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 are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/nmi.h>
52 #include <linux/kvm_para.h>
54 #include "workqueue_internal.h"
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
69 * Note that DISASSOCIATED should be flipped only while holding
70 * attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * A: pool->attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
136 * sched-RCU for reads.
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
142 * MD: wq_mayday_lock protected.
145 /* struct worker is defined in workqueue_internal.h */
148 spinlock_t lock; /* the pool lock */
149 int cpu; /* I: the associated cpu */
150 int node; /* I: the associated node ID */
151 int id; /* I: pool ID */
152 unsigned int flags; /* X: flags */
154 unsigned long watchdog_ts; /* L: watchdog timestamp */
156 struct list_head worklist; /* L: list of pending works */
157 int nr_workers; /* L: total number of workers */
159 /* nr_idle includes the ones off idle_list for rebinding */
160 int nr_idle; /* L: currently idle ones */
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
170 /* see manage_workers() for details on the two manager mutexes */
171 struct worker *manager; /* L: purely informational */
172 struct mutex attach_mutex; /* attach/detach exclusion */
173 struct list_head workers; /* A: attached workers */
174 struct completion *detach_completion; /* all workers detached */
176 struct ida worker_ida; /* worker IDs for task name */
178 struct workqueue_attrs *attrs; /* I: worker attributes */
179 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
180 int refcnt; /* PL: refcnt for unbound pools */
183 * The current concurrency level. As it's likely to be accessed
184 * from other CPUs during try_to_wake_up(), put it in a separate
187 atomic_t nr_running ____cacheline_aligned_in_smp;
190 * Destruction of pool is sched-RCU protected to allow dereferences
191 * from get_work_pool().
194 } ____cacheline_aligned_in_smp;
197 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
198 * of work_struct->data are used for flags and the remaining high bits
199 * point to the pwq; thus, pwqs need to be aligned at two's power of the
200 * number of flag bits.
202 struct pool_workqueue {
203 struct worker_pool *pool; /* I: the associated pool */
204 struct workqueue_struct *wq; /* I: the owning workqueue */
205 int work_color; /* L: current color */
206 int flush_color; /* L: flushing color */
207 int refcnt; /* L: reference count */
208 int nr_in_flight[WORK_NR_COLORS];
209 /* L: nr of in_flight works */
210 int nr_active; /* L: nr of active works */
211 int max_active; /* L: max active works */
212 struct list_head delayed_works; /* L: delayed works */
213 struct list_head pwqs_node; /* WR: node on wq->pwqs */
214 struct list_head mayday_node; /* MD: node on wq->maydays */
217 * Release of unbound pwq is punted to system_wq. See put_pwq()
218 * and pwq_unbound_release_workfn() for details. pool_workqueue
219 * itself is also sched-RCU protected so that the first pwq can be
220 * determined without grabbing wq->mutex.
222 struct work_struct unbound_release_work;
224 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
227 * Structure used to wait for workqueue flush.
230 struct list_head list; /* WQ: list of flushers */
231 int flush_color; /* WQ: flush color waiting for */
232 struct completion done; /* flush completion */
238 * The externally visible workqueue. It relays the issued work items to
239 * the appropriate worker_pool through its pool_workqueues.
241 struct workqueue_struct {
242 struct list_head pwqs; /* WR: all pwqs of this wq */
243 struct list_head list; /* PR: list of all workqueues */
245 struct mutex mutex; /* protects this wq */
246 int work_color; /* WQ: current work color */
247 int flush_color; /* WQ: current flush color */
248 atomic_t nr_pwqs_to_flush; /* flush in progress */
249 struct wq_flusher *first_flusher; /* WQ: first flusher */
250 struct list_head flusher_queue; /* WQ: flush waiters */
251 struct list_head flusher_overflow; /* WQ: flush overflow list */
253 struct list_head maydays; /* MD: pwqs requesting rescue */
254 struct worker *rescuer; /* I: rescue worker */
256 int nr_drainers; /* WQ: drain in progress */
257 int saved_max_active; /* WQ: saved pwq max_active */
259 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
260 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
263 struct wq_device *wq_dev; /* I: for sysfs interface */
265 #ifdef CONFIG_LOCKDEP
266 struct lockdep_map lockdep_map;
268 char name[WQ_NAME_LEN]; /* I: workqueue name */
271 * Destruction of workqueue_struct is sched-RCU protected to allow
272 * walking the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
283 static struct kmem_cache *pwq_cache;
285 static cpumask_var_t *wq_numa_possible_cpumask;
286 /* possible CPUs of each node */
288 static bool wq_disable_numa;
289 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
291 /* see the comment above the definition of WQ_POWER_EFFICIENT */
292 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
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 */
302 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
304 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
305 static bool workqueue_freezing; /* PL: have wqs started freezing? */
307 /* PL: allowable cpus for unbound wqs and work items */
308 static cpumask_var_t wq_unbound_cpumask;
310 /* CPU where unbound work was last round robin scheduled from this CPU */
311 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
314 * Local execution of unbound work items is no longer guaranteed. The
315 * following always forces round-robin CPU selection on unbound work items
316 * to uncover usages which depend on it.
318 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
319 static bool wq_debug_force_rr_cpu = true;
321 static bool wq_debug_force_rr_cpu = false;
323 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
325 /* the per-cpu worker pools */
326 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
328 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
330 /* PL: hash of all unbound pools keyed by pool->attrs */
331 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
333 /* I: attributes used when instantiating standard unbound pools on demand */
334 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
336 /* I: attributes used when instantiating ordered pools on demand */
337 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
339 struct workqueue_struct *system_wq __read_mostly;
340 EXPORT_SYMBOL(system_wq);
341 struct workqueue_struct *system_highpri_wq __read_mostly;
342 EXPORT_SYMBOL_GPL(system_highpri_wq);
343 struct workqueue_struct *system_long_wq __read_mostly;
344 EXPORT_SYMBOL_GPL(system_long_wq);
345 struct workqueue_struct *system_unbound_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_unbound_wq);
347 struct workqueue_struct *system_freezable_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_freezable_wq);
349 struct workqueue_struct *system_power_efficient_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
351 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
352 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
354 static int worker_thread(void *__worker);
355 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
357 #define CREATE_TRACE_POINTS
358 #include <trace/events/workqueue.h>
360 #define assert_rcu_or_pool_mutex() \
361 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
362 !lockdep_is_held(&wq_pool_mutex), \
363 "sched RCU or wq_pool_mutex should be held")
365 #define assert_rcu_or_wq_mutex(wq) \
366 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
367 !lockdep_is_held(&wq->mutex), \
368 "sched RCU or wq->mutex should be held")
370 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
371 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
372 !lockdep_is_held(&wq->mutex) && \
373 !lockdep_is_held(&wq_pool_mutex), \
374 "sched RCU, wq->mutex or wq_pool_mutex should be held")
376 #define for_each_cpu_worker_pool(pool, cpu) \
377 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
378 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
382 * for_each_pool - iterate through all worker_pools in the system
383 * @pool: iteration cursor
384 * @pi: integer used for iteration
386 * This must be called either with wq_pool_mutex held or sched RCU read
387 * locked. If the pool needs to be used beyond the locking in effect, the
388 * caller is responsible for guaranteeing that the pool stays online.
390 * The if/else clause exists only for the lockdep assertion and can be
393 #define for_each_pool(pool, pi) \
394 idr_for_each_entry(&worker_pool_idr, pool, pi) \
395 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
399 * for_each_pool_worker - iterate through all workers of a worker_pool
400 * @worker: iteration cursor
401 * @pool: worker_pool to iterate workers of
403 * This must be called with @pool->attach_mutex.
405 * The if/else clause exists only for the lockdep assertion and can be
408 #define for_each_pool_worker(worker, pool) \
409 list_for_each_entry((worker), &(pool)->workers, node) \
410 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
414 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
415 * @pwq: iteration cursor
416 * @wq: the target workqueue
418 * This must be called either with wq->mutex held or sched RCU read locked.
419 * If the pwq needs to be used beyond the locking in effect, the caller is
420 * responsible for guaranteeing that the pwq stays online.
422 * The if/else clause exists only for the lockdep assertion and can be
425 #define for_each_pwq(pwq, wq) \
426 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
427 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
430 #ifdef CONFIG_DEBUG_OBJECTS_WORK
432 static struct debug_obj_descr work_debug_descr;
434 static void *work_debug_hint(void *addr)
436 return ((struct work_struct *) addr)->func;
439 static bool work_is_static_object(void *addr)
441 struct work_struct *work = addr;
443 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
447 * fixup_init is called when:
448 * - an active object is initialized
450 static bool work_fixup_init(void *addr, enum debug_obj_state state)
452 struct work_struct *work = addr;
455 case ODEBUG_STATE_ACTIVE:
456 cancel_work_sync(work);
457 debug_object_init(work, &work_debug_descr);
465 * fixup_free is called when:
466 * - an active object is freed
468 static bool work_fixup_free(void *addr, enum debug_obj_state state)
470 struct work_struct *work = addr;
473 case ODEBUG_STATE_ACTIVE:
474 cancel_work_sync(work);
475 debug_object_free(work, &work_debug_descr);
482 static struct debug_obj_descr work_debug_descr = {
483 .name = "work_struct",
484 .debug_hint = work_debug_hint,
485 .is_static_object = work_is_static_object,
486 .fixup_init = work_fixup_init,
487 .fixup_free = work_fixup_free,
490 static inline void debug_work_activate(struct work_struct *work)
492 debug_object_activate(work, &work_debug_descr);
495 static inline void debug_work_deactivate(struct work_struct *work)
497 debug_object_deactivate(work, &work_debug_descr);
500 void __init_work(struct work_struct *work, int onstack)
503 debug_object_init_on_stack(work, &work_debug_descr);
505 debug_object_init(work, &work_debug_descr);
507 EXPORT_SYMBOL_GPL(__init_work);
509 void destroy_work_on_stack(struct work_struct *work)
511 debug_object_free(work, &work_debug_descr);
513 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
515 void destroy_delayed_work_on_stack(struct delayed_work *work)
517 destroy_timer_on_stack(&work->timer);
518 debug_object_free(&work->work, &work_debug_descr);
520 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
523 static inline void debug_work_activate(struct work_struct *work) { }
524 static inline void debug_work_deactivate(struct work_struct *work) { }
528 * worker_pool_assign_id - allocate ID and assing it to @pool
529 * @pool: the pool pointer of interest
531 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
532 * successfully, -errno on failure.
534 static int worker_pool_assign_id(struct worker_pool *pool)
538 lockdep_assert_held(&wq_pool_mutex);
540 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
550 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
551 * @wq: the target workqueue
554 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
556 * If the pwq needs to be used beyond the locking in effect, the caller is
557 * responsible for guaranteeing that the pwq stays online.
559 * Return: The unbound pool_workqueue for @node.
561 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
564 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
567 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
568 * delayed item is pending. The plan is to keep CPU -> NODE
569 * mapping valid and stable across CPU on/offlines. Once that
570 * happens, this workaround can be removed.
572 if (unlikely(node == NUMA_NO_NODE))
575 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
578 static unsigned int work_color_to_flags(int color)
580 return color << WORK_STRUCT_COLOR_SHIFT;
583 static int get_work_color(struct work_struct *work)
585 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
586 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
589 static int work_next_color(int color)
591 return (color + 1) % WORK_NR_COLORS;
595 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
596 * contain the pointer to the queued pwq. Once execution starts, the flag
597 * is cleared and the high bits contain OFFQ flags and pool ID.
599 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
600 * and clear_work_data() can be used to set the pwq, pool or clear
601 * work->data. These functions should only be called while the work is
602 * owned - ie. while the PENDING bit is set.
604 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
605 * corresponding to a work. Pool is available once the work has been
606 * queued anywhere after initialization until it is sync canceled. pwq is
607 * available only while the work item is queued.
609 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
610 * canceled. While being canceled, a work item may have its PENDING set
611 * but stay off timer and worklist for arbitrarily long and nobody should
612 * try to steal the PENDING bit.
614 static inline void set_work_data(struct work_struct *work, unsigned long data,
617 WARN_ON_ONCE(!work_pending(work));
618 atomic_long_set(&work->data, data | flags | work_static(work));
621 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
622 unsigned long extra_flags)
624 set_work_data(work, (unsigned long)pwq,
625 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
628 static void set_work_pool_and_keep_pending(struct work_struct *work,
631 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
632 WORK_STRUCT_PENDING);
635 static void set_work_pool_and_clear_pending(struct work_struct *work,
639 * The following wmb is paired with the implied mb in
640 * test_and_set_bit(PENDING) and ensures all updates to @work made
641 * here are visible to and precede any updates by the next PENDING
645 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
647 * The following mb guarantees that previous clear of a PENDING bit
648 * will not be reordered with any speculative LOADS or STORES from
649 * work->current_func, which is executed afterwards. This possible
650 * reordering can lead to a missed execution on attempt to qeueue
651 * the same @work. E.g. consider this case:
654 * ---------------------------- --------------------------------
656 * 1 STORE event_indicated
657 * 2 queue_work_on() {
658 * 3 test_and_set_bit(PENDING)
659 * 4 } set_..._and_clear_pending() {
660 * 5 set_work_data() # clear bit
662 * 7 work->current_func() {
663 * 8 LOAD event_indicated
666 * Without an explicit full barrier speculative LOAD on line 8 can
667 * be executed before CPU#0 does STORE on line 1. If that happens,
668 * CPU#0 observes the PENDING bit is still set and new execution of
669 * a @work is not queued in a hope, that CPU#1 will eventually
670 * finish the queued @work. Meanwhile CPU#1 does not see
671 * event_indicated is set, because speculative LOAD was executed
672 * before actual STORE.
677 static void clear_work_data(struct work_struct *work)
679 smp_wmb(); /* see set_work_pool_and_clear_pending() */
680 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
683 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
685 unsigned long data = atomic_long_read(&work->data);
687 if (data & WORK_STRUCT_PWQ)
688 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
694 * get_work_pool - return the worker_pool a given work was associated with
695 * @work: the work item of interest
697 * Pools are created and destroyed under wq_pool_mutex, and allows read
698 * access under sched-RCU read lock. As such, this function should be
699 * called under wq_pool_mutex or with preemption disabled.
701 * All fields of the returned pool are accessible as long as the above
702 * mentioned locking is in effect. If the returned pool needs to be used
703 * beyond the critical section, the caller is responsible for ensuring the
704 * returned pool is and stays online.
706 * Return: The worker_pool @work was last associated with. %NULL if none.
708 static struct worker_pool *get_work_pool(struct work_struct *work)
710 unsigned long data = atomic_long_read(&work->data);
713 assert_rcu_or_pool_mutex();
715 if (data & WORK_STRUCT_PWQ)
716 return ((struct pool_workqueue *)
717 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
719 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
720 if (pool_id == WORK_OFFQ_POOL_NONE)
723 return idr_find(&worker_pool_idr, pool_id);
727 * get_work_pool_id - return the worker pool ID a given work is associated with
728 * @work: the work item of interest
730 * Return: The worker_pool ID @work was last associated with.
731 * %WORK_OFFQ_POOL_NONE if none.
733 static int get_work_pool_id(struct work_struct *work)
735 unsigned long data = atomic_long_read(&work->data);
737 if (data & WORK_STRUCT_PWQ)
738 return ((struct pool_workqueue *)
739 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
741 return data >> WORK_OFFQ_POOL_SHIFT;
744 static void mark_work_canceling(struct work_struct *work)
746 unsigned long pool_id = get_work_pool_id(work);
748 pool_id <<= WORK_OFFQ_POOL_SHIFT;
749 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
752 static bool work_is_canceling(struct work_struct *work)
754 unsigned long data = atomic_long_read(&work->data);
756 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
760 * Policy functions. These define the policies on how the global worker
761 * pools are managed. Unless noted otherwise, these functions assume that
762 * they're being called with pool->lock held.
765 static bool __need_more_worker(struct worker_pool *pool)
767 return !atomic_read(&pool->nr_running);
771 * Need to wake up a worker? Called from anything but currently
774 * Note that, because unbound workers never contribute to nr_running, this
775 * function will always return %true for unbound pools as long as the
776 * worklist isn't empty.
778 static bool need_more_worker(struct worker_pool *pool)
780 return !list_empty(&pool->worklist) && __need_more_worker(pool);
783 /* Can I start working? Called from busy but !running workers. */
784 static bool may_start_working(struct worker_pool *pool)
786 return pool->nr_idle;
789 /* Do I need to keep working? Called from currently running workers. */
790 static bool keep_working(struct worker_pool *pool)
792 return !list_empty(&pool->worklist) &&
793 atomic_read(&pool->nr_running) <= 1;
796 /* Do we need a new worker? Called from manager. */
797 static bool need_to_create_worker(struct worker_pool *pool)
799 return need_more_worker(pool) && !may_start_working(pool);
802 /* Do we have too many workers and should some go away? */
803 static bool too_many_workers(struct worker_pool *pool)
805 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
806 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
807 int nr_busy = pool->nr_workers - nr_idle;
809 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
816 /* Return the first idle worker. Safe with preemption disabled */
817 static struct worker *first_idle_worker(struct worker_pool *pool)
819 if (unlikely(list_empty(&pool->idle_list)))
822 return list_first_entry(&pool->idle_list, struct worker, entry);
826 * wake_up_worker - wake up an idle worker
827 * @pool: worker pool to wake worker from
829 * Wake up the first idle worker of @pool.
832 * spin_lock_irq(pool->lock).
834 static void wake_up_worker(struct worker_pool *pool)
836 struct worker *worker = first_idle_worker(pool);
839 wake_up_process(worker->task);
843 * wq_worker_waking_up - a worker is waking up
844 * @task: task waking up
845 * @cpu: CPU @task is waking up to
847 * This function is called during try_to_wake_up() when a worker is
851 * spin_lock_irq(rq->lock)
853 void wq_worker_waking_up(struct task_struct *task, int cpu)
855 struct worker *worker = kthread_data(task);
857 if (!(worker->flags & WORKER_NOT_RUNNING)) {
858 WARN_ON_ONCE(worker->pool->cpu != cpu);
859 atomic_inc(&worker->pool->nr_running);
864 * wq_worker_sleeping - a worker is going to sleep
865 * @task: task going to sleep
867 * This function is called during schedule() when a busy worker is
868 * going to sleep. Worker on the same cpu can be woken up by
869 * returning pointer to its task.
872 * spin_lock_irq(rq->lock)
875 * Worker task on @cpu to wake up, %NULL if none.
877 struct task_struct *wq_worker_sleeping(struct task_struct *task)
879 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
880 struct worker_pool *pool;
883 * Rescuers, which may not have all the fields set up like normal
884 * workers, also reach here, let's not access anything before
885 * checking NOT_RUNNING.
887 if (worker->flags & WORKER_NOT_RUNNING)
892 /* this can only happen on the local cpu */
893 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
897 * The counterpart of the following dec_and_test, implied mb,
898 * worklist not empty test sequence is in insert_work().
899 * Please read comment there.
901 * NOT_RUNNING is clear. This means that we're bound to and
902 * running on the local cpu w/ rq lock held and preemption
903 * disabled, which in turn means that none else could be
904 * manipulating idle_list, so dereferencing idle_list without pool
907 if (atomic_dec_and_test(&pool->nr_running) &&
908 !list_empty(&pool->worklist))
909 to_wakeup = first_idle_worker(pool);
910 return to_wakeup ? to_wakeup->task : NULL;
914 * worker_set_flags - set worker flags and adjust nr_running accordingly
916 * @flags: flags to set
918 * Set @flags in @worker->flags and adjust nr_running accordingly.
921 * spin_lock_irq(pool->lock)
923 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
925 struct worker_pool *pool = worker->pool;
927 WARN_ON_ONCE(worker->task != current);
929 /* If transitioning into NOT_RUNNING, adjust nr_running. */
930 if ((flags & WORKER_NOT_RUNNING) &&
931 !(worker->flags & WORKER_NOT_RUNNING)) {
932 atomic_dec(&pool->nr_running);
935 worker->flags |= flags;
939 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
941 * @flags: flags to clear
943 * Clear @flags in @worker->flags and adjust nr_running accordingly.
946 * spin_lock_irq(pool->lock)
948 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
950 struct worker_pool *pool = worker->pool;
951 unsigned int oflags = worker->flags;
953 WARN_ON_ONCE(worker->task != current);
955 worker->flags &= ~flags;
958 * If transitioning out of NOT_RUNNING, increment nr_running. Note
959 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
960 * of multiple flags, not a single flag.
962 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
963 if (!(worker->flags & WORKER_NOT_RUNNING))
964 atomic_inc(&pool->nr_running);
968 * find_worker_executing_work - find worker which is executing a work
969 * @pool: pool of interest
970 * @work: work to find worker for
972 * Find a worker which is executing @work on @pool by searching
973 * @pool->busy_hash which is keyed by the address of @work. For a worker
974 * to match, its current execution should match the address of @work and
975 * its work function. This is to avoid unwanted dependency between
976 * unrelated work executions through a work item being recycled while still
979 * This is a bit tricky. A work item may be freed once its execution
980 * starts and nothing prevents the freed area from being recycled for
981 * another work item. If the same work item address ends up being reused
982 * before the original execution finishes, workqueue will identify the
983 * recycled work item as currently executing and make it wait until the
984 * current execution finishes, introducing an unwanted dependency.
986 * This function checks the work item address and work function to avoid
987 * false positives. Note that this isn't complete as one may construct a
988 * work function which can introduce dependency onto itself through a
989 * recycled work item. Well, if somebody wants to shoot oneself in the
990 * foot that badly, there's only so much we can do, and if such deadlock
991 * actually occurs, it should be easy to locate the culprit work function.
994 * spin_lock_irq(pool->lock).
997 * Pointer to worker which is executing @work if found, %NULL
1000 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1001 struct work_struct *work)
1003 struct worker *worker;
1005 hash_for_each_possible(pool->busy_hash, worker, hentry,
1006 (unsigned long)work)
1007 if (worker->current_work == work &&
1008 worker->current_func == work->func)
1015 * move_linked_works - move linked works to a list
1016 * @work: start of series of works to be scheduled
1017 * @head: target list to append @work to
1018 * @nextp: out parameter for nested worklist walking
1020 * Schedule linked works starting from @work to @head. Work series to
1021 * be scheduled starts at @work and includes any consecutive work with
1022 * WORK_STRUCT_LINKED set in its predecessor.
1024 * If @nextp is not NULL, it's updated to point to the next work of
1025 * the last scheduled work. This allows move_linked_works() to be
1026 * nested inside outer list_for_each_entry_safe().
1029 * spin_lock_irq(pool->lock).
1031 static void move_linked_works(struct work_struct *work, struct list_head *head,
1032 struct work_struct **nextp)
1034 struct work_struct *n;
1037 * Linked worklist will always end before the end of the list,
1038 * use NULL for list head.
1040 list_for_each_entry_safe_from(work, n, NULL, entry) {
1041 list_move_tail(&work->entry, head);
1042 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1047 * If we're already inside safe list traversal and have moved
1048 * multiple works to the scheduled queue, the next position
1049 * needs to be updated.
1056 * get_pwq - get an extra reference on the specified pool_workqueue
1057 * @pwq: pool_workqueue to get
1059 * Obtain an extra reference on @pwq. The caller should guarantee that
1060 * @pwq has positive refcnt and be holding the matching pool->lock.
1062 static void get_pwq(struct pool_workqueue *pwq)
1064 lockdep_assert_held(&pwq->pool->lock);
1065 WARN_ON_ONCE(pwq->refcnt <= 0);
1070 * put_pwq - put a pool_workqueue reference
1071 * @pwq: pool_workqueue to put
1073 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1074 * destruction. The caller should be holding the matching pool->lock.
1076 static void put_pwq(struct pool_workqueue *pwq)
1078 lockdep_assert_held(&pwq->pool->lock);
1079 if (likely(--pwq->refcnt))
1081 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1084 * @pwq can't be released under pool->lock, bounce to
1085 * pwq_unbound_release_workfn(). This never recurses on the same
1086 * pool->lock as this path is taken only for unbound workqueues and
1087 * the release work item is scheduled on a per-cpu workqueue. To
1088 * avoid lockdep warning, unbound pool->locks are given lockdep
1089 * subclass of 1 in get_unbound_pool().
1091 schedule_work(&pwq->unbound_release_work);
1095 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1096 * @pwq: pool_workqueue to put (can be %NULL)
1098 * put_pwq() with locking. This function also allows %NULL @pwq.
1100 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1104 * As both pwqs and pools are sched-RCU protected, the
1105 * following lock operations are safe.
1107 spin_lock_irq(&pwq->pool->lock);
1109 spin_unlock_irq(&pwq->pool->lock);
1113 static void pwq_activate_delayed_work(struct work_struct *work)
1115 struct pool_workqueue *pwq = get_work_pwq(work);
1117 trace_workqueue_activate_work(work);
1118 if (list_empty(&pwq->pool->worklist))
1119 pwq->pool->watchdog_ts = jiffies;
1120 move_linked_works(work, &pwq->pool->worklist, NULL);
1121 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1125 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1127 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1128 struct work_struct, entry);
1130 pwq_activate_delayed_work(work);
1134 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1135 * @pwq: pwq of interest
1136 * @color: color of work which left the queue
1138 * A work either has completed or is removed from pending queue,
1139 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1142 * spin_lock_irq(pool->lock).
1144 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1146 /* uncolored work items don't participate in flushing or nr_active */
1147 if (color == WORK_NO_COLOR)
1150 pwq->nr_in_flight[color]--;
1153 if (!list_empty(&pwq->delayed_works)) {
1154 /* one down, submit a delayed one */
1155 if (pwq->nr_active < pwq->max_active)
1156 pwq_activate_first_delayed(pwq);
1159 /* is flush in progress and are we at the flushing tip? */
1160 if (likely(pwq->flush_color != color))
1163 /* are there still in-flight works? */
1164 if (pwq->nr_in_flight[color])
1167 /* this pwq is done, clear flush_color */
1168 pwq->flush_color = -1;
1171 * If this was the last pwq, wake up the first flusher. It
1172 * will handle the rest.
1174 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1175 complete(&pwq->wq->first_flusher->done);
1181 * try_to_grab_pending - steal work item from worklist and disable irq
1182 * @work: work item to steal
1183 * @is_dwork: @work is a delayed_work
1184 * @flags: place to store irq state
1186 * Try to grab PENDING bit of @work. This function can handle @work in any
1187 * stable state - idle, on timer or on worklist.
1190 * 1 if @work was pending and we successfully stole PENDING
1191 * 0 if @work was idle and we claimed PENDING
1192 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1193 * -ENOENT if someone else is canceling @work, this state may persist
1194 * for arbitrarily long
1197 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1198 * interrupted while holding PENDING and @work off queue, irq must be
1199 * disabled on entry. This, combined with delayed_work->timer being
1200 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1202 * On successful return, >= 0, irq is disabled and the caller is
1203 * responsible for releasing it using local_irq_restore(*@flags).
1205 * This function is safe to call from any context including IRQ handler.
1207 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1208 unsigned long *flags)
1210 struct worker_pool *pool;
1211 struct pool_workqueue *pwq;
1213 local_irq_save(*flags);
1215 /* try to steal the timer if it exists */
1217 struct delayed_work *dwork = to_delayed_work(work);
1220 * dwork->timer is irqsafe. If del_timer() fails, it's
1221 * guaranteed that the timer is not queued anywhere and not
1222 * running on the local CPU.
1224 if (likely(del_timer(&dwork->timer)))
1228 /* try to claim PENDING the normal way */
1229 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1233 * The queueing is in progress, or it is already queued. Try to
1234 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1236 pool = get_work_pool(work);
1240 spin_lock(&pool->lock);
1242 * work->data is guaranteed to point to pwq only while the work
1243 * item is queued on pwq->wq, and both updating work->data to point
1244 * to pwq on queueing and to pool on dequeueing are done under
1245 * pwq->pool->lock. This in turn guarantees that, if work->data
1246 * points to pwq which is associated with a locked pool, the work
1247 * item is currently queued on that pool.
1249 pwq = get_work_pwq(work);
1250 if (pwq && pwq->pool == pool) {
1251 debug_work_deactivate(work);
1254 * A delayed work item cannot be grabbed directly because
1255 * it might have linked NO_COLOR work items which, if left
1256 * on the delayed_list, will confuse pwq->nr_active
1257 * management later on and cause stall. Make sure the work
1258 * item is activated before grabbing.
1260 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1261 pwq_activate_delayed_work(work);
1263 list_del_init(&work->entry);
1264 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1266 /* work->data points to pwq iff queued, point to pool */
1267 set_work_pool_and_keep_pending(work, pool->id);
1269 spin_unlock(&pool->lock);
1272 spin_unlock(&pool->lock);
1274 local_irq_restore(*flags);
1275 if (work_is_canceling(work))
1282 * insert_work - insert a work into a pool
1283 * @pwq: pwq @work belongs to
1284 * @work: work to insert
1285 * @head: insertion point
1286 * @extra_flags: extra WORK_STRUCT_* flags to set
1288 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1289 * work_struct flags.
1292 * spin_lock_irq(pool->lock).
1294 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1295 struct list_head *head, unsigned int extra_flags)
1297 struct worker_pool *pool = pwq->pool;
1299 /* we own @work, set data and link */
1300 set_work_pwq(work, pwq, extra_flags);
1301 list_add_tail(&work->entry, head);
1305 * Ensure either wq_worker_sleeping() sees the above
1306 * list_add_tail() or we see zero nr_running to avoid workers lying
1307 * around lazily while there are works to be processed.
1311 if (__need_more_worker(pool))
1312 wake_up_worker(pool);
1316 * Test whether @work is being queued from another work executing on the
1319 static bool is_chained_work(struct workqueue_struct *wq)
1321 struct worker *worker;
1323 worker = current_wq_worker();
1325 * Return %true iff I'm a worker execuing a work item on @wq. If
1326 * I'm @worker, it's safe to dereference it without locking.
1328 return worker && worker->current_pwq->wq == wq;
1332 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1333 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1334 * avoid perturbing sensitive tasks.
1336 static int wq_select_unbound_cpu(int cpu)
1338 static bool printed_dbg_warning;
1341 if (likely(!wq_debug_force_rr_cpu)) {
1342 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1344 } else if (!printed_dbg_warning) {
1345 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1346 printed_dbg_warning = true;
1349 if (cpumask_empty(wq_unbound_cpumask))
1352 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1353 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1354 if (unlikely(new_cpu >= nr_cpu_ids)) {
1355 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1356 if (unlikely(new_cpu >= nr_cpu_ids))
1359 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1364 static void __queue_work(int cpu, struct workqueue_struct *wq,
1365 struct work_struct *work)
1367 struct pool_workqueue *pwq;
1368 struct worker_pool *last_pool;
1369 struct list_head *worklist;
1370 unsigned int work_flags;
1371 unsigned int req_cpu = cpu;
1374 * While a work item is PENDING && off queue, a task trying to
1375 * steal the PENDING will busy-loop waiting for it to either get
1376 * queued or lose PENDING. Grabbing PENDING and queueing should
1377 * happen with IRQ disabled.
1379 WARN_ON_ONCE(!irqs_disabled());
1382 /* if draining, only works from the same workqueue are allowed */
1383 if (unlikely(wq->flags & __WQ_DRAINING) &&
1384 WARN_ON_ONCE(!is_chained_work(wq)))
1387 /* pwq which will be used unless @work is executing elsewhere */
1388 if (wq->flags & WQ_UNBOUND) {
1389 if (req_cpu == WORK_CPU_UNBOUND)
1390 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1391 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1393 if (req_cpu == WORK_CPU_UNBOUND)
1394 cpu = raw_smp_processor_id();
1395 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1399 * If @work was previously on a different pool, it might still be
1400 * running there, in which case the work needs to be queued on that
1401 * pool to guarantee non-reentrancy.
1403 last_pool = get_work_pool(work);
1404 if (last_pool && last_pool != pwq->pool) {
1405 struct worker *worker;
1407 spin_lock(&last_pool->lock);
1409 worker = find_worker_executing_work(last_pool, work);
1411 if (worker && worker->current_pwq->wq == wq) {
1412 pwq = worker->current_pwq;
1414 /* meh... not running there, queue here */
1415 spin_unlock(&last_pool->lock);
1416 spin_lock(&pwq->pool->lock);
1419 spin_lock(&pwq->pool->lock);
1423 * pwq is determined and locked. For unbound pools, we could have
1424 * raced with pwq release and it could already be dead. If its
1425 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1426 * without another pwq replacing it in the numa_pwq_tbl or while
1427 * work items are executing on it, so the retrying is guaranteed to
1428 * make forward-progress.
1430 if (unlikely(!pwq->refcnt)) {
1431 if (wq->flags & WQ_UNBOUND) {
1432 spin_unlock(&pwq->pool->lock);
1437 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1441 /* pwq determined, queue */
1442 trace_workqueue_queue_work(req_cpu, pwq, work);
1444 if (WARN_ON(!list_empty(&work->entry))) {
1445 spin_unlock(&pwq->pool->lock);
1449 pwq->nr_in_flight[pwq->work_color]++;
1450 work_flags = work_color_to_flags(pwq->work_color);
1452 if (likely(pwq->nr_active < pwq->max_active)) {
1453 trace_workqueue_activate_work(work);
1455 worklist = &pwq->pool->worklist;
1456 if (list_empty(worklist))
1457 pwq->pool->watchdog_ts = jiffies;
1459 work_flags |= WORK_STRUCT_DELAYED;
1460 worklist = &pwq->delayed_works;
1463 debug_work_activate(work);
1464 insert_work(pwq, work, worklist, work_flags);
1466 spin_unlock(&pwq->pool->lock);
1470 * queue_work_on - queue work on specific cpu
1471 * @cpu: CPU number to execute work on
1472 * @wq: workqueue to use
1473 * @work: work to queue
1475 * We queue the work to a specific CPU, the caller must ensure it
1478 * Return: %false if @work was already on a queue, %true otherwise.
1480 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1481 struct work_struct *work)
1484 unsigned long flags;
1486 local_irq_save(flags);
1488 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1489 __queue_work(cpu, wq, work);
1493 local_irq_restore(flags);
1496 EXPORT_SYMBOL(queue_work_on);
1498 void delayed_work_timer_fn(unsigned long __data)
1500 struct delayed_work *dwork = (struct delayed_work *)__data;
1502 /* should have been called from irqsafe timer with irq already off */
1503 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1505 EXPORT_SYMBOL(delayed_work_timer_fn);
1507 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1508 struct delayed_work *dwork, unsigned long delay)
1510 struct timer_list *timer = &dwork->timer;
1511 struct work_struct *work = &dwork->work;
1514 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1515 timer->data != (unsigned long)dwork);
1516 WARN_ON_ONCE(timer_pending(timer));
1517 WARN_ON_ONCE(!list_empty(&work->entry));
1520 * If @delay is 0, queue @dwork->work immediately. This is for
1521 * both optimization and correctness. The earliest @timer can
1522 * expire is on the closest next tick and delayed_work users depend
1523 * on that there's no such delay when @delay is 0.
1526 __queue_work(cpu, wq, &dwork->work);
1530 timer_stats_timer_set_start_info(&dwork->timer);
1534 timer->expires = jiffies + delay;
1536 if (unlikely(cpu != WORK_CPU_UNBOUND))
1537 add_timer_on(timer, cpu);
1543 * queue_delayed_work_on - queue work on specific CPU after delay
1544 * @cpu: CPU number to execute work on
1545 * @wq: workqueue to use
1546 * @dwork: work to queue
1547 * @delay: number of jiffies to wait before queueing
1549 * Return: %false if @work was already on a queue, %true otherwise. If
1550 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1553 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1554 struct delayed_work *dwork, unsigned long delay)
1556 struct work_struct *work = &dwork->work;
1558 unsigned long flags;
1560 /* read the comment in __queue_work() */
1561 local_irq_save(flags);
1563 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1564 __queue_delayed_work(cpu, wq, dwork, delay);
1568 local_irq_restore(flags);
1571 EXPORT_SYMBOL(queue_delayed_work_on);
1574 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1575 * @cpu: CPU number to execute work on
1576 * @wq: workqueue to use
1577 * @dwork: work to queue
1578 * @delay: number of jiffies to wait before queueing
1580 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1581 * modify @dwork's timer so that it expires after @delay. If @delay is
1582 * zero, @work is guaranteed to be scheduled immediately regardless of its
1585 * Return: %false if @dwork was idle and queued, %true if @dwork was
1586 * pending and its timer was modified.
1588 * This function is safe to call from any context including IRQ handler.
1589 * See try_to_grab_pending() for details.
1591 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1592 struct delayed_work *dwork, unsigned long delay)
1594 unsigned long flags;
1598 ret = try_to_grab_pending(&dwork->work, true, &flags);
1599 } while (unlikely(ret == -EAGAIN));
1601 if (likely(ret >= 0)) {
1602 __queue_delayed_work(cpu, wq, dwork, delay);
1603 local_irq_restore(flags);
1606 /* -ENOENT from try_to_grab_pending() becomes %true */
1609 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1612 * worker_enter_idle - enter idle state
1613 * @worker: worker which is entering idle state
1615 * @worker is entering idle state. Update stats and idle timer if
1619 * spin_lock_irq(pool->lock).
1621 static void worker_enter_idle(struct worker *worker)
1623 struct worker_pool *pool = worker->pool;
1625 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1626 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1627 (worker->hentry.next || worker->hentry.pprev)))
1630 /* can't use worker_set_flags(), also called from create_worker() */
1631 worker->flags |= WORKER_IDLE;
1633 worker->last_active = jiffies;
1635 /* idle_list is LIFO */
1636 list_add(&worker->entry, &pool->idle_list);
1638 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1639 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1642 * Sanity check nr_running. Because wq_unbind_fn() releases
1643 * pool->lock between setting %WORKER_UNBOUND and zapping
1644 * nr_running, the warning may trigger spuriously. Check iff
1645 * unbind is not in progress.
1647 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1648 pool->nr_workers == pool->nr_idle &&
1649 atomic_read(&pool->nr_running));
1653 * worker_leave_idle - leave idle state
1654 * @worker: worker which is leaving idle state
1656 * @worker is leaving idle state. Update stats.
1659 * spin_lock_irq(pool->lock).
1661 static void worker_leave_idle(struct worker *worker)
1663 struct worker_pool *pool = worker->pool;
1665 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1667 worker_clr_flags(worker, WORKER_IDLE);
1669 list_del_init(&worker->entry);
1672 static struct worker *alloc_worker(int node)
1674 struct worker *worker;
1676 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1678 INIT_LIST_HEAD(&worker->entry);
1679 INIT_LIST_HEAD(&worker->scheduled);
1680 INIT_LIST_HEAD(&worker->node);
1681 /* on creation a worker is in !idle && prep state */
1682 worker->flags = WORKER_PREP;
1688 * worker_attach_to_pool() - attach a worker to a pool
1689 * @worker: worker to be attached
1690 * @pool: the target pool
1692 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1693 * cpu-binding of @worker are kept coordinated with the pool across
1696 static void worker_attach_to_pool(struct worker *worker,
1697 struct worker_pool *pool)
1699 mutex_lock(&pool->attach_mutex);
1702 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1703 * online CPUs. It'll be re-applied when any of the CPUs come up.
1705 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1708 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1709 * stable across this function. See the comments above the
1710 * flag definition for details.
1712 if (pool->flags & POOL_DISASSOCIATED)
1713 worker->flags |= WORKER_UNBOUND;
1715 list_add_tail(&worker->node, &pool->workers);
1717 mutex_unlock(&pool->attach_mutex);
1721 * worker_detach_from_pool() - detach a worker from its pool
1722 * @worker: worker which is attached to its pool
1723 * @pool: the pool @worker is attached to
1725 * Undo the attaching which had been done in worker_attach_to_pool(). The
1726 * caller worker shouldn't access to the pool after detached except it has
1727 * other reference to the pool.
1729 static void worker_detach_from_pool(struct worker *worker,
1730 struct worker_pool *pool)
1732 struct completion *detach_completion = NULL;
1734 mutex_lock(&pool->attach_mutex);
1735 list_del(&worker->node);
1736 if (list_empty(&pool->workers))
1737 detach_completion = pool->detach_completion;
1738 mutex_unlock(&pool->attach_mutex);
1740 /* clear leftover flags without pool->lock after it is detached */
1741 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1743 if (detach_completion)
1744 complete(detach_completion);
1748 * create_worker - create a new workqueue worker
1749 * @pool: pool the new worker will belong to
1751 * Create and start a new worker which is attached to @pool.
1754 * Might sleep. Does GFP_KERNEL allocations.
1757 * Pointer to the newly created worker.
1759 static struct worker *create_worker(struct worker_pool *pool)
1761 struct worker *worker = NULL;
1765 /* ID is needed to determine kthread name */
1766 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1770 worker = alloc_worker(pool->node);
1774 worker->pool = pool;
1778 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1779 pool->attrs->nice < 0 ? "H" : "");
1781 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1783 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1784 "kworker/%s", id_buf);
1785 if (IS_ERR(worker->task))
1788 set_user_nice(worker->task, pool->attrs->nice);
1789 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1791 /* successful, attach the worker to the pool */
1792 worker_attach_to_pool(worker, pool);
1794 /* start the newly created worker */
1795 spin_lock_irq(&pool->lock);
1796 worker->pool->nr_workers++;
1797 worker_enter_idle(worker);
1798 wake_up_process(worker->task);
1799 spin_unlock_irq(&pool->lock);
1805 ida_simple_remove(&pool->worker_ida, id);
1811 * destroy_worker - destroy a workqueue worker
1812 * @worker: worker to be destroyed
1814 * Destroy @worker and adjust @pool stats accordingly. The worker should
1818 * spin_lock_irq(pool->lock).
1820 static void destroy_worker(struct worker *worker)
1822 struct worker_pool *pool = worker->pool;
1824 lockdep_assert_held(&pool->lock);
1826 /* sanity check frenzy */
1827 if (WARN_ON(worker->current_work) ||
1828 WARN_ON(!list_empty(&worker->scheduled)) ||
1829 WARN_ON(!(worker->flags & WORKER_IDLE)))
1835 list_del_init(&worker->entry);
1836 worker->flags |= WORKER_DIE;
1837 wake_up_process(worker->task);
1840 static void idle_worker_timeout(unsigned long __pool)
1842 struct worker_pool *pool = (void *)__pool;
1844 spin_lock_irq(&pool->lock);
1846 while (too_many_workers(pool)) {
1847 struct worker *worker;
1848 unsigned long expires;
1850 /* idle_list is kept in LIFO order, check the last one */
1851 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1852 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1854 if (time_before(jiffies, expires)) {
1855 mod_timer(&pool->idle_timer, expires);
1859 destroy_worker(worker);
1862 spin_unlock_irq(&pool->lock);
1865 static void send_mayday(struct work_struct *work)
1867 struct pool_workqueue *pwq = get_work_pwq(work);
1868 struct workqueue_struct *wq = pwq->wq;
1870 lockdep_assert_held(&wq_mayday_lock);
1875 /* mayday mayday mayday */
1876 if (list_empty(&pwq->mayday_node)) {
1878 * If @pwq is for an unbound wq, its base ref may be put at
1879 * any time due to an attribute change. Pin @pwq until the
1880 * rescuer is done with it.
1883 list_add_tail(&pwq->mayday_node, &wq->maydays);
1884 wake_up_process(wq->rescuer->task);
1888 static void pool_mayday_timeout(unsigned long __pool)
1890 struct worker_pool *pool = (void *)__pool;
1891 struct work_struct *work;
1893 spin_lock_irq(&pool->lock);
1894 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1896 if (need_to_create_worker(pool)) {
1898 * We've been trying to create a new worker but
1899 * haven't been successful. We might be hitting an
1900 * allocation deadlock. Send distress signals to
1903 list_for_each_entry(work, &pool->worklist, entry)
1907 spin_unlock(&wq_mayday_lock);
1908 spin_unlock_irq(&pool->lock);
1910 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1914 * maybe_create_worker - create a new worker if necessary
1915 * @pool: pool to create a new worker for
1917 * Create a new worker for @pool if necessary. @pool is guaranteed to
1918 * have at least one idle worker on return from this function. If
1919 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1920 * sent to all rescuers with works scheduled on @pool to resolve
1921 * possible allocation deadlock.
1923 * On return, need_to_create_worker() is guaranteed to be %false and
1924 * may_start_working() %true.
1927 * spin_lock_irq(pool->lock) which may be released and regrabbed
1928 * multiple times. Does GFP_KERNEL allocations. Called only from
1931 static void maybe_create_worker(struct worker_pool *pool)
1932 __releases(&pool->lock)
1933 __acquires(&pool->lock)
1936 spin_unlock_irq(&pool->lock);
1938 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1939 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1942 if (create_worker(pool) || !need_to_create_worker(pool))
1945 schedule_timeout_interruptible(CREATE_COOLDOWN);
1947 if (!need_to_create_worker(pool))
1951 del_timer_sync(&pool->mayday_timer);
1952 spin_lock_irq(&pool->lock);
1954 * This is necessary even after a new worker was just successfully
1955 * created as @pool->lock was dropped and the new worker might have
1956 * already become busy.
1958 if (need_to_create_worker(pool))
1963 * manage_workers - manage worker pool
1966 * Assume the manager role and manage the worker pool @worker belongs
1967 * to. At any given time, there can be only zero or one manager per
1968 * pool. The exclusion is handled automatically by this function.
1970 * The caller can safely start processing works on false return. On
1971 * true return, it's guaranteed that need_to_create_worker() is false
1972 * and may_start_working() is true.
1975 * spin_lock_irq(pool->lock) which may be released and regrabbed
1976 * multiple times. Does GFP_KERNEL allocations.
1979 * %false if the pool doesn't need management and the caller can safely
1980 * start processing works, %true if management function was performed and
1981 * the conditions that the caller verified before calling the function may
1982 * no longer be true.
1984 static bool manage_workers(struct worker *worker)
1986 struct worker_pool *pool = worker->pool;
1988 if (pool->flags & POOL_MANAGER_ACTIVE)
1991 pool->flags |= POOL_MANAGER_ACTIVE;
1992 pool->manager = worker;
1994 maybe_create_worker(pool);
1996 pool->manager = NULL;
1997 pool->flags &= ~POOL_MANAGER_ACTIVE;
1998 wake_up(&wq_manager_wait);
2003 * process_one_work - process single work
2005 * @work: work to process
2007 * Process @work. This function contains all the logics necessary to
2008 * process a single work including synchronization against and
2009 * interaction with other workers on the same cpu, queueing and
2010 * flushing. As long as context requirement is met, any worker can
2011 * call this function to process a work.
2014 * spin_lock_irq(pool->lock) which is released and regrabbed.
2016 static void process_one_work(struct worker *worker, struct work_struct *work)
2017 __releases(&pool->lock)
2018 __acquires(&pool->lock)
2020 struct pool_workqueue *pwq = get_work_pwq(work);
2021 struct worker_pool *pool = worker->pool;
2022 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2024 struct worker *collision;
2025 #ifdef CONFIG_LOCKDEP
2027 * It is permissible to free the struct work_struct from
2028 * inside the function that is called from it, this we need to
2029 * take into account for lockdep too. To avoid bogus "held
2030 * lock freed" warnings as well as problems when looking into
2031 * work->lockdep_map, make a copy and use that here.
2033 struct lockdep_map lockdep_map;
2035 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2037 /* ensure we're on the correct CPU */
2038 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2039 raw_smp_processor_id() != pool->cpu);
2042 * A single work shouldn't be executed concurrently by
2043 * multiple workers on a single cpu. Check whether anyone is
2044 * already processing the work. If so, defer the work to the
2045 * currently executing one.
2047 collision = find_worker_executing_work(pool, work);
2048 if (unlikely(collision)) {
2049 move_linked_works(work, &collision->scheduled, NULL);
2053 /* claim and dequeue */
2054 debug_work_deactivate(work);
2055 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2056 worker->current_work = work;
2057 worker->current_func = work->func;
2058 worker->current_pwq = pwq;
2059 work_color = get_work_color(work);
2061 list_del_init(&work->entry);
2064 * CPU intensive works don't participate in concurrency management.
2065 * They're the scheduler's responsibility. This takes @worker out
2066 * of concurrency management and the next code block will chain
2067 * execution of the pending work items.
2069 if (unlikely(cpu_intensive))
2070 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2073 * Wake up another worker if necessary. The condition is always
2074 * false for normal per-cpu workers since nr_running would always
2075 * be >= 1 at this point. This is used to chain execution of the
2076 * pending work items for WORKER_NOT_RUNNING workers such as the
2077 * UNBOUND and CPU_INTENSIVE ones.
2079 if (need_more_worker(pool))
2080 wake_up_worker(pool);
2083 * Record the last pool and clear PENDING which should be the last
2084 * update to @work. Also, do this inside @pool->lock so that
2085 * PENDING and queued state changes happen together while IRQ is
2088 set_work_pool_and_clear_pending(work, pool->id);
2090 spin_unlock_irq(&pool->lock);
2092 lock_map_acquire_read(&pwq->wq->lockdep_map);
2093 lock_map_acquire(&lockdep_map);
2094 trace_workqueue_execute_start(work);
2095 worker->current_func(work);
2097 * While we must be careful to not use "work" after this, the trace
2098 * point will only record its address.
2100 trace_workqueue_execute_end(work);
2101 lock_map_release(&lockdep_map);
2102 lock_map_release(&pwq->wq->lockdep_map);
2104 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2105 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2106 " last function: %pf\n",
2107 current->comm, preempt_count(), task_pid_nr(current),
2108 worker->current_func);
2109 debug_show_held_locks(current);
2114 * The following prevents a kworker from hogging CPU on !PREEMPT
2115 * kernels, where a requeueing work item waiting for something to
2116 * happen could deadlock with stop_machine as such work item could
2117 * indefinitely requeue itself while all other CPUs are trapped in
2118 * stop_machine. At the same time, report a quiescent RCU state so
2119 * the same condition doesn't freeze RCU.
2121 cond_resched_rcu_qs();
2123 spin_lock_irq(&pool->lock);
2125 /* clear cpu intensive status */
2126 if (unlikely(cpu_intensive))
2127 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2129 /* we're done with it, release */
2130 hash_del(&worker->hentry);
2131 worker->current_work = NULL;
2132 worker->current_func = NULL;
2133 worker->current_pwq = NULL;
2134 worker->desc_valid = false;
2135 pwq_dec_nr_in_flight(pwq, work_color);
2139 * process_scheduled_works - process scheduled works
2142 * Process all scheduled works. Please note that the scheduled list
2143 * may change while processing a work, so this function repeatedly
2144 * fetches a work from the top and executes it.
2147 * spin_lock_irq(pool->lock) which may be released and regrabbed
2150 static void process_scheduled_works(struct worker *worker)
2152 while (!list_empty(&worker->scheduled)) {
2153 struct work_struct *work = list_first_entry(&worker->scheduled,
2154 struct work_struct, entry);
2155 process_one_work(worker, work);
2160 * worker_thread - the worker thread function
2163 * The worker thread function. All workers belong to a worker_pool -
2164 * either a per-cpu one or dynamic unbound one. These workers process all
2165 * work items regardless of their specific target workqueue. The only
2166 * exception is work items which belong to workqueues with a rescuer which
2167 * will be explained in rescuer_thread().
2171 static int worker_thread(void *__worker)
2173 struct worker *worker = __worker;
2174 struct worker_pool *pool = worker->pool;
2176 /* tell the scheduler that this is a workqueue worker */
2177 worker->task->flags |= PF_WQ_WORKER;
2179 spin_lock_irq(&pool->lock);
2181 /* am I supposed to die? */
2182 if (unlikely(worker->flags & WORKER_DIE)) {
2183 spin_unlock_irq(&pool->lock);
2184 WARN_ON_ONCE(!list_empty(&worker->entry));
2185 worker->task->flags &= ~PF_WQ_WORKER;
2187 set_task_comm(worker->task, "kworker/dying");
2188 ida_simple_remove(&pool->worker_ida, worker->id);
2189 worker_detach_from_pool(worker, pool);
2194 worker_leave_idle(worker);
2196 /* no more worker necessary? */
2197 if (!need_more_worker(pool))
2200 /* do we need to manage? */
2201 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2205 * ->scheduled list can only be filled while a worker is
2206 * preparing to process a work or actually processing it.
2207 * Make sure nobody diddled with it while I was sleeping.
2209 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2212 * Finish PREP stage. We're guaranteed to have at least one idle
2213 * worker or that someone else has already assumed the manager
2214 * role. This is where @worker starts participating in concurrency
2215 * management if applicable and concurrency management is restored
2216 * after being rebound. See rebind_workers() for details.
2218 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2221 struct work_struct *work =
2222 list_first_entry(&pool->worklist,
2223 struct work_struct, entry);
2225 pool->watchdog_ts = jiffies;
2227 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2228 /* optimization path, not strictly necessary */
2229 process_one_work(worker, work);
2230 if (unlikely(!list_empty(&worker->scheduled)))
2231 process_scheduled_works(worker);
2233 move_linked_works(work, &worker->scheduled, NULL);
2234 process_scheduled_works(worker);
2236 } while (keep_working(pool));
2238 worker_set_flags(worker, WORKER_PREP);
2241 * pool->lock is held and there's no work to process and no need to
2242 * manage, sleep. Workers are woken up only while holding
2243 * pool->lock or from local cpu, so setting the current state
2244 * before releasing pool->lock is enough to prevent losing any
2247 worker_enter_idle(worker);
2248 __set_current_state(TASK_INTERRUPTIBLE);
2249 spin_unlock_irq(&pool->lock);
2255 * rescuer_thread - the rescuer thread function
2258 * Workqueue rescuer thread function. There's one rescuer for each
2259 * workqueue which has WQ_MEM_RECLAIM set.
2261 * Regular work processing on a pool may block trying to create a new
2262 * worker which uses GFP_KERNEL allocation which has slight chance of
2263 * developing into deadlock if some works currently on the same queue
2264 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2265 * the problem rescuer solves.
2267 * When such condition is possible, the pool summons rescuers of all
2268 * workqueues which have works queued on the pool and let them process
2269 * those works so that forward progress can be guaranteed.
2271 * This should happen rarely.
2275 static int rescuer_thread(void *__rescuer)
2277 struct worker *rescuer = __rescuer;
2278 struct workqueue_struct *wq = rescuer->rescue_wq;
2279 struct list_head *scheduled = &rescuer->scheduled;
2282 set_user_nice(current, RESCUER_NICE_LEVEL);
2285 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2286 * doesn't participate in concurrency management.
2288 rescuer->task->flags |= PF_WQ_WORKER;
2290 set_current_state(TASK_INTERRUPTIBLE);
2293 * By the time the rescuer is requested to stop, the workqueue
2294 * shouldn't have any work pending, but @wq->maydays may still have
2295 * pwq(s) queued. This can happen by non-rescuer workers consuming
2296 * all the work items before the rescuer got to them. Go through
2297 * @wq->maydays processing before acting on should_stop so that the
2298 * list is always empty on exit.
2300 should_stop = kthread_should_stop();
2302 /* see whether any pwq is asking for help */
2303 spin_lock_irq(&wq_mayday_lock);
2305 while (!list_empty(&wq->maydays)) {
2306 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2307 struct pool_workqueue, mayday_node);
2308 struct worker_pool *pool = pwq->pool;
2309 struct work_struct *work, *n;
2312 __set_current_state(TASK_RUNNING);
2313 list_del_init(&pwq->mayday_node);
2315 spin_unlock_irq(&wq_mayday_lock);
2317 worker_attach_to_pool(rescuer, pool);
2319 spin_lock_irq(&pool->lock);
2320 rescuer->pool = pool;
2323 * Slurp in all works issued via this workqueue and
2326 WARN_ON_ONCE(!list_empty(scheduled));
2327 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2328 if (get_work_pwq(work) == pwq) {
2330 pool->watchdog_ts = jiffies;
2331 move_linked_works(work, scheduled, &n);
2336 if (!list_empty(scheduled)) {
2337 process_scheduled_works(rescuer);
2340 * The above execution of rescued work items could
2341 * have created more to rescue through
2342 * pwq_activate_first_delayed() or chained
2343 * queueing. Let's put @pwq back on mayday list so
2344 * that such back-to-back work items, which may be
2345 * being used to relieve memory pressure, don't
2346 * incur MAYDAY_INTERVAL delay inbetween.
2348 if (need_to_create_worker(pool)) {
2349 spin_lock(&wq_mayday_lock);
2351 * Queue iff we aren't racing destruction
2352 * and somebody else hasn't queued it already.
2354 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2356 list_add_tail(&pwq->mayday_node, &wq->maydays);
2358 spin_unlock(&wq_mayday_lock);
2363 * Put the reference grabbed by send_mayday(). @pool won't
2364 * go away while we're still attached to it.
2369 * Leave this pool. If need_more_worker() is %true, notify a
2370 * regular worker; otherwise, we end up with 0 concurrency
2371 * and stalling the execution.
2373 if (need_more_worker(pool))
2374 wake_up_worker(pool);
2376 rescuer->pool = NULL;
2377 spin_unlock_irq(&pool->lock);
2379 worker_detach_from_pool(rescuer, pool);
2381 spin_lock_irq(&wq_mayday_lock);
2384 spin_unlock_irq(&wq_mayday_lock);
2387 __set_current_state(TASK_RUNNING);
2388 rescuer->task->flags &= ~PF_WQ_WORKER;
2392 /* rescuers should never participate in concurrency management */
2393 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2399 * check_flush_dependency - check for flush dependency sanity
2400 * @target_wq: workqueue being flushed
2401 * @target_work: work item being flushed (NULL for workqueue flushes)
2403 * %current is trying to flush the whole @target_wq or @target_work on it.
2404 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2405 * reclaiming memory or running on a workqueue which doesn't have
2406 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2409 static void check_flush_dependency(struct workqueue_struct *target_wq,
2410 struct work_struct *target_work)
2412 work_func_t target_func = target_work ? target_work->func : NULL;
2413 struct worker *worker;
2415 if (target_wq->flags & WQ_MEM_RECLAIM)
2418 worker = current_wq_worker();
2420 WARN_ONCE(current->flags & PF_MEMALLOC,
2421 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2422 current->pid, current->comm, target_wq->name, target_func);
2423 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2424 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2425 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2426 worker->current_pwq->wq->name, worker->current_func,
2427 target_wq->name, target_func);
2431 struct work_struct work;
2432 struct completion done;
2433 struct task_struct *task; /* purely informational */
2436 static void wq_barrier_func(struct work_struct *work)
2438 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2439 complete(&barr->done);
2443 * insert_wq_barrier - insert a barrier work
2444 * @pwq: pwq to insert barrier into
2445 * @barr: wq_barrier to insert
2446 * @target: target work to attach @barr to
2447 * @worker: worker currently executing @target, NULL if @target is not executing
2449 * @barr is linked to @target such that @barr is completed only after
2450 * @target finishes execution. Please note that the ordering
2451 * guarantee is observed only with respect to @target and on the local
2454 * Currently, a queued barrier can't be canceled. This is because
2455 * try_to_grab_pending() can't determine whether the work to be
2456 * grabbed is at the head of the queue and thus can't clear LINKED
2457 * flag of the previous work while there must be a valid next work
2458 * after a work with LINKED flag set.
2460 * Note that when @worker is non-NULL, @target may be modified
2461 * underneath us, so we can't reliably determine pwq from @target.
2464 * spin_lock_irq(pool->lock).
2466 static void insert_wq_barrier(struct pool_workqueue *pwq,
2467 struct wq_barrier *barr,
2468 struct work_struct *target, struct worker *worker)
2470 struct list_head *head;
2471 unsigned int linked = 0;
2474 * debugobject calls are safe here even with pool->lock locked
2475 * as we know for sure that this will not trigger any of the
2476 * checks and call back into the fixup functions where we
2479 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2480 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2481 init_completion(&barr->done);
2482 barr->task = current;
2485 * If @target is currently being executed, schedule the
2486 * barrier to the worker; otherwise, put it after @target.
2489 head = worker->scheduled.next;
2491 unsigned long *bits = work_data_bits(target);
2493 head = target->entry.next;
2494 /* there can already be other linked works, inherit and set */
2495 linked = *bits & WORK_STRUCT_LINKED;
2496 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2499 debug_work_activate(&barr->work);
2500 insert_work(pwq, &barr->work, head,
2501 work_color_to_flags(WORK_NO_COLOR) | linked);
2505 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2506 * @wq: workqueue being flushed
2507 * @flush_color: new flush color, < 0 for no-op
2508 * @work_color: new work color, < 0 for no-op
2510 * Prepare pwqs for workqueue flushing.
2512 * If @flush_color is non-negative, flush_color on all pwqs should be
2513 * -1. If no pwq has in-flight commands at the specified color, all
2514 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2515 * has in flight commands, its pwq->flush_color is set to
2516 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2517 * wakeup logic is armed and %true is returned.
2519 * The caller should have initialized @wq->first_flusher prior to
2520 * calling this function with non-negative @flush_color. If
2521 * @flush_color is negative, no flush color update is done and %false
2524 * If @work_color is non-negative, all pwqs should have the same
2525 * work_color which is previous to @work_color and all will be
2526 * advanced to @work_color.
2529 * mutex_lock(wq->mutex).
2532 * %true if @flush_color >= 0 and there's something to flush. %false
2535 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2536 int flush_color, int work_color)
2539 struct pool_workqueue *pwq;
2541 if (flush_color >= 0) {
2542 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2543 atomic_set(&wq->nr_pwqs_to_flush, 1);
2546 for_each_pwq(pwq, wq) {
2547 struct worker_pool *pool = pwq->pool;
2549 spin_lock_irq(&pool->lock);
2551 if (flush_color >= 0) {
2552 WARN_ON_ONCE(pwq->flush_color != -1);
2554 if (pwq->nr_in_flight[flush_color]) {
2555 pwq->flush_color = flush_color;
2556 atomic_inc(&wq->nr_pwqs_to_flush);
2561 if (work_color >= 0) {
2562 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2563 pwq->work_color = work_color;
2566 spin_unlock_irq(&pool->lock);
2569 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2570 complete(&wq->first_flusher->done);
2576 * flush_workqueue - ensure that any scheduled work has run to completion.
2577 * @wq: workqueue to flush
2579 * This function sleeps until all work items which were queued on entry
2580 * have finished execution, but it is not livelocked by new incoming ones.
2582 void flush_workqueue(struct workqueue_struct *wq)
2584 struct wq_flusher this_flusher = {
2585 .list = LIST_HEAD_INIT(this_flusher.list),
2587 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2591 lock_map_acquire(&wq->lockdep_map);
2592 lock_map_release(&wq->lockdep_map);
2594 mutex_lock(&wq->mutex);
2597 * Start-to-wait phase
2599 next_color = work_next_color(wq->work_color);
2601 if (next_color != wq->flush_color) {
2603 * Color space is not full. The current work_color
2604 * becomes our flush_color and work_color is advanced
2607 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2608 this_flusher.flush_color = wq->work_color;
2609 wq->work_color = next_color;
2611 if (!wq->first_flusher) {
2612 /* no flush in progress, become the first flusher */
2613 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2615 wq->first_flusher = &this_flusher;
2617 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2619 /* nothing to flush, done */
2620 wq->flush_color = next_color;
2621 wq->first_flusher = NULL;
2626 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2627 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2628 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2632 * Oops, color space is full, wait on overflow queue.
2633 * The next flush completion will assign us
2634 * flush_color and transfer to flusher_queue.
2636 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2639 check_flush_dependency(wq, NULL);
2641 mutex_unlock(&wq->mutex);
2643 wait_for_completion(&this_flusher.done);
2646 * Wake-up-and-cascade phase
2648 * First flushers are responsible for cascading flushes and
2649 * handling overflow. Non-first flushers can simply return.
2651 if (wq->first_flusher != &this_flusher)
2654 mutex_lock(&wq->mutex);
2656 /* we might have raced, check again with mutex held */
2657 if (wq->first_flusher != &this_flusher)
2660 wq->first_flusher = NULL;
2662 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2663 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2666 struct wq_flusher *next, *tmp;
2668 /* complete all the flushers sharing the current flush color */
2669 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2670 if (next->flush_color != wq->flush_color)
2672 list_del_init(&next->list);
2673 complete(&next->done);
2676 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2677 wq->flush_color != work_next_color(wq->work_color));
2679 /* this flush_color is finished, advance by one */
2680 wq->flush_color = work_next_color(wq->flush_color);
2682 /* one color has been freed, handle overflow queue */
2683 if (!list_empty(&wq->flusher_overflow)) {
2685 * Assign the same color to all overflowed
2686 * flushers, advance work_color and append to
2687 * flusher_queue. This is the start-to-wait
2688 * phase for these overflowed flushers.
2690 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2691 tmp->flush_color = wq->work_color;
2693 wq->work_color = work_next_color(wq->work_color);
2695 list_splice_tail_init(&wq->flusher_overflow,
2696 &wq->flusher_queue);
2697 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2700 if (list_empty(&wq->flusher_queue)) {
2701 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2706 * Need to flush more colors. Make the next flusher
2707 * the new first flusher and arm pwqs.
2709 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2710 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2712 list_del_init(&next->list);
2713 wq->first_flusher = next;
2715 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2719 * Meh... this color is already done, clear first
2720 * flusher and repeat cascading.
2722 wq->first_flusher = NULL;
2726 mutex_unlock(&wq->mutex);
2728 EXPORT_SYMBOL(flush_workqueue);
2731 * drain_workqueue - drain a workqueue
2732 * @wq: workqueue to drain
2734 * Wait until the workqueue becomes empty. While draining is in progress,
2735 * only chain queueing is allowed. IOW, only currently pending or running
2736 * work items on @wq can queue further work items on it. @wq is flushed
2737 * repeatedly until it becomes empty. The number of flushing is determined
2738 * by the depth of chaining and should be relatively short. Whine if it
2741 void drain_workqueue(struct workqueue_struct *wq)
2743 unsigned int flush_cnt = 0;
2744 struct pool_workqueue *pwq;
2747 * __queue_work() needs to test whether there are drainers, is much
2748 * hotter than drain_workqueue() and already looks at @wq->flags.
2749 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2751 mutex_lock(&wq->mutex);
2752 if (!wq->nr_drainers++)
2753 wq->flags |= __WQ_DRAINING;
2754 mutex_unlock(&wq->mutex);
2756 flush_workqueue(wq);
2758 mutex_lock(&wq->mutex);
2760 for_each_pwq(pwq, wq) {
2763 spin_lock_irq(&pwq->pool->lock);
2764 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2765 spin_unlock_irq(&pwq->pool->lock);
2770 if (++flush_cnt == 10 ||
2771 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2772 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2773 wq->name, flush_cnt);
2775 mutex_unlock(&wq->mutex);
2779 if (!--wq->nr_drainers)
2780 wq->flags &= ~__WQ_DRAINING;
2781 mutex_unlock(&wq->mutex);
2783 EXPORT_SYMBOL_GPL(drain_workqueue);
2785 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2787 struct worker *worker = NULL;
2788 struct worker_pool *pool;
2789 struct pool_workqueue *pwq;
2793 local_irq_disable();
2794 pool = get_work_pool(work);
2800 spin_lock(&pool->lock);
2801 /* see the comment in try_to_grab_pending() with the same code */
2802 pwq = get_work_pwq(work);
2804 if (unlikely(pwq->pool != pool))
2807 worker = find_worker_executing_work(pool, work);
2810 pwq = worker->current_pwq;
2813 check_flush_dependency(pwq->wq, work);
2815 insert_wq_barrier(pwq, barr, work, worker);
2816 spin_unlock_irq(&pool->lock);
2819 * If @max_active is 1 or rescuer is in use, flushing another work
2820 * item on the same workqueue may lead to deadlock. Make sure the
2821 * flusher is not running on the same workqueue by verifying write
2824 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2825 lock_map_acquire(&pwq->wq->lockdep_map);
2827 lock_map_acquire_read(&pwq->wq->lockdep_map);
2828 lock_map_release(&pwq->wq->lockdep_map);
2832 spin_unlock_irq(&pool->lock);
2837 * flush_work - wait for a work to finish executing the last queueing instance
2838 * @work: the work to flush
2840 * Wait until @work has finished execution. @work is guaranteed to be idle
2841 * on return if it hasn't been requeued since flush started.
2844 * %true if flush_work() waited for the work to finish execution,
2845 * %false if it was already idle.
2847 bool flush_work(struct work_struct *work)
2849 struct wq_barrier barr;
2851 lock_map_acquire(&work->lockdep_map);
2852 lock_map_release(&work->lockdep_map);
2854 if (start_flush_work(work, &barr)) {
2855 wait_for_completion(&barr.done);
2856 destroy_work_on_stack(&barr.work);
2862 EXPORT_SYMBOL_GPL(flush_work);
2866 struct work_struct *work;
2869 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2871 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2873 if (cwait->work != key)
2875 return autoremove_wake_function(wait, mode, sync, key);
2878 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2880 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2881 unsigned long flags;
2885 ret = try_to_grab_pending(work, is_dwork, &flags);
2887 * If someone else is already canceling, wait for it to
2888 * finish. flush_work() doesn't work for PREEMPT_NONE
2889 * because we may get scheduled between @work's completion
2890 * and the other canceling task resuming and clearing
2891 * CANCELING - flush_work() will return false immediately
2892 * as @work is no longer busy, try_to_grab_pending() will
2893 * return -ENOENT as @work is still being canceled and the
2894 * other canceling task won't be able to clear CANCELING as
2895 * we're hogging the CPU.
2897 * Let's wait for completion using a waitqueue. As this
2898 * may lead to the thundering herd problem, use a custom
2899 * wake function which matches @work along with exclusive
2902 if (unlikely(ret == -ENOENT)) {
2903 struct cwt_wait cwait;
2905 init_wait(&cwait.wait);
2906 cwait.wait.func = cwt_wakefn;
2909 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2910 TASK_UNINTERRUPTIBLE);
2911 if (work_is_canceling(work))
2913 finish_wait(&cancel_waitq, &cwait.wait);
2915 } while (unlikely(ret < 0));
2917 /* tell other tasks trying to grab @work to back off */
2918 mark_work_canceling(work);
2919 local_irq_restore(flags);
2922 clear_work_data(work);
2925 * Paired with prepare_to_wait() above so that either
2926 * waitqueue_active() is visible here or !work_is_canceling() is
2930 if (waitqueue_active(&cancel_waitq))
2931 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2937 * cancel_work_sync - cancel a work and wait for it to finish
2938 * @work: the work to cancel
2940 * Cancel @work and wait for its execution to finish. This function
2941 * can be used even if the work re-queues itself or migrates to
2942 * another workqueue. On return from this function, @work is
2943 * guaranteed to be not pending or executing on any CPU.
2945 * cancel_work_sync(&delayed_work->work) must not be used for
2946 * delayed_work's. Use cancel_delayed_work_sync() instead.
2948 * The caller must ensure that the workqueue on which @work was last
2949 * queued can't be destroyed before this function returns.
2952 * %true if @work was pending, %false otherwise.
2954 bool cancel_work_sync(struct work_struct *work)
2956 return __cancel_work_timer(work, false);
2958 EXPORT_SYMBOL_GPL(cancel_work_sync);
2961 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2962 * @dwork: the delayed work to flush
2964 * Delayed timer is cancelled and the pending work is queued for
2965 * immediate execution. Like flush_work(), this function only
2966 * considers the last queueing instance of @dwork.
2969 * %true if flush_work() waited for the work to finish execution,
2970 * %false if it was already idle.
2972 bool flush_delayed_work(struct delayed_work *dwork)
2974 local_irq_disable();
2975 if (del_timer_sync(&dwork->timer))
2976 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2978 return flush_work(&dwork->work);
2980 EXPORT_SYMBOL(flush_delayed_work);
2982 static bool __cancel_work(struct work_struct *work, bool is_dwork)
2984 unsigned long flags;
2988 ret = try_to_grab_pending(work, is_dwork, &flags);
2989 } while (unlikely(ret == -EAGAIN));
2991 if (unlikely(ret < 0))
2994 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
2995 local_irq_restore(flags);
3000 * See cancel_delayed_work()
3002 bool cancel_work(struct work_struct *work)
3004 return __cancel_work(work, false);
3008 * cancel_delayed_work - cancel a delayed work
3009 * @dwork: delayed_work to cancel
3011 * Kill off a pending delayed_work.
3013 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3017 * The work callback function may still be running on return, unless
3018 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3019 * use cancel_delayed_work_sync() to wait on it.
3021 * This function is safe to call from any context including IRQ handler.
3023 bool cancel_delayed_work(struct delayed_work *dwork)
3025 return __cancel_work(&dwork->work, true);
3027 EXPORT_SYMBOL(cancel_delayed_work);
3030 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3031 * @dwork: the delayed work cancel
3033 * This is cancel_work_sync() for delayed works.
3036 * %true if @dwork was pending, %false otherwise.
3038 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3040 return __cancel_work_timer(&dwork->work, true);
3042 EXPORT_SYMBOL(cancel_delayed_work_sync);
3045 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3046 * @func: the function to call
3048 * schedule_on_each_cpu() executes @func on each online CPU using the
3049 * system workqueue and blocks until all CPUs have completed.
3050 * schedule_on_each_cpu() is very slow.
3053 * 0 on success, -errno on failure.
3055 int schedule_on_each_cpu(work_func_t func)
3058 struct work_struct __percpu *works;
3060 works = alloc_percpu(struct work_struct);
3066 for_each_online_cpu(cpu) {
3067 struct work_struct *work = per_cpu_ptr(works, cpu);
3069 INIT_WORK(work, func);
3070 schedule_work_on(cpu, work);
3073 for_each_online_cpu(cpu)
3074 flush_work(per_cpu_ptr(works, cpu));
3082 * execute_in_process_context - reliably execute the routine with user context
3083 * @fn: the function to execute
3084 * @ew: guaranteed storage for the execute work structure (must
3085 * be available when the work executes)
3087 * Executes the function immediately if process context is available,
3088 * otherwise schedules the function for delayed execution.
3090 * Return: 0 - function was executed
3091 * 1 - function was scheduled for execution
3093 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3095 if (!in_interrupt()) {
3100 INIT_WORK(&ew->work, fn);
3101 schedule_work(&ew->work);
3105 EXPORT_SYMBOL_GPL(execute_in_process_context);
3108 * free_workqueue_attrs - free a workqueue_attrs
3109 * @attrs: workqueue_attrs to free
3111 * Undo alloc_workqueue_attrs().
3113 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3116 free_cpumask_var(attrs->cpumask);
3122 * alloc_workqueue_attrs - allocate a workqueue_attrs
3123 * @gfp_mask: allocation mask to use
3125 * Allocate a new workqueue_attrs, initialize with default settings and
3128 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3130 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3132 struct workqueue_attrs *attrs;
3134 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3137 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3140 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3143 free_workqueue_attrs(attrs);
3147 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3148 const struct workqueue_attrs *from)
3150 to->nice = from->nice;
3151 cpumask_copy(to->cpumask, from->cpumask);
3153 * Unlike hash and equality test, this function doesn't ignore
3154 * ->no_numa as it is used for both pool and wq attrs. Instead,
3155 * get_unbound_pool() explicitly clears ->no_numa after copying.
3157 to->no_numa = from->no_numa;
3160 /* hash value of the content of @attr */
3161 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3165 hash = jhash_1word(attrs->nice, hash);
3166 hash = jhash(cpumask_bits(attrs->cpumask),
3167 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3171 /* content equality test */
3172 static bool wqattrs_equal(const struct workqueue_attrs *a,
3173 const struct workqueue_attrs *b)
3175 if (a->nice != b->nice)
3177 if (!cpumask_equal(a->cpumask, b->cpumask))
3183 * init_worker_pool - initialize a newly zalloc'd worker_pool
3184 * @pool: worker_pool to initialize
3186 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3188 * Return: 0 on success, -errno on failure. Even on failure, all fields
3189 * inside @pool proper are initialized and put_unbound_pool() can be called
3190 * on @pool safely to release it.
3192 static int init_worker_pool(struct worker_pool *pool)
3194 spin_lock_init(&pool->lock);
3197 pool->node = NUMA_NO_NODE;
3198 pool->flags |= POOL_DISASSOCIATED;
3199 pool->watchdog_ts = jiffies;
3200 INIT_LIST_HEAD(&pool->worklist);
3201 INIT_LIST_HEAD(&pool->idle_list);
3202 hash_init(pool->busy_hash);
3204 init_timer_deferrable(&pool->idle_timer);
3205 pool->idle_timer.function = idle_worker_timeout;
3206 pool->idle_timer.data = (unsigned long)pool;
3208 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3209 (unsigned long)pool);
3211 mutex_init(&pool->attach_mutex);
3212 INIT_LIST_HEAD(&pool->workers);
3214 ida_init(&pool->worker_ida);
3215 INIT_HLIST_NODE(&pool->hash_node);
3218 /* shouldn't fail above this point */
3219 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3225 static void rcu_free_wq(struct rcu_head *rcu)
3227 struct workqueue_struct *wq =
3228 container_of(rcu, struct workqueue_struct, rcu);
3230 if (!(wq->flags & WQ_UNBOUND))
3231 free_percpu(wq->cpu_pwqs);
3233 free_workqueue_attrs(wq->unbound_attrs);
3239 static void rcu_free_pool(struct rcu_head *rcu)
3241 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3243 ida_destroy(&pool->worker_ida);
3244 free_workqueue_attrs(pool->attrs);
3249 * put_unbound_pool - put a worker_pool
3250 * @pool: worker_pool to put
3252 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3253 * safe manner. get_unbound_pool() calls this function on its failure path
3254 * and this function should be able to release pools which went through,
3255 * successfully or not, init_worker_pool().
3257 * Should be called with wq_pool_mutex held.
3259 static void put_unbound_pool(struct worker_pool *pool)
3261 DECLARE_COMPLETION_ONSTACK(detach_completion);
3262 struct worker *worker;
3264 lockdep_assert_held(&wq_pool_mutex);
3270 if (WARN_ON(!(pool->cpu < 0)) ||
3271 WARN_ON(!list_empty(&pool->worklist)))
3274 /* release id and unhash */
3276 idr_remove(&worker_pool_idr, pool->id);
3277 hash_del(&pool->hash_node);
3280 * Become the manager and destroy all workers. This prevents
3281 * @pool's workers from blocking on attach_mutex. We're the last
3282 * manager and @pool gets freed with the flag set.
3284 spin_lock_irq(&pool->lock);
3285 wait_event_lock_irq(wq_manager_wait,
3286 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3287 pool->flags |= POOL_MANAGER_ACTIVE;
3289 while ((worker = first_idle_worker(pool)))
3290 destroy_worker(worker);
3291 WARN_ON(pool->nr_workers || pool->nr_idle);
3292 spin_unlock_irq(&pool->lock);
3294 mutex_lock(&pool->attach_mutex);
3295 if (!list_empty(&pool->workers))
3296 pool->detach_completion = &detach_completion;
3297 mutex_unlock(&pool->attach_mutex);
3299 if (pool->detach_completion)
3300 wait_for_completion(pool->detach_completion);
3302 /* shut down the timers */
3303 del_timer_sync(&pool->idle_timer);
3304 del_timer_sync(&pool->mayday_timer);
3306 /* sched-RCU protected to allow dereferences from get_work_pool() */
3307 call_rcu_sched(&pool->rcu, rcu_free_pool);
3311 * get_unbound_pool - get a worker_pool with the specified attributes
3312 * @attrs: the attributes of the worker_pool to get
3314 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3315 * reference count and return it. If there already is a matching
3316 * worker_pool, it will be used; otherwise, this function attempts to
3319 * Should be called with wq_pool_mutex held.
3321 * Return: On success, a worker_pool with the same attributes as @attrs.
3322 * On failure, %NULL.
3324 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3326 u32 hash = wqattrs_hash(attrs);
3327 struct worker_pool *pool;
3329 int target_node = NUMA_NO_NODE;
3331 lockdep_assert_held(&wq_pool_mutex);
3333 /* do we already have a matching pool? */
3334 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3335 if (wqattrs_equal(pool->attrs, attrs)) {
3341 /* if cpumask is contained inside a NUMA node, we belong to that node */
3342 if (wq_numa_enabled) {
3343 for_each_node(node) {
3344 if (cpumask_subset(attrs->cpumask,
3345 wq_numa_possible_cpumask[node])) {
3352 /* nope, create a new one */
3353 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3354 if (!pool || init_worker_pool(pool) < 0)
3357 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3358 copy_workqueue_attrs(pool->attrs, attrs);
3359 pool->node = target_node;
3362 * no_numa isn't a worker_pool attribute, always clear it. See
3363 * 'struct workqueue_attrs' comments for detail.
3365 pool->attrs->no_numa = false;
3367 if (worker_pool_assign_id(pool) < 0)
3370 /* create and start the initial worker */
3371 if (!create_worker(pool))
3375 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3380 put_unbound_pool(pool);
3384 static void rcu_free_pwq(struct rcu_head *rcu)
3386 kmem_cache_free(pwq_cache,
3387 container_of(rcu, struct pool_workqueue, rcu));
3391 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3392 * and needs to be destroyed.
3394 static void pwq_unbound_release_workfn(struct work_struct *work)
3396 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3397 unbound_release_work);
3398 struct workqueue_struct *wq = pwq->wq;
3399 struct worker_pool *pool = pwq->pool;
3400 bool is_last = false;
3403 * when @pwq is not linked, it doesn't hold any reference to the
3404 * @wq, and @wq is invalid to access.
3406 if (!list_empty(&pwq->pwqs_node)) {
3407 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3410 mutex_lock(&wq->mutex);
3411 list_del_rcu(&pwq->pwqs_node);
3412 is_last = list_empty(&wq->pwqs);
3413 mutex_unlock(&wq->mutex);
3416 mutex_lock(&wq_pool_mutex);
3417 put_unbound_pool(pool);
3418 mutex_unlock(&wq_pool_mutex);
3420 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3423 * If we're the last pwq going away, @wq is already dead and no one
3424 * is gonna access it anymore. Schedule RCU free.
3427 call_rcu_sched(&wq->rcu, rcu_free_wq);
3431 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3432 * @pwq: target pool_workqueue
3434 * If @pwq isn't freezing, set @pwq->max_active to the associated
3435 * workqueue's saved_max_active and activate delayed work items
3436 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3438 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3440 struct workqueue_struct *wq = pwq->wq;
3441 bool freezable = wq->flags & WQ_FREEZABLE;
3443 /* for @wq->saved_max_active */
3444 lockdep_assert_held(&wq->mutex);
3446 /* fast exit for non-freezable wqs */
3447 if (!freezable && pwq->max_active == wq->saved_max_active)
3450 spin_lock_irq(&pwq->pool->lock);
3453 * During [un]freezing, the caller is responsible for ensuring that
3454 * this function is called at least once after @workqueue_freezing
3455 * is updated and visible.
3457 if (!freezable || !workqueue_freezing) {
3460 pwq->max_active = wq->saved_max_active;
3462 while (!list_empty(&pwq->delayed_works) &&
3463 pwq->nr_active < pwq->max_active) {
3464 pwq_activate_first_delayed(pwq);
3469 * Need to kick a worker after thawed or an unbound wq's
3470 * max_active is bumped. In realtime scenarios, always kicking a
3471 * worker will cause interference on the isolated cpu cores, so
3472 * let's kick iff work items were activated.
3475 wake_up_worker(pwq->pool);
3477 pwq->max_active = 0;
3480 spin_unlock_irq(&pwq->pool->lock);
3483 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3484 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3485 struct worker_pool *pool)
3487 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3489 memset(pwq, 0, sizeof(*pwq));
3493 pwq->flush_color = -1;
3495 INIT_LIST_HEAD(&pwq->delayed_works);
3496 INIT_LIST_HEAD(&pwq->pwqs_node);
3497 INIT_LIST_HEAD(&pwq->mayday_node);
3498 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3501 /* sync @pwq with the current state of its associated wq and link it */
3502 static void link_pwq(struct pool_workqueue *pwq)
3504 struct workqueue_struct *wq = pwq->wq;
3506 lockdep_assert_held(&wq->mutex);
3508 /* may be called multiple times, ignore if already linked */
3509 if (!list_empty(&pwq->pwqs_node))
3512 /* set the matching work_color */
3513 pwq->work_color = wq->work_color;
3515 /* sync max_active to the current setting */
3516 pwq_adjust_max_active(pwq);
3519 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3522 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3523 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3524 const struct workqueue_attrs *attrs)
3526 struct worker_pool *pool;
3527 struct pool_workqueue *pwq;
3529 lockdep_assert_held(&wq_pool_mutex);
3531 pool = get_unbound_pool(attrs);
3535 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3537 put_unbound_pool(pool);
3541 init_pwq(pwq, wq, pool);
3546 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3547 * @attrs: the wq_attrs of the default pwq of the target workqueue
3548 * @node: the target NUMA node
3549 * @cpu_going_down: if >= 0, the CPU to consider as offline
3550 * @cpumask: outarg, the resulting cpumask
3552 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3553 * @cpu_going_down is >= 0, that cpu is considered offline during
3554 * calculation. The result is stored in @cpumask.
3556 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3557 * enabled and @node has online CPUs requested by @attrs, the returned
3558 * cpumask is the intersection of the possible CPUs of @node and
3561 * The caller is responsible for ensuring that the cpumask of @node stays
3564 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3567 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3568 int cpu_going_down, cpumask_t *cpumask)
3570 if (!wq_numa_enabled || attrs->no_numa)
3573 /* does @node have any online CPUs @attrs wants? */
3574 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3575 if (cpu_going_down >= 0)
3576 cpumask_clear_cpu(cpu_going_down, cpumask);
3578 if (cpumask_empty(cpumask))
3581 /* yeap, return possible CPUs in @node that @attrs wants */
3582 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3583 return !cpumask_equal(cpumask, attrs->cpumask);
3586 cpumask_copy(cpumask, attrs->cpumask);
3590 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3591 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3593 struct pool_workqueue *pwq)
3595 struct pool_workqueue *old_pwq;
3597 lockdep_assert_held(&wq_pool_mutex);
3598 lockdep_assert_held(&wq->mutex);
3600 /* link_pwq() can handle duplicate calls */
3603 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3604 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3608 /* context to store the prepared attrs & pwqs before applying */
3609 struct apply_wqattrs_ctx {
3610 struct workqueue_struct *wq; /* target workqueue */
3611 struct workqueue_attrs *attrs; /* attrs to apply */
3612 struct list_head list; /* queued for batching commit */
3613 struct pool_workqueue *dfl_pwq;
3614 struct pool_workqueue *pwq_tbl[];
3617 /* free the resources after success or abort */
3618 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3624 put_pwq_unlocked(ctx->pwq_tbl[node]);
3625 put_pwq_unlocked(ctx->dfl_pwq);
3627 free_workqueue_attrs(ctx->attrs);
3633 /* allocate the attrs and pwqs for later installation */
3634 static struct apply_wqattrs_ctx *
3635 apply_wqattrs_prepare(struct workqueue_struct *wq,
3636 const struct workqueue_attrs *attrs)
3638 struct apply_wqattrs_ctx *ctx;
3639 struct workqueue_attrs *new_attrs, *tmp_attrs;
3642 lockdep_assert_held(&wq_pool_mutex);
3644 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3647 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3648 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3649 if (!ctx || !new_attrs || !tmp_attrs)
3653 * Calculate the attrs of the default pwq.
3654 * If the user configured cpumask doesn't overlap with the
3655 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3657 copy_workqueue_attrs(new_attrs, attrs);
3658 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3659 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3660 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3663 * We may create multiple pwqs with differing cpumasks. Make a
3664 * copy of @new_attrs which will be modified and used to obtain
3667 copy_workqueue_attrs(tmp_attrs, new_attrs);
3670 * If something goes wrong during CPU up/down, we'll fall back to
3671 * the default pwq covering whole @attrs->cpumask. Always create
3672 * it even if we don't use it immediately.
3674 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3678 for_each_node(node) {
3679 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3680 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3681 if (!ctx->pwq_tbl[node])
3684 ctx->dfl_pwq->refcnt++;
3685 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3689 /* save the user configured attrs and sanitize it. */
3690 copy_workqueue_attrs(new_attrs, attrs);
3691 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3692 ctx->attrs = new_attrs;
3695 free_workqueue_attrs(tmp_attrs);
3699 free_workqueue_attrs(tmp_attrs);
3700 free_workqueue_attrs(new_attrs);
3701 apply_wqattrs_cleanup(ctx);
3705 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3706 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3710 /* all pwqs have been created successfully, let's install'em */
3711 mutex_lock(&ctx->wq->mutex);
3713 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3715 /* save the previous pwq and install the new one */
3717 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3718 ctx->pwq_tbl[node]);
3720 /* @dfl_pwq might not have been used, ensure it's linked */
3721 link_pwq(ctx->dfl_pwq);
3722 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3724 mutex_unlock(&ctx->wq->mutex);
3727 static void apply_wqattrs_lock(void)
3729 /* CPUs should stay stable across pwq creations and installations */
3731 mutex_lock(&wq_pool_mutex);
3734 static void apply_wqattrs_unlock(void)
3736 mutex_unlock(&wq_pool_mutex);
3740 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3741 const struct workqueue_attrs *attrs)
3743 struct apply_wqattrs_ctx *ctx;
3745 /* only unbound workqueues can change attributes */
3746 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3749 /* creating multiple pwqs breaks ordering guarantee */
3750 if (!list_empty(&wq->pwqs)) {
3751 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3754 wq->flags &= ~__WQ_ORDERED;
3757 ctx = apply_wqattrs_prepare(wq, attrs);
3761 /* the ctx has been prepared successfully, let's commit it */
3762 apply_wqattrs_commit(ctx);
3763 apply_wqattrs_cleanup(ctx);
3769 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3770 * @wq: the target workqueue
3771 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3773 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3774 * machines, this function maps a separate pwq to each NUMA node with
3775 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3776 * NUMA node it was issued on. Older pwqs are released as in-flight work
3777 * items finish. Note that a work item which repeatedly requeues itself
3778 * back-to-back will stay on its current pwq.
3780 * Performs GFP_KERNEL allocations.
3782 * Return: 0 on success and -errno on failure.
3784 int apply_workqueue_attrs(struct workqueue_struct *wq,
3785 const struct workqueue_attrs *attrs)
3789 apply_wqattrs_lock();
3790 ret = apply_workqueue_attrs_locked(wq, attrs);
3791 apply_wqattrs_unlock();
3797 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3798 * @wq: the target workqueue
3799 * @cpu: the CPU coming up or going down
3800 * @online: whether @cpu is coming up or going down
3802 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3803 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3806 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3807 * falls back to @wq->dfl_pwq which may not be optimal but is always
3810 * Note that when the last allowed CPU of a NUMA node goes offline for a
3811 * workqueue with a cpumask spanning multiple nodes, the workers which were
3812 * already executing the work items for the workqueue will lose their CPU
3813 * affinity and may execute on any CPU. This is similar to how per-cpu
3814 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3815 * affinity, it's the user's responsibility to flush the work item from
3818 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3821 int node = cpu_to_node(cpu);
3822 int cpu_off = online ? -1 : cpu;
3823 struct pool_workqueue *old_pwq = NULL, *pwq;
3824 struct workqueue_attrs *target_attrs;
3827 lockdep_assert_held(&wq_pool_mutex);
3829 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3830 wq->unbound_attrs->no_numa)
3834 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3835 * Let's use a preallocated one. The following buf is protected by
3836 * CPU hotplug exclusion.
3838 target_attrs = wq_update_unbound_numa_attrs_buf;
3839 cpumask = target_attrs->cpumask;
3841 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3842 pwq = unbound_pwq_by_node(wq, node);
3845 * Let's determine what needs to be done. If the target cpumask is
3846 * different from the default pwq's, we need to compare it to @pwq's
3847 * and create a new one if they don't match. If the target cpumask
3848 * equals the default pwq's, the default pwq should be used.
3850 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3851 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3857 /* create a new pwq */
3858 pwq = alloc_unbound_pwq(wq, target_attrs);
3860 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3865 /* Install the new pwq. */
3866 mutex_lock(&wq->mutex);
3867 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3871 mutex_lock(&wq->mutex);
3872 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3873 get_pwq(wq->dfl_pwq);
3874 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3875 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3877 mutex_unlock(&wq->mutex);
3878 put_pwq_unlocked(old_pwq);
3881 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3883 bool highpri = wq->flags & WQ_HIGHPRI;
3886 if (!(wq->flags & WQ_UNBOUND)) {
3887 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3891 for_each_possible_cpu(cpu) {
3892 struct pool_workqueue *pwq =
3893 per_cpu_ptr(wq->cpu_pwqs, cpu);
3894 struct worker_pool *cpu_pools =
3895 per_cpu(cpu_worker_pools, cpu);
3897 init_pwq(pwq, wq, &cpu_pools[highpri]);
3899 mutex_lock(&wq->mutex);
3901 mutex_unlock(&wq->mutex);
3904 } else if (wq->flags & __WQ_ORDERED) {
3905 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3906 /* there should only be single pwq for ordering guarantee */
3907 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3908 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3909 "ordering guarantee broken for workqueue %s\n", wq->name);
3912 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3916 static int wq_clamp_max_active(int max_active, unsigned int flags,
3919 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3921 if (max_active < 1 || max_active > lim)
3922 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3923 max_active, name, 1, lim);
3925 return clamp_val(max_active, 1, lim);
3928 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3931 struct lock_class_key *key,
3932 const char *lock_name, ...)
3934 size_t tbl_size = 0;
3936 struct workqueue_struct *wq;
3937 struct pool_workqueue *pwq;
3940 * Unbound && max_active == 1 used to imply ordered, which is no
3941 * longer the case on NUMA machines due to per-node pools. While
3942 * alloc_ordered_workqueue() is the right way to create an ordered
3943 * workqueue, keep the previous behavior to avoid subtle breakages
3946 if ((flags & WQ_UNBOUND) && max_active == 1)
3947 flags |= __WQ_ORDERED;
3949 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3950 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3951 flags |= WQ_UNBOUND;
3953 /* allocate wq and format name */
3954 if (flags & WQ_UNBOUND)
3955 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3957 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3961 if (flags & WQ_UNBOUND) {
3962 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3963 if (!wq->unbound_attrs)
3967 va_start(args, lock_name);
3968 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3971 max_active = max_active ?: WQ_DFL_ACTIVE;
3972 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3976 wq->saved_max_active = max_active;
3977 mutex_init(&wq->mutex);
3978 atomic_set(&wq->nr_pwqs_to_flush, 0);
3979 INIT_LIST_HEAD(&wq->pwqs);
3980 INIT_LIST_HEAD(&wq->flusher_queue);
3981 INIT_LIST_HEAD(&wq->flusher_overflow);
3982 INIT_LIST_HEAD(&wq->maydays);
3984 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3985 INIT_LIST_HEAD(&wq->list);
3987 if (alloc_and_link_pwqs(wq) < 0)
3991 * Workqueues which may be used during memory reclaim should
3992 * have a rescuer to guarantee forward progress.
3994 if (flags & WQ_MEM_RECLAIM) {
3995 struct worker *rescuer;
3997 rescuer = alloc_worker(NUMA_NO_NODE);
4001 rescuer->rescue_wq = wq;
4002 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4004 if (IS_ERR(rescuer->task)) {
4009 wq->rescuer = rescuer;
4010 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4011 wake_up_process(rescuer->task);
4014 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4018 * wq_pool_mutex protects global freeze state and workqueues list.
4019 * Grab it, adjust max_active and add the new @wq to workqueues
4022 mutex_lock(&wq_pool_mutex);
4024 mutex_lock(&wq->mutex);
4025 for_each_pwq(pwq, wq)
4026 pwq_adjust_max_active(pwq);
4027 mutex_unlock(&wq->mutex);
4029 list_add_tail_rcu(&wq->list, &workqueues);
4031 mutex_unlock(&wq_pool_mutex);
4036 free_workqueue_attrs(wq->unbound_attrs);
4040 destroy_workqueue(wq);
4043 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4046 * destroy_workqueue - safely terminate a workqueue
4047 * @wq: target workqueue
4049 * Safely destroy a workqueue. All work currently pending will be done first.
4051 void destroy_workqueue(struct workqueue_struct *wq)
4053 struct pool_workqueue *pwq;
4057 * Remove it from sysfs first so that sanity check failure doesn't
4058 * lead to sysfs name conflicts.
4060 workqueue_sysfs_unregister(wq);
4062 /* drain it before proceeding with destruction */
4063 drain_workqueue(wq);
4065 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4067 struct worker *rescuer = wq->rescuer;
4069 /* this prevents new queueing */
4070 spin_lock_irq(&wq_mayday_lock);
4072 spin_unlock_irq(&wq_mayday_lock);
4074 /* rescuer will empty maydays list before exiting */
4075 kthread_stop(rescuer->task);
4080 mutex_lock(&wq->mutex);
4081 for_each_pwq(pwq, wq) {
4084 for (i = 0; i < WORK_NR_COLORS; i++) {
4085 if (WARN_ON(pwq->nr_in_flight[i])) {
4086 mutex_unlock(&wq->mutex);
4091 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4092 WARN_ON(pwq->nr_active) ||
4093 WARN_ON(!list_empty(&pwq->delayed_works))) {
4094 mutex_unlock(&wq->mutex);
4098 mutex_unlock(&wq->mutex);
4101 * wq list is used to freeze wq, remove from list after
4102 * flushing is complete in case freeze races us.
4104 mutex_lock(&wq_pool_mutex);
4105 list_del_rcu(&wq->list);
4106 mutex_unlock(&wq_pool_mutex);
4108 if (!(wq->flags & WQ_UNBOUND)) {
4110 * The base ref is never dropped on per-cpu pwqs. Directly
4111 * schedule RCU free.
4113 call_rcu_sched(&wq->rcu, rcu_free_wq);
4116 * We're the sole accessor of @wq at this point. Directly
4117 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4118 * @wq will be freed when the last pwq is released.
4120 for_each_node(node) {
4121 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4122 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4123 put_pwq_unlocked(pwq);
4127 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4128 * put. Don't access it afterwards.
4132 put_pwq_unlocked(pwq);
4135 EXPORT_SYMBOL_GPL(destroy_workqueue);
4138 * workqueue_set_max_active - adjust max_active of a workqueue
4139 * @wq: target workqueue
4140 * @max_active: new max_active value.
4142 * Set max_active of @wq to @max_active.
4145 * Don't call from IRQ context.
4147 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4149 struct pool_workqueue *pwq;
4151 /* disallow meddling with max_active for ordered workqueues */
4152 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4155 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4157 mutex_lock(&wq->mutex);
4159 wq->flags &= ~__WQ_ORDERED;
4160 wq->saved_max_active = max_active;
4162 for_each_pwq(pwq, wq)
4163 pwq_adjust_max_active(pwq);
4165 mutex_unlock(&wq->mutex);
4167 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4170 * current_work - retrieve %current task's work struct
4172 * Determine if %current task is a workqueue worker and what it's working on.
4173 * Useful to find out the context that the %current task is running in.
4175 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4177 struct work_struct *current_work(void)
4179 struct worker *worker = current_wq_worker();
4181 return worker ? worker->current_work : NULL;
4183 EXPORT_SYMBOL(current_work);
4186 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4188 * Determine whether %current is a workqueue rescuer. Can be used from
4189 * work functions to determine whether it's being run off the rescuer task.
4191 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4193 bool current_is_workqueue_rescuer(void)
4195 struct worker *worker = current_wq_worker();
4197 return worker && worker->rescue_wq;
4201 * workqueue_congested - test whether a workqueue is congested
4202 * @cpu: CPU in question
4203 * @wq: target workqueue
4205 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4206 * no synchronization around this function and the test result is
4207 * unreliable and only useful as advisory hints or for debugging.
4209 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4210 * Note that both per-cpu and unbound workqueues may be associated with
4211 * multiple pool_workqueues which have separate congested states. A
4212 * workqueue being congested on one CPU doesn't mean the workqueue is also
4213 * contested on other CPUs / NUMA nodes.
4216 * %true if congested, %false otherwise.
4218 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4220 struct pool_workqueue *pwq;
4223 rcu_read_lock_sched();
4225 if (cpu == WORK_CPU_UNBOUND)
4226 cpu = smp_processor_id();
4228 if (!(wq->flags & WQ_UNBOUND))
4229 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4231 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4233 ret = !list_empty(&pwq->delayed_works);
4234 rcu_read_unlock_sched();
4238 EXPORT_SYMBOL_GPL(workqueue_congested);
4241 * work_busy - test whether a work is currently pending or running
4242 * @work: the work to be tested
4244 * Test whether @work is currently pending or running. There is no
4245 * synchronization around this function and the test result is
4246 * unreliable and only useful as advisory hints or for debugging.
4249 * OR'd bitmask of WORK_BUSY_* bits.
4251 unsigned int work_busy(struct work_struct *work)
4253 struct worker_pool *pool;
4254 unsigned long flags;
4255 unsigned int ret = 0;
4257 if (work_pending(work))
4258 ret |= WORK_BUSY_PENDING;
4260 local_irq_save(flags);
4261 pool = get_work_pool(work);
4263 spin_lock(&pool->lock);
4264 if (find_worker_executing_work(pool, work))
4265 ret |= WORK_BUSY_RUNNING;
4266 spin_unlock(&pool->lock);
4268 local_irq_restore(flags);
4272 EXPORT_SYMBOL_GPL(work_busy);
4275 * set_worker_desc - set description for the current work item
4276 * @fmt: printf-style format string
4277 * @...: arguments for the format string
4279 * This function can be called by a running work function to describe what
4280 * the work item is about. If the worker task gets dumped, this
4281 * information will be printed out together to help debugging. The
4282 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4284 void set_worker_desc(const char *fmt, ...)
4286 struct worker *worker = current_wq_worker();
4290 va_start(args, fmt);
4291 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4293 worker->desc_valid = true;
4298 * print_worker_info - print out worker information and description
4299 * @log_lvl: the log level to use when printing
4300 * @task: target task
4302 * If @task is a worker and currently executing a work item, print out the
4303 * name of the workqueue being serviced and worker description set with
4304 * set_worker_desc() by the currently executing work item.
4306 * This function can be safely called on any task as long as the
4307 * task_struct itself is accessible. While safe, this function isn't
4308 * synchronized and may print out mixups or garbages of limited length.
4310 void print_worker_info(const char *log_lvl, struct task_struct *task)
4312 work_func_t *fn = NULL;
4313 char name[WQ_NAME_LEN] = { };
4314 char desc[WORKER_DESC_LEN] = { };
4315 struct pool_workqueue *pwq = NULL;
4316 struct workqueue_struct *wq = NULL;
4317 bool desc_valid = false;
4318 struct worker *worker;
4320 if (!(task->flags & PF_WQ_WORKER))
4324 * This function is called without any synchronization and @task
4325 * could be in any state. Be careful with dereferences.
4327 worker = kthread_probe_data(task);
4330 * Carefully copy the associated workqueue's workfn and name. Keep
4331 * the original last '\0' in case the original contains garbage.
4333 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4334 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4335 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4336 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4338 /* copy worker description */
4339 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4341 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4343 if (fn || name[0] || desc[0]) {
4344 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4346 pr_cont(" (%s)", desc);
4351 static void pr_cont_pool_info(struct worker_pool *pool)
4353 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4354 if (pool->node != NUMA_NO_NODE)
4355 pr_cont(" node=%d", pool->node);
4356 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4359 static void pr_cont_work(bool comma, struct work_struct *work)
4361 if (work->func == wq_barrier_func) {
4362 struct wq_barrier *barr;
4364 barr = container_of(work, struct wq_barrier, work);
4366 pr_cont("%s BAR(%d)", comma ? "," : "",
4367 task_pid_nr(barr->task));
4369 pr_cont("%s %pf", comma ? "," : "", work->func);
4373 static void show_pwq(struct pool_workqueue *pwq)
4375 struct worker_pool *pool = pwq->pool;
4376 struct work_struct *work;
4377 struct worker *worker;
4378 bool has_in_flight = false, has_pending = false;
4381 pr_info(" pwq %d:", pool->id);
4382 pr_cont_pool_info(pool);
4384 pr_cont(" active=%d/%d refcnt=%d%s\n",
4385 pwq->nr_active, pwq->max_active, pwq->refcnt,
4386 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4388 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4389 if (worker->current_pwq == pwq) {
4390 has_in_flight = true;
4394 if (has_in_flight) {
4397 pr_info(" in-flight:");
4398 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4399 if (worker->current_pwq != pwq)
4402 pr_cont("%s %d%s:%pf", comma ? "," : "",
4403 task_pid_nr(worker->task),
4404 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4405 worker->current_func);
4406 list_for_each_entry(work, &worker->scheduled, entry)
4407 pr_cont_work(false, work);
4413 list_for_each_entry(work, &pool->worklist, entry) {
4414 if (get_work_pwq(work) == pwq) {
4422 pr_info(" pending:");
4423 list_for_each_entry(work, &pool->worklist, entry) {
4424 if (get_work_pwq(work) != pwq)
4427 pr_cont_work(comma, work);
4428 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4433 if (!list_empty(&pwq->delayed_works)) {
4436 pr_info(" delayed:");
4437 list_for_each_entry(work, &pwq->delayed_works, entry) {
4438 pr_cont_work(comma, work);
4439 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4446 * show_workqueue_state - dump workqueue state
4448 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4449 * all busy workqueues and pools.
4451 void show_workqueue_state(void)
4453 struct workqueue_struct *wq;
4454 struct worker_pool *pool;
4455 unsigned long flags;
4458 rcu_read_lock_sched();
4460 pr_info("Showing busy workqueues and worker pools:\n");
4462 list_for_each_entry_rcu(wq, &workqueues, list) {
4463 struct pool_workqueue *pwq;
4466 for_each_pwq(pwq, wq) {
4467 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4475 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4477 for_each_pwq(pwq, wq) {
4478 spin_lock_irqsave(&pwq->pool->lock, flags);
4479 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4481 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4483 * We could be printing a lot from atomic context, e.g.
4484 * sysrq-t -> show_workqueue_state(). Avoid triggering
4487 touch_nmi_watchdog();
4491 for_each_pool(pool, pi) {
4492 struct worker *worker;
4495 spin_lock_irqsave(&pool->lock, flags);
4496 if (pool->nr_workers == pool->nr_idle)
4499 pr_info("pool %d:", pool->id);
4500 pr_cont_pool_info(pool);
4501 pr_cont(" hung=%us workers=%d",
4502 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4505 pr_cont(" manager: %d",
4506 task_pid_nr(pool->manager->task));
4507 list_for_each_entry(worker, &pool->idle_list, entry) {
4508 pr_cont(" %s%d", first ? "idle: " : "",
4509 task_pid_nr(worker->task));
4514 spin_unlock_irqrestore(&pool->lock, flags);
4516 * We could be printing a lot from atomic context, e.g.
4517 * sysrq-t -> show_workqueue_state(). Avoid triggering
4520 touch_nmi_watchdog();
4523 rcu_read_unlock_sched();
4529 * There are two challenges in supporting CPU hotplug. Firstly, there
4530 * are a lot of assumptions on strong associations among work, pwq and
4531 * pool which make migrating pending and scheduled works very
4532 * difficult to implement without impacting hot paths. Secondly,
4533 * worker pools serve mix of short, long and very long running works making
4534 * blocked draining impractical.
4536 * This is solved by allowing the pools to be disassociated from the CPU
4537 * running as an unbound one and allowing it to be reattached later if the
4538 * cpu comes back online.
4541 static void wq_unbind_fn(struct work_struct *work)
4543 int cpu = smp_processor_id();
4544 struct worker_pool *pool;
4545 struct worker *worker;
4547 for_each_cpu_worker_pool(pool, cpu) {
4548 mutex_lock(&pool->attach_mutex);
4549 spin_lock_irq(&pool->lock);
4552 * We've blocked all attach/detach operations. Make all workers
4553 * unbound and set DISASSOCIATED. Before this, all workers
4554 * except for the ones which are still executing works from
4555 * before the last CPU down must be on the cpu. After
4556 * this, they may become diasporas.
4558 for_each_pool_worker(worker, pool)
4559 worker->flags |= WORKER_UNBOUND;
4561 pool->flags |= POOL_DISASSOCIATED;
4563 spin_unlock_irq(&pool->lock);
4564 mutex_unlock(&pool->attach_mutex);
4567 * Call schedule() so that we cross rq->lock and thus can
4568 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4569 * This is necessary as scheduler callbacks may be invoked
4575 * Sched callbacks are disabled now. Zap nr_running.
4576 * After this, nr_running stays zero and need_more_worker()
4577 * and keep_working() are always true as long as the
4578 * worklist is not empty. This pool now behaves as an
4579 * unbound (in terms of concurrency management) pool which
4580 * are served by workers tied to the pool.
4582 atomic_set(&pool->nr_running, 0);
4585 * With concurrency management just turned off, a busy
4586 * worker blocking could lead to lengthy stalls. Kick off
4587 * unbound chain execution of currently pending work items.
4589 spin_lock_irq(&pool->lock);
4590 wake_up_worker(pool);
4591 spin_unlock_irq(&pool->lock);
4596 * rebind_workers - rebind all workers of a pool to the associated CPU
4597 * @pool: pool of interest
4599 * @pool->cpu is coming online. Rebind all workers to the CPU.
4601 static void rebind_workers(struct worker_pool *pool)
4603 struct worker *worker;
4605 lockdep_assert_held(&pool->attach_mutex);
4608 * Restore CPU affinity of all workers. As all idle workers should
4609 * be on the run-queue of the associated CPU before any local
4610 * wake-ups for concurrency management happen, restore CPU affinity
4611 * of all workers first and then clear UNBOUND. As we're called
4612 * from CPU_ONLINE, the following shouldn't fail.
4614 for_each_pool_worker(worker, pool)
4615 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4616 pool->attrs->cpumask) < 0);
4618 spin_lock_irq(&pool->lock);
4621 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4622 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4623 * being reworked and this can go away in time.
4625 if (!(pool->flags & POOL_DISASSOCIATED)) {
4626 spin_unlock_irq(&pool->lock);
4630 pool->flags &= ~POOL_DISASSOCIATED;
4632 for_each_pool_worker(worker, pool) {
4633 unsigned int worker_flags = worker->flags;
4636 * A bound idle worker should actually be on the runqueue
4637 * of the associated CPU for local wake-ups targeting it to
4638 * work. Kick all idle workers so that they migrate to the
4639 * associated CPU. Doing this in the same loop as
4640 * replacing UNBOUND with REBOUND is safe as no worker will
4641 * be bound before @pool->lock is released.
4643 if (worker_flags & WORKER_IDLE)
4644 wake_up_process(worker->task);
4647 * We want to clear UNBOUND but can't directly call
4648 * worker_clr_flags() or adjust nr_running. Atomically
4649 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4650 * @worker will clear REBOUND using worker_clr_flags() when
4651 * it initiates the next execution cycle thus restoring
4652 * concurrency management. Note that when or whether
4653 * @worker clears REBOUND doesn't affect correctness.
4655 * ACCESS_ONCE() is necessary because @worker->flags may be
4656 * tested without holding any lock in
4657 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4658 * fail incorrectly leading to premature concurrency
4659 * management operations.
4661 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4662 worker_flags |= WORKER_REBOUND;
4663 worker_flags &= ~WORKER_UNBOUND;
4664 ACCESS_ONCE(worker->flags) = worker_flags;
4667 spin_unlock_irq(&pool->lock);
4671 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4672 * @pool: unbound pool of interest
4673 * @cpu: the CPU which is coming up
4675 * An unbound pool may end up with a cpumask which doesn't have any online
4676 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4677 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4678 * online CPU before, cpus_allowed of all its workers should be restored.
4680 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4682 static cpumask_t cpumask;
4683 struct worker *worker;
4685 lockdep_assert_held(&pool->attach_mutex);
4687 /* is @cpu allowed for @pool? */
4688 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4691 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4693 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4694 for_each_pool_worker(worker, pool)
4695 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4698 int workqueue_prepare_cpu(unsigned int cpu)
4700 struct worker_pool *pool;
4702 for_each_cpu_worker_pool(pool, cpu) {
4703 if (pool->nr_workers)
4705 if (!create_worker(pool))
4711 int workqueue_online_cpu(unsigned int cpu)
4713 struct worker_pool *pool;
4714 struct workqueue_struct *wq;
4717 mutex_lock(&wq_pool_mutex);
4719 for_each_pool(pool, pi) {
4720 mutex_lock(&pool->attach_mutex);
4722 if (pool->cpu == cpu)
4723 rebind_workers(pool);
4724 else if (pool->cpu < 0)
4725 restore_unbound_workers_cpumask(pool, cpu);
4727 mutex_unlock(&pool->attach_mutex);
4730 /* update NUMA affinity of unbound workqueues */
4731 list_for_each_entry(wq, &workqueues, list)
4732 wq_update_unbound_numa(wq, cpu, true);
4734 mutex_unlock(&wq_pool_mutex);
4738 int workqueue_offline_cpu(unsigned int cpu)
4740 struct work_struct unbind_work;
4741 struct workqueue_struct *wq;
4743 /* unbinding per-cpu workers should happen on the local CPU */
4744 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4745 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4747 /* update NUMA affinity of unbound workqueues */
4748 mutex_lock(&wq_pool_mutex);
4749 list_for_each_entry(wq, &workqueues, list)
4750 wq_update_unbound_numa(wq, cpu, false);
4751 mutex_unlock(&wq_pool_mutex);
4753 /* wait for per-cpu unbinding to finish */
4754 flush_work(&unbind_work);
4755 destroy_work_on_stack(&unbind_work);
4761 struct work_for_cpu {
4762 struct work_struct work;
4768 static void work_for_cpu_fn(struct work_struct *work)
4770 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4772 wfc->ret = wfc->fn(wfc->arg);
4776 * work_on_cpu - run a function in thread context on a particular cpu
4777 * @cpu: the cpu to run on
4778 * @fn: the function to run
4779 * @arg: the function arg
4781 * It is up to the caller to ensure that the cpu doesn't go offline.
4782 * The caller must not hold any locks which would prevent @fn from completing.
4784 * Return: The value @fn returns.
4786 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4788 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4790 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4791 schedule_work_on(cpu, &wfc.work);
4792 flush_work(&wfc.work);
4793 destroy_work_on_stack(&wfc.work);
4796 EXPORT_SYMBOL_GPL(work_on_cpu);
4797 #endif /* CONFIG_SMP */
4799 #ifdef CONFIG_FREEZER
4802 * freeze_workqueues_begin - begin freezing workqueues
4804 * Start freezing workqueues. After this function returns, all freezable
4805 * workqueues will queue new works to their delayed_works list instead of
4809 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4811 void freeze_workqueues_begin(void)
4813 struct workqueue_struct *wq;
4814 struct pool_workqueue *pwq;
4816 mutex_lock(&wq_pool_mutex);
4818 WARN_ON_ONCE(workqueue_freezing);
4819 workqueue_freezing = true;
4821 list_for_each_entry(wq, &workqueues, list) {
4822 mutex_lock(&wq->mutex);
4823 for_each_pwq(pwq, wq)
4824 pwq_adjust_max_active(pwq);
4825 mutex_unlock(&wq->mutex);
4828 mutex_unlock(&wq_pool_mutex);
4832 * freeze_workqueues_busy - are freezable workqueues still busy?
4834 * Check whether freezing is complete. This function must be called
4835 * between freeze_workqueues_begin() and thaw_workqueues().
4838 * Grabs and releases wq_pool_mutex.
4841 * %true if some freezable workqueues are still busy. %false if freezing
4844 bool freeze_workqueues_busy(void)
4847 struct workqueue_struct *wq;
4848 struct pool_workqueue *pwq;
4850 mutex_lock(&wq_pool_mutex);
4852 WARN_ON_ONCE(!workqueue_freezing);
4854 list_for_each_entry(wq, &workqueues, list) {
4855 if (!(wq->flags & WQ_FREEZABLE))
4858 * nr_active is monotonically decreasing. It's safe
4859 * to peek without lock.
4861 rcu_read_lock_sched();
4862 for_each_pwq(pwq, wq) {
4863 WARN_ON_ONCE(pwq->nr_active < 0);
4864 if (pwq->nr_active) {
4866 rcu_read_unlock_sched();
4870 rcu_read_unlock_sched();
4873 mutex_unlock(&wq_pool_mutex);
4878 * thaw_workqueues - thaw workqueues
4880 * Thaw workqueues. Normal queueing is restored and all collected
4881 * frozen works are transferred to their respective pool worklists.
4884 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4886 void thaw_workqueues(void)
4888 struct workqueue_struct *wq;
4889 struct pool_workqueue *pwq;
4891 mutex_lock(&wq_pool_mutex);
4893 if (!workqueue_freezing)
4896 workqueue_freezing = false;
4898 /* restore max_active and repopulate worklist */
4899 list_for_each_entry(wq, &workqueues, list) {
4900 mutex_lock(&wq->mutex);
4901 for_each_pwq(pwq, wq)
4902 pwq_adjust_max_active(pwq);
4903 mutex_unlock(&wq->mutex);
4907 mutex_unlock(&wq_pool_mutex);
4909 #endif /* CONFIG_FREEZER */
4911 static int workqueue_apply_unbound_cpumask(void)
4915 struct workqueue_struct *wq;
4916 struct apply_wqattrs_ctx *ctx, *n;
4918 lockdep_assert_held(&wq_pool_mutex);
4920 list_for_each_entry(wq, &workqueues, list) {
4921 if (!(wq->flags & WQ_UNBOUND))
4923 /* creating multiple pwqs breaks ordering guarantee */
4924 if (wq->flags & __WQ_ORDERED)
4927 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4933 list_add_tail(&ctx->list, &ctxs);
4936 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4938 apply_wqattrs_commit(ctx);
4939 apply_wqattrs_cleanup(ctx);
4946 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4947 * @cpumask: the cpumask to set
4949 * The low-level workqueues cpumask is a global cpumask that limits
4950 * the affinity of all unbound workqueues. This function check the @cpumask
4951 * and apply it to all unbound workqueues and updates all pwqs of them.
4953 * Retun: 0 - Success
4954 * -EINVAL - Invalid @cpumask
4955 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4957 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4960 cpumask_var_t saved_cpumask;
4962 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4965 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4966 if (!cpumask_empty(cpumask)) {
4967 apply_wqattrs_lock();
4969 /* save the old wq_unbound_cpumask. */
4970 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4972 /* update wq_unbound_cpumask at first and apply it to wqs. */
4973 cpumask_copy(wq_unbound_cpumask, cpumask);
4974 ret = workqueue_apply_unbound_cpumask();
4976 /* restore the wq_unbound_cpumask when failed. */
4978 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4980 apply_wqattrs_unlock();
4983 free_cpumask_var(saved_cpumask);
4989 * Workqueues with WQ_SYSFS flag set is visible to userland via
4990 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4991 * following attributes.
4993 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4994 * max_active RW int : maximum number of in-flight work items
4996 * Unbound workqueues have the following extra attributes.
4998 * id RO int : the associated pool ID
4999 * nice RW int : nice value of the workers
5000 * cpumask RW mask : bitmask of allowed CPUs for the workers
5003 struct workqueue_struct *wq;
5007 static struct workqueue_struct *dev_to_wq(struct device *dev)
5009 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5014 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5017 struct workqueue_struct *wq = dev_to_wq(dev);
5019 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5021 static DEVICE_ATTR_RO(per_cpu);
5023 static ssize_t max_active_show(struct device *dev,
5024 struct device_attribute *attr, char *buf)
5026 struct workqueue_struct *wq = dev_to_wq(dev);
5028 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5031 static ssize_t max_active_store(struct device *dev,
5032 struct device_attribute *attr, const char *buf,
5035 struct workqueue_struct *wq = dev_to_wq(dev);
5038 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5041 workqueue_set_max_active(wq, val);
5044 static DEVICE_ATTR_RW(max_active);
5046 static struct attribute *wq_sysfs_attrs[] = {
5047 &dev_attr_per_cpu.attr,
5048 &dev_attr_max_active.attr,
5051 ATTRIBUTE_GROUPS(wq_sysfs);
5053 static ssize_t wq_pool_ids_show(struct device *dev,
5054 struct device_attribute *attr, char *buf)
5056 struct workqueue_struct *wq = dev_to_wq(dev);
5057 const char *delim = "";
5058 int node, written = 0;
5060 rcu_read_lock_sched();
5061 for_each_node(node) {
5062 written += scnprintf(buf + written, PAGE_SIZE - written,
5063 "%s%d:%d", delim, node,
5064 unbound_pwq_by_node(wq, node)->pool->id);
5067 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5068 rcu_read_unlock_sched();
5073 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5076 struct workqueue_struct *wq = dev_to_wq(dev);
5079 mutex_lock(&wq->mutex);
5080 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5081 mutex_unlock(&wq->mutex);
5086 /* prepare workqueue_attrs for sysfs store operations */
5087 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5089 struct workqueue_attrs *attrs;
5091 lockdep_assert_held(&wq_pool_mutex);
5093 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5097 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5101 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5102 const char *buf, size_t count)
5104 struct workqueue_struct *wq = dev_to_wq(dev);
5105 struct workqueue_attrs *attrs;
5108 apply_wqattrs_lock();
5110 attrs = wq_sysfs_prep_attrs(wq);
5114 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5115 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5116 ret = apply_workqueue_attrs_locked(wq, attrs);
5121 apply_wqattrs_unlock();
5122 free_workqueue_attrs(attrs);
5123 return ret ?: count;
5126 static ssize_t wq_cpumask_show(struct device *dev,
5127 struct device_attribute *attr, char *buf)
5129 struct workqueue_struct *wq = dev_to_wq(dev);
5132 mutex_lock(&wq->mutex);
5133 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5134 cpumask_pr_args(wq->unbound_attrs->cpumask));
5135 mutex_unlock(&wq->mutex);
5139 static ssize_t wq_cpumask_store(struct device *dev,
5140 struct device_attribute *attr,
5141 const char *buf, size_t count)
5143 struct workqueue_struct *wq = dev_to_wq(dev);
5144 struct workqueue_attrs *attrs;
5147 apply_wqattrs_lock();
5149 attrs = wq_sysfs_prep_attrs(wq);
5153 ret = cpumask_parse(buf, attrs->cpumask);
5155 ret = apply_workqueue_attrs_locked(wq, attrs);
5158 apply_wqattrs_unlock();
5159 free_workqueue_attrs(attrs);
5160 return ret ?: count;
5163 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5166 struct workqueue_struct *wq = dev_to_wq(dev);
5169 mutex_lock(&wq->mutex);
5170 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5171 !wq->unbound_attrs->no_numa);
5172 mutex_unlock(&wq->mutex);
5177 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5178 const char *buf, size_t count)
5180 struct workqueue_struct *wq = dev_to_wq(dev);
5181 struct workqueue_attrs *attrs;
5182 int v, ret = -ENOMEM;
5184 apply_wqattrs_lock();
5186 attrs = wq_sysfs_prep_attrs(wq);
5191 if (sscanf(buf, "%d", &v) == 1) {
5192 attrs->no_numa = !v;
5193 ret = apply_workqueue_attrs_locked(wq, attrs);
5197 apply_wqattrs_unlock();
5198 free_workqueue_attrs(attrs);
5199 return ret ?: count;
5202 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5203 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5204 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5205 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5206 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5210 static struct bus_type wq_subsys = {
5211 .name = "workqueue",
5212 .dev_groups = wq_sysfs_groups,
5215 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5216 struct device_attribute *attr, char *buf)
5220 mutex_lock(&wq_pool_mutex);
5221 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5222 cpumask_pr_args(wq_unbound_cpumask));
5223 mutex_unlock(&wq_pool_mutex);
5228 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5229 struct device_attribute *attr, const char *buf, size_t count)
5231 cpumask_var_t cpumask;
5234 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5237 ret = cpumask_parse(buf, cpumask);
5239 ret = workqueue_set_unbound_cpumask(cpumask);
5241 free_cpumask_var(cpumask);
5242 return ret ? ret : count;
5245 static struct device_attribute wq_sysfs_cpumask_attr =
5246 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5247 wq_unbound_cpumask_store);
5249 static int __init wq_sysfs_init(void)
5253 err = subsys_virtual_register(&wq_subsys, NULL);
5257 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5259 core_initcall(wq_sysfs_init);
5261 static void wq_device_release(struct device *dev)
5263 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5269 * workqueue_sysfs_register - make a workqueue visible in sysfs
5270 * @wq: the workqueue to register
5272 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5273 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5274 * which is the preferred method.
5276 * Workqueue user should use this function directly iff it wants to apply
5277 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5278 * apply_workqueue_attrs() may race against userland updating the
5281 * Return: 0 on success, -errno on failure.
5283 int workqueue_sysfs_register(struct workqueue_struct *wq)
5285 struct wq_device *wq_dev;
5289 * Adjusting max_active or creating new pwqs by applying
5290 * attributes breaks ordering guarantee. Disallow exposing ordered
5293 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5296 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5301 wq_dev->dev.bus = &wq_subsys;
5302 wq_dev->dev.release = wq_device_release;
5303 dev_set_name(&wq_dev->dev, "%s", wq->name);
5306 * unbound_attrs are created separately. Suppress uevent until
5307 * everything is ready.
5309 dev_set_uevent_suppress(&wq_dev->dev, true);
5311 ret = device_register(&wq_dev->dev);
5313 put_device(&wq_dev->dev);
5318 if (wq->flags & WQ_UNBOUND) {
5319 struct device_attribute *attr;
5321 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5322 ret = device_create_file(&wq_dev->dev, attr);
5324 device_unregister(&wq_dev->dev);
5331 dev_set_uevent_suppress(&wq_dev->dev, false);
5332 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5337 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5338 * @wq: the workqueue to unregister
5340 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5342 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5344 struct wq_device *wq_dev = wq->wq_dev;
5350 device_unregister(&wq_dev->dev);
5352 #else /* CONFIG_SYSFS */
5353 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5354 #endif /* CONFIG_SYSFS */
5357 * Workqueue watchdog.
5359 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5360 * flush dependency, a concurrency managed work item which stays RUNNING
5361 * indefinitely. Workqueue stalls can be very difficult to debug as the
5362 * usual warning mechanisms don't trigger and internal workqueue state is
5365 * Workqueue watchdog monitors all worker pools periodically and dumps
5366 * state if some pools failed to make forward progress for a while where
5367 * forward progress is defined as the first item on ->worklist changing.
5369 * This mechanism is controlled through the kernel parameter
5370 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5371 * corresponding sysfs parameter file.
5373 #ifdef CONFIG_WQ_WATCHDOG
5375 static void wq_watchdog_timer_fn(unsigned long data);
5377 static unsigned long wq_watchdog_thresh = 30;
5378 static struct timer_list wq_watchdog_timer =
5379 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5381 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5382 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5384 static void wq_watchdog_reset_touched(void)
5388 wq_watchdog_touched = jiffies;
5389 for_each_possible_cpu(cpu)
5390 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5393 static void wq_watchdog_timer_fn(unsigned long data)
5395 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5396 bool lockup_detected = false;
5397 unsigned long now = jiffies;
5398 struct worker_pool *pool;
5406 for_each_pool(pool, pi) {
5407 unsigned long pool_ts, touched, ts;
5409 if (list_empty(&pool->worklist))
5413 * If a virtual machine is stopped by the host it can look to
5414 * the watchdog like a stall.
5416 kvm_check_and_clear_guest_paused();
5418 /* get the latest of pool and touched timestamps */
5419 pool_ts = READ_ONCE(pool->watchdog_ts);
5420 touched = READ_ONCE(wq_watchdog_touched);
5422 if (time_after(pool_ts, touched))
5427 if (pool->cpu >= 0) {
5428 unsigned long cpu_touched =
5429 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5431 if (time_after(cpu_touched, ts))
5436 if (time_after(now, ts + thresh)) {
5437 lockup_detected = true;
5438 pr_emerg("BUG: workqueue lockup - pool");
5439 pr_cont_pool_info(pool);
5440 pr_cont(" stuck for %us!\n",
5441 jiffies_to_msecs(now - pool_ts) / 1000);
5447 if (lockup_detected)
5448 show_workqueue_state();
5450 wq_watchdog_reset_touched();
5451 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5454 void wq_watchdog_touch(int cpu)
5457 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5459 wq_watchdog_touched = jiffies;
5462 static void wq_watchdog_set_thresh(unsigned long thresh)
5464 wq_watchdog_thresh = 0;
5465 del_timer_sync(&wq_watchdog_timer);
5468 wq_watchdog_thresh = thresh;
5469 wq_watchdog_reset_touched();
5470 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5474 static int wq_watchdog_param_set_thresh(const char *val,
5475 const struct kernel_param *kp)
5477 unsigned long thresh;
5480 ret = kstrtoul(val, 0, &thresh);
5485 wq_watchdog_set_thresh(thresh);
5487 wq_watchdog_thresh = thresh;
5492 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5493 .set = wq_watchdog_param_set_thresh,
5494 .get = param_get_ulong,
5497 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5500 static void wq_watchdog_init(void)
5502 wq_watchdog_set_thresh(wq_watchdog_thresh);
5505 #else /* CONFIG_WQ_WATCHDOG */
5507 static inline void wq_watchdog_init(void) { }
5509 #endif /* CONFIG_WQ_WATCHDOG */
5511 static void __init wq_numa_init(void)
5516 if (num_possible_nodes() <= 1)
5519 if (wq_disable_numa) {
5520 pr_info("workqueue: NUMA affinity support disabled\n");
5524 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5525 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5528 * We want masks of possible CPUs of each node which isn't readily
5529 * available. Build one from cpu_to_node() which should have been
5530 * fully initialized by now.
5532 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5536 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5537 node_online(node) ? node : NUMA_NO_NODE));
5539 for_each_possible_cpu(cpu) {
5540 node = cpu_to_node(cpu);
5541 if (WARN_ON(node == NUMA_NO_NODE)) {
5542 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5543 /* happens iff arch is bonkers, let's just proceed */
5546 cpumask_set_cpu(cpu, tbl[node]);
5549 wq_numa_possible_cpumask = tbl;
5550 wq_numa_enabled = true;
5553 static int __init init_workqueues(void)
5555 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5558 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5560 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5561 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5563 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5567 /* initialize CPU pools */
5568 for_each_possible_cpu(cpu) {
5569 struct worker_pool *pool;
5572 for_each_cpu_worker_pool(pool, cpu) {
5573 BUG_ON(init_worker_pool(pool));
5575 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5576 pool->attrs->nice = std_nice[i++];
5577 pool->node = cpu_to_node(cpu);
5580 mutex_lock(&wq_pool_mutex);
5581 BUG_ON(worker_pool_assign_id(pool));
5582 mutex_unlock(&wq_pool_mutex);
5586 /* create the initial worker */
5587 for_each_online_cpu(cpu) {
5588 struct worker_pool *pool;
5590 for_each_cpu_worker_pool(pool, cpu) {
5591 pool->flags &= ~POOL_DISASSOCIATED;
5592 BUG_ON(!create_worker(pool));
5596 /* create default unbound and ordered wq attrs */
5597 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5598 struct workqueue_attrs *attrs;
5600 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5601 attrs->nice = std_nice[i];
5602 unbound_std_wq_attrs[i] = attrs;
5605 * An ordered wq should have only one pwq as ordering is
5606 * guaranteed by max_active which is enforced by pwqs.
5607 * Turn off NUMA so that dfl_pwq is used for all nodes.
5609 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5610 attrs->nice = std_nice[i];
5611 attrs->no_numa = true;
5612 ordered_wq_attrs[i] = attrs;
5615 system_wq = alloc_workqueue("events", 0, 0);
5616 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5617 system_long_wq = alloc_workqueue("events_long", 0, 0);
5618 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5619 WQ_UNBOUND_MAX_ACTIVE);
5620 system_freezable_wq = alloc_workqueue("events_freezable",
5622 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5623 WQ_POWER_EFFICIENT, 0);
5624 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5625 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5627 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5628 !system_unbound_wq || !system_freezable_wq ||
5629 !system_power_efficient_wq ||
5630 !system_freezable_power_efficient_wq);
5636 early_initcall(init_workqueues);