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);
317 struct workqueue_struct *system_power_efficient_wq __read_mostly;
318 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
319 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
320 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
322 static int worker_thread(void *__worker);
323 static void copy_workqueue_attrs(struct workqueue_attrs *to,
324 const struct workqueue_attrs *from);
326 #define CREATE_TRACE_POINTS
327 #include <trace/events/workqueue.h>
329 #define assert_rcu_or_pool_mutex() \
330 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
331 lockdep_is_held(&wq_pool_mutex), \
332 "sched RCU or wq_pool_mutex should be held")
334 #define assert_rcu_or_wq_mutex(wq) \
335 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
336 lockdep_is_held(&wq->mutex), \
337 "sched RCU or wq->mutex should be held")
339 #ifdef CONFIG_LOCKDEP
340 #define assert_manager_or_pool_lock(pool) \
341 WARN_ONCE(debug_locks && \
342 !lockdep_is_held(&(pool)->manager_mutex) && \
343 !lockdep_is_held(&(pool)->lock), \
344 "pool->manager_mutex or ->lock should be held")
346 #define assert_manager_or_pool_lock(pool) do { } while (0)
349 #define for_each_cpu_worker_pool(pool, cpu) \
350 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
351 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
355 * for_each_pool - iterate through all worker_pools in the system
356 * @pool: iteration cursor
357 * @pi: integer used for iteration
359 * This must be called either with wq_pool_mutex held or sched RCU read
360 * locked. If the pool needs to be used beyond the locking in effect, the
361 * caller is responsible for guaranteeing that the pool stays online.
363 * The if/else clause exists only for the lockdep assertion and can be
366 #define for_each_pool(pool, pi) \
367 idr_for_each_entry(&worker_pool_idr, pool, pi) \
368 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
372 * for_each_pool_worker - iterate through all workers of a worker_pool
373 * @worker: iteration cursor
374 * @wi: integer used for iteration
375 * @pool: worker_pool to iterate workers of
377 * This must be called with either @pool->manager_mutex or ->lock held.
379 * The if/else clause exists only for the lockdep assertion and can be
382 #define for_each_pool_worker(worker, wi, pool) \
383 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
384 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
388 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
389 * @pwq: iteration cursor
390 * @wq: the target workqueue
392 * This must be called either with wq->mutex held or sched RCU read locked.
393 * If the pwq needs to be used beyond the locking in effect, the caller is
394 * responsible for guaranteeing that the pwq stays online.
396 * The if/else clause exists only for the lockdep assertion and can be
399 #define for_each_pwq(pwq, wq) \
400 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
401 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
404 #ifdef CONFIG_DEBUG_OBJECTS_WORK
406 static struct debug_obj_descr work_debug_descr;
408 static void *work_debug_hint(void *addr)
410 return ((struct work_struct *) addr)->func;
414 * fixup_init is called when:
415 * - an active object is initialized
417 static int work_fixup_init(void *addr, enum debug_obj_state state)
419 struct work_struct *work = addr;
422 case ODEBUG_STATE_ACTIVE:
423 cancel_work_sync(work);
424 debug_object_init(work, &work_debug_descr);
432 * fixup_activate is called when:
433 * - an active object is activated
434 * - an unknown object is activated (might be a statically initialized object)
436 static int work_fixup_activate(void *addr, enum debug_obj_state state)
438 struct work_struct *work = addr;
442 case ODEBUG_STATE_NOTAVAILABLE:
444 * This is not really a fixup. The work struct was
445 * statically initialized. We just make sure that it
446 * is tracked in the object tracker.
448 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
449 debug_object_init(work, &work_debug_descr);
450 debug_object_activate(work, &work_debug_descr);
456 case ODEBUG_STATE_ACTIVE:
465 * fixup_free is called when:
466 * - an active object is freed
468 static int work_fixup_free(void *addr, enum debug_obj_state state)
470 struct work_struct *work = addr;
473 case ODEBUG_STATE_ACTIVE:
474 cancel_work_sync(work);
475 debug_object_free(work, &work_debug_descr);
482 static struct debug_obj_descr work_debug_descr = {
483 .name = "work_struct",
484 .debug_hint = work_debug_hint,
485 .fixup_init = work_fixup_init,
486 .fixup_activate = work_fixup_activate,
487 .fixup_free = work_fixup_free,
490 static inline void debug_work_activate(struct work_struct *work)
492 debug_object_activate(work, &work_debug_descr);
495 static inline void debug_work_deactivate(struct work_struct *work)
497 debug_object_deactivate(work, &work_debug_descr);
500 void __init_work(struct work_struct *work, int onstack)
503 debug_object_init_on_stack(work, &work_debug_descr);
505 debug_object_init(work, &work_debug_descr);
507 EXPORT_SYMBOL_GPL(__init_work);
509 void destroy_work_on_stack(struct work_struct *work)
511 debug_object_free(work, &work_debug_descr);
513 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
516 static inline void debug_work_activate(struct work_struct *work) { }
517 static inline void debug_work_deactivate(struct work_struct *work) { }
520 /* allocate ID and assign it to @pool */
521 static int worker_pool_assign_id(struct worker_pool *pool)
525 lockdep_assert_held(&wq_pool_mutex);
527 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
536 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
537 * @wq: the target workqueue
540 * This must be called either with pwq_lock held or sched RCU read locked.
541 * If the pwq needs to be used beyond the locking in effect, the caller is
542 * responsible for guaranteeing that the pwq stays online.
544 * Return: The unbound pool_workqueue for @node.
546 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
549 assert_rcu_or_wq_mutex(wq);
550 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
553 static unsigned int work_color_to_flags(int color)
555 return color << WORK_STRUCT_COLOR_SHIFT;
558 static int get_work_color(struct work_struct *work)
560 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
561 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
564 static int work_next_color(int color)
566 return (color + 1) % WORK_NR_COLORS;
570 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
571 * contain the pointer to the queued pwq. Once execution starts, the flag
572 * is cleared and the high bits contain OFFQ flags and pool ID.
574 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
575 * and clear_work_data() can be used to set the pwq, pool or clear
576 * work->data. These functions should only be called while the work is
577 * owned - ie. while the PENDING bit is set.
579 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
580 * corresponding to a work. Pool is available once the work has been
581 * queued anywhere after initialization until it is sync canceled. pwq is
582 * available only while the work item is queued.
584 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
585 * canceled. While being canceled, a work item may have its PENDING set
586 * but stay off timer and worklist for arbitrarily long and nobody should
587 * try to steal the PENDING bit.
589 static inline void set_work_data(struct work_struct *work, unsigned long data,
592 WARN_ON_ONCE(!work_pending(work));
593 atomic_long_set(&work->data, data | flags | work_static(work));
596 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
597 unsigned long extra_flags)
599 set_work_data(work, (unsigned long)pwq,
600 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
603 static void set_work_pool_and_keep_pending(struct work_struct *work,
606 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
607 WORK_STRUCT_PENDING);
610 static void set_work_pool_and_clear_pending(struct work_struct *work,
614 * The following wmb is paired with the implied mb in
615 * test_and_set_bit(PENDING) and ensures all updates to @work made
616 * here are visible to and precede any updates by the next PENDING
620 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
623 static void clear_work_data(struct work_struct *work)
625 smp_wmb(); /* see set_work_pool_and_clear_pending() */
626 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
629 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
631 unsigned long data = atomic_long_read(&work->data);
633 if (data & WORK_STRUCT_PWQ)
634 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
640 * get_work_pool - return the worker_pool a given work was associated with
641 * @work: the work item of interest
643 * Pools are created and destroyed under wq_pool_mutex, and allows read
644 * access under sched-RCU read lock. As such, this function should be
645 * called under wq_pool_mutex or with preemption disabled.
647 * All fields of the returned pool are accessible as long as the above
648 * mentioned locking is in effect. If the returned pool needs to be used
649 * beyond the critical section, the caller is responsible for ensuring the
650 * returned pool is and stays online.
652 * Return: The worker_pool @work was last associated with. %NULL if none.
654 static struct worker_pool *get_work_pool(struct work_struct *work)
656 unsigned long data = atomic_long_read(&work->data);
659 assert_rcu_or_pool_mutex();
661 if (data & WORK_STRUCT_PWQ)
662 return ((struct pool_workqueue *)
663 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
665 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
666 if (pool_id == WORK_OFFQ_POOL_NONE)
669 return idr_find(&worker_pool_idr, pool_id);
673 * get_work_pool_id - return the worker pool ID a given work is associated with
674 * @work: the work item of interest
676 * Return: The worker_pool ID @work was last associated with.
677 * %WORK_OFFQ_POOL_NONE if none.
679 static int get_work_pool_id(struct work_struct *work)
681 unsigned long data = atomic_long_read(&work->data);
683 if (data & WORK_STRUCT_PWQ)
684 return ((struct pool_workqueue *)
685 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
687 return data >> WORK_OFFQ_POOL_SHIFT;
690 static void mark_work_canceling(struct work_struct *work)
692 unsigned long pool_id = get_work_pool_id(work);
694 pool_id <<= WORK_OFFQ_POOL_SHIFT;
695 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
698 static bool work_is_canceling(struct work_struct *work)
700 unsigned long data = atomic_long_read(&work->data);
702 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
706 * Policy functions. These define the policies on how the global worker
707 * pools are managed. Unless noted otherwise, these functions assume that
708 * they're being called with pool->lock held.
711 static bool __need_more_worker(struct worker_pool *pool)
713 return !atomic_read(&pool->nr_running);
717 * Need to wake up a worker? Called from anything but currently
720 * Note that, because unbound workers never contribute to nr_running, this
721 * function will always return %true for unbound pools as long as the
722 * worklist isn't empty.
724 static bool need_more_worker(struct worker_pool *pool)
726 return !list_empty(&pool->worklist) && __need_more_worker(pool);
729 /* Can I start working? Called from busy but !running workers. */
730 static bool may_start_working(struct worker_pool *pool)
732 return pool->nr_idle;
735 /* Do I need to keep working? Called from currently running workers. */
736 static bool keep_working(struct worker_pool *pool)
738 return !list_empty(&pool->worklist) &&
739 atomic_read(&pool->nr_running) <= 1;
742 /* Do we need a new worker? Called from manager. */
743 static bool need_to_create_worker(struct worker_pool *pool)
745 return need_more_worker(pool) && !may_start_working(pool);
748 /* Do I need to be the manager? */
749 static bool need_to_manage_workers(struct worker_pool *pool)
751 return need_to_create_worker(pool) ||
752 (pool->flags & POOL_MANAGE_WORKERS);
755 /* Do we have too many workers and should some go away? */
756 static bool too_many_workers(struct worker_pool *pool)
758 bool managing = mutex_is_locked(&pool->manager_arb);
759 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
760 int nr_busy = pool->nr_workers - nr_idle;
763 * nr_idle and idle_list may disagree if idle rebinding is in
764 * progress. Never return %true if idle_list is empty.
766 if (list_empty(&pool->idle_list))
769 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
776 /* Return the first worker. Safe with preemption disabled */
777 static struct worker *first_worker(struct worker_pool *pool)
779 if (unlikely(list_empty(&pool->idle_list)))
782 return list_first_entry(&pool->idle_list, struct worker, entry);
786 * wake_up_worker - wake up an idle worker
787 * @pool: worker pool to wake worker from
789 * Wake up the first idle worker of @pool.
792 * spin_lock_irq(pool->lock).
794 static void wake_up_worker(struct worker_pool *pool)
796 struct worker *worker = first_worker(pool);
799 wake_up_process(worker->task);
803 * wq_worker_waking_up - a worker is waking up
804 * @task: task waking up
805 * @cpu: CPU @task is waking up to
807 * This function is called during try_to_wake_up() when a worker is
811 * spin_lock_irq(rq->lock)
813 void wq_worker_waking_up(struct task_struct *task, int cpu)
815 struct worker *worker = kthread_data(task);
817 if (!(worker->flags & WORKER_NOT_RUNNING)) {
818 WARN_ON_ONCE(worker->pool->cpu != cpu);
819 atomic_inc(&worker->pool->nr_running);
824 * wq_worker_sleeping - a worker is going to sleep
825 * @task: task going to sleep
826 * @cpu: CPU in question, must be the current CPU number
828 * This function is called during schedule() when a busy worker is
829 * going to sleep. Worker on the same cpu can be woken up by
830 * returning pointer to its task.
833 * spin_lock_irq(rq->lock)
836 * Worker task on @cpu to wake up, %NULL if none.
838 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
840 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
841 struct worker_pool *pool;
844 * Rescuers, which may not have all the fields set up like normal
845 * workers, also reach here, let's not access anything before
846 * checking NOT_RUNNING.
848 if (worker->flags & WORKER_NOT_RUNNING)
853 /* this can only happen on the local cpu */
854 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
858 * The counterpart of the following dec_and_test, implied mb,
859 * worklist not empty test sequence is in insert_work().
860 * Please read comment there.
862 * NOT_RUNNING is clear. This means that we're bound to and
863 * running on the local cpu w/ rq lock held and preemption
864 * disabled, which in turn means that none else could be
865 * manipulating idle_list, so dereferencing idle_list without pool
868 if (atomic_dec_and_test(&pool->nr_running) &&
869 !list_empty(&pool->worklist))
870 to_wakeup = first_worker(pool);
871 return to_wakeup ? to_wakeup->task : NULL;
875 * worker_set_flags - set worker flags and adjust nr_running accordingly
877 * @flags: flags to set
878 * @wakeup: wakeup an idle worker if necessary
880 * Set @flags in @worker->flags and adjust nr_running accordingly. If
881 * nr_running becomes zero and @wakeup is %true, an idle worker is
885 * spin_lock_irq(pool->lock)
887 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
890 struct worker_pool *pool = worker->pool;
892 WARN_ON_ONCE(worker->task != current);
895 * If transitioning into NOT_RUNNING, adjust nr_running and
896 * wake up an idle worker as necessary if requested by
899 if ((flags & WORKER_NOT_RUNNING) &&
900 !(worker->flags & WORKER_NOT_RUNNING)) {
902 if (atomic_dec_and_test(&pool->nr_running) &&
903 !list_empty(&pool->worklist))
904 wake_up_worker(pool);
906 atomic_dec(&pool->nr_running);
909 worker->flags |= flags;
913 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
915 * @flags: flags to clear
917 * Clear @flags in @worker->flags and adjust nr_running accordingly.
920 * spin_lock_irq(pool->lock)
922 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
924 struct worker_pool *pool = worker->pool;
925 unsigned int oflags = worker->flags;
927 WARN_ON_ONCE(worker->task != current);
929 worker->flags &= ~flags;
932 * If transitioning out of NOT_RUNNING, increment nr_running. Note
933 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
934 * of multiple flags, not a single flag.
936 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
937 if (!(worker->flags & WORKER_NOT_RUNNING))
938 atomic_inc(&pool->nr_running);
942 * find_worker_executing_work - find worker which is executing a work
943 * @pool: pool of interest
944 * @work: work to find worker for
946 * Find a worker which is executing @work on @pool by searching
947 * @pool->busy_hash which is keyed by the address of @work. For a worker
948 * to match, its current execution should match the address of @work and
949 * its work function. This is to avoid unwanted dependency between
950 * unrelated work executions through a work item being recycled while still
953 * This is a bit tricky. A work item may be freed once its execution
954 * starts and nothing prevents the freed area from being recycled for
955 * another work item. If the same work item address ends up being reused
956 * before the original execution finishes, workqueue will identify the
957 * recycled work item as currently executing and make it wait until the
958 * current execution finishes, introducing an unwanted dependency.
960 * This function checks the work item address and work function to avoid
961 * false positives. Note that this isn't complete as one may construct a
962 * work function which can introduce dependency onto itself through a
963 * recycled work item. Well, if somebody wants to shoot oneself in the
964 * foot that badly, there's only so much we can do, and if such deadlock
965 * actually occurs, it should be easy to locate the culprit work function.
968 * spin_lock_irq(pool->lock).
971 * Pointer to worker which is executing @work if found, %NULL
974 static struct worker *find_worker_executing_work(struct worker_pool *pool,
975 struct work_struct *work)
977 struct worker *worker;
979 hash_for_each_possible(pool->busy_hash, worker, hentry,
981 if (worker->current_work == work &&
982 worker->current_func == work->func)
989 * move_linked_works - move linked works to a list
990 * @work: start of series of works to be scheduled
991 * @head: target list to append @work to
992 * @nextp: out paramter for nested worklist walking
994 * Schedule linked works starting from @work to @head. Work series to
995 * be scheduled starts at @work and includes any consecutive work with
996 * WORK_STRUCT_LINKED set in its predecessor.
998 * If @nextp is not NULL, it's updated to point to the next work of
999 * the last scheduled work. This allows move_linked_works() to be
1000 * nested inside outer list_for_each_entry_safe().
1003 * spin_lock_irq(pool->lock).
1005 static void move_linked_works(struct work_struct *work, struct list_head *head,
1006 struct work_struct **nextp)
1008 struct work_struct *n;
1011 * Linked worklist will always end before the end of the list,
1012 * use NULL for list head.
1014 list_for_each_entry_safe_from(work, n, NULL, entry) {
1015 list_move_tail(&work->entry, head);
1016 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1021 * If we're already inside safe list traversal and have moved
1022 * multiple works to the scheduled queue, the next position
1023 * needs to be updated.
1030 * get_pwq - get an extra reference on the specified pool_workqueue
1031 * @pwq: pool_workqueue to get
1033 * Obtain an extra reference on @pwq. The caller should guarantee that
1034 * @pwq has positive refcnt and be holding the matching pool->lock.
1036 static void get_pwq(struct pool_workqueue *pwq)
1038 lockdep_assert_held(&pwq->pool->lock);
1039 WARN_ON_ONCE(pwq->refcnt <= 0);
1044 * put_pwq - put a pool_workqueue reference
1045 * @pwq: pool_workqueue to put
1047 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1048 * destruction. The caller should be holding the matching pool->lock.
1050 static void put_pwq(struct pool_workqueue *pwq)
1052 lockdep_assert_held(&pwq->pool->lock);
1053 if (likely(--pwq->refcnt))
1055 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1058 * @pwq can't be released under pool->lock, bounce to
1059 * pwq_unbound_release_workfn(). This never recurses on the same
1060 * pool->lock as this path is taken only for unbound workqueues and
1061 * the release work item is scheduled on a per-cpu workqueue. To
1062 * avoid lockdep warning, unbound pool->locks are given lockdep
1063 * subclass of 1 in get_unbound_pool().
1065 schedule_work(&pwq->unbound_release_work);
1069 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1070 * @pwq: pool_workqueue to put (can be %NULL)
1072 * put_pwq() with locking. This function also allows %NULL @pwq.
1074 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1078 * As both pwqs and pools are sched-RCU protected, the
1079 * following lock operations are safe.
1081 spin_lock_irq(&pwq->pool->lock);
1083 spin_unlock_irq(&pwq->pool->lock);
1087 static void pwq_activate_delayed_work(struct work_struct *work)
1089 struct pool_workqueue *pwq = get_work_pwq(work);
1091 trace_workqueue_activate_work(work);
1092 move_linked_works(work, &pwq->pool->worklist, NULL);
1093 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1097 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1099 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1100 struct work_struct, entry);
1102 pwq_activate_delayed_work(work);
1106 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1107 * @pwq: pwq of interest
1108 * @color: color of work which left the queue
1110 * A work either has completed or is removed from pending queue,
1111 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1114 * spin_lock_irq(pool->lock).
1116 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1118 /* uncolored work items don't participate in flushing or nr_active */
1119 if (color == WORK_NO_COLOR)
1122 pwq->nr_in_flight[color]--;
1125 if (!list_empty(&pwq->delayed_works)) {
1126 /* one down, submit a delayed one */
1127 if (pwq->nr_active < pwq->max_active)
1128 pwq_activate_first_delayed(pwq);
1131 /* is flush in progress and are we at the flushing tip? */
1132 if (likely(pwq->flush_color != color))
1135 /* are there still in-flight works? */
1136 if (pwq->nr_in_flight[color])
1139 /* this pwq is done, clear flush_color */
1140 pwq->flush_color = -1;
1143 * If this was the last pwq, wake up the first flusher. It
1144 * will handle the rest.
1146 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1147 complete(&pwq->wq->first_flusher->done);
1153 * try_to_grab_pending - steal work item from worklist and disable irq
1154 * @work: work item to steal
1155 * @is_dwork: @work is a delayed_work
1156 * @flags: place to store irq state
1158 * Try to grab PENDING bit of @work. This function can handle @work in any
1159 * stable state - idle, on timer or on worklist.
1162 * 1 if @work was pending and we successfully stole PENDING
1163 * 0 if @work was idle and we claimed PENDING
1164 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1165 * -ENOENT if someone else is canceling @work, this state may persist
1166 * for arbitrarily long
1169 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1170 * interrupted while holding PENDING and @work off queue, irq must be
1171 * disabled on entry. This, combined with delayed_work->timer being
1172 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1174 * On successful return, >= 0, irq is disabled and the caller is
1175 * responsible for releasing it using local_irq_restore(*@flags).
1177 * This function is safe to call from any context including IRQ handler.
1179 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1180 unsigned long *flags)
1182 struct worker_pool *pool;
1183 struct pool_workqueue *pwq;
1185 local_irq_save(*flags);
1187 /* try to steal the timer if it exists */
1189 struct delayed_work *dwork = to_delayed_work(work);
1192 * dwork->timer is irqsafe. If del_timer() fails, it's
1193 * guaranteed that the timer is not queued anywhere and not
1194 * running on the local CPU.
1196 if (likely(del_timer(&dwork->timer)))
1200 /* try to claim PENDING the normal way */
1201 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1205 * The queueing is in progress, or it is already queued. Try to
1206 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1208 pool = get_work_pool(work);
1212 spin_lock(&pool->lock);
1214 * work->data is guaranteed to point to pwq only while the work
1215 * item is queued on pwq->wq, and both updating work->data to point
1216 * to pwq on queueing and to pool on dequeueing are done under
1217 * pwq->pool->lock. This in turn guarantees that, if work->data
1218 * points to pwq which is associated with a locked pool, the work
1219 * item is currently queued on that pool.
1221 pwq = get_work_pwq(work);
1222 if (pwq && pwq->pool == pool) {
1223 debug_work_deactivate(work);
1226 * A delayed work item cannot be grabbed directly because
1227 * it might have linked NO_COLOR work items which, if left
1228 * on the delayed_list, will confuse pwq->nr_active
1229 * management later on and cause stall. Make sure the work
1230 * item is activated before grabbing.
1232 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1233 pwq_activate_delayed_work(work);
1235 list_del_init(&work->entry);
1236 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1238 /* work->data points to pwq iff queued, point to pool */
1239 set_work_pool_and_keep_pending(work, pool->id);
1241 spin_unlock(&pool->lock);
1244 spin_unlock(&pool->lock);
1246 local_irq_restore(*flags);
1247 if (work_is_canceling(work))
1254 * insert_work - insert a work into a pool
1255 * @pwq: pwq @work belongs to
1256 * @work: work to insert
1257 * @head: insertion point
1258 * @extra_flags: extra WORK_STRUCT_* flags to set
1260 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1261 * work_struct flags.
1264 * spin_lock_irq(pool->lock).
1266 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1267 struct list_head *head, unsigned int extra_flags)
1269 struct worker_pool *pool = pwq->pool;
1271 /* we own @work, set data and link */
1272 set_work_pwq(work, pwq, extra_flags);
1273 list_add_tail(&work->entry, head);
1277 * Ensure either wq_worker_sleeping() sees the above
1278 * list_add_tail() or we see zero nr_running to avoid workers lying
1279 * around lazily while there are works to be processed.
1283 if (__need_more_worker(pool))
1284 wake_up_worker(pool);
1288 * Test whether @work is being queued from another work executing on the
1291 static bool is_chained_work(struct workqueue_struct *wq)
1293 struct worker *worker;
1295 worker = current_wq_worker();
1297 * Return %true iff I'm a worker execuing a work item on @wq. If
1298 * I'm @worker, it's safe to dereference it without locking.
1300 return worker && worker->current_pwq->wq == wq;
1303 static void __queue_work(int cpu, struct workqueue_struct *wq,
1304 struct work_struct *work)
1306 struct pool_workqueue *pwq;
1307 struct worker_pool *last_pool;
1308 struct list_head *worklist;
1309 unsigned int work_flags;
1310 unsigned int req_cpu = cpu;
1313 * While a work item is PENDING && off queue, a task trying to
1314 * steal the PENDING will busy-loop waiting for it to either get
1315 * queued or lose PENDING. Grabbing PENDING and queueing should
1316 * happen with IRQ disabled.
1318 WARN_ON_ONCE(!irqs_disabled());
1320 debug_work_activate(work);
1322 /* if dying, only works from the same workqueue are allowed */
1323 if (unlikely(wq->flags & __WQ_DRAINING) &&
1324 WARN_ON_ONCE(!is_chained_work(wq)))
1327 if (req_cpu == WORK_CPU_UNBOUND)
1328 cpu = raw_smp_processor_id();
1330 /* pwq which will be used unless @work is executing elsewhere */
1331 if (!(wq->flags & WQ_UNBOUND))
1332 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1334 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1337 * If @work was previously on a different pool, it might still be
1338 * running there, in which case the work needs to be queued on that
1339 * pool to guarantee non-reentrancy.
1341 last_pool = get_work_pool(work);
1342 if (last_pool && last_pool != pwq->pool) {
1343 struct worker *worker;
1345 spin_lock(&last_pool->lock);
1347 worker = find_worker_executing_work(last_pool, work);
1349 if (worker && worker->current_pwq->wq == wq) {
1350 pwq = worker->current_pwq;
1352 /* meh... not running there, queue here */
1353 spin_unlock(&last_pool->lock);
1354 spin_lock(&pwq->pool->lock);
1357 spin_lock(&pwq->pool->lock);
1361 * pwq is determined and locked. For unbound pools, we could have
1362 * raced with pwq release and it could already be dead. If its
1363 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1364 * without another pwq replacing it in the numa_pwq_tbl or while
1365 * work items are executing on it, so the retrying is guaranteed to
1366 * make forward-progress.
1368 if (unlikely(!pwq->refcnt)) {
1369 if (wq->flags & WQ_UNBOUND) {
1370 spin_unlock(&pwq->pool->lock);
1375 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1379 /* pwq determined, queue */
1380 trace_workqueue_queue_work(req_cpu, pwq, work);
1382 if (WARN_ON(!list_empty(&work->entry))) {
1383 spin_unlock(&pwq->pool->lock);
1387 pwq->nr_in_flight[pwq->work_color]++;
1388 work_flags = work_color_to_flags(pwq->work_color);
1390 if (likely(pwq->nr_active < pwq->max_active)) {
1391 trace_workqueue_activate_work(work);
1393 worklist = &pwq->pool->worklist;
1395 work_flags |= WORK_STRUCT_DELAYED;
1396 worklist = &pwq->delayed_works;
1399 insert_work(pwq, work, worklist, work_flags);
1401 spin_unlock(&pwq->pool->lock);
1405 * queue_work_on - queue work on specific cpu
1406 * @cpu: CPU number to execute work on
1407 * @wq: workqueue to use
1408 * @work: work to queue
1410 * We queue the work to a specific CPU, the caller must ensure it
1413 * Return: %false if @work was already on a queue, %true otherwise.
1415 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1416 struct work_struct *work)
1419 unsigned long flags;
1421 local_irq_save(flags);
1423 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1424 __queue_work(cpu, wq, work);
1428 local_irq_restore(flags);
1431 EXPORT_SYMBOL(queue_work_on);
1433 void delayed_work_timer_fn(unsigned long __data)
1435 struct delayed_work *dwork = (struct delayed_work *)__data;
1437 /* should have been called from irqsafe timer with irq already off */
1438 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1440 EXPORT_SYMBOL(delayed_work_timer_fn);
1442 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1443 struct delayed_work *dwork, unsigned long delay)
1445 struct timer_list *timer = &dwork->timer;
1446 struct work_struct *work = &dwork->work;
1448 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1449 timer->data != (unsigned long)dwork);
1450 WARN_ON_ONCE(timer_pending(timer));
1451 WARN_ON_ONCE(!list_empty(&work->entry));
1454 * If @delay is 0, queue @dwork->work immediately. This is for
1455 * both optimization and correctness. The earliest @timer can
1456 * expire is on the closest next tick and delayed_work users depend
1457 * on that there's no such delay when @delay is 0.
1460 __queue_work(cpu, wq, &dwork->work);
1464 timer_stats_timer_set_start_info(&dwork->timer);
1468 timer->expires = jiffies + delay;
1470 if (unlikely(cpu != WORK_CPU_UNBOUND))
1471 add_timer_on(timer, cpu);
1477 * queue_delayed_work_on - queue work on specific CPU after delay
1478 * @cpu: CPU number to execute work on
1479 * @wq: workqueue to use
1480 * @dwork: work to queue
1481 * @delay: number of jiffies to wait before queueing
1483 * Return: %false if @work was already on a queue, %true otherwise. If
1484 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1487 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1488 struct delayed_work *dwork, unsigned long delay)
1490 struct work_struct *work = &dwork->work;
1492 unsigned long flags;
1494 /* read the comment in __queue_work() */
1495 local_irq_save(flags);
1497 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1498 __queue_delayed_work(cpu, wq, dwork, delay);
1502 local_irq_restore(flags);
1505 EXPORT_SYMBOL(queue_delayed_work_on);
1508 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1509 * @cpu: CPU number to execute work on
1510 * @wq: workqueue to use
1511 * @dwork: work to queue
1512 * @delay: number of jiffies to wait before queueing
1514 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1515 * modify @dwork's timer so that it expires after @delay. If @delay is
1516 * zero, @work is guaranteed to be scheduled immediately regardless of its
1519 * Return: %false if @dwork was idle and queued, %true if @dwork was
1520 * pending and its timer was modified.
1522 * This function is safe to call from any context including IRQ handler.
1523 * See try_to_grab_pending() for details.
1525 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1526 struct delayed_work *dwork, unsigned long delay)
1528 unsigned long flags;
1532 ret = try_to_grab_pending(&dwork->work, true, &flags);
1533 } while (unlikely(ret == -EAGAIN));
1535 if (likely(ret >= 0)) {
1536 __queue_delayed_work(cpu, wq, dwork, delay);
1537 local_irq_restore(flags);
1540 /* -ENOENT from try_to_grab_pending() becomes %true */
1543 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1546 * worker_enter_idle - enter idle state
1547 * @worker: worker which is entering idle state
1549 * @worker is entering idle state. Update stats and idle timer if
1553 * spin_lock_irq(pool->lock).
1555 static void worker_enter_idle(struct worker *worker)
1557 struct worker_pool *pool = worker->pool;
1559 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1560 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1561 (worker->hentry.next || worker->hentry.pprev)))
1564 /* can't use worker_set_flags(), also called from start_worker() */
1565 worker->flags |= WORKER_IDLE;
1567 worker->last_active = jiffies;
1569 /* idle_list is LIFO */
1570 list_add(&worker->entry, &pool->idle_list);
1572 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1573 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1576 * Sanity check nr_running. Because wq_unbind_fn() releases
1577 * pool->lock between setting %WORKER_UNBOUND and zapping
1578 * nr_running, the warning may trigger spuriously. Check iff
1579 * unbind is not in progress.
1581 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1582 pool->nr_workers == pool->nr_idle &&
1583 atomic_read(&pool->nr_running));
1587 * worker_leave_idle - leave idle state
1588 * @worker: worker which is leaving idle state
1590 * @worker is leaving idle state. Update stats.
1593 * spin_lock_irq(pool->lock).
1595 static void worker_leave_idle(struct worker *worker)
1597 struct worker_pool *pool = worker->pool;
1599 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1601 worker_clr_flags(worker, WORKER_IDLE);
1603 list_del_init(&worker->entry);
1607 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1608 * @pool: target worker_pool
1610 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1612 * Works which are scheduled while the cpu is online must at least be
1613 * scheduled to a worker which is bound to the cpu so that if they are
1614 * flushed from cpu callbacks while cpu is going down, they are
1615 * guaranteed to execute on the cpu.
1617 * This function is to be used by unbound workers and rescuers to bind
1618 * themselves to the target cpu and may race with cpu going down or
1619 * coming online. kthread_bind() can't be used because it may put the
1620 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1621 * verbatim as it's best effort and blocking and pool may be
1622 * [dis]associated in the meantime.
1624 * This function tries set_cpus_allowed() and locks pool and verifies the
1625 * binding against %POOL_DISASSOCIATED which is set during
1626 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1627 * enters idle state or fetches works without dropping lock, it can
1628 * guarantee the scheduling requirement described in the first paragraph.
1631 * Might sleep. Called without any lock but returns with pool->lock
1635 * %true if the associated pool is online (@worker is successfully
1636 * bound), %false if offline.
1638 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1639 __acquires(&pool->lock)
1643 * The following call may fail, succeed or succeed
1644 * without actually migrating the task to the cpu if
1645 * it races with cpu hotunplug operation. Verify
1646 * against POOL_DISASSOCIATED.
1648 if (!(pool->flags & POOL_DISASSOCIATED))
1649 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1651 spin_lock_irq(&pool->lock);
1652 if (pool->flags & POOL_DISASSOCIATED)
1654 if (task_cpu(current) == pool->cpu &&
1655 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1657 spin_unlock_irq(&pool->lock);
1660 * We've raced with CPU hot[un]plug. Give it a breather
1661 * and retry migration. cond_resched() is required here;
1662 * otherwise, we might deadlock against cpu_stop trying to
1663 * bring down the CPU on non-preemptive kernel.
1670 static struct worker *alloc_worker(void)
1672 struct worker *worker;
1674 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1676 INIT_LIST_HEAD(&worker->entry);
1677 INIT_LIST_HEAD(&worker->scheduled);
1678 /* on creation a worker is in !idle && prep state */
1679 worker->flags = WORKER_PREP;
1685 * create_worker - create a new workqueue worker
1686 * @pool: pool the new worker will belong to
1688 * Create a new worker which is bound to @pool. The returned worker
1689 * can be started by calling start_worker() or destroyed using
1693 * Might sleep. Does GFP_KERNEL allocations.
1696 * Pointer to the newly created worker.
1698 static struct worker *create_worker(struct worker_pool *pool)
1700 struct worker *worker = NULL;
1704 lockdep_assert_held(&pool->manager_mutex);
1707 * ID is needed to determine kthread name. Allocate ID first
1708 * without installing the pointer.
1710 idr_preload(GFP_KERNEL);
1711 spin_lock_irq(&pool->lock);
1713 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1715 spin_unlock_irq(&pool->lock);
1720 worker = alloc_worker();
1724 worker->pool = pool;
1728 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1729 pool->attrs->nice < 0 ? "H" : "");
1731 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1733 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1734 "kworker/%s", id_buf);
1735 if (IS_ERR(worker->task))
1739 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1740 * online CPUs. It'll be re-applied when any of the CPUs come up.
1742 set_user_nice(worker->task, pool->attrs->nice);
1743 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1745 /* prevent userland from meddling with cpumask of workqueue workers */
1746 worker->task->flags |= PF_NO_SETAFFINITY;
1749 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1750 * remains stable across this function. See the comments above the
1751 * flag definition for details.
1753 if (pool->flags & POOL_DISASSOCIATED)
1754 worker->flags |= WORKER_UNBOUND;
1756 /* successful, commit the pointer to idr */
1757 spin_lock_irq(&pool->lock);
1758 idr_replace(&pool->worker_idr, worker, worker->id);
1759 spin_unlock_irq(&pool->lock);
1765 spin_lock_irq(&pool->lock);
1766 idr_remove(&pool->worker_idr, id);
1767 spin_unlock_irq(&pool->lock);
1774 * start_worker - start a newly created worker
1775 * @worker: worker to start
1777 * Make the pool aware of @worker and start it.
1780 * spin_lock_irq(pool->lock).
1782 static void start_worker(struct worker *worker)
1784 worker->flags |= WORKER_STARTED;
1785 worker->pool->nr_workers++;
1786 worker_enter_idle(worker);
1787 wake_up_process(worker->task);
1791 * create_and_start_worker - create and start a worker for a pool
1792 * @pool: the target pool
1794 * Grab the managership of @pool and create and start a new worker for it.
1796 * Return: 0 on success. A negative error code otherwise.
1798 static int create_and_start_worker(struct worker_pool *pool)
1800 struct worker *worker;
1802 mutex_lock(&pool->manager_mutex);
1804 worker = create_worker(pool);
1806 spin_lock_irq(&pool->lock);
1807 start_worker(worker);
1808 spin_unlock_irq(&pool->lock);
1811 mutex_unlock(&pool->manager_mutex);
1813 return worker ? 0 : -ENOMEM;
1817 * destroy_worker - destroy a workqueue worker
1818 * @worker: worker to be destroyed
1820 * Destroy @worker and adjust @pool stats accordingly.
1823 * spin_lock_irq(pool->lock) which is released and regrabbed.
1825 static void destroy_worker(struct worker *worker)
1827 struct worker_pool *pool = worker->pool;
1829 lockdep_assert_held(&pool->manager_mutex);
1830 lockdep_assert_held(&pool->lock);
1832 /* sanity check frenzy */
1833 if (WARN_ON(worker->current_work) ||
1834 WARN_ON(!list_empty(&worker->scheduled)))
1837 if (worker->flags & WORKER_STARTED)
1839 if (worker->flags & WORKER_IDLE)
1842 list_del_init(&worker->entry);
1843 worker->flags |= WORKER_DIE;
1845 idr_remove(&pool->worker_idr, worker->id);
1847 spin_unlock_irq(&pool->lock);
1849 kthread_stop(worker->task);
1852 spin_lock_irq(&pool->lock);
1855 static void idle_worker_timeout(unsigned long __pool)
1857 struct worker_pool *pool = (void *)__pool;
1859 spin_lock_irq(&pool->lock);
1861 if (too_many_workers(pool)) {
1862 struct worker *worker;
1863 unsigned long expires;
1865 /* idle_list is kept in LIFO order, check the last one */
1866 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1867 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1869 if (time_before(jiffies, expires))
1870 mod_timer(&pool->idle_timer, expires);
1872 /* it's been idle for too long, wake up manager */
1873 pool->flags |= POOL_MANAGE_WORKERS;
1874 wake_up_worker(pool);
1878 spin_unlock_irq(&pool->lock);
1881 static void send_mayday(struct work_struct *work)
1883 struct pool_workqueue *pwq = get_work_pwq(work);
1884 struct workqueue_struct *wq = pwq->wq;
1886 lockdep_assert_held(&wq_mayday_lock);
1891 /* mayday mayday mayday */
1892 if (list_empty(&pwq->mayday_node)) {
1893 list_add_tail(&pwq->mayday_node, &wq->maydays);
1894 wake_up_process(wq->rescuer->task);
1898 static void pool_mayday_timeout(unsigned long __pool)
1900 struct worker_pool *pool = (void *)__pool;
1901 struct work_struct *work;
1903 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1904 spin_lock(&pool->lock);
1906 if (need_to_create_worker(pool)) {
1908 * We've been trying to create a new worker but
1909 * haven't been successful. We might be hitting an
1910 * allocation deadlock. Send distress signals to
1913 list_for_each_entry(work, &pool->worklist, entry)
1917 spin_unlock(&pool->lock);
1918 spin_unlock_irq(&wq_mayday_lock);
1920 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1924 * maybe_create_worker - create a new worker if necessary
1925 * @pool: pool to create a new worker for
1927 * Create a new worker for @pool if necessary. @pool is guaranteed to
1928 * have at least one idle worker on return from this function. If
1929 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1930 * sent to all rescuers with works scheduled on @pool to resolve
1931 * possible allocation deadlock.
1933 * On return, need_to_create_worker() is guaranteed to be %false and
1934 * may_start_working() %true.
1937 * spin_lock_irq(pool->lock) which may be released and regrabbed
1938 * multiple times. Does GFP_KERNEL allocations. Called only from
1942 * %false if no action was taken and pool->lock stayed locked, %true
1945 static bool maybe_create_worker(struct worker_pool *pool)
1946 __releases(&pool->lock)
1947 __acquires(&pool->lock)
1949 if (!need_to_create_worker(pool))
1952 spin_unlock_irq(&pool->lock);
1954 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1955 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1958 struct worker *worker;
1960 worker = create_worker(pool);
1962 del_timer_sync(&pool->mayday_timer);
1963 spin_lock_irq(&pool->lock);
1964 start_worker(worker);
1965 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1970 if (!need_to_create_worker(pool))
1973 __set_current_state(TASK_INTERRUPTIBLE);
1974 schedule_timeout(CREATE_COOLDOWN);
1976 if (!need_to_create_worker(pool))
1980 del_timer_sync(&pool->mayday_timer);
1981 spin_lock_irq(&pool->lock);
1982 if (need_to_create_worker(pool))
1988 * maybe_destroy_worker - destroy workers which have been idle for a while
1989 * @pool: pool to destroy workers for
1991 * Destroy @pool workers which have been idle for longer than
1992 * IDLE_WORKER_TIMEOUT.
1995 * spin_lock_irq(pool->lock) which may be released and regrabbed
1996 * multiple times. Called only from manager.
1999 * %false if no action was taken and pool->lock stayed locked, %true
2002 static bool maybe_destroy_workers(struct worker_pool *pool)
2006 while (too_many_workers(pool)) {
2007 struct worker *worker;
2008 unsigned long expires;
2010 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2011 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2013 if (time_before(jiffies, expires)) {
2014 mod_timer(&pool->idle_timer, expires);
2018 destroy_worker(worker);
2026 * manage_workers - manage worker pool
2029 * Assume the manager role and manage the worker pool @worker belongs
2030 * to. At any given time, there can be only zero or one manager per
2031 * pool. The exclusion is handled automatically by this function.
2033 * The caller can safely start processing works on false return. On
2034 * true return, it's guaranteed that need_to_create_worker() is false
2035 * and may_start_working() is true.
2038 * spin_lock_irq(pool->lock) which may be released and regrabbed
2039 * multiple times. Does GFP_KERNEL allocations.
2042 * spin_lock_irq(pool->lock) which may be released and regrabbed
2043 * multiple times. Does GFP_KERNEL allocations.
2045 static bool manage_workers(struct worker *worker)
2047 struct worker_pool *pool = worker->pool;
2051 * Managership is governed by two mutexes - manager_arb and
2052 * manager_mutex. manager_arb handles arbitration of manager role.
2053 * Anyone who successfully grabs manager_arb wins the arbitration
2054 * and becomes the manager. mutex_trylock() on pool->manager_arb
2055 * failure while holding pool->lock reliably indicates that someone
2056 * else is managing the pool and the worker which failed trylock
2057 * can proceed to executing work items. This means that anyone
2058 * grabbing manager_arb is responsible for actually performing
2059 * manager duties. If manager_arb is grabbed and released without
2060 * actual management, the pool may stall indefinitely.
2062 * manager_mutex is used for exclusion of actual management
2063 * operations. The holder of manager_mutex can be sure that none
2064 * of management operations, including creation and destruction of
2065 * workers, won't take place until the mutex is released. Because
2066 * manager_mutex doesn't interfere with manager role arbitration,
2067 * it is guaranteed that the pool's management, while may be
2068 * delayed, won't be disturbed by someone else grabbing
2071 if (!mutex_trylock(&pool->manager_arb))
2075 * With manager arbitration won, manager_mutex would be free in
2076 * most cases. trylock first without dropping @pool->lock.
2078 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2079 spin_unlock_irq(&pool->lock);
2080 mutex_lock(&pool->manager_mutex);
2081 spin_lock_irq(&pool->lock);
2085 pool->flags &= ~POOL_MANAGE_WORKERS;
2088 * Destroy and then create so that may_start_working() is true
2091 ret |= maybe_destroy_workers(pool);
2092 ret |= maybe_create_worker(pool);
2094 mutex_unlock(&pool->manager_mutex);
2095 mutex_unlock(&pool->manager_arb);
2100 * process_one_work - process single work
2102 * @work: work to process
2104 * Process @work. This function contains all the logics necessary to
2105 * process a single work including synchronization against and
2106 * interaction with other workers on the same cpu, queueing and
2107 * flushing. As long as context requirement is met, any worker can
2108 * call this function to process a work.
2111 * spin_lock_irq(pool->lock) which is released and regrabbed.
2113 static void process_one_work(struct worker *worker, struct work_struct *work)
2114 __releases(&pool->lock)
2115 __acquires(&pool->lock)
2117 struct pool_workqueue *pwq = get_work_pwq(work);
2118 struct worker_pool *pool = worker->pool;
2119 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2121 struct worker *collision;
2122 #ifdef CONFIG_LOCKDEP
2124 * It is permissible to free the struct work_struct from
2125 * inside the function that is called from it, this we need to
2126 * take into account for lockdep too. To avoid bogus "held
2127 * lock freed" warnings as well as problems when looking into
2128 * work->lockdep_map, make a copy and use that here.
2130 struct lockdep_map lockdep_map;
2132 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2135 * Ensure we're on the correct CPU. DISASSOCIATED test is
2136 * necessary to avoid spurious warnings from rescuers servicing the
2137 * unbound or a disassociated pool.
2139 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2140 !(pool->flags & POOL_DISASSOCIATED) &&
2141 raw_smp_processor_id() != pool->cpu);
2144 * A single work shouldn't be executed concurrently by
2145 * multiple workers on a single cpu. Check whether anyone is
2146 * already processing the work. If so, defer the work to the
2147 * currently executing one.
2149 collision = find_worker_executing_work(pool, work);
2150 if (unlikely(collision)) {
2151 move_linked_works(work, &collision->scheduled, NULL);
2155 /* claim and dequeue */
2156 debug_work_deactivate(work);
2157 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2158 worker->current_work = work;
2159 worker->current_func = work->func;
2160 worker->current_pwq = pwq;
2161 work_color = get_work_color(work);
2163 list_del_init(&work->entry);
2166 * CPU intensive works don't participate in concurrency
2167 * management. They're the scheduler's responsibility.
2169 if (unlikely(cpu_intensive))
2170 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2173 * Unbound pool isn't concurrency managed and work items should be
2174 * executed ASAP. Wake up another worker if necessary.
2176 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2177 wake_up_worker(pool);
2180 * Record the last pool and clear PENDING which should be the last
2181 * update to @work. Also, do this inside @pool->lock so that
2182 * PENDING and queued state changes happen together while IRQ is
2185 set_work_pool_and_clear_pending(work, pool->id);
2187 spin_unlock_irq(&pool->lock);
2189 lock_map_acquire_read(&pwq->wq->lockdep_map);
2190 lock_map_acquire(&lockdep_map);
2191 trace_workqueue_execute_start(work);
2192 worker->current_func(work);
2194 * While we must be careful to not use "work" after this, the trace
2195 * point will only record its address.
2197 trace_workqueue_execute_end(work);
2198 lock_map_release(&lockdep_map);
2199 lock_map_release(&pwq->wq->lockdep_map);
2201 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2202 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2203 " last function: %pf\n",
2204 current->comm, preempt_count(), task_pid_nr(current),
2205 worker->current_func);
2206 debug_show_held_locks(current);
2210 spin_lock_irq(&pool->lock);
2212 /* clear cpu intensive status */
2213 if (unlikely(cpu_intensive))
2214 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2216 /* we're done with it, release */
2217 hash_del(&worker->hentry);
2218 worker->current_work = NULL;
2219 worker->current_func = NULL;
2220 worker->current_pwq = NULL;
2221 worker->desc_valid = false;
2222 pwq_dec_nr_in_flight(pwq, work_color);
2226 * process_scheduled_works - process scheduled works
2229 * Process all scheduled works. Please note that the scheduled list
2230 * may change while processing a work, so this function repeatedly
2231 * fetches a work from the top and executes it.
2234 * spin_lock_irq(pool->lock) which may be released and regrabbed
2237 static void process_scheduled_works(struct worker *worker)
2239 while (!list_empty(&worker->scheduled)) {
2240 struct work_struct *work = list_first_entry(&worker->scheduled,
2241 struct work_struct, entry);
2242 process_one_work(worker, work);
2247 * worker_thread - the worker thread function
2250 * The worker thread function. All workers belong to a worker_pool -
2251 * either a per-cpu one or dynamic unbound one. These workers process all
2252 * work items regardless of their specific target workqueue. The only
2253 * exception is work items which belong to workqueues with a rescuer which
2254 * will be explained in rescuer_thread().
2258 static int worker_thread(void *__worker)
2260 struct worker *worker = __worker;
2261 struct worker_pool *pool = worker->pool;
2263 /* tell the scheduler that this is a workqueue worker */
2264 worker->task->flags |= PF_WQ_WORKER;
2266 spin_lock_irq(&pool->lock);
2268 /* am I supposed to die? */
2269 if (unlikely(worker->flags & WORKER_DIE)) {
2270 spin_unlock_irq(&pool->lock);
2271 WARN_ON_ONCE(!list_empty(&worker->entry));
2272 worker->task->flags &= ~PF_WQ_WORKER;
2276 worker_leave_idle(worker);
2278 /* no more worker necessary? */
2279 if (!need_more_worker(pool))
2282 /* do we need to manage? */
2283 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2287 * ->scheduled list can only be filled while a worker is
2288 * preparing to process a work or actually processing it.
2289 * Make sure nobody diddled with it while I was sleeping.
2291 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2294 * Finish PREP stage. We're guaranteed to have at least one idle
2295 * worker or that someone else has already assumed the manager
2296 * role. This is where @worker starts participating in concurrency
2297 * management if applicable and concurrency management is restored
2298 * after being rebound. See rebind_workers() for details.
2300 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2303 struct work_struct *work =
2304 list_first_entry(&pool->worklist,
2305 struct work_struct, entry);
2307 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2308 /* optimization path, not strictly necessary */
2309 process_one_work(worker, work);
2310 if (unlikely(!list_empty(&worker->scheduled)))
2311 process_scheduled_works(worker);
2313 move_linked_works(work, &worker->scheduled, NULL);
2314 process_scheduled_works(worker);
2316 } while (keep_working(pool));
2318 worker_set_flags(worker, WORKER_PREP, false);
2320 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2324 * pool->lock is held and there's no work to process and no need to
2325 * manage, sleep. Workers are woken up only while holding
2326 * pool->lock or from local cpu, so setting the current state
2327 * before releasing pool->lock is enough to prevent losing any
2330 worker_enter_idle(worker);
2331 __set_current_state(TASK_INTERRUPTIBLE);
2332 spin_unlock_irq(&pool->lock);
2338 * rescuer_thread - the rescuer thread function
2341 * Workqueue rescuer thread function. There's one rescuer for each
2342 * workqueue which has WQ_MEM_RECLAIM set.
2344 * Regular work processing on a pool may block trying to create a new
2345 * worker which uses GFP_KERNEL allocation which has slight chance of
2346 * developing into deadlock if some works currently on the same queue
2347 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2348 * the problem rescuer solves.
2350 * When such condition is possible, the pool summons rescuers of all
2351 * workqueues which have works queued on the pool and let them process
2352 * those works so that forward progress can be guaranteed.
2354 * This should happen rarely.
2358 static int rescuer_thread(void *__rescuer)
2360 struct worker *rescuer = __rescuer;
2361 struct workqueue_struct *wq = rescuer->rescue_wq;
2362 struct list_head *scheduled = &rescuer->scheduled;
2364 set_user_nice(current, RESCUER_NICE_LEVEL);
2367 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2368 * doesn't participate in concurrency management.
2370 rescuer->task->flags |= PF_WQ_WORKER;
2372 set_current_state(TASK_INTERRUPTIBLE);
2374 if (kthread_should_stop()) {
2375 __set_current_state(TASK_RUNNING);
2376 rescuer->task->flags &= ~PF_WQ_WORKER;
2380 /* see whether any pwq is asking for help */
2381 spin_lock_irq(&wq_mayday_lock);
2383 while (!list_empty(&wq->maydays)) {
2384 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2385 struct pool_workqueue, mayday_node);
2386 struct worker_pool *pool = pwq->pool;
2387 struct work_struct *work, *n;
2389 __set_current_state(TASK_RUNNING);
2390 list_del_init(&pwq->mayday_node);
2392 spin_unlock_irq(&wq_mayday_lock);
2394 /* migrate to the target cpu if possible */
2395 worker_maybe_bind_and_lock(pool);
2396 rescuer->pool = pool;
2399 * Slurp in all works issued via this workqueue and
2402 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2403 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2404 if (get_work_pwq(work) == pwq)
2405 move_linked_works(work, scheduled, &n);
2407 process_scheduled_works(rescuer);
2410 * Leave this pool. If keep_working() is %true, notify a
2411 * regular worker; otherwise, we end up with 0 concurrency
2412 * and stalling the execution.
2414 if (keep_working(pool))
2415 wake_up_worker(pool);
2417 rescuer->pool = NULL;
2418 spin_unlock(&pool->lock);
2419 spin_lock(&wq_mayday_lock);
2422 spin_unlock_irq(&wq_mayday_lock);
2424 /* rescuers should never participate in concurrency management */
2425 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2431 struct work_struct work;
2432 struct completion done;
2435 static void wq_barrier_func(struct work_struct *work)
2437 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2438 complete(&barr->done);
2442 * insert_wq_barrier - insert a barrier work
2443 * @pwq: pwq to insert barrier into
2444 * @barr: wq_barrier to insert
2445 * @target: target work to attach @barr to
2446 * @worker: worker currently executing @target, NULL if @target is not executing
2448 * @barr is linked to @target such that @barr is completed only after
2449 * @target finishes execution. Please note that the ordering
2450 * guarantee is observed only with respect to @target and on the local
2453 * Currently, a queued barrier can't be canceled. This is because
2454 * try_to_grab_pending() can't determine whether the work to be
2455 * grabbed is at the head of the queue and thus can't clear LINKED
2456 * flag of the previous work while there must be a valid next work
2457 * after a work with LINKED flag set.
2459 * Note that when @worker is non-NULL, @target may be modified
2460 * underneath us, so we can't reliably determine pwq from @target.
2463 * spin_lock_irq(pool->lock).
2465 static void insert_wq_barrier(struct pool_workqueue *pwq,
2466 struct wq_barrier *barr,
2467 struct work_struct *target, struct worker *worker)
2469 struct list_head *head;
2470 unsigned int linked = 0;
2473 * debugobject calls are safe here even with pool->lock locked
2474 * as we know for sure that this will not trigger any of the
2475 * checks and call back into the fixup functions where we
2478 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2479 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2480 init_completion(&barr->done);
2483 * If @target is currently being executed, schedule the
2484 * barrier to the worker; otherwise, put it after @target.
2487 head = worker->scheduled.next;
2489 unsigned long *bits = work_data_bits(target);
2491 head = target->entry.next;
2492 /* there can already be other linked works, inherit and set */
2493 linked = *bits & WORK_STRUCT_LINKED;
2494 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2497 debug_work_activate(&barr->work);
2498 insert_work(pwq, &barr->work, head,
2499 work_color_to_flags(WORK_NO_COLOR) | linked);
2503 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2504 * @wq: workqueue being flushed
2505 * @flush_color: new flush color, < 0 for no-op
2506 * @work_color: new work color, < 0 for no-op
2508 * Prepare pwqs for workqueue flushing.
2510 * If @flush_color is non-negative, flush_color on all pwqs should be
2511 * -1. If no pwq has in-flight commands at the specified color, all
2512 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2513 * has in flight commands, its pwq->flush_color is set to
2514 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2515 * wakeup logic is armed and %true is returned.
2517 * The caller should have initialized @wq->first_flusher prior to
2518 * calling this function with non-negative @flush_color. If
2519 * @flush_color is negative, no flush color update is done and %false
2522 * If @work_color is non-negative, all pwqs should have the same
2523 * work_color which is previous to @work_color and all will be
2524 * advanced to @work_color.
2527 * mutex_lock(wq->mutex).
2530 * %true if @flush_color >= 0 and there's something to flush. %false
2533 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2534 int flush_color, int work_color)
2537 struct pool_workqueue *pwq;
2539 if (flush_color >= 0) {
2540 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2541 atomic_set(&wq->nr_pwqs_to_flush, 1);
2544 for_each_pwq(pwq, wq) {
2545 struct worker_pool *pool = pwq->pool;
2547 spin_lock_irq(&pool->lock);
2549 if (flush_color >= 0) {
2550 WARN_ON_ONCE(pwq->flush_color != -1);
2552 if (pwq->nr_in_flight[flush_color]) {
2553 pwq->flush_color = flush_color;
2554 atomic_inc(&wq->nr_pwqs_to_flush);
2559 if (work_color >= 0) {
2560 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2561 pwq->work_color = work_color;
2564 spin_unlock_irq(&pool->lock);
2567 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2568 complete(&wq->first_flusher->done);
2574 * flush_workqueue - ensure that any scheduled work has run to completion.
2575 * @wq: workqueue to flush
2577 * This function sleeps until all work items which were queued on entry
2578 * have finished execution, but it is not livelocked by new incoming ones.
2580 void flush_workqueue(struct workqueue_struct *wq)
2582 struct wq_flusher this_flusher = {
2583 .list = LIST_HEAD_INIT(this_flusher.list),
2585 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2589 lock_map_acquire(&wq->lockdep_map);
2590 lock_map_release(&wq->lockdep_map);
2592 mutex_lock(&wq->mutex);
2595 * Start-to-wait phase
2597 next_color = work_next_color(wq->work_color);
2599 if (next_color != wq->flush_color) {
2601 * Color space is not full. The current work_color
2602 * becomes our flush_color and work_color is advanced
2605 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2606 this_flusher.flush_color = wq->work_color;
2607 wq->work_color = next_color;
2609 if (!wq->first_flusher) {
2610 /* no flush in progress, become the first flusher */
2611 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2613 wq->first_flusher = &this_flusher;
2615 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2617 /* nothing to flush, done */
2618 wq->flush_color = next_color;
2619 wq->first_flusher = NULL;
2624 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2625 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2626 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2630 * Oops, color space is full, wait on overflow queue.
2631 * The next flush completion will assign us
2632 * flush_color and transfer to flusher_queue.
2634 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2637 mutex_unlock(&wq->mutex);
2639 wait_for_completion(&this_flusher.done);
2642 * Wake-up-and-cascade phase
2644 * First flushers are responsible for cascading flushes and
2645 * handling overflow. Non-first flushers can simply return.
2647 if (wq->first_flusher != &this_flusher)
2650 mutex_lock(&wq->mutex);
2652 /* we might have raced, check again with mutex held */
2653 if (wq->first_flusher != &this_flusher)
2656 wq->first_flusher = NULL;
2658 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2659 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2662 struct wq_flusher *next, *tmp;
2664 /* complete all the flushers sharing the current flush color */
2665 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2666 if (next->flush_color != wq->flush_color)
2668 list_del_init(&next->list);
2669 complete(&next->done);
2672 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2673 wq->flush_color != work_next_color(wq->work_color));
2675 /* this flush_color is finished, advance by one */
2676 wq->flush_color = work_next_color(wq->flush_color);
2678 /* one color has been freed, handle overflow queue */
2679 if (!list_empty(&wq->flusher_overflow)) {
2681 * Assign the same color to all overflowed
2682 * flushers, advance work_color and append to
2683 * flusher_queue. This is the start-to-wait
2684 * phase for these overflowed flushers.
2686 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2687 tmp->flush_color = wq->work_color;
2689 wq->work_color = work_next_color(wq->work_color);
2691 list_splice_tail_init(&wq->flusher_overflow,
2692 &wq->flusher_queue);
2693 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2696 if (list_empty(&wq->flusher_queue)) {
2697 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2702 * Need to flush more colors. Make the next flusher
2703 * the new first flusher and arm pwqs.
2705 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2706 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2708 list_del_init(&next->list);
2709 wq->first_flusher = next;
2711 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2715 * Meh... this color is already done, clear first
2716 * flusher and repeat cascading.
2718 wq->first_flusher = NULL;
2722 mutex_unlock(&wq->mutex);
2724 EXPORT_SYMBOL_GPL(flush_workqueue);
2727 * drain_workqueue - drain a workqueue
2728 * @wq: workqueue to drain
2730 * Wait until the workqueue becomes empty. While draining is in progress,
2731 * only chain queueing is allowed. IOW, only currently pending or running
2732 * work items on @wq can queue further work items on it. @wq is flushed
2733 * repeatedly until it becomes empty. The number of flushing is detemined
2734 * by the depth of chaining and should be relatively short. Whine if it
2737 void drain_workqueue(struct workqueue_struct *wq)
2739 unsigned int flush_cnt = 0;
2740 struct pool_workqueue *pwq;
2743 * __queue_work() needs to test whether there are drainers, is much
2744 * hotter than drain_workqueue() and already looks at @wq->flags.
2745 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2747 mutex_lock(&wq->mutex);
2748 if (!wq->nr_drainers++)
2749 wq->flags |= __WQ_DRAINING;
2750 mutex_unlock(&wq->mutex);
2752 flush_workqueue(wq);
2754 mutex_lock(&wq->mutex);
2756 for_each_pwq(pwq, wq) {
2759 spin_lock_irq(&pwq->pool->lock);
2760 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2761 spin_unlock_irq(&pwq->pool->lock);
2766 if (++flush_cnt == 10 ||
2767 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2768 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2769 wq->name, flush_cnt);
2771 mutex_unlock(&wq->mutex);
2775 if (!--wq->nr_drainers)
2776 wq->flags &= ~__WQ_DRAINING;
2777 mutex_unlock(&wq->mutex);
2779 EXPORT_SYMBOL_GPL(drain_workqueue);
2781 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2783 struct worker *worker = NULL;
2784 struct worker_pool *pool;
2785 struct pool_workqueue *pwq;
2789 local_irq_disable();
2790 pool = get_work_pool(work);
2796 spin_lock(&pool->lock);
2797 /* see the comment in try_to_grab_pending() with the same code */
2798 pwq = get_work_pwq(work);
2800 if (unlikely(pwq->pool != pool))
2803 worker = find_worker_executing_work(pool, work);
2806 pwq = worker->current_pwq;
2809 insert_wq_barrier(pwq, barr, work, worker);
2810 spin_unlock_irq(&pool->lock);
2813 * If @max_active is 1 or rescuer is in use, flushing another work
2814 * item on the same workqueue may lead to deadlock. Make sure the
2815 * flusher is not running on the same workqueue by verifying write
2818 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2819 lock_map_acquire(&pwq->wq->lockdep_map);
2821 lock_map_acquire_read(&pwq->wq->lockdep_map);
2822 lock_map_release(&pwq->wq->lockdep_map);
2826 spin_unlock_irq(&pool->lock);
2831 * flush_work - wait for a work to finish executing the last queueing instance
2832 * @work: the work to flush
2834 * Wait until @work has finished execution. @work is guaranteed to be idle
2835 * on return if it hasn't been requeued since flush started.
2838 * %true if flush_work() waited for the work to finish execution,
2839 * %false if it was already idle.
2841 bool flush_work(struct work_struct *work)
2843 struct wq_barrier barr;
2845 lock_map_acquire(&work->lockdep_map);
2846 lock_map_release(&work->lockdep_map);
2848 if (start_flush_work(work, &barr)) {
2849 wait_for_completion(&barr.done);
2850 destroy_work_on_stack(&barr.work);
2856 EXPORT_SYMBOL_GPL(flush_work);
2858 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2860 unsigned long flags;
2864 ret = try_to_grab_pending(work, is_dwork, &flags);
2866 * If someone else is canceling, wait for the same event it
2867 * would be waiting for before retrying.
2869 if (unlikely(ret == -ENOENT))
2871 } while (unlikely(ret < 0));
2873 /* tell other tasks trying to grab @work to back off */
2874 mark_work_canceling(work);
2875 local_irq_restore(flags);
2878 clear_work_data(work);
2883 * cancel_work_sync - cancel a work and wait for it to finish
2884 * @work: the work to cancel
2886 * Cancel @work and wait for its execution to finish. This function
2887 * can be used even if the work re-queues itself or migrates to
2888 * another workqueue. On return from this function, @work is
2889 * guaranteed to be not pending or executing on any CPU.
2891 * cancel_work_sync(&delayed_work->work) must not be used for
2892 * delayed_work's. Use cancel_delayed_work_sync() instead.
2894 * The caller must ensure that the workqueue on which @work was last
2895 * queued can't be destroyed before this function returns.
2898 * %true if @work was pending, %false otherwise.
2900 bool cancel_work_sync(struct work_struct *work)
2902 return __cancel_work_timer(work, false);
2904 EXPORT_SYMBOL_GPL(cancel_work_sync);
2907 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2908 * @dwork: the delayed work to flush
2910 * Delayed timer is cancelled and the pending work is queued for
2911 * immediate execution. Like flush_work(), this function only
2912 * considers the last queueing instance of @dwork.
2915 * %true if flush_work() waited for the work to finish execution,
2916 * %false if it was already idle.
2918 bool flush_delayed_work(struct delayed_work *dwork)
2920 local_irq_disable();
2921 if (del_timer_sync(&dwork->timer))
2922 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2924 return flush_work(&dwork->work);
2926 EXPORT_SYMBOL(flush_delayed_work);
2929 * cancel_delayed_work - cancel a delayed work
2930 * @dwork: delayed_work to cancel
2932 * Kill off a pending delayed_work.
2934 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2938 * The work callback function may still be running on return, unless
2939 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2940 * use cancel_delayed_work_sync() to wait on it.
2942 * This function is safe to call from any context including IRQ handler.
2944 bool cancel_delayed_work(struct delayed_work *dwork)
2946 unsigned long flags;
2950 ret = try_to_grab_pending(&dwork->work, true, &flags);
2951 } while (unlikely(ret == -EAGAIN));
2953 if (unlikely(ret < 0))
2956 set_work_pool_and_clear_pending(&dwork->work,
2957 get_work_pool_id(&dwork->work));
2958 local_irq_restore(flags);
2961 EXPORT_SYMBOL(cancel_delayed_work);
2964 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2965 * @dwork: the delayed work cancel
2967 * This is cancel_work_sync() for delayed works.
2970 * %true if @dwork was pending, %false otherwise.
2972 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2974 return __cancel_work_timer(&dwork->work, true);
2976 EXPORT_SYMBOL(cancel_delayed_work_sync);
2979 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2980 * @func: the function to call
2982 * schedule_on_each_cpu() executes @func on each online CPU using the
2983 * system workqueue and blocks until all CPUs have completed.
2984 * schedule_on_each_cpu() is very slow.
2987 * 0 on success, -errno on failure.
2989 int schedule_on_each_cpu(work_func_t func)
2992 struct work_struct __percpu *works;
2994 works = alloc_percpu(struct work_struct);
3000 for_each_online_cpu(cpu) {
3001 struct work_struct *work = per_cpu_ptr(works, cpu);
3003 INIT_WORK(work, func);
3004 schedule_work_on(cpu, work);
3007 for_each_online_cpu(cpu)
3008 flush_work(per_cpu_ptr(works, cpu));
3016 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3018 * Forces execution of the kernel-global workqueue and blocks until its
3021 * Think twice before calling this function! It's very easy to get into
3022 * trouble if you don't take great care. Either of the following situations
3023 * will lead to deadlock:
3025 * One of the work items currently on the workqueue needs to acquire
3026 * a lock held by your code or its caller.
3028 * Your code is running in the context of a work routine.
3030 * They will be detected by lockdep when they occur, but the first might not
3031 * occur very often. It depends on what work items are on the workqueue and
3032 * what locks they need, which you have no control over.
3034 * In most situations flushing the entire workqueue is overkill; you merely
3035 * need to know that a particular work item isn't queued and isn't running.
3036 * In such cases you should use cancel_delayed_work_sync() or
3037 * cancel_work_sync() instead.
3039 void flush_scheduled_work(void)
3041 flush_workqueue(system_wq);
3043 EXPORT_SYMBOL(flush_scheduled_work);
3046 * execute_in_process_context - reliably execute the routine with user context
3047 * @fn: the function to execute
3048 * @ew: guaranteed storage for the execute work structure (must
3049 * be available when the work executes)
3051 * Executes the function immediately if process context is available,
3052 * otherwise schedules the function for delayed execution.
3054 * Return: 0 - function was executed
3055 * 1 - function was scheduled for execution
3057 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3059 if (!in_interrupt()) {
3064 INIT_WORK(&ew->work, fn);
3065 schedule_work(&ew->work);
3069 EXPORT_SYMBOL_GPL(execute_in_process_context);
3073 * Workqueues with WQ_SYSFS flag set is visible to userland via
3074 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3075 * following attributes.
3077 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3078 * max_active RW int : maximum number of in-flight work items
3080 * Unbound workqueues have the following extra attributes.
3082 * id RO int : the associated pool ID
3083 * nice RW int : nice value of the workers
3084 * cpumask RW mask : bitmask of allowed CPUs for the workers
3087 struct workqueue_struct *wq;
3091 static struct workqueue_struct *dev_to_wq(struct device *dev)
3093 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3098 static ssize_t wq_per_cpu_show(struct device *dev,
3099 struct device_attribute *attr, char *buf)
3101 struct workqueue_struct *wq = dev_to_wq(dev);
3103 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3106 static ssize_t wq_max_active_show(struct device *dev,
3107 struct device_attribute *attr, char *buf)
3109 struct workqueue_struct *wq = dev_to_wq(dev);
3111 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3114 static ssize_t wq_max_active_store(struct device *dev,
3115 struct device_attribute *attr,
3116 const char *buf, size_t count)
3118 struct workqueue_struct *wq = dev_to_wq(dev);
3121 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3124 workqueue_set_max_active(wq, val);
3128 static struct device_attribute wq_sysfs_attrs[] = {
3129 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3130 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3134 static ssize_t wq_pool_ids_show(struct device *dev,
3135 struct device_attribute *attr, char *buf)
3137 struct workqueue_struct *wq = dev_to_wq(dev);
3138 const char *delim = "";
3139 int node, written = 0;
3141 rcu_read_lock_sched();
3142 for_each_node(node) {
3143 written += scnprintf(buf + written, PAGE_SIZE - written,
3144 "%s%d:%d", delim, node,
3145 unbound_pwq_by_node(wq, node)->pool->id);
3148 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3149 rcu_read_unlock_sched();
3154 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3157 struct workqueue_struct *wq = dev_to_wq(dev);
3160 mutex_lock(&wq->mutex);
3161 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3162 mutex_unlock(&wq->mutex);
3167 /* prepare workqueue_attrs for sysfs store operations */
3168 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3170 struct workqueue_attrs *attrs;
3172 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3176 mutex_lock(&wq->mutex);
3177 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3178 mutex_unlock(&wq->mutex);
3182 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3183 const char *buf, size_t count)
3185 struct workqueue_struct *wq = dev_to_wq(dev);
3186 struct workqueue_attrs *attrs;
3189 attrs = wq_sysfs_prep_attrs(wq);
3193 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3194 attrs->nice >= -20 && attrs->nice <= 19)
3195 ret = apply_workqueue_attrs(wq, attrs);
3199 free_workqueue_attrs(attrs);
3200 return ret ?: count;
3203 static ssize_t wq_cpumask_show(struct device *dev,
3204 struct device_attribute *attr, char *buf)
3206 struct workqueue_struct *wq = dev_to_wq(dev);
3209 mutex_lock(&wq->mutex);
3210 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3211 mutex_unlock(&wq->mutex);
3213 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3217 static ssize_t wq_cpumask_store(struct device *dev,
3218 struct device_attribute *attr,
3219 const char *buf, size_t count)
3221 struct workqueue_struct *wq = dev_to_wq(dev);
3222 struct workqueue_attrs *attrs;
3225 attrs = wq_sysfs_prep_attrs(wq);
3229 ret = cpumask_parse(buf, attrs->cpumask);
3231 ret = apply_workqueue_attrs(wq, attrs);
3233 free_workqueue_attrs(attrs);
3234 return ret ?: count;
3237 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3240 struct workqueue_struct *wq = dev_to_wq(dev);
3243 mutex_lock(&wq->mutex);
3244 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3245 !wq->unbound_attrs->no_numa);
3246 mutex_unlock(&wq->mutex);
3251 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3252 const char *buf, size_t count)
3254 struct workqueue_struct *wq = dev_to_wq(dev);
3255 struct workqueue_attrs *attrs;
3258 attrs = wq_sysfs_prep_attrs(wq);
3263 if (sscanf(buf, "%d", &v) == 1) {
3264 attrs->no_numa = !v;
3265 ret = apply_workqueue_attrs(wq, attrs);
3268 free_workqueue_attrs(attrs);
3269 return ret ?: count;
3272 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3273 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3274 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3275 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3276 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3280 static struct bus_type wq_subsys = {
3281 .name = "workqueue",
3282 .dev_attrs = wq_sysfs_attrs,
3285 static int __init wq_sysfs_init(void)
3287 return subsys_virtual_register(&wq_subsys, NULL);
3289 core_initcall(wq_sysfs_init);
3291 static void wq_device_release(struct device *dev)
3293 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3299 * workqueue_sysfs_register - make a workqueue visible in sysfs
3300 * @wq: the workqueue to register
3302 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3303 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3304 * which is the preferred method.
3306 * Workqueue user should use this function directly iff it wants to apply
3307 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3308 * apply_workqueue_attrs() may race against userland updating the
3311 * Return: 0 on success, -errno on failure.
3313 int workqueue_sysfs_register(struct workqueue_struct *wq)
3315 struct wq_device *wq_dev;
3319 * Adjusting max_active or creating new pwqs by applyting
3320 * attributes breaks ordering guarantee. Disallow exposing ordered
3323 if (WARN_ON(wq->flags & __WQ_ORDERED))
3326 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3331 wq_dev->dev.bus = &wq_subsys;
3332 wq_dev->dev.init_name = wq->name;
3333 wq_dev->dev.release = wq_device_release;
3336 * unbound_attrs are created separately. Suppress uevent until
3337 * everything is ready.
3339 dev_set_uevent_suppress(&wq_dev->dev, true);
3341 ret = device_register(&wq_dev->dev);
3348 if (wq->flags & WQ_UNBOUND) {
3349 struct device_attribute *attr;
3351 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3352 ret = device_create_file(&wq_dev->dev, attr);
3354 device_unregister(&wq_dev->dev);
3361 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3366 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3367 * @wq: the workqueue to unregister
3369 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3371 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3373 struct wq_device *wq_dev = wq->wq_dev;
3379 device_unregister(&wq_dev->dev);
3381 #else /* CONFIG_SYSFS */
3382 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3383 #endif /* CONFIG_SYSFS */
3386 * free_workqueue_attrs - free a workqueue_attrs
3387 * @attrs: workqueue_attrs to free
3389 * Undo alloc_workqueue_attrs().
3391 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3394 free_cpumask_var(attrs->cpumask);
3400 * alloc_workqueue_attrs - allocate a workqueue_attrs
3401 * @gfp_mask: allocation mask to use
3403 * Allocate a new workqueue_attrs, initialize with default settings and
3406 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3408 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3410 struct workqueue_attrs *attrs;
3412 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3415 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3418 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3421 free_workqueue_attrs(attrs);
3425 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3426 const struct workqueue_attrs *from)
3428 to->nice = from->nice;
3429 cpumask_copy(to->cpumask, from->cpumask);
3432 /* hash value of the content of @attr */
3433 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3437 hash = jhash_1word(attrs->nice, hash);
3438 hash = jhash(cpumask_bits(attrs->cpumask),
3439 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3443 /* content equality test */
3444 static bool wqattrs_equal(const struct workqueue_attrs *a,
3445 const struct workqueue_attrs *b)
3447 if (a->nice != b->nice)
3449 if (!cpumask_equal(a->cpumask, b->cpumask))
3455 * init_worker_pool - initialize a newly zalloc'd worker_pool
3456 * @pool: worker_pool to initialize
3458 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3460 * Return: 0 on success, -errno on failure. Even on failure, all fields
3461 * inside @pool proper are initialized and put_unbound_pool() can be called
3462 * on @pool safely to release it.
3464 static int init_worker_pool(struct worker_pool *pool)
3466 spin_lock_init(&pool->lock);
3469 pool->node = NUMA_NO_NODE;
3470 pool->flags |= POOL_DISASSOCIATED;
3471 INIT_LIST_HEAD(&pool->worklist);
3472 INIT_LIST_HEAD(&pool->idle_list);
3473 hash_init(pool->busy_hash);
3475 init_timer_deferrable(&pool->idle_timer);
3476 pool->idle_timer.function = idle_worker_timeout;
3477 pool->idle_timer.data = (unsigned long)pool;
3479 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3480 (unsigned long)pool);
3482 mutex_init(&pool->manager_arb);
3483 mutex_init(&pool->manager_mutex);
3484 idr_init(&pool->worker_idr);
3486 INIT_HLIST_NODE(&pool->hash_node);
3489 /* shouldn't fail above this point */
3490 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3496 static void rcu_free_pool(struct rcu_head *rcu)
3498 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3500 idr_destroy(&pool->worker_idr);
3501 free_workqueue_attrs(pool->attrs);
3506 * put_unbound_pool - put a worker_pool
3507 * @pool: worker_pool to put
3509 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3510 * safe manner. get_unbound_pool() calls this function on its failure path
3511 * and this function should be able to release pools which went through,
3512 * successfully or not, init_worker_pool().
3514 * Should be called with wq_pool_mutex held.
3516 static void put_unbound_pool(struct worker_pool *pool)
3518 struct worker *worker;
3520 lockdep_assert_held(&wq_pool_mutex);
3526 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3527 WARN_ON(!list_empty(&pool->worklist)))
3530 /* release id and unhash */
3532 idr_remove(&worker_pool_idr, pool->id);
3533 hash_del(&pool->hash_node);
3536 * Become the manager and destroy all workers. Grabbing
3537 * manager_arb prevents @pool's workers from blocking on
3540 mutex_lock(&pool->manager_arb);
3541 mutex_lock(&pool->manager_mutex);
3542 spin_lock_irq(&pool->lock);
3544 while ((worker = first_worker(pool)))
3545 destroy_worker(worker);
3546 WARN_ON(pool->nr_workers || pool->nr_idle);
3548 spin_unlock_irq(&pool->lock);
3549 mutex_unlock(&pool->manager_mutex);
3550 mutex_unlock(&pool->manager_arb);
3552 /* shut down the timers */
3553 del_timer_sync(&pool->idle_timer);
3554 del_timer_sync(&pool->mayday_timer);
3556 /* sched-RCU protected to allow dereferences from get_work_pool() */
3557 call_rcu_sched(&pool->rcu, rcu_free_pool);
3561 * get_unbound_pool - get a worker_pool with the specified attributes
3562 * @attrs: the attributes of the worker_pool to get
3564 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3565 * reference count and return it. If there already is a matching
3566 * worker_pool, it will be used; otherwise, this function attempts to
3569 * Should be called with wq_pool_mutex held.
3571 * Return: On success, a worker_pool with the same attributes as @attrs.
3572 * On failure, %NULL.
3574 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3576 u32 hash = wqattrs_hash(attrs);
3577 struct worker_pool *pool;
3580 lockdep_assert_held(&wq_pool_mutex);
3582 /* do we already have a matching pool? */
3583 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3584 if (wqattrs_equal(pool->attrs, attrs)) {
3590 /* nope, create a new one */
3591 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3592 if (!pool || init_worker_pool(pool) < 0)
3595 if (workqueue_freezing)
3596 pool->flags |= POOL_FREEZING;
3598 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3599 copy_workqueue_attrs(pool->attrs, attrs);
3601 /* if cpumask is contained inside a NUMA node, we belong to that node */
3602 if (wq_numa_enabled) {
3603 for_each_node(node) {
3604 if (cpumask_subset(pool->attrs->cpumask,
3605 wq_numa_possible_cpumask[node])) {
3612 if (worker_pool_assign_id(pool) < 0)
3615 /* create and start the initial worker */
3616 if (create_and_start_worker(pool) < 0)
3620 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3625 put_unbound_pool(pool);
3629 static void rcu_free_pwq(struct rcu_head *rcu)
3631 kmem_cache_free(pwq_cache,
3632 container_of(rcu, struct pool_workqueue, rcu));
3636 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3637 * and needs to be destroyed.
3639 static void pwq_unbound_release_workfn(struct work_struct *work)
3641 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3642 unbound_release_work);
3643 struct workqueue_struct *wq = pwq->wq;
3644 struct worker_pool *pool = pwq->pool;
3647 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3651 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3652 * necessary on release but do it anyway. It's easier to verify
3653 * and consistent with the linking path.
3655 mutex_lock(&wq->mutex);
3656 list_del_rcu(&pwq->pwqs_node);
3657 is_last = list_empty(&wq->pwqs);
3658 mutex_unlock(&wq->mutex);
3660 mutex_lock(&wq_pool_mutex);
3661 put_unbound_pool(pool);
3662 mutex_unlock(&wq_pool_mutex);
3664 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3667 * If we're the last pwq going away, @wq is already dead and no one
3668 * is gonna access it anymore. Free it.
3671 free_workqueue_attrs(wq->unbound_attrs);
3677 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3678 * @pwq: target pool_workqueue
3680 * If @pwq isn't freezing, set @pwq->max_active to the associated
3681 * workqueue's saved_max_active and activate delayed work items
3682 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3684 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3686 struct workqueue_struct *wq = pwq->wq;
3687 bool freezable = wq->flags & WQ_FREEZABLE;
3689 /* for @wq->saved_max_active */
3690 lockdep_assert_held(&wq->mutex);
3692 /* fast exit for non-freezable wqs */
3693 if (!freezable && pwq->max_active == wq->saved_max_active)
3696 spin_lock_irq(&pwq->pool->lock);
3698 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3699 pwq->max_active = wq->saved_max_active;
3701 while (!list_empty(&pwq->delayed_works) &&
3702 pwq->nr_active < pwq->max_active)
3703 pwq_activate_first_delayed(pwq);
3706 * Need to kick a worker after thawed or an unbound wq's
3707 * max_active is bumped. It's a slow path. Do it always.
3709 wake_up_worker(pwq->pool);
3711 pwq->max_active = 0;
3714 spin_unlock_irq(&pwq->pool->lock);
3717 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3718 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3719 struct worker_pool *pool)
3721 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3723 memset(pwq, 0, sizeof(*pwq));
3727 pwq->flush_color = -1;
3729 INIT_LIST_HEAD(&pwq->delayed_works);
3730 INIT_LIST_HEAD(&pwq->pwqs_node);
3731 INIT_LIST_HEAD(&pwq->mayday_node);
3732 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3735 /* sync @pwq with the current state of its associated wq and link it */
3736 static void link_pwq(struct pool_workqueue *pwq)
3738 struct workqueue_struct *wq = pwq->wq;
3740 lockdep_assert_held(&wq->mutex);
3742 /* may be called multiple times, ignore if already linked */
3743 if (!list_empty(&pwq->pwqs_node))
3747 * Set the matching work_color. This is synchronized with
3748 * wq->mutex to avoid confusing flush_workqueue().
3750 pwq->work_color = wq->work_color;
3752 /* sync max_active to the current setting */
3753 pwq_adjust_max_active(pwq);
3756 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3759 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3760 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3761 const struct workqueue_attrs *attrs)
3763 struct worker_pool *pool;
3764 struct pool_workqueue *pwq;
3766 lockdep_assert_held(&wq_pool_mutex);
3768 pool = get_unbound_pool(attrs);
3772 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3774 put_unbound_pool(pool);
3778 init_pwq(pwq, wq, pool);
3782 /* undo alloc_unbound_pwq(), used only in the error path */
3783 static void free_unbound_pwq(struct pool_workqueue *pwq)
3785 lockdep_assert_held(&wq_pool_mutex);
3788 put_unbound_pool(pwq->pool);
3789 kmem_cache_free(pwq_cache, pwq);
3794 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3795 * @attrs: the wq_attrs of interest
3796 * @node: the target NUMA node
3797 * @cpu_going_down: if >= 0, the CPU to consider as offline
3798 * @cpumask: outarg, the resulting cpumask
3800 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3801 * @cpu_going_down is >= 0, that cpu is considered offline during
3802 * calculation. The result is stored in @cpumask.
3804 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3805 * enabled and @node has online CPUs requested by @attrs, the returned
3806 * cpumask is the intersection of the possible CPUs of @node and
3809 * The caller is responsible for ensuring that the cpumask of @node stays
3812 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3815 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3816 int cpu_going_down, cpumask_t *cpumask)
3818 if (!wq_numa_enabled || attrs->no_numa)
3821 /* does @node have any online CPUs @attrs wants? */
3822 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3823 if (cpu_going_down >= 0)
3824 cpumask_clear_cpu(cpu_going_down, cpumask);
3826 if (cpumask_empty(cpumask))
3829 /* yeap, return possible CPUs in @node that @attrs wants */
3830 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3831 return !cpumask_equal(cpumask, attrs->cpumask);
3834 cpumask_copy(cpumask, attrs->cpumask);
3838 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3839 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3841 struct pool_workqueue *pwq)
3843 struct pool_workqueue *old_pwq;
3845 lockdep_assert_held(&wq->mutex);
3847 /* link_pwq() can handle duplicate calls */
3850 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3851 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3856 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3857 * @wq: the target workqueue
3858 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3860 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3861 * machines, this function maps a separate pwq to each NUMA node with
3862 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3863 * NUMA node it was issued on. Older pwqs are released as in-flight work
3864 * items finish. Note that a work item which repeatedly requeues itself
3865 * back-to-back will stay on its current pwq.
3867 * Performs GFP_KERNEL allocations.
3869 * Return: 0 on success and -errno on failure.
3871 int apply_workqueue_attrs(struct workqueue_struct *wq,
3872 const struct workqueue_attrs *attrs)
3874 struct workqueue_attrs *new_attrs, *tmp_attrs;
3875 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3878 /* only unbound workqueues can change attributes */
3879 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3882 /* creating multiple pwqs breaks ordering guarantee */
3883 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3886 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3887 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3888 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3889 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3892 /* make a copy of @attrs and sanitize it */
3893 copy_workqueue_attrs(new_attrs, attrs);
3894 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3897 * We may create multiple pwqs with differing cpumasks. Make a
3898 * copy of @new_attrs which will be modified and used to obtain
3901 copy_workqueue_attrs(tmp_attrs, new_attrs);
3904 * CPUs should stay stable across pwq creations and installations.
3905 * Pin CPUs, determine the target cpumask for each node and create
3910 mutex_lock(&wq_pool_mutex);
3913 * If something goes wrong during CPU up/down, we'll fall back to
3914 * the default pwq covering whole @attrs->cpumask. Always create
3915 * it even if we don't use it immediately.
3917 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3921 for_each_node(node) {
3922 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3923 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3928 pwq_tbl[node] = dfl_pwq;
3932 mutex_unlock(&wq_pool_mutex);
3934 /* all pwqs have been created successfully, let's install'em */
3935 mutex_lock(&wq->mutex);
3937 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3939 /* save the previous pwq and install the new one */
3941 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3943 /* @dfl_pwq might not have been used, ensure it's linked */
3945 swap(wq->dfl_pwq, dfl_pwq);
3947 mutex_unlock(&wq->mutex);
3949 /* put the old pwqs */
3951 put_pwq_unlocked(pwq_tbl[node]);
3952 put_pwq_unlocked(dfl_pwq);
3958 free_workqueue_attrs(tmp_attrs);
3959 free_workqueue_attrs(new_attrs);
3964 free_unbound_pwq(dfl_pwq);
3966 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3967 free_unbound_pwq(pwq_tbl[node]);
3968 mutex_unlock(&wq_pool_mutex);
3976 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3977 * @wq: the target workqueue
3978 * @cpu: the CPU coming up or going down
3979 * @online: whether @cpu is coming up or going down
3981 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3982 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3985 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3986 * falls back to @wq->dfl_pwq which may not be optimal but is always
3989 * Note that when the last allowed CPU of a NUMA node goes offline for a
3990 * workqueue with a cpumask spanning multiple nodes, the workers which were
3991 * already executing the work items for the workqueue will lose their CPU
3992 * affinity and may execute on any CPU. This is similar to how per-cpu
3993 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3994 * affinity, it's the user's responsibility to flush the work item from
3997 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4000 int node = cpu_to_node(cpu);
4001 int cpu_off = online ? -1 : cpu;
4002 struct pool_workqueue *old_pwq = NULL, *pwq;
4003 struct workqueue_attrs *target_attrs;
4006 lockdep_assert_held(&wq_pool_mutex);
4008 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4012 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4013 * Let's use a preallocated one. The following buf is protected by
4014 * CPU hotplug exclusion.
4016 target_attrs = wq_update_unbound_numa_attrs_buf;
4017 cpumask = target_attrs->cpumask;
4019 mutex_lock(&wq->mutex);
4020 if (wq->unbound_attrs->no_numa)
4023 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4024 pwq = unbound_pwq_by_node(wq, node);
4027 * Let's determine what needs to be done. If the target cpumask is
4028 * different from wq's, we need to compare it to @pwq's and create
4029 * a new one if they don't match. If the target cpumask equals
4030 * wq's, the default pwq should be used. If @pwq is already the
4031 * default one, nothing to do; otherwise, install the default one.
4033 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4034 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4037 if (pwq == wq->dfl_pwq)
4043 mutex_unlock(&wq->mutex);
4045 /* create a new pwq */
4046 pwq = alloc_unbound_pwq(wq, target_attrs);
4048 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4054 * Install the new pwq. As this function is called only from CPU
4055 * hotplug callbacks and applying a new attrs is wrapped with
4056 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4059 mutex_lock(&wq->mutex);
4060 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4064 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4065 get_pwq(wq->dfl_pwq);
4066 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4067 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4069 mutex_unlock(&wq->mutex);
4070 put_pwq_unlocked(old_pwq);
4073 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4075 bool highpri = wq->flags & WQ_HIGHPRI;
4078 if (!(wq->flags & WQ_UNBOUND)) {
4079 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4083 for_each_possible_cpu(cpu) {
4084 struct pool_workqueue *pwq =
4085 per_cpu_ptr(wq->cpu_pwqs, cpu);
4086 struct worker_pool *cpu_pools =
4087 per_cpu(cpu_worker_pools, cpu);
4089 init_pwq(pwq, wq, &cpu_pools[highpri]);
4091 mutex_lock(&wq->mutex);
4093 mutex_unlock(&wq->mutex);
4097 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4101 static int wq_clamp_max_active(int max_active, unsigned int flags,
4104 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4106 if (max_active < 1 || max_active > lim)
4107 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4108 max_active, name, 1, lim);
4110 return clamp_val(max_active, 1, lim);
4113 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4116 struct lock_class_key *key,
4117 const char *lock_name, ...)
4119 size_t tbl_size = 0;
4121 struct workqueue_struct *wq;
4122 struct pool_workqueue *pwq;
4124 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4125 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4126 flags |= WQ_UNBOUND;
4128 /* allocate wq and format name */
4129 if (flags & WQ_UNBOUND)
4130 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4132 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4136 if (flags & WQ_UNBOUND) {
4137 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4138 if (!wq->unbound_attrs)
4142 va_start(args, lock_name);
4143 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4146 max_active = max_active ?: WQ_DFL_ACTIVE;
4147 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4151 wq->saved_max_active = max_active;
4152 mutex_init(&wq->mutex);
4153 atomic_set(&wq->nr_pwqs_to_flush, 0);
4154 INIT_LIST_HEAD(&wq->pwqs);
4155 INIT_LIST_HEAD(&wq->flusher_queue);
4156 INIT_LIST_HEAD(&wq->flusher_overflow);
4157 INIT_LIST_HEAD(&wq->maydays);
4159 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4160 INIT_LIST_HEAD(&wq->list);
4162 if (alloc_and_link_pwqs(wq) < 0)
4166 * Workqueues which may be used during memory reclaim should
4167 * have a rescuer to guarantee forward progress.
4169 if (flags & WQ_MEM_RECLAIM) {
4170 struct worker *rescuer;
4172 rescuer = alloc_worker();
4176 rescuer->rescue_wq = wq;
4177 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4179 if (IS_ERR(rescuer->task)) {
4184 wq->rescuer = rescuer;
4185 rescuer->task->flags |= PF_NO_SETAFFINITY;
4186 wake_up_process(rescuer->task);
4189 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4193 * wq_pool_mutex protects global freeze state and workqueues list.
4194 * Grab it, adjust max_active and add the new @wq to workqueues
4197 mutex_lock(&wq_pool_mutex);
4199 mutex_lock(&wq->mutex);
4200 for_each_pwq(pwq, wq)
4201 pwq_adjust_max_active(pwq);
4202 mutex_unlock(&wq->mutex);
4204 list_add(&wq->list, &workqueues);
4206 mutex_unlock(&wq_pool_mutex);
4211 free_workqueue_attrs(wq->unbound_attrs);
4215 destroy_workqueue(wq);
4218 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4221 * destroy_workqueue - safely terminate a workqueue
4222 * @wq: target workqueue
4224 * Safely destroy a workqueue. All work currently pending will be done first.
4226 void destroy_workqueue(struct workqueue_struct *wq)
4228 struct pool_workqueue *pwq;
4231 /* drain it before proceeding with destruction */
4232 drain_workqueue(wq);
4235 mutex_lock(&wq->mutex);
4236 for_each_pwq(pwq, wq) {
4239 for (i = 0; i < WORK_NR_COLORS; i++) {
4240 if (WARN_ON(pwq->nr_in_flight[i])) {
4241 mutex_unlock(&wq->mutex);
4246 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4247 WARN_ON(pwq->nr_active) ||
4248 WARN_ON(!list_empty(&pwq->delayed_works))) {
4249 mutex_unlock(&wq->mutex);
4253 mutex_unlock(&wq->mutex);
4256 * wq list is used to freeze wq, remove from list after
4257 * flushing is complete in case freeze races us.
4259 mutex_lock(&wq_pool_mutex);
4260 list_del_init(&wq->list);
4261 mutex_unlock(&wq_pool_mutex);
4263 workqueue_sysfs_unregister(wq);
4266 kthread_stop(wq->rescuer->task);
4271 if (!(wq->flags & WQ_UNBOUND)) {
4273 * The base ref is never dropped on per-cpu pwqs. Directly
4274 * free the pwqs and wq.
4276 free_percpu(wq->cpu_pwqs);
4280 * We're the sole accessor of @wq at this point. Directly
4281 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4282 * @wq will be freed when the last pwq is released.
4284 for_each_node(node) {
4285 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4286 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4287 put_pwq_unlocked(pwq);
4291 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4292 * put. Don't access it afterwards.
4296 put_pwq_unlocked(pwq);
4299 EXPORT_SYMBOL_GPL(destroy_workqueue);
4302 * workqueue_set_max_active - adjust max_active of a workqueue
4303 * @wq: target workqueue
4304 * @max_active: new max_active value.
4306 * Set max_active of @wq to @max_active.
4309 * Don't call from IRQ context.
4311 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4313 struct pool_workqueue *pwq;
4315 /* disallow meddling with max_active for ordered workqueues */
4316 if (WARN_ON(wq->flags & __WQ_ORDERED))
4319 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4321 mutex_lock(&wq->mutex);
4323 wq->saved_max_active = max_active;
4325 for_each_pwq(pwq, wq)
4326 pwq_adjust_max_active(pwq);
4328 mutex_unlock(&wq->mutex);
4330 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4333 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4335 * Determine whether %current is a workqueue rescuer. Can be used from
4336 * work functions to determine whether it's being run off the rescuer task.
4338 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4340 bool current_is_workqueue_rescuer(void)
4342 struct worker *worker = current_wq_worker();
4344 return worker && worker->rescue_wq;
4348 * workqueue_congested - test whether a workqueue is congested
4349 * @cpu: CPU in question
4350 * @wq: target workqueue
4352 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4353 * no synchronization around this function and the test result is
4354 * unreliable and only useful as advisory hints or for debugging.
4356 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4357 * Note that both per-cpu and unbound workqueues may be associated with
4358 * multiple pool_workqueues which have separate congested states. A
4359 * workqueue being congested on one CPU doesn't mean the workqueue is also
4360 * contested on other CPUs / NUMA nodes.
4363 * %true if congested, %false otherwise.
4365 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4367 struct pool_workqueue *pwq;
4370 rcu_read_lock_sched();
4372 if (cpu == WORK_CPU_UNBOUND)
4373 cpu = smp_processor_id();
4375 if (!(wq->flags & WQ_UNBOUND))
4376 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4378 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4380 ret = !list_empty(&pwq->delayed_works);
4381 rcu_read_unlock_sched();
4385 EXPORT_SYMBOL_GPL(workqueue_congested);
4388 * work_busy - test whether a work is currently pending or running
4389 * @work: the work to be tested
4391 * Test whether @work is currently pending or running. There is no
4392 * synchronization around this function and the test result is
4393 * unreliable and only useful as advisory hints or for debugging.
4396 * OR'd bitmask of WORK_BUSY_* bits.
4398 unsigned int work_busy(struct work_struct *work)
4400 struct worker_pool *pool;
4401 unsigned long flags;
4402 unsigned int ret = 0;
4404 if (work_pending(work))
4405 ret |= WORK_BUSY_PENDING;
4407 local_irq_save(flags);
4408 pool = get_work_pool(work);
4410 spin_lock(&pool->lock);
4411 if (find_worker_executing_work(pool, work))
4412 ret |= WORK_BUSY_RUNNING;
4413 spin_unlock(&pool->lock);
4415 local_irq_restore(flags);
4419 EXPORT_SYMBOL_GPL(work_busy);
4422 * set_worker_desc - set description for the current work item
4423 * @fmt: printf-style format string
4424 * @...: arguments for the format string
4426 * This function can be called by a running work function to describe what
4427 * the work item is about. If the worker task gets dumped, this
4428 * information will be printed out together to help debugging. The
4429 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4431 void set_worker_desc(const char *fmt, ...)
4433 struct worker *worker = current_wq_worker();
4437 va_start(args, fmt);
4438 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4440 worker->desc_valid = true;
4445 * print_worker_info - print out worker information and description
4446 * @log_lvl: the log level to use when printing
4447 * @task: target task
4449 * If @task is a worker and currently executing a work item, print out the
4450 * name of the workqueue being serviced and worker description set with
4451 * set_worker_desc() by the currently executing work item.
4453 * This function can be safely called on any task as long as the
4454 * task_struct itself is accessible. While safe, this function isn't
4455 * synchronized and may print out mixups or garbages of limited length.
4457 void print_worker_info(const char *log_lvl, struct task_struct *task)
4459 work_func_t *fn = NULL;
4460 char name[WQ_NAME_LEN] = { };
4461 char desc[WORKER_DESC_LEN] = { };
4462 struct pool_workqueue *pwq = NULL;
4463 struct workqueue_struct *wq = NULL;
4464 bool desc_valid = false;
4465 struct worker *worker;
4467 if (!(task->flags & PF_WQ_WORKER))
4471 * This function is called without any synchronization and @task
4472 * could be in any state. Be careful with dereferences.
4474 worker = probe_kthread_data(task);
4477 * Carefully copy the associated workqueue's workfn and name. Keep
4478 * the original last '\0' in case the original contains garbage.
4480 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4481 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4482 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4483 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4485 /* copy worker description */
4486 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4488 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4490 if (fn || name[0] || desc[0]) {
4491 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4493 pr_cont(" (%s)", desc);
4501 * There are two challenges in supporting CPU hotplug. Firstly, there
4502 * are a lot of assumptions on strong associations among work, pwq and
4503 * pool which make migrating pending and scheduled works very
4504 * difficult to implement without impacting hot paths. Secondly,
4505 * worker pools serve mix of short, long and very long running works making
4506 * blocked draining impractical.
4508 * This is solved by allowing the pools to be disassociated from the CPU
4509 * running as an unbound one and allowing it to be reattached later if the
4510 * cpu comes back online.
4513 static void wq_unbind_fn(struct work_struct *work)
4515 int cpu = smp_processor_id();
4516 struct worker_pool *pool;
4517 struct worker *worker;
4520 for_each_cpu_worker_pool(pool, cpu) {
4521 WARN_ON_ONCE(cpu != smp_processor_id());
4523 mutex_lock(&pool->manager_mutex);
4524 spin_lock_irq(&pool->lock);
4527 * We've blocked all manager operations. Make all workers
4528 * unbound and set DISASSOCIATED. Before this, all workers
4529 * except for the ones which are still executing works from
4530 * before the last CPU down must be on the cpu. After
4531 * this, they may become diasporas.
4533 for_each_pool_worker(worker, wi, pool)
4534 worker->flags |= WORKER_UNBOUND;
4536 pool->flags |= POOL_DISASSOCIATED;
4538 spin_unlock_irq(&pool->lock);
4539 mutex_unlock(&pool->manager_mutex);
4542 * Call schedule() so that we cross rq->lock and thus can
4543 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4544 * This is necessary as scheduler callbacks may be invoked
4550 * Sched callbacks are disabled now. Zap nr_running.
4551 * After this, nr_running stays zero and need_more_worker()
4552 * and keep_working() are always true as long as the
4553 * worklist is not empty. This pool now behaves as an
4554 * unbound (in terms of concurrency management) pool which
4555 * are served by workers tied to the pool.
4557 atomic_set(&pool->nr_running, 0);
4560 * With concurrency management just turned off, a busy
4561 * worker blocking could lead to lengthy stalls. Kick off
4562 * unbound chain execution of currently pending work items.
4564 spin_lock_irq(&pool->lock);
4565 wake_up_worker(pool);
4566 spin_unlock_irq(&pool->lock);
4571 * rebind_workers - rebind all workers of a pool to the associated CPU
4572 * @pool: pool of interest
4574 * @pool->cpu is coming online. Rebind all workers to the CPU.
4576 static void rebind_workers(struct worker_pool *pool)
4578 struct worker *worker;
4581 lockdep_assert_held(&pool->manager_mutex);
4584 * Restore CPU affinity of all workers. As all idle workers should
4585 * be on the run-queue of the associated CPU before any local
4586 * wake-ups for concurrency management happen, restore CPU affinty
4587 * of all workers first and then clear UNBOUND. As we're called
4588 * from CPU_ONLINE, the following shouldn't fail.
4590 for_each_pool_worker(worker, wi, pool)
4591 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4592 pool->attrs->cpumask) < 0);
4594 spin_lock_irq(&pool->lock);
4596 for_each_pool_worker(worker, wi, pool) {
4597 unsigned int worker_flags = worker->flags;
4600 * A bound idle worker should actually be on the runqueue
4601 * of the associated CPU for local wake-ups targeting it to
4602 * work. Kick all idle workers so that they migrate to the
4603 * associated CPU. Doing this in the same loop as
4604 * replacing UNBOUND with REBOUND is safe as no worker will
4605 * be bound before @pool->lock is released.
4607 if (worker_flags & WORKER_IDLE)
4608 wake_up_process(worker->task);
4611 * We want to clear UNBOUND but can't directly call
4612 * worker_clr_flags() or adjust nr_running. Atomically
4613 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4614 * @worker will clear REBOUND using worker_clr_flags() when
4615 * it initiates the next execution cycle thus restoring
4616 * concurrency management. Note that when or whether
4617 * @worker clears REBOUND doesn't affect correctness.
4619 * ACCESS_ONCE() is necessary because @worker->flags may be
4620 * tested without holding any lock in
4621 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4622 * fail incorrectly leading to premature concurrency
4623 * management operations.
4625 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4626 worker_flags |= WORKER_REBOUND;
4627 worker_flags &= ~WORKER_UNBOUND;
4628 ACCESS_ONCE(worker->flags) = worker_flags;
4631 spin_unlock_irq(&pool->lock);
4635 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4636 * @pool: unbound pool of interest
4637 * @cpu: the CPU which is coming up
4639 * An unbound pool may end up with a cpumask which doesn't have any online
4640 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4641 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4642 * online CPU before, cpus_allowed of all its workers should be restored.
4644 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4646 static cpumask_t cpumask;
4647 struct worker *worker;
4650 lockdep_assert_held(&pool->manager_mutex);
4652 /* is @cpu allowed for @pool? */
4653 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4656 /* is @cpu the only online CPU? */
4657 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4658 if (cpumask_weight(&cpumask) != 1)
4661 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4662 for_each_pool_worker(worker, wi, pool)
4663 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4664 pool->attrs->cpumask) < 0);
4668 * Workqueues should be brought up before normal priority CPU notifiers.
4669 * This will be registered high priority CPU notifier.
4671 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4672 unsigned long action,
4675 int cpu = (unsigned long)hcpu;
4676 struct worker_pool *pool;
4677 struct workqueue_struct *wq;
4680 switch (action & ~CPU_TASKS_FROZEN) {
4681 case CPU_UP_PREPARE:
4682 for_each_cpu_worker_pool(pool, cpu) {
4683 if (pool->nr_workers)
4685 if (create_and_start_worker(pool) < 0)
4690 case CPU_DOWN_FAILED:
4692 mutex_lock(&wq_pool_mutex);
4694 for_each_pool(pool, pi) {
4695 mutex_lock(&pool->manager_mutex);
4697 if (pool->cpu == cpu) {
4698 spin_lock_irq(&pool->lock);
4699 pool->flags &= ~POOL_DISASSOCIATED;
4700 spin_unlock_irq(&pool->lock);
4702 rebind_workers(pool);
4703 } else if (pool->cpu < 0) {
4704 restore_unbound_workers_cpumask(pool, cpu);
4707 mutex_unlock(&pool->manager_mutex);
4710 /* update NUMA affinity of unbound workqueues */
4711 list_for_each_entry(wq, &workqueues, list)
4712 wq_update_unbound_numa(wq, cpu, true);
4714 mutex_unlock(&wq_pool_mutex);
4721 * Workqueues should be brought down after normal priority CPU notifiers.
4722 * This will be registered as low priority CPU notifier.
4724 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4725 unsigned long action,
4728 int cpu = (unsigned long)hcpu;
4729 struct work_struct unbind_work;
4730 struct workqueue_struct *wq;
4732 switch (action & ~CPU_TASKS_FROZEN) {
4733 case CPU_DOWN_PREPARE:
4734 /* unbinding per-cpu workers should happen on the local CPU */
4735 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4736 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4738 /* update NUMA affinity of unbound workqueues */
4739 mutex_lock(&wq_pool_mutex);
4740 list_for_each_entry(wq, &workqueues, list)
4741 wq_update_unbound_numa(wq, cpu, false);
4742 mutex_unlock(&wq_pool_mutex);
4744 /* wait for per-cpu unbinding to finish */
4745 flush_work(&unbind_work);
4753 struct work_for_cpu {
4754 struct work_struct work;
4760 static void work_for_cpu_fn(struct work_struct *work)
4762 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4764 wfc->ret = wfc->fn(wfc->arg);
4768 * work_on_cpu - run a function in user context on a particular cpu
4769 * @cpu: the cpu to run on
4770 * @fn: the function to run
4771 * @arg: the function arg
4773 * It is up to the caller to ensure that the cpu doesn't go offline.
4774 * The caller must not hold any locks which would prevent @fn from completing.
4776 * Return: The value @fn returns.
4778 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4780 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4782 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4783 schedule_work_on(cpu, &wfc.work);
4784 flush_work(&wfc.work);
4787 EXPORT_SYMBOL_GPL(work_on_cpu);
4788 #endif /* CONFIG_SMP */
4790 #ifdef CONFIG_FREEZER
4793 * freeze_workqueues_begin - begin freezing workqueues
4795 * Start freezing workqueues. After this function returns, all freezable
4796 * workqueues will queue new works to their delayed_works list instead of
4800 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4802 void freeze_workqueues_begin(void)
4804 struct worker_pool *pool;
4805 struct workqueue_struct *wq;
4806 struct pool_workqueue *pwq;
4809 mutex_lock(&wq_pool_mutex);
4811 WARN_ON_ONCE(workqueue_freezing);
4812 workqueue_freezing = true;
4815 for_each_pool(pool, pi) {
4816 spin_lock_irq(&pool->lock);
4817 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4818 pool->flags |= POOL_FREEZING;
4819 spin_unlock_irq(&pool->lock);
4822 list_for_each_entry(wq, &workqueues, list) {
4823 mutex_lock(&wq->mutex);
4824 for_each_pwq(pwq, wq)
4825 pwq_adjust_max_active(pwq);
4826 mutex_unlock(&wq->mutex);
4829 mutex_unlock(&wq_pool_mutex);
4833 * freeze_workqueues_busy - are freezable workqueues still busy?
4835 * Check whether freezing is complete. This function must be called
4836 * between freeze_workqueues_begin() and thaw_workqueues().
4839 * Grabs and releases wq_pool_mutex.
4842 * %true if some freezable workqueues are still busy. %false if freezing
4845 bool freeze_workqueues_busy(void)
4848 struct workqueue_struct *wq;
4849 struct pool_workqueue *pwq;
4851 mutex_lock(&wq_pool_mutex);
4853 WARN_ON_ONCE(!workqueue_freezing);
4855 list_for_each_entry(wq, &workqueues, list) {
4856 if (!(wq->flags & WQ_FREEZABLE))
4859 * nr_active is monotonically decreasing. It's safe
4860 * to peek without lock.
4862 rcu_read_lock_sched();
4863 for_each_pwq(pwq, wq) {
4864 WARN_ON_ONCE(pwq->nr_active < 0);
4865 if (pwq->nr_active) {
4867 rcu_read_unlock_sched();
4871 rcu_read_unlock_sched();
4874 mutex_unlock(&wq_pool_mutex);
4879 * thaw_workqueues - thaw workqueues
4881 * Thaw workqueues. Normal queueing is restored and all collected
4882 * frozen works are transferred to their respective pool worklists.
4885 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4887 void thaw_workqueues(void)
4889 struct workqueue_struct *wq;
4890 struct pool_workqueue *pwq;
4891 struct worker_pool *pool;
4894 mutex_lock(&wq_pool_mutex);
4896 if (!workqueue_freezing)
4899 /* clear FREEZING */
4900 for_each_pool(pool, pi) {
4901 spin_lock_irq(&pool->lock);
4902 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4903 pool->flags &= ~POOL_FREEZING;
4904 spin_unlock_irq(&pool->lock);
4907 /* restore max_active and repopulate worklist */
4908 list_for_each_entry(wq, &workqueues, list) {
4909 mutex_lock(&wq->mutex);
4910 for_each_pwq(pwq, wq)
4911 pwq_adjust_max_active(pwq);
4912 mutex_unlock(&wq->mutex);
4915 workqueue_freezing = false;
4917 mutex_unlock(&wq_pool_mutex);
4919 #endif /* CONFIG_FREEZER */
4921 static void __init wq_numa_init(void)
4926 /* determine NUMA pwq table len - highest node id + 1 */
4928 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4930 if (num_possible_nodes() <= 1)
4933 if (wq_disable_numa) {
4934 pr_info("workqueue: NUMA affinity support disabled\n");
4938 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4939 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4942 * We want masks of possible CPUs of each node which isn't readily
4943 * available. Build one from cpu_to_node() which should have been
4944 * fully initialized by now.
4946 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4950 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4951 node_online(node) ? node : NUMA_NO_NODE));
4953 for_each_possible_cpu(cpu) {
4954 node = cpu_to_node(cpu);
4955 if (WARN_ON(node == NUMA_NO_NODE)) {
4956 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4957 /* happens iff arch is bonkers, let's just proceed */
4960 cpumask_set_cpu(cpu, tbl[node]);
4963 wq_numa_possible_cpumask = tbl;
4964 wq_numa_enabled = true;
4967 static int __init init_workqueues(void)
4969 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4972 /* make sure we have enough bits for OFFQ pool ID */
4973 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4974 WORK_CPU_END * NR_STD_WORKER_POOLS);
4976 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4978 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4980 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4981 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4985 /* initialize CPU pools */
4986 for_each_possible_cpu(cpu) {
4987 struct worker_pool *pool;
4990 for_each_cpu_worker_pool(pool, cpu) {
4991 BUG_ON(init_worker_pool(pool));
4993 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4994 pool->attrs->nice = std_nice[i++];
4995 pool->node = cpu_to_node(cpu);
4998 mutex_lock(&wq_pool_mutex);
4999 BUG_ON(worker_pool_assign_id(pool));
5000 mutex_unlock(&wq_pool_mutex);
5004 /* create the initial worker */
5005 for_each_online_cpu(cpu) {
5006 struct worker_pool *pool;
5008 for_each_cpu_worker_pool(pool, cpu) {
5009 pool->flags &= ~POOL_DISASSOCIATED;
5010 BUG_ON(create_and_start_worker(pool) < 0);
5014 /* create default unbound wq attrs */
5015 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5016 struct workqueue_attrs *attrs;
5018 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5019 attrs->nice = std_nice[i];
5020 unbound_std_wq_attrs[i] = attrs;
5023 system_wq = alloc_workqueue("events", 0, 0);
5024 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5025 system_long_wq = alloc_workqueue("events_long", 0, 0);
5026 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5027 WQ_UNBOUND_MAX_ACTIVE);
5028 system_freezable_wq = alloc_workqueue("events_freezable",
5030 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5031 WQ_POWER_EFFICIENT, 0);
5032 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5033 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5035 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5036 !system_unbound_wq || !system_freezable_wq ||
5037 !system_power_efficient_wq ||
5038 !system_freezable_power_efficient_wq);
5041 early_initcall(init_workqueues);