2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
51 #ifdef CONFIG_SEC_DEBUG
52 #include <soc/sprd/sec_debug.h>
55 #ifdef CONFIG_SPRD_DEBUG
56 #include <soc/sprd/sprd_debug.h>
58 #if defined(CONFIG_SYSTEM_LOAD_ANALYZER)
59 #include <linux/load_analyzer.h>
62 #include "workqueue_internal.h"
68 * A bound pool is either associated or disassociated with its CPU.
69 * While associated (!DISASSOCIATED), all workers are bound to the
70 * CPU and none has %WORKER_UNBOUND set and concurrency management
73 * While DISASSOCIATED, the cpu may be offline and all workers have
74 * %WORKER_UNBOUND set and concurrency management disabled, and may
75 * be executing on any CPU. The pool behaves as an unbound one.
77 * Note that DISASSOCIATED should be flipped only while holding
78 * manager_mutex to avoid changing binding state while
79 * create_worker() is in progress.
81 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
82 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
83 POOL_FREEZING = 1 << 3, /* freeze in progress */
86 WORKER_STARTED = 1 << 0, /* started */
87 WORKER_DIE = 1 << 1, /* die die die */
88 WORKER_IDLE = 1 << 2, /* is idle */
89 WORKER_PREP = 1 << 3, /* preparing to run works */
90 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
91 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
92 WORKER_REBOUND = 1 << 8, /* worker was rebound */
94 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
95 WORKER_UNBOUND | WORKER_REBOUND,
97 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
99 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
100 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
102 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
103 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
105 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
106 /* call for help after 10ms
108 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
109 CREATE_COOLDOWN = HZ, /* time to breath after fail */
112 * Rescue workers are used only on emergencies and shared by
113 * all cpus. Give -20.
115 RESCUER_NICE_LEVEL = -20,
116 HIGHPRI_NICE_LEVEL = -20,
122 * Structure fields follow one of the following exclusion rules.
124 * I: Modifiable by initialization/destruction paths and read-only for
127 * P: Preemption protected. Disabling preemption is enough and should
128 * only be modified and accessed from the local cpu.
130 * L: pool->lock protected. Access with pool->lock held.
132 * X: During normal operation, modification requires pool->lock and should
133 * be done only from local cpu. Either disabling preemption on local
134 * cpu or grabbing pool->lock is enough for read access. If
135 * POOL_DISASSOCIATED is set, it's identical to L.
137 * MG: pool->manager_mutex and pool->lock protected. Writes require both
138 * locks. Reads can happen under either lock.
140 * PL: wq_pool_mutex protected.
142 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
144 * WQ: wq->mutex protected.
146 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
148 * MD: wq_mayday_lock protected.
151 /* struct worker is defined in workqueue_internal.h */
154 spinlock_t lock; /* the pool lock */
155 int cpu; /* I: the associated cpu */
156 int node; /* I: the associated node ID */
157 int id; /* I: pool ID */
158 unsigned int flags; /* X: flags */
160 struct list_head worklist; /* L: list of pending works */
161 int nr_workers; /* L: total number of workers */
163 /* nr_idle includes the ones off idle_list for rebinding */
164 int nr_idle; /* L: currently idle ones */
166 struct list_head idle_list; /* X: list of idle workers */
167 struct timer_list idle_timer; /* L: worker idle timeout */
168 struct timer_list mayday_timer; /* L: SOS timer for workers */
170 /* a workers is either on busy_hash or idle_list, or the manager */
171 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
172 /* L: hash of busy workers */
174 /* see manage_workers() for details on the two manager mutexes */
175 struct mutex manager_arb; /* manager arbitration */
176 struct mutex manager_mutex; /* manager exclusion */
177 struct idr worker_idr; /* MG: worker IDs and iteration */
179 struct workqueue_attrs *attrs; /* I: worker attributes */
180 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
181 int refcnt; /* PL: refcnt for unbound pools */
184 * The current concurrency level. As it's likely to be accessed
185 * from other CPUs during try_to_wake_up(), put it in a separate
188 atomic_t nr_running ____cacheline_aligned_in_smp;
191 * Destruction of pool is sched-RCU protected to allow dereferences
192 * from get_work_pool().
195 } ____cacheline_aligned_in_smp;
198 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
199 * of work_struct->data are used for flags and the remaining high bits
200 * point to the pwq; thus, pwqs need to be aligned at two's power of the
201 * number of flag bits.
203 struct pool_workqueue {
204 struct worker_pool *pool; /* I: the associated pool */
205 struct workqueue_struct *wq; /* I: the owning workqueue */
206 int work_color; /* L: current color */
207 int flush_color; /* L: flushing color */
208 int refcnt; /* L: reference count */
209 int nr_in_flight[WORK_NR_COLORS];
210 /* L: nr of in_flight works */
211 int nr_active; /* L: nr of active works */
212 int max_active; /* L: max active works */
213 struct list_head delayed_works; /* L: delayed works */
214 struct list_head pwqs_node; /* WR: node on wq->pwqs */
215 struct list_head mayday_node; /* MD: node on wq->maydays */
218 * Release of unbound pwq is punted to system_wq. See put_pwq()
219 * and pwq_unbound_release_workfn() for details. pool_workqueue
220 * itself is also sched-RCU protected so that the first pwq can be
221 * determined without grabbing wq->mutex.
223 struct work_struct unbound_release_work;
225 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
228 * Structure used to wait for workqueue flush.
231 struct list_head list; /* WQ: list of flushers */
232 int flush_color; /* WQ: flush color waiting for */
233 struct completion done; /* flush completion */
239 * The externally visible workqueue. It relays the issued work items to
240 * the appropriate worker_pool through its pool_workqueues.
242 struct workqueue_struct {
243 struct list_head pwqs; /* WR: all pwqs of this wq */
244 struct list_head list; /* PL: list of all workqueues */
246 struct mutex mutex; /* protects this wq */
247 int work_color; /* WQ: current work color */
248 int flush_color; /* WQ: current flush color */
249 atomic_t nr_pwqs_to_flush; /* flush in progress */
250 struct wq_flusher *first_flusher; /* WQ: first flusher */
251 struct list_head flusher_queue; /* WQ: flush waiters */
252 struct list_head flusher_overflow; /* WQ: flush overflow list */
254 struct list_head maydays; /* MD: pwqs requesting rescue */
255 struct worker *rescuer; /* I: rescue worker */
257 int nr_drainers; /* WQ: drain in progress */
258 int saved_max_active; /* WQ: saved pwq max_active */
260 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
261 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
264 struct wq_device *wq_dev; /* I: for sysfs interface */
266 #ifdef CONFIG_LOCKDEP
267 struct lockdep_map lockdep_map;
269 char name[WQ_NAME_LEN]; /* I: workqueue name */
271 /* hot fields used during command issue, aligned to cacheline */
272 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
273 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
274 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
277 static struct kmem_cache *pwq_cache;
279 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
280 static cpumask_var_t *wq_numa_possible_cpumask;
281 /* possible CPUs of each node */
283 static bool wq_disable_numa;
284 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
286 /* see the comment above the definition of WQ_POWER_EFFICIENT */
287 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
288 static bool wq_power_efficient = true;
290 static bool wq_power_efficient;
293 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
295 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
301 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
303 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
304 static bool workqueue_freezing; /* PL: have wqs started freezing? */
306 /* the per-cpu worker pools */
307 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
310 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
312 /* PL: hash of all unbound pools keyed by pool->attrs */
313 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
315 /* I: attributes used when instantiating standard unbound pools on demand */
316 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
318 /* I: attributes used when instantiating ordered pools on demand */
319 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
321 struct workqueue_struct *system_wq __read_mostly;
322 EXPORT_SYMBOL(system_wq);
323 struct workqueue_struct *system_highpri_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_highpri_wq);
325 struct workqueue_struct *system_long_wq __read_mostly;
326 EXPORT_SYMBOL_GPL(system_long_wq);
327 struct workqueue_struct *system_unbound_wq __read_mostly;
328 EXPORT_SYMBOL_GPL(system_unbound_wq);
329 struct workqueue_struct *system_freezable_wq __read_mostly;
330 EXPORT_SYMBOL_GPL(system_freezable_wq);
331 struct workqueue_struct *system_power_efficient_wq __read_mostly;
332 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
333 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
334 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
336 static int worker_thread(void *__worker);
337 static void copy_workqueue_attrs(struct workqueue_attrs *to,
338 const struct workqueue_attrs *from);
340 #define CREATE_TRACE_POINTS
341 #include <trace/events/workqueue.h>
343 #define assert_rcu_or_pool_mutex() \
344 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
345 lockdep_is_held(&wq_pool_mutex), \
346 "sched RCU or wq_pool_mutex should be held")
348 #define assert_rcu_or_wq_mutex(wq) \
349 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
350 lockdep_is_held(&wq->mutex), \
351 "sched RCU or wq->mutex should be held")
353 #ifdef CONFIG_LOCKDEP
354 #define assert_manager_or_pool_lock(pool) \
355 WARN_ONCE(debug_locks && \
356 !lockdep_is_held(&(pool)->manager_mutex) && \
357 !lockdep_is_held(&(pool)->lock), \
358 "pool->manager_mutex or ->lock should be held")
360 #define assert_manager_or_pool_lock(pool) do { } while (0)
363 #define for_each_cpu_worker_pool(pool, cpu) \
364 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
365 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
369 * for_each_pool - iterate through all worker_pools in the system
370 * @pool: iteration cursor
371 * @pi: integer used for iteration
373 * This must be called either with wq_pool_mutex held or sched RCU read
374 * locked. If the pool needs to be used beyond the locking in effect, the
375 * caller is responsible for guaranteeing that the pool stays online.
377 * The if/else clause exists only for the lockdep assertion and can be
380 #define for_each_pool(pool, pi) \
381 idr_for_each_entry(&worker_pool_idr, pool, pi) \
382 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
386 * for_each_pool_worker - iterate through all workers of a worker_pool
387 * @worker: iteration cursor
388 * @wi: integer used for iteration
389 * @pool: worker_pool to iterate workers of
391 * This must be called with either @pool->manager_mutex or ->lock held.
393 * The if/else clause exists only for the lockdep assertion and can be
396 #define for_each_pool_worker(worker, wi, pool) \
397 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
398 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
402 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
403 * @pwq: iteration cursor
404 * @wq: the target workqueue
406 * This must be called either with wq->mutex held or sched RCU read locked.
407 * If the pwq needs to be used beyond the locking in effect, the caller is
408 * responsible for guaranteeing that the pwq stays online.
410 * The if/else clause exists only for the lockdep assertion and can be
413 #define for_each_pwq(pwq, wq) \
414 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
415 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
418 #ifdef CONFIG_DEBUG_OBJECTS_WORK
420 static struct debug_obj_descr work_debug_descr;
422 static void *work_debug_hint(void *addr)
424 return ((struct work_struct *) addr)->func;
428 * fixup_init is called when:
429 * - an active object is initialized
431 static int work_fixup_init(void *addr, enum debug_obj_state state)
433 struct work_struct *work = addr;
436 case ODEBUG_STATE_ACTIVE:
437 cancel_work_sync(work);
438 debug_object_init(work, &work_debug_descr);
446 * fixup_activate is called when:
447 * - an active object is activated
448 * - an unknown object is activated (might be a statically initialized object)
450 static int work_fixup_activate(void *addr, enum debug_obj_state state)
452 struct work_struct *work = addr;
456 case ODEBUG_STATE_NOTAVAILABLE:
458 * This is not really a fixup. The work struct was
459 * statically initialized. We just make sure that it
460 * is tracked in the object tracker.
462 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
463 debug_object_init(work, &work_debug_descr);
464 debug_object_activate(work, &work_debug_descr);
470 case ODEBUG_STATE_ACTIVE:
479 * fixup_free is called when:
480 * - an active object is freed
482 static int work_fixup_free(void *addr, enum debug_obj_state state)
484 struct work_struct *work = addr;
487 case ODEBUG_STATE_ACTIVE:
488 cancel_work_sync(work);
489 debug_object_free(work, &work_debug_descr);
496 static struct debug_obj_descr work_debug_descr = {
497 .name = "work_struct",
498 .debug_hint = work_debug_hint,
499 .fixup_init = work_fixup_init,
500 .fixup_activate = work_fixup_activate,
501 .fixup_free = work_fixup_free,
504 static inline void debug_work_activate(struct work_struct *work)
506 debug_object_activate(work, &work_debug_descr);
509 static inline void debug_work_deactivate(struct work_struct *work)
511 debug_object_deactivate(work, &work_debug_descr);
514 void __init_work(struct work_struct *work, int onstack)
517 debug_object_init_on_stack(work, &work_debug_descr);
519 debug_object_init(work, &work_debug_descr);
521 EXPORT_SYMBOL_GPL(__init_work);
523 void destroy_work_on_stack(struct work_struct *work)
525 debug_object_free(work, &work_debug_descr);
527 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
530 static inline void debug_work_activate(struct work_struct *work) { }
531 static inline void debug_work_deactivate(struct work_struct *work) { }
534 /* allocate ID and assign it to @pool */
535 static int worker_pool_assign_id(struct worker_pool *pool)
539 lockdep_assert_held(&wq_pool_mutex);
541 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
550 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
551 * @wq: the target workqueue
554 * This must be called either with pwq_lock held or sched RCU read locked.
555 * If the pwq needs to be used beyond the locking in effect, the caller is
556 * responsible for guaranteeing that the pwq stays online.
558 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
561 assert_rcu_or_wq_mutex(wq);
562 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
565 static unsigned int work_color_to_flags(int color)
567 return color << WORK_STRUCT_COLOR_SHIFT;
570 static int get_work_color(struct work_struct *work)
572 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
573 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
576 static int work_next_color(int color)
578 return (color + 1) % WORK_NR_COLORS;
582 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
583 * contain the pointer to the queued pwq. Once execution starts, the flag
584 * is cleared and the high bits contain OFFQ flags and pool ID.
586 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
587 * and clear_work_data() can be used to set the pwq, pool or clear
588 * work->data. These functions should only be called while the work is
589 * owned - ie. while the PENDING bit is set.
591 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
592 * corresponding to a work. Pool is available once the work has been
593 * queued anywhere after initialization until it is sync canceled. pwq is
594 * available only while the work item is queued.
596 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
597 * canceled. While being canceled, a work item may have its PENDING set
598 * but stay off timer and worklist for arbitrarily long and nobody should
599 * try to steal the PENDING bit.
601 static inline void set_work_data(struct work_struct *work, unsigned long data,
604 WARN_ON_ONCE(!work_pending(work));
605 atomic_long_set(&work->data, data | flags | work_static(work));
608 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
609 unsigned long extra_flags)
611 set_work_data(work, (unsigned long)pwq,
612 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
615 static void set_work_pool_and_keep_pending(struct work_struct *work,
618 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
619 WORK_STRUCT_PENDING);
622 static void set_work_pool_and_clear_pending(struct work_struct *work,
626 * The following wmb is paired with the implied mb in
627 * test_and_set_bit(PENDING) and ensures all updates to @work made
628 * here are visible to and precede any updates by the next PENDING
632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
635 static void clear_work_data(struct work_struct *work)
637 smp_wmb(); /* see set_work_pool_and_clear_pending() */
638 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
641 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
643 unsigned long data = atomic_long_read(&work->data);
645 if (data & WORK_STRUCT_PWQ)
646 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
652 * get_work_pool - return the worker_pool a given work was associated with
653 * @work: the work item of interest
655 * Return the worker_pool @work was last associated with. %NULL if none.
657 * Pools are created and destroyed under wq_pool_mutex, and allows read
658 * access under sched-RCU read lock. As such, this function should be
659 * called under wq_pool_mutex or with preemption disabled.
661 * All fields of the returned pool are accessible as long as the above
662 * mentioned locking is in effect. If the returned pool needs to be used
663 * beyond the critical section, the caller is responsible for ensuring the
664 * returned pool is and stays online.
666 static struct worker_pool *get_work_pool(struct work_struct *work)
668 unsigned long data = atomic_long_read(&work->data);
671 assert_rcu_or_pool_mutex();
673 if (data & WORK_STRUCT_PWQ)
674 return ((struct pool_workqueue *)
675 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
677 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
678 if (pool_id == WORK_OFFQ_POOL_NONE)
681 return idr_find(&worker_pool_idr, pool_id);
685 * get_work_pool_id - return the worker pool ID a given work is associated with
686 * @work: the work item of interest
688 * Return the worker_pool ID @work was last associated with.
689 * %WORK_OFFQ_POOL_NONE if none.
691 static int get_work_pool_id(struct work_struct *work)
693 unsigned long data = atomic_long_read(&work->data);
695 if (data & WORK_STRUCT_PWQ)
696 return ((struct pool_workqueue *)
697 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
699 return data >> WORK_OFFQ_POOL_SHIFT;
702 static void mark_work_canceling(struct work_struct *work)
704 unsigned long pool_id = get_work_pool_id(work);
706 pool_id <<= WORK_OFFQ_POOL_SHIFT;
707 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
710 static bool work_is_canceling(struct work_struct *work)
712 unsigned long data = atomic_long_read(&work->data);
714 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
718 * Policy functions. These define the policies on how the global worker
719 * pools are managed. Unless noted otherwise, these functions assume that
720 * they're being called with pool->lock held.
723 static bool __need_more_worker(struct worker_pool *pool)
725 return !atomic_read(&pool->nr_running);
729 * Need to wake up a worker? Called from anything but currently
732 * Note that, because unbound workers never contribute to nr_running, this
733 * function will always return %true for unbound pools as long as the
734 * worklist isn't empty.
736 static bool need_more_worker(struct worker_pool *pool)
738 return !list_empty(&pool->worklist) && __need_more_worker(pool);
741 /* Can I start working? Called from busy but !running workers. */
742 static bool may_start_working(struct worker_pool *pool)
744 return pool->nr_idle;
747 /* Do I need to keep working? Called from currently running workers. */
748 static bool keep_working(struct worker_pool *pool)
750 return !list_empty(&pool->worklist) &&
751 atomic_read(&pool->nr_running) <= 1;
754 /* Do we need a new worker? Called from manager. */
755 static bool need_to_create_worker(struct worker_pool *pool)
757 return need_more_worker(pool) && !may_start_working(pool);
760 /* Do I need to be the manager? */
761 static bool need_to_manage_workers(struct worker_pool *pool)
763 return need_to_create_worker(pool) ||
764 (pool->flags & POOL_MANAGE_WORKERS);
767 /* Do we have too many workers and should some go away? */
768 static bool too_many_workers(struct worker_pool *pool)
770 bool managing = mutex_is_locked(&pool->manager_arb);
771 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
772 int nr_busy = pool->nr_workers - nr_idle;
775 * nr_idle and idle_list may disagree if idle rebinding is in
776 * progress. Never return %true if idle_list is empty.
778 if (list_empty(&pool->idle_list))
781 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
788 /* Return the first worker. Safe with preemption disabled */
789 static struct worker *first_worker(struct worker_pool *pool)
791 if (unlikely(list_empty(&pool->idle_list)))
794 return list_first_entry(&pool->idle_list, struct worker, entry);
798 * wake_up_worker - wake up an idle worker
799 * @pool: worker pool to wake worker from
801 * Wake up the first idle worker of @pool.
804 * spin_lock_irq(pool->lock).
806 static void wake_up_worker(struct worker_pool *pool)
808 struct worker *worker = first_worker(pool);
811 wake_up_process(worker->task);
815 * wq_worker_waking_up - a worker is waking up
816 * @task: task waking up
817 * @cpu: CPU @task is waking up to
819 * This function is called during try_to_wake_up() when a worker is
823 * spin_lock_irq(rq->lock)
825 void wq_worker_waking_up(struct task_struct *task, int cpu)
827 struct worker *worker = kthread_data(task);
829 if (!(worker->flags & WORKER_NOT_RUNNING)) {
830 WARN_ON_ONCE(worker->pool->cpu != cpu);
831 atomic_inc(&worker->pool->nr_running);
836 * wq_worker_sleeping - a worker is going to sleep
837 * @task: task going to sleep
838 * @cpu: CPU in question, must be the current CPU number
840 * This function is called during schedule() when a busy worker is
841 * going to sleep. Worker on the same cpu can be woken up by
842 * returning pointer to its task.
845 * spin_lock_irq(rq->lock)
848 * Worker task on @cpu to wake up, %NULL if none.
850 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
852 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
853 struct worker_pool *pool;
856 * Rescuers, which may not have all the fields set up like normal
857 * workers, also reach here, let's not access anything before
858 * checking NOT_RUNNING.
860 if (worker->flags & WORKER_NOT_RUNNING)
865 /* this can only happen on the local cpu */
866 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
870 * The counterpart of the following dec_and_test, implied mb,
871 * worklist not empty test sequence is in insert_work().
872 * Please read comment there.
874 * NOT_RUNNING is clear. This means that we're bound to and
875 * running on the local cpu w/ rq lock held and preemption
876 * disabled, which in turn means that none else could be
877 * manipulating idle_list, so dereferencing idle_list without pool
880 if (atomic_dec_and_test(&pool->nr_running) &&
881 !list_empty(&pool->worklist))
882 to_wakeup = first_worker(pool);
883 return to_wakeup ? to_wakeup->task : NULL;
887 * worker_set_flags - set worker flags and adjust nr_running accordingly
889 * @flags: flags to set
890 * @wakeup: wakeup an idle worker if necessary
892 * Set @flags in @worker->flags and adjust nr_running accordingly. If
893 * nr_running becomes zero and @wakeup is %true, an idle worker is
897 * spin_lock_irq(pool->lock)
899 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
902 struct worker_pool *pool = worker->pool;
904 WARN_ON_ONCE(worker->task != current);
907 * If transitioning into NOT_RUNNING, adjust nr_running and
908 * wake up an idle worker as necessary if requested by
911 if ((flags & WORKER_NOT_RUNNING) &&
912 !(worker->flags & WORKER_NOT_RUNNING)) {
914 if (atomic_dec_and_test(&pool->nr_running) &&
915 !list_empty(&pool->worklist))
916 wake_up_worker(pool);
918 atomic_dec(&pool->nr_running);
921 worker->flags |= flags;
925 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
927 * @flags: flags to clear
929 * Clear @flags in @worker->flags and adjust nr_running accordingly.
932 * spin_lock_irq(pool->lock)
934 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
936 struct worker_pool *pool = worker->pool;
937 unsigned int oflags = worker->flags;
939 WARN_ON_ONCE(worker->task != current);
941 worker->flags &= ~flags;
944 * If transitioning out of NOT_RUNNING, increment nr_running. Note
945 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
946 * of multiple flags, not a single flag.
948 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
949 if (!(worker->flags & WORKER_NOT_RUNNING))
950 atomic_inc(&pool->nr_running);
954 * find_worker_executing_work - find worker which is executing a work
955 * @pool: pool of interest
956 * @work: work to find worker for
958 * Find a worker which is executing @work on @pool by searching
959 * @pool->busy_hash which is keyed by the address of @work. For a worker
960 * to match, its current execution should match the address of @work and
961 * its work function. This is to avoid unwanted dependency between
962 * unrelated work executions through a work item being recycled while still
965 * This is a bit tricky. A work item may be freed once its execution
966 * starts and nothing prevents the freed area from being recycled for
967 * another work item. If the same work item address ends up being reused
968 * before the original execution finishes, workqueue will identify the
969 * recycled work item as currently executing and make it wait until the
970 * current execution finishes, introducing an unwanted dependency.
972 * This function checks the work item address and work function to avoid
973 * false positives. Note that this isn't complete as one may construct a
974 * work function which can introduce dependency onto itself through a
975 * recycled work item. Well, if somebody wants to shoot oneself in the
976 * foot that badly, there's only so much we can do, and if such deadlock
977 * actually occurs, it should be easy to locate the culprit work function.
980 * spin_lock_irq(pool->lock).
983 * Pointer to worker which is executing @work if found, NULL
986 static struct worker *find_worker_executing_work(struct worker_pool *pool,
987 struct work_struct *work)
989 struct worker *worker;
991 hash_for_each_possible(pool->busy_hash, worker, hentry,
993 if (worker->current_work == work &&
994 worker->current_func == work->func)
1001 * move_linked_works - move linked works to a list
1002 * @work: start of series of works to be scheduled
1003 * @head: target list to append @work to
1004 * @nextp: out paramter for nested worklist walking
1006 * Schedule linked works starting from @work to @head. Work series to
1007 * be scheduled starts at @work and includes any consecutive work with
1008 * WORK_STRUCT_LINKED set in its predecessor.
1010 * If @nextp is not NULL, it's updated to point to the next work of
1011 * the last scheduled work. This allows move_linked_works() to be
1012 * nested inside outer list_for_each_entry_safe().
1015 * spin_lock_irq(pool->lock).
1017 static void move_linked_works(struct work_struct *work, struct list_head *head,
1018 struct work_struct **nextp)
1020 struct work_struct *n;
1023 * Linked worklist will always end before the end of the list,
1024 * use NULL for list head.
1026 list_for_each_entry_safe_from(work, n, NULL, entry) {
1027 list_move_tail(&work->entry, head);
1028 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1033 * If we're already inside safe list traversal and have moved
1034 * multiple works to the scheduled queue, the next position
1035 * needs to be updated.
1042 * get_pwq - get an extra reference on the specified pool_workqueue
1043 * @pwq: pool_workqueue to get
1045 * Obtain an extra reference on @pwq. The caller should guarantee that
1046 * @pwq has positive refcnt and be holding the matching pool->lock.
1048 static void get_pwq(struct pool_workqueue *pwq)
1050 lockdep_assert_held(&pwq->pool->lock);
1051 WARN_ON_ONCE(pwq->refcnt <= 0);
1056 * put_pwq - put a pool_workqueue reference
1057 * @pwq: pool_workqueue to put
1059 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1060 * destruction. The caller should be holding the matching pool->lock.
1062 static void put_pwq(struct pool_workqueue *pwq)
1064 lockdep_assert_held(&pwq->pool->lock);
1065 if (likely(--pwq->refcnt))
1067 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1070 * @pwq can't be released under pool->lock, bounce to
1071 * pwq_unbound_release_workfn(). This never recurses on the same
1072 * pool->lock as this path is taken only for unbound workqueues and
1073 * the release work item is scheduled on a per-cpu workqueue. To
1074 * avoid lockdep warning, unbound pool->locks are given lockdep
1075 * subclass of 1 in get_unbound_pool().
1077 schedule_work(&pwq->unbound_release_work);
1081 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1082 * @pwq: pool_workqueue to put (can be %NULL)
1084 * put_pwq() with locking. This function also allows %NULL @pwq.
1086 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1090 * As both pwqs and pools are sched-RCU protected, the
1091 * following lock operations are safe.
1093 spin_lock_irq(&pwq->pool->lock);
1095 spin_unlock_irq(&pwq->pool->lock);
1099 static void pwq_activate_delayed_work(struct work_struct *work)
1101 struct pool_workqueue *pwq = get_work_pwq(work);
1103 trace_workqueue_activate_work(work);
1104 move_linked_works(work, &pwq->pool->worklist, NULL);
1105 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1109 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1111 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1112 struct work_struct, entry);
1114 pwq_activate_delayed_work(work);
1118 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1119 * @pwq: pwq of interest
1120 * @color: color of work which left the queue
1122 * A work either has completed or is removed from pending queue,
1123 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1126 * spin_lock_irq(pool->lock).
1128 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1130 /* uncolored work items don't participate in flushing or nr_active */
1131 if (color == WORK_NO_COLOR)
1134 pwq->nr_in_flight[color]--;
1137 if (!list_empty(&pwq->delayed_works)) {
1138 /* one down, submit a delayed one */
1139 if (pwq->nr_active < pwq->max_active)
1140 pwq_activate_first_delayed(pwq);
1143 /* is flush in progress and are we at the flushing tip? */
1144 if (likely(pwq->flush_color != color))
1147 /* are there still in-flight works? */
1148 if (pwq->nr_in_flight[color])
1151 /* this pwq is done, clear flush_color */
1152 pwq->flush_color = -1;
1155 * If this was the last pwq, wake up the first flusher. It
1156 * will handle the rest.
1158 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1159 complete(&pwq->wq->first_flusher->done);
1165 * try_to_grab_pending - steal work item from worklist and disable irq
1166 * @work: work item to steal
1167 * @is_dwork: @work is a delayed_work
1168 * @flags: place to store irq state
1170 * Try to grab PENDING bit of @work. This function can handle @work in any
1171 * stable state - idle, on timer or on worklist. Return values are
1173 * 1 if @work was pending and we successfully stole PENDING
1174 * 0 if @work was idle and we claimed PENDING
1175 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1176 * -ENOENT if someone else is canceling @work, this state may persist
1177 * for arbitrarily long
1179 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1180 * interrupted while holding PENDING and @work off queue, irq must be
1181 * disabled on entry. This, combined with delayed_work->timer being
1182 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1184 * On successful return, >= 0, irq is disabled and the caller is
1185 * responsible for releasing it using local_irq_restore(*@flags).
1187 * This function is safe to call from any context including IRQ handler.
1189 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1190 unsigned long *flags)
1192 struct worker_pool *pool;
1193 struct pool_workqueue *pwq;
1195 local_irq_save(*flags);
1197 /* try to steal the timer if it exists */
1199 struct delayed_work *dwork = to_delayed_work(work);
1202 * dwork->timer is irqsafe. If del_timer() fails, it's
1203 * guaranteed that the timer is not queued anywhere and not
1204 * running on the local CPU.
1206 if (likely(del_timer(&dwork->timer)))
1210 /* try to claim PENDING the normal way */
1211 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1215 * The queueing is in progress, or it is already queued. Try to
1216 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1218 pool = get_work_pool(work);
1222 spin_lock(&pool->lock);
1224 * work->data is guaranteed to point to pwq only while the work
1225 * item is queued on pwq->wq, and both updating work->data to point
1226 * to pwq on queueing and to pool on dequeueing are done under
1227 * pwq->pool->lock. This in turn guarantees that, if work->data
1228 * points to pwq which is associated with a locked pool, the work
1229 * item is currently queued on that pool.
1231 pwq = get_work_pwq(work);
1232 if (pwq && pwq->pool == pool) {
1233 debug_work_deactivate(work);
1236 * A delayed work item cannot be grabbed directly because
1237 * it might have linked NO_COLOR work items which, if left
1238 * on the delayed_list, will confuse pwq->nr_active
1239 * management later on and cause stall. Make sure the work
1240 * item is activated before grabbing.
1242 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1243 pwq_activate_delayed_work(work);
1245 list_del_init(&work->entry);
1246 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1248 /* work->data points to pwq iff queued, point to pool */
1249 set_work_pool_and_keep_pending(work, pool->id);
1251 spin_unlock(&pool->lock);
1254 spin_unlock(&pool->lock);
1256 local_irq_restore(*flags);
1257 if (work_is_canceling(work))
1264 * insert_work - insert a work into a pool
1265 * @pwq: pwq @work belongs to
1266 * @work: work to insert
1267 * @head: insertion point
1268 * @extra_flags: extra WORK_STRUCT_* flags to set
1270 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1271 * work_struct flags.
1274 * spin_lock_irq(pool->lock).
1276 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1277 struct list_head *head, unsigned int extra_flags)
1279 struct worker_pool *pool = pwq->pool;
1281 /* we own @work, set data and link */
1282 set_work_pwq(work, pwq, extra_flags);
1283 list_add_tail(&work->entry, head);
1287 * Ensure either wq_worker_sleeping() sees the above
1288 * list_add_tail() or we see zero nr_running to avoid workers lying
1289 * around lazily while there are works to be processed.
1293 if (__need_more_worker(pool))
1294 wake_up_worker(pool);
1298 * Test whether @work is being queued from another work executing on the
1301 static bool is_chained_work(struct workqueue_struct *wq)
1303 struct worker *worker;
1305 worker = current_wq_worker();
1307 * Return %true iff I'm a worker execuing a work item on @wq. If
1308 * I'm @worker, it's safe to dereference it without locking.
1310 return worker && worker->current_pwq->wq == wq;
1313 static void __queue_work(int cpu, struct workqueue_struct *wq,
1314 struct work_struct *work)
1316 struct pool_workqueue *pwq;
1317 struct worker_pool *last_pool;
1318 struct list_head *worklist;
1319 unsigned int work_flags;
1320 unsigned int req_cpu = cpu;
1323 * While a work item is PENDING && off queue, a task trying to
1324 * steal the PENDING will busy-loop waiting for it to either get
1325 * queued or lose PENDING. Grabbing PENDING and queueing should
1326 * happen with IRQ disabled.
1328 WARN_ON_ONCE(!irqs_disabled());
1330 debug_work_activate(work);
1332 /* if dying, only works from the same workqueue are allowed */
1333 if (unlikely(wq->flags & __WQ_DRAINING) &&
1334 WARN_ON_ONCE(!is_chained_work(wq)))
1337 if (req_cpu == WORK_CPU_UNBOUND)
1338 cpu = raw_smp_processor_id();
1340 /* pwq which will be used unless @work is executing elsewhere */
1341 if (!(wq->flags & WQ_UNBOUND))
1342 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1344 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1347 * If @work was previously on a different pool, it might still be
1348 * running there, in which case the work needs to be queued on that
1349 * pool to guarantee non-reentrancy.
1351 last_pool = get_work_pool(work);
1352 if (last_pool && last_pool != pwq->pool) {
1353 struct worker *worker;
1355 spin_lock(&last_pool->lock);
1357 worker = find_worker_executing_work(last_pool, work);
1359 if (worker && worker->current_pwq->wq == wq) {
1360 pwq = worker->current_pwq;
1362 /* meh... not running there, queue here */
1363 spin_unlock(&last_pool->lock);
1364 spin_lock(&pwq->pool->lock);
1367 spin_lock(&pwq->pool->lock);
1371 * pwq is determined and locked. For unbound pools, we could have
1372 * raced with pwq release and it could already be dead. If its
1373 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1374 * without another pwq replacing it in the numa_pwq_tbl or while
1375 * work items are executing on it, so the retrying is guaranteed to
1376 * make forward-progress.
1378 if (unlikely(!pwq->refcnt)) {
1379 if (wq->flags & WQ_UNBOUND) {
1380 spin_unlock(&pwq->pool->lock);
1385 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1389 /* pwq determined, queue */
1390 trace_workqueue_queue_work(req_cpu, pwq, work);
1392 if (WARN_ON(!list_empty(&work->entry))) {
1393 spin_unlock(&pwq->pool->lock);
1397 pwq->nr_in_flight[pwq->work_color]++;
1398 work_flags = work_color_to_flags(pwq->work_color);
1400 if (likely(pwq->nr_active < pwq->max_active)) {
1401 trace_workqueue_activate_work(work);
1403 worklist = &pwq->pool->worklist;
1405 work_flags |= WORK_STRUCT_DELAYED;
1406 worklist = &pwq->delayed_works;
1409 insert_work(pwq, work, worklist, work_flags);
1411 spin_unlock(&pwq->pool->lock);
1415 * queue_work_on - queue work on specific cpu
1416 * @cpu: CPU number to execute work on
1417 * @wq: workqueue to use
1418 * @work: work to queue
1420 * Returns %false if @work was already on a queue, %true otherwise.
1422 * We queue the work to a specific CPU, the caller must ensure it
1425 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1426 struct work_struct *work)
1429 unsigned long flags;
1431 local_irq_save(flags);
1433 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1434 __queue_work(cpu, wq, work);
1438 local_irq_restore(flags);
1441 EXPORT_SYMBOL(queue_work_on);
1443 void delayed_work_timer_fn(unsigned long __data)
1445 struct delayed_work *dwork = (struct delayed_work *)__data;
1447 /* should have been called from irqsafe timer with irq already off */
1448 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1450 EXPORT_SYMBOL(delayed_work_timer_fn);
1452 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1453 struct delayed_work *dwork, unsigned long delay)
1455 struct timer_list *timer = &dwork->timer;
1456 struct work_struct *work = &dwork->work;
1458 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1459 timer->data != (unsigned long)dwork);
1460 WARN_ON_ONCE(timer_pending(timer));
1461 WARN_ON_ONCE(!list_empty(&work->entry));
1464 * If @delay is 0, queue @dwork->work immediately. This is for
1465 * both optimization and correctness. The earliest @timer can
1466 * expire is on the closest next tick and delayed_work users depend
1467 * on that there's no such delay when @delay is 0.
1470 __queue_work(cpu, wq, &dwork->work);
1474 timer_stats_timer_set_start_info(&dwork->timer);
1478 timer->expires = jiffies + delay;
1480 if (unlikely(cpu != WORK_CPU_UNBOUND))
1481 add_timer_on(timer, cpu);
1487 * queue_delayed_work_on - queue work on specific CPU after delay
1488 * @cpu: CPU number to execute work on
1489 * @wq: workqueue to use
1490 * @dwork: work to queue
1491 * @delay: number of jiffies to wait before queueing
1493 * Returns %false if @work was already on a queue, %true otherwise. If
1494 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1497 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1498 struct delayed_work *dwork, unsigned long delay)
1500 struct work_struct *work = &dwork->work;
1502 unsigned long flags;
1504 /* read the comment in __queue_work() */
1505 local_irq_save(flags);
1507 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1508 __queue_delayed_work(cpu, wq, dwork, delay);
1512 local_irq_restore(flags);
1515 EXPORT_SYMBOL(queue_delayed_work_on);
1518 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1519 * @cpu: CPU number to execute work on
1520 * @wq: workqueue to use
1521 * @dwork: work to queue
1522 * @delay: number of jiffies to wait before queueing
1524 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1525 * modify @dwork's timer so that it expires after @delay. If @delay is
1526 * zero, @work is guaranteed to be scheduled immediately regardless of its
1529 * Returns %false if @dwork was idle and queued, %true if @dwork was
1530 * pending and its timer was modified.
1532 * This function is safe to call from any context including IRQ handler.
1533 * See try_to_grab_pending() for details.
1535 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1536 struct delayed_work *dwork, unsigned long delay)
1538 unsigned long flags;
1542 ret = try_to_grab_pending(&dwork->work, true, &flags);
1543 } while (unlikely(ret == -EAGAIN));
1545 if (likely(ret >= 0)) {
1546 __queue_delayed_work(cpu, wq, dwork, delay);
1547 local_irq_restore(flags);
1550 /* -ENOENT from try_to_grab_pending() becomes %true */
1553 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1556 * worker_enter_idle - enter idle state
1557 * @worker: worker which is entering idle state
1559 * @worker is entering idle state. Update stats and idle timer if
1563 * spin_lock_irq(pool->lock).
1565 static void worker_enter_idle(struct worker *worker)
1567 struct worker_pool *pool = worker->pool;
1569 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1570 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1571 (worker->hentry.next || worker->hentry.pprev)))
1574 /* can't use worker_set_flags(), also called from start_worker() */
1575 worker->flags |= WORKER_IDLE;
1577 worker->last_active = jiffies;
1579 /* idle_list is LIFO */
1580 list_add(&worker->entry, &pool->idle_list);
1582 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1583 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1586 * Sanity check nr_running. Because wq_unbind_fn() releases
1587 * pool->lock between setting %WORKER_UNBOUND and zapping
1588 * nr_running, the warning may trigger spuriously. Check iff
1589 * unbind is not in progress.
1591 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1592 pool->nr_workers == pool->nr_idle &&
1593 atomic_read(&pool->nr_running));
1597 * worker_leave_idle - leave idle state
1598 * @worker: worker which is leaving idle state
1600 * @worker is leaving idle state. Update stats.
1603 * spin_lock_irq(pool->lock).
1605 static void worker_leave_idle(struct worker *worker)
1607 struct worker_pool *pool = worker->pool;
1609 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1611 worker_clr_flags(worker, WORKER_IDLE);
1613 list_del_init(&worker->entry);
1617 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1618 * @pool: target worker_pool
1620 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1622 * Works which are scheduled while the cpu is online must at least be
1623 * scheduled to a worker which is bound to the cpu so that if they are
1624 * flushed from cpu callbacks while cpu is going down, they are
1625 * guaranteed to execute on the cpu.
1627 * This function is to be used by unbound workers and rescuers to bind
1628 * themselves to the target cpu and may race with cpu going down or
1629 * coming online. kthread_bind() can't be used because it may put the
1630 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1631 * verbatim as it's best effort and blocking and pool may be
1632 * [dis]associated in the meantime.
1634 * This function tries set_cpus_allowed() and locks pool and verifies the
1635 * binding against %POOL_DISASSOCIATED which is set during
1636 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1637 * enters idle state or fetches works without dropping lock, it can
1638 * guarantee the scheduling requirement described in the first paragraph.
1641 * Might sleep. Called without any lock but returns with pool->lock
1645 * %true if the associated pool is online (@worker is successfully
1646 * bound), %false if offline.
1648 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1649 __acquires(&pool->lock)
1653 * The following call may fail, succeed or succeed
1654 * without actually migrating the task to the cpu if
1655 * it races with cpu hotunplug operation. Verify
1656 * against POOL_DISASSOCIATED.
1658 if (!(pool->flags & POOL_DISASSOCIATED))
1659 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1661 spin_lock_irq(&pool->lock);
1662 if (pool->flags & POOL_DISASSOCIATED)
1664 if (task_cpu(current) == pool->cpu &&
1665 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1667 spin_unlock_irq(&pool->lock);
1670 * We've raced with CPU hot[un]plug. Give it a breather
1671 * and retry migration. cond_resched() is required here;
1672 * otherwise, we might deadlock against cpu_stop trying to
1673 * bring down the CPU on non-preemptive kernel.
1680 static struct worker *alloc_worker(void)
1682 struct worker *worker;
1684 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1686 INIT_LIST_HEAD(&worker->entry);
1687 INIT_LIST_HEAD(&worker->scheduled);
1688 /* on creation a worker is in !idle && prep state */
1689 worker->flags = WORKER_PREP;
1695 * create_worker - create a new workqueue worker
1696 * @pool: pool the new worker will belong to
1698 * Create a new worker which is bound to @pool. The returned worker
1699 * can be started by calling start_worker() or destroyed using
1703 * Might sleep. Does GFP_KERNEL allocations.
1706 * Pointer to the newly created worker.
1708 static struct worker *create_worker(struct worker_pool *pool)
1710 struct worker *worker = NULL;
1714 lockdep_assert_held(&pool->manager_mutex);
1717 * ID is needed to determine kthread name. Allocate ID first
1718 * without installing the pointer.
1720 idr_preload(GFP_KERNEL);
1721 spin_lock_irq(&pool->lock);
1723 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1725 spin_unlock_irq(&pool->lock);
1730 worker = alloc_worker();
1734 worker->pool = pool;
1738 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1739 pool->attrs->nice < 0 ? "H" : "");
1741 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1743 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1744 "kworker/%s", id_buf);
1745 if (IS_ERR(worker->task))
1749 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1750 * online CPUs. It'll be re-applied when any of the CPUs come up.
1752 set_user_nice(worker->task, pool->attrs->nice);
1753 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1755 /* prevent userland from meddling with cpumask of workqueue workers */
1756 worker->task->flags |= PF_NO_SETAFFINITY;
1759 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1760 * remains stable across this function. See the comments above the
1761 * flag definition for details.
1763 if (pool->flags & POOL_DISASSOCIATED)
1764 worker->flags |= WORKER_UNBOUND;
1766 /* successful, commit the pointer to idr */
1767 spin_lock_irq(&pool->lock);
1768 idr_replace(&pool->worker_idr, worker, worker->id);
1769 spin_unlock_irq(&pool->lock);
1775 spin_lock_irq(&pool->lock);
1776 idr_remove(&pool->worker_idr, id);
1777 spin_unlock_irq(&pool->lock);
1784 * start_worker - start a newly created worker
1785 * @worker: worker to start
1787 * Make the pool aware of @worker and start it.
1790 * spin_lock_irq(pool->lock).
1792 static void start_worker(struct worker *worker)
1794 worker->flags |= WORKER_STARTED;
1795 worker->pool->nr_workers++;
1796 worker_enter_idle(worker);
1797 wake_up_process(worker->task);
1801 * create_and_start_worker - create and start a worker for a pool
1802 * @pool: the target pool
1804 * Grab the managership of @pool and create and start a new worker for it.
1806 static int create_and_start_worker(struct worker_pool *pool)
1808 struct worker *worker;
1810 mutex_lock(&pool->manager_mutex);
1812 worker = create_worker(pool);
1814 spin_lock_irq(&pool->lock);
1815 start_worker(worker);
1816 spin_unlock_irq(&pool->lock);
1819 mutex_unlock(&pool->manager_mutex);
1821 return worker ? 0 : -ENOMEM;
1825 * destroy_worker - destroy a workqueue worker
1826 * @worker: worker to be destroyed
1828 * Destroy @worker and adjust @pool stats accordingly.
1831 * spin_lock_irq(pool->lock) which is released and regrabbed.
1833 static void destroy_worker(struct worker *worker)
1835 struct worker_pool *pool = worker->pool;
1837 lockdep_assert_held(&pool->manager_mutex);
1838 lockdep_assert_held(&pool->lock);
1840 /* sanity check frenzy */
1841 if (WARN_ON(worker->current_work) ||
1842 WARN_ON(!list_empty(&worker->scheduled)))
1845 if (worker->flags & WORKER_STARTED)
1847 if (worker->flags & WORKER_IDLE)
1851 * Once WORKER_DIE is set, the kworker may destroy itself at any
1852 * point. Pin to ensure the task stays until we're done with it.
1854 get_task_struct(worker->task);
1856 list_del_init(&worker->entry);
1857 worker->flags |= WORKER_DIE;
1859 idr_remove(&pool->worker_idr, worker->id);
1861 spin_unlock_irq(&pool->lock);
1863 kthread_stop(worker->task);
1864 put_task_struct(worker->task);
1867 spin_lock_irq(&pool->lock);
1870 static void idle_worker_timeout(unsigned long __pool)
1872 struct worker_pool *pool = (void *)__pool;
1874 spin_lock_irq(&pool->lock);
1876 if (too_many_workers(pool)) {
1877 struct worker *worker;
1878 unsigned long expires;
1880 /* idle_list is kept in LIFO order, check the last one */
1881 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1882 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1884 if (time_before(jiffies, expires))
1885 mod_timer(&pool->idle_timer, expires);
1887 /* it's been idle for too long, wake up manager */
1888 pool->flags |= POOL_MANAGE_WORKERS;
1889 wake_up_worker(pool);
1893 spin_unlock_irq(&pool->lock);
1896 static void send_mayday(struct work_struct *work)
1898 struct pool_workqueue *pwq = get_work_pwq(work);
1899 struct workqueue_struct *wq = pwq->wq;
1901 lockdep_assert_held(&wq_mayday_lock);
1906 /* mayday mayday mayday */
1907 if (list_empty(&pwq->mayday_node)) {
1909 * If @pwq is for an unbound wq, its base ref may be put at
1910 * any time due to an attribute change. Pin @pwq until the
1911 * rescuer is done with it.
1914 list_add_tail(&pwq->mayday_node, &wq->maydays);
1915 wake_up_process(wq->rescuer->task);
1919 static void pool_mayday_timeout(unsigned long __pool)
1921 struct worker_pool *pool = (void *)__pool;
1922 struct work_struct *work;
1924 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1925 spin_lock(&pool->lock);
1927 if (need_to_create_worker(pool)) {
1929 * We've been trying to create a new worker but
1930 * haven't been successful. We might be hitting an
1931 * allocation deadlock. Send distress signals to
1934 list_for_each_entry(work, &pool->worklist, entry)
1938 spin_unlock(&pool->lock);
1939 spin_unlock_irq(&wq_mayday_lock);
1941 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1945 * maybe_create_worker - create a new worker if necessary
1946 * @pool: pool to create a new worker for
1948 * Create a new worker for @pool if necessary. @pool is guaranteed to
1949 * have at least one idle worker on return from this function. If
1950 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1951 * sent to all rescuers with works scheduled on @pool to resolve
1952 * possible allocation deadlock.
1954 * On return, need_to_create_worker() is guaranteed to be %false and
1955 * may_start_working() %true.
1958 * spin_lock_irq(pool->lock) which may be released and regrabbed
1959 * multiple times. Does GFP_KERNEL allocations. Called only from
1963 * %false if no action was taken and pool->lock stayed locked, %true
1966 static bool maybe_create_worker(struct worker_pool *pool)
1967 __releases(&pool->lock)
1968 __acquires(&pool->lock)
1970 if (!need_to_create_worker(pool))
1973 spin_unlock_irq(&pool->lock);
1975 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1976 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1979 struct worker *worker;
1981 worker = create_worker(pool);
1983 del_timer_sync(&pool->mayday_timer);
1984 spin_lock_irq(&pool->lock);
1985 start_worker(worker);
1986 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1991 if (!need_to_create_worker(pool))
1994 __set_current_state(TASK_INTERRUPTIBLE);
1995 schedule_timeout(CREATE_COOLDOWN);
1997 if (!need_to_create_worker(pool))
2001 del_timer_sync(&pool->mayday_timer);
2002 spin_lock_irq(&pool->lock);
2003 if (need_to_create_worker(pool))
2009 * maybe_destroy_worker - destroy workers which have been idle for a while
2010 * @pool: pool to destroy workers for
2012 * Destroy @pool workers which have been idle for longer than
2013 * IDLE_WORKER_TIMEOUT.
2016 * spin_lock_irq(pool->lock) which may be released and regrabbed
2017 * multiple times. Called only from manager.
2020 * %false if no action was taken and pool->lock stayed locked, %true
2023 static bool maybe_destroy_workers(struct worker_pool *pool)
2027 while (too_many_workers(pool)) {
2028 struct worker *worker;
2029 unsigned long expires;
2031 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2032 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2034 if (time_before(jiffies, expires)) {
2035 mod_timer(&pool->idle_timer, expires);
2039 destroy_worker(worker);
2047 * manage_workers - manage worker pool
2050 * Assume the manager role and manage the worker pool @worker belongs
2051 * to. At any given time, there can be only zero or one manager per
2052 * pool. The exclusion is handled automatically by this function.
2054 * The caller can safely start processing works on false return. On
2055 * true return, it's guaranteed that need_to_create_worker() is false
2056 * and may_start_working() is true.
2059 * spin_lock_irq(pool->lock) which may be released and regrabbed
2060 * multiple times. Does GFP_KERNEL allocations.
2063 * spin_lock_irq(pool->lock) which may be released and regrabbed
2064 * multiple times. Does GFP_KERNEL allocations.
2066 static bool manage_workers(struct worker *worker)
2068 struct worker_pool *pool = worker->pool;
2072 * Managership is governed by two mutexes - manager_arb and
2073 * manager_mutex. manager_arb handles arbitration of manager role.
2074 * Anyone who successfully grabs manager_arb wins the arbitration
2075 * and becomes the manager. mutex_trylock() on pool->manager_arb
2076 * failure while holding pool->lock reliably indicates that someone
2077 * else is managing the pool and the worker which failed trylock
2078 * can proceed to executing work items. This means that anyone
2079 * grabbing manager_arb is responsible for actually performing
2080 * manager duties. If manager_arb is grabbed and released without
2081 * actual management, the pool may stall indefinitely.
2083 * manager_mutex is used for exclusion of actual management
2084 * operations. The holder of manager_mutex can be sure that none
2085 * of management operations, including creation and destruction of
2086 * workers, won't take place until the mutex is released. Because
2087 * manager_mutex doesn't interfere with manager role arbitration,
2088 * it is guaranteed that the pool's management, while may be
2089 * delayed, won't be disturbed by someone else grabbing
2092 if (!mutex_trylock(&pool->manager_arb))
2096 * With manager arbitration won, manager_mutex would be free in
2097 * most cases. trylock first without dropping @pool->lock.
2099 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2100 spin_unlock_irq(&pool->lock);
2101 mutex_lock(&pool->manager_mutex);
2102 spin_lock_irq(&pool->lock);
2106 pool->flags &= ~POOL_MANAGE_WORKERS;
2109 * Destroy and then create so that may_start_working() is true
2112 ret |= maybe_destroy_workers(pool);
2113 ret |= maybe_create_worker(pool);
2115 mutex_unlock(&pool->manager_mutex);
2116 mutex_unlock(&pool->manager_arb);
2121 * process_one_work - process single work
2123 * @work: work to process
2125 * Process @work. This function contains all the logics necessary to
2126 * process a single work including synchronization against and
2127 * interaction with other workers on the same cpu, queueing and
2128 * flushing. As long as context requirement is met, any worker can
2129 * call this function to process a work.
2132 * spin_lock_irq(pool->lock) which is released and regrabbed.
2134 static void process_one_work(struct worker *worker, struct work_struct *work)
2135 __releases(&pool->lock)
2136 __acquires(&pool->lock)
2138 struct pool_workqueue *pwq = get_work_pwq(work);
2139 struct worker_pool *pool = worker->pool;
2140 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2142 struct worker *collision;
2143 #ifdef CONFIG_LOCKDEP
2145 * It is permissible to free the struct work_struct from
2146 * inside the function that is called from it, this we need to
2147 * take into account for lockdep too. To avoid bogus "held
2148 * lock freed" warnings as well as problems when looking into
2149 * work->lockdep_map, make a copy and use that here.
2151 struct lockdep_map lockdep_map;
2153 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2156 * Ensure we're on the correct CPU. DISASSOCIATED test is
2157 * necessary to avoid spurious warnings from rescuers servicing the
2158 * unbound or a disassociated pool.
2160 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2161 !(pool->flags & POOL_DISASSOCIATED) &&
2162 raw_smp_processor_id() != pool->cpu);
2165 * A single work shouldn't be executed concurrently by
2166 * multiple workers on a single cpu. Check whether anyone is
2167 * already processing the work. If so, defer the work to the
2168 * currently executing one.
2170 collision = find_worker_executing_work(pool, work);
2171 if (unlikely(collision)) {
2172 move_linked_works(work, &collision->scheduled, NULL);
2176 /* claim and dequeue */
2177 debug_work_deactivate(work);
2178 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2179 worker->current_work = work;
2180 worker->current_func = work->func;
2181 worker->current_pwq = pwq;
2182 work_color = get_work_color(work);
2184 list_del_init(&work->entry);
2187 * CPU intensive works don't participate in concurrency
2188 * management. They're the scheduler's responsibility.
2190 if (unlikely(cpu_intensive))
2191 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2194 * Unbound pool isn't concurrency managed and work items should be
2195 * executed ASAP. Wake up another worker if necessary.
2197 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2198 wake_up_worker(pool);
2201 * Record the last pool and clear PENDING which should be the last
2202 * update to @work. Also, do this inside @pool->lock so that
2203 * PENDING and queued state changes happen together while IRQ is
2206 set_work_pool_and_clear_pending(work, pool->id);
2208 spin_unlock_irq(&pool->lock);
2210 lock_map_acquire_read(&pwq->wq->lockdep_map);
2211 lock_map_acquire(&lockdep_map);
2212 trace_workqueue_execute_start(work);
2214 sec_debug_work_log(worker, work, worker->current_func, 1);
2216 #ifdef CONFIG_SPRD_DEBUG
2217 #ifndef CONFIG_64BIT
2218 sprd_debug_work_log(worker, work, worker->current_func);
2222 #if defined(CONFIG_SYSTEM_LOAD_ANALYZER)
2224 u64 work_start_time, work_end_time, work_delta_time;
2225 work_start_time = get_load_analyzer_time();
2227 worker->current_func(work);
2228 sec_debug_work_log(worker, work, worker->current_func, 2);
2230 #if defined(CONFIG_SYSTEM_LOAD_ANALYZER)
2231 work_end_time = get_load_analyzer_time();
2232 work_delta_time = work_end_time -work_start_time;
2233 __slp_store_work_history(work ,work->func ,work_start_time, work_end_time);
2237 * While we must be careful to not use "work" after this, the trace
2238 * point will only record its address.
2240 trace_workqueue_execute_end(work);
2241 lock_map_release(&lockdep_map);
2242 lock_map_release(&pwq->wq->lockdep_map);
2244 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2245 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2246 " last function: %pf\n",
2247 current->comm, preempt_count(), task_pid_nr(current),
2248 worker->current_func);
2249 debug_show_held_locks(current);
2254 * The following prevents a kworker from hogging CPU on !PREEMPT
2255 * kernels, where a requeueing work item waiting for something to
2256 * happen could deadlock with stop_machine as such work item could
2257 * indefinitely requeue itself while all other CPUs are trapped in
2262 spin_lock_irq(&pool->lock);
2264 /* clear cpu intensive status */
2265 if (unlikely(cpu_intensive))
2266 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2268 /* we're done with it, release */
2269 hash_del(&worker->hentry);
2270 worker->current_work = NULL;
2271 worker->current_func = NULL;
2272 worker->current_pwq = NULL;
2273 worker->desc_valid = false;
2274 pwq_dec_nr_in_flight(pwq, work_color);
2278 * process_scheduled_works - process scheduled works
2281 * Process all scheduled works. Please note that the scheduled list
2282 * may change while processing a work, so this function repeatedly
2283 * fetches a work from the top and executes it.
2286 * spin_lock_irq(pool->lock) which may be released and regrabbed
2289 static void process_scheduled_works(struct worker *worker)
2291 while (!list_empty(&worker->scheduled)) {
2292 struct work_struct *work = list_first_entry(&worker->scheduled,
2293 struct work_struct, entry);
2294 process_one_work(worker, work);
2299 * worker_thread - the worker thread function
2302 * The worker thread function. All workers belong to a worker_pool -
2303 * either a per-cpu one or dynamic unbound one. These workers process all
2304 * work items regardless of their specific target workqueue. The only
2305 * exception is work items which belong to workqueues with a rescuer which
2306 * will be explained in rescuer_thread().
2308 static int worker_thread(void *__worker)
2310 struct worker *worker = __worker;
2311 struct worker_pool *pool = worker->pool;
2313 /* tell the scheduler that this is a workqueue worker */
2314 worker->task->flags |= PF_WQ_WORKER;
2316 spin_lock_irq(&pool->lock);
2318 /* am I supposed to die? */
2319 if (unlikely(worker->flags & WORKER_DIE)) {
2320 spin_unlock_irq(&pool->lock);
2321 WARN_ON_ONCE(!list_empty(&worker->entry));
2322 worker->task->flags &= ~PF_WQ_WORKER;
2326 worker_leave_idle(worker);
2328 /* no more worker necessary? */
2329 if (!need_more_worker(pool))
2332 /* do we need to manage? */
2333 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2337 * ->scheduled list can only be filled while a worker is
2338 * preparing to process a work or actually processing it.
2339 * Make sure nobody diddled with it while I was sleeping.
2341 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2344 * Finish PREP stage. We're guaranteed to have at least one idle
2345 * worker or that someone else has already assumed the manager
2346 * role. This is where @worker starts participating in concurrency
2347 * management if applicable and concurrency management is restored
2348 * after being rebound. See rebind_workers() for details.
2350 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2353 struct work_struct *work =
2354 list_first_entry(&pool->worklist,
2355 struct work_struct, entry);
2357 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2358 /* optimization path, not strictly necessary */
2359 process_one_work(worker, work);
2360 if (unlikely(!list_empty(&worker->scheduled)))
2361 process_scheduled_works(worker);
2363 move_linked_works(work, &worker->scheduled, NULL);
2364 process_scheduled_works(worker);
2366 } while (keep_working(pool));
2368 worker_set_flags(worker, WORKER_PREP, false);
2370 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2374 * pool->lock is held and there's no work to process and no need to
2375 * manage, sleep. Workers are woken up only while holding
2376 * pool->lock or from local cpu, so setting the current state
2377 * before releasing pool->lock is enough to prevent losing any
2380 worker_enter_idle(worker);
2381 __set_current_state(TASK_INTERRUPTIBLE);
2382 spin_unlock_irq(&pool->lock);
2388 * rescuer_thread - the rescuer thread function
2391 * Workqueue rescuer thread function. There's one rescuer for each
2392 * workqueue which has WQ_MEM_RECLAIM set.
2394 * Regular work processing on a pool may block trying to create a new
2395 * worker which uses GFP_KERNEL allocation which has slight chance of
2396 * developing into deadlock if some works currently on the same queue
2397 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2398 * the problem rescuer solves.
2400 * When such condition is possible, the pool summons rescuers of all
2401 * workqueues which have works queued on the pool and let them process
2402 * those works so that forward progress can be guaranteed.
2404 * This should happen rarely.
2406 static int rescuer_thread(void *__rescuer)
2408 struct worker *rescuer = __rescuer;
2409 struct workqueue_struct *wq = rescuer->rescue_wq;
2410 struct list_head *scheduled = &rescuer->scheduled;
2413 set_user_nice(current, RESCUER_NICE_LEVEL);
2416 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2417 * doesn't participate in concurrency management.
2419 rescuer->task->flags |= PF_WQ_WORKER;
2421 set_current_state(TASK_INTERRUPTIBLE);
2424 * By the time the rescuer is requested to stop, the workqueue
2425 * shouldn't have any work pending, but @wq->maydays may still have
2426 * pwq(s) queued. This can happen by non-rescuer workers consuming
2427 * all the work items before the rescuer got to them. Go through
2428 * @wq->maydays processing before acting on should_stop so that the
2429 * list is always empty on exit.
2431 should_stop = kthread_should_stop();
2433 /* see whether any pwq is asking for help */
2434 spin_lock_irq(&wq_mayday_lock);
2436 while (!list_empty(&wq->maydays)) {
2437 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2438 struct pool_workqueue, mayday_node);
2439 struct worker_pool *pool = pwq->pool;
2440 struct work_struct *work, *n;
2442 __set_current_state(TASK_RUNNING);
2443 list_del_init(&pwq->mayday_node);
2445 spin_unlock_irq(&wq_mayday_lock);
2447 /* migrate to the target cpu if possible */
2448 worker_maybe_bind_and_lock(pool);
2449 rescuer->pool = pool;
2452 * Slurp in all works issued via this workqueue and
2455 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2456 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2457 if (get_work_pwq(work) == pwq)
2458 move_linked_works(work, scheduled, &n);
2460 process_scheduled_works(rescuer);
2463 * Put the reference grabbed by send_mayday(). @pool won't
2464 * go away while we're holding its lock.
2469 * Leave this pool. If keep_working() is %true, notify a
2470 * regular worker; otherwise, we end up with 0 concurrency
2471 * and stalling the execution.
2473 if (keep_working(pool))
2474 wake_up_worker(pool);
2476 rescuer->pool = NULL;
2477 spin_unlock(&pool->lock);
2478 spin_lock(&wq_mayday_lock);
2481 spin_unlock_irq(&wq_mayday_lock);
2484 __set_current_state(TASK_RUNNING);
2485 rescuer->task->flags &= ~PF_WQ_WORKER;
2489 /* rescuers should never participate in concurrency management */
2490 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2496 struct work_struct work;
2497 struct completion done;
2500 static void wq_barrier_func(struct work_struct *work)
2502 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2503 complete(&barr->done);
2507 * insert_wq_barrier - insert a barrier work
2508 * @pwq: pwq to insert barrier into
2509 * @barr: wq_barrier to insert
2510 * @target: target work to attach @barr to
2511 * @worker: worker currently executing @target, NULL if @target is not executing
2513 * @barr is linked to @target such that @barr is completed only after
2514 * @target finishes execution. Please note that the ordering
2515 * guarantee is observed only with respect to @target and on the local
2518 * Currently, a queued barrier can't be canceled. This is because
2519 * try_to_grab_pending() can't determine whether the work to be
2520 * grabbed is at the head of the queue and thus can't clear LINKED
2521 * flag of the previous work while there must be a valid next work
2522 * after a work with LINKED flag set.
2524 * Note that when @worker is non-NULL, @target may be modified
2525 * underneath us, so we can't reliably determine pwq from @target.
2528 * spin_lock_irq(pool->lock).
2530 static void insert_wq_barrier(struct pool_workqueue *pwq,
2531 struct wq_barrier *barr,
2532 struct work_struct *target, struct worker *worker)
2534 struct list_head *head;
2535 unsigned int linked = 0;
2538 * debugobject calls are safe here even with pool->lock locked
2539 * as we know for sure that this will not trigger any of the
2540 * checks and call back into the fixup functions where we
2543 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2544 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2545 init_completion(&barr->done);
2548 * If @target is currently being executed, schedule the
2549 * barrier to the worker; otherwise, put it after @target.
2552 head = worker->scheduled.next;
2554 unsigned long *bits = work_data_bits(target);
2556 head = target->entry.next;
2557 /* there can already be other linked works, inherit and set */
2558 linked = *bits & WORK_STRUCT_LINKED;
2559 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2562 debug_work_activate(&barr->work);
2563 insert_work(pwq, &barr->work, head,
2564 work_color_to_flags(WORK_NO_COLOR) | linked);
2568 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2569 * @wq: workqueue being flushed
2570 * @flush_color: new flush color, < 0 for no-op
2571 * @work_color: new work color, < 0 for no-op
2573 * Prepare pwqs for workqueue flushing.
2575 * If @flush_color is non-negative, flush_color on all pwqs should be
2576 * -1. If no pwq has in-flight commands at the specified color, all
2577 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2578 * has in flight commands, its pwq->flush_color is set to
2579 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2580 * wakeup logic is armed and %true is returned.
2582 * The caller should have initialized @wq->first_flusher prior to
2583 * calling this function with non-negative @flush_color. If
2584 * @flush_color is negative, no flush color update is done and %false
2587 * If @work_color is non-negative, all pwqs should have the same
2588 * work_color which is previous to @work_color and all will be
2589 * advanced to @work_color.
2592 * mutex_lock(wq->mutex).
2595 * %true if @flush_color >= 0 and there's something to flush. %false
2598 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2599 int flush_color, int work_color)
2602 struct pool_workqueue *pwq;
2604 if (flush_color >= 0) {
2605 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2606 atomic_set(&wq->nr_pwqs_to_flush, 1);
2609 for_each_pwq(pwq, wq) {
2610 struct worker_pool *pool = pwq->pool;
2612 spin_lock_irq(&pool->lock);
2614 if (flush_color >= 0) {
2615 WARN_ON_ONCE(pwq->flush_color != -1);
2617 if (pwq->nr_in_flight[flush_color]) {
2618 pwq->flush_color = flush_color;
2619 atomic_inc(&wq->nr_pwqs_to_flush);
2624 if (work_color >= 0) {
2625 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2626 pwq->work_color = work_color;
2629 spin_unlock_irq(&pool->lock);
2632 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2633 complete(&wq->first_flusher->done);
2639 * flush_workqueue - ensure that any scheduled work has run to completion.
2640 * @wq: workqueue to flush
2642 * This function sleeps until all work items which were queued on entry
2643 * have finished execution, but it is not livelocked by new incoming ones.
2645 void flush_workqueue(struct workqueue_struct *wq)
2647 struct wq_flusher this_flusher = {
2648 .list = LIST_HEAD_INIT(this_flusher.list),
2650 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2654 lock_map_acquire(&wq->lockdep_map);
2655 lock_map_release(&wq->lockdep_map);
2657 mutex_lock(&wq->mutex);
2660 * Start-to-wait phase
2662 next_color = work_next_color(wq->work_color);
2664 if (next_color != wq->flush_color) {
2666 * Color space is not full. The current work_color
2667 * becomes our flush_color and work_color is advanced
2670 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2671 this_flusher.flush_color = wq->work_color;
2672 wq->work_color = next_color;
2674 if (!wq->first_flusher) {
2675 /* no flush in progress, become the first flusher */
2676 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2678 wq->first_flusher = &this_flusher;
2680 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2682 /* nothing to flush, done */
2683 wq->flush_color = next_color;
2684 wq->first_flusher = NULL;
2689 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2690 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2691 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2695 * Oops, color space is full, wait on overflow queue.
2696 * The next flush completion will assign us
2697 * flush_color and transfer to flusher_queue.
2699 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2702 mutex_unlock(&wq->mutex);
2704 wait_for_completion(&this_flusher.done);
2707 * Wake-up-and-cascade phase
2709 * First flushers are responsible for cascading flushes and
2710 * handling overflow. Non-first flushers can simply return.
2712 if (wq->first_flusher != &this_flusher)
2715 mutex_lock(&wq->mutex);
2717 /* we might have raced, check again with mutex held */
2718 if (wq->first_flusher != &this_flusher)
2721 wq->first_flusher = NULL;
2723 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2724 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2727 struct wq_flusher *next, *tmp;
2729 /* complete all the flushers sharing the current flush color */
2730 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2731 if (next->flush_color != wq->flush_color)
2733 list_del_init(&next->list);
2734 complete(&next->done);
2737 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2738 wq->flush_color != work_next_color(wq->work_color));
2740 /* this flush_color is finished, advance by one */
2741 wq->flush_color = work_next_color(wq->flush_color);
2743 /* one color has been freed, handle overflow queue */
2744 if (!list_empty(&wq->flusher_overflow)) {
2746 * Assign the same color to all overflowed
2747 * flushers, advance work_color and append to
2748 * flusher_queue. This is the start-to-wait
2749 * phase for these overflowed flushers.
2751 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2752 tmp->flush_color = wq->work_color;
2754 wq->work_color = work_next_color(wq->work_color);
2756 list_splice_tail_init(&wq->flusher_overflow,
2757 &wq->flusher_queue);
2758 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2761 if (list_empty(&wq->flusher_queue)) {
2762 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2767 * Need to flush more colors. Make the next flusher
2768 * the new first flusher and arm pwqs.
2770 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2771 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2773 list_del_init(&next->list);
2774 wq->first_flusher = next;
2776 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2780 * Meh... this color is already done, clear first
2781 * flusher and repeat cascading.
2783 wq->first_flusher = NULL;
2787 mutex_unlock(&wq->mutex);
2789 EXPORT_SYMBOL_GPL(flush_workqueue);
2792 * drain_workqueue - drain a workqueue
2793 * @wq: workqueue to drain
2795 * Wait until the workqueue becomes empty. While draining is in progress,
2796 * only chain queueing is allowed. IOW, only currently pending or running
2797 * work items on @wq can queue further work items on it. @wq is flushed
2798 * repeatedly until it becomes empty. The number of flushing is detemined
2799 * by the depth of chaining and should be relatively short. Whine if it
2802 void drain_workqueue(struct workqueue_struct *wq)
2804 unsigned int flush_cnt = 0;
2805 struct pool_workqueue *pwq;
2808 * __queue_work() needs to test whether there are drainers, is much
2809 * hotter than drain_workqueue() and already looks at @wq->flags.
2810 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2812 mutex_lock(&wq->mutex);
2813 if (!wq->nr_drainers++)
2814 wq->flags |= __WQ_DRAINING;
2815 mutex_unlock(&wq->mutex);
2817 flush_workqueue(wq);
2819 mutex_lock(&wq->mutex);
2821 for_each_pwq(pwq, wq) {
2824 spin_lock_irq(&pwq->pool->lock);
2825 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2826 spin_unlock_irq(&pwq->pool->lock);
2831 if (++flush_cnt == 10 ||
2832 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2833 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2834 wq->name, flush_cnt);
2836 mutex_unlock(&wq->mutex);
2840 if (!--wq->nr_drainers)
2841 wq->flags &= ~__WQ_DRAINING;
2842 mutex_unlock(&wq->mutex);
2844 EXPORT_SYMBOL_GPL(drain_workqueue);
2846 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2848 struct worker *worker = NULL;
2849 struct worker_pool *pool;
2850 struct pool_workqueue *pwq;
2854 local_irq_disable();
2855 pool = get_work_pool(work);
2861 spin_lock(&pool->lock);
2862 /* see the comment in try_to_grab_pending() with the same code */
2863 pwq = get_work_pwq(work);
2865 if (unlikely(pwq->pool != pool))
2868 worker = find_worker_executing_work(pool, work);
2871 pwq = worker->current_pwq;
2874 insert_wq_barrier(pwq, barr, work, worker);
2875 spin_unlock_irq(&pool->lock);
2878 * If @max_active is 1 or rescuer is in use, flushing another work
2879 * item on the same workqueue may lead to deadlock. Make sure the
2880 * flusher is not running on the same workqueue by verifying write
2883 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2884 lock_map_acquire(&pwq->wq->lockdep_map);
2886 lock_map_acquire_read(&pwq->wq->lockdep_map);
2887 lock_map_release(&pwq->wq->lockdep_map);
2891 spin_unlock_irq(&pool->lock);
2896 * flush_work - wait for a work to finish executing the last queueing instance
2897 * @work: the work to flush
2899 * Wait until @work has finished execution. @work is guaranteed to be idle
2900 * on return if it hasn't been requeued since flush started.
2903 * %true if flush_work() waited for the work to finish execution,
2904 * %false if it was already idle.
2906 bool flush_work(struct work_struct *work)
2908 struct wq_barrier barr;
2910 lock_map_acquire(&work->lockdep_map);
2911 lock_map_release(&work->lockdep_map);
2913 if (start_flush_work(work, &barr)) {
2914 wait_for_completion(&barr.done);
2915 destroy_work_on_stack(&barr.work);
2921 EXPORT_SYMBOL_GPL(flush_work);
2923 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2925 unsigned long flags;
2929 ret = try_to_grab_pending(work, is_dwork, &flags);
2931 * If someone else is canceling, wait for the same event it
2932 * would be waiting for before retrying.
2934 if (unlikely(ret == -ENOENT))
2936 } while (unlikely(ret < 0));
2938 /* tell other tasks trying to grab @work to back off */
2939 mark_work_canceling(work);
2940 local_irq_restore(flags);
2943 clear_work_data(work);
2948 * cancel_work_sync - cancel a work and wait for it to finish
2949 * @work: the work to cancel
2951 * Cancel @work and wait for its execution to finish. This function
2952 * can be used even if the work re-queues itself or migrates to
2953 * another workqueue. On return from this function, @work is
2954 * guaranteed to be not pending or executing on any CPU.
2956 * cancel_work_sync(&delayed_work->work) must not be used for
2957 * delayed_work's. Use cancel_delayed_work_sync() instead.
2959 * The caller must ensure that the workqueue on which @work was last
2960 * queued can't be destroyed before this function returns.
2963 * %true if @work was pending, %false otherwise.
2965 bool cancel_work_sync(struct work_struct *work)
2967 return __cancel_work_timer(work, false);
2969 EXPORT_SYMBOL_GPL(cancel_work_sync);
2972 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2973 * @dwork: the delayed work to flush
2975 * Delayed timer is cancelled and the pending work is queued for
2976 * immediate execution. Like flush_work(), this function only
2977 * considers the last queueing instance of @dwork.
2980 * %true if flush_work() waited for the work to finish execution,
2981 * %false if it was already idle.
2983 bool flush_delayed_work(struct delayed_work *dwork)
2985 local_irq_disable();
2986 if (del_timer_sync(&dwork->timer))
2987 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2989 return flush_work(&dwork->work);
2991 EXPORT_SYMBOL(flush_delayed_work);
2994 * cancel_delayed_work - cancel a delayed work
2995 * @dwork: delayed_work to cancel
2997 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2998 * and canceled; %false if wasn't pending. Note that the work callback
2999 * function may still be running on return, unless it returns %true and the
3000 * work doesn't re-arm itself. Explicitly flush or use
3001 * cancel_delayed_work_sync() to wait on it.
3003 * This function is safe to call from any context including IRQ handler.
3005 bool cancel_delayed_work(struct delayed_work *dwork)
3007 unsigned long flags;
3011 ret = try_to_grab_pending(&dwork->work, true, &flags);
3012 } while (unlikely(ret == -EAGAIN));
3014 if (unlikely(ret < 0))
3017 set_work_pool_and_clear_pending(&dwork->work,
3018 get_work_pool_id(&dwork->work));
3019 local_irq_restore(flags);
3022 EXPORT_SYMBOL(cancel_delayed_work);
3025 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3026 * @dwork: the delayed work cancel
3028 * This is cancel_work_sync() for delayed works.
3031 * %true if @dwork was pending, %false otherwise.
3033 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3035 return __cancel_work_timer(&dwork->work, true);
3037 EXPORT_SYMBOL(cancel_delayed_work_sync);
3040 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3041 * @func: the function to call
3043 * schedule_on_each_cpu() executes @func on each online CPU using the
3044 * system workqueue and blocks until all CPUs have completed.
3045 * schedule_on_each_cpu() is very slow.
3048 * 0 on success, -errno on failure.
3050 int schedule_on_each_cpu(work_func_t func)
3053 struct work_struct __percpu *works;
3055 works = alloc_percpu(struct work_struct);
3061 for_each_online_cpu(cpu) {
3062 struct work_struct *work = per_cpu_ptr(works, cpu);
3064 INIT_WORK(work, func);
3065 schedule_work_on(cpu, work);
3068 for_each_online_cpu(cpu)
3069 flush_work(per_cpu_ptr(works, cpu));
3077 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3079 * Forces execution of the kernel-global workqueue and blocks until its
3082 * Think twice before calling this function! It's very easy to get into
3083 * trouble if you don't take great care. Either of the following situations
3084 * will lead to deadlock:
3086 * One of the work items currently on the workqueue needs to acquire
3087 * a lock held by your code or its caller.
3089 * Your code is running in the context of a work routine.
3091 * They will be detected by lockdep when they occur, but the first might not
3092 * occur very often. It depends on what work items are on the workqueue and
3093 * what locks they need, which you have no control over.
3095 * In most situations flushing the entire workqueue is overkill; you merely
3096 * need to know that a particular work item isn't queued and isn't running.
3097 * In such cases you should use cancel_delayed_work_sync() or
3098 * cancel_work_sync() instead.
3100 void flush_scheduled_work(void)
3102 flush_workqueue(system_wq);
3104 EXPORT_SYMBOL(flush_scheduled_work);
3107 * execute_in_process_context - reliably execute the routine with user context
3108 * @fn: the function to execute
3109 * @ew: guaranteed storage for the execute work structure (must
3110 * be available when the work executes)
3112 * Executes the function immediately if process context is available,
3113 * otherwise schedules the function for delayed execution.
3115 * Returns: 0 - function was executed
3116 * 1 - function was scheduled for execution
3118 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3120 if (!in_interrupt()) {
3125 INIT_WORK(&ew->work, fn);
3126 schedule_work(&ew->work);
3130 EXPORT_SYMBOL_GPL(execute_in_process_context);
3134 * Workqueues with WQ_SYSFS flag set is visible to userland via
3135 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3136 * following attributes.
3138 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3139 * max_active RW int : maximum number of in-flight work items
3141 * Unbound workqueues have the following extra attributes.
3143 * id RO int : the associated pool ID
3144 * nice RW int : nice value of the workers
3145 * cpumask RW mask : bitmask of allowed CPUs for the workers
3148 struct workqueue_struct *wq;
3152 static struct workqueue_struct *dev_to_wq(struct device *dev)
3154 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3159 static ssize_t wq_per_cpu_show(struct device *dev,
3160 struct device_attribute *attr, char *buf)
3162 struct workqueue_struct *wq = dev_to_wq(dev);
3164 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3167 static ssize_t wq_max_active_show(struct device *dev,
3168 struct device_attribute *attr, char *buf)
3170 struct workqueue_struct *wq = dev_to_wq(dev);
3172 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3175 static ssize_t wq_max_active_store(struct device *dev,
3176 struct device_attribute *attr,
3177 const char *buf, size_t count)
3179 struct workqueue_struct *wq = dev_to_wq(dev);
3182 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3185 workqueue_set_max_active(wq, val);
3189 static struct device_attribute wq_sysfs_attrs[] = {
3190 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3191 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3195 static ssize_t wq_pool_ids_show(struct device *dev,
3196 struct device_attribute *attr, char *buf)
3198 struct workqueue_struct *wq = dev_to_wq(dev);
3199 const char *delim = "";
3200 int node, written = 0;
3202 rcu_read_lock_sched();
3203 for_each_node(node) {
3204 written += scnprintf(buf + written, PAGE_SIZE - written,
3205 "%s%d:%d", delim, node,
3206 unbound_pwq_by_node(wq, node)->pool->id);
3209 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3210 rcu_read_unlock_sched();
3215 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3218 struct workqueue_struct *wq = dev_to_wq(dev);
3221 mutex_lock(&wq->mutex);
3222 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3223 mutex_unlock(&wq->mutex);
3228 /* prepare workqueue_attrs for sysfs store operations */
3229 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3231 struct workqueue_attrs *attrs;
3233 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3237 mutex_lock(&wq->mutex);
3238 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3239 mutex_unlock(&wq->mutex);
3243 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3244 const char *buf, size_t count)
3246 struct workqueue_struct *wq = dev_to_wq(dev);
3247 struct workqueue_attrs *attrs;
3250 attrs = wq_sysfs_prep_attrs(wq);
3254 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3255 attrs->nice >= -20 && attrs->nice <= 19)
3256 ret = apply_workqueue_attrs(wq, attrs);
3260 free_workqueue_attrs(attrs);
3261 return ret ?: count;
3264 static ssize_t wq_cpumask_show(struct device *dev,
3265 struct device_attribute *attr, char *buf)
3267 struct workqueue_struct *wq = dev_to_wq(dev);
3270 mutex_lock(&wq->mutex);
3271 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3272 mutex_unlock(&wq->mutex);
3274 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3278 static ssize_t wq_cpumask_store(struct device *dev,
3279 struct device_attribute *attr,
3280 const char *buf, size_t count)
3282 struct workqueue_struct *wq = dev_to_wq(dev);
3283 struct workqueue_attrs *attrs;
3286 attrs = wq_sysfs_prep_attrs(wq);
3290 ret = cpumask_parse(buf, attrs->cpumask);
3292 ret = apply_workqueue_attrs(wq, attrs);
3294 free_workqueue_attrs(attrs);
3295 return ret ?: count;
3298 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3301 struct workqueue_struct *wq = dev_to_wq(dev);
3304 mutex_lock(&wq->mutex);
3305 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3306 !wq->unbound_attrs->no_numa);
3307 mutex_unlock(&wq->mutex);
3312 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3313 const char *buf, size_t count)
3315 struct workqueue_struct *wq = dev_to_wq(dev);
3316 struct workqueue_attrs *attrs;
3319 attrs = wq_sysfs_prep_attrs(wq);
3324 if (sscanf(buf, "%d", &v) == 1) {
3325 attrs->no_numa = !v;
3326 ret = apply_workqueue_attrs(wq, attrs);
3329 free_workqueue_attrs(attrs);
3330 return ret ?: count;
3333 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3334 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3335 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3336 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3337 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3341 static struct bus_type wq_subsys = {
3342 .name = "workqueue",
3343 .dev_attrs = wq_sysfs_attrs,
3346 static int __init wq_sysfs_init(void)
3348 return subsys_virtual_register(&wq_subsys, NULL);
3350 core_initcall(wq_sysfs_init);
3352 static void wq_device_release(struct device *dev)
3354 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3360 * workqueue_sysfs_register - make a workqueue visible in sysfs
3361 * @wq: the workqueue to register
3363 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3364 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3365 * which is the preferred method.
3367 * Workqueue user should use this function directly iff it wants to apply
3368 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3369 * apply_workqueue_attrs() may race against userland updating the
3372 * Returns 0 on success, -errno on failure.
3374 int workqueue_sysfs_register(struct workqueue_struct *wq)
3376 struct wq_device *wq_dev;
3380 * Adjusting max_active or creating new pwqs by applyting
3381 * attributes breaks ordering guarantee. Disallow exposing ordered
3384 if (WARN_ON(wq->flags & __WQ_ORDERED))
3387 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3392 wq_dev->dev.bus = &wq_subsys;
3393 wq_dev->dev.init_name = wq->name;
3394 wq_dev->dev.release = wq_device_release;
3397 * unbound_attrs are created separately. Suppress uevent until
3398 * everything is ready.
3400 dev_set_uevent_suppress(&wq_dev->dev, true);
3402 ret = device_register(&wq_dev->dev);
3409 if (wq->flags & WQ_UNBOUND) {
3410 struct device_attribute *attr;
3412 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3413 ret = device_create_file(&wq_dev->dev, attr);
3415 device_unregister(&wq_dev->dev);
3422 dev_set_uevent_suppress(&wq_dev->dev, false);
3423 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3428 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3429 * @wq: the workqueue to unregister
3431 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3433 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3435 struct wq_device *wq_dev = wq->wq_dev;
3441 device_unregister(&wq_dev->dev);
3443 #else /* CONFIG_SYSFS */
3444 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3445 #endif /* CONFIG_SYSFS */
3448 * free_workqueue_attrs - free a workqueue_attrs
3449 * @attrs: workqueue_attrs to free
3451 * Undo alloc_workqueue_attrs().
3453 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3456 free_cpumask_var(attrs->cpumask);
3462 * alloc_workqueue_attrs - allocate a workqueue_attrs
3463 * @gfp_mask: allocation mask to use
3465 * Allocate a new workqueue_attrs, initialize with default settings and
3466 * return it. Returns NULL on failure.
3468 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3470 struct workqueue_attrs *attrs;
3472 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3475 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3478 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3481 free_workqueue_attrs(attrs);
3485 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3486 const struct workqueue_attrs *from)
3488 to->nice = from->nice;
3489 cpumask_copy(to->cpumask, from->cpumask);
3491 * Unlike hash and equality test, this function doesn't ignore
3492 * ->no_numa as it is used for both pool and wq attrs. Instead,
3493 * get_unbound_pool() explicitly clears ->no_numa after copying.
3495 to->no_numa = from->no_numa;
3498 /* hash value of the content of @attr */
3499 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3503 hash = jhash_1word(attrs->nice, hash);
3504 hash = jhash(cpumask_bits(attrs->cpumask),
3505 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3509 /* content equality test */
3510 static bool wqattrs_equal(const struct workqueue_attrs *a,
3511 const struct workqueue_attrs *b)
3513 if (a->nice != b->nice)
3515 if (!cpumask_equal(a->cpumask, b->cpumask))
3521 * init_worker_pool - initialize a newly zalloc'd worker_pool
3522 * @pool: worker_pool to initialize
3524 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3525 * Returns 0 on success, -errno on failure. Even on failure, all fields
3526 * inside @pool proper are initialized and put_unbound_pool() can be called
3527 * on @pool safely to release it.
3529 static int init_worker_pool(struct worker_pool *pool)
3531 spin_lock_init(&pool->lock);
3534 pool->node = NUMA_NO_NODE;
3535 pool->flags |= POOL_DISASSOCIATED;
3536 INIT_LIST_HEAD(&pool->worklist);
3537 INIT_LIST_HEAD(&pool->idle_list);
3538 hash_init(pool->busy_hash);
3540 init_timer_deferrable(&pool->idle_timer);
3541 pool->idle_timer.function = idle_worker_timeout;
3542 pool->idle_timer.data = (unsigned long)pool;
3544 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3545 (unsigned long)pool);
3547 mutex_init(&pool->manager_arb);
3548 mutex_init(&pool->manager_mutex);
3549 idr_init(&pool->worker_idr);
3551 INIT_HLIST_NODE(&pool->hash_node);
3554 /* shouldn't fail above this point */
3555 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3561 static void rcu_free_pool(struct rcu_head *rcu)
3563 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3565 idr_destroy(&pool->worker_idr);
3566 free_workqueue_attrs(pool->attrs);
3571 * put_unbound_pool - put a worker_pool
3572 * @pool: worker_pool to put
3574 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3575 * safe manner. get_unbound_pool() calls this function on its failure path
3576 * and this function should be able to release pools which went through,
3577 * successfully or not, init_worker_pool().
3579 * Should be called with wq_pool_mutex held.
3581 static void put_unbound_pool(struct worker_pool *pool)
3583 struct worker *worker;
3585 lockdep_assert_held(&wq_pool_mutex);
3591 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3592 WARN_ON(!list_empty(&pool->worklist)))
3595 /* release id and unhash */
3597 idr_remove(&worker_pool_idr, pool->id);
3598 hash_del(&pool->hash_node);
3601 * Become the manager and destroy all workers. Grabbing
3602 * manager_arb prevents @pool's workers from blocking on
3605 mutex_lock(&pool->manager_arb);
3606 mutex_lock(&pool->manager_mutex);
3607 spin_lock_irq(&pool->lock);
3609 while ((worker = first_worker(pool)))
3610 destroy_worker(worker);
3611 WARN_ON(pool->nr_workers || pool->nr_idle);
3613 spin_unlock_irq(&pool->lock);
3614 mutex_unlock(&pool->manager_mutex);
3615 mutex_unlock(&pool->manager_arb);
3617 /* shut down the timers */
3618 del_timer_sync(&pool->idle_timer);
3619 del_timer_sync(&pool->mayday_timer);
3621 /* sched-RCU protected to allow dereferences from get_work_pool() */
3622 call_rcu_sched(&pool->rcu, rcu_free_pool);
3626 * get_unbound_pool - get a worker_pool with the specified attributes
3627 * @attrs: the attributes of the worker_pool to get
3629 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3630 * reference count and return it. If there already is a matching
3631 * worker_pool, it will be used; otherwise, this function attempts to
3632 * create a new one. On failure, returns NULL.
3634 * Should be called with wq_pool_mutex held.
3636 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3638 u32 hash = wqattrs_hash(attrs);
3639 struct worker_pool *pool;
3642 lockdep_assert_held(&wq_pool_mutex);
3644 /* do we already have a matching pool? */
3645 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3646 if (wqattrs_equal(pool->attrs, attrs)) {
3652 /* nope, create a new one */
3653 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3654 if (!pool || init_worker_pool(pool) < 0)
3657 if (workqueue_freezing)
3658 pool->flags |= POOL_FREEZING;
3660 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3661 copy_workqueue_attrs(pool->attrs, attrs);
3664 * no_numa isn't a worker_pool attribute, always clear it. See
3665 * 'struct workqueue_attrs' comments for detail.
3667 pool->attrs->no_numa = false;
3669 /* if cpumask is contained inside a NUMA node, we belong to that node */
3670 if (wq_numa_enabled) {
3671 for_each_node(node) {
3672 if (cpumask_subset(pool->attrs->cpumask,
3673 wq_numa_possible_cpumask[node])) {
3680 if (worker_pool_assign_id(pool) < 0)
3683 /* create and start the initial worker */
3684 if (create_and_start_worker(pool) < 0)
3688 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3693 put_unbound_pool(pool);
3697 static void rcu_free_pwq(struct rcu_head *rcu)
3699 kmem_cache_free(pwq_cache,
3700 container_of(rcu, struct pool_workqueue, rcu));
3704 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3705 * and needs to be destroyed.
3707 static void pwq_unbound_release_workfn(struct work_struct *work)
3709 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3710 unbound_release_work);
3711 struct workqueue_struct *wq = pwq->wq;
3712 struct worker_pool *pool = pwq->pool;
3715 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3719 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3720 * necessary on release but do it anyway. It's easier to verify
3721 * and consistent with the linking path.
3723 mutex_lock(&wq->mutex);
3724 list_del_rcu(&pwq->pwqs_node);
3725 is_last = list_empty(&wq->pwqs);
3726 mutex_unlock(&wq->mutex);
3728 mutex_lock(&wq_pool_mutex);
3729 put_unbound_pool(pool);
3730 mutex_unlock(&wq_pool_mutex);
3732 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3735 * If we're the last pwq going away, @wq is already dead and no one
3736 * is gonna access it anymore. Free it.
3739 free_workqueue_attrs(wq->unbound_attrs);
3745 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3746 * @pwq: target pool_workqueue
3748 * If @pwq isn't freezing, set @pwq->max_active to the associated
3749 * workqueue's saved_max_active and activate delayed work items
3750 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3752 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3754 struct workqueue_struct *wq = pwq->wq;
3755 bool freezable = wq->flags & WQ_FREEZABLE;
3757 /* for @wq->saved_max_active */
3758 lockdep_assert_held(&wq->mutex);
3760 /* fast exit for non-freezable wqs */
3761 if (!freezable && pwq->max_active == wq->saved_max_active)
3764 spin_lock_irq(&pwq->pool->lock);
3766 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3767 pwq->max_active = wq->saved_max_active;
3769 while (!list_empty(&pwq->delayed_works) &&
3770 pwq->nr_active < pwq->max_active)
3771 pwq_activate_first_delayed(pwq);
3774 * Need to kick a worker after thawed or an unbound wq's
3775 * max_active is bumped. It's a slow path. Do it always.
3777 wake_up_worker(pwq->pool);
3779 pwq->max_active = 0;
3782 spin_unlock_irq(&pwq->pool->lock);
3785 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3786 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3787 struct worker_pool *pool)
3789 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3791 memset(pwq, 0, sizeof(*pwq));
3795 pwq->flush_color = -1;
3797 INIT_LIST_HEAD(&pwq->delayed_works);
3798 INIT_LIST_HEAD(&pwq->pwqs_node);
3799 INIT_LIST_HEAD(&pwq->mayday_node);
3800 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3803 /* sync @pwq with the current state of its associated wq and link it */
3804 static void link_pwq(struct pool_workqueue *pwq)
3806 struct workqueue_struct *wq = pwq->wq;
3808 lockdep_assert_held(&wq->mutex);
3810 /* may be called multiple times, ignore if already linked */
3811 if (!list_empty(&pwq->pwqs_node))
3815 * Set the matching work_color. This is synchronized with
3816 * wq->mutex to avoid confusing flush_workqueue().
3818 pwq->work_color = wq->work_color;
3820 /* sync max_active to the current setting */
3821 pwq_adjust_max_active(pwq);
3824 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3827 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3828 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3829 const struct workqueue_attrs *attrs)
3831 struct worker_pool *pool;
3832 struct pool_workqueue *pwq;
3834 lockdep_assert_held(&wq_pool_mutex);
3836 pool = get_unbound_pool(attrs);
3840 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3842 put_unbound_pool(pool);
3846 init_pwq(pwq, wq, pool);
3850 /* undo alloc_unbound_pwq(), used only in the error path */
3851 static void free_unbound_pwq(struct pool_workqueue *pwq)
3853 lockdep_assert_held(&wq_pool_mutex);
3856 put_unbound_pool(pwq->pool);
3857 kmem_cache_free(pwq_cache, pwq);
3862 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3863 * @attrs: the wq_attrs of interest
3864 * @node: the target NUMA node
3865 * @cpu_going_down: if >= 0, the CPU to consider as offline
3866 * @cpumask: outarg, the resulting cpumask
3868 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3869 * @cpu_going_down is >= 0, that cpu is considered offline during
3870 * calculation. The result is stored in @cpumask. This function returns
3871 * %true if the resulting @cpumask is different from @attrs->cpumask,
3874 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3875 * enabled and @node has online CPUs requested by @attrs, the returned
3876 * cpumask is the intersection of the possible CPUs of @node and
3879 * The caller is responsible for ensuring that the cpumask of @node stays
3882 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3883 int cpu_going_down, cpumask_t *cpumask)
3885 if (!wq_numa_enabled || attrs->no_numa)
3888 /* does @node have any online CPUs @attrs wants? */
3889 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3890 if (cpu_going_down >= 0)
3891 cpumask_clear_cpu(cpu_going_down, cpumask);
3893 if (cpumask_empty(cpumask))
3896 /* yeap, return possible CPUs in @node that @attrs wants */
3897 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3898 return !cpumask_equal(cpumask, attrs->cpumask);
3901 cpumask_copy(cpumask, attrs->cpumask);
3905 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3906 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3908 struct pool_workqueue *pwq)
3910 struct pool_workqueue *old_pwq;
3912 lockdep_assert_held(&wq->mutex);
3914 /* link_pwq() can handle duplicate calls */
3917 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3918 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3923 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3924 * @wq: the target workqueue
3925 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3927 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3928 * machines, this function maps a separate pwq to each NUMA node with
3929 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3930 * NUMA node it was issued on. Older pwqs are released as in-flight work
3931 * items finish. Note that a work item which repeatedly requeues itself
3932 * back-to-back will stay on its current pwq.
3934 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3937 int apply_workqueue_attrs(struct workqueue_struct *wq,
3938 const struct workqueue_attrs *attrs)
3940 struct workqueue_attrs *new_attrs, *tmp_attrs;
3941 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3944 /* only unbound workqueues can change attributes */
3945 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3948 /* creating multiple pwqs breaks ordering guarantee */
3949 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3952 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3953 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3954 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3955 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3958 /* make a copy of @attrs and sanitize it */
3959 copy_workqueue_attrs(new_attrs, attrs);
3960 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3963 * We may create multiple pwqs with differing cpumasks. Make a
3964 * copy of @new_attrs which will be modified and used to obtain
3967 copy_workqueue_attrs(tmp_attrs, new_attrs);
3970 * CPUs should stay stable across pwq creations and installations.
3971 * Pin CPUs, determine the target cpumask for each node and create
3976 mutex_lock(&wq_pool_mutex);
3979 * If something goes wrong during CPU up/down, we'll fall back to
3980 * the default pwq covering whole @attrs->cpumask. Always create
3981 * it even if we don't use it immediately.
3983 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3987 for_each_node(node) {
3988 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3989 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3994 pwq_tbl[node] = dfl_pwq;
3998 mutex_unlock(&wq_pool_mutex);
4000 /* all pwqs have been created successfully, let's install'em */
4001 mutex_lock(&wq->mutex);
4003 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
4005 /* save the previous pwq and install the new one */
4007 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
4009 /* @dfl_pwq might not have been used, ensure it's linked */
4011 swap(wq->dfl_pwq, dfl_pwq);
4013 mutex_unlock(&wq->mutex);
4015 /* put the old pwqs */
4017 put_pwq_unlocked(pwq_tbl[node]);
4018 put_pwq_unlocked(dfl_pwq);
4024 free_workqueue_attrs(tmp_attrs);
4025 free_workqueue_attrs(new_attrs);
4030 free_unbound_pwq(dfl_pwq);
4032 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
4033 free_unbound_pwq(pwq_tbl[node]);
4034 mutex_unlock(&wq_pool_mutex);
4042 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4043 * @wq: the target workqueue
4044 * @cpu: the CPU coming up or going down
4045 * @online: whether @cpu is coming up or going down
4047 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4048 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4051 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4052 * falls back to @wq->dfl_pwq which may not be optimal but is always
4055 * Note that when the last allowed CPU of a NUMA node goes offline for a
4056 * workqueue with a cpumask spanning multiple nodes, the workers which were
4057 * already executing the work items for the workqueue will lose their CPU
4058 * affinity and may execute on any CPU. This is similar to how per-cpu
4059 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4060 * affinity, it's the user's responsibility to flush the work item from
4063 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4066 int node = cpu_to_node(cpu);
4067 int cpu_off = online ? -1 : cpu;
4068 struct pool_workqueue *old_pwq = NULL, *pwq;
4069 struct workqueue_attrs *target_attrs;
4072 lockdep_assert_held(&wq_pool_mutex);
4074 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4078 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4079 * Let's use a preallocated one. The following buf is protected by
4080 * CPU hotplug exclusion.
4082 target_attrs = wq_update_unbound_numa_attrs_buf;
4083 cpumask = target_attrs->cpumask;
4085 mutex_lock(&wq->mutex);
4086 if (wq->unbound_attrs->no_numa)
4089 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4090 pwq = unbound_pwq_by_node(wq, node);
4093 * Let's determine what needs to be done. If the target cpumask is
4094 * different from wq's, we need to compare it to @pwq's and create
4095 * a new one if they don't match. If the target cpumask equals
4096 * wq's, the default pwq should be used. If @pwq is already the
4097 * default one, nothing to do; otherwise, install the default one.
4099 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4100 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4103 if (pwq == wq->dfl_pwq)
4109 mutex_unlock(&wq->mutex);
4111 /* create a new pwq */
4112 pwq = alloc_unbound_pwq(wq, target_attrs);
4114 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4116 mutex_lock(&wq->mutex);
4121 * Install the new pwq. As this function is called only from CPU
4122 * hotplug callbacks and applying a new attrs is wrapped with
4123 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4126 mutex_lock(&wq->mutex);
4127 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4131 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4132 get_pwq(wq->dfl_pwq);
4133 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4134 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4136 mutex_unlock(&wq->mutex);
4137 put_pwq_unlocked(old_pwq);
4140 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4142 bool highpri = wq->flags & WQ_HIGHPRI;
4145 if (!(wq->flags & WQ_UNBOUND)) {
4146 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4150 for_each_possible_cpu(cpu) {
4151 struct pool_workqueue *pwq =
4152 per_cpu_ptr(wq->cpu_pwqs, cpu);
4153 struct worker_pool *cpu_pools =
4154 per_cpu(cpu_worker_pools, cpu);
4156 init_pwq(pwq, wq, &cpu_pools[highpri]);
4158 mutex_lock(&wq->mutex);
4160 mutex_unlock(&wq->mutex);
4163 } else if (wq->flags & __WQ_ORDERED) {
4164 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4165 /* there should only be single pwq for ordering guarantee */
4166 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4167 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4168 "ordering guarantee broken for workqueue %s\n", wq->name);
4171 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4175 static int wq_clamp_max_active(int max_active, unsigned int flags,
4178 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4180 if (max_active < 1 || max_active > lim)
4181 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4182 max_active, name, 1, lim);
4184 return clamp_val(max_active, 1, lim);
4187 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4190 struct lock_class_key *key,
4191 const char *lock_name, ...)
4193 size_t tbl_size = 0;
4195 struct workqueue_struct *wq;
4196 struct pool_workqueue *pwq;
4198 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4199 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4200 flags |= WQ_UNBOUND;
4202 /* allocate wq and format name */
4203 if (flags & WQ_UNBOUND)
4204 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4206 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4210 if (flags & WQ_UNBOUND) {
4211 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4212 if (!wq->unbound_attrs)
4216 va_start(args, lock_name);
4217 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4220 max_active = max_active ?: WQ_DFL_ACTIVE;
4221 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4225 wq->saved_max_active = max_active;
4226 mutex_init(&wq->mutex);
4227 atomic_set(&wq->nr_pwqs_to_flush, 0);
4228 INIT_LIST_HEAD(&wq->pwqs);
4229 INIT_LIST_HEAD(&wq->flusher_queue);
4230 INIT_LIST_HEAD(&wq->flusher_overflow);
4231 INIT_LIST_HEAD(&wq->maydays);
4233 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4234 INIT_LIST_HEAD(&wq->list);
4236 if (alloc_and_link_pwqs(wq) < 0)
4240 * Workqueues which may be used during memory reclaim should
4241 * have a rescuer to guarantee forward progress.
4243 if (flags & WQ_MEM_RECLAIM) {
4244 struct worker *rescuer;
4246 rescuer = alloc_worker();
4250 rescuer->rescue_wq = wq;
4251 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4253 if (IS_ERR(rescuer->task)) {
4258 wq->rescuer = rescuer;
4259 rescuer->task->flags |= PF_NO_SETAFFINITY;
4260 wake_up_process(rescuer->task);
4263 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4267 * wq_pool_mutex protects global freeze state and workqueues list.
4268 * Grab it, adjust max_active and add the new @wq to workqueues
4271 mutex_lock(&wq_pool_mutex);
4273 mutex_lock(&wq->mutex);
4274 for_each_pwq(pwq, wq)
4275 pwq_adjust_max_active(pwq);
4276 mutex_unlock(&wq->mutex);
4278 list_add(&wq->list, &workqueues);
4280 mutex_unlock(&wq_pool_mutex);
4285 free_workqueue_attrs(wq->unbound_attrs);
4289 destroy_workqueue(wq);
4292 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4295 * destroy_workqueue - safely terminate a workqueue
4296 * @wq: target workqueue
4298 * Safely destroy a workqueue. All work currently pending will be done first.
4300 void destroy_workqueue(struct workqueue_struct *wq)
4302 struct pool_workqueue *pwq;
4305 /* drain it before proceeding with destruction */
4306 drain_workqueue(wq);
4309 mutex_lock(&wq->mutex);
4310 for_each_pwq(pwq, wq) {
4313 for (i = 0; i < WORK_NR_COLORS; i++) {
4314 if (WARN_ON(pwq->nr_in_flight[i])) {
4315 mutex_unlock(&wq->mutex);
4320 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4321 WARN_ON(pwq->nr_active) ||
4322 WARN_ON(!list_empty(&pwq->delayed_works))) {
4323 mutex_unlock(&wq->mutex);
4327 mutex_unlock(&wq->mutex);
4330 * wq list is used to freeze wq, remove from list after
4331 * flushing is complete in case freeze races us.
4333 mutex_lock(&wq_pool_mutex);
4334 list_del_init(&wq->list);
4335 mutex_unlock(&wq_pool_mutex);
4337 workqueue_sysfs_unregister(wq);
4340 kthread_stop(wq->rescuer->task);
4345 if (!(wq->flags & WQ_UNBOUND)) {
4347 * The base ref is never dropped on per-cpu pwqs. Directly
4348 * free the pwqs and wq.
4350 free_percpu(wq->cpu_pwqs);
4354 * We're the sole accessor of @wq at this point. Directly
4355 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4356 * @wq will be freed when the last pwq is released.
4358 for_each_node(node) {
4359 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4360 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4361 put_pwq_unlocked(pwq);
4365 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4366 * put. Don't access it afterwards.
4370 put_pwq_unlocked(pwq);
4373 EXPORT_SYMBOL_GPL(destroy_workqueue);
4376 * workqueue_set_max_active - adjust max_active of a workqueue
4377 * @wq: target workqueue
4378 * @max_active: new max_active value.
4380 * Set max_active of @wq to @max_active.
4383 * Don't call from IRQ context.
4385 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4387 struct pool_workqueue *pwq;
4389 /* disallow meddling with max_active for ordered workqueues */
4390 if (WARN_ON(wq->flags & __WQ_ORDERED))
4393 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4395 mutex_lock(&wq->mutex);
4397 wq->saved_max_active = max_active;
4399 for_each_pwq(pwq, wq)
4400 pwq_adjust_max_active(pwq);
4402 mutex_unlock(&wq->mutex);
4404 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4407 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4409 * Determine whether %current is a workqueue rescuer. Can be used from
4410 * work functions to determine whether it's being run off the rescuer task.
4412 bool current_is_workqueue_rescuer(void)
4414 struct worker *worker = current_wq_worker();
4416 return worker && worker->rescue_wq;
4420 * workqueue_congested - test whether a workqueue is congested
4421 * @cpu: CPU in question
4422 * @wq: target workqueue
4424 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4425 * no synchronization around this function and the test result is
4426 * unreliable and only useful as advisory hints or for debugging.
4428 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4429 * Note that both per-cpu and unbound workqueues may be associated with
4430 * multiple pool_workqueues which have separate congested states. A
4431 * workqueue being congested on one CPU doesn't mean the workqueue is also
4432 * contested on other CPUs / NUMA nodes.
4435 * %true if congested, %false otherwise.
4437 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4439 struct pool_workqueue *pwq;
4442 rcu_read_lock_sched();
4444 if (cpu == WORK_CPU_UNBOUND)
4445 cpu = smp_processor_id();
4447 if (!(wq->flags & WQ_UNBOUND))
4448 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4450 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4452 ret = !list_empty(&pwq->delayed_works);
4453 rcu_read_unlock_sched();
4457 EXPORT_SYMBOL_GPL(workqueue_congested);
4460 * work_busy - test whether a work is currently pending or running
4461 * @work: the work to be tested
4463 * Test whether @work is currently pending or running. There is no
4464 * synchronization around this function and the test result is
4465 * unreliable and only useful as advisory hints or for debugging.
4468 * OR'd bitmask of WORK_BUSY_* bits.
4470 unsigned int work_busy(struct work_struct *work)
4472 struct worker_pool *pool;
4473 unsigned long flags;
4474 unsigned int ret = 0;
4476 if (work_pending(work))
4477 ret |= WORK_BUSY_PENDING;
4479 local_irq_save(flags);
4480 pool = get_work_pool(work);
4482 spin_lock(&pool->lock);
4483 if (find_worker_executing_work(pool, work))
4484 ret |= WORK_BUSY_RUNNING;
4485 spin_unlock(&pool->lock);
4487 local_irq_restore(flags);
4491 EXPORT_SYMBOL_GPL(work_busy);
4494 * set_worker_desc - set description for the current work item
4495 * @fmt: printf-style format string
4496 * @...: arguments for the format string
4498 * This function can be called by a running work function to describe what
4499 * the work item is about. If the worker task gets dumped, this
4500 * information will be printed out together to help debugging. The
4501 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4503 void set_worker_desc(const char *fmt, ...)
4505 struct worker *worker = current_wq_worker();
4509 va_start(args, fmt);
4510 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4512 worker->desc_valid = true;
4517 * print_worker_info - print out worker information and description
4518 * @log_lvl: the log level to use when printing
4519 * @task: target task
4521 * If @task is a worker and currently executing a work item, print out the
4522 * name of the workqueue being serviced and worker description set with
4523 * set_worker_desc() by the currently executing work item.
4525 * This function can be safely called on any task as long as the
4526 * task_struct itself is accessible. While safe, this function isn't
4527 * synchronized and may print out mixups or garbages of limited length.
4529 void print_worker_info(const char *log_lvl, struct task_struct *task)
4531 work_func_t *fn = NULL;
4532 char name[WQ_NAME_LEN] = { };
4533 char desc[WORKER_DESC_LEN] = { };
4534 struct pool_workqueue *pwq = NULL;
4535 struct workqueue_struct *wq = NULL;
4536 bool desc_valid = false;
4537 struct worker *worker;
4539 if (!(task->flags & PF_WQ_WORKER))
4543 * This function is called without any synchronization and @task
4544 * could be in any state. Be careful with dereferences.
4546 worker = probe_kthread_data(task);
4549 * Carefully copy the associated workqueue's workfn and name. Keep
4550 * the original last '\0' in case the original contains garbage.
4552 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4553 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4554 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4555 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4557 /* copy worker description */
4558 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4560 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4562 if (fn || name[0] || desc[0]) {
4563 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4565 pr_cont(" (%s)", desc);
4573 * There are two challenges in supporting CPU hotplug. Firstly, there
4574 * are a lot of assumptions on strong associations among work, pwq and
4575 * pool which make migrating pending and scheduled works very
4576 * difficult to implement without impacting hot paths. Secondly,
4577 * worker pools serve mix of short, long and very long running works making
4578 * blocked draining impractical.
4580 * This is solved by allowing the pools to be disassociated from the CPU
4581 * running as an unbound one and allowing it to be reattached later if the
4582 * cpu comes back online.
4585 static void wq_unbind_fn(struct work_struct *work)
4587 int cpu = smp_processor_id();
4588 struct worker_pool *pool;
4589 struct worker *worker;
4592 for_each_cpu_worker_pool(pool, cpu) {
4593 WARN_ON_ONCE(cpu != smp_processor_id());
4595 mutex_lock(&pool->manager_mutex);
4596 spin_lock_irq(&pool->lock);
4599 * We've blocked all manager operations. Make all workers
4600 * unbound and set DISASSOCIATED. Before this, all workers
4601 * except for the ones which are still executing works from
4602 * before the last CPU down must be on the cpu. After
4603 * this, they may become diasporas.
4605 for_each_pool_worker(worker, wi, pool)
4606 worker->flags |= WORKER_UNBOUND;
4608 pool->flags |= POOL_DISASSOCIATED;
4610 spin_unlock_irq(&pool->lock);
4611 mutex_unlock(&pool->manager_mutex);
4614 * Call schedule() so that we cross rq->lock and thus can
4615 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4616 * This is necessary as scheduler callbacks may be invoked
4622 * Sched callbacks are disabled now. Zap nr_running.
4623 * After this, nr_running stays zero and need_more_worker()
4624 * and keep_working() are always true as long as the
4625 * worklist is not empty. This pool now behaves as an
4626 * unbound (in terms of concurrency management) pool which
4627 * are served by workers tied to the pool.
4629 atomic_set(&pool->nr_running, 0);
4632 * With concurrency management just turned off, a busy
4633 * worker blocking could lead to lengthy stalls. Kick off
4634 * unbound chain execution of currently pending work items.
4636 spin_lock_irq(&pool->lock);
4637 wake_up_worker(pool);
4638 spin_unlock_irq(&pool->lock);
4643 * rebind_workers - rebind all workers of a pool to the associated CPU
4644 * @pool: pool of interest
4646 * @pool->cpu is coming online. Rebind all workers to the CPU.
4648 static void rebind_workers(struct worker_pool *pool)
4650 struct worker *worker;
4653 lockdep_assert_held(&pool->manager_mutex);
4656 * Restore CPU affinity of all workers. As all idle workers should
4657 * be on the run-queue of the associated CPU before any local
4658 * wake-ups for concurrency management happen, restore CPU affinty
4659 * of all workers first and then clear UNBOUND. As we're called
4660 * from CPU_ONLINE, the following shouldn't fail.
4662 for_each_pool_worker(worker, wi, pool)
4663 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4664 pool->attrs->cpumask) < 0);
4666 spin_lock_irq(&pool->lock);
4668 for_each_pool_worker(worker, wi, pool) {
4669 unsigned int worker_flags = worker->flags;
4672 * A bound idle worker should actually be on the runqueue
4673 * of the associated CPU for local wake-ups targeting it to
4674 * work. Kick all idle workers so that they migrate to the
4675 * associated CPU. Doing this in the same loop as
4676 * replacing UNBOUND with REBOUND is safe as no worker will
4677 * be bound before @pool->lock is released.
4679 if (worker_flags & WORKER_IDLE)
4680 wake_up_process(worker->task);
4683 * We want to clear UNBOUND but can't directly call
4684 * worker_clr_flags() or adjust nr_running. Atomically
4685 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4686 * @worker will clear REBOUND using worker_clr_flags() when
4687 * it initiates the next execution cycle thus restoring
4688 * concurrency management. Note that when or whether
4689 * @worker clears REBOUND doesn't affect correctness.
4691 * ACCESS_ONCE() is necessary because @worker->flags may be
4692 * tested without holding any lock in
4693 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4694 * fail incorrectly leading to premature concurrency
4695 * management operations.
4697 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4698 worker_flags |= WORKER_REBOUND;
4699 worker_flags &= ~WORKER_UNBOUND;
4700 ACCESS_ONCE(worker->flags) = worker_flags;
4703 spin_unlock_irq(&pool->lock);
4707 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4708 * @pool: unbound pool of interest
4709 * @cpu: the CPU which is coming up
4711 * An unbound pool may end up with a cpumask which doesn't have any online
4712 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4713 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4714 * online CPU before, cpus_allowed of all its workers should be restored.
4716 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4718 static cpumask_t cpumask;
4719 struct worker *worker;
4722 lockdep_assert_held(&pool->manager_mutex);
4724 /* is @cpu allowed for @pool? */
4725 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4728 /* is @cpu the only online CPU? */
4729 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4730 if (cpumask_weight(&cpumask) != 1)
4733 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4734 for_each_pool_worker(worker, wi, pool)
4735 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4736 pool->attrs->cpumask) < 0);
4740 * Workqueues should be brought up before normal priority CPU notifiers.
4741 * This will be registered high priority CPU notifier.
4743 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4744 unsigned long action,
4747 int cpu = (unsigned long)hcpu;
4748 struct worker_pool *pool;
4749 struct workqueue_struct *wq;
4752 switch (action & ~CPU_TASKS_FROZEN) {
4753 case CPU_UP_PREPARE:
4754 for_each_cpu_worker_pool(pool, cpu) {
4755 if (pool->nr_workers)
4757 if (create_and_start_worker(pool) < 0)
4762 case CPU_DOWN_FAILED:
4764 mutex_lock(&wq_pool_mutex);
4766 for_each_pool(pool, pi) {
4767 mutex_lock(&pool->manager_mutex);
4769 if (pool->cpu == cpu) {
4770 spin_lock_irq(&pool->lock);
4771 pool->flags &= ~POOL_DISASSOCIATED;
4772 spin_unlock_irq(&pool->lock);
4774 rebind_workers(pool);
4775 } else if (pool->cpu < 0) {
4776 restore_unbound_workers_cpumask(pool, cpu);
4779 mutex_unlock(&pool->manager_mutex);
4782 /* update NUMA affinity of unbound workqueues */
4783 list_for_each_entry(wq, &workqueues, list)
4784 wq_update_unbound_numa(wq, cpu, true);
4786 mutex_unlock(&wq_pool_mutex);
4793 * Workqueues should be brought down after normal priority CPU notifiers.
4794 * This will be registered as low priority CPU notifier.
4796 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4797 unsigned long action,
4800 int cpu = (unsigned long)hcpu;
4801 struct work_struct unbind_work;
4802 struct workqueue_struct *wq;
4804 switch (action & ~CPU_TASKS_FROZEN) {
4805 case CPU_DOWN_PREPARE:
4806 /* unbinding per-cpu workers should happen on the local CPU */
4807 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4808 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4810 /* update NUMA affinity of unbound workqueues */
4811 mutex_lock(&wq_pool_mutex);
4812 list_for_each_entry(wq, &workqueues, list)
4813 wq_update_unbound_numa(wq, cpu, false);
4814 mutex_unlock(&wq_pool_mutex);
4816 /* wait for per-cpu unbinding to finish */
4817 flush_work(&unbind_work);
4825 struct work_for_cpu {
4826 struct work_struct work;
4832 static void work_for_cpu_fn(struct work_struct *work)
4834 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4836 wfc->ret = wfc->fn(wfc->arg);
4840 * work_on_cpu - run a function in user context on a particular cpu
4841 * @cpu: the cpu to run on
4842 * @fn: the function to run
4843 * @arg: the function arg
4845 * This will return the value @fn returns.
4846 * It is up to the caller to ensure that the cpu doesn't go offline.
4847 * The caller must not hold any locks which would prevent @fn from completing.
4849 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4851 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4853 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4854 schedule_work_on(cpu, &wfc.work);
4855 flush_work(&wfc.work);
4858 EXPORT_SYMBOL_GPL(work_on_cpu);
4859 #endif /* CONFIG_SMP */
4861 #ifdef CONFIG_FREEZER
4864 * freeze_workqueues_begin - begin freezing workqueues
4866 * Start freezing workqueues. After this function returns, all freezable
4867 * workqueues will queue new works to their delayed_works list instead of
4871 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4873 void freeze_workqueues_begin(void)
4875 struct worker_pool *pool;
4876 struct workqueue_struct *wq;
4877 struct pool_workqueue *pwq;
4880 mutex_lock(&wq_pool_mutex);
4882 WARN_ON_ONCE(workqueue_freezing);
4883 workqueue_freezing = true;
4886 for_each_pool(pool, pi) {
4887 spin_lock_irq(&pool->lock);
4888 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4889 pool->flags |= POOL_FREEZING;
4890 spin_unlock_irq(&pool->lock);
4893 list_for_each_entry(wq, &workqueues, list) {
4894 mutex_lock(&wq->mutex);
4895 for_each_pwq(pwq, wq)
4896 pwq_adjust_max_active(pwq);
4897 mutex_unlock(&wq->mutex);
4900 mutex_unlock(&wq_pool_mutex);
4904 * freeze_workqueues_busy - are freezable workqueues still busy?
4906 * Check whether freezing is complete. This function must be called
4907 * between freeze_workqueues_begin() and thaw_workqueues().
4910 * Grabs and releases wq_pool_mutex.
4913 * %true if some freezable workqueues are still busy. %false if freezing
4916 bool freeze_workqueues_busy(void)
4919 struct workqueue_struct *wq;
4920 struct pool_workqueue *pwq;
4922 mutex_lock(&wq_pool_mutex);
4924 WARN_ON_ONCE(!workqueue_freezing);
4926 list_for_each_entry(wq, &workqueues, list) {
4927 if (!(wq->flags & WQ_FREEZABLE))
4930 * nr_active is monotonically decreasing. It's safe
4931 * to peek without lock.
4933 rcu_read_lock_sched();
4934 for_each_pwq(pwq, wq) {
4935 WARN_ON_ONCE(pwq->nr_active < 0);
4936 if (pwq->nr_active) {
4938 rcu_read_unlock_sched();
4942 rcu_read_unlock_sched();
4945 mutex_unlock(&wq_pool_mutex);
4950 * thaw_workqueues - thaw workqueues
4952 * Thaw workqueues. Normal queueing is restored and all collected
4953 * frozen works are transferred to their respective pool worklists.
4956 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4958 void thaw_workqueues(void)
4960 struct workqueue_struct *wq;
4961 struct pool_workqueue *pwq;
4962 struct worker_pool *pool;
4965 mutex_lock(&wq_pool_mutex);
4967 if (!workqueue_freezing)
4970 /* clear FREEZING */
4971 for_each_pool(pool, pi) {
4972 spin_lock_irq(&pool->lock);
4973 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4974 pool->flags &= ~POOL_FREEZING;
4975 spin_unlock_irq(&pool->lock);
4978 /* restore max_active and repopulate worklist */
4979 list_for_each_entry(wq, &workqueues, list) {
4980 mutex_lock(&wq->mutex);
4981 for_each_pwq(pwq, wq)
4982 pwq_adjust_max_active(pwq);
4983 mutex_unlock(&wq->mutex);
4986 workqueue_freezing = false;
4988 mutex_unlock(&wq_pool_mutex);
4990 #endif /* CONFIG_FREEZER */
4992 static void __init wq_numa_init(void)
4997 /* determine NUMA pwq table len - highest node id + 1 */
4999 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
5001 if (num_possible_nodes() <= 1)
5004 if (wq_disable_numa) {
5005 pr_info("workqueue: NUMA affinity support disabled\n");
5009 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5010 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5013 * We want masks of possible CPUs of each node which isn't readily
5014 * available. Build one from cpu_to_node() which should have been
5015 * fully initialized by now.
5017 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
5021 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5022 node_online(node) ? node : NUMA_NO_NODE));
5024 for_each_possible_cpu(cpu) {
5025 node = cpu_to_node(cpu);
5026 if (WARN_ON(node == NUMA_NO_NODE)) {
5027 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5028 /* happens iff arch is bonkers, let's just proceed */
5031 cpumask_set_cpu(cpu, tbl[node]);
5034 wq_numa_possible_cpumask = tbl;
5035 wq_numa_enabled = true;
5038 static int __init init_workqueues(void)
5040 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5043 /* make sure we have enough bits for OFFQ pool ID */
5044 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
5045 WORK_CPU_END * NR_STD_WORKER_POOLS);
5047 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5049 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5051 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5052 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5056 /* initialize CPU pools */
5057 for_each_possible_cpu(cpu) {
5058 struct worker_pool *pool;
5061 for_each_cpu_worker_pool(pool, cpu) {
5062 BUG_ON(init_worker_pool(pool));
5064 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5065 pool->attrs->nice = std_nice[i++];
5066 pool->node = cpu_to_node(cpu);
5069 mutex_lock(&wq_pool_mutex);
5070 BUG_ON(worker_pool_assign_id(pool));
5071 mutex_unlock(&wq_pool_mutex);
5075 /* create the initial worker */
5076 for_each_online_cpu(cpu) {
5077 struct worker_pool *pool;
5079 for_each_cpu_worker_pool(pool, cpu) {
5080 pool->flags &= ~POOL_DISASSOCIATED;
5081 BUG_ON(create_and_start_worker(pool) < 0);
5085 /* create default unbound and ordered wq attrs */
5086 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5087 struct workqueue_attrs *attrs;
5089 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5090 attrs->nice = std_nice[i];
5091 unbound_std_wq_attrs[i] = attrs;
5094 * An ordered wq should have only one pwq as ordering is
5095 * guaranteed by max_active which is enforced by pwqs.
5096 * Turn off NUMA so that dfl_pwq is used for all nodes.
5098 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5099 attrs->nice = std_nice[i];
5100 attrs->no_numa = true;
5101 ordered_wq_attrs[i] = attrs;
5104 system_wq = alloc_workqueue("events", 0, 0);
5105 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5106 system_long_wq = alloc_workqueue("events_long", 0, 0);
5107 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5108 WQ_UNBOUND_MAX_ACTIVE);
5109 system_freezable_wq = alloc_workqueue("events_freezable",
5111 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5112 WQ_POWER_EFFICIENT, 0);
5113 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5114 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5116 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5117 !system_unbound_wq || !system_freezable_wq ||
5118 !system_power_efficient_wq ||
5119 !system_freezable_power_efficient_wq);
5122 early_initcall(init_workqueues);