2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
51 #include "workqueue_internal.h"
57 * A bound pool is either associated or disassociated with its CPU.
58 * While associated (!DISASSOCIATED), all workers are bound to the
59 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * While DISASSOCIATED, the cpu may be offline and all workers have
63 * %WORKER_UNBOUND set and concurrency management disabled, and may
64 * be executing on any CPU. The pool behaves as an unbound one.
66 * Note that DISASSOCIATED should be flipped only while holding
67 * manager_mutex to avoid changing binding state while
68 * create_worker() is in progress.
70 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72 POOL_FREEZING = 1 << 3, /* freeze in progress */
75 WORKER_STARTED = 1 << 0, /* started */
76 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
78 WORKER_PREP = 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give -20.
104 RESCUER_NICE_LEVEL = -20,
105 HIGHPRI_NICE_LEVEL = -20,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * MG: pool->manager_mutex and pool->lock protected. Writes require both
127 * locks. Reads can happen under either lock.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * WQ: wq->mutex protected.
135 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
137 * MD: wq_mayday_lock protected.
140 /* struct worker is defined in workqueue_internal.h */
143 spinlock_t lock; /* the pool lock */
144 int cpu; /* I: the associated cpu */
145 int node; /* I: the associated node ID */
146 int id; /* I: pool ID */
147 unsigned int flags; /* X: flags */
149 struct list_head worklist; /* L: list of pending works */
150 int nr_workers; /* L: total number of workers */
152 /* nr_idle includes the ones off idle_list for rebinding */
153 int nr_idle; /* L: currently idle ones */
155 struct list_head idle_list; /* X: list of idle workers */
156 struct timer_list idle_timer; /* L: worker idle timeout */
157 struct timer_list mayday_timer; /* L: SOS timer for workers */
159 /* a workers is either on busy_hash or idle_list, or the manager */
160 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
161 /* L: hash of busy workers */
163 /* see manage_workers() for details on the two manager mutexes */
164 struct mutex manager_arb; /* manager arbitration */
165 struct mutex manager_mutex; /* manager exclusion */
166 struct idr worker_idr; /* MG: worker IDs and iteration */
168 struct workqueue_attrs *attrs; /* I: worker attributes */
169 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
170 int refcnt; /* PL: refcnt for unbound pools */
173 * The current concurrency level. As it's likely to be accessed
174 * from other CPUs during try_to_wake_up(), put it in a separate
177 atomic_t nr_running ____cacheline_aligned_in_smp;
180 * Destruction of pool is sched-RCU protected to allow dereferences
181 * from get_work_pool().
184 } ____cacheline_aligned_in_smp;
187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 * of work_struct->data are used for flags and the remaining high bits
189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 * number of flag bits.
192 struct pool_workqueue {
193 struct worker_pool *pool; /* I: the associated pool */
194 struct workqueue_struct *wq; /* I: the owning workqueue */
195 int work_color; /* L: current color */
196 int flush_color; /* L: flushing color */
197 int refcnt; /* L: reference count */
198 int nr_in_flight[WORK_NR_COLORS];
199 /* L: nr of in_flight works */
200 int nr_active; /* L: nr of active works */
201 int max_active; /* L: max active works */
202 struct list_head delayed_works; /* L: delayed works */
203 struct list_head pwqs_node; /* WR: node on wq->pwqs */
204 struct list_head mayday_node; /* MD: node on wq->maydays */
207 * Release of unbound pwq is punted to system_wq. See put_pwq()
208 * and pwq_unbound_release_workfn() for details. pool_workqueue
209 * itself is also sched-RCU protected so that the first pwq can be
210 * determined without grabbing wq->mutex.
212 struct work_struct unbound_release_work;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
217 * Structure used to wait for workqueue flush.
220 struct list_head list; /* WQ: list of flushers */
221 int flush_color; /* WQ: flush color waiting for */
222 struct completion done; /* flush completion */
228 * The externally visible workqueue. It relays the issued work items to
229 * the appropriate worker_pool through its pool_workqueues.
231 struct workqueue_struct {
232 struct list_head pwqs; /* WR: all pwqs of this wq */
233 struct list_head list; /* PL: list of all workqueues */
235 struct mutex mutex; /* protects this wq */
236 int work_color; /* WQ: current work color */
237 int flush_color; /* WQ: current flush color */
238 atomic_t nr_pwqs_to_flush; /* flush in progress */
239 struct wq_flusher *first_flusher; /* WQ: first flusher */
240 struct list_head flusher_queue; /* WQ: flush waiters */
241 struct list_head flusher_overflow; /* WQ: flush overflow list */
243 struct list_head maydays; /* MD: pwqs requesting rescue */
244 struct worker *rescuer; /* I: rescue worker */
246 int nr_drainers; /* WQ: drain in progress */
247 int saved_max_active; /* WQ: saved pwq max_active */
249 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
253 struct wq_device *wq_dev; /* I: for sysfs interface */
255 #ifdef CONFIG_LOCKDEP
256 struct lockdep_map lockdep_map;
258 char name[WQ_NAME_LEN]; /* I: workqueue name */
260 /* hot fields used during command issue, aligned to cacheline */
261 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
262 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
263 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
266 static struct kmem_cache *pwq_cache;
268 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
269 static cpumask_var_t *wq_numa_possible_cpumask;
270 /* possible CPUs of each node */
272 static bool wq_disable_numa;
273 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
275 /* see the comment above the definition of WQ_POWER_EFFICIENT */
276 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
277 static bool wq_power_efficient = true;
279 static bool wq_power_efficient;
282 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
284 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
286 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
287 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
289 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
290 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
292 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
293 static bool workqueue_freezing; /* PL: have wqs started freezing? */
295 /* the per-cpu worker pools */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
299 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
301 /* PL: hash of all unbound pools keyed by pool->attrs */
302 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
304 /* I: attributes used when instantiating standard unbound pools on demand */
305 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
307 struct workqueue_struct *system_wq __read_mostly;
308 EXPORT_SYMBOL(system_wq);
309 struct workqueue_struct *system_highpri_wq __read_mostly;
310 EXPORT_SYMBOL_GPL(system_highpri_wq);
311 struct workqueue_struct *system_long_wq __read_mostly;
312 EXPORT_SYMBOL_GPL(system_long_wq);
313 struct workqueue_struct *system_unbound_wq __read_mostly;
314 EXPORT_SYMBOL_GPL(system_unbound_wq);
315 struct workqueue_struct *system_freezable_wq __read_mostly;
316 EXPORT_SYMBOL_GPL(system_freezable_wq);
317 struct workqueue_struct *system_power_efficient_wq __read_mostly;
318 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
319 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
320 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
322 static int worker_thread(void *__worker);
323 static void copy_workqueue_attrs(struct workqueue_attrs *to,
324 const struct workqueue_attrs *from);
326 #define CREATE_TRACE_POINTS
327 #include <trace/events/workqueue.h>
329 #define assert_rcu_or_pool_mutex() \
330 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
331 lockdep_is_held(&wq_pool_mutex), \
332 "sched RCU or wq_pool_mutex should be held")
334 #define assert_rcu_or_wq_mutex(wq) \
335 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
336 lockdep_is_held(&wq->mutex), \
337 "sched RCU or wq->mutex should be held")
339 #ifdef CONFIG_LOCKDEP
340 #define assert_manager_or_pool_lock(pool) \
341 WARN_ONCE(debug_locks && \
342 !lockdep_is_held(&(pool)->manager_mutex) && \
343 !lockdep_is_held(&(pool)->lock), \
344 "pool->manager_mutex or ->lock should be held")
346 #define assert_manager_or_pool_lock(pool) do { } while (0)
349 #define for_each_cpu_worker_pool(pool, cpu) \
350 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
351 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
355 * for_each_pool - iterate through all worker_pools in the system
356 * @pool: iteration cursor
357 * @pi: integer used for iteration
359 * This must be called either with wq_pool_mutex held or sched RCU read
360 * locked. If the pool needs to be used beyond the locking in effect, the
361 * caller is responsible for guaranteeing that the pool stays online.
363 * The if/else clause exists only for the lockdep assertion and can be
366 #define for_each_pool(pool, pi) \
367 idr_for_each_entry(&worker_pool_idr, pool, pi) \
368 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
372 * for_each_pool_worker - iterate through all workers of a worker_pool
373 * @worker: iteration cursor
374 * @wi: integer used for iteration
375 * @pool: worker_pool to iterate workers of
377 * This must be called with either @pool->manager_mutex or ->lock held.
379 * The if/else clause exists only for the lockdep assertion and can be
382 #define for_each_pool_worker(worker, wi, pool) \
383 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
384 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
388 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
389 * @pwq: iteration cursor
390 * @wq: the target workqueue
392 * This must be called either with wq->mutex held or sched RCU read locked.
393 * If the pwq needs to be used beyond the locking in effect, the caller is
394 * responsible for guaranteeing that the pwq stays online.
396 * The if/else clause exists only for the lockdep assertion and can be
399 #define for_each_pwq(pwq, wq) \
400 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
401 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
404 #ifdef CONFIG_DEBUG_OBJECTS_WORK
406 static struct debug_obj_descr work_debug_descr;
408 static void *work_debug_hint(void *addr)
410 return ((struct work_struct *) addr)->func;
414 * fixup_init is called when:
415 * - an active object is initialized
417 static int work_fixup_init(void *addr, enum debug_obj_state state)
419 struct work_struct *work = addr;
422 case ODEBUG_STATE_ACTIVE:
423 cancel_work_sync(work);
424 debug_object_init(work, &work_debug_descr);
432 * fixup_activate is called when:
433 * - an active object is activated
434 * - an unknown object is activated (might be a statically initialized object)
436 static int work_fixup_activate(void *addr, enum debug_obj_state state)
438 struct work_struct *work = addr;
442 case ODEBUG_STATE_NOTAVAILABLE:
444 * This is not really a fixup. The work struct was
445 * statically initialized. We just make sure that it
446 * is tracked in the object tracker.
448 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
449 debug_object_init(work, &work_debug_descr);
450 debug_object_activate(work, &work_debug_descr);
456 case ODEBUG_STATE_ACTIVE:
465 * fixup_free is called when:
466 * - an active object is freed
468 static int work_fixup_free(void *addr, enum debug_obj_state state)
470 struct work_struct *work = addr;
473 case ODEBUG_STATE_ACTIVE:
474 cancel_work_sync(work);
475 debug_object_free(work, &work_debug_descr);
482 static struct debug_obj_descr work_debug_descr = {
483 .name = "work_struct",
484 .debug_hint = work_debug_hint,
485 .fixup_init = work_fixup_init,
486 .fixup_activate = work_fixup_activate,
487 .fixup_free = work_fixup_free,
490 static inline void debug_work_activate(struct work_struct *work)
492 debug_object_activate(work, &work_debug_descr);
495 static inline void debug_work_deactivate(struct work_struct *work)
497 debug_object_deactivate(work, &work_debug_descr);
500 void __init_work(struct work_struct *work, int onstack)
503 debug_object_init_on_stack(work, &work_debug_descr);
505 debug_object_init(work, &work_debug_descr);
507 EXPORT_SYMBOL_GPL(__init_work);
509 void destroy_work_on_stack(struct work_struct *work)
511 debug_object_free(work, &work_debug_descr);
513 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
516 static inline void debug_work_activate(struct work_struct *work) { }
517 static inline void debug_work_deactivate(struct work_struct *work) { }
520 /* allocate ID and assign it to @pool */
521 static int worker_pool_assign_id(struct worker_pool *pool)
525 lockdep_assert_held(&wq_pool_mutex);
527 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
536 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
537 * @wq: the target workqueue
540 * This must be called either with pwq_lock held or sched RCU read locked.
541 * If the pwq needs to be used beyond the locking in effect, the caller is
542 * responsible for guaranteeing that the pwq stays online.
544 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
547 assert_rcu_or_wq_mutex(wq);
548 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
551 static unsigned int work_color_to_flags(int color)
553 return color << WORK_STRUCT_COLOR_SHIFT;
556 static int get_work_color(struct work_struct *work)
558 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
559 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
562 static int work_next_color(int color)
564 return (color + 1) % WORK_NR_COLORS;
568 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
569 * contain the pointer to the queued pwq. Once execution starts, the flag
570 * is cleared and the high bits contain OFFQ flags and pool ID.
572 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
573 * and clear_work_data() can be used to set the pwq, pool or clear
574 * work->data. These functions should only be called while the work is
575 * owned - ie. while the PENDING bit is set.
577 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
578 * corresponding to a work. Pool is available once the work has been
579 * queued anywhere after initialization until it is sync canceled. pwq is
580 * available only while the work item is queued.
582 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
583 * canceled. While being canceled, a work item may have its PENDING set
584 * but stay off timer and worklist for arbitrarily long and nobody should
585 * try to steal the PENDING bit.
587 static inline void set_work_data(struct work_struct *work, unsigned long data,
590 WARN_ON_ONCE(!work_pending(work));
591 atomic_long_set(&work->data, data | flags | work_static(work));
594 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
595 unsigned long extra_flags)
597 set_work_data(work, (unsigned long)pwq,
598 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
601 static void set_work_pool_and_keep_pending(struct work_struct *work,
604 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
605 WORK_STRUCT_PENDING);
608 static void set_work_pool_and_clear_pending(struct work_struct *work,
612 * The following wmb is paired with the implied mb in
613 * test_and_set_bit(PENDING) and ensures all updates to @work made
614 * here are visible to and precede any updates by the next PENDING
618 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
621 static void clear_work_data(struct work_struct *work)
623 smp_wmb(); /* see set_work_pool_and_clear_pending() */
624 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
627 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
629 unsigned long data = atomic_long_read(&work->data);
631 if (data & WORK_STRUCT_PWQ)
632 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
638 * get_work_pool - return the worker_pool a given work was associated with
639 * @work: the work item of interest
641 * Return the worker_pool @work was last associated with. %NULL if none.
643 * Pools are created and destroyed under wq_pool_mutex, and allows read
644 * access under sched-RCU read lock. As such, this function should be
645 * called under wq_pool_mutex or with preemption disabled.
647 * All fields of the returned pool are accessible as long as the above
648 * mentioned locking is in effect. If the returned pool needs to be used
649 * beyond the critical section, the caller is responsible for ensuring the
650 * returned pool is and stays online.
652 static struct worker_pool *get_work_pool(struct work_struct *work)
654 unsigned long data = atomic_long_read(&work->data);
657 assert_rcu_or_pool_mutex();
659 if (data & WORK_STRUCT_PWQ)
660 return ((struct pool_workqueue *)
661 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
663 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
664 if (pool_id == WORK_OFFQ_POOL_NONE)
667 return idr_find(&worker_pool_idr, pool_id);
671 * get_work_pool_id - return the worker pool ID a given work is associated with
672 * @work: the work item of interest
674 * Return the worker_pool ID @work was last associated with.
675 * %WORK_OFFQ_POOL_NONE if none.
677 static int get_work_pool_id(struct work_struct *work)
679 unsigned long data = atomic_long_read(&work->data);
681 if (data & WORK_STRUCT_PWQ)
682 return ((struct pool_workqueue *)
683 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
685 return data >> WORK_OFFQ_POOL_SHIFT;
688 static void mark_work_canceling(struct work_struct *work)
690 unsigned long pool_id = get_work_pool_id(work);
692 pool_id <<= WORK_OFFQ_POOL_SHIFT;
693 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
696 static bool work_is_canceling(struct work_struct *work)
698 unsigned long data = atomic_long_read(&work->data);
700 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
704 * Policy functions. These define the policies on how the global worker
705 * pools are managed. Unless noted otherwise, these functions assume that
706 * they're being called with pool->lock held.
709 static bool __need_more_worker(struct worker_pool *pool)
711 return !atomic_read(&pool->nr_running);
715 * Need to wake up a worker? Called from anything but currently
718 * Note that, because unbound workers never contribute to nr_running, this
719 * function will always return %true for unbound pools as long as the
720 * worklist isn't empty.
722 static bool need_more_worker(struct worker_pool *pool)
724 return !list_empty(&pool->worklist) && __need_more_worker(pool);
727 /* Can I start working? Called from busy but !running workers. */
728 static bool may_start_working(struct worker_pool *pool)
730 return pool->nr_idle;
733 /* Do I need to keep working? Called from currently running workers. */
734 static bool keep_working(struct worker_pool *pool)
736 return !list_empty(&pool->worklist) &&
737 atomic_read(&pool->nr_running) <= 1;
740 /* Do we need a new worker? Called from manager. */
741 static bool need_to_create_worker(struct worker_pool *pool)
743 return need_more_worker(pool) && !may_start_working(pool);
746 /* Do I need to be the manager? */
747 static bool need_to_manage_workers(struct worker_pool *pool)
749 return need_to_create_worker(pool) ||
750 (pool->flags & POOL_MANAGE_WORKERS);
753 /* Do we have too many workers and should some go away? */
754 static bool too_many_workers(struct worker_pool *pool)
756 bool managing = mutex_is_locked(&pool->manager_arb);
757 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
758 int nr_busy = pool->nr_workers - nr_idle;
761 * nr_idle and idle_list may disagree if idle rebinding is in
762 * progress. Never return %true if idle_list is empty.
764 if (list_empty(&pool->idle_list))
767 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
774 /* Return the first worker. Safe with preemption disabled */
775 static struct worker *first_worker(struct worker_pool *pool)
777 if (unlikely(list_empty(&pool->idle_list)))
780 return list_first_entry(&pool->idle_list, struct worker, entry);
784 * wake_up_worker - wake up an idle worker
785 * @pool: worker pool to wake worker from
787 * Wake up the first idle worker of @pool.
790 * spin_lock_irq(pool->lock).
792 static void wake_up_worker(struct worker_pool *pool)
794 struct worker *worker = first_worker(pool);
797 wake_up_process(worker->task);
801 * wq_worker_waking_up - a worker is waking up
802 * @task: task waking up
803 * @cpu: CPU @task is waking up to
805 * This function is called during try_to_wake_up() when a worker is
809 * spin_lock_irq(rq->lock)
811 void wq_worker_waking_up(struct task_struct *task, int cpu)
813 struct worker *worker = kthread_data(task);
815 if (!(worker->flags & WORKER_NOT_RUNNING)) {
816 WARN_ON_ONCE(worker->pool->cpu != cpu);
817 atomic_inc(&worker->pool->nr_running);
822 * wq_worker_sleeping - a worker is going to sleep
823 * @task: task going to sleep
824 * @cpu: CPU in question, must be the current CPU number
826 * This function is called during schedule() when a busy worker is
827 * going to sleep. Worker on the same cpu can be woken up by
828 * returning pointer to its task.
831 * spin_lock_irq(rq->lock)
834 * Worker task on @cpu to wake up, %NULL if none.
836 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
838 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
839 struct worker_pool *pool;
842 * Rescuers, which may not have all the fields set up like normal
843 * workers, also reach here, let's not access anything before
844 * checking NOT_RUNNING.
846 if (worker->flags & WORKER_NOT_RUNNING)
851 /* this can only happen on the local cpu */
852 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
856 * The counterpart of the following dec_and_test, implied mb,
857 * worklist not empty test sequence is in insert_work().
858 * Please read comment there.
860 * NOT_RUNNING is clear. This means that we're bound to and
861 * running on the local cpu w/ rq lock held and preemption
862 * disabled, which in turn means that none else could be
863 * manipulating idle_list, so dereferencing idle_list without pool
866 if (atomic_dec_and_test(&pool->nr_running) &&
867 !list_empty(&pool->worklist))
868 to_wakeup = first_worker(pool);
869 return to_wakeup ? to_wakeup->task : NULL;
873 * worker_set_flags - set worker flags and adjust nr_running accordingly
875 * @flags: flags to set
876 * @wakeup: wakeup an idle worker if necessary
878 * Set @flags in @worker->flags and adjust nr_running accordingly. If
879 * nr_running becomes zero and @wakeup is %true, an idle worker is
883 * spin_lock_irq(pool->lock)
885 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
888 struct worker_pool *pool = worker->pool;
890 WARN_ON_ONCE(worker->task != current);
893 * If transitioning into NOT_RUNNING, adjust nr_running and
894 * wake up an idle worker as necessary if requested by
897 if ((flags & WORKER_NOT_RUNNING) &&
898 !(worker->flags & WORKER_NOT_RUNNING)) {
900 if (atomic_dec_and_test(&pool->nr_running) &&
901 !list_empty(&pool->worklist))
902 wake_up_worker(pool);
904 atomic_dec(&pool->nr_running);
907 worker->flags |= flags;
911 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
913 * @flags: flags to clear
915 * Clear @flags in @worker->flags and adjust nr_running accordingly.
918 * spin_lock_irq(pool->lock)
920 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
922 struct worker_pool *pool = worker->pool;
923 unsigned int oflags = worker->flags;
925 WARN_ON_ONCE(worker->task != current);
927 worker->flags &= ~flags;
930 * If transitioning out of NOT_RUNNING, increment nr_running. Note
931 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
932 * of multiple flags, not a single flag.
934 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
935 if (!(worker->flags & WORKER_NOT_RUNNING))
936 atomic_inc(&pool->nr_running);
940 * find_worker_executing_work - find worker which is executing a work
941 * @pool: pool of interest
942 * @work: work to find worker for
944 * Find a worker which is executing @work on @pool by searching
945 * @pool->busy_hash which is keyed by the address of @work. For a worker
946 * to match, its current execution should match the address of @work and
947 * its work function. This is to avoid unwanted dependency between
948 * unrelated work executions through a work item being recycled while still
951 * This is a bit tricky. A work item may be freed once its execution
952 * starts and nothing prevents the freed area from being recycled for
953 * another work item. If the same work item address ends up being reused
954 * before the original execution finishes, workqueue will identify the
955 * recycled work item as currently executing and make it wait until the
956 * current execution finishes, introducing an unwanted dependency.
958 * This function checks the work item address and work function to avoid
959 * false positives. Note that this isn't complete as one may construct a
960 * work function which can introduce dependency onto itself through a
961 * recycled work item. Well, if somebody wants to shoot oneself in the
962 * foot that badly, there's only so much we can do, and if such deadlock
963 * actually occurs, it should be easy to locate the culprit work function.
966 * spin_lock_irq(pool->lock).
969 * Pointer to worker which is executing @work if found, NULL
972 static struct worker *find_worker_executing_work(struct worker_pool *pool,
973 struct work_struct *work)
975 struct worker *worker;
977 hash_for_each_possible(pool->busy_hash, worker, hentry,
979 if (worker->current_work == work &&
980 worker->current_func == work->func)
987 * move_linked_works - move linked works to a list
988 * @work: start of series of works to be scheduled
989 * @head: target list to append @work to
990 * @nextp: out paramter for nested worklist walking
992 * Schedule linked works starting from @work to @head. Work series to
993 * be scheduled starts at @work and includes any consecutive work with
994 * WORK_STRUCT_LINKED set in its predecessor.
996 * If @nextp is not NULL, it's updated to point to the next work of
997 * the last scheduled work. This allows move_linked_works() to be
998 * nested inside outer list_for_each_entry_safe().
1001 * spin_lock_irq(pool->lock).
1003 static void move_linked_works(struct work_struct *work, struct list_head *head,
1004 struct work_struct **nextp)
1006 struct work_struct *n;
1009 * Linked worklist will always end before the end of the list,
1010 * use NULL for list head.
1012 list_for_each_entry_safe_from(work, n, NULL, entry) {
1013 list_move_tail(&work->entry, head);
1014 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1019 * If we're already inside safe list traversal and have moved
1020 * multiple works to the scheduled queue, the next position
1021 * needs to be updated.
1028 * get_pwq - get an extra reference on the specified pool_workqueue
1029 * @pwq: pool_workqueue to get
1031 * Obtain an extra reference on @pwq. The caller should guarantee that
1032 * @pwq has positive refcnt and be holding the matching pool->lock.
1034 static void get_pwq(struct pool_workqueue *pwq)
1036 lockdep_assert_held(&pwq->pool->lock);
1037 WARN_ON_ONCE(pwq->refcnt <= 0);
1042 * put_pwq - put a pool_workqueue reference
1043 * @pwq: pool_workqueue to put
1045 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1046 * destruction. The caller should be holding the matching pool->lock.
1048 static void put_pwq(struct pool_workqueue *pwq)
1050 lockdep_assert_held(&pwq->pool->lock);
1051 if (likely(--pwq->refcnt))
1053 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1056 * @pwq can't be released under pool->lock, bounce to
1057 * pwq_unbound_release_workfn(). This never recurses on the same
1058 * pool->lock as this path is taken only for unbound workqueues and
1059 * the release work item is scheduled on a per-cpu workqueue. To
1060 * avoid lockdep warning, unbound pool->locks are given lockdep
1061 * subclass of 1 in get_unbound_pool().
1063 schedule_work(&pwq->unbound_release_work);
1067 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1068 * @pwq: pool_workqueue to put (can be %NULL)
1070 * put_pwq() with locking. This function also allows %NULL @pwq.
1072 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1076 * As both pwqs and pools are sched-RCU protected, the
1077 * following lock operations are safe.
1079 spin_lock_irq(&pwq->pool->lock);
1081 spin_unlock_irq(&pwq->pool->lock);
1085 static void pwq_activate_delayed_work(struct work_struct *work)
1087 struct pool_workqueue *pwq = get_work_pwq(work);
1089 trace_workqueue_activate_work(work);
1090 move_linked_works(work, &pwq->pool->worklist, NULL);
1091 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1095 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1097 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1098 struct work_struct, entry);
1100 pwq_activate_delayed_work(work);
1104 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1105 * @pwq: pwq of interest
1106 * @color: color of work which left the queue
1108 * A work either has completed or is removed from pending queue,
1109 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1112 * spin_lock_irq(pool->lock).
1114 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1116 /* uncolored work items don't participate in flushing or nr_active */
1117 if (color == WORK_NO_COLOR)
1120 pwq->nr_in_flight[color]--;
1123 if (!list_empty(&pwq->delayed_works)) {
1124 /* one down, submit a delayed one */
1125 if (pwq->nr_active < pwq->max_active)
1126 pwq_activate_first_delayed(pwq);
1129 /* is flush in progress and are we at the flushing tip? */
1130 if (likely(pwq->flush_color != color))
1133 /* are there still in-flight works? */
1134 if (pwq->nr_in_flight[color])
1137 /* this pwq is done, clear flush_color */
1138 pwq->flush_color = -1;
1141 * If this was the last pwq, wake up the first flusher. It
1142 * will handle the rest.
1144 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1145 complete(&pwq->wq->first_flusher->done);
1151 * try_to_grab_pending - steal work item from worklist and disable irq
1152 * @work: work item to steal
1153 * @is_dwork: @work is a delayed_work
1154 * @flags: place to store irq state
1156 * Try to grab PENDING bit of @work. This function can handle @work in any
1157 * stable state - idle, on timer or on worklist. Return values are
1159 * 1 if @work was pending and we successfully stole PENDING
1160 * 0 if @work was idle and we claimed PENDING
1161 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1162 * -ENOENT if someone else is canceling @work, this state may persist
1163 * for arbitrarily long
1165 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1166 * interrupted while holding PENDING and @work off queue, irq must be
1167 * disabled on entry. This, combined with delayed_work->timer being
1168 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1170 * On successful return, >= 0, irq is disabled and the caller is
1171 * responsible for releasing it using local_irq_restore(*@flags).
1173 * This function is safe to call from any context including IRQ handler.
1175 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1176 unsigned long *flags)
1178 struct worker_pool *pool;
1179 struct pool_workqueue *pwq;
1181 local_irq_save(*flags);
1183 /* try to steal the timer if it exists */
1185 struct delayed_work *dwork = to_delayed_work(work);
1188 * dwork->timer is irqsafe. If del_timer() fails, it's
1189 * guaranteed that the timer is not queued anywhere and not
1190 * running on the local CPU.
1192 if (likely(del_timer(&dwork->timer)))
1196 /* try to claim PENDING the normal way */
1197 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1201 * The queueing is in progress, or it is already queued. Try to
1202 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1204 pool = get_work_pool(work);
1208 spin_lock(&pool->lock);
1210 * work->data is guaranteed to point to pwq only while the work
1211 * item is queued on pwq->wq, and both updating work->data to point
1212 * to pwq on queueing and to pool on dequeueing are done under
1213 * pwq->pool->lock. This in turn guarantees that, if work->data
1214 * points to pwq which is associated with a locked pool, the work
1215 * item is currently queued on that pool.
1217 pwq = get_work_pwq(work);
1218 if (pwq && pwq->pool == pool) {
1219 debug_work_deactivate(work);
1222 * A delayed work item cannot be grabbed directly because
1223 * it might have linked NO_COLOR work items which, if left
1224 * on the delayed_list, will confuse pwq->nr_active
1225 * management later on and cause stall. Make sure the work
1226 * item is activated before grabbing.
1228 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1229 pwq_activate_delayed_work(work);
1231 list_del_init(&work->entry);
1232 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1234 /* work->data points to pwq iff queued, point to pool */
1235 set_work_pool_and_keep_pending(work, pool->id);
1237 spin_unlock(&pool->lock);
1240 spin_unlock(&pool->lock);
1242 local_irq_restore(*flags);
1243 if (work_is_canceling(work))
1250 * insert_work - insert a work into a pool
1251 * @pwq: pwq @work belongs to
1252 * @work: work to insert
1253 * @head: insertion point
1254 * @extra_flags: extra WORK_STRUCT_* flags to set
1256 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1257 * work_struct flags.
1260 * spin_lock_irq(pool->lock).
1262 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1263 struct list_head *head, unsigned int extra_flags)
1265 struct worker_pool *pool = pwq->pool;
1267 /* we own @work, set data and link */
1268 set_work_pwq(work, pwq, extra_flags);
1269 list_add_tail(&work->entry, head);
1273 * Ensure either wq_worker_sleeping() sees the above
1274 * list_add_tail() or we see zero nr_running to avoid workers lying
1275 * around lazily while there are works to be processed.
1279 if (__need_more_worker(pool))
1280 wake_up_worker(pool);
1284 * Test whether @work is being queued from another work executing on the
1287 static bool is_chained_work(struct workqueue_struct *wq)
1289 struct worker *worker;
1291 worker = current_wq_worker();
1293 * Return %true iff I'm a worker execuing a work item on @wq. If
1294 * I'm @worker, it's safe to dereference it without locking.
1296 return worker && worker->current_pwq->wq == wq;
1299 static void __queue_work(int cpu, struct workqueue_struct *wq,
1300 struct work_struct *work)
1302 struct pool_workqueue *pwq;
1303 struct worker_pool *last_pool;
1304 struct list_head *worklist;
1305 unsigned int work_flags;
1306 unsigned int req_cpu = cpu;
1309 * While a work item is PENDING && off queue, a task trying to
1310 * steal the PENDING will busy-loop waiting for it to either get
1311 * queued or lose PENDING. Grabbing PENDING and queueing should
1312 * happen with IRQ disabled.
1314 WARN_ON_ONCE(!irqs_disabled());
1316 debug_work_activate(work);
1318 /* if dying, only works from the same workqueue are allowed */
1319 if (unlikely(wq->flags & __WQ_DRAINING) &&
1320 WARN_ON_ONCE(!is_chained_work(wq)))
1323 if (req_cpu == WORK_CPU_UNBOUND)
1324 cpu = raw_smp_processor_id();
1326 /* pwq which will be used unless @work is executing elsewhere */
1327 if (!(wq->flags & WQ_UNBOUND))
1328 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1330 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1333 * If @work was previously on a different pool, it might still be
1334 * running there, in which case the work needs to be queued on that
1335 * pool to guarantee non-reentrancy.
1337 last_pool = get_work_pool(work);
1338 if (last_pool && last_pool != pwq->pool) {
1339 struct worker *worker;
1341 spin_lock(&last_pool->lock);
1343 worker = find_worker_executing_work(last_pool, work);
1345 if (worker && worker->current_pwq->wq == wq) {
1346 pwq = worker->current_pwq;
1348 /* meh... not running there, queue here */
1349 spin_unlock(&last_pool->lock);
1350 spin_lock(&pwq->pool->lock);
1353 spin_lock(&pwq->pool->lock);
1357 * pwq is determined and locked. For unbound pools, we could have
1358 * raced with pwq release and it could already be dead. If its
1359 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1360 * without another pwq replacing it in the numa_pwq_tbl or while
1361 * work items are executing on it, so the retrying is guaranteed to
1362 * make forward-progress.
1364 if (unlikely(!pwq->refcnt)) {
1365 if (wq->flags & WQ_UNBOUND) {
1366 spin_unlock(&pwq->pool->lock);
1371 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1375 /* pwq determined, queue */
1376 trace_workqueue_queue_work(req_cpu, pwq, work);
1378 if (WARN_ON(!list_empty(&work->entry))) {
1379 spin_unlock(&pwq->pool->lock);
1383 pwq->nr_in_flight[pwq->work_color]++;
1384 work_flags = work_color_to_flags(pwq->work_color);
1386 if (likely(pwq->nr_active < pwq->max_active)) {
1387 trace_workqueue_activate_work(work);
1389 worklist = &pwq->pool->worklist;
1391 work_flags |= WORK_STRUCT_DELAYED;
1392 worklist = &pwq->delayed_works;
1395 insert_work(pwq, work, worklist, work_flags);
1397 spin_unlock(&pwq->pool->lock);
1401 * queue_work_on - queue work on specific cpu
1402 * @cpu: CPU number to execute work on
1403 * @wq: workqueue to use
1404 * @work: work to queue
1406 * Returns %false if @work was already on a queue, %true otherwise.
1408 * We queue the work to a specific CPU, the caller must ensure it
1411 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1412 struct work_struct *work)
1415 unsigned long flags;
1417 local_irq_save(flags);
1419 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1420 __queue_work(cpu, wq, work);
1424 local_irq_restore(flags);
1427 EXPORT_SYMBOL(queue_work_on);
1429 void delayed_work_timer_fn(unsigned long __data)
1431 struct delayed_work *dwork = (struct delayed_work *)__data;
1433 /* should have been called from irqsafe timer with irq already off */
1434 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1436 EXPORT_SYMBOL(delayed_work_timer_fn);
1438 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1439 struct delayed_work *dwork, unsigned long delay)
1441 struct timer_list *timer = &dwork->timer;
1442 struct work_struct *work = &dwork->work;
1444 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1445 timer->data != (unsigned long)dwork);
1446 WARN_ON_ONCE(timer_pending(timer));
1447 WARN_ON_ONCE(!list_empty(&work->entry));
1450 * If @delay is 0, queue @dwork->work immediately. This is for
1451 * both optimization and correctness. The earliest @timer can
1452 * expire is on the closest next tick and delayed_work users depend
1453 * on that there's no such delay when @delay is 0.
1456 __queue_work(cpu, wq, &dwork->work);
1460 timer_stats_timer_set_start_info(&dwork->timer);
1464 timer->expires = jiffies + delay;
1466 if (unlikely(cpu != WORK_CPU_UNBOUND))
1467 add_timer_on(timer, cpu);
1473 * queue_delayed_work_on - queue work on specific CPU after delay
1474 * @cpu: CPU number to execute work on
1475 * @wq: workqueue to use
1476 * @dwork: work to queue
1477 * @delay: number of jiffies to wait before queueing
1479 * Returns %false if @work was already on a queue, %true otherwise. If
1480 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1483 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1484 struct delayed_work *dwork, unsigned long delay)
1486 struct work_struct *work = &dwork->work;
1488 unsigned long flags;
1490 /* read the comment in __queue_work() */
1491 local_irq_save(flags);
1493 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1494 __queue_delayed_work(cpu, wq, dwork, delay);
1498 local_irq_restore(flags);
1501 EXPORT_SYMBOL(queue_delayed_work_on);
1504 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1505 * @cpu: CPU number to execute work on
1506 * @wq: workqueue to use
1507 * @dwork: work to queue
1508 * @delay: number of jiffies to wait before queueing
1510 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1511 * modify @dwork's timer so that it expires after @delay. If @delay is
1512 * zero, @work is guaranteed to be scheduled immediately regardless of its
1515 * Returns %false if @dwork was idle and queued, %true if @dwork was
1516 * pending and its timer was modified.
1518 * This function is safe to call from any context including IRQ handler.
1519 * See try_to_grab_pending() for details.
1521 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1522 struct delayed_work *dwork, unsigned long delay)
1524 unsigned long flags;
1528 ret = try_to_grab_pending(&dwork->work, true, &flags);
1529 } while (unlikely(ret == -EAGAIN));
1531 if (likely(ret >= 0)) {
1532 __queue_delayed_work(cpu, wq, dwork, delay);
1533 local_irq_restore(flags);
1536 /* -ENOENT from try_to_grab_pending() becomes %true */
1539 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1542 * worker_enter_idle - enter idle state
1543 * @worker: worker which is entering idle state
1545 * @worker is entering idle state. Update stats and idle timer if
1549 * spin_lock_irq(pool->lock).
1551 static void worker_enter_idle(struct worker *worker)
1553 struct worker_pool *pool = worker->pool;
1555 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1556 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1557 (worker->hentry.next || worker->hentry.pprev)))
1560 /* can't use worker_set_flags(), also called from start_worker() */
1561 worker->flags |= WORKER_IDLE;
1563 worker->last_active = jiffies;
1565 /* idle_list is LIFO */
1566 list_add(&worker->entry, &pool->idle_list);
1568 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1569 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1572 * Sanity check nr_running. Because wq_unbind_fn() releases
1573 * pool->lock between setting %WORKER_UNBOUND and zapping
1574 * nr_running, the warning may trigger spuriously. Check iff
1575 * unbind is not in progress.
1577 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1578 pool->nr_workers == pool->nr_idle &&
1579 atomic_read(&pool->nr_running));
1583 * worker_leave_idle - leave idle state
1584 * @worker: worker which is leaving idle state
1586 * @worker is leaving idle state. Update stats.
1589 * spin_lock_irq(pool->lock).
1591 static void worker_leave_idle(struct worker *worker)
1593 struct worker_pool *pool = worker->pool;
1595 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1597 worker_clr_flags(worker, WORKER_IDLE);
1599 list_del_init(&worker->entry);
1603 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1604 * @pool: target worker_pool
1606 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1608 * Works which are scheduled while the cpu is online must at least be
1609 * scheduled to a worker which is bound to the cpu so that if they are
1610 * flushed from cpu callbacks while cpu is going down, they are
1611 * guaranteed to execute on the cpu.
1613 * This function is to be used by unbound workers and rescuers to bind
1614 * themselves to the target cpu and may race with cpu going down or
1615 * coming online. kthread_bind() can't be used because it may put the
1616 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1617 * verbatim as it's best effort and blocking and pool may be
1618 * [dis]associated in the meantime.
1620 * This function tries set_cpus_allowed() and locks pool and verifies the
1621 * binding against %POOL_DISASSOCIATED which is set during
1622 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1623 * enters idle state or fetches works without dropping lock, it can
1624 * guarantee the scheduling requirement described in the first paragraph.
1627 * Might sleep. Called without any lock but returns with pool->lock
1631 * %true if the associated pool is online (@worker is successfully
1632 * bound), %false if offline.
1634 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1635 __acquires(&pool->lock)
1639 * The following call may fail, succeed or succeed
1640 * without actually migrating the task to the cpu if
1641 * it races with cpu hotunplug operation. Verify
1642 * against POOL_DISASSOCIATED.
1644 if (!(pool->flags & POOL_DISASSOCIATED))
1645 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1647 spin_lock_irq(&pool->lock);
1648 if (pool->flags & POOL_DISASSOCIATED)
1650 if (task_cpu(current) == pool->cpu &&
1651 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1653 spin_unlock_irq(&pool->lock);
1656 * We've raced with CPU hot[un]plug. Give it a breather
1657 * and retry migration. cond_resched() is required here;
1658 * otherwise, we might deadlock against cpu_stop trying to
1659 * bring down the CPU on non-preemptive kernel.
1666 static struct worker *alloc_worker(void)
1668 struct worker *worker;
1670 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1672 INIT_LIST_HEAD(&worker->entry);
1673 INIT_LIST_HEAD(&worker->scheduled);
1674 /* on creation a worker is in !idle && prep state */
1675 worker->flags = WORKER_PREP;
1681 * create_worker - create a new workqueue worker
1682 * @pool: pool the new worker will belong to
1684 * Create a new worker which is bound to @pool. The returned worker
1685 * can be started by calling start_worker() or destroyed using
1689 * Might sleep. Does GFP_KERNEL allocations.
1692 * Pointer to the newly created worker.
1694 static struct worker *create_worker(struct worker_pool *pool)
1696 struct worker *worker = NULL;
1700 lockdep_assert_held(&pool->manager_mutex);
1703 * ID is needed to determine kthread name. Allocate ID first
1704 * without installing the pointer.
1706 idr_preload(GFP_KERNEL);
1707 spin_lock_irq(&pool->lock);
1709 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1711 spin_unlock_irq(&pool->lock);
1716 worker = alloc_worker();
1720 worker->pool = pool;
1724 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1725 pool->attrs->nice < 0 ? "H" : "");
1727 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1729 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1730 "kworker/%s", id_buf);
1731 if (IS_ERR(worker->task))
1735 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1736 * online CPUs. It'll be re-applied when any of the CPUs come up.
1738 set_user_nice(worker->task, pool->attrs->nice);
1739 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1741 /* prevent userland from meddling with cpumask of workqueue workers */
1742 worker->task->flags |= PF_NO_SETAFFINITY;
1745 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1746 * remains stable across this function. See the comments above the
1747 * flag definition for details.
1749 if (pool->flags & POOL_DISASSOCIATED)
1750 worker->flags |= WORKER_UNBOUND;
1752 /* successful, commit the pointer to idr */
1753 spin_lock_irq(&pool->lock);
1754 idr_replace(&pool->worker_idr, worker, worker->id);
1755 spin_unlock_irq(&pool->lock);
1761 spin_lock_irq(&pool->lock);
1762 idr_remove(&pool->worker_idr, id);
1763 spin_unlock_irq(&pool->lock);
1770 * start_worker - start a newly created worker
1771 * @worker: worker to start
1773 * Make the pool aware of @worker and start it.
1776 * spin_lock_irq(pool->lock).
1778 static void start_worker(struct worker *worker)
1780 worker->flags |= WORKER_STARTED;
1781 worker->pool->nr_workers++;
1782 worker_enter_idle(worker);
1783 wake_up_process(worker->task);
1787 * create_and_start_worker - create and start a worker for a pool
1788 * @pool: the target pool
1790 * Grab the managership of @pool and create and start a new worker for it.
1792 static int create_and_start_worker(struct worker_pool *pool)
1794 struct worker *worker;
1796 mutex_lock(&pool->manager_mutex);
1798 worker = create_worker(pool);
1800 spin_lock_irq(&pool->lock);
1801 start_worker(worker);
1802 spin_unlock_irq(&pool->lock);
1805 mutex_unlock(&pool->manager_mutex);
1807 return worker ? 0 : -ENOMEM;
1811 * destroy_worker - destroy a workqueue worker
1812 * @worker: worker to be destroyed
1814 * Destroy @worker and adjust @pool stats accordingly.
1817 * spin_lock_irq(pool->lock) which is released and regrabbed.
1819 static void destroy_worker(struct worker *worker)
1821 struct worker_pool *pool = worker->pool;
1823 lockdep_assert_held(&pool->manager_mutex);
1824 lockdep_assert_held(&pool->lock);
1826 /* sanity check frenzy */
1827 if (WARN_ON(worker->current_work) ||
1828 WARN_ON(!list_empty(&worker->scheduled)))
1831 if (worker->flags & WORKER_STARTED)
1833 if (worker->flags & WORKER_IDLE)
1836 list_del_init(&worker->entry);
1837 worker->flags |= WORKER_DIE;
1839 idr_remove(&pool->worker_idr, worker->id);
1841 spin_unlock_irq(&pool->lock);
1843 kthread_stop(worker->task);
1846 spin_lock_irq(&pool->lock);
1849 static void idle_worker_timeout(unsigned long __pool)
1851 struct worker_pool *pool = (void *)__pool;
1853 spin_lock_irq(&pool->lock);
1855 if (too_many_workers(pool)) {
1856 struct worker *worker;
1857 unsigned long expires;
1859 /* idle_list is kept in LIFO order, check the last one */
1860 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1861 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1863 if (time_before(jiffies, expires))
1864 mod_timer(&pool->idle_timer, expires);
1866 /* it's been idle for too long, wake up manager */
1867 pool->flags |= POOL_MANAGE_WORKERS;
1868 wake_up_worker(pool);
1872 spin_unlock_irq(&pool->lock);
1875 static void send_mayday(struct work_struct *work)
1877 struct pool_workqueue *pwq = get_work_pwq(work);
1878 struct workqueue_struct *wq = pwq->wq;
1880 lockdep_assert_held(&wq_mayday_lock);
1885 /* mayday mayday mayday */
1886 if (list_empty(&pwq->mayday_node)) {
1887 list_add_tail(&pwq->mayday_node, &wq->maydays);
1888 wake_up_process(wq->rescuer->task);
1892 static void pool_mayday_timeout(unsigned long __pool)
1894 struct worker_pool *pool = (void *)__pool;
1895 struct work_struct *work;
1897 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1898 spin_lock(&pool->lock);
1900 if (need_to_create_worker(pool)) {
1902 * We've been trying to create a new worker but
1903 * haven't been successful. We might be hitting an
1904 * allocation deadlock. Send distress signals to
1907 list_for_each_entry(work, &pool->worklist, entry)
1911 spin_unlock(&pool->lock);
1912 spin_unlock_irq(&wq_mayday_lock);
1914 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1918 * maybe_create_worker - create a new worker if necessary
1919 * @pool: pool to create a new worker for
1921 * Create a new worker for @pool if necessary. @pool is guaranteed to
1922 * have at least one idle worker on return from this function. If
1923 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1924 * sent to all rescuers with works scheduled on @pool to resolve
1925 * possible allocation deadlock.
1927 * On return, need_to_create_worker() is guaranteed to be %false and
1928 * may_start_working() %true.
1931 * spin_lock_irq(pool->lock) which may be released and regrabbed
1932 * multiple times. Does GFP_KERNEL allocations. Called only from
1936 * %false if no action was taken and pool->lock stayed locked, %true
1939 static bool maybe_create_worker(struct worker_pool *pool)
1940 __releases(&pool->lock)
1941 __acquires(&pool->lock)
1943 if (!need_to_create_worker(pool))
1946 spin_unlock_irq(&pool->lock);
1948 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1949 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1952 struct worker *worker;
1954 worker = create_worker(pool);
1956 del_timer_sync(&pool->mayday_timer);
1957 spin_lock_irq(&pool->lock);
1958 start_worker(worker);
1959 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1964 if (!need_to_create_worker(pool))
1967 __set_current_state(TASK_INTERRUPTIBLE);
1968 schedule_timeout(CREATE_COOLDOWN);
1970 if (!need_to_create_worker(pool))
1974 del_timer_sync(&pool->mayday_timer);
1975 spin_lock_irq(&pool->lock);
1976 if (need_to_create_worker(pool))
1982 * maybe_destroy_worker - destroy workers which have been idle for a while
1983 * @pool: pool to destroy workers for
1985 * Destroy @pool workers which have been idle for longer than
1986 * IDLE_WORKER_TIMEOUT.
1989 * spin_lock_irq(pool->lock) which may be released and regrabbed
1990 * multiple times. Called only from manager.
1993 * %false if no action was taken and pool->lock stayed locked, %true
1996 static bool maybe_destroy_workers(struct worker_pool *pool)
2000 while (too_many_workers(pool)) {
2001 struct worker *worker;
2002 unsigned long expires;
2004 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2005 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2007 if (time_before(jiffies, expires)) {
2008 mod_timer(&pool->idle_timer, expires);
2012 destroy_worker(worker);
2020 * manage_workers - manage worker pool
2023 * Assume the manager role and manage the worker pool @worker belongs
2024 * to. At any given time, there can be only zero or one manager per
2025 * pool. The exclusion is handled automatically by this function.
2027 * The caller can safely start processing works on false return. On
2028 * true return, it's guaranteed that need_to_create_worker() is false
2029 * and may_start_working() is true.
2032 * spin_lock_irq(pool->lock) which may be released and regrabbed
2033 * multiple times. Does GFP_KERNEL allocations.
2036 * spin_lock_irq(pool->lock) which may be released and regrabbed
2037 * multiple times. Does GFP_KERNEL allocations.
2039 static bool manage_workers(struct worker *worker)
2041 struct worker_pool *pool = worker->pool;
2045 * Managership is governed by two mutexes - manager_arb and
2046 * manager_mutex. manager_arb handles arbitration of manager role.
2047 * Anyone who successfully grabs manager_arb wins the arbitration
2048 * and becomes the manager. mutex_trylock() on pool->manager_arb
2049 * failure while holding pool->lock reliably indicates that someone
2050 * else is managing the pool and the worker which failed trylock
2051 * can proceed to executing work items. This means that anyone
2052 * grabbing manager_arb is responsible for actually performing
2053 * manager duties. If manager_arb is grabbed and released without
2054 * actual management, the pool may stall indefinitely.
2056 * manager_mutex is used for exclusion of actual management
2057 * operations. The holder of manager_mutex can be sure that none
2058 * of management operations, including creation and destruction of
2059 * workers, won't take place until the mutex is released. Because
2060 * manager_mutex doesn't interfere with manager role arbitration,
2061 * it is guaranteed that the pool's management, while may be
2062 * delayed, won't be disturbed by someone else grabbing
2065 if (!mutex_trylock(&pool->manager_arb))
2069 * With manager arbitration won, manager_mutex would be free in
2070 * most cases. trylock first without dropping @pool->lock.
2072 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2073 spin_unlock_irq(&pool->lock);
2074 mutex_lock(&pool->manager_mutex);
2075 spin_lock_irq(&pool->lock);
2079 pool->flags &= ~POOL_MANAGE_WORKERS;
2082 * Destroy and then create so that may_start_working() is true
2085 ret |= maybe_destroy_workers(pool);
2086 ret |= maybe_create_worker(pool);
2088 mutex_unlock(&pool->manager_mutex);
2089 mutex_unlock(&pool->manager_arb);
2094 * process_one_work - process single work
2096 * @work: work to process
2098 * Process @work. This function contains all the logics necessary to
2099 * process a single work including synchronization against and
2100 * interaction with other workers on the same cpu, queueing and
2101 * flushing. As long as context requirement is met, any worker can
2102 * call this function to process a work.
2105 * spin_lock_irq(pool->lock) which is released and regrabbed.
2107 static void process_one_work(struct worker *worker, struct work_struct *work)
2108 __releases(&pool->lock)
2109 __acquires(&pool->lock)
2111 struct pool_workqueue *pwq = get_work_pwq(work);
2112 struct worker_pool *pool = worker->pool;
2113 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2115 struct worker *collision;
2116 #ifdef CONFIG_LOCKDEP
2118 * It is permissible to free the struct work_struct from
2119 * inside the function that is called from it, this we need to
2120 * take into account for lockdep too. To avoid bogus "held
2121 * lock freed" warnings as well as problems when looking into
2122 * work->lockdep_map, make a copy and use that here.
2124 struct lockdep_map lockdep_map;
2126 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2129 * Ensure we're on the correct CPU. DISASSOCIATED test is
2130 * necessary to avoid spurious warnings from rescuers servicing the
2131 * unbound or a disassociated pool.
2133 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2134 !(pool->flags & POOL_DISASSOCIATED) &&
2135 raw_smp_processor_id() != pool->cpu);
2138 * A single work shouldn't be executed concurrently by
2139 * multiple workers on a single cpu. Check whether anyone is
2140 * already processing the work. If so, defer the work to the
2141 * currently executing one.
2143 collision = find_worker_executing_work(pool, work);
2144 if (unlikely(collision)) {
2145 move_linked_works(work, &collision->scheduled, NULL);
2149 /* claim and dequeue */
2150 debug_work_deactivate(work);
2151 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2152 worker->current_work = work;
2153 worker->current_func = work->func;
2154 worker->current_pwq = pwq;
2155 work_color = get_work_color(work);
2157 list_del_init(&work->entry);
2160 * CPU intensive works don't participate in concurrency
2161 * management. They're the scheduler's responsibility.
2163 if (unlikely(cpu_intensive))
2164 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2167 * Unbound pool isn't concurrency managed and work items should be
2168 * executed ASAP. Wake up another worker if necessary.
2170 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2171 wake_up_worker(pool);
2174 * Record the last pool and clear PENDING which should be the last
2175 * update to @work. Also, do this inside @pool->lock so that
2176 * PENDING and queued state changes happen together while IRQ is
2179 set_work_pool_and_clear_pending(work, pool->id);
2181 spin_unlock_irq(&pool->lock);
2183 lock_map_acquire_read(&pwq->wq->lockdep_map);
2184 lock_map_acquire(&lockdep_map);
2185 trace_workqueue_execute_start(work);
2186 worker->current_func(work);
2188 * While we must be careful to not use "work" after this, the trace
2189 * point will only record its address.
2191 trace_workqueue_execute_end(work);
2192 lock_map_release(&lockdep_map);
2193 lock_map_release(&pwq->wq->lockdep_map);
2195 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2196 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2197 " last function: %pf\n",
2198 current->comm, preempt_count(), task_pid_nr(current),
2199 worker->current_func);
2200 debug_show_held_locks(current);
2204 spin_lock_irq(&pool->lock);
2206 /* clear cpu intensive status */
2207 if (unlikely(cpu_intensive))
2208 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2210 /* we're done with it, release */
2211 hash_del(&worker->hentry);
2212 worker->current_work = NULL;
2213 worker->current_func = NULL;
2214 worker->current_pwq = NULL;
2215 worker->desc_valid = false;
2216 pwq_dec_nr_in_flight(pwq, work_color);
2220 * process_scheduled_works - process scheduled works
2223 * Process all scheduled works. Please note that the scheduled list
2224 * may change while processing a work, so this function repeatedly
2225 * fetches a work from the top and executes it.
2228 * spin_lock_irq(pool->lock) which may be released and regrabbed
2231 static void process_scheduled_works(struct worker *worker)
2233 while (!list_empty(&worker->scheduled)) {
2234 struct work_struct *work = list_first_entry(&worker->scheduled,
2235 struct work_struct, entry);
2236 process_one_work(worker, work);
2241 * worker_thread - the worker thread function
2244 * The worker thread function. All workers belong to a worker_pool -
2245 * either a per-cpu one or dynamic unbound one. These workers process all
2246 * work items regardless of their specific target workqueue. The only
2247 * exception is work items which belong to workqueues with a rescuer which
2248 * will be explained in rescuer_thread().
2250 static int worker_thread(void *__worker)
2252 struct worker *worker = __worker;
2253 struct worker_pool *pool = worker->pool;
2255 /* tell the scheduler that this is a workqueue worker */
2256 worker->task->flags |= PF_WQ_WORKER;
2258 spin_lock_irq(&pool->lock);
2260 /* am I supposed to die? */
2261 if (unlikely(worker->flags & WORKER_DIE)) {
2262 spin_unlock_irq(&pool->lock);
2263 WARN_ON_ONCE(!list_empty(&worker->entry));
2264 worker->task->flags &= ~PF_WQ_WORKER;
2268 worker_leave_idle(worker);
2270 /* no more worker necessary? */
2271 if (!need_more_worker(pool))
2274 /* do we need to manage? */
2275 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2279 * ->scheduled list can only be filled while a worker is
2280 * preparing to process a work or actually processing it.
2281 * Make sure nobody diddled with it while I was sleeping.
2283 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2286 * Finish PREP stage. We're guaranteed to have at least one idle
2287 * worker or that someone else has already assumed the manager
2288 * role. This is where @worker starts participating in concurrency
2289 * management if applicable and concurrency management is restored
2290 * after being rebound. See rebind_workers() for details.
2292 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2295 struct work_struct *work =
2296 list_first_entry(&pool->worklist,
2297 struct work_struct, entry);
2299 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2300 /* optimization path, not strictly necessary */
2301 process_one_work(worker, work);
2302 if (unlikely(!list_empty(&worker->scheduled)))
2303 process_scheduled_works(worker);
2305 move_linked_works(work, &worker->scheduled, NULL);
2306 process_scheduled_works(worker);
2308 } while (keep_working(pool));
2310 worker_set_flags(worker, WORKER_PREP, false);
2312 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2316 * pool->lock is held and there's no work to process and no need to
2317 * manage, sleep. Workers are woken up only while holding
2318 * pool->lock or from local cpu, so setting the current state
2319 * before releasing pool->lock is enough to prevent losing any
2322 worker_enter_idle(worker);
2323 __set_current_state(TASK_INTERRUPTIBLE);
2324 spin_unlock_irq(&pool->lock);
2330 * rescuer_thread - the rescuer thread function
2333 * Workqueue rescuer thread function. There's one rescuer for each
2334 * workqueue which has WQ_MEM_RECLAIM set.
2336 * Regular work processing on a pool may block trying to create a new
2337 * worker which uses GFP_KERNEL allocation which has slight chance of
2338 * developing into deadlock if some works currently on the same queue
2339 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2340 * the problem rescuer solves.
2342 * When such condition is possible, the pool summons rescuers of all
2343 * workqueues which have works queued on the pool and let them process
2344 * those works so that forward progress can be guaranteed.
2346 * This should happen rarely.
2348 static int rescuer_thread(void *__rescuer)
2350 struct worker *rescuer = __rescuer;
2351 struct workqueue_struct *wq = rescuer->rescue_wq;
2352 struct list_head *scheduled = &rescuer->scheduled;
2354 set_user_nice(current, RESCUER_NICE_LEVEL);
2357 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2358 * doesn't participate in concurrency management.
2360 rescuer->task->flags |= PF_WQ_WORKER;
2362 set_current_state(TASK_INTERRUPTIBLE);
2364 if (kthread_should_stop()) {
2365 __set_current_state(TASK_RUNNING);
2366 rescuer->task->flags &= ~PF_WQ_WORKER;
2370 /* see whether any pwq is asking for help */
2371 spin_lock_irq(&wq_mayday_lock);
2373 while (!list_empty(&wq->maydays)) {
2374 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2375 struct pool_workqueue, mayday_node);
2376 struct worker_pool *pool = pwq->pool;
2377 struct work_struct *work, *n;
2379 __set_current_state(TASK_RUNNING);
2380 list_del_init(&pwq->mayday_node);
2382 spin_unlock_irq(&wq_mayday_lock);
2384 /* migrate to the target cpu if possible */
2385 worker_maybe_bind_and_lock(pool);
2386 rescuer->pool = pool;
2389 * Slurp in all works issued via this workqueue and
2392 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2393 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2394 if (get_work_pwq(work) == pwq)
2395 move_linked_works(work, scheduled, &n);
2397 process_scheduled_works(rescuer);
2400 * Leave this pool. If keep_working() is %true, notify a
2401 * regular worker; otherwise, we end up with 0 concurrency
2402 * and stalling the execution.
2404 if (keep_working(pool))
2405 wake_up_worker(pool);
2407 rescuer->pool = NULL;
2408 spin_unlock(&pool->lock);
2409 spin_lock(&wq_mayday_lock);
2412 spin_unlock_irq(&wq_mayday_lock);
2414 /* rescuers should never participate in concurrency management */
2415 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2421 struct work_struct work;
2422 struct completion done;
2425 static void wq_barrier_func(struct work_struct *work)
2427 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2428 complete(&barr->done);
2432 * insert_wq_barrier - insert a barrier work
2433 * @pwq: pwq to insert barrier into
2434 * @barr: wq_barrier to insert
2435 * @target: target work to attach @barr to
2436 * @worker: worker currently executing @target, NULL if @target is not executing
2438 * @barr is linked to @target such that @barr is completed only after
2439 * @target finishes execution. Please note that the ordering
2440 * guarantee is observed only with respect to @target and on the local
2443 * Currently, a queued barrier can't be canceled. This is because
2444 * try_to_grab_pending() can't determine whether the work to be
2445 * grabbed is at the head of the queue and thus can't clear LINKED
2446 * flag of the previous work while there must be a valid next work
2447 * after a work with LINKED flag set.
2449 * Note that when @worker is non-NULL, @target may be modified
2450 * underneath us, so we can't reliably determine pwq from @target.
2453 * spin_lock_irq(pool->lock).
2455 static void insert_wq_barrier(struct pool_workqueue *pwq,
2456 struct wq_barrier *barr,
2457 struct work_struct *target, struct worker *worker)
2459 struct list_head *head;
2460 unsigned int linked = 0;
2463 * debugobject calls are safe here even with pool->lock locked
2464 * as we know for sure that this will not trigger any of the
2465 * checks and call back into the fixup functions where we
2468 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2469 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2470 init_completion(&barr->done);
2473 * If @target is currently being executed, schedule the
2474 * barrier to the worker; otherwise, put it after @target.
2477 head = worker->scheduled.next;
2479 unsigned long *bits = work_data_bits(target);
2481 head = target->entry.next;
2482 /* there can already be other linked works, inherit and set */
2483 linked = *bits & WORK_STRUCT_LINKED;
2484 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2487 debug_work_activate(&barr->work);
2488 insert_work(pwq, &barr->work, head,
2489 work_color_to_flags(WORK_NO_COLOR) | linked);
2493 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2494 * @wq: workqueue being flushed
2495 * @flush_color: new flush color, < 0 for no-op
2496 * @work_color: new work color, < 0 for no-op
2498 * Prepare pwqs for workqueue flushing.
2500 * If @flush_color is non-negative, flush_color on all pwqs should be
2501 * -1. If no pwq has in-flight commands at the specified color, all
2502 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2503 * has in flight commands, its pwq->flush_color is set to
2504 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2505 * wakeup logic is armed and %true is returned.
2507 * The caller should have initialized @wq->first_flusher prior to
2508 * calling this function with non-negative @flush_color. If
2509 * @flush_color is negative, no flush color update is done and %false
2512 * If @work_color is non-negative, all pwqs should have the same
2513 * work_color which is previous to @work_color and all will be
2514 * advanced to @work_color.
2517 * mutex_lock(wq->mutex).
2520 * %true if @flush_color >= 0 and there's something to flush. %false
2523 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2524 int flush_color, int work_color)
2527 struct pool_workqueue *pwq;
2529 if (flush_color >= 0) {
2530 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2531 atomic_set(&wq->nr_pwqs_to_flush, 1);
2534 for_each_pwq(pwq, wq) {
2535 struct worker_pool *pool = pwq->pool;
2537 spin_lock_irq(&pool->lock);
2539 if (flush_color >= 0) {
2540 WARN_ON_ONCE(pwq->flush_color != -1);
2542 if (pwq->nr_in_flight[flush_color]) {
2543 pwq->flush_color = flush_color;
2544 atomic_inc(&wq->nr_pwqs_to_flush);
2549 if (work_color >= 0) {
2550 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2551 pwq->work_color = work_color;
2554 spin_unlock_irq(&pool->lock);
2557 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2558 complete(&wq->first_flusher->done);
2564 * flush_workqueue - ensure that any scheduled work has run to completion.
2565 * @wq: workqueue to flush
2567 * This function sleeps until all work items which were queued on entry
2568 * have finished execution, but it is not livelocked by new incoming ones.
2570 void flush_workqueue(struct workqueue_struct *wq)
2572 struct wq_flusher this_flusher = {
2573 .list = LIST_HEAD_INIT(this_flusher.list),
2575 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2579 lock_map_acquire(&wq->lockdep_map);
2580 lock_map_release(&wq->lockdep_map);
2582 mutex_lock(&wq->mutex);
2585 * Start-to-wait phase
2587 next_color = work_next_color(wq->work_color);
2589 if (next_color != wq->flush_color) {
2591 * Color space is not full. The current work_color
2592 * becomes our flush_color and work_color is advanced
2595 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2596 this_flusher.flush_color = wq->work_color;
2597 wq->work_color = next_color;
2599 if (!wq->first_flusher) {
2600 /* no flush in progress, become the first flusher */
2601 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2603 wq->first_flusher = &this_flusher;
2605 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2607 /* nothing to flush, done */
2608 wq->flush_color = next_color;
2609 wq->first_flusher = NULL;
2614 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2615 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2616 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2620 * Oops, color space is full, wait on overflow queue.
2621 * The next flush completion will assign us
2622 * flush_color and transfer to flusher_queue.
2624 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2627 mutex_unlock(&wq->mutex);
2629 wait_for_completion(&this_flusher.done);
2632 * Wake-up-and-cascade phase
2634 * First flushers are responsible for cascading flushes and
2635 * handling overflow. Non-first flushers can simply return.
2637 if (wq->first_flusher != &this_flusher)
2640 mutex_lock(&wq->mutex);
2642 /* we might have raced, check again with mutex held */
2643 if (wq->first_flusher != &this_flusher)
2646 wq->first_flusher = NULL;
2648 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2649 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2652 struct wq_flusher *next, *tmp;
2654 /* complete all the flushers sharing the current flush color */
2655 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2656 if (next->flush_color != wq->flush_color)
2658 list_del_init(&next->list);
2659 complete(&next->done);
2662 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2663 wq->flush_color != work_next_color(wq->work_color));
2665 /* this flush_color is finished, advance by one */
2666 wq->flush_color = work_next_color(wq->flush_color);
2668 /* one color has been freed, handle overflow queue */
2669 if (!list_empty(&wq->flusher_overflow)) {
2671 * Assign the same color to all overflowed
2672 * flushers, advance work_color and append to
2673 * flusher_queue. This is the start-to-wait
2674 * phase for these overflowed flushers.
2676 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2677 tmp->flush_color = wq->work_color;
2679 wq->work_color = work_next_color(wq->work_color);
2681 list_splice_tail_init(&wq->flusher_overflow,
2682 &wq->flusher_queue);
2683 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2686 if (list_empty(&wq->flusher_queue)) {
2687 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2692 * Need to flush more colors. Make the next flusher
2693 * the new first flusher and arm pwqs.
2695 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2696 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2698 list_del_init(&next->list);
2699 wq->first_flusher = next;
2701 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2705 * Meh... this color is already done, clear first
2706 * flusher and repeat cascading.
2708 wq->first_flusher = NULL;
2712 mutex_unlock(&wq->mutex);
2714 EXPORT_SYMBOL_GPL(flush_workqueue);
2717 * drain_workqueue - drain a workqueue
2718 * @wq: workqueue to drain
2720 * Wait until the workqueue becomes empty. While draining is in progress,
2721 * only chain queueing is allowed. IOW, only currently pending or running
2722 * work items on @wq can queue further work items on it. @wq is flushed
2723 * repeatedly until it becomes empty. The number of flushing is detemined
2724 * by the depth of chaining and should be relatively short. Whine if it
2727 void drain_workqueue(struct workqueue_struct *wq)
2729 unsigned int flush_cnt = 0;
2730 struct pool_workqueue *pwq;
2733 * __queue_work() needs to test whether there are drainers, is much
2734 * hotter than drain_workqueue() and already looks at @wq->flags.
2735 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2737 mutex_lock(&wq->mutex);
2738 if (!wq->nr_drainers++)
2739 wq->flags |= __WQ_DRAINING;
2740 mutex_unlock(&wq->mutex);
2742 flush_workqueue(wq);
2744 mutex_lock(&wq->mutex);
2746 for_each_pwq(pwq, wq) {
2749 spin_lock_irq(&pwq->pool->lock);
2750 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2751 spin_unlock_irq(&pwq->pool->lock);
2756 if (++flush_cnt == 10 ||
2757 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2758 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2759 wq->name, flush_cnt);
2761 mutex_unlock(&wq->mutex);
2765 if (!--wq->nr_drainers)
2766 wq->flags &= ~__WQ_DRAINING;
2767 mutex_unlock(&wq->mutex);
2769 EXPORT_SYMBOL_GPL(drain_workqueue);
2771 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2773 struct worker *worker = NULL;
2774 struct worker_pool *pool;
2775 struct pool_workqueue *pwq;
2779 local_irq_disable();
2780 pool = get_work_pool(work);
2786 spin_lock(&pool->lock);
2787 /* see the comment in try_to_grab_pending() with the same code */
2788 pwq = get_work_pwq(work);
2790 if (unlikely(pwq->pool != pool))
2793 worker = find_worker_executing_work(pool, work);
2796 pwq = worker->current_pwq;
2799 insert_wq_barrier(pwq, barr, work, worker);
2800 spin_unlock_irq(&pool->lock);
2803 * If @max_active is 1 or rescuer is in use, flushing another work
2804 * item on the same workqueue may lead to deadlock. Make sure the
2805 * flusher is not running on the same workqueue by verifying write
2808 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2809 lock_map_acquire(&pwq->wq->lockdep_map);
2811 lock_map_acquire_read(&pwq->wq->lockdep_map);
2812 lock_map_release(&pwq->wq->lockdep_map);
2816 spin_unlock_irq(&pool->lock);
2820 static bool __flush_work(struct work_struct *work)
2822 struct wq_barrier barr;
2824 if (start_flush_work(work, &barr)) {
2825 wait_for_completion(&barr.done);
2826 destroy_work_on_stack(&barr.work);
2834 * flush_work - wait for a work to finish executing the last queueing instance
2835 * @work: the work to flush
2837 * Wait until @work has finished execution. @work is guaranteed to be idle
2838 * on return if it hasn't been requeued since flush started.
2841 * %true if flush_work() waited for the work to finish execution,
2842 * %false if it was already idle.
2844 bool flush_work(struct work_struct *work)
2846 lock_map_acquire(&work->lockdep_map);
2847 lock_map_release(&work->lockdep_map);
2849 return __flush_work(work);
2851 EXPORT_SYMBOL_GPL(flush_work);
2853 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2855 unsigned long flags;
2859 ret = try_to_grab_pending(work, is_dwork, &flags);
2861 * If someone else is canceling, wait for the same event it
2862 * would be waiting for before retrying.
2864 if (unlikely(ret == -ENOENT))
2866 } while (unlikely(ret < 0));
2868 /* tell other tasks trying to grab @work to back off */
2869 mark_work_canceling(work);
2870 local_irq_restore(flags);
2873 clear_work_data(work);
2878 * cancel_work_sync - cancel a work and wait for it to finish
2879 * @work: the work to cancel
2881 * Cancel @work and wait for its execution to finish. This function
2882 * can be used even if the work re-queues itself or migrates to
2883 * another workqueue. On return from this function, @work is
2884 * guaranteed to be not pending or executing on any CPU.
2886 * cancel_work_sync(&delayed_work->work) must not be used for
2887 * delayed_work's. Use cancel_delayed_work_sync() instead.
2889 * The caller must ensure that the workqueue on which @work was last
2890 * queued can't be destroyed before this function returns.
2893 * %true if @work was pending, %false otherwise.
2895 bool cancel_work_sync(struct work_struct *work)
2897 return __cancel_work_timer(work, false);
2899 EXPORT_SYMBOL_GPL(cancel_work_sync);
2902 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2903 * @dwork: the delayed work to flush
2905 * Delayed timer is cancelled and the pending work is queued for
2906 * immediate execution. Like flush_work(), this function only
2907 * considers the last queueing instance of @dwork.
2910 * %true if flush_work() waited for the work to finish execution,
2911 * %false if it was already idle.
2913 bool flush_delayed_work(struct delayed_work *dwork)
2915 local_irq_disable();
2916 if (del_timer_sync(&dwork->timer))
2917 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2919 return flush_work(&dwork->work);
2921 EXPORT_SYMBOL(flush_delayed_work);
2924 * cancel_delayed_work - cancel a delayed work
2925 * @dwork: delayed_work to cancel
2927 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2928 * and canceled; %false if wasn't pending. Note that the work callback
2929 * function may still be running on return, unless it returns %true and the
2930 * work doesn't re-arm itself. Explicitly flush or use
2931 * cancel_delayed_work_sync() to wait on it.
2933 * This function is safe to call from any context including IRQ handler.
2935 bool cancel_delayed_work(struct delayed_work *dwork)
2937 unsigned long flags;
2941 ret = try_to_grab_pending(&dwork->work, true, &flags);
2942 } while (unlikely(ret == -EAGAIN));
2944 if (unlikely(ret < 0))
2947 set_work_pool_and_clear_pending(&dwork->work,
2948 get_work_pool_id(&dwork->work));
2949 local_irq_restore(flags);
2952 EXPORT_SYMBOL(cancel_delayed_work);
2955 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2956 * @dwork: the delayed work cancel
2958 * This is cancel_work_sync() for delayed works.
2961 * %true if @dwork was pending, %false otherwise.
2963 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2965 return __cancel_work_timer(&dwork->work, true);
2967 EXPORT_SYMBOL(cancel_delayed_work_sync);
2970 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2971 * @func: the function to call
2973 * schedule_on_each_cpu() executes @func on each online CPU using the
2974 * system workqueue and blocks until all CPUs have completed.
2975 * schedule_on_each_cpu() is very slow.
2978 * 0 on success, -errno on failure.
2980 int schedule_on_each_cpu(work_func_t func)
2983 struct work_struct __percpu *works;
2985 works = alloc_percpu(struct work_struct);
2991 for_each_online_cpu(cpu) {
2992 struct work_struct *work = per_cpu_ptr(works, cpu);
2994 INIT_WORK(work, func);
2995 schedule_work_on(cpu, work);
2998 for_each_online_cpu(cpu)
2999 flush_work(per_cpu_ptr(works, cpu));
3007 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3009 * Forces execution of the kernel-global workqueue and blocks until its
3012 * Think twice before calling this function! It's very easy to get into
3013 * trouble if you don't take great care. Either of the following situations
3014 * will lead to deadlock:
3016 * One of the work items currently on the workqueue needs to acquire
3017 * a lock held by your code or its caller.
3019 * Your code is running in the context of a work routine.
3021 * They will be detected by lockdep when they occur, but the first might not
3022 * occur very often. It depends on what work items are on the workqueue and
3023 * what locks they need, which you have no control over.
3025 * In most situations flushing the entire workqueue is overkill; you merely
3026 * need to know that a particular work item isn't queued and isn't running.
3027 * In such cases you should use cancel_delayed_work_sync() or
3028 * cancel_work_sync() instead.
3030 void flush_scheduled_work(void)
3032 flush_workqueue(system_wq);
3034 EXPORT_SYMBOL(flush_scheduled_work);
3037 * execute_in_process_context - reliably execute the routine with user context
3038 * @fn: the function to execute
3039 * @ew: guaranteed storage for the execute work structure (must
3040 * be available when the work executes)
3042 * Executes the function immediately if process context is available,
3043 * otherwise schedules the function for delayed execution.
3045 * Returns: 0 - function was executed
3046 * 1 - function was scheduled for execution
3048 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3050 if (!in_interrupt()) {
3055 INIT_WORK(&ew->work, fn);
3056 schedule_work(&ew->work);
3060 EXPORT_SYMBOL_GPL(execute_in_process_context);
3064 * Workqueues with WQ_SYSFS flag set is visible to userland via
3065 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3066 * following attributes.
3068 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3069 * max_active RW int : maximum number of in-flight work items
3071 * Unbound workqueues have the following extra attributes.
3073 * id RO int : the associated pool ID
3074 * nice RW int : nice value of the workers
3075 * cpumask RW mask : bitmask of allowed CPUs for the workers
3078 struct workqueue_struct *wq;
3082 static struct workqueue_struct *dev_to_wq(struct device *dev)
3084 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3089 static ssize_t wq_per_cpu_show(struct device *dev,
3090 struct device_attribute *attr, char *buf)
3092 struct workqueue_struct *wq = dev_to_wq(dev);
3094 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3097 static ssize_t wq_max_active_show(struct device *dev,
3098 struct device_attribute *attr, char *buf)
3100 struct workqueue_struct *wq = dev_to_wq(dev);
3102 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3105 static ssize_t wq_max_active_store(struct device *dev,
3106 struct device_attribute *attr,
3107 const char *buf, size_t count)
3109 struct workqueue_struct *wq = dev_to_wq(dev);
3112 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3115 workqueue_set_max_active(wq, val);
3119 static struct device_attribute wq_sysfs_attrs[] = {
3120 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3121 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3125 static ssize_t wq_pool_ids_show(struct device *dev,
3126 struct device_attribute *attr, char *buf)
3128 struct workqueue_struct *wq = dev_to_wq(dev);
3129 const char *delim = "";
3130 int node, written = 0;
3132 rcu_read_lock_sched();
3133 for_each_node(node) {
3134 written += scnprintf(buf + written, PAGE_SIZE - written,
3135 "%s%d:%d", delim, node,
3136 unbound_pwq_by_node(wq, node)->pool->id);
3139 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3140 rcu_read_unlock_sched();
3145 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3148 struct workqueue_struct *wq = dev_to_wq(dev);
3151 mutex_lock(&wq->mutex);
3152 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3153 mutex_unlock(&wq->mutex);
3158 /* prepare workqueue_attrs for sysfs store operations */
3159 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3161 struct workqueue_attrs *attrs;
3163 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3167 mutex_lock(&wq->mutex);
3168 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3169 mutex_unlock(&wq->mutex);
3173 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3174 const char *buf, size_t count)
3176 struct workqueue_struct *wq = dev_to_wq(dev);
3177 struct workqueue_attrs *attrs;
3180 attrs = wq_sysfs_prep_attrs(wq);
3184 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3185 attrs->nice >= -20 && attrs->nice <= 19)
3186 ret = apply_workqueue_attrs(wq, attrs);
3190 free_workqueue_attrs(attrs);
3191 return ret ?: count;
3194 static ssize_t wq_cpumask_show(struct device *dev,
3195 struct device_attribute *attr, char *buf)
3197 struct workqueue_struct *wq = dev_to_wq(dev);
3200 mutex_lock(&wq->mutex);
3201 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3202 mutex_unlock(&wq->mutex);
3204 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3208 static ssize_t wq_cpumask_store(struct device *dev,
3209 struct device_attribute *attr,
3210 const char *buf, size_t count)
3212 struct workqueue_struct *wq = dev_to_wq(dev);
3213 struct workqueue_attrs *attrs;
3216 attrs = wq_sysfs_prep_attrs(wq);
3220 ret = cpumask_parse(buf, attrs->cpumask);
3222 ret = apply_workqueue_attrs(wq, attrs);
3224 free_workqueue_attrs(attrs);
3225 return ret ?: count;
3228 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3231 struct workqueue_struct *wq = dev_to_wq(dev);
3234 mutex_lock(&wq->mutex);
3235 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3236 !wq->unbound_attrs->no_numa);
3237 mutex_unlock(&wq->mutex);
3242 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3243 const char *buf, size_t count)
3245 struct workqueue_struct *wq = dev_to_wq(dev);
3246 struct workqueue_attrs *attrs;
3249 attrs = wq_sysfs_prep_attrs(wq);
3254 if (sscanf(buf, "%d", &v) == 1) {
3255 attrs->no_numa = !v;
3256 ret = apply_workqueue_attrs(wq, attrs);
3259 free_workqueue_attrs(attrs);
3260 return ret ?: count;
3263 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3264 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3265 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3266 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3267 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3271 static struct bus_type wq_subsys = {
3272 .name = "workqueue",
3273 .dev_attrs = wq_sysfs_attrs,
3276 static int __init wq_sysfs_init(void)
3278 return subsys_virtual_register(&wq_subsys, NULL);
3280 core_initcall(wq_sysfs_init);
3282 static void wq_device_release(struct device *dev)
3284 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3290 * workqueue_sysfs_register - make a workqueue visible in sysfs
3291 * @wq: the workqueue to register
3293 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3294 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3295 * which is the preferred method.
3297 * Workqueue user should use this function directly iff it wants to apply
3298 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3299 * apply_workqueue_attrs() may race against userland updating the
3302 * Returns 0 on success, -errno on failure.
3304 int workqueue_sysfs_register(struct workqueue_struct *wq)
3306 struct wq_device *wq_dev;
3310 * Adjusting max_active or creating new pwqs by applyting
3311 * attributes breaks ordering guarantee. Disallow exposing ordered
3314 if (WARN_ON(wq->flags & __WQ_ORDERED))
3317 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3322 wq_dev->dev.bus = &wq_subsys;
3323 wq_dev->dev.init_name = wq->name;
3324 wq_dev->dev.release = wq_device_release;
3327 * unbound_attrs are created separately. Suppress uevent until
3328 * everything is ready.
3330 dev_set_uevent_suppress(&wq_dev->dev, true);
3332 ret = device_register(&wq_dev->dev);
3339 if (wq->flags & WQ_UNBOUND) {
3340 struct device_attribute *attr;
3342 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3343 ret = device_create_file(&wq_dev->dev, attr);
3345 device_unregister(&wq_dev->dev);
3352 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3357 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3358 * @wq: the workqueue to unregister
3360 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3362 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3364 struct wq_device *wq_dev = wq->wq_dev;
3370 device_unregister(&wq_dev->dev);
3372 #else /* CONFIG_SYSFS */
3373 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3374 #endif /* CONFIG_SYSFS */
3377 * free_workqueue_attrs - free a workqueue_attrs
3378 * @attrs: workqueue_attrs to free
3380 * Undo alloc_workqueue_attrs().
3382 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3385 free_cpumask_var(attrs->cpumask);
3391 * alloc_workqueue_attrs - allocate a workqueue_attrs
3392 * @gfp_mask: allocation mask to use
3394 * Allocate a new workqueue_attrs, initialize with default settings and
3395 * return it. Returns NULL on failure.
3397 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3399 struct workqueue_attrs *attrs;
3401 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3404 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3407 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3410 free_workqueue_attrs(attrs);
3414 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3415 const struct workqueue_attrs *from)
3417 to->nice = from->nice;
3418 cpumask_copy(to->cpumask, from->cpumask);
3421 /* hash value of the content of @attr */
3422 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3426 hash = jhash_1word(attrs->nice, hash);
3427 hash = jhash(cpumask_bits(attrs->cpumask),
3428 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3432 /* content equality test */
3433 static bool wqattrs_equal(const struct workqueue_attrs *a,
3434 const struct workqueue_attrs *b)
3436 if (a->nice != b->nice)
3438 if (!cpumask_equal(a->cpumask, b->cpumask))
3444 * init_worker_pool - initialize a newly zalloc'd worker_pool
3445 * @pool: worker_pool to initialize
3447 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3448 * Returns 0 on success, -errno on failure. Even on failure, all fields
3449 * inside @pool proper are initialized and put_unbound_pool() can be called
3450 * on @pool safely to release it.
3452 static int init_worker_pool(struct worker_pool *pool)
3454 spin_lock_init(&pool->lock);
3457 pool->node = NUMA_NO_NODE;
3458 pool->flags |= POOL_DISASSOCIATED;
3459 INIT_LIST_HEAD(&pool->worklist);
3460 INIT_LIST_HEAD(&pool->idle_list);
3461 hash_init(pool->busy_hash);
3463 init_timer_deferrable(&pool->idle_timer);
3464 pool->idle_timer.function = idle_worker_timeout;
3465 pool->idle_timer.data = (unsigned long)pool;
3467 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3468 (unsigned long)pool);
3470 mutex_init(&pool->manager_arb);
3471 mutex_init(&pool->manager_mutex);
3472 idr_init(&pool->worker_idr);
3474 INIT_HLIST_NODE(&pool->hash_node);
3477 /* shouldn't fail above this point */
3478 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3484 static void rcu_free_pool(struct rcu_head *rcu)
3486 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3488 idr_destroy(&pool->worker_idr);
3489 free_workqueue_attrs(pool->attrs);
3494 * put_unbound_pool - put a worker_pool
3495 * @pool: worker_pool to put
3497 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3498 * safe manner. get_unbound_pool() calls this function on its failure path
3499 * and this function should be able to release pools which went through,
3500 * successfully or not, init_worker_pool().
3502 * Should be called with wq_pool_mutex held.
3504 static void put_unbound_pool(struct worker_pool *pool)
3506 struct worker *worker;
3508 lockdep_assert_held(&wq_pool_mutex);
3514 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3515 WARN_ON(!list_empty(&pool->worklist)))
3518 /* release id and unhash */
3520 idr_remove(&worker_pool_idr, pool->id);
3521 hash_del(&pool->hash_node);
3524 * Become the manager and destroy all workers. Grabbing
3525 * manager_arb prevents @pool's workers from blocking on
3528 mutex_lock(&pool->manager_arb);
3529 mutex_lock(&pool->manager_mutex);
3530 spin_lock_irq(&pool->lock);
3532 while ((worker = first_worker(pool)))
3533 destroy_worker(worker);
3534 WARN_ON(pool->nr_workers || pool->nr_idle);
3536 spin_unlock_irq(&pool->lock);
3537 mutex_unlock(&pool->manager_mutex);
3538 mutex_unlock(&pool->manager_arb);
3540 /* shut down the timers */
3541 del_timer_sync(&pool->idle_timer);
3542 del_timer_sync(&pool->mayday_timer);
3544 /* sched-RCU protected to allow dereferences from get_work_pool() */
3545 call_rcu_sched(&pool->rcu, rcu_free_pool);
3549 * get_unbound_pool - get a worker_pool with the specified attributes
3550 * @attrs: the attributes of the worker_pool to get
3552 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3553 * reference count and return it. If there already is a matching
3554 * worker_pool, it will be used; otherwise, this function attempts to
3555 * create a new one. On failure, returns NULL.
3557 * Should be called with wq_pool_mutex held.
3559 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3561 u32 hash = wqattrs_hash(attrs);
3562 struct worker_pool *pool;
3565 lockdep_assert_held(&wq_pool_mutex);
3567 /* do we already have a matching pool? */
3568 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3569 if (wqattrs_equal(pool->attrs, attrs)) {
3575 /* nope, create a new one */
3576 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3577 if (!pool || init_worker_pool(pool) < 0)
3580 if (workqueue_freezing)
3581 pool->flags |= POOL_FREEZING;
3583 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3584 copy_workqueue_attrs(pool->attrs, attrs);
3586 /* if cpumask is contained inside a NUMA node, we belong to that node */
3587 if (wq_numa_enabled) {
3588 for_each_node(node) {
3589 if (cpumask_subset(pool->attrs->cpumask,
3590 wq_numa_possible_cpumask[node])) {
3597 if (worker_pool_assign_id(pool) < 0)
3600 /* create and start the initial worker */
3601 if (create_and_start_worker(pool) < 0)
3605 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3610 put_unbound_pool(pool);
3614 static void rcu_free_pwq(struct rcu_head *rcu)
3616 kmem_cache_free(pwq_cache,
3617 container_of(rcu, struct pool_workqueue, rcu));
3621 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3622 * and needs to be destroyed.
3624 static void pwq_unbound_release_workfn(struct work_struct *work)
3626 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3627 unbound_release_work);
3628 struct workqueue_struct *wq = pwq->wq;
3629 struct worker_pool *pool = pwq->pool;
3632 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3636 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3637 * necessary on release but do it anyway. It's easier to verify
3638 * and consistent with the linking path.
3640 mutex_lock(&wq->mutex);
3641 list_del_rcu(&pwq->pwqs_node);
3642 is_last = list_empty(&wq->pwqs);
3643 mutex_unlock(&wq->mutex);
3645 mutex_lock(&wq_pool_mutex);
3646 put_unbound_pool(pool);
3647 mutex_unlock(&wq_pool_mutex);
3649 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3652 * If we're the last pwq going away, @wq is already dead and no one
3653 * is gonna access it anymore. Free it.
3656 free_workqueue_attrs(wq->unbound_attrs);
3662 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3663 * @pwq: target pool_workqueue
3665 * If @pwq isn't freezing, set @pwq->max_active to the associated
3666 * workqueue's saved_max_active and activate delayed work items
3667 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3669 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3671 struct workqueue_struct *wq = pwq->wq;
3672 bool freezable = wq->flags & WQ_FREEZABLE;
3674 /* for @wq->saved_max_active */
3675 lockdep_assert_held(&wq->mutex);
3677 /* fast exit for non-freezable wqs */
3678 if (!freezable && pwq->max_active == wq->saved_max_active)
3681 spin_lock_irq(&pwq->pool->lock);
3683 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3684 pwq->max_active = wq->saved_max_active;
3686 while (!list_empty(&pwq->delayed_works) &&
3687 pwq->nr_active < pwq->max_active)
3688 pwq_activate_first_delayed(pwq);
3691 * Need to kick a worker after thawed or an unbound wq's
3692 * max_active is bumped. It's a slow path. Do it always.
3694 wake_up_worker(pwq->pool);
3696 pwq->max_active = 0;
3699 spin_unlock_irq(&pwq->pool->lock);
3702 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3703 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3704 struct worker_pool *pool)
3706 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3708 memset(pwq, 0, sizeof(*pwq));
3712 pwq->flush_color = -1;
3714 INIT_LIST_HEAD(&pwq->delayed_works);
3715 INIT_LIST_HEAD(&pwq->pwqs_node);
3716 INIT_LIST_HEAD(&pwq->mayday_node);
3717 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3720 /* sync @pwq with the current state of its associated wq and link it */
3721 static void link_pwq(struct pool_workqueue *pwq)
3723 struct workqueue_struct *wq = pwq->wq;
3725 lockdep_assert_held(&wq->mutex);
3727 /* may be called multiple times, ignore if already linked */
3728 if (!list_empty(&pwq->pwqs_node))
3732 * Set the matching work_color. This is synchronized with
3733 * wq->mutex to avoid confusing flush_workqueue().
3735 pwq->work_color = wq->work_color;
3737 /* sync max_active to the current setting */
3738 pwq_adjust_max_active(pwq);
3741 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3744 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3745 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3746 const struct workqueue_attrs *attrs)
3748 struct worker_pool *pool;
3749 struct pool_workqueue *pwq;
3751 lockdep_assert_held(&wq_pool_mutex);
3753 pool = get_unbound_pool(attrs);
3757 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3759 put_unbound_pool(pool);
3763 init_pwq(pwq, wq, pool);
3767 /* undo alloc_unbound_pwq(), used only in the error path */
3768 static void free_unbound_pwq(struct pool_workqueue *pwq)
3770 lockdep_assert_held(&wq_pool_mutex);
3773 put_unbound_pool(pwq->pool);
3774 kmem_cache_free(pwq_cache, pwq);
3779 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3780 * @attrs: the wq_attrs of interest
3781 * @node: the target NUMA node
3782 * @cpu_going_down: if >= 0, the CPU to consider as offline
3783 * @cpumask: outarg, the resulting cpumask
3785 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3786 * @cpu_going_down is >= 0, that cpu is considered offline during
3787 * calculation. The result is stored in @cpumask. This function returns
3788 * %true if the resulting @cpumask is different from @attrs->cpumask,
3791 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3792 * enabled and @node has online CPUs requested by @attrs, the returned
3793 * cpumask is the intersection of the possible CPUs of @node and
3796 * The caller is responsible for ensuring that the cpumask of @node stays
3799 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3800 int cpu_going_down, cpumask_t *cpumask)
3802 if (!wq_numa_enabled || attrs->no_numa)
3805 /* does @node have any online CPUs @attrs wants? */
3806 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3807 if (cpu_going_down >= 0)
3808 cpumask_clear_cpu(cpu_going_down, cpumask);
3810 if (cpumask_empty(cpumask))
3813 /* yeap, return possible CPUs in @node that @attrs wants */
3814 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3815 return !cpumask_equal(cpumask, attrs->cpumask);
3818 cpumask_copy(cpumask, attrs->cpumask);
3822 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3823 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3825 struct pool_workqueue *pwq)
3827 struct pool_workqueue *old_pwq;
3829 lockdep_assert_held(&wq->mutex);
3831 /* link_pwq() can handle duplicate calls */
3834 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3835 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3840 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3841 * @wq: the target workqueue
3842 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3844 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3845 * machines, this function maps a separate pwq to each NUMA node with
3846 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3847 * NUMA node it was issued on. Older pwqs are released as in-flight work
3848 * items finish. Note that a work item which repeatedly requeues itself
3849 * back-to-back will stay on its current pwq.
3851 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3854 int apply_workqueue_attrs(struct workqueue_struct *wq,
3855 const struct workqueue_attrs *attrs)
3857 struct workqueue_attrs *new_attrs, *tmp_attrs;
3858 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3861 /* only unbound workqueues can change attributes */
3862 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3865 /* creating multiple pwqs breaks ordering guarantee */
3866 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3869 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3870 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3871 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3872 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3875 /* make a copy of @attrs and sanitize it */
3876 copy_workqueue_attrs(new_attrs, attrs);
3877 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3880 * We may create multiple pwqs with differing cpumasks. Make a
3881 * copy of @new_attrs which will be modified and used to obtain
3884 copy_workqueue_attrs(tmp_attrs, new_attrs);
3887 * CPUs should stay stable across pwq creations and installations.
3888 * Pin CPUs, determine the target cpumask for each node and create
3893 mutex_lock(&wq_pool_mutex);
3896 * If something goes wrong during CPU up/down, we'll fall back to
3897 * the default pwq covering whole @attrs->cpumask. Always create
3898 * it even if we don't use it immediately.
3900 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3904 for_each_node(node) {
3905 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3906 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3911 pwq_tbl[node] = dfl_pwq;
3915 mutex_unlock(&wq_pool_mutex);
3917 /* all pwqs have been created successfully, let's install'em */
3918 mutex_lock(&wq->mutex);
3920 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3922 /* save the previous pwq and install the new one */
3924 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3926 /* @dfl_pwq might not have been used, ensure it's linked */
3928 swap(wq->dfl_pwq, dfl_pwq);
3930 mutex_unlock(&wq->mutex);
3932 /* put the old pwqs */
3934 put_pwq_unlocked(pwq_tbl[node]);
3935 put_pwq_unlocked(dfl_pwq);
3941 free_workqueue_attrs(tmp_attrs);
3942 free_workqueue_attrs(new_attrs);
3947 free_unbound_pwq(dfl_pwq);
3949 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3950 free_unbound_pwq(pwq_tbl[node]);
3951 mutex_unlock(&wq_pool_mutex);
3959 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3960 * @wq: the target workqueue
3961 * @cpu: the CPU coming up or going down
3962 * @online: whether @cpu is coming up or going down
3964 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3965 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3968 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3969 * falls back to @wq->dfl_pwq which may not be optimal but is always
3972 * Note that when the last allowed CPU of a NUMA node goes offline for a
3973 * workqueue with a cpumask spanning multiple nodes, the workers which were
3974 * already executing the work items for the workqueue will lose their CPU
3975 * affinity and may execute on any CPU. This is similar to how per-cpu
3976 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3977 * affinity, it's the user's responsibility to flush the work item from
3980 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3983 int node = cpu_to_node(cpu);
3984 int cpu_off = online ? -1 : cpu;
3985 struct pool_workqueue *old_pwq = NULL, *pwq;
3986 struct workqueue_attrs *target_attrs;
3989 lockdep_assert_held(&wq_pool_mutex);
3991 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3995 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3996 * Let's use a preallocated one. The following buf is protected by
3997 * CPU hotplug exclusion.
3999 target_attrs = wq_update_unbound_numa_attrs_buf;
4000 cpumask = target_attrs->cpumask;
4002 mutex_lock(&wq->mutex);
4003 if (wq->unbound_attrs->no_numa)
4006 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4007 pwq = unbound_pwq_by_node(wq, node);
4010 * Let's determine what needs to be done. If the target cpumask is
4011 * different from wq's, we need to compare it to @pwq's and create
4012 * a new one if they don't match. If the target cpumask equals
4013 * wq's, the default pwq should be used. If @pwq is already the
4014 * default one, nothing to do; otherwise, install the default one.
4016 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4017 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4020 if (pwq == wq->dfl_pwq)
4026 mutex_unlock(&wq->mutex);
4028 /* create a new pwq */
4029 pwq = alloc_unbound_pwq(wq, target_attrs);
4031 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4037 * Install the new pwq. As this function is called only from CPU
4038 * hotplug callbacks and applying a new attrs is wrapped with
4039 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4042 mutex_lock(&wq->mutex);
4043 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4047 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4048 get_pwq(wq->dfl_pwq);
4049 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4050 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4052 mutex_unlock(&wq->mutex);
4053 put_pwq_unlocked(old_pwq);
4056 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4058 bool highpri = wq->flags & WQ_HIGHPRI;
4061 if (!(wq->flags & WQ_UNBOUND)) {
4062 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4066 for_each_possible_cpu(cpu) {
4067 struct pool_workqueue *pwq =
4068 per_cpu_ptr(wq->cpu_pwqs, cpu);
4069 struct worker_pool *cpu_pools =
4070 per_cpu(cpu_worker_pools, cpu);
4072 init_pwq(pwq, wq, &cpu_pools[highpri]);
4074 mutex_lock(&wq->mutex);
4076 mutex_unlock(&wq->mutex);
4080 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4084 static int wq_clamp_max_active(int max_active, unsigned int flags,
4087 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4089 if (max_active < 1 || max_active > lim)
4090 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4091 max_active, name, 1, lim);
4093 return clamp_val(max_active, 1, lim);
4096 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4099 struct lock_class_key *key,
4100 const char *lock_name, ...)
4102 size_t tbl_size = 0;
4104 struct workqueue_struct *wq;
4105 struct pool_workqueue *pwq;
4107 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4108 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4109 flags |= WQ_UNBOUND;
4111 /* allocate wq and format name */
4112 if (flags & WQ_UNBOUND)
4113 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4115 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4119 if (flags & WQ_UNBOUND) {
4120 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4121 if (!wq->unbound_attrs)
4125 va_start(args, lock_name);
4126 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4129 max_active = max_active ?: WQ_DFL_ACTIVE;
4130 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4134 wq->saved_max_active = max_active;
4135 mutex_init(&wq->mutex);
4136 atomic_set(&wq->nr_pwqs_to_flush, 0);
4137 INIT_LIST_HEAD(&wq->pwqs);
4138 INIT_LIST_HEAD(&wq->flusher_queue);
4139 INIT_LIST_HEAD(&wq->flusher_overflow);
4140 INIT_LIST_HEAD(&wq->maydays);
4142 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4143 INIT_LIST_HEAD(&wq->list);
4145 if (alloc_and_link_pwqs(wq) < 0)
4149 * Workqueues which may be used during memory reclaim should
4150 * have a rescuer to guarantee forward progress.
4152 if (flags & WQ_MEM_RECLAIM) {
4153 struct worker *rescuer;
4155 rescuer = alloc_worker();
4159 rescuer->rescue_wq = wq;
4160 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4162 if (IS_ERR(rescuer->task)) {
4167 wq->rescuer = rescuer;
4168 rescuer->task->flags |= PF_NO_SETAFFINITY;
4169 wake_up_process(rescuer->task);
4172 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4176 * wq_pool_mutex protects global freeze state and workqueues list.
4177 * Grab it, adjust max_active and add the new @wq to workqueues
4180 mutex_lock(&wq_pool_mutex);
4182 mutex_lock(&wq->mutex);
4183 for_each_pwq(pwq, wq)
4184 pwq_adjust_max_active(pwq);
4185 mutex_unlock(&wq->mutex);
4187 list_add(&wq->list, &workqueues);
4189 mutex_unlock(&wq_pool_mutex);
4194 free_workqueue_attrs(wq->unbound_attrs);
4198 destroy_workqueue(wq);
4201 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4204 * destroy_workqueue - safely terminate a workqueue
4205 * @wq: target workqueue
4207 * Safely destroy a workqueue. All work currently pending will be done first.
4209 void destroy_workqueue(struct workqueue_struct *wq)
4211 struct pool_workqueue *pwq;
4214 /* drain it before proceeding with destruction */
4215 drain_workqueue(wq);
4218 mutex_lock(&wq->mutex);
4219 for_each_pwq(pwq, wq) {
4222 for (i = 0; i < WORK_NR_COLORS; i++) {
4223 if (WARN_ON(pwq->nr_in_flight[i])) {
4224 mutex_unlock(&wq->mutex);
4229 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4230 WARN_ON(pwq->nr_active) ||
4231 WARN_ON(!list_empty(&pwq->delayed_works))) {
4232 mutex_unlock(&wq->mutex);
4236 mutex_unlock(&wq->mutex);
4239 * wq list is used to freeze wq, remove from list after
4240 * flushing is complete in case freeze races us.
4242 mutex_lock(&wq_pool_mutex);
4243 list_del_init(&wq->list);
4244 mutex_unlock(&wq_pool_mutex);
4246 workqueue_sysfs_unregister(wq);
4249 kthread_stop(wq->rescuer->task);
4254 if (!(wq->flags & WQ_UNBOUND)) {
4256 * The base ref is never dropped on per-cpu pwqs. Directly
4257 * free the pwqs and wq.
4259 free_percpu(wq->cpu_pwqs);
4263 * We're the sole accessor of @wq at this point. Directly
4264 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4265 * @wq will be freed when the last pwq is released.
4267 for_each_node(node) {
4268 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4269 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4270 put_pwq_unlocked(pwq);
4274 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4275 * put. Don't access it afterwards.
4279 put_pwq_unlocked(pwq);
4282 EXPORT_SYMBOL_GPL(destroy_workqueue);
4285 * workqueue_set_max_active - adjust max_active of a workqueue
4286 * @wq: target workqueue
4287 * @max_active: new max_active value.
4289 * Set max_active of @wq to @max_active.
4292 * Don't call from IRQ context.
4294 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4296 struct pool_workqueue *pwq;
4298 /* disallow meddling with max_active for ordered workqueues */
4299 if (WARN_ON(wq->flags & __WQ_ORDERED))
4302 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4304 mutex_lock(&wq->mutex);
4306 wq->saved_max_active = max_active;
4308 for_each_pwq(pwq, wq)
4309 pwq_adjust_max_active(pwq);
4311 mutex_unlock(&wq->mutex);
4313 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4316 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4318 * Determine whether %current is a workqueue rescuer. Can be used from
4319 * work functions to determine whether it's being run off the rescuer task.
4321 bool current_is_workqueue_rescuer(void)
4323 struct worker *worker = current_wq_worker();
4325 return worker && worker->rescue_wq;
4329 * workqueue_congested - test whether a workqueue is congested
4330 * @cpu: CPU in question
4331 * @wq: target workqueue
4333 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4334 * no synchronization around this function and the test result is
4335 * unreliable and only useful as advisory hints or for debugging.
4337 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4338 * Note that both per-cpu and unbound workqueues may be associated with
4339 * multiple pool_workqueues which have separate congested states. A
4340 * workqueue being congested on one CPU doesn't mean the workqueue is also
4341 * contested on other CPUs / NUMA nodes.
4344 * %true if congested, %false otherwise.
4346 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4348 struct pool_workqueue *pwq;
4351 rcu_read_lock_sched();
4353 if (cpu == WORK_CPU_UNBOUND)
4354 cpu = smp_processor_id();
4356 if (!(wq->flags & WQ_UNBOUND))
4357 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4359 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4361 ret = !list_empty(&pwq->delayed_works);
4362 rcu_read_unlock_sched();
4366 EXPORT_SYMBOL_GPL(workqueue_congested);
4369 * work_busy - test whether a work is currently pending or running
4370 * @work: the work to be tested
4372 * Test whether @work is currently pending or running. There is no
4373 * synchronization around this function and the test result is
4374 * unreliable and only useful as advisory hints or for debugging.
4377 * OR'd bitmask of WORK_BUSY_* bits.
4379 unsigned int work_busy(struct work_struct *work)
4381 struct worker_pool *pool;
4382 unsigned long flags;
4383 unsigned int ret = 0;
4385 if (work_pending(work))
4386 ret |= WORK_BUSY_PENDING;
4388 local_irq_save(flags);
4389 pool = get_work_pool(work);
4391 spin_lock(&pool->lock);
4392 if (find_worker_executing_work(pool, work))
4393 ret |= WORK_BUSY_RUNNING;
4394 spin_unlock(&pool->lock);
4396 local_irq_restore(flags);
4400 EXPORT_SYMBOL_GPL(work_busy);
4403 * set_worker_desc - set description for the current work item
4404 * @fmt: printf-style format string
4405 * @...: arguments for the format string
4407 * This function can be called by a running work function to describe what
4408 * the work item is about. If the worker task gets dumped, this
4409 * information will be printed out together to help debugging. The
4410 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4412 void set_worker_desc(const char *fmt, ...)
4414 struct worker *worker = current_wq_worker();
4418 va_start(args, fmt);
4419 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4421 worker->desc_valid = true;
4426 * print_worker_info - print out worker information and description
4427 * @log_lvl: the log level to use when printing
4428 * @task: target task
4430 * If @task is a worker and currently executing a work item, print out the
4431 * name of the workqueue being serviced and worker description set with
4432 * set_worker_desc() by the currently executing work item.
4434 * This function can be safely called on any task as long as the
4435 * task_struct itself is accessible. While safe, this function isn't
4436 * synchronized and may print out mixups or garbages of limited length.
4438 void print_worker_info(const char *log_lvl, struct task_struct *task)
4440 work_func_t *fn = NULL;
4441 char name[WQ_NAME_LEN] = { };
4442 char desc[WORKER_DESC_LEN] = { };
4443 struct pool_workqueue *pwq = NULL;
4444 struct workqueue_struct *wq = NULL;
4445 bool desc_valid = false;
4446 struct worker *worker;
4448 if (!(task->flags & PF_WQ_WORKER))
4452 * This function is called without any synchronization and @task
4453 * could be in any state. Be careful with dereferences.
4455 worker = probe_kthread_data(task);
4458 * Carefully copy the associated workqueue's workfn and name. Keep
4459 * the original last '\0' in case the original contains garbage.
4461 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4462 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4463 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4464 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4466 /* copy worker description */
4467 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4469 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4471 if (fn || name[0] || desc[0]) {
4472 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4474 pr_cont(" (%s)", desc);
4482 * There are two challenges in supporting CPU hotplug. Firstly, there
4483 * are a lot of assumptions on strong associations among work, pwq and
4484 * pool which make migrating pending and scheduled works very
4485 * difficult to implement without impacting hot paths. Secondly,
4486 * worker pools serve mix of short, long and very long running works making
4487 * blocked draining impractical.
4489 * This is solved by allowing the pools to be disassociated from the CPU
4490 * running as an unbound one and allowing it to be reattached later if the
4491 * cpu comes back online.
4494 static void wq_unbind_fn(struct work_struct *work)
4496 int cpu = smp_processor_id();
4497 struct worker_pool *pool;
4498 struct worker *worker;
4501 for_each_cpu_worker_pool(pool, cpu) {
4502 WARN_ON_ONCE(cpu != smp_processor_id());
4504 mutex_lock(&pool->manager_mutex);
4505 spin_lock_irq(&pool->lock);
4508 * We've blocked all manager operations. Make all workers
4509 * unbound and set DISASSOCIATED. Before this, all workers
4510 * except for the ones which are still executing works from
4511 * before the last CPU down must be on the cpu. After
4512 * this, they may become diasporas.
4514 for_each_pool_worker(worker, wi, pool)
4515 worker->flags |= WORKER_UNBOUND;
4517 pool->flags |= POOL_DISASSOCIATED;
4519 spin_unlock_irq(&pool->lock);
4520 mutex_unlock(&pool->manager_mutex);
4523 * Call schedule() so that we cross rq->lock and thus can
4524 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4525 * This is necessary as scheduler callbacks may be invoked
4531 * Sched callbacks are disabled now. Zap nr_running.
4532 * After this, nr_running stays zero and need_more_worker()
4533 * and keep_working() are always true as long as the
4534 * worklist is not empty. This pool now behaves as an
4535 * unbound (in terms of concurrency management) pool which
4536 * are served by workers tied to the pool.
4538 atomic_set(&pool->nr_running, 0);
4541 * With concurrency management just turned off, a busy
4542 * worker blocking could lead to lengthy stalls. Kick off
4543 * unbound chain execution of currently pending work items.
4545 spin_lock_irq(&pool->lock);
4546 wake_up_worker(pool);
4547 spin_unlock_irq(&pool->lock);
4552 * rebind_workers - rebind all workers of a pool to the associated CPU
4553 * @pool: pool of interest
4555 * @pool->cpu is coming online. Rebind all workers to the CPU.
4557 static void rebind_workers(struct worker_pool *pool)
4559 struct worker *worker;
4562 lockdep_assert_held(&pool->manager_mutex);
4565 * Restore CPU affinity of all workers. As all idle workers should
4566 * be on the run-queue of the associated CPU before any local
4567 * wake-ups for concurrency management happen, restore CPU affinty
4568 * of all workers first and then clear UNBOUND. As we're called
4569 * from CPU_ONLINE, the following shouldn't fail.
4571 for_each_pool_worker(worker, wi, pool)
4572 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4573 pool->attrs->cpumask) < 0);
4575 spin_lock_irq(&pool->lock);
4577 for_each_pool_worker(worker, wi, pool) {
4578 unsigned int worker_flags = worker->flags;
4581 * A bound idle worker should actually be on the runqueue
4582 * of the associated CPU for local wake-ups targeting it to
4583 * work. Kick all idle workers so that they migrate to the
4584 * associated CPU. Doing this in the same loop as
4585 * replacing UNBOUND with REBOUND is safe as no worker will
4586 * be bound before @pool->lock is released.
4588 if (worker_flags & WORKER_IDLE)
4589 wake_up_process(worker->task);
4592 * We want to clear UNBOUND but can't directly call
4593 * worker_clr_flags() or adjust nr_running. Atomically
4594 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4595 * @worker will clear REBOUND using worker_clr_flags() when
4596 * it initiates the next execution cycle thus restoring
4597 * concurrency management. Note that when or whether
4598 * @worker clears REBOUND doesn't affect correctness.
4600 * ACCESS_ONCE() is necessary because @worker->flags may be
4601 * tested without holding any lock in
4602 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4603 * fail incorrectly leading to premature concurrency
4604 * management operations.
4606 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4607 worker_flags |= WORKER_REBOUND;
4608 worker_flags &= ~WORKER_UNBOUND;
4609 ACCESS_ONCE(worker->flags) = worker_flags;
4612 spin_unlock_irq(&pool->lock);
4616 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4617 * @pool: unbound pool of interest
4618 * @cpu: the CPU which is coming up
4620 * An unbound pool may end up with a cpumask which doesn't have any online
4621 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4622 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4623 * online CPU before, cpus_allowed of all its workers should be restored.
4625 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4627 static cpumask_t cpumask;
4628 struct worker *worker;
4631 lockdep_assert_held(&pool->manager_mutex);
4633 /* is @cpu allowed for @pool? */
4634 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4637 /* is @cpu the only online CPU? */
4638 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4639 if (cpumask_weight(&cpumask) != 1)
4642 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4643 for_each_pool_worker(worker, wi, pool)
4644 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4645 pool->attrs->cpumask) < 0);
4649 * Workqueues should be brought up before normal priority CPU notifiers.
4650 * This will be registered high priority CPU notifier.
4652 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4653 unsigned long action,
4656 int cpu = (unsigned long)hcpu;
4657 struct worker_pool *pool;
4658 struct workqueue_struct *wq;
4661 switch (action & ~CPU_TASKS_FROZEN) {
4662 case CPU_UP_PREPARE:
4663 for_each_cpu_worker_pool(pool, cpu) {
4664 if (pool->nr_workers)
4666 if (create_and_start_worker(pool) < 0)
4671 case CPU_DOWN_FAILED:
4673 mutex_lock(&wq_pool_mutex);
4675 for_each_pool(pool, pi) {
4676 mutex_lock(&pool->manager_mutex);
4678 if (pool->cpu == cpu) {
4679 spin_lock_irq(&pool->lock);
4680 pool->flags &= ~POOL_DISASSOCIATED;
4681 spin_unlock_irq(&pool->lock);
4683 rebind_workers(pool);
4684 } else if (pool->cpu < 0) {
4685 restore_unbound_workers_cpumask(pool, cpu);
4688 mutex_unlock(&pool->manager_mutex);
4691 /* update NUMA affinity of unbound workqueues */
4692 list_for_each_entry(wq, &workqueues, list)
4693 wq_update_unbound_numa(wq, cpu, true);
4695 mutex_unlock(&wq_pool_mutex);
4702 * Workqueues should be brought down after normal priority CPU notifiers.
4703 * This will be registered as low priority CPU notifier.
4705 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4706 unsigned long action,
4709 int cpu = (unsigned long)hcpu;
4710 struct work_struct unbind_work;
4711 struct workqueue_struct *wq;
4713 switch (action & ~CPU_TASKS_FROZEN) {
4714 case CPU_DOWN_PREPARE:
4715 /* unbinding per-cpu workers should happen on the local CPU */
4716 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4717 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4719 /* update NUMA affinity of unbound workqueues */
4720 mutex_lock(&wq_pool_mutex);
4721 list_for_each_entry(wq, &workqueues, list)
4722 wq_update_unbound_numa(wq, cpu, false);
4723 mutex_unlock(&wq_pool_mutex);
4725 /* wait for per-cpu unbinding to finish */
4726 flush_work(&unbind_work);
4734 struct work_for_cpu {
4735 struct work_struct work;
4741 static void work_for_cpu_fn(struct work_struct *work)
4743 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4745 wfc->ret = wfc->fn(wfc->arg);
4749 * work_on_cpu - run a function in user context on a particular cpu
4750 * @cpu: the cpu to run on
4751 * @fn: the function to run
4752 * @arg: the function arg
4754 * This will return the value @fn returns.
4755 * It is up to the caller to ensure that the cpu doesn't go offline.
4756 * The caller must not hold any locks which would prevent @fn from completing.
4758 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4760 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4762 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4763 schedule_work_on(cpu, &wfc.work);
4766 * The work item is on-stack and can't lead to deadlock through
4767 * flushing. Use __flush_work() to avoid spurious lockdep warnings
4768 * when work_on_cpu()s are nested.
4770 __flush_work(&wfc.work);
4774 EXPORT_SYMBOL_GPL(work_on_cpu);
4775 #endif /* CONFIG_SMP */
4777 #ifdef CONFIG_FREEZER
4780 * freeze_workqueues_begin - begin freezing workqueues
4782 * Start freezing workqueues. After this function returns, all freezable
4783 * workqueues will queue new works to their delayed_works list instead of
4787 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4789 void freeze_workqueues_begin(void)
4791 struct worker_pool *pool;
4792 struct workqueue_struct *wq;
4793 struct pool_workqueue *pwq;
4796 mutex_lock(&wq_pool_mutex);
4798 WARN_ON_ONCE(workqueue_freezing);
4799 workqueue_freezing = true;
4802 for_each_pool(pool, pi) {
4803 spin_lock_irq(&pool->lock);
4804 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4805 pool->flags |= POOL_FREEZING;
4806 spin_unlock_irq(&pool->lock);
4809 list_for_each_entry(wq, &workqueues, list) {
4810 mutex_lock(&wq->mutex);
4811 for_each_pwq(pwq, wq)
4812 pwq_adjust_max_active(pwq);
4813 mutex_unlock(&wq->mutex);
4816 mutex_unlock(&wq_pool_mutex);
4820 * freeze_workqueues_busy - are freezable workqueues still busy?
4822 * Check whether freezing is complete. This function must be called
4823 * between freeze_workqueues_begin() and thaw_workqueues().
4826 * Grabs and releases wq_pool_mutex.
4829 * %true if some freezable workqueues are still busy. %false if freezing
4832 bool freeze_workqueues_busy(void)
4835 struct workqueue_struct *wq;
4836 struct pool_workqueue *pwq;
4838 mutex_lock(&wq_pool_mutex);
4840 WARN_ON_ONCE(!workqueue_freezing);
4842 list_for_each_entry(wq, &workqueues, list) {
4843 if (!(wq->flags & WQ_FREEZABLE))
4846 * nr_active is monotonically decreasing. It's safe
4847 * to peek without lock.
4849 rcu_read_lock_sched();
4850 for_each_pwq(pwq, wq) {
4851 WARN_ON_ONCE(pwq->nr_active < 0);
4852 if (pwq->nr_active) {
4854 rcu_read_unlock_sched();
4858 rcu_read_unlock_sched();
4861 mutex_unlock(&wq_pool_mutex);
4866 * thaw_workqueues - thaw workqueues
4868 * Thaw workqueues. Normal queueing is restored and all collected
4869 * frozen works are transferred to their respective pool worklists.
4872 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4874 void thaw_workqueues(void)
4876 struct workqueue_struct *wq;
4877 struct pool_workqueue *pwq;
4878 struct worker_pool *pool;
4881 mutex_lock(&wq_pool_mutex);
4883 if (!workqueue_freezing)
4886 /* clear FREEZING */
4887 for_each_pool(pool, pi) {
4888 spin_lock_irq(&pool->lock);
4889 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4890 pool->flags &= ~POOL_FREEZING;
4891 spin_unlock_irq(&pool->lock);
4894 /* restore max_active and repopulate worklist */
4895 list_for_each_entry(wq, &workqueues, list) {
4896 mutex_lock(&wq->mutex);
4897 for_each_pwq(pwq, wq)
4898 pwq_adjust_max_active(pwq);
4899 mutex_unlock(&wq->mutex);
4902 workqueue_freezing = false;
4904 mutex_unlock(&wq_pool_mutex);
4906 #endif /* CONFIG_FREEZER */
4908 static void __init wq_numa_init(void)
4913 /* determine NUMA pwq table len - highest node id + 1 */
4915 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4917 if (num_possible_nodes() <= 1)
4920 if (wq_disable_numa) {
4921 pr_info("workqueue: NUMA affinity support disabled\n");
4925 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4926 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4929 * We want masks of possible CPUs of each node which isn't readily
4930 * available. Build one from cpu_to_node() which should have been
4931 * fully initialized by now.
4933 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4937 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4938 node_online(node) ? node : NUMA_NO_NODE));
4940 for_each_possible_cpu(cpu) {
4941 node = cpu_to_node(cpu);
4942 if (WARN_ON(node == NUMA_NO_NODE)) {
4943 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4944 /* happens iff arch is bonkers, let's just proceed */
4947 cpumask_set_cpu(cpu, tbl[node]);
4950 wq_numa_possible_cpumask = tbl;
4951 wq_numa_enabled = true;
4954 static int __init init_workqueues(void)
4956 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4959 /* make sure we have enough bits for OFFQ pool ID */
4960 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4961 WORK_CPU_END * NR_STD_WORKER_POOLS);
4963 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4965 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4967 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4968 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4972 /* initialize CPU pools */
4973 for_each_possible_cpu(cpu) {
4974 struct worker_pool *pool;
4977 for_each_cpu_worker_pool(pool, cpu) {
4978 BUG_ON(init_worker_pool(pool));
4980 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4981 pool->attrs->nice = std_nice[i++];
4982 pool->node = cpu_to_node(cpu);
4985 mutex_lock(&wq_pool_mutex);
4986 BUG_ON(worker_pool_assign_id(pool));
4987 mutex_unlock(&wq_pool_mutex);
4991 /* create the initial worker */
4992 for_each_online_cpu(cpu) {
4993 struct worker_pool *pool;
4995 for_each_cpu_worker_pool(pool, cpu) {
4996 pool->flags &= ~POOL_DISASSOCIATED;
4997 BUG_ON(create_and_start_worker(pool) < 0);
5001 /* create default unbound wq attrs */
5002 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5003 struct workqueue_attrs *attrs;
5005 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5006 attrs->nice = std_nice[i];
5007 unbound_std_wq_attrs[i] = attrs;
5010 system_wq = alloc_workqueue("events", 0, 0);
5011 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5012 system_long_wq = alloc_workqueue("events_long", 0, 0);
5013 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5014 WQ_UNBOUND_MAX_ACTIVE);
5015 system_freezable_wq = alloc_workqueue("events_freezable",
5017 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5018 WQ_POWER_EFFICIENT, 0);
5019 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5020 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5022 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5023 !system_unbound_wq || !system_freezable_wq ||
5024 !system_power_efficient_wq ||
5025 !system_freezable_power_efficient_wq);
5028 early_initcall(init_workqueues);