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 is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
51 #include "workqueue_internal.h"
57 * A bound pool is either associated or disassociated with its CPU.
58 * While associated (!DISASSOCIATED), all workers are bound to the
59 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * While DISASSOCIATED, the cpu may be offline and all workers have
63 * %WORKER_UNBOUND set and concurrency management disabled, and may
64 * be executing on any CPU. The pool behaves as an unbound one.
66 * Note that DISASSOCIATED should be flipped only while holding
67 * manager_mutex to avoid changing binding state while
68 * create_worker() is in progress.
70 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72 POOL_FREEZING = 1 << 3, /* freeze in progress */
75 WORKER_STARTED = 1 << 0, /* started */
76 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
78 WORKER_PREP = 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give -20.
104 RESCUER_NICE_LEVEL = -20,
105 HIGHPRI_NICE_LEVEL = -20,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * MG: pool->manager_mutex and pool->lock protected. Writes require both
127 * locks. Reads can happen under either lock.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * WQ: wq->mutex protected.
135 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
137 * MD: wq_mayday_lock protected.
140 /* struct worker is defined in workqueue_internal.h */
143 spinlock_t lock; /* the pool lock */
144 int cpu; /* I: the associated cpu */
145 int node; /* I: the associated node ID */
146 int id; /* I: pool ID */
147 unsigned int flags; /* X: flags */
149 struct list_head worklist; /* L: list of pending works */
150 int nr_workers; /* L: total number of workers */
152 /* nr_idle includes the ones off idle_list for rebinding */
153 int nr_idle; /* L: currently idle ones */
155 struct list_head idle_list; /* X: list of idle workers */
156 struct timer_list idle_timer; /* L: worker idle timeout */
157 struct timer_list mayday_timer; /* L: SOS timer for workers */
159 /* a workers is either on busy_hash or idle_list, or the manager */
160 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
161 /* L: hash of busy workers */
163 /* see manage_workers() for details on the two manager mutexes */
164 struct mutex manager_arb; /* manager arbitration */
165 struct mutex manager_mutex; /* manager exclusion */
166 struct idr worker_idr; /* MG: worker IDs and iteration */
168 struct workqueue_attrs *attrs; /* I: worker attributes */
169 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
170 int refcnt; /* PL: refcnt for unbound pools */
173 * The current concurrency level. As it's likely to be accessed
174 * from other CPUs during try_to_wake_up(), put it in a separate
177 atomic_t nr_running ____cacheline_aligned_in_smp;
180 * Destruction of pool is sched-RCU protected to allow dereferences
181 * from get_work_pool().
184 } ____cacheline_aligned_in_smp;
187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 * of work_struct->data are used for flags and the remaining high bits
189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 * number of flag bits.
192 struct pool_workqueue {
193 struct worker_pool *pool; /* I: the associated pool */
194 struct workqueue_struct *wq; /* I: the owning workqueue */
195 int work_color; /* L: current color */
196 int flush_color; /* L: flushing color */
197 int refcnt; /* L: reference count */
198 int nr_in_flight[WORK_NR_COLORS];
199 /* L: nr of in_flight works */
200 int nr_active; /* L: nr of active works */
201 int max_active; /* L: max active works */
202 struct list_head delayed_works; /* L: delayed works */
203 struct list_head pwqs_node; /* WR: node on wq->pwqs */
204 struct list_head mayday_node; /* MD: node on wq->maydays */
207 * Release of unbound pwq is punted to system_wq. See put_pwq()
208 * and pwq_unbound_release_workfn() for details. pool_workqueue
209 * itself is also sched-RCU protected so that the first pwq can be
210 * determined without grabbing wq->mutex.
212 struct work_struct unbound_release_work;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
217 * Structure used to wait for workqueue flush.
220 struct list_head list; /* WQ: list of flushers */
221 int flush_color; /* WQ: flush color waiting for */
222 struct completion done; /* flush completion */
228 * The externally visible workqueue. It relays the issued work items to
229 * the appropriate worker_pool through its pool_workqueues.
231 struct workqueue_struct {
232 struct list_head pwqs; /* WR: all pwqs of this wq */
233 struct list_head list; /* PL: list of all workqueues */
235 struct mutex mutex; /* protects this wq */
236 int work_color; /* WQ: current work color */
237 int flush_color; /* WQ: current flush color */
238 atomic_t nr_pwqs_to_flush; /* flush in progress */
239 struct wq_flusher *first_flusher; /* WQ: first flusher */
240 struct list_head flusher_queue; /* WQ: flush waiters */
241 struct list_head flusher_overflow; /* WQ: flush overflow list */
243 struct list_head maydays; /* MD: pwqs requesting rescue */
244 struct worker *rescuer; /* I: rescue worker */
246 int nr_drainers; /* WQ: drain in progress */
247 int saved_max_active; /* WQ: saved pwq max_active */
249 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
253 struct wq_device *wq_dev; /* I: for sysfs interface */
255 #ifdef CONFIG_LOCKDEP
256 struct lockdep_map lockdep_map;
258 char name[WQ_NAME_LEN]; /* I: workqueue name */
260 /* hot fields used during command issue, aligned to cacheline */
261 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
262 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
263 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
266 static struct kmem_cache *pwq_cache;
268 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
269 static cpumask_var_t *wq_numa_possible_cpumask;
270 /* possible CPUs of each node */
272 static bool wq_disable_numa;
273 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
275 /* see the comment above the definition of WQ_POWER_EFFICIENT */
276 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
277 static bool wq_power_efficient = true;
279 static bool wq_power_efficient;
282 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
284 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
286 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
287 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
289 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
290 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
292 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
293 static bool workqueue_freezing; /* PL: have wqs started freezing? */
295 /* the per-cpu worker pools */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
299 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
301 /* PL: hash of all unbound pools keyed by pool->attrs */
302 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
304 /* I: attributes used when instantiating standard unbound pools on demand */
305 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
307 struct workqueue_struct *system_wq __read_mostly;
308 EXPORT_SYMBOL(system_wq);
309 struct workqueue_struct *system_highpri_wq __read_mostly;
310 EXPORT_SYMBOL_GPL(system_highpri_wq);
311 struct workqueue_struct *system_long_wq __read_mostly;
312 EXPORT_SYMBOL_GPL(system_long_wq);
313 struct workqueue_struct *system_unbound_wq __read_mostly;
314 EXPORT_SYMBOL_GPL(system_unbound_wq);
315 struct workqueue_struct *system_freezable_wq __read_mostly;
316 EXPORT_SYMBOL_GPL(system_freezable_wq);
318 static int worker_thread(void *__worker);
319 static void copy_workqueue_attrs(struct workqueue_attrs *to,
320 const struct workqueue_attrs *from);
322 #define CREATE_TRACE_POINTS
323 #include <trace/events/workqueue.h>
325 #define assert_rcu_or_pool_mutex() \
326 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
327 lockdep_is_held(&wq_pool_mutex), \
328 "sched RCU or wq_pool_mutex should be held")
330 #define assert_rcu_or_wq_mutex(wq) \
331 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
332 lockdep_is_held(&wq->mutex), \
333 "sched RCU or wq->mutex should be held")
335 #ifdef CONFIG_LOCKDEP
336 #define assert_manager_or_pool_lock(pool) \
337 WARN_ONCE(debug_locks && \
338 !lockdep_is_held(&(pool)->manager_mutex) && \
339 !lockdep_is_held(&(pool)->lock), \
340 "pool->manager_mutex or ->lock should be held")
342 #define assert_manager_or_pool_lock(pool) do { } while (0)
345 #define for_each_cpu_worker_pool(pool, cpu) \
346 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
347 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
351 * for_each_pool - iterate through all worker_pools in the system
352 * @pool: iteration cursor
353 * @pi: integer used for iteration
355 * This must be called either with wq_pool_mutex held or sched RCU read
356 * locked. If the pool needs to be used beyond the locking in effect, the
357 * caller is responsible for guaranteeing that the pool stays online.
359 * The if/else clause exists only for the lockdep assertion and can be
362 #define for_each_pool(pool, pi) \
363 idr_for_each_entry(&worker_pool_idr, pool, pi) \
364 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
368 * for_each_pool_worker - iterate through all workers of a worker_pool
369 * @worker: iteration cursor
370 * @wi: integer used for iteration
371 * @pool: worker_pool to iterate workers of
373 * This must be called with either @pool->manager_mutex or ->lock held.
375 * The if/else clause exists only for the lockdep assertion and can be
378 #define for_each_pool_worker(worker, wi, pool) \
379 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
380 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
384 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
385 * @pwq: iteration cursor
386 * @wq: the target workqueue
388 * This must be called either with wq->mutex held or sched RCU read locked.
389 * If the pwq needs to be used beyond the locking in effect, the caller is
390 * responsible for guaranteeing that the pwq stays online.
392 * The if/else clause exists only for the lockdep assertion and can be
395 #define for_each_pwq(pwq, wq) \
396 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
397 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
400 #ifdef CONFIG_DEBUG_OBJECTS_WORK
402 static struct debug_obj_descr work_debug_descr;
404 static void *work_debug_hint(void *addr)
406 return ((struct work_struct *) addr)->func;
410 * fixup_init is called when:
411 * - an active object is initialized
413 static int work_fixup_init(void *addr, enum debug_obj_state state)
415 struct work_struct *work = addr;
418 case ODEBUG_STATE_ACTIVE:
419 cancel_work_sync(work);
420 debug_object_init(work, &work_debug_descr);
428 * fixup_activate is called when:
429 * - an active object is activated
430 * - an unknown object is activated (might be a statically initialized object)
432 static int work_fixup_activate(void *addr, enum debug_obj_state state)
434 struct work_struct *work = addr;
438 case ODEBUG_STATE_NOTAVAILABLE:
440 * This is not really a fixup. The work struct was
441 * statically initialized. We just make sure that it
442 * is tracked in the object tracker.
444 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
445 debug_object_init(work, &work_debug_descr);
446 debug_object_activate(work, &work_debug_descr);
452 case ODEBUG_STATE_ACTIVE:
461 * fixup_free is called when:
462 * - an active object is freed
464 static int work_fixup_free(void *addr, enum debug_obj_state state)
466 struct work_struct *work = addr;
469 case ODEBUG_STATE_ACTIVE:
470 cancel_work_sync(work);
471 debug_object_free(work, &work_debug_descr);
478 static struct debug_obj_descr work_debug_descr = {
479 .name = "work_struct",
480 .debug_hint = work_debug_hint,
481 .fixup_init = work_fixup_init,
482 .fixup_activate = work_fixup_activate,
483 .fixup_free = work_fixup_free,
486 static inline void debug_work_activate(struct work_struct *work)
488 debug_object_activate(work, &work_debug_descr);
491 static inline void debug_work_deactivate(struct work_struct *work)
493 debug_object_deactivate(work, &work_debug_descr);
496 void __init_work(struct work_struct *work, int onstack)
499 debug_object_init_on_stack(work, &work_debug_descr);
501 debug_object_init(work, &work_debug_descr);
503 EXPORT_SYMBOL_GPL(__init_work);
505 void destroy_work_on_stack(struct work_struct *work)
507 debug_object_free(work, &work_debug_descr);
509 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
512 static inline void debug_work_activate(struct work_struct *work) { }
513 static inline void debug_work_deactivate(struct work_struct *work) { }
516 /* allocate ID and assign it to @pool */
517 static int worker_pool_assign_id(struct worker_pool *pool)
521 lockdep_assert_held(&wq_pool_mutex);
523 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
532 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
533 * @wq: the target workqueue
536 * This must be called either with pwq_lock held or sched RCU read locked.
537 * If the pwq needs to be used beyond the locking in effect, the caller is
538 * responsible for guaranteeing that the pwq stays online.
540 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
543 assert_rcu_or_wq_mutex(wq);
544 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
547 static unsigned int work_color_to_flags(int color)
549 return color << WORK_STRUCT_COLOR_SHIFT;
552 static int get_work_color(struct work_struct *work)
554 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
555 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
558 static int work_next_color(int color)
560 return (color + 1) % WORK_NR_COLORS;
564 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
565 * contain the pointer to the queued pwq. Once execution starts, the flag
566 * is cleared and the high bits contain OFFQ flags and pool ID.
568 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
569 * and clear_work_data() can be used to set the pwq, pool or clear
570 * work->data. These functions should only be called while the work is
571 * owned - ie. while the PENDING bit is set.
573 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
574 * corresponding to a work. Pool is available once the work has been
575 * queued anywhere after initialization until it is sync canceled. pwq is
576 * available only while the work item is queued.
578 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
579 * canceled. While being canceled, a work item may have its PENDING set
580 * but stay off timer and worklist for arbitrarily long and nobody should
581 * try to steal the PENDING bit.
583 static inline void set_work_data(struct work_struct *work, unsigned long data,
586 WARN_ON_ONCE(!work_pending(work));
587 atomic_long_set(&work->data, data | flags | work_static(work));
590 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
591 unsigned long extra_flags)
593 set_work_data(work, (unsigned long)pwq,
594 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
597 static void set_work_pool_and_keep_pending(struct work_struct *work,
600 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
601 WORK_STRUCT_PENDING);
604 static void set_work_pool_and_clear_pending(struct work_struct *work,
608 * The following wmb is paired with the implied mb in
609 * test_and_set_bit(PENDING) and ensures all updates to @work made
610 * here are visible to and precede any updates by the next PENDING
614 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
617 static void clear_work_data(struct work_struct *work)
619 smp_wmb(); /* see set_work_pool_and_clear_pending() */
620 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
623 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
625 unsigned long data = atomic_long_read(&work->data);
627 if (data & WORK_STRUCT_PWQ)
628 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
634 * get_work_pool - return the worker_pool a given work was associated with
635 * @work: the work item of interest
637 * Return the worker_pool @work was last associated with. %NULL if none.
639 * Pools are created and destroyed under wq_pool_mutex, and allows read
640 * access under sched-RCU read lock. As such, this function should be
641 * called under wq_pool_mutex or with preemption disabled.
643 * All fields of the returned pool are accessible as long as the above
644 * mentioned locking is in effect. If the returned pool needs to be used
645 * beyond the critical section, the caller is responsible for ensuring the
646 * returned pool is and stays online.
648 static struct worker_pool *get_work_pool(struct work_struct *work)
650 unsigned long data = atomic_long_read(&work->data);
653 assert_rcu_or_pool_mutex();
655 if (data & WORK_STRUCT_PWQ)
656 return ((struct pool_workqueue *)
657 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
659 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
660 if (pool_id == WORK_OFFQ_POOL_NONE)
663 return idr_find(&worker_pool_idr, pool_id);
667 * get_work_pool_id - return the worker pool ID a given work is associated with
668 * @work: the work item of interest
670 * Return the worker_pool ID @work was last associated with.
671 * %WORK_OFFQ_POOL_NONE if none.
673 static int get_work_pool_id(struct work_struct *work)
675 unsigned long data = atomic_long_read(&work->data);
677 if (data & WORK_STRUCT_PWQ)
678 return ((struct pool_workqueue *)
679 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
681 return data >> WORK_OFFQ_POOL_SHIFT;
684 static void mark_work_canceling(struct work_struct *work)
686 unsigned long pool_id = get_work_pool_id(work);
688 pool_id <<= WORK_OFFQ_POOL_SHIFT;
689 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
692 static bool work_is_canceling(struct work_struct *work)
694 unsigned long data = atomic_long_read(&work->data);
696 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
700 * Policy functions. These define the policies on how the global worker
701 * pools are managed. Unless noted otherwise, these functions assume that
702 * they're being called with pool->lock held.
705 static bool __need_more_worker(struct worker_pool *pool)
707 return !atomic_read(&pool->nr_running);
711 * Need to wake up a worker? Called from anything but currently
714 * Note that, because unbound workers never contribute to nr_running, this
715 * function will always return %true for unbound pools as long as the
716 * worklist isn't empty.
718 static bool need_more_worker(struct worker_pool *pool)
720 return !list_empty(&pool->worklist) && __need_more_worker(pool);
723 /* Can I start working? Called from busy but !running workers. */
724 static bool may_start_working(struct worker_pool *pool)
726 return pool->nr_idle;
729 /* Do I need to keep working? Called from currently running workers. */
730 static bool keep_working(struct worker_pool *pool)
732 return !list_empty(&pool->worklist) &&
733 atomic_read(&pool->nr_running) <= 1;
736 /* Do we need a new worker? Called from manager. */
737 static bool need_to_create_worker(struct worker_pool *pool)
739 return need_more_worker(pool) && !may_start_working(pool);
742 /* Do I need to be the manager? */
743 static bool need_to_manage_workers(struct worker_pool *pool)
745 return need_to_create_worker(pool) ||
746 (pool->flags & POOL_MANAGE_WORKERS);
749 /* Do we have too many workers and should some go away? */
750 static bool too_many_workers(struct worker_pool *pool)
752 bool managing = mutex_is_locked(&pool->manager_arb);
753 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
754 int nr_busy = pool->nr_workers - nr_idle;
757 * nr_idle and idle_list may disagree if idle rebinding is in
758 * progress. Never return %true if idle_list is empty.
760 if (list_empty(&pool->idle_list))
763 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
770 /* Return the first worker. Safe with preemption disabled */
771 static struct worker *first_worker(struct worker_pool *pool)
773 if (unlikely(list_empty(&pool->idle_list)))
776 return list_first_entry(&pool->idle_list, struct worker, entry);
780 * wake_up_worker - wake up an idle worker
781 * @pool: worker pool to wake worker from
783 * Wake up the first idle worker of @pool.
786 * spin_lock_irq(pool->lock).
788 static void wake_up_worker(struct worker_pool *pool)
790 struct worker *worker = first_worker(pool);
793 wake_up_process(worker->task);
797 * wq_worker_waking_up - a worker is waking up
798 * @task: task waking up
799 * @cpu: CPU @task is waking up to
801 * This function is called during try_to_wake_up() when a worker is
805 * spin_lock_irq(rq->lock)
807 void wq_worker_waking_up(struct task_struct *task, int cpu)
809 struct worker *worker = kthread_data(task);
811 if (!(worker->flags & WORKER_NOT_RUNNING)) {
812 WARN_ON_ONCE(worker->pool->cpu != cpu);
813 atomic_inc(&worker->pool->nr_running);
818 * wq_worker_sleeping - a worker is going to sleep
819 * @task: task going to sleep
820 * @cpu: CPU in question, must be the current CPU number
822 * This function is called during schedule() when a busy worker is
823 * going to sleep. Worker on the same cpu can be woken up by
824 * returning pointer to its task.
827 * spin_lock_irq(rq->lock)
830 * Worker task on @cpu to wake up, %NULL if none.
832 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
834 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
835 struct worker_pool *pool;
838 * Rescuers, which may not have all the fields set up like normal
839 * workers, also reach here, let's not access anything before
840 * checking NOT_RUNNING.
842 if (worker->flags & WORKER_NOT_RUNNING)
847 /* this can only happen on the local cpu */
848 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
852 * The counterpart of the following dec_and_test, implied mb,
853 * worklist not empty test sequence is in insert_work().
854 * Please read comment there.
856 * NOT_RUNNING is clear. This means that we're bound to and
857 * running on the local cpu w/ rq lock held and preemption
858 * disabled, which in turn means that none else could be
859 * manipulating idle_list, so dereferencing idle_list without pool
862 if (atomic_dec_and_test(&pool->nr_running) &&
863 !list_empty(&pool->worklist))
864 to_wakeup = first_worker(pool);
865 return to_wakeup ? to_wakeup->task : NULL;
869 * worker_set_flags - set worker flags and adjust nr_running accordingly
871 * @flags: flags to set
872 * @wakeup: wakeup an idle worker if necessary
874 * Set @flags in @worker->flags and adjust nr_running accordingly. If
875 * nr_running becomes zero and @wakeup is %true, an idle worker is
879 * spin_lock_irq(pool->lock)
881 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
884 struct worker_pool *pool = worker->pool;
886 WARN_ON_ONCE(worker->task != current);
889 * If transitioning into NOT_RUNNING, adjust nr_running and
890 * wake up an idle worker as necessary if requested by
893 if ((flags & WORKER_NOT_RUNNING) &&
894 !(worker->flags & WORKER_NOT_RUNNING)) {
896 if (atomic_dec_and_test(&pool->nr_running) &&
897 !list_empty(&pool->worklist))
898 wake_up_worker(pool);
900 atomic_dec(&pool->nr_running);
903 worker->flags |= flags;
907 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
909 * @flags: flags to clear
911 * Clear @flags in @worker->flags and adjust nr_running accordingly.
914 * spin_lock_irq(pool->lock)
916 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
918 struct worker_pool *pool = worker->pool;
919 unsigned int oflags = worker->flags;
921 WARN_ON_ONCE(worker->task != current);
923 worker->flags &= ~flags;
926 * If transitioning out of NOT_RUNNING, increment nr_running. Note
927 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
928 * of multiple flags, not a single flag.
930 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
931 if (!(worker->flags & WORKER_NOT_RUNNING))
932 atomic_inc(&pool->nr_running);
936 * find_worker_executing_work - find worker which is executing a work
937 * @pool: pool of interest
938 * @work: work to find worker for
940 * Find a worker which is executing @work on @pool by searching
941 * @pool->busy_hash which is keyed by the address of @work. For a worker
942 * to match, its current execution should match the address of @work and
943 * its work function. This is to avoid unwanted dependency between
944 * unrelated work executions through a work item being recycled while still
947 * This is a bit tricky. A work item may be freed once its execution
948 * starts and nothing prevents the freed area from being recycled for
949 * another work item. If the same work item address ends up being reused
950 * before the original execution finishes, workqueue will identify the
951 * recycled work item as currently executing and make it wait until the
952 * current execution finishes, introducing an unwanted dependency.
954 * This function checks the work item address and work function to avoid
955 * false positives. Note that this isn't complete as one may construct a
956 * work function which can introduce dependency onto itself through a
957 * recycled work item. Well, if somebody wants to shoot oneself in the
958 * foot that badly, there's only so much we can do, and if such deadlock
959 * actually occurs, it should be easy to locate the culprit work function.
962 * spin_lock_irq(pool->lock).
965 * Pointer to worker which is executing @work if found, NULL
968 static struct worker *find_worker_executing_work(struct worker_pool *pool,
969 struct work_struct *work)
971 struct worker *worker;
973 hash_for_each_possible(pool->busy_hash, worker, hentry,
975 if (worker->current_work == work &&
976 worker->current_func == work->func)
983 * move_linked_works - move linked works to a list
984 * @work: start of series of works to be scheduled
985 * @head: target list to append @work to
986 * @nextp: out paramter for nested worklist walking
988 * Schedule linked works starting from @work to @head. Work series to
989 * be scheduled starts at @work and includes any consecutive work with
990 * WORK_STRUCT_LINKED set in its predecessor.
992 * If @nextp is not NULL, it's updated to point to the next work of
993 * the last scheduled work. This allows move_linked_works() to be
994 * nested inside outer list_for_each_entry_safe().
997 * spin_lock_irq(pool->lock).
999 static void move_linked_works(struct work_struct *work, struct list_head *head,
1000 struct work_struct **nextp)
1002 struct work_struct *n;
1005 * Linked worklist will always end before the end of the list,
1006 * use NULL for list head.
1008 list_for_each_entry_safe_from(work, n, NULL, entry) {
1009 list_move_tail(&work->entry, head);
1010 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1015 * If we're already inside safe list traversal and have moved
1016 * multiple works to the scheduled queue, the next position
1017 * needs to be updated.
1024 * get_pwq - get an extra reference on the specified pool_workqueue
1025 * @pwq: pool_workqueue to get
1027 * Obtain an extra reference on @pwq. The caller should guarantee that
1028 * @pwq has positive refcnt and be holding the matching pool->lock.
1030 static void get_pwq(struct pool_workqueue *pwq)
1032 lockdep_assert_held(&pwq->pool->lock);
1033 WARN_ON_ONCE(pwq->refcnt <= 0);
1038 * put_pwq - put a pool_workqueue reference
1039 * @pwq: pool_workqueue to put
1041 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1042 * destruction. The caller should be holding the matching pool->lock.
1044 static void put_pwq(struct pool_workqueue *pwq)
1046 lockdep_assert_held(&pwq->pool->lock);
1047 if (likely(--pwq->refcnt))
1049 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1052 * @pwq can't be released under pool->lock, bounce to
1053 * pwq_unbound_release_workfn(). This never recurses on the same
1054 * pool->lock as this path is taken only for unbound workqueues and
1055 * the release work item is scheduled on a per-cpu workqueue. To
1056 * avoid lockdep warning, unbound pool->locks are given lockdep
1057 * subclass of 1 in get_unbound_pool().
1059 schedule_work(&pwq->unbound_release_work);
1063 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1064 * @pwq: pool_workqueue to put (can be %NULL)
1066 * put_pwq() with locking. This function also allows %NULL @pwq.
1068 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1072 * As both pwqs and pools are sched-RCU protected, the
1073 * following lock operations are safe.
1075 spin_lock_irq(&pwq->pool->lock);
1077 spin_unlock_irq(&pwq->pool->lock);
1081 static void pwq_activate_delayed_work(struct work_struct *work)
1083 struct pool_workqueue *pwq = get_work_pwq(work);
1085 trace_workqueue_activate_work(work);
1086 move_linked_works(work, &pwq->pool->worklist, NULL);
1087 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1091 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1093 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1094 struct work_struct, entry);
1096 pwq_activate_delayed_work(work);
1100 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1101 * @pwq: pwq of interest
1102 * @color: color of work which left the queue
1104 * A work either has completed or is removed from pending queue,
1105 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1108 * spin_lock_irq(pool->lock).
1110 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1112 /* uncolored work items don't participate in flushing or nr_active */
1113 if (color == WORK_NO_COLOR)
1116 pwq->nr_in_flight[color]--;
1119 if (!list_empty(&pwq->delayed_works)) {
1120 /* one down, submit a delayed one */
1121 if (pwq->nr_active < pwq->max_active)
1122 pwq_activate_first_delayed(pwq);
1125 /* is flush in progress and are we at the flushing tip? */
1126 if (likely(pwq->flush_color != color))
1129 /* are there still in-flight works? */
1130 if (pwq->nr_in_flight[color])
1133 /* this pwq is done, clear flush_color */
1134 pwq->flush_color = -1;
1137 * If this was the last pwq, wake up the first flusher. It
1138 * will handle the rest.
1140 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1141 complete(&pwq->wq->first_flusher->done);
1147 * try_to_grab_pending - steal work item from worklist and disable irq
1148 * @work: work item to steal
1149 * @is_dwork: @work is a delayed_work
1150 * @flags: place to store irq state
1152 * Try to grab PENDING bit of @work. This function can handle @work in any
1153 * stable state - idle, on timer or on worklist. Return values are
1155 * 1 if @work was pending and we successfully stole PENDING
1156 * 0 if @work was idle and we claimed PENDING
1157 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1158 * -ENOENT if someone else is canceling @work, this state may persist
1159 * for arbitrarily long
1161 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1162 * interrupted while holding PENDING and @work off queue, irq must be
1163 * disabled on entry. This, combined with delayed_work->timer being
1164 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1166 * On successful return, >= 0, irq is disabled and the caller is
1167 * responsible for releasing it using local_irq_restore(*@flags).
1169 * This function is safe to call from any context including IRQ handler.
1171 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1172 unsigned long *flags)
1174 struct worker_pool *pool;
1175 struct pool_workqueue *pwq;
1177 local_irq_save(*flags);
1179 /* try to steal the timer if it exists */
1181 struct delayed_work *dwork = to_delayed_work(work);
1184 * dwork->timer is irqsafe. If del_timer() fails, it's
1185 * guaranteed that the timer is not queued anywhere and not
1186 * running on the local CPU.
1188 if (likely(del_timer(&dwork->timer)))
1192 /* try to claim PENDING the normal way */
1193 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1197 * The queueing is in progress, or it is already queued. Try to
1198 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1200 pool = get_work_pool(work);
1204 spin_lock(&pool->lock);
1206 * work->data is guaranteed to point to pwq only while the work
1207 * item is queued on pwq->wq, and both updating work->data to point
1208 * to pwq on queueing and to pool on dequeueing are done under
1209 * pwq->pool->lock. This in turn guarantees that, if work->data
1210 * points to pwq which is associated with a locked pool, the work
1211 * item is currently queued on that pool.
1213 pwq = get_work_pwq(work);
1214 if (pwq && pwq->pool == pool) {
1215 debug_work_deactivate(work);
1218 * A delayed work item cannot be grabbed directly because
1219 * it might have linked NO_COLOR work items which, if left
1220 * on the delayed_list, will confuse pwq->nr_active
1221 * management later on and cause stall. Make sure the work
1222 * item is activated before grabbing.
1224 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1225 pwq_activate_delayed_work(work);
1227 list_del_init(&work->entry);
1228 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1230 /* work->data points to pwq iff queued, point to pool */
1231 set_work_pool_and_keep_pending(work, pool->id);
1233 spin_unlock(&pool->lock);
1236 spin_unlock(&pool->lock);
1238 local_irq_restore(*flags);
1239 if (work_is_canceling(work))
1246 * insert_work - insert a work into a pool
1247 * @pwq: pwq @work belongs to
1248 * @work: work to insert
1249 * @head: insertion point
1250 * @extra_flags: extra WORK_STRUCT_* flags to set
1252 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1253 * work_struct flags.
1256 * spin_lock_irq(pool->lock).
1258 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1259 struct list_head *head, unsigned int extra_flags)
1261 struct worker_pool *pool = pwq->pool;
1263 /* we own @work, set data and link */
1264 set_work_pwq(work, pwq, extra_flags);
1265 list_add_tail(&work->entry, head);
1269 * Ensure either wq_worker_sleeping() sees the above
1270 * list_add_tail() or we see zero nr_running to avoid workers lying
1271 * around lazily while there are works to be processed.
1275 if (__need_more_worker(pool))
1276 wake_up_worker(pool);
1280 * Test whether @work is being queued from another work executing on the
1283 static bool is_chained_work(struct workqueue_struct *wq)
1285 struct worker *worker;
1287 worker = current_wq_worker();
1289 * Return %true iff I'm a worker execuing a work item on @wq. If
1290 * I'm @worker, it's safe to dereference it without locking.
1292 return worker && worker->current_pwq->wq == wq;
1295 static void __queue_work(int cpu, struct workqueue_struct *wq,
1296 struct work_struct *work)
1298 struct pool_workqueue *pwq;
1299 struct worker_pool *last_pool;
1300 struct list_head *worklist;
1301 unsigned int work_flags;
1302 unsigned int req_cpu = cpu;
1305 * While a work item is PENDING && off queue, a task trying to
1306 * steal the PENDING will busy-loop waiting for it to either get
1307 * queued or lose PENDING. Grabbing PENDING and queueing should
1308 * happen with IRQ disabled.
1310 WARN_ON_ONCE(!irqs_disabled());
1312 debug_work_activate(work);
1314 /* if dying, only works from the same workqueue are allowed */
1315 if (unlikely(wq->flags & __WQ_DRAINING) &&
1316 WARN_ON_ONCE(!is_chained_work(wq)))
1319 if (req_cpu == WORK_CPU_UNBOUND)
1320 cpu = raw_smp_processor_id();
1322 /* pwq which will be used unless @work is executing elsewhere */
1323 if (!(wq->flags & WQ_UNBOUND))
1324 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1326 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1329 * If @work was previously on a different pool, it might still be
1330 * running there, in which case the work needs to be queued on that
1331 * pool to guarantee non-reentrancy.
1333 last_pool = get_work_pool(work);
1334 if (last_pool && last_pool != pwq->pool) {
1335 struct worker *worker;
1337 spin_lock(&last_pool->lock);
1339 worker = find_worker_executing_work(last_pool, work);
1341 if (worker && worker->current_pwq->wq == wq) {
1342 pwq = worker->current_pwq;
1344 /* meh... not running there, queue here */
1345 spin_unlock(&last_pool->lock);
1346 spin_lock(&pwq->pool->lock);
1349 spin_lock(&pwq->pool->lock);
1353 * pwq is determined and locked. For unbound pools, we could have
1354 * raced with pwq release and it could already be dead. If its
1355 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1356 * without another pwq replacing it in the numa_pwq_tbl or while
1357 * work items are executing on it, so the retrying is guaranteed to
1358 * make forward-progress.
1360 if (unlikely(!pwq->refcnt)) {
1361 if (wq->flags & WQ_UNBOUND) {
1362 spin_unlock(&pwq->pool->lock);
1367 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1371 /* pwq determined, queue */
1372 trace_workqueue_queue_work(req_cpu, pwq, work);
1374 if (WARN_ON(!list_empty(&work->entry))) {
1375 spin_unlock(&pwq->pool->lock);
1379 pwq->nr_in_flight[pwq->work_color]++;
1380 work_flags = work_color_to_flags(pwq->work_color);
1382 if (likely(pwq->nr_active < pwq->max_active)) {
1383 trace_workqueue_activate_work(work);
1385 worklist = &pwq->pool->worklist;
1387 work_flags |= WORK_STRUCT_DELAYED;
1388 worklist = &pwq->delayed_works;
1391 insert_work(pwq, work, worklist, work_flags);
1393 spin_unlock(&pwq->pool->lock);
1397 * queue_work_on - queue work on specific cpu
1398 * @cpu: CPU number to execute work on
1399 * @wq: workqueue to use
1400 * @work: work to queue
1402 * Returns %false if @work was already on a queue, %true otherwise.
1404 * We queue the work to a specific CPU, the caller must ensure it
1407 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1408 struct work_struct *work)
1411 unsigned long flags;
1413 local_irq_save(flags);
1415 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1416 __queue_work(cpu, wq, work);
1420 local_irq_restore(flags);
1423 EXPORT_SYMBOL(queue_work_on);
1425 void delayed_work_timer_fn(unsigned long __data)
1427 struct delayed_work *dwork = (struct delayed_work *)__data;
1429 /* should have been called from irqsafe timer with irq already off */
1430 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1432 EXPORT_SYMBOL(delayed_work_timer_fn);
1434 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1435 struct delayed_work *dwork, unsigned long delay)
1437 struct timer_list *timer = &dwork->timer;
1438 struct work_struct *work = &dwork->work;
1440 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1441 timer->data != (unsigned long)dwork);
1442 WARN_ON_ONCE(timer_pending(timer));
1443 WARN_ON_ONCE(!list_empty(&work->entry));
1446 * If @delay is 0, queue @dwork->work immediately. This is for
1447 * both optimization and correctness. The earliest @timer can
1448 * expire is on the closest next tick and delayed_work users depend
1449 * on that there's no such delay when @delay is 0.
1452 __queue_work(cpu, wq, &dwork->work);
1456 timer_stats_timer_set_start_info(&dwork->timer);
1460 timer->expires = jiffies + delay;
1462 if (unlikely(cpu != WORK_CPU_UNBOUND))
1463 add_timer_on(timer, cpu);
1469 * queue_delayed_work_on - queue work on specific CPU after delay
1470 * @cpu: CPU number to execute work on
1471 * @wq: workqueue to use
1472 * @dwork: work to queue
1473 * @delay: number of jiffies to wait before queueing
1475 * Returns %false if @work was already on a queue, %true otherwise. If
1476 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1479 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1480 struct delayed_work *dwork, unsigned long delay)
1482 struct work_struct *work = &dwork->work;
1484 unsigned long flags;
1486 /* read the comment in __queue_work() */
1487 local_irq_save(flags);
1489 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1490 __queue_delayed_work(cpu, wq, dwork, delay);
1494 local_irq_restore(flags);
1497 EXPORT_SYMBOL(queue_delayed_work_on);
1500 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1501 * @cpu: CPU number to execute work on
1502 * @wq: workqueue to use
1503 * @dwork: work to queue
1504 * @delay: number of jiffies to wait before queueing
1506 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1507 * modify @dwork's timer so that it expires after @delay. If @delay is
1508 * zero, @work is guaranteed to be scheduled immediately regardless of its
1511 * Returns %false if @dwork was idle and queued, %true if @dwork was
1512 * pending and its timer was modified.
1514 * This function is safe to call from any context including IRQ handler.
1515 * See try_to_grab_pending() for details.
1517 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1518 struct delayed_work *dwork, unsigned long delay)
1520 unsigned long flags;
1524 ret = try_to_grab_pending(&dwork->work, true, &flags);
1525 } while (unlikely(ret == -EAGAIN));
1527 if (likely(ret >= 0)) {
1528 __queue_delayed_work(cpu, wq, dwork, delay);
1529 local_irq_restore(flags);
1532 /* -ENOENT from try_to_grab_pending() becomes %true */
1535 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1538 * worker_enter_idle - enter idle state
1539 * @worker: worker which is entering idle state
1541 * @worker is entering idle state. Update stats and idle timer if
1545 * spin_lock_irq(pool->lock).
1547 static void worker_enter_idle(struct worker *worker)
1549 struct worker_pool *pool = worker->pool;
1551 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1552 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1553 (worker->hentry.next || worker->hentry.pprev)))
1556 /* can't use worker_set_flags(), also called from start_worker() */
1557 worker->flags |= WORKER_IDLE;
1559 worker->last_active = jiffies;
1561 /* idle_list is LIFO */
1562 list_add(&worker->entry, &pool->idle_list);
1564 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1565 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1568 * Sanity check nr_running. Because wq_unbind_fn() releases
1569 * pool->lock between setting %WORKER_UNBOUND and zapping
1570 * nr_running, the warning may trigger spuriously. Check iff
1571 * unbind is not in progress.
1573 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1574 pool->nr_workers == pool->nr_idle &&
1575 atomic_read(&pool->nr_running));
1579 * worker_leave_idle - leave idle state
1580 * @worker: worker which is leaving idle state
1582 * @worker is leaving idle state. Update stats.
1585 * spin_lock_irq(pool->lock).
1587 static void worker_leave_idle(struct worker *worker)
1589 struct worker_pool *pool = worker->pool;
1591 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1593 worker_clr_flags(worker, WORKER_IDLE);
1595 list_del_init(&worker->entry);
1599 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1600 * @pool: target worker_pool
1602 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1604 * Works which are scheduled while the cpu is online must at least be
1605 * scheduled to a worker which is bound to the cpu so that if they are
1606 * flushed from cpu callbacks while cpu is going down, they are
1607 * guaranteed to execute on the cpu.
1609 * This function is to be used by unbound workers and rescuers to bind
1610 * themselves to the target cpu and may race with cpu going down or
1611 * coming online. kthread_bind() can't be used because it may put the
1612 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1613 * verbatim as it's best effort and blocking and pool may be
1614 * [dis]associated in the meantime.
1616 * This function tries set_cpus_allowed() and locks pool and verifies the
1617 * binding against %POOL_DISASSOCIATED which is set during
1618 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1619 * enters idle state or fetches works without dropping lock, it can
1620 * guarantee the scheduling requirement described in the first paragraph.
1623 * Might sleep. Called without any lock but returns with pool->lock
1627 * %true if the associated pool is online (@worker is successfully
1628 * bound), %false if offline.
1630 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1631 __acquires(&pool->lock)
1635 * The following call may fail, succeed or succeed
1636 * without actually migrating the task to the cpu if
1637 * it races with cpu hotunplug operation. Verify
1638 * against POOL_DISASSOCIATED.
1640 if (!(pool->flags & POOL_DISASSOCIATED))
1641 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1643 spin_lock_irq(&pool->lock);
1644 if (pool->flags & POOL_DISASSOCIATED)
1646 if (task_cpu(current) == pool->cpu &&
1647 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1649 spin_unlock_irq(&pool->lock);
1652 * We've raced with CPU hot[un]plug. Give it a breather
1653 * and retry migration. cond_resched() is required here;
1654 * otherwise, we might deadlock against cpu_stop trying to
1655 * bring down the CPU on non-preemptive kernel.
1662 static struct worker *alloc_worker(void)
1664 struct worker *worker;
1666 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1668 INIT_LIST_HEAD(&worker->entry);
1669 INIT_LIST_HEAD(&worker->scheduled);
1670 /* on creation a worker is in !idle && prep state */
1671 worker->flags = WORKER_PREP;
1677 * create_worker - create a new workqueue worker
1678 * @pool: pool the new worker will belong to
1680 * Create a new worker which is bound to @pool. The returned worker
1681 * can be started by calling start_worker() or destroyed using
1685 * Might sleep. Does GFP_KERNEL allocations.
1688 * Pointer to the newly created worker.
1690 static struct worker *create_worker(struct worker_pool *pool)
1692 struct worker *worker = NULL;
1696 lockdep_assert_held(&pool->manager_mutex);
1699 * ID is needed to determine kthread name. Allocate ID first
1700 * without installing the pointer.
1702 idr_preload(GFP_KERNEL);
1703 spin_lock_irq(&pool->lock);
1705 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1707 spin_unlock_irq(&pool->lock);
1712 worker = alloc_worker();
1716 worker->pool = pool;
1720 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1721 pool->attrs->nice < 0 ? "H" : "");
1723 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1725 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1726 "kworker/%s", id_buf);
1727 if (IS_ERR(worker->task))
1731 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1732 * online CPUs. It'll be re-applied when any of the CPUs come up.
1734 set_user_nice(worker->task, pool->attrs->nice);
1735 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1737 /* prevent userland from meddling with cpumask of workqueue workers */
1738 worker->task->flags |= PF_NO_SETAFFINITY;
1741 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1742 * remains stable across this function. See the comments above the
1743 * flag definition for details.
1745 if (pool->flags & POOL_DISASSOCIATED)
1746 worker->flags |= WORKER_UNBOUND;
1748 /* successful, commit the pointer to idr */
1749 spin_lock_irq(&pool->lock);
1750 idr_replace(&pool->worker_idr, worker, worker->id);
1751 spin_unlock_irq(&pool->lock);
1757 spin_lock_irq(&pool->lock);
1758 idr_remove(&pool->worker_idr, id);
1759 spin_unlock_irq(&pool->lock);
1766 * start_worker - start a newly created worker
1767 * @worker: worker to start
1769 * Make the pool aware of @worker and start it.
1772 * spin_lock_irq(pool->lock).
1774 static void start_worker(struct worker *worker)
1776 worker->flags |= WORKER_STARTED;
1777 worker->pool->nr_workers++;
1778 worker_enter_idle(worker);
1779 wake_up_process(worker->task);
1783 * create_and_start_worker - create and start a worker for a pool
1784 * @pool: the target pool
1786 * Grab the managership of @pool and create and start a new worker for it.
1788 static int create_and_start_worker(struct worker_pool *pool)
1790 struct worker *worker;
1792 mutex_lock(&pool->manager_mutex);
1794 worker = create_worker(pool);
1796 spin_lock_irq(&pool->lock);
1797 start_worker(worker);
1798 spin_unlock_irq(&pool->lock);
1801 mutex_unlock(&pool->manager_mutex);
1803 return worker ? 0 : -ENOMEM;
1807 * destroy_worker - destroy a workqueue worker
1808 * @worker: worker to be destroyed
1810 * Destroy @worker and adjust @pool stats accordingly.
1813 * spin_lock_irq(pool->lock) which is released and regrabbed.
1815 static void destroy_worker(struct worker *worker)
1817 struct worker_pool *pool = worker->pool;
1819 lockdep_assert_held(&pool->manager_mutex);
1820 lockdep_assert_held(&pool->lock);
1822 /* sanity check frenzy */
1823 if (WARN_ON(worker->current_work) ||
1824 WARN_ON(!list_empty(&worker->scheduled)))
1827 if (worker->flags & WORKER_STARTED)
1829 if (worker->flags & WORKER_IDLE)
1832 list_del_init(&worker->entry);
1833 worker->flags |= WORKER_DIE;
1835 idr_remove(&pool->worker_idr, worker->id);
1837 spin_unlock_irq(&pool->lock);
1839 kthread_stop(worker->task);
1842 spin_lock_irq(&pool->lock);
1845 static void idle_worker_timeout(unsigned long __pool)
1847 struct worker_pool *pool = (void *)__pool;
1849 spin_lock_irq(&pool->lock);
1851 if (too_many_workers(pool)) {
1852 struct worker *worker;
1853 unsigned long expires;
1855 /* idle_list is kept in LIFO order, check the last one */
1856 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1857 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1859 if (time_before(jiffies, expires))
1860 mod_timer(&pool->idle_timer, expires);
1862 /* it's been idle for too long, wake up manager */
1863 pool->flags |= POOL_MANAGE_WORKERS;
1864 wake_up_worker(pool);
1868 spin_unlock_irq(&pool->lock);
1871 static void send_mayday(struct work_struct *work)
1873 struct pool_workqueue *pwq = get_work_pwq(work);
1874 struct workqueue_struct *wq = pwq->wq;
1876 lockdep_assert_held(&wq_mayday_lock);
1881 /* mayday mayday mayday */
1882 if (list_empty(&pwq->mayday_node)) {
1883 list_add_tail(&pwq->mayday_node, &wq->maydays);
1884 wake_up_process(wq->rescuer->task);
1888 static void pool_mayday_timeout(unsigned long __pool)
1890 struct worker_pool *pool = (void *)__pool;
1891 struct work_struct *work;
1893 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1894 spin_lock(&pool->lock);
1896 if (need_to_create_worker(pool)) {
1898 * We've been trying to create a new worker but
1899 * haven't been successful. We might be hitting an
1900 * allocation deadlock. Send distress signals to
1903 list_for_each_entry(work, &pool->worklist, entry)
1907 spin_unlock(&pool->lock);
1908 spin_unlock_irq(&wq_mayday_lock);
1910 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1914 * maybe_create_worker - create a new worker if necessary
1915 * @pool: pool to create a new worker for
1917 * Create a new worker for @pool if necessary. @pool is guaranteed to
1918 * have at least one idle worker on return from this function. If
1919 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1920 * sent to all rescuers with works scheduled on @pool to resolve
1921 * possible allocation deadlock.
1923 * On return, need_to_create_worker() is guaranteed to be %false and
1924 * may_start_working() %true.
1927 * spin_lock_irq(pool->lock) which may be released and regrabbed
1928 * multiple times. Does GFP_KERNEL allocations. Called only from
1932 * %false if no action was taken and pool->lock stayed locked, %true
1935 static bool maybe_create_worker(struct worker_pool *pool)
1936 __releases(&pool->lock)
1937 __acquires(&pool->lock)
1939 if (!need_to_create_worker(pool))
1942 spin_unlock_irq(&pool->lock);
1944 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1945 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1948 struct worker *worker;
1950 worker = create_worker(pool);
1952 del_timer_sync(&pool->mayday_timer);
1953 spin_lock_irq(&pool->lock);
1954 start_worker(worker);
1955 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1960 if (!need_to_create_worker(pool))
1963 __set_current_state(TASK_INTERRUPTIBLE);
1964 schedule_timeout(CREATE_COOLDOWN);
1966 if (!need_to_create_worker(pool))
1970 del_timer_sync(&pool->mayday_timer);
1971 spin_lock_irq(&pool->lock);
1972 if (need_to_create_worker(pool))
1978 * maybe_destroy_worker - destroy workers which have been idle for a while
1979 * @pool: pool to destroy workers for
1981 * Destroy @pool workers which have been idle for longer than
1982 * IDLE_WORKER_TIMEOUT.
1985 * spin_lock_irq(pool->lock) which may be released and regrabbed
1986 * multiple times. Called only from manager.
1989 * %false if no action was taken and pool->lock stayed locked, %true
1992 static bool maybe_destroy_workers(struct worker_pool *pool)
1996 while (too_many_workers(pool)) {
1997 struct worker *worker;
1998 unsigned long expires;
2000 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2001 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2003 if (time_before(jiffies, expires)) {
2004 mod_timer(&pool->idle_timer, expires);
2008 destroy_worker(worker);
2016 * manage_workers - manage worker pool
2019 * Assume the manager role and manage the worker pool @worker belongs
2020 * to. At any given time, there can be only zero or one manager per
2021 * pool. The exclusion is handled automatically by this function.
2023 * The caller can safely start processing works on false return. On
2024 * true return, it's guaranteed that need_to_create_worker() is false
2025 * and may_start_working() is true.
2028 * spin_lock_irq(pool->lock) which may be released and regrabbed
2029 * multiple times. Does GFP_KERNEL allocations.
2032 * spin_lock_irq(pool->lock) which may be released and regrabbed
2033 * multiple times. Does GFP_KERNEL allocations.
2035 static bool manage_workers(struct worker *worker)
2037 struct worker_pool *pool = worker->pool;
2041 * Managership is governed by two mutexes - manager_arb and
2042 * manager_mutex. manager_arb handles arbitration of manager role.
2043 * Anyone who successfully grabs manager_arb wins the arbitration
2044 * and becomes the manager. mutex_trylock() on pool->manager_arb
2045 * failure while holding pool->lock reliably indicates that someone
2046 * else is managing the pool and the worker which failed trylock
2047 * can proceed to executing work items. This means that anyone
2048 * grabbing manager_arb is responsible for actually performing
2049 * manager duties. If manager_arb is grabbed and released without
2050 * actual management, the pool may stall indefinitely.
2052 * manager_mutex is used for exclusion of actual management
2053 * operations. The holder of manager_mutex can be sure that none
2054 * of management operations, including creation and destruction of
2055 * workers, won't take place until the mutex is released. Because
2056 * manager_mutex doesn't interfere with manager role arbitration,
2057 * it is guaranteed that the pool's management, while may be
2058 * delayed, won't be disturbed by someone else grabbing
2061 if (!mutex_trylock(&pool->manager_arb))
2065 * With manager arbitration won, manager_mutex would be free in
2066 * most cases. trylock first without dropping @pool->lock.
2068 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2069 spin_unlock_irq(&pool->lock);
2070 mutex_lock(&pool->manager_mutex);
2071 spin_lock_irq(&pool->lock);
2075 pool->flags &= ~POOL_MANAGE_WORKERS;
2078 * Destroy and then create so that may_start_working() is true
2081 ret |= maybe_destroy_workers(pool);
2082 ret |= maybe_create_worker(pool);
2084 mutex_unlock(&pool->manager_mutex);
2085 mutex_unlock(&pool->manager_arb);
2090 * process_one_work - process single work
2092 * @work: work to process
2094 * Process @work. This function contains all the logics necessary to
2095 * process a single work including synchronization against and
2096 * interaction with other workers on the same cpu, queueing and
2097 * flushing. As long as context requirement is met, any worker can
2098 * call this function to process a work.
2101 * spin_lock_irq(pool->lock) which is released and regrabbed.
2103 static void process_one_work(struct worker *worker, struct work_struct *work)
2104 __releases(&pool->lock)
2105 __acquires(&pool->lock)
2107 struct pool_workqueue *pwq = get_work_pwq(work);
2108 struct worker_pool *pool = worker->pool;
2109 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2111 struct worker *collision;
2112 #ifdef CONFIG_LOCKDEP
2114 * It is permissible to free the struct work_struct from
2115 * inside the function that is called from it, this we need to
2116 * take into account for lockdep too. To avoid bogus "held
2117 * lock freed" warnings as well as problems when looking into
2118 * work->lockdep_map, make a copy and use that here.
2120 struct lockdep_map lockdep_map;
2122 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2125 * Ensure we're on the correct CPU. DISASSOCIATED test is
2126 * necessary to avoid spurious warnings from rescuers servicing the
2127 * unbound or a disassociated pool.
2129 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2130 !(pool->flags & POOL_DISASSOCIATED) &&
2131 raw_smp_processor_id() != pool->cpu);
2134 * A single work shouldn't be executed concurrently by
2135 * multiple workers on a single cpu. Check whether anyone is
2136 * already processing the work. If so, defer the work to the
2137 * currently executing one.
2139 collision = find_worker_executing_work(pool, work);
2140 if (unlikely(collision)) {
2141 move_linked_works(work, &collision->scheduled, NULL);
2145 /* claim and dequeue */
2146 debug_work_deactivate(work);
2147 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2148 worker->current_work = work;
2149 worker->current_func = work->func;
2150 worker->current_pwq = pwq;
2151 work_color = get_work_color(work);
2153 list_del_init(&work->entry);
2156 * CPU intensive works don't participate in concurrency
2157 * management. They're the scheduler's responsibility.
2159 if (unlikely(cpu_intensive))
2160 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2163 * Unbound pool isn't concurrency managed and work items should be
2164 * executed ASAP. Wake up another worker if necessary.
2166 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2167 wake_up_worker(pool);
2170 * Record the last pool and clear PENDING which should be the last
2171 * update to @work. Also, do this inside @pool->lock so that
2172 * PENDING and queued state changes happen together while IRQ is
2175 set_work_pool_and_clear_pending(work, pool->id);
2177 spin_unlock_irq(&pool->lock);
2179 lock_map_acquire_read(&pwq->wq->lockdep_map);
2180 lock_map_acquire(&lockdep_map);
2181 trace_workqueue_execute_start(work);
2182 worker->current_func(work);
2184 * While we must be careful to not use "work" after this, the trace
2185 * point will only record its address.
2187 trace_workqueue_execute_end(work);
2188 lock_map_release(&lockdep_map);
2189 lock_map_release(&pwq->wq->lockdep_map);
2191 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2192 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2193 " last function: %pf\n",
2194 current->comm, preempt_count(), task_pid_nr(current),
2195 worker->current_func);
2196 debug_show_held_locks(current);
2200 spin_lock_irq(&pool->lock);
2202 /* clear cpu intensive status */
2203 if (unlikely(cpu_intensive))
2204 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2206 /* we're done with it, release */
2207 hash_del(&worker->hentry);
2208 worker->current_work = NULL;
2209 worker->current_func = NULL;
2210 worker->current_pwq = NULL;
2211 worker->desc_valid = false;
2212 pwq_dec_nr_in_flight(pwq, work_color);
2216 * process_scheduled_works - process scheduled works
2219 * Process all scheduled works. Please note that the scheduled list
2220 * may change while processing a work, so this function repeatedly
2221 * fetches a work from the top and executes it.
2224 * spin_lock_irq(pool->lock) which may be released and regrabbed
2227 static void process_scheduled_works(struct worker *worker)
2229 while (!list_empty(&worker->scheduled)) {
2230 struct work_struct *work = list_first_entry(&worker->scheduled,
2231 struct work_struct, entry);
2232 process_one_work(worker, work);
2237 * worker_thread - the worker thread function
2240 * The worker thread function. All workers belong to a worker_pool -
2241 * either a per-cpu one or dynamic unbound one. These workers process all
2242 * work items regardless of their specific target workqueue. The only
2243 * exception is work items which belong to workqueues with a rescuer which
2244 * will be explained in rescuer_thread().
2246 static int worker_thread(void *__worker)
2248 struct worker *worker = __worker;
2249 struct worker_pool *pool = worker->pool;
2251 /* tell the scheduler that this is a workqueue worker */
2252 worker->task->flags |= PF_WQ_WORKER;
2254 spin_lock_irq(&pool->lock);
2256 /* am I supposed to die? */
2257 if (unlikely(worker->flags & WORKER_DIE)) {
2258 spin_unlock_irq(&pool->lock);
2259 WARN_ON_ONCE(!list_empty(&worker->entry));
2260 worker->task->flags &= ~PF_WQ_WORKER;
2264 worker_leave_idle(worker);
2266 /* no more worker necessary? */
2267 if (!need_more_worker(pool))
2270 /* do we need to manage? */
2271 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2275 * ->scheduled list can only be filled while a worker is
2276 * preparing to process a work or actually processing it.
2277 * Make sure nobody diddled with it while I was sleeping.
2279 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2282 * Finish PREP stage. We're guaranteed to have at least one idle
2283 * worker or that someone else has already assumed the manager
2284 * role. This is where @worker starts participating in concurrency
2285 * management if applicable and concurrency management is restored
2286 * after being rebound. See rebind_workers() for details.
2288 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2291 struct work_struct *work =
2292 list_first_entry(&pool->worklist,
2293 struct work_struct, entry);
2295 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2296 /* optimization path, not strictly necessary */
2297 process_one_work(worker, work);
2298 if (unlikely(!list_empty(&worker->scheduled)))
2299 process_scheduled_works(worker);
2301 move_linked_works(work, &worker->scheduled, NULL);
2302 process_scheduled_works(worker);
2304 } while (keep_working(pool));
2306 worker_set_flags(worker, WORKER_PREP, false);
2308 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2312 * pool->lock is held and there's no work to process and no need to
2313 * manage, sleep. Workers are woken up only while holding
2314 * pool->lock or from local cpu, so setting the current state
2315 * before releasing pool->lock is enough to prevent losing any
2318 worker_enter_idle(worker);
2319 __set_current_state(TASK_INTERRUPTIBLE);
2320 spin_unlock_irq(&pool->lock);
2326 * rescuer_thread - the rescuer thread function
2329 * Workqueue rescuer thread function. There's one rescuer for each
2330 * workqueue which has WQ_MEM_RECLAIM set.
2332 * Regular work processing on a pool may block trying to create a new
2333 * worker which uses GFP_KERNEL allocation which has slight chance of
2334 * developing into deadlock if some works currently on the same queue
2335 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2336 * the problem rescuer solves.
2338 * When such condition is possible, the pool summons rescuers of all
2339 * workqueues which have works queued on the pool and let them process
2340 * those works so that forward progress can be guaranteed.
2342 * This should happen rarely.
2344 static int rescuer_thread(void *__rescuer)
2346 struct worker *rescuer = __rescuer;
2347 struct workqueue_struct *wq = rescuer->rescue_wq;
2348 struct list_head *scheduled = &rescuer->scheduled;
2350 set_user_nice(current, RESCUER_NICE_LEVEL);
2353 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2354 * doesn't participate in concurrency management.
2356 rescuer->task->flags |= PF_WQ_WORKER;
2358 set_current_state(TASK_INTERRUPTIBLE);
2360 if (kthread_should_stop()) {
2361 __set_current_state(TASK_RUNNING);
2362 rescuer->task->flags &= ~PF_WQ_WORKER;
2366 /* see whether any pwq is asking for help */
2367 spin_lock_irq(&wq_mayday_lock);
2369 while (!list_empty(&wq->maydays)) {
2370 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2371 struct pool_workqueue, mayday_node);
2372 struct worker_pool *pool = pwq->pool;
2373 struct work_struct *work, *n;
2375 __set_current_state(TASK_RUNNING);
2376 list_del_init(&pwq->mayday_node);
2378 spin_unlock_irq(&wq_mayday_lock);
2380 /* migrate to the target cpu if possible */
2381 worker_maybe_bind_and_lock(pool);
2382 rescuer->pool = pool;
2385 * Slurp in all works issued via this workqueue and
2388 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2389 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2390 if (get_work_pwq(work) == pwq)
2391 move_linked_works(work, scheduled, &n);
2393 process_scheduled_works(rescuer);
2396 * Leave this pool. If keep_working() is %true, notify a
2397 * regular worker; otherwise, we end up with 0 concurrency
2398 * and stalling the execution.
2400 if (keep_working(pool))
2401 wake_up_worker(pool);
2403 rescuer->pool = NULL;
2404 spin_unlock(&pool->lock);
2405 spin_lock(&wq_mayday_lock);
2408 spin_unlock_irq(&wq_mayday_lock);
2410 /* rescuers should never participate in concurrency management */
2411 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2417 struct work_struct work;
2418 struct completion done;
2421 static void wq_barrier_func(struct work_struct *work)
2423 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2424 complete(&barr->done);
2428 * insert_wq_barrier - insert a barrier work
2429 * @pwq: pwq to insert barrier into
2430 * @barr: wq_barrier to insert
2431 * @target: target work to attach @barr to
2432 * @worker: worker currently executing @target, NULL if @target is not executing
2434 * @barr is linked to @target such that @barr is completed only after
2435 * @target finishes execution. Please note that the ordering
2436 * guarantee is observed only with respect to @target and on the local
2439 * Currently, a queued barrier can't be canceled. This is because
2440 * try_to_grab_pending() can't determine whether the work to be
2441 * grabbed is at the head of the queue and thus can't clear LINKED
2442 * flag of the previous work while there must be a valid next work
2443 * after a work with LINKED flag set.
2445 * Note that when @worker is non-NULL, @target may be modified
2446 * underneath us, so we can't reliably determine pwq from @target.
2449 * spin_lock_irq(pool->lock).
2451 static void insert_wq_barrier(struct pool_workqueue *pwq,
2452 struct wq_barrier *barr,
2453 struct work_struct *target, struct worker *worker)
2455 struct list_head *head;
2456 unsigned int linked = 0;
2459 * debugobject calls are safe here even with pool->lock locked
2460 * as we know for sure that this will not trigger any of the
2461 * checks and call back into the fixup functions where we
2464 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2465 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2466 init_completion(&barr->done);
2469 * If @target is currently being executed, schedule the
2470 * barrier to the worker; otherwise, put it after @target.
2473 head = worker->scheduled.next;
2475 unsigned long *bits = work_data_bits(target);
2477 head = target->entry.next;
2478 /* there can already be other linked works, inherit and set */
2479 linked = *bits & WORK_STRUCT_LINKED;
2480 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2483 debug_work_activate(&barr->work);
2484 insert_work(pwq, &barr->work, head,
2485 work_color_to_flags(WORK_NO_COLOR) | linked);
2489 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2490 * @wq: workqueue being flushed
2491 * @flush_color: new flush color, < 0 for no-op
2492 * @work_color: new work color, < 0 for no-op
2494 * Prepare pwqs for workqueue flushing.
2496 * If @flush_color is non-negative, flush_color on all pwqs should be
2497 * -1. If no pwq has in-flight commands at the specified color, all
2498 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2499 * has in flight commands, its pwq->flush_color is set to
2500 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2501 * wakeup logic is armed and %true is returned.
2503 * The caller should have initialized @wq->first_flusher prior to
2504 * calling this function with non-negative @flush_color. If
2505 * @flush_color is negative, no flush color update is done and %false
2508 * If @work_color is non-negative, all pwqs should have the same
2509 * work_color which is previous to @work_color and all will be
2510 * advanced to @work_color.
2513 * mutex_lock(wq->mutex).
2516 * %true if @flush_color >= 0 and there's something to flush. %false
2519 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2520 int flush_color, int work_color)
2523 struct pool_workqueue *pwq;
2525 if (flush_color >= 0) {
2526 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2527 atomic_set(&wq->nr_pwqs_to_flush, 1);
2530 for_each_pwq(pwq, wq) {
2531 struct worker_pool *pool = pwq->pool;
2533 spin_lock_irq(&pool->lock);
2535 if (flush_color >= 0) {
2536 WARN_ON_ONCE(pwq->flush_color != -1);
2538 if (pwq->nr_in_flight[flush_color]) {
2539 pwq->flush_color = flush_color;
2540 atomic_inc(&wq->nr_pwqs_to_flush);
2545 if (work_color >= 0) {
2546 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2547 pwq->work_color = work_color;
2550 spin_unlock_irq(&pool->lock);
2553 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2554 complete(&wq->first_flusher->done);
2560 * flush_workqueue - ensure that any scheduled work has run to completion.
2561 * @wq: workqueue to flush
2563 * This function sleeps until all work items which were queued on entry
2564 * have finished execution, but it is not livelocked by new incoming ones.
2566 void flush_workqueue(struct workqueue_struct *wq)
2568 struct wq_flusher this_flusher = {
2569 .list = LIST_HEAD_INIT(this_flusher.list),
2571 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2575 lock_map_acquire(&wq->lockdep_map);
2576 lock_map_release(&wq->lockdep_map);
2578 mutex_lock(&wq->mutex);
2581 * Start-to-wait phase
2583 next_color = work_next_color(wq->work_color);
2585 if (next_color != wq->flush_color) {
2587 * Color space is not full. The current work_color
2588 * becomes our flush_color and work_color is advanced
2591 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2592 this_flusher.flush_color = wq->work_color;
2593 wq->work_color = next_color;
2595 if (!wq->first_flusher) {
2596 /* no flush in progress, become the first flusher */
2597 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2599 wq->first_flusher = &this_flusher;
2601 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2603 /* nothing to flush, done */
2604 wq->flush_color = next_color;
2605 wq->first_flusher = NULL;
2610 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2611 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2612 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2616 * Oops, color space is full, wait on overflow queue.
2617 * The next flush completion will assign us
2618 * flush_color and transfer to flusher_queue.
2620 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2623 mutex_unlock(&wq->mutex);
2625 wait_for_completion(&this_flusher.done);
2628 * Wake-up-and-cascade phase
2630 * First flushers are responsible for cascading flushes and
2631 * handling overflow. Non-first flushers can simply return.
2633 if (wq->first_flusher != &this_flusher)
2636 mutex_lock(&wq->mutex);
2638 /* we might have raced, check again with mutex held */
2639 if (wq->first_flusher != &this_flusher)
2642 wq->first_flusher = NULL;
2644 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2645 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2648 struct wq_flusher *next, *tmp;
2650 /* complete all the flushers sharing the current flush color */
2651 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2652 if (next->flush_color != wq->flush_color)
2654 list_del_init(&next->list);
2655 complete(&next->done);
2658 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2659 wq->flush_color != work_next_color(wq->work_color));
2661 /* this flush_color is finished, advance by one */
2662 wq->flush_color = work_next_color(wq->flush_color);
2664 /* one color has been freed, handle overflow queue */
2665 if (!list_empty(&wq->flusher_overflow)) {
2667 * Assign the same color to all overflowed
2668 * flushers, advance work_color and append to
2669 * flusher_queue. This is the start-to-wait
2670 * phase for these overflowed flushers.
2672 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2673 tmp->flush_color = wq->work_color;
2675 wq->work_color = work_next_color(wq->work_color);
2677 list_splice_tail_init(&wq->flusher_overflow,
2678 &wq->flusher_queue);
2679 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2682 if (list_empty(&wq->flusher_queue)) {
2683 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2688 * Need to flush more colors. Make the next flusher
2689 * the new first flusher and arm pwqs.
2691 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2692 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2694 list_del_init(&next->list);
2695 wq->first_flusher = next;
2697 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2701 * Meh... this color is already done, clear first
2702 * flusher and repeat cascading.
2704 wq->first_flusher = NULL;
2708 mutex_unlock(&wq->mutex);
2710 EXPORT_SYMBOL_GPL(flush_workqueue);
2713 * drain_workqueue - drain a workqueue
2714 * @wq: workqueue to drain
2716 * Wait until the workqueue becomes empty. While draining is in progress,
2717 * only chain queueing is allowed. IOW, only currently pending or running
2718 * work items on @wq can queue further work items on it. @wq is flushed
2719 * repeatedly until it becomes empty. The number of flushing is detemined
2720 * by the depth of chaining and should be relatively short. Whine if it
2723 void drain_workqueue(struct workqueue_struct *wq)
2725 unsigned int flush_cnt = 0;
2726 struct pool_workqueue *pwq;
2729 * __queue_work() needs to test whether there are drainers, is much
2730 * hotter than drain_workqueue() and already looks at @wq->flags.
2731 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2733 mutex_lock(&wq->mutex);
2734 if (!wq->nr_drainers++)
2735 wq->flags |= __WQ_DRAINING;
2736 mutex_unlock(&wq->mutex);
2738 flush_workqueue(wq);
2740 mutex_lock(&wq->mutex);
2742 for_each_pwq(pwq, wq) {
2745 spin_lock_irq(&pwq->pool->lock);
2746 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2747 spin_unlock_irq(&pwq->pool->lock);
2752 if (++flush_cnt == 10 ||
2753 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2754 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2755 wq->name, flush_cnt);
2757 mutex_unlock(&wq->mutex);
2761 if (!--wq->nr_drainers)
2762 wq->flags &= ~__WQ_DRAINING;
2763 mutex_unlock(&wq->mutex);
2765 EXPORT_SYMBOL_GPL(drain_workqueue);
2767 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2769 struct worker *worker = NULL;
2770 struct worker_pool *pool;
2771 struct pool_workqueue *pwq;
2775 local_irq_disable();
2776 pool = get_work_pool(work);
2782 spin_lock(&pool->lock);
2783 /* see the comment in try_to_grab_pending() with the same code */
2784 pwq = get_work_pwq(work);
2786 if (unlikely(pwq->pool != pool))
2789 worker = find_worker_executing_work(pool, work);
2792 pwq = worker->current_pwq;
2795 insert_wq_barrier(pwq, barr, work, worker);
2796 spin_unlock_irq(&pool->lock);
2799 * If @max_active is 1 or rescuer is in use, flushing another work
2800 * item on the same workqueue may lead to deadlock. Make sure the
2801 * flusher is not running on the same workqueue by verifying write
2804 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2805 lock_map_acquire(&pwq->wq->lockdep_map);
2807 lock_map_acquire_read(&pwq->wq->lockdep_map);
2808 lock_map_release(&pwq->wq->lockdep_map);
2812 spin_unlock_irq(&pool->lock);
2817 * flush_work - wait for a work to finish executing the last queueing instance
2818 * @work: the work to flush
2820 * Wait until @work has finished execution. @work is guaranteed to be idle
2821 * on return if it hasn't been requeued since flush started.
2824 * %true if flush_work() waited for the work to finish execution,
2825 * %false if it was already idle.
2827 bool flush_work(struct work_struct *work)
2829 struct wq_barrier barr;
2831 lock_map_acquire(&work->lockdep_map);
2832 lock_map_release(&work->lockdep_map);
2834 if (start_flush_work(work, &barr)) {
2835 wait_for_completion(&barr.done);
2836 destroy_work_on_stack(&barr.work);
2842 EXPORT_SYMBOL_GPL(flush_work);
2844 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2846 unsigned long flags;
2850 ret = try_to_grab_pending(work, is_dwork, &flags);
2852 * If someone else is canceling, wait for the same event it
2853 * would be waiting for before retrying.
2855 if (unlikely(ret == -ENOENT))
2857 } while (unlikely(ret < 0));
2859 /* tell other tasks trying to grab @work to back off */
2860 mark_work_canceling(work);
2861 local_irq_restore(flags);
2864 clear_work_data(work);
2869 * cancel_work_sync - cancel a work and wait for it to finish
2870 * @work: the work to cancel
2872 * Cancel @work and wait for its execution to finish. This function
2873 * can be used even if the work re-queues itself or migrates to
2874 * another workqueue. On return from this function, @work is
2875 * guaranteed to be not pending or executing on any CPU.
2877 * cancel_work_sync(&delayed_work->work) must not be used for
2878 * delayed_work's. Use cancel_delayed_work_sync() instead.
2880 * The caller must ensure that the workqueue on which @work was last
2881 * queued can't be destroyed before this function returns.
2884 * %true if @work was pending, %false otherwise.
2886 bool cancel_work_sync(struct work_struct *work)
2888 return __cancel_work_timer(work, false);
2890 EXPORT_SYMBOL_GPL(cancel_work_sync);
2893 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2894 * @dwork: the delayed work to flush
2896 * Delayed timer is cancelled and the pending work is queued for
2897 * immediate execution. Like flush_work(), this function only
2898 * considers the last queueing instance of @dwork.
2901 * %true if flush_work() waited for the work to finish execution,
2902 * %false if it was already idle.
2904 bool flush_delayed_work(struct delayed_work *dwork)
2906 local_irq_disable();
2907 if (del_timer_sync(&dwork->timer))
2908 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2910 return flush_work(&dwork->work);
2912 EXPORT_SYMBOL(flush_delayed_work);
2915 * cancel_delayed_work - cancel a delayed work
2916 * @dwork: delayed_work to cancel
2918 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2919 * and canceled; %false if wasn't pending. Note that the work callback
2920 * function may still be running on return, unless it returns %true and the
2921 * work doesn't re-arm itself. Explicitly flush or use
2922 * cancel_delayed_work_sync() to wait on it.
2924 * This function is safe to call from any context including IRQ handler.
2926 bool cancel_delayed_work(struct delayed_work *dwork)
2928 unsigned long flags;
2932 ret = try_to_grab_pending(&dwork->work, true, &flags);
2933 } while (unlikely(ret == -EAGAIN));
2935 if (unlikely(ret < 0))
2938 set_work_pool_and_clear_pending(&dwork->work,
2939 get_work_pool_id(&dwork->work));
2940 local_irq_restore(flags);
2943 EXPORT_SYMBOL(cancel_delayed_work);
2946 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2947 * @dwork: the delayed work cancel
2949 * This is cancel_work_sync() for delayed works.
2952 * %true if @dwork was pending, %false otherwise.
2954 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2956 return __cancel_work_timer(&dwork->work, true);
2958 EXPORT_SYMBOL(cancel_delayed_work_sync);
2961 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2962 * @func: the function to call
2964 * schedule_on_each_cpu() executes @func on each online CPU using the
2965 * system workqueue and blocks until all CPUs have completed.
2966 * schedule_on_each_cpu() is very slow.
2969 * 0 on success, -errno on failure.
2971 int schedule_on_each_cpu(work_func_t func)
2974 struct work_struct __percpu *works;
2976 works = alloc_percpu(struct work_struct);
2982 for_each_online_cpu(cpu) {
2983 struct work_struct *work = per_cpu_ptr(works, cpu);
2985 INIT_WORK(work, func);
2986 schedule_work_on(cpu, work);
2989 for_each_online_cpu(cpu)
2990 flush_work(per_cpu_ptr(works, cpu));
2998 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3000 * Forces execution of the kernel-global workqueue and blocks until its
3003 * Think twice before calling this function! It's very easy to get into
3004 * trouble if you don't take great care. Either of the following situations
3005 * will lead to deadlock:
3007 * One of the work items currently on the workqueue needs to acquire
3008 * a lock held by your code or its caller.
3010 * Your code is running in the context of a work routine.
3012 * They will be detected by lockdep when they occur, but the first might not
3013 * occur very often. It depends on what work items are on the workqueue and
3014 * what locks they need, which you have no control over.
3016 * In most situations flushing the entire workqueue is overkill; you merely
3017 * need to know that a particular work item isn't queued and isn't running.
3018 * In such cases you should use cancel_delayed_work_sync() or
3019 * cancel_work_sync() instead.
3021 void flush_scheduled_work(void)
3023 flush_workqueue(system_wq);
3025 EXPORT_SYMBOL(flush_scheduled_work);
3028 * execute_in_process_context - reliably execute the routine with user context
3029 * @fn: the function to execute
3030 * @ew: guaranteed storage for the execute work structure (must
3031 * be available when the work executes)
3033 * Executes the function immediately if process context is available,
3034 * otherwise schedules the function for delayed execution.
3036 * Returns: 0 - function was executed
3037 * 1 - function was scheduled for execution
3039 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3041 if (!in_interrupt()) {
3046 INIT_WORK(&ew->work, fn);
3047 schedule_work(&ew->work);
3051 EXPORT_SYMBOL_GPL(execute_in_process_context);
3055 * Workqueues with WQ_SYSFS flag set is visible to userland via
3056 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3057 * following attributes.
3059 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3060 * max_active RW int : maximum number of in-flight work items
3062 * Unbound workqueues have the following extra attributes.
3064 * id RO int : the associated pool ID
3065 * nice RW int : nice value of the workers
3066 * cpumask RW mask : bitmask of allowed CPUs for the workers
3069 struct workqueue_struct *wq;
3073 static struct workqueue_struct *dev_to_wq(struct device *dev)
3075 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3080 static ssize_t wq_per_cpu_show(struct device *dev,
3081 struct device_attribute *attr, char *buf)
3083 struct workqueue_struct *wq = dev_to_wq(dev);
3085 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3088 static ssize_t wq_max_active_show(struct device *dev,
3089 struct device_attribute *attr, char *buf)
3091 struct workqueue_struct *wq = dev_to_wq(dev);
3093 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3096 static ssize_t wq_max_active_store(struct device *dev,
3097 struct device_attribute *attr,
3098 const char *buf, size_t count)
3100 struct workqueue_struct *wq = dev_to_wq(dev);
3103 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3106 workqueue_set_max_active(wq, val);
3110 static struct device_attribute wq_sysfs_attrs[] = {
3111 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3112 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3116 static ssize_t wq_pool_ids_show(struct device *dev,
3117 struct device_attribute *attr, char *buf)
3119 struct workqueue_struct *wq = dev_to_wq(dev);
3120 const char *delim = "";
3121 int node, written = 0;
3123 rcu_read_lock_sched();
3124 for_each_node(node) {
3125 written += scnprintf(buf + written, PAGE_SIZE - written,
3126 "%s%d:%d", delim, node,
3127 unbound_pwq_by_node(wq, node)->pool->id);
3130 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3131 rcu_read_unlock_sched();
3136 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3139 struct workqueue_struct *wq = dev_to_wq(dev);
3142 mutex_lock(&wq->mutex);
3143 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3144 mutex_unlock(&wq->mutex);
3149 /* prepare workqueue_attrs for sysfs store operations */
3150 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3152 struct workqueue_attrs *attrs;
3154 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3158 mutex_lock(&wq->mutex);
3159 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3160 mutex_unlock(&wq->mutex);
3164 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3165 const char *buf, size_t count)
3167 struct workqueue_struct *wq = dev_to_wq(dev);
3168 struct workqueue_attrs *attrs;
3171 attrs = wq_sysfs_prep_attrs(wq);
3175 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3176 attrs->nice >= -20 && attrs->nice <= 19)
3177 ret = apply_workqueue_attrs(wq, attrs);
3181 free_workqueue_attrs(attrs);
3182 return ret ?: count;
3185 static ssize_t wq_cpumask_show(struct device *dev,
3186 struct device_attribute *attr, char *buf)
3188 struct workqueue_struct *wq = dev_to_wq(dev);
3191 mutex_lock(&wq->mutex);
3192 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3193 mutex_unlock(&wq->mutex);
3195 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3199 static ssize_t wq_cpumask_store(struct device *dev,
3200 struct device_attribute *attr,
3201 const char *buf, size_t count)
3203 struct workqueue_struct *wq = dev_to_wq(dev);
3204 struct workqueue_attrs *attrs;
3207 attrs = wq_sysfs_prep_attrs(wq);
3211 ret = cpumask_parse(buf, attrs->cpumask);
3213 ret = apply_workqueue_attrs(wq, attrs);
3215 free_workqueue_attrs(attrs);
3216 return ret ?: count;
3219 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3222 struct workqueue_struct *wq = dev_to_wq(dev);
3225 mutex_lock(&wq->mutex);
3226 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3227 !wq->unbound_attrs->no_numa);
3228 mutex_unlock(&wq->mutex);
3233 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3234 const char *buf, size_t count)
3236 struct workqueue_struct *wq = dev_to_wq(dev);
3237 struct workqueue_attrs *attrs;
3240 attrs = wq_sysfs_prep_attrs(wq);
3245 if (sscanf(buf, "%d", &v) == 1) {
3246 attrs->no_numa = !v;
3247 ret = apply_workqueue_attrs(wq, attrs);
3250 free_workqueue_attrs(attrs);
3251 return ret ?: count;
3254 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3255 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3256 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3257 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3258 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3262 static struct bus_type wq_subsys = {
3263 .name = "workqueue",
3264 .dev_attrs = wq_sysfs_attrs,
3267 static int __init wq_sysfs_init(void)
3269 return subsys_virtual_register(&wq_subsys, NULL);
3271 core_initcall(wq_sysfs_init);
3273 static void wq_device_release(struct device *dev)
3275 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3281 * workqueue_sysfs_register - make a workqueue visible in sysfs
3282 * @wq: the workqueue to register
3284 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3285 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3286 * which is the preferred method.
3288 * Workqueue user should use this function directly iff it wants to apply
3289 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3290 * apply_workqueue_attrs() may race against userland updating the
3293 * Returns 0 on success, -errno on failure.
3295 int workqueue_sysfs_register(struct workqueue_struct *wq)
3297 struct wq_device *wq_dev;
3301 * Adjusting max_active or creating new pwqs by applyting
3302 * attributes breaks ordering guarantee. Disallow exposing ordered
3305 if (WARN_ON(wq->flags & __WQ_ORDERED))
3308 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3313 wq_dev->dev.bus = &wq_subsys;
3314 wq_dev->dev.init_name = wq->name;
3315 wq_dev->dev.release = wq_device_release;
3318 * unbound_attrs are created separately. Suppress uevent until
3319 * everything is ready.
3321 dev_set_uevent_suppress(&wq_dev->dev, true);
3323 ret = device_register(&wq_dev->dev);
3330 if (wq->flags & WQ_UNBOUND) {
3331 struct device_attribute *attr;
3333 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3334 ret = device_create_file(&wq_dev->dev, attr);
3336 device_unregister(&wq_dev->dev);
3343 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3348 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3349 * @wq: the workqueue to unregister
3351 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3353 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3355 struct wq_device *wq_dev = wq->wq_dev;
3361 device_unregister(&wq_dev->dev);
3363 #else /* CONFIG_SYSFS */
3364 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3365 #endif /* CONFIG_SYSFS */
3368 * free_workqueue_attrs - free a workqueue_attrs
3369 * @attrs: workqueue_attrs to free
3371 * Undo alloc_workqueue_attrs().
3373 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3376 free_cpumask_var(attrs->cpumask);
3382 * alloc_workqueue_attrs - allocate a workqueue_attrs
3383 * @gfp_mask: allocation mask to use
3385 * Allocate a new workqueue_attrs, initialize with default settings and
3386 * return it. Returns NULL on failure.
3388 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3390 struct workqueue_attrs *attrs;
3392 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3395 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3398 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3401 free_workqueue_attrs(attrs);
3405 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3406 const struct workqueue_attrs *from)
3408 to->nice = from->nice;
3409 cpumask_copy(to->cpumask, from->cpumask);
3412 /* hash value of the content of @attr */
3413 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3417 hash = jhash_1word(attrs->nice, hash);
3418 hash = jhash(cpumask_bits(attrs->cpumask),
3419 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3423 /* content equality test */
3424 static bool wqattrs_equal(const struct workqueue_attrs *a,
3425 const struct workqueue_attrs *b)
3427 if (a->nice != b->nice)
3429 if (!cpumask_equal(a->cpumask, b->cpumask))
3435 * init_worker_pool - initialize a newly zalloc'd worker_pool
3436 * @pool: worker_pool to initialize
3438 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3439 * Returns 0 on success, -errno on failure. Even on failure, all fields
3440 * inside @pool proper are initialized and put_unbound_pool() can be called
3441 * on @pool safely to release it.
3443 static int init_worker_pool(struct worker_pool *pool)
3445 spin_lock_init(&pool->lock);
3448 pool->node = NUMA_NO_NODE;
3449 pool->flags |= POOL_DISASSOCIATED;
3450 INIT_LIST_HEAD(&pool->worklist);
3451 INIT_LIST_HEAD(&pool->idle_list);
3452 hash_init(pool->busy_hash);
3454 init_timer_deferrable(&pool->idle_timer);
3455 pool->idle_timer.function = idle_worker_timeout;
3456 pool->idle_timer.data = (unsigned long)pool;
3458 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3459 (unsigned long)pool);
3461 mutex_init(&pool->manager_arb);
3462 mutex_init(&pool->manager_mutex);
3463 idr_init(&pool->worker_idr);
3465 INIT_HLIST_NODE(&pool->hash_node);
3468 /* shouldn't fail above this point */
3469 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3475 static void rcu_free_pool(struct rcu_head *rcu)
3477 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3479 idr_destroy(&pool->worker_idr);
3480 free_workqueue_attrs(pool->attrs);
3485 * put_unbound_pool - put a worker_pool
3486 * @pool: worker_pool to put
3488 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3489 * safe manner. get_unbound_pool() calls this function on its failure path
3490 * and this function should be able to release pools which went through,
3491 * successfully or not, init_worker_pool().
3493 * Should be called with wq_pool_mutex held.
3495 static void put_unbound_pool(struct worker_pool *pool)
3497 struct worker *worker;
3499 lockdep_assert_held(&wq_pool_mutex);
3505 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3506 WARN_ON(!list_empty(&pool->worklist)))
3509 /* release id and unhash */
3511 idr_remove(&worker_pool_idr, pool->id);
3512 hash_del(&pool->hash_node);
3515 * Become the manager and destroy all workers. Grabbing
3516 * manager_arb prevents @pool's workers from blocking on
3519 mutex_lock(&pool->manager_arb);
3520 mutex_lock(&pool->manager_mutex);
3521 spin_lock_irq(&pool->lock);
3523 while ((worker = first_worker(pool)))
3524 destroy_worker(worker);
3525 WARN_ON(pool->nr_workers || pool->nr_idle);
3527 spin_unlock_irq(&pool->lock);
3528 mutex_unlock(&pool->manager_mutex);
3529 mutex_unlock(&pool->manager_arb);
3531 /* shut down the timers */
3532 del_timer_sync(&pool->idle_timer);
3533 del_timer_sync(&pool->mayday_timer);
3535 /* sched-RCU protected to allow dereferences from get_work_pool() */
3536 call_rcu_sched(&pool->rcu, rcu_free_pool);
3540 * get_unbound_pool - get a worker_pool with the specified attributes
3541 * @attrs: the attributes of the worker_pool to get
3543 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3544 * reference count and return it. If there already is a matching
3545 * worker_pool, it will be used; otherwise, this function attempts to
3546 * create a new one. On failure, returns NULL.
3548 * Should be called with wq_pool_mutex held.
3550 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3552 u32 hash = wqattrs_hash(attrs);
3553 struct worker_pool *pool;
3556 lockdep_assert_held(&wq_pool_mutex);
3558 /* do we already have a matching pool? */
3559 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3560 if (wqattrs_equal(pool->attrs, attrs)) {
3566 /* nope, create a new one */
3567 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3568 if (!pool || init_worker_pool(pool) < 0)
3571 if (workqueue_freezing)
3572 pool->flags |= POOL_FREEZING;
3574 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3575 copy_workqueue_attrs(pool->attrs, attrs);
3577 /* if cpumask is contained inside a NUMA node, we belong to that node */
3578 if (wq_numa_enabled) {
3579 for_each_node(node) {
3580 if (cpumask_subset(pool->attrs->cpumask,
3581 wq_numa_possible_cpumask[node])) {
3588 if (worker_pool_assign_id(pool) < 0)
3591 /* create and start the initial worker */
3592 if (create_and_start_worker(pool) < 0)
3596 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3601 put_unbound_pool(pool);
3605 static void rcu_free_pwq(struct rcu_head *rcu)
3607 kmem_cache_free(pwq_cache,
3608 container_of(rcu, struct pool_workqueue, rcu));
3612 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3613 * and needs to be destroyed.
3615 static void pwq_unbound_release_workfn(struct work_struct *work)
3617 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3618 unbound_release_work);
3619 struct workqueue_struct *wq = pwq->wq;
3620 struct worker_pool *pool = pwq->pool;
3623 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3627 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3628 * necessary on release but do it anyway. It's easier to verify
3629 * and consistent with the linking path.
3631 mutex_lock(&wq->mutex);
3632 list_del_rcu(&pwq->pwqs_node);
3633 is_last = list_empty(&wq->pwqs);
3634 mutex_unlock(&wq->mutex);
3636 mutex_lock(&wq_pool_mutex);
3637 put_unbound_pool(pool);
3638 mutex_unlock(&wq_pool_mutex);
3640 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3643 * If we're the last pwq going away, @wq is already dead and no one
3644 * is gonna access it anymore. Free it.
3647 free_workqueue_attrs(wq->unbound_attrs);
3653 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3654 * @pwq: target pool_workqueue
3656 * If @pwq isn't freezing, set @pwq->max_active to the associated
3657 * workqueue's saved_max_active and activate delayed work items
3658 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3660 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3662 struct workqueue_struct *wq = pwq->wq;
3663 bool freezable = wq->flags & WQ_FREEZABLE;
3665 /* for @wq->saved_max_active */
3666 lockdep_assert_held(&wq->mutex);
3668 /* fast exit for non-freezable wqs */
3669 if (!freezable && pwq->max_active == wq->saved_max_active)
3672 spin_lock_irq(&pwq->pool->lock);
3674 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3675 pwq->max_active = wq->saved_max_active;
3677 while (!list_empty(&pwq->delayed_works) &&
3678 pwq->nr_active < pwq->max_active)
3679 pwq_activate_first_delayed(pwq);
3682 * Need to kick a worker after thawed or an unbound wq's
3683 * max_active is bumped. It's a slow path. Do it always.
3685 wake_up_worker(pwq->pool);
3687 pwq->max_active = 0;
3690 spin_unlock_irq(&pwq->pool->lock);
3693 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3694 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3695 struct worker_pool *pool)
3697 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3699 memset(pwq, 0, sizeof(*pwq));
3703 pwq->flush_color = -1;
3705 INIT_LIST_HEAD(&pwq->delayed_works);
3706 INIT_LIST_HEAD(&pwq->pwqs_node);
3707 INIT_LIST_HEAD(&pwq->mayday_node);
3708 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3711 /* sync @pwq with the current state of its associated wq and link it */
3712 static void link_pwq(struct pool_workqueue *pwq)
3714 struct workqueue_struct *wq = pwq->wq;
3716 lockdep_assert_held(&wq->mutex);
3718 /* may be called multiple times, ignore if already linked */
3719 if (!list_empty(&pwq->pwqs_node))
3723 * Set the matching work_color. This is synchronized with
3724 * wq->mutex to avoid confusing flush_workqueue().
3726 pwq->work_color = wq->work_color;
3728 /* sync max_active to the current setting */
3729 pwq_adjust_max_active(pwq);
3732 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3735 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3736 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3737 const struct workqueue_attrs *attrs)
3739 struct worker_pool *pool;
3740 struct pool_workqueue *pwq;
3742 lockdep_assert_held(&wq_pool_mutex);
3744 pool = get_unbound_pool(attrs);
3748 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3750 put_unbound_pool(pool);
3754 init_pwq(pwq, wq, pool);
3758 /* undo alloc_unbound_pwq(), used only in the error path */
3759 static void free_unbound_pwq(struct pool_workqueue *pwq)
3761 lockdep_assert_held(&wq_pool_mutex);
3764 put_unbound_pool(pwq->pool);
3765 kmem_cache_free(pwq_cache, pwq);
3770 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3771 * @attrs: the wq_attrs of interest
3772 * @node: the target NUMA node
3773 * @cpu_going_down: if >= 0, the CPU to consider as offline
3774 * @cpumask: outarg, the resulting cpumask
3776 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3777 * @cpu_going_down is >= 0, that cpu is considered offline during
3778 * calculation. The result is stored in @cpumask. This function returns
3779 * %true if the resulting @cpumask is different from @attrs->cpumask,
3782 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3783 * enabled and @node has online CPUs requested by @attrs, the returned
3784 * cpumask is the intersection of the possible CPUs of @node and
3787 * The caller is responsible for ensuring that the cpumask of @node stays
3790 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3791 int cpu_going_down, cpumask_t *cpumask)
3793 if (!wq_numa_enabled || attrs->no_numa)
3796 /* does @node have any online CPUs @attrs wants? */
3797 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3798 if (cpu_going_down >= 0)
3799 cpumask_clear_cpu(cpu_going_down, cpumask);
3801 if (cpumask_empty(cpumask))
3804 /* yeap, return possible CPUs in @node that @attrs wants */
3805 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3806 return !cpumask_equal(cpumask, attrs->cpumask);
3809 cpumask_copy(cpumask, attrs->cpumask);
3813 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3814 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3816 struct pool_workqueue *pwq)
3818 struct pool_workqueue *old_pwq;
3820 lockdep_assert_held(&wq->mutex);
3822 /* link_pwq() can handle duplicate calls */
3825 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3826 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3831 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3832 * @wq: the target workqueue
3833 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3835 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3836 * machines, this function maps a separate pwq to each NUMA node with
3837 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3838 * NUMA node it was issued on. Older pwqs are released as in-flight work
3839 * items finish. Note that a work item which repeatedly requeues itself
3840 * back-to-back will stay on its current pwq.
3842 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3845 int apply_workqueue_attrs(struct workqueue_struct *wq,
3846 const struct workqueue_attrs *attrs)
3848 struct workqueue_attrs *new_attrs, *tmp_attrs;
3849 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3852 /* only unbound workqueues can change attributes */
3853 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3856 /* creating multiple pwqs breaks ordering guarantee */
3857 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3860 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3861 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3862 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3863 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3866 /* make a copy of @attrs and sanitize it */
3867 copy_workqueue_attrs(new_attrs, attrs);
3868 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3871 * We may create multiple pwqs with differing cpumasks. Make a
3872 * copy of @new_attrs which will be modified and used to obtain
3875 copy_workqueue_attrs(tmp_attrs, new_attrs);
3878 * CPUs should stay stable across pwq creations and installations.
3879 * Pin CPUs, determine the target cpumask for each node and create
3884 mutex_lock(&wq_pool_mutex);
3887 * If something goes wrong during CPU up/down, we'll fall back to
3888 * the default pwq covering whole @attrs->cpumask. Always create
3889 * it even if we don't use it immediately.
3891 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3895 for_each_node(node) {
3896 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3897 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3902 pwq_tbl[node] = dfl_pwq;
3906 mutex_unlock(&wq_pool_mutex);
3908 /* all pwqs have been created successfully, let's install'em */
3909 mutex_lock(&wq->mutex);
3911 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3913 /* save the previous pwq and install the new one */
3915 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3917 /* @dfl_pwq might not have been used, ensure it's linked */
3919 swap(wq->dfl_pwq, dfl_pwq);
3921 mutex_unlock(&wq->mutex);
3923 /* put the old pwqs */
3925 put_pwq_unlocked(pwq_tbl[node]);
3926 put_pwq_unlocked(dfl_pwq);
3932 free_workqueue_attrs(tmp_attrs);
3933 free_workqueue_attrs(new_attrs);
3938 free_unbound_pwq(dfl_pwq);
3940 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3941 free_unbound_pwq(pwq_tbl[node]);
3942 mutex_unlock(&wq_pool_mutex);
3950 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3951 * @wq: the target workqueue
3952 * @cpu: the CPU coming up or going down
3953 * @online: whether @cpu is coming up or going down
3955 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3956 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3959 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3960 * falls back to @wq->dfl_pwq which may not be optimal but is always
3963 * Note that when the last allowed CPU of a NUMA node goes offline for a
3964 * workqueue with a cpumask spanning multiple nodes, the workers which were
3965 * already executing the work items for the workqueue will lose their CPU
3966 * affinity and may execute on any CPU. This is similar to how per-cpu
3967 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3968 * affinity, it's the user's responsibility to flush the work item from
3971 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3974 int node = cpu_to_node(cpu);
3975 int cpu_off = online ? -1 : cpu;
3976 struct pool_workqueue *old_pwq = NULL, *pwq;
3977 struct workqueue_attrs *target_attrs;
3980 lockdep_assert_held(&wq_pool_mutex);
3982 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3986 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3987 * Let's use a preallocated one. The following buf is protected by
3988 * CPU hotplug exclusion.
3990 target_attrs = wq_update_unbound_numa_attrs_buf;
3991 cpumask = target_attrs->cpumask;
3993 mutex_lock(&wq->mutex);
3994 if (wq->unbound_attrs->no_numa)
3997 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3998 pwq = unbound_pwq_by_node(wq, node);
4001 * Let's determine what needs to be done. If the target cpumask is
4002 * different from wq's, we need to compare it to @pwq's and create
4003 * a new one if they don't match. If the target cpumask equals
4004 * wq's, the default pwq should be used. If @pwq is already the
4005 * default one, nothing to do; otherwise, install the default one.
4007 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4008 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4011 if (pwq == wq->dfl_pwq)
4017 mutex_unlock(&wq->mutex);
4019 /* create a new pwq */
4020 pwq = alloc_unbound_pwq(wq, target_attrs);
4022 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4028 * Install the new pwq. As this function is called only from CPU
4029 * hotplug callbacks and applying a new attrs is wrapped with
4030 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4033 mutex_lock(&wq->mutex);
4034 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4038 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4039 get_pwq(wq->dfl_pwq);
4040 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4041 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4043 mutex_unlock(&wq->mutex);
4044 put_pwq_unlocked(old_pwq);
4047 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4049 bool highpri = wq->flags & WQ_HIGHPRI;
4052 if (!(wq->flags & WQ_UNBOUND)) {
4053 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4057 for_each_possible_cpu(cpu) {
4058 struct pool_workqueue *pwq =
4059 per_cpu_ptr(wq->cpu_pwqs, cpu);
4060 struct worker_pool *cpu_pools =
4061 per_cpu(cpu_worker_pools, cpu);
4063 init_pwq(pwq, wq, &cpu_pools[highpri]);
4065 mutex_lock(&wq->mutex);
4067 mutex_unlock(&wq->mutex);
4071 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4075 static int wq_clamp_max_active(int max_active, unsigned int flags,
4078 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4080 if (max_active < 1 || max_active > lim)
4081 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4082 max_active, name, 1, lim);
4084 return clamp_val(max_active, 1, lim);
4087 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4090 struct lock_class_key *key,
4091 const char *lock_name, ...)
4093 size_t tbl_size = 0;
4095 struct workqueue_struct *wq;
4096 struct pool_workqueue *pwq;
4098 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4099 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4100 flags |= WQ_UNBOUND;
4102 /* allocate wq and format name */
4103 if (flags & WQ_UNBOUND)
4104 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4106 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4110 if (flags & WQ_UNBOUND) {
4111 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4112 if (!wq->unbound_attrs)
4116 va_start(args, lock_name);
4117 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4120 max_active = max_active ?: WQ_DFL_ACTIVE;
4121 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4125 wq->saved_max_active = max_active;
4126 mutex_init(&wq->mutex);
4127 atomic_set(&wq->nr_pwqs_to_flush, 0);
4128 INIT_LIST_HEAD(&wq->pwqs);
4129 INIT_LIST_HEAD(&wq->flusher_queue);
4130 INIT_LIST_HEAD(&wq->flusher_overflow);
4131 INIT_LIST_HEAD(&wq->maydays);
4133 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4134 INIT_LIST_HEAD(&wq->list);
4136 if (alloc_and_link_pwqs(wq) < 0)
4140 * Workqueues which may be used during memory reclaim should
4141 * have a rescuer to guarantee forward progress.
4143 if (flags & WQ_MEM_RECLAIM) {
4144 struct worker *rescuer;
4146 rescuer = alloc_worker();
4150 rescuer->rescue_wq = wq;
4151 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4153 if (IS_ERR(rescuer->task)) {
4158 wq->rescuer = rescuer;
4159 rescuer->task->flags |= PF_NO_SETAFFINITY;
4160 wake_up_process(rescuer->task);
4163 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4167 * wq_pool_mutex protects global freeze state and workqueues list.
4168 * Grab it, adjust max_active and add the new @wq to workqueues
4171 mutex_lock(&wq_pool_mutex);
4173 mutex_lock(&wq->mutex);
4174 for_each_pwq(pwq, wq)
4175 pwq_adjust_max_active(pwq);
4176 mutex_unlock(&wq->mutex);
4178 list_add(&wq->list, &workqueues);
4180 mutex_unlock(&wq_pool_mutex);
4185 free_workqueue_attrs(wq->unbound_attrs);
4189 destroy_workqueue(wq);
4192 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4195 * destroy_workqueue - safely terminate a workqueue
4196 * @wq: target workqueue
4198 * Safely destroy a workqueue. All work currently pending will be done first.
4200 void destroy_workqueue(struct workqueue_struct *wq)
4202 struct pool_workqueue *pwq;
4205 /* drain it before proceeding with destruction */
4206 drain_workqueue(wq);
4209 mutex_lock(&wq->mutex);
4210 for_each_pwq(pwq, wq) {
4213 for (i = 0; i < WORK_NR_COLORS; i++) {
4214 if (WARN_ON(pwq->nr_in_flight[i])) {
4215 mutex_unlock(&wq->mutex);
4220 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4221 WARN_ON(pwq->nr_active) ||
4222 WARN_ON(!list_empty(&pwq->delayed_works))) {
4223 mutex_unlock(&wq->mutex);
4227 mutex_unlock(&wq->mutex);
4230 * wq list is used to freeze wq, remove from list after
4231 * flushing is complete in case freeze races us.
4233 mutex_lock(&wq_pool_mutex);
4234 list_del_init(&wq->list);
4235 mutex_unlock(&wq_pool_mutex);
4237 workqueue_sysfs_unregister(wq);
4240 kthread_stop(wq->rescuer->task);
4245 if (!(wq->flags & WQ_UNBOUND)) {
4247 * The base ref is never dropped on per-cpu pwqs. Directly
4248 * free the pwqs and wq.
4250 free_percpu(wq->cpu_pwqs);
4254 * We're the sole accessor of @wq at this point. Directly
4255 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4256 * @wq will be freed when the last pwq is released.
4258 for_each_node(node) {
4259 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4260 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4261 put_pwq_unlocked(pwq);
4265 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4266 * put. Don't access it afterwards.
4270 put_pwq_unlocked(pwq);
4273 EXPORT_SYMBOL_GPL(destroy_workqueue);
4276 * workqueue_set_max_active - adjust max_active of a workqueue
4277 * @wq: target workqueue
4278 * @max_active: new max_active value.
4280 * Set max_active of @wq to @max_active.
4283 * Don't call from IRQ context.
4285 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4287 struct pool_workqueue *pwq;
4289 /* disallow meddling with max_active for ordered workqueues */
4290 if (WARN_ON(wq->flags & __WQ_ORDERED))
4293 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4295 mutex_lock(&wq->mutex);
4297 wq->saved_max_active = max_active;
4299 for_each_pwq(pwq, wq)
4300 pwq_adjust_max_active(pwq);
4302 mutex_unlock(&wq->mutex);
4304 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4307 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4309 * Determine whether %current is a workqueue rescuer. Can be used from
4310 * work functions to determine whether it's being run off the rescuer task.
4312 bool current_is_workqueue_rescuer(void)
4314 struct worker *worker = current_wq_worker();
4316 return worker && worker->rescue_wq;
4320 * workqueue_congested - test whether a workqueue is congested
4321 * @cpu: CPU in question
4322 * @wq: target workqueue
4324 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4325 * no synchronization around this function and the test result is
4326 * unreliable and only useful as advisory hints or for debugging.
4328 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4329 * Note that both per-cpu and unbound workqueues may be associated with
4330 * multiple pool_workqueues which have separate congested states. A
4331 * workqueue being congested on one CPU doesn't mean the workqueue is also
4332 * contested on other CPUs / NUMA nodes.
4335 * %true if congested, %false otherwise.
4337 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4339 struct pool_workqueue *pwq;
4342 rcu_read_lock_sched();
4344 if (cpu == WORK_CPU_UNBOUND)
4345 cpu = smp_processor_id();
4347 if (!(wq->flags & WQ_UNBOUND))
4348 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4350 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4352 ret = !list_empty(&pwq->delayed_works);
4353 rcu_read_unlock_sched();
4357 EXPORT_SYMBOL_GPL(workqueue_congested);
4360 * work_busy - test whether a work is currently pending or running
4361 * @work: the work to be tested
4363 * Test whether @work is currently pending or running. There is no
4364 * synchronization around this function and the test result is
4365 * unreliable and only useful as advisory hints or for debugging.
4368 * OR'd bitmask of WORK_BUSY_* bits.
4370 unsigned int work_busy(struct work_struct *work)
4372 struct worker_pool *pool;
4373 unsigned long flags;
4374 unsigned int ret = 0;
4376 if (work_pending(work))
4377 ret |= WORK_BUSY_PENDING;
4379 local_irq_save(flags);
4380 pool = get_work_pool(work);
4382 spin_lock(&pool->lock);
4383 if (find_worker_executing_work(pool, work))
4384 ret |= WORK_BUSY_RUNNING;
4385 spin_unlock(&pool->lock);
4387 local_irq_restore(flags);
4391 EXPORT_SYMBOL_GPL(work_busy);
4394 * set_worker_desc - set description for the current work item
4395 * @fmt: printf-style format string
4396 * @...: arguments for the format string
4398 * This function can be called by a running work function to describe what
4399 * the work item is about. If the worker task gets dumped, this
4400 * information will be printed out together to help debugging. The
4401 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4403 void set_worker_desc(const char *fmt, ...)
4405 struct worker *worker = current_wq_worker();
4409 va_start(args, fmt);
4410 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4412 worker->desc_valid = true;
4417 * print_worker_info - print out worker information and description
4418 * @log_lvl: the log level to use when printing
4419 * @task: target task
4421 * If @task is a worker and currently executing a work item, print out the
4422 * name of the workqueue being serviced and worker description set with
4423 * set_worker_desc() by the currently executing work item.
4425 * This function can be safely called on any task as long as the
4426 * task_struct itself is accessible. While safe, this function isn't
4427 * synchronized and may print out mixups or garbages of limited length.
4429 void print_worker_info(const char *log_lvl, struct task_struct *task)
4431 work_func_t *fn = NULL;
4432 char name[WQ_NAME_LEN] = { };
4433 char desc[WORKER_DESC_LEN] = { };
4434 struct pool_workqueue *pwq = NULL;
4435 struct workqueue_struct *wq = NULL;
4436 bool desc_valid = false;
4437 struct worker *worker;
4439 if (!(task->flags & PF_WQ_WORKER))
4443 * This function is called without any synchronization and @task
4444 * could be in any state. Be careful with dereferences.
4446 worker = probe_kthread_data(task);
4449 * Carefully copy the associated workqueue's workfn and name. Keep
4450 * the original last '\0' in case the original contains garbage.
4452 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4453 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4454 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4455 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4457 /* copy worker description */
4458 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4460 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4462 if (fn || name[0] || desc[0]) {
4463 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4465 pr_cont(" (%s)", desc);
4473 * There are two challenges in supporting CPU hotplug. Firstly, there
4474 * are a lot of assumptions on strong associations among work, pwq and
4475 * pool which make migrating pending and scheduled works very
4476 * difficult to implement without impacting hot paths. Secondly,
4477 * worker pools serve mix of short, long and very long running works making
4478 * blocked draining impractical.
4480 * This is solved by allowing the pools to be disassociated from the CPU
4481 * running as an unbound one and allowing it to be reattached later if the
4482 * cpu comes back online.
4485 static void wq_unbind_fn(struct work_struct *work)
4487 int cpu = smp_processor_id();
4488 struct worker_pool *pool;
4489 struct worker *worker;
4492 for_each_cpu_worker_pool(pool, cpu) {
4493 WARN_ON_ONCE(cpu != smp_processor_id());
4495 mutex_lock(&pool->manager_mutex);
4496 spin_lock_irq(&pool->lock);
4499 * We've blocked all manager operations. Make all workers
4500 * unbound and set DISASSOCIATED. Before this, all workers
4501 * except for the ones which are still executing works from
4502 * before the last CPU down must be on the cpu. After
4503 * this, they may become diasporas.
4505 for_each_pool_worker(worker, wi, pool)
4506 worker->flags |= WORKER_UNBOUND;
4508 pool->flags |= POOL_DISASSOCIATED;
4510 spin_unlock_irq(&pool->lock);
4511 mutex_unlock(&pool->manager_mutex);
4514 * Call schedule() so that we cross rq->lock and thus can
4515 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4516 * This is necessary as scheduler callbacks may be invoked
4522 * Sched callbacks are disabled now. Zap nr_running.
4523 * After this, nr_running stays zero and need_more_worker()
4524 * and keep_working() are always true as long as the
4525 * worklist is not empty. This pool now behaves as an
4526 * unbound (in terms of concurrency management) pool which
4527 * are served by workers tied to the pool.
4529 atomic_set(&pool->nr_running, 0);
4532 * With concurrency management just turned off, a busy
4533 * worker blocking could lead to lengthy stalls. Kick off
4534 * unbound chain execution of currently pending work items.
4536 spin_lock_irq(&pool->lock);
4537 wake_up_worker(pool);
4538 spin_unlock_irq(&pool->lock);
4543 * rebind_workers - rebind all workers of a pool to the associated CPU
4544 * @pool: pool of interest
4546 * @pool->cpu is coming online. Rebind all workers to the CPU.
4548 static void rebind_workers(struct worker_pool *pool)
4550 struct worker *worker;
4553 lockdep_assert_held(&pool->manager_mutex);
4556 * Restore CPU affinity of all workers. As all idle workers should
4557 * be on the run-queue of the associated CPU before any local
4558 * wake-ups for concurrency management happen, restore CPU affinty
4559 * of all workers first and then clear UNBOUND. As we're called
4560 * from CPU_ONLINE, the following shouldn't fail.
4562 for_each_pool_worker(worker, wi, pool)
4563 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4564 pool->attrs->cpumask) < 0);
4566 spin_lock_irq(&pool->lock);
4568 for_each_pool_worker(worker, wi, pool) {
4569 unsigned int worker_flags = worker->flags;
4572 * A bound idle worker should actually be on the runqueue
4573 * of the associated CPU for local wake-ups targeting it to
4574 * work. Kick all idle workers so that they migrate to the
4575 * associated CPU. Doing this in the same loop as
4576 * replacing UNBOUND with REBOUND is safe as no worker will
4577 * be bound before @pool->lock is released.
4579 if (worker_flags & WORKER_IDLE)
4580 wake_up_process(worker->task);
4583 * We want to clear UNBOUND but can't directly call
4584 * worker_clr_flags() or adjust nr_running. Atomically
4585 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4586 * @worker will clear REBOUND using worker_clr_flags() when
4587 * it initiates the next execution cycle thus restoring
4588 * concurrency management. Note that when or whether
4589 * @worker clears REBOUND doesn't affect correctness.
4591 * ACCESS_ONCE() is necessary because @worker->flags may be
4592 * tested without holding any lock in
4593 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4594 * fail incorrectly leading to premature concurrency
4595 * management operations.
4597 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4598 worker_flags |= WORKER_REBOUND;
4599 worker_flags &= ~WORKER_UNBOUND;
4600 ACCESS_ONCE(worker->flags) = worker_flags;
4603 spin_unlock_irq(&pool->lock);
4607 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4608 * @pool: unbound pool of interest
4609 * @cpu: the CPU which is coming up
4611 * An unbound pool may end up with a cpumask which doesn't have any online
4612 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4613 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4614 * online CPU before, cpus_allowed of all its workers should be restored.
4616 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4618 static cpumask_t cpumask;
4619 struct worker *worker;
4622 lockdep_assert_held(&pool->manager_mutex);
4624 /* is @cpu allowed for @pool? */
4625 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4628 /* is @cpu the only online CPU? */
4629 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4630 if (cpumask_weight(&cpumask) != 1)
4633 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4634 for_each_pool_worker(worker, wi, pool)
4635 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4636 pool->attrs->cpumask) < 0);
4640 * Workqueues should be brought up before normal priority CPU notifiers.
4641 * This will be registered high priority CPU notifier.
4643 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4644 unsigned long action,
4647 int cpu = (unsigned long)hcpu;
4648 struct worker_pool *pool;
4649 struct workqueue_struct *wq;
4652 switch (action & ~CPU_TASKS_FROZEN) {
4653 case CPU_UP_PREPARE:
4654 for_each_cpu_worker_pool(pool, cpu) {
4655 if (pool->nr_workers)
4657 if (create_and_start_worker(pool) < 0)
4662 case CPU_DOWN_FAILED:
4664 mutex_lock(&wq_pool_mutex);
4666 for_each_pool(pool, pi) {
4667 mutex_lock(&pool->manager_mutex);
4669 if (pool->cpu == cpu) {
4670 spin_lock_irq(&pool->lock);
4671 pool->flags &= ~POOL_DISASSOCIATED;
4672 spin_unlock_irq(&pool->lock);
4674 rebind_workers(pool);
4675 } else if (pool->cpu < 0) {
4676 restore_unbound_workers_cpumask(pool, cpu);
4679 mutex_unlock(&pool->manager_mutex);
4682 /* update NUMA affinity of unbound workqueues */
4683 list_for_each_entry(wq, &workqueues, list)
4684 wq_update_unbound_numa(wq, cpu, true);
4686 mutex_unlock(&wq_pool_mutex);
4693 * Workqueues should be brought down after normal priority CPU notifiers.
4694 * This will be registered as low priority CPU notifier.
4696 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4697 unsigned long action,
4700 int cpu = (unsigned long)hcpu;
4701 struct work_struct unbind_work;
4702 struct workqueue_struct *wq;
4704 switch (action & ~CPU_TASKS_FROZEN) {
4705 case CPU_DOWN_PREPARE:
4706 /* unbinding per-cpu workers should happen on the local CPU */
4707 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4708 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4710 /* update NUMA affinity of unbound workqueues */
4711 mutex_lock(&wq_pool_mutex);
4712 list_for_each_entry(wq, &workqueues, list)
4713 wq_update_unbound_numa(wq, cpu, false);
4714 mutex_unlock(&wq_pool_mutex);
4716 /* wait for per-cpu unbinding to finish */
4717 flush_work(&unbind_work);
4725 struct work_for_cpu {
4726 struct work_struct work;
4732 static void work_for_cpu_fn(struct work_struct *work)
4734 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4736 wfc->ret = wfc->fn(wfc->arg);
4740 * work_on_cpu - run a function in user context on a particular cpu
4741 * @cpu: the cpu to run on
4742 * @fn: the function to run
4743 * @arg: the function arg
4745 * This will return the value @fn returns.
4746 * It is up to the caller to ensure that the cpu doesn't go offline.
4747 * The caller must not hold any locks which would prevent @fn from completing.
4749 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4751 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4753 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4754 schedule_work_on(cpu, &wfc.work);
4755 flush_work(&wfc.work);
4758 EXPORT_SYMBOL_GPL(work_on_cpu);
4759 #endif /* CONFIG_SMP */
4761 #ifdef CONFIG_FREEZER
4764 * freeze_workqueues_begin - begin freezing workqueues
4766 * Start freezing workqueues. After this function returns, all freezable
4767 * workqueues will queue new works to their delayed_works list instead of
4771 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4773 void freeze_workqueues_begin(void)
4775 struct worker_pool *pool;
4776 struct workqueue_struct *wq;
4777 struct pool_workqueue *pwq;
4780 mutex_lock(&wq_pool_mutex);
4782 WARN_ON_ONCE(workqueue_freezing);
4783 workqueue_freezing = true;
4786 for_each_pool(pool, pi) {
4787 spin_lock_irq(&pool->lock);
4788 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4789 pool->flags |= POOL_FREEZING;
4790 spin_unlock_irq(&pool->lock);
4793 list_for_each_entry(wq, &workqueues, list) {
4794 mutex_lock(&wq->mutex);
4795 for_each_pwq(pwq, wq)
4796 pwq_adjust_max_active(pwq);
4797 mutex_unlock(&wq->mutex);
4800 mutex_unlock(&wq_pool_mutex);
4804 * freeze_workqueues_busy - are freezable workqueues still busy?
4806 * Check whether freezing is complete. This function must be called
4807 * between freeze_workqueues_begin() and thaw_workqueues().
4810 * Grabs and releases wq_pool_mutex.
4813 * %true if some freezable workqueues are still busy. %false if freezing
4816 bool freeze_workqueues_busy(void)
4819 struct workqueue_struct *wq;
4820 struct pool_workqueue *pwq;
4822 mutex_lock(&wq_pool_mutex);
4824 WARN_ON_ONCE(!workqueue_freezing);
4826 list_for_each_entry(wq, &workqueues, list) {
4827 if (!(wq->flags & WQ_FREEZABLE))
4830 * nr_active is monotonically decreasing. It's safe
4831 * to peek without lock.
4833 rcu_read_lock_sched();
4834 for_each_pwq(pwq, wq) {
4835 WARN_ON_ONCE(pwq->nr_active < 0);
4836 if (pwq->nr_active) {
4838 rcu_read_unlock_sched();
4842 rcu_read_unlock_sched();
4845 mutex_unlock(&wq_pool_mutex);
4850 * thaw_workqueues - thaw workqueues
4852 * Thaw workqueues. Normal queueing is restored and all collected
4853 * frozen works are transferred to their respective pool worklists.
4856 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4858 void thaw_workqueues(void)
4860 struct workqueue_struct *wq;
4861 struct pool_workqueue *pwq;
4862 struct worker_pool *pool;
4865 mutex_lock(&wq_pool_mutex);
4867 if (!workqueue_freezing)
4870 /* clear FREEZING */
4871 for_each_pool(pool, pi) {
4872 spin_lock_irq(&pool->lock);
4873 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4874 pool->flags &= ~POOL_FREEZING;
4875 spin_unlock_irq(&pool->lock);
4878 /* restore max_active and repopulate worklist */
4879 list_for_each_entry(wq, &workqueues, list) {
4880 mutex_lock(&wq->mutex);
4881 for_each_pwq(pwq, wq)
4882 pwq_adjust_max_active(pwq);
4883 mutex_unlock(&wq->mutex);
4886 workqueue_freezing = false;
4888 mutex_unlock(&wq_pool_mutex);
4890 #endif /* CONFIG_FREEZER */
4892 static void __init wq_numa_init(void)
4897 /* determine NUMA pwq table len - highest node id + 1 */
4899 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4901 if (num_possible_nodes() <= 1)
4904 if (wq_disable_numa) {
4905 pr_info("workqueue: NUMA affinity support disabled\n");
4909 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4910 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4913 * We want masks of possible CPUs of each node which isn't readily
4914 * available. Build one from cpu_to_node() which should have been
4915 * fully initialized by now.
4917 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4921 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4922 node_online(node) ? node : NUMA_NO_NODE));
4924 for_each_possible_cpu(cpu) {
4925 node = cpu_to_node(cpu);
4926 if (WARN_ON(node == NUMA_NO_NODE)) {
4927 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4928 /* happens iff arch is bonkers, let's just proceed */
4931 cpumask_set_cpu(cpu, tbl[node]);
4934 wq_numa_possible_cpumask = tbl;
4935 wq_numa_enabled = true;
4938 static int __init init_workqueues(void)
4940 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4943 /* make sure we have enough bits for OFFQ pool ID */
4944 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4945 WORK_CPU_END * NR_STD_WORKER_POOLS);
4947 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4949 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4951 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4952 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4956 /* initialize CPU pools */
4957 for_each_possible_cpu(cpu) {
4958 struct worker_pool *pool;
4961 for_each_cpu_worker_pool(pool, cpu) {
4962 BUG_ON(init_worker_pool(pool));
4964 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4965 pool->attrs->nice = std_nice[i++];
4966 pool->node = cpu_to_node(cpu);
4969 mutex_lock(&wq_pool_mutex);
4970 BUG_ON(worker_pool_assign_id(pool));
4971 mutex_unlock(&wq_pool_mutex);
4975 /* create the initial worker */
4976 for_each_online_cpu(cpu) {
4977 struct worker_pool *pool;
4979 for_each_cpu_worker_pool(pool, cpu) {
4980 pool->flags &= ~POOL_DISASSOCIATED;
4981 BUG_ON(create_and_start_worker(pool) < 0);
4985 /* create default unbound wq attrs */
4986 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4987 struct workqueue_attrs *attrs;
4989 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4990 attrs->nice = std_nice[i];
4991 unbound_std_wq_attrs[i] = attrs;
4994 system_wq = alloc_workqueue("events", 0, 0);
4995 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4996 system_long_wq = alloc_workqueue("events_long", 0, 0);
4997 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4998 WQ_UNBOUND_MAX_ACTIVE);
4999 system_freezable_wq = alloc_workqueue("events_freezable",
5001 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5002 !system_unbound_wq || !system_freezable_wq);
5005 early_initcall(init_workqueues);