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/core-api/workqueue.rst 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>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
72 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
75 WORKER_DIE = 1 << 1, /* die die die */
76 WORKER_IDLE = 1 << 2, /* is idle */
77 WORKER_PREP = 1 << 3, /* preparing to run works */
78 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
79 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
80 WORKER_REBOUND = 1 << 8, /* worker was rebound */
82 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
83 WORKER_UNBOUND | WORKER_REBOUND,
85 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
87 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
88 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
90 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
91 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
93 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
94 /* call for help after 10ms
96 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
97 CREATE_COOLDOWN = HZ, /* time to breath after fail */
100 * Rescue workers are used only on emergencies and shared by
101 * all cpus. Give MIN_NICE.
103 RESCUER_NICE_LEVEL = MIN_NICE,
104 HIGHPRI_NICE_LEVEL = MIN_NICE,
110 * Structure fields follow one of the following exclusion rules.
112 * I: Modifiable by initialization/destruction paths and read-only for
115 * P: Preemption protected. Disabling preemption is enough and should
116 * only be modified and accessed from the local cpu.
118 * L: pool->lock protected. Access with pool->lock held.
120 * X: During normal operation, modification requires pool->lock and should
121 * be done only from local cpu. Either disabling preemption on local
122 * cpu or grabbing pool->lock is enough for read access. If
123 * POOL_DISASSOCIATED is set, it's identical to L.
125 * A: pool->attach_mutex protected.
127 * PL: wq_pool_mutex protected.
129 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
133 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
134 * sched-RCU for reads.
136 * WQ: wq->mutex protected.
138 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
140 * MD: wq_mayday_lock protected.
143 /* struct worker is defined in workqueue_internal.h */
146 spinlock_t lock; /* the pool lock */
147 int cpu; /* I: the associated cpu */
148 int node; /* I: the associated node ID */
149 int id; /* I: pool ID */
150 unsigned int flags; /* X: flags */
152 unsigned long watchdog_ts; /* L: watchdog timestamp */
154 struct list_head worklist; /* L: list of pending works */
155 int nr_workers; /* L: total number of workers */
157 /* nr_idle includes the ones off idle_list for rebinding */
158 int nr_idle; /* L: currently idle ones */
160 struct list_head idle_list; /* X: list of idle workers */
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for workers */
164 /* a workers is either on busy_hash or idle_list, or the manager */
165 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
166 /* L: hash of busy workers */
168 /* see manage_workers() for details on the two manager mutexes */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
174 struct ida worker_ida; /* worker IDs for task name */
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp;
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
225 * Structure used to wait for workqueue flush.
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
261 struct wq_device *wq_dev; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache *pwq_cache;
283 static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa;
287 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293 static bool wq_online; /* can kworkers be created yet? */
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 lockdep_assert_irqs_disabled();
1381 debug_work_activate(work);
1383 /* if draining, only works from the same workqueue are allowed */
1384 if (unlikely(wq->flags & __WQ_DRAINING) &&
1385 WARN_ON_ONCE(!is_chained_work(wq)))
1388 if (req_cpu == WORK_CPU_UNBOUND)
1389 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1391 /* pwq which will be used unless @work is executing elsewhere */
1392 if (!(wq->flags & WQ_UNBOUND))
1393 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1395 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1398 * If @work was previously on a different pool, it might still be
1399 * running there, in which case the work needs to be queued on that
1400 * pool to guarantee non-reentrancy.
1402 last_pool = get_work_pool(work);
1403 if (last_pool && last_pool != pwq->pool) {
1404 struct worker *worker;
1406 spin_lock(&last_pool->lock);
1408 worker = find_worker_executing_work(last_pool, work);
1410 if (worker && worker->current_pwq->wq == wq) {
1411 pwq = worker->current_pwq;
1413 /* meh... not running there, queue here */
1414 spin_unlock(&last_pool->lock);
1415 spin_lock(&pwq->pool->lock);
1418 spin_lock(&pwq->pool->lock);
1422 * pwq is determined and locked. For unbound pools, we could have
1423 * raced with pwq release and it could already be dead. If its
1424 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1425 * without another pwq replacing it in the numa_pwq_tbl or while
1426 * work items are executing on it, so the retrying is guaranteed to
1427 * make forward-progress.
1429 if (unlikely(!pwq->refcnt)) {
1430 if (wq->flags & WQ_UNBOUND) {
1431 spin_unlock(&pwq->pool->lock);
1436 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1440 /* pwq determined, queue */
1441 trace_workqueue_queue_work(req_cpu, pwq, work);
1443 if (WARN_ON(!list_empty(&work->entry))) {
1444 spin_unlock(&pwq->pool->lock);
1448 pwq->nr_in_flight[pwq->work_color]++;
1449 work_flags = work_color_to_flags(pwq->work_color);
1451 if (likely(pwq->nr_active < pwq->max_active)) {
1452 trace_workqueue_activate_work(work);
1454 worklist = &pwq->pool->worklist;
1455 if (list_empty(worklist))
1456 pwq->pool->watchdog_ts = jiffies;
1458 work_flags |= WORK_STRUCT_DELAYED;
1459 worklist = &pwq->delayed_works;
1462 insert_work(pwq, work, worklist, work_flags);
1464 spin_unlock(&pwq->pool->lock);
1468 * queue_work_on - queue work on specific cpu
1469 * @cpu: CPU number to execute work on
1470 * @wq: workqueue to use
1471 * @work: work to queue
1473 * We queue the work to a specific CPU, the caller must ensure it
1476 * Return: %false if @work was already on a queue, %true otherwise.
1478 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1479 struct work_struct *work)
1482 unsigned long flags;
1484 local_irq_save(flags);
1486 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1487 __queue_work(cpu, wq, work);
1491 local_irq_restore(flags);
1494 EXPORT_SYMBOL(queue_work_on);
1496 void delayed_work_timer_fn(unsigned long __data)
1498 struct delayed_work *dwork = (struct delayed_work *)__data;
1500 /* should have been called from irqsafe timer with irq already off */
1501 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1503 EXPORT_SYMBOL(delayed_work_timer_fn);
1505 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1506 struct delayed_work *dwork, unsigned long delay)
1508 struct timer_list *timer = &dwork->timer;
1509 struct work_struct *work = &dwork->work;
1512 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1513 timer->data != (unsigned long)dwork);
1514 WARN_ON_ONCE(timer_pending(timer));
1515 WARN_ON_ONCE(!list_empty(&work->entry));
1518 * If @delay is 0, queue @dwork->work immediately. This is for
1519 * both optimization and correctness. The earliest @timer can
1520 * expire is on the closest next tick and delayed_work users depend
1521 * on that there's no such delay when @delay is 0.
1524 __queue_work(cpu, wq, &dwork->work);
1530 timer->expires = jiffies + delay;
1532 if (unlikely(cpu != WORK_CPU_UNBOUND))
1533 add_timer_on(timer, cpu);
1539 * queue_delayed_work_on - queue work on specific CPU after delay
1540 * @cpu: CPU number to execute work on
1541 * @wq: workqueue to use
1542 * @dwork: work to queue
1543 * @delay: number of jiffies to wait before queueing
1545 * Return: %false if @work was already on a queue, %true otherwise. If
1546 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1549 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1550 struct delayed_work *dwork, unsigned long delay)
1552 struct work_struct *work = &dwork->work;
1554 unsigned long flags;
1556 /* read the comment in __queue_work() */
1557 local_irq_save(flags);
1559 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1560 __queue_delayed_work(cpu, wq, dwork, delay);
1564 local_irq_restore(flags);
1567 EXPORT_SYMBOL(queue_delayed_work_on);
1570 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1571 * @cpu: CPU number to execute work on
1572 * @wq: workqueue to use
1573 * @dwork: work to queue
1574 * @delay: number of jiffies to wait before queueing
1576 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1577 * modify @dwork's timer so that it expires after @delay. If @delay is
1578 * zero, @work is guaranteed to be scheduled immediately regardless of its
1581 * Return: %false if @dwork was idle and queued, %true if @dwork was
1582 * pending and its timer was modified.
1584 * This function is safe to call from any context including IRQ handler.
1585 * See try_to_grab_pending() for details.
1587 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1588 struct delayed_work *dwork, unsigned long delay)
1590 unsigned long flags;
1594 ret = try_to_grab_pending(&dwork->work, true, &flags);
1595 } while (unlikely(ret == -EAGAIN));
1597 if (likely(ret >= 0)) {
1598 __queue_delayed_work(cpu, wq, dwork, delay);
1599 local_irq_restore(flags);
1602 /* -ENOENT from try_to_grab_pending() becomes %true */
1605 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1608 * worker_enter_idle - enter idle state
1609 * @worker: worker which is entering idle state
1611 * @worker is entering idle state. Update stats and idle timer if
1615 * spin_lock_irq(pool->lock).
1617 static void worker_enter_idle(struct worker *worker)
1619 struct worker_pool *pool = worker->pool;
1621 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1622 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1623 (worker->hentry.next || worker->hentry.pprev)))
1626 /* can't use worker_set_flags(), also called from create_worker() */
1627 worker->flags |= WORKER_IDLE;
1629 worker->last_active = jiffies;
1631 /* idle_list is LIFO */
1632 list_add(&worker->entry, &pool->idle_list);
1634 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1635 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1638 * Sanity check nr_running. Because wq_unbind_fn() releases
1639 * pool->lock between setting %WORKER_UNBOUND and zapping
1640 * nr_running, the warning may trigger spuriously. Check iff
1641 * unbind is not in progress.
1643 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1644 pool->nr_workers == pool->nr_idle &&
1645 atomic_read(&pool->nr_running));
1649 * worker_leave_idle - leave idle state
1650 * @worker: worker which is leaving idle state
1652 * @worker is leaving idle state. Update stats.
1655 * spin_lock_irq(pool->lock).
1657 static void worker_leave_idle(struct worker *worker)
1659 struct worker_pool *pool = worker->pool;
1661 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1663 worker_clr_flags(worker, WORKER_IDLE);
1665 list_del_init(&worker->entry);
1668 static struct worker *alloc_worker(int node)
1670 struct worker *worker;
1672 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1674 INIT_LIST_HEAD(&worker->entry);
1675 INIT_LIST_HEAD(&worker->scheduled);
1676 INIT_LIST_HEAD(&worker->node);
1677 /* on creation a worker is in !idle && prep state */
1678 worker->flags = WORKER_PREP;
1684 * worker_attach_to_pool() - attach a worker to a pool
1685 * @worker: worker to be attached
1686 * @pool: the target pool
1688 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1689 * cpu-binding of @worker are kept coordinated with the pool across
1692 static void worker_attach_to_pool(struct worker *worker,
1693 struct worker_pool *pool)
1695 mutex_lock(&pool->attach_mutex);
1698 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1699 * online CPUs. It'll be re-applied when any of the CPUs come up.
1701 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1704 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1705 * stable across this function. See the comments above the
1706 * flag definition for details.
1708 if (pool->flags & POOL_DISASSOCIATED)
1709 worker->flags |= WORKER_UNBOUND;
1711 list_add_tail(&worker->node, &pool->workers);
1713 mutex_unlock(&pool->attach_mutex);
1717 * worker_detach_from_pool() - detach a worker from its pool
1718 * @worker: worker which is attached to its pool
1719 * @pool: the pool @worker is attached to
1721 * Undo the attaching which had been done in worker_attach_to_pool(). The
1722 * caller worker shouldn't access to the pool after detached except it has
1723 * other reference to the pool.
1725 static void worker_detach_from_pool(struct worker *worker,
1726 struct worker_pool *pool)
1728 struct completion *detach_completion = NULL;
1730 mutex_lock(&pool->attach_mutex);
1731 list_del(&worker->node);
1732 if (list_empty(&pool->workers))
1733 detach_completion = pool->detach_completion;
1734 mutex_unlock(&pool->attach_mutex);
1736 /* clear leftover flags without pool->lock after it is detached */
1737 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1739 if (detach_completion)
1740 complete(detach_completion);
1744 * create_worker - create a new workqueue worker
1745 * @pool: pool the new worker will belong to
1747 * Create and start a new worker which is attached to @pool.
1750 * Might sleep. Does GFP_KERNEL allocations.
1753 * Pointer to the newly created worker.
1755 static struct worker *create_worker(struct worker_pool *pool)
1757 struct worker *worker = NULL;
1761 /* ID is needed to determine kthread name */
1762 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1766 worker = alloc_worker(pool->node);
1770 worker->pool = pool;
1774 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1775 pool->attrs->nice < 0 ? "H" : "");
1777 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1779 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1780 "kworker/%s", id_buf);
1781 if (IS_ERR(worker->task))
1784 set_user_nice(worker->task, pool->attrs->nice);
1785 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1787 /* successful, attach the worker to the pool */
1788 worker_attach_to_pool(worker, pool);
1790 /* start the newly created worker */
1791 spin_lock_irq(&pool->lock);
1792 worker->pool->nr_workers++;
1793 worker_enter_idle(worker);
1794 wake_up_process(worker->task);
1795 spin_unlock_irq(&pool->lock);
1801 ida_simple_remove(&pool->worker_ida, id);
1807 * destroy_worker - destroy a workqueue worker
1808 * @worker: worker to be destroyed
1810 * Destroy @worker and adjust @pool stats accordingly. The worker should
1814 * spin_lock_irq(pool->lock).
1816 static void destroy_worker(struct worker *worker)
1818 struct worker_pool *pool = worker->pool;
1820 lockdep_assert_held(&pool->lock);
1822 /* sanity check frenzy */
1823 if (WARN_ON(worker->current_work) ||
1824 WARN_ON(!list_empty(&worker->scheduled)) ||
1825 WARN_ON(!(worker->flags & WORKER_IDLE)))
1831 list_del_init(&worker->entry);
1832 worker->flags |= WORKER_DIE;
1833 wake_up_process(worker->task);
1836 static void idle_worker_timeout(unsigned long __pool)
1838 struct worker_pool *pool = (void *)__pool;
1840 spin_lock_irq(&pool->lock);
1842 while (too_many_workers(pool)) {
1843 struct worker *worker;
1844 unsigned long expires;
1846 /* idle_list is kept in LIFO order, check the last one */
1847 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1848 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1850 if (time_before(jiffies, expires)) {
1851 mod_timer(&pool->idle_timer, expires);
1855 destroy_worker(worker);
1858 spin_unlock_irq(&pool->lock);
1861 static void send_mayday(struct work_struct *work)
1863 struct pool_workqueue *pwq = get_work_pwq(work);
1864 struct workqueue_struct *wq = pwq->wq;
1866 lockdep_assert_held(&wq_mayday_lock);
1871 /* mayday mayday mayday */
1872 if (list_empty(&pwq->mayday_node)) {
1874 * If @pwq is for an unbound wq, its base ref may be put at
1875 * any time due to an attribute change. Pin @pwq until the
1876 * rescuer is done with it.
1879 list_add_tail(&pwq->mayday_node, &wq->maydays);
1880 wake_up_process(wq->rescuer->task);
1884 static void pool_mayday_timeout(unsigned long __pool)
1886 struct worker_pool *pool = (void *)__pool;
1887 struct work_struct *work;
1889 spin_lock_irq(&pool->lock);
1890 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1892 if (need_to_create_worker(pool)) {
1894 * We've been trying to create a new worker but
1895 * haven't been successful. We might be hitting an
1896 * allocation deadlock. Send distress signals to
1899 list_for_each_entry(work, &pool->worklist, entry)
1903 spin_unlock(&wq_mayday_lock);
1904 spin_unlock_irq(&pool->lock);
1906 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1910 * maybe_create_worker - create a new worker if necessary
1911 * @pool: pool to create a new worker for
1913 * Create a new worker for @pool if necessary. @pool is guaranteed to
1914 * have at least one idle worker on return from this function. If
1915 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1916 * sent to all rescuers with works scheduled on @pool to resolve
1917 * possible allocation deadlock.
1919 * On return, need_to_create_worker() is guaranteed to be %false and
1920 * may_start_working() %true.
1923 * spin_lock_irq(pool->lock) which may be released and regrabbed
1924 * multiple times. Does GFP_KERNEL allocations. Called only from
1927 static void maybe_create_worker(struct worker_pool *pool)
1928 __releases(&pool->lock)
1929 __acquires(&pool->lock)
1932 spin_unlock_irq(&pool->lock);
1934 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1935 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1938 if (create_worker(pool) || !need_to_create_worker(pool))
1941 schedule_timeout_interruptible(CREATE_COOLDOWN);
1943 if (!need_to_create_worker(pool))
1947 del_timer_sync(&pool->mayday_timer);
1948 spin_lock_irq(&pool->lock);
1950 * This is necessary even after a new worker was just successfully
1951 * created as @pool->lock was dropped and the new worker might have
1952 * already become busy.
1954 if (need_to_create_worker(pool))
1959 * manage_workers - manage worker pool
1962 * Assume the manager role and manage the worker pool @worker belongs
1963 * to. At any given time, there can be only zero or one manager per
1964 * pool. The exclusion is handled automatically by this function.
1966 * The caller can safely start processing works on false return. On
1967 * true return, it's guaranteed that need_to_create_worker() is false
1968 * and may_start_working() is true.
1971 * spin_lock_irq(pool->lock) which may be released and regrabbed
1972 * multiple times. Does GFP_KERNEL allocations.
1975 * %false if the pool doesn't need management and the caller can safely
1976 * start processing works, %true if management function was performed and
1977 * the conditions that the caller verified before calling the function may
1978 * no longer be true.
1980 static bool manage_workers(struct worker *worker)
1982 struct worker_pool *pool = worker->pool;
1984 if (pool->flags & POOL_MANAGER_ACTIVE)
1987 pool->flags |= POOL_MANAGER_ACTIVE;
1988 pool->manager = worker;
1990 maybe_create_worker(pool);
1992 pool->manager = NULL;
1993 pool->flags &= ~POOL_MANAGER_ACTIVE;
1994 wake_up(&wq_manager_wait);
1999 * process_one_work - process single work
2001 * @work: work to process
2003 * Process @work. This function contains all the logics necessary to
2004 * process a single work including synchronization against and
2005 * interaction with other workers on the same cpu, queueing and
2006 * flushing. As long as context requirement is met, any worker can
2007 * call this function to process a work.
2010 * spin_lock_irq(pool->lock) which is released and regrabbed.
2012 static void process_one_work(struct worker *worker, struct work_struct *work)
2013 __releases(&pool->lock)
2014 __acquires(&pool->lock)
2016 struct pool_workqueue *pwq = get_work_pwq(work);
2017 struct worker_pool *pool = worker->pool;
2018 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2020 struct worker *collision;
2021 #ifdef CONFIG_LOCKDEP
2023 * It is permissible to free the struct work_struct from
2024 * inside the function that is called from it, this we need to
2025 * take into account for lockdep too. To avoid bogus "held
2026 * lock freed" warnings as well as problems when looking into
2027 * work->lockdep_map, make a copy and use that here.
2029 struct lockdep_map lockdep_map;
2031 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2033 /* ensure we're on the correct CPU */
2034 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2035 raw_smp_processor_id() != pool->cpu);
2038 * A single work shouldn't be executed concurrently by
2039 * multiple workers on a single cpu. Check whether anyone is
2040 * already processing the work. If so, defer the work to the
2041 * currently executing one.
2043 collision = find_worker_executing_work(pool, work);
2044 if (unlikely(collision)) {
2045 move_linked_works(work, &collision->scheduled, NULL);
2049 /* claim and dequeue */
2050 debug_work_deactivate(work);
2051 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2052 worker->current_work = work;
2053 worker->current_func = work->func;
2054 worker->current_pwq = pwq;
2055 work_color = get_work_color(work);
2057 list_del_init(&work->entry);
2060 * CPU intensive works don't participate in concurrency management.
2061 * They're the scheduler's responsibility. This takes @worker out
2062 * of concurrency management and the next code block will chain
2063 * execution of the pending work items.
2065 if (unlikely(cpu_intensive))
2066 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2069 * Wake up another worker if necessary. The condition is always
2070 * false for normal per-cpu workers since nr_running would always
2071 * be >= 1 at this point. This is used to chain execution of the
2072 * pending work items for WORKER_NOT_RUNNING workers such as the
2073 * UNBOUND and CPU_INTENSIVE ones.
2075 if (need_more_worker(pool))
2076 wake_up_worker(pool);
2079 * Record the last pool and clear PENDING which should be the last
2080 * update to @work. Also, do this inside @pool->lock so that
2081 * PENDING and queued state changes happen together while IRQ is
2084 set_work_pool_and_clear_pending(work, pool->id);
2086 spin_unlock_irq(&pool->lock);
2088 lock_map_acquire(&pwq->wq->lockdep_map);
2089 lock_map_acquire(&lockdep_map);
2091 * Strictly speaking we should mark the invariant state without holding
2092 * any locks, that is, before these two lock_map_acquire()'s.
2094 * However, that would result in:
2101 * Which would create W1->C->W1 dependencies, even though there is no
2102 * actual deadlock possible. There are two solutions, using a
2103 * read-recursive acquire on the work(queue) 'locks', but this will then
2104 * hit the lockdep limitation on recursive locks, or simply discard
2107 * AFAICT there is no possible deadlock scenario between the
2108 * flush_work() and complete() primitives (except for single-threaded
2109 * workqueues), so hiding them isn't a problem.
2111 lockdep_invariant_state(true);
2112 trace_workqueue_execute_start(work);
2113 worker->current_func(work);
2115 * While we must be careful to not use "work" after this, the trace
2116 * point will only record its address.
2118 trace_workqueue_execute_end(work);
2119 lock_map_release(&lockdep_map);
2120 lock_map_release(&pwq->wq->lockdep_map);
2122 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2123 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2124 " last function: %pf\n",
2125 current->comm, preempt_count(), task_pid_nr(current),
2126 worker->current_func);
2127 debug_show_held_locks(current);
2132 * The following prevents a kworker from hogging CPU on !PREEMPT
2133 * kernels, where a requeueing work item waiting for something to
2134 * happen could deadlock with stop_machine as such work item could
2135 * indefinitely requeue itself while all other CPUs are trapped in
2136 * stop_machine. At the same time, report a quiescent RCU state so
2137 * the same condition doesn't freeze RCU.
2139 cond_resched_rcu_qs();
2141 spin_lock_irq(&pool->lock);
2143 /* clear cpu intensive status */
2144 if (unlikely(cpu_intensive))
2145 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2147 /* we're done with it, release */
2148 hash_del(&worker->hentry);
2149 worker->current_work = NULL;
2150 worker->current_func = NULL;
2151 worker->current_pwq = NULL;
2152 worker->desc_valid = false;
2153 pwq_dec_nr_in_flight(pwq, work_color);
2157 * process_scheduled_works - process scheduled works
2160 * Process all scheduled works. Please note that the scheduled list
2161 * may change while processing a work, so this function repeatedly
2162 * fetches a work from the top and executes it.
2165 * spin_lock_irq(pool->lock) which may be released and regrabbed
2168 static void process_scheduled_works(struct worker *worker)
2170 while (!list_empty(&worker->scheduled)) {
2171 struct work_struct *work = list_first_entry(&worker->scheduled,
2172 struct work_struct, entry);
2173 process_one_work(worker, work);
2178 * worker_thread - the worker thread function
2181 * The worker thread function. All workers belong to a worker_pool -
2182 * either a per-cpu one or dynamic unbound one. These workers process all
2183 * work items regardless of their specific target workqueue. The only
2184 * exception is work items which belong to workqueues with a rescuer which
2185 * will be explained in rescuer_thread().
2189 static int worker_thread(void *__worker)
2191 struct worker *worker = __worker;
2192 struct worker_pool *pool = worker->pool;
2194 /* tell the scheduler that this is a workqueue worker */
2195 worker->task->flags |= PF_WQ_WORKER;
2197 spin_lock_irq(&pool->lock);
2199 /* am I supposed to die? */
2200 if (unlikely(worker->flags & WORKER_DIE)) {
2201 spin_unlock_irq(&pool->lock);
2202 WARN_ON_ONCE(!list_empty(&worker->entry));
2203 worker->task->flags &= ~PF_WQ_WORKER;
2205 set_task_comm(worker->task, "kworker/dying");
2206 ida_simple_remove(&pool->worker_ida, worker->id);
2207 worker_detach_from_pool(worker, pool);
2212 worker_leave_idle(worker);
2214 /* no more worker necessary? */
2215 if (!need_more_worker(pool))
2218 /* do we need to manage? */
2219 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2223 * ->scheduled list can only be filled while a worker is
2224 * preparing to process a work or actually processing it.
2225 * Make sure nobody diddled with it while I was sleeping.
2227 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2230 * Finish PREP stage. We're guaranteed to have at least one idle
2231 * worker or that someone else has already assumed the manager
2232 * role. This is where @worker starts participating in concurrency
2233 * management if applicable and concurrency management is restored
2234 * after being rebound. See rebind_workers() for details.
2236 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2239 struct work_struct *work =
2240 list_first_entry(&pool->worklist,
2241 struct work_struct, entry);
2243 pool->watchdog_ts = jiffies;
2245 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2246 /* optimization path, not strictly necessary */
2247 process_one_work(worker, work);
2248 if (unlikely(!list_empty(&worker->scheduled)))
2249 process_scheduled_works(worker);
2251 move_linked_works(work, &worker->scheduled, NULL);
2252 process_scheduled_works(worker);
2254 } while (keep_working(pool));
2256 worker_set_flags(worker, WORKER_PREP);
2259 * pool->lock is held and there's no work to process and no need to
2260 * manage, sleep. Workers are woken up only while holding
2261 * pool->lock or from local cpu, so setting the current state
2262 * before releasing pool->lock is enough to prevent losing any
2265 worker_enter_idle(worker);
2266 __set_current_state(TASK_IDLE);
2267 spin_unlock_irq(&pool->lock);
2273 * rescuer_thread - the rescuer thread function
2276 * Workqueue rescuer thread function. There's one rescuer for each
2277 * workqueue which has WQ_MEM_RECLAIM set.
2279 * Regular work processing on a pool may block trying to create a new
2280 * worker which uses GFP_KERNEL allocation which has slight chance of
2281 * developing into deadlock if some works currently on the same queue
2282 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2283 * the problem rescuer solves.
2285 * When such condition is possible, the pool summons rescuers of all
2286 * workqueues which have works queued on the pool and let them process
2287 * those works so that forward progress can be guaranteed.
2289 * This should happen rarely.
2293 static int rescuer_thread(void *__rescuer)
2295 struct worker *rescuer = __rescuer;
2296 struct workqueue_struct *wq = rescuer->rescue_wq;
2297 struct list_head *scheduled = &rescuer->scheduled;
2300 set_user_nice(current, RESCUER_NICE_LEVEL);
2303 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2304 * doesn't participate in concurrency management.
2306 rescuer->task->flags |= PF_WQ_WORKER;
2308 set_current_state(TASK_IDLE);
2311 * By the time the rescuer is requested to stop, the workqueue
2312 * shouldn't have any work pending, but @wq->maydays may still have
2313 * pwq(s) queued. This can happen by non-rescuer workers consuming
2314 * all the work items before the rescuer got to them. Go through
2315 * @wq->maydays processing before acting on should_stop so that the
2316 * list is always empty on exit.
2318 should_stop = kthread_should_stop();
2320 /* see whether any pwq is asking for help */
2321 spin_lock_irq(&wq_mayday_lock);
2323 while (!list_empty(&wq->maydays)) {
2324 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2325 struct pool_workqueue, mayday_node);
2326 struct worker_pool *pool = pwq->pool;
2327 struct work_struct *work, *n;
2330 __set_current_state(TASK_RUNNING);
2331 list_del_init(&pwq->mayday_node);
2333 spin_unlock_irq(&wq_mayday_lock);
2335 worker_attach_to_pool(rescuer, pool);
2337 spin_lock_irq(&pool->lock);
2338 rescuer->pool = pool;
2341 * Slurp in all works issued via this workqueue and
2344 WARN_ON_ONCE(!list_empty(scheduled));
2345 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2346 if (get_work_pwq(work) == pwq) {
2348 pool->watchdog_ts = jiffies;
2349 move_linked_works(work, scheduled, &n);
2354 if (!list_empty(scheduled)) {
2355 process_scheduled_works(rescuer);
2358 * The above execution of rescued work items could
2359 * have created more to rescue through
2360 * pwq_activate_first_delayed() or chained
2361 * queueing. Let's put @pwq back on mayday list so
2362 * that such back-to-back work items, which may be
2363 * being used to relieve memory pressure, don't
2364 * incur MAYDAY_INTERVAL delay inbetween.
2366 if (need_to_create_worker(pool)) {
2367 spin_lock(&wq_mayday_lock);
2369 list_move_tail(&pwq->mayday_node, &wq->maydays);
2370 spin_unlock(&wq_mayday_lock);
2375 * Put the reference grabbed by send_mayday(). @pool won't
2376 * go away while we're still attached to it.
2381 * Leave this pool. If need_more_worker() is %true, notify a
2382 * regular worker; otherwise, we end up with 0 concurrency
2383 * and stalling the execution.
2385 if (need_more_worker(pool))
2386 wake_up_worker(pool);
2388 rescuer->pool = NULL;
2389 spin_unlock_irq(&pool->lock);
2391 worker_detach_from_pool(rescuer, pool);
2393 spin_lock_irq(&wq_mayday_lock);
2396 spin_unlock_irq(&wq_mayday_lock);
2399 __set_current_state(TASK_RUNNING);
2400 rescuer->task->flags &= ~PF_WQ_WORKER;
2404 /* rescuers should never participate in concurrency management */
2405 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2411 * check_flush_dependency - check for flush dependency sanity
2412 * @target_wq: workqueue being flushed
2413 * @target_work: work item being flushed (NULL for workqueue flushes)
2415 * %current is trying to flush the whole @target_wq or @target_work on it.
2416 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2417 * reclaiming memory or running on a workqueue which doesn't have
2418 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2421 static void check_flush_dependency(struct workqueue_struct *target_wq,
2422 struct work_struct *target_work)
2424 work_func_t target_func = target_work ? target_work->func : NULL;
2425 struct worker *worker;
2427 if (target_wq->flags & WQ_MEM_RECLAIM)
2430 worker = current_wq_worker();
2432 WARN_ONCE(current->flags & PF_MEMALLOC,
2433 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2434 current->pid, current->comm, target_wq->name, target_func);
2435 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2436 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2437 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2438 worker->current_pwq->wq->name, worker->current_func,
2439 target_wq->name, target_func);
2443 struct work_struct work;
2444 struct completion done;
2445 struct task_struct *task; /* purely informational */
2448 static void wq_barrier_func(struct work_struct *work)
2450 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2451 complete(&barr->done);
2455 * insert_wq_barrier - insert a barrier work
2456 * @pwq: pwq to insert barrier into
2457 * @barr: wq_barrier to insert
2458 * @target: target work to attach @barr to
2459 * @worker: worker currently executing @target, NULL if @target is not executing
2461 * @barr is linked to @target such that @barr is completed only after
2462 * @target finishes execution. Please note that the ordering
2463 * guarantee is observed only with respect to @target and on the local
2466 * Currently, a queued barrier can't be canceled. This is because
2467 * try_to_grab_pending() can't determine whether the work to be
2468 * grabbed is at the head of the queue and thus can't clear LINKED
2469 * flag of the previous work while there must be a valid next work
2470 * after a work with LINKED flag set.
2472 * Note that when @worker is non-NULL, @target may be modified
2473 * underneath us, so we can't reliably determine pwq from @target.
2476 * spin_lock_irq(pool->lock).
2478 static void insert_wq_barrier(struct pool_workqueue *pwq,
2479 struct wq_barrier *barr,
2480 struct work_struct *target, struct worker *worker)
2482 struct list_head *head;
2483 unsigned int linked = 0;
2486 * debugobject calls are safe here even with pool->lock locked
2487 * as we know for sure that this will not trigger any of the
2488 * checks and call back into the fixup functions where we
2491 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2492 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2494 init_completion_map(&barr->done, &target->lockdep_map);
2496 barr->task = current;
2499 * If @target is currently being executed, schedule the
2500 * barrier to the worker; otherwise, put it after @target.
2503 head = worker->scheduled.next;
2505 unsigned long *bits = work_data_bits(target);
2507 head = target->entry.next;
2508 /* there can already be other linked works, inherit and set */
2509 linked = *bits & WORK_STRUCT_LINKED;
2510 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2513 debug_work_activate(&barr->work);
2514 insert_work(pwq, &barr->work, head,
2515 work_color_to_flags(WORK_NO_COLOR) | linked);
2519 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2520 * @wq: workqueue being flushed
2521 * @flush_color: new flush color, < 0 for no-op
2522 * @work_color: new work color, < 0 for no-op
2524 * Prepare pwqs for workqueue flushing.
2526 * If @flush_color is non-negative, flush_color on all pwqs should be
2527 * -1. If no pwq has in-flight commands at the specified color, all
2528 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2529 * has in flight commands, its pwq->flush_color is set to
2530 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2531 * wakeup logic is armed and %true is returned.
2533 * The caller should have initialized @wq->first_flusher prior to
2534 * calling this function with non-negative @flush_color. If
2535 * @flush_color is negative, no flush color update is done and %false
2538 * If @work_color is non-negative, all pwqs should have the same
2539 * work_color which is previous to @work_color and all will be
2540 * advanced to @work_color.
2543 * mutex_lock(wq->mutex).
2546 * %true if @flush_color >= 0 and there's something to flush. %false
2549 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2550 int flush_color, int work_color)
2553 struct pool_workqueue *pwq;
2555 if (flush_color >= 0) {
2556 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2557 atomic_set(&wq->nr_pwqs_to_flush, 1);
2560 for_each_pwq(pwq, wq) {
2561 struct worker_pool *pool = pwq->pool;
2563 spin_lock_irq(&pool->lock);
2565 if (flush_color >= 0) {
2566 WARN_ON_ONCE(pwq->flush_color != -1);
2568 if (pwq->nr_in_flight[flush_color]) {
2569 pwq->flush_color = flush_color;
2570 atomic_inc(&wq->nr_pwqs_to_flush);
2575 if (work_color >= 0) {
2576 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2577 pwq->work_color = work_color;
2580 spin_unlock_irq(&pool->lock);
2583 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2584 complete(&wq->first_flusher->done);
2590 * flush_workqueue - ensure that any scheduled work has run to completion.
2591 * @wq: workqueue to flush
2593 * This function sleeps until all work items which were queued on entry
2594 * have finished execution, but it is not livelocked by new incoming ones.
2596 void flush_workqueue(struct workqueue_struct *wq)
2598 struct wq_flusher this_flusher = {
2599 .list = LIST_HEAD_INIT(this_flusher.list),
2601 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2605 if (WARN_ON(!wq_online))
2608 mutex_lock(&wq->mutex);
2611 * Start-to-wait phase
2613 next_color = work_next_color(wq->work_color);
2615 if (next_color != wq->flush_color) {
2617 * Color space is not full. The current work_color
2618 * becomes our flush_color and work_color is advanced
2621 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2622 this_flusher.flush_color = wq->work_color;
2623 wq->work_color = next_color;
2625 if (!wq->first_flusher) {
2626 /* no flush in progress, become the first flusher */
2627 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2629 wq->first_flusher = &this_flusher;
2631 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2633 /* nothing to flush, done */
2634 wq->flush_color = next_color;
2635 wq->first_flusher = NULL;
2640 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2641 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2642 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2646 * Oops, color space is full, wait on overflow queue.
2647 * The next flush completion will assign us
2648 * flush_color and transfer to flusher_queue.
2650 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2653 check_flush_dependency(wq, NULL);
2655 mutex_unlock(&wq->mutex);
2657 wait_for_completion(&this_flusher.done);
2660 * Wake-up-and-cascade phase
2662 * First flushers are responsible for cascading flushes and
2663 * handling overflow. Non-first flushers can simply return.
2665 if (wq->first_flusher != &this_flusher)
2668 mutex_lock(&wq->mutex);
2670 /* we might have raced, check again with mutex held */
2671 if (wq->first_flusher != &this_flusher)
2674 wq->first_flusher = NULL;
2676 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2677 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2680 struct wq_flusher *next, *tmp;
2682 /* complete all the flushers sharing the current flush color */
2683 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2684 if (next->flush_color != wq->flush_color)
2686 list_del_init(&next->list);
2687 complete(&next->done);
2690 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2691 wq->flush_color != work_next_color(wq->work_color));
2693 /* this flush_color is finished, advance by one */
2694 wq->flush_color = work_next_color(wq->flush_color);
2696 /* one color has been freed, handle overflow queue */
2697 if (!list_empty(&wq->flusher_overflow)) {
2699 * Assign the same color to all overflowed
2700 * flushers, advance work_color and append to
2701 * flusher_queue. This is the start-to-wait
2702 * phase for these overflowed flushers.
2704 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2705 tmp->flush_color = wq->work_color;
2707 wq->work_color = work_next_color(wq->work_color);
2709 list_splice_tail_init(&wq->flusher_overflow,
2710 &wq->flusher_queue);
2711 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2714 if (list_empty(&wq->flusher_queue)) {
2715 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2720 * Need to flush more colors. Make the next flusher
2721 * the new first flusher and arm pwqs.
2723 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2724 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2726 list_del_init(&next->list);
2727 wq->first_flusher = next;
2729 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2733 * Meh... this color is already done, clear first
2734 * flusher and repeat cascading.
2736 wq->first_flusher = NULL;
2740 mutex_unlock(&wq->mutex);
2742 EXPORT_SYMBOL(flush_workqueue);
2745 * drain_workqueue - drain a workqueue
2746 * @wq: workqueue to drain
2748 * Wait until the workqueue becomes empty. While draining is in progress,
2749 * only chain queueing is allowed. IOW, only currently pending or running
2750 * work items on @wq can queue further work items on it. @wq is flushed
2751 * repeatedly until it becomes empty. The number of flushing is determined
2752 * by the depth of chaining and should be relatively short. Whine if it
2755 void drain_workqueue(struct workqueue_struct *wq)
2757 unsigned int flush_cnt = 0;
2758 struct pool_workqueue *pwq;
2761 * __queue_work() needs to test whether there are drainers, is much
2762 * hotter than drain_workqueue() and already looks at @wq->flags.
2763 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2765 mutex_lock(&wq->mutex);
2766 if (!wq->nr_drainers++)
2767 wq->flags |= __WQ_DRAINING;
2768 mutex_unlock(&wq->mutex);
2770 flush_workqueue(wq);
2772 mutex_lock(&wq->mutex);
2774 for_each_pwq(pwq, wq) {
2777 spin_lock_irq(&pwq->pool->lock);
2778 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2779 spin_unlock_irq(&pwq->pool->lock);
2784 if (++flush_cnt == 10 ||
2785 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2786 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2787 wq->name, flush_cnt);
2789 mutex_unlock(&wq->mutex);
2793 if (!--wq->nr_drainers)
2794 wq->flags &= ~__WQ_DRAINING;
2795 mutex_unlock(&wq->mutex);
2797 EXPORT_SYMBOL_GPL(drain_workqueue);
2799 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2801 struct worker *worker = NULL;
2802 struct worker_pool *pool;
2803 struct pool_workqueue *pwq;
2807 local_irq_disable();
2808 pool = get_work_pool(work);
2814 spin_lock(&pool->lock);
2815 /* see the comment in try_to_grab_pending() with the same code */
2816 pwq = get_work_pwq(work);
2818 if (unlikely(pwq->pool != pool))
2821 worker = find_worker_executing_work(pool, work);
2824 pwq = worker->current_pwq;
2827 check_flush_dependency(pwq->wq, work);
2829 insert_wq_barrier(pwq, barr, work, worker);
2830 spin_unlock_irq(&pool->lock);
2833 * Force a lock recursion deadlock when using flush_work() inside a
2834 * single-threaded or rescuer equipped workqueue.
2836 * For single threaded workqueues the deadlock happens when the work
2837 * is after the work issuing the flush_work(). For rescuer equipped
2838 * workqueues the deadlock happens when the rescuer stalls, blocking
2841 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
2842 lock_map_acquire(&pwq->wq->lockdep_map);
2843 lock_map_release(&pwq->wq->lockdep_map);
2848 spin_unlock_irq(&pool->lock);
2853 * flush_work - wait for a work to finish executing the last queueing instance
2854 * @work: the work to flush
2856 * Wait until @work has finished execution. @work is guaranteed to be idle
2857 * on return if it hasn't been requeued since flush started.
2860 * %true if flush_work() waited for the work to finish execution,
2861 * %false if it was already idle.
2863 bool flush_work(struct work_struct *work)
2865 struct wq_barrier barr;
2867 if (WARN_ON(!wq_online))
2870 if (start_flush_work(work, &barr)) {
2871 wait_for_completion(&barr.done);
2872 destroy_work_on_stack(&barr.work);
2878 EXPORT_SYMBOL_GPL(flush_work);
2881 wait_queue_entry_t wait;
2882 struct work_struct *work;
2885 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2887 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2889 if (cwait->work != key)
2891 return autoremove_wake_function(wait, mode, sync, key);
2894 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2896 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2897 unsigned long flags;
2901 ret = try_to_grab_pending(work, is_dwork, &flags);
2903 * If someone else is already canceling, wait for it to
2904 * finish. flush_work() doesn't work for PREEMPT_NONE
2905 * because we may get scheduled between @work's completion
2906 * and the other canceling task resuming and clearing
2907 * CANCELING - flush_work() will return false immediately
2908 * as @work is no longer busy, try_to_grab_pending() will
2909 * return -ENOENT as @work is still being canceled and the
2910 * other canceling task won't be able to clear CANCELING as
2911 * we're hogging the CPU.
2913 * Let's wait for completion using a waitqueue. As this
2914 * may lead to the thundering herd problem, use a custom
2915 * wake function which matches @work along with exclusive
2918 if (unlikely(ret == -ENOENT)) {
2919 struct cwt_wait cwait;
2921 init_wait(&cwait.wait);
2922 cwait.wait.func = cwt_wakefn;
2925 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2926 TASK_UNINTERRUPTIBLE);
2927 if (work_is_canceling(work))
2929 finish_wait(&cancel_waitq, &cwait.wait);
2931 } while (unlikely(ret < 0));
2933 /* tell other tasks trying to grab @work to back off */
2934 mark_work_canceling(work);
2935 local_irq_restore(flags);
2938 * This allows canceling during early boot. We know that @work
2944 clear_work_data(work);
2947 * Paired with prepare_to_wait() above so that either
2948 * waitqueue_active() is visible here or !work_is_canceling() is
2952 if (waitqueue_active(&cancel_waitq))
2953 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2959 * cancel_work_sync - cancel a work and wait for it to finish
2960 * @work: the work to cancel
2962 * Cancel @work and wait for its execution to finish. This function
2963 * can be used even if the work re-queues itself or migrates to
2964 * another workqueue. On return from this function, @work is
2965 * guaranteed to be not pending or executing on any CPU.
2967 * cancel_work_sync(&delayed_work->work) must not be used for
2968 * delayed_work's. Use cancel_delayed_work_sync() instead.
2970 * The caller must ensure that the workqueue on which @work was last
2971 * queued can't be destroyed before this function returns.
2974 * %true if @work was pending, %false otherwise.
2976 bool cancel_work_sync(struct work_struct *work)
2978 return __cancel_work_timer(work, false);
2980 EXPORT_SYMBOL_GPL(cancel_work_sync);
2983 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2984 * @dwork: the delayed work to flush
2986 * Delayed timer is cancelled and the pending work is queued for
2987 * immediate execution. Like flush_work(), this function only
2988 * considers the last queueing instance of @dwork.
2991 * %true if flush_work() waited for the work to finish execution,
2992 * %false if it was already idle.
2994 bool flush_delayed_work(struct delayed_work *dwork)
2996 local_irq_disable();
2997 if (del_timer_sync(&dwork->timer))
2998 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3000 return flush_work(&dwork->work);
3002 EXPORT_SYMBOL(flush_delayed_work);
3004 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3006 unsigned long flags;
3010 ret = try_to_grab_pending(work, is_dwork, &flags);
3011 } while (unlikely(ret == -EAGAIN));
3013 if (unlikely(ret < 0))
3016 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3017 local_irq_restore(flags);
3022 * See cancel_delayed_work()
3024 bool cancel_work(struct work_struct *work)
3026 return __cancel_work(work, false);
3030 * cancel_delayed_work - cancel a delayed work
3031 * @dwork: delayed_work to cancel
3033 * Kill off a pending delayed_work.
3035 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3039 * The work callback function may still be running on return, unless
3040 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3041 * use cancel_delayed_work_sync() to wait on it.
3043 * This function is safe to call from any context including IRQ handler.
3045 bool cancel_delayed_work(struct delayed_work *dwork)
3047 return __cancel_work(&dwork->work, true);
3049 EXPORT_SYMBOL(cancel_delayed_work);
3052 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3053 * @dwork: the delayed work cancel
3055 * This is cancel_work_sync() for delayed works.
3058 * %true if @dwork was pending, %false otherwise.
3060 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3062 return __cancel_work_timer(&dwork->work, true);
3064 EXPORT_SYMBOL(cancel_delayed_work_sync);
3067 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3068 * @func: the function to call
3070 * schedule_on_each_cpu() executes @func on each online CPU using the
3071 * system workqueue and blocks until all CPUs have completed.
3072 * schedule_on_each_cpu() is very slow.
3075 * 0 on success, -errno on failure.
3077 int schedule_on_each_cpu(work_func_t func)
3080 struct work_struct __percpu *works;
3082 works = alloc_percpu(struct work_struct);
3088 for_each_online_cpu(cpu) {
3089 struct work_struct *work = per_cpu_ptr(works, cpu);
3091 INIT_WORK(work, func);
3092 schedule_work_on(cpu, work);
3095 for_each_online_cpu(cpu)
3096 flush_work(per_cpu_ptr(works, cpu));
3104 * execute_in_process_context - reliably execute the routine with user context
3105 * @fn: the function to execute
3106 * @ew: guaranteed storage for the execute work structure (must
3107 * be available when the work executes)
3109 * Executes the function immediately if process context is available,
3110 * otherwise schedules the function for delayed execution.
3112 * Return: 0 - function was executed
3113 * 1 - function was scheduled for execution
3115 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3117 if (!in_interrupt()) {
3122 INIT_WORK(&ew->work, fn);
3123 schedule_work(&ew->work);
3127 EXPORT_SYMBOL_GPL(execute_in_process_context);
3130 * free_workqueue_attrs - free a workqueue_attrs
3131 * @attrs: workqueue_attrs to free
3133 * Undo alloc_workqueue_attrs().
3135 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3138 free_cpumask_var(attrs->cpumask);
3144 * alloc_workqueue_attrs - allocate a workqueue_attrs
3145 * @gfp_mask: allocation mask to use
3147 * Allocate a new workqueue_attrs, initialize with default settings and
3150 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3152 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3154 struct workqueue_attrs *attrs;
3156 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3159 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3162 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3165 free_workqueue_attrs(attrs);
3169 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3170 const struct workqueue_attrs *from)
3172 to->nice = from->nice;
3173 cpumask_copy(to->cpumask, from->cpumask);
3175 * Unlike hash and equality test, this function doesn't ignore
3176 * ->no_numa as it is used for both pool and wq attrs. Instead,
3177 * get_unbound_pool() explicitly clears ->no_numa after copying.
3179 to->no_numa = from->no_numa;
3182 /* hash value of the content of @attr */
3183 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3187 hash = jhash_1word(attrs->nice, hash);
3188 hash = jhash(cpumask_bits(attrs->cpumask),
3189 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3193 /* content equality test */
3194 static bool wqattrs_equal(const struct workqueue_attrs *a,
3195 const struct workqueue_attrs *b)
3197 if (a->nice != b->nice)
3199 if (!cpumask_equal(a->cpumask, b->cpumask))
3205 * init_worker_pool - initialize a newly zalloc'd worker_pool
3206 * @pool: worker_pool to initialize
3208 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3210 * Return: 0 on success, -errno on failure. Even on failure, all fields
3211 * inside @pool proper are initialized and put_unbound_pool() can be called
3212 * on @pool safely to release it.
3214 static int init_worker_pool(struct worker_pool *pool)
3216 spin_lock_init(&pool->lock);
3219 pool->node = NUMA_NO_NODE;
3220 pool->flags |= POOL_DISASSOCIATED;
3221 pool->watchdog_ts = jiffies;
3222 INIT_LIST_HEAD(&pool->worklist);
3223 INIT_LIST_HEAD(&pool->idle_list);
3224 hash_init(pool->busy_hash);
3226 setup_deferrable_timer(&pool->idle_timer, idle_worker_timeout,
3227 (unsigned long)pool);
3229 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3230 (unsigned long)pool);
3232 mutex_init(&pool->attach_mutex);
3233 INIT_LIST_HEAD(&pool->workers);
3235 ida_init(&pool->worker_ida);
3236 INIT_HLIST_NODE(&pool->hash_node);
3239 /* shouldn't fail above this point */
3240 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3246 static void rcu_free_wq(struct rcu_head *rcu)
3248 struct workqueue_struct *wq =
3249 container_of(rcu, struct workqueue_struct, rcu);
3251 if (!(wq->flags & WQ_UNBOUND))
3252 free_percpu(wq->cpu_pwqs);
3254 free_workqueue_attrs(wq->unbound_attrs);
3260 static void rcu_free_pool(struct rcu_head *rcu)
3262 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3264 ida_destroy(&pool->worker_ida);
3265 free_workqueue_attrs(pool->attrs);
3270 * put_unbound_pool - put a worker_pool
3271 * @pool: worker_pool to put
3273 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3274 * safe manner. get_unbound_pool() calls this function on its failure path
3275 * and this function should be able to release pools which went through,
3276 * successfully or not, init_worker_pool().
3278 * Should be called with wq_pool_mutex held.
3280 static void put_unbound_pool(struct worker_pool *pool)
3282 DECLARE_COMPLETION_ONSTACK(detach_completion);
3283 struct worker *worker;
3285 lockdep_assert_held(&wq_pool_mutex);
3291 if (WARN_ON(!(pool->cpu < 0)) ||
3292 WARN_ON(!list_empty(&pool->worklist)))
3295 /* release id and unhash */
3297 idr_remove(&worker_pool_idr, pool->id);
3298 hash_del(&pool->hash_node);
3301 * Become the manager and destroy all workers. This prevents
3302 * @pool's workers from blocking on attach_mutex. We're the last
3303 * manager and @pool gets freed with the flag set.
3305 spin_lock_irq(&pool->lock);
3306 wait_event_lock_irq(wq_manager_wait,
3307 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3308 pool->flags |= POOL_MANAGER_ACTIVE;
3310 while ((worker = first_idle_worker(pool)))
3311 destroy_worker(worker);
3312 WARN_ON(pool->nr_workers || pool->nr_idle);
3313 spin_unlock_irq(&pool->lock);
3315 mutex_lock(&pool->attach_mutex);
3316 if (!list_empty(&pool->workers))
3317 pool->detach_completion = &detach_completion;
3318 mutex_unlock(&pool->attach_mutex);
3320 if (pool->detach_completion)
3321 wait_for_completion(pool->detach_completion);
3323 /* shut down the timers */
3324 del_timer_sync(&pool->idle_timer);
3325 del_timer_sync(&pool->mayday_timer);
3327 /* sched-RCU protected to allow dereferences from get_work_pool() */
3328 call_rcu_sched(&pool->rcu, rcu_free_pool);
3332 * get_unbound_pool - get a worker_pool with the specified attributes
3333 * @attrs: the attributes of the worker_pool to get
3335 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3336 * reference count and return it. If there already is a matching
3337 * worker_pool, it will be used; otherwise, this function attempts to
3340 * Should be called with wq_pool_mutex held.
3342 * Return: On success, a worker_pool with the same attributes as @attrs.
3343 * On failure, %NULL.
3345 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3347 u32 hash = wqattrs_hash(attrs);
3348 struct worker_pool *pool;
3350 int target_node = NUMA_NO_NODE;
3352 lockdep_assert_held(&wq_pool_mutex);
3354 /* do we already have a matching pool? */
3355 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3356 if (wqattrs_equal(pool->attrs, attrs)) {
3362 /* if cpumask is contained inside a NUMA node, we belong to that node */
3363 if (wq_numa_enabled) {
3364 for_each_node(node) {
3365 if (cpumask_subset(attrs->cpumask,
3366 wq_numa_possible_cpumask[node])) {
3373 /* nope, create a new one */
3374 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3375 if (!pool || init_worker_pool(pool) < 0)
3378 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3379 copy_workqueue_attrs(pool->attrs, attrs);
3380 pool->node = target_node;
3383 * no_numa isn't a worker_pool attribute, always clear it. See
3384 * 'struct workqueue_attrs' comments for detail.
3386 pool->attrs->no_numa = false;
3388 if (worker_pool_assign_id(pool) < 0)
3391 /* create and start the initial worker */
3392 if (wq_online && !create_worker(pool))
3396 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3401 put_unbound_pool(pool);
3405 static void rcu_free_pwq(struct rcu_head *rcu)
3407 kmem_cache_free(pwq_cache,
3408 container_of(rcu, struct pool_workqueue, rcu));
3412 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3413 * and needs to be destroyed.
3415 static void pwq_unbound_release_workfn(struct work_struct *work)
3417 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3418 unbound_release_work);
3419 struct workqueue_struct *wq = pwq->wq;
3420 struct worker_pool *pool = pwq->pool;
3423 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3426 mutex_lock(&wq->mutex);
3427 list_del_rcu(&pwq->pwqs_node);
3428 is_last = list_empty(&wq->pwqs);
3429 mutex_unlock(&wq->mutex);
3431 mutex_lock(&wq_pool_mutex);
3432 put_unbound_pool(pool);
3433 mutex_unlock(&wq_pool_mutex);
3435 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3438 * If we're the last pwq going away, @wq is already dead and no one
3439 * is gonna access it anymore. Schedule RCU free.
3442 call_rcu_sched(&wq->rcu, rcu_free_wq);
3446 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3447 * @pwq: target pool_workqueue
3449 * If @pwq isn't freezing, set @pwq->max_active to the associated
3450 * workqueue's saved_max_active and activate delayed work items
3451 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3453 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3455 struct workqueue_struct *wq = pwq->wq;
3456 bool freezable = wq->flags & WQ_FREEZABLE;
3457 unsigned long flags;
3459 /* for @wq->saved_max_active */
3460 lockdep_assert_held(&wq->mutex);
3462 /* fast exit for non-freezable wqs */
3463 if (!freezable && pwq->max_active == wq->saved_max_active)
3466 /* this function can be called during early boot w/ irq disabled */
3467 spin_lock_irqsave(&pwq->pool->lock, flags);
3470 * During [un]freezing, the caller is responsible for ensuring that
3471 * this function is called at least once after @workqueue_freezing
3472 * is updated and visible.
3474 if (!freezable || !workqueue_freezing) {
3475 pwq->max_active = wq->saved_max_active;
3477 while (!list_empty(&pwq->delayed_works) &&
3478 pwq->nr_active < pwq->max_active)
3479 pwq_activate_first_delayed(pwq);
3482 * Need to kick a worker after thawed or an unbound wq's
3483 * max_active is bumped. It's a slow path. Do it always.
3485 wake_up_worker(pwq->pool);
3487 pwq->max_active = 0;
3490 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3493 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3494 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3495 struct worker_pool *pool)
3497 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3499 memset(pwq, 0, sizeof(*pwq));
3503 pwq->flush_color = -1;
3505 INIT_LIST_HEAD(&pwq->delayed_works);
3506 INIT_LIST_HEAD(&pwq->pwqs_node);
3507 INIT_LIST_HEAD(&pwq->mayday_node);
3508 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3511 /* sync @pwq with the current state of its associated wq and link it */
3512 static void link_pwq(struct pool_workqueue *pwq)
3514 struct workqueue_struct *wq = pwq->wq;
3516 lockdep_assert_held(&wq->mutex);
3518 /* may be called multiple times, ignore if already linked */
3519 if (!list_empty(&pwq->pwqs_node))
3522 /* set the matching work_color */
3523 pwq->work_color = wq->work_color;
3525 /* sync max_active to the current setting */
3526 pwq_adjust_max_active(pwq);
3529 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3532 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3533 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3534 const struct workqueue_attrs *attrs)
3536 struct worker_pool *pool;
3537 struct pool_workqueue *pwq;
3539 lockdep_assert_held(&wq_pool_mutex);
3541 pool = get_unbound_pool(attrs);
3545 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3547 put_unbound_pool(pool);
3551 init_pwq(pwq, wq, pool);
3556 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3557 * @attrs: the wq_attrs of the default pwq of the target workqueue
3558 * @node: the target NUMA node
3559 * @cpu_going_down: if >= 0, the CPU to consider as offline
3560 * @cpumask: outarg, the resulting cpumask
3562 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3563 * @cpu_going_down is >= 0, that cpu is considered offline during
3564 * calculation. The result is stored in @cpumask.
3566 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3567 * enabled and @node has online CPUs requested by @attrs, the returned
3568 * cpumask is the intersection of the possible CPUs of @node and
3571 * The caller is responsible for ensuring that the cpumask of @node stays
3574 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3577 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3578 int cpu_going_down, cpumask_t *cpumask)
3580 if (!wq_numa_enabled || attrs->no_numa)
3583 /* does @node have any online CPUs @attrs wants? */
3584 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3585 if (cpu_going_down >= 0)
3586 cpumask_clear_cpu(cpu_going_down, cpumask);
3588 if (cpumask_empty(cpumask))
3591 /* yeap, return possible CPUs in @node that @attrs wants */
3592 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3594 if (cpumask_empty(cpumask)) {
3595 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3596 "possible intersect\n");
3600 return !cpumask_equal(cpumask, attrs->cpumask);
3603 cpumask_copy(cpumask, attrs->cpumask);
3607 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3608 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3610 struct pool_workqueue *pwq)
3612 struct pool_workqueue *old_pwq;
3614 lockdep_assert_held(&wq_pool_mutex);
3615 lockdep_assert_held(&wq->mutex);
3617 /* link_pwq() can handle duplicate calls */
3620 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3621 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3625 /* context to store the prepared attrs & pwqs before applying */
3626 struct apply_wqattrs_ctx {
3627 struct workqueue_struct *wq; /* target workqueue */
3628 struct workqueue_attrs *attrs; /* attrs to apply */
3629 struct list_head list; /* queued for batching commit */
3630 struct pool_workqueue *dfl_pwq;
3631 struct pool_workqueue *pwq_tbl[];
3634 /* free the resources after success or abort */
3635 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3641 put_pwq_unlocked(ctx->pwq_tbl[node]);
3642 put_pwq_unlocked(ctx->dfl_pwq);
3644 free_workqueue_attrs(ctx->attrs);
3650 /* allocate the attrs and pwqs for later installation */
3651 static struct apply_wqattrs_ctx *
3652 apply_wqattrs_prepare(struct workqueue_struct *wq,
3653 const struct workqueue_attrs *attrs)
3655 struct apply_wqattrs_ctx *ctx;
3656 struct workqueue_attrs *new_attrs, *tmp_attrs;
3659 lockdep_assert_held(&wq_pool_mutex);
3661 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3664 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3665 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3666 if (!ctx || !new_attrs || !tmp_attrs)
3670 * Calculate the attrs of the default pwq.
3671 * If the user configured cpumask doesn't overlap with the
3672 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3674 copy_workqueue_attrs(new_attrs, attrs);
3675 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3676 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3677 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3680 * We may create multiple pwqs with differing cpumasks. Make a
3681 * copy of @new_attrs which will be modified and used to obtain
3684 copy_workqueue_attrs(tmp_attrs, new_attrs);
3687 * If something goes wrong during CPU up/down, we'll fall back to
3688 * the default pwq covering whole @attrs->cpumask. Always create
3689 * it even if we don't use it immediately.
3691 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3695 for_each_node(node) {
3696 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3697 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3698 if (!ctx->pwq_tbl[node])
3701 ctx->dfl_pwq->refcnt++;
3702 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3706 /* save the user configured attrs and sanitize it. */
3707 copy_workqueue_attrs(new_attrs, attrs);
3708 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3709 ctx->attrs = new_attrs;
3712 free_workqueue_attrs(tmp_attrs);
3716 free_workqueue_attrs(tmp_attrs);
3717 free_workqueue_attrs(new_attrs);
3718 apply_wqattrs_cleanup(ctx);
3722 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3723 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3727 /* all pwqs have been created successfully, let's install'em */
3728 mutex_lock(&ctx->wq->mutex);
3730 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3732 /* save the previous pwq and install the new one */
3734 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3735 ctx->pwq_tbl[node]);
3737 /* @dfl_pwq might not have been used, ensure it's linked */
3738 link_pwq(ctx->dfl_pwq);
3739 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3741 mutex_unlock(&ctx->wq->mutex);
3744 static void apply_wqattrs_lock(void)
3746 /* CPUs should stay stable across pwq creations and installations */
3748 mutex_lock(&wq_pool_mutex);
3751 static void apply_wqattrs_unlock(void)
3753 mutex_unlock(&wq_pool_mutex);
3757 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3758 const struct workqueue_attrs *attrs)
3760 struct apply_wqattrs_ctx *ctx;
3762 /* only unbound workqueues can change attributes */
3763 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3766 /* creating multiple pwqs breaks ordering guarantee */
3767 if (!list_empty(&wq->pwqs)) {
3768 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3771 wq->flags &= ~__WQ_ORDERED;
3774 ctx = apply_wqattrs_prepare(wq, attrs);
3778 /* the ctx has been prepared successfully, let's commit it */
3779 apply_wqattrs_commit(ctx);
3780 apply_wqattrs_cleanup(ctx);
3786 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3787 * @wq: the target workqueue
3788 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3790 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3791 * machines, this function maps a separate pwq to each NUMA node with
3792 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3793 * NUMA node it was issued on. Older pwqs are released as in-flight work
3794 * items finish. Note that a work item which repeatedly requeues itself
3795 * back-to-back will stay on its current pwq.
3797 * Performs GFP_KERNEL allocations.
3799 * Return: 0 on success and -errno on failure.
3801 int apply_workqueue_attrs(struct workqueue_struct *wq,
3802 const struct workqueue_attrs *attrs)
3806 apply_wqattrs_lock();
3807 ret = apply_workqueue_attrs_locked(wq, attrs);
3808 apply_wqattrs_unlock();
3814 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3815 * @wq: the target workqueue
3816 * @cpu: the CPU coming up or going down
3817 * @online: whether @cpu is coming up or going down
3819 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3820 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3823 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3824 * falls back to @wq->dfl_pwq which may not be optimal but is always
3827 * Note that when the last allowed CPU of a NUMA node goes offline for a
3828 * workqueue with a cpumask spanning multiple nodes, the workers which were
3829 * already executing the work items for the workqueue will lose their CPU
3830 * affinity and may execute on any CPU. This is similar to how per-cpu
3831 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3832 * affinity, it's the user's responsibility to flush the work item from
3835 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3838 int node = cpu_to_node(cpu);
3839 int cpu_off = online ? -1 : cpu;
3840 struct pool_workqueue *old_pwq = NULL, *pwq;
3841 struct workqueue_attrs *target_attrs;
3844 lockdep_assert_held(&wq_pool_mutex);
3846 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3847 wq->unbound_attrs->no_numa)
3851 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3852 * Let's use a preallocated one. The following buf is protected by
3853 * CPU hotplug exclusion.
3855 target_attrs = wq_update_unbound_numa_attrs_buf;
3856 cpumask = target_attrs->cpumask;
3858 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3859 pwq = unbound_pwq_by_node(wq, node);
3862 * Let's determine what needs to be done. If the target cpumask is
3863 * different from the default pwq's, we need to compare it to @pwq's
3864 * and create a new one if they don't match. If the target cpumask
3865 * equals the default pwq's, the default pwq should be used.
3867 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3868 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3874 /* create a new pwq */
3875 pwq = alloc_unbound_pwq(wq, target_attrs);
3877 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3882 /* Install the new pwq. */
3883 mutex_lock(&wq->mutex);
3884 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3888 mutex_lock(&wq->mutex);
3889 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3890 get_pwq(wq->dfl_pwq);
3891 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3892 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3894 mutex_unlock(&wq->mutex);
3895 put_pwq_unlocked(old_pwq);
3898 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3900 bool highpri = wq->flags & WQ_HIGHPRI;
3903 if (!(wq->flags & WQ_UNBOUND)) {
3904 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3908 for_each_possible_cpu(cpu) {
3909 struct pool_workqueue *pwq =
3910 per_cpu_ptr(wq->cpu_pwqs, cpu);
3911 struct worker_pool *cpu_pools =
3912 per_cpu(cpu_worker_pools, cpu);
3914 init_pwq(pwq, wq, &cpu_pools[highpri]);
3916 mutex_lock(&wq->mutex);
3918 mutex_unlock(&wq->mutex);
3921 } else if (wq->flags & __WQ_ORDERED) {
3922 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3923 /* there should only be single pwq for ordering guarantee */
3924 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3925 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3926 "ordering guarantee broken for workqueue %s\n", wq->name);
3929 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3933 static int wq_clamp_max_active(int max_active, unsigned int flags,
3936 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3938 if (max_active < 1 || max_active > lim)
3939 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3940 max_active, name, 1, lim);
3942 return clamp_val(max_active, 1, lim);
3945 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3948 struct lock_class_key *key,
3949 const char *lock_name, ...)
3951 size_t tbl_size = 0;
3953 struct workqueue_struct *wq;
3954 struct pool_workqueue *pwq;
3957 * Unbound && max_active == 1 used to imply ordered, which is no
3958 * longer the case on NUMA machines due to per-node pools. While
3959 * alloc_ordered_workqueue() is the right way to create an ordered
3960 * workqueue, keep the previous behavior to avoid subtle breakages
3963 if ((flags & WQ_UNBOUND) && max_active == 1)
3964 flags |= __WQ_ORDERED;
3966 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3967 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3968 flags |= WQ_UNBOUND;
3970 /* allocate wq and format name */
3971 if (flags & WQ_UNBOUND)
3972 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3974 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3978 if (flags & WQ_UNBOUND) {
3979 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3980 if (!wq->unbound_attrs)
3984 va_start(args, lock_name);
3985 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3988 max_active = max_active ?: WQ_DFL_ACTIVE;
3989 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3993 wq->saved_max_active = max_active;
3994 mutex_init(&wq->mutex);
3995 atomic_set(&wq->nr_pwqs_to_flush, 0);
3996 INIT_LIST_HEAD(&wq->pwqs);
3997 INIT_LIST_HEAD(&wq->flusher_queue);
3998 INIT_LIST_HEAD(&wq->flusher_overflow);
3999 INIT_LIST_HEAD(&wq->maydays);
4001 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4002 INIT_LIST_HEAD(&wq->list);
4004 if (alloc_and_link_pwqs(wq) < 0)
4008 * Workqueues which may be used during memory reclaim should
4009 * have a rescuer to guarantee forward progress.
4011 if (flags & WQ_MEM_RECLAIM) {
4012 struct worker *rescuer;
4014 rescuer = alloc_worker(NUMA_NO_NODE);
4018 rescuer->rescue_wq = wq;
4019 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4021 if (IS_ERR(rescuer->task)) {
4026 wq->rescuer = rescuer;
4027 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4028 wake_up_process(rescuer->task);
4031 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4035 * wq_pool_mutex protects global freeze state and workqueues list.
4036 * Grab it, adjust max_active and add the new @wq to workqueues
4039 mutex_lock(&wq_pool_mutex);
4041 mutex_lock(&wq->mutex);
4042 for_each_pwq(pwq, wq)
4043 pwq_adjust_max_active(pwq);
4044 mutex_unlock(&wq->mutex);
4046 list_add_tail_rcu(&wq->list, &workqueues);
4048 mutex_unlock(&wq_pool_mutex);
4053 free_workqueue_attrs(wq->unbound_attrs);
4057 destroy_workqueue(wq);
4060 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4063 * destroy_workqueue - safely terminate a workqueue
4064 * @wq: target workqueue
4066 * Safely destroy a workqueue. All work currently pending will be done first.
4068 void destroy_workqueue(struct workqueue_struct *wq)
4070 struct pool_workqueue *pwq;
4073 /* drain it before proceeding with destruction */
4074 drain_workqueue(wq);
4077 mutex_lock(&wq->mutex);
4078 for_each_pwq(pwq, wq) {
4081 for (i = 0; i < WORK_NR_COLORS; i++) {
4082 if (WARN_ON(pwq->nr_in_flight[i])) {
4083 mutex_unlock(&wq->mutex);
4084 show_workqueue_state();
4089 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4090 WARN_ON(pwq->nr_active) ||
4091 WARN_ON(!list_empty(&pwq->delayed_works))) {
4092 mutex_unlock(&wq->mutex);
4093 show_workqueue_state();
4097 mutex_unlock(&wq->mutex);
4100 * wq list is used to freeze wq, remove from list after
4101 * flushing is complete in case freeze races us.
4103 mutex_lock(&wq_pool_mutex);
4104 list_del_rcu(&wq->list);
4105 mutex_unlock(&wq_pool_mutex);
4107 workqueue_sysfs_unregister(wq);
4110 kthread_stop(wq->rescuer->task);
4112 if (!(wq->flags & WQ_UNBOUND)) {
4114 * The base ref is never dropped on per-cpu pwqs. Directly
4115 * schedule RCU free.
4117 call_rcu_sched(&wq->rcu, rcu_free_wq);
4120 * We're the sole accessor of @wq at this point. Directly
4121 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4122 * @wq will be freed when the last pwq is released.
4124 for_each_node(node) {
4125 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4126 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4127 put_pwq_unlocked(pwq);
4131 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4132 * put. Don't access it afterwards.
4136 put_pwq_unlocked(pwq);
4139 EXPORT_SYMBOL_GPL(destroy_workqueue);
4142 * workqueue_set_max_active - adjust max_active of a workqueue
4143 * @wq: target workqueue
4144 * @max_active: new max_active value.
4146 * Set max_active of @wq to @max_active.
4149 * Don't call from IRQ context.
4151 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4153 struct pool_workqueue *pwq;
4155 /* disallow meddling with max_active for ordered workqueues */
4156 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4159 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4161 mutex_lock(&wq->mutex);
4163 wq->flags &= ~__WQ_ORDERED;
4164 wq->saved_max_active = max_active;
4166 for_each_pwq(pwq, wq)
4167 pwq_adjust_max_active(pwq);
4169 mutex_unlock(&wq->mutex);
4171 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4174 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4176 * Determine whether %current is a workqueue rescuer. Can be used from
4177 * work functions to determine whether it's being run off the rescuer task.
4179 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4181 bool current_is_workqueue_rescuer(void)
4183 struct worker *worker = current_wq_worker();
4185 return worker && worker->rescue_wq;
4189 * workqueue_congested - test whether a workqueue is congested
4190 * @cpu: CPU in question
4191 * @wq: target workqueue
4193 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4194 * no synchronization around this function and the test result is
4195 * unreliable and only useful as advisory hints or for debugging.
4197 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4198 * Note that both per-cpu and unbound workqueues may be associated with
4199 * multiple pool_workqueues which have separate congested states. A
4200 * workqueue being congested on one CPU doesn't mean the workqueue is also
4201 * contested on other CPUs / NUMA nodes.
4204 * %true if congested, %false otherwise.
4206 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4208 struct pool_workqueue *pwq;
4211 rcu_read_lock_sched();
4213 if (cpu == WORK_CPU_UNBOUND)
4214 cpu = smp_processor_id();
4216 if (!(wq->flags & WQ_UNBOUND))
4217 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4219 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4221 ret = !list_empty(&pwq->delayed_works);
4222 rcu_read_unlock_sched();
4226 EXPORT_SYMBOL_GPL(workqueue_congested);
4229 * work_busy - test whether a work is currently pending or running
4230 * @work: the work to be tested
4232 * Test whether @work is currently pending or running. There is no
4233 * synchronization around this function and the test result is
4234 * unreliable and only useful as advisory hints or for debugging.
4237 * OR'd bitmask of WORK_BUSY_* bits.
4239 unsigned int work_busy(struct work_struct *work)
4241 struct worker_pool *pool;
4242 unsigned long flags;
4243 unsigned int ret = 0;
4245 if (work_pending(work))
4246 ret |= WORK_BUSY_PENDING;
4248 local_irq_save(flags);
4249 pool = get_work_pool(work);
4251 spin_lock(&pool->lock);
4252 if (find_worker_executing_work(pool, work))
4253 ret |= WORK_BUSY_RUNNING;
4254 spin_unlock(&pool->lock);
4256 local_irq_restore(flags);
4260 EXPORT_SYMBOL_GPL(work_busy);
4263 * set_worker_desc - set description for the current work item
4264 * @fmt: printf-style format string
4265 * @...: arguments for the format string
4267 * This function can be called by a running work function to describe what
4268 * the work item is about. If the worker task gets dumped, this
4269 * information will be printed out together to help debugging. The
4270 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4272 void set_worker_desc(const char *fmt, ...)
4274 struct worker *worker = current_wq_worker();
4278 va_start(args, fmt);
4279 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4281 worker->desc_valid = true;
4286 * print_worker_info - print out worker information and description
4287 * @log_lvl: the log level to use when printing
4288 * @task: target task
4290 * If @task is a worker and currently executing a work item, print out the
4291 * name of the workqueue being serviced and worker description set with
4292 * set_worker_desc() by the currently executing work item.
4294 * This function can be safely called on any task as long as the
4295 * task_struct itself is accessible. While safe, this function isn't
4296 * synchronized and may print out mixups or garbages of limited length.
4298 void print_worker_info(const char *log_lvl, struct task_struct *task)
4300 work_func_t *fn = NULL;
4301 char name[WQ_NAME_LEN] = { };
4302 char desc[WORKER_DESC_LEN] = { };
4303 struct pool_workqueue *pwq = NULL;
4304 struct workqueue_struct *wq = NULL;
4305 bool desc_valid = false;
4306 struct worker *worker;
4308 if (!(task->flags & PF_WQ_WORKER))
4312 * This function is called without any synchronization and @task
4313 * could be in any state. Be careful with dereferences.
4315 worker = kthread_probe_data(task);
4318 * Carefully copy the associated workqueue's workfn and name. Keep
4319 * the original last '\0' in case the original contains garbage.
4321 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4322 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4323 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4324 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4326 /* copy worker description */
4327 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4329 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4331 if (fn || name[0] || desc[0]) {
4332 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4334 pr_cont(" (%s)", desc);
4339 static void pr_cont_pool_info(struct worker_pool *pool)
4341 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4342 if (pool->node != NUMA_NO_NODE)
4343 pr_cont(" node=%d", pool->node);
4344 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4347 static void pr_cont_work(bool comma, struct work_struct *work)
4349 if (work->func == wq_barrier_func) {
4350 struct wq_barrier *barr;
4352 barr = container_of(work, struct wq_barrier, work);
4354 pr_cont("%s BAR(%d)", comma ? "," : "",
4355 task_pid_nr(barr->task));
4357 pr_cont("%s %pf", comma ? "," : "", work->func);
4361 static void show_pwq(struct pool_workqueue *pwq)
4363 struct worker_pool *pool = pwq->pool;
4364 struct work_struct *work;
4365 struct worker *worker;
4366 bool has_in_flight = false, has_pending = false;
4369 pr_info(" pwq %d:", pool->id);
4370 pr_cont_pool_info(pool);
4372 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4373 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4375 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4376 if (worker->current_pwq == pwq) {
4377 has_in_flight = true;
4381 if (has_in_flight) {
4384 pr_info(" in-flight:");
4385 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4386 if (worker->current_pwq != pwq)
4389 pr_cont("%s %d%s:%pf", comma ? "," : "",
4390 task_pid_nr(worker->task),
4391 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4392 worker->current_func);
4393 list_for_each_entry(work, &worker->scheduled, entry)
4394 pr_cont_work(false, work);
4400 list_for_each_entry(work, &pool->worklist, entry) {
4401 if (get_work_pwq(work) == pwq) {
4409 pr_info(" pending:");
4410 list_for_each_entry(work, &pool->worklist, entry) {
4411 if (get_work_pwq(work) != pwq)
4414 pr_cont_work(comma, work);
4415 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4420 if (!list_empty(&pwq->delayed_works)) {
4423 pr_info(" delayed:");
4424 list_for_each_entry(work, &pwq->delayed_works, entry) {
4425 pr_cont_work(comma, work);
4426 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4433 * show_workqueue_state - dump workqueue state
4435 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4436 * all busy workqueues and pools.
4438 void show_workqueue_state(void)
4440 struct workqueue_struct *wq;
4441 struct worker_pool *pool;
4442 unsigned long flags;
4445 rcu_read_lock_sched();
4447 pr_info("Showing busy workqueues and worker pools:\n");
4449 list_for_each_entry_rcu(wq, &workqueues, list) {
4450 struct pool_workqueue *pwq;
4453 for_each_pwq(pwq, wq) {
4454 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4462 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4464 for_each_pwq(pwq, wq) {
4465 spin_lock_irqsave(&pwq->pool->lock, flags);
4466 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4468 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4472 for_each_pool(pool, pi) {
4473 struct worker *worker;
4476 spin_lock_irqsave(&pool->lock, flags);
4477 if (pool->nr_workers == pool->nr_idle)
4480 pr_info("pool %d:", pool->id);
4481 pr_cont_pool_info(pool);
4482 pr_cont(" hung=%us workers=%d",
4483 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4486 pr_cont(" manager: %d",
4487 task_pid_nr(pool->manager->task));
4488 list_for_each_entry(worker, &pool->idle_list, entry) {
4489 pr_cont(" %s%d", first ? "idle: " : "",
4490 task_pid_nr(worker->task));
4495 spin_unlock_irqrestore(&pool->lock, flags);
4498 rcu_read_unlock_sched();
4504 * There are two challenges in supporting CPU hotplug. Firstly, there
4505 * are a lot of assumptions on strong associations among work, pwq and
4506 * pool which make migrating pending and scheduled works very
4507 * difficult to implement without impacting hot paths. Secondly,
4508 * worker pools serve mix of short, long and very long running works making
4509 * blocked draining impractical.
4511 * This is solved by allowing the pools to be disassociated from the CPU
4512 * running as an unbound one and allowing it to be reattached later if the
4513 * cpu comes back online.
4516 static void wq_unbind_fn(struct work_struct *work)
4518 int cpu = smp_processor_id();
4519 struct worker_pool *pool;
4520 struct worker *worker;
4522 for_each_cpu_worker_pool(pool, cpu) {
4523 mutex_lock(&pool->attach_mutex);
4524 spin_lock_irq(&pool->lock);
4527 * We've blocked all attach/detach operations. Make all workers
4528 * unbound and set DISASSOCIATED. Before this, all workers
4529 * except for the ones which are still executing works from
4530 * before the last CPU down must be on the cpu. After
4531 * this, they may become diasporas.
4533 for_each_pool_worker(worker, pool)
4534 worker->flags |= WORKER_UNBOUND;
4536 pool->flags |= POOL_DISASSOCIATED;
4538 spin_unlock_irq(&pool->lock);
4539 mutex_unlock(&pool->attach_mutex);
4542 * Call schedule() so that we cross rq->lock and thus can
4543 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4544 * This is necessary as scheduler callbacks may be invoked
4550 * Sched callbacks are disabled now. Zap nr_running.
4551 * After this, nr_running stays zero and need_more_worker()
4552 * and keep_working() are always true as long as the
4553 * worklist is not empty. This pool now behaves as an
4554 * unbound (in terms of concurrency management) pool which
4555 * are served by workers tied to the pool.
4557 atomic_set(&pool->nr_running, 0);
4560 * With concurrency management just turned off, a busy
4561 * worker blocking could lead to lengthy stalls. Kick off
4562 * unbound chain execution of currently pending work items.
4564 spin_lock_irq(&pool->lock);
4565 wake_up_worker(pool);
4566 spin_unlock_irq(&pool->lock);
4571 * rebind_workers - rebind all workers of a pool to the associated CPU
4572 * @pool: pool of interest
4574 * @pool->cpu is coming online. Rebind all workers to the CPU.
4576 static void rebind_workers(struct worker_pool *pool)
4578 struct worker *worker;
4580 lockdep_assert_held(&pool->attach_mutex);
4583 * Restore CPU affinity of all workers. As all idle workers should
4584 * be on the run-queue of the associated CPU before any local
4585 * wake-ups for concurrency management happen, restore CPU affinity
4586 * of all workers first and then clear UNBOUND. As we're called
4587 * from CPU_ONLINE, the following shouldn't fail.
4589 for_each_pool_worker(worker, pool)
4590 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4591 pool->attrs->cpumask) < 0);
4593 spin_lock_irq(&pool->lock);
4596 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4597 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4598 * being reworked and this can go away in time.
4600 if (!(pool->flags & POOL_DISASSOCIATED)) {
4601 spin_unlock_irq(&pool->lock);
4605 pool->flags &= ~POOL_DISASSOCIATED;
4607 for_each_pool_worker(worker, pool) {
4608 unsigned int worker_flags = worker->flags;
4611 * A bound idle worker should actually be on the runqueue
4612 * of the associated CPU for local wake-ups targeting it to
4613 * work. Kick all idle workers so that they migrate to the
4614 * associated CPU. Doing this in the same loop as
4615 * replacing UNBOUND with REBOUND is safe as no worker will
4616 * be bound before @pool->lock is released.
4618 if (worker_flags & WORKER_IDLE)
4619 wake_up_process(worker->task);
4622 * We want to clear UNBOUND but can't directly call
4623 * worker_clr_flags() or adjust nr_running. Atomically
4624 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4625 * @worker will clear REBOUND using worker_clr_flags() when
4626 * it initiates the next execution cycle thus restoring
4627 * concurrency management. Note that when or whether
4628 * @worker clears REBOUND doesn't affect correctness.
4630 * WRITE_ONCE() is necessary because @worker->flags may be
4631 * tested without holding any lock in
4632 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4633 * fail incorrectly leading to premature concurrency
4634 * management operations.
4636 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4637 worker_flags |= WORKER_REBOUND;
4638 worker_flags &= ~WORKER_UNBOUND;
4639 WRITE_ONCE(worker->flags, worker_flags);
4642 spin_unlock_irq(&pool->lock);
4646 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4647 * @pool: unbound pool of interest
4648 * @cpu: the CPU which is coming up
4650 * An unbound pool may end up with a cpumask which doesn't have any online
4651 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4652 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4653 * online CPU before, cpus_allowed of all its workers should be restored.
4655 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4657 static cpumask_t cpumask;
4658 struct worker *worker;
4660 lockdep_assert_held(&pool->attach_mutex);
4662 /* is @cpu allowed for @pool? */
4663 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4666 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4668 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4669 for_each_pool_worker(worker, pool)
4670 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4673 int workqueue_prepare_cpu(unsigned int cpu)
4675 struct worker_pool *pool;
4677 for_each_cpu_worker_pool(pool, cpu) {
4678 if (pool->nr_workers)
4680 if (!create_worker(pool))
4686 int workqueue_online_cpu(unsigned int cpu)
4688 struct worker_pool *pool;
4689 struct workqueue_struct *wq;
4692 mutex_lock(&wq_pool_mutex);
4694 for_each_pool(pool, pi) {
4695 mutex_lock(&pool->attach_mutex);
4697 if (pool->cpu == cpu)
4698 rebind_workers(pool);
4699 else if (pool->cpu < 0)
4700 restore_unbound_workers_cpumask(pool, cpu);
4702 mutex_unlock(&pool->attach_mutex);
4705 /* update NUMA affinity of unbound workqueues */
4706 list_for_each_entry(wq, &workqueues, list)
4707 wq_update_unbound_numa(wq, cpu, true);
4709 mutex_unlock(&wq_pool_mutex);
4713 int workqueue_offline_cpu(unsigned int cpu)
4715 struct work_struct unbind_work;
4716 struct workqueue_struct *wq;
4718 /* unbinding per-cpu workers should happen on the local CPU */
4719 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4720 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4722 /* update NUMA affinity of unbound workqueues */
4723 mutex_lock(&wq_pool_mutex);
4724 list_for_each_entry(wq, &workqueues, list)
4725 wq_update_unbound_numa(wq, cpu, false);
4726 mutex_unlock(&wq_pool_mutex);
4728 /* wait for per-cpu unbinding to finish */
4729 flush_work(&unbind_work);
4730 destroy_work_on_stack(&unbind_work);
4736 struct work_for_cpu {
4737 struct work_struct work;
4743 static void work_for_cpu_fn(struct work_struct *work)
4745 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4747 wfc->ret = wfc->fn(wfc->arg);
4751 * work_on_cpu - run a function in thread context on a particular cpu
4752 * @cpu: the cpu to run on
4753 * @fn: the function to run
4754 * @arg: the function arg
4756 * It is up to the caller to ensure that the cpu doesn't go offline.
4757 * The caller must not hold any locks which would prevent @fn from completing.
4759 * Return: The value @fn returns.
4761 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4763 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4765 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4766 schedule_work_on(cpu, &wfc.work);
4767 flush_work(&wfc.work);
4768 destroy_work_on_stack(&wfc.work);
4771 EXPORT_SYMBOL_GPL(work_on_cpu);
4774 * work_on_cpu_safe - run a function in thread context on a particular cpu
4775 * @cpu: the cpu to run on
4776 * @fn: the function to run
4777 * @arg: the function argument
4779 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4780 * any locks which would prevent @fn from completing.
4782 * Return: The value @fn returns.
4784 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4789 if (cpu_online(cpu))
4790 ret = work_on_cpu(cpu, fn, arg);
4794 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
4795 #endif /* CONFIG_SMP */
4797 #ifdef CONFIG_FREEZER
4800 * freeze_workqueues_begin - begin freezing workqueues
4802 * Start freezing workqueues. After this function returns, all freezable
4803 * workqueues will queue new works to their delayed_works list instead of
4807 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4809 void freeze_workqueues_begin(void)
4811 struct workqueue_struct *wq;
4812 struct pool_workqueue *pwq;
4814 mutex_lock(&wq_pool_mutex);
4816 WARN_ON_ONCE(workqueue_freezing);
4817 workqueue_freezing = true;
4819 list_for_each_entry(wq, &workqueues, list) {
4820 mutex_lock(&wq->mutex);
4821 for_each_pwq(pwq, wq)
4822 pwq_adjust_max_active(pwq);
4823 mutex_unlock(&wq->mutex);
4826 mutex_unlock(&wq_pool_mutex);
4830 * freeze_workqueues_busy - are freezable workqueues still busy?
4832 * Check whether freezing is complete. This function must be called
4833 * between freeze_workqueues_begin() and thaw_workqueues().
4836 * Grabs and releases wq_pool_mutex.
4839 * %true if some freezable workqueues are still busy. %false if freezing
4842 bool freeze_workqueues_busy(void)
4845 struct workqueue_struct *wq;
4846 struct pool_workqueue *pwq;
4848 mutex_lock(&wq_pool_mutex);
4850 WARN_ON_ONCE(!workqueue_freezing);
4852 list_for_each_entry(wq, &workqueues, list) {
4853 if (!(wq->flags & WQ_FREEZABLE))
4856 * nr_active is monotonically decreasing. It's safe
4857 * to peek without lock.
4859 rcu_read_lock_sched();
4860 for_each_pwq(pwq, wq) {
4861 WARN_ON_ONCE(pwq->nr_active < 0);
4862 if (pwq->nr_active) {
4864 rcu_read_unlock_sched();
4868 rcu_read_unlock_sched();
4871 mutex_unlock(&wq_pool_mutex);
4876 * thaw_workqueues - thaw workqueues
4878 * Thaw workqueues. Normal queueing is restored and all collected
4879 * frozen works are transferred to their respective pool worklists.
4882 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4884 void thaw_workqueues(void)
4886 struct workqueue_struct *wq;
4887 struct pool_workqueue *pwq;
4889 mutex_lock(&wq_pool_mutex);
4891 if (!workqueue_freezing)
4894 workqueue_freezing = false;
4896 /* restore max_active and repopulate worklist */
4897 list_for_each_entry(wq, &workqueues, list) {
4898 mutex_lock(&wq->mutex);
4899 for_each_pwq(pwq, wq)
4900 pwq_adjust_max_active(pwq);
4901 mutex_unlock(&wq->mutex);
4905 mutex_unlock(&wq_pool_mutex);
4907 #endif /* CONFIG_FREEZER */
4909 static int workqueue_apply_unbound_cpumask(void)
4913 struct workqueue_struct *wq;
4914 struct apply_wqattrs_ctx *ctx, *n;
4916 lockdep_assert_held(&wq_pool_mutex);
4918 list_for_each_entry(wq, &workqueues, list) {
4919 if (!(wq->flags & WQ_UNBOUND))
4921 /* creating multiple pwqs breaks ordering guarantee */
4922 if (wq->flags & __WQ_ORDERED)
4925 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4931 list_add_tail(&ctx->list, &ctxs);
4934 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4936 apply_wqattrs_commit(ctx);
4937 apply_wqattrs_cleanup(ctx);
4944 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4945 * @cpumask: the cpumask to set
4947 * The low-level workqueues cpumask is a global cpumask that limits
4948 * the affinity of all unbound workqueues. This function check the @cpumask
4949 * and apply it to all unbound workqueues and updates all pwqs of them.
4951 * Retun: 0 - Success
4952 * -EINVAL - Invalid @cpumask
4953 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4955 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4958 cpumask_var_t saved_cpumask;
4960 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4963 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4964 if (!cpumask_empty(cpumask)) {
4965 apply_wqattrs_lock();
4967 /* save the old wq_unbound_cpumask. */
4968 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4970 /* update wq_unbound_cpumask at first and apply it to wqs. */
4971 cpumask_copy(wq_unbound_cpumask, cpumask);
4972 ret = workqueue_apply_unbound_cpumask();
4974 /* restore the wq_unbound_cpumask when failed. */
4976 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4978 apply_wqattrs_unlock();
4981 free_cpumask_var(saved_cpumask);
4987 * Workqueues with WQ_SYSFS flag set is visible to userland via
4988 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4989 * following attributes.
4991 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4992 * max_active RW int : maximum number of in-flight work items
4994 * Unbound workqueues have the following extra attributes.
4996 * id RO int : the associated pool ID
4997 * nice RW int : nice value of the workers
4998 * cpumask RW mask : bitmask of allowed CPUs for the workers
5001 struct workqueue_struct *wq;
5005 static struct workqueue_struct *dev_to_wq(struct device *dev)
5007 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5012 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5015 struct workqueue_struct *wq = dev_to_wq(dev);
5017 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5019 static DEVICE_ATTR_RO(per_cpu);
5021 static ssize_t max_active_show(struct device *dev,
5022 struct device_attribute *attr, char *buf)
5024 struct workqueue_struct *wq = dev_to_wq(dev);
5026 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5029 static ssize_t max_active_store(struct device *dev,
5030 struct device_attribute *attr, const char *buf,
5033 struct workqueue_struct *wq = dev_to_wq(dev);
5036 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5039 workqueue_set_max_active(wq, val);
5042 static DEVICE_ATTR_RW(max_active);
5044 static struct attribute *wq_sysfs_attrs[] = {
5045 &dev_attr_per_cpu.attr,
5046 &dev_attr_max_active.attr,
5049 ATTRIBUTE_GROUPS(wq_sysfs);
5051 static ssize_t wq_pool_ids_show(struct device *dev,
5052 struct device_attribute *attr, char *buf)
5054 struct workqueue_struct *wq = dev_to_wq(dev);
5055 const char *delim = "";
5056 int node, written = 0;
5058 rcu_read_lock_sched();
5059 for_each_node(node) {
5060 written += scnprintf(buf + written, PAGE_SIZE - written,
5061 "%s%d:%d", delim, node,
5062 unbound_pwq_by_node(wq, node)->pool->id);
5065 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5066 rcu_read_unlock_sched();
5071 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5074 struct workqueue_struct *wq = dev_to_wq(dev);
5077 mutex_lock(&wq->mutex);
5078 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5079 mutex_unlock(&wq->mutex);
5084 /* prepare workqueue_attrs for sysfs store operations */
5085 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5087 struct workqueue_attrs *attrs;
5089 lockdep_assert_held(&wq_pool_mutex);
5091 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5095 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5099 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5100 const char *buf, size_t count)
5102 struct workqueue_struct *wq = dev_to_wq(dev);
5103 struct workqueue_attrs *attrs;
5106 apply_wqattrs_lock();
5108 attrs = wq_sysfs_prep_attrs(wq);
5112 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5113 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5114 ret = apply_workqueue_attrs_locked(wq, attrs);
5119 apply_wqattrs_unlock();
5120 free_workqueue_attrs(attrs);
5121 return ret ?: count;
5124 static ssize_t wq_cpumask_show(struct device *dev,
5125 struct device_attribute *attr, char *buf)
5127 struct workqueue_struct *wq = dev_to_wq(dev);
5130 mutex_lock(&wq->mutex);
5131 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5132 cpumask_pr_args(wq->unbound_attrs->cpumask));
5133 mutex_unlock(&wq->mutex);
5137 static ssize_t wq_cpumask_store(struct device *dev,
5138 struct device_attribute *attr,
5139 const char *buf, size_t count)
5141 struct workqueue_struct *wq = dev_to_wq(dev);
5142 struct workqueue_attrs *attrs;
5145 apply_wqattrs_lock();
5147 attrs = wq_sysfs_prep_attrs(wq);
5151 ret = cpumask_parse(buf, attrs->cpumask);
5153 ret = apply_workqueue_attrs_locked(wq, attrs);
5156 apply_wqattrs_unlock();
5157 free_workqueue_attrs(attrs);
5158 return ret ?: count;
5161 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5164 struct workqueue_struct *wq = dev_to_wq(dev);
5167 mutex_lock(&wq->mutex);
5168 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5169 !wq->unbound_attrs->no_numa);
5170 mutex_unlock(&wq->mutex);
5175 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5176 const char *buf, size_t count)
5178 struct workqueue_struct *wq = dev_to_wq(dev);
5179 struct workqueue_attrs *attrs;
5180 int v, ret = -ENOMEM;
5182 apply_wqattrs_lock();
5184 attrs = wq_sysfs_prep_attrs(wq);
5189 if (sscanf(buf, "%d", &v) == 1) {
5190 attrs->no_numa = !v;
5191 ret = apply_workqueue_attrs_locked(wq, attrs);
5195 apply_wqattrs_unlock();
5196 free_workqueue_attrs(attrs);
5197 return ret ?: count;
5200 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5201 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5202 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5203 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5204 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5208 static struct bus_type wq_subsys = {
5209 .name = "workqueue",
5210 .dev_groups = wq_sysfs_groups,
5213 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5214 struct device_attribute *attr, char *buf)
5218 mutex_lock(&wq_pool_mutex);
5219 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5220 cpumask_pr_args(wq_unbound_cpumask));
5221 mutex_unlock(&wq_pool_mutex);
5226 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5227 struct device_attribute *attr, const char *buf, size_t count)
5229 cpumask_var_t cpumask;
5232 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5235 ret = cpumask_parse(buf, cpumask);
5237 ret = workqueue_set_unbound_cpumask(cpumask);
5239 free_cpumask_var(cpumask);
5240 return ret ? ret : count;
5243 static struct device_attribute wq_sysfs_cpumask_attr =
5244 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5245 wq_unbound_cpumask_store);
5247 static int __init wq_sysfs_init(void)
5251 err = subsys_virtual_register(&wq_subsys, NULL);
5255 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5257 core_initcall(wq_sysfs_init);
5259 static void wq_device_release(struct device *dev)
5261 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5267 * workqueue_sysfs_register - make a workqueue visible in sysfs
5268 * @wq: the workqueue to register
5270 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5271 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5272 * which is the preferred method.
5274 * Workqueue user should use this function directly iff it wants to apply
5275 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5276 * apply_workqueue_attrs() may race against userland updating the
5279 * Return: 0 on success, -errno on failure.
5281 int workqueue_sysfs_register(struct workqueue_struct *wq)
5283 struct wq_device *wq_dev;
5287 * Adjusting max_active or creating new pwqs by applying
5288 * attributes breaks ordering guarantee. Disallow exposing ordered
5291 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5294 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5299 wq_dev->dev.bus = &wq_subsys;
5300 wq_dev->dev.release = wq_device_release;
5301 dev_set_name(&wq_dev->dev, "%s", wq->name);
5304 * unbound_attrs are created separately. Suppress uevent until
5305 * everything is ready.
5307 dev_set_uevent_suppress(&wq_dev->dev, true);
5309 ret = device_register(&wq_dev->dev);
5316 if (wq->flags & WQ_UNBOUND) {
5317 struct device_attribute *attr;
5319 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5320 ret = device_create_file(&wq_dev->dev, attr);
5322 device_unregister(&wq_dev->dev);
5329 dev_set_uevent_suppress(&wq_dev->dev, false);
5330 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5335 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5336 * @wq: the workqueue to unregister
5338 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5340 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5342 struct wq_device *wq_dev = wq->wq_dev;
5348 device_unregister(&wq_dev->dev);
5350 #else /* CONFIG_SYSFS */
5351 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5352 #endif /* CONFIG_SYSFS */
5355 * Workqueue watchdog.
5357 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5358 * flush dependency, a concurrency managed work item which stays RUNNING
5359 * indefinitely. Workqueue stalls can be very difficult to debug as the
5360 * usual warning mechanisms don't trigger and internal workqueue state is
5363 * Workqueue watchdog monitors all worker pools periodically and dumps
5364 * state if some pools failed to make forward progress for a while where
5365 * forward progress is defined as the first item on ->worklist changing.
5367 * This mechanism is controlled through the kernel parameter
5368 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5369 * corresponding sysfs parameter file.
5371 #ifdef CONFIG_WQ_WATCHDOG
5373 static void wq_watchdog_timer_fn(unsigned long data);
5375 static unsigned long wq_watchdog_thresh = 30;
5376 static struct timer_list wq_watchdog_timer =
5377 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5379 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5380 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5382 static void wq_watchdog_reset_touched(void)
5386 wq_watchdog_touched = jiffies;
5387 for_each_possible_cpu(cpu)
5388 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5391 static void wq_watchdog_timer_fn(unsigned long data)
5393 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5394 bool lockup_detected = false;
5395 struct worker_pool *pool;
5403 for_each_pool(pool, pi) {
5404 unsigned long pool_ts, touched, ts;
5406 if (list_empty(&pool->worklist))
5409 /* get the latest of pool and touched timestamps */
5410 pool_ts = READ_ONCE(pool->watchdog_ts);
5411 touched = READ_ONCE(wq_watchdog_touched);
5413 if (time_after(pool_ts, touched))
5418 if (pool->cpu >= 0) {
5419 unsigned long cpu_touched =
5420 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5422 if (time_after(cpu_touched, ts))
5427 if (time_after(jiffies, ts + thresh)) {
5428 lockup_detected = true;
5429 pr_emerg("BUG: workqueue lockup - pool");
5430 pr_cont_pool_info(pool);
5431 pr_cont(" stuck for %us!\n",
5432 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5438 if (lockup_detected)
5439 show_workqueue_state();
5441 wq_watchdog_reset_touched();
5442 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5445 void wq_watchdog_touch(int cpu)
5448 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5450 wq_watchdog_touched = jiffies;
5453 static void wq_watchdog_set_thresh(unsigned long thresh)
5455 wq_watchdog_thresh = 0;
5456 del_timer_sync(&wq_watchdog_timer);
5459 wq_watchdog_thresh = thresh;
5460 wq_watchdog_reset_touched();
5461 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5465 static int wq_watchdog_param_set_thresh(const char *val,
5466 const struct kernel_param *kp)
5468 unsigned long thresh;
5471 ret = kstrtoul(val, 0, &thresh);
5476 wq_watchdog_set_thresh(thresh);
5478 wq_watchdog_thresh = thresh;
5483 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5484 .set = wq_watchdog_param_set_thresh,
5485 .get = param_get_ulong,
5488 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5491 static void wq_watchdog_init(void)
5493 wq_watchdog_set_thresh(wq_watchdog_thresh);
5496 #else /* CONFIG_WQ_WATCHDOG */
5498 static inline void wq_watchdog_init(void) { }
5500 #endif /* CONFIG_WQ_WATCHDOG */
5502 static void __init wq_numa_init(void)
5507 if (num_possible_nodes() <= 1)
5510 if (wq_disable_numa) {
5511 pr_info("workqueue: NUMA affinity support disabled\n");
5515 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5516 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5519 * We want masks of possible CPUs of each node which isn't readily
5520 * available. Build one from cpu_to_node() which should have been
5521 * fully initialized by now.
5523 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5527 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5528 node_online(node) ? node : NUMA_NO_NODE));
5530 for_each_possible_cpu(cpu) {
5531 node = cpu_to_node(cpu);
5532 if (WARN_ON(node == NUMA_NO_NODE)) {
5533 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5534 /* happens iff arch is bonkers, let's just proceed */
5537 cpumask_set_cpu(cpu, tbl[node]);
5540 wq_numa_possible_cpumask = tbl;
5541 wq_numa_enabled = true;
5545 * workqueue_init_early - early init for workqueue subsystem
5547 * This is the first half of two-staged workqueue subsystem initialization
5548 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5549 * idr are up. It sets up all the data structures and system workqueues
5550 * and allows early boot code to create workqueues and queue/cancel work
5551 * items. Actual work item execution starts only after kthreads can be
5552 * created and scheduled right before early initcalls.
5554 int __init workqueue_init_early(void)
5556 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5559 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5561 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5562 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5564 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5566 /* initialize CPU pools */
5567 for_each_possible_cpu(cpu) {
5568 struct worker_pool *pool;
5571 for_each_cpu_worker_pool(pool, cpu) {
5572 BUG_ON(init_worker_pool(pool));
5574 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5575 pool->attrs->nice = std_nice[i++];
5576 pool->node = cpu_to_node(cpu);
5579 mutex_lock(&wq_pool_mutex);
5580 BUG_ON(worker_pool_assign_id(pool));
5581 mutex_unlock(&wq_pool_mutex);
5585 /* create default unbound and ordered wq attrs */
5586 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5587 struct workqueue_attrs *attrs;
5589 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5590 attrs->nice = std_nice[i];
5591 unbound_std_wq_attrs[i] = attrs;
5594 * An ordered wq should have only one pwq as ordering is
5595 * guaranteed by max_active which is enforced by pwqs.
5596 * Turn off NUMA so that dfl_pwq is used for all nodes.
5598 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5599 attrs->nice = std_nice[i];
5600 attrs->no_numa = true;
5601 ordered_wq_attrs[i] = attrs;
5604 system_wq = alloc_workqueue("events", 0, 0);
5605 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5606 system_long_wq = alloc_workqueue("events_long", 0, 0);
5607 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5608 WQ_UNBOUND_MAX_ACTIVE);
5609 system_freezable_wq = alloc_workqueue("events_freezable",
5611 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5612 WQ_POWER_EFFICIENT, 0);
5613 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5614 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5616 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5617 !system_unbound_wq || !system_freezable_wq ||
5618 !system_power_efficient_wq ||
5619 !system_freezable_power_efficient_wq);
5625 * workqueue_init - bring workqueue subsystem fully online
5627 * This is the latter half of two-staged workqueue subsystem initialization
5628 * and invoked as soon as kthreads can be created and scheduled.
5629 * Workqueues have been created and work items queued on them, but there
5630 * are no kworkers executing the work items yet. Populate the worker pools
5631 * with the initial workers and enable future kworker creations.
5633 int __init workqueue_init(void)
5635 struct workqueue_struct *wq;
5636 struct worker_pool *pool;
5640 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5641 * CPU to node mapping may not be available that early on some
5642 * archs such as power and arm64. As per-cpu pools created
5643 * previously could be missing node hint and unbound pools NUMA
5644 * affinity, fix them up.
5648 mutex_lock(&wq_pool_mutex);
5650 for_each_possible_cpu(cpu) {
5651 for_each_cpu_worker_pool(pool, cpu) {
5652 pool->node = cpu_to_node(cpu);
5656 list_for_each_entry(wq, &workqueues, list)
5657 wq_update_unbound_numa(wq, smp_processor_id(), true);
5659 mutex_unlock(&wq_pool_mutex);
5661 /* create the initial workers */
5662 for_each_online_cpu(cpu) {
5663 for_each_cpu_worker_pool(pool, cpu) {
5664 pool->flags &= ~POOL_DISASSOCIATED;
5665 BUG_ON(!create_worker(pool));
5669 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5670 BUG_ON(!create_worker(pool));