2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
135 * WQ: wq->mutex protected.
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
139 * MD: wq_mayday_lock protected.
142 /* struct worker is defined in workqueue_internal.h */
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
151 unsigned long watchdog_ts; /* L: watchdog timestamp */
153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle; /* L: currently idle ones */
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb; /* manager arbitration */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
174 struct ida worker_ida; /* worker IDs for task name */
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp;
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
225 * Structure used to wait for workqueue flush.
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
261 struct wq_device *wq_dev; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache *pwq_cache;
283 static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa;
287 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
295 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
298 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
301 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302 static bool workqueue_freezing; /* PL: have wqs started freezing? */
304 /* PL: allowable cpus for unbound wqs and work items */
305 static cpumask_var_t wq_unbound_cpumask;
307 /* CPU where unbound work was last round robin scheduled from this CPU */
308 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
311 * Local execution of unbound work items is no longer guaranteed. The
312 * following always forces round-robin CPU selection on unbound work items
313 * to uncover usages which depend on it.
315 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316 static bool wq_debug_force_rr_cpu = true;
318 static bool wq_debug_force_rr_cpu = false;
320 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
322 /* the per-cpu worker pools */
323 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
326 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
328 /* PL: hash of all unbound pools keyed by pool->attrs */
329 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
331 /* I: attributes used when instantiating standard unbound pools on demand */
332 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
334 /* I: attributes used when instantiating ordered pools on demand */
335 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
337 struct workqueue_struct *system_wq __read_mostly;
338 EXPORT_SYMBOL(system_wq);
339 struct workqueue_struct *system_highpri_wq __read_mostly;
340 EXPORT_SYMBOL_GPL(system_highpri_wq);
341 struct workqueue_struct *system_long_wq __read_mostly;
342 EXPORT_SYMBOL_GPL(system_long_wq);
343 struct workqueue_struct *system_unbound_wq __read_mostly;
344 EXPORT_SYMBOL_GPL(system_unbound_wq);
345 struct workqueue_struct *system_freezable_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_freezable_wq);
347 struct workqueue_struct *system_power_efficient_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
349 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
352 static int worker_thread(void *__worker);
353 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
355 #define CREATE_TRACE_POINTS
356 #include <trace/events/workqueue.h>
358 #define assert_rcu_or_pool_mutex() \
359 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
360 !lockdep_is_held(&wq_pool_mutex), \
361 "sched RCU or wq_pool_mutex should be held")
363 #define assert_rcu_or_wq_mutex(wq) \
364 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
365 !lockdep_is_held(&wq->mutex), \
366 "sched RCU or wq->mutex should be held")
368 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
369 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
370 !lockdep_is_held(&wq->mutex) && \
371 !lockdep_is_held(&wq_pool_mutex), \
372 "sched RCU, wq->mutex or wq_pool_mutex should be held")
374 #define for_each_cpu_worker_pool(pool, cpu) \
375 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
376 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
380 * for_each_pool - iterate through all worker_pools in the system
381 * @pool: iteration cursor
382 * @pi: integer used for iteration
384 * This must be called either with wq_pool_mutex held or sched RCU read
385 * locked. If the pool needs to be used beyond the locking in effect, the
386 * caller is responsible for guaranteeing that the pool stays online.
388 * The if/else clause exists only for the lockdep assertion and can be
391 #define for_each_pool(pool, pi) \
392 idr_for_each_entry(&worker_pool_idr, pool, pi) \
393 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
397 * for_each_pool_worker - iterate through all workers of a worker_pool
398 * @worker: iteration cursor
399 * @pool: worker_pool to iterate workers of
401 * This must be called with @pool->attach_mutex.
403 * The if/else clause exists only for the lockdep assertion and can be
406 #define for_each_pool_worker(worker, pool) \
407 list_for_each_entry((worker), &(pool)->workers, node) \
408 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
412 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
413 * @pwq: iteration cursor
414 * @wq: the target workqueue
416 * This must be called either with wq->mutex held or sched RCU read locked.
417 * If the pwq needs to be used beyond the locking in effect, the caller is
418 * responsible for guaranteeing that the pwq stays online.
420 * The if/else clause exists only for the lockdep assertion and can be
423 #define for_each_pwq(pwq, wq) \
424 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
425 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
428 #ifdef CONFIG_DEBUG_OBJECTS_WORK
430 static struct debug_obj_descr work_debug_descr;
432 static void *work_debug_hint(void *addr)
434 return ((struct work_struct *) addr)->func;
438 * fixup_init is called when:
439 * - an active object is initialized
441 static int work_fixup_init(void *addr, enum debug_obj_state state)
443 struct work_struct *work = addr;
446 case ODEBUG_STATE_ACTIVE:
447 cancel_work_sync(work);
448 debug_object_init(work, &work_debug_descr);
456 * fixup_activate is called when:
457 * - an active object is activated
458 * - an unknown object is activated (might be a statically initialized object)
460 static int work_fixup_activate(void *addr, enum debug_obj_state state)
462 struct work_struct *work = addr;
466 case ODEBUG_STATE_NOTAVAILABLE:
468 * This is not really a fixup. The work struct was
469 * statically initialized. We just make sure that it
470 * is tracked in the object tracker.
472 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
473 debug_object_init(work, &work_debug_descr);
474 debug_object_activate(work, &work_debug_descr);
480 case ODEBUG_STATE_ACTIVE:
489 * fixup_free is called when:
490 * - an active object is freed
492 static int work_fixup_free(void *addr, enum debug_obj_state state)
494 struct work_struct *work = addr;
497 case ODEBUG_STATE_ACTIVE:
498 cancel_work_sync(work);
499 debug_object_free(work, &work_debug_descr);
506 static struct debug_obj_descr work_debug_descr = {
507 .name = "work_struct",
508 .debug_hint = work_debug_hint,
509 .fixup_init = work_fixup_init,
510 .fixup_activate = work_fixup_activate,
511 .fixup_free = work_fixup_free,
514 static inline void debug_work_activate(struct work_struct *work)
516 debug_object_activate(work, &work_debug_descr);
519 static inline void debug_work_deactivate(struct work_struct *work)
521 debug_object_deactivate(work, &work_debug_descr);
524 void __init_work(struct work_struct *work, int onstack)
527 debug_object_init_on_stack(work, &work_debug_descr);
529 debug_object_init(work, &work_debug_descr);
531 EXPORT_SYMBOL_GPL(__init_work);
533 void destroy_work_on_stack(struct work_struct *work)
535 debug_object_free(work, &work_debug_descr);
537 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
539 void destroy_delayed_work_on_stack(struct delayed_work *work)
541 destroy_timer_on_stack(&work->timer);
542 debug_object_free(&work->work, &work_debug_descr);
544 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
547 static inline void debug_work_activate(struct work_struct *work) { }
548 static inline void debug_work_deactivate(struct work_struct *work) { }
552 * worker_pool_assign_id - allocate ID and assing it to @pool
553 * @pool: the pool pointer of interest
555 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
556 * successfully, -errno on failure.
558 static int worker_pool_assign_id(struct worker_pool *pool)
562 lockdep_assert_held(&wq_pool_mutex);
564 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
574 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
575 * @wq: the target workqueue
578 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
580 * If the pwq needs to be used beyond the locking in effect, the caller is
581 * responsible for guaranteeing that the pwq stays online.
583 * Return: The unbound pool_workqueue for @node.
585 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
588 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
589 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
592 static unsigned int work_color_to_flags(int color)
594 return color << WORK_STRUCT_COLOR_SHIFT;
597 static int get_work_color(struct work_struct *work)
599 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
600 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
603 static int work_next_color(int color)
605 return (color + 1) % WORK_NR_COLORS;
609 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
610 * contain the pointer to the queued pwq. Once execution starts, the flag
611 * is cleared and the high bits contain OFFQ flags and pool ID.
613 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
614 * and clear_work_data() can be used to set the pwq, pool or clear
615 * work->data. These functions should only be called while the work is
616 * owned - ie. while the PENDING bit is set.
618 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
619 * corresponding to a work. Pool is available once the work has been
620 * queued anywhere after initialization until it is sync canceled. pwq is
621 * available only while the work item is queued.
623 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
624 * canceled. While being canceled, a work item may have its PENDING set
625 * but stay off timer and worklist for arbitrarily long and nobody should
626 * try to steal the PENDING bit.
628 static inline void set_work_data(struct work_struct *work, unsigned long data,
631 WARN_ON_ONCE(!work_pending(work));
632 atomic_long_set(&work->data, data | flags | work_static(work));
635 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
636 unsigned long extra_flags)
638 set_work_data(work, (unsigned long)pwq,
639 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
642 static void set_work_pool_and_keep_pending(struct work_struct *work,
645 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
646 WORK_STRUCT_PENDING);
649 static void set_work_pool_and_clear_pending(struct work_struct *work,
653 * The following wmb is paired with the implied mb in
654 * test_and_set_bit(PENDING) and ensures all updates to @work made
655 * here are visible to and precede any updates by the next PENDING
659 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
662 static void clear_work_data(struct work_struct *work)
664 smp_wmb(); /* see set_work_pool_and_clear_pending() */
665 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
668 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
670 unsigned long data = atomic_long_read(&work->data);
672 if (data & WORK_STRUCT_PWQ)
673 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
679 * get_work_pool - return the worker_pool a given work was associated with
680 * @work: the work item of interest
682 * Pools are created and destroyed under wq_pool_mutex, and allows read
683 * access under sched-RCU read lock. As such, this function should be
684 * called under wq_pool_mutex or with preemption disabled.
686 * All fields of the returned pool are accessible as long as the above
687 * mentioned locking is in effect. If the returned pool needs to be used
688 * beyond the critical section, the caller is responsible for ensuring the
689 * returned pool is and stays online.
691 * Return: The worker_pool @work was last associated with. %NULL if none.
693 static struct worker_pool *get_work_pool(struct work_struct *work)
695 unsigned long data = atomic_long_read(&work->data);
698 assert_rcu_or_pool_mutex();
700 if (data & WORK_STRUCT_PWQ)
701 return ((struct pool_workqueue *)
702 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
704 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
705 if (pool_id == WORK_OFFQ_POOL_NONE)
708 return idr_find(&worker_pool_idr, pool_id);
712 * get_work_pool_id - return the worker pool ID a given work is associated with
713 * @work: the work item of interest
715 * Return: The worker_pool ID @work was last associated with.
716 * %WORK_OFFQ_POOL_NONE if none.
718 static int get_work_pool_id(struct work_struct *work)
720 unsigned long data = atomic_long_read(&work->data);
722 if (data & WORK_STRUCT_PWQ)
723 return ((struct pool_workqueue *)
724 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
726 return data >> WORK_OFFQ_POOL_SHIFT;
729 static void mark_work_canceling(struct work_struct *work)
731 unsigned long pool_id = get_work_pool_id(work);
733 pool_id <<= WORK_OFFQ_POOL_SHIFT;
734 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
737 static bool work_is_canceling(struct work_struct *work)
739 unsigned long data = atomic_long_read(&work->data);
741 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
745 * Policy functions. These define the policies on how the global worker
746 * pools are managed. Unless noted otherwise, these functions assume that
747 * they're being called with pool->lock held.
750 static bool __need_more_worker(struct worker_pool *pool)
752 return !atomic_read(&pool->nr_running);
756 * Need to wake up a worker? Called from anything but currently
759 * Note that, because unbound workers never contribute to nr_running, this
760 * function will always return %true for unbound pools as long as the
761 * worklist isn't empty.
763 static bool need_more_worker(struct worker_pool *pool)
765 return !list_empty(&pool->worklist) && __need_more_worker(pool);
768 /* Can I start working? Called from busy but !running workers. */
769 static bool may_start_working(struct worker_pool *pool)
771 return pool->nr_idle;
774 /* Do I need to keep working? Called from currently running workers. */
775 static bool keep_working(struct worker_pool *pool)
777 return !list_empty(&pool->worklist) &&
778 atomic_read(&pool->nr_running) <= 1;
781 /* Do we need a new worker? Called from manager. */
782 static bool need_to_create_worker(struct worker_pool *pool)
784 return need_more_worker(pool) && !may_start_working(pool);
787 /* Do we have too many workers and should some go away? */
788 static bool too_many_workers(struct worker_pool *pool)
790 bool managing = mutex_is_locked(&pool->manager_arb);
791 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
792 int nr_busy = pool->nr_workers - nr_idle;
794 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
801 /* Return the first idle worker. Safe with preemption disabled */
802 static struct worker *first_idle_worker(struct worker_pool *pool)
804 if (unlikely(list_empty(&pool->idle_list)))
807 return list_first_entry(&pool->idle_list, struct worker, entry);
811 * wake_up_worker - wake up an idle worker
812 * @pool: worker pool to wake worker from
814 * Wake up the first idle worker of @pool.
817 * spin_lock_irq(pool->lock).
819 static void wake_up_worker(struct worker_pool *pool)
821 struct worker *worker = first_idle_worker(pool);
824 wake_up_process(worker->task);
828 * wq_worker_waking_up - a worker is waking up
829 * @task: task waking up
830 * @cpu: CPU @task is waking up to
832 * This function is called during try_to_wake_up() when a worker is
836 * spin_lock_irq(rq->lock)
838 void wq_worker_waking_up(struct task_struct *task, int cpu)
840 struct worker *worker = kthread_data(task);
842 if (!(worker->flags & WORKER_NOT_RUNNING)) {
843 WARN_ON_ONCE(worker->pool->cpu != cpu);
844 atomic_inc(&worker->pool->nr_running);
849 * wq_worker_sleeping - a worker is going to sleep
850 * @task: task going to sleep
851 * @cpu: CPU in question, must be the current CPU number
853 * This function is called during schedule() when a busy worker is
854 * going to sleep. Worker on the same cpu can be woken up by
855 * returning pointer to its task.
858 * spin_lock_irq(rq->lock)
861 * Worker task on @cpu to wake up, %NULL if none.
863 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
865 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
866 struct worker_pool *pool;
869 * Rescuers, which may not have all the fields set up like normal
870 * workers, also reach here, let's not access anything before
871 * checking NOT_RUNNING.
873 if (worker->flags & WORKER_NOT_RUNNING)
878 /* this can only happen on the local cpu */
879 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
883 * The counterpart of the following dec_and_test, implied mb,
884 * worklist not empty test sequence is in insert_work().
885 * Please read comment there.
887 * NOT_RUNNING is clear. This means that we're bound to and
888 * running on the local cpu w/ rq lock held and preemption
889 * disabled, which in turn means that none else could be
890 * manipulating idle_list, so dereferencing idle_list without pool
893 if (atomic_dec_and_test(&pool->nr_running) &&
894 !list_empty(&pool->worklist))
895 to_wakeup = first_idle_worker(pool);
896 return to_wakeup ? to_wakeup->task : NULL;
900 * worker_set_flags - set worker flags and adjust nr_running accordingly
902 * @flags: flags to set
904 * Set @flags in @worker->flags and adjust nr_running accordingly.
907 * spin_lock_irq(pool->lock)
909 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
911 struct worker_pool *pool = worker->pool;
913 WARN_ON_ONCE(worker->task != current);
915 /* If transitioning into NOT_RUNNING, adjust nr_running. */
916 if ((flags & WORKER_NOT_RUNNING) &&
917 !(worker->flags & WORKER_NOT_RUNNING)) {
918 atomic_dec(&pool->nr_running);
921 worker->flags |= flags;
925 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
927 * @flags: flags to clear
929 * Clear @flags in @worker->flags and adjust nr_running accordingly.
932 * spin_lock_irq(pool->lock)
934 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
936 struct worker_pool *pool = worker->pool;
937 unsigned int oflags = worker->flags;
939 WARN_ON_ONCE(worker->task != current);
941 worker->flags &= ~flags;
944 * If transitioning out of NOT_RUNNING, increment nr_running. Note
945 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
946 * of multiple flags, not a single flag.
948 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
949 if (!(worker->flags & WORKER_NOT_RUNNING))
950 atomic_inc(&pool->nr_running);
954 * find_worker_executing_work - find worker which is executing a work
955 * @pool: pool of interest
956 * @work: work to find worker for
958 * Find a worker which is executing @work on @pool by searching
959 * @pool->busy_hash which is keyed by the address of @work. For a worker
960 * to match, its current execution should match the address of @work and
961 * its work function. This is to avoid unwanted dependency between
962 * unrelated work executions through a work item being recycled while still
965 * This is a bit tricky. A work item may be freed once its execution
966 * starts and nothing prevents the freed area from being recycled for
967 * another work item. If the same work item address ends up being reused
968 * before the original execution finishes, workqueue will identify the
969 * recycled work item as currently executing and make it wait until the
970 * current execution finishes, introducing an unwanted dependency.
972 * This function checks the work item address and work function to avoid
973 * false positives. Note that this isn't complete as one may construct a
974 * work function which can introduce dependency onto itself through a
975 * recycled work item. Well, if somebody wants to shoot oneself in the
976 * foot that badly, there's only so much we can do, and if such deadlock
977 * actually occurs, it should be easy to locate the culprit work function.
980 * spin_lock_irq(pool->lock).
983 * Pointer to worker which is executing @work if found, %NULL
986 static struct worker *find_worker_executing_work(struct worker_pool *pool,
987 struct work_struct *work)
989 struct worker *worker;
991 hash_for_each_possible(pool->busy_hash, worker, hentry,
993 if (worker->current_work == work &&
994 worker->current_func == work->func)
1001 * move_linked_works - move linked works to a list
1002 * @work: start of series of works to be scheduled
1003 * @head: target list to append @work to
1004 * @nextp: out parameter for nested worklist walking
1006 * Schedule linked works starting from @work to @head. Work series to
1007 * be scheduled starts at @work and includes any consecutive work with
1008 * WORK_STRUCT_LINKED set in its predecessor.
1010 * If @nextp is not NULL, it's updated to point to the next work of
1011 * the last scheduled work. This allows move_linked_works() to be
1012 * nested inside outer list_for_each_entry_safe().
1015 * spin_lock_irq(pool->lock).
1017 static void move_linked_works(struct work_struct *work, struct list_head *head,
1018 struct work_struct **nextp)
1020 struct work_struct *n;
1023 * Linked worklist will always end before the end of the list,
1024 * use NULL for list head.
1026 list_for_each_entry_safe_from(work, n, NULL, entry) {
1027 list_move_tail(&work->entry, head);
1028 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1033 * If we're already inside safe list traversal and have moved
1034 * multiple works to the scheduled queue, the next position
1035 * needs to be updated.
1042 * get_pwq - get an extra reference on the specified pool_workqueue
1043 * @pwq: pool_workqueue to get
1045 * Obtain an extra reference on @pwq. The caller should guarantee that
1046 * @pwq has positive refcnt and be holding the matching pool->lock.
1048 static void get_pwq(struct pool_workqueue *pwq)
1050 lockdep_assert_held(&pwq->pool->lock);
1051 WARN_ON_ONCE(pwq->refcnt <= 0);
1056 * put_pwq - put a pool_workqueue reference
1057 * @pwq: pool_workqueue to put
1059 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1060 * destruction. The caller should be holding the matching pool->lock.
1062 static void put_pwq(struct pool_workqueue *pwq)
1064 lockdep_assert_held(&pwq->pool->lock);
1065 if (likely(--pwq->refcnt))
1067 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1070 * @pwq can't be released under pool->lock, bounce to
1071 * pwq_unbound_release_workfn(). This never recurses on the same
1072 * pool->lock as this path is taken only for unbound workqueues and
1073 * the release work item is scheduled on a per-cpu workqueue. To
1074 * avoid lockdep warning, unbound pool->locks are given lockdep
1075 * subclass of 1 in get_unbound_pool().
1077 schedule_work(&pwq->unbound_release_work);
1081 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1082 * @pwq: pool_workqueue to put (can be %NULL)
1084 * put_pwq() with locking. This function also allows %NULL @pwq.
1086 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1090 * As both pwqs and pools are sched-RCU protected, the
1091 * following lock operations are safe.
1093 spin_lock_irq(&pwq->pool->lock);
1095 spin_unlock_irq(&pwq->pool->lock);
1099 static void pwq_activate_delayed_work(struct work_struct *work)
1101 struct pool_workqueue *pwq = get_work_pwq(work);
1103 trace_workqueue_activate_work(work);
1104 if (list_empty(&pwq->pool->worklist))
1105 pwq->pool->watchdog_ts = jiffies;
1106 move_linked_works(work, &pwq->pool->worklist, NULL);
1107 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1111 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1113 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1114 struct work_struct, entry);
1116 pwq_activate_delayed_work(work);
1120 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1121 * @pwq: pwq of interest
1122 * @color: color of work which left the queue
1124 * A work either has completed or is removed from pending queue,
1125 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1128 * spin_lock_irq(pool->lock).
1130 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1132 /* uncolored work items don't participate in flushing or nr_active */
1133 if (color == WORK_NO_COLOR)
1136 pwq->nr_in_flight[color]--;
1139 if (!list_empty(&pwq->delayed_works)) {
1140 /* one down, submit a delayed one */
1141 if (pwq->nr_active < pwq->max_active)
1142 pwq_activate_first_delayed(pwq);
1145 /* is flush in progress and are we at the flushing tip? */
1146 if (likely(pwq->flush_color != color))
1149 /* are there still in-flight works? */
1150 if (pwq->nr_in_flight[color])
1153 /* this pwq is done, clear flush_color */
1154 pwq->flush_color = -1;
1157 * If this was the last pwq, wake up the first flusher. It
1158 * will handle the rest.
1160 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1161 complete(&pwq->wq->first_flusher->done);
1167 * try_to_grab_pending - steal work item from worklist and disable irq
1168 * @work: work item to steal
1169 * @is_dwork: @work is a delayed_work
1170 * @flags: place to store irq state
1172 * Try to grab PENDING bit of @work. This function can handle @work in any
1173 * stable state - idle, on timer or on worklist.
1176 * 1 if @work was pending and we successfully stole PENDING
1177 * 0 if @work was idle and we claimed PENDING
1178 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1179 * -ENOENT if someone else is canceling @work, this state may persist
1180 * for arbitrarily long
1183 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1184 * interrupted while holding PENDING and @work off queue, irq must be
1185 * disabled on entry. This, combined with delayed_work->timer being
1186 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1188 * On successful return, >= 0, irq is disabled and the caller is
1189 * responsible for releasing it using local_irq_restore(*@flags).
1191 * This function is safe to call from any context including IRQ handler.
1193 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1194 unsigned long *flags)
1196 struct worker_pool *pool;
1197 struct pool_workqueue *pwq;
1199 local_irq_save(*flags);
1201 /* try to steal the timer if it exists */
1203 struct delayed_work *dwork = to_delayed_work(work);
1206 * dwork->timer is irqsafe. If del_timer() fails, it's
1207 * guaranteed that the timer is not queued anywhere and not
1208 * running on the local CPU.
1210 if (likely(del_timer(&dwork->timer)))
1214 /* try to claim PENDING the normal way */
1215 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1219 * The queueing is in progress, or it is already queued. Try to
1220 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1222 pool = get_work_pool(work);
1226 spin_lock(&pool->lock);
1228 * work->data is guaranteed to point to pwq only while the work
1229 * item is queued on pwq->wq, and both updating work->data to point
1230 * to pwq on queueing and to pool on dequeueing are done under
1231 * pwq->pool->lock. This in turn guarantees that, if work->data
1232 * points to pwq which is associated with a locked pool, the work
1233 * item is currently queued on that pool.
1235 pwq = get_work_pwq(work);
1236 if (pwq && pwq->pool == pool) {
1237 debug_work_deactivate(work);
1240 * A delayed work item cannot be grabbed directly because
1241 * it might have linked NO_COLOR work items which, if left
1242 * on the delayed_list, will confuse pwq->nr_active
1243 * management later on and cause stall. Make sure the work
1244 * item is activated before grabbing.
1246 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1247 pwq_activate_delayed_work(work);
1249 list_del_init(&work->entry);
1250 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1252 /* work->data points to pwq iff queued, point to pool */
1253 set_work_pool_and_keep_pending(work, pool->id);
1255 spin_unlock(&pool->lock);
1258 spin_unlock(&pool->lock);
1260 local_irq_restore(*flags);
1261 if (work_is_canceling(work))
1268 * insert_work - insert a work into a pool
1269 * @pwq: pwq @work belongs to
1270 * @work: work to insert
1271 * @head: insertion point
1272 * @extra_flags: extra WORK_STRUCT_* flags to set
1274 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1275 * work_struct flags.
1278 * spin_lock_irq(pool->lock).
1280 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1281 struct list_head *head, unsigned int extra_flags)
1283 struct worker_pool *pool = pwq->pool;
1285 /* we own @work, set data and link */
1286 set_work_pwq(work, pwq, extra_flags);
1287 list_add_tail(&work->entry, head);
1291 * Ensure either wq_worker_sleeping() sees the above
1292 * list_add_tail() or we see zero nr_running to avoid workers lying
1293 * around lazily while there are works to be processed.
1297 if (__need_more_worker(pool))
1298 wake_up_worker(pool);
1302 * Test whether @work is being queued from another work executing on the
1305 static bool is_chained_work(struct workqueue_struct *wq)
1307 struct worker *worker;
1309 worker = current_wq_worker();
1311 * Return %true iff I'm a worker execuing a work item on @wq. If
1312 * I'm @worker, it's safe to dereference it without locking.
1314 return worker && worker->current_pwq->wq == wq;
1318 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1319 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1320 * avoid perturbing sensitive tasks.
1322 static int wq_select_unbound_cpu(int cpu)
1324 static bool printed_dbg_warning;
1327 if (likely(!wq_debug_force_rr_cpu)) {
1328 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1330 } else if (!printed_dbg_warning) {
1331 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1332 printed_dbg_warning = true;
1335 if (cpumask_empty(wq_unbound_cpumask))
1338 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1339 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1340 if (unlikely(new_cpu >= nr_cpu_ids)) {
1341 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1342 if (unlikely(new_cpu >= nr_cpu_ids))
1345 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1350 static void __queue_work(int cpu, struct workqueue_struct *wq,
1351 struct work_struct *work)
1353 struct pool_workqueue *pwq;
1354 struct worker_pool *last_pool;
1355 struct list_head *worklist;
1356 unsigned int work_flags;
1357 unsigned int req_cpu = cpu;
1360 * While a work item is PENDING && off queue, a task trying to
1361 * steal the PENDING will busy-loop waiting for it to either get
1362 * queued or lose PENDING. Grabbing PENDING and queueing should
1363 * happen with IRQ disabled.
1365 WARN_ON_ONCE(!irqs_disabled());
1367 debug_work_activate(work);
1369 /* if draining, only works from the same workqueue are allowed */
1370 if (unlikely(wq->flags & __WQ_DRAINING) &&
1371 WARN_ON_ONCE(!is_chained_work(wq)))
1374 if (req_cpu == WORK_CPU_UNBOUND)
1375 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1377 /* pwq which will be used unless @work is executing elsewhere */
1378 if (!(wq->flags & WQ_UNBOUND))
1379 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1381 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1384 * If @work was previously on a different pool, it might still be
1385 * running there, in which case the work needs to be queued on that
1386 * pool to guarantee non-reentrancy.
1388 last_pool = get_work_pool(work);
1389 if (last_pool && last_pool != pwq->pool) {
1390 struct worker *worker;
1392 spin_lock(&last_pool->lock);
1394 worker = find_worker_executing_work(last_pool, work);
1396 if (worker && worker->current_pwq->wq == wq) {
1397 pwq = worker->current_pwq;
1399 /* meh... not running there, queue here */
1400 spin_unlock(&last_pool->lock);
1401 spin_lock(&pwq->pool->lock);
1404 spin_lock(&pwq->pool->lock);
1408 * pwq is determined and locked. For unbound pools, we could have
1409 * raced with pwq release and it could already be dead. If its
1410 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1411 * without another pwq replacing it in the numa_pwq_tbl or while
1412 * work items are executing on it, so the retrying is guaranteed to
1413 * make forward-progress.
1415 if (unlikely(!pwq->refcnt)) {
1416 if (wq->flags & WQ_UNBOUND) {
1417 spin_unlock(&pwq->pool->lock);
1422 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1426 /* pwq determined, queue */
1427 trace_workqueue_queue_work(req_cpu, pwq, work);
1429 if (WARN_ON(!list_empty(&work->entry))) {
1430 spin_unlock(&pwq->pool->lock);
1434 pwq->nr_in_flight[pwq->work_color]++;
1435 work_flags = work_color_to_flags(pwq->work_color);
1437 if (likely(pwq->nr_active < pwq->max_active)) {
1438 trace_workqueue_activate_work(work);
1440 worklist = &pwq->pool->worklist;
1441 if (list_empty(worklist))
1442 pwq->pool->watchdog_ts = jiffies;
1444 work_flags |= WORK_STRUCT_DELAYED;
1445 worklist = &pwq->delayed_works;
1448 insert_work(pwq, work, worklist, work_flags);
1450 spin_unlock(&pwq->pool->lock);
1454 * queue_work_on - queue work on specific cpu
1455 * @cpu: CPU number to execute work on
1456 * @wq: workqueue to use
1457 * @work: work to queue
1459 * We queue the work to a specific CPU, the caller must ensure it
1462 * Return: %false if @work was already on a queue, %true otherwise.
1464 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1465 struct work_struct *work)
1468 unsigned long flags;
1470 local_irq_save(flags);
1472 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1473 __queue_work(cpu, wq, work);
1477 local_irq_restore(flags);
1480 EXPORT_SYMBOL(queue_work_on);
1482 void delayed_work_timer_fn(unsigned long __data)
1484 struct delayed_work *dwork = (struct delayed_work *)__data;
1486 /* should have been called from irqsafe timer with irq already off */
1487 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1489 EXPORT_SYMBOL(delayed_work_timer_fn);
1491 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1492 struct delayed_work *dwork, unsigned long delay)
1494 struct timer_list *timer = &dwork->timer;
1495 struct work_struct *work = &dwork->work;
1497 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1498 timer->data != (unsigned long)dwork);
1499 WARN_ON_ONCE(timer_pending(timer));
1500 WARN_ON_ONCE(!list_empty(&work->entry));
1503 * If @delay is 0, queue @dwork->work immediately. This is for
1504 * both optimization and correctness. The earliest @timer can
1505 * expire is on the closest next tick and delayed_work users depend
1506 * on that there's no such delay when @delay is 0.
1509 __queue_work(cpu, wq, &dwork->work);
1513 timer_stats_timer_set_start_info(&dwork->timer);
1517 timer->expires = jiffies + delay;
1519 if (unlikely(cpu != WORK_CPU_UNBOUND))
1520 add_timer_on(timer, cpu);
1526 * queue_delayed_work_on - queue work on specific CPU after delay
1527 * @cpu: CPU number to execute work on
1528 * @wq: workqueue to use
1529 * @dwork: work to queue
1530 * @delay: number of jiffies to wait before queueing
1532 * Return: %false if @work was already on a queue, %true otherwise. If
1533 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1536 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1537 struct delayed_work *dwork, unsigned long delay)
1539 struct work_struct *work = &dwork->work;
1541 unsigned long flags;
1543 /* read the comment in __queue_work() */
1544 local_irq_save(flags);
1546 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1547 __queue_delayed_work(cpu, wq, dwork, delay);
1551 local_irq_restore(flags);
1554 EXPORT_SYMBOL(queue_delayed_work_on);
1557 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1558 * @cpu: CPU number to execute work on
1559 * @wq: workqueue to use
1560 * @dwork: work to queue
1561 * @delay: number of jiffies to wait before queueing
1563 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1564 * modify @dwork's timer so that it expires after @delay. If @delay is
1565 * zero, @work is guaranteed to be scheduled immediately regardless of its
1568 * Return: %false if @dwork was idle and queued, %true if @dwork was
1569 * pending and its timer was modified.
1571 * This function is safe to call from any context including IRQ handler.
1572 * See try_to_grab_pending() for details.
1574 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1575 struct delayed_work *dwork, unsigned long delay)
1577 unsigned long flags;
1581 ret = try_to_grab_pending(&dwork->work, true, &flags);
1582 } while (unlikely(ret == -EAGAIN));
1584 if (likely(ret >= 0)) {
1585 __queue_delayed_work(cpu, wq, dwork, delay);
1586 local_irq_restore(flags);
1589 /* -ENOENT from try_to_grab_pending() becomes %true */
1592 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1595 * worker_enter_idle - enter idle state
1596 * @worker: worker which is entering idle state
1598 * @worker is entering idle state. Update stats and idle timer if
1602 * spin_lock_irq(pool->lock).
1604 static void worker_enter_idle(struct worker *worker)
1606 struct worker_pool *pool = worker->pool;
1608 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1609 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1610 (worker->hentry.next || worker->hentry.pprev)))
1613 /* can't use worker_set_flags(), also called from create_worker() */
1614 worker->flags |= WORKER_IDLE;
1616 worker->last_active = jiffies;
1618 /* idle_list is LIFO */
1619 list_add(&worker->entry, &pool->idle_list);
1621 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1622 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1625 * Sanity check nr_running. Because wq_unbind_fn() releases
1626 * pool->lock between setting %WORKER_UNBOUND and zapping
1627 * nr_running, the warning may trigger spuriously. Check iff
1628 * unbind is not in progress.
1630 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1631 pool->nr_workers == pool->nr_idle &&
1632 atomic_read(&pool->nr_running));
1636 * worker_leave_idle - leave idle state
1637 * @worker: worker which is leaving idle state
1639 * @worker is leaving idle state. Update stats.
1642 * spin_lock_irq(pool->lock).
1644 static void worker_leave_idle(struct worker *worker)
1646 struct worker_pool *pool = worker->pool;
1648 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1650 worker_clr_flags(worker, WORKER_IDLE);
1652 list_del_init(&worker->entry);
1655 static struct worker *alloc_worker(int node)
1657 struct worker *worker;
1659 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1661 INIT_LIST_HEAD(&worker->entry);
1662 INIT_LIST_HEAD(&worker->scheduled);
1663 INIT_LIST_HEAD(&worker->node);
1664 /* on creation a worker is in !idle && prep state */
1665 worker->flags = WORKER_PREP;
1671 * worker_attach_to_pool() - attach a worker to a pool
1672 * @worker: worker to be attached
1673 * @pool: the target pool
1675 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1676 * cpu-binding of @worker are kept coordinated with the pool across
1679 static void worker_attach_to_pool(struct worker *worker,
1680 struct worker_pool *pool)
1682 mutex_lock(&pool->attach_mutex);
1685 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1686 * online CPUs. It'll be re-applied when any of the CPUs come up.
1688 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1691 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1692 * stable across this function. See the comments above the
1693 * flag definition for details.
1695 if (pool->flags & POOL_DISASSOCIATED)
1696 worker->flags |= WORKER_UNBOUND;
1698 list_add_tail(&worker->node, &pool->workers);
1700 mutex_unlock(&pool->attach_mutex);
1704 * worker_detach_from_pool() - detach a worker from its pool
1705 * @worker: worker which is attached to its pool
1706 * @pool: the pool @worker is attached to
1708 * Undo the attaching which had been done in worker_attach_to_pool(). The
1709 * caller worker shouldn't access to the pool after detached except it has
1710 * other reference to the pool.
1712 static void worker_detach_from_pool(struct worker *worker,
1713 struct worker_pool *pool)
1715 struct completion *detach_completion = NULL;
1717 mutex_lock(&pool->attach_mutex);
1718 list_del(&worker->node);
1719 if (list_empty(&pool->workers))
1720 detach_completion = pool->detach_completion;
1721 mutex_unlock(&pool->attach_mutex);
1723 /* clear leftover flags without pool->lock after it is detached */
1724 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1726 if (detach_completion)
1727 complete(detach_completion);
1731 * create_worker - create a new workqueue worker
1732 * @pool: pool the new worker will belong to
1734 * Create and start a new worker which is attached to @pool.
1737 * Might sleep. Does GFP_KERNEL allocations.
1740 * Pointer to the newly created worker.
1742 static struct worker *create_worker(struct worker_pool *pool)
1744 struct worker *worker = NULL;
1748 /* ID is needed to determine kthread name */
1749 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1753 worker = alloc_worker(pool->node);
1757 worker->pool = pool;
1761 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1762 pool->attrs->nice < 0 ? "H" : "");
1764 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1766 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1767 "kworker/%s", id_buf);
1768 if (IS_ERR(worker->task))
1771 set_user_nice(worker->task, pool->attrs->nice);
1772 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1774 /* successful, attach the worker to the pool */
1775 worker_attach_to_pool(worker, pool);
1777 /* start the newly created worker */
1778 spin_lock_irq(&pool->lock);
1779 worker->pool->nr_workers++;
1780 worker_enter_idle(worker);
1781 wake_up_process(worker->task);
1782 spin_unlock_irq(&pool->lock);
1788 ida_simple_remove(&pool->worker_ida, id);
1794 * destroy_worker - destroy a workqueue worker
1795 * @worker: worker to be destroyed
1797 * Destroy @worker and adjust @pool stats accordingly. The worker should
1801 * spin_lock_irq(pool->lock).
1803 static void destroy_worker(struct worker *worker)
1805 struct worker_pool *pool = worker->pool;
1807 lockdep_assert_held(&pool->lock);
1809 /* sanity check frenzy */
1810 if (WARN_ON(worker->current_work) ||
1811 WARN_ON(!list_empty(&worker->scheduled)) ||
1812 WARN_ON(!(worker->flags & WORKER_IDLE)))
1818 list_del_init(&worker->entry);
1819 worker->flags |= WORKER_DIE;
1820 wake_up_process(worker->task);
1823 static void idle_worker_timeout(unsigned long __pool)
1825 struct worker_pool *pool = (void *)__pool;
1827 spin_lock_irq(&pool->lock);
1829 while (too_many_workers(pool)) {
1830 struct worker *worker;
1831 unsigned long expires;
1833 /* idle_list is kept in LIFO order, check the last one */
1834 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1835 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1837 if (time_before(jiffies, expires)) {
1838 mod_timer(&pool->idle_timer, expires);
1842 destroy_worker(worker);
1845 spin_unlock_irq(&pool->lock);
1848 static void send_mayday(struct work_struct *work)
1850 struct pool_workqueue *pwq = get_work_pwq(work);
1851 struct workqueue_struct *wq = pwq->wq;
1853 lockdep_assert_held(&wq_mayday_lock);
1858 /* mayday mayday mayday */
1859 if (list_empty(&pwq->mayday_node)) {
1861 * If @pwq is for an unbound wq, its base ref may be put at
1862 * any time due to an attribute change. Pin @pwq until the
1863 * rescuer is done with it.
1866 list_add_tail(&pwq->mayday_node, &wq->maydays);
1867 wake_up_process(wq->rescuer->task);
1871 static void pool_mayday_timeout(unsigned long __pool)
1873 struct worker_pool *pool = (void *)__pool;
1874 struct work_struct *work;
1876 spin_lock_irq(&pool->lock);
1877 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1879 if (need_to_create_worker(pool)) {
1881 * We've been trying to create a new worker but
1882 * haven't been successful. We might be hitting an
1883 * allocation deadlock. Send distress signals to
1886 list_for_each_entry(work, &pool->worklist, entry)
1890 spin_unlock(&wq_mayday_lock);
1891 spin_unlock_irq(&pool->lock);
1893 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1897 * maybe_create_worker - create a new worker if necessary
1898 * @pool: pool to create a new worker for
1900 * Create a new worker for @pool if necessary. @pool is guaranteed to
1901 * have at least one idle worker on return from this function. If
1902 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1903 * sent to all rescuers with works scheduled on @pool to resolve
1904 * possible allocation deadlock.
1906 * On return, need_to_create_worker() is guaranteed to be %false and
1907 * may_start_working() %true.
1910 * spin_lock_irq(pool->lock) which may be released and regrabbed
1911 * multiple times. Does GFP_KERNEL allocations. Called only from
1914 static void maybe_create_worker(struct worker_pool *pool)
1915 __releases(&pool->lock)
1916 __acquires(&pool->lock)
1919 spin_unlock_irq(&pool->lock);
1921 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1922 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1925 if (create_worker(pool) || !need_to_create_worker(pool))
1928 schedule_timeout_interruptible(CREATE_COOLDOWN);
1930 if (!need_to_create_worker(pool))
1934 del_timer_sync(&pool->mayday_timer);
1935 spin_lock_irq(&pool->lock);
1937 * This is necessary even after a new worker was just successfully
1938 * created as @pool->lock was dropped and the new worker might have
1939 * already become busy.
1941 if (need_to_create_worker(pool))
1946 * manage_workers - manage worker pool
1949 * Assume the manager role and manage the worker pool @worker belongs
1950 * to. At any given time, there can be only zero or one manager per
1951 * pool. The exclusion is handled automatically by this function.
1953 * The caller can safely start processing works on false return. On
1954 * true return, it's guaranteed that need_to_create_worker() is false
1955 * and may_start_working() is true.
1958 * spin_lock_irq(pool->lock) which may be released and regrabbed
1959 * multiple times. Does GFP_KERNEL allocations.
1962 * %false if the pool doesn't need management and the caller can safely
1963 * start processing works, %true if management function was performed and
1964 * the conditions that the caller verified before calling the function may
1965 * no longer be true.
1967 static bool manage_workers(struct worker *worker)
1969 struct worker_pool *pool = worker->pool;
1972 * Anyone who successfully grabs manager_arb wins the arbitration
1973 * and becomes the manager. mutex_trylock() on pool->manager_arb
1974 * failure while holding pool->lock reliably indicates that someone
1975 * else is managing the pool and the worker which failed trylock
1976 * can proceed to executing work items. This means that anyone
1977 * grabbing manager_arb is responsible for actually performing
1978 * manager duties. If manager_arb is grabbed and released without
1979 * actual management, the pool may stall indefinitely.
1981 if (!mutex_trylock(&pool->manager_arb))
1983 pool->manager = worker;
1985 maybe_create_worker(pool);
1987 pool->manager = NULL;
1988 mutex_unlock(&pool->manager_arb);
1993 * process_one_work - process single work
1995 * @work: work to process
1997 * Process @work. This function contains all the logics necessary to
1998 * process a single work including synchronization against and
1999 * interaction with other workers on the same cpu, queueing and
2000 * flushing. As long as context requirement is met, any worker can
2001 * call this function to process a work.
2004 * spin_lock_irq(pool->lock) which is released and regrabbed.
2006 static void process_one_work(struct worker *worker, struct work_struct *work)
2007 __releases(&pool->lock)
2008 __acquires(&pool->lock)
2010 struct pool_workqueue *pwq = get_work_pwq(work);
2011 struct worker_pool *pool = worker->pool;
2012 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2014 struct worker *collision;
2015 #ifdef CONFIG_LOCKDEP
2017 * It is permissible to free the struct work_struct from
2018 * inside the function that is called from it, this we need to
2019 * take into account for lockdep too. To avoid bogus "held
2020 * lock freed" warnings as well as problems when looking into
2021 * work->lockdep_map, make a copy and use that here.
2023 struct lockdep_map lockdep_map;
2025 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2027 /* ensure we're on the correct CPU */
2028 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2029 raw_smp_processor_id() != pool->cpu);
2032 * A single work shouldn't be executed concurrently by
2033 * multiple workers on a single cpu. Check whether anyone is
2034 * already processing the work. If so, defer the work to the
2035 * currently executing one.
2037 collision = find_worker_executing_work(pool, work);
2038 if (unlikely(collision)) {
2039 move_linked_works(work, &collision->scheduled, NULL);
2043 /* claim and dequeue */
2044 debug_work_deactivate(work);
2045 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2046 worker->current_work = work;
2047 worker->current_func = work->func;
2048 worker->current_pwq = pwq;
2049 work_color = get_work_color(work);
2051 list_del_init(&work->entry);
2054 * CPU intensive works don't participate in concurrency management.
2055 * They're the scheduler's responsibility. This takes @worker out
2056 * of concurrency management and the next code block will chain
2057 * execution of the pending work items.
2059 if (unlikely(cpu_intensive))
2060 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2063 * Wake up another worker if necessary. The condition is always
2064 * false for normal per-cpu workers since nr_running would always
2065 * be >= 1 at this point. This is used to chain execution of the
2066 * pending work items for WORKER_NOT_RUNNING workers such as the
2067 * UNBOUND and CPU_INTENSIVE ones.
2069 if (need_more_worker(pool))
2070 wake_up_worker(pool);
2073 * Record the last pool and clear PENDING which should be the last
2074 * update to @work. Also, do this inside @pool->lock so that
2075 * PENDING and queued state changes happen together while IRQ is
2078 set_work_pool_and_clear_pending(work, pool->id);
2080 spin_unlock_irq(&pool->lock);
2082 lock_map_acquire_read(&pwq->wq->lockdep_map);
2083 lock_map_acquire(&lockdep_map);
2084 trace_workqueue_execute_start(work);
2085 worker->current_func(work);
2087 * While we must be careful to not use "work" after this, the trace
2088 * point will only record its address.
2090 trace_workqueue_execute_end(work);
2091 lock_map_release(&lockdep_map);
2092 lock_map_release(&pwq->wq->lockdep_map);
2094 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2095 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2096 " last function: %pf\n",
2097 current->comm, preempt_count(), task_pid_nr(current),
2098 worker->current_func);
2099 debug_show_held_locks(current);
2104 * The following prevents a kworker from hogging CPU on !PREEMPT
2105 * kernels, where a requeueing work item waiting for something to
2106 * happen could deadlock with stop_machine as such work item could
2107 * indefinitely requeue itself while all other CPUs are trapped in
2108 * stop_machine. At the same time, report a quiescent RCU state so
2109 * the same condition doesn't freeze RCU.
2111 cond_resched_rcu_qs();
2113 spin_lock_irq(&pool->lock);
2115 /* clear cpu intensive status */
2116 if (unlikely(cpu_intensive))
2117 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2119 /* we're done with it, release */
2120 hash_del(&worker->hentry);
2121 worker->current_work = NULL;
2122 worker->current_func = NULL;
2123 worker->current_pwq = NULL;
2124 worker->desc_valid = false;
2125 pwq_dec_nr_in_flight(pwq, work_color);
2129 * process_scheduled_works - process scheduled works
2132 * Process all scheduled works. Please note that the scheduled list
2133 * may change while processing a work, so this function repeatedly
2134 * fetches a work from the top and executes it.
2137 * spin_lock_irq(pool->lock) which may be released and regrabbed
2140 static void process_scheduled_works(struct worker *worker)
2142 while (!list_empty(&worker->scheduled)) {
2143 struct work_struct *work = list_first_entry(&worker->scheduled,
2144 struct work_struct, entry);
2145 process_one_work(worker, work);
2150 * worker_thread - the worker thread function
2153 * The worker thread function. All workers belong to a worker_pool -
2154 * either a per-cpu one or dynamic unbound one. These workers process all
2155 * work items regardless of their specific target workqueue. The only
2156 * exception is work items which belong to workqueues with a rescuer which
2157 * will be explained in rescuer_thread().
2161 static int worker_thread(void *__worker)
2163 struct worker *worker = __worker;
2164 struct worker_pool *pool = worker->pool;
2166 /* tell the scheduler that this is a workqueue worker */
2167 worker->task->flags |= PF_WQ_WORKER;
2169 spin_lock_irq(&pool->lock);
2171 /* am I supposed to die? */
2172 if (unlikely(worker->flags & WORKER_DIE)) {
2173 spin_unlock_irq(&pool->lock);
2174 WARN_ON_ONCE(!list_empty(&worker->entry));
2175 worker->task->flags &= ~PF_WQ_WORKER;
2177 set_task_comm(worker->task, "kworker/dying");
2178 ida_simple_remove(&pool->worker_ida, worker->id);
2179 worker_detach_from_pool(worker, pool);
2184 worker_leave_idle(worker);
2186 /* no more worker necessary? */
2187 if (!need_more_worker(pool))
2190 /* do we need to manage? */
2191 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2195 * ->scheduled list can only be filled while a worker is
2196 * preparing to process a work or actually processing it.
2197 * Make sure nobody diddled with it while I was sleeping.
2199 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2202 * Finish PREP stage. We're guaranteed to have at least one idle
2203 * worker or that someone else has already assumed the manager
2204 * role. This is where @worker starts participating in concurrency
2205 * management if applicable and concurrency management is restored
2206 * after being rebound. See rebind_workers() for details.
2208 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2211 struct work_struct *work =
2212 list_first_entry(&pool->worklist,
2213 struct work_struct, entry);
2215 pool->watchdog_ts = jiffies;
2217 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2218 /* optimization path, not strictly necessary */
2219 process_one_work(worker, work);
2220 if (unlikely(!list_empty(&worker->scheduled)))
2221 process_scheduled_works(worker);
2223 move_linked_works(work, &worker->scheduled, NULL);
2224 process_scheduled_works(worker);
2226 } while (keep_working(pool));
2228 worker_set_flags(worker, WORKER_PREP);
2231 * pool->lock is held and there's no work to process and no need to
2232 * manage, sleep. Workers are woken up only while holding
2233 * pool->lock or from local cpu, so setting the current state
2234 * before releasing pool->lock is enough to prevent losing any
2237 worker_enter_idle(worker);
2238 __set_current_state(TASK_INTERRUPTIBLE);
2239 spin_unlock_irq(&pool->lock);
2245 * rescuer_thread - the rescuer thread function
2248 * Workqueue rescuer thread function. There's one rescuer for each
2249 * workqueue which has WQ_MEM_RECLAIM set.
2251 * Regular work processing on a pool may block trying to create a new
2252 * worker which uses GFP_KERNEL allocation which has slight chance of
2253 * developing into deadlock if some works currently on the same queue
2254 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2255 * the problem rescuer solves.
2257 * When such condition is possible, the pool summons rescuers of all
2258 * workqueues which have works queued on the pool and let them process
2259 * those works so that forward progress can be guaranteed.
2261 * This should happen rarely.
2265 static int rescuer_thread(void *__rescuer)
2267 struct worker *rescuer = __rescuer;
2268 struct workqueue_struct *wq = rescuer->rescue_wq;
2269 struct list_head *scheduled = &rescuer->scheduled;
2272 set_user_nice(current, RESCUER_NICE_LEVEL);
2275 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2276 * doesn't participate in concurrency management.
2278 rescuer->task->flags |= PF_WQ_WORKER;
2280 set_current_state(TASK_INTERRUPTIBLE);
2283 * By the time the rescuer is requested to stop, the workqueue
2284 * shouldn't have any work pending, but @wq->maydays may still have
2285 * pwq(s) queued. This can happen by non-rescuer workers consuming
2286 * all the work items before the rescuer got to them. Go through
2287 * @wq->maydays processing before acting on should_stop so that the
2288 * list is always empty on exit.
2290 should_stop = kthread_should_stop();
2292 /* see whether any pwq is asking for help */
2293 spin_lock_irq(&wq_mayday_lock);
2295 while (!list_empty(&wq->maydays)) {
2296 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2297 struct pool_workqueue, mayday_node);
2298 struct worker_pool *pool = pwq->pool;
2299 struct work_struct *work, *n;
2302 __set_current_state(TASK_RUNNING);
2303 list_del_init(&pwq->mayday_node);
2305 spin_unlock_irq(&wq_mayday_lock);
2307 worker_attach_to_pool(rescuer, pool);
2309 spin_lock_irq(&pool->lock);
2310 rescuer->pool = pool;
2313 * Slurp in all works issued via this workqueue and
2316 WARN_ON_ONCE(!list_empty(scheduled));
2317 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2318 if (get_work_pwq(work) == pwq) {
2320 pool->watchdog_ts = jiffies;
2321 move_linked_works(work, scheduled, &n);
2326 if (!list_empty(scheduled)) {
2327 process_scheduled_works(rescuer);
2330 * The above execution of rescued work items could
2331 * have created more to rescue through
2332 * pwq_activate_first_delayed() or chained
2333 * queueing. Let's put @pwq back on mayday list so
2334 * that such back-to-back work items, which may be
2335 * being used to relieve memory pressure, don't
2336 * incur MAYDAY_INTERVAL delay inbetween.
2338 if (need_to_create_worker(pool)) {
2339 spin_lock(&wq_mayday_lock);
2341 list_move_tail(&pwq->mayday_node, &wq->maydays);
2342 spin_unlock(&wq_mayday_lock);
2347 * Put the reference grabbed by send_mayday(). @pool won't
2348 * go away while we're still attached to it.
2353 * Leave this pool. If need_more_worker() is %true, notify a
2354 * regular worker; otherwise, we end up with 0 concurrency
2355 * and stalling the execution.
2357 if (need_more_worker(pool))
2358 wake_up_worker(pool);
2360 rescuer->pool = NULL;
2361 spin_unlock_irq(&pool->lock);
2363 worker_detach_from_pool(rescuer, pool);
2365 spin_lock_irq(&wq_mayday_lock);
2368 spin_unlock_irq(&wq_mayday_lock);
2371 __set_current_state(TASK_RUNNING);
2372 rescuer->task->flags &= ~PF_WQ_WORKER;
2376 /* rescuers should never participate in concurrency management */
2377 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2383 * check_flush_dependency - check for flush dependency sanity
2384 * @target_wq: workqueue being flushed
2385 * @target_work: work item being flushed (NULL for workqueue flushes)
2387 * %current is trying to flush the whole @target_wq or @target_work on it.
2388 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2389 * reclaiming memory or running on a workqueue which doesn't have
2390 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2393 static void check_flush_dependency(struct workqueue_struct *target_wq,
2394 struct work_struct *target_work)
2396 work_func_t target_func = target_work ? target_work->func : NULL;
2397 struct worker *worker;
2399 if (target_wq->flags & WQ_MEM_RECLAIM)
2402 worker = current_wq_worker();
2404 WARN_ONCE(current->flags & PF_MEMALLOC,
2405 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2406 current->pid, current->comm, target_wq->name, target_func);
2407 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2408 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2409 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2410 worker->current_pwq->wq->name, worker->current_func,
2411 target_wq->name, target_func);
2415 struct work_struct work;
2416 struct completion done;
2417 struct task_struct *task; /* purely informational */
2420 static void wq_barrier_func(struct work_struct *work)
2422 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2423 complete(&barr->done);
2427 * insert_wq_barrier - insert a barrier work
2428 * @pwq: pwq to insert barrier into
2429 * @barr: wq_barrier to insert
2430 * @target: target work to attach @barr to
2431 * @worker: worker currently executing @target, NULL if @target is not executing
2433 * @barr is linked to @target such that @barr is completed only after
2434 * @target finishes execution. Please note that the ordering
2435 * guarantee is observed only with respect to @target and on the local
2438 * Currently, a queued barrier can't be canceled. This is because
2439 * try_to_grab_pending() can't determine whether the work to be
2440 * grabbed is at the head of the queue and thus can't clear LINKED
2441 * flag of the previous work while there must be a valid next work
2442 * after a work with LINKED flag set.
2444 * Note that when @worker is non-NULL, @target may be modified
2445 * underneath us, so we can't reliably determine pwq from @target.
2448 * spin_lock_irq(pool->lock).
2450 static void insert_wq_barrier(struct pool_workqueue *pwq,
2451 struct wq_barrier *barr,
2452 struct work_struct *target, struct worker *worker)
2454 struct list_head *head;
2455 unsigned int linked = 0;
2458 * debugobject calls are safe here even with pool->lock locked
2459 * as we know for sure that this will not trigger any of the
2460 * checks and call back into the fixup functions where we
2463 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2464 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2465 init_completion(&barr->done);
2466 barr->task = current;
2469 * If @target is currently being executed, schedule the
2470 * barrier to the worker; otherwise, put it after @target.
2473 head = worker->scheduled.next;
2475 unsigned long *bits = work_data_bits(target);
2477 head = target->entry.next;
2478 /* there can already be other linked works, inherit and set */
2479 linked = *bits & WORK_STRUCT_LINKED;
2480 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2483 debug_work_activate(&barr->work);
2484 insert_work(pwq, &barr->work, head,
2485 work_color_to_flags(WORK_NO_COLOR) | linked);
2489 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2490 * @wq: workqueue being flushed
2491 * @flush_color: new flush color, < 0 for no-op
2492 * @work_color: new work color, < 0 for no-op
2494 * Prepare pwqs for workqueue flushing.
2496 * If @flush_color is non-negative, flush_color on all pwqs should be
2497 * -1. If no pwq has in-flight commands at the specified color, all
2498 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2499 * has in flight commands, its pwq->flush_color is set to
2500 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2501 * wakeup logic is armed and %true is returned.
2503 * The caller should have initialized @wq->first_flusher prior to
2504 * calling this function with non-negative @flush_color. If
2505 * @flush_color is negative, no flush color update is done and %false
2508 * If @work_color is non-negative, all pwqs should have the same
2509 * work_color which is previous to @work_color and all will be
2510 * advanced to @work_color.
2513 * mutex_lock(wq->mutex).
2516 * %true if @flush_color >= 0 and there's something to flush. %false
2519 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2520 int flush_color, int work_color)
2523 struct pool_workqueue *pwq;
2525 if (flush_color >= 0) {
2526 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2527 atomic_set(&wq->nr_pwqs_to_flush, 1);
2530 for_each_pwq(pwq, wq) {
2531 struct worker_pool *pool = pwq->pool;
2533 spin_lock_irq(&pool->lock);
2535 if (flush_color >= 0) {
2536 WARN_ON_ONCE(pwq->flush_color != -1);
2538 if (pwq->nr_in_flight[flush_color]) {
2539 pwq->flush_color = flush_color;
2540 atomic_inc(&wq->nr_pwqs_to_flush);
2545 if (work_color >= 0) {
2546 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2547 pwq->work_color = work_color;
2550 spin_unlock_irq(&pool->lock);
2553 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2554 complete(&wq->first_flusher->done);
2560 * flush_workqueue - ensure that any scheduled work has run to completion.
2561 * @wq: workqueue to flush
2563 * This function sleeps until all work items which were queued on entry
2564 * have finished execution, but it is not livelocked by new incoming ones.
2566 void flush_workqueue(struct workqueue_struct *wq)
2568 struct wq_flusher this_flusher = {
2569 .list = LIST_HEAD_INIT(this_flusher.list),
2571 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2575 lock_map_acquire(&wq->lockdep_map);
2576 lock_map_release(&wq->lockdep_map);
2578 mutex_lock(&wq->mutex);
2581 * Start-to-wait phase
2583 next_color = work_next_color(wq->work_color);
2585 if (next_color != wq->flush_color) {
2587 * Color space is not full. The current work_color
2588 * becomes our flush_color and work_color is advanced
2591 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2592 this_flusher.flush_color = wq->work_color;
2593 wq->work_color = next_color;
2595 if (!wq->first_flusher) {
2596 /* no flush in progress, become the first flusher */
2597 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2599 wq->first_flusher = &this_flusher;
2601 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2603 /* nothing to flush, done */
2604 wq->flush_color = next_color;
2605 wq->first_flusher = NULL;
2610 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2611 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2612 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2616 * Oops, color space is full, wait on overflow queue.
2617 * The next flush completion will assign us
2618 * flush_color and transfer to flusher_queue.
2620 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2623 check_flush_dependency(wq, NULL);
2625 mutex_unlock(&wq->mutex);
2627 wait_for_completion(&this_flusher.done);
2630 * Wake-up-and-cascade phase
2632 * First flushers are responsible for cascading flushes and
2633 * handling overflow. Non-first flushers can simply return.
2635 if (wq->first_flusher != &this_flusher)
2638 mutex_lock(&wq->mutex);
2640 /* we might have raced, check again with mutex held */
2641 if (wq->first_flusher != &this_flusher)
2644 wq->first_flusher = NULL;
2646 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2647 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2650 struct wq_flusher *next, *tmp;
2652 /* complete all the flushers sharing the current flush color */
2653 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2654 if (next->flush_color != wq->flush_color)
2656 list_del_init(&next->list);
2657 complete(&next->done);
2660 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2661 wq->flush_color != work_next_color(wq->work_color));
2663 /* this flush_color is finished, advance by one */
2664 wq->flush_color = work_next_color(wq->flush_color);
2666 /* one color has been freed, handle overflow queue */
2667 if (!list_empty(&wq->flusher_overflow)) {
2669 * Assign the same color to all overflowed
2670 * flushers, advance work_color and append to
2671 * flusher_queue. This is the start-to-wait
2672 * phase for these overflowed flushers.
2674 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2675 tmp->flush_color = wq->work_color;
2677 wq->work_color = work_next_color(wq->work_color);
2679 list_splice_tail_init(&wq->flusher_overflow,
2680 &wq->flusher_queue);
2681 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2684 if (list_empty(&wq->flusher_queue)) {
2685 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2690 * Need to flush more colors. Make the next flusher
2691 * the new first flusher and arm pwqs.
2693 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2694 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2696 list_del_init(&next->list);
2697 wq->first_flusher = next;
2699 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2703 * Meh... this color is already done, clear first
2704 * flusher and repeat cascading.
2706 wq->first_flusher = NULL;
2710 mutex_unlock(&wq->mutex);
2712 EXPORT_SYMBOL(flush_workqueue);
2715 * drain_workqueue - drain a workqueue
2716 * @wq: workqueue to drain
2718 * Wait until the workqueue becomes empty. While draining is in progress,
2719 * only chain queueing is allowed. IOW, only currently pending or running
2720 * work items on @wq can queue further work items on it. @wq is flushed
2721 * repeatedly until it becomes empty. The number of flushing is determined
2722 * by the depth of chaining and should be relatively short. Whine if it
2725 void drain_workqueue(struct workqueue_struct *wq)
2727 unsigned int flush_cnt = 0;
2728 struct pool_workqueue *pwq;
2731 * __queue_work() needs to test whether there are drainers, is much
2732 * hotter than drain_workqueue() and already looks at @wq->flags.
2733 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2735 mutex_lock(&wq->mutex);
2736 if (!wq->nr_drainers++)
2737 wq->flags |= __WQ_DRAINING;
2738 mutex_unlock(&wq->mutex);
2740 flush_workqueue(wq);
2742 mutex_lock(&wq->mutex);
2744 for_each_pwq(pwq, wq) {
2747 spin_lock_irq(&pwq->pool->lock);
2748 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2749 spin_unlock_irq(&pwq->pool->lock);
2754 if (++flush_cnt == 10 ||
2755 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2756 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2757 wq->name, flush_cnt);
2759 mutex_unlock(&wq->mutex);
2763 if (!--wq->nr_drainers)
2764 wq->flags &= ~__WQ_DRAINING;
2765 mutex_unlock(&wq->mutex);
2767 EXPORT_SYMBOL_GPL(drain_workqueue);
2769 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2771 struct worker *worker = NULL;
2772 struct worker_pool *pool;
2773 struct pool_workqueue *pwq;
2777 local_irq_disable();
2778 pool = get_work_pool(work);
2784 spin_lock(&pool->lock);
2785 /* see the comment in try_to_grab_pending() with the same code */
2786 pwq = get_work_pwq(work);
2788 if (unlikely(pwq->pool != pool))
2791 worker = find_worker_executing_work(pool, work);
2794 pwq = worker->current_pwq;
2797 check_flush_dependency(pwq->wq, work);
2799 insert_wq_barrier(pwq, barr, work, worker);
2800 spin_unlock_irq(&pool->lock);
2803 * If @max_active is 1 or rescuer is in use, flushing another work
2804 * item on the same workqueue may lead to deadlock. Make sure the
2805 * flusher is not running on the same workqueue by verifying write
2808 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2809 lock_map_acquire(&pwq->wq->lockdep_map);
2811 lock_map_acquire_read(&pwq->wq->lockdep_map);
2812 lock_map_release(&pwq->wq->lockdep_map);
2816 spin_unlock_irq(&pool->lock);
2821 * flush_work - wait for a work to finish executing the last queueing instance
2822 * @work: the work to flush
2824 * Wait until @work has finished execution. @work is guaranteed to be idle
2825 * on return if it hasn't been requeued since flush started.
2828 * %true if flush_work() waited for the work to finish execution,
2829 * %false if it was already idle.
2831 bool flush_work(struct work_struct *work)
2833 struct wq_barrier barr;
2835 lock_map_acquire(&work->lockdep_map);
2836 lock_map_release(&work->lockdep_map);
2838 if (start_flush_work(work, &barr)) {
2839 wait_for_completion(&barr.done);
2840 destroy_work_on_stack(&barr.work);
2846 EXPORT_SYMBOL_GPL(flush_work);
2850 struct work_struct *work;
2853 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2855 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2857 if (cwait->work != key)
2859 return autoremove_wake_function(wait, mode, sync, key);
2862 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2864 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2865 unsigned long flags;
2869 ret = try_to_grab_pending(work, is_dwork, &flags);
2871 * If someone else is already canceling, wait for it to
2872 * finish. flush_work() doesn't work for PREEMPT_NONE
2873 * because we may get scheduled between @work's completion
2874 * and the other canceling task resuming and clearing
2875 * CANCELING - flush_work() will return false immediately
2876 * as @work is no longer busy, try_to_grab_pending() will
2877 * return -ENOENT as @work is still being canceled and the
2878 * other canceling task won't be able to clear CANCELING as
2879 * we're hogging the CPU.
2881 * Let's wait for completion using a waitqueue. As this
2882 * may lead to the thundering herd problem, use a custom
2883 * wake function which matches @work along with exclusive
2886 if (unlikely(ret == -ENOENT)) {
2887 struct cwt_wait cwait;
2889 init_wait(&cwait.wait);
2890 cwait.wait.func = cwt_wakefn;
2893 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2894 TASK_UNINTERRUPTIBLE);
2895 if (work_is_canceling(work))
2897 finish_wait(&cancel_waitq, &cwait.wait);
2899 } while (unlikely(ret < 0));
2901 /* tell other tasks trying to grab @work to back off */
2902 mark_work_canceling(work);
2903 local_irq_restore(flags);
2906 clear_work_data(work);
2909 * Paired with prepare_to_wait() above so that either
2910 * waitqueue_active() is visible here or !work_is_canceling() is
2914 if (waitqueue_active(&cancel_waitq))
2915 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2921 * cancel_work_sync - cancel a work and wait for it to finish
2922 * @work: the work to cancel
2924 * Cancel @work and wait for its execution to finish. This function
2925 * can be used even if the work re-queues itself or migrates to
2926 * another workqueue. On return from this function, @work is
2927 * guaranteed to be not pending or executing on any CPU.
2929 * cancel_work_sync(&delayed_work->work) must not be used for
2930 * delayed_work's. Use cancel_delayed_work_sync() instead.
2932 * The caller must ensure that the workqueue on which @work was last
2933 * queued can't be destroyed before this function returns.
2936 * %true if @work was pending, %false otherwise.
2938 bool cancel_work_sync(struct work_struct *work)
2940 return __cancel_work_timer(work, false);
2942 EXPORT_SYMBOL_GPL(cancel_work_sync);
2945 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2946 * @dwork: the delayed work to flush
2948 * Delayed timer is cancelled and the pending work is queued for
2949 * immediate execution. Like flush_work(), this function only
2950 * considers the last queueing instance of @dwork.
2953 * %true if flush_work() waited for the work to finish execution,
2954 * %false if it was already idle.
2956 bool flush_delayed_work(struct delayed_work *dwork)
2958 local_irq_disable();
2959 if (del_timer_sync(&dwork->timer))
2960 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2962 return flush_work(&dwork->work);
2964 EXPORT_SYMBOL(flush_delayed_work);
2967 * cancel_delayed_work - cancel a delayed work
2968 * @dwork: delayed_work to cancel
2970 * Kill off a pending delayed_work.
2972 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2976 * The work callback function may still be running on return, unless
2977 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2978 * use cancel_delayed_work_sync() to wait on it.
2980 * This function is safe to call from any context including IRQ handler.
2982 bool cancel_delayed_work(struct delayed_work *dwork)
2984 unsigned long flags;
2988 ret = try_to_grab_pending(&dwork->work, true, &flags);
2989 } while (unlikely(ret == -EAGAIN));
2991 if (unlikely(ret < 0))
2994 set_work_pool_and_clear_pending(&dwork->work,
2995 get_work_pool_id(&dwork->work));
2996 local_irq_restore(flags);
2999 EXPORT_SYMBOL(cancel_delayed_work);
3002 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3003 * @dwork: the delayed work cancel
3005 * This is cancel_work_sync() for delayed works.
3008 * %true if @dwork was pending, %false otherwise.
3010 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3012 return __cancel_work_timer(&dwork->work, true);
3014 EXPORT_SYMBOL(cancel_delayed_work_sync);
3017 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3018 * @func: the function to call
3020 * schedule_on_each_cpu() executes @func on each online CPU using the
3021 * system workqueue and blocks until all CPUs have completed.
3022 * schedule_on_each_cpu() is very slow.
3025 * 0 on success, -errno on failure.
3027 int schedule_on_each_cpu(work_func_t func)
3030 struct work_struct __percpu *works;
3032 works = alloc_percpu(struct work_struct);
3038 for_each_online_cpu(cpu) {
3039 struct work_struct *work = per_cpu_ptr(works, cpu);
3041 INIT_WORK(work, func);
3042 schedule_work_on(cpu, work);
3045 for_each_online_cpu(cpu)
3046 flush_work(per_cpu_ptr(works, cpu));
3054 * execute_in_process_context - reliably execute the routine with user context
3055 * @fn: the function to execute
3056 * @ew: guaranteed storage for the execute work structure (must
3057 * be available when the work executes)
3059 * Executes the function immediately if process context is available,
3060 * otherwise schedules the function for delayed execution.
3062 * Return: 0 - function was executed
3063 * 1 - function was scheduled for execution
3065 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3067 if (!in_interrupt()) {
3072 INIT_WORK(&ew->work, fn);
3073 schedule_work(&ew->work);
3077 EXPORT_SYMBOL_GPL(execute_in_process_context);
3080 * free_workqueue_attrs - free a workqueue_attrs
3081 * @attrs: workqueue_attrs to free
3083 * Undo alloc_workqueue_attrs().
3085 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3088 free_cpumask_var(attrs->cpumask);
3094 * alloc_workqueue_attrs - allocate a workqueue_attrs
3095 * @gfp_mask: allocation mask to use
3097 * Allocate a new workqueue_attrs, initialize with default settings and
3100 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3102 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3104 struct workqueue_attrs *attrs;
3106 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3109 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3112 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3115 free_workqueue_attrs(attrs);
3119 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3120 const struct workqueue_attrs *from)
3122 to->nice = from->nice;
3123 cpumask_copy(to->cpumask, from->cpumask);
3125 * Unlike hash and equality test, this function doesn't ignore
3126 * ->no_numa as it is used for both pool and wq attrs. Instead,
3127 * get_unbound_pool() explicitly clears ->no_numa after copying.
3129 to->no_numa = from->no_numa;
3132 /* hash value of the content of @attr */
3133 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3137 hash = jhash_1word(attrs->nice, hash);
3138 hash = jhash(cpumask_bits(attrs->cpumask),
3139 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3143 /* content equality test */
3144 static bool wqattrs_equal(const struct workqueue_attrs *a,
3145 const struct workqueue_attrs *b)
3147 if (a->nice != b->nice)
3149 if (!cpumask_equal(a->cpumask, b->cpumask))
3155 * init_worker_pool - initialize a newly zalloc'd worker_pool
3156 * @pool: worker_pool to initialize
3158 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3160 * Return: 0 on success, -errno on failure. Even on failure, all fields
3161 * inside @pool proper are initialized and put_unbound_pool() can be called
3162 * on @pool safely to release it.
3164 static int init_worker_pool(struct worker_pool *pool)
3166 spin_lock_init(&pool->lock);
3169 pool->node = NUMA_NO_NODE;
3170 pool->flags |= POOL_DISASSOCIATED;
3171 pool->watchdog_ts = jiffies;
3172 INIT_LIST_HEAD(&pool->worklist);
3173 INIT_LIST_HEAD(&pool->idle_list);
3174 hash_init(pool->busy_hash);
3176 init_timer_deferrable(&pool->idle_timer);
3177 pool->idle_timer.function = idle_worker_timeout;
3178 pool->idle_timer.data = (unsigned long)pool;
3180 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3181 (unsigned long)pool);
3183 mutex_init(&pool->manager_arb);
3184 mutex_init(&pool->attach_mutex);
3185 INIT_LIST_HEAD(&pool->workers);
3187 ida_init(&pool->worker_ida);
3188 INIT_HLIST_NODE(&pool->hash_node);
3191 /* shouldn't fail above this point */
3192 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3198 static void rcu_free_wq(struct rcu_head *rcu)
3200 struct workqueue_struct *wq =
3201 container_of(rcu, struct workqueue_struct, rcu);
3203 if (!(wq->flags & WQ_UNBOUND))
3204 free_percpu(wq->cpu_pwqs);
3206 free_workqueue_attrs(wq->unbound_attrs);
3212 static void rcu_free_pool(struct rcu_head *rcu)
3214 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3216 ida_destroy(&pool->worker_ida);
3217 free_workqueue_attrs(pool->attrs);
3222 * put_unbound_pool - put a worker_pool
3223 * @pool: worker_pool to put
3225 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3226 * safe manner. get_unbound_pool() calls this function on its failure path
3227 * and this function should be able to release pools which went through,
3228 * successfully or not, init_worker_pool().
3230 * Should be called with wq_pool_mutex held.
3232 static void put_unbound_pool(struct worker_pool *pool)
3234 DECLARE_COMPLETION_ONSTACK(detach_completion);
3235 struct worker *worker;
3237 lockdep_assert_held(&wq_pool_mutex);
3243 if (WARN_ON(!(pool->cpu < 0)) ||
3244 WARN_ON(!list_empty(&pool->worklist)))
3247 /* release id and unhash */
3249 idr_remove(&worker_pool_idr, pool->id);
3250 hash_del(&pool->hash_node);
3253 * Become the manager and destroy all workers. Grabbing
3254 * manager_arb prevents @pool's workers from blocking on
3257 mutex_lock(&pool->manager_arb);
3259 spin_lock_irq(&pool->lock);
3260 while ((worker = first_idle_worker(pool)))
3261 destroy_worker(worker);
3262 WARN_ON(pool->nr_workers || pool->nr_idle);
3263 spin_unlock_irq(&pool->lock);
3265 mutex_lock(&pool->attach_mutex);
3266 if (!list_empty(&pool->workers))
3267 pool->detach_completion = &detach_completion;
3268 mutex_unlock(&pool->attach_mutex);
3270 if (pool->detach_completion)
3271 wait_for_completion(pool->detach_completion);
3273 mutex_unlock(&pool->manager_arb);
3275 /* shut down the timers */
3276 del_timer_sync(&pool->idle_timer);
3277 del_timer_sync(&pool->mayday_timer);
3279 /* sched-RCU protected to allow dereferences from get_work_pool() */
3280 call_rcu_sched(&pool->rcu, rcu_free_pool);
3284 * get_unbound_pool - get a worker_pool with the specified attributes
3285 * @attrs: the attributes of the worker_pool to get
3287 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3288 * reference count and return it. If there already is a matching
3289 * worker_pool, it will be used; otherwise, this function attempts to
3292 * Should be called with wq_pool_mutex held.
3294 * Return: On success, a worker_pool with the same attributes as @attrs.
3295 * On failure, %NULL.
3297 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3299 u32 hash = wqattrs_hash(attrs);
3300 struct worker_pool *pool;
3302 int target_node = NUMA_NO_NODE;
3304 lockdep_assert_held(&wq_pool_mutex);
3306 /* do we already have a matching pool? */
3307 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3308 if (wqattrs_equal(pool->attrs, attrs)) {
3314 /* if cpumask is contained inside a NUMA node, we belong to that node */
3315 if (wq_numa_enabled) {
3316 for_each_node(node) {
3317 if (cpumask_subset(attrs->cpumask,
3318 wq_numa_possible_cpumask[node])) {
3325 /* nope, create a new one */
3326 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3327 if (!pool || init_worker_pool(pool) < 0)
3330 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3331 copy_workqueue_attrs(pool->attrs, attrs);
3332 pool->node = target_node;
3335 * no_numa isn't a worker_pool attribute, always clear it. See
3336 * 'struct workqueue_attrs' comments for detail.
3338 pool->attrs->no_numa = false;
3340 if (worker_pool_assign_id(pool) < 0)
3343 /* create and start the initial worker */
3344 if (!create_worker(pool))
3348 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3353 put_unbound_pool(pool);
3357 static void rcu_free_pwq(struct rcu_head *rcu)
3359 kmem_cache_free(pwq_cache,
3360 container_of(rcu, struct pool_workqueue, rcu));
3364 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3365 * and needs to be destroyed.
3367 static void pwq_unbound_release_workfn(struct work_struct *work)
3369 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3370 unbound_release_work);
3371 struct workqueue_struct *wq = pwq->wq;
3372 struct worker_pool *pool = pwq->pool;
3375 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3378 mutex_lock(&wq->mutex);
3379 list_del_rcu(&pwq->pwqs_node);
3380 is_last = list_empty(&wq->pwqs);
3381 mutex_unlock(&wq->mutex);
3383 mutex_lock(&wq_pool_mutex);
3384 put_unbound_pool(pool);
3385 mutex_unlock(&wq_pool_mutex);
3387 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3390 * If we're the last pwq going away, @wq is already dead and no one
3391 * is gonna access it anymore. Schedule RCU free.
3394 call_rcu_sched(&wq->rcu, rcu_free_wq);
3398 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3399 * @pwq: target pool_workqueue
3401 * If @pwq isn't freezing, set @pwq->max_active to the associated
3402 * workqueue's saved_max_active and activate delayed work items
3403 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3405 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3407 struct workqueue_struct *wq = pwq->wq;
3408 bool freezable = wq->flags & WQ_FREEZABLE;
3410 /* for @wq->saved_max_active */
3411 lockdep_assert_held(&wq->mutex);
3413 /* fast exit for non-freezable wqs */
3414 if (!freezable && pwq->max_active == wq->saved_max_active)
3417 spin_lock_irq(&pwq->pool->lock);
3420 * During [un]freezing, the caller is responsible for ensuring that
3421 * this function is called at least once after @workqueue_freezing
3422 * is updated and visible.
3424 if (!freezable || !workqueue_freezing) {
3425 pwq->max_active = wq->saved_max_active;
3427 while (!list_empty(&pwq->delayed_works) &&
3428 pwq->nr_active < pwq->max_active)
3429 pwq_activate_first_delayed(pwq);
3432 * Need to kick a worker after thawed or an unbound wq's
3433 * max_active is bumped. It's a slow path. Do it always.
3435 wake_up_worker(pwq->pool);
3437 pwq->max_active = 0;
3440 spin_unlock_irq(&pwq->pool->lock);
3443 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3444 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3445 struct worker_pool *pool)
3447 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3449 memset(pwq, 0, sizeof(*pwq));
3453 pwq->flush_color = -1;
3455 INIT_LIST_HEAD(&pwq->delayed_works);
3456 INIT_LIST_HEAD(&pwq->pwqs_node);
3457 INIT_LIST_HEAD(&pwq->mayday_node);
3458 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3461 /* sync @pwq with the current state of its associated wq and link it */
3462 static void link_pwq(struct pool_workqueue *pwq)
3464 struct workqueue_struct *wq = pwq->wq;
3466 lockdep_assert_held(&wq->mutex);
3468 /* may be called multiple times, ignore if already linked */
3469 if (!list_empty(&pwq->pwqs_node))
3472 /* set the matching work_color */
3473 pwq->work_color = wq->work_color;
3475 /* sync max_active to the current setting */
3476 pwq_adjust_max_active(pwq);
3479 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3482 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3483 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3484 const struct workqueue_attrs *attrs)
3486 struct worker_pool *pool;
3487 struct pool_workqueue *pwq;
3489 lockdep_assert_held(&wq_pool_mutex);
3491 pool = get_unbound_pool(attrs);
3495 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3497 put_unbound_pool(pool);
3501 init_pwq(pwq, wq, pool);
3506 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3507 * @attrs: the wq_attrs of the default pwq of the target workqueue
3508 * @node: the target NUMA node
3509 * @cpu_going_down: if >= 0, the CPU to consider as offline
3510 * @cpumask: outarg, the resulting cpumask
3512 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3513 * @cpu_going_down is >= 0, that cpu is considered offline during
3514 * calculation. The result is stored in @cpumask.
3516 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3517 * enabled and @node has online CPUs requested by @attrs, the returned
3518 * cpumask is the intersection of the possible CPUs of @node and
3521 * The caller is responsible for ensuring that the cpumask of @node stays
3524 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3527 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3528 int cpu_going_down, cpumask_t *cpumask)
3530 if (!wq_numa_enabled || attrs->no_numa)
3533 /* does @node have any online CPUs @attrs wants? */
3534 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3535 if (cpu_going_down >= 0)
3536 cpumask_clear_cpu(cpu_going_down, cpumask);
3538 if (cpumask_empty(cpumask))
3541 /* yeap, return possible CPUs in @node that @attrs wants */
3542 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3543 return !cpumask_equal(cpumask, attrs->cpumask);
3546 cpumask_copy(cpumask, attrs->cpumask);
3550 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3551 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3553 struct pool_workqueue *pwq)
3555 struct pool_workqueue *old_pwq;
3557 lockdep_assert_held(&wq_pool_mutex);
3558 lockdep_assert_held(&wq->mutex);
3560 /* link_pwq() can handle duplicate calls */
3563 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3564 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3568 /* context to store the prepared attrs & pwqs before applying */
3569 struct apply_wqattrs_ctx {
3570 struct workqueue_struct *wq; /* target workqueue */
3571 struct workqueue_attrs *attrs; /* attrs to apply */
3572 struct list_head list; /* queued for batching commit */
3573 struct pool_workqueue *dfl_pwq;
3574 struct pool_workqueue *pwq_tbl[];
3577 /* free the resources after success or abort */
3578 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3584 put_pwq_unlocked(ctx->pwq_tbl[node]);
3585 put_pwq_unlocked(ctx->dfl_pwq);
3587 free_workqueue_attrs(ctx->attrs);
3593 /* allocate the attrs and pwqs for later installation */
3594 static struct apply_wqattrs_ctx *
3595 apply_wqattrs_prepare(struct workqueue_struct *wq,
3596 const struct workqueue_attrs *attrs)
3598 struct apply_wqattrs_ctx *ctx;
3599 struct workqueue_attrs *new_attrs, *tmp_attrs;
3602 lockdep_assert_held(&wq_pool_mutex);
3604 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3607 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3608 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3609 if (!ctx || !new_attrs || !tmp_attrs)
3613 * Calculate the attrs of the default pwq.
3614 * If the user configured cpumask doesn't overlap with the
3615 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3617 copy_workqueue_attrs(new_attrs, attrs);
3618 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3619 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3620 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3623 * We may create multiple pwqs with differing cpumasks. Make a
3624 * copy of @new_attrs which will be modified and used to obtain
3627 copy_workqueue_attrs(tmp_attrs, new_attrs);
3630 * If something goes wrong during CPU up/down, we'll fall back to
3631 * the default pwq covering whole @attrs->cpumask. Always create
3632 * it even if we don't use it immediately.
3634 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3638 for_each_node(node) {
3639 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3640 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3641 if (!ctx->pwq_tbl[node])
3644 ctx->dfl_pwq->refcnt++;
3645 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3649 /* save the user configured attrs and sanitize it. */
3650 copy_workqueue_attrs(new_attrs, attrs);
3651 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3652 ctx->attrs = new_attrs;
3655 free_workqueue_attrs(tmp_attrs);
3659 free_workqueue_attrs(tmp_attrs);
3660 free_workqueue_attrs(new_attrs);
3661 apply_wqattrs_cleanup(ctx);
3665 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3666 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3670 /* all pwqs have been created successfully, let's install'em */
3671 mutex_lock(&ctx->wq->mutex);
3673 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3675 /* save the previous pwq and install the new one */
3677 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3678 ctx->pwq_tbl[node]);
3680 /* @dfl_pwq might not have been used, ensure it's linked */
3681 link_pwq(ctx->dfl_pwq);
3682 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3684 mutex_unlock(&ctx->wq->mutex);
3687 static void apply_wqattrs_lock(void)
3689 /* CPUs should stay stable across pwq creations and installations */
3691 mutex_lock(&wq_pool_mutex);
3694 static void apply_wqattrs_unlock(void)
3696 mutex_unlock(&wq_pool_mutex);
3700 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3701 const struct workqueue_attrs *attrs)
3703 struct apply_wqattrs_ctx *ctx;
3705 /* only unbound workqueues can change attributes */
3706 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3709 /* creating multiple pwqs breaks ordering guarantee */
3710 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3713 ctx = apply_wqattrs_prepare(wq, attrs);
3717 /* the ctx has been prepared successfully, let's commit it */
3718 apply_wqattrs_commit(ctx);
3719 apply_wqattrs_cleanup(ctx);
3725 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3726 * @wq: the target workqueue
3727 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3729 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3730 * machines, this function maps a separate pwq to each NUMA node with
3731 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3732 * NUMA node it was issued on. Older pwqs are released as in-flight work
3733 * items finish. Note that a work item which repeatedly requeues itself
3734 * back-to-back will stay on its current pwq.
3736 * Performs GFP_KERNEL allocations.
3738 * Return: 0 on success and -errno on failure.
3740 int apply_workqueue_attrs(struct workqueue_struct *wq,
3741 const struct workqueue_attrs *attrs)
3745 apply_wqattrs_lock();
3746 ret = apply_workqueue_attrs_locked(wq, attrs);
3747 apply_wqattrs_unlock();
3753 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3754 * @wq: the target workqueue
3755 * @cpu: the CPU coming up or going down
3756 * @online: whether @cpu is coming up or going down
3758 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3759 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3762 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3763 * falls back to @wq->dfl_pwq which may not be optimal but is always
3766 * Note that when the last allowed CPU of a NUMA node goes offline for a
3767 * workqueue with a cpumask spanning multiple nodes, the workers which were
3768 * already executing the work items for the workqueue will lose their CPU
3769 * affinity and may execute on any CPU. This is similar to how per-cpu
3770 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3771 * affinity, it's the user's responsibility to flush the work item from
3774 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3777 int node = cpu_to_node(cpu);
3778 int cpu_off = online ? -1 : cpu;
3779 struct pool_workqueue *old_pwq = NULL, *pwq;
3780 struct workqueue_attrs *target_attrs;
3783 lockdep_assert_held(&wq_pool_mutex);
3785 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3786 wq->unbound_attrs->no_numa)
3790 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3791 * Let's use a preallocated one. The following buf is protected by
3792 * CPU hotplug exclusion.
3794 target_attrs = wq_update_unbound_numa_attrs_buf;
3795 cpumask = target_attrs->cpumask;
3797 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3798 pwq = unbound_pwq_by_node(wq, node);
3801 * Let's determine what needs to be done. If the target cpumask is
3802 * different from the default pwq's, we need to compare it to @pwq's
3803 * and create a new one if they don't match. If the target cpumask
3804 * equals the default pwq's, the default pwq should be used.
3806 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3807 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3813 /* create a new pwq */
3814 pwq = alloc_unbound_pwq(wq, target_attrs);
3816 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3821 /* Install the new pwq. */
3822 mutex_lock(&wq->mutex);
3823 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3827 mutex_lock(&wq->mutex);
3828 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3829 get_pwq(wq->dfl_pwq);
3830 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3831 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3833 mutex_unlock(&wq->mutex);
3834 put_pwq_unlocked(old_pwq);
3837 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3839 bool highpri = wq->flags & WQ_HIGHPRI;
3842 if (!(wq->flags & WQ_UNBOUND)) {
3843 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3847 for_each_possible_cpu(cpu) {
3848 struct pool_workqueue *pwq =
3849 per_cpu_ptr(wq->cpu_pwqs, cpu);
3850 struct worker_pool *cpu_pools =
3851 per_cpu(cpu_worker_pools, cpu);
3853 init_pwq(pwq, wq, &cpu_pools[highpri]);
3855 mutex_lock(&wq->mutex);
3857 mutex_unlock(&wq->mutex);
3860 } else if (wq->flags & __WQ_ORDERED) {
3861 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3862 /* there should only be single pwq for ordering guarantee */
3863 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3864 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3865 "ordering guarantee broken for workqueue %s\n", wq->name);
3868 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3872 static int wq_clamp_max_active(int max_active, unsigned int flags,
3875 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3877 if (max_active < 1 || max_active > lim)
3878 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3879 max_active, name, 1, lim);
3881 return clamp_val(max_active, 1, lim);
3884 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3887 struct lock_class_key *key,
3888 const char *lock_name, ...)
3890 size_t tbl_size = 0;
3892 struct workqueue_struct *wq;
3893 struct pool_workqueue *pwq;
3895 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3896 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3897 flags |= WQ_UNBOUND;
3899 /* allocate wq and format name */
3900 if (flags & WQ_UNBOUND)
3901 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3903 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3907 if (flags & WQ_UNBOUND) {
3908 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3909 if (!wq->unbound_attrs)
3913 va_start(args, lock_name);
3914 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3917 max_active = max_active ?: WQ_DFL_ACTIVE;
3918 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3922 wq->saved_max_active = max_active;
3923 mutex_init(&wq->mutex);
3924 atomic_set(&wq->nr_pwqs_to_flush, 0);
3925 INIT_LIST_HEAD(&wq->pwqs);
3926 INIT_LIST_HEAD(&wq->flusher_queue);
3927 INIT_LIST_HEAD(&wq->flusher_overflow);
3928 INIT_LIST_HEAD(&wq->maydays);
3930 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3931 INIT_LIST_HEAD(&wq->list);
3933 if (alloc_and_link_pwqs(wq) < 0)
3937 * Workqueues which may be used during memory reclaim should
3938 * have a rescuer to guarantee forward progress.
3940 if (flags & WQ_MEM_RECLAIM) {
3941 struct worker *rescuer;
3943 rescuer = alloc_worker(NUMA_NO_NODE);
3947 rescuer->rescue_wq = wq;
3948 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3950 if (IS_ERR(rescuer->task)) {
3955 wq->rescuer = rescuer;
3956 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3957 wake_up_process(rescuer->task);
3960 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3964 * wq_pool_mutex protects global freeze state and workqueues list.
3965 * Grab it, adjust max_active and add the new @wq to workqueues
3968 mutex_lock(&wq_pool_mutex);
3970 mutex_lock(&wq->mutex);
3971 for_each_pwq(pwq, wq)
3972 pwq_adjust_max_active(pwq);
3973 mutex_unlock(&wq->mutex);
3975 list_add_tail_rcu(&wq->list, &workqueues);
3977 mutex_unlock(&wq_pool_mutex);
3982 free_workqueue_attrs(wq->unbound_attrs);
3986 destroy_workqueue(wq);
3989 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3992 * destroy_workqueue - safely terminate a workqueue
3993 * @wq: target workqueue
3995 * Safely destroy a workqueue. All work currently pending will be done first.
3997 void destroy_workqueue(struct workqueue_struct *wq)
3999 struct pool_workqueue *pwq;
4002 /* drain it before proceeding with destruction */
4003 drain_workqueue(wq);
4006 mutex_lock(&wq->mutex);
4007 for_each_pwq(pwq, wq) {
4010 for (i = 0; i < WORK_NR_COLORS; i++) {
4011 if (WARN_ON(pwq->nr_in_flight[i])) {
4012 mutex_unlock(&wq->mutex);
4017 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4018 WARN_ON(pwq->nr_active) ||
4019 WARN_ON(!list_empty(&pwq->delayed_works))) {
4020 mutex_unlock(&wq->mutex);
4024 mutex_unlock(&wq->mutex);
4027 * wq list is used to freeze wq, remove from list after
4028 * flushing is complete in case freeze races us.
4030 mutex_lock(&wq_pool_mutex);
4031 list_del_rcu(&wq->list);
4032 mutex_unlock(&wq_pool_mutex);
4034 workqueue_sysfs_unregister(wq);
4037 kthread_stop(wq->rescuer->task);
4039 if (!(wq->flags & WQ_UNBOUND)) {
4041 * The base ref is never dropped on per-cpu pwqs. Directly
4042 * schedule RCU free.
4044 call_rcu_sched(&wq->rcu, rcu_free_wq);
4047 * We're the sole accessor of @wq at this point. Directly
4048 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4049 * @wq will be freed when the last pwq is released.
4051 for_each_node(node) {
4052 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4053 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4054 put_pwq_unlocked(pwq);
4058 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4059 * put. Don't access it afterwards.
4063 put_pwq_unlocked(pwq);
4066 EXPORT_SYMBOL_GPL(destroy_workqueue);
4069 * workqueue_set_max_active - adjust max_active of a workqueue
4070 * @wq: target workqueue
4071 * @max_active: new max_active value.
4073 * Set max_active of @wq to @max_active.
4076 * Don't call from IRQ context.
4078 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4080 struct pool_workqueue *pwq;
4082 /* disallow meddling with max_active for ordered workqueues */
4083 if (WARN_ON(wq->flags & __WQ_ORDERED))
4086 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4088 mutex_lock(&wq->mutex);
4090 wq->saved_max_active = max_active;
4092 for_each_pwq(pwq, wq)
4093 pwq_adjust_max_active(pwq);
4095 mutex_unlock(&wq->mutex);
4097 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4100 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4102 * Determine whether %current is a workqueue rescuer. Can be used from
4103 * work functions to determine whether it's being run off the rescuer task.
4105 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4107 bool current_is_workqueue_rescuer(void)
4109 struct worker *worker = current_wq_worker();
4111 return worker && worker->rescue_wq;
4115 * workqueue_congested - test whether a workqueue is congested
4116 * @cpu: CPU in question
4117 * @wq: target workqueue
4119 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4120 * no synchronization around this function and the test result is
4121 * unreliable and only useful as advisory hints or for debugging.
4123 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4124 * Note that both per-cpu and unbound workqueues may be associated with
4125 * multiple pool_workqueues which have separate congested states. A
4126 * workqueue being congested on one CPU doesn't mean the workqueue is also
4127 * contested on other CPUs / NUMA nodes.
4130 * %true if congested, %false otherwise.
4132 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4134 struct pool_workqueue *pwq;
4137 rcu_read_lock_sched();
4139 if (cpu == WORK_CPU_UNBOUND)
4140 cpu = smp_processor_id();
4142 if (!(wq->flags & WQ_UNBOUND))
4143 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4145 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4147 ret = !list_empty(&pwq->delayed_works);
4148 rcu_read_unlock_sched();
4152 EXPORT_SYMBOL_GPL(workqueue_congested);
4155 * work_busy - test whether a work is currently pending or running
4156 * @work: the work to be tested
4158 * Test whether @work is currently pending or running. There is no
4159 * synchronization around this function and the test result is
4160 * unreliable and only useful as advisory hints or for debugging.
4163 * OR'd bitmask of WORK_BUSY_* bits.
4165 unsigned int work_busy(struct work_struct *work)
4167 struct worker_pool *pool;
4168 unsigned long flags;
4169 unsigned int ret = 0;
4171 if (work_pending(work))
4172 ret |= WORK_BUSY_PENDING;
4174 local_irq_save(flags);
4175 pool = get_work_pool(work);
4177 spin_lock(&pool->lock);
4178 if (find_worker_executing_work(pool, work))
4179 ret |= WORK_BUSY_RUNNING;
4180 spin_unlock(&pool->lock);
4182 local_irq_restore(flags);
4186 EXPORT_SYMBOL_GPL(work_busy);
4189 * set_worker_desc - set description for the current work item
4190 * @fmt: printf-style format string
4191 * @...: arguments for the format string
4193 * This function can be called by a running work function to describe what
4194 * the work item is about. If the worker task gets dumped, this
4195 * information will be printed out together to help debugging. The
4196 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4198 void set_worker_desc(const char *fmt, ...)
4200 struct worker *worker = current_wq_worker();
4204 va_start(args, fmt);
4205 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4207 worker->desc_valid = true;
4212 * print_worker_info - print out worker information and description
4213 * @log_lvl: the log level to use when printing
4214 * @task: target task
4216 * If @task is a worker and currently executing a work item, print out the
4217 * name of the workqueue being serviced and worker description set with
4218 * set_worker_desc() by the currently executing work item.
4220 * This function can be safely called on any task as long as the
4221 * task_struct itself is accessible. While safe, this function isn't
4222 * synchronized and may print out mixups or garbages of limited length.
4224 void print_worker_info(const char *log_lvl, struct task_struct *task)
4226 work_func_t *fn = NULL;
4227 char name[WQ_NAME_LEN] = { };
4228 char desc[WORKER_DESC_LEN] = { };
4229 struct pool_workqueue *pwq = NULL;
4230 struct workqueue_struct *wq = NULL;
4231 bool desc_valid = false;
4232 struct worker *worker;
4234 if (!(task->flags & PF_WQ_WORKER))
4238 * This function is called without any synchronization and @task
4239 * could be in any state. Be careful with dereferences.
4241 worker = probe_kthread_data(task);
4244 * Carefully copy the associated workqueue's workfn and name. Keep
4245 * the original last '\0' in case the original contains garbage.
4247 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4248 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4249 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4250 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4252 /* copy worker description */
4253 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4255 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4257 if (fn || name[0] || desc[0]) {
4258 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4260 pr_cont(" (%s)", desc);
4265 static void pr_cont_pool_info(struct worker_pool *pool)
4267 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4268 if (pool->node != NUMA_NO_NODE)
4269 pr_cont(" node=%d", pool->node);
4270 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4273 static void pr_cont_work(bool comma, struct work_struct *work)
4275 if (work->func == wq_barrier_func) {
4276 struct wq_barrier *barr;
4278 barr = container_of(work, struct wq_barrier, work);
4280 pr_cont("%s BAR(%d)", comma ? "," : "",
4281 task_pid_nr(barr->task));
4283 pr_cont("%s %pf", comma ? "," : "", work->func);
4287 static void show_pwq(struct pool_workqueue *pwq)
4289 struct worker_pool *pool = pwq->pool;
4290 struct work_struct *work;
4291 struct worker *worker;
4292 bool has_in_flight = false, has_pending = false;
4295 pr_info(" pwq %d:", pool->id);
4296 pr_cont_pool_info(pool);
4298 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4299 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4301 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4302 if (worker->current_pwq == pwq) {
4303 has_in_flight = true;
4307 if (has_in_flight) {
4310 pr_info(" in-flight:");
4311 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4312 if (worker->current_pwq != pwq)
4315 pr_cont("%s %d%s:%pf", comma ? "," : "",
4316 task_pid_nr(worker->task),
4317 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4318 worker->current_func);
4319 list_for_each_entry(work, &worker->scheduled, entry)
4320 pr_cont_work(false, work);
4326 list_for_each_entry(work, &pool->worklist, entry) {
4327 if (get_work_pwq(work) == pwq) {
4335 pr_info(" pending:");
4336 list_for_each_entry(work, &pool->worklist, entry) {
4337 if (get_work_pwq(work) != pwq)
4340 pr_cont_work(comma, work);
4341 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4346 if (!list_empty(&pwq->delayed_works)) {
4349 pr_info(" delayed:");
4350 list_for_each_entry(work, &pwq->delayed_works, entry) {
4351 pr_cont_work(comma, work);
4352 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4359 * show_workqueue_state - dump workqueue state
4361 * Called from a sysrq handler and prints out all busy workqueues and
4364 void show_workqueue_state(void)
4366 struct workqueue_struct *wq;
4367 struct worker_pool *pool;
4368 unsigned long flags;
4371 rcu_read_lock_sched();
4373 pr_info("Showing busy workqueues and worker pools:\n");
4375 list_for_each_entry_rcu(wq, &workqueues, list) {
4376 struct pool_workqueue *pwq;
4379 for_each_pwq(pwq, wq) {
4380 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4388 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4390 for_each_pwq(pwq, wq) {
4391 spin_lock_irqsave(&pwq->pool->lock, flags);
4392 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4394 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4398 for_each_pool(pool, pi) {
4399 struct worker *worker;
4402 spin_lock_irqsave(&pool->lock, flags);
4403 if (pool->nr_workers == pool->nr_idle)
4406 pr_info("pool %d:", pool->id);
4407 pr_cont_pool_info(pool);
4408 pr_cont(" hung=%us workers=%d",
4409 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4412 pr_cont(" manager: %d",
4413 task_pid_nr(pool->manager->task));
4414 list_for_each_entry(worker, &pool->idle_list, entry) {
4415 pr_cont(" %s%d", first ? "idle: " : "",
4416 task_pid_nr(worker->task));
4421 spin_unlock_irqrestore(&pool->lock, flags);
4424 rcu_read_unlock_sched();
4430 * There are two challenges in supporting CPU hotplug. Firstly, there
4431 * are a lot of assumptions on strong associations among work, pwq and
4432 * pool which make migrating pending and scheduled works very
4433 * difficult to implement without impacting hot paths. Secondly,
4434 * worker pools serve mix of short, long and very long running works making
4435 * blocked draining impractical.
4437 * This is solved by allowing the pools to be disassociated from the CPU
4438 * running as an unbound one and allowing it to be reattached later if the
4439 * cpu comes back online.
4442 static void wq_unbind_fn(struct work_struct *work)
4444 int cpu = smp_processor_id();
4445 struct worker_pool *pool;
4446 struct worker *worker;
4448 for_each_cpu_worker_pool(pool, cpu) {
4449 mutex_lock(&pool->attach_mutex);
4450 spin_lock_irq(&pool->lock);
4453 * We've blocked all attach/detach operations. Make all workers
4454 * unbound and set DISASSOCIATED. Before this, all workers
4455 * except for the ones which are still executing works from
4456 * before the last CPU down must be on the cpu. After
4457 * this, they may become diasporas.
4459 for_each_pool_worker(worker, pool)
4460 worker->flags |= WORKER_UNBOUND;
4462 pool->flags |= POOL_DISASSOCIATED;
4464 spin_unlock_irq(&pool->lock);
4465 mutex_unlock(&pool->attach_mutex);
4468 * Call schedule() so that we cross rq->lock and thus can
4469 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4470 * This is necessary as scheduler callbacks may be invoked
4476 * Sched callbacks are disabled now. Zap nr_running.
4477 * After this, nr_running stays zero and need_more_worker()
4478 * and keep_working() are always true as long as the
4479 * worklist is not empty. This pool now behaves as an
4480 * unbound (in terms of concurrency management) pool which
4481 * are served by workers tied to the pool.
4483 atomic_set(&pool->nr_running, 0);
4486 * With concurrency management just turned off, a busy
4487 * worker blocking could lead to lengthy stalls. Kick off
4488 * unbound chain execution of currently pending work items.
4490 spin_lock_irq(&pool->lock);
4491 wake_up_worker(pool);
4492 spin_unlock_irq(&pool->lock);
4497 * rebind_workers - rebind all workers of a pool to the associated CPU
4498 * @pool: pool of interest
4500 * @pool->cpu is coming online. Rebind all workers to the CPU.
4502 static void rebind_workers(struct worker_pool *pool)
4504 struct worker *worker;
4506 lockdep_assert_held(&pool->attach_mutex);
4509 * Restore CPU affinity of all workers. As all idle workers should
4510 * be on the run-queue of the associated CPU before any local
4511 * wake-ups for concurrency management happen, restore CPU affinity
4512 * of all workers first and then clear UNBOUND. As we're called
4513 * from CPU_ONLINE, the following shouldn't fail.
4515 for_each_pool_worker(worker, pool)
4516 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4517 pool->attrs->cpumask) < 0);
4519 spin_lock_irq(&pool->lock);
4520 pool->flags &= ~POOL_DISASSOCIATED;
4522 for_each_pool_worker(worker, pool) {
4523 unsigned int worker_flags = worker->flags;
4526 * A bound idle worker should actually be on the runqueue
4527 * of the associated CPU for local wake-ups targeting it to
4528 * work. Kick all idle workers so that they migrate to the
4529 * associated CPU. Doing this in the same loop as
4530 * replacing UNBOUND with REBOUND is safe as no worker will
4531 * be bound before @pool->lock is released.
4533 if (worker_flags & WORKER_IDLE)
4534 wake_up_process(worker->task);
4537 * We want to clear UNBOUND but can't directly call
4538 * worker_clr_flags() or adjust nr_running. Atomically
4539 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4540 * @worker will clear REBOUND using worker_clr_flags() when
4541 * it initiates the next execution cycle thus restoring
4542 * concurrency management. Note that when or whether
4543 * @worker clears REBOUND doesn't affect correctness.
4545 * ACCESS_ONCE() is necessary because @worker->flags may be
4546 * tested without holding any lock in
4547 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4548 * fail incorrectly leading to premature concurrency
4549 * management operations.
4551 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4552 worker_flags |= WORKER_REBOUND;
4553 worker_flags &= ~WORKER_UNBOUND;
4554 ACCESS_ONCE(worker->flags) = worker_flags;
4557 spin_unlock_irq(&pool->lock);
4561 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4562 * @pool: unbound pool of interest
4563 * @cpu: the CPU which is coming up
4565 * An unbound pool may end up with a cpumask which doesn't have any online
4566 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4567 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4568 * online CPU before, cpus_allowed of all its workers should be restored.
4570 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4572 static cpumask_t cpumask;
4573 struct worker *worker;
4575 lockdep_assert_held(&pool->attach_mutex);
4577 /* is @cpu allowed for @pool? */
4578 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4581 /* is @cpu the only online CPU? */
4582 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4583 if (cpumask_weight(&cpumask) != 1)
4586 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4587 for_each_pool_worker(worker, pool)
4588 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4589 pool->attrs->cpumask) < 0);
4593 * Workqueues should be brought up before normal priority CPU notifiers.
4594 * This will be registered high priority CPU notifier.
4596 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4597 unsigned long action,
4600 int cpu = (unsigned long)hcpu;
4601 struct worker_pool *pool;
4602 struct workqueue_struct *wq;
4605 switch (action & ~CPU_TASKS_FROZEN) {
4606 case CPU_UP_PREPARE:
4607 for_each_cpu_worker_pool(pool, cpu) {
4608 if (pool->nr_workers)
4610 if (!create_worker(pool))
4615 case CPU_DOWN_FAILED:
4617 mutex_lock(&wq_pool_mutex);
4619 for_each_pool(pool, pi) {
4620 mutex_lock(&pool->attach_mutex);
4622 if (pool->cpu == cpu)
4623 rebind_workers(pool);
4624 else if (pool->cpu < 0)
4625 restore_unbound_workers_cpumask(pool, cpu);
4627 mutex_unlock(&pool->attach_mutex);
4630 /* update NUMA affinity of unbound workqueues */
4631 list_for_each_entry(wq, &workqueues, list)
4632 wq_update_unbound_numa(wq, cpu, true);
4634 mutex_unlock(&wq_pool_mutex);
4641 * Workqueues should be brought down after normal priority CPU notifiers.
4642 * This will be registered as low priority CPU notifier.
4644 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4645 unsigned long action,
4648 int cpu = (unsigned long)hcpu;
4649 struct work_struct unbind_work;
4650 struct workqueue_struct *wq;
4652 switch (action & ~CPU_TASKS_FROZEN) {
4653 case CPU_DOWN_PREPARE:
4654 /* unbinding per-cpu workers should happen on the local CPU */
4655 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4656 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4658 /* update NUMA affinity of unbound workqueues */
4659 mutex_lock(&wq_pool_mutex);
4660 list_for_each_entry(wq, &workqueues, list)
4661 wq_update_unbound_numa(wq, cpu, false);
4662 mutex_unlock(&wq_pool_mutex);
4664 /* wait for per-cpu unbinding to finish */
4665 flush_work(&unbind_work);
4666 destroy_work_on_stack(&unbind_work);
4674 struct work_for_cpu {
4675 struct work_struct work;
4681 static void work_for_cpu_fn(struct work_struct *work)
4683 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4685 wfc->ret = wfc->fn(wfc->arg);
4689 * work_on_cpu - run a function in user context on a particular cpu
4690 * @cpu: the cpu to run on
4691 * @fn: the function to run
4692 * @arg: the function arg
4694 * It is up to the caller to ensure that the cpu doesn't go offline.
4695 * The caller must not hold any locks which would prevent @fn from completing.
4697 * Return: The value @fn returns.
4699 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4701 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4703 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4704 schedule_work_on(cpu, &wfc.work);
4705 flush_work(&wfc.work);
4706 destroy_work_on_stack(&wfc.work);
4709 EXPORT_SYMBOL_GPL(work_on_cpu);
4710 #endif /* CONFIG_SMP */
4712 #ifdef CONFIG_FREEZER
4715 * freeze_workqueues_begin - begin freezing workqueues
4717 * Start freezing workqueues. After this function returns, all freezable
4718 * workqueues will queue new works to their delayed_works list instead of
4722 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4724 void freeze_workqueues_begin(void)
4726 struct workqueue_struct *wq;
4727 struct pool_workqueue *pwq;
4729 mutex_lock(&wq_pool_mutex);
4731 WARN_ON_ONCE(workqueue_freezing);
4732 workqueue_freezing = true;
4734 list_for_each_entry(wq, &workqueues, list) {
4735 mutex_lock(&wq->mutex);
4736 for_each_pwq(pwq, wq)
4737 pwq_adjust_max_active(pwq);
4738 mutex_unlock(&wq->mutex);
4741 mutex_unlock(&wq_pool_mutex);
4745 * freeze_workqueues_busy - are freezable workqueues still busy?
4747 * Check whether freezing is complete. This function must be called
4748 * between freeze_workqueues_begin() and thaw_workqueues().
4751 * Grabs and releases wq_pool_mutex.
4754 * %true if some freezable workqueues are still busy. %false if freezing
4757 bool freeze_workqueues_busy(void)
4760 struct workqueue_struct *wq;
4761 struct pool_workqueue *pwq;
4763 mutex_lock(&wq_pool_mutex);
4765 WARN_ON_ONCE(!workqueue_freezing);
4767 list_for_each_entry(wq, &workqueues, list) {
4768 if (!(wq->flags & WQ_FREEZABLE))
4771 * nr_active is monotonically decreasing. It's safe
4772 * to peek without lock.
4774 rcu_read_lock_sched();
4775 for_each_pwq(pwq, wq) {
4776 WARN_ON_ONCE(pwq->nr_active < 0);
4777 if (pwq->nr_active) {
4779 rcu_read_unlock_sched();
4783 rcu_read_unlock_sched();
4786 mutex_unlock(&wq_pool_mutex);
4791 * thaw_workqueues - thaw workqueues
4793 * Thaw workqueues. Normal queueing is restored and all collected
4794 * frozen works are transferred to their respective pool worklists.
4797 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4799 void thaw_workqueues(void)
4801 struct workqueue_struct *wq;
4802 struct pool_workqueue *pwq;
4804 mutex_lock(&wq_pool_mutex);
4806 if (!workqueue_freezing)
4809 workqueue_freezing = false;
4811 /* restore max_active and repopulate worklist */
4812 list_for_each_entry(wq, &workqueues, list) {
4813 mutex_lock(&wq->mutex);
4814 for_each_pwq(pwq, wq)
4815 pwq_adjust_max_active(pwq);
4816 mutex_unlock(&wq->mutex);
4820 mutex_unlock(&wq_pool_mutex);
4822 #endif /* CONFIG_FREEZER */
4824 static int workqueue_apply_unbound_cpumask(void)
4828 struct workqueue_struct *wq;
4829 struct apply_wqattrs_ctx *ctx, *n;
4831 lockdep_assert_held(&wq_pool_mutex);
4833 list_for_each_entry(wq, &workqueues, list) {
4834 if (!(wq->flags & WQ_UNBOUND))
4836 /* creating multiple pwqs breaks ordering guarantee */
4837 if (wq->flags & __WQ_ORDERED)
4840 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4846 list_add_tail(&ctx->list, &ctxs);
4849 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4851 apply_wqattrs_commit(ctx);
4852 apply_wqattrs_cleanup(ctx);
4859 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4860 * @cpumask: the cpumask to set
4862 * The low-level workqueues cpumask is a global cpumask that limits
4863 * the affinity of all unbound workqueues. This function check the @cpumask
4864 * and apply it to all unbound workqueues and updates all pwqs of them.
4866 * Retun: 0 - Success
4867 * -EINVAL - Invalid @cpumask
4868 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4870 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4873 cpumask_var_t saved_cpumask;
4875 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4878 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4879 if (!cpumask_empty(cpumask)) {
4880 apply_wqattrs_lock();
4882 /* save the old wq_unbound_cpumask. */
4883 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4885 /* update wq_unbound_cpumask at first and apply it to wqs. */
4886 cpumask_copy(wq_unbound_cpumask, cpumask);
4887 ret = workqueue_apply_unbound_cpumask();
4889 /* restore the wq_unbound_cpumask when failed. */
4891 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4893 apply_wqattrs_unlock();
4896 free_cpumask_var(saved_cpumask);
4902 * Workqueues with WQ_SYSFS flag set is visible to userland via
4903 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4904 * following attributes.
4906 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4907 * max_active RW int : maximum number of in-flight work items
4909 * Unbound workqueues have the following extra attributes.
4911 * id RO int : the associated pool ID
4912 * nice RW int : nice value of the workers
4913 * cpumask RW mask : bitmask of allowed CPUs for the workers
4916 struct workqueue_struct *wq;
4920 static struct workqueue_struct *dev_to_wq(struct device *dev)
4922 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4927 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4930 struct workqueue_struct *wq = dev_to_wq(dev);
4932 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4934 static DEVICE_ATTR_RO(per_cpu);
4936 static ssize_t max_active_show(struct device *dev,
4937 struct device_attribute *attr, char *buf)
4939 struct workqueue_struct *wq = dev_to_wq(dev);
4941 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4944 static ssize_t max_active_store(struct device *dev,
4945 struct device_attribute *attr, const char *buf,
4948 struct workqueue_struct *wq = dev_to_wq(dev);
4951 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4954 workqueue_set_max_active(wq, val);
4957 static DEVICE_ATTR_RW(max_active);
4959 static struct attribute *wq_sysfs_attrs[] = {
4960 &dev_attr_per_cpu.attr,
4961 &dev_attr_max_active.attr,
4964 ATTRIBUTE_GROUPS(wq_sysfs);
4966 static ssize_t wq_pool_ids_show(struct device *dev,
4967 struct device_attribute *attr, char *buf)
4969 struct workqueue_struct *wq = dev_to_wq(dev);
4970 const char *delim = "";
4971 int node, written = 0;
4973 rcu_read_lock_sched();
4974 for_each_node(node) {
4975 written += scnprintf(buf + written, PAGE_SIZE - written,
4976 "%s%d:%d", delim, node,
4977 unbound_pwq_by_node(wq, node)->pool->id);
4980 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4981 rcu_read_unlock_sched();
4986 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4989 struct workqueue_struct *wq = dev_to_wq(dev);
4992 mutex_lock(&wq->mutex);
4993 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4994 mutex_unlock(&wq->mutex);
4999 /* prepare workqueue_attrs for sysfs store operations */
5000 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5002 struct workqueue_attrs *attrs;
5004 lockdep_assert_held(&wq_pool_mutex);
5006 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5010 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5014 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5015 const char *buf, size_t count)
5017 struct workqueue_struct *wq = dev_to_wq(dev);
5018 struct workqueue_attrs *attrs;
5021 apply_wqattrs_lock();
5023 attrs = wq_sysfs_prep_attrs(wq);
5027 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5028 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5029 ret = apply_workqueue_attrs_locked(wq, attrs);
5034 apply_wqattrs_unlock();
5035 free_workqueue_attrs(attrs);
5036 return ret ?: count;
5039 static ssize_t wq_cpumask_show(struct device *dev,
5040 struct device_attribute *attr, char *buf)
5042 struct workqueue_struct *wq = dev_to_wq(dev);
5045 mutex_lock(&wq->mutex);
5046 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5047 cpumask_pr_args(wq->unbound_attrs->cpumask));
5048 mutex_unlock(&wq->mutex);
5052 static ssize_t wq_cpumask_store(struct device *dev,
5053 struct device_attribute *attr,
5054 const char *buf, size_t count)
5056 struct workqueue_struct *wq = dev_to_wq(dev);
5057 struct workqueue_attrs *attrs;
5060 apply_wqattrs_lock();
5062 attrs = wq_sysfs_prep_attrs(wq);
5066 ret = cpumask_parse(buf, attrs->cpumask);
5068 ret = apply_workqueue_attrs_locked(wq, attrs);
5071 apply_wqattrs_unlock();
5072 free_workqueue_attrs(attrs);
5073 return ret ?: count;
5076 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5079 struct workqueue_struct *wq = dev_to_wq(dev);
5082 mutex_lock(&wq->mutex);
5083 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5084 !wq->unbound_attrs->no_numa);
5085 mutex_unlock(&wq->mutex);
5090 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5091 const char *buf, size_t count)
5093 struct workqueue_struct *wq = dev_to_wq(dev);
5094 struct workqueue_attrs *attrs;
5095 int v, ret = -ENOMEM;
5097 apply_wqattrs_lock();
5099 attrs = wq_sysfs_prep_attrs(wq);
5104 if (sscanf(buf, "%d", &v) == 1) {
5105 attrs->no_numa = !v;
5106 ret = apply_workqueue_attrs_locked(wq, attrs);
5110 apply_wqattrs_unlock();
5111 free_workqueue_attrs(attrs);
5112 return ret ?: count;
5115 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5116 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5117 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5118 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5119 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5123 static struct bus_type wq_subsys = {
5124 .name = "workqueue",
5125 .dev_groups = wq_sysfs_groups,
5128 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5129 struct device_attribute *attr, char *buf)
5133 mutex_lock(&wq_pool_mutex);
5134 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5135 cpumask_pr_args(wq_unbound_cpumask));
5136 mutex_unlock(&wq_pool_mutex);
5141 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5142 struct device_attribute *attr, const char *buf, size_t count)
5144 cpumask_var_t cpumask;
5147 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5150 ret = cpumask_parse(buf, cpumask);
5152 ret = workqueue_set_unbound_cpumask(cpumask);
5154 free_cpumask_var(cpumask);
5155 return ret ? ret : count;
5158 static struct device_attribute wq_sysfs_cpumask_attr =
5159 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5160 wq_unbound_cpumask_store);
5162 static int __init wq_sysfs_init(void)
5166 err = subsys_virtual_register(&wq_subsys, NULL);
5170 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5172 core_initcall(wq_sysfs_init);
5174 static void wq_device_release(struct device *dev)
5176 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5182 * workqueue_sysfs_register - make a workqueue visible in sysfs
5183 * @wq: the workqueue to register
5185 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5186 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5187 * which is the preferred method.
5189 * Workqueue user should use this function directly iff it wants to apply
5190 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5191 * apply_workqueue_attrs() may race against userland updating the
5194 * Return: 0 on success, -errno on failure.
5196 int workqueue_sysfs_register(struct workqueue_struct *wq)
5198 struct wq_device *wq_dev;
5202 * Adjusting max_active or creating new pwqs by applying
5203 * attributes breaks ordering guarantee. Disallow exposing ordered
5206 if (WARN_ON(wq->flags & __WQ_ORDERED))
5209 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5214 wq_dev->dev.bus = &wq_subsys;
5215 wq_dev->dev.init_name = wq->name;
5216 wq_dev->dev.release = wq_device_release;
5219 * unbound_attrs are created separately. Suppress uevent until
5220 * everything is ready.
5222 dev_set_uevent_suppress(&wq_dev->dev, true);
5224 ret = device_register(&wq_dev->dev);
5231 if (wq->flags & WQ_UNBOUND) {
5232 struct device_attribute *attr;
5234 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5235 ret = device_create_file(&wq_dev->dev, attr);
5237 device_unregister(&wq_dev->dev);
5244 dev_set_uevent_suppress(&wq_dev->dev, false);
5245 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5250 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5251 * @wq: the workqueue to unregister
5253 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5255 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5257 struct wq_device *wq_dev = wq->wq_dev;
5263 device_unregister(&wq_dev->dev);
5265 #else /* CONFIG_SYSFS */
5266 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5267 #endif /* CONFIG_SYSFS */
5270 * Workqueue watchdog.
5272 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5273 * flush dependency, a concurrency managed work item which stays RUNNING
5274 * indefinitely. Workqueue stalls can be very difficult to debug as the
5275 * usual warning mechanisms don't trigger and internal workqueue state is
5278 * Workqueue watchdog monitors all worker pools periodically and dumps
5279 * state if some pools failed to make forward progress for a while where
5280 * forward progress is defined as the first item on ->worklist changing.
5282 * This mechanism is controlled through the kernel parameter
5283 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5284 * corresponding sysfs parameter file.
5286 #ifdef CONFIG_WQ_WATCHDOG
5288 static void wq_watchdog_timer_fn(unsigned long data);
5290 static unsigned long wq_watchdog_thresh = 30;
5291 static struct timer_list wq_watchdog_timer =
5292 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5294 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5295 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5297 static void wq_watchdog_reset_touched(void)
5301 wq_watchdog_touched = jiffies;
5302 for_each_possible_cpu(cpu)
5303 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5306 static void wq_watchdog_timer_fn(unsigned long data)
5308 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5309 bool lockup_detected = false;
5310 struct worker_pool *pool;
5318 for_each_pool(pool, pi) {
5319 unsigned long pool_ts, touched, ts;
5321 if (list_empty(&pool->worklist))
5324 /* get the latest of pool and touched timestamps */
5325 pool_ts = READ_ONCE(pool->watchdog_ts);
5326 touched = READ_ONCE(wq_watchdog_touched);
5328 if (time_after(pool_ts, touched))
5333 if (pool->cpu >= 0) {
5334 unsigned long cpu_touched =
5335 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5337 if (time_after(cpu_touched, ts))
5342 if (time_after(jiffies, ts + thresh)) {
5343 lockup_detected = true;
5344 pr_emerg("BUG: workqueue lockup - pool");
5345 pr_cont_pool_info(pool);
5346 pr_cont(" stuck for %us!\n",
5347 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5353 if (lockup_detected)
5354 show_workqueue_state();
5356 wq_watchdog_reset_touched();
5357 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5360 void wq_watchdog_touch(int cpu)
5363 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5365 wq_watchdog_touched = jiffies;
5368 static void wq_watchdog_set_thresh(unsigned long thresh)
5370 wq_watchdog_thresh = 0;
5371 del_timer_sync(&wq_watchdog_timer);
5374 wq_watchdog_thresh = thresh;
5375 wq_watchdog_reset_touched();
5376 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5380 static int wq_watchdog_param_set_thresh(const char *val,
5381 const struct kernel_param *kp)
5383 unsigned long thresh;
5386 ret = kstrtoul(val, 0, &thresh);
5391 wq_watchdog_set_thresh(thresh);
5393 wq_watchdog_thresh = thresh;
5398 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5399 .set = wq_watchdog_param_set_thresh,
5400 .get = param_get_ulong,
5403 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5406 static void wq_watchdog_init(void)
5408 wq_watchdog_set_thresh(wq_watchdog_thresh);
5411 #else /* CONFIG_WQ_WATCHDOG */
5413 static inline void wq_watchdog_init(void) { }
5415 #endif /* CONFIG_WQ_WATCHDOG */
5417 static void __init wq_numa_init(void)
5422 if (num_possible_nodes() <= 1)
5425 if (wq_disable_numa) {
5426 pr_info("workqueue: NUMA affinity support disabled\n");
5430 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5431 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5434 * We want masks of possible CPUs of each node which isn't readily
5435 * available. Build one from cpu_to_node() which should have been
5436 * fully initialized by now.
5438 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5442 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5443 node_online(node) ? node : NUMA_NO_NODE));
5445 for_each_possible_cpu(cpu) {
5446 node = cpu_to_node(cpu);
5447 if (WARN_ON(node == NUMA_NO_NODE)) {
5448 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5449 /* happens iff arch is bonkers, let's just proceed */
5452 cpumask_set_cpu(cpu, tbl[node]);
5455 wq_numa_possible_cpumask = tbl;
5456 wq_numa_enabled = true;
5459 static int __init init_workqueues(void)
5461 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5464 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5466 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5467 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5469 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5471 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5472 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5476 /* initialize CPU pools */
5477 for_each_possible_cpu(cpu) {
5478 struct worker_pool *pool;
5481 for_each_cpu_worker_pool(pool, cpu) {
5482 BUG_ON(init_worker_pool(pool));
5484 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5485 pool->attrs->nice = std_nice[i++];
5486 pool->node = cpu_to_node(cpu);
5489 mutex_lock(&wq_pool_mutex);
5490 BUG_ON(worker_pool_assign_id(pool));
5491 mutex_unlock(&wq_pool_mutex);
5495 /* create the initial worker */
5496 for_each_online_cpu(cpu) {
5497 struct worker_pool *pool;
5499 for_each_cpu_worker_pool(pool, cpu) {
5500 pool->flags &= ~POOL_DISASSOCIATED;
5501 BUG_ON(!create_worker(pool));
5505 /* create default unbound and ordered wq attrs */
5506 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5507 struct workqueue_attrs *attrs;
5509 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5510 attrs->nice = std_nice[i];
5511 unbound_std_wq_attrs[i] = attrs;
5514 * An ordered wq should have only one pwq as ordering is
5515 * guaranteed by max_active which is enforced by pwqs.
5516 * Turn off NUMA so that dfl_pwq is used for all nodes.
5518 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5519 attrs->nice = std_nice[i];
5520 attrs->no_numa = true;
5521 ordered_wq_attrs[i] = attrs;
5524 system_wq = alloc_workqueue("events", 0, 0);
5525 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5526 system_long_wq = alloc_workqueue("events_long", 0, 0);
5527 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5528 WQ_UNBOUND_MAX_ACTIVE);
5529 system_freezable_wq = alloc_workqueue("events_freezable",
5531 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5532 WQ_POWER_EFFICIENT, 0);
5533 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5534 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5536 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5537 !system_unbound_wq || !system_freezable_wq ||
5538 !system_power_efficient_wq ||
5539 !system_freezable_power_efficient_wq);
5545 early_initcall(init_workqueues);