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
263 struct lock_class_key key;
264 struct lockdep_map lockdep_map;
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache *pwq_cache;
283 static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa;
287 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293 static bool wq_online; /* can kworkers be created yet? */
295 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
301 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
302 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
303 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
305 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
306 static bool workqueue_freezing; /* PL: have wqs started freezing? */
308 /* PL: allowable cpus for unbound wqs and work items */
309 static cpumask_var_t wq_unbound_cpumask;
311 /* CPU where unbound work was last round robin scheduled from this CPU */
312 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
315 * Local execution of unbound work items is no longer guaranteed. The
316 * following always forces round-robin CPU selection on unbound work items
317 * to uncover usages which depend on it.
319 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
320 static bool wq_debug_force_rr_cpu = true;
322 static bool wq_debug_force_rr_cpu = false;
324 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
326 /* the per-cpu worker pools */
327 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
329 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
331 /* PL: hash of all unbound pools keyed by pool->attrs */
332 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
334 /* I: attributes used when instantiating standard unbound pools on demand */
335 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
337 /* I: attributes used when instantiating ordered pools on demand */
338 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
340 struct workqueue_struct *system_wq __read_mostly;
341 EXPORT_SYMBOL(system_wq);
342 struct workqueue_struct *system_highpri_wq __read_mostly;
343 EXPORT_SYMBOL_GPL(system_highpri_wq);
344 struct workqueue_struct *system_long_wq __read_mostly;
345 EXPORT_SYMBOL_GPL(system_long_wq);
346 struct workqueue_struct *system_unbound_wq __read_mostly;
347 EXPORT_SYMBOL_GPL(system_unbound_wq);
348 struct workqueue_struct *system_freezable_wq __read_mostly;
349 EXPORT_SYMBOL_GPL(system_freezable_wq);
350 struct workqueue_struct *system_power_efficient_wq __read_mostly;
351 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
352 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
353 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
355 static int worker_thread(void *__worker);
356 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
358 #define CREATE_TRACE_POINTS
359 #include <trace/events/workqueue.h>
361 #define assert_rcu_or_pool_mutex() \
362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
363 !lockdep_is_held(&wq_pool_mutex), \
364 "sched RCU or wq_pool_mutex should be held")
366 #define assert_rcu_or_wq_mutex(wq) \
367 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
368 !lockdep_is_held(&wq->mutex), \
369 "sched RCU or wq->mutex should be held")
371 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
373 !lockdep_is_held(&wq->mutex) && \
374 !lockdep_is_held(&wq_pool_mutex), \
375 "sched RCU, wq->mutex or wq_pool_mutex should be held")
377 #define for_each_cpu_worker_pool(pool, cpu) \
378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
383 * for_each_pool - iterate through all worker_pools in the system
384 * @pool: iteration cursor
385 * @pi: integer used for iteration
387 * This must be called either with wq_pool_mutex held or sched RCU read
388 * locked. If the pool needs to be used beyond the locking in effect, the
389 * caller is responsible for guaranteeing that the pool stays online.
391 * The if/else clause exists only for the lockdep assertion and can be
394 #define for_each_pool(pool, pi) \
395 idr_for_each_entry(&worker_pool_idr, pool, pi) \
396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
400 * for_each_pool_worker - iterate through all workers of a worker_pool
401 * @worker: iteration cursor
402 * @pool: worker_pool to iterate workers of
404 * This must be called with wq_pool_attach_mutex.
406 * The if/else clause exists only for the lockdep assertion and can be
409 #define for_each_pool_worker(worker, pool) \
410 list_for_each_entry((worker), &(pool)->workers, node) \
411 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
416 * @pwq: iteration cursor
417 * @wq: the target workqueue
419 * This must be called either with wq->mutex held or sched RCU read locked.
420 * If the pwq needs to be used beyond the locking in effect, the caller is
421 * responsible for guaranteeing that the pwq stays online.
423 * The if/else clause exists only for the lockdep assertion and can be
426 #define for_each_pwq(pwq, wq) \
427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
431 #ifdef CONFIG_DEBUG_OBJECTS_WORK
433 static struct debug_obj_descr work_debug_descr;
435 static void *work_debug_hint(void *addr)
437 return ((struct work_struct *) addr)->func;
440 static bool work_is_static_object(void *addr)
442 struct work_struct *work = addr;
444 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
448 * fixup_init is called when:
449 * - an active object is initialized
451 static bool work_fixup_init(void *addr, enum debug_obj_state state)
453 struct work_struct *work = addr;
456 case ODEBUG_STATE_ACTIVE:
457 cancel_work_sync(work);
458 debug_object_init(work, &work_debug_descr);
466 * fixup_free is called when:
467 * - an active object is freed
469 static bool work_fixup_free(void *addr, enum debug_obj_state state)
471 struct work_struct *work = addr;
474 case ODEBUG_STATE_ACTIVE:
475 cancel_work_sync(work);
476 debug_object_free(work, &work_debug_descr);
483 static struct debug_obj_descr work_debug_descr = {
484 .name = "work_struct",
485 .debug_hint = work_debug_hint,
486 .is_static_object = work_is_static_object,
487 .fixup_init = work_fixup_init,
488 .fixup_free = work_fixup_free,
491 static inline void debug_work_activate(struct work_struct *work)
493 debug_object_activate(work, &work_debug_descr);
496 static inline void debug_work_deactivate(struct work_struct *work)
498 debug_object_deactivate(work, &work_debug_descr);
501 void __init_work(struct work_struct *work, int onstack)
504 debug_object_init_on_stack(work, &work_debug_descr);
506 debug_object_init(work, &work_debug_descr);
508 EXPORT_SYMBOL_GPL(__init_work);
510 void destroy_work_on_stack(struct work_struct *work)
512 debug_object_free(work, &work_debug_descr);
514 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
516 void destroy_delayed_work_on_stack(struct delayed_work *work)
518 destroy_timer_on_stack(&work->timer);
519 debug_object_free(&work->work, &work_debug_descr);
521 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
524 static inline void debug_work_activate(struct work_struct *work) { }
525 static inline void debug_work_deactivate(struct work_struct *work) { }
529 * worker_pool_assign_id - allocate ID and assing it to @pool
530 * @pool: the pool pointer of interest
532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
533 * successfully, -errno on failure.
535 static int worker_pool_assign_id(struct worker_pool *pool)
539 lockdep_assert_held(&wq_pool_mutex);
541 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
552 * @wq: the target workqueue
555 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
557 * If the pwq needs to be used beyond the locking in effect, the caller is
558 * responsible for guaranteeing that the pwq stays online.
560 * Return: The unbound pool_workqueue for @node.
562 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
565 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
569 * delayed item is pending. The plan is to keep CPU -> NODE
570 * mapping valid and stable across CPU on/offlines. Once that
571 * happens, this workaround can be removed.
573 if (unlikely(node == NUMA_NO_NODE))
576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
579 static unsigned int work_color_to_flags(int color)
581 return color << WORK_STRUCT_COLOR_SHIFT;
584 static int get_work_color(struct work_struct *work)
586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
587 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
590 static int work_next_color(int color)
592 return (color + 1) % WORK_NR_COLORS;
596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597 * contain the pointer to the queued pwq. Once execution starts, the flag
598 * is cleared and the high bits contain OFFQ flags and pool ID.
600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601 * and clear_work_data() can be used to set the pwq, pool or clear
602 * work->data. These functions should only be called while the work is
603 * owned - ie. while the PENDING bit is set.
605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
606 * corresponding to a work. Pool is available once the work has been
607 * queued anywhere after initialization until it is sync canceled. pwq is
608 * available only while the work item is queued.
610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
611 * canceled. While being canceled, a work item may have its PENDING set
612 * but stay off timer and worklist for arbitrarily long and nobody should
613 * try to steal the PENDING bit.
615 static inline void set_work_data(struct work_struct *work, unsigned long data,
618 WARN_ON_ONCE(!work_pending(work));
619 atomic_long_set(&work->data, data | flags | work_static(work));
622 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
623 unsigned long extra_flags)
625 set_work_data(work, (unsigned long)pwq,
626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
629 static void set_work_pool_and_keep_pending(struct work_struct *work,
632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
633 WORK_STRUCT_PENDING);
636 static void set_work_pool_and_clear_pending(struct work_struct *work,
640 * The following wmb is paired with the implied mb in
641 * test_and_set_bit(PENDING) and ensures all updates to @work made
642 * here are visible to and precede any updates by the next PENDING
646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
648 * The following mb guarantees that previous clear of a PENDING bit
649 * will not be reordered with any speculative LOADS or STORES from
650 * work->current_func, which is executed afterwards. This possible
651 * reordering can lead to a missed execution on attempt to qeueue
652 * the same @work. E.g. consider this case:
655 * ---------------------------- --------------------------------
657 * 1 STORE event_indicated
658 * 2 queue_work_on() {
659 * 3 test_and_set_bit(PENDING)
660 * 4 } set_..._and_clear_pending() {
661 * 5 set_work_data() # clear bit
663 * 7 work->current_func() {
664 * 8 LOAD event_indicated
667 * Without an explicit full barrier speculative LOAD on line 8 can
668 * be executed before CPU#0 does STORE on line 1. If that happens,
669 * CPU#0 observes the PENDING bit is still set and new execution of
670 * a @work is not queued in a hope, that CPU#1 will eventually
671 * finish the queued @work. Meanwhile CPU#1 does not see
672 * event_indicated is set, because speculative LOAD was executed
673 * before actual STORE.
678 static void clear_work_data(struct work_struct *work)
680 smp_wmb(); /* see set_work_pool_and_clear_pending() */
681 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
684 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
686 unsigned long data = atomic_long_read(&work->data);
688 if (data & WORK_STRUCT_PWQ)
689 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
695 * get_work_pool - return the worker_pool a given work was associated with
696 * @work: the work item of interest
698 * Pools are created and destroyed under wq_pool_mutex, and allows read
699 * access under sched-RCU read lock. As such, this function should be
700 * called under wq_pool_mutex or with preemption disabled.
702 * All fields of the returned pool are accessible as long as the above
703 * mentioned locking is in effect. If the returned pool needs to be used
704 * beyond the critical section, the caller is responsible for ensuring the
705 * returned pool is and stays online.
707 * Return: The worker_pool @work was last associated with. %NULL if none.
709 static struct worker_pool *get_work_pool(struct work_struct *work)
711 unsigned long data = atomic_long_read(&work->data);
714 assert_rcu_or_pool_mutex();
716 if (data & WORK_STRUCT_PWQ)
717 return ((struct pool_workqueue *)
718 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
720 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
721 if (pool_id == WORK_OFFQ_POOL_NONE)
724 return idr_find(&worker_pool_idr, pool_id);
728 * get_work_pool_id - return the worker pool ID a given work is associated with
729 * @work: the work item of interest
731 * Return: The worker_pool ID @work was last associated with.
732 * %WORK_OFFQ_POOL_NONE if none.
734 static int get_work_pool_id(struct work_struct *work)
736 unsigned long data = atomic_long_read(&work->data);
738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
742 return data >> WORK_OFFQ_POOL_SHIFT;
745 static void mark_work_canceling(struct work_struct *work)
747 unsigned long pool_id = get_work_pool_id(work);
749 pool_id <<= WORK_OFFQ_POOL_SHIFT;
750 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
753 static bool work_is_canceling(struct work_struct *work)
755 unsigned long data = atomic_long_read(&work->data);
757 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
761 * Policy functions. These define the policies on how the global worker
762 * pools are managed. Unless noted otherwise, these functions assume that
763 * they're being called with pool->lock held.
766 static bool __need_more_worker(struct worker_pool *pool)
768 return !atomic_read(&pool->nr_running);
772 * Need to wake up a worker? Called from anything but currently
775 * Note that, because unbound workers never contribute to nr_running, this
776 * function will always return %true for unbound pools as long as the
777 * worklist isn't empty.
779 static bool need_more_worker(struct worker_pool *pool)
781 return !list_empty(&pool->worklist) && __need_more_worker(pool);
784 /* Can I start working? Called from busy but !running workers. */
785 static bool may_start_working(struct worker_pool *pool)
787 return pool->nr_idle;
790 /* Do I need to keep working? Called from currently running workers. */
791 static bool keep_working(struct worker_pool *pool)
793 return !list_empty(&pool->worklist) &&
794 atomic_read(&pool->nr_running) <= 1;
797 /* Do we need a new worker? Called from manager. */
798 static bool need_to_create_worker(struct worker_pool *pool)
800 return need_more_worker(pool) && !may_start_working(pool);
803 /* Do we have too many workers and should some go away? */
804 static bool too_many_workers(struct worker_pool *pool)
806 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
807 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
808 int nr_busy = pool->nr_workers - nr_idle;
810 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
817 /* Return the first idle worker. Safe with preemption disabled */
818 static struct worker *first_idle_worker(struct worker_pool *pool)
820 if (unlikely(list_empty(&pool->idle_list)))
823 return list_first_entry(&pool->idle_list, struct worker, entry);
827 * wake_up_worker - wake up an idle worker
828 * @pool: worker pool to wake worker from
830 * Wake up the first idle worker of @pool.
833 * spin_lock_irq(pool->lock).
835 static void wake_up_worker(struct worker_pool *pool)
837 struct worker *worker = first_idle_worker(pool);
840 wake_up_process(worker->task);
844 * wq_worker_waking_up - a worker is waking up
845 * @task: task waking up
846 * @cpu: CPU @task is waking up to
848 * This function is called during try_to_wake_up() when a worker is
852 * spin_lock_irq(rq->lock)
854 void wq_worker_waking_up(struct task_struct *task, int cpu)
856 struct worker *worker = kthread_data(task);
858 if (!(worker->flags & WORKER_NOT_RUNNING)) {
859 WARN_ON_ONCE(worker->pool->cpu != cpu);
860 atomic_inc(&worker->pool->nr_running);
865 * wq_worker_sleeping - a worker is going to sleep
866 * @task: task going to sleep
868 * This function is called during schedule() when a busy worker is
869 * going to sleep. Worker on the same cpu can be woken up by
870 * returning pointer to its task.
873 * spin_lock_irq(rq->lock)
876 * Worker task on @cpu to wake up, %NULL if none.
878 struct task_struct *wq_worker_sleeping(struct task_struct *task)
880 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
881 struct worker_pool *pool;
884 * Rescuers, which may not have all the fields set up like normal
885 * workers, also reach here, let's not access anything before
886 * checking NOT_RUNNING.
888 if (worker->flags & WORKER_NOT_RUNNING)
893 /* this can only happen on the local cpu */
894 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
898 * The counterpart of the following dec_and_test, implied mb,
899 * worklist not empty test sequence is in insert_work().
900 * Please read comment there.
902 * NOT_RUNNING is clear. This means that we're bound to and
903 * running on the local cpu w/ rq lock held and preemption
904 * disabled, which in turn means that none else could be
905 * manipulating idle_list, so dereferencing idle_list without pool
908 if (atomic_dec_and_test(&pool->nr_running) &&
909 !list_empty(&pool->worklist))
910 to_wakeup = first_idle_worker(pool);
911 return to_wakeup ? to_wakeup->task : NULL;
915 * wq_worker_last_func - retrieve worker's last work function
917 * Determine the last function a worker executed. This is called from
918 * the scheduler to get a worker's last known identity.
921 * spin_lock_irq(rq->lock)
924 * The last work function %current executed as a worker, NULL if it
925 * hasn't executed any work yet.
927 work_func_t wq_worker_last_func(struct task_struct *task)
929 struct worker *worker = kthread_data(task);
931 return worker->last_func;
935 * worker_set_flags - set worker flags and adjust nr_running accordingly
937 * @flags: flags to set
939 * Set @flags in @worker->flags and adjust nr_running accordingly.
942 * spin_lock_irq(pool->lock)
944 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
946 struct worker_pool *pool = worker->pool;
948 WARN_ON_ONCE(worker->task != current);
950 /* If transitioning into NOT_RUNNING, adjust nr_running. */
951 if ((flags & WORKER_NOT_RUNNING) &&
952 !(worker->flags & WORKER_NOT_RUNNING)) {
953 atomic_dec(&pool->nr_running);
956 worker->flags |= flags;
960 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
962 * @flags: flags to clear
964 * Clear @flags in @worker->flags and adjust nr_running accordingly.
967 * spin_lock_irq(pool->lock)
969 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
971 struct worker_pool *pool = worker->pool;
972 unsigned int oflags = worker->flags;
974 WARN_ON_ONCE(worker->task != current);
976 worker->flags &= ~flags;
979 * If transitioning out of NOT_RUNNING, increment nr_running. Note
980 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
981 * of multiple flags, not a single flag.
983 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
984 if (!(worker->flags & WORKER_NOT_RUNNING))
985 atomic_inc(&pool->nr_running);
989 * find_worker_executing_work - find worker which is executing a work
990 * @pool: pool of interest
991 * @work: work to find worker for
993 * Find a worker which is executing @work on @pool by searching
994 * @pool->busy_hash which is keyed by the address of @work. For a worker
995 * to match, its current execution should match the address of @work and
996 * its work function. This is to avoid unwanted dependency between
997 * unrelated work executions through a work item being recycled while still
1000 * This is a bit tricky. A work item may be freed once its execution
1001 * starts and nothing prevents the freed area from being recycled for
1002 * another work item. If the same work item address ends up being reused
1003 * before the original execution finishes, workqueue will identify the
1004 * recycled work item as currently executing and make it wait until the
1005 * current execution finishes, introducing an unwanted dependency.
1007 * This function checks the work item address and work function to avoid
1008 * false positives. Note that this isn't complete as one may construct a
1009 * work function which can introduce dependency onto itself through a
1010 * recycled work item. Well, if somebody wants to shoot oneself in the
1011 * foot that badly, there's only so much we can do, and if such deadlock
1012 * actually occurs, it should be easy to locate the culprit work function.
1015 * spin_lock_irq(pool->lock).
1018 * Pointer to worker which is executing @work if found, %NULL
1021 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1022 struct work_struct *work)
1024 struct worker *worker;
1026 hash_for_each_possible(pool->busy_hash, worker, hentry,
1027 (unsigned long)work)
1028 if (worker->current_work == work &&
1029 worker->current_func == work->func)
1036 * move_linked_works - move linked works to a list
1037 * @work: start of series of works to be scheduled
1038 * @head: target list to append @work to
1039 * @nextp: out parameter for nested worklist walking
1041 * Schedule linked works starting from @work to @head. Work series to
1042 * be scheduled starts at @work and includes any consecutive work with
1043 * WORK_STRUCT_LINKED set in its predecessor.
1045 * If @nextp is not NULL, it's updated to point to the next work of
1046 * the last scheduled work. This allows move_linked_works() to be
1047 * nested inside outer list_for_each_entry_safe().
1050 * spin_lock_irq(pool->lock).
1052 static void move_linked_works(struct work_struct *work, struct list_head *head,
1053 struct work_struct **nextp)
1055 struct work_struct *n;
1058 * Linked worklist will always end before the end of the list,
1059 * use NULL for list head.
1061 list_for_each_entry_safe_from(work, n, NULL, entry) {
1062 list_move_tail(&work->entry, head);
1063 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1068 * If we're already inside safe list traversal and have moved
1069 * multiple works to the scheduled queue, the next position
1070 * needs to be updated.
1077 * get_pwq - get an extra reference on the specified pool_workqueue
1078 * @pwq: pool_workqueue to get
1080 * Obtain an extra reference on @pwq. The caller should guarantee that
1081 * @pwq has positive refcnt and be holding the matching pool->lock.
1083 static void get_pwq(struct pool_workqueue *pwq)
1085 lockdep_assert_held(&pwq->pool->lock);
1086 WARN_ON_ONCE(pwq->refcnt <= 0);
1091 * put_pwq - put a pool_workqueue reference
1092 * @pwq: pool_workqueue to put
1094 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1095 * destruction. The caller should be holding the matching pool->lock.
1097 static void put_pwq(struct pool_workqueue *pwq)
1099 lockdep_assert_held(&pwq->pool->lock);
1100 if (likely(--pwq->refcnt))
1102 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1105 * @pwq can't be released under pool->lock, bounce to
1106 * pwq_unbound_release_workfn(). This never recurses on the same
1107 * pool->lock as this path is taken only for unbound workqueues and
1108 * the release work item is scheduled on a per-cpu workqueue. To
1109 * avoid lockdep warning, unbound pool->locks are given lockdep
1110 * subclass of 1 in get_unbound_pool().
1112 schedule_work(&pwq->unbound_release_work);
1116 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1117 * @pwq: pool_workqueue to put (can be %NULL)
1119 * put_pwq() with locking. This function also allows %NULL @pwq.
1121 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1125 * As both pwqs and pools are sched-RCU protected, the
1126 * following lock operations are safe.
1128 spin_lock_irq(&pwq->pool->lock);
1130 spin_unlock_irq(&pwq->pool->lock);
1134 static void pwq_activate_delayed_work(struct work_struct *work)
1136 struct pool_workqueue *pwq = get_work_pwq(work);
1138 trace_workqueue_activate_work(work);
1139 if (list_empty(&pwq->pool->worklist))
1140 pwq->pool->watchdog_ts = jiffies;
1141 move_linked_works(work, &pwq->pool->worklist, NULL);
1142 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1146 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1148 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1149 struct work_struct, entry);
1151 pwq_activate_delayed_work(work);
1155 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1156 * @pwq: pwq of interest
1157 * @color: color of work which left the queue
1159 * A work either has completed or is removed from pending queue,
1160 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1163 * spin_lock_irq(pool->lock).
1165 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1167 /* uncolored work items don't participate in flushing or nr_active */
1168 if (color == WORK_NO_COLOR)
1171 pwq->nr_in_flight[color]--;
1174 if (!list_empty(&pwq->delayed_works)) {
1175 /* one down, submit a delayed one */
1176 if (pwq->nr_active < pwq->max_active)
1177 pwq_activate_first_delayed(pwq);
1180 /* is flush in progress and are we at the flushing tip? */
1181 if (likely(pwq->flush_color != color))
1184 /* are there still in-flight works? */
1185 if (pwq->nr_in_flight[color])
1188 /* this pwq is done, clear flush_color */
1189 pwq->flush_color = -1;
1192 * If this was the last pwq, wake up the first flusher. It
1193 * will handle the rest.
1195 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1196 complete(&pwq->wq->first_flusher->done);
1202 * try_to_grab_pending - steal work item from worklist and disable irq
1203 * @work: work item to steal
1204 * @is_dwork: @work is a delayed_work
1205 * @flags: place to store irq state
1207 * Try to grab PENDING bit of @work. This function can handle @work in any
1208 * stable state - idle, on timer or on worklist.
1211 * 1 if @work was pending and we successfully stole PENDING
1212 * 0 if @work was idle and we claimed PENDING
1213 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1214 * -ENOENT if someone else is canceling @work, this state may persist
1215 * for arbitrarily long
1218 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1219 * interrupted while holding PENDING and @work off queue, irq must be
1220 * disabled on entry. This, combined with delayed_work->timer being
1221 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1223 * On successful return, >= 0, irq is disabled and the caller is
1224 * responsible for releasing it using local_irq_restore(*@flags).
1226 * This function is safe to call from any context including IRQ handler.
1228 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1229 unsigned long *flags)
1231 struct worker_pool *pool;
1232 struct pool_workqueue *pwq;
1234 local_irq_save(*flags);
1236 /* try to steal the timer if it exists */
1238 struct delayed_work *dwork = to_delayed_work(work);
1241 * dwork->timer is irqsafe. If del_timer() fails, it's
1242 * guaranteed that the timer is not queued anywhere and not
1243 * running on the local CPU.
1245 if (likely(del_timer(&dwork->timer)))
1249 /* try to claim PENDING the normal way */
1250 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1254 * The queueing is in progress, or it is already queued. Try to
1255 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1257 pool = get_work_pool(work);
1261 spin_lock(&pool->lock);
1263 * work->data is guaranteed to point to pwq only while the work
1264 * item is queued on pwq->wq, and both updating work->data to point
1265 * to pwq on queueing and to pool on dequeueing are done under
1266 * pwq->pool->lock. This in turn guarantees that, if work->data
1267 * points to pwq which is associated with a locked pool, the work
1268 * item is currently queued on that pool.
1270 pwq = get_work_pwq(work);
1271 if (pwq && pwq->pool == pool) {
1272 debug_work_deactivate(work);
1275 * A delayed work item cannot be grabbed directly because
1276 * it might have linked NO_COLOR work items which, if left
1277 * on the delayed_list, will confuse pwq->nr_active
1278 * management later on and cause stall. Make sure the work
1279 * item is activated before grabbing.
1281 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1282 pwq_activate_delayed_work(work);
1284 list_del_init(&work->entry);
1285 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1287 /* work->data points to pwq iff queued, point to pool */
1288 set_work_pool_and_keep_pending(work, pool->id);
1290 spin_unlock(&pool->lock);
1293 spin_unlock(&pool->lock);
1295 local_irq_restore(*flags);
1296 if (work_is_canceling(work))
1303 * insert_work - insert a work into a pool
1304 * @pwq: pwq @work belongs to
1305 * @work: work to insert
1306 * @head: insertion point
1307 * @extra_flags: extra WORK_STRUCT_* flags to set
1309 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1310 * work_struct flags.
1313 * spin_lock_irq(pool->lock).
1315 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1316 struct list_head *head, unsigned int extra_flags)
1318 struct worker_pool *pool = pwq->pool;
1320 /* we own @work, set data and link */
1321 set_work_pwq(work, pwq, extra_flags);
1322 list_add_tail(&work->entry, head);
1326 * Ensure either wq_worker_sleeping() sees the above
1327 * list_add_tail() or we see zero nr_running to avoid workers lying
1328 * around lazily while there are works to be processed.
1332 if (__need_more_worker(pool))
1333 wake_up_worker(pool);
1337 * Test whether @work is being queued from another work executing on the
1340 static bool is_chained_work(struct workqueue_struct *wq)
1342 struct worker *worker;
1344 worker = current_wq_worker();
1346 * Return %true iff I'm a worker execuing a work item on @wq. If
1347 * I'm @worker, it's safe to dereference it without locking.
1349 return worker && worker->current_pwq->wq == wq;
1353 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1354 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1355 * avoid perturbing sensitive tasks.
1357 static int wq_select_unbound_cpu(int cpu)
1359 static bool printed_dbg_warning;
1362 if (likely(!wq_debug_force_rr_cpu)) {
1363 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1365 } else if (!printed_dbg_warning) {
1366 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1367 printed_dbg_warning = true;
1370 if (cpumask_empty(wq_unbound_cpumask))
1373 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1374 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1375 if (unlikely(new_cpu >= nr_cpu_ids)) {
1376 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1377 if (unlikely(new_cpu >= nr_cpu_ids))
1380 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1385 static void __queue_work(int cpu, struct workqueue_struct *wq,
1386 struct work_struct *work)
1388 struct pool_workqueue *pwq;
1389 struct worker_pool *last_pool;
1390 struct list_head *worklist;
1391 unsigned int work_flags;
1392 unsigned int req_cpu = cpu;
1395 * While a work item is PENDING && off queue, a task trying to
1396 * steal the PENDING will busy-loop waiting for it to either get
1397 * queued or lose PENDING. Grabbing PENDING and queueing should
1398 * happen with IRQ disabled.
1400 lockdep_assert_irqs_disabled();
1402 debug_work_activate(work);
1404 /* if draining, only works from the same workqueue are allowed */
1405 if (unlikely(wq->flags & __WQ_DRAINING) &&
1406 WARN_ON_ONCE(!is_chained_work(wq)))
1409 if (req_cpu == WORK_CPU_UNBOUND)
1410 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1412 /* pwq which will be used unless @work is executing elsewhere */
1413 if (!(wq->flags & WQ_UNBOUND))
1414 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1416 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1419 * If @work was previously on a different pool, it might still be
1420 * running there, in which case the work needs to be queued on that
1421 * pool to guarantee non-reentrancy.
1423 last_pool = get_work_pool(work);
1424 if (last_pool && last_pool != pwq->pool) {
1425 struct worker *worker;
1427 spin_lock(&last_pool->lock);
1429 worker = find_worker_executing_work(last_pool, work);
1431 if (worker && worker->current_pwq->wq == wq) {
1432 pwq = worker->current_pwq;
1434 /* meh... not running there, queue here */
1435 spin_unlock(&last_pool->lock);
1436 spin_lock(&pwq->pool->lock);
1439 spin_lock(&pwq->pool->lock);
1443 * pwq is determined and locked. For unbound pools, we could have
1444 * raced with pwq release and it could already be dead. If its
1445 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1446 * without another pwq replacing it in the numa_pwq_tbl or while
1447 * work items are executing on it, so the retrying is guaranteed to
1448 * make forward-progress.
1450 if (unlikely(!pwq->refcnt)) {
1451 if (wq->flags & WQ_UNBOUND) {
1452 spin_unlock(&pwq->pool->lock);
1457 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1461 /* pwq determined, queue */
1462 trace_workqueue_queue_work(req_cpu, pwq, work);
1464 if (WARN_ON(!list_empty(&work->entry))) {
1465 spin_unlock(&pwq->pool->lock);
1469 pwq->nr_in_flight[pwq->work_color]++;
1470 work_flags = work_color_to_flags(pwq->work_color);
1472 if (likely(pwq->nr_active < pwq->max_active)) {
1473 trace_workqueue_activate_work(work);
1475 worklist = &pwq->pool->worklist;
1476 if (list_empty(worklist))
1477 pwq->pool->watchdog_ts = jiffies;
1479 work_flags |= WORK_STRUCT_DELAYED;
1480 worklist = &pwq->delayed_works;
1483 insert_work(pwq, work, worklist, work_flags);
1485 spin_unlock(&pwq->pool->lock);
1489 * queue_work_on - queue work on specific cpu
1490 * @cpu: CPU number to execute work on
1491 * @wq: workqueue to use
1492 * @work: work to queue
1494 * We queue the work to a specific CPU, the caller must ensure it
1497 * Return: %false if @work was already on a queue, %true otherwise.
1499 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1500 struct work_struct *work)
1503 unsigned long flags;
1505 local_irq_save(flags);
1507 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1508 __queue_work(cpu, wq, work);
1512 local_irq_restore(flags);
1515 EXPORT_SYMBOL(queue_work_on);
1518 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1519 * @node: NUMA node ID that we want to select a CPU from
1521 * This function will attempt to find a "random" cpu available on a given
1522 * node. If there are no CPUs available on the given node it will return
1523 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1524 * available CPU if we need to schedule this work.
1526 static int workqueue_select_cpu_near(int node)
1530 /* No point in doing this if NUMA isn't enabled for workqueues */
1531 if (!wq_numa_enabled)
1532 return WORK_CPU_UNBOUND;
1534 /* Delay binding to CPU if node is not valid or online */
1535 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1536 return WORK_CPU_UNBOUND;
1538 /* Use local node/cpu if we are already there */
1539 cpu = raw_smp_processor_id();
1540 if (node == cpu_to_node(cpu))
1543 /* Use "random" otherwise know as "first" online CPU of node */
1544 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1546 /* If CPU is valid return that, otherwise just defer */
1547 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1551 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1552 * @node: NUMA node that we are targeting the work for
1553 * @wq: workqueue to use
1554 * @work: work to queue
1556 * We queue the work to a "random" CPU within a given NUMA node. The basic
1557 * idea here is to provide a way to somehow associate work with a given
1560 * This function will only make a best effort attempt at getting this onto
1561 * the right NUMA node. If no node is requested or the requested node is
1562 * offline then we just fall back to standard queue_work behavior.
1564 * Currently the "random" CPU ends up being the first available CPU in the
1565 * intersection of cpu_online_mask and the cpumask of the node, unless we
1566 * are running on the node. In that case we just use the current CPU.
1568 * Return: %false if @work was already on a queue, %true otherwise.
1570 bool queue_work_node(int node, struct workqueue_struct *wq,
1571 struct work_struct *work)
1573 unsigned long flags;
1577 * This current implementation is specific to unbound workqueues.
1578 * Specifically we only return the first available CPU for a given
1579 * node instead of cycling through individual CPUs within the node.
1581 * If this is used with a per-cpu workqueue then the logic in
1582 * workqueue_select_cpu_near would need to be updated to allow for
1583 * some round robin type logic.
1585 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1587 local_irq_save(flags);
1589 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1590 int cpu = workqueue_select_cpu_near(node);
1592 __queue_work(cpu, wq, work);
1596 local_irq_restore(flags);
1599 EXPORT_SYMBOL_GPL(queue_work_node);
1601 void delayed_work_timer_fn(struct timer_list *t)
1603 struct delayed_work *dwork = from_timer(dwork, t, timer);
1605 /* should have been called from irqsafe timer with irq already off */
1606 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1608 EXPORT_SYMBOL(delayed_work_timer_fn);
1610 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1611 struct delayed_work *dwork, unsigned long delay)
1613 struct timer_list *timer = &dwork->timer;
1614 struct work_struct *work = &dwork->work;
1617 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1618 WARN_ON_ONCE(timer_pending(timer));
1619 WARN_ON_ONCE(!list_empty(&work->entry));
1622 * If @delay is 0, queue @dwork->work immediately. This is for
1623 * both optimization and correctness. The earliest @timer can
1624 * expire is on the closest next tick and delayed_work users depend
1625 * on that there's no such delay when @delay is 0.
1628 __queue_work(cpu, wq, &dwork->work);
1634 timer->expires = jiffies + delay;
1636 if (unlikely(cpu != WORK_CPU_UNBOUND))
1637 add_timer_on(timer, cpu);
1643 * queue_delayed_work_on - queue work on specific CPU after delay
1644 * @cpu: CPU number to execute work on
1645 * @wq: workqueue to use
1646 * @dwork: work to queue
1647 * @delay: number of jiffies to wait before queueing
1649 * Return: %false if @work was already on a queue, %true otherwise. If
1650 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1653 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1654 struct delayed_work *dwork, unsigned long delay)
1656 struct work_struct *work = &dwork->work;
1658 unsigned long flags;
1660 /* read the comment in __queue_work() */
1661 local_irq_save(flags);
1663 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1664 __queue_delayed_work(cpu, wq, dwork, delay);
1668 local_irq_restore(flags);
1671 EXPORT_SYMBOL(queue_delayed_work_on);
1674 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1675 * @cpu: CPU number to execute work on
1676 * @wq: workqueue to use
1677 * @dwork: work to queue
1678 * @delay: number of jiffies to wait before queueing
1680 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1681 * modify @dwork's timer so that it expires after @delay. If @delay is
1682 * zero, @work is guaranteed to be scheduled immediately regardless of its
1685 * Return: %false if @dwork was idle and queued, %true if @dwork was
1686 * pending and its timer was modified.
1688 * This function is safe to call from any context including IRQ handler.
1689 * See try_to_grab_pending() for details.
1691 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1692 struct delayed_work *dwork, unsigned long delay)
1694 unsigned long flags;
1698 ret = try_to_grab_pending(&dwork->work, true, &flags);
1699 } while (unlikely(ret == -EAGAIN));
1701 if (likely(ret >= 0)) {
1702 __queue_delayed_work(cpu, wq, dwork, delay);
1703 local_irq_restore(flags);
1706 /* -ENOENT from try_to_grab_pending() becomes %true */
1709 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1711 static void rcu_work_rcufn(struct rcu_head *rcu)
1713 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1715 /* read the comment in __queue_work() */
1716 local_irq_disable();
1717 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1722 * queue_rcu_work - queue work after a RCU grace period
1723 * @wq: workqueue to use
1724 * @rwork: work to queue
1726 * Return: %false if @rwork was already pending, %true otherwise. Note
1727 * that a full RCU grace period is guaranteed only after a %true return.
1728 * While @rwork is guarnateed to be executed after a %false return, the
1729 * execution may happen before a full RCU grace period has passed.
1731 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1733 struct work_struct *work = &rwork->work;
1735 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1737 call_rcu(&rwork->rcu, rcu_work_rcufn);
1743 EXPORT_SYMBOL(queue_rcu_work);
1746 * worker_enter_idle - enter idle state
1747 * @worker: worker which is entering idle state
1749 * @worker is entering idle state. Update stats and idle timer if
1753 * spin_lock_irq(pool->lock).
1755 static void worker_enter_idle(struct worker *worker)
1757 struct worker_pool *pool = worker->pool;
1759 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1760 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1761 (worker->hentry.next || worker->hentry.pprev)))
1764 /* can't use worker_set_flags(), also called from create_worker() */
1765 worker->flags |= WORKER_IDLE;
1767 worker->last_active = jiffies;
1769 /* idle_list is LIFO */
1770 list_add(&worker->entry, &pool->idle_list);
1772 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1773 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1776 * Sanity check nr_running. Because unbind_workers() releases
1777 * pool->lock between setting %WORKER_UNBOUND and zapping
1778 * nr_running, the warning may trigger spuriously. Check iff
1779 * unbind is not in progress.
1781 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1782 pool->nr_workers == pool->nr_idle &&
1783 atomic_read(&pool->nr_running));
1787 * worker_leave_idle - leave idle state
1788 * @worker: worker which is leaving idle state
1790 * @worker is leaving idle state. Update stats.
1793 * spin_lock_irq(pool->lock).
1795 static void worker_leave_idle(struct worker *worker)
1797 struct worker_pool *pool = worker->pool;
1799 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1801 worker_clr_flags(worker, WORKER_IDLE);
1803 list_del_init(&worker->entry);
1806 static struct worker *alloc_worker(int node)
1808 struct worker *worker;
1810 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1812 INIT_LIST_HEAD(&worker->entry);
1813 INIT_LIST_HEAD(&worker->scheduled);
1814 INIT_LIST_HEAD(&worker->node);
1815 /* on creation a worker is in !idle && prep state */
1816 worker->flags = WORKER_PREP;
1822 * worker_attach_to_pool() - attach a worker to a pool
1823 * @worker: worker to be attached
1824 * @pool: the target pool
1826 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1827 * cpu-binding of @worker are kept coordinated with the pool across
1830 static void worker_attach_to_pool(struct worker *worker,
1831 struct worker_pool *pool)
1833 mutex_lock(&wq_pool_attach_mutex);
1836 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1837 * online CPUs. It'll be re-applied when any of the CPUs come up.
1839 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1842 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1843 * stable across this function. See the comments above the flag
1844 * definition for details.
1846 if (pool->flags & POOL_DISASSOCIATED)
1847 worker->flags |= WORKER_UNBOUND;
1849 list_add_tail(&worker->node, &pool->workers);
1850 worker->pool = pool;
1852 mutex_unlock(&wq_pool_attach_mutex);
1856 * worker_detach_from_pool() - detach a worker from its pool
1857 * @worker: worker which is attached to its pool
1859 * Undo the attaching which had been done in worker_attach_to_pool(). The
1860 * caller worker shouldn't access to the pool after detached except it has
1861 * other reference to the pool.
1863 static void worker_detach_from_pool(struct worker *worker)
1865 struct worker_pool *pool = worker->pool;
1866 struct completion *detach_completion = NULL;
1868 mutex_lock(&wq_pool_attach_mutex);
1870 list_del(&worker->node);
1871 worker->pool = NULL;
1873 if (list_empty(&pool->workers))
1874 detach_completion = pool->detach_completion;
1875 mutex_unlock(&wq_pool_attach_mutex);
1877 /* clear leftover flags without pool->lock after it is detached */
1878 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1880 if (detach_completion)
1881 complete(detach_completion);
1885 * create_worker - create a new workqueue worker
1886 * @pool: pool the new worker will belong to
1888 * Create and start a new worker which is attached to @pool.
1891 * Might sleep. Does GFP_KERNEL allocations.
1894 * Pointer to the newly created worker.
1896 static struct worker *create_worker(struct worker_pool *pool)
1898 struct worker *worker = NULL;
1902 /* ID is needed to determine kthread name */
1903 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1907 worker = alloc_worker(pool->node);
1914 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1915 pool->attrs->nice < 0 ? "H" : "");
1917 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1919 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1920 "kworker/%s", id_buf);
1921 if (IS_ERR(worker->task))
1924 set_user_nice(worker->task, pool->attrs->nice);
1925 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1927 /* successful, attach the worker to the pool */
1928 worker_attach_to_pool(worker, pool);
1930 /* start the newly created worker */
1931 spin_lock_irq(&pool->lock);
1932 worker->pool->nr_workers++;
1933 worker_enter_idle(worker);
1934 wake_up_process(worker->task);
1935 spin_unlock_irq(&pool->lock);
1941 ida_simple_remove(&pool->worker_ida, id);
1947 * destroy_worker - destroy a workqueue worker
1948 * @worker: worker to be destroyed
1950 * Destroy @worker and adjust @pool stats accordingly. The worker should
1954 * spin_lock_irq(pool->lock).
1956 static void destroy_worker(struct worker *worker)
1958 struct worker_pool *pool = worker->pool;
1960 lockdep_assert_held(&pool->lock);
1962 /* sanity check frenzy */
1963 if (WARN_ON(worker->current_work) ||
1964 WARN_ON(!list_empty(&worker->scheduled)) ||
1965 WARN_ON(!(worker->flags & WORKER_IDLE)))
1971 list_del_init(&worker->entry);
1972 worker->flags |= WORKER_DIE;
1973 wake_up_process(worker->task);
1976 static void idle_worker_timeout(struct timer_list *t)
1978 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1980 spin_lock_irq(&pool->lock);
1982 while (too_many_workers(pool)) {
1983 struct worker *worker;
1984 unsigned long expires;
1986 /* idle_list is kept in LIFO order, check the last one */
1987 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1988 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1990 if (time_before(jiffies, expires)) {
1991 mod_timer(&pool->idle_timer, expires);
1995 destroy_worker(worker);
1998 spin_unlock_irq(&pool->lock);
2001 static void send_mayday(struct work_struct *work)
2003 struct pool_workqueue *pwq = get_work_pwq(work);
2004 struct workqueue_struct *wq = pwq->wq;
2006 lockdep_assert_held(&wq_mayday_lock);
2011 /* mayday mayday mayday */
2012 if (list_empty(&pwq->mayday_node)) {
2014 * If @pwq is for an unbound wq, its base ref may be put at
2015 * any time due to an attribute change. Pin @pwq until the
2016 * rescuer is done with it.
2019 list_add_tail(&pwq->mayday_node, &wq->maydays);
2020 wake_up_process(wq->rescuer->task);
2024 static void pool_mayday_timeout(struct timer_list *t)
2026 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2027 struct work_struct *work;
2029 spin_lock_irq(&pool->lock);
2030 spin_lock(&wq_mayday_lock); /* for wq->maydays */
2032 if (need_to_create_worker(pool)) {
2034 * We've been trying to create a new worker but
2035 * haven't been successful. We might be hitting an
2036 * allocation deadlock. Send distress signals to
2039 list_for_each_entry(work, &pool->worklist, entry)
2043 spin_unlock(&wq_mayday_lock);
2044 spin_unlock_irq(&pool->lock);
2046 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2050 * maybe_create_worker - create a new worker if necessary
2051 * @pool: pool to create a new worker for
2053 * Create a new worker for @pool if necessary. @pool is guaranteed to
2054 * have at least one idle worker on return from this function. If
2055 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2056 * sent to all rescuers with works scheduled on @pool to resolve
2057 * possible allocation deadlock.
2059 * On return, need_to_create_worker() is guaranteed to be %false and
2060 * may_start_working() %true.
2063 * spin_lock_irq(pool->lock) which may be released and regrabbed
2064 * multiple times. Does GFP_KERNEL allocations. Called only from
2067 static void maybe_create_worker(struct worker_pool *pool)
2068 __releases(&pool->lock)
2069 __acquires(&pool->lock)
2072 spin_unlock_irq(&pool->lock);
2074 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2075 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2078 if (create_worker(pool) || !need_to_create_worker(pool))
2081 schedule_timeout_interruptible(CREATE_COOLDOWN);
2083 if (!need_to_create_worker(pool))
2087 del_timer_sync(&pool->mayday_timer);
2088 spin_lock_irq(&pool->lock);
2090 * This is necessary even after a new worker was just successfully
2091 * created as @pool->lock was dropped and the new worker might have
2092 * already become busy.
2094 if (need_to_create_worker(pool))
2099 * manage_workers - manage worker pool
2102 * Assume the manager role and manage the worker pool @worker belongs
2103 * to. At any given time, there can be only zero or one manager per
2104 * pool. The exclusion is handled automatically by this function.
2106 * The caller can safely start processing works on false return. On
2107 * true return, it's guaranteed that need_to_create_worker() is false
2108 * and may_start_working() is true.
2111 * spin_lock_irq(pool->lock) which may be released and regrabbed
2112 * multiple times. Does GFP_KERNEL allocations.
2115 * %false if the pool doesn't need management and the caller can safely
2116 * start processing works, %true if management function was performed and
2117 * the conditions that the caller verified before calling the function may
2118 * no longer be true.
2120 static bool manage_workers(struct worker *worker)
2122 struct worker_pool *pool = worker->pool;
2124 if (pool->flags & POOL_MANAGER_ACTIVE)
2127 pool->flags |= POOL_MANAGER_ACTIVE;
2128 pool->manager = worker;
2130 maybe_create_worker(pool);
2132 pool->manager = NULL;
2133 pool->flags &= ~POOL_MANAGER_ACTIVE;
2134 wake_up(&wq_manager_wait);
2139 * process_one_work - process single work
2141 * @work: work to process
2143 * Process @work. This function contains all the logics necessary to
2144 * process a single work including synchronization against and
2145 * interaction with other workers on the same cpu, queueing and
2146 * flushing. As long as context requirement is met, any worker can
2147 * call this function to process a work.
2150 * spin_lock_irq(pool->lock) which is released and regrabbed.
2152 static void process_one_work(struct worker *worker, struct work_struct *work)
2153 __releases(&pool->lock)
2154 __acquires(&pool->lock)
2156 struct pool_workqueue *pwq = get_work_pwq(work);
2157 struct worker_pool *pool = worker->pool;
2158 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2160 struct worker *collision;
2161 #ifdef CONFIG_LOCKDEP
2163 * It is permissible to free the struct work_struct from
2164 * inside the function that is called from it, this we need to
2165 * take into account for lockdep too. To avoid bogus "held
2166 * lock freed" warnings as well as problems when looking into
2167 * work->lockdep_map, make a copy and use that here.
2169 struct lockdep_map lockdep_map;
2171 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2173 /* ensure we're on the correct CPU */
2174 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2175 raw_smp_processor_id() != pool->cpu);
2178 * A single work shouldn't be executed concurrently by
2179 * multiple workers on a single cpu. Check whether anyone is
2180 * already processing the work. If so, defer the work to the
2181 * currently executing one.
2183 collision = find_worker_executing_work(pool, work);
2184 if (unlikely(collision)) {
2185 move_linked_works(work, &collision->scheduled, NULL);
2189 /* claim and dequeue */
2190 debug_work_deactivate(work);
2191 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2192 worker->current_work = work;
2193 worker->current_func = work->func;
2194 worker->current_pwq = pwq;
2195 work_color = get_work_color(work);
2198 * Record wq name for cmdline and debug reporting, may get
2199 * overridden through set_worker_desc().
2201 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2203 list_del_init(&work->entry);
2206 * CPU intensive works don't participate in concurrency management.
2207 * They're the scheduler's responsibility. This takes @worker out
2208 * of concurrency management and the next code block will chain
2209 * execution of the pending work items.
2211 if (unlikely(cpu_intensive))
2212 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2215 * Wake up another worker if necessary. The condition is always
2216 * false for normal per-cpu workers since nr_running would always
2217 * be >= 1 at this point. This is used to chain execution of the
2218 * pending work items for WORKER_NOT_RUNNING workers such as the
2219 * UNBOUND and CPU_INTENSIVE ones.
2221 if (need_more_worker(pool))
2222 wake_up_worker(pool);
2225 * Record the last pool and clear PENDING which should be the last
2226 * update to @work. Also, do this inside @pool->lock so that
2227 * PENDING and queued state changes happen together while IRQ is
2230 set_work_pool_and_clear_pending(work, pool->id);
2232 spin_unlock_irq(&pool->lock);
2234 lock_map_acquire(&pwq->wq->lockdep_map);
2235 lock_map_acquire(&lockdep_map);
2237 * Strictly speaking we should mark the invariant state without holding
2238 * any locks, that is, before these two lock_map_acquire()'s.
2240 * However, that would result in:
2247 * Which would create W1->C->W1 dependencies, even though there is no
2248 * actual deadlock possible. There are two solutions, using a
2249 * read-recursive acquire on the work(queue) 'locks', but this will then
2250 * hit the lockdep limitation on recursive locks, or simply discard
2253 * AFAICT there is no possible deadlock scenario between the
2254 * flush_work() and complete() primitives (except for single-threaded
2255 * workqueues), so hiding them isn't a problem.
2257 lockdep_invariant_state(true);
2258 trace_workqueue_execute_start(work);
2259 worker->current_func(work);
2261 * While we must be careful to not use "work" after this, the trace
2262 * point will only record its address.
2264 trace_workqueue_execute_end(work);
2265 lock_map_release(&lockdep_map);
2266 lock_map_release(&pwq->wq->lockdep_map);
2268 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2269 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2270 " last function: %pf\n",
2271 current->comm, preempt_count(), task_pid_nr(current),
2272 worker->current_func);
2273 debug_show_held_locks(current);
2278 * The following prevents a kworker from hogging CPU on !PREEMPT
2279 * kernels, where a requeueing work item waiting for something to
2280 * happen could deadlock with stop_machine as such work item could
2281 * indefinitely requeue itself while all other CPUs are trapped in
2282 * stop_machine. At the same time, report a quiescent RCU state so
2283 * the same condition doesn't freeze RCU.
2287 spin_lock_irq(&pool->lock);
2289 /* clear cpu intensive status */
2290 if (unlikely(cpu_intensive))
2291 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2293 /* tag the worker for identification in schedule() */
2294 worker->last_func = worker->current_func;
2296 /* we're done with it, release */
2297 hash_del(&worker->hentry);
2298 worker->current_work = NULL;
2299 worker->current_func = NULL;
2300 worker->current_pwq = NULL;
2301 pwq_dec_nr_in_flight(pwq, work_color);
2305 * process_scheduled_works - process scheduled works
2308 * Process all scheduled works. Please note that the scheduled list
2309 * may change while processing a work, so this function repeatedly
2310 * fetches a work from the top and executes it.
2313 * spin_lock_irq(pool->lock) which may be released and regrabbed
2316 static void process_scheduled_works(struct worker *worker)
2318 while (!list_empty(&worker->scheduled)) {
2319 struct work_struct *work = list_first_entry(&worker->scheduled,
2320 struct work_struct, entry);
2321 process_one_work(worker, work);
2325 static void set_pf_worker(bool val)
2327 mutex_lock(&wq_pool_attach_mutex);
2329 current->flags |= PF_WQ_WORKER;
2331 current->flags &= ~PF_WQ_WORKER;
2332 mutex_unlock(&wq_pool_attach_mutex);
2336 * worker_thread - the worker thread function
2339 * The worker thread function. All workers belong to a worker_pool -
2340 * either a per-cpu one or dynamic unbound one. These workers process all
2341 * work items regardless of their specific target workqueue. The only
2342 * exception is work items which belong to workqueues with a rescuer which
2343 * will be explained in rescuer_thread().
2347 static int worker_thread(void *__worker)
2349 struct worker *worker = __worker;
2350 struct worker_pool *pool = worker->pool;
2352 /* tell the scheduler that this is a workqueue worker */
2353 set_pf_worker(true);
2355 spin_lock_irq(&pool->lock);
2357 /* am I supposed to die? */
2358 if (unlikely(worker->flags & WORKER_DIE)) {
2359 spin_unlock_irq(&pool->lock);
2360 WARN_ON_ONCE(!list_empty(&worker->entry));
2361 set_pf_worker(false);
2363 set_task_comm(worker->task, "kworker/dying");
2364 ida_simple_remove(&pool->worker_ida, worker->id);
2365 worker_detach_from_pool(worker);
2370 worker_leave_idle(worker);
2372 /* no more worker necessary? */
2373 if (!need_more_worker(pool))
2376 /* do we need to manage? */
2377 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2381 * ->scheduled list can only be filled while a worker is
2382 * preparing to process a work or actually processing it.
2383 * Make sure nobody diddled with it while I was sleeping.
2385 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2388 * Finish PREP stage. We're guaranteed to have at least one idle
2389 * worker or that someone else has already assumed the manager
2390 * role. This is where @worker starts participating in concurrency
2391 * management if applicable and concurrency management is restored
2392 * after being rebound. See rebind_workers() for details.
2394 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2397 struct work_struct *work =
2398 list_first_entry(&pool->worklist,
2399 struct work_struct, entry);
2401 pool->watchdog_ts = jiffies;
2403 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2404 /* optimization path, not strictly necessary */
2405 process_one_work(worker, work);
2406 if (unlikely(!list_empty(&worker->scheduled)))
2407 process_scheduled_works(worker);
2409 move_linked_works(work, &worker->scheduled, NULL);
2410 process_scheduled_works(worker);
2412 } while (keep_working(pool));
2414 worker_set_flags(worker, WORKER_PREP);
2417 * pool->lock is held and there's no work to process and no need to
2418 * manage, sleep. Workers are woken up only while holding
2419 * pool->lock or from local cpu, so setting the current state
2420 * before releasing pool->lock is enough to prevent losing any
2423 worker_enter_idle(worker);
2424 __set_current_state(TASK_IDLE);
2425 spin_unlock_irq(&pool->lock);
2431 * rescuer_thread - the rescuer thread function
2434 * Workqueue rescuer thread function. There's one rescuer for each
2435 * workqueue which has WQ_MEM_RECLAIM set.
2437 * Regular work processing on a pool may block trying to create a new
2438 * worker which uses GFP_KERNEL allocation which has slight chance of
2439 * developing into deadlock if some works currently on the same queue
2440 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2441 * the problem rescuer solves.
2443 * When such condition is possible, the pool summons rescuers of all
2444 * workqueues which have works queued on the pool and let them process
2445 * those works so that forward progress can be guaranteed.
2447 * This should happen rarely.
2451 static int rescuer_thread(void *__rescuer)
2453 struct worker *rescuer = __rescuer;
2454 struct workqueue_struct *wq = rescuer->rescue_wq;
2455 struct list_head *scheduled = &rescuer->scheduled;
2458 set_user_nice(current, RESCUER_NICE_LEVEL);
2461 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2462 * doesn't participate in concurrency management.
2464 set_pf_worker(true);
2466 set_current_state(TASK_IDLE);
2469 * By the time the rescuer is requested to stop, the workqueue
2470 * shouldn't have any work pending, but @wq->maydays may still have
2471 * pwq(s) queued. This can happen by non-rescuer workers consuming
2472 * all the work items before the rescuer got to them. Go through
2473 * @wq->maydays processing before acting on should_stop so that the
2474 * list is always empty on exit.
2476 should_stop = kthread_should_stop();
2478 /* see whether any pwq is asking for help */
2479 spin_lock_irq(&wq_mayday_lock);
2481 while (!list_empty(&wq->maydays)) {
2482 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2483 struct pool_workqueue, mayday_node);
2484 struct worker_pool *pool = pwq->pool;
2485 struct work_struct *work, *n;
2488 __set_current_state(TASK_RUNNING);
2489 list_del_init(&pwq->mayday_node);
2491 spin_unlock_irq(&wq_mayday_lock);
2493 worker_attach_to_pool(rescuer, pool);
2495 spin_lock_irq(&pool->lock);
2498 * Slurp in all works issued via this workqueue and
2501 WARN_ON_ONCE(!list_empty(scheduled));
2502 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2503 if (get_work_pwq(work) == pwq) {
2505 pool->watchdog_ts = jiffies;
2506 move_linked_works(work, scheduled, &n);
2511 if (!list_empty(scheduled)) {
2512 process_scheduled_works(rescuer);
2515 * The above execution of rescued work items could
2516 * have created more to rescue through
2517 * pwq_activate_first_delayed() or chained
2518 * queueing. Let's put @pwq back on mayday list so
2519 * that such back-to-back work items, which may be
2520 * being used to relieve memory pressure, don't
2521 * incur MAYDAY_INTERVAL delay inbetween.
2523 if (need_to_create_worker(pool)) {
2524 spin_lock(&wq_mayday_lock);
2526 list_move_tail(&pwq->mayday_node, &wq->maydays);
2527 spin_unlock(&wq_mayday_lock);
2532 * Put the reference grabbed by send_mayday(). @pool won't
2533 * go away while we're still attached to it.
2538 * Leave this pool. If need_more_worker() is %true, notify a
2539 * regular worker; otherwise, we end up with 0 concurrency
2540 * and stalling the execution.
2542 if (need_more_worker(pool))
2543 wake_up_worker(pool);
2545 spin_unlock_irq(&pool->lock);
2547 worker_detach_from_pool(rescuer);
2549 spin_lock_irq(&wq_mayday_lock);
2552 spin_unlock_irq(&wq_mayday_lock);
2555 __set_current_state(TASK_RUNNING);
2556 set_pf_worker(false);
2560 /* rescuers should never participate in concurrency management */
2561 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2567 * check_flush_dependency - check for flush dependency sanity
2568 * @target_wq: workqueue being flushed
2569 * @target_work: work item being flushed (NULL for workqueue flushes)
2571 * %current is trying to flush the whole @target_wq or @target_work on it.
2572 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2573 * reclaiming memory or running on a workqueue which doesn't have
2574 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2577 static void check_flush_dependency(struct workqueue_struct *target_wq,
2578 struct work_struct *target_work)
2580 work_func_t target_func = target_work ? target_work->func : NULL;
2581 struct worker *worker;
2583 if (target_wq->flags & WQ_MEM_RECLAIM)
2586 worker = current_wq_worker();
2588 WARN_ONCE(current->flags & PF_MEMALLOC,
2589 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2590 current->pid, current->comm, target_wq->name, target_func);
2591 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2592 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2593 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2594 worker->current_pwq->wq->name, worker->current_func,
2595 target_wq->name, target_func);
2599 struct work_struct work;
2600 struct completion done;
2601 struct task_struct *task; /* purely informational */
2604 static void wq_barrier_func(struct work_struct *work)
2606 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2607 complete(&barr->done);
2611 * insert_wq_barrier - insert a barrier work
2612 * @pwq: pwq to insert barrier into
2613 * @barr: wq_barrier to insert
2614 * @target: target work to attach @barr to
2615 * @worker: worker currently executing @target, NULL if @target is not executing
2617 * @barr is linked to @target such that @barr is completed only after
2618 * @target finishes execution. Please note that the ordering
2619 * guarantee is observed only with respect to @target and on the local
2622 * Currently, a queued barrier can't be canceled. This is because
2623 * try_to_grab_pending() can't determine whether the work to be
2624 * grabbed is at the head of the queue and thus can't clear LINKED
2625 * flag of the previous work while there must be a valid next work
2626 * after a work with LINKED flag set.
2628 * Note that when @worker is non-NULL, @target may be modified
2629 * underneath us, so we can't reliably determine pwq from @target.
2632 * spin_lock_irq(pool->lock).
2634 static void insert_wq_barrier(struct pool_workqueue *pwq,
2635 struct wq_barrier *barr,
2636 struct work_struct *target, struct worker *worker)
2638 struct list_head *head;
2639 unsigned int linked = 0;
2642 * debugobject calls are safe here even with pool->lock locked
2643 * as we know for sure that this will not trigger any of the
2644 * checks and call back into the fixup functions where we
2647 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2648 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2650 init_completion_map(&barr->done, &target->lockdep_map);
2652 barr->task = current;
2655 * If @target is currently being executed, schedule the
2656 * barrier to the worker; otherwise, put it after @target.
2659 head = worker->scheduled.next;
2661 unsigned long *bits = work_data_bits(target);
2663 head = target->entry.next;
2664 /* there can already be other linked works, inherit and set */
2665 linked = *bits & WORK_STRUCT_LINKED;
2666 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2669 debug_work_activate(&barr->work);
2670 insert_work(pwq, &barr->work, head,
2671 work_color_to_flags(WORK_NO_COLOR) | linked);
2675 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2676 * @wq: workqueue being flushed
2677 * @flush_color: new flush color, < 0 for no-op
2678 * @work_color: new work color, < 0 for no-op
2680 * Prepare pwqs for workqueue flushing.
2682 * If @flush_color is non-negative, flush_color on all pwqs should be
2683 * -1. If no pwq has in-flight commands at the specified color, all
2684 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2685 * has in flight commands, its pwq->flush_color is set to
2686 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2687 * wakeup logic is armed and %true is returned.
2689 * The caller should have initialized @wq->first_flusher prior to
2690 * calling this function with non-negative @flush_color. If
2691 * @flush_color is negative, no flush color update is done and %false
2694 * If @work_color is non-negative, all pwqs should have the same
2695 * work_color which is previous to @work_color and all will be
2696 * advanced to @work_color.
2699 * mutex_lock(wq->mutex).
2702 * %true if @flush_color >= 0 and there's something to flush. %false
2705 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2706 int flush_color, int work_color)
2709 struct pool_workqueue *pwq;
2711 if (flush_color >= 0) {
2712 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2713 atomic_set(&wq->nr_pwqs_to_flush, 1);
2716 for_each_pwq(pwq, wq) {
2717 struct worker_pool *pool = pwq->pool;
2719 spin_lock_irq(&pool->lock);
2721 if (flush_color >= 0) {
2722 WARN_ON_ONCE(pwq->flush_color != -1);
2724 if (pwq->nr_in_flight[flush_color]) {
2725 pwq->flush_color = flush_color;
2726 atomic_inc(&wq->nr_pwqs_to_flush);
2731 if (work_color >= 0) {
2732 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2733 pwq->work_color = work_color;
2736 spin_unlock_irq(&pool->lock);
2739 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2740 complete(&wq->first_flusher->done);
2746 * flush_workqueue - ensure that any scheduled work has run to completion.
2747 * @wq: workqueue to flush
2749 * This function sleeps until all work items which were queued on entry
2750 * have finished execution, but it is not livelocked by new incoming ones.
2752 void flush_workqueue(struct workqueue_struct *wq)
2754 struct wq_flusher this_flusher = {
2755 .list = LIST_HEAD_INIT(this_flusher.list),
2757 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2761 if (WARN_ON(!wq_online))
2764 lock_map_acquire(&wq->lockdep_map);
2765 lock_map_release(&wq->lockdep_map);
2767 mutex_lock(&wq->mutex);
2770 * Start-to-wait phase
2772 next_color = work_next_color(wq->work_color);
2774 if (next_color != wq->flush_color) {
2776 * Color space is not full. The current work_color
2777 * becomes our flush_color and work_color is advanced
2780 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2781 this_flusher.flush_color = wq->work_color;
2782 wq->work_color = next_color;
2784 if (!wq->first_flusher) {
2785 /* no flush in progress, become the first flusher */
2786 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2788 wq->first_flusher = &this_flusher;
2790 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2792 /* nothing to flush, done */
2793 wq->flush_color = next_color;
2794 wq->first_flusher = NULL;
2799 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2800 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2801 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2805 * Oops, color space is full, wait on overflow queue.
2806 * The next flush completion will assign us
2807 * flush_color and transfer to flusher_queue.
2809 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2812 check_flush_dependency(wq, NULL);
2814 mutex_unlock(&wq->mutex);
2816 wait_for_completion(&this_flusher.done);
2819 * Wake-up-and-cascade phase
2821 * First flushers are responsible for cascading flushes and
2822 * handling overflow. Non-first flushers can simply return.
2824 if (wq->first_flusher != &this_flusher)
2827 mutex_lock(&wq->mutex);
2829 /* we might have raced, check again with mutex held */
2830 if (wq->first_flusher != &this_flusher)
2833 wq->first_flusher = NULL;
2835 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2836 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2839 struct wq_flusher *next, *tmp;
2841 /* complete all the flushers sharing the current flush color */
2842 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2843 if (next->flush_color != wq->flush_color)
2845 list_del_init(&next->list);
2846 complete(&next->done);
2849 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2850 wq->flush_color != work_next_color(wq->work_color));
2852 /* this flush_color is finished, advance by one */
2853 wq->flush_color = work_next_color(wq->flush_color);
2855 /* one color has been freed, handle overflow queue */
2856 if (!list_empty(&wq->flusher_overflow)) {
2858 * Assign the same color to all overflowed
2859 * flushers, advance work_color and append to
2860 * flusher_queue. This is the start-to-wait
2861 * phase for these overflowed flushers.
2863 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2864 tmp->flush_color = wq->work_color;
2866 wq->work_color = work_next_color(wq->work_color);
2868 list_splice_tail_init(&wq->flusher_overflow,
2869 &wq->flusher_queue);
2870 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2873 if (list_empty(&wq->flusher_queue)) {
2874 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2879 * Need to flush more colors. Make the next flusher
2880 * the new first flusher and arm pwqs.
2882 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2883 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2885 list_del_init(&next->list);
2886 wq->first_flusher = next;
2888 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2892 * Meh... this color is already done, clear first
2893 * flusher and repeat cascading.
2895 wq->first_flusher = NULL;
2899 mutex_unlock(&wq->mutex);
2901 EXPORT_SYMBOL(flush_workqueue);
2904 * drain_workqueue - drain a workqueue
2905 * @wq: workqueue to drain
2907 * Wait until the workqueue becomes empty. While draining is in progress,
2908 * only chain queueing is allowed. IOW, only currently pending or running
2909 * work items on @wq can queue further work items on it. @wq is flushed
2910 * repeatedly until it becomes empty. The number of flushing is determined
2911 * by the depth of chaining and should be relatively short. Whine if it
2914 void drain_workqueue(struct workqueue_struct *wq)
2916 unsigned int flush_cnt = 0;
2917 struct pool_workqueue *pwq;
2920 * __queue_work() needs to test whether there are drainers, is much
2921 * hotter than drain_workqueue() and already looks at @wq->flags.
2922 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2924 mutex_lock(&wq->mutex);
2925 if (!wq->nr_drainers++)
2926 wq->flags |= __WQ_DRAINING;
2927 mutex_unlock(&wq->mutex);
2929 flush_workqueue(wq);
2931 mutex_lock(&wq->mutex);
2933 for_each_pwq(pwq, wq) {
2936 spin_lock_irq(&pwq->pool->lock);
2937 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2938 spin_unlock_irq(&pwq->pool->lock);
2943 if (++flush_cnt == 10 ||
2944 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2945 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2946 wq->name, flush_cnt);
2948 mutex_unlock(&wq->mutex);
2952 if (!--wq->nr_drainers)
2953 wq->flags &= ~__WQ_DRAINING;
2954 mutex_unlock(&wq->mutex);
2956 EXPORT_SYMBOL_GPL(drain_workqueue);
2958 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2961 struct worker *worker = NULL;
2962 struct worker_pool *pool;
2963 struct pool_workqueue *pwq;
2967 local_irq_disable();
2968 pool = get_work_pool(work);
2974 spin_lock(&pool->lock);
2975 /* see the comment in try_to_grab_pending() with the same code */
2976 pwq = get_work_pwq(work);
2978 if (unlikely(pwq->pool != pool))
2981 worker = find_worker_executing_work(pool, work);
2984 pwq = worker->current_pwq;
2987 check_flush_dependency(pwq->wq, work);
2989 insert_wq_barrier(pwq, barr, work, worker);
2990 spin_unlock_irq(&pool->lock);
2993 * Force a lock recursion deadlock when using flush_work() inside a
2994 * single-threaded or rescuer equipped workqueue.
2996 * For single threaded workqueues the deadlock happens when the work
2997 * is after the work issuing the flush_work(). For rescuer equipped
2998 * workqueues the deadlock happens when the rescuer stalls, blocking
3002 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3003 lock_map_acquire(&pwq->wq->lockdep_map);
3004 lock_map_release(&pwq->wq->lockdep_map);
3009 spin_unlock_irq(&pool->lock);
3013 static bool __flush_work(struct work_struct *work, bool from_cancel)
3015 struct wq_barrier barr;
3017 if (WARN_ON(!wq_online))
3021 lock_map_acquire(&work->lockdep_map);
3022 lock_map_release(&work->lockdep_map);
3025 if (start_flush_work(work, &barr, from_cancel)) {
3026 wait_for_completion(&barr.done);
3027 destroy_work_on_stack(&barr.work);
3035 * flush_work - wait for a work to finish executing the last queueing instance
3036 * @work: the work to flush
3038 * Wait until @work has finished execution. @work is guaranteed to be idle
3039 * on return if it hasn't been requeued since flush started.
3042 * %true if flush_work() waited for the work to finish execution,
3043 * %false if it was already idle.
3045 bool flush_work(struct work_struct *work)
3047 return __flush_work(work, false);
3049 EXPORT_SYMBOL_GPL(flush_work);
3052 wait_queue_entry_t wait;
3053 struct work_struct *work;
3056 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3058 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3060 if (cwait->work != key)
3062 return autoremove_wake_function(wait, mode, sync, key);
3065 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3067 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3068 unsigned long flags;
3072 ret = try_to_grab_pending(work, is_dwork, &flags);
3074 * If someone else is already canceling, wait for it to
3075 * finish. flush_work() doesn't work for PREEMPT_NONE
3076 * because we may get scheduled between @work's completion
3077 * and the other canceling task resuming and clearing
3078 * CANCELING - flush_work() will return false immediately
3079 * as @work is no longer busy, try_to_grab_pending() will
3080 * return -ENOENT as @work is still being canceled and the
3081 * other canceling task won't be able to clear CANCELING as
3082 * we're hogging the CPU.
3084 * Let's wait for completion using a waitqueue. As this
3085 * may lead to the thundering herd problem, use a custom
3086 * wake function which matches @work along with exclusive
3089 if (unlikely(ret == -ENOENT)) {
3090 struct cwt_wait cwait;
3092 init_wait(&cwait.wait);
3093 cwait.wait.func = cwt_wakefn;
3096 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3097 TASK_UNINTERRUPTIBLE);
3098 if (work_is_canceling(work))
3100 finish_wait(&cancel_waitq, &cwait.wait);
3102 } while (unlikely(ret < 0));
3104 /* tell other tasks trying to grab @work to back off */
3105 mark_work_canceling(work);
3106 local_irq_restore(flags);
3109 * This allows canceling during early boot. We know that @work
3113 __flush_work(work, true);
3115 clear_work_data(work);
3118 * Paired with prepare_to_wait() above so that either
3119 * waitqueue_active() is visible here or !work_is_canceling() is
3123 if (waitqueue_active(&cancel_waitq))
3124 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3130 * cancel_work_sync - cancel a work and wait for it to finish
3131 * @work: the work to cancel
3133 * Cancel @work and wait for its execution to finish. This function
3134 * can be used even if the work re-queues itself or migrates to
3135 * another workqueue. On return from this function, @work is
3136 * guaranteed to be not pending or executing on any CPU.
3138 * cancel_work_sync(&delayed_work->work) must not be used for
3139 * delayed_work's. Use cancel_delayed_work_sync() instead.
3141 * The caller must ensure that the workqueue on which @work was last
3142 * queued can't be destroyed before this function returns.
3145 * %true if @work was pending, %false otherwise.
3147 bool cancel_work_sync(struct work_struct *work)
3149 return __cancel_work_timer(work, false);
3151 EXPORT_SYMBOL_GPL(cancel_work_sync);
3154 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3155 * @dwork: the delayed work to flush
3157 * Delayed timer is cancelled and the pending work is queued for
3158 * immediate execution. Like flush_work(), this function only
3159 * considers the last queueing instance of @dwork.
3162 * %true if flush_work() waited for the work to finish execution,
3163 * %false if it was already idle.
3165 bool flush_delayed_work(struct delayed_work *dwork)
3167 local_irq_disable();
3168 if (del_timer_sync(&dwork->timer))
3169 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3171 return flush_work(&dwork->work);
3173 EXPORT_SYMBOL(flush_delayed_work);
3176 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3177 * @rwork: the rcu work to flush
3180 * %true if flush_rcu_work() waited for the work to finish execution,
3181 * %false if it was already idle.
3183 bool flush_rcu_work(struct rcu_work *rwork)
3185 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3187 flush_work(&rwork->work);
3190 return flush_work(&rwork->work);
3193 EXPORT_SYMBOL(flush_rcu_work);
3195 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3197 unsigned long flags;
3201 ret = try_to_grab_pending(work, is_dwork, &flags);
3202 } while (unlikely(ret == -EAGAIN));
3204 if (unlikely(ret < 0))
3207 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3208 local_irq_restore(flags);
3213 * cancel_delayed_work - cancel a delayed work
3214 * @dwork: delayed_work to cancel
3216 * Kill off a pending delayed_work.
3218 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3222 * The work callback function may still be running on return, unless
3223 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3224 * use cancel_delayed_work_sync() to wait on it.
3226 * This function is safe to call from any context including IRQ handler.
3228 bool cancel_delayed_work(struct delayed_work *dwork)
3230 return __cancel_work(&dwork->work, true);
3232 EXPORT_SYMBOL(cancel_delayed_work);
3235 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3236 * @dwork: the delayed work cancel
3238 * This is cancel_work_sync() for delayed works.
3241 * %true if @dwork was pending, %false otherwise.
3243 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3245 return __cancel_work_timer(&dwork->work, true);
3247 EXPORT_SYMBOL(cancel_delayed_work_sync);
3250 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3251 * @func: the function to call
3253 * schedule_on_each_cpu() executes @func on each online CPU using the
3254 * system workqueue and blocks until all CPUs have completed.
3255 * schedule_on_each_cpu() is very slow.
3258 * 0 on success, -errno on failure.
3260 int schedule_on_each_cpu(work_func_t func)
3263 struct work_struct __percpu *works;
3265 works = alloc_percpu(struct work_struct);
3271 for_each_online_cpu(cpu) {
3272 struct work_struct *work = per_cpu_ptr(works, cpu);
3274 INIT_WORK(work, func);
3275 schedule_work_on(cpu, work);
3278 for_each_online_cpu(cpu)
3279 flush_work(per_cpu_ptr(works, cpu));
3287 * execute_in_process_context - reliably execute the routine with user context
3288 * @fn: the function to execute
3289 * @ew: guaranteed storage for the execute work structure (must
3290 * be available when the work executes)
3292 * Executes the function immediately if process context is available,
3293 * otherwise schedules the function for delayed execution.
3295 * Return: 0 - function was executed
3296 * 1 - function was scheduled for execution
3298 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3300 if (!in_interrupt()) {
3305 INIT_WORK(&ew->work, fn);
3306 schedule_work(&ew->work);
3310 EXPORT_SYMBOL_GPL(execute_in_process_context);
3313 * free_workqueue_attrs - free a workqueue_attrs
3314 * @attrs: workqueue_attrs to free
3316 * Undo alloc_workqueue_attrs().
3318 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3321 free_cpumask_var(attrs->cpumask);
3327 * alloc_workqueue_attrs - allocate a workqueue_attrs
3328 * @gfp_mask: allocation mask to use
3330 * Allocate a new workqueue_attrs, initialize with default settings and
3333 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3335 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3337 struct workqueue_attrs *attrs;
3339 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3342 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3345 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3348 free_workqueue_attrs(attrs);
3352 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3353 const struct workqueue_attrs *from)
3355 to->nice = from->nice;
3356 cpumask_copy(to->cpumask, from->cpumask);
3358 * Unlike hash and equality test, this function doesn't ignore
3359 * ->no_numa as it is used for both pool and wq attrs. Instead,
3360 * get_unbound_pool() explicitly clears ->no_numa after copying.
3362 to->no_numa = from->no_numa;
3365 /* hash value of the content of @attr */
3366 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3370 hash = jhash_1word(attrs->nice, hash);
3371 hash = jhash(cpumask_bits(attrs->cpumask),
3372 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3376 /* content equality test */
3377 static bool wqattrs_equal(const struct workqueue_attrs *a,
3378 const struct workqueue_attrs *b)
3380 if (a->nice != b->nice)
3382 if (!cpumask_equal(a->cpumask, b->cpumask))
3388 * init_worker_pool - initialize a newly zalloc'd worker_pool
3389 * @pool: worker_pool to initialize
3391 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3393 * Return: 0 on success, -errno on failure. Even on failure, all fields
3394 * inside @pool proper are initialized and put_unbound_pool() can be called
3395 * on @pool safely to release it.
3397 static int init_worker_pool(struct worker_pool *pool)
3399 spin_lock_init(&pool->lock);
3402 pool->node = NUMA_NO_NODE;
3403 pool->flags |= POOL_DISASSOCIATED;
3404 pool->watchdog_ts = jiffies;
3405 INIT_LIST_HEAD(&pool->worklist);
3406 INIT_LIST_HEAD(&pool->idle_list);
3407 hash_init(pool->busy_hash);
3409 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3411 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3413 INIT_LIST_HEAD(&pool->workers);
3415 ida_init(&pool->worker_ida);
3416 INIT_HLIST_NODE(&pool->hash_node);
3419 /* shouldn't fail above this point */
3420 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3426 #ifdef CONFIG_LOCKDEP
3427 static void wq_init_lockdep(struct workqueue_struct *wq)
3431 lockdep_register_key(&wq->key);
3432 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3434 lock_name = wq->name;
3435 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3438 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3440 lockdep_unregister_key(&wq->key);
3443 static void wq_free_lockdep(struct workqueue_struct *wq)
3445 if (wq->lock_name != wq->name)
3446 kfree(wq->lock_name);
3449 static void wq_init_lockdep(struct workqueue_struct *wq)
3453 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3457 static void wq_free_lockdep(struct workqueue_struct *wq)
3462 static void rcu_free_wq(struct rcu_head *rcu)
3464 struct workqueue_struct *wq =
3465 container_of(rcu, struct workqueue_struct, rcu);
3467 wq_free_lockdep(wq);
3469 if (!(wq->flags & WQ_UNBOUND))
3470 free_percpu(wq->cpu_pwqs);
3472 free_workqueue_attrs(wq->unbound_attrs);
3478 static void rcu_free_pool(struct rcu_head *rcu)
3480 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3482 ida_destroy(&pool->worker_ida);
3483 free_workqueue_attrs(pool->attrs);
3488 * put_unbound_pool - put a worker_pool
3489 * @pool: worker_pool to put
3491 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3492 * safe manner. get_unbound_pool() calls this function on its failure path
3493 * and this function should be able to release pools which went through,
3494 * successfully or not, init_worker_pool().
3496 * Should be called with wq_pool_mutex held.
3498 static void put_unbound_pool(struct worker_pool *pool)
3500 DECLARE_COMPLETION_ONSTACK(detach_completion);
3501 struct worker *worker;
3503 lockdep_assert_held(&wq_pool_mutex);
3509 if (WARN_ON(!(pool->cpu < 0)) ||
3510 WARN_ON(!list_empty(&pool->worklist)))
3513 /* release id and unhash */
3515 idr_remove(&worker_pool_idr, pool->id);
3516 hash_del(&pool->hash_node);
3519 * Become the manager and destroy all workers. This prevents
3520 * @pool's workers from blocking on attach_mutex. We're the last
3521 * manager and @pool gets freed with the flag set.
3523 spin_lock_irq(&pool->lock);
3524 wait_event_lock_irq(wq_manager_wait,
3525 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3526 pool->flags |= POOL_MANAGER_ACTIVE;
3528 while ((worker = first_idle_worker(pool)))
3529 destroy_worker(worker);
3530 WARN_ON(pool->nr_workers || pool->nr_idle);
3531 spin_unlock_irq(&pool->lock);
3533 mutex_lock(&wq_pool_attach_mutex);
3534 if (!list_empty(&pool->workers))
3535 pool->detach_completion = &detach_completion;
3536 mutex_unlock(&wq_pool_attach_mutex);
3538 if (pool->detach_completion)
3539 wait_for_completion(pool->detach_completion);
3541 /* shut down the timers */
3542 del_timer_sync(&pool->idle_timer);
3543 del_timer_sync(&pool->mayday_timer);
3545 /* sched-RCU protected to allow dereferences from get_work_pool() */
3546 call_rcu(&pool->rcu, rcu_free_pool);
3550 * get_unbound_pool - get a worker_pool with the specified attributes
3551 * @attrs: the attributes of the worker_pool to get
3553 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3554 * reference count and return it. If there already is a matching
3555 * worker_pool, it will be used; otherwise, this function attempts to
3558 * Should be called with wq_pool_mutex held.
3560 * Return: On success, a worker_pool with the same attributes as @attrs.
3561 * On failure, %NULL.
3563 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3565 u32 hash = wqattrs_hash(attrs);
3566 struct worker_pool *pool;
3568 int target_node = NUMA_NO_NODE;
3570 lockdep_assert_held(&wq_pool_mutex);
3572 /* do we already have a matching pool? */
3573 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3574 if (wqattrs_equal(pool->attrs, attrs)) {
3580 /* if cpumask is contained inside a NUMA node, we belong to that node */
3581 if (wq_numa_enabled) {
3582 for_each_node(node) {
3583 if (cpumask_subset(attrs->cpumask,
3584 wq_numa_possible_cpumask[node])) {
3591 /* nope, create a new one */
3592 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3593 if (!pool || init_worker_pool(pool) < 0)
3596 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3597 copy_workqueue_attrs(pool->attrs, attrs);
3598 pool->node = target_node;
3601 * no_numa isn't a worker_pool attribute, always clear it. See
3602 * 'struct workqueue_attrs' comments for detail.
3604 pool->attrs->no_numa = false;
3606 if (worker_pool_assign_id(pool) < 0)
3609 /* create and start the initial worker */
3610 if (wq_online && !create_worker(pool))
3614 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3619 put_unbound_pool(pool);
3623 static void rcu_free_pwq(struct rcu_head *rcu)
3625 kmem_cache_free(pwq_cache,
3626 container_of(rcu, struct pool_workqueue, rcu));
3630 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3631 * and needs to be destroyed.
3633 static void pwq_unbound_release_workfn(struct work_struct *work)
3635 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3636 unbound_release_work);
3637 struct workqueue_struct *wq = pwq->wq;
3638 struct worker_pool *pool = pwq->pool;
3641 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3644 mutex_lock(&wq->mutex);
3645 list_del_rcu(&pwq->pwqs_node);
3646 is_last = list_empty(&wq->pwqs);
3647 mutex_unlock(&wq->mutex);
3649 mutex_lock(&wq_pool_mutex);
3650 put_unbound_pool(pool);
3651 mutex_unlock(&wq_pool_mutex);
3653 call_rcu(&pwq->rcu, rcu_free_pwq);
3656 * If we're the last pwq going away, @wq is already dead and no one
3657 * is gonna access it anymore. Schedule RCU free.
3660 wq_unregister_lockdep(wq);
3661 call_rcu(&wq->rcu, rcu_free_wq);
3666 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3667 * @pwq: target pool_workqueue
3669 * If @pwq isn't freezing, set @pwq->max_active to the associated
3670 * workqueue's saved_max_active and activate delayed work items
3671 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3673 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3675 struct workqueue_struct *wq = pwq->wq;
3676 bool freezable = wq->flags & WQ_FREEZABLE;
3677 unsigned long flags;
3679 /* for @wq->saved_max_active */
3680 lockdep_assert_held(&wq->mutex);
3682 /* fast exit for non-freezable wqs */
3683 if (!freezable && pwq->max_active == wq->saved_max_active)
3686 /* this function can be called during early boot w/ irq disabled */
3687 spin_lock_irqsave(&pwq->pool->lock, flags);
3690 * During [un]freezing, the caller is responsible for ensuring that
3691 * this function is called at least once after @workqueue_freezing
3692 * is updated and visible.
3694 if (!freezable || !workqueue_freezing) {
3695 pwq->max_active = wq->saved_max_active;
3697 while (!list_empty(&pwq->delayed_works) &&
3698 pwq->nr_active < pwq->max_active)
3699 pwq_activate_first_delayed(pwq);
3702 * Need to kick a worker after thawed or an unbound wq's
3703 * max_active is bumped. It's a slow path. Do it always.
3705 wake_up_worker(pwq->pool);
3707 pwq->max_active = 0;
3710 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3713 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3714 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3715 struct worker_pool *pool)
3717 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3719 memset(pwq, 0, sizeof(*pwq));
3723 pwq->flush_color = -1;
3725 INIT_LIST_HEAD(&pwq->delayed_works);
3726 INIT_LIST_HEAD(&pwq->pwqs_node);
3727 INIT_LIST_HEAD(&pwq->mayday_node);
3728 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3731 /* sync @pwq with the current state of its associated wq and link it */
3732 static void link_pwq(struct pool_workqueue *pwq)
3734 struct workqueue_struct *wq = pwq->wq;
3736 lockdep_assert_held(&wq->mutex);
3738 /* may be called multiple times, ignore if already linked */
3739 if (!list_empty(&pwq->pwqs_node))
3742 /* set the matching work_color */
3743 pwq->work_color = wq->work_color;
3745 /* sync max_active to the current setting */
3746 pwq_adjust_max_active(pwq);
3749 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3752 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3753 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3754 const struct workqueue_attrs *attrs)
3756 struct worker_pool *pool;
3757 struct pool_workqueue *pwq;
3759 lockdep_assert_held(&wq_pool_mutex);
3761 pool = get_unbound_pool(attrs);
3765 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3767 put_unbound_pool(pool);
3771 init_pwq(pwq, wq, pool);
3776 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3777 * @attrs: the wq_attrs of the default pwq of the target workqueue
3778 * @node: the target NUMA node
3779 * @cpu_going_down: if >= 0, the CPU to consider as offline
3780 * @cpumask: outarg, the resulting cpumask
3782 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3783 * @cpu_going_down is >= 0, that cpu is considered offline during
3784 * calculation. The result is stored in @cpumask.
3786 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3787 * enabled and @node has online CPUs requested by @attrs, the returned
3788 * cpumask is the intersection of the possible CPUs of @node and
3791 * The caller is responsible for ensuring that the cpumask of @node stays
3794 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3797 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3798 int cpu_going_down, cpumask_t *cpumask)
3800 if (!wq_numa_enabled || attrs->no_numa)
3803 /* does @node have any online CPUs @attrs wants? */
3804 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3805 if (cpu_going_down >= 0)
3806 cpumask_clear_cpu(cpu_going_down, cpumask);
3808 if (cpumask_empty(cpumask))
3811 /* yeap, return possible CPUs in @node that @attrs wants */
3812 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3814 if (cpumask_empty(cpumask)) {
3815 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3816 "possible intersect\n");
3820 return !cpumask_equal(cpumask, attrs->cpumask);
3823 cpumask_copy(cpumask, attrs->cpumask);
3827 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3828 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3830 struct pool_workqueue *pwq)
3832 struct pool_workqueue *old_pwq;
3834 lockdep_assert_held(&wq_pool_mutex);
3835 lockdep_assert_held(&wq->mutex);
3837 /* link_pwq() can handle duplicate calls */
3840 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3841 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3845 /* context to store the prepared attrs & pwqs before applying */
3846 struct apply_wqattrs_ctx {
3847 struct workqueue_struct *wq; /* target workqueue */
3848 struct workqueue_attrs *attrs; /* attrs to apply */
3849 struct list_head list; /* queued for batching commit */
3850 struct pool_workqueue *dfl_pwq;
3851 struct pool_workqueue *pwq_tbl[];
3854 /* free the resources after success or abort */
3855 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3861 put_pwq_unlocked(ctx->pwq_tbl[node]);
3862 put_pwq_unlocked(ctx->dfl_pwq);
3864 free_workqueue_attrs(ctx->attrs);
3870 /* allocate the attrs and pwqs for later installation */
3871 static struct apply_wqattrs_ctx *
3872 apply_wqattrs_prepare(struct workqueue_struct *wq,
3873 const struct workqueue_attrs *attrs)
3875 struct apply_wqattrs_ctx *ctx;
3876 struct workqueue_attrs *new_attrs, *tmp_attrs;
3879 lockdep_assert_held(&wq_pool_mutex);
3881 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3883 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3884 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3885 if (!ctx || !new_attrs || !tmp_attrs)
3889 * Calculate the attrs of the default pwq.
3890 * If the user configured cpumask doesn't overlap with the
3891 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3893 copy_workqueue_attrs(new_attrs, attrs);
3894 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3895 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3896 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3899 * We may create multiple pwqs with differing cpumasks. Make a
3900 * copy of @new_attrs which will be modified and used to obtain
3903 copy_workqueue_attrs(tmp_attrs, new_attrs);
3906 * If something goes wrong during CPU up/down, we'll fall back to
3907 * the default pwq covering whole @attrs->cpumask. Always create
3908 * it even if we don't use it immediately.
3910 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3914 for_each_node(node) {
3915 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3916 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3917 if (!ctx->pwq_tbl[node])
3920 ctx->dfl_pwq->refcnt++;
3921 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3925 /* save the user configured attrs and sanitize it. */
3926 copy_workqueue_attrs(new_attrs, attrs);
3927 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3928 ctx->attrs = new_attrs;
3931 free_workqueue_attrs(tmp_attrs);
3935 free_workqueue_attrs(tmp_attrs);
3936 free_workqueue_attrs(new_attrs);
3937 apply_wqattrs_cleanup(ctx);
3941 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3942 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3946 /* all pwqs have been created successfully, let's install'em */
3947 mutex_lock(&ctx->wq->mutex);
3949 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3951 /* save the previous pwq and install the new one */
3953 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3954 ctx->pwq_tbl[node]);
3956 /* @dfl_pwq might not have been used, ensure it's linked */
3957 link_pwq(ctx->dfl_pwq);
3958 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3960 mutex_unlock(&ctx->wq->mutex);
3963 static void apply_wqattrs_lock(void)
3965 /* CPUs should stay stable across pwq creations and installations */
3967 mutex_lock(&wq_pool_mutex);
3970 static void apply_wqattrs_unlock(void)
3972 mutex_unlock(&wq_pool_mutex);
3976 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3977 const struct workqueue_attrs *attrs)
3979 struct apply_wqattrs_ctx *ctx;
3981 /* only unbound workqueues can change attributes */
3982 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3985 /* creating multiple pwqs breaks ordering guarantee */
3986 if (!list_empty(&wq->pwqs)) {
3987 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3990 wq->flags &= ~__WQ_ORDERED;
3993 ctx = apply_wqattrs_prepare(wq, attrs);
3997 /* the ctx has been prepared successfully, let's commit it */
3998 apply_wqattrs_commit(ctx);
3999 apply_wqattrs_cleanup(ctx);
4005 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4006 * @wq: the target workqueue
4007 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4009 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4010 * machines, this function maps a separate pwq to each NUMA node with
4011 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4012 * NUMA node it was issued on. Older pwqs are released as in-flight work
4013 * items finish. Note that a work item which repeatedly requeues itself
4014 * back-to-back will stay on its current pwq.
4016 * Performs GFP_KERNEL allocations.
4018 * Return: 0 on success and -errno on failure.
4020 int apply_workqueue_attrs(struct workqueue_struct *wq,
4021 const struct workqueue_attrs *attrs)
4025 apply_wqattrs_lock();
4026 ret = apply_workqueue_attrs_locked(wq, attrs);
4027 apply_wqattrs_unlock();
4031 EXPORT_SYMBOL_GPL(apply_workqueue_attrs);
4034 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4035 * @wq: the target workqueue
4036 * @cpu: the CPU coming up or going down
4037 * @online: whether @cpu is coming up or going down
4039 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4040 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4043 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4044 * falls back to @wq->dfl_pwq which may not be optimal but is always
4047 * Note that when the last allowed CPU of a NUMA node goes offline for a
4048 * workqueue with a cpumask spanning multiple nodes, the workers which were
4049 * already executing the work items for the workqueue will lose their CPU
4050 * affinity and may execute on any CPU. This is similar to how per-cpu
4051 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4052 * affinity, it's the user's responsibility to flush the work item from
4055 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4058 int node = cpu_to_node(cpu);
4059 int cpu_off = online ? -1 : cpu;
4060 struct pool_workqueue *old_pwq = NULL, *pwq;
4061 struct workqueue_attrs *target_attrs;
4064 lockdep_assert_held(&wq_pool_mutex);
4066 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4067 wq->unbound_attrs->no_numa)
4071 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4072 * Let's use a preallocated one. The following buf is protected by
4073 * CPU hotplug exclusion.
4075 target_attrs = wq_update_unbound_numa_attrs_buf;
4076 cpumask = target_attrs->cpumask;
4078 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4079 pwq = unbound_pwq_by_node(wq, node);
4082 * Let's determine what needs to be done. If the target cpumask is
4083 * different from the default pwq's, we need to compare it to @pwq's
4084 * and create a new one if they don't match. If the target cpumask
4085 * equals the default pwq's, the default pwq should be used.
4087 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4088 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4094 /* create a new pwq */
4095 pwq = alloc_unbound_pwq(wq, target_attrs);
4097 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4102 /* Install the new pwq. */
4103 mutex_lock(&wq->mutex);
4104 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4108 mutex_lock(&wq->mutex);
4109 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4110 get_pwq(wq->dfl_pwq);
4111 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4112 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4114 mutex_unlock(&wq->mutex);
4115 put_pwq_unlocked(old_pwq);
4118 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4120 bool highpri = wq->flags & WQ_HIGHPRI;
4123 if (!(wq->flags & WQ_UNBOUND)) {
4124 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4128 for_each_possible_cpu(cpu) {
4129 struct pool_workqueue *pwq =
4130 per_cpu_ptr(wq->cpu_pwqs, cpu);
4131 struct worker_pool *cpu_pools =
4132 per_cpu(cpu_worker_pools, cpu);
4134 init_pwq(pwq, wq, &cpu_pools[highpri]);
4136 mutex_lock(&wq->mutex);
4138 mutex_unlock(&wq->mutex);
4141 } else if (wq->flags & __WQ_ORDERED) {
4142 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4143 /* there should only be single pwq for ordering guarantee */
4144 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4145 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4146 "ordering guarantee broken for workqueue %s\n", wq->name);
4149 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4153 static int wq_clamp_max_active(int max_active, unsigned int flags,
4156 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4158 if (max_active < 1 || max_active > lim)
4159 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4160 max_active, name, 1, lim);
4162 return clamp_val(max_active, 1, lim);
4166 * Workqueues which may be used during memory reclaim should have a rescuer
4167 * to guarantee forward progress.
4169 static int init_rescuer(struct workqueue_struct *wq)
4171 struct worker *rescuer;
4174 if (!(wq->flags & WQ_MEM_RECLAIM))
4177 rescuer = alloc_worker(NUMA_NO_NODE);
4181 rescuer->rescue_wq = wq;
4182 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4183 ret = PTR_ERR_OR_ZERO(rescuer->task);
4189 wq->rescuer = rescuer;
4190 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4191 wake_up_process(rescuer->task);
4196 struct workqueue_struct *alloc_workqueue(const char *fmt,
4198 int max_active, ...)
4200 size_t tbl_size = 0;
4202 struct workqueue_struct *wq;
4203 struct pool_workqueue *pwq;
4206 * Unbound && max_active == 1 used to imply ordered, which is no
4207 * longer the case on NUMA machines due to per-node pools. While
4208 * alloc_ordered_workqueue() is the right way to create an ordered
4209 * workqueue, keep the previous behavior to avoid subtle breakages
4212 if ((flags & WQ_UNBOUND) && max_active == 1)
4213 flags |= __WQ_ORDERED;
4215 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4216 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4217 flags |= WQ_UNBOUND;
4219 /* allocate wq and format name */
4220 if (flags & WQ_UNBOUND)
4221 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4223 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4227 if (flags & WQ_UNBOUND) {
4228 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4229 if (!wq->unbound_attrs)
4233 va_start(args, max_active);
4234 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4237 max_active = max_active ?: WQ_DFL_ACTIVE;
4238 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4242 wq->saved_max_active = max_active;
4243 mutex_init(&wq->mutex);
4244 atomic_set(&wq->nr_pwqs_to_flush, 0);
4245 INIT_LIST_HEAD(&wq->pwqs);
4246 INIT_LIST_HEAD(&wq->flusher_queue);
4247 INIT_LIST_HEAD(&wq->flusher_overflow);
4248 INIT_LIST_HEAD(&wq->maydays);
4250 wq_init_lockdep(wq);
4251 INIT_LIST_HEAD(&wq->list);
4253 if (alloc_and_link_pwqs(wq) < 0)
4256 if (wq_online && init_rescuer(wq) < 0)
4259 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4263 * wq_pool_mutex protects global freeze state and workqueues list.
4264 * Grab it, adjust max_active and add the new @wq to workqueues
4267 mutex_lock(&wq_pool_mutex);
4269 mutex_lock(&wq->mutex);
4270 for_each_pwq(pwq, wq)
4271 pwq_adjust_max_active(pwq);
4272 mutex_unlock(&wq->mutex);
4274 list_add_tail_rcu(&wq->list, &workqueues);
4276 mutex_unlock(&wq_pool_mutex);
4281 free_workqueue_attrs(wq->unbound_attrs);
4285 destroy_workqueue(wq);
4288 EXPORT_SYMBOL_GPL(alloc_workqueue);
4291 * destroy_workqueue - safely terminate a workqueue
4292 * @wq: target workqueue
4294 * Safely destroy a workqueue. All work currently pending will be done first.
4296 void destroy_workqueue(struct workqueue_struct *wq)
4298 struct pool_workqueue *pwq;
4301 /* drain it before proceeding with destruction */
4302 drain_workqueue(wq);
4305 mutex_lock(&wq->mutex);
4306 for_each_pwq(pwq, wq) {
4309 for (i = 0; i < WORK_NR_COLORS; i++) {
4310 if (WARN_ON(pwq->nr_in_flight[i])) {
4311 mutex_unlock(&wq->mutex);
4312 show_workqueue_state();
4317 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4318 WARN_ON(pwq->nr_active) ||
4319 WARN_ON(!list_empty(&pwq->delayed_works))) {
4320 mutex_unlock(&wq->mutex);
4321 show_workqueue_state();
4325 mutex_unlock(&wq->mutex);
4328 * wq list is used to freeze wq, remove from list after
4329 * flushing is complete in case freeze races us.
4331 mutex_lock(&wq_pool_mutex);
4332 list_del_rcu(&wq->list);
4333 mutex_unlock(&wq_pool_mutex);
4335 workqueue_sysfs_unregister(wq);
4338 kthread_stop(wq->rescuer->task);
4340 if (!(wq->flags & WQ_UNBOUND)) {
4341 wq_unregister_lockdep(wq);
4343 * The base ref is never dropped on per-cpu pwqs. Directly
4344 * schedule RCU free.
4346 call_rcu(&wq->rcu, rcu_free_wq);
4349 * We're the sole accessor of @wq at this point. Directly
4350 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4351 * @wq will be freed when the last pwq is released.
4353 for_each_node(node) {
4354 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4355 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4356 put_pwq_unlocked(pwq);
4360 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4361 * put. Don't access it afterwards.
4365 put_pwq_unlocked(pwq);
4368 EXPORT_SYMBOL_GPL(destroy_workqueue);
4371 * workqueue_set_max_active - adjust max_active of a workqueue
4372 * @wq: target workqueue
4373 * @max_active: new max_active value.
4375 * Set max_active of @wq to @max_active.
4378 * Don't call from IRQ context.
4380 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4382 struct pool_workqueue *pwq;
4384 /* disallow meddling with max_active for ordered workqueues */
4385 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4388 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4390 mutex_lock(&wq->mutex);
4392 wq->flags &= ~__WQ_ORDERED;
4393 wq->saved_max_active = max_active;
4395 for_each_pwq(pwq, wq)
4396 pwq_adjust_max_active(pwq);
4398 mutex_unlock(&wq->mutex);
4400 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4403 * current_work - retrieve %current task's work struct
4405 * Determine if %current task is a workqueue worker and what it's working on.
4406 * Useful to find out the context that the %current task is running in.
4408 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4410 struct work_struct *current_work(void)
4412 struct worker *worker = current_wq_worker();
4414 return worker ? worker->current_work : NULL;
4416 EXPORT_SYMBOL(current_work);
4419 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4421 * Determine whether %current is a workqueue rescuer. Can be used from
4422 * work functions to determine whether it's being run off the rescuer task.
4424 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4426 bool current_is_workqueue_rescuer(void)
4428 struct worker *worker = current_wq_worker();
4430 return worker && worker->rescue_wq;
4434 * workqueue_congested - test whether a workqueue is congested
4435 * @cpu: CPU in question
4436 * @wq: target workqueue
4438 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4439 * no synchronization around this function and the test result is
4440 * unreliable and only useful as advisory hints or for debugging.
4442 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4443 * Note that both per-cpu and unbound workqueues may be associated with
4444 * multiple pool_workqueues which have separate congested states. A
4445 * workqueue being congested on one CPU doesn't mean the workqueue is also
4446 * contested on other CPUs / NUMA nodes.
4449 * %true if congested, %false otherwise.
4451 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4453 struct pool_workqueue *pwq;
4456 rcu_read_lock_sched();
4458 if (cpu == WORK_CPU_UNBOUND)
4459 cpu = smp_processor_id();
4461 if (!(wq->flags & WQ_UNBOUND))
4462 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4464 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4466 ret = !list_empty(&pwq->delayed_works);
4467 rcu_read_unlock_sched();
4471 EXPORT_SYMBOL_GPL(workqueue_congested);
4474 * work_busy - test whether a work is currently pending or running
4475 * @work: the work to be tested
4477 * Test whether @work is currently pending or running. There is no
4478 * synchronization around this function and the test result is
4479 * unreliable and only useful as advisory hints or for debugging.
4482 * OR'd bitmask of WORK_BUSY_* bits.
4484 unsigned int work_busy(struct work_struct *work)
4486 struct worker_pool *pool;
4487 unsigned long flags;
4488 unsigned int ret = 0;
4490 if (work_pending(work))
4491 ret |= WORK_BUSY_PENDING;
4493 local_irq_save(flags);
4494 pool = get_work_pool(work);
4496 spin_lock(&pool->lock);
4497 if (find_worker_executing_work(pool, work))
4498 ret |= WORK_BUSY_RUNNING;
4499 spin_unlock(&pool->lock);
4501 local_irq_restore(flags);
4505 EXPORT_SYMBOL_GPL(work_busy);
4508 * set_worker_desc - set description for the current work item
4509 * @fmt: printf-style format string
4510 * @...: arguments for the format string
4512 * This function can be called by a running work function to describe what
4513 * the work item is about. If the worker task gets dumped, this
4514 * information will be printed out together to help debugging. The
4515 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4517 void set_worker_desc(const char *fmt, ...)
4519 struct worker *worker = current_wq_worker();
4523 va_start(args, fmt);
4524 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4528 EXPORT_SYMBOL_GPL(set_worker_desc);
4531 * print_worker_info - print out worker information and description
4532 * @log_lvl: the log level to use when printing
4533 * @task: target task
4535 * If @task is a worker and currently executing a work item, print out the
4536 * name of the workqueue being serviced and worker description set with
4537 * set_worker_desc() by the currently executing work item.
4539 * This function can be safely called on any task as long as the
4540 * task_struct itself is accessible. While safe, this function isn't
4541 * synchronized and may print out mixups or garbages of limited length.
4543 void print_worker_info(const char *log_lvl, struct task_struct *task)
4545 work_func_t *fn = NULL;
4546 char name[WQ_NAME_LEN] = { };
4547 char desc[WORKER_DESC_LEN] = { };
4548 struct pool_workqueue *pwq = NULL;
4549 struct workqueue_struct *wq = NULL;
4550 struct worker *worker;
4552 if (!(task->flags & PF_WQ_WORKER))
4556 * This function is called without any synchronization and @task
4557 * could be in any state. Be careful with dereferences.
4559 worker = kthread_probe_data(task);
4562 * Carefully copy the associated workqueue's workfn, name and desc.
4563 * Keep the original last '\0' in case the original is garbage.
4565 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4566 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4567 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4568 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4569 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4571 if (fn || name[0] || desc[0]) {
4572 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4573 if (strcmp(name, desc))
4574 pr_cont(" (%s)", desc);
4579 static void pr_cont_pool_info(struct worker_pool *pool)
4581 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4582 if (pool->node != NUMA_NO_NODE)
4583 pr_cont(" node=%d", pool->node);
4584 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4587 static void pr_cont_work(bool comma, struct work_struct *work)
4589 if (work->func == wq_barrier_func) {
4590 struct wq_barrier *barr;
4592 barr = container_of(work, struct wq_barrier, work);
4594 pr_cont("%s BAR(%d)", comma ? "," : "",
4595 task_pid_nr(barr->task));
4597 pr_cont("%s %pf", comma ? "," : "", work->func);
4601 static void show_pwq(struct pool_workqueue *pwq)
4603 struct worker_pool *pool = pwq->pool;
4604 struct work_struct *work;
4605 struct worker *worker;
4606 bool has_in_flight = false, has_pending = false;
4609 pr_info(" pwq %d:", pool->id);
4610 pr_cont_pool_info(pool);
4612 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4613 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4615 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4616 if (worker->current_pwq == pwq) {
4617 has_in_flight = true;
4621 if (has_in_flight) {
4624 pr_info(" in-flight:");
4625 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4626 if (worker->current_pwq != pwq)
4629 pr_cont("%s %d%s:%pf", comma ? "," : "",
4630 task_pid_nr(worker->task),
4631 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4632 worker->current_func);
4633 list_for_each_entry(work, &worker->scheduled, entry)
4634 pr_cont_work(false, work);
4640 list_for_each_entry(work, &pool->worklist, entry) {
4641 if (get_work_pwq(work) == pwq) {
4649 pr_info(" pending:");
4650 list_for_each_entry(work, &pool->worklist, entry) {
4651 if (get_work_pwq(work) != pwq)
4654 pr_cont_work(comma, work);
4655 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4660 if (!list_empty(&pwq->delayed_works)) {
4663 pr_info(" delayed:");
4664 list_for_each_entry(work, &pwq->delayed_works, entry) {
4665 pr_cont_work(comma, work);
4666 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4673 * show_workqueue_state - dump workqueue state
4675 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4676 * all busy workqueues and pools.
4678 void show_workqueue_state(void)
4680 struct workqueue_struct *wq;
4681 struct worker_pool *pool;
4682 unsigned long flags;
4685 rcu_read_lock_sched();
4687 pr_info("Showing busy workqueues and worker pools:\n");
4689 list_for_each_entry_rcu(wq, &workqueues, list) {
4690 struct pool_workqueue *pwq;
4693 for_each_pwq(pwq, wq) {
4694 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4702 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4704 for_each_pwq(pwq, wq) {
4705 spin_lock_irqsave(&pwq->pool->lock, flags);
4706 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4708 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4710 * We could be printing a lot from atomic context, e.g.
4711 * sysrq-t -> show_workqueue_state(). Avoid triggering
4714 touch_nmi_watchdog();
4718 for_each_pool(pool, pi) {
4719 struct worker *worker;
4722 spin_lock_irqsave(&pool->lock, flags);
4723 if (pool->nr_workers == pool->nr_idle)
4726 pr_info("pool %d:", pool->id);
4727 pr_cont_pool_info(pool);
4728 pr_cont(" hung=%us workers=%d",
4729 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4732 pr_cont(" manager: %d",
4733 task_pid_nr(pool->manager->task));
4734 list_for_each_entry(worker, &pool->idle_list, entry) {
4735 pr_cont(" %s%d", first ? "idle: " : "",
4736 task_pid_nr(worker->task));
4741 spin_unlock_irqrestore(&pool->lock, flags);
4743 * We could be printing a lot from atomic context, e.g.
4744 * sysrq-t -> show_workqueue_state(). Avoid triggering
4747 touch_nmi_watchdog();
4750 rcu_read_unlock_sched();
4753 /* used to show worker information through /proc/PID/{comm,stat,status} */
4754 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4758 /* always show the actual comm */
4759 off = strscpy(buf, task->comm, size);
4763 /* stabilize PF_WQ_WORKER and worker pool association */
4764 mutex_lock(&wq_pool_attach_mutex);
4766 if (task->flags & PF_WQ_WORKER) {
4767 struct worker *worker = kthread_data(task);
4768 struct worker_pool *pool = worker->pool;
4771 spin_lock_irq(&pool->lock);
4773 * ->desc tracks information (wq name or
4774 * set_worker_desc()) for the latest execution. If
4775 * current, prepend '+', otherwise '-'.
4777 if (worker->desc[0] != '\0') {
4778 if (worker->current_work)
4779 scnprintf(buf + off, size - off, "+%s",
4782 scnprintf(buf + off, size - off, "-%s",
4785 spin_unlock_irq(&pool->lock);
4789 mutex_unlock(&wq_pool_attach_mutex);
4797 * There are two challenges in supporting CPU hotplug. Firstly, there
4798 * are a lot of assumptions on strong associations among work, pwq and
4799 * pool which make migrating pending and scheduled works very
4800 * difficult to implement without impacting hot paths. Secondly,
4801 * worker pools serve mix of short, long and very long running works making
4802 * blocked draining impractical.
4804 * This is solved by allowing the pools to be disassociated from the CPU
4805 * running as an unbound one and allowing it to be reattached later if the
4806 * cpu comes back online.
4809 static void unbind_workers(int cpu)
4811 struct worker_pool *pool;
4812 struct worker *worker;
4814 for_each_cpu_worker_pool(pool, cpu) {
4815 mutex_lock(&wq_pool_attach_mutex);
4816 spin_lock_irq(&pool->lock);
4819 * We've blocked all attach/detach operations. Make all workers
4820 * unbound and set DISASSOCIATED. Before this, all workers
4821 * except for the ones which are still executing works from
4822 * before the last CPU down must be on the cpu. After
4823 * this, they may become diasporas.
4825 for_each_pool_worker(worker, pool)
4826 worker->flags |= WORKER_UNBOUND;
4828 pool->flags |= POOL_DISASSOCIATED;
4830 spin_unlock_irq(&pool->lock);
4831 mutex_unlock(&wq_pool_attach_mutex);
4834 * Call schedule() so that we cross rq->lock and thus can
4835 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4836 * This is necessary as scheduler callbacks may be invoked
4842 * Sched callbacks are disabled now. Zap nr_running.
4843 * After this, nr_running stays zero and need_more_worker()
4844 * and keep_working() are always true as long as the
4845 * worklist is not empty. This pool now behaves as an
4846 * unbound (in terms of concurrency management) pool which
4847 * are served by workers tied to the pool.
4849 atomic_set(&pool->nr_running, 0);
4852 * With concurrency management just turned off, a busy
4853 * worker blocking could lead to lengthy stalls. Kick off
4854 * unbound chain execution of currently pending work items.
4856 spin_lock_irq(&pool->lock);
4857 wake_up_worker(pool);
4858 spin_unlock_irq(&pool->lock);
4863 * rebind_workers - rebind all workers of a pool to the associated CPU
4864 * @pool: pool of interest
4866 * @pool->cpu is coming online. Rebind all workers to the CPU.
4868 static void rebind_workers(struct worker_pool *pool)
4870 struct worker *worker;
4872 lockdep_assert_held(&wq_pool_attach_mutex);
4875 * Restore CPU affinity of all workers. As all idle workers should
4876 * be on the run-queue of the associated CPU before any local
4877 * wake-ups for concurrency management happen, restore CPU affinity
4878 * of all workers first and then clear UNBOUND. As we're called
4879 * from CPU_ONLINE, the following shouldn't fail.
4881 for_each_pool_worker(worker, pool)
4882 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4883 pool->attrs->cpumask) < 0);
4885 spin_lock_irq(&pool->lock);
4887 pool->flags &= ~POOL_DISASSOCIATED;
4889 for_each_pool_worker(worker, pool) {
4890 unsigned int worker_flags = worker->flags;
4893 * A bound idle worker should actually be on the runqueue
4894 * of the associated CPU for local wake-ups targeting it to
4895 * work. Kick all idle workers so that they migrate to the
4896 * associated CPU. Doing this in the same loop as
4897 * replacing UNBOUND with REBOUND is safe as no worker will
4898 * be bound before @pool->lock is released.
4900 if (worker_flags & WORKER_IDLE)
4901 wake_up_process(worker->task);
4904 * We want to clear UNBOUND but can't directly call
4905 * worker_clr_flags() or adjust nr_running. Atomically
4906 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4907 * @worker will clear REBOUND using worker_clr_flags() when
4908 * it initiates the next execution cycle thus restoring
4909 * concurrency management. Note that when or whether
4910 * @worker clears REBOUND doesn't affect correctness.
4912 * WRITE_ONCE() is necessary because @worker->flags may be
4913 * tested without holding any lock in
4914 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4915 * fail incorrectly leading to premature concurrency
4916 * management operations.
4918 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4919 worker_flags |= WORKER_REBOUND;
4920 worker_flags &= ~WORKER_UNBOUND;
4921 WRITE_ONCE(worker->flags, worker_flags);
4924 spin_unlock_irq(&pool->lock);
4928 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4929 * @pool: unbound pool of interest
4930 * @cpu: the CPU which is coming up
4932 * An unbound pool may end up with a cpumask which doesn't have any online
4933 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4934 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4935 * online CPU before, cpus_allowed of all its workers should be restored.
4937 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4939 static cpumask_t cpumask;
4940 struct worker *worker;
4942 lockdep_assert_held(&wq_pool_attach_mutex);
4944 /* is @cpu allowed for @pool? */
4945 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4948 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4950 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4951 for_each_pool_worker(worker, pool)
4952 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4955 int workqueue_prepare_cpu(unsigned int cpu)
4957 struct worker_pool *pool;
4959 for_each_cpu_worker_pool(pool, cpu) {
4960 if (pool->nr_workers)
4962 if (!create_worker(pool))
4968 int workqueue_online_cpu(unsigned int cpu)
4970 struct worker_pool *pool;
4971 struct workqueue_struct *wq;
4974 mutex_lock(&wq_pool_mutex);
4976 for_each_pool(pool, pi) {
4977 mutex_lock(&wq_pool_attach_mutex);
4979 if (pool->cpu == cpu)
4980 rebind_workers(pool);
4981 else if (pool->cpu < 0)
4982 restore_unbound_workers_cpumask(pool, cpu);
4984 mutex_unlock(&wq_pool_attach_mutex);
4987 /* update NUMA affinity of unbound workqueues */
4988 list_for_each_entry(wq, &workqueues, list)
4989 wq_update_unbound_numa(wq, cpu, true);
4991 mutex_unlock(&wq_pool_mutex);
4995 int workqueue_offline_cpu(unsigned int cpu)
4997 struct workqueue_struct *wq;
4999 /* unbinding per-cpu workers should happen on the local CPU */
5000 if (WARN_ON(cpu != smp_processor_id()))
5003 unbind_workers(cpu);
5005 /* update NUMA affinity of unbound workqueues */
5006 mutex_lock(&wq_pool_mutex);
5007 list_for_each_entry(wq, &workqueues, list)
5008 wq_update_unbound_numa(wq, cpu, false);
5009 mutex_unlock(&wq_pool_mutex);
5014 struct work_for_cpu {
5015 struct work_struct work;
5021 static void work_for_cpu_fn(struct work_struct *work)
5023 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5025 wfc->ret = wfc->fn(wfc->arg);
5029 * work_on_cpu - run a function in thread context on a particular cpu
5030 * @cpu: the cpu to run on
5031 * @fn: the function to run
5032 * @arg: the function arg
5034 * It is up to the caller to ensure that the cpu doesn't go offline.
5035 * The caller must not hold any locks which would prevent @fn from completing.
5037 * Return: The value @fn returns.
5039 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5041 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5043 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5044 schedule_work_on(cpu, &wfc.work);
5045 flush_work(&wfc.work);
5046 destroy_work_on_stack(&wfc.work);
5049 EXPORT_SYMBOL_GPL(work_on_cpu);
5052 * work_on_cpu_safe - run a function in thread context on a particular cpu
5053 * @cpu: the cpu to run on
5054 * @fn: the function to run
5055 * @arg: the function argument
5057 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5058 * any locks which would prevent @fn from completing.
5060 * Return: The value @fn returns.
5062 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5067 if (cpu_online(cpu))
5068 ret = work_on_cpu(cpu, fn, arg);
5072 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5073 #endif /* CONFIG_SMP */
5075 #ifdef CONFIG_FREEZER
5078 * freeze_workqueues_begin - begin freezing workqueues
5080 * Start freezing workqueues. After this function returns, all freezable
5081 * workqueues will queue new works to their delayed_works list instead of
5085 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5087 void freeze_workqueues_begin(void)
5089 struct workqueue_struct *wq;
5090 struct pool_workqueue *pwq;
5092 mutex_lock(&wq_pool_mutex);
5094 WARN_ON_ONCE(workqueue_freezing);
5095 workqueue_freezing = true;
5097 list_for_each_entry(wq, &workqueues, list) {
5098 mutex_lock(&wq->mutex);
5099 for_each_pwq(pwq, wq)
5100 pwq_adjust_max_active(pwq);
5101 mutex_unlock(&wq->mutex);
5104 mutex_unlock(&wq_pool_mutex);
5108 * freeze_workqueues_busy - are freezable workqueues still busy?
5110 * Check whether freezing is complete. This function must be called
5111 * between freeze_workqueues_begin() and thaw_workqueues().
5114 * Grabs and releases wq_pool_mutex.
5117 * %true if some freezable workqueues are still busy. %false if freezing
5120 bool freeze_workqueues_busy(void)
5123 struct workqueue_struct *wq;
5124 struct pool_workqueue *pwq;
5126 mutex_lock(&wq_pool_mutex);
5128 WARN_ON_ONCE(!workqueue_freezing);
5130 list_for_each_entry(wq, &workqueues, list) {
5131 if (!(wq->flags & WQ_FREEZABLE))
5134 * nr_active is monotonically decreasing. It's safe
5135 * to peek without lock.
5137 rcu_read_lock_sched();
5138 for_each_pwq(pwq, wq) {
5139 WARN_ON_ONCE(pwq->nr_active < 0);
5140 if (pwq->nr_active) {
5142 rcu_read_unlock_sched();
5146 rcu_read_unlock_sched();
5149 mutex_unlock(&wq_pool_mutex);
5154 * thaw_workqueues - thaw workqueues
5156 * Thaw workqueues. Normal queueing is restored and all collected
5157 * frozen works are transferred to their respective pool worklists.
5160 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5162 void thaw_workqueues(void)
5164 struct workqueue_struct *wq;
5165 struct pool_workqueue *pwq;
5167 mutex_lock(&wq_pool_mutex);
5169 if (!workqueue_freezing)
5172 workqueue_freezing = false;
5174 /* restore max_active and repopulate worklist */
5175 list_for_each_entry(wq, &workqueues, list) {
5176 mutex_lock(&wq->mutex);
5177 for_each_pwq(pwq, wq)
5178 pwq_adjust_max_active(pwq);
5179 mutex_unlock(&wq->mutex);
5183 mutex_unlock(&wq_pool_mutex);
5185 #endif /* CONFIG_FREEZER */
5187 static int workqueue_apply_unbound_cpumask(void)
5191 struct workqueue_struct *wq;
5192 struct apply_wqattrs_ctx *ctx, *n;
5194 lockdep_assert_held(&wq_pool_mutex);
5196 list_for_each_entry(wq, &workqueues, list) {
5197 if (!(wq->flags & WQ_UNBOUND))
5199 /* creating multiple pwqs breaks ordering guarantee */
5200 if (wq->flags & __WQ_ORDERED)
5203 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5209 list_add_tail(&ctx->list, &ctxs);
5212 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5214 apply_wqattrs_commit(ctx);
5215 apply_wqattrs_cleanup(ctx);
5222 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5223 * @cpumask: the cpumask to set
5225 * The low-level workqueues cpumask is a global cpumask that limits
5226 * the affinity of all unbound workqueues. This function check the @cpumask
5227 * and apply it to all unbound workqueues and updates all pwqs of them.
5229 * Retun: 0 - Success
5230 * -EINVAL - Invalid @cpumask
5231 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5233 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5236 cpumask_var_t saved_cpumask;
5238 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5242 * Not excluding isolated cpus on purpose.
5243 * If the user wishes to include them, we allow that.
5245 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5246 if (!cpumask_empty(cpumask)) {
5247 apply_wqattrs_lock();
5249 /* save the old wq_unbound_cpumask. */
5250 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5252 /* update wq_unbound_cpumask at first and apply it to wqs. */
5253 cpumask_copy(wq_unbound_cpumask, cpumask);
5254 ret = workqueue_apply_unbound_cpumask();
5256 /* restore the wq_unbound_cpumask when failed. */
5258 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5260 apply_wqattrs_unlock();
5263 free_cpumask_var(saved_cpumask);
5269 * Workqueues with WQ_SYSFS flag set is visible to userland via
5270 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5271 * following attributes.
5273 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5274 * max_active RW int : maximum number of in-flight work items
5276 * Unbound workqueues have the following extra attributes.
5278 * pool_ids RO int : the associated pool IDs for each node
5279 * nice RW int : nice value of the workers
5280 * cpumask RW mask : bitmask of allowed CPUs for the workers
5281 * numa RW bool : whether enable NUMA affinity
5284 struct workqueue_struct *wq;
5288 static struct workqueue_struct *dev_to_wq(struct device *dev)
5290 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5295 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5298 struct workqueue_struct *wq = dev_to_wq(dev);
5300 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5302 static DEVICE_ATTR_RO(per_cpu);
5304 static ssize_t max_active_show(struct device *dev,
5305 struct device_attribute *attr, char *buf)
5307 struct workqueue_struct *wq = dev_to_wq(dev);
5309 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5312 static ssize_t max_active_store(struct device *dev,
5313 struct device_attribute *attr, const char *buf,
5316 struct workqueue_struct *wq = dev_to_wq(dev);
5319 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5322 workqueue_set_max_active(wq, val);
5325 static DEVICE_ATTR_RW(max_active);
5327 static struct attribute *wq_sysfs_attrs[] = {
5328 &dev_attr_per_cpu.attr,
5329 &dev_attr_max_active.attr,
5332 ATTRIBUTE_GROUPS(wq_sysfs);
5334 static ssize_t wq_pool_ids_show(struct device *dev,
5335 struct device_attribute *attr, char *buf)
5337 struct workqueue_struct *wq = dev_to_wq(dev);
5338 const char *delim = "";
5339 int node, written = 0;
5341 rcu_read_lock_sched();
5342 for_each_node(node) {
5343 written += scnprintf(buf + written, PAGE_SIZE - written,
5344 "%s%d:%d", delim, node,
5345 unbound_pwq_by_node(wq, node)->pool->id);
5348 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5349 rcu_read_unlock_sched();
5354 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5357 struct workqueue_struct *wq = dev_to_wq(dev);
5360 mutex_lock(&wq->mutex);
5361 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5362 mutex_unlock(&wq->mutex);
5367 /* prepare workqueue_attrs for sysfs store operations */
5368 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5370 struct workqueue_attrs *attrs;
5372 lockdep_assert_held(&wq_pool_mutex);
5374 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5378 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5382 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5383 const char *buf, size_t count)
5385 struct workqueue_struct *wq = dev_to_wq(dev);
5386 struct workqueue_attrs *attrs;
5389 apply_wqattrs_lock();
5391 attrs = wq_sysfs_prep_attrs(wq);
5395 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5396 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5397 ret = apply_workqueue_attrs_locked(wq, attrs);
5402 apply_wqattrs_unlock();
5403 free_workqueue_attrs(attrs);
5404 return ret ?: count;
5407 static ssize_t wq_cpumask_show(struct device *dev,
5408 struct device_attribute *attr, char *buf)
5410 struct workqueue_struct *wq = dev_to_wq(dev);
5413 mutex_lock(&wq->mutex);
5414 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5415 cpumask_pr_args(wq->unbound_attrs->cpumask));
5416 mutex_unlock(&wq->mutex);
5420 static ssize_t wq_cpumask_store(struct device *dev,
5421 struct device_attribute *attr,
5422 const char *buf, size_t count)
5424 struct workqueue_struct *wq = dev_to_wq(dev);
5425 struct workqueue_attrs *attrs;
5428 apply_wqattrs_lock();
5430 attrs = wq_sysfs_prep_attrs(wq);
5434 ret = cpumask_parse(buf, attrs->cpumask);
5436 ret = apply_workqueue_attrs_locked(wq, attrs);
5439 apply_wqattrs_unlock();
5440 free_workqueue_attrs(attrs);
5441 return ret ?: count;
5444 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5447 struct workqueue_struct *wq = dev_to_wq(dev);
5450 mutex_lock(&wq->mutex);
5451 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5452 !wq->unbound_attrs->no_numa);
5453 mutex_unlock(&wq->mutex);
5458 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5459 const char *buf, size_t count)
5461 struct workqueue_struct *wq = dev_to_wq(dev);
5462 struct workqueue_attrs *attrs;
5463 int v, ret = -ENOMEM;
5465 apply_wqattrs_lock();
5467 attrs = wq_sysfs_prep_attrs(wq);
5472 if (sscanf(buf, "%d", &v) == 1) {
5473 attrs->no_numa = !v;
5474 ret = apply_workqueue_attrs_locked(wq, attrs);
5478 apply_wqattrs_unlock();
5479 free_workqueue_attrs(attrs);
5480 return ret ?: count;
5483 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5484 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5485 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5486 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5487 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5491 static struct bus_type wq_subsys = {
5492 .name = "workqueue",
5493 .dev_groups = wq_sysfs_groups,
5496 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5497 struct device_attribute *attr, char *buf)
5501 mutex_lock(&wq_pool_mutex);
5502 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5503 cpumask_pr_args(wq_unbound_cpumask));
5504 mutex_unlock(&wq_pool_mutex);
5509 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5510 struct device_attribute *attr, const char *buf, size_t count)
5512 cpumask_var_t cpumask;
5515 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5518 ret = cpumask_parse(buf, cpumask);
5520 ret = workqueue_set_unbound_cpumask(cpumask);
5522 free_cpumask_var(cpumask);
5523 return ret ? ret : count;
5526 static struct device_attribute wq_sysfs_cpumask_attr =
5527 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5528 wq_unbound_cpumask_store);
5530 static int __init wq_sysfs_init(void)
5534 err = subsys_virtual_register(&wq_subsys, NULL);
5538 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5540 core_initcall(wq_sysfs_init);
5542 static void wq_device_release(struct device *dev)
5544 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5550 * workqueue_sysfs_register - make a workqueue visible in sysfs
5551 * @wq: the workqueue to register
5553 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5554 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5555 * which is the preferred method.
5557 * Workqueue user should use this function directly iff it wants to apply
5558 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5559 * apply_workqueue_attrs() may race against userland updating the
5562 * Return: 0 on success, -errno on failure.
5564 int workqueue_sysfs_register(struct workqueue_struct *wq)
5566 struct wq_device *wq_dev;
5570 * Adjusting max_active or creating new pwqs by applying
5571 * attributes breaks ordering guarantee. Disallow exposing ordered
5574 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5577 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5582 wq_dev->dev.bus = &wq_subsys;
5583 wq_dev->dev.release = wq_device_release;
5584 dev_set_name(&wq_dev->dev, "%s", wq->name);
5587 * unbound_attrs are created separately. Suppress uevent until
5588 * everything is ready.
5590 dev_set_uevent_suppress(&wq_dev->dev, true);
5592 ret = device_register(&wq_dev->dev);
5594 put_device(&wq_dev->dev);
5599 if (wq->flags & WQ_UNBOUND) {
5600 struct device_attribute *attr;
5602 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5603 ret = device_create_file(&wq_dev->dev, attr);
5605 device_unregister(&wq_dev->dev);
5612 dev_set_uevent_suppress(&wq_dev->dev, false);
5613 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5618 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5619 * @wq: the workqueue to unregister
5621 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5623 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5625 struct wq_device *wq_dev = wq->wq_dev;
5631 device_unregister(&wq_dev->dev);
5633 #else /* CONFIG_SYSFS */
5634 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5635 #endif /* CONFIG_SYSFS */
5638 * Workqueue watchdog.
5640 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5641 * flush dependency, a concurrency managed work item which stays RUNNING
5642 * indefinitely. Workqueue stalls can be very difficult to debug as the
5643 * usual warning mechanisms don't trigger and internal workqueue state is
5646 * Workqueue watchdog monitors all worker pools periodically and dumps
5647 * state if some pools failed to make forward progress for a while where
5648 * forward progress is defined as the first item on ->worklist changing.
5650 * This mechanism is controlled through the kernel parameter
5651 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5652 * corresponding sysfs parameter file.
5654 #ifdef CONFIG_WQ_WATCHDOG
5656 static unsigned long wq_watchdog_thresh = 30;
5657 static struct timer_list wq_watchdog_timer;
5659 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5660 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5662 static void wq_watchdog_reset_touched(void)
5666 wq_watchdog_touched = jiffies;
5667 for_each_possible_cpu(cpu)
5668 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5671 static void wq_watchdog_timer_fn(struct timer_list *unused)
5673 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5674 bool lockup_detected = false;
5675 struct worker_pool *pool;
5683 for_each_pool(pool, pi) {
5684 unsigned long pool_ts, touched, ts;
5686 if (list_empty(&pool->worklist))
5689 /* get the latest of pool and touched timestamps */
5690 pool_ts = READ_ONCE(pool->watchdog_ts);
5691 touched = READ_ONCE(wq_watchdog_touched);
5693 if (time_after(pool_ts, touched))
5698 if (pool->cpu >= 0) {
5699 unsigned long cpu_touched =
5700 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5702 if (time_after(cpu_touched, ts))
5707 if (time_after(jiffies, ts + thresh)) {
5708 lockup_detected = true;
5709 pr_emerg("BUG: workqueue lockup - pool");
5710 pr_cont_pool_info(pool);
5711 pr_cont(" stuck for %us!\n",
5712 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5718 if (lockup_detected)
5719 show_workqueue_state();
5721 wq_watchdog_reset_touched();
5722 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5725 notrace void wq_watchdog_touch(int cpu)
5728 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5730 wq_watchdog_touched = jiffies;
5733 static void wq_watchdog_set_thresh(unsigned long thresh)
5735 wq_watchdog_thresh = 0;
5736 del_timer_sync(&wq_watchdog_timer);
5739 wq_watchdog_thresh = thresh;
5740 wq_watchdog_reset_touched();
5741 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5745 static int wq_watchdog_param_set_thresh(const char *val,
5746 const struct kernel_param *kp)
5748 unsigned long thresh;
5751 ret = kstrtoul(val, 0, &thresh);
5756 wq_watchdog_set_thresh(thresh);
5758 wq_watchdog_thresh = thresh;
5763 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5764 .set = wq_watchdog_param_set_thresh,
5765 .get = param_get_ulong,
5768 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5771 static void wq_watchdog_init(void)
5773 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5774 wq_watchdog_set_thresh(wq_watchdog_thresh);
5777 #else /* CONFIG_WQ_WATCHDOG */
5779 static inline void wq_watchdog_init(void) { }
5781 #endif /* CONFIG_WQ_WATCHDOG */
5783 static void __init wq_numa_init(void)
5788 if (num_possible_nodes() <= 1)
5791 if (wq_disable_numa) {
5792 pr_info("workqueue: NUMA affinity support disabled\n");
5796 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5797 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5800 * We want masks of possible CPUs of each node which isn't readily
5801 * available. Build one from cpu_to_node() which should have been
5802 * fully initialized by now.
5804 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5808 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5809 node_online(node) ? node : NUMA_NO_NODE));
5811 for_each_possible_cpu(cpu) {
5812 node = cpu_to_node(cpu);
5813 if (WARN_ON(node == NUMA_NO_NODE)) {
5814 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5815 /* happens iff arch is bonkers, let's just proceed */
5818 cpumask_set_cpu(cpu, tbl[node]);
5821 wq_numa_possible_cpumask = tbl;
5822 wq_numa_enabled = true;
5826 * workqueue_init_early - early init for workqueue subsystem
5828 * This is the first half of two-staged workqueue subsystem initialization
5829 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5830 * idr are up. It sets up all the data structures and system workqueues
5831 * and allows early boot code to create workqueues and queue/cancel work
5832 * items. Actual work item execution starts only after kthreads can be
5833 * created and scheduled right before early initcalls.
5835 int __init workqueue_init_early(void)
5837 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5838 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5841 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5843 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5844 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5846 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5848 /* initialize CPU pools */
5849 for_each_possible_cpu(cpu) {
5850 struct worker_pool *pool;
5853 for_each_cpu_worker_pool(pool, cpu) {
5854 BUG_ON(init_worker_pool(pool));
5856 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5857 pool->attrs->nice = std_nice[i++];
5858 pool->node = cpu_to_node(cpu);
5861 mutex_lock(&wq_pool_mutex);
5862 BUG_ON(worker_pool_assign_id(pool));
5863 mutex_unlock(&wq_pool_mutex);
5867 /* create default unbound and ordered wq attrs */
5868 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5869 struct workqueue_attrs *attrs;
5871 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5872 attrs->nice = std_nice[i];
5873 unbound_std_wq_attrs[i] = attrs;
5876 * An ordered wq should have only one pwq as ordering is
5877 * guaranteed by max_active which is enforced by pwqs.
5878 * Turn off NUMA so that dfl_pwq is used for all nodes.
5880 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5881 attrs->nice = std_nice[i];
5882 attrs->no_numa = true;
5883 ordered_wq_attrs[i] = attrs;
5886 system_wq = alloc_workqueue("events", 0, 0);
5887 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5888 system_long_wq = alloc_workqueue("events_long", 0, 0);
5889 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5890 WQ_UNBOUND_MAX_ACTIVE);
5891 system_freezable_wq = alloc_workqueue("events_freezable",
5893 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5894 WQ_POWER_EFFICIENT, 0);
5895 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5896 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5898 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5899 !system_unbound_wq || !system_freezable_wq ||
5900 !system_power_efficient_wq ||
5901 !system_freezable_power_efficient_wq);
5907 * workqueue_init - bring workqueue subsystem fully online
5909 * This is the latter half of two-staged workqueue subsystem initialization
5910 * and invoked as soon as kthreads can be created and scheduled.
5911 * Workqueues have been created and work items queued on them, but there
5912 * are no kworkers executing the work items yet. Populate the worker pools
5913 * with the initial workers and enable future kworker creations.
5915 int __init workqueue_init(void)
5917 struct workqueue_struct *wq;
5918 struct worker_pool *pool;
5922 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5923 * CPU to node mapping may not be available that early on some
5924 * archs such as power and arm64. As per-cpu pools created
5925 * previously could be missing node hint and unbound pools NUMA
5926 * affinity, fix them up.
5928 * Also, while iterating workqueues, create rescuers if requested.
5932 mutex_lock(&wq_pool_mutex);
5934 for_each_possible_cpu(cpu) {
5935 for_each_cpu_worker_pool(pool, cpu) {
5936 pool->node = cpu_to_node(cpu);
5940 list_for_each_entry(wq, &workqueues, list) {
5941 wq_update_unbound_numa(wq, smp_processor_id(), true);
5942 WARN(init_rescuer(wq),
5943 "workqueue: failed to create early rescuer for %s",
5947 mutex_unlock(&wq_pool_mutex);
5949 /* create the initial workers */
5950 for_each_online_cpu(cpu) {
5951 for_each_cpu_worker_pool(pool, cpu) {
5952 pool->flags &= ~POOL_DISASSOCIATED;
5953 BUG_ON(!create_worker(pool));
5957 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5958 BUG_ON(!create_worker(pool));