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>
52 #include <linux/kvm_para.h>
54 #include "workqueue_internal.h"
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
69 * Note that DISASSOCIATED should be flipped only while holding
70 * wq_pool_attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * A: wq_pool_attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
136 * sched-RCU for reads.
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
142 * MD: wq_mayday_lock protected.
145 /* struct worker is defined in workqueue_internal.h */
148 spinlock_t lock; /* the pool lock */
149 int cpu; /* I: the associated cpu */
150 int node; /* I: the associated node ID */
151 int id; /* I: pool ID */
152 unsigned int flags; /* X: flags */
154 unsigned long watchdog_ts; /* L: watchdog timestamp */
156 struct list_head worklist; /* L: list of pending works */
158 int nr_workers; /* L: total number of workers */
159 int nr_idle; /* L: currently idle workers */
161 struct list_head idle_list; /* X: list of idle workers */
162 struct timer_list idle_timer; /* L: worker idle timeout */
163 struct timer_list mayday_timer; /* L: SOS timer for workers */
165 /* a workers is either on busy_hash or idle_list, or the manager */
166 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
167 /* L: hash of busy workers */
169 struct worker *manager; /* L: purely informational */
170 struct list_head workers; /* A: attached workers */
171 struct completion *detach_completion; /* all workers detached */
173 struct ida worker_ida; /* worker IDs for task name */
175 struct workqueue_attrs *attrs; /* I: worker attributes */
176 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
177 int refcnt; /* PL: refcnt for unbound pools */
180 * The current concurrency level. As it's likely to be accessed
181 * from other CPUs during try_to_wake_up(), put it in a separate
184 atomic_t nr_running ____cacheline_aligned_in_smp;
187 * Destruction of pool is sched-RCU protected to allow dereferences
188 * from get_work_pool().
191 } ____cacheline_aligned_in_smp;
194 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
195 * of work_struct->data are used for flags and the remaining high bits
196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
197 * number of flag bits.
199 struct pool_workqueue {
200 struct worker_pool *pool; /* I: the associated pool */
201 struct workqueue_struct *wq; /* I: the owning workqueue */
202 int work_color; /* L: current color */
203 int flush_color; /* L: flushing color */
204 int refcnt; /* L: reference count */
205 int nr_in_flight[WORK_NR_COLORS];
206 /* L: nr of in_flight works */
207 int nr_active; /* L: nr of active works */
208 int max_active; /* L: max active works */
209 struct list_head delayed_works; /* L: delayed works */
210 struct list_head pwqs_node; /* WR: node on wq->pwqs */
211 struct list_head mayday_node; /* MD: node on wq->maydays */
214 * Release of unbound pwq is punted to system_wq. See put_pwq()
215 * and pwq_unbound_release_workfn() for details. pool_workqueue
216 * itself is also sched-RCU protected so that the first pwq can be
217 * determined without grabbing wq->mutex.
219 struct work_struct unbound_release_work;
221 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
224 * Structure used to wait for workqueue flush.
227 struct list_head list; /* WQ: list of flushers */
228 int flush_color; /* WQ: flush color waiting for */
229 struct completion done; /* flush completion */
235 * The externally visible workqueue. It relays the issued work items to
236 * the appropriate worker_pool through its pool_workqueues.
238 struct workqueue_struct {
239 struct list_head pwqs; /* WR: all pwqs of this wq */
240 struct list_head list; /* PR: list of all workqueues */
242 struct mutex mutex; /* protects this wq */
243 int work_color; /* WQ: current work color */
244 int flush_color; /* WQ: current flush color */
245 atomic_t nr_pwqs_to_flush; /* flush in progress */
246 struct wq_flusher *first_flusher; /* WQ: first flusher */
247 struct list_head flusher_queue; /* WQ: flush waiters */
248 struct list_head flusher_overflow; /* WQ: flush overflow list */
250 struct list_head maydays; /* MD: pwqs requesting rescue */
251 struct worker *rescuer; /* I: rescue worker */
253 int nr_drainers; /* WQ: drain in progress */
254 int saved_max_active; /* WQ: saved pwq max_active */
256 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
257 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
260 struct wq_device *wq_dev; /* I: for sysfs interface */
262 #ifdef CONFIG_LOCKDEP
263 struct lockdep_map lockdep_map;
265 char name[WQ_NAME_LEN]; /* I: workqueue name */
268 * Destruction of workqueue_struct is sched-RCU protected to allow
269 * walking the workqueues list without grabbing wq_pool_mutex.
270 * This is used to dump all workqueues from sysrq.
274 /* hot fields used during command issue, aligned to cacheline */
275 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
276 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
277 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
280 static struct kmem_cache *pwq_cache;
282 static cpumask_var_t *wq_numa_possible_cpumask;
283 /* possible CPUs of each node */
285 static bool wq_disable_numa;
286 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288 /* see the comment above the definition of WQ_POWER_EFFICIENT */
289 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
290 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292 static bool wq_online; /* can kworkers be created yet? */
294 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
296 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
297 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
299 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
300 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
301 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
302 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
304 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
305 static bool workqueue_freezing; /* PL: have wqs started freezing? */
307 /* PL: allowable cpus for unbound wqs and work items */
308 static cpumask_var_t wq_unbound_cpumask;
310 /* CPU where unbound work was last round robin scheduled from this CPU */
311 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
314 * Local execution of unbound work items is no longer guaranteed. The
315 * following always forces round-robin CPU selection on unbound work items
316 * to uncover usages which depend on it.
318 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
319 static bool wq_debug_force_rr_cpu = true;
321 static bool wq_debug_force_rr_cpu = false;
323 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
325 /* the per-cpu worker pools */
326 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
328 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
330 /* PL: hash of all unbound pools keyed by pool->attrs */
331 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
333 /* I: attributes used when instantiating standard unbound pools on demand */
334 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
336 /* I: attributes used when instantiating ordered pools on demand */
337 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
339 struct workqueue_struct *system_wq __read_mostly;
340 EXPORT_SYMBOL(system_wq);
341 struct workqueue_struct *system_highpri_wq __read_mostly;
342 EXPORT_SYMBOL_GPL(system_highpri_wq);
343 struct workqueue_struct *system_long_wq __read_mostly;
344 EXPORT_SYMBOL_GPL(system_long_wq);
345 struct workqueue_struct *system_unbound_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_unbound_wq);
347 struct workqueue_struct *system_freezable_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_freezable_wq);
349 struct workqueue_struct *system_power_efficient_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
351 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
352 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
354 static int worker_thread(void *__worker);
355 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
357 #define CREATE_TRACE_POINTS
358 #include <trace/events/workqueue.h>
360 #define assert_rcu_or_pool_mutex() \
361 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
362 !lockdep_is_held(&wq_pool_mutex), \
363 "sched RCU or wq_pool_mutex should be held")
365 #define assert_rcu_or_wq_mutex(wq) \
366 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
367 !lockdep_is_held(&wq->mutex), \
368 "sched RCU or wq->mutex should be held")
370 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
371 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
372 !lockdep_is_held(&wq->mutex) && \
373 !lockdep_is_held(&wq_pool_mutex), \
374 "sched RCU, wq->mutex or wq_pool_mutex should be held")
376 #define for_each_cpu_worker_pool(pool, cpu) \
377 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
378 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
382 * for_each_pool - iterate through all worker_pools in the system
383 * @pool: iteration cursor
384 * @pi: integer used for iteration
386 * This must be called either with wq_pool_mutex held or sched RCU read
387 * locked. If the pool needs to be used beyond the locking in effect, the
388 * caller is responsible for guaranteeing that the pool stays online.
390 * The if/else clause exists only for the lockdep assertion and can be
393 #define for_each_pool(pool, pi) \
394 idr_for_each_entry(&worker_pool_idr, pool, pi) \
395 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
399 * for_each_pool_worker - iterate through all workers of a worker_pool
400 * @worker: iteration cursor
401 * @pool: worker_pool to iterate workers of
403 * This must be called with wq_pool_attach_mutex.
405 * The if/else clause exists only for the lockdep assertion and can be
408 #define for_each_pool_worker(worker, pool) \
409 list_for_each_entry((worker), &(pool)->workers, node) \
410 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
414 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
415 * @pwq: iteration cursor
416 * @wq: the target workqueue
418 * This must be called either with wq->mutex held or sched RCU read locked.
419 * If the pwq needs to be used beyond the locking in effect, the caller is
420 * responsible for guaranteeing that the pwq stays online.
422 * The if/else clause exists only for the lockdep assertion and can be
425 #define for_each_pwq(pwq, wq) \
426 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
427 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
430 #ifdef CONFIG_DEBUG_OBJECTS_WORK
432 static struct debug_obj_descr work_debug_descr;
434 static void *work_debug_hint(void *addr)
436 return ((struct work_struct *) addr)->func;
439 static bool work_is_static_object(void *addr)
441 struct work_struct *work = addr;
443 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
447 * fixup_init is called when:
448 * - an active object is initialized
450 static bool work_fixup_init(void *addr, enum debug_obj_state state)
452 struct work_struct *work = addr;
455 case ODEBUG_STATE_ACTIVE:
456 cancel_work_sync(work);
457 debug_object_init(work, &work_debug_descr);
465 * fixup_free is called when:
466 * - an active object is freed
468 static bool work_fixup_free(void *addr, enum debug_obj_state state)
470 struct work_struct *work = addr;
473 case ODEBUG_STATE_ACTIVE:
474 cancel_work_sync(work);
475 debug_object_free(work, &work_debug_descr);
482 static struct debug_obj_descr work_debug_descr = {
483 .name = "work_struct",
484 .debug_hint = work_debug_hint,
485 .is_static_object = work_is_static_object,
486 .fixup_init = work_fixup_init,
487 .fixup_free = work_fixup_free,
490 static inline void debug_work_activate(struct work_struct *work)
492 debug_object_activate(work, &work_debug_descr);
495 static inline void debug_work_deactivate(struct work_struct *work)
497 debug_object_deactivate(work, &work_debug_descr);
500 void __init_work(struct work_struct *work, int onstack)
503 debug_object_init_on_stack(work, &work_debug_descr);
505 debug_object_init(work, &work_debug_descr);
507 EXPORT_SYMBOL_GPL(__init_work);
509 void destroy_work_on_stack(struct work_struct *work)
511 debug_object_free(work, &work_debug_descr);
513 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
515 void destroy_delayed_work_on_stack(struct delayed_work *work)
517 destroy_timer_on_stack(&work->timer);
518 debug_object_free(&work->work, &work_debug_descr);
520 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
523 static inline void debug_work_activate(struct work_struct *work) { }
524 static inline void debug_work_deactivate(struct work_struct *work) { }
528 * worker_pool_assign_id - allocate ID and assing it to @pool
529 * @pool: the pool pointer of interest
531 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
532 * successfully, -errno on failure.
534 static int worker_pool_assign_id(struct worker_pool *pool)
538 lockdep_assert_held(&wq_pool_mutex);
540 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
550 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
551 * @wq: the target workqueue
554 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
556 * If the pwq needs to be used beyond the locking in effect, the caller is
557 * responsible for guaranteeing that the pwq stays online.
559 * Return: The unbound pool_workqueue for @node.
561 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
564 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
567 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
568 * delayed item is pending. The plan is to keep CPU -> NODE
569 * mapping valid and stable across CPU on/offlines. Once that
570 * happens, this workaround can be removed.
572 if (unlikely(node == NUMA_NO_NODE))
575 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
578 static unsigned int work_color_to_flags(int color)
580 return color << WORK_STRUCT_COLOR_SHIFT;
583 static int get_work_color(struct work_struct *work)
585 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
586 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
589 static int work_next_color(int color)
591 return (color + 1) % WORK_NR_COLORS;
595 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
596 * contain the pointer to the queued pwq. Once execution starts, the flag
597 * is cleared and the high bits contain OFFQ flags and pool ID.
599 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
600 * and clear_work_data() can be used to set the pwq, pool or clear
601 * work->data. These functions should only be called while the work is
602 * owned - ie. while the PENDING bit is set.
604 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
605 * corresponding to a work. Pool is available once the work has been
606 * queued anywhere after initialization until it is sync canceled. pwq is
607 * available only while the work item is queued.
609 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
610 * canceled. While being canceled, a work item may have its PENDING set
611 * but stay off timer and worklist for arbitrarily long and nobody should
612 * try to steal the PENDING bit.
614 static inline void set_work_data(struct work_struct *work, unsigned long data,
617 WARN_ON_ONCE(!work_pending(work));
618 atomic_long_set(&work->data, data | flags | work_static(work));
621 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
622 unsigned long extra_flags)
624 set_work_data(work, (unsigned long)pwq,
625 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
628 static void set_work_pool_and_keep_pending(struct work_struct *work,
631 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
632 WORK_STRUCT_PENDING);
635 static void set_work_pool_and_clear_pending(struct work_struct *work,
639 * The following wmb is paired with the implied mb in
640 * test_and_set_bit(PENDING) and ensures all updates to @work made
641 * here are visible to and precede any updates by the next PENDING
645 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
647 * The following mb guarantees that previous clear of a PENDING bit
648 * will not be reordered with any speculative LOADS or STORES from
649 * work->current_func, which is executed afterwards. This possible
650 * reordering can lead to a missed execution on attempt to qeueue
651 * the same @work. E.g. consider this case:
654 * ---------------------------- --------------------------------
656 * 1 STORE event_indicated
657 * 2 queue_work_on() {
658 * 3 test_and_set_bit(PENDING)
659 * 4 } set_..._and_clear_pending() {
660 * 5 set_work_data() # clear bit
662 * 7 work->current_func() {
663 * 8 LOAD event_indicated
666 * Without an explicit full barrier speculative LOAD on line 8 can
667 * be executed before CPU#0 does STORE on line 1. If that happens,
668 * CPU#0 observes the PENDING bit is still set and new execution of
669 * a @work is not queued in a hope, that CPU#1 will eventually
670 * finish the queued @work. Meanwhile CPU#1 does not see
671 * event_indicated is set, because speculative LOAD was executed
672 * before actual STORE.
677 static void clear_work_data(struct work_struct *work)
679 smp_wmb(); /* see set_work_pool_and_clear_pending() */
680 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
683 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
685 unsigned long data = atomic_long_read(&work->data);
687 if (data & WORK_STRUCT_PWQ)
688 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
694 * get_work_pool - return the worker_pool a given work was associated with
695 * @work: the work item of interest
697 * Pools are created and destroyed under wq_pool_mutex, and allows read
698 * access under sched-RCU read lock. As such, this function should be
699 * called under wq_pool_mutex or with preemption disabled.
701 * All fields of the returned pool are accessible as long as the above
702 * mentioned locking is in effect. If the returned pool needs to be used
703 * beyond the critical section, the caller is responsible for ensuring the
704 * returned pool is and stays online.
706 * Return: The worker_pool @work was last associated with. %NULL if none.
708 static struct worker_pool *get_work_pool(struct work_struct *work)
710 unsigned long data = atomic_long_read(&work->data);
713 assert_rcu_or_pool_mutex();
715 if (data & WORK_STRUCT_PWQ)
716 return ((struct pool_workqueue *)
717 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
719 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
720 if (pool_id == WORK_OFFQ_POOL_NONE)
723 return idr_find(&worker_pool_idr, pool_id);
727 * get_work_pool_id - return the worker pool ID a given work is associated with
728 * @work: the work item of interest
730 * Return: The worker_pool ID @work was last associated with.
731 * %WORK_OFFQ_POOL_NONE if none.
733 static int get_work_pool_id(struct work_struct *work)
735 unsigned long data = atomic_long_read(&work->data);
737 if (data & WORK_STRUCT_PWQ)
738 return ((struct pool_workqueue *)
739 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
741 return data >> WORK_OFFQ_POOL_SHIFT;
744 static void mark_work_canceling(struct work_struct *work)
746 unsigned long pool_id = get_work_pool_id(work);
748 pool_id <<= WORK_OFFQ_POOL_SHIFT;
749 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
752 static bool work_is_canceling(struct work_struct *work)
754 unsigned long data = atomic_long_read(&work->data);
756 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
760 * Policy functions. These define the policies on how the global worker
761 * pools are managed. Unless noted otherwise, these functions assume that
762 * they're being called with pool->lock held.
765 static bool __need_more_worker(struct worker_pool *pool)
767 return !atomic_read(&pool->nr_running);
771 * Need to wake up a worker? Called from anything but currently
774 * Note that, because unbound workers never contribute to nr_running, this
775 * function will always return %true for unbound pools as long as the
776 * worklist isn't empty.
778 static bool need_more_worker(struct worker_pool *pool)
780 return !list_empty(&pool->worklist) && __need_more_worker(pool);
783 /* Can I start working? Called from busy but !running workers. */
784 static bool may_start_working(struct worker_pool *pool)
786 return pool->nr_idle;
789 /* Do I need to keep working? Called from currently running workers. */
790 static bool keep_working(struct worker_pool *pool)
792 return !list_empty(&pool->worklist) &&
793 atomic_read(&pool->nr_running) <= 1;
796 /* Do we need a new worker? Called from manager. */
797 static bool need_to_create_worker(struct worker_pool *pool)
799 return need_more_worker(pool) && !may_start_working(pool);
802 /* Do we have too many workers and should some go away? */
803 static bool too_many_workers(struct worker_pool *pool)
805 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
806 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
807 int nr_busy = pool->nr_workers - nr_idle;
809 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
816 /* Return the first idle worker. Safe with preemption disabled */
817 static struct worker *first_idle_worker(struct worker_pool *pool)
819 if (unlikely(list_empty(&pool->idle_list)))
822 return list_first_entry(&pool->idle_list, struct worker, entry);
826 * wake_up_worker - wake up an idle worker
827 * @pool: worker pool to wake worker from
829 * Wake up the first idle worker of @pool.
832 * spin_lock_irq(pool->lock).
834 static void wake_up_worker(struct worker_pool *pool)
836 struct worker *worker = first_idle_worker(pool);
839 wake_up_process(worker->task);
843 * wq_worker_waking_up - a worker is waking up
844 * @task: task waking up
845 * @cpu: CPU @task is waking up to
847 * This function is called during try_to_wake_up() when a worker is
851 * spin_lock_irq(rq->lock)
853 void wq_worker_waking_up(struct task_struct *task, int cpu)
855 struct worker *worker = kthread_data(task);
857 if (!(worker->flags & WORKER_NOT_RUNNING)) {
858 WARN_ON_ONCE(worker->pool->cpu != cpu);
859 atomic_inc(&worker->pool->nr_running);
864 * wq_worker_sleeping - a worker is going to sleep
865 * @task: task going to sleep
867 * This function is called during schedule() when a busy worker is
868 * going to sleep. Worker on the same cpu can be woken up by
869 * returning pointer to its task.
872 * spin_lock_irq(rq->lock)
875 * Worker task on @cpu to wake up, %NULL if none.
877 struct task_struct *wq_worker_sleeping(struct task_struct *task)
879 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
880 struct worker_pool *pool;
883 * Rescuers, which may not have all the fields set up like normal
884 * workers, also reach here, let's not access anything before
885 * checking NOT_RUNNING.
887 if (worker->flags & WORKER_NOT_RUNNING)
892 /* this can only happen on the local cpu */
893 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
897 * The counterpart of the following dec_and_test, implied mb,
898 * worklist not empty test sequence is in insert_work().
899 * Please read comment there.
901 * NOT_RUNNING is clear. This means that we're bound to and
902 * running on the local cpu w/ rq lock held and preemption
903 * disabled, which in turn means that none else could be
904 * manipulating idle_list, so dereferencing idle_list without pool
907 if (atomic_dec_and_test(&pool->nr_running) &&
908 !list_empty(&pool->worklist))
909 to_wakeup = first_idle_worker(pool);
910 return to_wakeup ? to_wakeup->task : NULL;
914 * worker_set_flags - set worker flags and adjust nr_running accordingly
916 * @flags: flags to set
918 * Set @flags in @worker->flags and adjust nr_running accordingly.
921 * spin_lock_irq(pool->lock)
923 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
925 struct worker_pool *pool = worker->pool;
927 WARN_ON_ONCE(worker->task != current);
929 /* If transitioning into NOT_RUNNING, adjust nr_running. */
930 if ((flags & WORKER_NOT_RUNNING) &&
931 !(worker->flags & WORKER_NOT_RUNNING)) {
932 atomic_dec(&pool->nr_running);
935 worker->flags |= flags;
939 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
941 * @flags: flags to clear
943 * Clear @flags in @worker->flags and adjust nr_running accordingly.
946 * spin_lock_irq(pool->lock)
948 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
950 struct worker_pool *pool = worker->pool;
951 unsigned int oflags = worker->flags;
953 WARN_ON_ONCE(worker->task != current);
955 worker->flags &= ~flags;
958 * If transitioning out of NOT_RUNNING, increment nr_running. Note
959 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
960 * of multiple flags, not a single flag.
962 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
963 if (!(worker->flags & WORKER_NOT_RUNNING))
964 atomic_inc(&pool->nr_running);
968 * find_worker_executing_work - find worker which is executing a work
969 * @pool: pool of interest
970 * @work: work to find worker for
972 * Find a worker which is executing @work on @pool by searching
973 * @pool->busy_hash which is keyed by the address of @work. For a worker
974 * to match, its current execution should match the address of @work and
975 * its work function. This is to avoid unwanted dependency between
976 * unrelated work executions through a work item being recycled while still
979 * This is a bit tricky. A work item may be freed once its execution
980 * starts and nothing prevents the freed area from being recycled for
981 * another work item. If the same work item address ends up being reused
982 * before the original execution finishes, workqueue will identify the
983 * recycled work item as currently executing and make it wait until the
984 * current execution finishes, introducing an unwanted dependency.
986 * This function checks the work item address and work function to avoid
987 * false positives. Note that this isn't complete as one may construct a
988 * work function which can introduce dependency onto itself through a
989 * recycled work item. Well, if somebody wants to shoot oneself in the
990 * foot that badly, there's only so much we can do, and if such deadlock
991 * actually occurs, it should be easy to locate the culprit work function.
994 * spin_lock_irq(pool->lock).
997 * Pointer to worker which is executing @work if found, %NULL
1000 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1001 struct work_struct *work)
1003 struct worker *worker;
1005 hash_for_each_possible(pool->busy_hash, worker, hentry,
1006 (unsigned long)work)
1007 if (worker->current_work == work &&
1008 worker->current_func == work->func)
1015 * move_linked_works - move linked works to a list
1016 * @work: start of series of works to be scheduled
1017 * @head: target list to append @work to
1018 * @nextp: out parameter for nested worklist walking
1020 * Schedule linked works starting from @work to @head. Work series to
1021 * be scheduled starts at @work and includes any consecutive work with
1022 * WORK_STRUCT_LINKED set in its predecessor.
1024 * If @nextp is not NULL, it's updated to point to the next work of
1025 * the last scheduled work. This allows move_linked_works() to be
1026 * nested inside outer list_for_each_entry_safe().
1029 * spin_lock_irq(pool->lock).
1031 static void move_linked_works(struct work_struct *work, struct list_head *head,
1032 struct work_struct **nextp)
1034 struct work_struct *n;
1037 * Linked worklist will always end before the end of the list,
1038 * use NULL for list head.
1040 list_for_each_entry_safe_from(work, n, NULL, entry) {
1041 list_move_tail(&work->entry, head);
1042 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1047 * If we're already inside safe list traversal and have moved
1048 * multiple works to the scheduled queue, the next position
1049 * needs to be updated.
1056 * get_pwq - get an extra reference on the specified pool_workqueue
1057 * @pwq: pool_workqueue to get
1059 * Obtain an extra reference on @pwq. The caller should guarantee that
1060 * @pwq has positive refcnt and be holding the matching pool->lock.
1062 static void get_pwq(struct pool_workqueue *pwq)
1064 lockdep_assert_held(&pwq->pool->lock);
1065 WARN_ON_ONCE(pwq->refcnt <= 0);
1070 * put_pwq - put a pool_workqueue reference
1071 * @pwq: pool_workqueue to put
1073 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1074 * destruction. The caller should be holding the matching pool->lock.
1076 static void put_pwq(struct pool_workqueue *pwq)
1078 lockdep_assert_held(&pwq->pool->lock);
1079 if (likely(--pwq->refcnt))
1081 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1084 * @pwq can't be released under pool->lock, bounce to
1085 * pwq_unbound_release_workfn(). This never recurses on the same
1086 * pool->lock as this path is taken only for unbound workqueues and
1087 * the release work item is scheduled on a per-cpu workqueue. To
1088 * avoid lockdep warning, unbound pool->locks are given lockdep
1089 * subclass of 1 in get_unbound_pool().
1091 schedule_work(&pwq->unbound_release_work);
1095 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1096 * @pwq: pool_workqueue to put (can be %NULL)
1098 * put_pwq() with locking. This function also allows %NULL @pwq.
1100 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1104 * As both pwqs and pools are sched-RCU protected, the
1105 * following lock operations are safe.
1107 spin_lock_irq(&pwq->pool->lock);
1109 spin_unlock_irq(&pwq->pool->lock);
1113 static void pwq_activate_delayed_work(struct work_struct *work)
1115 struct pool_workqueue *pwq = get_work_pwq(work);
1117 trace_workqueue_activate_work(work);
1118 if (list_empty(&pwq->pool->worklist))
1119 pwq->pool->watchdog_ts = jiffies;
1120 move_linked_works(work, &pwq->pool->worklist, NULL);
1121 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1125 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1127 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1128 struct work_struct, entry);
1130 pwq_activate_delayed_work(work);
1134 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1135 * @pwq: pwq of interest
1136 * @color: color of work which left the queue
1138 * A work either has completed or is removed from pending queue,
1139 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1142 * spin_lock_irq(pool->lock).
1144 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1146 /* uncolored work items don't participate in flushing or nr_active */
1147 if (color == WORK_NO_COLOR)
1150 pwq->nr_in_flight[color]--;
1153 if (!list_empty(&pwq->delayed_works)) {
1154 /* one down, submit a delayed one */
1155 if (pwq->nr_active < pwq->max_active)
1156 pwq_activate_first_delayed(pwq);
1159 /* is flush in progress and are we at the flushing tip? */
1160 if (likely(pwq->flush_color != color))
1163 /* are there still in-flight works? */
1164 if (pwq->nr_in_flight[color])
1167 /* this pwq is done, clear flush_color */
1168 pwq->flush_color = -1;
1171 * If this was the last pwq, wake up the first flusher. It
1172 * will handle the rest.
1174 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1175 complete(&pwq->wq->first_flusher->done);
1181 * try_to_grab_pending - steal work item from worklist and disable irq
1182 * @work: work item to steal
1183 * @is_dwork: @work is a delayed_work
1184 * @flags: place to store irq state
1186 * Try to grab PENDING bit of @work. This function can handle @work in any
1187 * stable state - idle, on timer or on worklist.
1190 * 1 if @work was pending and we successfully stole PENDING
1191 * 0 if @work was idle and we claimed PENDING
1192 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1193 * -ENOENT if someone else is canceling @work, this state may persist
1194 * for arbitrarily long
1197 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1198 * interrupted while holding PENDING and @work off queue, irq must be
1199 * disabled on entry. This, combined with delayed_work->timer being
1200 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1202 * On successful return, >= 0, irq is disabled and the caller is
1203 * responsible for releasing it using local_irq_restore(*@flags).
1205 * This function is safe to call from any context including IRQ handler.
1207 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1208 unsigned long *flags)
1210 struct worker_pool *pool;
1211 struct pool_workqueue *pwq;
1213 local_irq_save(*flags);
1215 /* try to steal the timer if it exists */
1217 struct delayed_work *dwork = to_delayed_work(work);
1220 * dwork->timer is irqsafe. If del_timer() fails, it's
1221 * guaranteed that the timer is not queued anywhere and not
1222 * running on the local CPU.
1224 if (likely(del_timer(&dwork->timer)))
1228 /* try to claim PENDING the normal way */
1229 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1233 * The queueing is in progress, or it is already queued. Try to
1234 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1236 pool = get_work_pool(work);
1240 spin_lock(&pool->lock);
1242 * work->data is guaranteed to point to pwq only while the work
1243 * item is queued on pwq->wq, and both updating work->data to point
1244 * to pwq on queueing and to pool on dequeueing are done under
1245 * pwq->pool->lock. This in turn guarantees that, if work->data
1246 * points to pwq which is associated with a locked pool, the work
1247 * item is currently queued on that pool.
1249 pwq = get_work_pwq(work);
1250 if (pwq && pwq->pool == pool) {
1251 debug_work_deactivate(work);
1254 * A delayed work item cannot be grabbed directly because
1255 * it might have linked NO_COLOR work items which, if left
1256 * on the delayed_list, will confuse pwq->nr_active
1257 * management later on and cause stall. Make sure the work
1258 * item is activated before grabbing.
1260 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1261 pwq_activate_delayed_work(work);
1263 list_del_init(&work->entry);
1264 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1266 /* work->data points to pwq iff queued, point to pool */
1267 set_work_pool_and_keep_pending(work, pool->id);
1269 spin_unlock(&pool->lock);
1272 spin_unlock(&pool->lock);
1274 local_irq_restore(*flags);
1275 if (work_is_canceling(work))
1282 * insert_work - insert a work into a pool
1283 * @pwq: pwq @work belongs to
1284 * @work: work to insert
1285 * @head: insertion point
1286 * @extra_flags: extra WORK_STRUCT_* flags to set
1288 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1289 * work_struct flags.
1292 * spin_lock_irq(pool->lock).
1294 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1295 struct list_head *head, unsigned int extra_flags)
1297 struct worker_pool *pool = pwq->pool;
1299 /* we own @work, set data and link */
1300 set_work_pwq(work, pwq, extra_flags);
1301 list_add_tail(&work->entry, head);
1305 * Ensure either wq_worker_sleeping() sees the above
1306 * list_add_tail() or we see zero nr_running to avoid workers lying
1307 * around lazily while there are works to be processed.
1311 if (__need_more_worker(pool))
1312 wake_up_worker(pool);
1316 * Test whether @work is being queued from another work executing on the
1319 static bool is_chained_work(struct workqueue_struct *wq)
1321 struct worker *worker;
1323 worker = current_wq_worker();
1325 * Return %true iff I'm a worker execuing a work item on @wq. If
1326 * I'm @worker, it's safe to dereference it without locking.
1328 return worker && worker->current_pwq->wq == wq;
1332 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1333 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1334 * avoid perturbing sensitive tasks.
1336 static int wq_select_unbound_cpu(int cpu)
1338 static bool printed_dbg_warning;
1341 if (likely(!wq_debug_force_rr_cpu)) {
1342 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1344 } else if (!printed_dbg_warning) {
1345 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1346 printed_dbg_warning = true;
1349 if (cpumask_empty(wq_unbound_cpumask))
1352 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1353 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1354 if (unlikely(new_cpu >= nr_cpu_ids)) {
1355 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1356 if (unlikely(new_cpu >= nr_cpu_ids))
1359 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1364 static void __queue_work(int cpu, struct workqueue_struct *wq,
1365 struct work_struct *work)
1367 struct pool_workqueue *pwq;
1368 struct worker_pool *last_pool;
1369 struct list_head *worklist;
1370 unsigned int work_flags;
1371 unsigned int req_cpu = cpu;
1374 * While a work item is PENDING && off queue, a task trying to
1375 * steal the PENDING will busy-loop waiting for it to either get
1376 * queued or lose PENDING. Grabbing PENDING and queueing should
1377 * happen with IRQ disabled.
1379 lockdep_assert_irqs_disabled();
1382 /* if draining, only works from the same workqueue are allowed */
1383 if (unlikely(wq->flags & __WQ_DRAINING) &&
1384 WARN_ON_ONCE(!is_chained_work(wq)))
1387 /* pwq which will be used unless @work is executing elsewhere */
1388 if (wq->flags & WQ_UNBOUND) {
1389 if (req_cpu == WORK_CPU_UNBOUND)
1390 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1391 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1393 if (req_cpu == WORK_CPU_UNBOUND)
1394 cpu = raw_smp_processor_id();
1395 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1399 * If @work was previously on a different pool, it might still be
1400 * running there, in which case the work needs to be queued on that
1401 * pool to guarantee non-reentrancy.
1403 last_pool = get_work_pool(work);
1404 if (last_pool && last_pool != pwq->pool) {
1405 struct worker *worker;
1407 spin_lock(&last_pool->lock);
1409 worker = find_worker_executing_work(last_pool, work);
1411 if (worker && worker->current_pwq->wq == wq) {
1412 pwq = worker->current_pwq;
1414 /* meh... not running there, queue here */
1415 spin_unlock(&last_pool->lock);
1416 spin_lock(&pwq->pool->lock);
1419 spin_lock(&pwq->pool->lock);
1423 * pwq is determined and locked. For unbound pools, we could have
1424 * raced with pwq release and it could already be dead. If its
1425 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1426 * without another pwq replacing it in the numa_pwq_tbl or while
1427 * work items are executing on it, so the retrying is guaranteed to
1428 * make forward-progress.
1430 if (unlikely(!pwq->refcnt)) {
1431 if (wq->flags & WQ_UNBOUND) {
1432 spin_unlock(&pwq->pool->lock);
1437 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1441 /* pwq determined, queue */
1442 trace_workqueue_queue_work(req_cpu, pwq, work);
1444 if (WARN_ON(!list_empty(&work->entry))) {
1445 spin_unlock(&pwq->pool->lock);
1449 pwq->nr_in_flight[pwq->work_color]++;
1450 work_flags = work_color_to_flags(pwq->work_color);
1452 if (likely(pwq->nr_active < pwq->max_active)) {
1453 trace_workqueue_activate_work(work);
1455 worklist = &pwq->pool->worklist;
1456 if (list_empty(worklist))
1457 pwq->pool->watchdog_ts = jiffies;
1459 work_flags |= WORK_STRUCT_DELAYED;
1460 worklist = &pwq->delayed_works;
1463 debug_work_activate(work);
1464 insert_work(pwq, work, worklist, work_flags);
1466 spin_unlock(&pwq->pool->lock);
1470 * queue_work_on - queue work on specific cpu
1471 * @cpu: CPU number to execute work on
1472 * @wq: workqueue to use
1473 * @work: work to queue
1475 * We queue the work to a specific CPU, the caller must ensure it
1478 * Return: %false if @work was already on a queue, %true otherwise.
1480 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1481 struct work_struct *work)
1484 unsigned long flags;
1486 local_irq_save(flags);
1488 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1489 __queue_work(cpu, wq, work);
1493 local_irq_restore(flags);
1496 EXPORT_SYMBOL(queue_work_on);
1498 void delayed_work_timer_fn(struct timer_list *t)
1500 struct delayed_work *dwork = from_timer(dwork, t, timer);
1502 /* should have been called from irqsafe timer with irq already off */
1503 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1505 EXPORT_SYMBOL(delayed_work_timer_fn);
1507 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1508 struct delayed_work *dwork, unsigned long delay)
1510 struct timer_list *timer = &dwork->timer;
1511 struct work_struct *work = &dwork->work;
1514 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1515 WARN_ON_ONCE(timer_pending(timer));
1516 WARN_ON_ONCE(!list_empty(&work->entry));
1519 * If @delay is 0, queue @dwork->work immediately. This is for
1520 * both optimization and correctness. The earliest @timer can
1521 * expire is on the closest next tick and delayed_work users depend
1522 * on that there's no such delay when @delay is 0.
1525 __queue_work(cpu, wq, &dwork->work);
1531 timer->expires = jiffies + delay;
1533 if (unlikely(cpu != WORK_CPU_UNBOUND))
1534 add_timer_on(timer, cpu);
1540 * queue_delayed_work_on - queue work on specific CPU after delay
1541 * @cpu: CPU number to execute work on
1542 * @wq: workqueue to use
1543 * @dwork: work to queue
1544 * @delay: number of jiffies to wait before queueing
1546 * Return: %false if @work was already on a queue, %true otherwise. If
1547 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1550 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1551 struct delayed_work *dwork, unsigned long delay)
1553 struct work_struct *work = &dwork->work;
1555 unsigned long flags;
1557 /* read the comment in __queue_work() */
1558 local_irq_save(flags);
1560 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1561 __queue_delayed_work(cpu, wq, dwork, delay);
1565 local_irq_restore(flags);
1568 EXPORT_SYMBOL(queue_delayed_work_on);
1571 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1572 * @cpu: CPU number to execute work on
1573 * @wq: workqueue to use
1574 * @dwork: work to queue
1575 * @delay: number of jiffies to wait before queueing
1577 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1578 * modify @dwork's timer so that it expires after @delay. If @delay is
1579 * zero, @work is guaranteed to be scheduled immediately regardless of its
1582 * Return: %false if @dwork was idle and queued, %true if @dwork was
1583 * pending and its timer was modified.
1585 * This function is safe to call from any context including IRQ handler.
1586 * See try_to_grab_pending() for details.
1588 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1589 struct delayed_work *dwork, unsigned long delay)
1591 unsigned long flags;
1595 ret = try_to_grab_pending(&dwork->work, true, &flags);
1596 } while (unlikely(ret == -EAGAIN));
1598 if (likely(ret >= 0)) {
1599 __queue_delayed_work(cpu, wq, dwork, delay);
1600 local_irq_restore(flags);
1603 /* -ENOENT from try_to_grab_pending() becomes %true */
1606 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1608 static void rcu_work_rcufn(struct rcu_head *rcu)
1610 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1612 /* read the comment in __queue_work() */
1613 local_irq_disable();
1614 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1619 * queue_rcu_work - queue work after a RCU grace period
1620 * @wq: workqueue to use
1621 * @rwork: work to queue
1623 * Return: %false if @rwork was already pending, %true otherwise. Note
1624 * that a full RCU grace period is guaranteed only after a %true return.
1625 * While @rwork is guarnateed to be executed after a %false return, the
1626 * execution may happen before a full RCU grace period has passed.
1628 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1630 struct work_struct *work = &rwork->work;
1632 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1634 call_rcu(&rwork->rcu, rcu_work_rcufn);
1640 EXPORT_SYMBOL(queue_rcu_work);
1643 * worker_enter_idle - enter idle state
1644 * @worker: worker which is entering idle state
1646 * @worker is entering idle state. Update stats and idle timer if
1650 * spin_lock_irq(pool->lock).
1652 static void worker_enter_idle(struct worker *worker)
1654 struct worker_pool *pool = worker->pool;
1656 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1657 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1658 (worker->hentry.next || worker->hentry.pprev)))
1661 /* can't use worker_set_flags(), also called from create_worker() */
1662 worker->flags |= WORKER_IDLE;
1664 worker->last_active = jiffies;
1666 /* idle_list is LIFO */
1667 list_add(&worker->entry, &pool->idle_list);
1669 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1670 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1673 * Sanity check nr_running. Because unbind_workers() releases
1674 * pool->lock between setting %WORKER_UNBOUND and zapping
1675 * nr_running, the warning may trigger spuriously. Check iff
1676 * unbind is not in progress.
1678 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1679 pool->nr_workers == pool->nr_idle &&
1680 atomic_read(&pool->nr_running));
1684 * worker_leave_idle - leave idle state
1685 * @worker: worker which is leaving idle state
1687 * @worker is leaving idle state. Update stats.
1690 * spin_lock_irq(pool->lock).
1692 static void worker_leave_idle(struct worker *worker)
1694 struct worker_pool *pool = worker->pool;
1696 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1698 worker_clr_flags(worker, WORKER_IDLE);
1700 list_del_init(&worker->entry);
1703 static struct worker *alloc_worker(int node)
1705 struct worker *worker;
1707 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1709 INIT_LIST_HEAD(&worker->entry);
1710 INIT_LIST_HEAD(&worker->scheduled);
1711 INIT_LIST_HEAD(&worker->node);
1712 /* on creation a worker is in !idle && prep state */
1713 worker->flags = WORKER_PREP;
1719 * worker_attach_to_pool() - attach a worker to a pool
1720 * @worker: worker to be attached
1721 * @pool: the target pool
1723 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1724 * cpu-binding of @worker are kept coordinated with the pool across
1727 static void worker_attach_to_pool(struct worker *worker,
1728 struct worker_pool *pool)
1730 mutex_lock(&wq_pool_attach_mutex);
1733 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1734 * stable across this function. See the comments above the flag
1735 * definition for details.
1737 if (pool->flags & POOL_DISASSOCIATED)
1738 worker->flags |= WORKER_UNBOUND;
1740 if (worker->rescue_wq)
1741 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1743 list_add_tail(&worker->node, &pool->workers);
1744 worker->pool = pool;
1746 mutex_unlock(&wq_pool_attach_mutex);
1750 * worker_detach_from_pool() - detach a worker from its pool
1751 * @worker: worker which is attached to its pool
1753 * Undo the attaching which had been done in worker_attach_to_pool(). The
1754 * caller worker shouldn't access to the pool after detached except it has
1755 * other reference to the pool.
1757 static void worker_detach_from_pool(struct worker *worker)
1759 struct worker_pool *pool = worker->pool;
1760 struct completion *detach_completion = NULL;
1762 mutex_lock(&wq_pool_attach_mutex);
1764 list_del(&worker->node);
1765 worker->pool = NULL;
1767 if (list_empty(&pool->workers))
1768 detach_completion = pool->detach_completion;
1769 mutex_unlock(&wq_pool_attach_mutex);
1771 /* clear leftover flags without pool->lock after it is detached */
1772 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1774 if (detach_completion)
1775 complete(detach_completion);
1779 * create_worker - create a new workqueue worker
1780 * @pool: pool the new worker will belong to
1782 * Create and start a new worker which is attached to @pool.
1785 * Might sleep. Does GFP_KERNEL allocations.
1788 * Pointer to the newly created worker.
1790 static struct worker *create_worker(struct worker_pool *pool)
1792 struct worker *worker = NULL;
1796 /* ID is needed to determine kthread name */
1797 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1801 worker = alloc_worker(pool->node);
1808 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1809 pool->attrs->nice < 0 ? "H" : "");
1811 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1813 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1814 "kworker/%s", id_buf);
1815 if (IS_ERR(worker->task))
1818 set_user_nice(worker->task, pool->attrs->nice);
1819 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1821 /* successful, attach the worker to the pool */
1822 worker_attach_to_pool(worker, pool);
1824 /* start the newly created worker */
1825 spin_lock_irq(&pool->lock);
1826 worker->pool->nr_workers++;
1827 worker_enter_idle(worker);
1828 wake_up_process(worker->task);
1829 spin_unlock_irq(&pool->lock);
1835 ida_simple_remove(&pool->worker_ida, id);
1841 * destroy_worker - destroy a workqueue worker
1842 * @worker: worker to be destroyed
1844 * Destroy @worker and adjust @pool stats accordingly. The worker should
1848 * spin_lock_irq(pool->lock).
1850 static void destroy_worker(struct worker *worker)
1852 struct worker_pool *pool = worker->pool;
1854 lockdep_assert_held(&pool->lock);
1856 /* sanity check frenzy */
1857 if (WARN_ON(worker->current_work) ||
1858 WARN_ON(!list_empty(&worker->scheduled)) ||
1859 WARN_ON(!(worker->flags & WORKER_IDLE)))
1865 list_del_init(&worker->entry);
1866 worker->flags |= WORKER_DIE;
1867 wake_up_process(worker->task);
1870 static void idle_worker_timeout(struct timer_list *t)
1872 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1874 spin_lock_irq(&pool->lock);
1876 while (too_many_workers(pool)) {
1877 struct worker *worker;
1878 unsigned long expires;
1880 /* idle_list is kept in LIFO order, check the last one */
1881 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1882 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1884 if (time_before(jiffies, expires)) {
1885 mod_timer(&pool->idle_timer, expires);
1889 destroy_worker(worker);
1892 spin_unlock_irq(&pool->lock);
1895 static void send_mayday(struct work_struct *work)
1897 struct pool_workqueue *pwq = get_work_pwq(work);
1898 struct workqueue_struct *wq = pwq->wq;
1900 lockdep_assert_held(&wq_mayday_lock);
1905 /* mayday mayday mayday */
1906 if (list_empty(&pwq->mayday_node)) {
1908 * If @pwq is for an unbound wq, its base ref may be put at
1909 * any time due to an attribute change. Pin @pwq until the
1910 * rescuer is done with it.
1913 list_add_tail(&pwq->mayday_node, &wq->maydays);
1914 wake_up_process(wq->rescuer->task);
1918 static void pool_mayday_timeout(struct timer_list *t)
1920 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
1921 struct work_struct *work;
1923 spin_lock_irq(&pool->lock);
1924 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1926 if (need_to_create_worker(pool)) {
1928 * We've been trying to create a new worker but
1929 * haven't been successful. We might be hitting an
1930 * allocation deadlock. Send distress signals to
1933 list_for_each_entry(work, &pool->worklist, entry)
1937 spin_unlock(&wq_mayday_lock);
1938 spin_unlock_irq(&pool->lock);
1940 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1944 * maybe_create_worker - create a new worker if necessary
1945 * @pool: pool to create a new worker for
1947 * Create a new worker for @pool if necessary. @pool is guaranteed to
1948 * have at least one idle worker on return from this function. If
1949 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1950 * sent to all rescuers with works scheduled on @pool to resolve
1951 * possible allocation deadlock.
1953 * On return, need_to_create_worker() is guaranteed to be %false and
1954 * may_start_working() %true.
1957 * spin_lock_irq(pool->lock) which may be released and regrabbed
1958 * multiple times. Does GFP_KERNEL allocations. Called only from
1961 static void maybe_create_worker(struct worker_pool *pool)
1962 __releases(&pool->lock)
1963 __acquires(&pool->lock)
1966 spin_unlock_irq(&pool->lock);
1968 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1969 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1972 if (create_worker(pool) || !need_to_create_worker(pool))
1975 schedule_timeout_interruptible(CREATE_COOLDOWN);
1977 if (!need_to_create_worker(pool))
1981 del_timer_sync(&pool->mayday_timer);
1982 spin_lock_irq(&pool->lock);
1984 * This is necessary even after a new worker was just successfully
1985 * created as @pool->lock was dropped and the new worker might have
1986 * already become busy.
1988 if (need_to_create_worker(pool))
1993 * manage_workers - manage worker pool
1996 * Assume the manager role and manage the worker pool @worker belongs
1997 * to. At any given time, there can be only zero or one manager per
1998 * pool. The exclusion is handled automatically by this function.
2000 * The caller can safely start processing works on false return. On
2001 * true return, it's guaranteed that need_to_create_worker() is false
2002 * and may_start_working() is true.
2005 * spin_lock_irq(pool->lock) which may be released and regrabbed
2006 * multiple times. Does GFP_KERNEL allocations.
2009 * %false if the pool doesn't need management and the caller can safely
2010 * start processing works, %true if management function was performed and
2011 * the conditions that the caller verified before calling the function may
2012 * no longer be true.
2014 static bool manage_workers(struct worker *worker)
2016 struct worker_pool *pool = worker->pool;
2018 if (pool->flags & POOL_MANAGER_ACTIVE)
2021 pool->flags |= POOL_MANAGER_ACTIVE;
2022 pool->manager = worker;
2024 maybe_create_worker(pool);
2026 pool->manager = NULL;
2027 pool->flags &= ~POOL_MANAGER_ACTIVE;
2028 wake_up(&wq_manager_wait);
2033 * process_one_work - process single work
2035 * @work: work to process
2037 * Process @work. This function contains all the logics necessary to
2038 * process a single work including synchronization against and
2039 * interaction with other workers on the same cpu, queueing and
2040 * flushing. As long as context requirement is met, any worker can
2041 * call this function to process a work.
2044 * spin_lock_irq(pool->lock) which is released and regrabbed.
2046 static void process_one_work(struct worker *worker, struct work_struct *work)
2047 __releases(&pool->lock)
2048 __acquires(&pool->lock)
2050 struct pool_workqueue *pwq = get_work_pwq(work);
2051 struct worker_pool *pool = worker->pool;
2052 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2054 struct worker *collision;
2055 #ifdef CONFIG_LOCKDEP
2057 * It is permissible to free the struct work_struct from
2058 * inside the function that is called from it, this we need to
2059 * take into account for lockdep too. To avoid bogus "held
2060 * lock freed" warnings as well as problems when looking into
2061 * work->lockdep_map, make a copy and use that here.
2063 struct lockdep_map lockdep_map;
2065 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2067 /* ensure we're on the correct CPU */
2068 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2069 raw_smp_processor_id() != pool->cpu);
2072 * A single work shouldn't be executed concurrently by
2073 * multiple workers on a single cpu. Check whether anyone is
2074 * already processing the work. If so, defer the work to the
2075 * currently executing one.
2077 collision = find_worker_executing_work(pool, work);
2078 if (unlikely(collision)) {
2079 move_linked_works(work, &collision->scheduled, NULL);
2083 /* claim and dequeue */
2084 debug_work_deactivate(work);
2085 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2086 worker->current_work = work;
2087 worker->current_func = work->func;
2088 worker->current_pwq = pwq;
2089 work_color = get_work_color(work);
2092 * Record wq name for cmdline and debug reporting, may get
2093 * overridden through set_worker_desc().
2095 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2097 list_del_init(&work->entry);
2100 * CPU intensive works don't participate in concurrency management.
2101 * They're the scheduler's responsibility. This takes @worker out
2102 * of concurrency management and the next code block will chain
2103 * execution of the pending work items.
2105 if (unlikely(cpu_intensive))
2106 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2109 * Wake up another worker if necessary. The condition is always
2110 * false for normal per-cpu workers since nr_running would always
2111 * be >= 1 at this point. This is used to chain execution of the
2112 * pending work items for WORKER_NOT_RUNNING workers such as the
2113 * UNBOUND and CPU_INTENSIVE ones.
2115 if (need_more_worker(pool))
2116 wake_up_worker(pool);
2119 * Record the last pool and clear PENDING which should be the last
2120 * update to @work. Also, do this inside @pool->lock so that
2121 * PENDING and queued state changes happen together while IRQ is
2124 set_work_pool_and_clear_pending(work, pool->id);
2126 spin_unlock_irq(&pool->lock);
2128 lock_map_acquire(&pwq->wq->lockdep_map);
2129 lock_map_acquire(&lockdep_map);
2131 * Strictly speaking we should mark the invariant state without holding
2132 * any locks, that is, before these two lock_map_acquire()'s.
2134 * However, that would result in:
2141 * Which would create W1->C->W1 dependencies, even though there is no
2142 * actual deadlock possible. There are two solutions, using a
2143 * read-recursive acquire on the work(queue) 'locks', but this will then
2144 * hit the lockdep limitation on recursive locks, or simply discard
2147 * AFAICT there is no possible deadlock scenario between the
2148 * flush_work() and complete() primitives (except for single-threaded
2149 * workqueues), so hiding them isn't a problem.
2151 lockdep_invariant_state(true);
2152 trace_workqueue_execute_start(work);
2153 worker->current_func(work);
2155 * While we must be careful to not use "work" after this, the trace
2156 * point will only record its address.
2158 trace_workqueue_execute_end(work);
2159 lock_map_release(&lockdep_map);
2160 lock_map_release(&pwq->wq->lockdep_map);
2162 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2163 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2164 " last function: %pf\n",
2165 current->comm, preempt_count(), task_pid_nr(current),
2166 worker->current_func);
2167 debug_show_held_locks(current);
2172 * The following prevents a kworker from hogging CPU on !PREEMPT
2173 * kernels, where a requeueing work item waiting for something to
2174 * happen could deadlock with stop_machine as such work item could
2175 * indefinitely requeue itself while all other CPUs are trapped in
2176 * stop_machine. At the same time, report a quiescent RCU state so
2177 * the same condition doesn't freeze RCU.
2181 spin_lock_irq(&pool->lock);
2183 /* clear cpu intensive status */
2184 if (unlikely(cpu_intensive))
2185 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2187 /* we're done with it, release */
2188 hash_del(&worker->hentry);
2189 worker->current_work = NULL;
2190 worker->current_func = NULL;
2191 worker->current_pwq = NULL;
2192 pwq_dec_nr_in_flight(pwq, work_color);
2196 * process_scheduled_works - process scheduled works
2199 * Process all scheduled works. Please note that the scheduled list
2200 * may change while processing a work, so this function repeatedly
2201 * fetches a work from the top and executes it.
2204 * spin_lock_irq(pool->lock) which may be released and regrabbed
2207 static void process_scheduled_works(struct worker *worker)
2209 while (!list_empty(&worker->scheduled)) {
2210 struct work_struct *work = list_first_entry(&worker->scheduled,
2211 struct work_struct, entry);
2212 process_one_work(worker, work);
2216 static void set_pf_worker(bool val)
2218 mutex_lock(&wq_pool_attach_mutex);
2220 current->flags |= PF_WQ_WORKER;
2222 current->flags &= ~PF_WQ_WORKER;
2223 mutex_unlock(&wq_pool_attach_mutex);
2227 * worker_thread - the worker thread function
2230 * The worker thread function. All workers belong to a worker_pool -
2231 * either a per-cpu one or dynamic unbound one. These workers process all
2232 * work items regardless of their specific target workqueue. The only
2233 * exception is work items which belong to workqueues with a rescuer which
2234 * will be explained in rescuer_thread().
2238 static int worker_thread(void *__worker)
2240 struct worker *worker = __worker;
2241 struct worker_pool *pool = worker->pool;
2243 /* tell the scheduler that this is a workqueue worker */
2244 set_pf_worker(true);
2246 spin_lock_irq(&pool->lock);
2248 /* am I supposed to die? */
2249 if (unlikely(worker->flags & WORKER_DIE)) {
2250 spin_unlock_irq(&pool->lock);
2251 WARN_ON_ONCE(!list_empty(&worker->entry));
2252 set_pf_worker(false);
2254 set_task_comm(worker->task, "kworker/dying");
2255 ida_simple_remove(&pool->worker_ida, worker->id);
2256 worker_detach_from_pool(worker);
2261 worker_leave_idle(worker);
2263 /* no more worker necessary? */
2264 if (!need_more_worker(pool))
2267 /* do we need to manage? */
2268 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2272 * ->scheduled list can only be filled while a worker is
2273 * preparing to process a work or actually processing it.
2274 * Make sure nobody diddled with it while I was sleeping.
2276 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2279 * Finish PREP stage. We're guaranteed to have at least one idle
2280 * worker or that someone else has already assumed the manager
2281 * role. This is where @worker starts participating in concurrency
2282 * management if applicable and concurrency management is restored
2283 * after being rebound. See rebind_workers() for details.
2285 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2288 struct work_struct *work =
2289 list_first_entry(&pool->worklist,
2290 struct work_struct, entry);
2292 pool->watchdog_ts = jiffies;
2294 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2295 /* optimization path, not strictly necessary */
2296 process_one_work(worker, work);
2297 if (unlikely(!list_empty(&worker->scheduled)))
2298 process_scheduled_works(worker);
2300 move_linked_works(work, &worker->scheduled, NULL);
2301 process_scheduled_works(worker);
2303 } while (keep_working(pool));
2305 worker_set_flags(worker, WORKER_PREP);
2308 * pool->lock is held and there's no work to process and no need to
2309 * manage, sleep. Workers are woken up only while holding
2310 * pool->lock or from local cpu, so setting the current state
2311 * before releasing pool->lock is enough to prevent losing any
2314 worker_enter_idle(worker);
2315 __set_current_state(TASK_IDLE);
2316 spin_unlock_irq(&pool->lock);
2322 * rescuer_thread - the rescuer thread function
2325 * Workqueue rescuer thread function. There's one rescuer for each
2326 * workqueue which has WQ_MEM_RECLAIM set.
2328 * Regular work processing on a pool may block trying to create a new
2329 * worker which uses GFP_KERNEL allocation which has slight chance of
2330 * developing into deadlock if some works currently on the same queue
2331 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2332 * the problem rescuer solves.
2334 * When such condition is possible, the pool summons rescuers of all
2335 * workqueues which have works queued on the pool and let them process
2336 * those works so that forward progress can be guaranteed.
2338 * This should happen rarely.
2342 static int rescuer_thread(void *__rescuer)
2344 struct worker *rescuer = __rescuer;
2345 struct workqueue_struct *wq = rescuer->rescue_wq;
2346 struct list_head *scheduled = &rescuer->scheduled;
2349 set_user_nice(current, RESCUER_NICE_LEVEL);
2352 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2353 * doesn't participate in concurrency management.
2355 set_pf_worker(true);
2357 set_current_state(TASK_IDLE);
2360 * By the time the rescuer is requested to stop, the workqueue
2361 * shouldn't have any work pending, but @wq->maydays may still have
2362 * pwq(s) queued. This can happen by non-rescuer workers consuming
2363 * all the work items before the rescuer got to them. Go through
2364 * @wq->maydays processing before acting on should_stop so that the
2365 * list is always empty on exit.
2367 should_stop = kthread_should_stop();
2369 /* see whether any pwq is asking for help */
2370 spin_lock_irq(&wq_mayday_lock);
2372 while (!list_empty(&wq->maydays)) {
2373 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2374 struct pool_workqueue, mayday_node);
2375 struct worker_pool *pool = pwq->pool;
2376 struct work_struct *work, *n;
2379 __set_current_state(TASK_RUNNING);
2380 list_del_init(&pwq->mayday_node);
2382 spin_unlock_irq(&wq_mayday_lock);
2384 worker_attach_to_pool(rescuer, pool);
2386 spin_lock_irq(&pool->lock);
2389 * Slurp in all works issued via this workqueue and
2392 WARN_ON_ONCE(!list_empty(scheduled));
2393 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2394 if (get_work_pwq(work) == pwq) {
2396 pool->watchdog_ts = jiffies;
2397 move_linked_works(work, scheduled, &n);
2402 if (!list_empty(scheduled)) {
2403 process_scheduled_works(rescuer);
2406 * The above execution of rescued work items could
2407 * have created more to rescue through
2408 * pwq_activate_first_delayed() or chained
2409 * queueing. Let's put @pwq back on mayday list so
2410 * that such back-to-back work items, which may be
2411 * being used to relieve memory pressure, don't
2412 * incur MAYDAY_INTERVAL delay inbetween.
2414 if (need_to_create_worker(pool)) {
2415 spin_lock(&wq_mayday_lock);
2417 * Queue iff we aren't racing destruction
2418 * and somebody else hasn't queued it already.
2420 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2422 list_add_tail(&pwq->mayday_node, &wq->maydays);
2424 spin_unlock(&wq_mayday_lock);
2429 * Put the reference grabbed by send_mayday(). @pool won't
2430 * go away while we're still attached to it.
2435 * Leave this pool. If need_more_worker() is %true, notify a
2436 * regular worker; otherwise, we end up with 0 concurrency
2437 * and stalling the execution.
2439 if (need_more_worker(pool))
2440 wake_up_worker(pool);
2442 spin_unlock_irq(&pool->lock);
2444 worker_detach_from_pool(rescuer);
2446 spin_lock_irq(&wq_mayday_lock);
2449 spin_unlock_irq(&wq_mayday_lock);
2452 __set_current_state(TASK_RUNNING);
2453 set_pf_worker(false);
2457 /* rescuers should never participate in concurrency management */
2458 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2464 * check_flush_dependency - check for flush dependency sanity
2465 * @target_wq: workqueue being flushed
2466 * @target_work: work item being flushed (NULL for workqueue flushes)
2468 * %current is trying to flush the whole @target_wq or @target_work on it.
2469 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2470 * reclaiming memory or running on a workqueue which doesn't have
2471 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2474 static void check_flush_dependency(struct workqueue_struct *target_wq,
2475 struct work_struct *target_work)
2477 work_func_t target_func = target_work ? target_work->func : NULL;
2478 struct worker *worker;
2480 if (target_wq->flags & WQ_MEM_RECLAIM)
2483 worker = current_wq_worker();
2485 WARN_ONCE(current->flags & PF_MEMALLOC,
2486 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2487 current->pid, current->comm, target_wq->name, target_func);
2488 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2489 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2490 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2491 worker->current_pwq->wq->name, worker->current_func,
2492 target_wq->name, target_func);
2496 struct work_struct work;
2497 struct completion done;
2498 struct task_struct *task; /* purely informational */
2501 static void wq_barrier_func(struct work_struct *work)
2503 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2504 complete(&barr->done);
2508 * insert_wq_barrier - insert a barrier work
2509 * @pwq: pwq to insert barrier into
2510 * @barr: wq_barrier to insert
2511 * @target: target work to attach @barr to
2512 * @worker: worker currently executing @target, NULL if @target is not executing
2514 * @barr is linked to @target such that @barr is completed only after
2515 * @target finishes execution. Please note that the ordering
2516 * guarantee is observed only with respect to @target and on the local
2519 * Currently, a queued barrier can't be canceled. This is because
2520 * try_to_grab_pending() can't determine whether the work to be
2521 * grabbed is at the head of the queue and thus can't clear LINKED
2522 * flag of the previous work while there must be a valid next work
2523 * after a work with LINKED flag set.
2525 * Note that when @worker is non-NULL, @target may be modified
2526 * underneath us, so we can't reliably determine pwq from @target.
2529 * spin_lock_irq(pool->lock).
2531 static void insert_wq_barrier(struct pool_workqueue *pwq,
2532 struct wq_barrier *barr,
2533 struct work_struct *target, struct worker *worker)
2535 struct list_head *head;
2536 unsigned int linked = 0;
2539 * debugobject calls are safe here even with pool->lock locked
2540 * as we know for sure that this will not trigger any of the
2541 * checks and call back into the fixup functions where we
2544 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2545 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2547 init_completion_map(&barr->done, &target->lockdep_map);
2549 barr->task = current;
2552 * If @target is currently being executed, schedule the
2553 * barrier to the worker; otherwise, put it after @target.
2556 head = worker->scheduled.next;
2558 unsigned long *bits = work_data_bits(target);
2560 head = target->entry.next;
2561 /* there can already be other linked works, inherit and set */
2562 linked = *bits & WORK_STRUCT_LINKED;
2563 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2566 debug_work_activate(&barr->work);
2567 insert_work(pwq, &barr->work, head,
2568 work_color_to_flags(WORK_NO_COLOR) | linked);
2572 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2573 * @wq: workqueue being flushed
2574 * @flush_color: new flush color, < 0 for no-op
2575 * @work_color: new work color, < 0 for no-op
2577 * Prepare pwqs for workqueue flushing.
2579 * If @flush_color is non-negative, flush_color on all pwqs should be
2580 * -1. If no pwq has in-flight commands at the specified color, all
2581 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2582 * has in flight commands, its pwq->flush_color is set to
2583 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2584 * wakeup logic is armed and %true is returned.
2586 * The caller should have initialized @wq->first_flusher prior to
2587 * calling this function with non-negative @flush_color. If
2588 * @flush_color is negative, no flush color update is done and %false
2591 * If @work_color is non-negative, all pwqs should have the same
2592 * work_color which is previous to @work_color and all will be
2593 * advanced to @work_color.
2596 * mutex_lock(wq->mutex).
2599 * %true if @flush_color >= 0 and there's something to flush. %false
2602 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2603 int flush_color, int work_color)
2606 struct pool_workqueue *pwq;
2608 if (flush_color >= 0) {
2609 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2610 atomic_set(&wq->nr_pwqs_to_flush, 1);
2613 for_each_pwq(pwq, wq) {
2614 struct worker_pool *pool = pwq->pool;
2616 spin_lock_irq(&pool->lock);
2618 if (flush_color >= 0) {
2619 WARN_ON_ONCE(pwq->flush_color != -1);
2621 if (pwq->nr_in_flight[flush_color]) {
2622 pwq->flush_color = flush_color;
2623 atomic_inc(&wq->nr_pwqs_to_flush);
2628 if (work_color >= 0) {
2629 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2630 pwq->work_color = work_color;
2633 spin_unlock_irq(&pool->lock);
2636 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2637 complete(&wq->first_flusher->done);
2643 * flush_workqueue - ensure that any scheduled work has run to completion.
2644 * @wq: workqueue to flush
2646 * This function sleeps until all work items which were queued on entry
2647 * have finished execution, but it is not livelocked by new incoming ones.
2649 void flush_workqueue(struct workqueue_struct *wq)
2651 struct wq_flusher this_flusher = {
2652 .list = LIST_HEAD_INIT(this_flusher.list),
2654 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2658 if (WARN_ON(!wq_online))
2661 lock_map_acquire(&wq->lockdep_map);
2662 lock_map_release(&wq->lockdep_map);
2664 mutex_lock(&wq->mutex);
2667 * Start-to-wait phase
2669 next_color = work_next_color(wq->work_color);
2671 if (next_color != wq->flush_color) {
2673 * Color space is not full. The current work_color
2674 * becomes our flush_color and work_color is advanced
2677 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2678 this_flusher.flush_color = wq->work_color;
2679 wq->work_color = next_color;
2681 if (!wq->first_flusher) {
2682 /* no flush in progress, become the first flusher */
2683 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2685 wq->first_flusher = &this_flusher;
2687 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2689 /* nothing to flush, done */
2690 wq->flush_color = next_color;
2691 wq->first_flusher = NULL;
2696 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2697 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2698 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2702 * Oops, color space is full, wait on overflow queue.
2703 * The next flush completion will assign us
2704 * flush_color and transfer to flusher_queue.
2706 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2709 check_flush_dependency(wq, NULL);
2711 mutex_unlock(&wq->mutex);
2713 wait_for_completion(&this_flusher.done);
2716 * Wake-up-and-cascade phase
2718 * First flushers are responsible for cascading flushes and
2719 * handling overflow. Non-first flushers can simply return.
2721 if (wq->first_flusher != &this_flusher)
2724 mutex_lock(&wq->mutex);
2726 /* we might have raced, check again with mutex held */
2727 if (wq->first_flusher != &this_flusher)
2730 wq->first_flusher = NULL;
2732 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2733 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2736 struct wq_flusher *next, *tmp;
2738 /* complete all the flushers sharing the current flush color */
2739 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2740 if (next->flush_color != wq->flush_color)
2742 list_del_init(&next->list);
2743 complete(&next->done);
2746 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2747 wq->flush_color != work_next_color(wq->work_color));
2749 /* this flush_color is finished, advance by one */
2750 wq->flush_color = work_next_color(wq->flush_color);
2752 /* one color has been freed, handle overflow queue */
2753 if (!list_empty(&wq->flusher_overflow)) {
2755 * Assign the same color to all overflowed
2756 * flushers, advance work_color and append to
2757 * flusher_queue. This is the start-to-wait
2758 * phase for these overflowed flushers.
2760 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2761 tmp->flush_color = wq->work_color;
2763 wq->work_color = work_next_color(wq->work_color);
2765 list_splice_tail_init(&wq->flusher_overflow,
2766 &wq->flusher_queue);
2767 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2770 if (list_empty(&wq->flusher_queue)) {
2771 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2776 * Need to flush more colors. Make the next flusher
2777 * the new first flusher and arm pwqs.
2779 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2780 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2782 list_del_init(&next->list);
2783 wq->first_flusher = next;
2785 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2789 * Meh... this color is already done, clear first
2790 * flusher and repeat cascading.
2792 wq->first_flusher = NULL;
2796 mutex_unlock(&wq->mutex);
2798 EXPORT_SYMBOL(flush_workqueue);
2801 * drain_workqueue - drain a workqueue
2802 * @wq: workqueue to drain
2804 * Wait until the workqueue becomes empty. While draining is in progress,
2805 * only chain queueing is allowed. IOW, only currently pending or running
2806 * work items on @wq can queue further work items on it. @wq is flushed
2807 * repeatedly until it becomes empty. The number of flushing is determined
2808 * by the depth of chaining and should be relatively short. Whine if it
2811 void drain_workqueue(struct workqueue_struct *wq)
2813 unsigned int flush_cnt = 0;
2814 struct pool_workqueue *pwq;
2817 * __queue_work() needs to test whether there are drainers, is much
2818 * hotter than drain_workqueue() and already looks at @wq->flags.
2819 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2821 mutex_lock(&wq->mutex);
2822 if (!wq->nr_drainers++)
2823 wq->flags |= __WQ_DRAINING;
2824 mutex_unlock(&wq->mutex);
2826 flush_workqueue(wq);
2828 mutex_lock(&wq->mutex);
2830 for_each_pwq(pwq, wq) {
2833 spin_lock_irq(&pwq->pool->lock);
2834 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2835 spin_unlock_irq(&pwq->pool->lock);
2840 if (++flush_cnt == 10 ||
2841 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2842 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2843 wq->name, flush_cnt);
2845 mutex_unlock(&wq->mutex);
2849 if (!--wq->nr_drainers)
2850 wq->flags &= ~__WQ_DRAINING;
2851 mutex_unlock(&wq->mutex);
2853 EXPORT_SYMBOL_GPL(drain_workqueue);
2855 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2858 struct worker *worker = NULL;
2859 struct worker_pool *pool;
2860 struct pool_workqueue *pwq;
2864 local_irq_disable();
2865 pool = get_work_pool(work);
2871 spin_lock(&pool->lock);
2872 /* see the comment in try_to_grab_pending() with the same code */
2873 pwq = get_work_pwq(work);
2875 if (unlikely(pwq->pool != pool))
2878 worker = find_worker_executing_work(pool, work);
2881 pwq = worker->current_pwq;
2884 check_flush_dependency(pwq->wq, work);
2886 insert_wq_barrier(pwq, barr, work, worker);
2887 spin_unlock_irq(&pool->lock);
2890 * Force a lock recursion deadlock when using flush_work() inside a
2891 * single-threaded or rescuer equipped workqueue.
2893 * For single threaded workqueues the deadlock happens when the work
2894 * is after the work issuing the flush_work(). For rescuer equipped
2895 * workqueues the deadlock happens when the rescuer stalls, blocking
2899 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
2900 lock_map_acquire(&pwq->wq->lockdep_map);
2901 lock_map_release(&pwq->wq->lockdep_map);
2906 spin_unlock_irq(&pool->lock);
2910 static bool __flush_work(struct work_struct *work, bool from_cancel)
2912 struct wq_barrier barr;
2914 if (WARN_ON(!wq_online))
2917 if (WARN_ON(!work->func))
2921 lock_map_acquire(&work->lockdep_map);
2922 lock_map_release(&work->lockdep_map);
2925 if (start_flush_work(work, &barr, from_cancel)) {
2926 wait_for_completion(&barr.done);
2927 destroy_work_on_stack(&barr.work);
2935 * flush_work - wait for a work to finish executing the last queueing instance
2936 * @work: the work to flush
2938 * Wait until @work has finished execution. @work is guaranteed to be idle
2939 * on return if it hasn't been requeued since flush started.
2942 * %true if flush_work() waited for the work to finish execution,
2943 * %false if it was already idle.
2945 bool flush_work(struct work_struct *work)
2947 return __flush_work(work, false);
2949 EXPORT_SYMBOL_GPL(flush_work);
2952 wait_queue_entry_t wait;
2953 struct work_struct *work;
2956 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2958 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2960 if (cwait->work != key)
2962 return autoremove_wake_function(wait, mode, sync, key);
2965 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2967 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2968 unsigned long flags;
2972 ret = try_to_grab_pending(work, is_dwork, &flags);
2974 * If someone else is already canceling, wait for it to
2975 * finish. flush_work() doesn't work for PREEMPT_NONE
2976 * because we may get scheduled between @work's completion
2977 * and the other canceling task resuming and clearing
2978 * CANCELING - flush_work() will return false immediately
2979 * as @work is no longer busy, try_to_grab_pending() will
2980 * return -ENOENT as @work is still being canceled and the
2981 * other canceling task won't be able to clear CANCELING as
2982 * we're hogging the CPU.
2984 * Let's wait for completion using a waitqueue. As this
2985 * may lead to the thundering herd problem, use a custom
2986 * wake function which matches @work along with exclusive
2989 if (unlikely(ret == -ENOENT)) {
2990 struct cwt_wait cwait;
2992 init_wait(&cwait.wait);
2993 cwait.wait.func = cwt_wakefn;
2996 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2997 TASK_UNINTERRUPTIBLE);
2998 if (work_is_canceling(work))
3000 finish_wait(&cancel_waitq, &cwait.wait);
3002 } while (unlikely(ret < 0));
3004 /* tell other tasks trying to grab @work to back off */
3005 mark_work_canceling(work);
3006 local_irq_restore(flags);
3009 * This allows canceling during early boot. We know that @work
3013 __flush_work(work, true);
3015 clear_work_data(work);
3018 * Paired with prepare_to_wait() above so that either
3019 * waitqueue_active() is visible here or !work_is_canceling() is
3023 if (waitqueue_active(&cancel_waitq))
3024 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3030 * cancel_work_sync - cancel a work and wait for it to finish
3031 * @work: the work to cancel
3033 * Cancel @work and wait for its execution to finish. This function
3034 * can be used even if the work re-queues itself or migrates to
3035 * another workqueue. On return from this function, @work is
3036 * guaranteed to be not pending or executing on any CPU.
3038 * cancel_work_sync(&delayed_work->work) must not be used for
3039 * delayed_work's. Use cancel_delayed_work_sync() instead.
3041 * The caller must ensure that the workqueue on which @work was last
3042 * queued can't be destroyed before this function returns.
3045 * %true if @work was pending, %false otherwise.
3047 bool cancel_work_sync(struct work_struct *work)
3049 return __cancel_work_timer(work, false);
3051 EXPORT_SYMBOL_GPL(cancel_work_sync);
3054 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3055 * @dwork: the delayed work to flush
3057 * Delayed timer is cancelled and the pending work is queued for
3058 * immediate execution. Like flush_work(), this function only
3059 * considers the last queueing instance of @dwork.
3062 * %true if flush_work() waited for the work to finish execution,
3063 * %false if it was already idle.
3065 bool flush_delayed_work(struct delayed_work *dwork)
3067 local_irq_disable();
3068 if (del_timer_sync(&dwork->timer))
3069 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3071 return flush_work(&dwork->work);
3073 EXPORT_SYMBOL(flush_delayed_work);
3076 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3077 * @rwork: the rcu work to flush
3080 * %true if flush_rcu_work() waited for the work to finish execution,
3081 * %false if it was already idle.
3083 bool flush_rcu_work(struct rcu_work *rwork)
3085 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3087 flush_work(&rwork->work);
3090 return flush_work(&rwork->work);
3093 EXPORT_SYMBOL(flush_rcu_work);
3095 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3097 unsigned long flags;
3101 ret = try_to_grab_pending(work, is_dwork, &flags);
3102 } while (unlikely(ret == -EAGAIN));
3104 if (unlikely(ret < 0))
3107 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3108 local_irq_restore(flags);
3113 * cancel_delayed_work - cancel a delayed work
3114 * @dwork: delayed_work to cancel
3116 * Kill off a pending delayed_work.
3118 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3122 * The work callback function may still be running on return, unless
3123 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3124 * use cancel_delayed_work_sync() to wait on it.
3126 * This function is safe to call from any context including IRQ handler.
3128 bool cancel_delayed_work(struct delayed_work *dwork)
3130 return __cancel_work(&dwork->work, true);
3132 EXPORT_SYMBOL(cancel_delayed_work);
3135 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3136 * @dwork: the delayed work cancel
3138 * This is cancel_work_sync() for delayed works.
3141 * %true if @dwork was pending, %false otherwise.
3143 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3145 return __cancel_work_timer(&dwork->work, true);
3147 EXPORT_SYMBOL(cancel_delayed_work_sync);
3150 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3151 * @func: the function to call
3153 * schedule_on_each_cpu() executes @func on each online CPU using the
3154 * system workqueue and blocks until all CPUs have completed.
3155 * schedule_on_each_cpu() is very slow.
3158 * 0 on success, -errno on failure.
3160 int schedule_on_each_cpu(work_func_t func)
3163 struct work_struct __percpu *works;
3165 works = alloc_percpu(struct work_struct);
3171 for_each_online_cpu(cpu) {
3172 struct work_struct *work = per_cpu_ptr(works, cpu);
3174 INIT_WORK(work, func);
3175 schedule_work_on(cpu, work);
3178 for_each_online_cpu(cpu)
3179 flush_work(per_cpu_ptr(works, cpu));
3187 * execute_in_process_context - reliably execute the routine with user context
3188 * @fn: the function to execute
3189 * @ew: guaranteed storage for the execute work structure (must
3190 * be available when the work executes)
3192 * Executes the function immediately if process context is available,
3193 * otherwise schedules the function for delayed execution.
3195 * Return: 0 - function was executed
3196 * 1 - function was scheduled for execution
3198 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3200 if (!in_interrupt()) {
3205 INIT_WORK(&ew->work, fn);
3206 schedule_work(&ew->work);
3210 EXPORT_SYMBOL_GPL(execute_in_process_context);
3213 * free_workqueue_attrs - free a workqueue_attrs
3214 * @attrs: workqueue_attrs to free
3216 * Undo alloc_workqueue_attrs().
3218 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3221 free_cpumask_var(attrs->cpumask);
3227 * alloc_workqueue_attrs - allocate a workqueue_attrs
3228 * @gfp_mask: allocation mask to use
3230 * Allocate a new workqueue_attrs, initialize with default settings and
3233 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3235 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3237 struct workqueue_attrs *attrs;
3239 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3242 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3245 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3248 free_workqueue_attrs(attrs);
3252 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3253 const struct workqueue_attrs *from)
3255 to->nice = from->nice;
3256 cpumask_copy(to->cpumask, from->cpumask);
3258 * Unlike hash and equality test, this function doesn't ignore
3259 * ->no_numa as it is used for both pool and wq attrs. Instead,
3260 * get_unbound_pool() explicitly clears ->no_numa after copying.
3262 to->no_numa = from->no_numa;
3265 /* hash value of the content of @attr */
3266 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3270 hash = jhash_1word(attrs->nice, hash);
3271 hash = jhash(cpumask_bits(attrs->cpumask),
3272 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3276 /* content equality test */
3277 static bool wqattrs_equal(const struct workqueue_attrs *a,
3278 const struct workqueue_attrs *b)
3280 if (a->nice != b->nice)
3282 if (!cpumask_equal(a->cpumask, b->cpumask))
3288 * init_worker_pool - initialize a newly zalloc'd worker_pool
3289 * @pool: worker_pool to initialize
3291 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3293 * Return: 0 on success, -errno on failure. Even on failure, all fields
3294 * inside @pool proper are initialized and put_unbound_pool() can be called
3295 * on @pool safely to release it.
3297 static int init_worker_pool(struct worker_pool *pool)
3299 spin_lock_init(&pool->lock);
3302 pool->node = NUMA_NO_NODE;
3303 pool->flags |= POOL_DISASSOCIATED;
3304 pool->watchdog_ts = jiffies;
3305 INIT_LIST_HEAD(&pool->worklist);
3306 INIT_LIST_HEAD(&pool->idle_list);
3307 hash_init(pool->busy_hash);
3309 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3311 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3313 INIT_LIST_HEAD(&pool->workers);
3315 ida_init(&pool->worker_ida);
3316 INIT_HLIST_NODE(&pool->hash_node);
3319 /* shouldn't fail above this point */
3320 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3326 static void rcu_free_wq(struct rcu_head *rcu)
3328 struct workqueue_struct *wq =
3329 container_of(rcu, struct workqueue_struct, rcu);
3331 if (!(wq->flags & WQ_UNBOUND))
3332 free_percpu(wq->cpu_pwqs);
3334 free_workqueue_attrs(wq->unbound_attrs);
3340 static void rcu_free_pool(struct rcu_head *rcu)
3342 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3344 ida_destroy(&pool->worker_ida);
3345 free_workqueue_attrs(pool->attrs);
3350 * put_unbound_pool - put a worker_pool
3351 * @pool: worker_pool to put
3353 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3354 * safe manner. get_unbound_pool() calls this function on its failure path
3355 * and this function should be able to release pools which went through,
3356 * successfully or not, init_worker_pool().
3358 * Should be called with wq_pool_mutex held.
3360 static void put_unbound_pool(struct worker_pool *pool)
3362 DECLARE_COMPLETION_ONSTACK(detach_completion);
3363 struct worker *worker;
3365 lockdep_assert_held(&wq_pool_mutex);
3371 if (WARN_ON(!(pool->cpu < 0)) ||
3372 WARN_ON(!list_empty(&pool->worklist)))
3375 /* release id and unhash */
3377 idr_remove(&worker_pool_idr, pool->id);
3378 hash_del(&pool->hash_node);
3381 * Become the manager and destroy all workers. This prevents
3382 * @pool's workers from blocking on attach_mutex. We're the last
3383 * manager and @pool gets freed with the flag set.
3385 spin_lock_irq(&pool->lock);
3386 wait_event_lock_irq(wq_manager_wait,
3387 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3388 pool->flags |= POOL_MANAGER_ACTIVE;
3390 while ((worker = first_idle_worker(pool)))
3391 destroy_worker(worker);
3392 WARN_ON(pool->nr_workers || pool->nr_idle);
3393 spin_unlock_irq(&pool->lock);
3395 mutex_lock(&wq_pool_attach_mutex);
3396 if (!list_empty(&pool->workers))
3397 pool->detach_completion = &detach_completion;
3398 mutex_unlock(&wq_pool_attach_mutex);
3400 if (pool->detach_completion)
3401 wait_for_completion(pool->detach_completion);
3403 /* shut down the timers */
3404 del_timer_sync(&pool->idle_timer);
3405 del_timer_sync(&pool->mayday_timer);
3407 /* sched-RCU protected to allow dereferences from get_work_pool() */
3408 call_rcu_sched(&pool->rcu, rcu_free_pool);
3412 * get_unbound_pool - get a worker_pool with the specified attributes
3413 * @attrs: the attributes of the worker_pool to get
3415 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3416 * reference count and return it. If there already is a matching
3417 * worker_pool, it will be used; otherwise, this function attempts to
3420 * Should be called with wq_pool_mutex held.
3422 * Return: On success, a worker_pool with the same attributes as @attrs.
3423 * On failure, %NULL.
3425 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3427 u32 hash = wqattrs_hash(attrs);
3428 struct worker_pool *pool;
3430 int target_node = NUMA_NO_NODE;
3432 lockdep_assert_held(&wq_pool_mutex);
3434 /* do we already have a matching pool? */
3435 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3436 if (wqattrs_equal(pool->attrs, attrs)) {
3442 /* if cpumask is contained inside a NUMA node, we belong to that node */
3443 if (wq_numa_enabled) {
3444 for_each_node(node) {
3445 if (cpumask_subset(attrs->cpumask,
3446 wq_numa_possible_cpumask[node])) {
3453 /* nope, create a new one */
3454 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3455 if (!pool || init_worker_pool(pool) < 0)
3458 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3459 copy_workqueue_attrs(pool->attrs, attrs);
3460 pool->node = target_node;
3463 * no_numa isn't a worker_pool attribute, always clear it. See
3464 * 'struct workqueue_attrs' comments for detail.
3466 pool->attrs->no_numa = false;
3468 if (worker_pool_assign_id(pool) < 0)
3471 /* create and start the initial worker */
3472 if (wq_online && !create_worker(pool))
3476 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3481 put_unbound_pool(pool);
3485 static void rcu_free_pwq(struct rcu_head *rcu)
3487 kmem_cache_free(pwq_cache,
3488 container_of(rcu, struct pool_workqueue, rcu));
3492 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3493 * and needs to be destroyed.
3495 static void pwq_unbound_release_workfn(struct work_struct *work)
3497 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3498 unbound_release_work);
3499 struct workqueue_struct *wq = pwq->wq;
3500 struct worker_pool *pool = pwq->pool;
3501 bool is_last = false;
3504 * when @pwq is not linked, it doesn't hold any reference to the
3505 * @wq, and @wq is invalid to access.
3507 if (!list_empty(&pwq->pwqs_node)) {
3508 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3511 mutex_lock(&wq->mutex);
3512 list_del_rcu(&pwq->pwqs_node);
3513 is_last = list_empty(&wq->pwqs);
3514 mutex_unlock(&wq->mutex);
3517 mutex_lock(&wq_pool_mutex);
3518 put_unbound_pool(pool);
3519 mutex_unlock(&wq_pool_mutex);
3521 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3524 * If we're the last pwq going away, @wq is already dead and no one
3525 * is gonna access it anymore. Schedule RCU free.
3528 call_rcu_sched(&wq->rcu, rcu_free_wq);
3532 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3533 * @pwq: target pool_workqueue
3535 * If @pwq isn't freezing, set @pwq->max_active to the associated
3536 * workqueue's saved_max_active and activate delayed work items
3537 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3539 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3541 struct workqueue_struct *wq = pwq->wq;
3542 bool freezable = wq->flags & WQ_FREEZABLE;
3543 unsigned long flags;
3545 /* for @wq->saved_max_active */
3546 lockdep_assert_held(&wq->mutex);
3548 /* fast exit for non-freezable wqs */
3549 if (!freezable && pwq->max_active == wq->saved_max_active)
3552 /* this function can be called during early boot w/ irq disabled */
3553 spin_lock_irqsave(&pwq->pool->lock, flags);
3556 * During [un]freezing, the caller is responsible for ensuring that
3557 * this function is called at least once after @workqueue_freezing
3558 * is updated and visible.
3560 if (!freezable || !workqueue_freezing) {
3563 pwq->max_active = wq->saved_max_active;
3565 while (!list_empty(&pwq->delayed_works) &&
3566 pwq->nr_active < pwq->max_active) {
3567 pwq_activate_first_delayed(pwq);
3572 * Need to kick a worker after thawed or an unbound wq's
3573 * max_active is bumped. In realtime scenarios, always kicking a
3574 * worker will cause interference on the isolated cpu cores, so
3575 * let's kick iff work items were activated.
3578 wake_up_worker(pwq->pool);
3580 pwq->max_active = 0;
3583 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3586 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3587 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3588 struct worker_pool *pool)
3590 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3592 memset(pwq, 0, sizeof(*pwq));
3596 pwq->flush_color = -1;
3598 INIT_LIST_HEAD(&pwq->delayed_works);
3599 INIT_LIST_HEAD(&pwq->pwqs_node);
3600 INIT_LIST_HEAD(&pwq->mayday_node);
3601 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3604 /* sync @pwq with the current state of its associated wq and link it */
3605 static void link_pwq(struct pool_workqueue *pwq)
3607 struct workqueue_struct *wq = pwq->wq;
3609 lockdep_assert_held(&wq->mutex);
3611 /* may be called multiple times, ignore if already linked */
3612 if (!list_empty(&pwq->pwqs_node))
3615 /* set the matching work_color */
3616 pwq->work_color = wq->work_color;
3618 /* sync max_active to the current setting */
3619 pwq_adjust_max_active(pwq);
3622 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3625 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3626 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3627 const struct workqueue_attrs *attrs)
3629 struct worker_pool *pool;
3630 struct pool_workqueue *pwq;
3632 lockdep_assert_held(&wq_pool_mutex);
3634 pool = get_unbound_pool(attrs);
3638 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3640 put_unbound_pool(pool);
3644 init_pwq(pwq, wq, pool);
3649 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3650 * @attrs: the wq_attrs of the default pwq of the target workqueue
3651 * @node: the target NUMA node
3652 * @cpu_going_down: if >= 0, the CPU to consider as offline
3653 * @cpumask: outarg, the resulting cpumask
3655 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3656 * @cpu_going_down is >= 0, that cpu is considered offline during
3657 * calculation. The result is stored in @cpumask.
3659 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3660 * enabled and @node has online CPUs requested by @attrs, the returned
3661 * cpumask is the intersection of the possible CPUs of @node and
3664 * The caller is responsible for ensuring that the cpumask of @node stays
3667 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3670 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3671 int cpu_going_down, cpumask_t *cpumask)
3673 if (!wq_numa_enabled || attrs->no_numa)
3676 /* does @node have any online CPUs @attrs wants? */
3677 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3678 if (cpu_going_down >= 0)
3679 cpumask_clear_cpu(cpu_going_down, cpumask);
3681 if (cpumask_empty(cpumask))
3684 /* yeap, return possible CPUs in @node that @attrs wants */
3685 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3687 if (cpumask_empty(cpumask)) {
3688 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3689 "possible intersect\n");
3693 return !cpumask_equal(cpumask, attrs->cpumask);
3696 cpumask_copy(cpumask, attrs->cpumask);
3700 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3701 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3703 struct pool_workqueue *pwq)
3705 struct pool_workqueue *old_pwq;
3707 lockdep_assert_held(&wq_pool_mutex);
3708 lockdep_assert_held(&wq->mutex);
3710 /* link_pwq() can handle duplicate calls */
3713 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3714 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3718 /* context to store the prepared attrs & pwqs before applying */
3719 struct apply_wqattrs_ctx {
3720 struct workqueue_struct *wq; /* target workqueue */
3721 struct workqueue_attrs *attrs; /* attrs to apply */
3722 struct list_head list; /* queued for batching commit */
3723 struct pool_workqueue *dfl_pwq;
3724 struct pool_workqueue *pwq_tbl[];
3727 /* free the resources after success or abort */
3728 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3734 put_pwq_unlocked(ctx->pwq_tbl[node]);
3735 put_pwq_unlocked(ctx->dfl_pwq);
3737 free_workqueue_attrs(ctx->attrs);
3743 /* allocate the attrs and pwqs for later installation */
3744 static struct apply_wqattrs_ctx *
3745 apply_wqattrs_prepare(struct workqueue_struct *wq,
3746 const struct workqueue_attrs *attrs)
3748 struct apply_wqattrs_ctx *ctx;
3749 struct workqueue_attrs *new_attrs, *tmp_attrs;
3752 lockdep_assert_held(&wq_pool_mutex);
3754 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3756 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3757 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3758 if (!ctx || !new_attrs || !tmp_attrs)
3762 * Calculate the attrs of the default pwq.
3763 * If the user configured cpumask doesn't overlap with the
3764 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3766 copy_workqueue_attrs(new_attrs, attrs);
3767 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3768 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3769 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3772 * We may create multiple pwqs with differing cpumasks. Make a
3773 * copy of @new_attrs which will be modified and used to obtain
3776 copy_workqueue_attrs(tmp_attrs, new_attrs);
3779 * If something goes wrong during CPU up/down, we'll fall back to
3780 * the default pwq covering whole @attrs->cpumask. Always create
3781 * it even if we don't use it immediately.
3783 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3787 for_each_node(node) {
3788 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3789 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3790 if (!ctx->pwq_tbl[node])
3793 ctx->dfl_pwq->refcnt++;
3794 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3798 /* save the user configured attrs and sanitize it. */
3799 copy_workqueue_attrs(new_attrs, attrs);
3800 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3801 ctx->attrs = new_attrs;
3804 free_workqueue_attrs(tmp_attrs);
3808 free_workqueue_attrs(tmp_attrs);
3809 free_workqueue_attrs(new_attrs);
3810 apply_wqattrs_cleanup(ctx);
3814 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3815 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3819 /* all pwqs have been created successfully, let's install'em */
3820 mutex_lock(&ctx->wq->mutex);
3822 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3824 /* save the previous pwq and install the new one */
3826 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3827 ctx->pwq_tbl[node]);
3829 /* @dfl_pwq might not have been used, ensure it's linked */
3830 link_pwq(ctx->dfl_pwq);
3831 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3833 mutex_unlock(&ctx->wq->mutex);
3836 static void apply_wqattrs_lock(void)
3838 /* CPUs should stay stable across pwq creations and installations */
3840 mutex_lock(&wq_pool_mutex);
3843 static void apply_wqattrs_unlock(void)
3845 mutex_unlock(&wq_pool_mutex);
3849 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3850 const struct workqueue_attrs *attrs)
3852 struct apply_wqattrs_ctx *ctx;
3854 /* only unbound workqueues can change attributes */
3855 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3858 /* creating multiple pwqs breaks ordering guarantee */
3859 if (!list_empty(&wq->pwqs)) {
3860 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3863 wq->flags &= ~__WQ_ORDERED;
3866 ctx = apply_wqattrs_prepare(wq, attrs);
3870 /* the ctx has been prepared successfully, let's commit it */
3871 apply_wqattrs_commit(ctx);
3872 apply_wqattrs_cleanup(ctx);
3878 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3879 * @wq: the target workqueue
3880 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3882 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3883 * machines, this function maps a separate pwq to each NUMA node with
3884 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3885 * NUMA node it was issued on. Older pwqs are released as in-flight work
3886 * items finish. Note that a work item which repeatedly requeues itself
3887 * back-to-back will stay on its current pwq.
3889 * Performs GFP_KERNEL allocations.
3891 * Return: 0 on success and -errno on failure.
3893 int apply_workqueue_attrs(struct workqueue_struct *wq,
3894 const struct workqueue_attrs *attrs)
3898 apply_wqattrs_lock();
3899 ret = apply_workqueue_attrs_locked(wq, attrs);
3900 apply_wqattrs_unlock();
3904 EXPORT_SYMBOL_GPL(apply_workqueue_attrs);
3907 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3908 * @wq: the target workqueue
3909 * @cpu: the CPU coming up or going down
3910 * @online: whether @cpu is coming up or going down
3912 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3913 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3916 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3917 * falls back to @wq->dfl_pwq which may not be optimal but is always
3920 * Note that when the last allowed CPU of a NUMA node goes offline for a
3921 * workqueue with a cpumask spanning multiple nodes, the workers which were
3922 * already executing the work items for the workqueue will lose their CPU
3923 * affinity and may execute on any CPU. This is similar to how per-cpu
3924 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3925 * affinity, it's the user's responsibility to flush the work item from
3928 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3931 int node = cpu_to_node(cpu);
3932 int cpu_off = online ? -1 : cpu;
3933 struct pool_workqueue *old_pwq = NULL, *pwq;
3934 struct workqueue_attrs *target_attrs;
3937 lockdep_assert_held(&wq_pool_mutex);
3939 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3940 wq->unbound_attrs->no_numa)
3944 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3945 * Let's use a preallocated one. The following buf is protected by
3946 * CPU hotplug exclusion.
3948 target_attrs = wq_update_unbound_numa_attrs_buf;
3949 cpumask = target_attrs->cpumask;
3951 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3952 pwq = unbound_pwq_by_node(wq, node);
3955 * Let's determine what needs to be done. If the target cpumask is
3956 * different from the default pwq's, we need to compare it to @pwq's
3957 * and create a new one if they don't match. If the target cpumask
3958 * equals the default pwq's, the default pwq should be used.
3960 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3961 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3967 /* create a new pwq */
3968 pwq = alloc_unbound_pwq(wq, target_attrs);
3970 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3975 /* Install the new pwq. */
3976 mutex_lock(&wq->mutex);
3977 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3981 mutex_lock(&wq->mutex);
3982 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3983 get_pwq(wq->dfl_pwq);
3984 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3985 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3987 mutex_unlock(&wq->mutex);
3988 put_pwq_unlocked(old_pwq);
3991 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3993 bool highpri = wq->flags & WQ_HIGHPRI;
3996 if (!(wq->flags & WQ_UNBOUND)) {
3997 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4001 for_each_possible_cpu(cpu) {
4002 struct pool_workqueue *pwq =
4003 per_cpu_ptr(wq->cpu_pwqs, cpu);
4004 struct worker_pool *cpu_pools =
4005 per_cpu(cpu_worker_pools, cpu);
4007 init_pwq(pwq, wq, &cpu_pools[highpri]);
4009 mutex_lock(&wq->mutex);
4011 mutex_unlock(&wq->mutex);
4014 } else if (wq->flags & __WQ_ORDERED) {
4015 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4016 /* there should only be single pwq for ordering guarantee */
4017 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4018 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4019 "ordering guarantee broken for workqueue %s\n", wq->name);
4022 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4026 static int wq_clamp_max_active(int max_active, unsigned int flags,
4029 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4031 if (max_active < 1 || max_active > lim)
4032 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4033 max_active, name, 1, lim);
4035 return clamp_val(max_active, 1, lim);
4039 * Workqueues which may be used during memory reclaim should have a rescuer
4040 * to guarantee forward progress.
4042 static int init_rescuer(struct workqueue_struct *wq)
4044 struct worker *rescuer;
4047 if (!(wq->flags & WQ_MEM_RECLAIM))
4050 rescuer = alloc_worker(NUMA_NO_NODE);
4054 rescuer->rescue_wq = wq;
4055 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4056 ret = PTR_ERR_OR_ZERO(rescuer->task);
4062 wq->rescuer = rescuer;
4063 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4064 wake_up_process(rescuer->task);
4069 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4072 struct lock_class_key *key,
4073 const char *lock_name, ...)
4075 size_t tbl_size = 0;
4077 struct workqueue_struct *wq;
4078 struct pool_workqueue *pwq;
4081 * Unbound && max_active == 1 used to imply ordered, which is no
4082 * longer the case on NUMA machines due to per-node pools. While
4083 * alloc_ordered_workqueue() is the right way to create an ordered
4084 * workqueue, keep the previous behavior to avoid subtle breakages
4087 if ((flags & WQ_UNBOUND) && max_active == 1)
4088 flags |= __WQ_ORDERED;
4090 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4091 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4092 flags |= WQ_UNBOUND;
4094 /* allocate wq and format name */
4095 if (flags & WQ_UNBOUND)
4096 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4098 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4102 if (flags & WQ_UNBOUND) {
4103 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4104 if (!wq->unbound_attrs)
4108 va_start(args, lock_name);
4109 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4112 max_active = max_active ?: WQ_DFL_ACTIVE;
4113 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4117 wq->saved_max_active = max_active;
4118 mutex_init(&wq->mutex);
4119 atomic_set(&wq->nr_pwqs_to_flush, 0);
4120 INIT_LIST_HEAD(&wq->pwqs);
4121 INIT_LIST_HEAD(&wq->flusher_queue);
4122 INIT_LIST_HEAD(&wq->flusher_overflow);
4123 INIT_LIST_HEAD(&wq->maydays);
4125 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4126 INIT_LIST_HEAD(&wq->list);
4128 if (alloc_and_link_pwqs(wq) < 0)
4131 if (wq_online && init_rescuer(wq) < 0)
4134 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4138 * wq_pool_mutex protects global freeze state and workqueues list.
4139 * Grab it, adjust max_active and add the new @wq to workqueues
4142 mutex_lock(&wq_pool_mutex);
4144 mutex_lock(&wq->mutex);
4145 for_each_pwq(pwq, wq)
4146 pwq_adjust_max_active(pwq);
4147 mutex_unlock(&wq->mutex);
4149 list_add_tail_rcu(&wq->list, &workqueues);
4151 mutex_unlock(&wq_pool_mutex);
4156 free_workqueue_attrs(wq->unbound_attrs);
4160 destroy_workqueue(wq);
4163 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4166 * destroy_workqueue - safely terminate a workqueue
4167 * @wq: target workqueue
4169 * Safely destroy a workqueue. All work currently pending will be done first.
4171 void destroy_workqueue(struct workqueue_struct *wq)
4173 struct pool_workqueue *pwq;
4177 * Remove it from sysfs first so that sanity check failure doesn't
4178 * lead to sysfs name conflicts.
4180 workqueue_sysfs_unregister(wq);
4182 /* drain it before proceeding with destruction */
4183 drain_workqueue(wq);
4185 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4187 struct worker *rescuer = wq->rescuer;
4189 /* this prevents new queueing */
4190 spin_lock_irq(&wq_mayday_lock);
4192 spin_unlock_irq(&wq_mayday_lock);
4194 /* rescuer will empty maydays list before exiting */
4195 kthread_stop(rescuer->task);
4200 mutex_lock(&wq->mutex);
4201 for_each_pwq(pwq, wq) {
4204 for (i = 0; i < WORK_NR_COLORS; i++) {
4205 if (WARN_ON(pwq->nr_in_flight[i])) {
4206 mutex_unlock(&wq->mutex);
4207 show_workqueue_state();
4212 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4213 WARN_ON(pwq->nr_active) ||
4214 WARN_ON(!list_empty(&pwq->delayed_works))) {
4215 mutex_unlock(&wq->mutex);
4216 show_workqueue_state();
4220 mutex_unlock(&wq->mutex);
4223 * wq list is used to freeze wq, remove from list after
4224 * flushing is complete in case freeze races us.
4226 mutex_lock(&wq_pool_mutex);
4227 list_del_rcu(&wq->list);
4228 mutex_unlock(&wq_pool_mutex);
4230 if (!(wq->flags & WQ_UNBOUND)) {
4232 * The base ref is never dropped on per-cpu pwqs. Directly
4233 * schedule RCU free.
4235 call_rcu_sched(&wq->rcu, rcu_free_wq);
4238 * We're the sole accessor of @wq at this point. Directly
4239 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4240 * @wq will be freed when the last pwq is released.
4242 for_each_node(node) {
4243 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4244 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4245 put_pwq_unlocked(pwq);
4249 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4250 * put. Don't access it afterwards.
4254 put_pwq_unlocked(pwq);
4257 EXPORT_SYMBOL_GPL(destroy_workqueue);
4260 * workqueue_set_max_active - adjust max_active of a workqueue
4261 * @wq: target workqueue
4262 * @max_active: new max_active value.
4264 * Set max_active of @wq to @max_active.
4267 * Don't call from IRQ context.
4269 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4271 struct pool_workqueue *pwq;
4273 /* disallow meddling with max_active for ordered workqueues */
4274 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4277 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4279 mutex_lock(&wq->mutex);
4281 wq->flags &= ~__WQ_ORDERED;
4282 wq->saved_max_active = max_active;
4284 for_each_pwq(pwq, wq)
4285 pwq_adjust_max_active(pwq);
4287 mutex_unlock(&wq->mutex);
4289 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4292 * current_work - retrieve %current task's work struct
4294 * Determine if %current task is a workqueue worker and what it's working on.
4295 * Useful to find out the context that the %current task is running in.
4297 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4299 struct work_struct *current_work(void)
4301 struct worker *worker = current_wq_worker();
4303 return worker ? worker->current_work : NULL;
4305 EXPORT_SYMBOL(current_work);
4308 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4310 * Determine whether %current is a workqueue rescuer. Can be used from
4311 * work functions to determine whether it's being run off the rescuer task.
4313 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4315 bool current_is_workqueue_rescuer(void)
4317 struct worker *worker = current_wq_worker();
4319 return worker && worker->rescue_wq;
4323 * workqueue_congested - test whether a workqueue is congested
4324 * @cpu: CPU in question
4325 * @wq: target workqueue
4327 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4328 * no synchronization around this function and the test result is
4329 * unreliable and only useful as advisory hints or for debugging.
4331 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4332 * Note that both per-cpu and unbound workqueues may be associated with
4333 * multiple pool_workqueues which have separate congested states. A
4334 * workqueue being congested on one CPU doesn't mean the workqueue is also
4335 * contested on other CPUs / NUMA nodes.
4338 * %true if congested, %false otherwise.
4340 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4342 struct pool_workqueue *pwq;
4345 rcu_read_lock_sched();
4347 if (cpu == WORK_CPU_UNBOUND)
4348 cpu = smp_processor_id();
4350 if (!(wq->flags & WQ_UNBOUND))
4351 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4353 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4355 ret = !list_empty(&pwq->delayed_works);
4356 rcu_read_unlock_sched();
4360 EXPORT_SYMBOL_GPL(workqueue_congested);
4363 * work_busy - test whether a work is currently pending or running
4364 * @work: the work to be tested
4366 * Test whether @work is currently pending or running. There is no
4367 * synchronization around this function and the test result is
4368 * unreliable and only useful as advisory hints or for debugging.
4371 * OR'd bitmask of WORK_BUSY_* bits.
4373 unsigned int work_busy(struct work_struct *work)
4375 struct worker_pool *pool;
4376 unsigned long flags;
4377 unsigned int ret = 0;
4379 if (work_pending(work))
4380 ret |= WORK_BUSY_PENDING;
4382 local_irq_save(flags);
4383 pool = get_work_pool(work);
4385 spin_lock(&pool->lock);
4386 if (find_worker_executing_work(pool, work))
4387 ret |= WORK_BUSY_RUNNING;
4388 spin_unlock(&pool->lock);
4390 local_irq_restore(flags);
4394 EXPORT_SYMBOL_GPL(work_busy);
4397 * set_worker_desc - set description for the current work item
4398 * @fmt: printf-style format string
4399 * @...: arguments for the format string
4401 * This function can be called by a running work function to describe what
4402 * the work item is about. If the worker task gets dumped, this
4403 * information will be printed out together to help debugging. The
4404 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4406 void set_worker_desc(const char *fmt, ...)
4408 struct worker *worker = current_wq_worker();
4412 va_start(args, fmt);
4413 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4417 EXPORT_SYMBOL_GPL(set_worker_desc);
4420 * print_worker_info - print out worker information and description
4421 * @log_lvl: the log level to use when printing
4422 * @task: target task
4424 * If @task is a worker and currently executing a work item, print out the
4425 * name of the workqueue being serviced and worker description set with
4426 * set_worker_desc() by the currently executing work item.
4428 * This function can be safely called on any task as long as the
4429 * task_struct itself is accessible. While safe, this function isn't
4430 * synchronized and may print out mixups or garbages of limited length.
4432 void print_worker_info(const char *log_lvl, struct task_struct *task)
4434 work_func_t *fn = NULL;
4435 char name[WQ_NAME_LEN] = { };
4436 char desc[WORKER_DESC_LEN] = { };
4437 struct pool_workqueue *pwq = NULL;
4438 struct workqueue_struct *wq = NULL;
4439 struct worker *worker;
4441 if (!(task->flags & PF_WQ_WORKER))
4445 * This function is called without any synchronization and @task
4446 * could be in any state. Be careful with dereferences.
4448 worker = kthread_probe_data(task);
4451 * Carefully copy the associated workqueue's workfn, name and desc.
4452 * Keep the original last '\0' in case the original is garbage.
4454 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4455 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4456 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4457 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4458 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4460 if (fn || name[0] || desc[0]) {
4461 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4462 if (strcmp(name, desc))
4463 pr_cont(" (%s)", desc);
4468 static void pr_cont_pool_info(struct worker_pool *pool)
4470 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4471 if (pool->node != NUMA_NO_NODE)
4472 pr_cont(" node=%d", pool->node);
4473 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4476 static void pr_cont_work(bool comma, struct work_struct *work)
4478 if (work->func == wq_barrier_func) {
4479 struct wq_barrier *barr;
4481 barr = container_of(work, struct wq_barrier, work);
4483 pr_cont("%s BAR(%d)", comma ? "," : "",
4484 task_pid_nr(barr->task));
4486 pr_cont("%s %pf", comma ? "," : "", work->func);
4490 static void show_pwq(struct pool_workqueue *pwq)
4492 struct worker_pool *pool = pwq->pool;
4493 struct work_struct *work;
4494 struct worker *worker;
4495 bool has_in_flight = false, has_pending = false;
4498 pr_info(" pwq %d:", pool->id);
4499 pr_cont_pool_info(pool);
4501 pr_cont(" active=%d/%d refcnt=%d%s\n",
4502 pwq->nr_active, pwq->max_active, pwq->refcnt,
4503 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4505 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4506 if (worker->current_pwq == pwq) {
4507 has_in_flight = true;
4511 if (has_in_flight) {
4514 pr_info(" in-flight:");
4515 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4516 if (worker->current_pwq != pwq)
4519 pr_cont("%s %d%s:%pf", comma ? "," : "",
4520 task_pid_nr(worker->task),
4521 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4522 worker->current_func);
4523 list_for_each_entry(work, &worker->scheduled, entry)
4524 pr_cont_work(false, work);
4530 list_for_each_entry(work, &pool->worklist, entry) {
4531 if (get_work_pwq(work) == pwq) {
4539 pr_info(" pending:");
4540 list_for_each_entry(work, &pool->worklist, entry) {
4541 if (get_work_pwq(work) != pwq)
4544 pr_cont_work(comma, work);
4545 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4550 if (!list_empty(&pwq->delayed_works)) {
4553 pr_info(" delayed:");
4554 list_for_each_entry(work, &pwq->delayed_works, entry) {
4555 pr_cont_work(comma, work);
4556 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4563 * show_workqueue_state - dump workqueue state
4565 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4566 * all busy workqueues and pools.
4568 void show_workqueue_state(void)
4570 struct workqueue_struct *wq;
4571 struct worker_pool *pool;
4572 unsigned long flags;
4575 rcu_read_lock_sched();
4577 pr_info("Showing busy workqueues and worker pools:\n");
4579 list_for_each_entry_rcu(wq, &workqueues, list) {
4580 struct pool_workqueue *pwq;
4583 for_each_pwq(pwq, wq) {
4584 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4592 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4594 for_each_pwq(pwq, wq) {
4595 spin_lock_irqsave(&pwq->pool->lock, flags);
4596 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4598 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4600 * We could be printing a lot from atomic context, e.g.
4601 * sysrq-t -> show_workqueue_state(). Avoid triggering
4604 touch_nmi_watchdog();
4608 for_each_pool(pool, pi) {
4609 struct worker *worker;
4612 spin_lock_irqsave(&pool->lock, flags);
4613 if (pool->nr_workers == pool->nr_idle)
4616 pr_info("pool %d:", pool->id);
4617 pr_cont_pool_info(pool);
4618 pr_cont(" hung=%us workers=%d",
4619 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4622 pr_cont(" manager: %d",
4623 task_pid_nr(pool->manager->task));
4624 list_for_each_entry(worker, &pool->idle_list, entry) {
4625 pr_cont(" %s%d", first ? "idle: " : "",
4626 task_pid_nr(worker->task));
4631 spin_unlock_irqrestore(&pool->lock, flags);
4633 * We could be printing a lot from atomic context, e.g.
4634 * sysrq-t -> show_workqueue_state(). Avoid triggering
4637 touch_nmi_watchdog();
4640 rcu_read_unlock_sched();
4643 /* used to show worker information through /proc/PID/{comm,stat,status} */
4644 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4648 /* always show the actual comm */
4649 off = strscpy(buf, task->comm, size);
4653 /* stabilize PF_WQ_WORKER and worker pool association */
4654 mutex_lock(&wq_pool_attach_mutex);
4656 if (task->flags & PF_WQ_WORKER) {
4657 struct worker *worker = kthread_data(task);
4658 struct worker_pool *pool = worker->pool;
4661 spin_lock_irq(&pool->lock);
4663 * ->desc tracks information (wq name or
4664 * set_worker_desc()) for the latest execution. If
4665 * current, prepend '+', otherwise '-'.
4667 if (worker->desc[0] != '\0') {
4668 if (worker->current_work)
4669 scnprintf(buf + off, size - off, "+%s",
4672 scnprintf(buf + off, size - off, "-%s",
4675 spin_unlock_irq(&pool->lock);
4679 mutex_unlock(&wq_pool_attach_mutex);
4687 * There are two challenges in supporting CPU hotplug. Firstly, there
4688 * are a lot of assumptions on strong associations among work, pwq and
4689 * pool which make migrating pending and scheduled works very
4690 * difficult to implement without impacting hot paths. Secondly,
4691 * worker pools serve mix of short, long and very long running works making
4692 * blocked draining impractical.
4694 * This is solved by allowing the pools to be disassociated from the CPU
4695 * running as an unbound one and allowing it to be reattached later if the
4696 * cpu comes back online.
4699 static void unbind_workers(int cpu)
4701 struct worker_pool *pool;
4702 struct worker *worker;
4704 for_each_cpu_worker_pool(pool, cpu) {
4705 mutex_lock(&wq_pool_attach_mutex);
4706 spin_lock_irq(&pool->lock);
4709 * We've blocked all attach/detach operations. Make all workers
4710 * unbound and set DISASSOCIATED. Before this, all workers
4711 * except for the ones which are still executing works from
4712 * before the last CPU down must be on the cpu. After
4713 * this, they may become diasporas.
4715 for_each_pool_worker(worker, pool)
4716 worker->flags |= WORKER_UNBOUND;
4718 pool->flags |= POOL_DISASSOCIATED;
4720 spin_unlock_irq(&pool->lock);
4721 mutex_unlock(&wq_pool_attach_mutex);
4724 * Call schedule() so that we cross rq->lock and thus can
4725 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4726 * This is necessary as scheduler callbacks may be invoked
4732 * Sched callbacks are disabled now. Zap nr_running.
4733 * After this, nr_running stays zero and need_more_worker()
4734 * and keep_working() are always true as long as the
4735 * worklist is not empty. This pool now behaves as an
4736 * unbound (in terms of concurrency management) pool which
4737 * are served by workers tied to the pool.
4739 atomic_set(&pool->nr_running, 0);
4742 * With concurrency management just turned off, a busy
4743 * worker blocking could lead to lengthy stalls. Kick off
4744 * unbound chain execution of currently pending work items.
4746 spin_lock_irq(&pool->lock);
4747 wake_up_worker(pool);
4748 spin_unlock_irq(&pool->lock);
4753 * rebind_workers - rebind all workers of a pool to the associated CPU
4754 * @pool: pool of interest
4756 * @pool->cpu is coming online. Rebind all workers to the CPU.
4758 static void rebind_workers(struct worker_pool *pool)
4760 struct worker *worker;
4762 lockdep_assert_held(&wq_pool_attach_mutex);
4765 * Restore CPU affinity of all workers. As all idle workers should
4766 * be on the run-queue of the associated CPU before any local
4767 * wake-ups for concurrency management happen, restore CPU affinity
4768 * of all workers first and then clear UNBOUND. As we're called
4769 * from CPU_ONLINE, the following shouldn't fail.
4771 for_each_pool_worker(worker, pool)
4772 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4773 pool->attrs->cpumask) < 0);
4775 spin_lock_irq(&pool->lock);
4777 pool->flags &= ~POOL_DISASSOCIATED;
4779 for_each_pool_worker(worker, pool) {
4780 unsigned int worker_flags = worker->flags;
4783 * A bound idle worker should actually be on the runqueue
4784 * of the associated CPU for local wake-ups targeting it to
4785 * work. Kick all idle workers so that they migrate to the
4786 * associated CPU. Doing this in the same loop as
4787 * replacing UNBOUND with REBOUND is safe as no worker will
4788 * be bound before @pool->lock is released.
4790 if (worker_flags & WORKER_IDLE)
4791 wake_up_process(worker->task);
4794 * We want to clear UNBOUND but can't directly call
4795 * worker_clr_flags() or adjust nr_running. Atomically
4796 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4797 * @worker will clear REBOUND using worker_clr_flags() when
4798 * it initiates the next execution cycle thus restoring
4799 * concurrency management. Note that when or whether
4800 * @worker clears REBOUND doesn't affect correctness.
4802 * WRITE_ONCE() is necessary because @worker->flags may be
4803 * tested without holding any lock in
4804 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4805 * fail incorrectly leading to premature concurrency
4806 * management operations.
4808 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4809 worker_flags |= WORKER_REBOUND;
4810 worker_flags &= ~WORKER_UNBOUND;
4811 WRITE_ONCE(worker->flags, worker_flags);
4814 spin_unlock_irq(&pool->lock);
4818 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4819 * @pool: unbound pool of interest
4820 * @cpu: the CPU which is coming up
4822 * An unbound pool may end up with a cpumask which doesn't have any online
4823 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4824 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4825 * online CPU before, cpus_allowed of all its workers should be restored.
4827 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4829 static cpumask_t cpumask;
4830 struct worker *worker;
4832 lockdep_assert_held(&wq_pool_attach_mutex);
4834 /* is @cpu allowed for @pool? */
4835 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4838 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4840 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4841 for_each_pool_worker(worker, pool)
4842 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4845 int workqueue_prepare_cpu(unsigned int cpu)
4847 struct worker_pool *pool;
4849 for_each_cpu_worker_pool(pool, cpu) {
4850 if (pool->nr_workers)
4852 if (!create_worker(pool))
4858 int workqueue_online_cpu(unsigned int cpu)
4860 struct worker_pool *pool;
4861 struct workqueue_struct *wq;
4864 mutex_lock(&wq_pool_mutex);
4866 for_each_pool(pool, pi) {
4867 mutex_lock(&wq_pool_attach_mutex);
4869 if (pool->cpu == cpu)
4870 rebind_workers(pool);
4871 else if (pool->cpu < 0)
4872 restore_unbound_workers_cpumask(pool, cpu);
4874 mutex_unlock(&wq_pool_attach_mutex);
4877 /* update NUMA affinity of unbound workqueues */
4878 list_for_each_entry(wq, &workqueues, list)
4879 wq_update_unbound_numa(wq, cpu, true);
4881 mutex_unlock(&wq_pool_mutex);
4885 int workqueue_offline_cpu(unsigned int cpu)
4887 struct workqueue_struct *wq;
4889 /* unbinding per-cpu workers should happen on the local CPU */
4890 if (WARN_ON(cpu != smp_processor_id()))
4893 unbind_workers(cpu);
4895 /* update NUMA affinity of unbound workqueues */
4896 mutex_lock(&wq_pool_mutex);
4897 list_for_each_entry(wq, &workqueues, list)
4898 wq_update_unbound_numa(wq, cpu, false);
4899 mutex_unlock(&wq_pool_mutex);
4904 struct work_for_cpu {
4905 struct work_struct work;
4911 static void work_for_cpu_fn(struct work_struct *work)
4913 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4915 wfc->ret = wfc->fn(wfc->arg);
4919 * work_on_cpu - run a function in thread context on a particular cpu
4920 * @cpu: the cpu to run on
4921 * @fn: the function to run
4922 * @arg: the function arg
4924 * It is up to the caller to ensure that the cpu doesn't go offline.
4925 * The caller must not hold any locks which would prevent @fn from completing.
4927 * Return: The value @fn returns.
4929 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4931 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4933 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4934 schedule_work_on(cpu, &wfc.work);
4935 flush_work(&wfc.work);
4936 destroy_work_on_stack(&wfc.work);
4939 EXPORT_SYMBOL_GPL(work_on_cpu);
4942 * work_on_cpu_safe - run a function in thread context on a particular cpu
4943 * @cpu: the cpu to run on
4944 * @fn: the function to run
4945 * @arg: the function argument
4947 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4948 * any locks which would prevent @fn from completing.
4950 * Return: The value @fn returns.
4952 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4957 if (cpu_online(cpu))
4958 ret = work_on_cpu(cpu, fn, arg);
4962 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
4963 #endif /* CONFIG_SMP */
4965 #ifdef CONFIG_FREEZER
4968 * freeze_workqueues_begin - begin freezing workqueues
4970 * Start freezing workqueues. After this function returns, all freezable
4971 * workqueues will queue new works to their delayed_works list instead of
4975 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4977 void freeze_workqueues_begin(void)
4979 struct workqueue_struct *wq;
4980 struct pool_workqueue *pwq;
4982 mutex_lock(&wq_pool_mutex);
4984 WARN_ON_ONCE(workqueue_freezing);
4985 workqueue_freezing = true;
4987 list_for_each_entry(wq, &workqueues, list) {
4988 mutex_lock(&wq->mutex);
4989 for_each_pwq(pwq, wq)
4990 pwq_adjust_max_active(pwq);
4991 mutex_unlock(&wq->mutex);
4994 mutex_unlock(&wq_pool_mutex);
4998 * freeze_workqueues_busy - are freezable workqueues still busy?
5000 * Check whether freezing is complete. This function must be called
5001 * between freeze_workqueues_begin() and thaw_workqueues().
5004 * Grabs and releases wq_pool_mutex.
5007 * %true if some freezable workqueues are still busy. %false if freezing
5010 bool freeze_workqueues_busy(void)
5013 struct workqueue_struct *wq;
5014 struct pool_workqueue *pwq;
5016 mutex_lock(&wq_pool_mutex);
5018 WARN_ON_ONCE(!workqueue_freezing);
5020 list_for_each_entry(wq, &workqueues, list) {
5021 if (!(wq->flags & WQ_FREEZABLE))
5024 * nr_active is monotonically decreasing. It's safe
5025 * to peek without lock.
5027 rcu_read_lock_sched();
5028 for_each_pwq(pwq, wq) {
5029 WARN_ON_ONCE(pwq->nr_active < 0);
5030 if (pwq->nr_active) {
5032 rcu_read_unlock_sched();
5036 rcu_read_unlock_sched();
5039 mutex_unlock(&wq_pool_mutex);
5044 * thaw_workqueues - thaw workqueues
5046 * Thaw workqueues. Normal queueing is restored and all collected
5047 * frozen works are transferred to their respective pool worklists.
5050 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5052 void thaw_workqueues(void)
5054 struct workqueue_struct *wq;
5055 struct pool_workqueue *pwq;
5057 mutex_lock(&wq_pool_mutex);
5059 if (!workqueue_freezing)
5062 workqueue_freezing = false;
5064 /* restore max_active and repopulate worklist */
5065 list_for_each_entry(wq, &workqueues, list) {
5066 mutex_lock(&wq->mutex);
5067 for_each_pwq(pwq, wq)
5068 pwq_adjust_max_active(pwq);
5069 mutex_unlock(&wq->mutex);
5073 mutex_unlock(&wq_pool_mutex);
5075 #endif /* CONFIG_FREEZER */
5077 static int workqueue_apply_unbound_cpumask(void)
5081 struct workqueue_struct *wq;
5082 struct apply_wqattrs_ctx *ctx, *n;
5084 lockdep_assert_held(&wq_pool_mutex);
5086 list_for_each_entry(wq, &workqueues, list) {
5087 if (!(wq->flags & WQ_UNBOUND))
5089 /* creating multiple pwqs breaks ordering guarantee */
5090 if (wq->flags & __WQ_ORDERED)
5093 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5099 list_add_tail(&ctx->list, &ctxs);
5102 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5104 apply_wqattrs_commit(ctx);
5105 apply_wqattrs_cleanup(ctx);
5112 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5113 * @cpumask: the cpumask to set
5115 * The low-level workqueues cpumask is a global cpumask that limits
5116 * the affinity of all unbound workqueues. This function check the @cpumask
5117 * and apply it to all unbound workqueues and updates all pwqs of them.
5119 * Retun: 0 - Success
5120 * -EINVAL - Invalid @cpumask
5121 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5123 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5126 cpumask_var_t saved_cpumask;
5129 * Not excluding isolated cpus on purpose.
5130 * If the user wishes to include them, we allow that.
5132 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5133 if (!cpumask_empty(cpumask)) {
5134 apply_wqattrs_lock();
5135 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5140 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) {
5145 /* save the old wq_unbound_cpumask. */
5146 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5148 /* update wq_unbound_cpumask at first and apply it to wqs. */
5149 cpumask_copy(wq_unbound_cpumask, cpumask);
5150 ret = workqueue_apply_unbound_cpumask();
5152 /* restore the wq_unbound_cpumask when failed. */
5154 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5156 free_cpumask_var(saved_cpumask);
5158 apply_wqattrs_unlock();
5166 * Workqueues with WQ_SYSFS flag set is visible to userland via
5167 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5168 * following attributes.
5170 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5171 * max_active RW int : maximum number of in-flight work items
5173 * Unbound workqueues have the following extra attributes.
5175 * pool_ids RO int : the associated pool IDs for each node
5176 * nice RW int : nice value of the workers
5177 * cpumask RW mask : bitmask of allowed CPUs for the workers
5178 * numa RW bool : whether enable NUMA affinity
5181 struct workqueue_struct *wq;
5185 static struct workqueue_struct *dev_to_wq(struct device *dev)
5187 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5192 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5195 struct workqueue_struct *wq = dev_to_wq(dev);
5197 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5199 static DEVICE_ATTR_RO(per_cpu);
5201 static ssize_t max_active_show(struct device *dev,
5202 struct device_attribute *attr, char *buf)
5204 struct workqueue_struct *wq = dev_to_wq(dev);
5206 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5209 static ssize_t max_active_store(struct device *dev,
5210 struct device_attribute *attr, const char *buf,
5213 struct workqueue_struct *wq = dev_to_wq(dev);
5216 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5219 workqueue_set_max_active(wq, val);
5222 static DEVICE_ATTR_RW(max_active);
5224 static struct attribute *wq_sysfs_attrs[] = {
5225 &dev_attr_per_cpu.attr,
5226 &dev_attr_max_active.attr,
5229 ATTRIBUTE_GROUPS(wq_sysfs);
5231 static ssize_t wq_pool_ids_show(struct device *dev,
5232 struct device_attribute *attr, char *buf)
5234 struct workqueue_struct *wq = dev_to_wq(dev);
5235 const char *delim = "";
5236 int node, written = 0;
5238 rcu_read_lock_sched();
5239 for_each_node(node) {
5240 written += scnprintf(buf + written, PAGE_SIZE - written,
5241 "%s%d:%d", delim, node,
5242 unbound_pwq_by_node(wq, node)->pool->id);
5245 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5246 rcu_read_unlock_sched();
5251 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5254 struct workqueue_struct *wq = dev_to_wq(dev);
5257 mutex_lock(&wq->mutex);
5258 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5259 mutex_unlock(&wq->mutex);
5264 /* prepare workqueue_attrs for sysfs store operations */
5265 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5267 struct workqueue_attrs *attrs;
5269 lockdep_assert_held(&wq_pool_mutex);
5271 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5275 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5279 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5280 const char *buf, size_t count)
5282 struct workqueue_struct *wq = dev_to_wq(dev);
5283 struct workqueue_attrs *attrs;
5286 apply_wqattrs_lock();
5288 attrs = wq_sysfs_prep_attrs(wq);
5292 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5293 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5294 ret = apply_workqueue_attrs_locked(wq, attrs);
5299 apply_wqattrs_unlock();
5300 free_workqueue_attrs(attrs);
5301 return ret ?: count;
5304 static ssize_t wq_cpumask_show(struct device *dev,
5305 struct device_attribute *attr, char *buf)
5307 struct workqueue_struct *wq = dev_to_wq(dev);
5310 mutex_lock(&wq->mutex);
5311 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5312 cpumask_pr_args(wq->unbound_attrs->cpumask));
5313 mutex_unlock(&wq->mutex);
5317 static ssize_t wq_cpumask_store(struct device *dev,
5318 struct device_attribute *attr,
5319 const char *buf, size_t count)
5321 struct workqueue_struct *wq = dev_to_wq(dev);
5322 struct workqueue_attrs *attrs;
5325 apply_wqattrs_lock();
5327 attrs = wq_sysfs_prep_attrs(wq);
5331 ret = cpumask_parse(buf, attrs->cpumask);
5333 ret = apply_workqueue_attrs_locked(wq, attrs);
5336 apply_wqattrs_unlock();
5337 free_workqueue_attrs(attrs);
5338 return ret ?: count;
5341 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5344 struct workqueue_struct *wq = dev_to_wq(dev);
5347 mutex_lock(&wq->mutex);
5348 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5349 !wq->unbound_attrs->no_numa);
5350 mutex_unlock(&wq->mutex);
5355 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5356 const char *buf, size_t count)
5358 struct workqueue_struct *wq = dev_to_wq(dev);
5359 struct workqueue_attrs *attrs;
5360 int v, ret = -ENOMEM;
5362 apply_wqattrs_lock();
5364 attrs = wq_sysfs_prep_attrs(wq);
5369 if (sscanf(buf, "%d", &v) == 1) {
5370 attrs->no_numa = !v;
5371 ret = apply_workqueue_attrs_locked(wq, attrs);
5375 apply_wqattrs_unlock();
5376 free_workqueue_attrs(attrs);
5377 return ret ?: count;
5380 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5381 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5382 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5383 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5384 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5388 static struct bus_type wq_subsys = {
5389 .name = "workqueue",
5390 .dev_groups = wq_sysfs_groups,
5393 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5394 struct device_attribute *attr, char *buf)
5398 mutex_lock(&wq_pool_mutex);
5399 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5400 cpumask_pr_args(wq_unbound_cpumask));
5401 mutex_unlock(&wq_pool_mutex);
5406 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5407 struct device_attribute *attr, const char *buf, size_t count)
5409 cpumask_var_t cpumask;
5412 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5415 ret = cpumask_parse(buf, cpumask);
5417 ret = workqueue_set_unbound_cpumask(cpumask);
5419 free_cpumask_var(cpumask);
5420 return ret ? ret : count;
5423 static struct device_attribute wq_sysfs_cpumask_attr =
5424 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5425 wq_unbound_cpumask_store);
5427 static int __init wq_sysfs_init(void)
5431 err = subsys_virtual_register(&wq_subsys, NULL);
5435 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5437 core_initcall(wq_sysfs_init);
5439 static void wq_device_release(struct device *dev)
5441 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5447 * workqueue_sysfs_register - make a workqueue visible in sysfs
5448 * @wq: the workqueue to register
5450 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5451 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5452 * which is the preferred method.
5454 * Workqueue user should use this function directly iff it wants to apply
5455 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5456 * apply_workqueue_attrs() may race against userland updating the
5459 * Return: 0 on success, -errno on failure.
5461 int workqueue_sysfs_register(struct workqueue_struct *wq)
5463 struct wq_device *wq_dev;
5467 * Adjusting max_active or creating new pwqs by applying
5468 * attributes breaks ordering guarantee. Disallow exposing ordered
5471 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5474 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5479 wq_dev->dev.bus = &wq_subsys;
5480 wq_dev->dev.release = wq_device_release;
5481 dev_set_name(&wq_dev->dev, "%s", wq->name);
5484 * unbound_attrs are created separately. Suppress uevent until
5485 * everything is ready.
5487 dev_set_uevent_suppress(&wq_dev->dev, true);
5489 ret = device_register(&wq_dev->dev);
5491 put_device(&wq_dev->dev);
5496 if (wq->flags & WQ_UNBOUND) {
5497 struct device_attribute *attr;
5499 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5500 ret = device_create_file(&wq_dev->dev, attr);
5502 device_unregister(&wq_dev->dev);
5509 dev_set_uevent_suppress(&wq_dev->dev, false);
5510 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5515 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5516 * @wq: the workqueue to unregister
5518 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5520 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5522 struct wq_device *wq_dev = wq->wq_dev;
5528 device_unregister(&wq_dev->dev);
5530 #else /* CONFIG_SYSFS */
5531 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5532 #endif /* CONFIG_SYSFS */
5535 * Workqueue watchdog.
5537 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5538 * flush dependency, a concurrency managed work item which stays RUNNING
5539 * indefinitely. Workqueue stalls can be very difficult to debug as the
5540 * usual warning mechanisms don't trigger and internal workqueue state is
5543 * Workqueue watchdog monitors all worker pools periodically and dumps
5544 * state if some pools failed to make forward progress for a while where
5545 * forward progress is defined as the first item on ->worklist changing.
5547 * This mechanism is controlled through the kernel parameter
5548 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5549 * corresponding sysfs parameter file.
5551 #ifdef CONFIG_WQ_WATCHDOG
5553 static unsigned long wq_watchdog_thresh = 30;
5554 static struct timer_list wq_watchdog_timer;
5556 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5557 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5559 static void wq_watchdog_reset_touched(void)
5563 wq_watchdog_touched = jiffies;
5564 for_each_possible_cpu(cpu)
5565 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5568 static void wq_watchdog_timer_fn(struct timer_list *unused)
5570 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5571 bool lockup_detected = false;
5572 unsigned long now = jiffies;
5573 struct worker_pool *pool;
5581 for_each_pool(pool, pi) {
5582 unsigned long pool_ts, touched, ts;
5584 if (list_empty(&pool->worklist))
5588 * If a virtual machine is stopped by the host it can look to
5589 * the watchdog like a stall.
5591 kvm_check_and_clear_guest_paused();
5593 /* get the latest of pool and touched timestamps */
5594 pool_ts = READ_ONCE(pool->watchdog_ts);
5595 touched = READ_ONCE(wq_watchdog_touched);
5597 if (time_after(pool_ts, touched))
5602 if (pool->cpu >= 0) {
5603 unsigned long cpu_touched =
5604 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5606 if (time_after(cpu_touched, ts))
5611 if (time_after(now, ts + thresh)) {
5612 lockup_detected = true;
5613 pr_emerg("BUG: workqueue lockup - pool");
5614 pr_cont_pool_info(pool);
5615 pr_cont(" stuck for %us!\n",
5616 jiffies_to_msecs(now - pool_ts) / 1000);
5622 if (lockup_detected)
5623 show_workqueue_state();
5625 wq_watchdog_reset_touched();
5626 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5629 notrace void wq_watchdog_touch(int cpu)
5632 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5634 wq_watchdog_touched = jiffies;
5637 static void wq_watchdog_set_thresh(unsigned long thresh)
5639 wq_watchdog_thresh = 0;
5640 del_timer_sync(&wq_watchdog_timer);
5643 wq_watchdog_thresh = thresh;
5644 wq_watchdog_reset_touched();
5645 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5649 static int wq_watchdog_param_set_thresh(const char *val,
5650 const struct kernel_param *kp)
5652 unsigned long thresh;
5655 ret = kstrtoul(val, 0, &thresh);
5660 wq_watchdog_set_thresh(thresh);
5662 wq_watchdog_thresh = thresh;
5667 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5668 .set = wq_watchdog_param_set_thresh,
5669 .get = param_get_ulong,
5672 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5675 static void wq_watchdog_init(void)
5677 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5678 wq_watchdog_set_thresh(wq_watchdog_thresh);
5681 #else /* CONFIG_WQ_WATCHDOG */
5683 static inline void wq_watchdog_init(void) { }
5685 #endif /* CONFIG_WQ_WATCHDOG */
5687 static void __init wq_numa_init(void)
5692 if (num_possible_nodes() <= 1)
5695 if (wq_disable_numa) {
5696 pr_info("workqueue: NUMA affinity support disabled\n");
5700 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5701 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5704 * We want masks of possible CPUs of each node which isn't readily
5705 * available. Build one from cpu_to_node() which should have been
5706 * fully initialized by now.
5708 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5712 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5713 node_online(node) ? node : NUMA_NO_NODE));
5715 for_each_possible_cpu(cpu) {
5716 node = cpu_to_node(cpu);
5717 if (WARN_ON(node == NUMA_NO_NODE)) {
5718 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5719 /* happens iff arch is bonkers, let's just proceed */
5722 cpumask_set_cpu(cpu, tbl[node]);
5725 wq_numa_possible_cpumask = tbl;
5726 wq_numa_enabled = true;
5730 * workqueue_init_early - early init for workqueue subsystem
5732 * This is the first half of two-staged workqueue subsystem initialization
5733 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5734 * idr are up. It sets up all the data structures and system workqueues
5735 * and allows early boot code to create workqueues and queue/cancel work
5736 * items. Actual work item execution starts only after kthreads can be
5737 * created and scheduled right before early initcalls.
5739 int __init workqueue_init_early(void)
5741 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5742 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5745 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5747 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5748 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5750 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5752 /* initialize CPU pools */
5753 for_each_possible_cpu(cpu) {
5754 struct worker_pool *pool;
5757 for_each_cpu_worker_pool(pool, cpu) {
5758 BUG_ON(init_worker_pool(pool));
5760 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5761 pool->attrs->nice = std_nice[i++];
5762 pool->node = cpu_to_node(cpu);
5765 mutex_lock(&wq_pool_mutex);
5766 BUG_ON(worker_pool_assign_id(pool));
5767 mutex_unlock(&wq_pool_mutex);
5771 /* create default unbound and ordered wq attrs */
5772 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5773 struct workqueue_attrs *attrs;
5775 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5776 attrs->nice = std_nice[i];
5777 unbound_std_wq_attrs[i] = attrs;
5780 * An ordered wq should have only one pwq as ordering is
5781 * guaranteed by max_active which is enforced by pwqs.
5782 * Turn off NUMA so that dfl_pwq is used for all nodes.
5784 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5785 attrs->nice = std_nice[i];
5786 attrs->no_numa = true;
5787 ordered_wq_attrs[i] = attrs;
5790 system_wq = alloc_workqueue("events", 0, 0);
5791 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5792 system_long_wq = alloc_workqueue("events_long", 0, 0);
5793 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5794 WQ_UNBOUND_MAX_ACTIVE);
5795 system_freezable_wq = alloc_workqueue("events_freezable",
5797 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5798 WQ_POWER_EFFICIENT, 0);
5799 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5800 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5802 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5803 !system_unbound_wq || !system_freezable_wq ||
5804 !system_power_efficient_wq ||
5805 !system_freezable_power_efficient_wq);
5811 * workqueue_init - bring workqueue subsystem fully online
5813 * This is the latter half of two-staged workqueue subsystem initialization
5814 * and invoked as soon as kthreads can be created and scheduled.
5815 * Workqueues have been created and work items queued on them, but there
5816 * are no kworkers executing the work items yet. Populate the worker pools
5817 * with the initial workers and enable future kworker creations.
5819 int __init workqueue_init(void)
5821 struct workqueue_struct *wq;
5822 struct worker_pool *pool;
5826 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5827 * CPU to node mapping may not be available that early on some
5828 * archs such as power and arm64. As per-cpu pools created
5829 * previously could be missing node hint and unbound pools NUMA
5830 * affinity, fix them up.
5832 * Also, while iterating workqueues, create rescuers if requested.
5836 mutex_lock(&wq_pool_mutex);
5838 for_each_possible_cpu(cpu) {
5839 for_each_cpu_worker_pool(pool, cpu) {
5840 pool->node = cpu_to_node(cpu);
5844 list_for_each_entry(wq, &workqueues, list) {
5845 wq_update_unbound_numa(wq, smp_processor_id(), true);
5846 WARN(init_rescuer(wq),
5847 "workqueue: failed to create early rescuer for %s",
5851 mutex_unlock(&wq_pool_mutex);
5853 /* create the initial workers */
5854 for_each_online_cpu(cpu) {
5855 for_each_cpu_worker_pool(pool, cpu) {
5856 pool->flags &= ~POOL_DISASSOCIATED;
5857 BUG_ON(!create_worker(pool));
5861 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5862 BUG_ON(!create_worker(pool));