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>
49 #include "workqueue_internal.h"
55 * A bound pool is either associated or disassociated with its CPU.
56 * While associated (!DISASSOCIATED), all workers are bound to the
57 * CPU and none has %WORKER_UNBOUND set and concurrency management
60 * While DISASSOCIATED, the cpu may be offline and all workers have
61 * %WORKER_UNBOUND set and concurrency management disabled, and may
62 * be executing on any CPU. The pool behaves as an unbound one.
64 * Note that DISASSOCIATED should be flipped only while holding
65 * manager_mutex to avoid changing binding state while
66 * create_worker() is in progress.
68 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
69 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
70 POOL_FREEZING = 1 << 3, /* freeze in progress */
73 WORKER_STARTED = 1 << 0, /* started */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give -20.
102 RESCUER_NICE_LEVEL = -20,
103 HIGHPRI_NICE_LEVEL = -20,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * MG: pool->manager_mutex and pool->lock protected. Writes require both
125 * locks. Reads can happen under either lock.
127 * PL: wq_pool_mutex protected.
129 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131 * WQ: wq->mutex protected.
133 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
135 * MD: wq_mayday_lock protected.
138 /* struct worker is defined in workqueue_internal.h */
141 spinlock_t lock; /* the pool lock */
142 int cpu; /* I: the associated cpu */
143 int node; /* I: the associated node ID */
144 int id; /* I: pool ID */
145 unsigned int flags; /* X: flags */
147 struct list_head worklist; /* L: list of pending works */
148 int nr_workers; /* L: total number of workers */
150 /* nr_idle includes the ones off idle_list for rebinding */
151 int nr_idle; /* L: currently idle ones */
153 struct list_head idle_list; /* X: list of idle workers */
154 struct timer_list idle_timer; /* L: worker idle timeout */
155 struct timer_list mayday_timer; /* L: SOS timer for workers */
157 /* a workers is either on busy_hash or idle_list, or the manager */
158 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
159 /* L: hash of busy workers */
161 /* see manage_workers() for details on the two manager mutexes */
162 struct mutex manager_arb; /* manager arbitration */
163 struct mutex manager_mutex; /* manager exclusion */
164 struct idr worker_idr; /* MG: worker IDs and iteration */
166 struct workqueue_attrs *attrs; /* I: worker attributes */
167 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
168 int refcnt; /* PL: refcnt for unbound pools */
171 * The current concurrency level. As it's likely to be accessed
172 * from other CPUs during try_to_wake_up(), put it in a separate
175 atomic_t nr_running ____cacheline_aligned_in_smp;
178 * Destruction of pool is sched-RCU protected to allow dereferences
179 * from get_work_pool().
182 } ____cacheline_aligned_in_smp;
185 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
186 * of work_struct->data are used for flags and the remaining high bits
187 * point to the pwq; thus, pwqs need to be aligned at two's power of the
188 * number of flag bits.
190 struct pool_workqueue {
191 struct worker_pool *pool; /* I: the associated pool */
192 struct workqueue_struct *wq; /* I: the owning workqueue */
193 int work_color; /* L: current color */
194 int flush_color; /* L: flushing color */
195 int refcnt; /* L: reference count */
196 int nr_in_flight[WORK_NR_COLORS];
197 /* L: nr of in_flight works */
198 int nr_active; /* L: nr of active works */
199 int max_active; /* L: max active works */
200 struct list_head delayed_works; /* L: delayed works */
201 struct list_head pwqs_node; /* WR: node on wq->pwqs */
202 struct list_head mayday_node; /* MD: node on wq->maydays */
205 * Release of unbound pwq is punted to system_wq. See put_pwq()
206 * and pwq_unbound_release_workfn() for details. pool_workqueue
207 * itself is also sched-RCU protected so that the first pwq can be
208 * determined without grabbing wq->mutex.
210 struct work_struct unbound_release_work;
212 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
215 * Structure used to wait for workqueue flush.
218 struct list_head list; /* WQ: list of flushers */
219 int flush_color; /* WQ: flush color waiting for */
220 struct completion done; /* flush completion */
226 * The externally visible workqueue. It relays the issued work items to
227 * the appropriate worker_pool through its pool_workqueues.
229 struct workqueue_struct {
230 struct list_head pwqs; /* WR: all pwqs of this wq */
231 struct list_head list; /* PL: list of all workqueues */
233 struct mutex mutex; /* protects this wq */
234 int work_color; /* WQ: current work color */
235 int flush_color; /* WQ: current flush color */
236 atomic_t nr_pwqs_to_flush; /* flush in progress */
237 struct wq_flusher *first_flusher; /* WQ: first flusher */
238 struct list_head flusher_queue; /* WQ: flush waiters */
239 struct list_head flusher_overflow; /* WQ: flush overflow list */
241 struct list_head maydays; /* MD: pwqs requesting rescue */
242 struct worker *rescuer; /* I: rescue worker */
244 int nr_drainers; /* WQ: drain in progress */
245 int saved_max_active; /* WQ: saved pwq max_active */
247 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 struct wq_device *wq_dev; /* I: for sysfs interface */
252 #ifdef CONFIG_LOCKDEP
253 struct lockdep_map lockdep_map;
255 char name[WQ_NAME_LEN]; /* I: workqueue name */
257 /* hot fields used during command issue, aligned to cacheline */
258 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
259 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
260 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
263 static struct kmem_cache *pwq_cache;
265 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
266 static cpumask_var_t *wq_numa_possible_cpumask;
267 /* possible CPUs of each node */
269 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
271 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
272 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
274 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
275 static bool workqueue_freezing; /* PL: have wqs started freezing? */
277 /* the per-cpu worker pools */
278 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
281 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
283 /* PL: hash of all unbound pools keyed by pool->attrs */
284 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
286 /* I: attributes used when instantiating standard unbound pools on demand */
287 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
289 struct workqueue_struct *system_wq __read_mostly;
290 EXPORT_SYMBOL_GPL(system_wq);
291 struct workqueue_struct *system_highpri_wq __read_mostly;
292 EXPORT_SYMBOL_GPL(system_highpri_wq);
293 struct workqueue_struct *system_long_wq __read_mostly;
294 EXPORT_SYMBOL_GPL(system_long_wq);
295 struct workqueue_struct *system_unbound_wq __read_mostly;
296 EXPORT_SYMBOL_GPL(system_unbound_wq);
297 struct workqueue_struct *system_freezable_wq __read_mostly;
298 EXPORT_SYMBOL_GPL(system_freezable_wq);
300 static int worker_thread(void *__worker);
301 static void copy_workqueue_attrs(struct workqueue_attrs *to,
302 const struct workqueue_attrs *from);
304 #define CREATE_TRACE_POINTS
305 #include <trace/events/workqueue.h>
307 #define assert_rcu_or_pool_mutex() \
308 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
309 lockdep_is_held(&wq_pool_mutex), \
310 "sched RCU or wq_pool_mutex should be held")
312 #define assert_rcu_or_wq_mutex(wq) \
313 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
314 lockdep_is_held(&wq->mutex), \
315 "sched RCU or wq->mutex should be held")
317 #ifdef CONFIG_LOCKDEP
318 #define assert_manager_or_pool_lock(pool) \
319 WARN_ONCE(debug_locks && \
320 !lockdep_is_held(&(pool)->manager_mutex) && \
321 !lockdep_is_held(&(pool)->lock), \
322 "pool->manager_mutex or ->lock should be held")
324 #define assert_manager_or_pool_lock(pool) do { } while (0)
327 #define for_each_cpu_worker_pool(pool, cpu) \
328 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
329 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
333 * for_each_pool - iterate through all worker_pools in the system
334 * @pool: iteration cursor
335 * @pi: integer used for iteration
337 * This must be called either with wq_pool_mutex held or sched RCU read
338 * locked. If the pool needs to be used beyond the locking in effect, the
339 * caller is responsible for guaranteeing that the pool stays online.
341 * The if/else clause exists only for the lockdep assertion and can be
344 #define for_each_pool(pool, pi) \
345 idr_for_each_entry(&worker_pool_idr, pool, pi) \
346 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
350 * for_each_pool_worker - iterate through all workers of a worker_pool
351 * @worker: iteration cursor
352 * @wi: integer used for iteration
353 * @pool: worker_pool to iterate workers of
355 * This must be called with either @pool->manager_mutex or ->lock held.
357 * The if/else clause exists only for the lockdep assertion and can be
360 #define for_each_pool_worker(worker, wi, pool) \
361 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
362 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
366 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
367 * @pwq: iteration cursor
368 * @wq: the target workqueue
370 * This must be called either with wq->mutex held or sched RCU read locked.
371 * If the pwq needs to be used beyond the locking in effect, the caller is
372 * responsible for guaranteeing that the pwq stays online.
374 * The if/else clause exists only for the lockdep assertion and can be
377 #define for_each_pwq(pwq, wq) \
378 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
379 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
382 #ifdef CONFIG_DEBUG_OBJECTS_WORK
384 static struct debug_obj_descr work_debug_descr;
386 static void *work_debug_hint(void *addr)
388 return ((struct work_struct *) addr)->func;
392 * fixup_init is called when:
393 * - an active object is initialized
395 static int work_fixup_init(void *addr, enum debug_obj_state state)
397 struct work_struct *work = addr;
400 case ODEBUG_STATE_ACTIVE:
401 cancel_work_sync(work);
402 debug_object_init(work, &work_debug_descr);
410 * fixup_activate is called when:
411 * - an active object is activated
412 * - an unknown object is activated (might be a statically initialized object)
414 static int work_fixup_activate(void *addr, enum debug_obj_state state)
416 struct work_struct *work = addr;
420 case ODEBUG_STATE_NOTAVAILABLE:
422 * This is not really a fixup. The work struct was
423 * statically initialized. We just make sure that it
424 * is tracked in the object tracker.
426 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
427 debug_object_init(work, &work_debug_descr);
428 debug_object_activate(work, &work_debug_descr);
434 case ODEBUG_STATE_ACTIVE:
443 * fixup_free is called when:
444 * - an active object is freed
446 static int work_fixup_free(void *addr, enum debug_obj_state state)
448 struct work_struct *work = addr;
451 case ODEBUG_STATE_ACTIVE:
452 cancel_work_sync(work);
453 debug_object_free(work, &work_debug_descr);
460 static struct debug_obj_descr work_debug_descr = {
461 .name = "work_struct",
462 .debug_hint = work_debug_hint,
463 .fixup_init = work_fixup_init,
464 .fixup_activate = work_fixup_activate,
465 .fixup_free = work_fixup_free,
468 static inline void debug_work_activate(struct work_struct *work)
470 debug_object_activate(work, &work_debug_descr);
473 static inline void debug_work_deactivate(struct work_struct *work)
475 debug_object_deactivate(work, &work_debug_descr);
478 void __init_work(struct work_struct *work, int onstack)
481 debug_object_init_on_stack(work, &work_debug_descr);
483 debug_object_init(work, &work_debug_descr);
485 EXPORT_SYMBOL_GPL(__init_work);
487 void destroy_work_on_stack(struct work_struct *work)
489 debug_object_free(work, &work_debug_descr);
491 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
494 static inline void debug_work_activate(struct work_struct *work) { }
495 static inline void debug_work_deactivate(struct work_struct *work) { }
498 /* allocate ID and assign it to @pool */
499 static int worker_pool_assign_id(struct worker_pool *pool)
503 lockdep_assert_held(&wq_pool_mutex);
506 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
508 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
509 } while (ret == -EAGAIN);
515 * first_pwq - return the first pool_workqueue of the specified workqueue
516 * @wq: the target workqueue
518 * This must be called either with wq->mutex held or sched RCU read locked.
519 * If the pwq needs to be used beyond the locking in effect, the caller is
520 * responsible for guaranteeing that the pwq stays online.
522 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
524 assert_rcu_or_wq_mutex(wq);
525 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
530 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
531 * @wq: the target workqueue
534 * This must be called either with pwq_lock held or sched RCU read locked.
535 * If the pwq needs to be used beyond the locking in effect, the caller is
536 * responsible for guaranteeing that the pwq stays online.
538 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
541 assert_rcu_or_wq_mutex(wq);
542 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
545 static unsigned int work_color_to_flags(int color)
547 return color << WORK_STRUCT_COLOR_SHIFT;
550 static int get_work_color(struct work_struct *work)
552 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
553 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
556 static int work_next_color(int color)
558 return (color + 1) % WORK_NR_COLORS;
562 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
563 * contain the pointer to the queued pwq. Once execution starts, the flag
564 * is cleared and the high bits contain OFFQ flags and pool ID.
566 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
567 * and clear_work_data() can be used to set the pwq, pool or clear
568 * work->data. These functions should only be called while the work is
569 * owned - ie. while the PENDING bit is set.
571 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
572 * corresponding to a work. Pool is available once the work has been
573 * queued anywhere after initialization until it is sync canceled. pwq is
574 * available only while the work item is queued.
576 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
577 * canceled. While being canceled, a work item may have its PENDING set
578 * but stay off timer and worklist for arbitrarily long and nobody should
579 * try to steal the PENDING bit.
581 static inline void set_work_data(struct work_struct *work, unsigned long data,
584 WARN_ON_ONCE(!work_pending(work));
585 atomic_long_set(&work->data, data | flags | work_static(work));
588 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
589 unsigned long extra_flags)
591 set_work_data(work, (unsigned long)pwq,
592 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
595 static void set_work_pool_and_keep_pending(struct work_struct *work,
598 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
599 WORK_STRUCT_PENDING);
602 static void set_work_pool_and_clear_pending(struct work_struct *work,
606 * The following wmb is paired with the implied mb in
607 * test_and_set_bit(PENDING) and ensures all updates to @work made
608 * here are visible to and precede any updates by the next PENDING
612 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
615 static void clear_work_data(struct work_struct *work)
617 smp_wmb(); /* see set_work_pool_and_clear_pending() */
618 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
621 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
623 unsigned long data = atomic_long_read(&work->data);
625 if (data & WORK_STRUCT_PWQ)
626 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
632 * get_work_pool - return the worker_pool a given work was associated with
633 * @work: the work item of interest
635 * Return the worker_pool @work was last associated with. %NULL if none.
637 * Pools are created and destroyed under wq_pool_mutex, and allows read
638 * access under sched-RCU read lock. As such, this function should be
639 * called under wq_pool_mutex or with preemption disabled.
641 * All fields of the returned pool are accessible as long as the above
642 * mentioned locking is in effect. If the returned pool needs to be used
643 * beyond the critical section, the caller is responsible for ensuring the
644 * returned pool is and stays online.
646 static struct worker_pool *get_work_pool(struct work_struct *work)
648 unsigned long data = atomic_long_read(&work->data);
651 assert_rcu_or_pool_mutex();
653 if (data & WORK_STRUCT_PWQ)
654 return ((struct pool_workqueue *)
655 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
657 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
658 if (pool_id == WORK_OFFQ_POOL_NONE)
661 return idr_find(&worker_pool_idr, pool_id);
665 * get_work_pool_id - return the worker pool ID a given work is associated with
666 * @work: the work item of interest
668 * Return the worker_pool ID @work was last associated with.
669 * %WORK_OFFQ_POOL_NONE if none.
671 static int get_work_pool_id(struct work_struct *work)
673 unsigned long data = atomic_long_read(&work->data);
675 if (data & WORK_STRUCT_PWQ)
676 return ((struct pool_workqueue *)
677 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
679 return data >> WORK_OFFQ_POOL_SHIFT;
682 static void mark_work_canceling(struct work_struct *work)
684 unsigned long pool_id = get_work_pool_id(work);
686 pool_id <<= WORK_OFFQ_POOL_SHIFT;
687 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
690 static bool work_is_canceling(struct work_struct *work)
692 unsigned long data = atomic_long_read(&work->data);
694 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
698 * Policy functions. These define the policies on how the global worker
699 * pools are managed. Unless noted otherwise, these functions assume that
700 * they're being called with pool->lock held.
703 static bool __need_more_worker(struct worker_pool *pool)
705 return !atomic_read(&pool->nr_running);
709 * Need to wake up a worker? Called from anything but currently
712 * Note that, because unbound workers never contribute to nr_running, this
713 * function will always return %true for unbound pools as long as the
714 * worklist isn't empty.
716 static bool need_more_worker(struct worker_pool *pool)
718 return !list_empty(&pool->worklist) && __need_more_worker(pool);
721 /* Can I start working? Called from busy but !running workers. */
722 static bool may_start_working(struct worker_pool *pool)
724 return pool->nr_idle;
727 /* Do I need to keep working? Called from currently running workers. */
728 static bool keep_working(struct worker_pool *pool)
730 return !list_empty(&pool->worklist) &&
731 atomic_read(&pool->nr_running) <= 1;
734 /* Do we need a new worker? Called from manager. */
735 static bool need_to_create_worker(struct worker_pool *pool)
737 return need_more_worker(pool) && !may_start_working(pool);
740 /* Do I need to be the manager? */
741 static bool need_to_manage_workers(struct worker_pool *pool)
743 return need_to_create_worker(pool) ||
744 (pool->flags & POOL_MANAGE_WORKERS);
747 /* Do we have too many workers and should some go away? */
748 static bool too_many_workers(struct worker_pool *pool)
750 bool managing = mutex_is_locked(&pool->manager_arb);
751 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
752 int nr_busy = pool->nr_workers - nr_idle;
755 * nr_idle and idle_list may disagree if idle rebinding is in
756 * progress. Never return %true if idle_list is empty.
758 if (list_empty(&pool->idle_list))
761 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
768 /* Return the first worker. Safe with preemption disabled */
769 static struct worker *first_worker(struct worker_pool *pool)
771 if (unlikely(list_empty(&pool->idle_list)))
774 return list_first_entry(&pool->idle_list, struct worker, entry);
778 * wake_up_worker - wake up an idle worker
779 * @pool: worker pool to wake worker from
781 * Wake up the first idle worker of @pool.
784 * spin_lock_irq(pool->lock).
786 static void wake_up_worker(struct worker_pool *pool)
788 struct worker *worker = first_worker(pool);
791 wake_up_process(worker->task);
795 * wq_worker_waking_up - a worker is waking up
796 * @task: task waking up
797 * @cpu: CPU @task is waking up to
799 * This function is called during try_to_wake_up() when a worker is
803 * spin_lock_irq(rq->lock)
805 void wq_worker_waking_up(struct task_struct *task, int cpu)
807 struct worker *worker = kthread_data(task);
809 if (!(worker->flags & WORKER_NOT_RUNNING)) {
810 WARN_ON_ONCE(worker->pool->cpu != cpu);
811 atomic_inc(&worker->pool->nr_running);
816 * wq_worker_sleeping - a worker is going to sleep
817 * @task: task going to sleep
818 * @cpu: CPU in question, must be the current CPU number
820 * This function is called during schedule() when a busy worker is
821 * going to sleep. Worker on the same cpu can be woken up by
822 * returning pointer to its task.
825 * spin_lock_irq(rq->lock)
828 * Worker task on @cpu to wake up, %NULL if none.
830 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
832 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
833 struct worker_pool *pool;
836 * Rescuers, which may not have all the fields set up like normal
837 * workers, also reach here, let's not access anything before
838 * checking NOT_RUNNING.
840 if (worker->flags & WORKER_NOT_RUNNING)
845 /* this can only happen on the local cpu */
846 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
850 * The counterpart of the following dec_and_test, implied mb,
851 * worklist not empty test sequence is in insert_work().
852 * Please read comment there.
854 * NOT_RUNNING is clear. This means that we're bound to and
855 * running on the local cpu w/ rq lock held and preemption
856 * disabled, which in turn means that none else could be
857 * manipulating idle_list, so dereferencing idle_list without pool
860 if (atomic_dec_and_test(&pool->nr_running) &&
861 !list_empty(&pool->worklist))
862 to_wakeup = first_worker(pool);
863 return to_wakeup ? to_wakeup->task : NULL;
867 * worker_set_flags - set worker flags and adjust nr_running accordingly
869 * @flags: flags to set
870 * @wakeup: wakeup an idle worker if necessary
872 * Set @flags in @worker->flags and adjust nr_running accordingly. If
873 * nr_running becomes zero and @wakeup is %true, an idle worker is
877 * spin_lock_irq(pool->lock)
879 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
882 struct worker_pool *pool = worker->pool;
884 WARN_ON_ONCE(worker->task != current);
887 * If transitioning into NOT_RUNNING, adjust nr_running and
888 * wake up an idle worker as necessary if requested by
891 if ((flags & WORKER_NOT_RUNNING) &&
892 !(worker->flags & WORKER_NOT_RUNNING)) {
894 if (atomic_dec_and_test(&pool->nr_running) &&
895 !list_empty(&pool->worklist))
896 wake_up_worker(pool);
898 atomic_dec(&pool->nr_running);
901 worker->flags |= flags;
905 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
907 * @flags: flags to clear
909 * Clear @flags in @worker->flags and adjust nr_running accordingly.
912 * spin_lock_irq(pool->lock)
914 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
916 struct worker_pool *pool = worker->pool;
917 unsigned int oflags = worker->flags;
919 WARN_ON_ONCE(worker->task != current);
921 worker->flags &= ~flags;
924 * If transitioning out of NOT_RUNNING, increment nr_running. Note
925 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
926 * of multiple flags, not a single flag.
928 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
929 if (!(worker->flags & WORKER_NOT_RUNNING))
930 atomic_inc(&pool->nr_running);
934 * find_worker_executing_work - find worker which is executing a work
935 * @pool: pool of interest
936 * @work: work to find worker for
938 * Find a worker which is executing @work on @pool by searching
939 * @pool->busy_hash which is keyed by the address of @work. For a worker
940 * to match, its current execution should match the address of @work and
941 * its work function. This is to avoid unwanted dependency between
942 * unrelated work executions through a work item being recycled while still
945 * This is a bit tricky. A work item may be freed once its execution
946 * starts and nothing prevents the freed area from being recycled for
947 * another work item. If the same work item address ends up being reused
948 * before the original execution finishes, workqueue will identify the
949 * recycled work item as currently executing and make it wait until the
950 * current execution finishes, introducing an unwanted dependency.
952 * This function checks the work item address and work function to avoid
953 * false positives. Note that this isn't complete as one may construct a
954 * work function which can introduce dependency onto itself through a
955 * recycled work item. Well, if somebody wants to shoot oneself in the
956 * foot that badly, there's only so much we can do, and if such deadlock
957 * actually occurs, it should be easy to locate the culprit work function.
960 * spin_lock_irq(pool->lock).
963 * Pointer to worker which is executing @work if found, NULL
966 static struct worker *find_worker_executing_work(struct worker_pool *pool,
967 struct work_struct *work)
969 struct worker *worker;
971 hash_for_each_possible(pool->busy_hash, worker, hentry,
973 if (worker->current_work == work &&
974 worker->current_func == work->func)
981 * move_linked_works - move linked works to a list
982 * @work: start of series of works to be scheduled
983 * @head: target list to append @work to
984 * @nextp: out paramter for nested worklist walking
986 * Schedule linked works starting from @work to @head. Work series to
987 * be scheduled starts at @work and includes any consecutive work with
988 * WORK_STRUCT_LINKED set in its predecessor.
990 * If @nextp is not NULL, it's updated to point to the next work of
991 * the last scheduled work. This allows move_linked_works() to be
992 * nested inside outer list_for_each_entry_safe().
995 * spin_lock_irq(pool->lock).
997 static void move_linked_works(struct work_struct *work, struct list_head *head,
998 struct work_struct **nextp)
1000 struct work_struct *n;
1003 * Linked worklist will always end before the end of the list,
1004 * use NULL for list head.
1006 list_for_each_entry_safe_from(work, n, NULL, entry) {
1007 list_move_tail(&work->entry, head);
1008 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1013 * If we're already inside safe list traversal and have moved
1014 * multiple works to the scheduled queue, the next position
1015 * needs to be updated.
1022 * get_pwq - get an extra reference on the specified pool_workqueue
1023 * @pwq: pool_workqueue to get
1025 * Obtain an extra reference on @pwq. The caller should guarantee that
1026 * @pwq has positive refcnt and be holding the matching pool->lock.
1028 static void get_pwq(struct pool_workqueue *pwq)
1030 lockdep_assert_held(&pwq->pool->lock);
1031 WARN_ON_ONCE(pwq->refcnt <= 0);
1036 * put_pwq - put a pool_workqueue reference
1037 * @pwq: pool_workqueue to put
1039 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1040 * destruction. The caller should be holding the matching pool->lock.
1042 static void put_pwq(struct pool_workqueue *pwq)
1044 lockdep_assert_held(&pwq->pool->lock);
1045 if (likely(--pwq->refcnt))
1047 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1050 * @pwq can't be released under pool->lock, bounce to
1051 * pwq_unbound_release_workfn(). This never recurses on the same
1052 * pool->lock as this path is taken only for unbound workqueues and
1053 * the release work item is scheduled on a per-cpu workqueue. To
1054 * avoid lockdep warning, unbound pool->locks are given lockdep
1055 * subclass of 1 in get_unbound_pool().
1057 schedule_work(&pwq->unbound_release_work);
1060 static void pwq_activate_delayed_work(struct work_struct *work)
1062 struct pool_workqueue *pwq = get_work_pwq(work);
1064 trace_workqueue_activate_work(work);
1065 move_linked_works(work, &pwq->pool->worklist, NULL);
1066 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1070 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1072 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1073 struct work_struct, entry);
1075 pwq_activate_delayed_work(work);
1079 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1080 * @pwq: pwq of interest
1081 * @color: color of work which left the queue
1083 * A work either has completed or is removed from pending queue,
1084 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1087 * spin_lock_irq(pool->lock).
1089 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1091 /* uncolored work items don't participate in flushing or nr_active */
1092 if (color == WORK_NO_COLOR)
1095 pwq->nr_in_flight[color]--;
1098 if (!list_empty(&pwq->delayed_works)) {
1099 /* one down, submit a delayed one */
1100 if (pwq->nr_active < pwq->max_active)
1101 pwq_activate_first_delayed(pwq);
1104 /* is flush in progress and are we at the flushing tip? */
1105 if (likely(pwq->flush_color != color))
1108 /* are there still in-flight works? */
1109 if (pwq->nr_in_flight[color])
1112 /* this pwq is done, clear flush_color */
1113 pwq->flush_color = -1;
1116 * If this was the last pwq, wake up the first flusher. It
1117 * will handle the rest.
1119 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1120 complete(&pwq->wq->first_flusher->done);
1126 * try_to_grab_pending - steal work item from worklist and disable irq
1127 * @work: work item to steal
1128 * @is_dwork: @work is a delayed_work
1129 * @flags: place to store irq state
1131 * Try to grab PENDING bit of @work. This function can handle @work in any
1132 * stable state - idle, on timer or on worklist. Return values are
1134 * 1 if @work was pending and we successfully stole PENDING
1135 * 0 if @work was idle and we claimed PENDING
1136 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1137 * -ENOENT if someone else is canceling @work, this state may persist
1138 * for arbitrarily long
1140 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1141 * interrupted while holding PENDING and @work off queue, irq must be
1142 * disabled on entry. This, combined with delayed_work->timer being
1143 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1145 * On successful return, >= 0, irq is disabled and the caller is
1146 * responsible for releasing it using local_irq_restore(*@flags).
1148 * This function is safe to call from any context including IRQ handler.
1150 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1151 unsigned long *flags)
1153 struct worker_pool *pool;
1154 struct pool_workqueue *pwq;
1156 local_irq_save(*flags);
1158 /* try to steal the timer if it exists */
1160 struct delayed_work *dwork = to_delayed_work(work);
1163 * dwork->timer is irqsafe. If del_timer() fails, it's
1164 * guaranteed that the timer is not queued anywhere and not
1165 * running on the local CPU.
1167 if (likely(del_timer(&dwork->timer)))
1171 /* try to claim PENDING the normal way */
1172 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1176 * The queueing is in progress, or it is already queued. Try to
1177 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1179 pool = get_work_pool(work);
1183 spin_lock(&pool->lock);
1185 * work->data is guaranteed to point to pwq only while the work
1186 * item is queued on pwq->wq, and both updating work->data to point
1187 * to pwq on queueing and to pool on dequeueing are done under
1188 * pwq->pool->lock. This in turn guarantees that, if work->data
1189 * points to pwq which is associated with a locked pool, the work
1190 * item is currently queued on that pool.
1192 pwq = get_work_pwq(work);
1193 if (pwq && pwq->pool == pool) {
1194 debug_work_deactivate(work);
1197 * A delayed work item cannot be grabbed directly because
1198 * it might have linked NO_COLOR work items which, if left
1199 * on the delayed_list, will confuse pwq->nr_active
1200 * management later on and cause stall. Make sure the work
1201 * item is activated before grabbing.
1203 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1204 pwq_activate_delayed_work(work);
1206 list_del_init(&work->entry);
1207 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1209 /* work->data points to pwq iff queued, point to pool */
1210 set_work_pool_and_keep_pending(work, pool->id);
1212 spin_unlock(&pool->lock);
1215 spin_unlock(&pool->lock);
1217 local_irq_restore(*flags);
1218 if (work_is_canceling(work))
1225 * insert_work - insert a work into a pool
1226 * @pwq: pwq @work belongs to
1227 * @work: work to insert
1228 * @head: insertion point
1229 * @extra_flags: extra WORK_STRUCT_* flags to set
1231 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1232 * work_struct flags.
1235 * spin_lock_irq(pool->lock).
1237 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1238 struct list_head *head, unsigned int extra_flags)
1240 struct worker_pool *pool = pwq->pool;
1242 /* we own @work, set data and link */
1243 set_work_pwq(work, pwq, extra_flags);
1244 list_add_tail(&work->entry, head);
1248 * Ensure either wq_worker_sleeping() sees the above
1249 * list_add_tail() or we see zero nr_running to avoid workers lying
1250 * around lazily while there are works to be processed.
1254 if (__need_more_worker(pool))
1255 wake_up_worker(pool);
1259 * Test whether @work is being queued from another work executing on the
1262 static bool is_chained_work(struct workqueue_struct *wq)
1264 struct worker *worker;
1266 worker = current_wq_worker();
1268 * Return %true iff I'm a worker execuing a work item on @wq. If
1269 * I'm @worker, it's safe to dereference it without locking.
1271 return worker && worker->current_pwq->wq == wq;
1274 static void __queue_work(int cpu, struct workqueue_struct *wq,
1275 struct work_struct *work)
1277 struct pool_workqueue *pwq;
1278 struct worker_pool *last_pool;
1279 struct list_head *worklist;
1280 unsigned int work_flags;
1281 unsigned int req_cpu = cpu;
1284 * While a work item is PENDING && off queue, a task trying to
1285 * steal the PENDING will busy-loop waiting for it to either get
1286 * queued or lose PENDING. Grabbing PENDING and queueing should
1287 * happen with IRQ disabled.
1289 WARN_ON_ONCE(!irqs_disabled());
1291 debug_work_activate(work);
1293 /* if dying, only works from the same workqueue are allowed */
1294 if (unlikely(wq->flags & __WQ_DRAINING) &&
1295 WARN_ON_ONCE(!is_chained_work(wq)))
1298 if (req_cpu == WORK_CPU_UNBOUND)
1299 cpu = raw_smp_processor_id();
1301 /* pwq which will be used unless @work is executing elsewhere */
1302 if (!(wq->flags & WQ_UNBOUND))
1303 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1305 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1308 * If @work was previously on a different pool, it might still be
1309 * running there, in which case the work needs to be queued on that
1310 * pool to guarantee non-reentrancy.
1312 last_pool = get_work_pool(work);
1313 if (last_pool && last_pool != pwq->pool) {
1314 struct worker *worker;
1316 spin_lock(&last_pool->lock);
1318 worker = find_worker_executing_work(last_pool, work);
1320 if (worker && worker->current_pwq->wq == wq) {
1321 pwq = worker->current_pwq;
1323 /* meh... not running there, queue here */
1324 spin_unlock(&last_pool->lock);
1325 spin_lock(&pwq->pool->lock);
1328 spin_lock(&pwq->pool->lock);
1332 * pwq is determined and locked. For unbound pools, we could have
1333 * raced with pwq release and it could already be dead. If its
1334 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1335 * without another pwq replacing it in the numa_pwq_tbl or while
1336 * work items are executing on it, so the retrying is guaranteed to
1337 * make forward-progress.
1339 if (unlikely(!pwq->refcnt)) {
1340 if (wq->flags & WQ_UNBOUND) {
1341 spin_unlock(&pwq->pool->lock);
1346 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1350 /* pwq determined, queue */
1351 trace_workqueue_queue_work(req_cpu, pwq, work);
1353 if (WARN_ON(!list_empty(&work->entry))) {
1354 spin_unlock(&pwq->pool->lock);
1358 pwq->nr_in_flight[pwq->work_color]++;
1359 work_flags = work_color_to_flags(pwq->work_color);
1361 if (likely(pwq->nr_active < pwq->max_active)) {
1362 trace_workqueue_activate_work(work);
1364 worklist = &pwq->pool->worklist;
1366 work_flags |= WORK_STRUCT_DELAYED;
1367 worklist = &pwq->delayed_works;
1370 insert_work(pwq, work, worklist, work_flags);
1372 spin_unlock(&pwq->pool->lock);
1376 * queue_work_on - queue work on specific cpu
1377 * @cpu: CPU number to execute work on
1378 * @wq: workqueue to use
1379 * @work: work to queue
1381 * Returns %false if @work was already on a queue, %true otherwise.
1383 * We queue the work to a specific CPU, the caller must ensure it
1386 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1387 struct work_struct *work)
1390 unsigned long flags;
1392 local_irq_save(flags);
1394 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1395 __queue_work(cpu, wq, work);
1399 local_irq_restore(flags);
1402 EXPORT_SYMBOL_GPL(queue_work_on);
1404 void delayed_work_timer_fn(unsigned long __data)
1406 struct delayed_work *dwork = (struct delayed_work *)__data;
1408 /* should have been called from irqsafe timer with irq already off */
1409 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1411 EXPORT_SYMBOL(delayed_work_timer_fn);
1413 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1414 struct delayed_work *dwork, unsigned long delay)
1416 struct timer_list *timer = &dwork->timer;
1417 struct work_struct *work = &dwork->work;
1419 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1420 timer->data != (unsigned long)dwork);
1421 WARN_ON_ONCE(timer_pending(timer));
1422 WARN_ON_ONCE(!list_empty(&work->entry));
1425 * If @delay is 0, queue @dwork->work immediately. This is for
1426 * both optimization and correctness. The earliest @timer can
1427 * expire is on the closest next tick and delayed_work users depend
1428 * on that there's no such delay when @delay is 0.
1431 __queue_work(cpu, wq, &dwork->work);
1435 timer_stats_timer_set_start_info(&dwork->timer);
1439 timer->expires = jiffies + delay;
1441 if (unlikely(cpu != WORK_CPU_UNBOUND))
1442 add_timer_on(timer, cpu);
1448 * queue_delayed_work_on - queue work on specific CPU after delay
1449 * @cpu: CPU number to execute work on
1450 * @wq: workqueue to use
1451 * @dwork: work to queue
1452 * @delay: number of jiffies to wait before queueing
1454 * Returns %false if @work was already on a queue, %true otherwise. If
1455 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1458 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1459 struct delayed_work *dwork, unsigned long delay)
1461 struct work_struct *work = &dwork->work;
1463 unsigned long flags;
1465 /* read the comment in __queue_work() */
1466 local_irq_save(flags);
1468 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1469 __queue_delayed_work(cpu, wq, dwork, delay);
1473 local_irq_restore(flags);
1476 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1479 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1480 * @cpu: CPU number to execute work on
1481 * @wq: workqueue to use
1482 * @dwork: work to queue
1483 * @delay: number of jiffies to wait before queueing
1485 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1486 * modify @dwork's timer so that it expires after @delay. If @delay is
1487 * zero, @work is guaranteed to be scheduled immediately regardless of its
1490 * Returns %false if @dwork was idle and queued, %true if @dwork was
1491 * pending and its timer was modified.
1493 * This function is safe to call from any context including IRQ handler.
1494 * See try_to_grab_pending() for details.
1496 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1497 struct delayed_work *dwork, unsigned long delay)
1499 unsigned long flags;
1503 ret = try_to_grab_pending(&dwork->work, true, &flags);
1504 } while (unlikely(ret == -EAGAIN));
1506 if (likely(ret >= 0)) {
1507 __queue_delayed_work(cpu, wq, dwork, delay);
1508 local_irq_restore(flags);
1511 /* -ENOENT from try_to_grab_pending() becomes %true */
1514 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1517 * worker_enter_idle - enter idle state
1518 * @worker: worker which is entering idle state
1520 * @worker is entering idle state. Update stats and idle timer if
1524 * spin_lock_irq(pool->lock).
1526 static void worker_enter_idle(struct worker *worker)
1528 struct worker_pool *pool = worker->pool;
1530 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1531 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1532 (worker->hentry.next || worker->hentry.pprev)))
1535 /* can't use worker_set_flags(), also called from start_worker() */
1536 worker->flags |= WORKER_IDLE;
1538 worker->last_active = jiffies;
1540 /* idle_list is LIFO */
1541 list_add(&worker->entry, &pool->idle_list);
1543 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1544 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1547 * Sanity check nr_running. Because wq_unbind_fn() releases
1548 * pool->lock between setting %WORKER_UNBOUND and zapping
1549 * nr_running, the warning may trigger spuriously. Check iff
1550 * unbind is not in progress.
1552 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1553 pool->nr_workers == pool->nr_idle &&
1554 atomic_read(&pool->nr_running));
1558 * worker_leave_idle - leave idle state
1559 * @worker: worker which is leaving idle state
1561 * @worker is leaving idle state. Update stats.
1564 * spin_lock_irq(pool->lock).
1566 static void worker_leave_idle(struct worker *worker)
1568 struct worker_pool *pool = worker->pool;
1570 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1572 worker_clr_flags(worker, WORKER_IDLE);
1574 list_del_init(&worker->entry);
1578 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1579 * @pool: target worker_pool
1581 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1583 * Works which are scheduled while the cpu is online must at least be
1584 * scheduled to a worker which is bound to the cpu so that if they are
1585 * flushed from cpu callbacks while cpu is going down, they are
1586 * guaranteed to execute on the cpu.
1588 * This function is to be used by unbound workers and rescuers to bind
1589 * themselves to the target cpu and may race with cpu going down or
1590 * coming online. kthread_bind() can't be used because it may put the
1591 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1592 * verbatim as it's best effort and blocking and pool may be
1593 * [dis]associated in the meantime.
1595 * This function tries set_cpus_allowed() and locks pool and verifies the
1596 * binding against %POOL_DISASSOCIATED which is set during
1597 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1598 * enters idle state or fetches works without dropping lock, it can
1599 * guarantee the scheduling requirement described in the first paragraph.
1602 * Might sleep. Called without any lock but returns with pool->lock
1606 * %true if the associated pool is online (@worker is successfully
1607 * bound), %false if offline.
1609 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1610 __acquires(&pool->lock)
1614 * The following call may fail, succeed or succeed
1615 * without actually migrating the task to the cpu if
1616 * it races with cpu hotunplug operation. Verify
1617 * against POOL_DISASSOCIATED.
1619 if (!(pool->flags & POOL_DISASSOCIATED))
1620 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1622 spin_lock_irq(&pool->lock);
1623 if (pool->flags & POOL_DISASSOCIATED)
1625 if (task_cpu(current) == pool->cpu &&
1626 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1628 spin_unlock_irq(&pool->lock);
1631 * We've raced with CPU hot[un]plug. Give it a breather
1632 * and retry migration. cond_resched() is required here;
1633 * otherwise, we might deadlock against cpu_stop trying to
1634 * bring down the CPU on non-preemptive kernel.
1641 static struct worker *alloc_worker(void)
1643 struct worker *worker;
1645 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1647 INIT_LIST_HEAD(&worker->entry);
1648 INIT_LIST_HEAD(&worker->scheduled);
1649 /* on creation a worker is in !idle && prep state */
1650 worker->flags = WORKER_PREP;
1656 * create_worker - create a new workqueue worker
1657 * @pool: pool the new worker will belong to
1659 * Create a new worker which is bound to @pool. The returned worker
1660 * can be started by calling start_worker() or destroyed using
1664 * Might sleep. Does GFP_KERNEL allocations.
1667 * Pointer to the newly created worker.
1669 static struct worker *create_worker(struct worker_pool *pool)
1671 struct worker *worker = NULL;
1675 lockdep_assert_held(&pool->manager_mutex);
1678 * ID is needed to determine kthread name. Allocate ID first
1679 * without installing the pointer.
1681 idr_preload(GFP_KERNEL);
1682 spin_lock_irq(&pool->lock);
1684 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1686 spin_unlock_irq(&pool->lock);
1691 worker = alloc_worker();
1695 worker->pool = pool;
1699 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1700 pool->attrs->nice < 0 ? "H" : "");
1702 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1704 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1705 "kworker/%s", id_buf);
1706 if (IS_ERR(worker->task))
1710 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1711 * online CPUs. It'll be re-applied when any of the CPUs come up.
1713 set_user_nice(worker->task, pool->attrs->nice);
1714 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1716 /* prevent userland from meddling with cpumask of workqueue workers */
1717 worker->task->flags |= PF_NO_SETAFFINITY;
1720 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1721 * remains stable across this function. See the comments above the
1722 * flag definition for details.
1724 if (pool->flags & POOL_DISASSOCIATED)
1725 worker->flags |= WORKER_UNBOUND;
1727 /* successful, commit the pointer to idr */
1728 spin_lock_irq(&pool->lock);
1729 idr_replace(&pool->worker_idr, worker, worker->id);
1730 spin_unlock_irq(&pool->lock);
1736 spin_lock_irq(&pool->lock);
1737 idr_remove(&pool->worker_idr, id);
1738 spin_unlock_irq(&pool->lock);
1745 * start_worker - start a newly created worker
1746 * @worker: worker to start
1748 * Make the pool aware of @worker and start it.
1751 * spin_lock_irq(pool->lock).
1753 static void start_worker(struct worker *worker)
1755 worker->flags |= WORKER_STARTED;
1756 worker->pool->nr_workers++;
1757 worker_enter_idle(worker);
1758 wake_up_process(worker->task);
1762 * create_and_start_worker - create and start a worker for a pool
1763 * @pool: the target pool
1765 * Grab the managership of @pool and create and start a new worker for it.
1767 static int create_and_start_worker(struct worker_pool *pool)
1769 struct worker *worker;
1771 mutex_lock(&pool->manager_mutex);
1773 worker = create_worker(pool);
1775 spin_lock_irq(&pool->lock);
1776 start_worker(worker);
1777 spin_unlock_irq(&pool->lock);
1780 mutex_unlock(&pool->manager_mutex);
1782 return worker ? 0 : -ENOMEM;
1786 * destroy_worker - destroy a workqueue worker
1787 * @worker: worker to be destroyed
1789 * Destroy @worker and adjust @pool stats accordingly.
1792 * spin_lock_irq(pool->lock) which is released and regrabbed.
1794 static void destroy_worker(struct worker *worker)
1796 struct worker_pool *pool = worker->pool;
1798 lockdep_assert_held(&pool->manager_mutex);
1799 lockdep_assert_held(&pool->lock);
1801 /* sanity check frenzy */
1802 if (WARN_ON(worker->current_work) ||
1803 WARN_ON(!list_empty(&worker->scheduled)))
1806 if (worker->flags & WORKER_STARTED)
1808 if (worker->flags & WORKER_IDLE)
1811 list_del_init(&worker->entry);
1812 worker->flags |= WORKER_DIE;
1814 idr_remove(&pool->worker_idr, worker->id);
1816 spin_unlock_irq(&pool->lock);
1818 kthread_stop(worker->task);
1821 spin_lock_irq(&pool->lock);
1824 static void idle_worker_timeout(unsigned long __pool)
1826 struct worker_pool *pool = (void *)__pool;
1828 spin_lock_irq(&pool->lock);
1830 if (too_many_workers(pool)) {
1831 struct worker *worker;
1832 unsigned long expires;
1834 /* idle_list is kept in LIFO order, check the last one */
1835 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1836 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1838 if (time_before(jiffies, expires))
1839 mod_timer(&pool->idle_timer, expires);
1841 /* it's been idle for too long, wake up manager */
1842 pool->flags |= POOL_MANAGE_WORKERS;
1843 wake_up_worker(pool);
1847 spin_unlock_irq(&pool->lock);
1850 static void send_mayday(struct work_struct *work)
1852 struct pool_workqueue *pwq = get_work_pwq(work);
1853 struct workqueue_struct *wq = pwq->wq;
1855 lockdep_assert_held(&wq_mayday_lock);
1860 /* mayday mayday mayday */
1861 if (list_empty(&pwq->mayday_node)) {
1862 list_add_tail(&pwq->mayday_node, &wq->maydays);
1863 wake_up_process(wq->rescuer->task);
1867 static void pool_mayday_timeout(unsigned long __pool)
1869 struct worker_pool *pool = (void *)__pool;
1870 struct work_struct *work;
1872 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1873 spin_lock(&pool->lock);
1875 if (need_to_create_worker(pool)) {
1877 * We've been trying to create a new worker but
1878 * haven't been successful. We might be hitting an
1879 * allocation deadlock. Send distress signals to
1882 list_for_each_entry(work, &pool->worklist, entry)
1886 spin_unlock(&pool->lock);
1887 spin_unlock_irq(&wq_mayday_lock);
1889 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1893 * maybe_create_worker - create a new worker if necessary
1894 * @pool: pool to create a new worker for
1896 * Create a new worker for @pool if necessary. @pool is guaranteed to
1897 * have at least one idle worker on return from this function. If
1898 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1899 * sent to all rescuers with works scheduled on @pool to resolve
1900 * possible allocation deadlock.
1902 * On return, need_to_create_worker() is guaranteed to be %false and
1903 * may_start_working() %true.
1906 * spin_lock_irq(pool->lock) which may be released and regrabbed
1907 * multiple times. Does GFP_KERNEL allocations. Called only from
1911 * %false if no action was taken and pool->lock stayed locked, %true
1914 static bool maybe_create_worker(struct worker_pool *pool)
1915 __releases(&pool->lock)
1916 __acquires(&pool->lock)
1918 if (!need_to_create_worker(pool))
1921 spin_unlock_irq(&pool->lock);
1923 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1924 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1927 struct worker *worker;
1929 worker = create_worker(pool);
1931 del_timer_sync(&pool->mayday_timer);
1932 spin_lock_irq(&pool->lock);
1933 start_worker(worker);
1934 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1939 if (!need_to_create_worker(pool))
1942 __set_current_state(TASK_INTERRUPTIBLE);
1943 schedule_timeout(CREATE_COOLDOWN);
1945 if (!need_to_create_worker(pool))
1949 del_timer_sync(&pool->mayday_timer);
1950 spin_lock_irq(&pool->lock);
1951 if (need_to_create_worker(pool))
1957 * maybe_destroy_worker - destroy workers which have been idle for a while
1958 * @pool: pool to destroy workers for
1960 * Destroy @pool workers which have been idle for longer than
1961 * IDLE_WORKER_TIMEOUT.
1964 * spin_lock_irq(pool->lock) which may be released and regrabbed
1965 * multiple times. Called only from manager.
1968 * %false if no action was taken and pool->lock stayed locked, %true
1971 static bool maybe_destroy_workers(struct worker_pool *pool)
1975 while (too_many_workers(pool)) {
1976 struct worker *worker;
1977 unsigned long expires;
1979 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1980 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1982 if (time_before(jiffies, expires)) {
1983 mod_timer(&pool->idle_timer, expires);
1987 destroy_worker(worker);
1995 * manage_workers - manage worker pool
1998 * Assume the manager role and manage the worker pool @worker belongs
1999 * to. At any given time, there can be only zero or one manager per
2000 * pool. The exclusion is handled automatically by this function.
2002 * The caller can safely start processing works on false return. On
2003 * true return, it's guaranteed that need_to_create_worker() is false
2004 * and may_start_working() is true.
2007 * spin_lock_irq(pool->lock) which may be released and regrabbed
2008 * multiple times. Does GFP_KERNEL allocations.
2011 * spin_lock_irq(pool->lock) which may be released and regrabbed
2012 * multiple times. Does GFP_KERNEL allocations.
2014 static bool manage_workers(struct worker *worker)
2016 struct worker_pool *pool = worker->pool;
2020 * Managership is governed by two mutexes - manager_arb and
2021 * manager_mutex. manager_arb handles arbitration of manager role.
2022 * Anyone who successfully grabs manager_arb wins the arbitration
2023 * and becomes the manager. mutex_trylock() on pool->manager_arb
2024 * failure while holding pool->lock reliably indicates that someone
2025 * else is managing the pool and the worker which failed trylock
2026 * can proceed to executing work items. This means that anyone
2027 * grabbing manager_arb is responsible for actually performing
2028 * manager duties. If manager_arb is grabbed and released without
2029 * actual management, the pool may stall indefinitely.
2031 * manager_mutex is used for exclusion of actual management
2032 * operations. The holder of manager_mutex can be sure that none
2033 * of management operations, including creation and destruction of
2034 * workers, won't take place until the mutex is released. Because
2035 * manager_mutex doesn't interfere with manager role arbitration,
2036 * it is guaranteed that the pool's management, while may be
2037 * delayed, won't be disturbed by someone else grabbing
2040 if (!mutex_trylock(&pool->manager_arb))
2044 * With manager arbitration won, manager_mutex would be free in
2045 * most cases. trylock first without dropping @pool->lock.
2047 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2048 spin_unlock_irq(&pool->lock);
2049 mutex_lock(&pool->manager_mutex);
2053 pool->flags &= ~POOL_MANAGE_WORKERS;
2056 * Destroy and then create so that may_start_working() is true
2059 ret |= maybe_destroy_workers(pool);
2060 ret |= maybe_create_worker(pool);
2062 mutex_unlock(&pool->manager_mutex);
2063 mutex_unlock(&pool->manager_arb);
2068 * process_one_work - process single work
2070 * @work: work to process
2072 * Process @work. This function contains all the logics necessary to
2073 * process a single work including synchronization against and
2074 * interaction with other workers on the same cpu, queueing and
2075 * flushing. As long as context requirement is met, any worker can
2076 * call this function to process a work.
2079 * spin_lock_irq(pool->lock) which is released and regrabbed.
2081 static void process_one_work(struct worker *worker, struct work_struct *work)
2082 __releases(&pool->lock)
2083 __acquires(&pool->lock)
2085 struct pool_workqueue *pwq = get_work_pwq(work);
2086 struct worker_pool *pool = worker->pool;
2087 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2089 struct worker *collision;
2090 #ifdef CONFIG_LOCKDEP
2092 * It is permissible to free the struct work_struct from
2093 * inside the function that is called from it, this we need to
2094 * take into account for lockdep too. To avoid bogus "held
2095 * lock freed" warnings as well as problems when looking into
2096 * work->lockdep_map, make a copy and use that here.
2098 struct lockdep_map lockdep_map;
2100 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2103 * Ensure we're on the correct CPU. DISASSOCIATED test is
2104 * necessary to avoid spurious warnings from rescuers servicing the
2105 * unbound or a disassociated pool.
2107 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2108 !(pool->flags & POOL_DISASSOCIATED) &&
2109 raw_smp_processor_id() != pool->cpu);
2112 * A single work shouldn't be executed concurrently by
2113 * multiple workers on a single cpu. Check whether anyone is
2114 * already processing the work. If so, defer the work to the
2115 * currently executing one.
2117 collision = find_worker_executing_work(pool, work);
2118 if (unlikely(collision)) {
2119 move_linked_works(work, &collision->scheduled, NULL);
2123 /* claim and dequeue */
2124 debug_work_deactivate(work);
2125 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2126 worker->current_work = work;
2127 worker->current_func = work->func;
2128 worker->current_pwq = pwq;
2129 work_color = get_work_color(work);
2131 list_del_init(&work->entry);
2134 * CPU intensive works don't participate in concurrency
2135 * management. They're the scheduler's responsibility.
2137 if (unlikely(cpu_intensive))
2138 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2141 * Unbound pool isn't concurrency managed and work items should be
2142 * executed ASAP. Wake up another worker if necessary.
2144 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2145 wake_up_worker(pool);
2148 * Record the last pool and clear PENDING which should be the last
2149 * update to @work. Also, do this inside @pool->lock so that
2150 * PENDING and queued state changes happen together while IRQ is
2153 set_work_pool_and_clear_pending(work, pool->id);
2155 spin_unlock_irq(&pool->lock);
2157 lock_map_acquire_read(&pwq->wq->lockdep_map);
2158 lock_map_acquire(&lockdep_map);
2159 trace_workqueue_execute_start(work);
2160 worker->current_func(work);
2162 * While we must be careful to not use "work" after this, the trace
2163 * point will only record its address.
2165 trace_workqueue_execute_end(work);
2166 lock_map_release(&lockdep_map);
2167 lock_map_release(&pwq->wq->lockdep_map);
2169 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2170 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2171 " last function: %pf\n",
2172 current->comm, preempt_count(), task_pid_nr(current),
2173 worker->current_func);
2174 debug_show_held_locks(current);
2178 spin_lock_irq(&pool->lock);
2180 /* clear cpu intensive status */
2181 if (unlikely(cpu_intensive))
2182 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2184 /* we're done with it, release */
2185 hash_del(&worker->hentry);
2186 worker->current_work = NULL;
2187 worker->current_func = NULL;
2188 worker->current_pwq = NULL;
2189 pwq_dec_nr_in_flight(pwq, work_color);
2193 * process_scheduled_works - process scheduled works
2196 * Process all scheduled works. Please note that the scheduled list
2197 * may change while processing a work, so this function repeatedly
2198 * fetches a work from the top and executes it.
2201 * spin_lock_irq(pool->lock) which may be released and regrabbed
2204 static void process_scheduled_works(struct worker *worker)
2206 while (!list_empty(&worker->scheduled)) {
2207 struct work_struct *work = list_first_entry(&worker->scheduled,
2208 struct work_struct, entry);
2209 process_one_work(worker, work);
2214 * worker_thread - the worker thread function
2217 * The worker thread function. All workers belong to a worker_pool -
2218 * either a per-cpu one or dynamic unbound one. These workers process all
2219 * work items regardless of their specific target workqueue. The only
2220 * exception is work items which belong to workqueues with a rescuer which
2221 * will be explained in rescuer_thread().
2223 static int worker_thread(void *__worker)
2225 struct worker *worker = __worker;
2226 struct worker_pool *pool = worker->pool;
2228 /* tell the scheduler that this is a workqueue worker */
2229 worker->task->flags |= PF_WQ_WORKER;
2231 spin_lock_irq(&pool->lock);
2233 /* am I supposed to die? */
2234 if (unlikely(worker->flags & WORKER_DIE)) {
2235 spin_unlock_irq(&pool->lock);
2236 WARN_ON_ONCE(!list_empty(&worker->entry));
2237 worker->task->flags &= ~PF_WQ_WORKER;
2241 worker_leave_idle(worker);
2243 /* no more worker necessary? */
2244 if (!need_more_worker(pool))
2247 /* do we need to manage? */
2248 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2252 * ->scheduled list can only be filled while a worker is
2253 * preparing to process a work or actually processing it.
2254 * Make sure nobody diddled with it while I was sleeping.
2256 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2259 * Finish PREP stage. We're guaranteed to have at least one idle
2260 * worker or that someone else has already assumed the manager
2261 * role. This is where @worker starts participating in concurrency
2262 * management if applicable and concurrency management is restored
2263 * after being rebound. See rebind_workers() for details.
2265 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2268 struct work_struct *work =
2269 list_first_entry(&pool->worklist,
2270 struct work_struct, entry);
2272 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2273 /* optimization path, not strictly necessary */
2274 process_one_work(worker, work);
2275 if (unlikely(!list_empty(&worker->scheduled)))
2276 process_scheduled_works(worker);
2278 move_linked_works(work, &worker->scheduled, NULL);
2279 process_scheduled_works(worker);
2281 } while (keep_working(pool));
2283 worker_set_flags(worker, WORKER_PREP, false);
2285 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2289 * pool->lock is held and there's no work to process and no need to
2290 * manage, sleep. Workers are woken up only while holding
2291 * pool->lock or from local cpu, so setting the current state
2292 * before releasing pool->lock is enough to prevent losing any
2295 worker_enter_idle(worker);
2296 __set_current_state(TASK_INTERRUPTIBLE);
2297 spin_unlock_irq(&pool->lock);
2303 * rescuer_thread - the rescuer thread function
2306 * Workqueue rescuer thread function. There's one rescuer for each
2307 * workqueue which has WQ_MEM_RECLAIM set.
2309 * Regular work processing on a pool may block trying to create a new
2310 * worker which uses GFP_KERNEL allocation which has slight chance of
2311 * developing into deadlock if some works currently on the same queue
2312 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2313 * the problem rescuer solves.
2315 * When such condition is possible, the pool summons rescuers of all
2316 * workqueues which have works queued on the pool and let them process
2317 * those works so that forward progress can be guaranteed.
2319 * This should happen rarely.
2321 static int rescuer_thread(void *__rescuer)
2323 struct worker *rescuer = __rescuer;
2324 struct workqueue_struct *wq = rescuer->rescue_wq;
2325 struct list_head *scheduled = &rescuer->scheduled;
2327 set_user_nice(current, RESCUER_NICE_LEVEL);
2330 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2331 * doesn't participate in concurrency management.
2333 rescuer->task->flags |= PF_WQ_WORKER;
2335 set_current_state(TASK_INTERRUPTIBLE);
2337 if (kthread_should_stop()) {
2338 __set_current_state(TASK_RUNNING);
2339 rescuer->task->flags &= ~PF_WQ_WORKER;
2343 /* see whether any pwq is asking for help */
2344 spin_lock_irq(&wq_mayday_lock);
2346 while (!list_empty(&wq->maydays)) {
2347 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2348 struct pool_workqueue, mayday_node);
2349 struct worker_pool *pool = pwq->pool;
2350 struct work_struct *work, *n;
2352 __set_current_state(TASK_RUNNING);
2353 list_del_init(&pwq->mayday_node);
2355 spin_unlock_irq(&wq_mayday_lock);
2357 /* migrate to the target cpu if possible */
2358 worker_maybe_bind_and_lock(pool);
2359 rescuer->pool = pool;
2362 * Slurp in all works issued via this workqueue and
2365 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2366 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2367 if (get_work_pwq(work) == pwq)
2368 move_linked_works(work, scheduled, &n);
2370 process_scheduled_works(rescuer);
2373 * Leave this pool. If keep_working() is %true, notify a
2374 * regular worker; otherwise, we end up with 0 concurrency
2375 * and stalling the execution.
2377 if (keep_working(pool))
2378 wake_up_worker(pool);
2380 rescuer->pool = NULL;
2381 spin_unlock(&pool->lock);
2382 spin_lock(&wq_mayday_lock);
2385 spin_unlock_irq(&wq_mayday_lock);
2387 /* rescuers should never participate in concurrency management */
2388 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2394 struct work_struct work;
2395 struct completion done;
2398 static void wq_barrier_func(struct work_struct *work)
2400 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2401 complete(&barr->done);
2405 * insert_wq_barrier - insert a barrier work
2406 * @pwq: pwq to insert barrier into
2407 * @barr: wq_barrier to insert
2408 * @target: target work to attach @barr to
2409 * @worker: worker currently executing @target, NULL if @target is not executing
2411 * @barr is linked to @target such that @barr is completed only after
2412 * @target finishes execution. Please note that the ordering
2413 * guarantee is observed only with respect to @target and on the local
2416 * Currently, a queued barrier can't be canceled. This is because
2417 * try_to_grab_pending() can't determine whether the work to be
2418 * grabbed is at the head of the queue and thus can't clear LINKED
2419 * flag of the previous work while there must be a valid next work
2420 * after a work with LINKED flag set.
2422 * Note that when @worker is non-NULL, @target may be modified
2423 * underneath us, so we can't reliably determine pwq from @target.
2426 * spin_lock_irq(pool->lock).
2428 static void insert_wq_barrier(struct pool_workqueue *pwq,
2429 struct wq_barrier *barr,
2430 struct work_struct *target, struct worker *worker)
2432 struct list_head *head;
2433 unsigned int linked = 0;
2436 * debugobject calls are safe here even with pool->lock locked
2437 * as we know for sure that this will not trigger any of the
2438 * checks and call back into the fixup functions where we
2441 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2442 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2443 init_completion(&barr->done);
2446 * If @target is currently being executed, schedule the
2447 * barrier to the worker; otherwise, put it after @target.
2450 head = worker->scheduled.next;
2452 unsigned long *bits = work_data_bits(target);
2454 head = target->entry.next;
2455 /* there can already be other linked works, inherit and set */
2456 linked = *bits & WORK_STRUCT_LINKED;
2457 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2460 debug_work_activate(&barr->work);
2461 insert_work(pwq, &barr->work, head,
2462 work_color_to_flags(WORK_NO_COLOR) | linked);
2466 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2467 * @wq: workqueue being flushed
2468 * @flush_color: new flush color, < 0 for no-op
2469 * @work_color: new work color, < 0 for no-op
2471 * Prepare pwqs for workqueue flushing.
2473 * If @flush_color is non-negative, flush_color on all pwqs should be
2474 * -1. If no pwq has in-flight commands at the specified color, all
2475 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2476 * has in flight commands, its pwq->flush_color is set to
2477 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2478 * wakeup logic is armed and %true is returned.
2480 * The caller should have initialized @wq->first_flusher prior to
2481 * calling this function with non-negative @flush_color. If
2482 * @flush_color is negative, no flush color update is done and %false
2485 * If @work_color is non-negative, all pwqs should have the same
2486 * work_color which is previous to @work_color and all will be
2487 * advanced to @work_color.
2490 * mutex_lock(wq->mutex).
2493 * %true if @flush_color >= 0 and there's something to flush. %false
2496 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2497 int flush_color, int work_color)
2500 struct pool_workqueue *pwq;
2502 if (flush_color >= 0) {
2503 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2504 atomic_set(&wq->nr_pwqs_to_flush, 1);
2507 for_each_pwq(pwq, wq) {
2508 struct worker_pool *pool = pwq->pool;
2510 spin_lock_irq(&pool->lock);
2512 if (flush_color >= 0) {
2513 WARN_ON_ONCE(pwq->flush_color != -1);
2515 if (pwq->nr_in_flight[flush_color]) {
2516 pwq->flush_color = flush_color;
2517 atomic_inc(&wq->nr_pwqs_to_flush);
2522 if (work_color >= 0) {
2523 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2524 pwq->work_color = work_color;
2527 spin_unlock_irq(&pool->lock);
2530 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2531 complete(&wq->first_flusher->done);
2537 * flush_workqueue - ensure that any scheduled work has run to completion.
2538 * @wq: workqueue to flush
2540 * This function sleeps until all work items which were queued on entry
2541 * have finished execution, but it is not livelocked by new incoming ones.
2543 void flush_workqueue(struct workqueue_struct *wq)
2545 struct wq_flusher this_flusher = {
2546 .list = LIST_HEAD_INIT(this_flusher.list),
2548 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2552 lock_map_acquire(&wq->lockdep_map);
2553 lock_map_release(&wq->lockdep_map);
2555 mutex_lock(&wq->mutex);
2558 * Start-to-wait phase
2560 next_color = work_next_color(wq->work_color);
2562 if (next_color != wq->flush_color) {
2564 * Color space is not full. The current work_color
2565 * becomes our flush_color and work_color is advanced
2568 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2569 this_flusher.flush_color = wq->work_color;
2570 wq->work_color = next_color;
2572 if (!wq->first_flusher) {
2573 /* no flush in progress, become the first flusher */
2574 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2576 wq->first_flusher = &this_flusher;
2578 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2580 /* nothing to flush, done */
2581 wq->flush_color = next_color;
2582 wq->first_flusher = NULL;
2587 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2588 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2589 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2593 * Oops, color space is full, wait on overflow queue.
2594 * The next flush completion will assign us
2595 * flush_color and transfer to flusher_queue.
2597 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2600 mutex_unlock(&wq->mutex);
2602 wait_for_completion(&this_flusher.done);
2605 * Wake-up-and-cascade phase
2607 * First flushers are responsible for cascading flushes and
2608 * handling overflow. Non-first flushers can simply return.
2610 if (wq->first_flusher != &this_flusher)
2613 mutex_lock(&wq->mutex);
2615 /* we might have raced, check again with mutex held */
2616 if (wq->first_flusher != &this_flusher)
2619 wq->first_flusher = NULL;
2621 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2622 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2625 struct wq_flusher *next, *tmp;
2627 /* complete all the flushers sharing the current flush color */
2628 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2629 if (next->flush_color != wq->flush_color)
2631 list_del_init(&next->list);
2632 complete(&next->done);
2635 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2636 wq->flush_color != work_next_color(wq->work_color));
2638 /* this flush_color is finished, advance by one */
2639 wq->flush_color = work_next_color(wq->flush_color);
2641 /* one color has been freed, handle overflow queue */
2642 if (!list_empty(&wq->flusher_overflow)) {
2644 * Assign the same color to all overflowed
2645 * flushers, advance work_color and append to
2646 * flusher_queue. This is the start-to-wait
2647 * phase for these overflowed flushers.
2649 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2650 tmp->flush_color = wq->work_color;
2652 wq->work_color = work_next_color(wq->work_color);
2654 list_splice_tail_init(&wq->flusher_overflow,
2655 &wq->flusher_queue);
2656 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2659 if (list_empty(&wq->flusher_queue)) {
2660 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2665 * Need to flush more colors. Make the next flusher
2666 * the new first flusher and arm pwqs.
2668 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2669 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2671 list_del_init(&next->list);
2672 wq->first_flusher = next;
2674 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2678 * Meh... this color is already done, clear first
2679 * flusher and repeat cascading.
2681 wq->first_flusher = NULL;
2685 mutex_unlock(&wq->mutex);
2687 EXPORT_SYMBOL_GPL(flush_workqueue);
2690 * drain_workqueue - drain a workqueue
2691 * @wq: workqueue to drain
2693 * Wait until the workqueue becomes empty. While draining is in progress,
2694 * only chain queueing is allowed. IOW, only currently pending or running
2695 * work items on @wq can queue further work items on it. @wq is flushed
2696 * repeatedly until it becomes empty. The number of flushing is detemined
2697 * by the depth of chaining and should be relatively short. Whine if it
2700 void drain_workqueue(struct workqueue_struct *wq)
2702 unsigned int flush_cnt = 0;
2703 struct pool_workqueue *pwq;
2706 * __queue_work() needs to test whether there are drainers, is much
2707 * hotter than drain_workqueue() and already looks at @wq->flags.
2708 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2710 mutex_lock(&wq->mutex);
2711 if (!wq->nr_drainers++)
2712 wq->flags |= __WQ_DRAINING;
2713 mutex_unlock(&wq->mutex);
2715 flush_workqueue(wq);
2717 mutex_lock(&wq->mutex);
2719 for_each_pwq(pwq, wq) {
2722 spin_lock_irq(&pwq->pool->lock);
2723 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2724 spin_unlock_irq(&pwq->pool->lock);
2729 if (++flush_cnt == 10 ||
2730 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2731 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2732 wq->name, flush_cnt);
2734 mutex_unlock(&wq->mutex);
2738 if (!--wq->nr_drainers)
2739 wq->flags &= ~__WQ_DRAINING;
2740 mutex_unlock(&wq->mutex);
2742 EXPORT_SYMBOL_GPL(drain_workqueue);
2744 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2746 struct worker *worker = NULL;
2747 struct worker_pool *pool;
2748 struct pool_workqueue *pwq;
2752 local_irq_disable();
2753 pool = get_work_pool(work);
2759 spin_lock(&pool->lock);
2760 /* see the comment in try_to_grab_pending() with the same code */
2761 pwq = get_work_pwq(work);
2763 if (unlikely(pwq->pool != pool))
2766 worker = find_worker_executing_work(pool, work);
2769 pwq = worker->current_pwq;
2772 insert_wq_barrier(pwq, barr, work, worker);
2773 spin_unlock_irq(&pool->lock);
2776 * If @max_active is 1 or rescuer is in use, flushing another work
2777 * item on the same workqueue may lead to deadlock. Make sure the
2778 * flusher is not running on the same workqueue by verifying write
2781 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2782 lock_map_acquire(&pwq->wq->lockdep_map);
2784 lock_map_acquire_read(&pwq->wq->lockdep_map);
2785 lock_map_release(&pwq->wq->lockdep_map);
2789 spin_unlock_irq(&pool->lock);
2794 * flush_work - wait for a work to finish executing the last queueing instance
2795 * @work: the work to flush
2797 * Wait until @work has finished execution. @work is guaranteed to be idle
2798 * on return if it hasn't been requeued since flush started.
2801 * %true if flush_work() waited for the work to finish execution,
2802 * %false if it was already idle.
2804 bool flush_work(struct work_struct *work)
2806 struct wq_barrier barr;
2808 lock_map_acquire(&work->lockdep_map);
2809 lock_map_release(&work->lockdep_map);
2811 if (start_flush_work(work, &barr)) {
2812 wait_for_completion(&barr.done);
2813 destroy_work_on_stack(&barr.work);
2819 EXPORT_SYMBOL_GPL(flush_work);
2821 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2823 unsigned long flags;
2827 ret = try_to_grab_pending(work, is_dwork, &flags);
2829 * If someone else is canceling, wait for the same event it
2830 * would be waiting for before retrying.
2832 if (unlikely(ret == -ENOENT))
2834 } while (unlikely(ret < 0));
2836 /* tell other tasks trying to grab @work to back off */
2837 mark_work_canceling(work);
2838 local_irq_restore(flags);
2841 clear_work_data(work);
2846 * cancel_work_sync - cancel a work and wait for it to finish
2847 * @work: the work to cancel
2849 * Cancel @work and wait for its execution to finish. This function
2850 * can be used even if the work re-queues itself or migrates to
2851 * another workqueue. On return from this function, @work is
2852 * guaranteed to be not pending or executing on any CPU.
2854 * cancel_work_sync(&delayed_work->work) must not be used for
2855 * delayed_work's. Use cancel_delayed_work_sync() instead.
2857 * The caller must ensure that the workqueue on which @work was last
2858 * queued can't be destroyed before this function returns.
2861 * %true if @work was pending, %false otherwise.
2863 bool cancel_work_sync(struct work_struct *work)
2865 return __cancel_work_timer(work, false);
2867 EXPORT_SYMBOL_GPL(cancel_work_sync);
2870 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2871 * @dwork: the delayed work to flush
2873 * Delayed timer is cancelled and the pending work is queued for
2874 * immediate execution. Like flush_work(), this function only
2875 * considers the last queueing instance of @dwork.
2878 * %true if flush_work() waited for the work to finish execution,
2879 * %false if it was already idle.
2881 bool flush_delayed_work(struct delayed_work *dwork)
2883 local_irq_disable();
2884 if (del_timer_sync(&dwork->timer))
2885 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2887 return flush_work(&dwork->work);
2889 EXPORT_SYMBOL(flush_delayed_work);
2892 * cancel_delayed_work - cancel a delayed work
2893 * @dwork: delayed_work to cancel
2895 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2896 * and canceled; %false if wasn't pending. Note that the work callback
2897 * function may still be running on return, unless it returns %true and the
2898 * work doesn't re-arm itself. Explicitly flush or use
2899 * cancel_delayed_work_sync() to wait on it.
2901 * This function is safe to call from any context including IRQ handler.
2903 bool cancel_delayed_work(struct delayed_work *dwork)
2905 unsigned long flags;
2909 ret = try_to_grab_pending(&dwork->work, true, &flags);
2910 } while (unlikely(ret == -EAGAIN));
2912 if (unlikely(ret < 0))
2915 set_work_pool_and_clear_pending(&dwork->work,
2916 get_work_pool_id(&dwork->work));
2917 local_irq_restore(flags);
2920 EXPORT_SYMBOL(cancel_delayed_work);
2923 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2924 * @dwork: the delayed work cancel
2926 * This is cancel_work_sync() for delayed works.
2929 * %true if @dwork was pending, %false otherwise.
2931 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2933 return __cancel_work_timer(&dwork->work, true);
2935 EXPORT_SYMBOL(cancel_delayed_work_sync);
2938 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2939 * @func: the function to call
2941 * schedule_on_each_cpu() executes @func on each online CPU using the
2942 * system workqueue and blocks until all CPUs have completed.
2943 * schedule_on_each_cpu() is very slow.
2946 * 0 on success, -errno on failure.
2948 int schedule_on_each_cpu(work_func_t func)
2951 struct work_struct __percpu *works;
2953 works = alloc_percpu(struct work_struct);
2959 for_each_online_cpu(cpu) {
2960 struct work_struct *work = per_cpu_ptr(works, cpu);
2962 INIT_WORK(work, func);
2963 schedule_work_on(cpu, work);
2966 for_each_online_cpu(cpu)
2967 flush_work(per_cpu_ptr(works, cpu));
2975 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2977 * Forces execution of the kernel-global workqueue and blocks until its
2980 * Think twice before calling this function! It's very easy to get into
2981 * trouble if you don't take great care. Either of the following situations
2982 * will lead to deadlock:
2984 * One of the work items currently on the workqueue needs to acquire
2985 * a lock held by your code or its caller.
2987 * Your code is running in the context of a work routine.
2989 * They will be detected by lockdep when they occur, but the first might not
2990 * occur very often. It depends on what work items are on the workqueue and
2991 * what locks they need, which you have no control over.
2993 * In most situations flushing the entire workqueue is overkill; you merely
2994 * need to know that a particular work item isn't queued and isn't running.
2995 * In such cases you should use cancel_delayed_work_sync() or
2996 * cancel_work_sync() instead.
2998 void flush_scheduled_work(void)
3000 flush_workqueue(system_wq);
3002 EXPORT_SYMBOL(flush_scheduled_work);
3005 * execute_in_process_context - reliably execute the routine with user context
3006 * @fn: the function to execute
3007 * @ew: guaranteed storage for the execute work structure (must
3008 * be available when the work executes)
3010 * Executes the function immediately if process context is available,
3011 * otherwise schedules the function for delayed execution.
3013 * Returns: 0 - function was executed
3014 * 1 - function was scheduled for execution
3016 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3018 if (!in_interrupt()) {
3023 INIT_WORK(&ew->work, fn);
3024 schedule_work(&ew->work);
3028 EXPORT_SYMBOL_GPL(execute_in_process_context);
3032 * Workqueues with WQ_SYSFS flag set is visible to userland via
3033 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3034 * following attributes.
3036 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3037 * max_active RW int : maximum number of in-flight work items
3039 * Unbound workqueues have the following extra attributes.
3041 * id RO int : the associated pool ID
3042 * nice RW int : nice value of the workers
3043 * cpumask RW mask : bitmask of allowed CPUs for the workers
3046 struct workqueue_struct *wq;
3050 static struct workqueue_struct *dev_to_wq(struct device *dev)
3052 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3057 static ssize_t wq_per_cpu_show(struct device *dev,
3058 struct device_attribute *attr, char *buf)
3060 struct workqueue_struct *wq = dev_to_wq(dev);
3062 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3065 static ssize_t wq_max_active_show(struct device *dev,
3066 struct device_attribute *attr, char *buf)
3068 struct workqueue_struct *wq = dev_to_wq(dev);
3070 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3073 static ssize_t wq_max_active_store(struct device *dev,
3074 struct device_attribute *attr,
3075 const char *buf, size_t count)
3077 struct workqueue_struct *wq = dev_to_wq(dev);
3080 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3083 workqueue_set_max_active(wq, val);
3087 static struct device_attribute wq_sysfs_attrs[] = {
3088 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3089 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3093 static ssize_t wq_pool_id_show(struct device *dev,
3094 struct device_attribute *attr, char *buf)
3096 struct workqueue_struct *wq = dev_to_wq(dev);
3097 struct worker_pool *pool;
3100 rcu_read_lock_sched();
3101 pool = first_pwq(wq)->pool;
3102 written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id);
3103 rcu_read_unlock_sched();
3108 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3111 struct workqueue_struct *wq = dev_to_wq(dev);
3114 mutex_lock(&wq->mutex);
3115 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3116 mutex_unlock(&wq->mutex);
3121 /* prepare workqueue_attrs for sysfs store operations */
3122 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3124 struct workqueue_attrs *attrs;
3126 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3130 mutex_lock(&wq->mutex);
3131 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3132 mutex_unlock(&wq->mutex);
3136 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3137 const char *buf, size_t count)
3139 struct workqueue_struct *wq = dev_to_wq(dev);
3140 struct workqueue_attrs *attrs;
3143 attrs = wq_sysfs_prep_attrs(wq);
3147 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3148 attrs->nice >= -20 && attrs->nice <= 19)
3149 ret = apply_workqueue_attrs(wq, attrs);
3153 free_workqueue_attrs(attrs);
3154 return ret ?: count;
3157 static ssize_t wq_cpumask_show(struct device *dev,
3158 struct device_attribute *attr, char *buf)
3160 struct workqueue_struct *wq = dev_to_wq(dev);
3163 mutex_lock(&wq->mutex);
3164 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3165 mutex_unlock(&wq->mutex);
3167 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3171 static ssize_t wq_cpumask_store(struct device *dev,
3172 struct device_attribute *attr,
3173 const char *buf, size_t count)
3175 struct workqueue_struct *wq = dev_to_wq(dev);
3176 struct workqueue_attrs *attrs;
3179 attrs = wq_sysfs_prep_attrs(wq);
3183 ret = cpumask_parse(buf, attrs->cpumask);
3185 ret = apply_workqueue_attrs(wq, attrs);
3187 free_workqueue_attrs(attrs);
3188 return ret ?: count;
3191 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3192 __ATTR(pool_id, 0444, wq_pool_id_show, NULL),
3193 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3194 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3198 static struct bus_type wq_subsys = {
3199 .name = "workqueue",
3200 .dev_attrs = wq_sysfs_attrs,
3203 static int __init wq_sysfs_init(void)
3205 return subsys_virtual_register(&wq_subsys, NULL);
3207 core_initcall(wq_sysfs_init);
3209 static void wq_device_release(struct device *dev)
3211 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3217 * workqueue_sysfs_register - make a workqueue visible in sysfs
3218 * @wq: the workqueue to register
3220 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3221 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3222 * which is the preferred method.
3224 * Workqueue user should use this function directly iff it wants to apply
3225 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3226 * apply_workqueue_attrs() may race against userland updating the
3229 * Returns 0 on success, -errno on failure.
3231 int workqueue_sysfs_register(struct workqueue_struct *wq)
3233 struct wq_device *wq_dev;
3237 * Adjusting max_active or creating new pwqs by applyting
3238 * attributes breaks ordering guarantee. Disallow exposing ordered
3241 if (WARN_ON(wq->flags & __WQ_ORDERED))
3244 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3249 wq_dev->dev.bus = &wq_subsys;
3250 wq_dev->dev.init_name = wq->name;
3251 wq_dev->dev.release = wq_device_release;
3254 * unbound_attrs are created separately. Suppress uevent until
3255 * everything is ready.
3257 dev_set_uevent_suppress(&wq_dev->dev, true);
3259 ret = device_register(&wq_dev->dev);
3266 if (wq->flags & WQ_UNBOUND) {
3267 struct device_attribute *attr;
3269 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3270 ret = device_create_file(&wq_dev->dev, attr);
3272 device_unregister(&wq_dev->dev);
3279 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3284 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3285 * @wq: the workqueue to unregister
3287 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3289 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3291 struct wq_device *wq_dev = wq->wq_dev;
3297 device_unregister(&wq_dev->dev);
3299 #else /* CONFIG_SYSFS */
3300 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3301 #endif /* CONFIG_SYSFS */
3304 * free_workqueue_attrs - free a workqueue_attrs
3305 * @attrs: workqueue_attrs to free
3307 * Undo alloc_workqueue_attrs().
3309 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3312 free_cpumask_var(attrs->cpumask);
3318 * alloc_workqueue_attrs - allocate a workqueue_attrs
3319 * @gfp_mask: allocation mask to use
3321 * Allocate a new workqueue_attrs, initialize with default settings and
3322 * return it. Returns NULL on failure.
3324 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3326 struct workqueue_attrs *attrs;
3328 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3331 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3334 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3337 free_workqueue_attrs(attrs);
3341 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3342 const struct workqueue_attrs *from)
3344 to->nice = from->nice;
3345 cpumask_copy(to->cpumask, from->cpumask);
3348 /* hash value of the content of @attr */
3349 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3353 hash = jhash_1word(attrs->nice, hash);
3354 hash = jhash(cpumask_bits(attrs->cpumask),
3355 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3359 /* content equality test */
3360 static bool wqattrs_equal(const struct workqueue_attrs *a,
3361 const struct workqueue_attrs *b)
3363 if (a->nice != b->nice)
3365 if (!cpumask_equal(a->cpumask, b->cpumask))
3371 * init_worker_pool - initialize a newly zalloc'd worker_pool
3372 * @pool: worker_pool to initialize
3374 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3375 * Returns 0 on success, -errno on failure. Even on failure, all fields
3376 * inside @pool proper are initialized and put_unbound_pool() can be called
3377 * on @pool safely to release it.
3379 static int init_worker_pool(struct worker_pool *pool)
3381 spin_lock_init(&pool->lock);
3384 pool->node = NUMA_NO_NODE;
3385 pool->flags |= POOL_DISASSOCIATED;
3386 INIT_LIST_HEAD(&pool->worklist);
3387 INIT_LIST_HEAD(&pool->idle_list);
3388 hash_init(pool->busy_hash);
3390 init_timer_deferrable(&pool->idle_timer);
3391 pool->idle_timer.function = idle_worker_timeout;
3392 pool->idle_timer.data = (unsigned long)pool;
3394 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3395 (unsigned long)pool);
3397 mutex_init(&pool->manager_arb);
3398 mutex_init(&pool->manager_mutex);
3399 idr_init(&pool->worker_idr);
3401 INIT_HLIST_NODE(&pool->hash_node);
3404 /* shouldn't fail above this point */
3405 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3411 static void rcu_free_pool(struct rcu_head *rcu)
3413 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3415 idr_destroy(&pool->worker_idr);
3416 free_workqueue_attrs(pool->attrs);
3421 * put_unbound_pool - put a worker_pool
3422 * @pool: worker_pool to put
3424 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3425 * safe manner. get_unbound_pool() calls this function on its failure path
3426 * and this function should be able to release pools which went through,
3427 * successfully or not, init_worker_pool().
3429 * Should be called with wq_pool_mutex held.
3431 static void put_unbound_pool(struct worker_pool *pool)
3433 struct worker *worker;
3435 lockdep_assert_held(&wq_pool_mutex);
3441 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3442 WARN_ON(!list_empty(&pool->worklist)))
3445 /* release id and unhash */
3447 idr_remove(&worker_pool_idr, pool->id);
3448 hash_del(&pool->hash_node);
3451 * Become the manager and destroy all workers. Grabbing
3452 * manager_arb prevents @pool's workers from blocking on
3455 mutex_lock(&pool->manager_arb);
3456 mutex_lock(&pool->manager_mutex);
3457 spin_lock_irq(&pool->lock);
3459 while ((worker = first_worker(pool)))
3460 destroy_worker(worker);
3461 WARN_ON(pool->nr_workers || pool->nr_idle);
3463 spin_unlock_irq(&pool->lock);
3464 mutex_unlock(&pool->manager_mutex);
3465 mutex_unlock(&pool->manager_arb);
3467 /* shut down the timers */
3468 del_timer_sync(&pool->idle_timer);
3469 del_timer_sync(&pool->mayday_timer);
3471 /* sched-RCU protected to allow dereferences from get_work_pool() */
3472 call_rcu_sched(&pool->rcu, rcu_free_pool);
3476 * get_unbound_pool - get a worker_pool with the specified attributes
3477 * @attrs: the attributes of the worker_pool to get
3479 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3480 * reference count and return it. If there already is a matching
3481 * worker_pool, it will be used; otherwise, this function attempts to
3482 * create a new one. On failure, returns NULL.
3484 * Should be called with wq_pool_mutex held.
3486 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3488 u32 hash = wqattrs_hash(attrs);
3489 struct worker_pool *pool;
3492 lockdep_assert_held(&wq_pool_mutex);
3494 /* do we already have a matching pool? */
3495 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3496 if (wqattrs_equal(pool->attrs, attrs)) {
3502 /* nope, create a new one */
3503 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3504 if (!pool || init_worker_pool(pool) < 0)
3507 if (workqueue_freezing)
3508 pool->flags |= POOL_FREEZING;
3510 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3511 copy_workqueue_attrs(pool->attrs, attrs);
3513 /* if cpumask is contained inside a NUMA node, we belong to that node */
3514 if (wq_numa_enabled) {
3515 for_each_node(node) {
3516 if (cpumask_subset(pool->attrs->cpumask,
3517 wq_numa_possible_cpumask[node])) {
3524 if (worker_pool_assign_id(pool) < 0)
3527 /* create and start the initial worker */
3528 if (create_and_start_worker(pool) < 0)
3532 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3537 put_unbound_pool(pool);
3541 static void rcu_free_pwq(struct rcu_head *rcu)
3543 kmem_cache_free(pwq_cache,
3544 container_of(rcu, struct pool_workqueue, rcu));
3548 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3549 * and needs to be destroyed.
3551 static void pwq_unbound_release_workfn(struct work_struct *work)
3553 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3554 unbound_release_work);
3555 struct workqueue_struct *wq = pwq->wq;
3556 struct worker_pool *pool = pwq->pool;
3559 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3563 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3564 * necessary on release but do it anyway. It's easier to verify
3565 * and consistent with the linking path.
3567 mutex_lock(&wq->mutex);
3568 list_del_rcu(&pwq->pwqs_node);
3569 is_last = list_empty(&wq->pwqs);
3570 mutex_unlock(&wq->mutex);
3572 mutex_lock(&wq_pool_mutex);
3573 put_unbound_pool(pool);
3574 mutex_unlock(&wq_pool_mutex);
3576 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3579 * If we're the last pwq going away, @wq is already dead and no one
3580 * is gonna access it anymore. Free it.
3583 free_workqueue_attrs(wq->unbound_attrs);
3589 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3590 * @pwq: target pool_workqueue
3592 * If @pwq isn't freezing, set @pwq->max_active to the associated
3593 * workqueue's saved_max_active and activate delayed work items
3594 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3596 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3598 struct workqueue_struct *wq = pwq->wq;
3599 bool freezable = wq->flags & WQ_FREEZABLE;
3601 /* for @wq->saved_max_active */
3602 lockdep_assert_held(&wq->mutex);
3604 /* fast exit for non-freezable wqs */
3605 if (!freezable && pwq->max_active == wq->saved_max_active)
3608 spin_lock_irq(&pwq->pool->lock);
3610 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3611 pwq->max_active = wq->saved_max_active;
3613 while (!list_empty(&pwq->delayed_works) &&
3614 pwq->nr_active < pwq->max_active)
3615 pwq_activate_first_delayed(pwq);
3618 * Need to kick a worker after thawed or an unbound wq's
3619 * max_active is bumped. It's a slow path. Do it always.
3621 wake_up_worker(pwq->pool);
3623 pwq->max_active = 0;
3626 spin_unlock_irq(&pwq->pool->lock);
3629 static void init_and_link_pwq(struct pool_workqueue *pwq,
3630 struct workqueue_struct *wq,
3631 struct worker_pool *pool,
3632 struct pool_workqueue **p_last_pwq)
3636 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3640 pwq->flush_color = -1;
3642 INIT_LIST_HEAD(&pwq->delayed_works);
3643 INIT_LIST_HEAD(&pwq->mayday_node);
3644 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3646 mutex_lock(&wq->mutex);
3649 * Set the matching work_color. This is synchronized with
3650 * wq->mutex to avoid confusing flush_workqueue().
3653 *p_last_pwq = first_pwq(wq);
3654 pwq->work_color = wq->work_color;
3656 /* sync max_active to the current setting */
3657 pwq_adjust_max_active(pwq);
3660 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3662 if (wq->flags & WQ_UNBOUND) {
3663 copy_workqueue_attrs(wq->unbound_attrs, pool->attrs);
3665 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3668 mutex_unlock(&wq->mutex);
3672 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3673 * @wq: the target workqueue
3674 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3676 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3677 * current attributes, a new pwq is created and made the first pwq which
3678 * will serve all new work items. Older pwqs are released as in-flight
3679 * work items finish. Note that a work item which repeatedly requeues
3680 * itself back-to-back will stay on its current pwq.
3682 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3685 int apply_workqueue_attrs(struct workqueue_struct *wq,
3686 const struct workqueue_attrs *attrs)
3688 struct workqueue_attrs *new_attrs;
3689 struct pool_workqueue *pwq = NULL, *last_pwq;
3690 struct worker_pool *pool;
3693 /* only unbound workqueues can change attributes */
3694 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3697 /* creating multiple pwqs breaks ordering guarantee */
3698 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3701 /* make a copy of @attrs and sanitize it */
3702 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3706 copy_workqueue_attrs(new_attrs, attrs);
3707 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3709 mutex_lock(&wq_pool_mutex);
3711 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3713 mutex_unlock(&wq_pool_mutex);
3717 pool = get_unbound_pool(new_attrs);
3719 mutex_unlock(&wq_pool_mutex);
3723 mutex_unlock(&wq_pool_mutex);
3725 init_and_link_pwq(pwq, wq, pool, &last_pwq);
3727 spin_lock_irq(&last_pwq->pool->lock);
3729 spin_unlock_irq(&last_pwq->pool->lock);
3735 free_workqueue_attrs(new_attrs);
3739 kmem_cache_free(pwq_cache, pwq);
3744 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3746 bool highpri = wq->flags & WQ_HIGHPRI;
3749 if (!(wq->flags & WQ_UNBOUND)) {
3750 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3754 for_each_possible_cpu(cpu) {
3755 struct pool_workqueue *pwq =
3756 per_cpu_ptr(wq->cpu_pwqs, cpu);
3757 struct worker_pool *cpu_pools =
3758 per_cpu(cpu_worker_pools, cpu);
3760 init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL);
3764 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3768 static int wq_clamp_max_active(int max_active, unsigned int flags,
3771 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3773 if (max_active < 1 || max_active > lim)
3774 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3775 max_active, name, 1, lim);
3777 return clamp_val(max_active, 1, lim);
3780 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3783 struct lock_class_key *key,
3784 const char *lock_name, ...)
3786 size_t tbl_size = 0;
3788 struct workqueue_struct *wq;
3789 struct pool_workqueue *pwq;
3791 /* allocate wq and format name */
3792 if (flags & WQ_UNBOUND)
3793 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
3795 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3799 if (flags & WQ_UNBOUND) {
3800 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3801 if (!wq->unbound_attrs)
3805 va_start(args, lock_name);
3806 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3809 max_active = max_active ?: WQ_DFL_ACTIVE;
3810 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3814 wq->saved_max_active = max_active;
3815 mutex_init(&wq->mutex);
3816 atomic_set(&wq->nr_pwqs_to_flush, 0);
3817 INIT_LIST_HEAD(&wq->pwqs);
3818 INIT_LIST_HEAD(&wq->flusher_queue);
3819 INIT_LIST_HEAD(&wq->flusher_overflow);
3820 INIT_LIST_HEAD(&wq->maydays);
3822 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3823 INIT_LIST_HEAD(&wq->list);
3825 if (alloc_and_link_pwqs(wq) < 0)
3829 * Workqueues which may be used during memory reclaim should
3830 * have a rescuer to guarantee forward progress.
3832 if (flags & WQ_MEM_RECLAIM) {
3833 struct worker *rescuer;
3835 rescuer = alloc_worker();
3839 rescuer->rescue_wq = wq;
3840 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3842 if (IS_ERR(rescuer->task)) {
3847 wq->rescuer = rescuer;
3848 rescuer->task->flags |= PF_NO_SETAFFINITY;
3849 wake_up_process(rescuer->task);
3852 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3856 * wq_pool_mutex protects global freeze state and workqueues list.
3857 * Grab it, adjust max_active and add the new @wq to workqueues
3860 mutex_lock(&wq_pool_mutex);
3862 mutex_lock(&wq->mutex);
3863 for_each_pwq(pwq, wq)
3864 pwq_adjust_max_active(pwq);
3865 mutex_unlock(&wq->mutex);
3867 list_add(&wq->list, &workqueues);
3869 mutex_unlock(&wq_pool_mutex);
3874 free_workqueue_attrs(wq->unbound_attrs);
3878 destroy_workqueue(wq);
3881 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3884 * destroy_workqueue - safely terminate a workqueue
3885 * @wq: target workqueue
3887 * Safely destroy a workqueue. All work currently pending will be done first.
3889 void destroy_workqueue(struct workqueue_struct *wq)
3891 struct pool_workqueue *pwq;
3893 /* drain it before proceeding with destruction */
3894 drain_workqueue(wq);
3897 mutex_lock(&wq->mutex);
3898 for_each_pwq(pwq, wq) {
3901 for (i = 0; i < WORK_NR_COLORS; i++) {
3902 if (WARN_ON(pwq->nr_in_flight[i])) {
3903 mutex_unlock(&wq->mutex);
3908 if (WARN_ON(pwq->refcnt > 1) ||
3909 WARN_ON(pwq->nr_active) ||
3910 WARN_ON(!list_empty(&pwq->delayed_works))) {
3911 mutex_unlock(&wq->mutex);
3915 mutex_unlock(&wq->mutex);
3918 * wq list is used to freeze wq, remove from list after
3919 * flushing is complete in case freeze races us.
3921 mutex_lock(&wq_pool_mutex);
3922 list_del_init(&wq->list);
3923 mutex_unlock(&wq_pool_mutex);
3925 workqueue_sysfs_unregister(wq);
3928 kthread_stop(wq->rescuer->task);
3933 if (!(wq->flags & WQ_UNBOUND)) {
3935 * The base ref is never dropped on per-cpu pwqs. Directly
3936 * free the pwqs and wq.
3938 free_percpu(wq->cpu_pwqs);
3942 * We're the sole accessor of @wq at this point. Directly
3943 * access the first pwq and put the base ref. As both pwqs
3944 * and pools are sched-RCU protected, the lock operations
3945 * are safe. @wq will be freed when the last pwq is
3948 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3950 spin_lock_irq(&pwq->pool->lock);
3952 spin_unlock_irq(&pwq->pool->lock);
3955 EXPORT_SYMBOL_GPL(destroy_workqueue);
3958 * workqueue_set_max_active - adjust max_active of a workqueue
3959 * @wq: target workqueue
3960 * @max_active: new max_active value.
3962 * Set max_active of @wq to @max_active.
3965 * Don't call from IRQ context.
3967 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3969 struct pool_workqueue *pwq;
3971 /* disallow meddling with max_active for ordered workqueues */
3972 if (WARN_ON(wq->flags & __WQ_ORDERED))
3975 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3977 mutex_lock(&wq->mutex);
3979 wq->saved_max_active = max_active;
3981 for_each_pwq(pwq, wq)
3982 pwq_adjust_max_active(pwq);
3984 mutex_unlock(&wq->mutex);
3986 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3989 * current_is_workqueue_rescuer - is %current workqueue rescuer?
3991 * Determine whether %current is a workqueue rescuer. Can be used from
3992 * work functions to determine whether it's being run off the rescuer task.
3994 bool current_is_workqueue_rescuer(void)
3996 struct worker *worker = current_wq_worker();
3998 return worker && worker->rescue_wq;
4002 * workqueue_congested - test whether a workqueue is congested
4003 * @cpu: CPU in question
4004 * @wq: target workqueue
4006 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4007 * no synchronization around this function and the test result is
4008 * unreliable and only useful as advisory hints or for debugging.
4011 * %true if congested, %false otherwise.
4013 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4015 struct pool_workqueue *pwq;
4018 rcu_read_lock_sched();
4020 if (!(wq->flags & WQ_UNBOUND))
4021 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4023 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4025 ret = !list_empty(&pwq->delayed_works);
4026 rcu_read_unlock_sched();
4030 EXPORT_SYMBOL_GPL(workqueue_congested);
4033 * work_busy - test whether a work is currently pending or running
4034 * @work: the work to be tested
4036 * Test whether @work is currently pending or running. There is no
4037 * synchronization around this function and the test result is
4038 * unreliable and only useful as advisory hints or for debugging.
4041 * OR'd bitmask of WORK_BUSY_* bits.
4043 unsigned int work_busy(struct work_struct *work)
4045 struct worker_pool *pool;
4046 unsigned long flags;
4047 unsigned int ret = 0;
4049 if (work_pending(work))
4050 ret |= WORK_BUSY_PENDING;
4052 local_irq_save(flags);
4053 pool = get_work_pool(work);
4055 spin_lock(&pool->lock);
4056 if (find_worker_executing_work(pool, work))
4057 ret |= WORK_BUSY_RUNNING;
4058 spin_unlock(&pool->lock);
4060 local_irq_restore(flags);
4064 EXPORT_SYMBOL_GPL(work_busy);
4069 * There are two challenges in supporting CPU hotplug. Firstly, there
4070 * are a lot of assumptions on strong associations among work, pwq and
4071 * pool which make migrating pending and scheduled works very
4072 * difficult to implement without impacting hot paths. Secondly,
4073 * worker pools serve mix of short, long and very long running works making
4074 * blocked draining impractical.
4076 * This is solved by allowing the pools to be disassociated from the CPU
4077 * running as an unbound one and allowing it to be reattached later if the
4078 * cpu comes back online.
4081 static void wq_unbind_fn(struct work_struct *work)
4083 int cpu = smp_processor_id();
4084 struct worker_pool *pool;
4085 struct worker *worker;
4088 for_each_cpu_worker_pool(pool, cpu) {
4089 WARN_ON_ONCE(cpu != smp_processor_id());
4091 mutex_lock(&pool->manager_mutex);
4092 spin_lock_irq(&pool->lock);
4095 * We've blocked all manager operations. Make all workers
4096 * unbound and set DISASSOCIATED. Before this, all workers
4097 * except for the ones which are still executing works from
4098 * before the last CPU down must be on the cpu. After
4099 * this, they may become diasporas.
4101 for_each_pool_worker(worker, wi, pool)
4102 worker->flags |= WORKER_UNBOUND;
4104 pool->flags |= POOL_DISASSOCIATED;
4106 spin_unlock_irq(&pool->lock);
4107 mutex_unlock(&pool->manager_mutex);
4111 * Call schedule() so that we cross rq->lock and thus can guarantee
4112 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4113 * as scheduler callbacks may be invoked from other cpus.
4118 * Sched callbacks are disabled now. Zap nr_running. After this,
4119 * nr_running stays zero and need_more_worker() and keep_working()
4120 * are always true as long as the worklist is not empty. Pools on
4121 * @cpu now behave as unbound (in terms of concurrency management)
4122 * pools which are served by workers tied to the CPU.
4124 * On return from this function, the current worker would trigger
4125 * unbound chain execution of pending work items if other workers
4128 for_each_cpu_worker_pool(pool, cpu)
4129 atomic_set(&pool->nr_running, 0);
4133 * rebind_workers - rebind all workers of a pool to the associated CPU
4134 * @pool: pool of interest
4136 * @pool->cpu is coming online. Rebind all workers to the CPU.
4138 static void rebind_workers(struct worker_pool *pool)
4140 struct worker *worker;
4143 lockdep_assert_held(&pool->manager_mutex);
4146 * Restore CPU affinity of all workers. As all idle workers should
4147 * be on the run-queue of the associated CPU before any local
4148 * wake-ups for concurrency management happen, restore CPU affinty
4149 * of all workers first and then clear UNBOUND. As we're called
4150 * from CPU_ONLINE, the following shouldn't fail.
4152 for_each_pool_worker(worker, wi, pool)
4153 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4154 pool->attrs->cpumask) < 0);
4156 spin_lock_irq(&pool->lock);
4158 for_each_pool_worker(worker, wi, pool) {
4159 unsigned int worker_flags = worker->flags;
4162 * A bound idle worker should actually be on the runqueue
4163 * of the associated CPU for local wake-ups targeting it to
4164 * work. Kick all idle workers so that they migrate to the
4165 * associated CPU. Doing this in the same loop as
4166 * replacing UNBOUND with REBOUND is safe as no worker will
4167 * be bound before @pool->lock is released.
4169 if (worker_flags & WORKER_IDLE)
4170 wake_up_process(worker->task);
4173 * We want to clear UNBOUND but can't directly call
4174 * worker_clr_flags() or adjust nr_running. Atomically
4175 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4176 * @worker will clear REBOUND using worker_clr_flags() when
4177 * it initiates the next execution cycle thus restoring
4178 * concurrency management. Note that when or whether
4179 * @worker clears REBOUND doesn't affect correctness.
4181 * ACCESS_ONCE() is necessary because @worker->flags may be
4182 * tested without holding any lock in
4183 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4184 * fail incorrectly leading to premature concurrency
4185 * management operations.
4187 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4188 worker_flags |= WORKER_REBOUND;
4189 worker_flags &= ~WORKER_UNBOUND;
4190 ACCESS_ONCE(worker->flags) = worker_flags;
4193 spin_unlock_irq(&pool->lock);
4197 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4198 * @pool: unbound pool of interest
4199 * @cpu: the CPU which is coming up
4201 * An unbound pool may end up with a cpumask which doesn't have any online
4202 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4203 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4204 * online CPU before, cpus_allowed of all its workers should be restored.
4206 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4208 static cpumask_t cpumask;
4209 struct worker *worker;
4212 lockdep_assert_held(&pool->manager_mutex);
4214 /* is @cpu allowed for @pool? */
4215 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4218 /* is @cpu the only online CPU? */
4219 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4220 if (cpumask_weight(&cpumask) != 1)
4223 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4224 for_each_pool_worker(worker, wi, pool)
4225 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4226 pool->attrs->cpumask) < 0);
4230 * Workqueues should be brought up before normal priority CPU notifiers.
4231 * This will be registered high priority CPU notifier.
4233 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4234 unsigned long action,
4237 int cpu = (unsigned long)hcpu;
4238 struct worker_pool *pool;
4241 switch (action & ~CPU_TASKS_FROZEN) {
4242 case CPU_UP_PREPARE:
4243 for_each_cpu_worker_pool(pool, cpu) {
4244 if (pool->nr_workers)
4246 if (create_and_start_worker(pool) < 0)
4251 case CPU_DOWN_FAILED:
4253 mutex_lock(&wq_pool_mutex);
4255 for_each_pool(pool, pi) {
4256 mutex_lock(&pool->manager_mutex);
4258 if (pool->cpu == cpu) {
4259 spin_lock_irq(&pool->lock);
4260 pool->flags &= ~POOL_DISASSOCIATED;
4261 spin_unlock_irq(&pool->lock);
4263 rebind_workers(pool);
4264 } else if (pool->cpu < 0) {
4265 restore_unbound_workers_cpumask(pool, cpu);
4268 mutex_unlock(&pool->manager_mutex);
4271 mutex_unlock(&wq_pool_mutex);
4278 * Workqueues should be brought down after normal priority CPU notifiers.
4279 * This will be registered as low priority CPU notifier.
4281 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4282 unsigned long action,
4285 int cpu = (unsigned long)hcpu;
4286 struct work_struct unbind_work;
4288 switch (action & ~CPU_TASKS_FROZEN) {
4289 case CPU_DOWN_PREPARE:
4290 /* unbinding should happen on the local CPU */
4291 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4292 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4293 flush_work(&unbind_work);
4301 struct work_for_cpu {
4302 struct work_struct work;
4308 static void work_for_cpu_fn(struct work_struct *work)
4310 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4312 wfc->ret = wfc->fn(wfc->arg);
4316 * work_on_cpu - run a function in user context on a particular cpu
4317 * @cpu: the cpu to run on
4318 * @fn: the function to run
4319 * @arg: the function arg
4321 * This will return the value @fn returns.
4322 * It is up to the caller to ensure that the cpu doesn't go offline.
4323 * The caller must not hold any locks which would prevent @fn from completing.
4325 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4327 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4329 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4330 schedule_work_on(cpu, &wfc.work);
4331 flush_work(&wfc.work);
4334 EXPORT_SYMBOL_GPL(work_on_cpu);
4335 #endif /* CONFIG_SMP */
4337 #ifdef CONFIG_FREEZER
4340 * freeze_workqueues_begin - begin freezing workqueues
4342 * Start freezing workqueues. After this function returns, all freezable
4343 * workqueues will queue new works to their delayed_works list instead of
4347 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4349 void freeze_workqueues_begin(void)
4351 struct worker_pool *pool;
4352 struct workqueue_struct *wq;
4353 struct pool_workqueue *pwq;
4356 mutex_lock(&wq_pool_mutex);
4358 WARN_ON_ONCE(workqueue_freezing);
4359 workqueue_freezing = true;
4362 for_each_pool(pool, pi) {
4363 spin_lock_irq(&pool->lock);
4364 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4365 pool->flags |= POOL_FREEZING;
4366 spin_unlock_irq(&pool->lock);
4369 list_for_each_entry(wq, &workqueues, list) {
4370 mutex_lock(&wq->mutex);
4371 for_each_pwq(pwq, wq)
4372 pwq_adjust_max_active(pwq);
4373 mutex_unlock(&wq->mutex);
4376 mutex_unlock(&wq_pool_mutex);
4380 * freeze_workqueues_busy - are freezable workqueues still busy?
4382 * Check whether freezing is complete. This function must be called
4383 * between freeze_workqueues_begin() and thaw_workqueues().
4386 * Grabs and releases wq_pool_mutex.
4389 * %true if some freezable workqueues are still busy. %false if freezing
4392 bool freeze_workqueues_busy(void)
4395 struct workqueue_struct *wq;
4396 struct pool_workqueue *pwq;
4398 mutex_lock(&wq_pool_mutex);
4400 WARN_ON_ONCE(!workqueue_freezing);
4402 list_for_each_entry(wq, &workqueues, list) {
4403 if (!(wq->flags & WQ_FREEZABLE))
4406 * nr_active is monotonically decreasing. It's safe
4407 * to peek without lock.
4409 rcu_read_lock_sched();
4410 for_each_pwq(pwq, wq) {
4411 WARN_ON_ONCE(pwq->nr_active < 0);
4412 if (pwq->nr_active) {
4414 rcu_read_unlock_sched();
4418 rcu_read_unlock_sched();
4421 mutex_unlock(&wq_pool_mutex);
4426 * thaw_workqueues - thaw workqueues
4428 * Thaw workqueues. Normal queueing is restored and all collected
4429 * frozen works are transferred to their respective pool worklists.
4432 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4434 void thaw_workqueues(void)
4436 struct workqueue_struct *wq;
4437 struct pool_workqueue *pwq;
4438 struct worker_pool *pool;
4441 mutex_lock(&wq_pool_mutex);
4443 if (!workqueue_freezing)
4446 /* clear FREEZING */
4447 for_each_pool(pool, pi) {
4448 spin_lock_irq(&pool->lock);
4449 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4450 pool->flags &= ~POOL_FREEZING;
4451 spin_unlock_irq(&pool->lock);
4454 /* restore max_active and repopulate worklist */
4455 list_for_each_entry(wq, &workqueues, list) {
4456 mutex_lock(&wq->mutex);
4457 for_each_pwq(pwq, wq)
4458 pwq_adjust_max_active(pwq);
4459 mutex_unlock(&wq->mutex);
4462 workqueue_freezing = false;
4464 mutex_unlock(&wq_pool_mutex);
4466 #endif /* CONFIG_FREEZER */
4468 static void __init wq_numa_init(void)
4473 /* determine NUMA pwq table len - highest node id + 1 */
4475 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4477 if (num_possible_nodes() <= 1)
4481 * We want masks of possible CPUs of each node which isn't readily
4482 * available. Build one from cpu_to_node() which should have been
4483 * fully initialized by now.
4485 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4489 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node));
4491 for_each_possible_cpu(cpu) {
4492 node = cpu_to_node(cpu);
4493 if (WARN_ON(node == NUMA_NO_NODE)) {
4494 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4495 /* happens iff arch is bonkers, let's just proceed */
4498 cpumask_set_cpu(cpu, tbl[node]);
4501 wq_numa_possible_cpumask = tbl;
4502 wq_numa_enabled = true;
4505 static int __init init_workqueues(void)
4507 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4510 /* make sure we have enough bits for OFFQ pool ID */
4511 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4512 WORK_CPU_END * NR_STD_WORKER_POOLS);
4514 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4516 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4518 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4519 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4523 /* initialize CPU pools */
4524 for_each_possible_cpu(cpu) {
4525 struct worker_pool *pool;
4528 for_each_cpu_worker_pool(pool, cpu) {
4529 BUG_ON(init_worker_pool(pool));
4531 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4532 pool->attrs->nice = std_nice[i++];
4533 pool->node = cpu_to_node(cpu);
4536 mutex_lock(&wq_pool_mutex);
4537 BUG_ON(worker_pool_assign_id(pool));
4538 mutex_unlock(&wq_pool_mutex);
4542 /* create the initial worker */
4543 for_each_online_cpu(cpu) {
4544 struct worker_pool *pool;
4546 for_each_cpu_worker_pool(pool, cpu) {
4547 pool->flags &= ~POOL_DISASSOCIATED;
4548 BUG_ON(create_and_start_worker(pool) < 0);
4552 /* create default unbound wq attrs */
4553 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4554 struct workqueue_attrs *attrs;
4556 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4557 attrs->nice = std_nice[i];
4558 unbound_std_wq_attrs[i] = attrs;
4561 system_wq = alloc_workqueue("events", 0, 0);
4562 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4563 system_long_wq = alloc_workqueue("events_long", 0, 0);
4564 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4565 WQ_UNBOUND_MAX_ACTIVE);
4566 system_freezable_wq = alloc_workqueue("events_freezable",
4568 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4569 !system_unbound_wq || !system_freezable_wq);
4572 early_initcall(init_workqueues);