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 /* I: attributes used when instantiating ordered pools on demand */
308 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
310 struct workqueue_struct *system_wq __read_mostly;
311 EXPORT_SYMBOL(system_wq);
312 struct workqueue_struct *system_highpri_wq __read_mostly;
313 EXPORT_SYMBOL_GPL(system_highpri_wq);
314 struct workqueue_struct *system_long_wq __read_mostly;
315 EXPORT_SYMBOL_GPL(system_long_wq);
316 struct workqueue_struct *system_unbound_wq __read_mostly;
317 EXPORT_SYMBOL_GPL(system_unbound_wq);
318 struct workqueue_struct *system_freezable_wq __read_mostly;
319 EXPORT_SYMBOL_GPL(system_freezable_wq);
320 struct workqueue_struct *system_power_efficient_wq __read_mostly;
321 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
322 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
323 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
325 static int worker_thread(void *__worker);
326 static void copy_workqueue_attrs(struct workqueue_attrs *to,
327 const struct workqueue_attrs *from);
329 #define CREATE_TRACE_POINTS
330 #include <trace/events/workqueue.h>
332 #define assert_rcu_or_pool_mutex() \
333 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
334 lockdep_is_held(&wq_pool_mutex), \
335 "sched RCU or wq_pool_mutex should be held")
337 #define assert_rcu_or_wq_mutex(wq) \
338 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
339 lockdep_is_held(&wq->mutex), \
340 "sched RCU or wq->mutex should be held")
342 #ifdef CONFIG_LOCKDEP
343 #define assert_manager_or_pool_lock(pool) \
344 WARN_ONCE(debug_locks && \
345 !lockdep_is_held(&(pool)->manager_mutex) && \
346 !lockdep_is_held(&(pool)->lock), \
347 "pool->manager_mutex or ->lock should be held")
349 #define assert_manager_or_pool_lock(pool) do { } while (0)
352 #define for_each_cpu_worker_pool(pool, cpu) \
353 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
354 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
358 * for_each_pool - iterate through all worker_pools in the system
359 * @pool: iteration cursor
360 * @pi: integer used for iteration
362 * This must be called either with wq_pool_mutex held or sched RCU read
363 * locked. If the pool needs to be used beyond the locking in effect, the
364 * caller is responsible for guaranteeing that the pool stays online.
366 * The if/else clause exists only for the lockdep assertion and can be
369 #define for_each_pool(pool, pi) \
370 idr_for_each_entry(&worker_pool_idr, pool, pi) \
371 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
375 * for_each_pool_worker - iterate through all workers of a worker_pool
376 * @worker: iteration cursor
377 * @wi: integer used for iteration
378 * @pool: worker_pool to iterate workers of
380 * This must be called with either @pool->manager_mutex or ->lock held.
382 * The if/else clause exists only for the lockdep assertion and can be
385 #define for_each_pool_worker(worker, wi, pool) \
386 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
387 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
391 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
392 * @pwq: iteration cursor
393 * @wq: the target workqueue
395 * This must be called either with wq->mutex held or sched RCU read locked.
396 * If the pwq needs to be used beyond the locking in effect, the caller is
397 * responsible for guaranteeing that the pwq stays online.
399 * The if/else clause exists only for the lockdep assertion and can be
402 #define for_each_pwq(pwq, wq) \
403 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
404 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
407 #ifdef CONFIG_DEBUG_OBJECTS_WORK
409 static struct debug_obj_descr work_debug_descr;
411 static void *work_debug_hint(void *addr)
413 return ((struct work_struct *) addr)->func;
417 * fixup_init is called when:
418 * - an active object is initialized
420 static int work_fixup_init(void *addr, enum debug_obj_state state)
422 struct work_struct *work = addr;
425 case ODEBUG_STATE_ACTIVE:
426 cancel_work_sync(work);
427 debug_object_init(work, &work_debug_descr);
435 * fixup_activate is called when:
436 * - an active object is activated
437 * - an unknown object is activated (might be a statically initialized object)
439 static int work_fixup_activate(void *addr, enum debug_obj_state state)
441 struct work_struct *work = addr;
445 case ODEBUG_STATE_NOTAVAILABLE:
447 * This is not really a fixup. The work struct was
448 * statically initialized. We just make sure that it
449 * is tracked in the object tracker.
451 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
452 debug_object_init(work, &work_debug_descr);
453 debug_object_activate(work, &work_debug_descr);
459 case ODEBUG_STATE_ACTIVE:
468 * fixup_free is called when:
469 * - an active object is freed
471 static int work_fixup_free(void *addr, enum debug_obj_state state)
473 struct work_struct *work = addr;
476 case ODEBUG_STATE_ACTIVE:
477 cancel_work_sync(work);
478 debug_object_free(work, &work_debug_descr);
485 static struct debug_obj_descr work_debug_descr = {
486 .name = "work_struct",
487 .debug_hint = work_debug_hint,
488 .fixup_init = work_fixup_init,
489 .fixup_activate = work_fixup_activate,
490 .fixup_free = work_fixup_free,
493 static inline void debug_work_activate(struct work_struct *work)
495 debug_object_activate(work, &work_debug_descr);
498 static inline void debug_work_deactivate(struct work_struct *work)
500 debug_object_deactivate(work, &work_debug_descr);
503 void __init_work(struct work_struct *work, int onstack)
506 debug_object_init_on_stack(work, &work_debug_descr);
508 debug_object_init(work, &work_debug_descr);
510 EXPORT_SYMBOL_GPL(__init_work);
512 void destroy_work_on_stack(struct work_struct *work)
514 debug_object_free(work, &work_debug_descr);
516 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
519 static inline void debug_work_activate(struct work_struct *work) { }
520 static inline void debug_work_deactivate(struct work_struct *work) { }
523 /* allocate ID and assign it to @pool */
524 static int worker_pool_assign_id(struct worker_pool *pool)
528 lockdep_assert_held(&wq_pool_mutex);
530 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
539 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
540 * @wq: the target workqueue
543 * This must be called either with pwq_lock held or sched RCU read locked.
544 * If the pwq needs to be used beyond the locking in effect, the caller is
545 * responsible for guaranteeing that the pwq stays online.
547 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
550 assert_rcu_or_wq_mutex(wq);
551 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
554 static unsigned int work_color_to_flags(int color)
556 return color << WORK_STRUCT_COLOR_SHIFT;
559 static int get_work_color(struct work_struct *work)
561 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
562 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
565 static int work_next_color(int color)
567 return (color + 1) % WORK_NR_COLORS;
571 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
572 * contain the pointer to the queued pwq. Once execution starts, the flag
573 * is cleared and the high bits contain OFFQ flags and pool ID.
575 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
576 * and clear_work_data() can be used to set the pwq, pool or clear
577 * work->data. These functions should only be called while the work is
578 * owned - ie. while the PENDING bit is set.
580 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
581 * corresponding to a work. Pool is available once the work has been
582 * queued anywhere after initialization until it is sync canceled. pwq is
583 * available only while the work item is queued.
585 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
586 * canceled. While being canceled, a work item may have its PENDING set
587 * but stay off timer and worklist for arbitrarily long and nobody should
588 * try to steal the PENDING bit.
590 static inline void set_work_data(struct work_struct *work, unsigned long data,
593 WARN_ON_ONCE(!work_pending(work));
594 atomic_long_set(&work->data, data | flags | work_static(work));
597 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
598 unsigned long extra_flags)
600 set_work_data(work, (unsigned long)pwq,
601 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
604 static void set_work_pool_and_keep_pending(struct work_struct *work,
607 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
608 WORK_STRUCT_PENDING);
611 static void set_work_pool_and_clear_pending(struct work_struct *work,
615 * The following wmb is paired with the implied mb in
616 * test_and_set_bit(PENDING) and ensures all updates to @work made
617 * here are visible to and precede any updates by the next PENDING
621 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
624 static void clear_work_data(struct work_struct *work)
626 smp_wmb(); /* see set_work_pool_and_clear_pending() */
627 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
630 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
632 unsigned long data = atomic_long_read(&work->data);
634 if (data & WORK_STRUCT_PWQ)
635 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
641 * get_work_pool - return the worker_pool a given work was associated with
642 * @work: the work item of interest
644 * Return the worker_pool @work was last associated with. %NULL if none.
646 * Pools are created and destroyed under wq_pool_mutex, and allows read
647 * access under sched-RCU read lock. As such, this function should be
648 * called under wq_pool_mutex or with preemption disabled.
650 * All fields of the returned pool are accessible as long as the above
651 * mentioned locking is in effect. If the returned pool needs to be used
652 * beyond the critical section, the caller is responsible for ensuring the
653 * returned pool is and stays online.
655 static struct worker_pool *get_work_pool(struct work_struct *work)
657 unsigned long data = atomic_long_read(&work->data);
660 assert_rcu_or_pool_mutex();
662 if (data & WORK_STRUCT_PWQ)
663 return ((struct pool_workqueue *)
664 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
666 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
667 if (pool_id == WORK_OFFQ_POOL_NONE)
670 return idr_find(&worker_pool_idr, pool_id);
674 * get_work_pool_id - return the worker pool ID a given work is associated with
675 * @work: the work item of interest
677 * Return the worker_pool ID @work was last associated with.
678 * %WORK_OFFQ_POOL_NONE if none.
680 static int get_work_pool_id(struct work_struct *work)
682 unsigned long data = atomic_long_read(&work->data);
684 if (data & WORK_STRUCT_PWQ)
685 return ((struct pool_workqueue *)
686 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
688 return data >> WORK_OFFQ_POOL_SHIFT;
691 static void mark_work_canceling(struct work_struct *work)
693 unsigned long pool_id = get_work_pool_id(work);
695 pool_id <<= WORK_OFFQ_POOL_SHIFT;
696 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
699 static bool work_is_canceling(struct work_struct *work)
701 unsigned long data = atomic_long_read(&work->data);
703 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
707 * Policy functions. These define the policies on how the global worker
708 * pools are managed. Unless noted otherwise, these functions assume that
709 * they're being called with pool->lock held.
712 static bool __need_more_worker(struct worker_pool *pool)
714 return !atomic_read(&pool->nr_running);
718 * Need to wake up a worker? Called from anything but currently
721 * Note that, because unbound workers never contribute to nr_running, this
722 * function will always return %true for unbound pools as long as the
723 * worklist isn't empty.
725 static bool need_more_worker(struct worker_pool *pool)
727 return !list_empty(&pool->worklist) && __need_more_worker(pool);
730 /* Can I start working? Called from busy but !running workers. */
731 static bool may_start_working(struct worker_pool *pool)
733 return pool->nr_idle;
736 /* Do I need to keep working? Called from currently running workers. */
737 static bool keep_working(struct worker_pool *pool)
739 return !list_empty(&pool->worklist) &&
740 atomic_read(&pool->nr_running) <= 1;
743 /* Do we need a new worker? Called from manager. */
744 static bool need_to_create_worker(struct worker_pool *pool)
746 return need_more_worker(pool) && !may_start_working(pool);
749 /* Do I need to be the manager? */
750 static bool need_to_manage_workers(struct worker_pool *pool)
752 return need_to_create_worker(pool) ||
753 (pool->flags & POOL_MANAGE_WORKERS);
756 /* Do we have too many workers and should some go away? */
757 static bool too_many_workers(struct worker_pool *pool)
759 bool managing = mutex_is_locked(&pool->manager_arb);
760 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
761 int nr_busy = pool->nr_workers - nr_idle;
764 * nr_idle and idle_list may disagree if idle rebinding is in
765 * progress. Never return %true if idle_list is empty.
767 if (list_empty(&pool->idle_list))
770 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
777 /* Return the first worker. Safe with preemption disabled */
778 static struct worker *first_worker(struct worker_pool *pool)
780 if (unlikely(list_empty(&pool->idle_list)))
783 return list_first_entry(&pool->idle_list, struct worker, entry);
787 * wake_up_worker - wake up an idle worker
788 * @pool: worker pool to wake worker from
790 * Wake up the first idle worker of @pool.
793 * spin_lock_irq(pool->lock).
795 static void wake_up_worker(struct worker_pool *pool)
797 struct worker *worker = first_worker(pool);
800 wake_up_process(worker->task);
804 * wq_worker_waking_up - a worker is waking up
805 * @task: task waking up
806 * @cpu: CPU @task is waking up to
808 * This function is called during try_to_wake_up() when a worker is
812 * spin_lock_irq(rq->lock)
814 void wq_worker_waking_up(struct task_struct *task, int cpu)
816 struct worker *worker = kthread_data(task);
818 if (!(worker->flags & WORKER_NOT_RUNNING)) {
819 WARN_ON_ONCE(worker->pool->cpu != cpu);
820 atomic_inc(&worker->pool->nr_running);
825 * wq_worker_sleeping - a worker is going to sleep
826 * @task: task going to sleep
827 * @cpu: CPU in question, must be the current CPU number
829 * This function is called during schedule() when a busy worker is
830 * going to sleep. Worker on the same cpu can be woken up by
831 * returning pointer to its task.
834 * spin_lock_irq(rq->lock)
837 * Worker task on @cpu to wake up, %NULL if none.
839 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
841 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
842 struct worker_pool *pool;
845 * Rescuers, which may not have all the fields set up like normal
846 * workers, also reach here, let's not access anything before
847 * checking NOT_RUNNING.
849 if (worker->flags & WORKER_NOT_RUNNING)
854 /* this can only happen on the local cpu */
855 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
859 * The counterpart of the following dec_and_test, implied mb,
860 * worklist not empty test sequence is in insert_work().
861 * Please read comment there.
863 * NOT_RUNNING is clear. This means that we're bound to and
864 * running on the local cpu w/ rq lock held and preemption
865 * disabled, which in turn means that none else could be
866 * manipulating idle_list, so dereferencing idle_list without pool
869 if (atomic_dec_and_test(&pool->nr_running) &&
870 !list_empty(&pool->worklist))
871 to_wakeup = first_worker(pool);
872 return to_wakeup ? to_wakeup->task : NULL;
876 * worker_set_flags - set worker flags and adjust nr_running accordingly
878 * @flags: flags to set
879 * @wakeup: wakeup an idle worker if necessary
881 * Set @flags in @worker->flags and adjust nr_running accordingly. If
882 * nr_running becomes zero and @wakeup is %true, an idle worker is
886 * spin_lock_irq(pool->lock)
888 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
891 struct worker_pool *pool = worker->pool;
893 WARN_ON_ONCE(worker->task != current);
896 * If transitioning into NOT_RUNNING, adjust nr_running and
897 * wake up an idle worker as necessary if requested by
900 if ((flags & WORKER_NOT_RUNNING) &&
901 !(worker->flags & WORKER_NOT_RUNNING)) {
903 if (atomic_dec_and_test(&pool->nr_running) &&
904 !list_empty(&pool->worklist))
905 wake_up_worker(pool);
907 atomic_dec(&pool->nr_running);
910 worker->flags |= flags;
914 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
916 * @flags: flags to clear
918 * Clear @flags in @worker->flags and adjust nr_running accordingly.
921 * spin_lock_irq(pool->lock)
923 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
925 struct worker_pool *pool = worker->pool;
926 unsigned int oflags = worker->flags;
928 WARN_ON_ONCE(worker->task != current);
930 worker->flags &= ~flags;
933 * If transitioning out of NOT_RUNNING, increment nr_running. Note
934 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
935 * of multiple flags, not a single flag.
937 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
938 if (!(worker->flags & WORKER_NOT_RUNNING))
939 atomic_inc(&pool->nr_running);
943 * find_worker_executing_work - find worker which is executing a work
944 * @pool: pool of interest
945 * @work: work to find worker for
947 * Find a worker which is executing @work on @pool by searching
948 * @pool->busy_hash which is keyed by the address of @work. For a worker
949 * to match, its current execution should match the address of @work and
950 * its work function. This is to avoid unwanted dependency between
951 * unrelated work executions through a work item being recycled while still
954 * This is a bit tricky. A work item may be freed once its execution
955 * starts and nothing prevents the freed area from being recycled for
956 * another work item. If the same work item address ends up being reused
957 * before the original execution finishes, workqueue will identify the
958 * recycled work item as currently executing and make it wait until the
959 * current execution finishes, introducing an unwanted dependency.
961 * This function checks the work item address and work function to avoid
962 * false positives. Note that this isn't complete as one may construct a
963 * work function which can introduce dependency onto itself through a
964 * recycled work item. Well, if somebody wants to shoot oneself in the
965 * foot that badly, there's only so much we can do, and if such deadlock
966 * actually occurs, it should be easy to locate the culprit work function.
969 * spin_lock_irq(pool->lock).
972 * Pointer to worker which is executing @work if found, NULL
975 static struct worker *find_worker_executing_work(struct worker_pool *pool,
976 struct work_struct *work)
978 struct worker *worker;
980 hash_for_each_possible(pool->busy_hash, worker, hentry,
982 if (worker->current_work == work &&
983 worker->current_func == work->func)
990 * move_linked_works - move linked works to a list
991 * @work: start of series of works to be scheduled
992 * @head: target list to append @work to
993 * @nextp: out paramter for nested worklist walking
995 * Schedule linked works starting from @work to @head. Work series to
996 * be scheduled starts at @work and includes any consecutive work with
997 * WORK_STRUCT_LINKED set in its predecessor.
999 * If @nextp is not NULL, it's updated to point to the next work of
1000 * the last scheduled work. This allows move_linked_works() to be
1001 * nested inside outer list_for_each_entry_safe().
1004 * spin_lock_irq(pool->lock).
1006 static void move_linked_works(struct work_struct *work, struct list_head *head,
1007 struct work_struct **nextp)
1009 struct work_struct *n;
1012 * Linked worklist will always end before the end of the list,
1013 * use NULL for list head.
1015 list_for_each_entry_safe_from(work, n, NULL, entry) {
1016 list_move_tail(&work->entry, head);
1017 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1022 * If we're already inside safe list traversal and have moved
1023 * multiple works to the scheduled queue, the next position
1024 * needs to be updated.
1031 * get_pwq - get an extra reference on the specified pool_workqueue
1032 * @pwq: pool_workqueue to get
1034 * Obtain an extra reference on @pwq. The caller should guarantee that
1035 * @pwq has positive refcnt and be holding the matching pool->lock.
1037 static void get_pwq(struct pool_workqueue *pwq)
1039 lockdep_assert_held(&pwq->pool->lock);
1040 WARN_ON_ONCE(pwq->refcnt <= 0);
1045 * put_pwq - put a pool_workqueue reference
1046 * @pwq: pool_workqueue to put
1048 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1049 * destruction. The caller should be holding the matching pool->lock.
1051 static void put_pwq(struct pool_workqueue *pwq)
1053 lockdep_assert_held(&pwq->pool->lock);
1054 if (likely(--pwq->refcnt))
1056 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1059 * @pwq can't be released under pool->lock, bounce to
1060 * pwq_unbound_release_workfn(). This never recurses on the same
1061 * pool->lock as this path is taken only for unbound workqueues and
1062 * the release work item is scheduled on a per-cpu workqueue. To
1063 * avoid lockdep warning, unbound pool->locks are given lockdep
1064 * subclass of 1 in get_unbound_pool().
1066 schedule_work(&pwq->unbound_release_work);
1070 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1071 * @pwq: pool_workqueue to put (can be %NULL)
1073 * put_pwq() with locking. This function also allows %NULL @pwq.
1075 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1079 * As both pwqs and pools are sched-RCU protected, the
1080 * following lock operations are safe.
1082 spin_lock_irq(&pwq->pool->lock);
1084 spin_unlock_irq(&pwq->pool->lock);
1088 static void pwq_activate_delayed_work(struct work_struct *work)
1090 struct pool_workqueue *pwq = get_work_pwq(work);
1092 trace_workqueue_activate_work(work);
1093 move_linked_works(work, &pwq->pool->worklist, NULL);
1094 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1098 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1100 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1101 struct work_struct, entry);
1103 pwq_activate_delayed_work(work);
1107 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1108 * @pwq: pwq of interest
1109 * @color: color of work which left the queue
1111 * A work either has completed or is removed from pending queue,
1112 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1115 * spin_lock_irq(pool->lock).
1117 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1119 /* uncolored work items don't participate in flushing or nr_active */
1120 if (color == WORK_NO_COLOR)
1123 pwq->nr_in_flight[color]--;
1126 if (!list_empty(&pwq->delayed_works)) {
1127 /* one down, submit a delayed one */
1128 if (pwq->nr_active < pwq->max_active)
1129 pwq_activate_first_delayed(pwq);
1132 /* is flush in progress and are we at the flushing tip? */
1133 if (likely(pwq->flush_color != color))
1136 /* are there still in-flight works? */
1137 if (pwq->nr_in_flight[color])
1140 /* this pwq is done, clear flush_color */
1141 pwq->flush_color = -1;
1144 * If this was the last pwq, wake up the first flusher. It
1145 * will handle the rest.
1147 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1148 complete(&pwq->wq->first_flusher->done);
1154 * try_to_grab_pending - steal work item from worklist and disable irq
1155 * @work: work item to steal
1156 * @is_dwork: @work is a delayed_work
1157 * @flags: place to store irq state
1159 * Try to grab PENDING bit of @work. This function can handle @work in any
1160 * stable state - idle, on timer or on worklist. Return values are
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
1168 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1169 * interrupted while holding PENDING and @work off queue, irq must be
1170 * disabled on entry. This, combined with delayed_work->timer being
1171 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1173 * On successful return, >= 0, irq is disabled and the caller is
1174 * responsible for releasing it using local_irq_restore(*@flags).
1176 * This function is safe to call from any context including IRQ handler.
1178 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1179 unsigned long *flags)
1181 struct worker_pool *pool;
1182 struct pool_workqueue *pwq;
1184 local_irq_save(*flags);
1186 /* try to steal the timer if it exists */
1188 struct delayed_work *dwork = to_delayed_work(work);
1191 * dwork->timer is irqsafe. If del_timer() fails, it's
1192 * guaranteed that the timer is not queued anywhere and not
1193 * running on the local CPU.
1195 if (likely(del_timer(&dwork->timer)))
1199 /* try to claim PENDING the normal way */
1200 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1204 * The queueing is in progress, or it is already queued. Try to
1205 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1207 pool = get_work_pool(work);
1211 spin_lock(&pool->lock);
1213 * work->data is guaranteed to point to pwq only while the work
1214 * item is queued on pwq->wq, and both updating work->data to point
1215 * to pwq on queueing and to pool on dequeueing are done under
1216 * pwq->pool->lock. This in turn guarantees that, if work->data
1217 * points to pwq which is associated with a locked pool, the work
1218 * item is currently queued on that pool.
1220 pwq = get_work_pwq(work);
1221 if (pwq && pwq->pool == pool) {
1222 debug_work_deactivate(work);
1225 * A delayed work item cannot be grabbed directly because
1226 * it might have linked NO_COLOR work items which, if left
1227 * on the delayed_list, will confuse pwq->nr_active
1228 * management later on and cause stall. Make sure the work
1229 * item is activated before grabbing.
1231 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1232 pwq_activate_delayed_work(work);
1234 list_del_init(&work->entry);
1235 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1237 /* work->data points to pwq iff queued, point to pool */
1238 set_work_pool_and_keep_pending(work, pool->id);
1240 spin_unlock(&pool->lock);
1243 spin_unlock(&pool->lock);
1245 local_irq_restore(*flags);
1246 if (work_is_canceling(work))
1253 * insert_work - insert a work into a pool
1254 * @pwq: pwq @work belongs to
1255 * @work: work to insert
1256 * @head: insertion point
1257 * @extra_flags: extra WORK_STRUCT_* flags to set
1259 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1260 * work_struct flags.
1263 * spin_lock_irq(pool->lock).
1265 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1266 struct list_head *head, unsigned int extra_flags)
1268 struct worker_pool *pool = pwq->pool;
1270 /* we own @work, set data and link */
1271 set_work_pwq(work, pwq, extra_flags);
1272 list_add_tail(&work->entry, head);
1276 * Ensure either wq_worker_sleeping() sees the above
1277 * list_add_tail() or we see zero nr_running to avoid workers lying
1278 * around lazily while there are works to be processed.
1282 if (__need_more_worker(pool))
1283 wake_up_worker(pool);
1287 * Test whether @work is being queued from another work executing on the
1290 static bool is_chained_work(struct workqueue_struct *wq)
1292 struct worker *worker;
1294 worker = current_wq_worker();
1296 * Return %true iff I'm a worker execuing a work item on @wq. If
1297 * I'm @worker, it's safe to dereference it without locking.
1299 return worker && worker->current_pwq->wq == wq;
1302 static void __queue_work(int cpu, struct workqueue_struct *wq,
1303 struct work_struct *work)
1305 struct pool_workqueue *pwq;
1306 struct worker_pool *last_pool;
1307 struct list_head *worklist;
1308 unsigned int work_flags;
1309 unsigned int req_cpu = cpu;
1312 * While a work item is PENDING && off queue, a task trying to
1313 * steal the PENDING will busy-loop waiting for it to either get
1314 * queued or lose PENDING. Grabbing PENDING and queueing should
1315 * happen with IRQ disabled.
1317 WARN_ON_ONCE(!irqs_disabled());
1319 debug_work_activate(work);
1321 /* if dying, only works from the same workqueue are allowed */
1322 if (unlikely(wq->flags & __WQ_DRAINING) &&
1323 WARN_ON_ONCE(!is_chained_work(wq)))
1326 if (req_cpu == WORK_CPU_UNBOUND)
1327 cpu = raw_smp_processor_id();
1329 /* pwq which will be used unless @work is executing elsewhere */
1330 if (!(wq->flags & WQ_UNBOUND))
1331 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1333 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1336 * If @work was previously on a different pool, it might still be
1337 * running there, in which case the work needs to be queued on that
1338 * pool to guarantee non-reentrancy.
1340 last_pool = get_work_pool(work);
1341 if (last_pool && last_pool != pwq->pool) {
1342 struct worker *worker;
1344 spin_lock(&last_pool->lock);
1346 worker = find_worker_executing_work(last_pool, work);
1348 if (worker && worker->current_pwq->wq == wq) {
1349 pwq = worker->current_pwq;
1351 /* meh... not running there, queue here */
1352 spin_unlock(&last_pool->lock);
1353 spin_lock(&pwq->pool->lock);
1356 spin_lock(&pwq->pool->lock);
1360 * pwq is determined and locked. For unbound pools, we could have
1361 * raced with pwq release and it could already be dead. If its
1362 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1363 * without another pwq replacing it in the numa_pwq_tbl or while
1364 * work items are executing on it, so the retrying is guaranteed to
1365 * make forward-progress.
1367 if (unlikely(!pwq->refcnt)) {
1368 if (wq->flags & WQ_UNBOUND) {
1369 spin_unlock(&pwq->pool->lock);
1374 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1378 /* pwq determined, queue */
1379 trace_workqueue_queue_work(req_cpu, pwq, work);
1381 if (WARN_ON(!list_empty(&work->entry))) {
1382 spin_unlock(&pwq->pool->lock);
1386 pwq->nr_in_flight[pwq->work_color]++;
1387 work_flags = work_color_to_flags(pwq->work_color);
1389 if (likely(pwq->nr_active < pwq->max_active)) {
1390 trace_workqueue_activate_work(work);
1392 worklist = &pwq->pool->worklist;
1394 work_flags |= WORK_STRUCT_DELAYED;
1395 worklist = &pwq->delayed_works;
1398 insert_work(pwq, work, worklist, work_flags);
1400 spin_unlock(&pwq->pool->lock);
1404 * queue_work_on - queue work on specific cpu
1405 * @cpu: CPU number to execute work on
1406 * @wq: workqueue to use
1407 * @work: work to queue
1409 * Returns %false if @work was already on a queue, %true otherwise.
1411 * We queue the work to a specific CPU, the caller must ensure it
1414 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1415 struct work_struct *work)
1418 unsigned long flags;
1420 local_irq_save(flags);
1422 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1423 __queue_work(cpu, wq, work);
1427 local_irq_restore(flags);
1430 EXPORT_SYMBOL(queue_work_on);
1432 void delayed_work_timer_fn(unsigned long __data)
1434 struct delayed_work *dwork = (struct delayed_work *)__data;
1436 /* should have been called from irqsafe timer with irq already off */
1437 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1439 EXPORT_SYMBOL(delayed_work_timer_fn);
1441 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1442 struct delayed_work *dwork, unsigned long delay)
1444 struct timer_list *timer = &dwork->timer;
1445 struct work_struct *work = &dwork->work;
1447 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1448 timer->data != (unsigned long)dwork);
1449 WARN_ON_ONCE(timer_pending(timer));
1450 WARN_ON_ONCE(!list_empty(&work->entry));
1453 * If @delay is 0, queue @dwork->work immediately. This is for
1454 * both optimization and correctness. The earliest @timer can
1455 * expire is on the closest next tick and delayed_work users depend
1456 * on that there's no such delay when @delay is 0.
1459 __queue_work(cpu, wq, &dwork->work);
1463 timer_stats_timer_set_start_info(&dwork->timer);
1467 timer->expires = jiffies + delay;
1469 if (unlikely(cpu != WORK_CPU_UNBOUND))
1470 add_timer_on(timer, cpu);
1476 * queue_delayed_work_on - queue work on specific CPU after delay
1477 * @cpu: CPU number to execute work on
1478 * @wq: workqueue to use
1479 * @dwork: work to queue
1480 * @delay: number of jiffies to wait before queueing
1482 * Returns %false if @work was already on a queue, %true otherwise. If
1483 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1486 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1487 struct delayed_work *dwork, unsigned long delay)
1489 struct work_struct *work = &dwork->work;
1491 unsigned long flags;
1493 /* read the comment in __queue_work() */
1494 local_irq_save(flags);
1496 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1497 __queue_delayed_work(cpu, wq, dwork, delay);
1501 local_irq_restore(flags);
1504 EXPORT_SYMBOL(queue_delayed_work_on);
1507 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1508 * @cpu: CPU number to execute work on
1509 * @wq: workqueue to use
1510 * @dwork: work to queue
1511 * @delay: number of jiffies to wait before queueing
1513 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1514 * modify @dwork's timer so that it expires after @delay. If @delay is
1515 * zero, @work is guaranteed to be scheduled immediately regardless of its
1518 * Returns %false if @dwork was idle and queued, %true if @dwork was
1519 * pending and its timer was modified.
1521 * This function is safe to call from any context including IRQ handler.
1522 * See try_to_grab_pending() for details.
1524 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1525 struct delayed_work *dwork, unsigned long delay)
1527 unsigned long flags;
1531 ret = try_to_grab_pending(&dwork->work, true, &flags);
1532 } while (unlikely(ret == -EAGAIN));
1534 if (likely(ret >= 0)) {
1535 __queue_delayed_work(cpu, wq, dwork, delay);
1536 local_irq_restore(flags);
1539 /* -ENOENT from try_to_grab_pending() becomes %true */
1542 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1545 * worker_enter_idle - enter idle state
1546 * @worker: worker which is entering idle state
1548 * @worker is entering idle state. Update stats and idle timer if
1552 * spin_lock_irq(pool->lock).
1554 static void worker_enter_idle(struct worker *worker)
1556 struct worker_pool *pool = worker->pool;
1558 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1559 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1560 (worker->hentry.next || worker->hentry.pprev)))
1563 /* can't use worker_set_flags(), also called from start_worker() */
1564 worker->flags |= WORKER_IDLE;
1566 worker->last_active = jiffies;
1568 /* idle_list is LIFO */
1569 list_add(&worker->entry, &pool->idle_list);
1571 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1572 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1575 * Sanity check nr_running. Because wq_unbind_fn() releases
1576 * pool->lock between setting %WORKER_UNBOUND and zapping
1577 * nr_running, the warning may trigger spuriously. Check iff
1578 * unbind is not in progress.
1580 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1581 pool->nr_workers == pool->nr_idle &&
1582 atomic_read(&pool->nr_running));
1586 * worker_leave_idle - leave idle state
1587 * @worker: worker which is leaving idle state
1589 * @worker is leaving idle state. Update stats.
1592 * spin_lock_irq(pool->lock).
1594 static void worker_leave_idle(struct worker *worker)
1596 struct worker_pool *pool = worker->pool;
1598 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1600 worker_clr_flags(worker, WORKER_IDLE);
1602 list_del_init(&worker->entry);
1606 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1607 * @pool: target worker_pool
1609 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1611 * Works which are scheduled while the cpu is online must at least be
1612 * scheduled to a worker which is bound to the cpu so that if they are
1613 * flushed from cpu callbacks while cpu is going down, they are
1614 * guaranteed to execute on the cpu.
1616 * This function is to be used by unbound workers and rescuers to bind
1617 * themselves to the target cpu and may race with cpu going down or
1618 * coming online. kthread_bind() can't be used because it may put the
1619 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1620 * verbatim as it's best effort and blocking and pool may be
1621 * [dis]associated in the meantime.
1623 * This function tries set_cpus_allowed() and locks pool and verifies the
1624 * binding against %POOL_DISASSOCIATED which is set during
1625 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1626 * enters idle state or fetches works without dropping lock, it can
1627 * guarantee the scheduling requirement described in the first paragraph.
1630 * Might sleep. Called without any lock but returns with pool->lock
1634 * %true if the associated pool is online (@worker is successfully
1635 * bound), %false if offline.
1637 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1638 __acquires(&pool->lock)
1642 * The following call may fail, succeed or succeed
1643 * without actually migrating the task to the cpu if
1644 * it races with cpu hotunplug operation. Verify
1645 * against POOL_DISASSOCIATED.
1647 if (!(pool->flags & POOL_DISASSOCIATED))
1648 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1650 spin_lock_irq(&pool->lock);
1651 if (pool->flags & POOL_DISASSOCIATED)
1653 if (task_cpu(current) == pool->cpu &&
1654 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1656 spin_unlock_irq(&pool->lock);
1659 * We've raced with CPU hot[un]plug. Give it a breather
1660 * and retry migration. cond_resched() is required here;
1661 * otherwise, we might deadlock against cpu_stop trying to
1662 * bring down the CPU on non-preemptive kernel.
1669 static struct worker *alloc_worker(void)
1671 struct worker *worker;
1673 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1675 INIT_LIST_HEAD(&worker->entry);
1676 INIT_LIST_HEAD(&worker->scheduled);
1677 /* on creation a worker is in !idle && prep state */
1678 worker->flags = WORKER_PREP;
1684 * create_worker - create a new workqueue worker
1685 * @pool: pool the new worker will belong to
1687 * Create a new worker which is bound to @pool. The returned worker
1688 * can be started by calling start_worker() or destroyed using
1692 * Might sleep. Does GFP_KERNEL allocations.
1695 * Pointer to the newly created worker.
1697 static struct worker *create_worker(struct worker_pool *pool)
1699 struct worker *worker = NULL;
1703 lockdep_assert_held(&pool->manager_mutex);
1706 * ID is needed to determine kthread name. Allocate ID first
1707 * without installing the pointer.
1709 idr_preload(GFP_KERNEL);
1710 spin_lock_irq(&pool->lock);
1712 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1714 spin_unlock_irq(&pool->lock);
1719 worker = alloc_worker();
1723 worker->pool = pool;
1727 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1728 pool->attrs->nice < 0 ? "H" : "");
1730 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1732 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1733 "kworker/%s", id_buf);
1734 if (IS_ERR(worker->task))
1738 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1739 * online CPUs. It'll be re-applied when any of the CPUs come up.
1741 set_user_nice(worker->task, pool->attrs->nice);
1742 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1744 /* prevent userland from meddling with cpumask of workqueue workers */
1745 worker->task->flags |= PF_NO_SETAFFINITY;
1748 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1749 * remains stable across this function. See the comments above the
1750 * flag definition for details.
1752 if (pool->flags & POOL_DISASSOCIATED)
1753 worker->flags |= WORKER_UNBOUND;
1755 /* successful, commit the pointer to idr */
1756 spin_lock_irq(&pool->lock);
1757 idr_replace(&pool->worker_idr, worker, worker->id);
1758 spin_unlock_irq(&pool->lock);
1764 spin_lock_irq(&pool->lock);
1765 idr_remove(&pool->worker_idr, id);
1766 spin_unlock_irq(&pool->lock);
1773 * start_worker - start a newly created worker
1774 * @worker: worker to start
1776 * Make the pool aware of @worker and start it.
1779 * spin_lock_irq(pool->lock).
1781 static void start_worker(struct worker *worker)
1783 worker->flags |= WORKER_STARTED;
1784 worker->pool->nr_workers++;
1785 worker_enter_idle(worker);
1786 wake_up_process(worker->task);
1790 * create_and_start_worker - create and start a worker for a pool
1791 * @pool: the target pool
1793 * Grab the managership of @pool and create and start a new worker for it.
1795 static int create_and_start_worker(struct worker_pool *pool)
1797 struct worker *worker;
1799 mutex_lock(&pool->manager_mutex);
1801 worker = create_worker(pool);
1803 spin_lock_irq(&pool->lock);
1804 start_worker(worker);
1805 spin_unlock_irq(&pool->lock);
1808 mutex_unlock(&pool->manager_mutex);
1810 return worker ? 0 : -ENOMEM;
1814 * destroy_worker - destroy a workqueue worker
1815 * @worker: worker to be destroyed
1817 * Destroy @worker and adjust @pool stats accordingly.
1820 * spin_lock_irq(pool->lock) which is released and regrabbed.
1822 static void destroy_worker(struct worker *worker)
1824 struct worker_pool *pool = worker->pool;
1826 lockdep_assert_held(&pool->manager_mutex);
1827 lockdep_assert_held(&pool->lock);
1829 /* sanity check frenzy */
1830 if (WARN_ON(worker->current_work) ||
1831 WARN_ON(!list_empty(&worker->scheduled)))
1834 if (worker->flags & WORKER_STARTED)
1836 if (worker->flags & WORKER_IDLE)
1840 * Once WORKER_DIE is set, the kworker may destroy itself at any
1841 * point. Pin to ensure the task stays until we're done with it.
1843 get_task_struct(worker->task);
1845 list_del_init(&worker->entry);
1846 worker->flags |= WORKER_DIE;
1848 idr_remove(&pool->worker_idr, worker->id);
1850 spin_unlock_irq(&pool->lock);
1852 kthread_stop(worker->task);
1853 put_task_struct(worker->task);
1856 spin_lock_irq(&pool->lock);
1859 static void idle_worker_timeout(unsigned long __pool)
1861 struct worker_pool *pool = (void *)__pool;
1863 spin_lock_irq(&pool->lock);
1865 if (too_many_workers(pool)) {
1866 struct worker *worker;
1867 unsigned long expires;
1869 /* idle_list is kept in LIFO order, check the last one */
1870 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1871 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1873 if (time_before(jiffies, expires))
1874 mod_timer(&pool->idle_timer, expires);
1876 /* it's been idle for too long, wake up manager */
1877 pool->flags |= POOL_MANAGE_WORKERS;
1878 wake_up_worker(pool);
1882 spin_unlock_irq(&pool->lock);
1885 static void send_mayday(struct work_struct *work)
1887 struct pool_workqueue *pwq = get_work_pwq(work);
1888 struct workqueue_struct *wq = pwq->wq;
1890 lockdep_assert_held(&wq_mayday_lock);
1895 /* mayday mayday mayday */
1896 if (list_empty(&pwq->mayday_node)) {
1898 * If @pwq is for an unbound wq, its base ref may be put at
1899 * any time due to an attribute change. Pin @pwq until the
1900 * rescuer is done with it.
1903 list_add_tail(&pwq->mayday_node, &wq->maydays);
1904 wake_up_process(wq->rescuer->task);
1908 static void pool_mayday_timeout(unsigned long __pool)
1910 struct worker_pool *pool = (void *)__pool;
1911 struct work_struct *work;
1913 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1914 spin_lock(&pool->lock);
1916 if (need_to_create_worker(pool)) {
1918 * We've been trying to create a new worker but
1919 * haven't been successful. We might be hitting an
1920 * allocation deadlock. Send distress signals to
1923 list_for_each_entry(work, &pool->worklist, entry)
1927 spin_unlock(&pool->lock);
1928 spin_unlock_irq(&wq_mayday_lock);
1930 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1934 * maybe_create_worker - create a new worker if necessary
1935 * @pool: pool to create a new worker for
1937 * Create a new worker for @pool if necessary. @pool is guaranteed to
1938 * have at least one idle worker on return from this function. If
1939 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1940 * sent to all rescuers with works scheduled on @pool to resolve
1941 * possible allocation deadlock.
1943 * On return, need_to_create_worker() is guaranteed to be %false and
1944 * may_start_working() %true.
1947 * spin_lock_irq(pool->lock) which may be released and regrabbed
1948 * multiple times. Does GFP_KERNEL allocations. Called only from
1952 * %false if no action was taken and pool->lock stayed locked, %true
1955 static bool maybe_create_worker(struct worker_pool *pool)
1956 __releases(&pool->lock)
1957 __acquires(&pool->lock)
1959 if (!need_to_create_worker(pool))
1962 spin_unlock_irq(&pool->lock);
1964 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1965 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1968 struct worker *worker;
1970 worker = create_worker(pool);
1972 del_timer_sync(&pool->mayday_timer);
1973 spin_lock_irq(&pool->lock);
1974 start_worker(worker);
1975 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1980 if (!need_to_create_worker(pool))
1983 __set_current_state(TASK_INTERRUPTIBLE);
1984 schedule_timeout(CREATE_COOLDOWN);
1986 if (!need_to_create_worker(pool))
1990 del_timer_sync(&pool->mayday_timer);
1991 spin_lock_irq(&pool->lock);
1992 if (need_to_create_worker(pool))
1998 * maybe_destroy_worker - destroy workers which have been idle for a while
1999 * @pool: pool to destroy workers for
2001 * Destroy @pool workers which have been idle for longer than
2002 * IDLE_WORKER_TIMEOUT.
2005 * spin_lock_irq(pool->lock) which may be released and regrabbed
2006 * multiple times. Called only from manager.
2009 * %false if no action was taken and pool->lock stayed locked, %true
2012 static bool maybe_destroy_workers(struct worker_pool *pool)
2016 while (too_many_workers(pool)) {
2017 struct worker *worker;
2018 unsigned long expires;
2020 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2021 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2023 if (time_before(jiffies, expires)) {
2024 mod_timer(&pool->idle_timer, expires);
2028 destroy_worker(worker);
2036 * manage_workers - manage worker pool
2039 * Assume the manager role and manage the worker pool @worker belongs
2040 * to. At any given time, there can be only zero or one manager per
2041 * pool. The exclusion is handled automatically by this function.
2043 * The caller can safely start processing works on false return. On
2044 * true return, it's guaranteed that need_to_create_worker() is false
2045 * and may_start_working() is true.
2048 * spin_lock_irq(pool->lock) which may be released and regrabbed
2049 * multiple times. Does GFP_KERNEL allocations.
2052 * spin_lock_irq(pool->lock) which may be released and regrabbed
2053 * multiple times. Does GFP_KERNEL allocations.
2055 static bool manage_workers(struct worker *worker)
2057 struct worker_pool *pool = worker->pool;
2061 * Managership is governed by two mutexes - manager_arb and
2062 * manager_mutex. manager_arb handles arbitration of manager role.
2063 * Anyone who successfully grabs manager_arb wins the arbitration
2064 * and becomes the manager. mutex_trylock() on pool->manager_arb
2065 * failure while holding pool->lock reliably indicates that someone
2066 * else is managing the pool and the worker which failed trylock
2067 * can proceed to executing work items. This means that anyone
2068 * grabbing manager_arb is responsible for actually performing
2069 * manager duties. If manager_arb is grabbed and released without
2070 * actual management, the pool may stall indefinitely.
2072 * manager_mutex is used for exclusion of actual management
2073 * operations. The holder of manager_mutex can be sure that none
2074 * of management operations, including creation and destruction of
2075 * workers, won't take place until the mutex is released. Because
2076 * manager_mutex doesn't interfere with manager role arbitration,
2077 * it is guaranteed that the pool's management, while may be
2078 * delayed, won't be disturbed by someone else grabbing
2081 if (!mutex_trylock(&pool->manager_arb))
2085 * With manager arbitration won, manager_mutex would be free in
2086 * most cases. trylock first without dropping @pool->lock.
2088 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2089 spin_unlock_irq(&pool->lock);
2090 mutex_lock(&pool->manager_mutex);
2091 spin_lock_irq(&pool->lock);
2095 pool->flags &= ~POOL_MANAGE_WORKERS;
2098 * Destroy and then create so that may_start_working() is true
2101 ret |= maybe_destroy_workers(pool);
2102 ret |= maybe_create_worker(pool);
2104 mutex_unlock(&pool->manager_mutex);
2105 mutex_unlock(&pool->manager_arb);
2110 * process_one_work - process single work
2112 * @work: work to process
2114 * Process @work. This function contains all the logics necessary to
2115 * process a single work including synchronization against and
2116 * interaction with other workers on the same cpu, queueing and
2117 * flushing. As long as context requirement is met, any worker can
2118 * call this function to process a work.
2121 * spin_lock_irq(pool->lock) which is released and regrabbed.
2123 static void process_one_work(struct worker *worker, struct work_struct *work)
2124 __releases(&pool->lock)
2125 __acquires(&pool->lock)
2127 struct pool_workqueue *pwq = get_work_pwq(work);
2128 struct worker_pool *pool = worker->pool;
2129 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2131 struct worker *collision;
2132 #ifdef CONFIG_LOCKDEP
2134 * It is permissible to free the struct work_struct from
2135 * inside the function that is called from it, this we need to
2136 * take into account for lockdep too. To avoid bogus "held
2137 * lock freed" warnings as well as problems when looking into
2138 * work->lockdep_map, make a copy and use that here.
2140 struct lockdep_map lockdep_map;
2142 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2145 * Ensure we're on the correct CPU. DISASSOCIATED test is
2146 * necessary to avoid spurious warnings from rescuers servicing the
2147 * unbound or a disassociated pool.
2149 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2150 !(pool->flags & POOL_DISASSOCIATED) &&
2151 raw_smp_processor_id() != pool->cpu);
2154 * A single work shouldn't be executed concurrently by
2155 * multiple workers on a single cpu. Check whether anyone is
2156 * already processing the work. If so, defer the work to the
2157 * currently executing one.
2159 collision = find_worker_executing_work(pool, work);
2160 if (unlikely(collision)) {
2161 move_linked_works(work, &collision->scheduled, NULL);
2165 /* claim and dequeue */
2166 debug_work_deactivate(work);
2167 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2168 worker->current_work = work;
2169 worker->current_func = work->func;
2170 worker->current_pwq = pwq;
2171 work_color = get_work_color(work);
2173 list_del_init(&work->entry);
2176 * CPU intensive works don't participate in concurrency
2177 * management. They're the scheduler's responsibility.
2179 if (unlikely(cpu_intensive))
2180 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2183 * Unbound pool isn't concurrency managed and work items should be
2184 * executed ASAP. Wake up another worker if necessary.
2186 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2187 wake_up_worker(pool);
2190 * Record the last pool and clear PENDING which should be the last
2191 * update to @work. Also, do this inside @pool->lock so that
2192 * PENDING and queued state changes happen together while IRQ is
2195 set_work_pool_and_clear_pending(work, pool->id);
2197 spin_unlock_irq(&pool->lock);
2199 lock_map_acquire_read(&pwq->wq->lockdep_map);
2200 lock_map_acquire(&lockdep_map);
2201 trace_workqueue_execute_start(work);
2202 worker->current_func(work);
2204 * While we must be careful to not use "work" after this, the trace
2205 * point will only record its address.
2207 trace_workqueue_execute_end(work);
2208 lock_map_release(&lockdep_map);
2209 lock_map_release(&pwq->wq->lockdep_map);
2211 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2212 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2213 " last function: %pf\n",
2214 current->comm, preempt_count(), task_pid_nr(current),
2215 worker->current_func);
2216 debug_show_held_locks(current);
2221 * The following prevents a kworker from hogging CPU on !PREEMPT
2222 * kernels, where a requeueing work item waiting for something to
2223 * happen could deadlock with stop_machine as such work item could
2224 * indefinitely requeue itself while all other CPUs are trapped in
2229 spin_lock_irq(&pool->lock);
2231 /* clear cpu intensive status */
2232 if (unlikely(cpu_intensive))
2233 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2235 /* we're done with it, release */
2236 hash_del(&worker->hentry);
2237 worker->current_work = NULL;
2238 worker->current_func = NULL;
2239 worker->current_pwq = NULL;
2240 worker->desc_valid = false;
2241 pwq_dec_nr_in_flight(pwq, work_color);
2245 * process_scheduled_works - process scheduled works
2248 * Process all scheduled works. Please note that the scheduled list
2249 * may change while processing a work, so this function repeatedly
2250 * fetches a work from the top and executes it.
2253 * spin_lock_irq(pool->lock) which may be released and regrabbed
2256 static void process_scheduled_works(struct worker *worker)
2258 while (!list_empty(&worker->scheduled)) {
2259 struct work_struct *work = list_first_entry(&worker->scheduled,
2260 struct work_struct, entry);
2261 process_one_work(worker, work);
2266 * worker_thread - the worker thread function
2269 * The worker thread function. All workers belong to a worker_pool -
2270 * either a per-cpu one or dynamic unbound one. These workers process all
2271 * work items regardless of their specific target workqueue. The only
2272 * exception is work items which belong to workqueues with a rescuer which
2273 * will be explained in rescuer_thread().
2275 static int worker_thread(void *__worker)
2277 struct worker *worker = __worker;
2278 struct worker_pool *pool = worker->pool;
2280 /* tell the scheduler that this is a workqueue worker */
2281 worker->task->flags |= PF_WQ_WORKER;
2283 spin_lock_irq(&pool->lock);
2285 /* am I supposed to die? */
2286 if (unlikely(worker->flags & WORKER_DIE)) {
2287 spin_unlock_irq(&pool->lock);
2288 WARN_ON_ONCE(!list_empty(&worker->entry));
2289 worker->task->flags &= ~PF_WQ_WORKER;
2293 worker_leave_idle(worker);
2295 /* no more worker necessary? */
2296 if (!need_more_worker(pool))
2299 /* do we need to manage? */
2300 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2304 * ->scheduled list can only be filled while a worker is
2305 * preparing to process a work or actually processing it.
2306 * Make sure nobody diddled with it while I was sleeping.
2308 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2311 * Finish PREP stage. We're guaranteed to have at least one idle
2312 * worker or that someone else has already assumed the manager
2313 * role. This is where @worker starts participating in concurrency
2314 * management if applicable and concurrency management is restored
2315 * after being rebound. See rebind_workers() for details.
2317 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2320 struct work_struct *work =
2321 list_first_entry(&pool->worklist,
2322 struct work_struct, entry);
2324 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2325 /* optimization path, not strictly necessary */
2326 process_one_work(worker, work);
2327 if (unlikely(!list_empty(&worker->scheduled)))
2328 process_scheduled_works(worker);
2330 move_linked_works(work, &worker->scheduled, NULL);
2331 process_scheduled_works(worker);
2333 } while (keep_working(pool));
2335 worker_set_flags(worker, WORKER_PREP, false);
2337 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2341 * pool->lock is held and there's no work to process and no need to
2342 * manage, sleep. Workers are woken up only while holding
2343 * pool->lock or from local cpu, so setting the current state
2344 * before releasing pool->lock is enough to prevent losing any
2347 worker_enter_idle(worker);
2348 __set_current_state(TASK_INTERRUPTIBLE);
2349 spin_unlock_irq(&pool->lock);
2355 * rescuer_thread - the rescuer thread function
2358 * Workqueue rescuer thread function. There's one rescuer for each
2359 * workqueue which has WQ_MEM_RECLAIM set.
2361 * Regular work processing on a pool may block trying to create a new
2362 * worker which uses GFP_KERNEL allocation which has slight chance of
2363 * developing into deadlock if some works currently on the same queue
2364 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2365 * the problem rescuer solves.
2367 * When such condition is possible, the pool summons rescuers of all
2368 * workqueues which have works queued on the pool and let them process
2369 * those works so that forward progress can be guaranteed.
2371 * This should happen rarely.
2373 static int rescuer_thread(void *__rescuer)
2375 struct worker *rescuer = __rescuer;
2376 struct workqueue_struct *wq = rescuer->rescue_wq;
2377 struct list_head *scheduled = &rescuer->scheduled;
2380 set_user_nice(current, RESCUER_NICE_LEVEL);
2383 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2384 * doesn't participate in concurrency management.
2386 rescuer->task->flags |= PF_WQ_WORKER;
2388 set_current_state(TASK_INTERRUPTIBLE);
2391 * By the time the rescuer is requested to stop, the workqueue
2392 * shouldn't have any work pending, but @wq->maydays may still have
2393 * pwq(s) queued. This can happen by non-rescuer workers consuming
2394 * all the work items before the rescuer got to them. Go through
2395 * @wq->maydays processing before acting on should_stop so that the
2396 * list is always empty on exit.
2398 should_stop = kthread_should_stop();
2400 /* see whether any pwq is asking for help */
2401 spin_lock_irq(&wq_mayday_lock);
2403 while (!list_empty(&wq->maydays)) {
2404 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2405 struct pool_workqueue, mayday_node);
2406 struct worker_pool *pool = pwq->pool;
2407 struct work_struct *work, *n;
2409 __set_current_state(TASK_RUNNING);
2410 list_del_init(&pwq->mayday_node);
2412 spin_unlock_irq(&wq_mayday_lock);
2414 /* migrate to the target cpu if possible */
2415 worker_maybe_bind_and_lock(pool);
2416 rescuer->pool = pool;
2419 * Slurp in all works issued via this workqueue and
2422 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2423 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2424 if (get_work_pwq(work) == pwq)
2425 move_linked_works(work, scheduled, &n);
2427 process_scheduled_works(rescuer);
2430 * Put the reference grabbed by send_mayday(). @pool won't
2431 * go away while we're holding its lock.
2436 * Leave this pool. If keep_working() is %true, notify a
2437 * regular worker; otherwise, we end up with 0 concurrency
2438 * and stalling the execution.
2440 if (keep_working(pool))
2441 wake_up_worker(pool);
2443 rescuer->pool = NULL;
2444 spin_unlock(&pool->lock);
2445 spin_lock(&wq_mayday_lock);
2448 spin_unlock_irq(&wq_mayday_lock);
2451 __set_current_state(TASK_RUNNING);
2452 rescuer->task->flags &= ~PF_WQ_WORKER;
2456 /* rescuers should never participate in concurrency management */
2457 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2463 struct work_struct work;
2464 struct completion done;
2467 static void wq_barrier_func(struct work_struct *work)
2469 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2470 complete(&barr->done);
2474 * insert_wq_barrier - insert a barrier work
2475 * @pwq: pwq to insert barrier into
2476 * @barr: wq_barrier to insert
2477 * @target: target work to attach @barr to
2478 * @worker: worker currently executing @target, NULL if @target is not executing
2480 * @barr is linked to @target such that @barr is completed only after
2481 * @target finishes execution. Please note that the ordering
2482 * guarantee is observed only with respect to @target and on the local
2485 * Currently, a queued barrier can't be canceled. This is because
2486 * try_to_grab_pending() can't determine whether the work to be
2487 * grabbed is at the head of the queue and thus can't clear LINKED
2488 * flag of the previous work while there must be a valid next work
2489 * after a work with LINKED flag set.
2491 * Note that when @worker is non-NULL, @target may be modified
2492 * underneath us, so we can't reliably determine pwq from @target.
2495 * spin_lock_irq(pool->lock).
2497 static void insert_wq_barrier(struct pool_workqueue *pwq,
2498 struct wq_barrier *barr,
2499 struct work_struct *target, struct worker *worker)
2501 struct list_head *head;
2502 unsigned int linked = 0;
2505 * debugobject calls are safe here even with pool->lock locked
2506 * as we know for sure that this will not trigger any of the
2507 * checks and call back into the fixup functions where we
2510 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2511 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2512 init_completion(&barr->done);
2515 * If @target is currently being executed, schedule the
2516 * barrier to the worker; otherwise, put it after @target.
2519 head = worker->scheduled.next;
2521 unsigned long *bits = work_data_bits(target);
2523 head = target->entry.next;
2524 /* there can already be other linked works, inherit and set */
2525 linked = *bits & WORK_STRUCT_LINKED;
2526 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2529 debug_work_activate(&barr->work);
2530 insert_work(pwq, &barr->work, head,
2531 work_color_to_flags(WORK_NO_COLOR) | linked);
2535 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2536 * @wq: workqueue being flushed
2537 * @flush_color: new flush color, < 0 for no-op
2538 * @work_color: new work color, < 0 for no-op
2540 * Prepare pwqs for workqueue flushing.
2542 * If @flush_color is non-negative, flush_color on all pwqs should be
2543 * -1. If no pwq has in-flight commands at the specified color, all
2544 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2545 * has in flight commands, its pwq->flush_color is set to
2546 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2547 * wakeup logic is armed and %true is returned.
2549 * The caller should have initialized @wq->first_flusher prior to
2550 * calling this function with non-negative @flush_color. If
2551 * @flush_color is negative, no flush color update is done and %false
2554 * If @work_color is non-negative, all pwqs should have the same
2555 * work_color which is previous to @work_color and all will be
2556 * advanced to @work_color.
2559 * mutex_lock(wq->mutex).
2562 * %true if @flush_color >= 0 and there's something to flush. %false
2565 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2566 int flush_color, int work_color)
2569 struct pool_workqueue *pwq;
2571 if (flush_color >= 0) {
2572 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2573 atomic_set(&wq->nr_pwqs_to_flush, 1);
2576 for_each_pwq(pwq, wq) {
2577 struct worker_pool *pool = pwq->pool;
2579 spin_lock_irq(&pool->lock);
2581 if (flush_color >= 0) {
2582 WARN_ON_ONCE(pwq->flush_color != -1);
2584 if (pwq->nr_in_flight[flush_color]) {
2585 pwq->flush_color = flush_color;
2586 atomic_inc(&wq->nr_pwqs_to_flush);
2591 if (work_color >= 0) {
2592 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2593 pwq->work_color = work_color;
2596 spin_unlock_irq(&pool->lock);
2599 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2600 complete(&wq->first_flusher->done);
2606 * flush_workqueue - ensure that any scheduled work has run to completion.
2607 * @wq: workqueue to flush
2609 * This function sleeps until all work items which were queued on entry
2610 * have finished execution, but it is not livelocked by new incoming ones.
2612 void flush_workqueue(struct workqueue_struct *wq)
2614 struct wq_flusher this_flusher = {
2615 .list = LIST_HEAD_INIT(this_flusher.list),
2617 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2621 lock_map_acquire(&wq->lockdep_map);
2622 lock_map_release(&wq->lockdep_map);
2624 mutex_lock(&wq->mutex);
2627 * Start-to-wait phase
2629 next_color = work_next_color(wq->work_color);
2631 if (next_color != wq->flush_color) {
2633 * Color space is not full. The current work_color
2634 * becomes our flush_color and work_color is advanced
2637 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2638 this_flusher.flush_color = wq->work_color;
2639 wq->work_color = next_color;
2641 if (!wq->first_flusher) {
2642 /* no flush in progress, become the first flusher */
2643 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2645 wq->first_flusher = &this_flusher;
2647 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2649 /* nothing to flush, done */
2650 wq->flush_color = next_color;
2651 wq->first_flusher = NULL;
2656 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2657 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2658 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2662 * Oops, color space is full, wait on overflow queue.
2663 * The next flush completion will assign us
2664 * flush_color and transfer to flusher_queue.
2666 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2669 mutex_unlock(&wq->mutex);
2671 wait_for_completion(&this_flusher.done);
2674 * Wake-up-and-cascade phase
2676 * First flushers are responsible for cascading flushes and
2677 * handling overflow. Non-first flushers can simply return.
2679 if (wq->first_flusher != &this_flusher)
2682 mutex_lock(&wq->mutex);
2684 /* we might have raced, check again with mutex held */
2685 if (wq->first_flusher != &this_flusher)
2688 wq->first_flusher = NULL;
2690 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2691 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2694 struct wq_flusher *next, *tmp;
2696 /* complete all the flushers sharing the current flush color */
2697 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2698 if (next->flush_color != wq->flush_color)
2700 list_del_init(&next->list);
2701 complete(&next->done);
2704 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2705 wq->flush_color != work_next_color(wq->work_color));
2707 /* this flush_color is finished, advance by one */
2708 wq->flush_color = work_next_color(wq->flush_color);
2710 /* one color has been freed, handle overflow queue */
2711 if (!list_empty(&wq->flusher_overflow)) {
2713 * Assign the same color to all overflowed
2714 * flushers, advance work_color and append to
2715 * flusher_queue. This is the start-to-wait
2716 * phase for these overflowed flushers.
2718 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2719 tmp->flush_color = wq->work_color;
2721 wq->work_color = work_next_color(wq->work_color);
2723 list_splice_tail_init(&wq->flusher_overflow,
2724 &wq->flusher_queue);
2725 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2728 if (list_empty(&wq->flusher_queue)) {
2729 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2734 * Need to flush more colors. Make the next flusher
2735 * the new first flusher and arm pwqs.
2737 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2738 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2740 list_del_init(&next->list);
2741 wq->first_flusher = next;
2743 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2747 * Meh... this color is already done, clear first
2748 * flusher and repeat cascading.
2750 wq->first_flusher = NULL;
2754 mutex_unlock(&wq->mutex);
2756 EXPORT_SYMBOL_GPL(flush_workqueue);
2759 * drain_workqueue - drain a workqueue
2760 * @wq: workqueue to drain
2762 * Wait until the workqueue becomes empty. While draining is in progress,
2763 * only chain queueing is allowed. IOW, only currently pending or running
2764 * work items on @wq can queue further work items on it. @wq is flushed
2765 * repeatedly until it becomes empty. The number of flushing is detemined
2766 * by the depth of chaining and should be relatively short. Whine if it
2769 void drain_workqueue(struct workqueue_struct *wq)
2771 unsigned int flush_cnt = 0;
2772 struct pool_workqueue *pwq;
2775 * __queue_work() needs to test whether there are drainers, is much
2776 * hotter than drain_workqueue() and already looks at @wq->flags.
2777 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2779 mutex_lock(&wq->mutex);
2780 if (!wq->nr_drainers++)
2781 wq->flags |= __WQ_DRAINING;
2782 mutex_unlock(&wq->mutex);
2784 flush_workqueue(wq);
2786 mutex_lock(&wq->mutex);
2788 for_each_pwq(pwq, wq) {
2791 spin_lock_irq(&pwq->pool->lock);
2792 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2793 spin_unlock_irq(&pwq->pool->lock);
2798 if (++flush_cnt == 10 ||
2799 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2800 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2801 wq->name, flush_cnt);
2803 mutex_unlock(&wq->mutex);
2807 if (!--wq->nr_drainers)
2808 wq->flags &= ~__WQ_DRAINING;
2809 mutex_unlock(&wq->mutex);
2811 EXPORT_SYMBOL_GPL(drain_workqueue);
2813 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2815 struct worker *worker = NULL;
2816 struct worker_pool *pool;
2817 struct pool_workqueue *pwq;
2821 local_irq_disable();
2822 pool = get_work_pool(work);
2828 spin_lock(&pool->lock);
2829 /* see the comment in try_to_grab_pending() with the same code */
2830 pwq = get_work_pwq(work);
2832 if (unlikely(pwq->pool != pool))
2835 worker = find_worker_executing_work(pool, work);
2838 pwq = worker->current_pwq;
2841 insert_wq_barrier(pwq, barr, work, worker);
2842 spin_unlock_irq(&pool->lock);
2845 * If @max_active is 1 or rescuer is in use, flushing another work
2846 * item on the same workqueue may lead to deadlock. Make sure the
2847 * flusher is not running on the same workqueue by verifying write
2850 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2851 lock_map_acquire(&pwq->wq->lockdep_map);
2853 lock_map_acquire_read(&pwq->wq->lockdep_map);
2854 lock_map_release(&pwq->wq->lockdep_map);
2858 spin_unlock_irq(&pool->lock);
2863 * flush_work - wait for a work to finish executing the last queueing instance
2864 * @work: the work to flush
2866 * Wait until @work has finished execution. @work is guaranteed to be idle
2867 * on return if it hasn't been requeued since flush started.
2870 * %true if flush_work() waited for the work to finish execution,
2871 * %false if it was already idle.
2873 bool flush_work(struct work_struct *work)
2875 struct wq_barrier barr;
2877 lock_map_acquire(&work->lockdep_map);
2878 lock_map_release(&work->lockdep_map);
2880 if (start_flush_work(work, &barr)) {
2881 wait_for_completion(&barr.done);
2882 destroy_work_on_stack(&barr.work);
2888 EXPORT_SYMBOL_GPL(flush_work);
2890 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2892 unsigned long flags;
2896 ret = try_to_grab_pending(work, is_dwork, &flags);
2898 * If someone else is canceling, wait for the same event it
2899 * would be waiting for before retrying.
2901 if (unlikely(ret == -ENOENT))
2903 } while (unlikely(ret < 0));
2905 /* tell other tasks trying to grab @work to back off */
2906 mark_work_canceling(work);
2907 local_irq_restore(flags);
2910 clear_work_data(work);
2915 * cancel_work_sync - cancel a work and wait for it to finish
2916 * @work: the work to cancel
2918 * Cancel @work and wait for its execution to finish. This function
2919 * can be used even if the work re-queues itself or migrates to
2920 * another workqueue. On return from this function, @work is
2921 * guaranteed to be not pending or executing on any CPU.
2923 * cancel_work_sync(&delayed_work->work) must not be used for
2924 * delayed_work's. Use cancel_delayed_work_sync() instead.
2926 * The caller must ensure that the workqueue on which @work was last
2927 * queued can't be destroyed before this function returns.
2930 * %true if @work was pending, %false otherwise.
2932 bool cancel_work_sync(struct work_struct *work)
2934 return __cancel_work_timer(work, false);
2936 EXPORT_SYMBOL_GPL(cancel_work_sync);
2939 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2940 * @dwork: the delayed work to flush
2942 * Delayed timer is cancelled and the pending work is queued for
2943 * immediate execution. Like flush_work(), this function only
2944 * considers the last queueing instance of @dwork.
2947 * %true if flush_work() waited for the work to finish execution,
2948 * %false if it was already idle.
2950 bool flush_delayed_work(struct delayed_work *dwork)
2952 local_irq_disable();
2953 if (del_timer_sync(&dwork->timer))
2954 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2956 return flush_work(&dwork->work);
2958 EXPORT_SYMBOL(flush_delayed_work);
2961 * cancel_delayed_work - cancel a delayed work
2962 * @dwork: delayed_work to cancel
2964 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2965 * and canceled; %false if wasn't pending. Note that the work callback
2966 * function may still be running on return, unless it returns %true and the
2967 * work doesn't re-arm itself. Explicitly flush or use
2968 * cancel_delayed_work_sync() to wait on it.
2970 * This function is safe to call from any context including IRQ handler.
2972 bool cancel_delayed_work(struct delayed_work *dwork)
2974 unsigned long flags;
2978 ret = try_to_grab_pending(&dwork->work, true, &flags);
2979 } while (unlikely(ret == -EAGAIN));
2981 if (unlikely(ret < 0))
2984 set_work_pool_and_clear_pending(&dwork->work,
2985 get_work_pool_id(&dwork->work));
2986 local_irq_restore(flags);
2989 EXPORT_SYMBOL(cancel_delayed_work);
2992 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2993 * @dwork: the delayed work cancel
2995 * This is cancel_work_sync() for delayed works.
2998 * %true if @dwork was pending, %false otherwise.
3000 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3002 return __cancel_work_timer(&dwork->work, true);
3004 EXPORT_SYMBOL(cancel_delayed_work_sync);
3007 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3008 * @func: the function to call
3010 * schedule_on_each_cpu() executes @func on each online CPU using the
3011 * system workqueue and blocks until all CPUs have completed.
3012 * schedule_on_each_cpu() is very slow.
3015 * 0 on success, -errno on failure.
3017 int schedule_on_each_cpu(work_func_t func)
3020 struct work_struct __percpu *works;
3022 works = alloc_percpu(struct work_struct);
3028 for_each_online_cpu(cpu) {
3029 struct work_struct *work = per_cpu_ptr(works, cpu);
3031 INIT_WORK(work, func);
3032 schedule_work_on(cpu, work);
3035 for_each_online_cpu(cpu)
3036 flush_work(per_cpu_ptr(works, cpu));
3044 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3046 * Forces execution of the kernel-global workqueue and blocks until its
3049 * Think twice before calling this function! It's very easy to get into
3050 * trouble if you don't take great care. Either of the following situations
3051 * will lead to deadlock:
3053 * One of the work items currently on the workqueue needs to acquire
3054 * a lock held by your code or its caller.
3056 * Your code is running in the context of a work routine.
3058 * They will be detected by lockdep when they occur, but the first might not
3059 * occur very often. It depends on what work items are on the workqueue and
3060 * what locks they need, which you have no control over.
3062 * In most situations flushing the entire workqueue is overkill; you merely
3063 * need to know that a particular work item isn't queued and isn't running.
3064 * In such cases you should use cancel_delayed_work_sync() or
3065 * cancel_work_sync() instead.
3067 void flush_scheduled_work(void)
3069 flush_workqueue(system_wq);
3071 EXPORT_SYMBOL(flush_scheduled_work);
3074 * execute_in_process_context - reliably execute the routine with user context
3075 * @fn: the function to execute
3076 * @ew: guaranteed storage for the execute work structure (must
3077 * be available when the work executes)
3079 * Executes the function immediately if process context is available,
3080 * otherwise schedules the function for delayed execution.
3082 * Returns: 0 - function was executed
3083 * 1 - function was scheduled for execution
3085 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3087 if (!in_interrupt()) {
3092 INIT_WORK(&ew->work, fn);
3093 schedule_work(&ew->work);
3097 EXPORT_SYMBOL_GPL(execute_in_process_context);
3101 * Workqueues with WQ_SYSFS flag set is visible to userland via
3102 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3103 * following attributes.
3105 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3106 * max_active RW int : maximum number of in-flight work items
3108 * Unbound workqueues have the following extra attributes.
3110 * id RO int : the associated pool ID
3111 * nice RW int : nice value of the workers
3112 * cpumask RW mask : bitmask of allowed CPUs for the workers
3115 struct workqueue_struct *wq;
3119 static struct workqueue_struct *dev_to_wq(struct device *dev)
3121 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3126 static ssize_t wq_per_cpu_show(struct device *dev,
3127 struct device_attribute *attr, char *buf)
3129 struct workqueue_struct *wq = dev_to_wq(dev);
3131 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3134 static ssize_t wq_max_active_show(struct device *dev,
3135 struct device_attribute *attr, char *buf)
3137 struct workqueue_struct *wq = dev_to_wq(dev);
3139 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3142 static ssize_t wq_max_active_store(struct device *dev,
3143 struct device_attribute *attr,
3144 const char *buf, size_t count)
3146 struct workqueue_struct *wq = dev_to_wq(dev);
3149 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3152 workqueue_set_max_active(wq, val);
3156 static struct device_attribute wq_sysfs_attrs[] = {
3157 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3158 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3162 static ssize_t wq_pool_ids_show(struct device *dev,
3163 struct device_attribute *attr, char *buf)
3165 struct workqueue_struct *wq = dev_to_wq(dev);
3166 const char *delim = "";
3167 int node, written = 0;
3169 rcu_read_lock_sched();
3170 for_each_node(node) {
3171 written += scnprintf(buf + written, PAGE_SIZE - written,
3172 "%s%d:%d", delim, node,
3173 unbound_pwq_by_node(wq, node)->pool->id);
3176 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3177 rcu_read_unlock_sched();
3182 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3185 struct workqueue_struct *wq = dev_to_wq(dev);
3188 mutex_lock(&wq->mutex);
3189 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3190 mutex_unlock(&wq->mutex);
3195 /* prepare workqueue_attrs for sysfs store operations */
3196 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3198 struct workqueue_attrs *attrs;
3200 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3204 mutex_lock(&wq->mutex);
3205 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3206 mutex_unlock(&wq->mutex);
3210 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3211 const char *buf, size_t count)
3213 struct workqueue_struct *wq = dev_to_wq(dev);
3214 struct workqueue_attrs *attrs;
3217 attrs = wq_sysfs_prep_attrs(wq);
3221 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3222 attrs->nice >= -20 && attrs->nice <= 19)
3223 ret = apply_workqueue_attrs(wq, attrs);
3227 free_workqueue_attrs(attrs);
3228 return ret ?: count;
3231 static ssize_t wq_cpumask_show(struct device *dev,
3232 struct device_attribute *attr, char *buf)
3234 struct workqueue_struct *wq = dev_to_wq(dev);
3237 mutex_lock(&wq->mutex);
3238 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3239 mutex_unlock(&wq->mutex);
3241 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3245 static ssize_t wq_cpumask_store(struct device *dev,
3246 struct device_attribute *attr,
3247 const char *buf, size_t count)
3249 struct workqueue_struct *wq = dev_to_wq(dev);
3250 struct workqueue_attrs *attrs;
3253 attrs = wq_sysfs_prep_attrs(wq);
3257 ret = cpumask_parse(buf, attrs->cpumask);
3259 ret = apply_workqueue_attrs(wq, attrs);
3261 free_workqueue_attrs(attrs);
3262 return ret ?: count;
3265 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3268 struct workqueue_struct *wq = dev_to_wq(dev);
3271 mutex_lock(&wq->mutex);
3272 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3273 !wq->unbound_attrs->no_numa);
3274 mutex_unlock(&wq->mutex);
3279 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3280 const char *buf, size_t count)
3282 struct workqueue_struct *wq = dev_to_wq(dev);
3283 struct workqueue_attrs *attrs;
3286 attrs = wq_sysfs_prep_attrs(wq);
3291 if (sscanf(buf, "%d", &v) == 1) {
3292 attrs->no_numa = !v;
3293 ret = apply_workqueue_attrs(wq, attrs);
3296 free_workqueue_attrs(attrs);
3297 return ret ?: count;
3300 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3301 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3302 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3303 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3304 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3308 static struct bus_type wq_subsys = {
3309 .name = "workqueue",
3310 .dev_attrs = wq_sysfs_attrs,
3313 static int __init wq_sysfs_init(void)
3315 return subsys_virtual_register(&wq_subsys, NULL);
3317 core_initcall(wq_sysfs_init);
3319 static void wq_device_release(struct device *dev)
3321 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3327 * workqueue_sysfs_register - make a workqueue visible in sysfs
3328 * @wq: the workqueue to register
3330 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3331 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3332 * which is the preferred method.
3334 * Workqueue user should use this function directly iff it wants to apply
3335 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3336 * apply_workqueue_attrs() may race against userland updating the
3339 * Returns 0 on success, -errno on failure.
3341 int workqueue_sysfs_register(struct workqueue_struct *wq)
3343 struct wq_device *wq_dev;
3347 * Adjusting max_active or creating new pwqs by applyting
3348 * attributes breaks ordering guarantee. Disallow exposing ordered
3351 if (WARN_ON(wq->flags & __WQ_ORDERED))
3354 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3359 wq_dev->dev.bus = &wq_subsys;
3360 wq_dev->dev.init_name = wq->name;
3361 wq_dev->dev.release = wq_device_release;
3364 * unbound_attrs are created separately. Suppress uevent until
3365 * everything is ready.
3367 dev_set_uevent_suppress(&wq_dev->dev, true);
3369 ret = device_register(&wq_dev->dev);
3376 if (wq->flags & WQ_UNBOUND) {
3377 struct device_attribute *attr;
3379 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3380 ret = device_create_file(&wq_dev->dev, attr);
3382 device_unregister(&wq_dev->dev);
3389 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3394 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3395 * @wq: the workqueue to unregister
3397 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3399 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3401 struct wq_device *wq_dev = wq->wq_dev;
3407 device_unregister(&wq_dev->dev);
3409 #else /* CONFIG_SYSFS */
3410 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3411 #endif /* CONFIG_SYSFS */
3414 * free_workqueue_attrs - free a workqueue_attrs
3415 * @attrs: workqueue_attrs to free
3417 * Undo alloc_workqueue_attrs().
3419 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3422 free_cpumask_var(attrs->cpumask);
3428 * alloc_workqueue_attrs - allocate a workqueue_attrs
3429 * @gfp_mask: allocation mask to use
3431 * Allocate a new workqueue_attrs, initialize with default settings and
3432 * return it. Returns NULL on failure.
3434 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3436 struct workqueue_attrs *attrs;
3438 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3441 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3444 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3447 free_workqueue_attrs(attrs);
3451 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3452 const struct workqueue_attrs *from)
3454 to->nice = from->nice;
3455 cpumask_copy(to->cpumask, from->cpumask);
3457 * Unlike hash and equality test, this function doesn't ignore
3458 * ->no_numa as it is used for both pool and wq attrs. Instead,
3459 * get_unbound_pool() explicitly clears ->no_numa after copying.
3461 to->no_numa = from->no_numa;
3464 /* hash value of the content of @attr */
3465 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3469 hash = jhash_1word(attrs->nice, hash);
3470 hash = jhash(cpumask_bits(attrs->cpumask),
3471 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3475 /* content equality test */
3476 static bool wqattrs_equal(const struct workqueue_attrs *a,
3477 const struct workqueue_attrs *b)
3479 if (a->nice != b->nice)
3481 if (!cpumask_equal(a->cpumask, b->cpumask))
3487 * init_worker_pool - initialize a newly zalloc'd worker_pool
3488 * @pool: worker_pool to initialize
3490 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3491 * Returns 0 on success, -errno on failure. Even on failure, all fields
3492 * inside @pool proper are initialized and put_unbound_pool() can be called
3493 * on @pool safely to release it.
3495 static int init_worker_pool(struct worker_pool *pool)
3497 spin_lock_init(&pool->lock);
3500 pool->node = NUMA_NO_NODE;
3501 pool->flags |= POOL_DISASSOCIATED;
3502 INIT_LIST_HEAD(&pool->worklist);
3503 INIT_LIST_HEAD(&pool->idle_list);
3504 hash_init(pool->busy_hash);
3506 init_timer_deferrable(&pool->idle_timer);
3507 pool->idle_timer.function = idle_worker_timeout;
3508 pool->idle_timer.data = (unsigned long)pool;
3510 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3511 (unsigned long)pool);
3513 mutex_init(&pool->manager_arb);
3514 mutex_init(&pool->manager_mutex);
3515 idr_init(&pool->worker_idr);
3517 INIT_HLIST_NODE(&pool->hash_node);
3520 /* shouldn't fail above this point */
3521 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3527 static void rcu_free_pool(struct rcu_head *rcu)
3529 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3531 idr_destroy(&pool->worker_idr);
3532 free_workqueue_attrs(pool->attrs);
3537 * put_unbound_pool - put a worker_pool
3538 * @pool: worker_pool to put
3540 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3541 * safe manner. get_unbound_pool() calls this function on its failure path
3542 * and this function should be able to release pools which went through,
3543 * successfully or not, init_worker_pool().
3545 * Should be called with wq_pool_mutex held.
3547 static void put_unbound_pool(struct worker_pool *pool)
3549 struct worker *worker;
3551 lockdep_assert_held(&wq_pool_mutex);
3557 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3558 WARN_ON(!list_empty(&pool->worklist)))
3561 /* release id and unhash */
3563 idr_remove(&worker_pool_idr, pool->id);
3564 hash_del(&pool->hash_node);
3567 * Become the manager and destroy all workers. Grabbing
3568 * manager_arb prevents @pool's workers from blocking on
3571 mutex_lock(&pool->manager_arb);
3572 mutex_lock(&pool->manager_mutex);
3573 spin_lock_irq(&pool->lock);
3575 while ((worker = first_worker(pool)))
3576 destroy_worker(worker);
3577 WARN_ON(pool->nr_workers || pool->nr_idle);
3579 spin_unlock_irq(&pool->lock);
3580 mutex_unlock(&pool->manager_mutex);
3581 mutex_unlock(&pool->manager_arb);
3583 /* shut down the timers */
3584 del_timer_sync(&pool->idle_timer);
3585 del_timer_sync(&pool->mayday_timer);
3587 /* sched-RCU protected to allow dereferences from get_work_pool() */
3588 call_rcu_sched(&pool->rcu, rcu_free_pool);
3592 * get_unbound_pool - get a worker_pool with the specified attributes
3593 * @attrs: the attributes of the worker_pool to get
3595 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3596 * reference count and return it. If there already is a matching
3597 * worker_pool, it will be used; otherwise, this function attempts to
3598 * create a new one. On failure, returns NULL.
3600 * Should be called with wq_pool_mutex held.
3602 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3604 u32 hash = wqattrs_hash(attrs);
3605 struct worker_pool *pool;
3608 lockdep_assert_held(&wq_pool_mutex);
3610 /* do we already have a matching pool? */
3611 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3612 if (wqattrs_equal(pool->attrs, attrs)) {
3618 /* nope, create a new one */
3619 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3620 if (!pool || init_worker_pool(pool) < 0)
3623 if (workqueue_freezing)
3624 pool->flags |= POOL_FREEZING;
3626 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3627 copy_workqueue_attrs(pool->attrs, attrs);
3630 * no_numa isn't a worker_pool attribute, always clear it. See
3631 * 'struct workqueue_attrs' comments for detail.
3633 pool->attrs->no_numa = false;
3635 /* if cpumask is contained inside a NUMA node, we belong to that node */
3636 if (wq_numa_enabled) {
3637 for_each_node(node) {
3638 if (cpumask_subset(pool->attrs->cpumask,
3639 wq_numa_possible_cpumask[node])) {
3646 if (worker_pool_assign_id(pool) < 0)
3649 /* create and start the initial worker */
3650 if (create_and_start_worker(pool) < 0)
3654 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3659 put_unbound_pool(pool);
3663 static void rcu_free_pwq(struct rcu_head *rcu)
3665 kmem_cache_free(pwq_cache,
3666 container_of(rcu, struct pool_workqueue, rcu));
3670 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3671 * and needs to be destroyed.
3673 static void pwq_unbound_release_workfn(struct work_struct *work)
3675 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3676 unbound_release_work);
3677 struct workqueue_struct *wq = pwq->wq;
3678 struct worker_pool *pool = pwq->pool;
3681 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3685 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3686 * necessary on release but do it anyway. It's easier to verify
3687 * and consistent with the linking path.
3689 mutex_lock(&wq->mutex);
3690 list_del_rcu(&pwq->pwqs_node);
3691 is_last = list_empty(&wq->pwqs);
3692 mutex_unlock(&wq->mutex);
3694 mutex_lock(&wq_pool_mutex);
3695 put_unbound_pool(pool);
3696 mutex_unlock(&wq_pool_mutex);
3698 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3701 * If we're the last pwq going away, @wq is already dead and no one
3702 * is gonna access it anymore. Free it.
3705 free_workqueue_attrs(wq->unbound_attrs);
3711 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3712 * @pwq: target pool_workqueue
3714 * If @pwq isn't freezing, set @pwq->max_active to the associated
3715 * workqueue's saved_max_active and activate delayed work items
3716 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3718 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3720 struct workqueue_struct *wq = pwq->wq;
3721 bool freezable = wq->flags & WQ_FREEZABLE;
3723 /* for @wq->saved_max_active */
3724 lockdep_assert_held(&wq->mutex);
3726 /* fast exit for non-freezable wqs */
3727 if (!freezable && pwq->max_active == wq->saved_max_active)
3730 spin_lock_irq(&pwq->pool->lock);
3732 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3733 pwq->max_active = wq->saved_max_active;
3735 while (!list_empty(&pwq->delayed_works) &&
3736 pwq->nr_active < pwq->max_active)
3737 pwq_activate_first_delayed(pwq);
3740 * Need to kick a worker after thawed or an unbound wq's
3741 * max_active is bumped. It's a slow path. Do it always.
3743 wake_up_worker(pwq->pool);
3745 pwq->max_active = 0;
3748 spin_unlock_irq(&pwq->pool->lock);
3751 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3752 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3753 struct worker_pool *pool)
3755 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3757 memset(pwq, 0, sizeof(*pwq));
3761 pwq->flush_color = -1;
3763 INIT_LIST_HEAD(&pwq->delayed_works);
3764 INIT_LIST_HEAD(&pwq->pwqs_node);
3765 INIT_LIST_HEAD(&pwq->mayday_node);
3766 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3769 /* sync @pwq with the current state of its associated wq and link it */
3770 static void link_pwq(struct pool_workqueue *pwq)
3772 struct workqueue_struct *wq = pwq->wq;
3774 lockdep_assert_held(&wq->mutex);
3776 /* may be called multiple times, ignore if already linked */
3777 if (!list_empty(&pwq->pwqs_node))
3781 * Set the matching work_color. This is synchronized with
3782 * wq->mutex to avoid confusing flush_workqueue().
3784 pwq->work_color = wq->work_color;
3786 /* sync max_active to the current setting */
3787 pwq_adjust_max_active(pwq);
3790 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3793 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3794 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3795 const struct workqueue_attrs *attrs)
3797 struct worker_pool *pool;
3798 struct pool_workqueue *pwq;
3800 lockdep_assert_held(&wq_pool_mutex);
3802 pool = get_unbound_pool(attrs);
3806 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3808 put_unbound_pool(pool);
3812 init_pwq(pwq, wq, pool);
3816 /* undo alloc_unbound_pwq(), used only in the error path */
3817 static void free_unbound_pwq(struct pool_workqueue *pwq)
3819 lockdep_assert_held(&wq_pool_mutex);
3822 put_unbound_pool(pwq->pool);
3823 kmem_cache_free(pwq_cache, pwq);
3828 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3829 * @attrs: the wq_attrs of interest
3830 * @node: the target NUMA node
3831 * @cpu_going_down: if >= 0, the CPU to consider as offline
3832 * @cpumask: outarg, the resulting cpumask
3834 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3835 * @cpu_going_down is >= 0, that cpu is considered offline during
3836 * calculation. The result is stored in @cpumask. This function returns
3837 * %true if the resulting @cpumask is different from @attrs->cpumask,
3840 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3841 * enabled and @node has online CPUs requested by @attrs, the returned
3842 * cpumask is the intersection of the possible CPUs of @node and
3845 * The caller is responsible for ensuring that the cpumask of @node stays
3848 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3849 int cpu_going_down, cpumask_t *cpumask)
3851 if (!wq_numa_enabled || attrs->no_numa)
3854 /* does @node have any online CPUs @attrs wants? */
3855 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3856 if (cpu_going_down >= 0)
3857 cpumask_clear_cpu(cpu_going_down, cpumask);
3859 if (cpumask_empty(cpumask))
3862 /* yeap, return possible CPUs in @node that @attrs wants */
3863 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3864 return !cpumask_equal(cpumask, attrs->cpumask);
3867 cpumask_copy(cpumask, attrs->cpumask);
3871 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3872 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3874 struct pool_workqueue *pwq)
3876 struct pool_workqueue *old_pwq;
3878 lockdep_assert_held(&wq->mutex);
3880 /* link_pwq() can handle duplicate calls */
3883 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3884 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3889 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3890 * @wq: the target workqueue
3891 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3893 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3894 * machines, this function maps a separate pwq to each NUMA node with
3895 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3896 * NUMA node it was issued on. Older pwqs are released as in-flight work
3897 * items finish. Note that a work item which repeatedly requeues itself
3898 * back-to-back will stay on its current pwq.
3900 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3903 int apply_workqueue_attrs(struct workqueue_struct *wq,
3904 const struct workqueue_attrs *attrs)
3906 struct workqueue_attrs *new_attrs, *tmp_attrs;
3907 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3910 /* only unbound workqueues can change attributes */
3911 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3914 /* creating multiple pwqs breaks ordering guarantee */
3915 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3918 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3919 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3920 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3921 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3924 /* make a copy of @attrs and sanitize it */
3925 copy_workqueue_attrs(new_attrs, attrs);
3926 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3929 * We may create multiple pwqs with differing cpumasks. Make a
3930 * copy of @new_attrs which will be modified and used to obtain
3933 copy_workqueue_attrs(tmp_attrs, new_attrs);
3936 * CPUs should stay stable across pwq creations and installations.
3937 * Pin CPUs, determine the target cpumask for each node and create
3942 mutex_lock(&wq_pool_mutex);
3945 * If something goes wrong during CPU up/down, we'll fall back to
3946 * the default pwq covering whole @attrs->cpumask. Always create
3947 * it even if we don't use it immediately.
3949 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3953 for_each_node(node) {
3954 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3955 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3960 pwq_tbl[node] = dfl_pwq;
3964 mutex_unlock(&wq_pool_mutex);
3966 /* all pwqs have been created successfully, let's install'em */
3967 mutex_lock(&wq->mutex);
3969 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3971 /* save the previous pwq and install the new one */
3973 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3975 /* @dfl_pwq might not have been used, ensure it's linked */
3977 swap(wq->dfl_pwq, dfl_pwq);
3979 mutex_unlock(&wq->mutex);
3981 /* put the old pwqs */
3983 put_pwq_unlocked(pwq_tbl[node]);
3984 put_pwq_unlocked(dfl_pwq);
3990 free_workqueue_attrs(tmp_attrs);
3991 free_workqueue_attrs(new_attrs);
3996 free_unbound_pwq(dfl_pwq);
3998 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3999 free_unbound_pwq(pwq_tbl[node]);
4000 mutex_unlock(&wq_pool_mutex);
4008 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4009 * @wq: the target workqueue
4010 * @cpu: the CPU coming up or going down
4011 * @online: whether @cpu is coming up or going down
4013 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4014 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4017 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4018 * falls back to @wq->dfl_pwq which may not be optimal but is always
4021 * Note that when the last allowed CPU of a NUMA node goes offline for a
4022 * workqueue with a cpumask spanning multiple nodes, the workers which were
4023 * already executing the work items for the workqueue will lose their CPU
4024 * affinity and may execute on any CPU. This is similar to how per-cpu
4025 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4026 * affinity, it's the user's responsibility to flush the work item from
4029 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4032 int node = cpu_to_node(cpu);
4033 int cpu_off = online ? -1 : cpu;
4034 struct pool_workqueue *old_pwq = NULL, *pwq;
4035 struct workqueue_attrs *target_attrs;
4038 lockdep_assert_held(&wq_pool_mutex);
4040 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4044 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4045 * Let's use a preallocated one. The following buf is protected by
4046 * CPU hotplug exclusion.
4048 target_attrs = wq_update_unbound_numa_attrs_buf;
4049 cpumask = target_attrs->cpumask;
4051 mutex_lock(&wq->mutex);
4052 if (wq->unbound_attrs->no_numa)
4055 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4056 pwq = unbound_pwq_by_node(wq, node);
4059 * Let's determine what needs to be done. If the target cpumask is
4060 * different from wq's, we need to compare it to @pwq's and create
4061 * a new one if they don't match. If the target cpumask equals
4062 * wq's, the default pwq should be used. If @pwq is already the
4063 * default one, nothing to do; otherwise, install the default one.
4065 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4066 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4069 if (pwq == wq->dfl_pwq)
4075 mutex_unlock(&wq->mutex);
4077 /* create a new pwq */
4078 pwq = alloc_unbound_pwq(wq, target_attrs);
4080 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4082 mutex_lock(&wq->mutex);
4087 * Install the new pwq. As this function is called only from CPU
4088 * hotplug callbacks and applying a new attrs is wrapped with
4089 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4092 mutex_lock(&wq->mutex);
4093 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4097 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4098 get_pwq(wq->dfl_pwq);
4099 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4100 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4102 mutex_unlock(&wq->mutex);
4103 put_pwq_unlocked(old_pwq);
4106 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4108 bool highpri = wq->flags & WQ_HIGHPRI;
4111 if (!(wq->flags & WQ_UNBOUND)) {
4112 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4116 for_each_possible_cpu(cpu) {
4117 struct pool_workqueue *pwq =
4118 per_cpu_ptr(wq->cpu_pwqs, cpu);
4119 struct worker_pool *cpu_pools =
4120 per_cpu(cpu_worker_pools, cpu);
4122 init_pwq(pwq, wq, &cpu_pools[highpri]);
4124 mutex_lock(&wq->mutex);
4126 mutex_unlock(&wq->mutex);
4129 } else if (wq->flags & __WQ_ORDERED) {
4130 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4131 /* there should only be single pwq for ordering guarantee */
4132 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4133 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4134 "ordering guarantee broken for workqueue %s\n", wq->name);
4137 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4141 static int wq_clamp_max_active(int max_active, unsigned int flags,
4144 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4146 if (max_active < 1 || max_active > lim)
4147 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4148 max_active, name, 1, lim);
4150 return clamp_val(max_active, 1, lim);
4153 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4156 struct lock_class_key *key,
4157 const char *lock_name, ...)
4159 size_t tbl_size = 0;
4161 struct workqueue_struct *wq;
4162 struct pool_workqueue *pwq;
4164 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4165 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4166 flags |= WQ_UNBOUND;
4168 /* allocate wq and format name */
4169 if (flags & WQ_UNBOUND)
4170 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4172 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4176 if (flags & WQ_UNBOUND) {
4177 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4178 if (!wq->unbound_attrs)
4182 va_start(args, lock_name);
4183 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4186 max_active = max_active ?: WQ_DFL_ACTIVE;
4187 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4191 wq->saved_max_active = max_active;
4192 mutex_init(&wq->mutex);
4193 atomic_set(&wq->nr_pwqs_to_flush, 0);
4194 INIT_LIST_HEAD(&wq->pwqs);
4195 INIT_LIST_HEAD(&wq->flusher_queue);
4196 INIT_LIST_HEAD(&wq->flusher_overflow);
4197 INIT_LIST_HEAD(&wq->maydays);
4199 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4200 INIT_LIST_HEAD(&wq->list);
4202 if (alloc_and_link_pwqs(wq) < 0)
4206 * Workqueues which may be used during memory reclaim should
4207 * have a rescuer to guarantee forward progress.
4209 if (flags & WQ_MEM_RECLAIM) {
4210 struct worker *rescuer;
4212 rescuer = alloc_worker();
4216 rescuer->rescue_wq = wq;
4217 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4219 if (IS_ERR(rescuer->task)) {
4224 wq->rescuer = rescuer;
4225 rescuer->task->flags |= PF_NO_SETAFFINITY;
4226 wake_up_process(rescuer->task);
4229 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4233 * wq_pool_mutex protects global freeze state and workqueues list.
4234 * Grab it, adjust max_active and add the new @wq to workqueues
4237 mutex_lock(&wq_pool_mutex);
4239 mutex_lock(&wq->mutex);
4240 for_each_pwq(pwq, wq)
4241 pwq_adjust_max_active(pwq);
4242 mutex_unlock(&wq->mutex);
4244 list_add(&wq->list, &workqueues);
4246 mutex_unlock(&wq_pool_mutex);
4251 free_workqueue_attrs(wq->unbound_attrs);
4255 destroy_workqueue(wq);
4258 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4261 * destroy_workqueue - safely terminate a workqueue
4262 * @wq: target workqueue
4264 * Safely destroy a workqueue. All work currently pending will be done first.
4266 void destroy_workqueue(struct workqueue_struct *wq)
4268 struct pool_workqueue *pwq;
4271 /* drain it before proceeding with destruction */
4272 drain_workqueue(wq);
4275 mutex_lock(&wq->mutex);
4276 for_each_pwq(pwq, wq) {
4279 for (i = 0; i < WORK_NR_COLORS; i++) {
4280 if (WARN_ON(pwq->nr_in_flight[i])) {
4281 mutex_unlock(&wq->mutex);
4286 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4287 WARN_ON(pwq->nr_active) ||
4288 WARN_ON(!list_empty(&pwq->delayed_works))) {
4289 mutex_unlock(&wq->mutex);
4293 mutex_unlock(&wq->mutex);
4296 * wq list is used to freeze wq, remove from list after
4297 * flushing is complete in case freeze races us.
4299 mutex_lock(&wq_pool_mutex);
4300 list_del_init(&wq->list);
4301 mutex_unlock(&wq_pool_mutex);
4303 workqueue_sysfs_unregister(wq);
4306 kthread_stop(wq->rescuer->task);
4311 if (!(wq->flags & WQ_UNBOUND)) {
4313 * The base ref is never dropped on per-cpu pwqs. Directly
4314 * free the pwqs and wq.
4316 free_percpu(wq->cpu_pwqs);
4320 * We're the sole accessor of @wq at this point. Directly
4321 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4322 * @wq will be freed when the last pwq is released.
4324 for_each_node(node) {
4325 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4326 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4327 put_pwq_unlocked(pwq);
4331 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4332 * put. Don't access it afterwards.
4336 put_pwq_unlocked(pwq);
4339 EXPORT_SYMBOL_GPL(destroy_workqueue);
4342 * workqueue_set_max_active - adjust max_active of a workqueue
4343 * @wq: target workqueue
4344 * @max_active: new max_active value.
4346 * Set max_active of @wq to @max_active.
4349 * Don't call from IRQ context.
4351 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4353 struct pool_workqueue *pwq;
4355 /* disallow meddling with max_active for ordered workqueues */
4356 if (WARN_ON(wq->flags & __WQ_ORDERED))
4359 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4361 mutex_lock(&wq->mutex);
4363 wq->saved_max_active = max_active;
4365 for_each_pwq(pwq, wq)
4366 pwq_adjust_max_active(pwq);
4368 mutex_unlock(&wq->mutex);
4370 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4373 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4375 * Determine whether %current is a workqueue rescuer. Can be used from
4376 * work functions to determine whether it's being run off the rescuer task.
4378 bool current_is_workqueue_rescuer(void)
4380 struct worker *worker = current_wq_worker();
4382 return worker && worker->rescue_wq;
4386 * workqueue_congested - test whether a workqueue is congested
4387 * @cpu: CPU in question
4388 * @wq: target workqueue
4390 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4391 * no synchronization around this function and the test result is
4392 * unreliable and only useful as advisory hints or for debugging.
4394 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4395 * Note that both per-cpu and unbound workqueues may be associated with
4396 * multiple pool_workqueues which have separate congested states. A
4397 * workqueue being congested on one CPU doesn't mean the workqueue is also
4398 * contested on other CPUs / NUMA nodes.
4401 * %true if congested, %false otherwise.
4403 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4405 struct pool_workqueue *pwq;
4408 rcu_read_lock_sched();
4410 if (cpu == WORK_CPU_UNBOUND)
4411 cpu = smp_processor_id();
4413 if (!(wq->flags & WQ_UNBOUND))
4414 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4416 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4418 ret = !list_empty(&pwq->delayed_works);
4419 rcu_read_unlock_sched();
4423 EXPORT_SYMBOL_GPL(workqueue_congested);
4426 * work_busy - test whether a work is currently pending or running
4427 * @work: the work to be tested
4429 * Test whether @work is currently pending or running. There is no
4430 * synchronization around this function and the test result is
4431 * unreliable and only useful as advisory hints or for debugging.
4434 * OR'd bitmask of WORK_BUSY_* bits.
4436 unsigned int work_busy(struct work_struct *work)
4438 struct worker_pool *pool;
4439 unsigned long flags;
4440 unsigned int ret = 0;
4442 if (work_pending(work))
4443 ret |= WORK_BUSY_PENDING;
4445 local_irq_save(flags);
4446 pool = get_work_pool(work);
4448 spin_lock(&pool->lock);
4449 if (find_worker_executing_work(pool, work))
4450 ret |= WORK_BUSY_RUNNING;
4451 spin_unlock(&pool->lock);
4453 local_irq_restore(flags);
4457 EXPORT_SYMBOL_GPL(work_busy);
4460 * set_worker_desc - set description for the current work item
4461 * @fmt: printf-style format string
4462 * @...: arguments for the format string
4464 * This function can be called by a running work function to describe what
4465 * the work item is about. If the worker task gets dumped, this
4466 * information will be printed out together to help debugging. The
4467 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4469 void set_worker_desc(const char *fmt, ...)
4471 struct worker *worker = current_wq_worker();
4475 va_start(args, fmt);
4476 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4478 worker->desc_valid = true;
4483 * print_worker_info - print out worker information and description
4484 * @log_lvl: the log level to use when printing
4485 * @task: target task
4487 * If @task is a worker and currently executing a work item, print out the
4488 * name of the workqueue being serviced and worker description set with
4489 * set_worker_desc() by the currently executing work item.
4491 * This function can be safely called on any task as long as the
4492 * task_struct itself is accessible. While safe, this function isn't
4493 * synchronized and may print out mixups or garbages of limited length.
4495 void print_worker_info(const char *log_lvl, struct task_struct *task)
4497 work_func_t *fn = NULL;
4498 char name[WQ_NAME_LEN] = { };
4499 char desc[WORKER_DESC_LEN] = { };
4500 struct pool_workqueue *pwq = NULL;
4501 struct workqueue_struct *wq = NULL;
4502 bool desc_valid = false;
4503 struct worker *worker;
4505 if (!(task->flags & PF_WQ_WORKER))
4509 * This function is called without any synchronization and @task
4510 * could be in any state. Be careful with dereferences.
4512 worker = probe_kthread_data(task);
4515 * Carefully copy the associated workqueue's workfn and name. Keep
4516 * the original last '\0' in case the original contains garbage.
4518 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4519 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4520 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4521 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4523 /* copy worker description */
4524 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4526 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4528 if (fn || name[0] || desc[0]) {
4529 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4531 pr_cont(" (%s)", desc);
4539 * There are two challenges in supporting CPU hotplug. Firstly, there
4540 * are a lot of assumptions on strong associations among work, pwq and
4541 * pool which make migrating pending and scheduled works very
4542 * difficult to implement without impacting hot paths. Secondly,
4543 * worker pools serve mix of short, long and very long running works making
4544 * blocked draining impractical.
4546 * This is solved by allowing the pools to be disassociated from the CPU
4547 * running as an unbound one and allowing it to be reattached later if the
4548 * cpu comes back online.
4551 static void wq_unbind_fn(struct work_struct *work)
4553 int cpu = smp_processor_id();
4554 struct worker_pool *pool;
4555 struct worker *worker;
4558 for_each_cpu_worker_pool(pool, cpu) {
4559 WARN_ON_ONCE(cpu != smp_processor_id());
4561 mutex_lock(&pool->manager_mutex);
4562 spin_lock_irq(&pool->lock);
4565 * We've blocked all manager operations. Make all workers
4566 * unbound and set DISASSOCIATED. Before this, all workers
4567 * except for the ones which are still executing works from
4568 * before the last CPU down must be on the cpu. After
4569 * this, they may become diasporas.
4571 for_each_pool_worker(worker, wi, pool)
4572 worker->flags |= WORKER_UNBOUND;
4574 pool->flags |= POOL_DISASSOCIATED;
4576 spin_unlock_irq(&pool->lock);
4577 mutex_unlock(&pool->manager_mutex);
4580 * Call schedule() so that we cross rq->lock and thus can
4581 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4582 * This is necessary as scheduler callbacks may be invoked
4588 * Sched callbacks are disabled now. Zap nr_running.
4589 * After this, nr_running stays zero and need_more_worker()
4590 * and keep_working() are always true as long as the
4591 * worklist is not empty. This pool now behaves as an
4592 * unbound (in terms of concurrency management) pool which
4593 * are served by workers tied to the pool.
4595 atomic_set(&pool->nr_running, 0);
4598 * With concurrency management just turned off, a busy
4599 * worker blocking could lead to lengthy stalls. Kick off
4600 * unbound chain execution of currently pending work items.
4602 spin_lock_irq(&pool->lock);
4603 wake_up_worker(pool);
4604 spin_unlock_irq(&pool->lock);
4609 * rebind_workers - rebind all workers of a pool to the associated CPU
4610 * @pool: pool of interest
4612 * @pool->cpu is coming online. Rebind all workers to the CPU.
4614 static void rebind_workers(struct worker_pool *pool)
4616 struct worker *worker;
4619 lockdep_assert_held(&pool->manager_mutex);
4622 * Restore CPU affinity of all workers. As all idle workers should
4623 * be on the run-queue of the associated CPU before any local
4624 * wake-ups for concurrency management happen, restore CPU affinty
4625 * of all workers first and then clear UNBOUND. As we're called
4626 * from CPU_ONLINE, the following shouldn't fail.
4628 for_each_pool_worker(worker, wi, pool)
4629 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4630 pool->attrs->cpumask) < 0);
4632 spin_lock_irq(&pool->lock);
4634 for_each_pool_worker(worker, wi, pool) {
4635 unsigned int worker_flags = worker->flags;
4638 * A bound idle worker should actually be on the runqueue
4639 * of the associated CPU for local wake-ups targeting it to
4640 * work. Kick all idle workers so that they migrate to the
4641 * associated CPU. Doing this in the same loop as
4642 * replacing UNBOUND with REBOUND is safe as no worker will
4643 * be bound before @pool->lock is released.
4645 if (worker_flags & WORKER_IDLE)
4646 wake_up_process(worker->task);
4649 * We want to clear UNBOUND but can't directly call
4650 * worker_clr_flags() or adjust nr_running. Atomically
4651 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4652 * @worker will clear REBOUND using worker_clr_flags() when
4653 * it initiates the next execution cycle thus restoring
4654 * concurrency management. Note that when or whether
4655 * @worker clears REBOUND doesn't affect correctness.
4657 * ACCESS_ONCE() is necessary because @worker->flags may be
4658 * tested without holding any lock in
4659 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4660 * fail incorrectly leading to premature concurrency
4661 * management operations.
4663 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4664 worker_flags |= WORKER_REBOUND;
4665 worker_flags &= ~WORKER_UNBOUND;
4666 ACCESS_ONCE(worker->flags) = worker_flags;
4669 spin_unlock_irq(&pool->lock);
4673 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4674 * @pool: unbound pool of interest
4675 * @cpu: the CPU which is coming up
4677 * An unbound pool may end up with a cpumask which doesn't have any online
4678 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4679 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4680 * online CPU before, cpus_allowed of all its workers should be restored.
4682 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4684 static cpumask_t cpumask;
4685 struct worker *worker;
4688 lockdep_assert_held(&pool->manager_mutex);
4690 /* is @cpu allowed for @pool? */
4691 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4694 /* is @cpu the only online CPU? */
4695 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4696 if (cpumask_weight(&cpumask) != 1)
4699 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4700 for_each_pool_worker(worker, wi, pool)
4701 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4702 pool->attrs->cpumask) < 0);
4706 * Workqueues should be brought up before normal priority CPU notifiers.
4707 * This will be registered high priority CPU notifier.
4709 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4710 unsigned long action,
4713 int cpu = (unsigned long)hcpu;
4714 struct worker_pool *pool;
4715 struct workqueue_struct *wq;
4718 switch (action & ~CPU_TASKS_FROZEN) {
4719 case CPU_UP_PREPARE:
4720 for_each_cpu_worker_pool(pool, cpu) {
4721 if (pool->nr_workers)
4723 if (create_and_start_worker(pool) < 0)
4728 case CPU_DOWN_FAILED:
4730 mutex_lock(&wq_pool_mutex);
4732 for_each_pool(pool, pi) {
4733 mutex_lock(&pool->manager_mutex);
4735 if (pool->cpu == cpu) {
4736 spin_lock_irq(&pool->lock);
4737 pool->flags &= ~POOL_DISASSOCIATED;
4738 spin_unlock_irq(&pool->lock);
4740 rebind_workers(pool);
4741 } else if (pool->cpu < 0) {
4742 restore_unbound_workers_cpumask(pool, cpu);
4745 mutex_unlock(&pool->manager_mutex);
4748 /* update NUMA affinity of unbound workqueues */
4749 list_for_each_entry(wq, &workqueues, list)
4750 wq_update_unbound_numa(wq, cpu, true);
4752 mutex_unlock(&wq_pool_mutex);
4759 * Workqueues should be brought down after normal priority CPU notifiers.
4760 * This will be registered as low priority CPU notifier.
4762 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4763 unsigned long action,
4766 int cpu = (unsigned long)hcpu;
4767 struct work_struct unbind_work;
4768 struct workqueue_struct *wq;
4770 switch (action & ~CPU_TASKS_FROZEN) {
4771 case CPU_DOWN_PREPARE:
4772 /* unbinding per-cpu workers should happen on the local CPU */
4773 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4774 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4776 /* update NUMA affinity of unbound workqueues */
4777 mutex_lock(&wq_pool_mutex);
4778 list_for_each_entry(wq, &workqueues, list)
4779 wq_update_unbound_numa(wq, cpu, false);
4780 mutex_unlock(&wq_pool_mutex);
4782 /* wait for per-cpu unbinding to finish */
4783 flush_work(&unbind_work);
4791 struct work_for_cpu {
4792 struct work_struct work;
4798 static void work_for_cpu_fn(struct work_struct *work)
4800 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4802 wfc->ret = wfc->fn(wfc->arg);
4806 * work_on_cpu - run a function in user context on a particular cpu
4807 * @cpu: the cpu to run on
4808 * @fn: the function to run
4809 * @arg: the function arg
4811 * This will return the value @fn returns.
4812 * It is up to the caller to ensure that the cpu doesn't go offline.
4813 * The caller must not hold any locks which would prevent @fn from completing.
4815 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4817 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4819 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4820 schedule_work_on(cpu, &wfc.work);
4821 flush_work(&wfc.work);
4824 EXPORT_SYMBOL_GPL(work_on_cpu);
4825 #endif /* CONFIG_SMP */
4827 #ifdef CONFIG_FREEZER
4830 * freeze_workqueues_begin - begin freezing workqueues
4832 * Start freezing workqueues. After this function returns, all freezable
4833 * workqueues will queue new works to their delayed_works list instead of
4837 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4839 void freeze_workqueues_begin(void)
4841 struct worker_pool *pool;
4842 struct workqueue_struct *wq;
4843 struct pool_workqueue *pwq;
4846 mutex_lock(&wq_pool_mutex);
4848 WARN_ON_ONCE(workqueue_freezing);
4849 workqueue_freezing = true;
4852 for_each_pool(pool, pi) {
4853 spin_lock_irq(&pool->lock);
4854 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4855 pool->flags |= POOL_FREEZING;
4856 spin_unlock_irq(&pool->lock);
4859 list_for_each_entry(wq, &workqueues, list) {
4860 mutex_lock(&wq->mutex);
4861 for_each_pwq(pwq, wq)
4862 pwq_adjust_max_active(pwq);
4863 mutex_unlock(&wq->mutex);
4866 mutex_unlock(&wq_pool_mutex);
4870 * freeze_workqueues_busy - are freezable workqueues still busy?
4872 * Check whether freezing is complete. This function must be called
4873 * between freeze_workqueues_begin() and thaw_workqueues().
4876 * Grabs and releases wq_pool_mutex.
4879 * %true if some freezable workqueues are still busy. %false if freezing
4882 bool freeze_workqueues_busy(void)
4885 struct workqueue_struct *wq;
4886 struct pool_workqueue *pwq;
4888 mutex_lock(&wq_pool_mutex);
4890 WARN_ON_ONCE(!workqueue_freezing);
4892 list_for_each_entry(wq, &workqueues, list) {
4893 if (!(wq->flags & WQ_FREEZABLE))
4896 * nr_active is monotonically decreasing. It's safe
4897 * to peek without lock.
4899 rcu_read_lock_sched();
4900 for_each_pwq(pwq, wq) {
4901 WARN_ON_ONCE(pwq->nr_active < 0);
4902 if (pwq->nr_active) {
4904 rcu_read_unlock_sched();
4908 rcu_read_unlock_sched();
4911 mutex_unlock(&wq_pool_mutex);
4916 * thaw_workqueues - thaw workqueues
4918 * Thaw workqueues. Normal queueing is restored and all collected
4919 * frozen works are transferred to their respective pool worklists.
4922 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4924 void thaw_workqueues(void)
4926 struct workqueue_struct *wq;
4927 struct pool_workqueue *pwq;
4928 struct worker_pool *pool;
4931 mutex_lock(&wq_pool_mutex);
4933 if (!workqueue_freezing)
4936 /* clear FREEZING */
4937 for_each_pool(pool, pi) {
4938 spin_lock_irq(&pool->lock);
4939 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4940 pool->flags &= ~POOL_FREEZING;
4941 spin_unlock_irq(&pool->lock);
4944 /* restore max_active and repopulate worklist */
4945 list_for_each_entry(wq, &workqueues, list) {
4946 mutex_lock(&wq->mutex);
4947 for_each_pwq(pwq, wq)
4948 pwq_adjust_max_active(pwq);
4949 mutex_unlock(&wq->mutex);
4952 workqueue_freezing = false;
4954 mutex_unlock(&wq_pool_mutex);
4956 #endif /* CONFIG_FREEZER */
4958 static void __init wq_numa_init(void)
4963 /* determine NUMA pwq table len - highest node id + 1 */
4965 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4967 if (num_possible_nodes() <= 1)
4970 if (wq_disable_numa) {
4971 pr_info("workqueue: NUMA affinity support disabled\n");
4975 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4976 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4979 * We want masks of possible CPUs of each node which isn't readily
4980 * available. Build one from cpu_to_node() which should have been
4981 * fully initialized by now.
4983 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4987 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4988 node_online(node) ? node : NUMA_NO_NODE));
4990 for_each_possible_cpu(cpu) {
4991 node = cpu_to_node(cpu);
4992 if (WARN_ON(node == NUMA_NO_NODE)) {
4993 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4994 /* happens iff arch is bonkers, let's just proceed */
4997 cpumask_set_cpu(cpu, tbl[node]);
5000 wq_numa_possible_cpumask = tbl;
5001 wq_numa_enabled = true;
5004 static int __init init_workqueues(void)
5006 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5009 /* make sure we have enough bits for OFFQ pool ID */
5010 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
5011 WORK_CPU_END * NR_STD_WORKER_POOLS);
5013 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5015 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5017 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5018 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5022 /* initialize CPU pools */
5023 for_each_possible_cpu(cpu) {
5024 struct worker_pool *pool;
5027 for_each_cpu_worker_pool(pool, cpu) {
5028 BUG_ON(init_worker_pool(pool));
5030 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5031 pool->attrs->nice = std_nice[i++];
5032 pool->node = cpu_to_node(cpu);
5035 mutex_lock(&wq_pool_mutex);
5036 BUG_ON(worker_pool_assign_id(pool));
5037 mutex_unlock(&wq_pool_mutex);
5041 /* create the initial worker */
5042 for_each_online_cpu(cpu) {
5043 struct worker_pool *pool;
5045 for_each_cpu_worker_pool(pool, cpu) {
5046 pool->flags &= ~POOL_DISASSOCIATED;
5047 BUG_ON(create_and_start_worker(pool) < 0);
5051 /* create default unbound and ordered wq attrs */
5052 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5053 struct workqueue_attrs *attrs;
5055 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5056 attrs->nice = std_nice[i];
5057 unbound_std_wq_attrs[i] = attrs;
5060 * An ordered wq should have only one pwq as ordering is
5061 * guaranteed by max_active which is enforced by pwqs.
5062 * Turn off NUMA so that dfl_pwq is used for all nodes.
5064 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5065 attrs->nice = std_nice[i];
5066 attrs->no_numa = true;
5067 ordered_wq_attrs[i] = attrs;
5070 system_wq = alloc_workqueue("events", 0, 0);
5071 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5072 system_long_wq = alloc_workqueue("events_long", 0, 0);
5073 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5074 WQ_UNBOUND_MAX_ACTIVE);
5075 system_freezable_wq = alloc_workqueue("events_freezable",
5077 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5078 WQ_POWER_EFFICIENT, 0);
5079 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5080 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5082 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5083 !system_unbound_wq || !system_freezable_wq ||
5084 !system_power_efficient_wq ||
5085 !system_freezable_power_efficient_wq);
5088 early_initcall(init_workqueues);