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/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
50 #include <linux/sched/isolation.h>
51 #include <linux/nmi.h>
53 #include "workqueue_internal.h"
59 * A bound pool is either associated or disassociated with its CPU.
60 * While associated (!DISASSOCIATED), all workers are bound to the
61 * CPU and none has %WORKER_UNBOUND set and concurrency management
64 * While DISASSOCIATED, the cpu may be offline and all workers have
65 * %WORKER_UNBOUND set and concurrency management disabled, and may
66 * be executing on any CPU. The pool behaves as an unbound one.
68 * Note that DISASSOCIATED should be flipped only while holding
69 * wq_pool_attach_mutex to avoid changing binding state while
70 * worker_attach_to_pool() is in progress.
72 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
73 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
76 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
78 WORKER_PREP = 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give MIN_NICE.
104 RESCUER_NICE_LEVEL = MIN_NICE,
105 HIGHPRI_NICE_LEVEL = MIN_NICE,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * A: wq_pool_attach_mutex protected.
128 * PL: wq_pool_mutex protected.
130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
135 * sched-RCU for reads.
137 * WQ: wq->mutex protected.
139 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
141 * MD: wq_mayday_lock protected.
144 /* struct worker is defined in workqueue_internal.h */
147 spinlock_t lock; /* the pool lock */
148 int cpu; /* I: the associated cpu */
149 int node; /* I: the associated node ID */
150 int id; /* I: pool ID */
151 unsigned int flags; /* X: flags */
153 unsigned long watchdog_ts; /* L: watchdog timestamp */
155 struct list_head worklist; /* L: list of pending works */
157 int nr_workers; /* L: total number of workers */
158 int nr_idle; /* L: currently idle workers */
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 struct worker *manager; /* L: purely informational */
169 struct list_head workers; /* A: attached workers */
170 struct completion *detach_completion; /* all workers detached */
172 struct ida worker_ida; /* worker IDs for task name */
174 struct workqueue_attrs *attrs; /* I: worker attributes */
175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
176 int refcnt; /* PL: refcnt for unbound pools */
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
183 atomic_t nr_running ____cacheline_aligned_in_smp;
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
190 } ____cacheline_aligned_in_smp;
193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
198 struct pool_workqueue {
199 struct worker_pool *pool; /* I: the associated pool */
200 struct workqueue_struct *wq; /* I: the owning workqueue */
201 int work_color; /* L: current color */
202 int flush_color; /* L: flushing color */
203 int refcnt; /* L: reference count */
204 int nr_in_flight[WORK_NR_COLORS];
205 /* L: nr of in_flight works */
206 int nr_active; /* L: nr of active works */
207 int max_active; /* L: max active works */
208 struct list_head delayed_works; /* L: delayed works */
209 struct list_head pwqs_node; /* WR: node on wq->pwqs */
210 struct list_head mayday_node; /* MD: node on wq->maydays */
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
216 * determined without grabbing wq->mutex.
218 struct work_struct unbound_release_work;
220 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
223 * Structure used to wait for workqueue flush.
226 struct list_head list; /* WQ: list of flushers */
227 int flush_color; /* WQ: flush color waiting for */
228 struct completion done; /* flush completion */
234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
237 struct workqueue_struct {
238 struct list_head pwqs; /* WR: all pwqs of this wq */
239 struct list_head list; /* PR: list of all workqueues */
241 struct mutex mutex; /* protects this wq */
242 int work_color; /* WQ: current work color */
243 int flush_color; /* WQ: current flush color */
244 atomic_t nr_pwqs_to_flush; /* flush in progress */
245 struct wq_flusher *first_flusher; /* WQ: first flusher */
246 struct list_head flusher_queue; /* WQ: flush waiters */
247 struct list_head flusher_overflow; /* WQ: flush overflow list */
249 struct list_head maydays; /* MD: pwqs requesting rescue */
250 struct worker *rescuer; /* I: rescue worker */
252 int nr_drainers; /* WQ: drain in progress */
253 int saved_max_active; /* WQ: saved pwq max_active */
255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259 struct wq_device *wq_dev; /* I: for sysfs interface */
261 #ifdef CONFIG_LOCKDEP
262 struct lockdep_map lockdep_map;
264 char name[WQ_NAME_LEN]; /* I: workqueue name */
267 * Destruction of workqueue_struct is sched-RCU protected to allow
268 * walking the workqueues list without grabbing wq_pool_mutex.
269 * This is used to dump all workqueues from sysrq.
273 /* hot fields used during command issue, aligned to cacheline */
274 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
275 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
276 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279 static struct kmem_cache *pwq_cache;
281 static cpumask_var_t *wq_numa_possible_cpumask;
282 /* possible CPUs of each node */
284 static bool wq_disable_numa;
285 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
287 /* see the comment above the definition of WQ_POWER_EFFICIENT */
288 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
289 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
291 static bool wq_online; /* can kworkers be created yet? */
293 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
295 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
298 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
300 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
301 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
303 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
304 static bool workqueue_freezing; /* PL: have wqs started freezing? */
306 /* PL: allowable cpus for unbound wqs and work items */
307 static cpumask_var_t wq_unbound_cpumask;
309 /* CPU where unbound work was last round robin scheduled from this CPU */
310 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
313 * Local execution of unbound work items is no longer guaranteed. The
314 * following always forces round-robin CPU selection on unbound work items
315 * to uncover usages which depend on it.
317 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
318 static bool wq_debug_force_rr_cpu = true;
320 static bool wq_debug_force_rr_cpu = false;
322 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
324 /* the per-cpu worker pools */
325 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
327 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
329 /* PL: hash of all unbound pools keyed by pool->attrs */
330 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
332 /* I: attributes used when instantiating standard unbound pools on demand */
333 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
335 /* I: attributes used when instantiating ordered pools on demand */
336 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
338 struct workqueue_struct *system_wq __read_mostly;
339 EXPORT_SYMBOL(system_wq);
340 struct workqueue_struct *system_highpri_wq __read_mostly;
341 EXPORT_SYMBOL_GPL(system_highpri_wq);
342 struct workqueue_struct *system_long_wq __read_mostly;
343 EXPORT_SYMBOL_GPL(system_long_wq);
344 struct workqueue_struct *system_unbound_wq __read_mostly;
345 EXPORT_SYMBOL_GPL(system_unbound_wq);
346 struct workqueue_struct *system_freezable_wq __read_mostly;
347 EXPORT_SYMBOL_GPL(system_freezable_wq);
348 struct workqueue_struct *system_power_efficient_wq __read_mostly;
349 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
350 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
351 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
353 static int worker_thread(void *__worker);
354 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
356 #define CREATE_TRACE_POINTS
357 #include <trace/events/workqueue.h>
359 #define assert_rcu_or_pool_mutex() \
360 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
361 !lockdep_is_held(&wq_pool_mutex), \
362 "sched RCU or wq_pool_mutex should be held")
364 #define assert_rcu_or_wq_mutex(wq) \
365 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
366 !lockdep_is_held(&wq->mutex), \
367 "sched RCU or wq->mutex should be held")
369 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
370 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
371 !lockdep_is_held(&wq->mutex) && \
372 !lockdep_is_held(&wq_pool_mutex), \
373 "sched RCU, wq->mutex or wq_pool_mutex should be held")
375 #define for_each_cpu_worker_pool(pool, cpu) \
376 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
377 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
381 * for_each_pool - iterate through all worker_pools in the system
382 * @pool: iteration cursor
383 * @pi: integer used for iteration
385 * This must be called either with wq_pool_mutex held or sched RCU read
386 * locked. If the pool needs to be used beyond the locking in effect, the
387 * caller is responsible for guaranteeing that the pool stays online.
389 * The if/else clause exists only for the lockdep assertion and can be
392 #define for_each_pool(pool, pi) \
393 idr_for_each_entry(&worker_pool_idr, pool, pi) \
394 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
398 * for_each_pool_worker - iterate through all workers of a worker_pool
399 * @worker: iteration cursor
400 * @pool: worker_pool to iterate workers of
402 * This must be called with wq_pool_attach_mutex.
404 * The if/else clause exists only for the lockdep assertion and can be
407 #define for_each_pool_worker(worker, pool) \
408 list_for_each_entry((worker), &(pool)->workers, node) \
409 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
413 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
414 * @pwq: iteration cursor
415 * @wq: the target workqueue
417 * This must be called either with wq->mutex held or sched RCU read locked.
418 * If the pwq needs to be used beyond the locking in effect, the caller is
419 * responsible for guaranteeing that the pwq stays online.
421 * The if/else clause exists only for the lockdep assertion and can be
424 #define for_each_pwq(pwq, wq) \
425 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
426 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
429 #ifdef CONFIG_DEBUG_OBJECTS_WORK
431 static struct debug_obj_descr work_debug_descr;
433 static void *work_debug_hint(void *addr)
435 return ((struct work_struct *) addr)->func;
438 static bool work_is_static_object(void *addr)
440 struct work_struct *work = addr;
442 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
446 * fixup_init is called when:
447 * - an active object is initialized
449 static bool work_fixup_init(void *addr, enum debug_obj_state state)
451 struct work_struct *work = addr;
454 case ODEBUG_STATE_ACTIVE:
455 cancel_work_sync(work);
456 debug_object_init(work, &work_debug_descr);
464 * fixup_free is called when:
465 * - an active object is freed
467 static bool work_fixup_free(void *addr, enum debug_obj_state state)
469 struct work_struct *work = addr;
472 case ODEBUG_STATE_ACTIVE:
473 cancel_work_sync(work);
474 debug_object_free(work, &work_debug_descr);
481 static struct debug_obj_descr work_debug_descr = {
482 .name = "work_struct",
483 .debug_hint = work_debug_hint,
484 .is_static_object = work_is_static_object,
485 .fixup_init = work_fixup_init,
486 .fixup_free = work_fixup_free,
489 static inline void debug_work_activate(struct work_struct *work)
491 debug_object_activate(work, &work_debug_descr);
494 static inline void debug_work_deactivate(struct work_struct *work)
496 debug_object_deactivate(work, &work_debug_descr);
499 void __init_work(struct work_struct *work, int onstack)
502 debug_object_init_on_stack(work, &work_debug_descr);
504 debug_object_init(work, &work_debug_descr);
506 EXPORT_SYMBOL_GPL(__init_work);
508 void destroy_work_on_stack(struct work_struct *work)
510 debug_object_free(work, &work_debug_descr);
512 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
514 void destroy_delayed_work_on_stack(struct delayed_work *work)
516 destroy_timer_on_stack(&work->timer);
517 debug_object_free(&work->work, &work_debug_descr);
519 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
522 static inline void debug_work_activate(struct work_struct *work) { }
523 static inline void debug_work_deactivate(struct work_struct *work) { }
527 * worker_pool_assign_id - allocate ID and assing it to @pool
528 * @pool: the pool pointer of interest
530 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
531 * successfully, -errno on failure.
533 static int worker_pool_assign_id(struct worker_pool *pool)
537 lockdep_assert_held(&wq_pool_mutex);
539 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
549 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
550 * @wq: the target workqueue
553 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
555 * If the pwq needs to be used beyond the locking in effect, the caller is
556 * responsible for guaranteeing that the pwq stays online.
558 * Return: The unbound pool_workqueue for @node.
560 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
563 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
566 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
567 * delayed item is pending. The plan is to keep CPU -> NODE
568 * mapping valid and stable across CPU on/offlines. Once that
569 * happens, this workaround can be removed.
571 if (unlikely(node == NUMA_NO_NODE))
574 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
577 static unsigned int work_color_to_flags(int color)
579 return color << WORK_STRUCT_COLOR_SHIFT;
582 static int get_work_color(struct work_struct *work)
584 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
585 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
588 static int work_next_color(int color)
590 return (color + 1) % WORK_NR_COLORS;
594 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
595 * contain the pointer to the queued pwq. Once execution starts, the flag
596 * is cleared and the high bits contain OFFQ flags and pool ID.
598 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
599 * and clear_work_data() can be used to set the pwq, pool or clear
600 * work->data. These functions should only be called while the work is
601 * owned - ie. while the PENDING bit is set.
603 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
604 * corresponding to a work. Pool is available once the work has been
605 * queued anywhere after initialization until it is sync canceled. pwq is
606 * available only while the work item is queued.
608 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
609 * canceled. While being canceled, a work item may have its PENDING set
610 * but stay off timer and worklist for arbitrarily long and nobody should
611 * try to steal the PENDING bit.
613 static inline void set_work_data(struct work_struct *work, unsigned long data,
616 WARN_ON_ONCE(!work_pending(work));
617 atomic_long_set(&work->data, data | flags | work_static(work));
620 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
621 unsigned long extra_flags)
623 set_work_data(work, (unsigned long)pwq,
624 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
627 static void set_work_pool_and_keep_pending(struct work_struct *work,
630 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
631 WORK_STRUCT_PENDING);
634 static void set_work_pool_and_clear_pending(struct work_struct *work,
638 * The following wmb is paired with the implied mb in
639 * test_and_set_bit(PENDING) and ensures all updates to @work made
640 * here are visible to and precede any updates by the next PENDING
644 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
646 * The following mb guarantees that previous clear of a PENDING bit
647 * will not be reordered with any speculative LOADS or STORES from
648 * work->current_func, which is executed afterwards. This possible
649 * reordering can lead to a missed execution on attempt to qeueue
650 * the same @work. E.g. consider this case:
653 * ---------------------------- --------------------------------
655 * 1 STORE event_indicated
656 * 2 queue_work_on() {
657 * 3 test_and_set_bit(PENDING)
658 * 4 } set_..._and_clear_pending() {
659 * 5 set_work_data() # clear bit
661 * 7 work->current_func() {
662 * 8 LOAD event_indicated
665 * Without an explicit full barrier speculative LOAD on line 8 can
666 * be executed before CPU#0 does STORE on line 1. If that happens,
667 * CPU#0 observes the PENDING bit is still set and new execution of
668 * a @work is not queued in a hope, that CPU#1 will eventually
669 * finish the queued @work. Meanwhile CPU#1 does not see
670 * event_indicated is set, because speculative LOAD was executed
671 * before actual STORE.
676 static void clear_work_data(struct work_struct *work)
678 smp_wmb(); /* see set_work_pool_and_clear_pending() */
679 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
682 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
684 unsigned long data = atomic_long_read(&work->data);
686 if (data & WORK_STRUCT_PWQ)
687 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
693 * get_work_pool - return the worker_pool a given work was associated with
694 * @work: the work item of interest
696 * Pools are created and destroyed under wq_pool_mutex, and allows read
697 * access under sched-RCU read lock. As such, this function should be
698 * called under wq_pool_mutex or with preemption disabled.
700 * All fields of the returned pool are accessible as long as the above
701 * mentioned locking is in effect. If the returned pool needs to be used
702 * beyond the critical section, the caller is responsible for ensuring the
703 * returned pool is and stays online.
705 * Return: The worker_pool @work was last associated with. %NULL if none.
707 static struct worker_pool *get_work_pool(struct work_struct *work)
709 unsigned long data = atomic_long_read(&work->data);
712 assert_rcu_or_pool_mutex();
714 if (data & WORK_STRUCT_PWQ)
715 return ((struct pool_workqueue *)
716 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
718 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
719 if (pool_id == WORK_OFFQ_POOL_NONE)
722 return idr_find(&worker_pool_idr, pool_id);
726 * get_work_pool_id - return the worker pool ID a given work is associated with
727 * @work: the work item of interest
729 * Return: The worker_pool ID @work was last associated with.
730 * %WORK_OFFQ_POOL_NONE if none.
732 static int get_work_pool_id(struct work_struct *work)
734 unsigned long data = atomic_long_read(&work->data);
736 if (data & WORK_STRUCT_PWQ)
737 return ((struct pool_workqueue *)
738 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
740 return data >> WORK_OFFQ_POOL_SHIFT;
743 static void mark_work_canceling(struct work_struct *work)
745 unsigned long pool_id = get_work_pool_id(work);
747 pool_id <<= WORK_OFFQ_POOL_SHIFT;
748 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
751 static bool work_is_canceling(struct work_struct *work)
753 unsigned long data = atomic_long_read(&work->data);
755 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
759 * Policy functions. These define the policies on how the global worker
760 * pools are managed. Unless noted otherwise, these functions assume that
761 * they're being called with pool->lock held.
764 static bool __need_more_worker(struct worker_pool *pool)
766 return !atomic_read(&pool->nr_running);
770 * Need to wake up a worker? Called from anything but currently
773 * Note that, because unbound workers never contribute to nr_running, this
774 * function will always return %true for unbound pools as long as the
775 * worklist isn't empty.
777 static bool need_more_worker(struct worker_pool *pool)
779 return !list_empty(&pool->worklist) && __need_more_worker(pool);
782 /* Can I start working? Called from busy but !running workers. */
783 static bool may_start_working(struct worker_pool *pool)
785 return pool->nr_idle;
788 /* Do I need to keep working? Called from currently running workers. */
789 static bool keep_working(struct worker_pool *pool)
791 return !list_empty(&pool->worklist) &&
792 atomic_read(&pool->nr_running) <= 1;
795 /* Do we need a new worker? Called from manager. */
796 static bool need_to_create_worker(struct worker_pool *pool)
798 return need_more_worker(pool) && !may_start_working(pool);
801 /* Do we have too many workers and should some go away? */
802 static bool too_many_workers(struct worker_pool *pool)
804 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
805 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
806 int nr_busy = pool->nr_workers - nr_idle;
808 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
815 /* Return the first idle worker. Safe with preemption disabled */
816 static struct worker *first_idle_worker(struct worker_pool *pool)
818 if (unlikely(list_empty(&pool->idle_list)))
821 return list_first_entry(&pool->idle_list, struct worker, entry);
825 * wake_up_worker - wake up an idle worker
826 * @pool: worker pool to wake worker from
828 * Wake up the first idle worker of @pool.
831 * spin_lock_irq(pool->lock).
833 static void wake_up_worker(struct worker_pool *pool)
835 struct worker *worker = first_idle_worker(pool);
838 wake_up_process(worker->task);
842 * wq_worker_waking_up - a worker is waking up
843 * @task: task waking up
844 * @cpu: CPU @task is waking up to
846 * This function is called during try_to_wake_up() when a worker is
850 * spin_lock_irq(rq->lock)
852 void wq_worker_waking_up(struct task_struct *task, int cpu)
854 struct worker *worker = kthread_data(task);
856 if (!(worker->flags & WORKER_NOT_RUNNING)) {
857 WARN_ON_ONCE(worker->pool->cpu != cpu);
858 atomic_inc(&worker->pool->nr_running);
863 * wq_worker_sleeping - a worker is going to sleep
864 * @task: task going to sleep
866 * This function is called during schedule() when a busy worker is
867 * going to sleep. Worker on the same cpu can be woken up by
868 * returning pointer to its task.
871 * spin_lock_irq(rq->lock)
874 * Worker task on @cpu to wake up, %NULL if none.
876 struct task_struct *wq_worker_sleeping(struct task_struct *task)
878 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
879 struct worker_pool *pool;
882 * Rescuers, which may not have all the fields set up like normal
883 * workers, also reach here, let's not access anything before
884 * checking NOT_RUNNING.
886 if (worker->flags & WORKER_NOT_RUNNING)
891 /* this can only happen on the local cpu */
892 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
896 * The counterpart of the following dec_and_test, implied mb,
897 * worklist not empty test sequence is in insert_work().
898 * Please read comment there.
900 * NOT_RUNNING is clear. This means that we're bound to and
901 * running on the local cpu w/ rq lock held and preemption
902 * disabled, which in turn means that none else could be
903 * manipulating idle_list, so dereferencing idle_list without pool
906 if (atomic_dec_and_test(&pool->nr_running) &&
907 !list_empty(&pool->worklist))
908 to_wakeup = first_idle_worker(pool);
909 return to_wakeup ? to_wakeup->task : NULL;
913 * wq_worker_last_func - retrieve worker's last work function
915 * Determine the last function a worker executed. This is called from
916 * the scheduler to get a worker's last known identity.
919 * spin_lock_irq(rq->lock)
922 * The last work function %current executed as a worker, NULL if it
923 * hasn't executed any work yet.
925 work_func_t wq_worker_last_func(struct task_struct *task)
927 struct worker *worker = kthread_data(task);
929 return worker->last_func;
933 * worker_set_flags - set worker flags and adjust nr_running accordingly
935 * @flags: flags to set
937 * Set @flags in @worker->flags and adjust nr_running accordingly.
940 * spin_lock_irq(pool->lock)
942 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
944 struct worker_pool *pool = worker->pool;
946 WARN_ON_ONCE(worker->task != current);
948 /* If transitioning into NOT_RUNNING, adjust nr_running. */
949 if ((flags & WORKER_NOT_RUNNING) &&
950 !(worker->flags & WORKER_NOT_RUNNING)) {
951 atomic_dec(&pool->nr_running);
954 worker->flags |= flags;
958 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
960 * @flags: flags to clear
962 * Clear @flags in @worker->flags and adjust nr_running accordingly.
965 * spin_lock_irq(pool->lock)
967 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
969 struct worker_pool *pool = worker->pool;
970 unsigned int oflags = worker->flags;
972 WARN_ON_ONCE(worker->task != current);
974 worker->flags &= ~flags;
977 * If transitioning out of NOT_RUNNING, increment nr_running. Note
978 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
979 * of multiple flags, not a single flag.
981 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
982 if (!(worker->flags & WORKER_NOT_RUNNING))
983 atomic_inc(&pool->nr_running);
987 * find_worker_executing_work - find worker which is executing a work
988 * @pool: pool of interest
989 * @work: work to find worker for
991 * Find a worker which is executing @work on @pool by searching
992 * @pool->busy_hash which is keyed by the address of @work. For a worker
993 * to match, its current execution should match the address of @work and
994 * its work function. This is to avoid unwanted dependency between
995 * unrelated work executions through a work item being recycled while still
998 * This is a bit tricky. A work item may be freed once its execution
999 * starts and nothing prevents the freed area from being recycled for
1000 * another work item. If the same work item address ends up being reused
1001 * before the original execution finishes, workqueue will identify the
1002 * recycled work item as currently executing and make it wait until the
1003 * current execution finishes, introducing an unwanted dependency.
1005 * This function checks the work item address and work function to avoid
1006 * false positives. Note that this isn't complete as one may construct a
1007 * work function which can introduce dependency onto itself through a
1008 * recycled work item. Well, if somebody wants to shoot oneself in the
1009 * foot that badly, there's only so much we can do, and if such deadlock
1010 * actually occurs, it should be easy to locate the culprit work function.
1013 * spin_lock_irq(pool->lock).
1016 * Pointer to worker which is executing @work if found, %NULL
1019 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1020 struct work_struct *work)
1022 struct worker *worker;
1024 hash_for_each_possible(pool->busy_hash, worker, hentry,
1025 (unsigned long)work)
1026 if (worker->current_work == work &&
1027 worker->current_func == work->func)
1034 * move_linked_works - move linked works to a list
1035 * @work: start of series of works to be scheduled
1036 * @head: target list to append @work to
1037 * @nextp: out parameter for nested worklist walking
1039 * Schedule linked works starting from @work to @head. Work series to
1040 * be scheduled starts at @work and includes any consecutive work with
1041 * WORK_STRUCT_LINKED set in its predecessor.
1043 * If @nextp is not NULL, it's updated to point to the next work of
1044 * the last scheduled work. This allows move_linked_works() to be
1045 * nested inside outer list_for_each_entry_safe().
1048 * spin_lock_irq(pool->lock).
1050 static void move_linked_works(struct work_struct *work, struct list_head *head,
1051 struct work_struct **nextp)
1053 struct work_struct *n;
1056 * Linked worklist will always end before the end of the list,
1057 * use NULL for list head.
1059 list_for_each_entry_safe_from(work, n, NULL, entry) {
1060 list_move_tail(&work->entry, head);
1061 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1066 * If we're already inside safe list traversal and have moved
1067 * multiple works to the scheduled queue, the next position
1068 * needs to be updated.
1075 * get_pwq - get an extra reference on the specified pool_workqueue
1076 * @pwq: pool_workqueue to get
1078 * Obtain an extra reference on @pwq. The caller should guarantee that
1079 * @pwq has positive refcnt and be holding the matching pool->lock.
1081 static void get_pwq(struct pool_workqueue *pwq)
1083 lockdep_assert_held(&pwq->pool->lock);
1084 WARN_ON_ONCE(pwq->refcnt <= 0);
1089 * put_pwq - put a pool_workqueue reference
1090 * @pwq: pool_workqueue to put
1092 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1093 * destruction. The caller should be holding the matching pool->lock.
1095 static void put_pwq(struct pool_workqueue *pwq)
1097 lockdep_assert_held(&pwq->pool->lock);
1098 if (likely(--pwq->refcnt))
1100 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1103 * @pwq can't be released under pool->lock, bounce to
1104 * pwq_unbound_release_workfn(). This never recurses on the same
1105 * pool->lock as this path is taken only for unbound workqueues and
1106 * the release work item is scheduled on a per-cpu workqueue. To
1107 * avoid lockdep warning, unbound pool->locks are given lockdep
1108 * subclass of 1 in get_unbound_pool().
1110 schedule_work(&pwq->unbound_release_work);
1114 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1115 * @pwq: pool_workqueue to put (can be %NULL)
1117 * put_pwq() with locking. This function also allows %NULL @pwq.
1119 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1123 * As both pwqs and pools are sched-RCU protected, the
1124 * following lock operations are safe.
1126 spin_lock_irq(&pwq->pool->lock);
1128 spin_unlock_irq(&pwq->pool->lock);
1132 static void pwq_activate_delayed_work(struct work_struct *work)
1134 struct pool_workqueue *pwq = get_work_pwq(work);
1136 trace_workqueue_activate_work(work);
1137 if (list_empty(&pwq->pool->worklist))
1138 pwq->pool->watchdog_ts = jiffies;
1139 move_linked_works(work, &pwq->pool->worklist, NULL);
1140 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1144 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1146 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1147 struct work_struct, entry);
1149 pwq_activate_delayed_work(work);
1153 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1154 * @pwq: pwq of interest
1155 * @color: color of work which left the queue
1157 * A work either has completed or is removed from pending queue,
1158 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1161 * spin_lock_irq(pool->lock).
1163 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1165 /* uncolored work items don't participate in flushing or nr_active */
1166 if (color == WORK_NO_COLOR)
1169 pwq->nr_in_flight[color]--;
1172 if (!list_empty(&pwq->delayed_works)) {
1173 /* one down, submit a delayed one */
1174 if (pwq->nr_active < pwq->max_active)
1175 pwq_activate_first_delayed(pwq);
1178 /* is flush in progress and are we at the flushing tip? */
1179 if (likely(pwq->flush_color != color))
1182 /* are there still in-flight works? */
1183 if (pwq->nr_in_flight[color])
1186 /* this pwq is done, clear flush_color */
1187 pwq->flush_color = -1;
1190 * If this was the last pwq, wake up the first flusher. It
1191 * will handle the rest.
1193 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1194 complete(&pwq->wq->first_flusher->done);
1200 * try_to_grab_pending - steal work item from worklist and disable irq
1201 * @work: work item to steal
1202 * @is_dwork: @work is a delayed_work
1203 * @flags: place to store irq state
1205 * Try to grab PENDING bit of @work. This function can handle @work in any
1206 * stable state - idle, on timer or on worklist.
1209 * 1 if @work was pending and we successfully stole PENDING
1210 * 0 if @work was idle and we claimed PENDING
1211 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1212 * -ENOENT if someone else is canceling @work, this state may persist
1213 * for arbitrarily long
1216 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1217 * interrupted while holding PENDING and @work off queue, irq must be
1218 * disabled on entry. This, combined with delayed_work->timer being
1219 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1221 * On successful return, >= 0, irq is disabled and the caller is
1222 * responsible for releasing it using local_irq_restore(*@flags).
1224 * This function is safe to call from any context including IRQ handler.
1226 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1227 unsigned long *flags)
1229 struct worker_pool *pool;
1230 struct pool_workqueue *pwq;
1232 local_irq_save(*flags);
1234 /* try to steal the timer if it exists */
1236 struct delayed_work *dwork = to_delayed_work(work);
1239 * dwork->timer is irqsafe. If del_timer() fails, it's
1240 * guaranteed that the timer is not queued anywhere and not
1241 * running on the local CPU.
1243 if (likely(del_timer(&dwork->timer)))
1247 /* try to claim PENDING the normal way */
1248 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1252 * The queueing is in progress, or it is already queued. Try to
1253 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1255 pool = get_work_pool(work);
1259 spin_lock(&pool->lock);
1261 * work->data is guaranteed to point to pwq only while the work
1262 * item is queued on pwq->wq, and both updating work->data to point
1263 * to pwq on queueing and to pool on dequeueing are done under
1264 * pwq->pool->lock. This in turn guarantees that, if work->data
1265 * points to pwq which is associated with a locked pool, the work
1266 * item is currently queued on that pool.
1268 pwq = get_work_pwq(work);
1269 if (pwq && pwq->pool == pool) {
1270 debug_work_deactivate(work);
1273 * A delayed work item cannot be grabbed directly because
1274 * it might have linked NO_COLOR work items which, if left
1275 * on the delayed_list, will confuse pwq->nr_active
1276 * management later on and cause stall. Make sure the work
1277 * item is activated before grabbing.
1279 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1280 pwq_activate_delayed_work(work);
1282 list_del_init(&work->entry);
1283 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1285 /* work->data points to pwq iff queued, point to pool */
1286 set_work_pool_and_keep_pending(work, pool->id);
1288 spin_unlock(&pool->lock);
1291 spin_unlock(&pool->lock);
1293 local_irq_restore(*flags);
1294 if (work_is_canceling(work))
1301 * insert_work - insert a work into a pool
1302 * @pwq: pwq @work belongs to
1303 * @work: work to insert
1304 * @head: insertion point
1305 * @extra_flags: extra WORK_STRUCT_* flags to set
1307 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1308 * work_struct flags.
1311 * spin_lock_irq(pool->lock).
1313 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1314 struct list_head *head, unsigned int extra_flags)
1316 struct worker_pool *pool = pwq->pool;
1318 /* we own @work, set data and link */
1319 set_work_pwq(work, pwq, extra_flags);
1320 list_add_tail(&work->entry, head);
1324 * Ensure either wq_worker_sleeping() sees the above
1325 * list_add_tail() or we see zero nr_running to avoid workers lying
1326 * around lazily while there are works to be processed.
1330 if (__need_more_worker(pool))
1331 wake_up_worker(pool);
1335 * Test whether @work is being queued from another work executing on the
1338 static bool is_chained_work(struct workqueue_struct *wq)
1340 struct worker *worker;
1342 worker = current_wq_worker();
1344 * Return %true iff I'm a worker execuing a work item on @wq. If
1345 * I'm @worker, it's safe to dereference it without locking.
1347 return worker && worker->current_pwq->wq == wq;
1351 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1352 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1353 * avoid perturbing sensitive tasks.
1355 static int wq_select_unbound_cpu(int cpu)
1357 static bool printed_dbg_warning;
1360 if (likely(!wq_debug_force_rr_cpu)) {
1361 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1363 } else if (!printed_dbg_warning) {
1364 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1365 printed_dbg_warning = true;
1368 if (cpumask_empty(wq_unbound_cpumask))
1371 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1372 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1373 if (unlikely(new_cpu >= nr_cpu_ids)) {
1374 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1375 if (unlikely(new_cpu >= nr_cpu_ids))
1378 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1383 static void __queue_work(int cpu, struct workqueue_struct *wq,
1384 struct work_struct *work)
1386 struct pool_workqueue *pwq;
1387 struct worker_pool *last_pool;
1388 struct list_head *worklist;
1389 unsigned int work_flags;
1390 unsigned int req_cpu = cpu;
1393 * While a work item is PENDING && off queue, a task trying to
1394 * steal the PENDING will busy-loop waiting for it to either get
1395 * queued or lose PENDING. Grabbing PENDING and queueing should
1396 * happen with IRQ disabled.
1398 lockdep_assert_irqs_disabled();
1400 debug_work_activate(work);
1402 /* if draining, only works from the same workqueue are allowed */
1403 if (unlikely(wq->flags & __WQ_DRAINING) &&
1404 WARN_ON_ONCE(!is_chained_work(wq)))
1407 if (req_cpu == WORK_CPU_UNBOUND)
1408 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1410 /* pwq which will be used unless @work is executing elsewhere */
1411 if (!(wq->flags & WQ_UNBOUND))
1412 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1414 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1417 * If @work was previously on a different pool, it might still be
1418 * running there, in which case the work needs to be queued on that
1419 * pool to guarantee non-reentrancy.
1421 last_pool = get_work_pool(work);
1422 if (last_pool && last_pool != pwq->pool) {
1423 struct worker *worker;
1425 spin_lock(&last_pool->lock);
1427 worker = find_worker_executing_work(last_pool, work);
1429 if (worker && worker->current_pwq->wq == wq) {
1430 pwq = worker->current_pwq;
1432 /* meh... not running there, queue here */
1433 spin_unlock(&last_pool->lock);
1434 spin_lock(&pwq->pool->lock);
1437 spin_lock(&pwq->pool->lock);
1441 * pwq is determined and locked. For unbound pools, we could have
1442 * raced with pwq release and it could already be dead. If its
1443 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1444 * without another pwq replacing it in the numa_pwq_tbl or while
1445 * work items are executing on it, so the retrying is guaranteed to
1446 * make forward-progress.
1448 if (unlikely(!pwq->refcnt)) {
1449 if (wq->flags & WQ_UNBOUND) {
1450 spin_unlock(&pwq->pool->lock);
1455 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1459 /* pwq determined, queue */
1460 trace_workqueue_queue_work(req_cpu, pwq, work);
1462 if (WARN_ON(!list_empty(&work->entry))) {
1463 spin_unlock(&pwq->pool->lock);
1467 pwq->nr_in_flight[pwq->work_color]++;
1468 work_flags = work_color_to_flags(pwq->work_color);
1470 if (likely(pwq->nr_active < pwq->max_active)) {
1471 trace_workqueue_activate_work(work);
1473 worklist = &pwq->pool->worklist;
1474 if (list_empty(worklist))
1475 pwq->pool->watchdog_ts = jiffies;
1477 work_flags |= WORK_STRUCT_DELAYED;
1478 worklist = &pwq->delayed_works;
1481 insert_work(pwq, work, worklist, work_flags);
1483 spin_unlock(&pwq->pool->lock);
1487 * queue_work_on - queue work on specific cpu
1488 * @cpu: CPU number to execute work on
1489 * @wq: workqueue to use
1490 * @work: work to queue
1492 * We queue the work to a specific CPU, the caller must ensure it
1495 * Return: %false if @work was already on a queue, %true otherwise.
1497 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1498 struct work_struct *work)
1501 unsigned long flags;
1503 local_irq_save(flags);
1505 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1506 __queue_work(cpu, wq, work);
1510 local_irq_restore(flags);
1513 EXPORT_SYMBOL(queue_work_on);
1515 void delayed_work_timer_fn(struct timer_list *t)
1517 struct delayed_work *dwork = from_timer(dwork, t, timer);
1519 /* should have been called from irqsafe timer with irq already off */
1520 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1522 EXPORT_SYMBOL(delayed_work_timer_fn);
1524 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1525 struct delayed_work *dwork, unsigned long delay)
1527 struct timer_list *timer = &dwork->timer;
1528 struct work_struct *work = &dwork->work;
1531 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1532 WARN_ON_ONCE(timer_pending(timer));
1533 WARN_ON_ONCE(!list_empty(&work->entry));
1536 * If @delay is 0, queue @dwork->work immediately. This is for
1537 * both optimization and correctness. The earliest @timer can
1538 * expire is on the closest next tick and delayed_work users depend
1539 * on that there's no such delay when @delay is 0.
1542 __queue_work(cpu, wq, &dwork->work);
1548 timer->expires = jiffies + delay;
1550 if (unlikely(cpu != WORK_CPU_UNBOUND))
1551 add_timer_on(timer, cpu);
1557 * queue_delayed_work_on - queue work on specific CPU after delay
1558 * @cpu: CPU number to execute work on
1559 * @wq: workqueue to use
1560 * @dwork: work to queue
1561 * @delay: number of jiffies to wait before queueing
1563 * Return: %false if @work was already on a queue, %true otherwise. If
1564 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1567 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1568 struct delayed_work *dwork, unsigned long delay)
1570 struct work_struct *work = &dwork->work;
1572 unsigned long flags;
1574 /* read the comment in __queue_work() */
1575 local_irq_save(flags);
1577 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1578 __queue_delayed_work(cpu, wq, dwork, delay);
1582 local_irq_restore(flags);
1585 EXPORT_SYMBOL(queue_delayed_work_on);
1588 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1589 * @cpu: CPU number to execute work on
1590 * @wq: workqueue to use
1591 * @dwork: work to queue
1592 * @delay: number of jiffies to wait before queueing
1594 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1595 * modify @dwork's timer so that it expires after @delay. If @delay is
1596 * zero, @work is guaranteed to be scheduled immediately regardless of its
1599 * Return: %false if @dwork was idle and queued, %true if @dwork was
1600 * pending and its timer was modified.
1602 * This function is safe to call from any context including IRQ handler.
1603 * See try_to_grab_pending() for details.
1605 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1606 struct delayed_work *dwork, unsigned long delay)
1608 unsigned long flags;
1612 ret = try_to_grab_pending(&dwork->work, true, &flags);
1613 } while (unlikely(ret == -EAGAIN));
1615 if (likely(ret >= 0)) {
1616 __queue_delayed_work(cpu, wq, dwork, delay);
1617 local_irq_restore(flags);
1620 /* -ENOENT from try_to_grab_pending() becomes %true */
1623 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1625 static void rcu_work_rcufn(struct rcu_head *rcu)
1627 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1629 /* read the comment in __queue_work() */
1630 local_irq_disable();
1631 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1636 * queue_rcu_work - queue work after a RCU grace period
1637 * @wq: workqueue to use
1638 * @rwork: work to queue
1640 * Return: %false if @rwork was already pending, %true otherwise. Note
1641 * that a full RCU grace period is guaranteed only after a %true return.
1642 * While @rwork is guarnateed to be executed after a %false return, the
1643 * execution may happen before a full RCU grace period has passed.
1645 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1647 struct work_struct *work = &rwork->work;
1649 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1651 call_rcu(&rwork->rcu, rcu_work_rcufn);
1657 EXPORT_SYMBOL(queue_rcu_work);
1660 * worker_enter_idle - enter idle state
1661 * @worker: worker which is entering idle state
1663 * @worker is entering idle state. Update stats and idle timer if
1667 * spin_lock_irq(pool->lock).
1669 static void worker_enter_idle(struct worker *worker)
1671 struct worker_pool *pool = worker->pool;
1673 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1674 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1675 (worker->hentry.next || worker->hentry.pprev)))
1678 /* can't use worker_set_flags(), also called from create_worker() */
1679 worker->flags |= WORKER_IDLE;
1681 worker->last_active = jiffies;
1683 /* idle_list is LIFO */
1684 list_add(&worker->entry, &pool->idle_list);
1686 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1687 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1690 * Sanity check nr_running. Because unbind_workers() releases
1691 * pool->lock between setting %WORKER_UNBOUND and zapping
1692 * nr_running, the warning may trigger spuriously. Check iff
1693 * unbind is not in progress.
1695 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1696 pool->nr_workers == pool->nr_idle &&
1697 atomic_read(&pool->nr_running));
1701 * worker_leave_idle - leave idle state
1702 * @worker: worker which is leaving idle state
1704 * @worker is leaving idle state. Update stats.
1707 * spin_lock_irq(pool->lock).
1709 static void worker_leave_idle(struct worker *worker)
1711 struct worker_pool *pool = worker->pool;
1713 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1715 worker_clr_flags(worker, WORKER_IDLE);
1717 list_del_init(&worker->entry);
1720 static struct worker *alloc_worker(int node)
1722 struct worker *worker;
1724 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1726 INIT_LIST_HEAD(&worker->entry);
1727 INIT_LIST_HEAD(&worker->scheduled);
1728 INIT_LIST_HEAD(&worker->node);
1729 /* on creation a worker is in !idle && prep state */
1730 worker->flags = WORKER_PREP;
1736 * worker_attach_to_pool() - attach a worker to a pool
1737 * @worker: worker to be attached
1738 * @pool: the target pool
1740 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1741 * cpu-binding of @worker are kept coordinated with the pool across
1744 static void worker_attach_to_pool(struct worker *worker,
1745 struct worker_pool *pool)
1747 mutex_lock(&wq_pool_attach_mutex);
1750 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1751 * online CPUs. It'll be re-applied when any of the CPUs come up.
1753 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1756 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1757 * stable across this function. See the comments above the flag
1758 * definition for details.
1760 if (pool->flags & POOL_DISASSOCIATED)
1761 worker->flags |= WORKER_UNBOUND;
1763 list_add_tail(&worker->node, &pool->workers);
1764 worker->pool = pool;
1766 mutex_unlock(&wq_pool_attach_mutex);
1770 * worker_detach_from_pool() - detach a worker from its pool
1771 * @worker: worker which is attached to its pool
1773 * Undo the attaching which had been done in worker_attach_to_pool(). The
1774 * caller worker shouldn't access to the pool after detached except it has
1775 * other reference to the pool.
1777 static void worker_detach_from_pool(struct worker *worker)
1779 struct worker_pool *pool = worker->pool;
1780 struct completion *detach_completion = NULL;
1782 mutex_lock(&wq_pool_attach_mutex);
1784 list_del(&worker->node);
1785 worker->pool = NULL;
1787 if (list_empty(&pool->workers))
1788 detach_completion = pool->detach_completion;
1789 mutex_unlock(&wq_pool_attach_mutex);
1791 /* clear leftover flags without pool->lock after it is detached */
1792 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1794 if (detach_completion)
1795 complete(detach_completion);
1799 * create_worker - create a new workqueue worker
1800 * @pool: pool the new worker will belong to
1802 * Create and start a new worker which is attached to @pool.
1805 * Might sleep. Does GFP_KERNEL allocations.
1808 * Pointer to the newly created worker.
1810 static struct worker *create_worker(struct worker_pool *pool)
1812 struct worker *worker = NULL;
1816 /* ID is needed to determine kthread name */
1817 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1821 worker = alloc_worker(pool->node);
1828 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1829 pool->attrs->nice < 0 ? "H" : "");
1831 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1833 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1834 "kworker/%s", id_buf);
1835 if (IS_ERR(worker->task))
1838 set_user_nice(worker->task, pool->attrs->nice);
1839 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1841 /* successful, attach the worker to the pool */
1842 worker_attach_to_pool(worker, pool);
1844 /* start the newly created worker */
1845 spin_lock_irq(&pool->lock);
1846 worker->pool->nr_workers++;
1847 worker_enter_idle(worker);
1848 wake_up_process(worker->task);
1849 spin_unlock_irq(&pool->lock);
1855 ida_simple_remove(&pool->worker_ida, id);
1861 * destroy_worker - destroy a workqueue worker
1862 * @worker: worker to be destroyed
1864 * Destroy @worker and adjust @pool stats accordingly. The worker should
1868 * spin_lock_irq(pool->lock).
1870 static void destroy_worker(struct worker *worker)
1872 struct worker_pool *pool = worker->pool;
1874 lockdep_assert_held(&pool->lock);
1876 /* sanity check frenzy */
1877 if (WARN_ON(worker->current_work) ||
1878 WARN_ON(!list_empty(&worker->scheduled)) ||
1879 WARN_ON(!(worker->flags & WORKER_IDLE)))
1885 list_del_init(&worker->entry);
1886 worker->flags |= WORKER_DIE;
1887 wake_up_process(worker->task);
1890 static void idle_worker_timeout(struct timer_list *t)
1892 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1894 spin_lock_irq(&pool->lock);
1896 while (too_many_workers(pool)) {
1897 struct worker *worker;
1898 unsigned long expires;
1900 /* idle_list is kept in LIFO order, check the last one */
1901 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1902 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1904 if (time_before(jiffies, expires)) {
1905 mod_timer(&pool->idle_timer, expires);
1909 destroy_worker(worker);
1912 spin_unlock_irq(&pool->lock);
1915 static void send_mayday(struct work_struct *work)
1917 struct pool_workqueue *pwq = get_work_pwq(work);
1918 struct workqueue_struct *wq = pwq->wq;
1920 lockdep_assert_held(&wq_mayday_lock);
1925 /* mayday mayday mayday */
1926 if (list_empty(&pwq->mayday_node)) {
1928 * If @pwq is for an unbound wq, its base ref may be put at
1929 * any time due to an attribute change. Pin @pwq until the
1930 * rescuer is done with it.
1933 list_add_tail(&pwq->mayday_node, &wq->maydays);
1934 wake_up_process(wq->rescuer->task);
1938 static void pool_mayday_timeout(struct timer_list *t)
1940 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
1941 struct work_struct *work;
1943 spin_lock_irq(&pool->lock);
1944 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1946 if (need_to_create_worker(pool)) {
1948 * We've been trying to create a new worker but
1949 * haven't been successful. We might be hitting an
1950 * allocation deadlock. Send distress signals to
1953 list_for_each_entry(work, &pool->worklist, entry)
1957 spin_unlock(&wq_mayday_lock);
1958 spin_unlock_irq(&pool->lock);
1960 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1964 * maybe_create_worker - create a new worker if necessary
1965 * @pool: pool to create a new worker for
1967 * Create a new worker for @pool if necessary. @pool is guaranteed to
1968 * have at least one idle worker on return from this function. If
1969 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1970 * sent to all rescuers with works scheduled on @pool to resolve
1971 * possible allocation deadlock.
1973 * On return, need_to_create_worker() is guaranteed to be %false and
1974 * may_start_working() %true.
1977 * spin_lock_irq(pool->lock) which may be released and regrabbed
1978 * multiple times. Does GFP_KERNEL allocations. Called only from
1981 static void maybe_create_worker(struct worker_pool *pool)
1982 __releases(&pool->lock)
1983 __acquires(&pool->lock)
1986 spin_unlock_irq(&pool->lock);
1988 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1989 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1992 if (create_worker(pool) || !need_to_create_worker(pool))
1995 schedule_timeout_interruptible(CREATE_COOLDOWN);
1997 if (!need_to_create_worker(pool))
2001 del_timer_sync(&pool->mayday_timer);
2002 spin_lock_irq(&pool->lock);
2004 * This is necessary even after a new worker was just successfully
2005 * created as @pool->lock was dropped and the new worker might have
2006 * already become busy.
2008 if (need_to_create_worker(pool))
2013 * manage_workers - manage worker pool
2016 * Assume the manager role and manage the worker pool @worker belongs
2017 * to. At any given time, there can be only zero or one manager per
2018 * pool. The exclusion is handled automatically by this function.
2020 * The caller can safely start processing works on false return. On
2021 * true return, it's guaranteed that need_to_create_worker() is false
2022 * and may_start_working() is true.
2025 * spin_lock_irq(pool->lock) which may be released and regrabbed
2026 * multiple times. Does GFP_KERNEL allocations.
2029 * %false if the pool doesn't need management and the caller can safely
2030 * start processing works, %true if management function was performed and
2031 * the conditions that the caller verified before calling the function may
2032 * no longer be true.
2034 static bool manage_workers(struct worker *worker)
2036 struct worker_pool *pool = worker->pool;
2038 if (pool->flags & POOL_MANAGER_ACTIVE)
2041 pool->flags |= POOL_MANAGER_ACTIVE;
2042 pool->manager = worker;
2044 maybe_create_worker(pool);
2046 pool->manager = NULL;
2047 pool->flags &= ~POOL_MANAGER_ACTIVE;
2048 wake_up(&wq_manager_wait);
2053 * process_one_work - process single work
2055 * @work: work to process
2057 * Process @work. This function contains all the logics necessary to
2058 * process a single work including synchronization against and
2059 * interaction with other workers on the same cpu, queueing and
2060 * flushing. As long as context requirement is met, any worker can
2061 * call this function to process a work.
2064 * spin_lock_irq(pool->lock) which is released and regrabbed.
2066 static void process_one_work(struct worker *worker, struct work_struct *work)
2067 __releases(&pool->lock)
2068 __acquires(&pool->lock)
2070 struct pool_workqueue *pwq = get_work_pwq(work);
2071 struct worker_pool *pool = worker->pool;
2072 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2074 struct worker *collision;
2075 #ifdef CONFIG_LOCKDEP
2077 * It is permissible to free the struct work_struct from
2078 * inside the function that is called from it, this we need to
2079 * take into account for lockdep too. To avoid bogus "held
2080 * lock freed" warnings as well as problems when looking into
2081 * work->lockdep_map, make a copy and use that here.
2083 struct lockdep_map lockdep_map;
2085 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2087 /* ensure we're on the correct CPU */
2088 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2089 raw_smp_processor_id() != pool->cpu);
2092 * A single work shouldn't be executed concurrently by
2093 * multiple workers on a single cpu. Check whether anyone is
2094 * already processing the work. If so, defer the work to the
2095 * currently executing one.
2097 collision = find_worker_executing_work(pool, work);
2098 if (unlikely(collision)) {
2099 move_linked_works(work, &collision->scheduled, NULL);
2103 /* claim and dequeue */
2104 debug_work_deactivate(work);
2105 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2106 worker->current_work = work;
2107 worker->current_func = work->func;
2108 worker->current_pwq = pwq;
2109 work_color = get_work_color(work);
2112 * Record wq name for cmdline and debug reporting, may get
2113 * overridden through set_worker_desc().
2115 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2117 list_del_init(&work->entry);
2120 * CPU intensive works don't participate in concurrency management.
2121 * They're the scheduler's responsibility. This takes @worker out
2122 * of concurrency management and the next code block will chain
2123 * execution of the pending work items.
2125 if (unlikely(cpu_intensive))
2126 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2129 * Wake up another worker if necessary. The condition is always
2130 * false for normal per-cpu workers since nr_running would always
2131 * be >= 1 at this point. This is used to chain execution of the
2132 * pending work items for WORKER_NOT_RUNNING workers such as the
2133 * UNBOUND and CPU_INTENSIVE ones.
2135 if (need_more_worker(pool))
2136 wake_up_worker(pool);
2139 * Record the last pool and clear PENDING which should be the last
2140 * update to @work. Also, do this inside @pool->lock so that
2141 * PENDING and queued state changes happen together while IRQ is
2144 set_work_pool_and_clear_pending(work, pool->id);
2146 spin_unlock_irq(&pool->lock);
2148 lock_map_acquire(&pwq->wq->lockdep_map);
2149 lock_map_acquire(&lockdep_map);
2151 * Strictly speaking we should mark the invariant state without holding
2152 * any locks, that is, before these two lock_map_acquire()'s.
2154 * However, that would result in:
2161 * Which would create W1->C->W1 dependencies, even though there is no
2162 * actual deadlock possible. There are two solutions, using a
2163 * read-recursive acquire on the work(queue) 'locks', but this will then
2164 * hit the lockdep limitation on recursive locks, or simply discard
2167 * AFAICT there is no possible deadlock scenario between the
2168 * flush_work() and complete() primitives (except for single-threaded
2169 * workqueues), so hiding them isn't a problem.
2171 lockdep_invariant_state(true);
2172 trace_workqueue_execute_start(work);
2173 worker->current_func(work);
2175 * While we must be careful to not use "work" after this, the trace
2176 * point will only record its address.
2178 trace_workqueue_execute_end(work);
2179 lock_map_release(&lockdep_map);
2180 lock_map_release(&pwq->wq->lockdep_map);
2182 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2183 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2184 " last function: %pf\n",
2185 current->comm, preempt_count(), task_pid_nr(current),
2186 worker->current_func);
2187 debug_show_held_locks(current);
2192 * The following prevents a kworker from hogging CPU on !PREEMPT
2193 * kernels, where a requeueing work item waiting for something to
2194 * happen could deadlock with stop_machine as such work item could
2195 * indefinitely requeue itself while all other CPUs are trapped in
2196 * stop_machine. At the same time, report a quiescent RCU state so
2197 * the same condition doesn't freeze RCU.
2201 spin_lock_irq(&pool->lock);
2203 /* clear cpu intensive status */
2204 if (unlikely(cpu_intensive))
2205 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2207 /* tag the worker for identification in schedule() */
2208 worker->last_func = worker->current_func;
2210 /* we're done with it, release */
2211 hash_del(&worker->hentry);
2212 worker->current_work = NULL;
2213 worker->current_func = NULL;
2214 worker->current_pwq = NULL;
2215 pwq_dec_nr_in_flight(pwq, work_color);
2219 * process_scheduled_works - process scheduled works
2222 * Process all scheduled works. Please note that the scheduled list
2223 * may change while processing a work, so this function repeatedly
2224 * fetches a work from the top and executes it.
2227 * spin_lock_irq(pool->lock) which may be released and regrabbed
2230 static void process_scheduled_works(struct worker *worker)
2232 while (!list_empty(&worker->scheduled)) {
2233 struct work_struct *work = list_first_entry(&worker->scheduled,
2234 struct work_struct, entry);
2235 process_one_work(worker, work);
2239 static void set_pf_worker(bool val)
2241 mutex_lock(&wq_pool_attach_mutex);
2243 current->flags |= PF_WQ_WORKER;
2245 current->flags &= ~PF_WQ_WORKER;
2246 mutex_unlock(&wq_pool_attach_mutex);
2250 * worker_thread - the worker thread function
2253 * The worker thread function. All workers belong to a worker_pool -
2254 * either a per-cpu one or dynamic unbound one. These workers process all
2255 * work items regardless of their specific target workqueue. The only
2256 * exception is work items which belong to workqueues with a rescuer which
2257 * will be explained in rescuer_thread().
2261 static int worker_thread(void *__worker)
2263 struct worker *worker = __worker;
2264 struct worker_pool *pool = worker->pool;
2266 /* tell the scheduler that this is a workqueue worker */
2267 set_pf_worker(true);
2269 spin_lock_irq(&pool->lock);
2271 /* am I supposed to die? */
2272 if (unlikely(worker->flags & WORKER_DIE)) {
2273 spin_unlock_irq(&pool->lock);
2274 WARN_ON_ONCE(!list_empty(&worker->entry));
2275 set_pf_worker(false);
2277 set_task_comm(worker->task, "kworker/dying");
2278 ida_simple_remove(&pool->worker_ida, worker->id);
2279 worker_detach_from_pool(worker);
2284 worker_leave_idle(worker);
2286 /* no more worker necessary? */
2287 if (!need_more_worker(pool))
2290 /* do we need to manage? */
2291 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2295 * ->scheduled list can only be filled while a worker is
2296 * preparing to process a work or actually processing it.
2297 * Make sure nobody diddled with it while I was sleeping.
2299 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2302 * Finish PREP stage. We're guaranteed to have at least one idle
2303 * worker or that someone else has already assumed the manager
2304 * role. This is where @worker starts participating in concurrency
2305 * management if applicable and concurrency management is restored
2306 * after being rebound. See rebind_workers() for details.
2308 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2311 struct work_struct *work =
2312 list_first_entry(&pool->worklist,
2313 struct work_struct, entry);
2315 pool->watchdog_ts = jiffies;
2317 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2318 /* optimization path, not strictly necessary */
2319 process_one_work(worker, work);
2320 if (unlikely(!list_empty(&worker->scheduled)))
2321 process_scheduled_works(worker);
2323 move_linked_works(work, &worker->scheduled, NULL);
2324 process_scheduled_works(worker);
2326 } while (keep_working(pool));
2328 worker_set_flags(worker, WORKER_PREP);
2331 * pool->lock is held and there's no work to process and no need to
2332 * manage, sleep. Workers are woken up only while holding
2333 * pool->lock or from local cpu, so setting the current state
2334 * before releasing pool->lock is enough to prevent losing any
2337 worker_enter_idle(worker);
2338 __set_current_state(TASK_IDLE);
2339 spin_unlock_irq(&pool->lock);
2345 * rescuer_thread - the rescuer thread function
2348 * Workqueue rescuer thread function. There's one rescuer for each
2349 * workqueue which has WQ_MEM_RECLAIM set.
2351 * Regular work processing on a pool may block trying to create a new
2352 * worker which uses GFP_KERNEL allocation which has slight chance of
2353 * developing into deadlock if some works currently on the same queue
2354 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2355 * the problem rescuer solves.
2357 * When such condition is possible, the pool summons rescuers of all
2358 * workqueues which have works queued on the pool and let them process
2359 * those works so that forward progress can be guaranteed.
2361 * This should happen rarely.
2365 static int rescuer_thread(void *__rescuer)
2367 struct worker *rescuer = __rescuer;
2368 struct workqueue_struct *wq = rescuer->rescue_wq;
2369 struct list_head *scheduled = &rescuer->scheduled;
2372 set_user_nice(current, RESCUER_NICE_LEVEL);
2375 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2376 * doesn't participate in concurrency management.
2378 set_pf_worker(true);
2380 set_current_state(TASK_IDLE);
2383 * By the time the rescuer is requested to stop, the workqueue
2384 * shouldn't have any work pending, but @wq->maydays may still have
2385 * pwq(s) queued. This can happen by non-rescuer workers consuming
2386 * all the work items before the rescuer got to them. Go through
2387 * @wq->maydays processing before acting on should_stop so that the
2388 * list is always empty on exit.
2390 should_stop = kthread_should_stop();
2392 /* see whether any pwq is asking for help */
2393 spin_lock_irq(&wq_mayday_lock);
2395 while (!list_empty(&wq->maydays)) {
2396 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2397 struct pool_workqueue, mayday_node);
2398 struct worker_pool *pool = pwq->pool;
2399 struct work_struct *work, *n;
2402 __set_current_state(TASK_RUNNING);
2403 list_del_init(&pwq->mayday_node);
2405 spin_unlock_irq(&wq_mayday_lock);
2407 worker_attach_to_pool(rescuer, pool);
2409 spin_lock_irq(&pool->lock);
2412 * Slurp in all works issued via this workqueue and
2415 WARN_ON_ONCE(!list_empty(scheduled));
2416 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2417 if (get_work_pwq(work) == pwq) {
2419 pool->watchdog_ts = jiffies;
2420 move_linked_works(work, scheduled, &n);
2425 if (!list_empty(scheduled)) {
2426 process_scheduled_works(rescuer);
2429 * The above execution of rescued work items could
2430 * have created more to rescue through
2431 * pwq_activate_first_delayed() or chained
2432 * queueing. Let's put @pwq back on mayday list so
2433 * that such back-to-back work items, which may be
2434 * being used to relieve memory pressure, don't
2435 * incur MAYDAY_INTERVAL delay inbetween.
2437 if (need_to_create_worker(pool)) {
2438 spin_lock(&wq_mayday_lock);
2440 list_move_tail(&pwq->mayday_node, &wq->maydays);
2441 spin_unlock(&wq_mayday_lock);
2446 * Put the reference grabbed by send_mayday(). @pool won't
2447 * go away while we're still attached to it.
2452 * Leave this pool. If need_more_worker() is %true, notify a
2453 * regular worker; otherwise, we end up with 0 concurrency
2454 * and stalling the execution.
2456 if (need_more_worker(pool))
2457 wake_up_worker(pool);
2459 spin_unlock_irq(&pool->lock);
2461 worker_detach_from_pool(rescuer);
2463 spin_lock_irq(&wq_mayday_lock);
2466 spin_unlock_irq(&wq_mayday_lock);
2469 __set_current_state(TASK_RUNNING);
2470 set_pf_worker(false);
2474 /* rescuers should never participate in concurrency management */
2475 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2481 * check_flush_dependency - check for flush dependency sanity
2482 * @target_wq: workqueue being flushed
2483 * @target_work: work item being flushed (NULL for workqueue flushes)
2485 * %current is trying to flush the whole @target_wq or @target_work on it.
2486 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2487 * reclaiming memory or running on a workqueue which doesn't have
2488 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2491 static void check_flush_dependency(struct workqueue_struct *target_wq,
2492 struct work_struct *target_work)
2494 work_func_t target_func = target_work ? target_work->func : NULL;
2495 struct worker *worker;
2497 if (target_wq->flags & WQ_MEM_RECLAIM)
2500 worker = current_wq_worker();
2502 WARN_ONCE(current->flags & PF_MEMALLOC,
2503 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2504 current->pid, current->comm, target_wq->name, target_func);
2505 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2506 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2507 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2508 worker->current_pwq->wq->name, worker->current_func,
2509 target_wq->name, target_func);
2513 struct work_struct work;
2514 struct completion done;
2515 struct task_struct *task; /* purely informational */
2518 static void wq_barrier_func(struct work_struct *work)
2520 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2521 complete(&barr->done);
2525 * insert_wq_barrier - insert a barrier work
2526 * @pwq: pwq to insert barrier into
2527 * @barr: wq_barrier to insert
2528 * @target: target work to attach @barr to
2529 * @worker: worker currently executing @target, NULL if @target is not executing
2531 * @barr is linked to @target such that @barr is completed only after
2532 * @target finishes execution. Please note that the ordering
2533 * guarantee is observed only with respect to @target and on the local
2536 * Currently, a queued barrier can't be canceled. This is because
2537 * try_to_grab_pending() can't determine whether the work to be
2538 * grabbed is at the head of the queue and thus can't clear LINKED
2539 * flag of the previous work while there must be a valid next work
2540 * after a work with LINKED flag set.
2542 * Note that when @worker is non-NULL, @target may be modified
2543 * underneath us, so we can't reliably determine pwq from @target.
2546 * spin_lock_irq(pool->lock).
2548 static void insert_wq_barrier(struct pool_workqueue *pwq,
2549 struct wq_barrier *barr,
2550 struct work_struct *target, struct worker *worker)
2552 struct list_head *head;
2553 unsigned int linked = 0;
2556 * debugobject calls are safe here even with pool->lock locked
2557 * as we know for sure that this will not trigger any of the
2558 * checks and call back into the fixup functions where we
2561 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2562 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2564 init_completion_map(&barr->done, &target->lockdep_map);
2566 barr->task = current;
2569 * If @target is currently being executed, schedule the
2570 * barrier to the worker; otherwise, put it after @target.
2573 head = worker->scheduled.next;
2575 unsigned long *bits = work_data_bits(target);
2577 head = target->entry.next;
2578 /* there can already be other linked works, inherit and set */
2579 linked = *bits & WORK_STRUCT_LINKED;
2580 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2583 debug_work_activate(&barr->work);
2584 insert_work(pwq, &barr->work, head,
2585 work_color_to_flags(WORK_NO_COLOR) | linked);
2589 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2590 * @wq: workqueue being flushed
2591 * @flush_color: new flush color, < 0 for no-op
2592 * @work_color: new work color, < 0 for no-op
2594 * Prepare pwqs for workqueue flushing.
2596 * If @flush_color is non-negative, flush_color on all pwqs should be
2597 * -1. If no pwq has in-flight commands at the specified color, all
2598 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2599 * has in flight commands, its pwq->flush_color is set to
2600 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2601 * wakeup logic is armed and %true is returned.
2603 * The caller should have initialized @wq->first_flusher prior to
2604 * calling this function with non-negative @flush_color. If
2605 * @flush_color is negative, no flush color update is done and %false
2608 * If @work_color is non-negative, all pwqs should have the same
2609 * work_color which is previous to @work_color and all will be
2610 * advanced to @work_color.
2613 * mutex_lock(wq->mutex).
2616 * %true if @flush_color >= 0 and there's something to flush. %false
2619 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2620 int flush_color, int work_color)
2623 struct pool_workqueue *pwq;
2625 if (flush_color >= 0) {
2626 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2627 atomic_set(&wq->nr_pwqs_to_flush, 1);
2630 for_each_pwq(pwq, wq) {
2631 struct worker_pool *pool = pwq->pool;
2633 spin_lock_irq(&pool->lock);
2635 if (flush_color >= 0) {
2636 WARN_ON_ONCE(pwq->flush_color != -1);
2638 if (pwq->nr_in_flight[flush_color]) {
2639 pwq->flush_color = flush_color;
2640 atomic_inc(&wq->nr_pwqs_to_flush);
2645 if (work_color >= 0) {
2646 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2647 pwq->work_color = work_color;
2650 spin_unlock_irq(&pool->lock);
2653 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2654 complete(&wq->first_flusher->done);
2660 * flush_workqueue - ensure that any scheduled work has run to completion.
2661 * @wq: workqueue to flush
2663 * This function sleeps until all work items which were queued on entry
2664 * have finished execution, but it is not livelocked by new incoming ones.
2666 void flush_workqueue(struct workqueue_struct *wq)
2668 struct wq_flusher this_flusher = {
2669 .list = LIST_HEAD_INIT(this_flusher.list),
2671 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2675 if (WARN_ON(!wq_online))
2678 lock_map_acquire(&wq->lockdep_map);
2679 lock_map_release(&wq->lockdep_map);
2681 mutex_lock(&wq->mutex);
2684 * Start-to-wait phase
2686 next_color = work_next_color(wq->work_color);
2688 if (next_color != wq->flush_color) {
2690 * Color space is not full. The current work_color
2691 * becomes our flush_color and work_color is advanced
2694 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2695 this_flusher.flush_color = wq->work_color;
2696 wq->work_color = next_color;
2698 if (!wq->first_flusher) {
2699 /* no flush in progress, become the first flusher */
2700 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2702 wq->first_flusher = &this_flusher;
2704 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2706 /* nothing to flush, done */
2707 wq->flush_color = next_color;
2708 wq->first_flusher = NULL;
2713 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2714 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2715 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2719 * Oops, color space is full, wait on overflow queue.
2720 * The next flush completion will assign us
2721 * flush_color and transfer to flusher_queue.
2723 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2726 check_flush_dependency(wq, NULL);
2728 mutex_unlock(&wq->mutex);
2730 wait_for_completion(&this_flusher.done);
2733 * Wake-up-and-cascade phase
2735 * First flushers are responsible for cascading flushes and
2736 * handling overflow. Non-first flushers can simply return.
2738 if (wq->first_flusher != &this_flusher)
2741 mutex_lock(&wq->mutex);
2743 /* we might have raced, check again with mutex held */
2744 if (wq->first_flusher != &this_flusher)
2747 wq->first_flusher = NULL;
2749 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2750 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2753 struct wq_flusher *next, *tmp;
2755 /* complete all the flushers sharing the current flush color */
2756 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2757 if (next->flush_color != wq->flush_color)
2759 list_del_init(&next->list);
2760 complete(&next->done);
2763 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2764 wq->flush_color != work_next_color(wq->work_color));
2766 /* this flush_color is finished, advance by one */
2767 wq->flush_color = work_next_color(wq->flush_color);
2769 /* one color has been freed, handle overflow queue */
2770 if (!list_empty(&wq->flusher_overflow)) {
2772 * Assign the same color to all overflowed
2773 * flushers, advance work_color and append to
2774 * flusher_queue. This is the start-to-wait
2775 * phase for these overflowed flushers.
2777 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2778 tmp->flush_color = wq->work_color;
2780 wq->work_color = work_next_color(wq->work_color);
2782 list_splice_tail_init(&wq->flusher_overflow,
2783 &wq->flusher_queue);
2784 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2787 if (list_empty(&wq->flusher_queue)) {
2788 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2793 * Need to flush more colors. Make the next flusher
2794 * the new first flusher and arm pwqs.
2796 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2797 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2799 list_del_init(&next->list);
2800 wq->first_flusher = next;
2802 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2806 * Meh... this color is already done, clear first
2807 * flusher and repeat cascading.
2809 wq->first_flusher = NULL;
2813 mutex_unlock(&wq->mutex);
2815 EXPORT_SYMBOL(flush_workqueue);
2818 * drain_workqueue - drain a workqueue
2819 * @wq: workqueue to drain
2821 * Wait until the workqueue becomes empty. While draining is in progress,
2822 * only chain queueing is allowed. IOW, only currently pending or running
2823 * work items on @wq can queue further work items on it. @wq is flushed
2824 * repeatedly until it becomes empty. The number of flushing is determined
2825 * by the depth of chaining and should be relatively short. Whine if it
2828 void drain_workqueue(struct workqueue_struct *wq)
2830 unsigned int flush_cnt = 0;
2831 struct pool_workqueue *pwq;
2834 * __queue_work() needs to test whether there are drainers, is much
2835 * hotter than drain_workqueue() and already looks at @wq->flags.
2836 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2838 mutex_lock(&wq->mutex);
2839 if (!wq->nr_drainers++)
2840 wq->flags |= __WQ_DRAINING;
2841 mutex_unlock(&wq->mutex);
2843 flush_workqueue(wq);
2845 mutex_lock(&wq->mutex);
2847 for_each_pwq(pwq, wq) {
2850 spin_lock_irq(&pwq->pool->lock);
2851 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2852 spin_unlock_irq(&pwq->pool->lock);
2857 if (++flush_cnt == 10 ||
2858 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2859 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2860 wq->name, flush_cnt);
2862 mutex_unlock(&wq->mutex);
2866 if (!--wq->nr_drainers)
2867 wq->flags &= ~__WQ_DRAINING;
2868 mutex_unlock(&wq->mutex);
2870 EXPORT_SYMBOL_GPL(drain_workqueue);
2872 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2875 struct worker *worker = NULL;
2876 struct worker_pool *pool;
2877 struct pool_workqueue *pwq;
2881 local_irq_disable();
2882 pool = get_work_pool(work);
2888 spin_lock(&pool->lock);
2889 /* see the comment in try_to_grab_pending() with the same code */
2890 pwq = get_work_pwq(work);
2892 if (unlikely(pwq->pool != pool))
2895 worker = find_worker_executing_work(pool, work);
2898 pwq = worker->current_pwq;
2901 check_flush_dependency(pwq->wq, work);
2903 insert_wq_barrier(pwq, barr, work, worker);
2904 spin_unlock_irq(&pool->lock);
2907 * Force a lock recursion deadlock when using flush_work() inside a
2908 * single-threaded or rescuer equipped workqueue.
2910 * For single threaded workqueues the deadlock happens when the work
2911 * is after the work issuing the flush_work(). For rescuer equipped
2912 * workqueues the deadlock happens when the rescuer stalls, blocking
2916 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
2917 lock_map_acquire(&pwq->wq->lockdep_map);
2918 lock_map_release(&pwq->wq->lockdep_map);
2923 spin_unlock_irq(&pool->lock);
2927 static bool __flush_work(struct work_struct *work, bool from_cancel)
2929 struct wq_barrier barr;
2931 if (WARN_ON(!wq_online))
2935 lock_map_acquire(&work->lockdep_map);
2936 lock_map_release(&work->lockdep_map);
2939 if (start_flush_work(work, &barr, from_cancel)) {
2940 wait_for_completion(&barr.done);
2941 destroy_work_on_stack(&barr.work);
2949 * flush_work - wait for a work to finish executing the last queueing instance
2950 * @work: the work to flush
2952 * Wait until @work has finished execution. @work is guaranteed to be idle
2953 * on return if it hasn't been requeued since flush started.
2956 * %true if flush_work() waited for the work to finish execution,
2957 * %false if it was already idle.
2959 bool flush_work(struct work_struct *work)
2961 return __flush_work(work, false);
2963 EXPORT_SYMBOL_GPL(flush_work);
2966 wait_queue_entry_t wait;
2967 struct work_struct *work;
2970 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2972 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2974 if (cwait->work != key)
2976 return autoremove_wake_function(wait, mode, sync, key);
2979 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2981 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2982 unsigned long flags;
2986 ret = try_to_grab_pending(work, is_dwork, &flags);
2988 * If someone else is already canceling, wait for it to
2989 * finish. flush_work() doesn't work for PREEMPT_NONE
2990 * because we may get scheduled between @work's completion
2991 * and the other canceling task resuming and clearing
2992 * CANCELING - flush_work() will return false immediately
2993 * as @work is no longer busy, try_to_grab_pending() will
2994 * return -ENOENT as @work is still being canceled and the
2995 * other canceling task won't be able to clear CANCELING as
2996 * we're hogging the CPU.
2998 * Let's wait for completion using a waitqueue. As this
2999 * may lead to the thundering herd problem, use a custom
3000 * wake function which matches @work along with exclusive
3003 if (unlikely(ret == -ENOENT)) {
3004 struct cwt_wait cwait;
3006 init_wait(&cwait.wait);
3007 cwait.wait.func = cwt_wakefn;
3010 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3011 TASK_UNINTERRUPTIBLE);
3012 if (work_is_canceling(work))
3014 finish_wait(&cancel_waitq, &cwait.wait);
3016 } while (unlikely(ret < 0));
3018 /* tell other tasks trying to grab @work to back off */
3019 mark_work_canceling(work);
3020 local_irq_restore(flags);
3023 * This allows canceling during early boot. We know that @work
3027 __flush_work(work, true);
3029 clear_work_data(work);
3032 * Paired with prepare_to_wait() above so that either
3033 * waitqueue_active() is visible here or !work_is_canceling() is
3037 if (waitqueue_active(&cancel_waitq))
3038 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3044 * cancel_work_sync - cancel a work and wait for it to finish
3045 * @work: the work to cancel
3047 * Cancel @work and wait for its execution to finish. This function
3048 * can be used even if the work re-queues itself or migrates to
3049 * another workqueue. On return from this function, @work is
3050 * guaranteed to be not pending or executing on any CPU.
3052 * cancel_work_sync(&delayed_work->work) must not be used for
3053 * delayed_work's. Use cancel_delayed_work_sync() instead.
3055 * The caller must ensure that the workqueue on which @work was last
3056 * queued can't be destroyed before this function returns.
3059 * %true if @work was pending, %false otherwise.
3061 bool cancel_work_sync(struct work_struct *work)
3063 return __cancel_work_timer(work, false);
3065 EXPORT_SYMBOL_GPL(cancel_work_sync);
3068 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3069 * @dwork: the delayed work to flush
3071 * Delayed timer is cancelled and the pending work is queued for
3072 * immediate execution. Like flush_work(), this function only
3073 * considers the last queueing instance of @dwork.
3076 * %true if flush_work() waited for the work to finish execution,
3077 * %false if it was already idle.
3079 bool flush_delayed_work(struct delayed_work *dwork)
3081 local_irq_disable();
3082 if (del_timer_sync(&dwork->timer))
3083 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3085 return flush_work(&dwork->work);
3087 EXPORT_SYMBOL(flush_delayed_work);
3090 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3091 * @rwork: the rcu work to flush
3094 * %true if flush_rcu_work() waited for the work to finish execution,
3095 * %false if it was already idle.
3097 bool flush_rcu_work(struct rcu_work *rwork)
3099 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3101 flush_work(&rwork->work);
3104 return flush_work(&rwork->work);
3107 EXPORT_SYMBOL(flush_rcu_work);
3109 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3111 unsigned long flags;
3115 ret = try_to_grab_pending(work, is_dwork, &flags);
3116 } while (unlikely(ret == -EAGAIN));
3118 if (unlikely(ret < 0))
3121 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3122 local_irq_restore(flags);
3127 * cancel_delayed_work - cancel a delayed work
3128 * @dwork: delayed_work to cancel
3130 * Kill off a pending delayed_work.
3132 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3136 * The work callback function may still be running on return, unless
3137 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3138 * use cancel_delayed_work_sync() to wait on it.
3140 * This function is safe to call from any context including IRQ handler.
3142 bool cancel_delayed_work(struct delayed_work *dwork)
3144 return __cancel_work(&dwork->work, true);
3146 EXPORT_SYMBOL(cancel_delayed_work);
3149 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3150 * @dwork: the delayed work cancel
3152 * This is cancel_work_sync() for delayed works.
3155 * %true if @dwork was pending, %false otherwise.
3157 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3159 return __cancel_work_timer(&dwork->work, true);
3161 EXPORT_SYMBOL(cancel_delayed_work_sync);
3164 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3165 * @func: the function to call
3167 * schedule_on_each_cpu() executes @func on each online CPU using the
3168 * system workqueue and blocks until all CPUs have completed.
3169 * schedule_on_each_cpu() is very slow.
3172 * 0 on success, -errno on failure.
3174 int schedule_on_each_cpu(work_func_t func)
3177 struct work_struct __percpu *works;
3179 works = alloc_percpu(struct work_struct);
3185 for_each_online_cpu(cpu) {
3186 struct work_struct *work = per_cpu_ptr(works, cpu);
3188 INIT_WORK(work, func);
3189 schedule_work_on(cpu, work);
3192 for_each_online_cpu(cpu)
3193 flush_work(per_cpu_ptr(works, cpu));
3201 * execute_in_process_context - reliably execute the routine with user context
3202 * @fn: the function to execute
3203 * @ew: guaranteed storage for the execute work structure (must
3204 * be available when the work executes)
3206 * Executes the function immediately if process context is available,
3207 * otherwise schedules the function for delayed execution.
3209 * Return: 0 - function was executed
3210 * 1 - function was scheduled for execution
3212 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3214 if (!in_interrupt()) {
3219 INIT_WORK(&ew->work, fn);
3220 schedule_work(&ew->work);
3224 EXPORT_SYMBOL_GPL(execute_in_process_context);
3227 * free_workqueue_attrs - free a workqueue_attrs
3228 * @attrs: workqueue_attrs to free
3230 * Undo alloc_workqueue_attrs().
3232 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3235 free_cpumask_var(attrs->cpumask);
3241 * alloc_workqueue_attrs - allocate a workqueue_attrs
3242 * @gfp_mask: allocation mask to use
3244 * Allocate a new workqueue_attrs, initialize with default settings and
3247 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3249 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3251 struct workqueue_attrs *attrs;
3253 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3256 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3259 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3262 free_workqueue_attrs(attrs);
3266 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3267 const struct workqueue_attrs *from)
3269 to->nice = from->nice;
3270 cpumask_copy(to->cpumask, from->cpumask);
3272 * Unlike hash and equality test, this function doesn't ignore
3273 * ->no_numa as it is used for both pool and wq attrs. Instead,
3274 * get_unbound_pool() explicitly clears ->no_numa after copying.
3276 to->no_numa = from->no_numa;
3279 /* hash value of the content of @attr */
3280 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3284 hash = jhash_1word(attrs->nice, hash);
3285 hash = jhash(cpumask_bits(attrs->cpumask),
3286 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3290 /* content equality test */
3291 static bool wqattrs_equal(const struct workqueue_attrs *a,
3292 const struct workqueue_attrs *b)
3294 if (a->nice != b->nice)
3296 if (!cpumask_equal(a->cpumask, b->cpumask))
3302 * init_worker_pool - initialize a newly zalloc'd worker_pool
3303 * @pool: worker_pool to initialize
3305 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3307 * Return: 0 on success, -errno on failure. Even on failure, all fields
3308 * inside @pool proper are initialized and put_unbound_pool() can be called
3309 * on @pool safely to release it.
3311 static int init_worker_pool(struct worker_pool *pool)
3313 spin_lock_init(&pool->lock);
3316 pool->node = NUMA_NO_NODE;
3317 pool->flags |= POOL_DISASSOCIATED;
3318 pool->watchdog_ts = jiffies;
3319 INIT_LIST_HEAD(&pool->worklist);
3320 INIT_LIST_HEAD(&pool->idle_list);
3321 hash_init(pool->busy_hash);
3323 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3325 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3327 INIT_LIST_HEAD(&pool->workers);
3329 ida_init(&pool->worker_ida);
3330 INIT_HLIST_NODE(&pool->hash_node);
3333 /* shouldn't fail above this point */
3334 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3340 static void rcu_free_wq(struct rcu_head *rcu)
3342 struct workqueue_struct *wq =
3343 container_of(rcu, struct workqueue_struct, rcu);
3345 if (!(wq->flags & WQ_UNBOUND))
3346 free_percpu(wq->cpu_pwqs);
3348 free_workqueue_attrs(wq->unbound_attrs);
3354 static void rcu_free_pool(struct rcu_head *rcu)
3356 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3358 ida_destroy(&pool->worker_ida);
3359 free_workqueue_attrs(pool->attrs);
3364 * put_unbound_pool - put a worker_pool
3365 * @pool: worker_pool to put
3367 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3368 * safe manner. get_unbound_pool() calls this function on its failure path
3369 * and this function should be able to release pools which went through,
3370 * successfully or not, init_worker_pool().
3372 * Should be called with wq_pool_mutex held.
3374 static void put_unbound_pool(struct worker_pool *pool)
3376 DECLARE_COMPLETION_ONSTACK(detach_completion);
3377 struct worker *worker;
3379 lockdep_assert_held(&wq_pool_mutex);
3385 if (WARN_ON(!(pool->cpu < 0)) ||
3386 WARN_ON(!list_empty(&pool->worklist)))
3389 /* release id and unhash */
3391 idr_remove(&worker_pool_idr, pool->id);
3392 hash_del(&pool->hash_node);
3395 * Become the manager and destroy all workers. This prevents
3396 * @pool's workers from blocking on attach_mutex. We're the last
3397 * manager and @pool gets freed with the flag set.
3399 spin_lock_irq(&pool->lock);
3400 wait_event_lock_irq(wq_manager_wait,
3401 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3402 pool->flags |= POOL_MANAGER_ACTIVE;
3404 while ((worker = first_idle_worker(pool)))
3405 destroy_worker(worker);
3406 WARN_ON(pool->nr_workers || pool->nr_idle);
3407 spin_unlock_irq(&pool->lock);
3409 mutex_lock(&wq_pool_attach_mutex);
3410 if (!list_empty(&pool->workers))
3411 pool->detach_completion = &detach_completion;
3412 mutex_unlock(&wq_pool_attach_mutex);
3414 if (pool->detach_completion)
3415 wait_for_completion(pool->detach_completion);
3417 /* shut down the timers */
3418 del_timer_sync(&pool->idle_timer);
3419 del_timer_sync(&pool->mayday_timer);
3421 /* sched-RCU protected to allow dereferences from get_work_pool() */
3422 call_rcu(&pool->rcu, rcu_free_pool);
3426 * get_unbound_pool - get a worker_pool with the specified attributes
3427 * @attrs: the attributes of the worker_pool to get
3429 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3430 * reference count and return it. If there already is a matching
3431 * worker_pool, it will be used; otherwise, this function attempts to
3434 * Should be called with wq_pool_mutex held.
3436 * Return: On success, a worker_pool with the same attributes as @attrs.
3437 * On failure, %NULL.
3439 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3441 u32 hash = wqattrs_hash(attrs);
3442 struct worker_pool *pool;
3444 int target_node = NUMA_NO_NODE;
3446 lockdep_assert_held(&wq_pool_mutex);
3448 /* do we already have a matching pool? */
3449 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3450 if (wqattrs_equal(pool->attrs, attrs)) {
3456 /* if cpumask is contained inside a NUMA node, we belong to that node */
3457 if (wq_numa_enabled) {
3458 for_each_node(node) {
3459 if (cpumask_subset(attrs->cpumask,
3460 wq_numa_possible_cpumask[node])) {
3467 /* nope, create a new one */
3468 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3469 if (!pool || init_worker_pool(pool) < 0)
3472 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3473 copy_workqueue_attrs(pool->attrs, attrs);
3474 pool->node = target_node;
3477 * no_numa isn't a worker_pool attribute, always clear it. See
3478 * 'struct workqueue_attrs' comments for detail.
3480 pool->attrs->no_numa = false;
3482 if (worker_pool_assign_id(pool) < 0)
3485 /* create and start the initial worker */
3486 if (wq_online && !create_worker(pool))
3490 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3495 put_unbound_pool(pool);
3499 static void rcu_free_pwq(struct rcu_head *rcu)
3501 kmem_cache_free(pwq_cache,
3502 container_of(rcu, struct pool_workqueue, rcu));
3506 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3507 * and needs to be destroyed.
3509 static void pwq_unbound_release_workfn(struct work_struct *work)
3511 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3512 unbound_release_work);
3513 struct workqueue_struct *wq = pwq->wq;
3514 struct worker_pool *pool = pwq->pool;
3517 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3520 mutex_lock(&wq->mutex);
3521 list_del_rcu(&pwq->pwqs_node);
3522 is_last = list_empty(&wq->pwqs);
3523 mutex_unlock(&wq->mutex);
3525 mutex_lock(&wq_pool_mutex);
3526 put_unbound_pool(pool);
3527 mutex_unlock(&wq_pool_mutex);
3529 call_rcu(&pwq->rcu, rcu_free_pwq);
3532 * If we're the last pwq going away, @wq is already dead and no one
3533 * is gonna access it anymore. Schedule RCU free.
3536 call_rcu(&wq->rcu, rcu_free_wq);
3540 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3541 * @pwq: target pool_workqueue
3543 * If @pwq isn't freezing, set @pwq->max_active to the associated
3544 * workqueue's saved_max_active and activate delayed work items
3545 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3547 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3549 struct workqueue_struct *wq = pwq->wq;
3550 bool freezable = wq->flags & WQ_FREEZABLE;
3551 unsigned long flags;
3553 /* for @wq->saved_max_active */
3554 lockdep_assert_held(&wq->mutex);
3556 /* fast exit for non-freezable wqs */
3557 if (!freezable && pwq->max_active == wq->saved_max_active)
3560 /* this function can be called during early boot w/ irq disabled */
3561 spin_lock_irqsave(&pwq->pool->lock, flags);
3564 * During [un]freezing, the caller is responsible for ensuring that
3565 * this function is called at least once after @workqueue_freezing
3566 * is updated and visible.
3568 if (!freezable || !workqueue_freezing) {
3569 pwq->max_active = wq->saved_max_active;
3571 while (!list_empty(&pwq->delayed_works) &&
3572 pwq->nr_active < pwq->max_active)
3573 pwq_activate_first_delayed(pwq);
3576 * Need to kick a worker after thawed or an unbound wq's
3577 * max_active is bumped. It's a slow path. Do it always.
3579 wake_up_worker(pwq->pool);
3581 pwq->max_active = 0;
3584 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3587 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3588 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3589 struct worker_pool *pool)
3591 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3593 memset(pwq, 0, sizeof(*pwq));
3597 pwq->flush_color = -1;
3599 INIT_LIST_HEAD(&pwq->delayed_works);
3600 INIT_LIST_HEAD(&pwq->pwqs_node);
3601 INIT_LIST_HEAD(&pwq->mayday_node);
3602 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3605 /* sync @pwq with the current state of its associated wq and link it */
3606 static void link_pwq(struct pool_workqueue *pwq)
3608 struct workqueue_struct *wq = pwq->wq;
3610 lockdep_assert_held(&wq->mutex);
3612 /* may be called multiple times, ignore if already linked */
3613 if (!list_empty(&pwq->pwqs_node))
3616 /* set the matching work_color */
3617 pwq->work_color = wq->work_color;
3619 /* sync max_active to the current setting */
3620 pwq_adjust_max_active(pwq);
3623 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3626 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3627 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3628 const struct workqueue_attrs *attrs)
3630 struct worker_pool *pool;
3631 struct pool_workqueue *pwq;
3633 lockdep_assert_held(&wq_pool_mutex);
3635 pool = get_unbound_pool(attrs);
3639 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3641 put_unbound_pool(pool);
3645 init_pwq(pwq, wq, pool);
3650 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3651 * @attrs: the wq_attrs of the default pwq of the target workqueue
3652 * @node: the target NUMA node
3653 * @cpu_going_down: if >= 0, the CPU to consider as offline
3654 * @cpumask: outarg, the resulting cpumask
3656 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3657 * @cpu_going_down is >= 0, that cpu is considered offline during
3658 * calculation. The result is stored in @cpumask.
3660 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3661 * enabled and @node has online CPUs requested by @attrs, the returned
3662 * cpumask is the intersection of the possible CPUs of @node and
3665 * The caller is responsible for ensuring that the cpumask of @node stays
3668 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3671 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3672 int cpu_going_down, cpumask_t *cpumask)
3674 if (!wq_numa_enabled || attrs->no_numa)
3677 /* does @node have any online CPUs @attrs wants? */
3678 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3679 if (cpu_going_down >= 0)
3680 cpumask_clear_cpu(cpu_going_down, cpumask);
3682 if (cpumask_empty(cpumask))
3685 /* yeap, return possible CPUs in @node that @attrs wants */
3686 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3688 if (cpumask_empty(cpumask)) {
3689 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3690 "possible intersect\n");
3694 return !cpumask_equal(cpumask, attrs->cpumask);
3697 cpumask_copy(cpumask, attrs->cpumask);
3701 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3702 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3704 struct pool_workqueue *pwq)
3706 struct pool_workqueue *old_pwq;
3708 lockdep_assert_held(&wq_pool_mutex);
3709 lockdep_assert_held(&wq->mutex);
3711 /* link_pwq() can handle duplicate calls */
3714 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3715 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3719 /* context to store the prepared attrs & pwqs before applying */
3720 struct apply_wqattrs_ctx {
3721 struct workqueue_struct *wq; /* target workqueue */
3722 struct workqueue_attrs *attrs; /* attrs to apply */
3723 struct list_head list; /* queued for batching commit */
3724 struct pool_workqueue *dfl_pwq;
3725 struct pool_workqueue *pwq_tbl[];
3728 /* free the resources after success or abort */
3729 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3735 put_pwq_unlocked(ctx->pwq_tbl[node]);
3736 put_pwq_unlocked(ctx->dfl_pwq);
3738 free_workqueue_attrs(ctx->attrs);
3744 /* allocate the attrs and pwqs for later installation */
3745 static struct apply_wqattrs_ctx *
3746 apply_wqattrs_prepare(struct workqueue_struct *wq,
3747 const struct workqueue_attrs *attrs)
3749 struct apply_wqattrs_ctx *ctx;
3750 struct workqueue_attrs *new_attrs, *tmp_attrs;
3753 lockdep_assert_held(&wq_pool_mutex);
3755 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3757 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3758 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3759 if (!ctx || !new_attrs || !tmp_attrs)
3763 * Calculate the attrs of the default pwq.
3764 * If the user configured cpumask doesn't overlap with the
3765 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3767 copy_workqueue_attrs(new_attrs, attrs);
3768 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3769 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3770 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3773 * We may create multiple pwqs with differing cpumasks. Make a
3774 * copy of @new_attrs which will be modified and used to obtain
3777 copy_workqueue_attrs(tmp_attrs, new_attrs);
3780 * If something goes wrong during CPU up/down, we'll fall back to
3781 * the default pwq covering whole @attrs->cpumask. Always create
3782 * it even if we don't use it immediately.
3784 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3788 for_each_node(node) {
3789 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3790 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3791 if (!ctx->pwq_tbl[node])
3794 ctx->dfl_pwq->refcnt++;
3795 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3799 /* save the user configured attrs and sanitize it. */
3800 copy_workqueue_attrs(new_attrs, attrs);
3801 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3802 ctx->attrs = new_attrs;
3805 free_workqueue_attrs(tmp_attrs);
3809 free_workqueue_attrs(tmp_attrs);
3810 free_workqueue_attrs(new_attrs);
3811 apply_wqattrs_cleanup(ctx);
3815 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3816 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3820 /* all pwqs have been created successfully, let's install'em */
3821 mutex_lock(&ctx->wq->mutex);
3823 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3825 /* save the previous pwq and install the new one */
3827 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3828 ctx->pwq_tbl[node]);
3830 /* @dfl_pwq might not have been used, ensure it's linked */
3831 link_pwq(ctx->dfl_pwq);
3832 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3834 mutex_unlock(&ctx->wq->mutex);
3837 static void apply_wqattrs_lock(void)
3839 /* CPUs should stay stable across pwq creations and installations */
3841 mutex_lock(&wq_pool_mutex);
3844 static void apply_wqattrs_unlock(void)
3846 mutex_unlock(&wq_pool_mutex);
3850 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3851 const struct workqueue_attrs *attrs)
3853 struct apply_wqattrs_ctx *ctx;
3855 /* only unbound workqueues can change attributes */
3856 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3859 /* creating multiple pwqs breaks ordering guarantee */
3860 if (!list_empty(&wq->pwqs)) {
3861 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3864 wq->flags &= ~__WQ_ORDERED;
3867 ctx = apply_wqattrs_prepare(wq, attrs);
3871 /* the ctx has been prepared successfully, let's commit it */
3872 apply_wqattrs_commit(ctx);
3873 apply_wqattrs_cleanup(ctx);
3879 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3880 * @wq: the target workqueue
3881 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3883 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3884 * machines, this function maps a separate pwq to each NUMA node with
3885 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3886 * NUMA node it was issued on. Older pwqs are released as in-flight work
3887 * items finish. Note that a work item which repeatedly requeues itself
3888 * back-to-back will stay on its current pwq.
3890 * Performs GFP_KERNEL allocations.
3892 * Return: 0 on success and -errno on failure.
3894 int apply_workqueue_attrs(struct workqueue_struct *wq,
3895 const struct workqueue_attrs *attrs)
3899 apply_wqattrs_lock();
3900 ret = apply_workqueue_attrs_locked(wq, attrs);
3901 apply_wqattrs_unlock();
3905 EXPORT_SYMBOL_GPL(apply_workqueue_attrs);
3908 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3909 * @wq: the target workqueue
3910 * @cpu: the CPU coming up or going down
3911 * @online: whether @cpu is coming up or going down
3913 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3914 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3917 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3918 * falls back to @wq->dfl_pwq which may not be optimal but is always
3921 * Note that when the last allowed CPU of a NUMA node goes offline for a
3922 * workqueue with a cpumask spanning multiple nodes, the workers which were
3923 * already executing the work items for the workqueue will lose their CPU
3924 * affinity and may execute on any CPU. This is similar to how per-cpu
3925 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3926 * affinity, it's the user's responsibility to flush the work item from
3929 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3932 int node = cpu_to_node(cpu);
3933 int cpu_off = online ? -1 : cpu;
3934 struct pool_workqueue *old_pwq = NULL, *pwq;
3935 struct workqueue_attrs *target_attrs;
3938 lockdep_assert_held(&wq_pool_mutex);
3940 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3941 wq->unbound_attrs->no_numa)
3945 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3946 * Let's use a preallocated one. The following buf is protected by
3947 * CPU hotplug exclusion.
3949 target_attrs = wq_update_unbound_numa_attrs_buf;
3950 cpumask = target_attrs->cpumask;
3952 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3953 pwq = unbound_pwq_by_node(wq, node);
3956 * Let's determine what needs to be done. If the target cpumask is
3957 * different from the default pwq's, we need to compare it to @pwq's
3958 * and create a new one if they don't match. If the target cpumask
3959 * equals the default pwq's, the default pwq should be used.
3961 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3962 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3968 /* create a new pwq */
3969 pwq = alloc_unbound_pwq(wq, target_attrs);
3971 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3976 /* Install the new pwq. */
3977 mutex_lock(&wq->mutex);
3978 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3982 mutex_lock(&wq->mutex);
3983 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3984 get_pwq(wq->dfl_pwq);
3985 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3986 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3988 mutex_unlock(&wq->mutex);
3989 put_pwq_unlocked(old_pwq);
3992 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3994 bool highpri = wq->flags & WQ_HIGHPRI;
3997 if (!(wq->flags & WQ_UNBOUND)) {
3998 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4002 for_each_possible_cpu(cpu) {
4003 struct pool_workqueue *pwq =
4004 per_cpu_ptr(wq->cpu_pwqs, cpu);
4005 struct worker_pool *cpu_pools =
4006 per_cpu(cpu_worker_pools, cpu);
4008 init_pwq(pwq, wq, &cpu_pools[highpri]);
4010 mutex_lock(&wq->mutex);
4012 mutex_unlock(&wq->mutex);
4015 } else if (wq->flags & __WQ_ORDERED) {
4016 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4017 /* there should only be single pwq for ordering guarantee */
4018 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4019 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4020 "ordering guarantee broken for workqueue %s\n", wq->name);
4023 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4027 static int wq_clamp_max_active(int max_active, unsigned int flags,
4030 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4032 if (max_active < 1 || max_active > lim)
4033 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4034 max_active, name, 1, lim);
4036 return clamp_val(max_active, 1, lim);
4040 * Workqueues which may be used during memory reclaim should have a rescuer
4041 * to guarantee forward progress.
4043 static int init_rescuer(struct workqueue_struct *wq)
4045 struct worker *rescuer;
4048 if (!(wq->flags & WQ_MEM_RECLAIM))
4051 rescuer = alloc_worker(NUMA_NO_NODE);
4055 rescuer->rescue_wq = wq;
4056 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4057 ret = PTR_ERR_OR_ZERO(rescuer->task);
4063 wq->rescuer = rescuer;
4064 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4065 wake_up_process(rescuer->task);
4070 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4073 struct lock_class_key *key,
4074 const char *lock_name, ...)
4076 size_t tbl_size = 0;
4078 struct workqueue_struct *wq;
4079 struct pool_workqueue *pwq;
4082 * Unbound && max_active == 1 used to imply ordered, which is no
4083 * longer the case on NUMA machines due to per-node pools. While
4084 * alloc_ordered_workqueue() is the right way to create an ordered
4085 * workqueue, keep the previous behavior to avoid subtle breakages
4088 if ((flags & WQ_UNBOUND) && max_active == 1)
4089 flags |= __WQ_ORDERED;
4091 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4092 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4093 flags |= WQ_UNBOUND;
4095 /* allocate wq and format name */
4096 if (flags & WQ_UNBOUND)
4097 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4099 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4103 if (flags & WQ_UNBOUND) {
4104 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4105 if (!wq->unbound_attrs)
4109 va_start(args, lock_name);
4110 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4113 max_active = max_active ?: WQ_DFL_ACTIVE;
4114 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4118 wq->saved_max_active = max_active;
4119 mutex_init(&wq->mutex);
4120 atomic_set(&wq->nr_pwqs_to_flush, 0);
4121 INIT_LIST_HEAD(&wq->pwqs);
4122 INIT_LIST_HEAD(&wq->flusher_queue);
4123 INIT_LIST_HEAD(&wq->flusher_overflow);
4124 INIT_LIST_HEAD(&wq->maydays);
4126 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4127 INIT_LIST_HEAD(&wq->list);
4129 if (alloc_and_link_pwqs(wq) < 0)
4132 if (wq_online && init_rescuer(wq) < 0)
4135 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4139 * wq_pool_mutex protects global freeze state and workqueues list.
4140 * Grab it, adjust max_active and add the new @wq to workqueues
4143 mutex_lock(&wq_pool_mutex);
4145 mutex_lock(&wq->mutex);
4146 for_each_pwq(pwq, wq)
4147 pwq_adjust_max_active(pwq);
4148 mutex_unlock(&wq->mutex);
4150 list_add_tail_rcu(&wq->list, &workqueues);
4152 mutex_unlock(&wq_pool_mutex);
4157 free_workqueue_attrs(wq->unbound_attrs);
4161 destroy_workqueue(wq);
4164 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4167 * destroy_workqueue - safely terminate a workqueue
4168 * @wq: target workqueue
4170 * Safely destroy a workqueue. All work currently pending will be done first.
4172 void destroy_workqueue(struct workqueue_struct *wq)
4174 struct pool_workqueue *pwq;
4177 /* drain it before proceeding with destruction */
4178 drain_workqueue(wq);
4181 mutex_lock(&wq->mutex);
4182 for_each_pwq(pwq, wq) {
4185 for (i = 0; i < WORK_NR_COLORS; i++) {
4186 if (WARN_ON(pwq->nr_in_flight[i])) {
4187 mutex_unlock(&wq->mutex);
4188 show_workqueue_state();
4193 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4194 WARN_ON(pwq->nr_active) ||
4195 WARN_ON(!list_empty(&pwq->delayed_works))) {
4196 mutex_unlock(&wq->mutex);
4197 show_workqueue_state();
4201 mutex_unlock(&wq->mutex);
4204 * wq list is used to freeze wq, remove from list after
4205 * flushing is complete in case freeze races us.
4207 mutex_lock(&wq_pool_mutex);
4208 list_del_rcu(&wq->list);
4209 mutex_unlock(&wq_pool_mutex);
4211 workqueue_sysfs_unregister(wq);
4214 kthread_stop(wq->rescuer->task);
4216 if (!(wq->flags & WQ_UNBOUND)) {
4218 * The base ref is never dropped on per-cpu pwqs. Directly
4219 * schedule RCU free.
4221 call_rcu(&wq->rcu, rcu_free_wq);
4224 * We're the sole accessor of @wq at this point. Directly
4225 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4226 * @wq will be freed when the last pwq is released.
4228 for_each_node(node) {
4229 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4230 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4231 put_pwq_unlocked(pwq);
4235 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4236 * put. Don't access it afterwards.
4240 put_pwq_unlocked(pwq);
4243 EXPORT_SYMBOL_GPL(destroy_workqueue);
4246 * workqueue_set_max_active - adjust max_active of a workqueue
4247 * @wq: target workqueue
4248 * @max_active: new max_active value.
4250 * Set max_active of @wq to @max_active.
4253 * Don't call from IRQ context.
4255 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4257 struct pool_workqueue *pwq;
4259 /* disallow meddling with max_active for ordered workqueues */
4260 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4263 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4265 mutex_lock(&wq->mutex);
4267 wq->flags &= ~__WQ_ORDERED;
4268 wq->saved_max_active = max_active;
4270 for_each_pwq(pwq, wq)
4271 pwq_adjust_max_active(pwq);
4273 mutex_unlock(&wq->mutex);
4275 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4278 * current_work - retrieve %current task's work struct
4280 * Determine if %current task is a workqueue worker and what it's working on.
4281 * Useful to find out the context that the %current task is running in.
4283 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4285 struct work_struct *current_work(void)
4287 struct worker *worker = current_wq_worker();
4289 return worker ? worker->current_work : NULL;
4291 EXPORT_SYMBOL(current_work);
4294 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4296 * Determine whether %current is a workqueue rescuer. Can be used from
4297 * work functions to determine whether it's being run off the rescuer task.
4299 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4301 bool current_is_workqueue_rescuer(void)
4303 struct worker *worker = current_wq_worker();
4305 return worker && worker->rescue_wq;
4309 * workqueue_congested - test whether a workqueue is congested
4310 * @cpu: CPU in question
4311 * @wq: target workqueue
4313 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4314 * no synchronization around this function and the test result is
4315 * unreliable and only useful as advisory hints or for debugging.
4317 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4318 * Note that both per-cpu and unbound workqueues may be associated with
4319 * multiple pool_workqueues which have separate congested states. A
4320 * workqueue being congested on one CPU doesn't mean the workqueue is also
4321 * contested on other CPUs / NUMA nodes.
4324 * %true if congested, %false otherwise.
4326 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4328 struct pool_workqueue *pwq;
4331 rcu_read_lock_sched();
4333 if (cpu == WORK_CPU_UNBOUND)
4334 cpu = smp_processor_id();
4336 if (!(wq->flags & WQ_UNBOUND))
4337 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4339 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4341 ret = !list_empty(&pwq->delayed_works);
4342 rcu_read_unlock_sched();
4346 EXPORT_SYMBOL_GPL(workqueue_congested);
4349 * work_busy - test whether a work is currently pending or running
4350 * @work: the work to be tested
4352 * Test whether @work is currently pending or running. There is no
4353 * synchronization around this function and the test result is
4354 * unreliable and only useful as advisory hints or for debugging.
4357 * OR'd bitmask of WORK_BUSY_* bits.
4359 unsigned int work_busy(struct work_struct *work)
4361 struct worker_pool *pool;
4362 unsigned long flags;
4363 unsigned int ret = 0;
4365 if (work_pending(work))
4366 ret |= WORK_BUSY_PENDING;
4368 local_irq_save(flags);
4369 pool = get_work_pool(work);
4371 spin_lock(&pool->lock);
4372 if (find_worker_executing_work(pool, work))
4373 ret |= WORK_BUSY_RUNNING;
4374 spin_unlock(&pool->lock);
4376 local_irq_restore(flags);
4380 EXPORT_SYMBOL_GPL(work_busy);
4383 * set_worker_desc - set description for the current work item
4384 * @fmt: printf-style format string
4385 * @...: arguments for the format string
4387 * This function can be called by a running work function to describe what
4388 * the work item is about. If the worker task gets dumped, this
4389 * information will be printed out together to help debugging. The
4390 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4392 void set_worker_desc(const char *fmt, ...)
4394 struct worker *worker = current_wq_worker();
4398 va_start(args, fmt);
4399 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4403 EXPORT_SYMBOL_GPL(set_worker_desc);
4406 * print_worker_info - print out worker information and description
4407 * @log_lvl: the log level to use when printing
4408 * @task: target task
4410 * If @task is a worker and currently executing a work item, print out the
4411 * name of the workqueue being serviced and worker description set with
4412 * set_worker_desc() by the currently executing work item.
4414 * This function can be safely called on any task as long as the
4415 * task_struct itself is accessible. While safe, this function isn't
4416 * synchronized and may print out mixups or garbages of limited length.
4418 void print_worker_info(const char *log_lvl, struct task_struct *task)
4420 work_func_t *fn = NULL;
4421 char name[WQ_NAME_LEN] = { };
4422 char desc[WORKER_DESC_LEN] = { };
4423 struct pool_workqueue *pwq = NULL;
4424 struct workqueue_struct *wq = NULL;
4425 struct worker *worker;
4427 if (!(task->flags & PF_WQ_WORKER))
4431 * This function is called without any synchronization and @task
4432 * could be in any state. Be careful with dereferences.
4434 worker = kthread_probe_data(task);
4437 * Carefully copy the associated workqueue's workfn, name and desc.
4438 * Keep the original last '\0' in case the original is garbage.
4440 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4441 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4442 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4443 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4444 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4446 if (fn || name[0] || desc[0]) {
4447 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4448 if (strcmp(name, desc))
4449 pr_cont(" (%s)", desc);
4454 static void pr_cont_pool_info(struct worker_pool *pool)
4456 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4457 if (pool->node != NUMA_NO_NODE)
4458 pr_cont(" node=%d", pool->node);
4459 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4462 static void pr_cont_work(bool comma, struct work_struct *work)
4464 if (work->func == wq_barrier_func) {
4465 struct wq_barrier *barr;
4467 barr = container_of(work, struct wq_barrier, work);
4469 pr_cont("%s BAR(%d)", comma ? "," : "",
4470 task_pid_nr(barr->task));
4472 pr_cont("%s %pf", comma ? "," : "", work->func);
4476 static void show_pwq(struct pool_workqueue *pwq)
4478 struct worker_pool *pool = pwq->pool;
4479 struct work_struct *work;
4480 struct worker *worker;
4481 bool has_in_flight = false, has_pending = false;
4484 pr_info(" pwq %d:", pool->id);
4485 pr_cont_pool_info(pool);
4487 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4488 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4490 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4491 if (worker->current_pwq == pwq) {
4492 has_in_flight = true;
4496 if (has_in_flight) {
4499 pr_info(" in-flight:");
4500 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4501 if (worker->current_pwq != pwq)
4504 pr_cont("%s %d%s:%pf", comma ? "," : "",
4505 task_pid_nr(worker->task),
4506 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4507 worker->current_func);
4508 list_for_each_entry(work, &worker->scheduled, entry)
4509 pr_cont_work(false, work);
4515 list_for_each_entry(work, &pool->worklist, entry) {
4516 if (get_work_pwq(work) == pwq) {
4524 pr_info(" pending:");
4525 list_for_each_entry(work, &pool->worklist, entry) {
4526 if (get_work_pwq(work) != pwq)
4529 pr_cont_work(comma, work);
4530 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4535 if (!list_empty(&pwq->delayed_works)) {
4538 pr_info(" delayed:");
4539 list_for_each_entry(work, &pwq->delayed_works, entry) {
4540 pr_cont_work(comma, work);
4541 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4548 * show_workqueue_state - dump workqueue state
4550 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4551 * all busy workqueues and pools.
4553 void show_workqueue_state(void)
4555 struct workqueue_struct *wq;
4556 struct worker_pool *pool;
4557 unsigned long flags;
4560 rcu_read_lock_sched();
4562 pr_info("Showing busy workqueues and worker pools:\n");
4564 list_for_each_entry_rcu(wq, &workqueues, list) {
4565 struct pool_workqueue *pwq;
4568 for_each_pwq(pwq, wq) {
4569 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4577 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4579 for_each_pwq(pwq, wq) {
4580 spin_lock_irqsave(&pwq->pool->lock, flags);
4581 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4583 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4585 * We could be printing a lot from atomic context, e.g.
4586 * sysrq-t -> show_workqueue_state(). Avoid triggering
4589 touch_nmi_watchdog();
4593 for_each_pool(pool, pi) {
4594 struct worker *worker;
4597 spin_lock_irqsave(&pool->lock, flags);
4598 if (pool->nr_workers == pool->nr_idle)
4601 pr_info("pool %d:", pool->id);
4602 pr_cont_pool_info(pool);
4603 pr_cont(" hung=%us workers=%d",
4604 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4607 pr_cont(" manager: %d",
4608 task_pid_nr(pool->manager->task));
4609 list_for_each_entry(worker, &pool->idle_list, entry) {
4610 pr_cont(" %s%d", first ? "idle: " : "",
4611 task_pid_nr(worker->task));
4616 spin_unlock_irqrestore(&pool->lock, flags);
4618 * We could be printing a lot from atomic context, e.g.
4619 * sysrq-t -> show_workqueue_state(). Avoid triggering
4622 touch_nmi_watchdog();
4625 rcu_read_unlock_sched();
4628 /* used to show worker information through /proc/PID/{comm,stat,status} */
4629 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4633 /* always show the actual comm */
4634 off = strscpy(buf, task->comm, size);
4638 /* stabilize PF_WQ_WORKER and worker pool association */
4639 mutex_lock(&wq_pool_attach_mutex);
4641 if (task->flags & PF_WQ_WORKER) {
4642 struct worker *worker = kthread_data(task);
4643 struct worker_pool *pool = worker->pool;
4646 spin_lock_irq(&pool->lock);
4648 * ->desc tracks information (wq name or
4649 * set_worker_desc()) for the latest execution. If
4650 * current, prepend '+', otherwise '-'.
4652 if (worker->desc[0] != '\0') {
4653 if (worker->current_work)
4654 scnprintf(buf + off, size - off, "+%s",
4657 scnprintf(buf + off, size - off, "-%s",
4660 spin_unlock_irq(&pool->lock);
4664 mutex_unlock(&wq_pool_attach_mutex);
4672 * There are two challenges in supporting CPU hotplug. Firstly, there
4673 * are a lot of assumptions on strong associations among work, pwq and
4674 * pool which make migrating pending and scheduled works very
4675 * difficult to implement without impacting hot paths. Secondly,
4676 * worker pools serve mix of short, long and very long running works making
4677 * blocked draining impractical.
4679 * This is solved by allowing the pools to be disassociated from the CPU
4680 * running as an unbound one and allowing it to be reattached later if the
4681 * cpu comes back online.
4684 static void unbind_workers(int cpu)
4686 struct worker_pool *pool;
4687 struct worker *worker;
4689 for_each_cpu_worker_pool(pool, cpu) {
4690 mutex_lock(&wq_pool_attach_mutex);
4691 spin_lock_irq(&pool->lock);
4694 * We've blocked all attach/detach operations. Make all workers
4695 * unbound and set DISASSOCIATED. Before this, all workers
4696 * except for the ones which are still executing works from
4697 * before the last CPU down must be on the cpu. After
4698 * this, they may become diasporas.
4700 for_each_pool_worker(worker, pool)
4701 worker->flags |= WORKER_UNBOUND;
4703 pool->flags |= POOL_DISASSOCIATED;
4705 spin_unlock_irq(&pool->lock);
4706 mutex_unlock(&wq_pool_attach_mutex);
4709 * Call schedule() so that we cross rq->lock and thus can
4710 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4711 * This is necessary as scheduler callbacks may be invoked
4717 * Sched callbacks are disabled now. Zap nr_running.
4718 * After this, nr_running stays zero and need_more_worker()
4719 * and keep_working() are always true as long as the
4720 * worklist is not empty. This pool now behaves as an
4721 * unbound (in terms of concurrency management) pool which
4722 * are served by workers tied to the pool.
4724 atomic_set(&pool->nr_running, 0);
4727 * With concurrency management just turned off, a busy
4728 * worker blocking could lead to lengthy stalls. Kick off
4729 * unbound chain execution of currently pending work items.
4731 spin_lock_irq(&pool->lock);
4732 wake_up_worker(pool);
4733 spin_unlock_irq(&pool->lock);
4738 * rebind_workers - rebind all workers of a pool to the associated CPU
4739 * @pool: pool of interest
4741 * @pool->cpu is coming online. Rebind all workers to the CPU.
4743 static void rebind_workers(struct worker_pool *pool)
4745 struct worker *worker;
4747 lockdep_assert_held(&wq_pool_attach_mutex);
4750 * Restore CPU affinity of all workers. As all idle workers should
4751 * be on the run-queue of the associated CPU before any local
4752 * wake-ups for concurrency management happen, restore CPU affinity
4753 * of all workers first and then clear UNBOUND. As we're called
4754 * from CPU_ONLINE, the following shouldn't fail.
4756 for_each_pool_worker(worker, pool)
4757 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4758 pool->attrs->cpumask) < 0);
4760 spin_lock_irq(&pool->lock);
4762 pool->flags &= ~POOL_DISASSOCIATED;
4764 for_each_pool_worker(worker, pool) {
4765 unsigned int worker_flags = worker->flags;
4768 * A bound idle worker should actually be on the runqueue
4769 * of the associated CPU for local wake-ups targeting it to
4770 * work. Kick all idle workers so that they migrate to the
4771 * associated CPU. Doing this in the same loop as
4772 * replacing UNBOUND with REBOUND is safe as no worker will
4773 * be bound before @pool->lock is released.
4775 if (worker_flags & WORKER_IDLE)
4776 wake_up_process(worker->task);
4779 * We want to clear UNBOUND but can't directly call
4780 * worker_clr_flags() or adjust nr_running. Atomically
4781 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4782 * @worker will clear REBOUND using worker_clr_flags() when
4783 * it initiates the next execution cycle thus restoring
4784 * concurrency management. Note that when or whether
4785 * @worker clears REBOUND doesn't affect correctness.
4787 * WRITE_ONCE() is necessary because @worker->flags may be
4788 * tested without holding any lock in
4789 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4790 * fail incorrectly leading to premature concurrency
4791 * management operations.
4793 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4794 worker_flags |= WORKER_REBOUND;
4795 worker_flags &= ~WORKER_UNBOUND;
4796 WRITE_ONCE(worker->flags, worker_flags);
4799 spin_unlock_irq(&pool->lock);
4803 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4804 * @pool: unbound pool of interest
4805 * @cpu: the CPU which is coming up
4807 * An unbound pool may end up with a cpumask which doesn't have any online
4808 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4809 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4810 * online CPU before, cpus_allowed of all its workers should be restored.
4812 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4814 static cpumask_t cpumask;
4815 struct worker *worker;
4817 lockdep_assert_held(&wq_pool_attach_mutex);
4819 /* is @cpu allowed for @pool? */
4820 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4823 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4825 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4826 for_each_pool_worker(worker, pool)
4827 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4830 int workqueue_prepare_cpu(unsigned int cpu)
4832 struct worker_pool *pool;
4834 for_each_cpu_worker_pool(pool, cpu) {
4835 if (pool->nr_workers)
4837 if (!create_worker(pool))
4843 int workqueue_online_cpu(unsigned int cpu)
4845 struct worker_pool *pool;
4846 struct workqueue_struct *wq;
4849 mutex_lock(&wq_pool_mutex);
4851 for_each_pool(pool, pi) {
4852 mutex_lock(&wq_pool_attach_mutex);
4854 if (pool->cpu == cpu)
4855 rebind_workers(pool);
4856 else if (pool->cpu < 0)
4857 restore_unbound_workers_cpumask(pool, cpu);
4859 mutex_unlock(&wq_pool_attach_mutex);
4862 /* update NUMA affinity of unbound workqueues */
4863 list_for_each_entry(wq, &workqueues, list)
4864 wq_update_unbound_numa(wq, cpu, true);
4866 mutex_unlock(&wq_pool_mutex);
4870 int workqueue_offline_cpu(unsigned int cpu)
4872 struct workqueue_struct *wq;
4874 /* unbinding per-cpu workers should happen on the local CPU */
4875 if (WARN_ON(cpu != smp_processor_id()))
4878 unbind_workers(cpu);
4880 /* update NUMA affinity of unbound workqueues */
4881 mutex_lock(&wq_pool_mutex);
4882 list_for_each_entry(wq, &workqueues, list)
4883 wq_update_unbound_numa(wq, cpu, false);
4884 mutex_unlock(&wq_pool_mutex);
4889 struct work_for_cpu {
4890 struct work_struct work;
4896 static void work_for_cpu_fn(struct work_struct *work)
4898 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4900 wfc->ret = wfc->fn(wfc->arg);
4904 * work_on_cpu - run a function in thread context on a particular cpu
4905 * @cpu: the cpu to run on
4906 * @fn: the function to run
4907 * @arg: the function arg
4909 * It is up to the caller to ensure that the cpu doesn't go offline.
4910 * The caller must not hold any locks which would prevent @fn from completing.
4912 * Return: The value @fn returns.
4914 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4916 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4918 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4919 schedule_work_on(cpu, &wfc.work);
4920 flush_work(&wfc.work);
4921 destroy_work_on_stack(&wfc.work);
4924 EXPORT_SYMBOL_GPL(work_on_cpu);
4927 * work_on_cpu_safe - run a function in thread context on a particular cpu
4928 * @cpu: the cpu to run on
4929 * @fn: the function to run
4930 * @arg: the function argument
4932 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4933 * any locks which would prevent @fn from completing.
4935 * Return: The value @fn returns.
4937 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4942 if (cpu_online(cpu))
4943 ret = work_on_cpu(cpu, fn, arg);
4947 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
4948 #endif /* CONFIG_SMP */
4950 #ifdef CONFIG_FREEZER
4953 * freeze_workqueues_begin - begin freezing workqueues
4955 * Start freezing workqueues. After this function returns, all freezable
4956 * workqueues will queue new works to their delayed_works list instead of
4960 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4962 void freeze_workqueues_begin(void)
4964 struct workqueue_struct *wq;
4965 struct pool_workqueue *pwq;
4967 mutex_lock(&wq_pool_mutex);
4969 WARN_ON_ONCE(workqueue_freezing);
4970 workqueue_freezing = true;
4972 list_for_each_entry(wq, &workqueues, list) {
4973 mutex_lock(&wq->mutex);
4974 for_each_pwq(pwq, wq)
4975 pwq_adjust_max_active(pwq);
4976 mutex_unlock(&wq->mutex);
4979 mutex_unlock(&wq_pool_mutex);
4983 * freeze_workqueues_busy - are freezable workqueues still busy?
4985 * Check whether freezing is complete. This function must be called
4986 * between freeze_workqueues_begin() and thaw_workqueues().
4989 * Grabs and releases wq_pool_mutex.
4992 * %true if some freezable workqueues are still busy. %false if freezing
4995 bool freeze_workqueues_busy(void)
4998 struct workqueue_struct *wq;
4999 struct pool_workqueue *pwq;
5001 mutex_lock(&wq_pool_mutex);
5003 WARN_ON_ONCE(!workqueue_freezing);
5005 list_for_each_entry(wq, &workqueues, list) {
5006 if (!(wq->flags & WQ_FREEZABLE))
5009 * nr_active is monotonically decreasing. It's safe
5010 * to peek without lock.
5012 rcu_read_lock_sched();
5013 for_each_pwq(pwq, wq) {
5014 WARN_ON_ONCE(pwq->nr_active < 0);
5015 if (pwq->nr_active) {
5017 rcu_read_unlock_sched();
5021 rcu_read_unlock_sched();
5024 mutex_unlock(&wq_pool_mutex);
5029 * thaw_workqueues - thaw workqueues
5031 * Thaw workqueues. Normal queueing is restored and all collected
5032 * frozen works are transferred to their respective pool worklists.
5035 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5037 void thaw_workqueues(void)
5039 struct workqueue_struct *wq;
5040 struct pool_workqueue *pwq;
5042 mutex_lock(&wq_pool_mutex);
5044 if (!workqueue_freezing)
5047 workqueue_freezing = false;
5049 /* restore max_active and repopulate worklist */
5050 list_for_each_entry(wq, &workqueues, list) {
5051 mutex_lock(&wq->mutex);
5052 for_each_pwq(pwq, wq)
5053 pwq_adjust_max_active(pwq);
5054 mutex_unlock(&wq->mutex);
5058 mutex_unlock(&wq_pool_mutex);
5060 #endif /* CONFIG_FREEZER */
5062 static int workqueue_apply_unbound_cpumask(void)
5066 struct workqueue_struct *wq;
5067 struct apply_wqattrs_ctx *ctx, *n;
5069 lockdep_assert_held(&wq_pool_mutex);
5071 list_for_each_entry(wq, &workqueues, list) {
5072 if (!(wq->flags & WQ_UNBOUND))
5074 /* creating multiple pwqs breaks ordering guarantee */
5075 if (wq->flags & __WQ_ORDERED)
5078 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5084 list_add_tail(&ctx->list, &ctxs);
5087 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5089 apply_wqattrs_commit(ctx);
5090 apply_wqattrs_cleanup(ctx);
5097 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5098 * @cpumask: the cpumask to set
5100 * The low-level workqueues cpumask is a global cpumask that limits
5101 * the affinity of all unbound workqueues. This function check the @cpumask
5102 * and apply it to all unbound workqueues and updates all pwqs of them.
5104 * Retun: 0 - Success
5105 * -EINVAL - Invalid @cpumask
5106 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5108 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5111 cpumask_var_t saved_cpumask;
5113 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5117 * Not excluding isolated cpus on purpose.
5118 * If the user wishes to include them, we allow that.
5120 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5121 if (!cpumask_empty(cpumask)) {
5122 apply_wqattrs_lock();
5124 /* save the old wq_unbound_cpumask. */
5125 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5127 /* update wq_unbound_cpumask at first and apply it to wqs. */
5128 cpumask_copy(wq_unbound_cpumask, cpumask);
5129 ret = workqueue_apply_unbound_cpumask();
5131 /* restore the wq_unbound_cpumask when failed. */
5133 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5135 apply_wqattrs_unlock();
5138 free_cpumask_var(saved_cpumask);
5144 * Workqueues with WQ_SYSFS flag set is visible to userland via
5145 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5146 * following attributes.
5148 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5149 * max_active RW int : maximum number of in-flight work items
5151 * Unbound workqueues have the following extra attributes.
5153 * pool_ids RO int : the associated pool IDs for each node
5154 * nice RW int : nice value of the workers
5155 * cpumask RW mask : bitmask of allowed CPUs for the workers
5156 * numa RW bool : whether enable NUMA affinity
5159 struct workqueue_struct *wq;
5163 static struct workqueue_struct *dev_to_wq(struct device *dev)
5165 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5170 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5173 struct workqueue_struct *wq = dev_to_wq(dev);
5175 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5177 static DEVICE_ATTR_RO(per_cpu);
5179 static ssize_t max_active_show(struct device *dev,
5180 struct device_attribute *attr, char *buf)
5182 struct workqueue_struct *wq = dev_to_wq(dev);
5184 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5187 static ssize_t max_active_store(struct device *dev,
5188 struct device_attribute *attr, const char *buf,
5191 struct workqueue_struct *wq = dev_to_wq(dev);
5194 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5197 workqueue_set_max_active(wq, val);
5200 static DEVICE_ATTR_RW(max_active);
5202 static struct attribute *wq_sysfs_attrs[] = {
5203 &dev_attr_per_cpu.attr,
5204 &dev_attr_max_active.attr,
5207 ATTRIBUTE_GROUPS(wq_sysfs);
5209 static ssize_t wq_pool_ids_show(struct device *dev,
5210 struct device_attribute *attr, char *buf)
5212 struct workqueue_struct *wq = dev_to_wq(dev);
5213 const char *delim = "";
5214 int node, written = 0;
5216 rcu_read_lock_sched();
5217 for_each_node(node) {
5218 written += scnprintf(buf + written, PAGE_SIZE - written,
5219 "%s%d:%d", delim, node,
5220 unbound_pwq_by_node(wq, node)->pool->id);
5223 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5224 rcu_read_unlock_sched();
5229 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5232 struct workqueue_struct *wq = dev_to_wq(dev);
5235 mutex_lock(&wq->mutex);
5236 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5237 mutex_unlock(&wq->mutex);
5242 /* prepare workqueue_attrs for sysfs store operations */
5243 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5245 struct workqueue_attrs *attrs;
5247 lockdep_assert_held(&wq_pool_mutex);
5249 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5253 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5257 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5258 const char *buf, size_t count)
5260 struct workqueue_struct *wq = dev_to_wq(dev);
5261 struct workqueue_attrs *attrs;
5264 apply_wqattrs_lock();
5266 attrs = wq_sysfs_prep_attrs(wq);
5270 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5271 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5272 ret = apply_workqueue_attrs_locked(wq, attrs);
5277 apply_wqattrs_unlock();
5278 free_workqueue_attrs(attrs);
5279 return ret ?: count;
5282 static ssize_t wq_cpumask_show(struct device *dev,
5283 struct device_attribute *attr, char *buf)
5285 struct workqueue_struct *wq = dev_to_wq(dev);
5288 mutex_lock(&wq->mutex);
5289 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5290 cpumask_pr_args(wq->unbound_attrs->cpumask));
5291 mutex_unlock(&wq->mutex);
5295 static ssize_t wq_cpumask_store(struct device *dev,
5296 struct device_attribute *attr,
5297 const char *buf, size_t count)
5299 struct workqueue_struct *wq = dev_to_wq(dev);
5300 struct workqueue_attrs *attrs;
5303 apply_wqattrs_lock();
5305 attrs = wq_sysfs_prep_attrs(wq);
5309 ret = cpumask_parse(buf, attrs->cpumask);
5311 ret = apply_workqueue_attrs_locked(wq, attrs);
5314 apply_wqattrs_unlock();
5315 free_workqueue_attrs(attrs);
5316 return ret ?: count;
5319 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5322 struct workqueue_struct *wq = dev_to_wq(dev);
5325 mutex_lock(&wq->mutex);
5326 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5327 !wq->unbound_attrs->no_numa);
5328 mutex_unlock(&wq->mutex);
5333 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5334 const char *buf, size_t count)
5336 struct workqueue_struct *wq = dev_to_wq(dev);
5337 struct workqueue_attrs *attrs;
5338 int v, ret = -ENOMEM;
5340 apply_wqattrs_lock();
5342 attrs = wq_sysfs_prep_attrs(wq);
5347 if (sscanf(buf, "%d", &v) == 1) {
5348 attrs->no_numa = !v;
5349 ret = apply_workqueue_attrs_locked(wq, attrs);
5353 apply_wqattrs_unlock();
5354 free_workqueue_attrs(attrs);
5355 return ret ?: count;
5358 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5359 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5360 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5361 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5362 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5366 static struct bus_type wq_subsys = {
5367 .name = "workqueue",
5368 .dev_groups = wq_sysfs_groups,
5371 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5372 struct device_attribute *attr, char *buf)
5376 mutex_lock(&wq_pool_mutex);
5377 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5378 cpumask_pr_args(wq_unbound_cpumask));
5379 mutex_unlock(&wq_pool_mutex);
5384 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5385 struct device_attribute *attr, const char *buf, size_t count)
5387 cpumask_var_t cpumask;
5390 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5393 ret = cpumask_parse(buf, cpumask);
5395 ret = workqueue_set_unbound_cpumask(cpumask);
5397 free_cpumask_var(cpumask);
5398 return ret ? ret : count;
5401 static struct device_attribute wq_sysfs_cpumask_attr =
5402 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5403 wq_unbound_cpumask_store);
5405 static int __init wq_sysfs_init(void)
5409 err = subsys_virtual_register(&wq_subsys, NULL);
5413 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5415 core_initcall(wq_sysfs_init);
5417 static void wq_device_release(struct device *dev)
5419 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5425 * workqueue_sysfs_register - make a workqueue visible in sysfs
5426 * @wq: the workqueue to register
5428 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5429 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5430 * which is the preferred method.
5432 * Workqueue user should use this function directly iff it wants to apply
5433 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5434 * apply_workqueue_attrs() may race against userland updating the
5437 * Return: 0 on success, -errno on failure.
5439 int workqueue_sysfs_register(struct workqueue_struct *wq)
5441 struct wq_device *wq_dev;
5445 * Adjusting max_active or creating new pwqs by applying
5446 * attributes breaks ordering guarantee. Disallow exposing ordered
5449 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5452 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5457 wq_dev->dev.bus = &wq_subsys;
5458 wq_dev->dev.release = wq_device_release;
5459 dev_set_name(&wq_dev->dev, "%s", wq->name);
5462 * unbound_attrs are created separately. Suppress uevent until
5463 * everything is ready.
5465 dev_set_uevent_suppress(&wq_dev->dev, true);
5467 ret = device_register(&wq_dev->dev);
5469 put_device(&wq_dev->dev);
5474 if (wq->flags & WQ_UNBOUND) {
5475 struct device_attribute *attr;
5477 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5478 ret = device_create_file(&wq_dev->dev, attr);
5480 device_unregister(&wq_dev->dev);
5487 dev_set_uevent_suppress(&wq_dev->dev, false);
5488 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5493 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5494 * @wq: the workqueue to unregister
5496 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5498 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5500 struct wq_device *wq_dev = wq->wq_dev;
5506 device_unregister(&wq_dev->dev);
5508 #else /* CONFIG_SYSFS */
5509 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5510 #endif /* CONFIG_SYSFS */
5513 * Workqueue watchdog.
5515 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5516 * flush dependency, a concurrency managed work item which stays RUNNING
5517 * indefinitely. Workqueue stalls can be very difficult to debug as the
5518 * usual warning mechanisms don't trigger and internal workqueue state is
5521 * Workqueue watchdog monitors all worker pools periodically and dumps
5522 * state if some pools failed to make forward progress for a while where
5523 * forward progress is defined as the first item on ->worklist changing.
5525 * This mechanism is controlled through the kernel parameter
5526 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5527 * corresponding sysfs parameter file.
5529 #ifdef CONFIG_WQ_WATCHDOG
5531 static unsigned long wq_watchdog_thresh = 30;
5532 static struct timer_list wq_watchdog_timer;
5534 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5535 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5537 static void wq_watchdog_reset_touched(void)
5541 wq_watchdog_touched = jiffies;
5542 for_each_possible_cpu(cpu)
5543 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5546 static void wq_watchdog_timer_fn(struct timer_list *unused)
5548 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5549 bool lockup_detected = false;
5550 struct worker_pool *pool;
5558 for_each_pool(pool, pi) {
5559 unsigned long pool_ts, touched, ts;
5561 if (list_empty(&pool->worklist))
5564 /* get the latest of pool and touched timestamps */
5565 pool_ts = READ_ONCE(pool->watchdog_ts);
5566 touched = READ_ONCE(wq_watchdog_touched);
5568 if (time_after(pool_ts, touched))
5573 if (pool->cpu >= 0) {
5574 unsigned long cpu_touched =
5575 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5577 if (time_after(cpu_touched, ts))
5582 if (time_after(jiffies, ts + thresh)) {
5583 lockup_detected = true;
5584 pr_emerg("BUG: workqueue lockup - pool");
5585 pr_cont_pool_info(pool);
5586 pr_cont(" stuck for %us!\n",
5587 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5593 if (lockup_detected)
5594 show_workqueue_state();
5596 wq_watchdog_reset_touched();
5597 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5600 notrace void wq_watchdog_touch(int cpu)
5603 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5605 wq_watchdog_touched = jiffies;
5608 static void wq_watchdog_set_thresh(unsigned long thresh)
5610 wq_watchdog_thresh = 0;
5611 del_timer_sync(&wq_watchdog_timer);
5614 wq_watchdog_thresh = thresh;
5615 wq_watchdog_reset_touched();
5616 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5620 static int wq_watchdog_param_set_thresh(const char *val,
5621 const struct kernel_param *kp)
5623 unsigned long thresh;
5626 ret = kstrtoul(val, 0, &thresh);
5631 wq_watchdog_set_thresh(thresh);
5633 wq_watchdog_thresh = thresh;
5638 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5639 .set = wq_watchdog_param_set_thresh,
5640 .get = param_get_ulong,
5643 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5646 static void wq_watchdog_init(void)
5648 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5649 wq_watchdog_set_thresh(wq_watchdog_thresh);
5652 #else /* CONFIG_WQ_WATCHDOG */
5654 static inline void wq_watchdog_init(void) { }
5656 #endif /* CONFIG_WQ_WATCHDOG */
5658 static void __init wq_numa_init(void)
5663 if (num_possible_nodes() <= 1)
5666 if (wq_disable_numa) {
5667 pr_info("workqueue: NUMA affinity support disabled\n");
5671 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5672 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5675 * We want masks of possible CPUs of each node which isn't readily
5676 * available. Build one from cpu_to_node() which should have been
5677 * fully initialized by now.
5679 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5683 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5684 node_online(node) ? node : NUMA_NO_NODE));
5686 for_each_possible_cpu(cpu) {
5687 node = cpu_to_node(cpu);
5688 if (WARN_ON(node == NUMA_NO_NODE)) {
5689 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5690 /* happens iff arch is bonkers, let's just proceed */
5693 cpumask_set_cpu(cpu, tbl[node]);
5696 wq_numa_possible_cpumask = tbl;
5697 wq_numa_enabled = true;
5701 * workqueue_init_early - early init for workqueue subsystem
5703 * This is the first half of two-staged workqueue subsystem initialization
5704 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5705 * idr are up. It sets up all the data structures and system workqueues
5706 * and allows early boot code to create workqueues and queue/cancel work
5707 * items. Actual work item execution starts only after kthreads can be
5708 * created and scheduled right before early initcalls.
5710 int __init workqueue_init_early(void)
5712 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5713 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5716 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5718 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5719 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5721 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5723 /* initialize CPU pools */
5724 for_each_possible_cpu(cpu) {
5725 struct worker_pool *pool;
5728 for_each_cpu_worker_pool(pool, cpu) {
5729 BUG_ON(init_worker_pool(pool));
5731 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5732 pool->attrs->nice = std_nice[i++];
5733 pool->node = cpu_to_node(cpu);
5736 mutex_lock(&wq_pool_mutex);
5737 BUG_ON(worker_pool_assign_id(pool));
5738 mutex_unlock(&wq_pool_mutex);
5742 /* create default unbound and ordered wq attrs */
5743 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5744 struct workqueue_attrs *attrs;
5746 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5747 attrs->nice = std_nice[i];
5748 unbound_std_wq_attrs[i] = attrs;
5751 * An ordered wq should have only one pwq as ordering is
5752 * guaranteed by max_active which is enforced by pwqs.
5753 * Turn off NUMA so that dfl_pwq is used for all nodes.
5755 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5756 attrs->nice = std_nice[i];
5757 attrs->no_numa = true;
5758 ordered_wq_attrs[i] = attrs;
5761 system_wq = alloc_workqueue("events", 0, 0);
5762 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5763 system_long_wq = alloc_workqueue("events_long", 0, 0);
5764 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5765 WQ_UNBOUND_MAX_ACTIVE);
5766 system_freezable_wq = alloc_workqueue("events_freezable",
5768 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5769 WQ_POWER_EFFICIENT, 0);
5770 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5771 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5773 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5774 !system_unbound_wq || !system_freezable_wq ||
5775 !system_power_efficient_wq ||
5776 !system_freezable_power_efficient_wq);
5782 * workqueue_init - bring workqueue subsystem fully online
5784 * This is the latter half of two-staged workqueue subsystem initialization
5785 * and invoked as soon as kthreads can be created and scheduled.
5786 * Workqueues have been created and work items queued on them, but there
5787 * are no kworkers executing the work items yet. Populate the worker pools
5788 * with the initial workers and enable future kworker creations.
5790 int __init workqueue_init(void)
5792 struct workqueue_struct *wq;
5793 struct worker_pool *pool;
5797 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5798 * CPU to node mapping may not be available that early on some
5799 * archs such as power and arm64. As per-cpu pools created
5800 * previously could be missing node hint and unbound pools NUMA
5801 * affinity, fix them up.
5803 * Also, while iterating workqueues, create rescuers if requested.
5807 mutex_lock(&wq_pool_mutex);
5809 for_each_possible_cpu(cpu) {
5810 for_each_cpu_worker_pool(pool, cpu) {
5811 pool->node = cpu_to_node(cpu);
5815 list_for_each_entry(wq, &workqueues, list) {
5816 wq_update_unbound_numa(wq, smp_processor_id(), true);
5817 WARN(init_rescuer(wq),
5818 "workqueue: failed to create early rescuer for %s",
5822 mutex_unlock(&wq_pool_mutex);
5824 /* create the initial workers */
5825 for_each_online_cpu(cpu) {
5826 for_each_cpu_worker_pool(pool, cpu) {
5827 pool->flags &= ~POOL_DISASSOCIATED;
5828 BUG_ON(!create_worker(pool));
5832 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5833 BUG_ON(!create_worker(pool));