqs for later installation */
3531 static struct apply_wqattrs_ctx *
3532 apply_wqattrs_prepare(struct workqueue_struct *wq,
3533 const struct workqueue_attrs *attrs)
3535 struct apply_wqattrs_ctx *ctx;
3536 struct workqueue_attrs *new_attrs, *tmp_attrs;
3539 lockdep_assert_held(&wq_pool_mutex);
3541 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3544 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3545 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3546 if (!ctx || !new_attrs || !tmp_attrs)
3550 * Calculate the attrs of the default pwq.
3551 * If the user configured cpumask doesn't overlap with the
3552 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3554 copy_workqueue_attrs(new_attrs, attrs);
3555 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3556 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3557 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3560 * We may create multiple pwqs with differing cpumasks. Make a
3561 * copy of @new_attrs which will be modified and used to obtain
3564 copy_workqueue_attrs(tmp_attrs, new_attrs);
3567 * If something goes wrong during CPU up/down, we'll fall back to
3568 * the default pwq covering whole @attrs->cpumask. Always create
3569 * it even if we don't use it immediately.
3571 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3575 for_each_node(node) {
3576 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3577 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3578 if (!ctx->pwq_tbl[node])
3581 ctx->dfl_pwq->refcnt++;
3582 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3586 /* save the user configured attrs and sanitize it. */
3587 copy_workqueue_attrs(new_attrs, attrs);
3588 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3589 ctx->attrs = new_attrs;
3592 free_workqueue_attrs(tmp_attrs);
3596 free_workqueue_attrs(tmp_attrs);
3597 free_workqueue_attrs(new_attrs);
3598 apply_wqattrs_cleanup(ctx);
3602 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3603 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3607 /* all pwqs have been created successfully, let's install'em */
3608 mutex_lock(&ctx->wq->mutex);
3610 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3612 /* save the previous pwq and install the new one */
3614 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3615 ctx->pwq_tbl[node]);
3617 /* @dfl_pwq might not have been used, ensure it's linked */
3618 link_pwq(ctx->dfl_pwq);
3619 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3621 mutex_unlock(&ctx->wq->mutex);
3625 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3626 * @wq: the target workqueue
3627 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3629 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3630 * machines, this function maps a separate pwq to each NUMA node with
3631 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3632 * NUMA node it was issued on. Older pwqs are released as in-flight work
3633 * items finish. Note that a work item which repeatedly requeues itself
3634 * back-to-back will stay on its current pwq.
3636 * Performs GFP_KERNEL allocations.
3638 * Return: 0 on success and -errno on failure.
3640 int apply_workqueue_attrs(struct workqueue_struct *wq,
3641 const struct workqueue_attrs *attrs)
3643 struct apply_wqattrs_ctx *ctx;
3646 /* only unbound workqueues can change attributes */
3647 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3650 /* creating multiple pwqs breaks ordering guarantee */
3651 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3655 * CPUs should stay stable across pwq creations and installations.
3656 * Pin CPUs, determine the target cpumask for each node and create
3660 mutex_lock(&wq_pool_mutex);
3662 ctx = apply_wqattrs_prepare(wq, attrs);
3664 /* the ctx has been prepared successfully, let's commit it */
3666 apply_wqattrs_commit(ctx);
3670 mutex_unlock(&wq_pool_mutex);
3673 apply_wqattrs_cleanup(ctx);
3679 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3680 * @wq: the target workqueue
3681 * @cpu: the CPU coming up or going down
3682 * @online: whether @cpu is coming up or going down
3684 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3685 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3688 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3689 * falls back to @wq->dfl_pwq which may not be optimal but is always
3692 * Note that when the last allowed CPU of a NUMA node goes offline for a
3693 * workqueue with a cpumask spanning multiple nodes, the workers which were
3694 * already executing the work items for the workqueue will lose their CPU
3695 * affinity and may execute on any CPU. This is similar to how per-cpu
3696 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3697 * affinity, it's the user's responsibility to flush the work item from
3700 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3703 int node = cpu_to_node(cpu);
3704 int cpu_off = online ? -1 : cpu;
3705 struct pool_workqueue *old_pwq = NULL, *pwq;
3706 struct workqueue_attrs *target_attrs;
3709 lockdep_assert_held(&wq_pool_mutex);
3711 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3715 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3716 * Let's use a preallocated one. The following buf is protected by
3717 * CPU hotplug exclusion.
3719 target_attrs = wq_update_unbound_numa_attrs_buf;
3720 cpumask = target_attrs->cpumask;
3722 mutex_lock(&wq->mutex);
3723 if (wq->unbound_attrs->no_numa)
3726 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3727 pwq = unbound_pwq_by_node(wq, node);
3730 * Let's determine what needs to be done. If the target cpumask is
3731 * different from the default pwq's, we need to compare it to @pwq's
3732 * and create a new one if they don't match. If the target cpumask
3733 * equals the default pwq's, the default pwq should be used.
3735 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3736 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3742 mutex_unlock(&wq->mutex);
3744 /* create a new pwq */
3745 pwq = alloc_unbound_pwq(wq, target_attrs);
3747 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3749 mutex_lock(&wq->mutex);
3754 * Install the new pwq. As this function is called only from CPU
3755 * hotplug callbacks and applying a new attrs is wrapped with
3756 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3759 mutex_lock(&wq->mutex);
3760 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3764 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3765 get_pwq(wq->dfl_pwq);
3766 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3767 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3769 mutex_unlock(&wq->mutex);
3770 put_pwq_unlocked(old_pwq);
3773 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3775 bool highpri = wq->flags & WQ_HIGHPRI;
3778 if (!(wq->flags & WQ_UNBOUND)) {
3779 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3783 for_each_possible_cpu(cpu) {
3784 struct pool_workqueue *pwq =
3785 per_cpu_ptr(wq->cpu_pwqs, cpu);
3786 struct worker_pool *cpu_pools =
3787 per_cpu(cpu_worker_pools, cpu);
3789 init_pwq(pwq, wq, &cpu_pools[highpri]);
3791 mutex_lock(&wq->mutex);
3793 mutex_unlock(&wq->mutex);
3796 } else if (wq->flags & __WQ_ORDERED) {
3797 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3798 /* there should only be single pwq for ordering guarantee */
3799 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3800 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3801 "ordering guarantee broken for workqueue %s\n", wq->name);
3804 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3808 static int wq_clamp_max_active(int max_active, unsigned int flags,
3811 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3813 if (max_active < 1 || max_active > lim)
3814 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3815 max_active, name, 1, lim);
3817 return clamp_val(max_active, 1, lim);
3820 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3823 struct lock_class_key *key,
3824 const char *lock_name, ...)
3826 size_t tbl_size = 0;
3828 struct workqueue_struct *wq;
3829 struct pool_workqueue *pwq;
3831 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3832 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3833 flags |= WQ_UNBOUND;
3835 /* allocate wq and format name */
3836 if (flags & WQ_UNBOUND)
3837 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3839 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3843 if (flags & WQ_UNBOUND) {
3844 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3845 if (!wq->unbound_attrs)
3849 va_start(args, lock_name);
3850 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3853 max_active = max_active ?: WQ_DFL_ACTIVE;
3854 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3858 wq->saved_max_active = max_active;
3859 mutex_init(&wq->mutex);
3860 atomic_set(&wq->nr_pwqs_to_flush, 0);
3861 INIT_LIST_HEAD(&wq->pwqs);
3862 INIT_LIST_HEAD(&wq->flusher_queue);
3863 INIT_LIST_HEAD(&wq->flusher_overflow);
3864 INIT_LIST_HEAD(&wq->maydays);
3866 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3867 INIT_LIST_HEAD(&wq->list);
3869 if (alloc_and_link_pwqs(wq) < 0)
3873 * Workqueues which may be used during memory reclaim should
3874 * have a rescuer to guarantee forward progress.
3876 if (flags & WQ_MEM_RECLAIM) {
3877 struct worker *rescuer;
3879 rescuer = alloc_worker(NUMA_NO_NODE);
3883 rescuer->rescue_wq = wq;
3884 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3886 if (IS_ERR(rescuer->task)) {
3891 wq->rescuer = rescuer;
3892 rescuer->task->flags |= PF_NO_SETAFFINITY;
3893 wake_up_process(rescuer->task);
3896 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3900 * wq_pool_mutex protects global freeze state and workqueues list.
3901 * Grab it, adjust max_active and add the new @wq to workqueues
3904 mutex_lock(&wq_pool_mutex);
3906 mutex_lock(&wq->mutex);
3907 for_each_pwq(pwq, wq)
3908 pwq_adjust_max_active(pwq);
3909 mutex_unlock(&wq->mutex);
3911 list_add_tail_rcu(&wq->list, &workqueues);
3913 mutex_unlock(&wq_pool_mutex);
3918 free_workqueue_attrs(wq->unbound_attrs);
3922 destroy_workqueue(wq);
3925 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3928 * destroy_workqueue - safely terminate a workqueue
3929 * @wq: target workqueue
3931 * Safely destroy a workqueue. All work currently pending will be done first.
3933 void destroy_workqueue(struct workqueue_struct *wq)
3935 struct pool_workqueue *pwq;
3938 /* drain it before proceeding with destruction */
3939 drain_workqueue(wq);
3942 mutex_lock(&wq->mutex);
3943 for_each_pwq(pwq, wq) {
3946 for (i = 0; i < WORK_NR_COLORS; i++) {
3947 if (WARN_ON(pwq->nr_in_flight[i])) {
3948 mutex_unlock(&wq->mutex);
3953 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3954 WARN_ON(pwq->nr_active) ||
3955 WARN_ON(!list_empty(&pwq->delayed_works))) {
3956 mutex_unlock(&wq->mutex);
3960 mutex_unlock(&wq->mutex);
3963 * wq list is used to freeze wq, remove from list after
3964 * flushing is complete in case freeze races us.
3966 mutex_lock(&wq_pool_mutex);
3967 list_del_rcu(&wq->list);
3968 mutex_unlock(&wq_pool_mutex);
3970 workqueue_sysfs_unregister(wq);
3973 kthread_stop(wq->rescuer->task);
3975 if (!(wq->flags & WQ_UNBOUND)) {
3977 * The base ref is never dropped on per-cpu pwqs. Directly
3978 * schedule RCU free.
3980 call_rcu_sched(&wq->rcu, rcu_free_wq);
3983 * We're the sole accessor of @wq at this point. Directly
3984 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3985 * @wq will be freed when the last pwq is released.
3987 for_each_node(node) {
3988 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3989 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3990 put_pwq_unlocked(pwq);
3994 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
3995 * put. Don't access it afterwards.
3999 put_pwq_unlocked(pwq);
4002 EXPORT_SYMBOL_GPL(destroy_workqueue);
4005 * workqueue_set_max_active - adjust max_active of a workqueue
4006 * @wq: target workqueue
4007 * @max_active: new max_active value.
4009 * Set max_active of @wq to @max_active.
4012 * Don't call from IRQ context.
4014 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4016 struct pool_workqueue *pwq;
4018 /* disallow meddling with max_active for ordered workqueues */
4019 if (WARN_ON(wq->flags & __WQ_ORDERED))
4022 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4024 mutex_lock(&wq->mutex);
4026 wq->saved_max_active = max_active;
4028 for_each_pwq(pwq, wq)
4029 pwq_adjust_max_active(pwq);
4031 mutex_unlock(&wq->mutex);
4033 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4036 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4038 * Determine whether %current is a workqueue rescuer. Can be used from
4039 * work functions to determine whether it's being run off the rescuer task.
4041 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4043 bool current_is_workqueue_rescuer(void)
4045 struct worker *worker = current_wq_worker();
4047 return worker && worker->rescue_wq;
4051 * workqueue_congested - test whether a workqueue is congested
4052 * @cpu: CPU in question
4053 * @wq: target workqueue
4055 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4056 * no synchronization around this function and the test result is
4057 * unreliable and only useful as advisory hints or for debugging.
4059 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4060 * Note that both per-cpu and unbound workqueues may be associated with
4061 * multiple pool_workqueues which have separate congested states. A
4062 * workqueue being congested on one CPU doesn't mean the workqueue is also
4063 * contested on other CPUs / NUMA nodes.
4066 * %true if congested, %false otherwise.
4068 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4070 struct pool_workqueue *pwq;
4073 rcu_read_lock_sched();
4075 if (cpu == WORK_CPU_UNBOUND)
4076 cpu = smp_processor_id();
4078 if (!(wq->flags & WQ_UNBOUND))
4079 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4081 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4083 ret = !list_empty(&pwq->delayed_works);
4084 rcu_read_unlock_sched();
4088 EXPORT_SYMBOL_GPL(workqueue_congested);
4091 * work_busy - test whether a work is currently pending or running
4092 * @work: the work to be tested
4094 * Test whether @work is currently pending or running. There is no
4095 * synchronization around this function and the test result is
4096 * unreliable and only useful as advisory hints or for debugging.
4099 * OR'd bitmask of WORK_BUSY_* bits.
4101 unsigned int work_busy(struct work_struct *work)
4103 struct worker_pool *pool;
4104 unsigned long flags;
4105 unsigned int ret = 0;
4107 if (work_pending(work))
4108 ret |= WORK_BUSY_PENDING;
4110 local_irq_save(flags);
4111 pool = get_work_pool(work);
4113 spin_lock(&pool->lock);
4114 if (find_worker_executing_work(pool, work))
4115 ret |= WORK_BUSY_RUNNING;
4116 spin_unlock(&pool->lock);
4118 local_irq_restore(flags);
4122 EXPORT_SYMBOL_GPL(work_busy);
4125 * set_worker_desc - set description for the current work item
4126 * @fmt: printf-style format string
4127 * @...: arguments for the format string
4129 * This function can be called by a running work function to describe what
4130 * the work item is about. If the worker task gets dumped, this
4131 * information will be printed out together to help debugging. The
4132 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4134 void set_worker_desc(const char *fmt, ...)
4136 struct worker *worker = current_wq_worker();
4140 va_start(args, fmt);
4141 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4143 worker->desc_valid = true;
4148 * print_worker_info - print out worker information and description
4149 * @log_lvl: the log level to use when printing
4150 * @task: target task
4152 * If @task is a worker and currently executing a work item, print out the
4153 * name of the workqueue being serviced and worker description set with
4154 * set_worker_desc() by the currently executing work item.
4156 * This function can be safely called on any task as long as the
4157 * task_struct itself is accessible. While safe, this function isn't
4158 * synchronized and may print out mixups or garbages of limited length.
4160 void print_worker_info(const char *log_lvl, struct task_struct *task)
4162 work_func_t *fn = NULL;
4163 char name[WQ_NAME_LEN] = { };
4164 char desc[WORKER_DESC_LEN] = { };
4165 struct pool_workqueue *pwq = NULL;
4166 struct workqueue_struct *wq = NULL;
4167 bool desc_valid = false;
4168 struct worker *worker;
4170 if (!(task->flags & PF_WQ_WORKER))
4174 * This function is called without any synchronization and @task
4175 * could be in any state. Be careful with dereferences.
4177 worker = probe_kthread_data(task);
4180 * Carefully copy the associated workqueue's workfn and name. Keep
4181 * the original last '\0' in case the original contains garbage.
4183 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4184 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4185 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4186 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4188 /* copy worker description */
4189 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4191 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4193 if (fn || name[0] || desc[0]) {
4194 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4196 pr_cont(" (%s)", desc);
4201 static void pr_cont_pool_info(struct worker_pool *pool)
4203 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4204 if (pool->node != NUMA_NO_NODE)
4205 pr_cont(" node=%d", pool->node);
4206 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4209 static void pr_cont_work(bool comma, struct work_struct *work)
4211 if (work->func == wq_barrier_func) {
4212 struct wq_barrier *barr;
4214 barr = container_of(work, struct wq_barrier, work);
4216 pr_cont("%s BAR(%d)", comma ? "," : "",
4217 task_pid_nr(barr->task));
4219 pr_cont("%s %pf", comma ? "," : "", work->func);
4223 static void show_pwq(struct pool_workqueue *pwq)
4225 struct worker_pool *pool = pwq->pool;
4226 struct work_struct *work;
4227 struct worker *worker;
4228 bool has_in_flight = false, has_pending = false;
4231 pr_info(" pwq %d:", pool->id);
4232 pr_cont_pool_info(pool);
4234 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4235 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4237 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4238 if (worker->current_pwq == pwq) {
4239 has_in_flight = true;
4243 if (has_in_flight) {
4246 pr_info(" in-flight:");
4247 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4248 if (worker->current_pwq != pwq)
4251 pr_cont("%s %d%s:%pf", comma ? "," : "",
4252 task_pid_nr(worker->task),
4253 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4254 worker->current_func);
4255 list_for_each_entry(work, &worker->scheduled, entry)
4256 pr_cont_work(false, work);
4262 list_for_each_entry(work, &pool->worklist, entry) {
4263 if (get_work_pwq(work) == pwq) {
4271 pr_info(" pending:");
4272 list_for_each_entry(work, &pool->worklist, entry) {
4273 if (get_work_pwq(work) != pwq)
4276 pr_cont_work(comma, work);
4277 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4282 if (!list_empty(&pwq->delayed_works)) {
4285 pr_info(" delayed:");
4286 list_for_each_entry(work, &pwq->delayed_works, entry) {
4287 pr_cont_work(comma, work);
4288 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4295 * show_workqueue_state - dump workqueue state
4297 * Called from a sysrq handler and prints out all busy workqueues and
4300 void show_workqueue_state(void)
4302 struct workqueue_struct *wq;
4303 struct worker_pool *pool;
4304 unsigned long flags;
4307 rcu_read_lock_sched();
4309 pr_info("Showing busy workqueues and worker pools:\n");
4311 list_for_each_entry_rcu(wq, &workqueues, list) {
4312 struct pool_workqueue *pwq;
4315 for_each_pwq(pwq, wq) {
4316 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4324 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4326 for_each_pwq(pwq, wq) {
4327 spin_lock_irqsave(&pwq->pool->lock, flags);
4328 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4330 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4334 for_each_pool(pool, pi) {
4335 struct worker *worker;
4338 spin_lock_irqsave(&pool->lock, flags);
4339 if (pool->nr_workers == pool->nr_idle)
4342 pr_info("pool %d:", pool->id);
4343 pr_cont_pool_info(pool);
4344 pr_cont(" workers=%d", pool->nr_workers);
4346 pr_cont(" manager: %d",
4347 task_pid_nr(pool->manager->task));
4348 list_for_each_entry(worker, &pool->idle_list, entry) {
4349 pr_cont(" %s%d", first ? "idle: " : "",
4350 task_pid_nr(worker->task));
4355 spin_unlock_irqrestore(&pool->lock, flags);
4358 rcu_read_unlock_sched();
4364 * There are two challenges in supporting CPU hotplug. Firstly, there
4365 * are a lot of assumptions on strong associations among work, pwq and
4366 * pool which make migrating pending and scheduled works very
4367 * difficult to implement without impacting hot paths. Secondly,
4368 * worker pools serve mix of short, long and very long running works making
4369 * blocked draining impractical.
4371 * This is solved by allowing the pools to be disassociated from the CPU
4372 * running as an unbound one and allowing it to be reattached later if the
4373 * cpu comes back online.
4376 static void wq_unbind_fn(struct work_struct *work)
4378 int cpu = smp_processor_id();
4379 struct worker_pool *pool;
4380 struct worker *worker;
4382 for_each_cpu_worker_pool(pool, cpu) {
4383 mutex_lock(&pool->attach_mutex);
4384 spin_lock_irq(&pool->lock);
4387 * We've blocked all attach/detach operations. Make all workers
4388 * unbound and set DISASSOCIATED. Before this, all workers
4389 * except for the ones which are still executing works from
4390 * before the last CPU down must be on the cpu. After
4391 * this, they may become diasporas.
4393 for_each_pool_worker(worker, pool)
4394 worker->flags |= WORKER_UNBOUND;
4396 pool->flags |= POOL_DISASSOCIATED;
4398 spin_unlock_irq(&pool->lock);
4399 mutex_unlock(&pool->attach_mutex);
4402 * Call schedule() so that we cross rq->lock and thus can
4403 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4404 * This is necessary as scheduler callbacks may be invoked
4410 * Sched callbacks are disabled now. Zap nr_running.
4411 * After this, nr_running stays zero and need_more_worker()
4412 * and keep_working() are always true as long as the
4413 * worklist is not empty. This pool now behaves as an
4414 * unbound (in terms of concurrency management) pool which
4415 * are served by workers tied to the pool.
4417 atomic_set(&pool->nr_running, 0);
4420 * With concurrency management just turned off, a busy
4421 * worker blocking could lead to lengthy stalls. Kick off
4422 * unbound chain execution of currently pending work items.
4424 spin_lock_irq(&pool->lock);
4425 wake_up_worker(pool);
4426 spin_unlock_irq(&pool->lock);
4431 * rebind_workers - rebind all workers of a pool to the associated CPU
4432 * @pool: pool of interest
4434 * @pool->cpu is coming online. Rebind all workers to the CPU.
4436 static void rebind_workers(struct worker_pool *pool)
4438 struct worker *worker;
4440 lockdep_assert_held(&pool->attach_mutex);
4443 * Restore CPU affinity of all workers. As all idle workers should
4444 * be on the run-queue of the associated CPU before any local
4445 * wake-ups for concurrency management happen, restore CPU affinty
4446 * of all workers first and then clear UNBOUND. As we're called
4447 * from CPU_ONLINE, the following shouldn't fail.
4449 for_each_pool_worker(worker, pool)
4450 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4451 pool->attrs->cpumask) < 0);
4453 spin_lock_irq(&pool->lock);
4454 pool->flags &= ~POOL_DISASSOCIATED;
4456 for_each_pool_worker(worker, pool) {
4457 unsigned int worker_flags = worker->flags;
4460 * A bound idle worker should actually be on the runqueue
4461 * of the associated CPU for local wake-ups targeting it to
4462 * work. Kick all idle workers so that they migrate to the
4463 * associated CPU. Doing this in the same loop as
4464 * replacing UNBOUND with REBOUND is safe as no worker will
4465 * be bound before @pool->lock is released.
4467 if (worker_flags & WORKER_IDLE)
4468 wake_up_process(worker->task);
4471 * We want to clear UNBOUND but can't directly call
4472 * worker_clr_flags() or adjust nr_running. Atomically
4473 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4474 * @worker will clear REBOUND using worker_clr_flags() when
4475 * it initiates the next execution cycle thus restoring
4476 * concurrency management. Note that when or whether
4477 * @worker clears REBOUND doesn't affect correctness.
4479 * ACCESS_ONCE() is necessary because @worker->flags may be
4480 * tested without holding any lock in
4481 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4482 * fail incorrectly leading to premature concurrency
4483 * management operations.
4485 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4486 worker_flags |= WORKER_REBOUND;
4487 worker_flags &= ~WORKER_UNBOUND;
4488 ACCESS_ONCE(worker->flags) = worker_flags;
4491 spin_unlock_irq(&pool->lock);
4495 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4496 * @pool: unbound pool of interest
4497 * @cpu: the CPU which is coming up
4499 * An unbound pool may end up with a cpumask which doesn't have any online
4500 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4501 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4502 * online CPU before, cpus_allowed of all its workers should be restored.
4504 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4506 static cpumask_t cpumask;
4507 struct worker *worker;
4509 lockdep_assert_held(&pool->attach_mutex);
4511 /* is @cpu allowed for @pool? */
4512 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4515 /* is @cpu the only online CPU? */
4516 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4517 if (cpumask_weight(&cpumask) != 1)
4520 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4521 for_each_pool_worker(worker, pool)
4522 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4523 pool->attrs->cpumask) < 0);
4527 * Workqueues should be brought up before normal priority CPU notifiers.
4528 * This will be registered high priority CPU notifier.
4530 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4531 unsigned long action,
4534 int cpu = (unsigned long)hcpu;
4535 struct worker_pool *pool;
4536 struct workqueue_struct *wq;
4539 switch (action & ~CPU_TASKS_FROZEN) {
4540 case CPU_UP_PREPARE:
4541 for_each_cpu_worker_pool(pool, cpu) {
4542 if (pool->nr_workers)
4544 if (!create_worker(pool))
4549 case CPU_DOWN_FAILED:
4551 mutex_lock(&wq_pool_mutex);
4553 for_each_pool(pool, pi) {
4554 mutex_lock(&pool->attach_mutex);
4556 if (pool->cpu == cpu)
4557 rebind_workers(pool);
4558 else if (pool->cpu < 0)
4559 restore_unbound_workers_cpumask(pool, cpu);
4561 mutex_unlock(&pool->attach_mutex);
4564 /* update NUMA affinity of unbound workqueues */
4565 list_for_each_entry(wq, &workqueues, list)
4566 wq_update_unbound_numa(wq, cpu, true);
4568 mutex_unlock(&wq_pool_mutex);
4575 * Workqueues should be brought down after normal priority CPU notifiers.
4576 * This will be registered as low priority CPU notifier.
4578 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4579 unsigned long action,
4582 int cpu = (unsigned long)hcpu;
4583 struct work_struct unbind_work;
4584 struct workqueue_struct *wq;
4586 switch (action & ~CPU_TASKS_FROZEN) {
4587 case CPU_DOWN_PREPARE:
4588 /* unbinding per-cpu workers should happen on the local CPU */
4589 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4590 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4592 /* update NUMA affinity of unbound workqueues */
4593 mutex_lock(&wq_pool_mutex);
4594 list_for_each_entry(wq, &workqueues, list)
4595 wq_update_unbound_numa(wq, cpu, false);
4596 mutex_unlock(&wq_pool_mutex);
4598 /* wait for per-cpu unbinding to finish */
4599 flush_work(&unbind_work);
4600 destroy_work_on_stack(&unbind_work);
4608 struct work_for_cpu {
4609 struct work_struct work;
4615 static void work_for_cpu_fn(struct work_struct *work)
4617 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4619 wfc->ret = wfc->fn(wfc->arg);
4623 * work_on_cpu - run a function in user context on a particular cpu
4624 * @cpu: the cpu to run on
4625 * @fn: the function to run
4626 * @arg: the function arg
4628 * It is up to the caller to ensure that the cpu doesn't go offline.
4629 * The caller must not hold any locks which would prevent @fn from completing.
4631 * Return: The value @fn returns.
4633 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4635 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4637 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4638 schedule_work_on(cpu, &wfc.work);
4639 flush_work(&wfc.work);
4640 destroy_work_on_stack(&wfc.work);
4643 EXPORT_SYMBOL_GPL(work_on_cpu);
4644 #endif /* CONFIG_SMP */
4646 #ifdef CONFIG_FREEZER
4649 * freeze_workqueues_begin - begin freezing workqueues
4651 * Start freezing workqueues. After this function returns, all freezable
4652 * workqueues will queue new works to their delayed_works list instead of
4656 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4658 void freeze_workqueues_begin(void)
4660 struct workqueue_struct *wq;
4661 struct pool_workqueue *pwq;
4663 mutex_lock(&wq_pool_mutex);
4665 WARN_ON_ONCE(workqueue_freezing);
4666 workqueue_freezing = true;
4668 list_for_each_entry(wq, &workqueues, list) {
4669 mutex_lock(&wq->mutex);
4670 for_each_pwq(pwq, wq)
4671 pwq_adjust_max_active(pwq);
4672 mutex_unlock(&wq->mutex);
4675 mutex_unlock(&wq_pool_mutex);
4679 * freeze_workqueues_busy - are freezable workqueues still busy?
4681 * Check whether freezing is complete. This function must be called
4682 * between freeze_workqueues_begin() and thaw_workqueues().
4685 * Grabs and releases wq_pool_mutex.
4688 * %true if some freezable workqueues are still busy. %false if freezing
4691 bool freeze_workqueues_busy(void)
4694 struct workqueue_struct *wq;
4695 struct pool_workqueue *pwq;
4697 mutex_lock(&wq_pool_mutex);
4699 WARN_ON_ONCE(!workqueue_freezing);
4701 list_for_each_entry(wq, &workqueues, list) {
4702 if (!(wq->flags & WQ_FREEZABLE))
4705 * nr_active is monotonically decreasing. It's safe
4706 * to peek without lock.
4708 rcu_read_lock_sched();
4709 for_each_pwq(pwq, wq) {
4710 WARN_ON_ONCE(pwq->nr_active < 0);
4711 if (pwq->nr_active) {
4713 rcu_read_unlock_sched();
4717 rcu_read_unlock_sched();
4720 mutex_unlock(&wq_pool_mutex);
4725 * thaw_workqueues - thaw workqueues
4727 * Thaw workqueues. Normal queueing is restored and all collected
4728 * frozen works are transferred to their respective pool worklists.
4731 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4733 void thaw_workqueues(void)
4735 struct workqueue_struct *wq;
4736 struct pool_workqueue *pwq;
4738 mutex_lock(&wq_pool_mutex);
4740 if (!workqueue_freezing)
4743 workqueue_freezing = false;
4745 /* restore max_active and repopulate worklist */
4746 list_for_each_entry(wq, &workqueues, list) {
4747 mutex_lock(&wq->mutex);
4748 for_each_pwq(pwq, wq)
4749 pwq_adjust_max_active(pwq);
4750 mutex_unlock(&wq->mutex);
4754 mutex_unlock(&wq_pool_mutex);
4756 #endif /* CONFIG_FREEZER */
4758 static int workqueue_apply_unbound_cpumask(void)
4762 struct workqueue_struct *wq;
4763 struct apply_wqattrs_ctx *ctx, *n;
4765 lockdep_assert_held(&wq_pool_mutex);
4767 list_for_each_entry(wq, &workqueues, list) {
4768 if (!(wq->flags & WQ_UNBOUND))
4770 /* creating multiple pwqs breaks ordering guarantee */
4771 if (wq->flags & __WQ_ORDERED)
4774 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4780 list_add_tail(&ctx->list, &ctxs);
4783 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4785 apply_wqattrs_commit(ctx);
4786 apply_wqattrs_cleanup(ctx);
4793 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4794 * @cpumask: the cpumask to set
4796 * The low-level workqueues cpumask is a global cpumask that limits
4797 * the affinity of all unbound workqueues. This function check the @cpumask
4798 * and apply it to all unbound workqueues and updates all pwqs of them.
4800 * Retun: 0 - Success
4801 * -EINVAL - Invalid @cpumask
4802 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4804 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4807 cpumask_var_t saved_cpumask;
4809 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4813 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4814 if (!cpumask_empty(cpumask)) {
4815 mutex_lock(&wq_pool_mutex);
4817 /* save the old wq_unbound_cpumask. */
4818 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4820 /* update wq_unbound_cpumask at first and apply it to wqs. */
4821 cpumask_copy(wq_unbound_cpumask, cpumask);
4822 ret = workqueue_apply_unbound_cpumask();
4824 /* restore the wq_unbound_cpumask when failed. */
4826 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4828 mutex_unlock(&wq_pool_mutex);
4832 free_cpumask_var(saved_cpumask);
4838 * Workqueues with WQ_SYSFS flag set is visible to userland via
4839 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4840 * following attributes.
4842 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4843 * max_active RW int : maximum number of in-flight work items
4845 * Unbound workqueues have the following extra attributes.
4847 * id RO int : the associated pool ID
4848 * nice RW int : nice value of the workers
4849 * cpumask RW mask : bitmask of allowed CPUs for the workers
4852 struct workqueue_struct *wq;
4856 static struct workqueue_struct *dev_to_wq(struct device *dev)
4858 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4863 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4866 struct workqueue_struct *wq = dev_to_wq(dev);
4868 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4870 static DEVICE_ATTR_RO(per_cpu);
4872 static ssize_t max_active_show(struct device *dev,
4873 struct device_attribute *attr, char *buf)
4875 struct workqueue_struct *wq = dev_to_wq(dev);
4877 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4880 static ssize_t max_active_store(struct device *dev,
4881 struct device_attribute *attr, const char *buf,
4884 struct workqueue_struct *wq = dev_to_wq(dev);
4887 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4890 workqueue_set_max_active(wq, val);
4893 static DEVICE_ATTR_RW(max_active);
4895 static struct attribute *wq_sysfs_attrs[] = {
4896 &dev_attr_per_cpu.attr,
4897 &dev_attr_max_active.attr,
4900 ATTRIBUTE_GROUPS(wq_sysfs);
4902 static ssize_t wq_pool_ids_show(struct device *dev,
4903 struct device_attribute *attr, char *buf)
4905 struct workqueue_struct *wq = dev_to_wq(dev);
4906 const char *delim = "";
4907 int node, written = 0;
4909 rcu_read_lock_sched();
4910 for_each_node(node) {
4911 written += scnprintf(buf + written, PAGE_SIZE - written,
4912 "%s%d:%d", delim, node,
4913 unbound_pwq_by_node(wq, node)->pool->id);
4916 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4917 rcu_read_unlock_sched();
4922 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4925 struct workqueue_struct *wq = dev_to_wq(dev);
4928 mutex_lock(&wq->mutex);
4929 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4930 mutex_unlock(&wq->mutex);
4935 /* prepare workqueue_attrs for sysfs store operations */
4936 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4938 struct workqueue_attrs *attrs;
4940 attrs = alloc_workqueue_attrs(GFP_KERNEL);
4944 mutex_lock(&wq->mutex);
4945 copy_workqueue_attrs(attrs, wq->unbound_attrs);
4946 mutex_unlock(&wq->mutex);
4950 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4951 const char *buf, size_t count)
4953 struct workqueue_struct *wq = dev_to_wq(dev);
4954 struct workqueue_attrs *attrs;
4957 attrs = wq_sysfs_prep_attrs(wq);
4961 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4962 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4963 ret = apply_workqueue_attrs(wq, attrs);
4967 free_workqueue_attrs(attrs);
4968 return ret ?: count;
4971 static ssize_t wq_cpumask_show(struct device *dev,
4972 struct device_attribute *attr, char *buf)
4974 struct workqueue_struct *wq = dev_to_wq(dev);
4977 mutex_lock(&wq->mutex);
4978 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4979 cpumask_pr_args(wq->unbound_attrs->cpumask));
4980 mutex_unlock(&wq->mutex);
4984 static ssize_t wq_cpumask_store(struct device *dev,
4985 struct device_attribute *attr,
4986 const char *buf, size_t count)
4988 struct workqueue_struct *wq = dev_to_wq(dev);
4989 struct workqueue_attrs *attrs;
4992 attrs = wq_sysfs_prep_attrs(wq);
4996 ret = cpumask_parse(buf, attrs->cpumask);
4998 ret = apply_workqueue_attrs(wq, attrs);
5000 free_workqueue_attrs(attrs);
5001 return ret ?: count;
5004 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5007 struct workqueue_struct *wq = dev_to_wq(dev);
5010 mutex_lock(&wq->mutex);
5011 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5012 !wq->unbound_attrs->no_numa);
5013 mutex_unlock(&wq->mutex);
5018 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5019 const char *buf, size_t count)
5021 struct workqueue_struct *wq = dev_to_wq(dev);
5022 struct workqueue_attrs *attrs;
5025 attrs = wq_sysfs_prep_attrs(wq);
5030 if (sscanf(buf, "%d", &v) == 1) {
5031 attrs->no_numa = !v;
5032 ret = apply_workqueue_attrs(wq, attrs);
5035 free_workqueue_attrs(attrs);
5036 return ret ?: count;
5039 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5040 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5041 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5042 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5043 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5047 static struct bus_type wq_subsys = {
5048 .name = "workqueue",
5049 .dev_groups = wq_sysfs_groups,
5052 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5053 struct device_attribute *attr, char *buf)
5057 mutex_lock(&wq_pool_mutex);
5058 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5059 cpumask_pr_args(wq_unbound_cpumask));
5060 mutex_unlock(&wq_pool_mutex);
5065 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5066 struct device_attribute *attr, const char *buf, size_t count)
5068 cpumask_var_t cpumask;
5071 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5074 ret = cpumask_parse(buf, cpumask);
5076 ret = workqueue_set_unbound_cpumask(cpumask);
5078 free_cpumask_var(cpumask);
5079 return ret ? ret : count;
5082 static struct device_attribute wq_sysfs_cpumask_attr =
5083 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5084 wq_unbound_cpumask_store);
5086 static int __init wq_sysfs_init(void)
5090 err = subsys_virtual_register(&wq_subsys, NULL);
5094 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5096 core_initcall(wq_sysfs_init);
5098 static void wq_device_release(struct device *dev)
5100 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5106 * workqueue_sysfs_register - make a workqueue visible in sysfs
5107 * @wq: the workqueue to register
5109 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5110 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5111 * which is the preferred method.
5113 * Workqueue user should use this function directly iff it wants to apply
5114 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5115 * apply_workqueue_attrs() may race against userland updating the
5118 * Return: 0 on success, -errno on failure.
5120 int workqueue_sysfs_register(struct workqueue_struct *wq)
5122 struct wq_device *wq_dev;
5126 * Adjusting max_active or creating new pwqs by applyting
5127 * attributes breaks ordering guarantee. Disallow exposing ordered
5130 if (WARN_ON(wq->flags & __WQ_ORDERED))
5133 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5138 wq_dev->dev.bus = &wq_subsys;
5139 wq_dev->dev.init_name = wq->name;
5140 wq_dev->dev.release = wq_device_release;
5143 * unbound_attrs are created separately. Suppress uevent until
5144 * everything is ready.
5146 dev_set_uevent_suppress(&wq_dev->dev, true);
5148 ret = device_register(&wq_dev->dev);
5155 if (wq->flags & WQ_UNBOUND) {
5156 struct device_attribute *attr;
5158 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5159 ret = device_create_file(&wq_dev->dev, attr);
5161 device_unregister(&wq_dev->dev);
5168 dev_set_uevent_suppress(&wq_dev->dev, false);
5169 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5174 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5175 * @wq: the workqueue to unregister
5177 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5179 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5181 struct wq_device *wq_dev = wq->wq_dev;
5187 device_unregister(&wq_dev->dev);
5189 #else /* CONFIG_SYSFS */
5190 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5191 #endif /* CONFIG_SYSFS */
5193 static void __init wq_numa_init(void)
5198 if (num_possible_nodes() <= 1)
5201 if (wq_disable_numa) {
5202 pr_info("workqueue: NUMA affinity support disabled\n");
5206 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5207 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5210 * We want masks of possible CPUs of each node which isn't readily
5211 * available. Build one from cpu_to_node() which should have been
5212 * fully initialized by now.
5214 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5218 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5219 node_online(node) ? node : NUMA_NO_NODE));
5221 for_each_possible_cpu(cpu) {
5222 node = cpu_to_node(cpu);
5223 if (WARN_ON(node == NUMA_NO_NODE)) {
5224 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5225 /* happens iff arch is bonkers, let's just proceed */
5228 cpumask_set_cpu(cpu, tbl[node]);
5231 wq_numa_possible_cpumask = tbl;
5232 wq_numa_enabled = true;
5235 static int __init init_workqueues(void)
5237 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5240 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5242 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5243 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5245 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5247 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5248 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5252 /* initialize CPU pools */
5253 for_each_possible_cpu(cpu) {
5254 struct worker_pool *pool;
5257 for_each_cpu_worker_pool(pool, cpu) {
5258 BUG_ON(init_worker_pool(pool));
5260 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5261 pool->attrs->nice = std_nice[i++];
5262 pool->node = cpu_to_node(cpu);
5265 mutex_lock(&wq_pool_mutex);
5266 BUG_ON(worker_pool_assign_id(pool));
5267 mutex_unlock(&wq_pool_mutex);
5271 /* create the initial worker */
5272 for_each_online_cpu(cpu) {
5273 struct worker_pool *pool;
5275 for_each_cpu_worker_pool(pool, cpu) {
5276 pool->flags &= ~POOL_DISASSOCIATED;
5277 BUG_ON(!create_worker(pool));
5281 /* create default unbound and ordered wq attrs */
5282 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5283 struct workqueue_attrs *attrs;
5285 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5286 attrs->nice = std_nice[i];
5287 unbound_std_wq_attrs[i] = attrs;
5290 * An ordered wq should have only one pwq as ordering is
5291 * guaranteed by max_active which is enforced by pwqs.
5292 * Turn off NUMA so that dfl_pwq is used for all nodes.
5294 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5295 attrs->nice = std_nice[i];
5296 attrs->no_numa = true;
5297 ordered_wq_attrs[i] = attrs;
5300 system_wq = alloc_workqueue("events", 0, 0);
5301 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5302 system_long_wq = alloc_workqueue("events_long", 0, 0);
5303 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5304 WQ_UNBOUND_MAX_ACTIVE);
5305 system_freezable_wq = alloc_workqueue("events_freezable",
5307 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5308 WQ_POWER_EFFICIENT, 0);
5309 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5310 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5312 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5313 !system_unbound_wq || !system_freezable_wq ||
5314 !system_power_efficient_wq ||
5315 !system_freezable_power_efficient_wq);
5318 early_initcall(init_workqueues);