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 struct list_head worklist; /* L: list of pending works */
152 int nr_workers; /* L: total number of workers */
154 /* nr_idle includes the ones off idle_list for rebinding */
155 int nr_idle; /* L: currently idle ones */
157 struct list_head idle_list; /* X: list of idle workers */
158 struct timer_list idle_timer; /* L: worker idle timeout */
159 struct timer_list mayday_timer; /* L: SOS timer for workers */
161 /* a workers is either on busy_hash or idle_list, or the manager */
162 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
163 /* L: hash of busy workers */
165 /* see manage_workers() for details on the two manager mutexes */
166 struct mutex manager_arb; /* manager arbitration */
167 struct worker *manager; /* L: purely informational */
168 struct mutex attach_mutex; /* attach/detach exclusion */
169 struct list_head workers; /* A: attached workers */
170 struct completion *detach_completion; /* all workers detached */
172 struct ida worker_ida; /* worker IDs for task name */
174 struct workqueue_attrs *attrs; /* I: worker attributes */
175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
176 int refcnt; /* PL: refcnt for unbound pools */
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
183 atomic_t nr_running ____cacheline_aligned_in_smp;
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
190 } ____cacheline_aligned_in_smp;
193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
198 struct pool_workqueue {
199 struct worker_pool *pool; /* I: the associated pool */
200 struct workqueue_struct *wq; /* I: the owning workqueue */
201 int work_color; /* L: current color */
202 int flush_color; /* L: flushing color */
203 int refcnt; /* L: reference count */
204 int nr_in_flight[WORK_NR_COLORS];
205 /* L: nr of in_flight works */
206 int nr_active; /* L: nr of active works */
207 int max_active; /* L: max active works */
208 struct list_head delayed_works; /* L: delayed works */
209 struct list_head pwqs_node; /* WR: node on wq->pwqs */
210 struct list_head mayday_node; /* MD: node on wq->maydays */
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
216 * determined without grabbing wq->mutex.
218 struct work_struct unbound_release_work;
220 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
223 * Structure used to wait for workqueue flush.
226 struct list_head list; /* WQ: list of flushers */
227 int flush_color; /* WQ: flush color waiting for */
228 struct completion done; /* flush completion */
234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
237 struct workqueue_struct {
238 struct list_head pwqs; /* WR: all pwqs of this wq */
239 struct list_head list; /* PR: list of all workqueues */
241 struct mutex mutex; /* protects this wq */
242 int work_color; /* WQ: current work color */
243 int flush_color; /* WQ: current flush color */
244 atomic_t nr_pwqs_to_flush; /* flush in progress */
245 struct wq_flusher *first_flusher; /* WQ: first flusher */
246 struct list_head flusher_queue; /* WQ: flush waiters */
247 struct list_head flusher_overflow; /* WQ: flush overflow list */
249 struct list_head maydays; /* MD: pwqs requesting rescue */
250 struct worker *rescuer; /* I: rescue worker */
252 int nr_drainers; /* WQ: drain in progress */
253 int saved_max_active; /* WQ: saved pwq max_active */
255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259 struct wq_device *wq_dev; /* I: for sysfs interface */
261 #ifdef CONFIG_LOCKDEP
262 struct lockdep_map lockdep_map;
264 char name[WQ_NAME_LEN]; /* I: workqueue name */
267 * Destruction of workqueue_struct is sched-RCU protected to allow
268 * walking the workqueues list without grabbing wq_pool_mutex.
269 * This is used to dump all workqueues from sysrq.
273 /* hot fields used during command issue, aligned to cacheline */
274 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
275 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
276 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279 static struct kmem_cache *pwq_cache;
281 static cpumask_var_t *wq_numa_possible_cpumask;
282 /* possible CPUs of each node */
284 static bool wq_disable_numa;
285 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
287 /* see the comment above the definition of WQ_POWER_EFFICIENT */
288 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
289 static bool wq_power_efficient = true;
291 static bool wq_power_efficient;
294 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
296 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
298 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
301 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
302 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
304 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
305 static bool workqueue_freezing; /* PL: have wqs started freezing? */
307 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
309 /* the per-cpu worker pools */
310 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
313 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
315 /* PL: hash of all unbound pools keyed by pool->attrs */
316 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
318 /* I: attributes used when instantiating standard unbound pools on demand */
319 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
321 /* I: attributes used when instantiating ordered pools on demand */
322 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
324 struct workqueue_struct *system_wq __read_mostly;
325 EXPORT_SYMBOL(system_wq);
326 struct workqueue_struct *system_highpri_wq __read_mostly;
327 EXPORT_SYMBOL_GPL(system_highpri_wq);
328 struct workqueue_struct *system_long_wq __read_mostly;
329 EXPORT_SYMBOL_GPL(system_long_wq);
330 struct workqueue_struct *system_unbound_wq __read_mostly;
331 EXPORT_SYMBOL_GPL(system_unbound_wq);
332 struct workqueue_struct *system_freezable_wq __read_mostly;
333 EXPORT_SYMBOL_GPL(system_freezable_wq);
334 struct workqueue_struct *system_power_efficient_wq __read_mostly;
335 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
336 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
337 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
339 static int worker_thread(void *__worker);
340 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
342 #define CREATE_TRACE_POINTS
343 #include <trace/events/workqueue.h>
345 #define assert_rcu_or_pool_mutex() \
346 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
347 lockdep_is_held(&wq_pool_mutex), \
348 "sched RCU or wq_pool_mutex should be held")
350 #define assert_rcu_or_wq_mutex(wq) \
351 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
352 lockdep_is_held(&wq->mutex), \
353 "sched RCU or wq->mutex should be held")
355 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
356 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
357 lockdep_is_held(&wq->mutex) || \
358 lockdep_is_held(&wq_pool_mutex), \
359 "sched RCU, wq->mutex or wq_pool_mutex should be held")
361 #define for_each_cpu_worker_pool(pool, cpu) \
362 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
363 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
367 * for_each_pool - iterate through all worker_pools in the system
368 * @pool: iteration cursor
369 * @pi: integer used for iteration
371 * This must be called either with wq_pool_mutex held or sched RCU read
372 * locked. If the pool needs to be used beyond the locking in effect, the
373 * caller is responsible for guaranteeing that the pool stays online.
375 * The if/else clause exists only for the lockdep assertion and can be
378 #define for_each_pool(pool, pi) \
379 idr_for_each_entry(&worker_pool_idr, pool, pi) \
380 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
384 * for_each_pool_worker - iterate through all workers of a worker_pool
385 * @worker: iteration cursor
386 * @pool: worker_pool to iterate workers of
388 * This must be called with @pool->attach_mutex.
390 * The if/else clause exists only for the lockdep assertion and can be
393 #define for_each_pool_worker(worker, pool) \
394 list_for_each_entry((worker), &(pool)->workers, node) \
395 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
399 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
400 * @pwq: iteration cursor
401 * @wq: the target workqueue
403 * This must be called either with wq->mutex held or sched RCU read locked.
404 * If the pwq needs to be used beyond the locking in effect, the caller is
405 * responsible for guaranteeing that the pwq stays online.
407 * The if/else clause exists only for the lockdep assertion and can be
410 #define for_each_pwq(pwq, wq) \
411 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
412 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
415 #ifdef CONFIG_DEBUG_OBJECTS_WORK
417 static struct debug_obj_descr work_debug_descr;
419 static void *work_debug_hint(void *addr)
421 return ((struct work_struct *) addr)->func;
425 * fixup_init is called when:
426 * - an active object is initialized
428 static int work_fixup_init(void *addr, enum debug_obj_state state)
430 struct work_struct *work = addr;
433 case ODEBUG_STATE_ACTIVE:
434 cancel_work_sync(work);
435 debug_object_init(work, &work_debug_descr);
443 * fixup_activate is called when:
444 * - an active object is activated
445 * - an unknown object is activated (might be a statically initialized object)
447 static int work_fixup_activate(void *addr, enum debug_obj_state state)
449 struct work_struct *work = addr;
453 case ODEBUG_STATE_NOTAVAILABLE:
455 * This is not really a fixup. The work struct was
456 * statically initialized. We just make sure that it
457 * is tracked in the object tracker.
459 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
460 debug_object_init(work, &work_debug_descr);
461 debug_object_activate(work, &work_debug_descr);
467 case ODEBUG_STATE_ACTIVE:
476 * fixup_free is called when:
477 * - an active object is freed
479 static int work_fixup_free(void *addr, enum debug_obj_state state)
481 struct work_struct *work = addr;
484 case ODEBUG_STATE_ACTIVE:
485 cancel_work_sync(work);
486 debug_object_free(work, &work_debug_descr);
493 static struct debug_obj_descr work_debug_descr = {
494 .name = "work_struct",
495 .debug_hint = work_debug_hint,
496 .fixup_init = work_fixup_init,
497 .fixup_activate = work_fixup_activate,
498 .fixup_free = work_fixup_free,
501 static inline void debug_work_activate(struct work_struct *work)
503 debug_object_activate(work, &work_debug_descr);
506 static inline void debug_work_deactivate(struct work_struct *work)
508 debug_object_deactivate(work, &work_debug_descr);
511 void __init_work(struct work_struct *work, int onstack)
514 debug_object_init_on_stack(work, &work_debug_descr);
516 debug_object_init(work, &work_debug_descr);
518 EXPORT_SYMBOL_GPL(__init_work);
520 void destroy_work_on_stack(struct work_struct *work)
522 debug_object_free(work, &work_debug_descr);
524 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
526 void destroy_delayed_work_on_stack(struct delayed_work *work)
528 destroy_timer_on_stack(&work->timer);
529 debug_object_free(&work->work, &work_debug_descr);
531 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
534 static inline void debug_work_activate(struct work_struct *work) { }
535 static inline void debug_work_deactivate(struct work_struct *work) { }
539 * worker_pool_assign_id - allocate ID and assing it to @pool
540 * @pool: the pool pointer of interest
542 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
543 * successfully, -errno on failure.
545 static int worker_pool_assign_id(struct worker_pool *pool)
549 lockdep_assert_held(&wq_pool_mutex);
551 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
561 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
562 * @wq: the target workqueue
565 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
567 * If the pwq needs to be used beyond the locking in effect, the caller is
568 * responsible for guaranteeing that the pwq stays online.
570 * Return: The unbound pool_workqueue for @node.
572 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
575 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
579 static unsigned int work_color_to_flags(int color)
581 return color << WORK_STRUCT_COLOR_SHIFT;
584 static int get_work_color(struct work_struct *work)
586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
587 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
590 static int work_next_color(int color)
592 return (color + 1) % WORK_NR_COLORS;
596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597 * contain the pointer to the queued pwq. Once execution starts, the flag
598 * is cleared and the high bits contain OFFQ flags and pool ID.
600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601 * and clear_work_data() can be used to set the pwq, pool or clear
602 * work->data. These functions should only be called while the work is
603 * owned - ie. while the PENDING bit is set.
605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
606 * corresponding to a work. Pool is available once the work has been
607 * queued anywhere after initialization until it is sync canceled. pwq is
608 * available only while the work item is queued.
610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
611 * canceled. While being canceled, a work item may have its PENDING set
612 * but stay off timer and worklist for arbitrarily long and nobody should
613 * try to steal the PENDING bit.
615 static inline void set_work_data(struct work_struct *work, unsigned long data,
618 WARN_ON_ONCE(!work_pending(work));
619 atomic_long_set(&work->data, data | flags | work_static(work));
622 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
623 unsigned long extra_flags)
625 set_work_data(work, (unsigned long)pwq,
626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
629 static void set_work_pool_and_keep_pending(struct work_struct *work,
632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
633 WORK_STRUCT_PENDING);
636 static void set_work_pool_and_clear_pending(struct work_struct *work,
640 * The following wmb is paired with the implied mb in
641 * test_and_set_bit(PENDING) and ensures all updates to @work made
642 * here are visible to and precede any updates by the next PENDING
646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
649 static void clear_work_data(struct work_struct *work)
651 smp_wmb(); /* see set_work_pool_and_clear_pending() */
652 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
655 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
657 unsigned long data = atomic_long_read(&work->data);
659 if (data & WORK_STRUCT_PWQ)
660 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
666 * get_work_pool - return the worker_pool a given work was associated with
667 * @work: the work item of interest
669 * Pools are created and destroyed under wq_pool_mutex, and allows read
670 * access under sched-RCU read lock. As such, this function should be
671 * called under wq_pool_mutex or with preemption disabled.
673 * All fields of the returned pool are accessible as long as the above
674 * mentioned locking is in effect. If the returned pool needs to be used
675 * beyond the critical section, the caller is responsible for ensuring the
676 * returned pool is and stays online.
678 * Return: The worker_pool @work was last associated with. %NULL if none.
680 static struct worker_pool *get_work_pool(struct work_struct *work)
682 unsigned long data = atomic_long_read(&work->data);
685 assert_rcu_or_pool_mutex();
687 if (data & WORK_STRUCT_PWQ)
688 return ((struct pool_workqueue *)
689 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
691 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
692 if (pool_id == WORK_OFFQ_POOL_NONE)
695 return idr_find(&worker_pool_idr, pool_id);
699 * get_work_pool_id - return the worker pool ID a given work is associated with
700 * @work: the work item of interest
702 * Return: The worker_pool ID @work was last associated with.
703 * %WORK_OFFQ_POOL_NONE if none.
705 static int get_work_pool_id(struct work_struct *work)
707 unsigned long data = atomic_long_read(&work->data);
709 if (data & WORK_STRUCT_PWQ)
710 return ((struct pool_workqueue *)
711 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
713 return data >> WORK_OFFQ_POOL_SHIFT;
716 static void mark_work_canceling(struct work_struct *work)
718 unsigned long pool_id = get_work_pool_id(work);
720 pool_id <<= WORK_OFFQ_POOL_SHIFT;
721 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
724 static bool work_is_canceling(struct work_struct *work)
726 unsigned long data = atomic_long_read(&work->data);
728 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
732 * Policy functions. These define the policies on how the global worker
733 * pools are managed. Unless noted otherwise, these functions assume that
734 * they're being called with pool->lock held.
737 static bool __need_more_worker(struct worker_pool *pool)
739 return !atomic_read(&pool->nr_running);
743 * Need to wake up a worker? Called from anything but currently
746 * Note that, because unbound workers never contribute to nr_running, this
747 * function will always return %true for unbound pools as long as the
748 * worklist isn't empty.
750 static bool need_more_worker(struct worker_pool *pool)
752 return !list_empty(&pool->worklist) && __need_more_worker(pool);
755 /* Can I start working? Called from busy but !running workers. */
756 static bool may_start_working(struct worker_pool *pool)
758 return pool->nr_idle;
761 /* Do I need to keep working? Called from currently running workers. */
762 static bool keep_working(struct worker_pool *pool)
764 return !list_empty(&pool->worklist) &&
765 atomic_read(&pool->nr_running) <= 1;
768 /* Do we need a new worker? Called from manager. */
769 static bool need_to_create_worker(struct worker_pool *pool)
771 return need_more_worker(pool) && !may_start_working(pool);
774 /* Do we have too many workers and should some go away? */
775 static bool too_many_workers(struct worker_pool *pool)
777 bool managing = mutex_is_locked(&pool->manager_arb);
778 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
779 int nr_busy = pool->nr_workers - nr_idle;
781 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
788 /* Return the first idle worker. Safe with preemption disabled */
789 static struct worker *first_idle_worker(struct worker_pool *pool)
791 if (unlikely(list_empty(&pool->idle_list)))
794 return list_first_entry(&pool->idle_list, struct worker, entry);
798 * wake_up_worker - wake up an idle worker
799 * @pool: worker pool to wake worker from
801 * Wake up the first idle worker of @pool.
804 * spin_lock_irq(pool->lock).
806 static void wake_up_worker(struct worker_pool *pool)
808 struct worker *worker = first_idle_worker(pool);
811 wake_up_process(worker->task);
815 * wq_worker_waking_up - a worker is waking up
816 * @task: task waking up
817 * @cpu: CPU @task is waking up to
819 * This function is called during try_to_wake_up() when a worker is
823 * spin_lock_irq(rq->lock)
825 void wq_worker_waking_up(struct task_struct *task, int cpu)
827 struct worker *worker = kthread_data(task);
829 if (!(worker->flags & WORKER_NOT_RUNNING)) {
830 WARN_ON_ONCE(worker->pool->cpu != cpu);
831 atomic_inc(&worker->pool->nr_running);
836 * wq_worker_sleeping - a worker is going to sleep
837 * @task: task going to sleep
838 * @cpu: CPU in question, must be the current CPU number
840 * This function is called during schedule() when a busy worker is
841 * going to sleep. Worker on the same cpu can be woken up by
842 * returning pointer to its task.
845 * spin_lock_irq(rq->lock)
848 * Worker task on @cpu to wake up, %NULL if none.
850 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
852 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
853 struct worker_pool *pool;
856 * Rescuers, which may not have all the fields set up like normal
857 * workers, also reach here, let's not access anything before
858 * checking NOT_RUNNING.
860 if (worker->flags & WORKER_NOT_RUNNING)
865 /* this can only happen on the local cpu */
866 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
870 * The counterpart of the following dec_and_test, implied mb,
871 * worklist not empty test sequence is in insert_work().
872 * Please read comment there.
874 * NOT_RUNNING is clear. This means that we're bound to and
875 * running on the local cpu w/ rq lock held and preemption
876 * disabled, which in turn means that none else could be
877 * manipulating idle_list, so dereferencing idle_list without pool
880 if (atomic_dec_and_test(&pool->nr_running) &&
881 !list_empty(&pool->worklist))
882 to_wakeup = first_idle_worker(pool);
883 return to_wakeup ? to_wakeup->task : NULL;
887 * worker_set_flags - set worker flags and adjust nr_running accordingly
889 * @flags: flags to set
891 * Set @flags in @worker->flags and adjust nr_running accordingly.
894 * spin_lock_irq(pool->lock)
896 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
898 struct worker_pool *pool = worker->pool;
900 WARN_ON_ONCE(worker->task != current);
902 /* If transitioning into NOT_RUNNING, adjust nr_running. */
903 if ((flags & WORKER_NOT_RUNNING) &&
904 !(worker->flags & WORKER_NOT_RUNNING)) {
905 atomic_dec(&pool->nr_running);
908 worker->flags |= flags;
912 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
914 * @flags: flags to clear
916 * Clear @flags in @worker->flags and adjust nr_running accordingly.
919 * spin_lock_irq(pool->lock)
921 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
923 struct worker_pool *pool = worker->pool;
924 unsigned int oflags = worker->flags;
926 WARN_ON_ONCE(worker->task != current);
928 worker->flags &= ~flags;
931 * If transitioning out of NOT_RUNNING, increment nr_running. Note
932 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
933 * of multiple flags, not a single flag.
935 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
936 if (!(worker->flags & WORKER_NOT_RUNNING))
937 atomic_inc(&pool->nr_running);
941 * find_worker_executing_work - find worker which is executing a work
942 * @pool: pool of interest
943 * @work: work to find worker for
945 * Find a worker which is executing @work on @pool by searching
946 * @pool->busy_hash which is keyed by the address of @work. For a worker
947 * to match, its current execution should match the address of @work and
948 * its work function. This is to avoid unwanted dependency between
949 * unrelated work executions through a work item being recycled while still
952 * This is a bit tricky. A work item may be freed once its execution
953 * starts and nothing prevents the freed area from being recycled for
954 * another work item. If the same work item address ends up being reused
955 * before the original execution finishes, workqueue will identify the
956 * recycled work item as currently executing and make it wait until the
957 * current execution finishes, introducing an unwanted dependency.
959 * This function checks the work item address and work function to avoid
960 * false positives. Note that this isn't complete as one may construct a
961 * work function which can introduce dependency onto itself through a
962 * recycled work item. Well, if somebody wants to shoot oneself in the
963 * foot that badly, there's only so much we can do, and if such deadlock
964 * actually occurs, it should be easy to locate the culprit work function.
967 * spin_lock_irq(pool->lock).
970 * Pointer to worker which is executing @work if found, %NULL
973 static struct worker *find_worker_executing_work(struct worker_pool *pool,
974 struct work_struct *work)
976 struct worker *worker;
978 hash_for_each_possible(pool->busy_hash, worker, hentry,
980 if (worker->current_work == work &&
981 worker->current_func == work->func)
988 * move_linked_works - move linked works to a list
989 * @work: start of series of works to be scheduled
990 * @head: target list to append @work to
991 * @nextp: out parameter for nested worklist walking
993 * Schedule linked works starting from @work to @head. Work series to
994 * be scheduled starts at @work and includes any consecutive work with
995 * WORK_STRUCT_LINKED set in its predecessor.
997 * If @nextp is not NULL, it's updated to point to the next work of
998 * the last scheduled work. This allows move_linked_works() to be
999 * nested inside outer list_for_each_entry_safe().
1002 * spin_lock_irq(pool->lock).
1004 static void move_linked_works(struct work_struct *work, struct list_head *head,
1005 struct work_struct **nextp)
1007 struct work_struct *n;
1010 * Linked worklist will always end before the end of the list,
1011 * use NULL for list head.
1013 list_for_each_entry_safe_from(work, n, NULL, entry) {
1014 list_move_tail(&work->entry, head);
1015 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1020 * If we're already inside safe list traversal and have moved
1021 * multiple works to the scheduled queue, the next position
1022 * needs to be updated.
1029 * get_pwq - get an extra reference on the specified pool_workqueue
1030 * @pwq: pool_workqueue to get
1032 * Obtain an extra reference on @pwq. The caller should guarantee that
1033 * @pwq has positive refcnt and be holding the matching pool->lock.
1035 static void get_pwq(struct pool_workqueue *pwq)
1037 lockdep_assert_held(&pwq->pool->lock);
1038 WARN_ON_ONCE(pwq->refcnt <= 0);
1043 * put_pwq - put a pool_workqueue reference
1044 * @pwq: pool_workqueue to put
1046 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1047 * destruction. The caller should be holding the matching pool->lock.
1049 static void put_pwq(struct pool_workqueue *pwq)
1051 lockdep_assert_held(&pwq->pool->lock);
1052 if (likely(--pwq->refcnt))
1054 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1057 * @pwq can't be released under pool->lock, bounce to
1058 * pwq_unbound_release_workfn(). This never recurses on the same
1059 * pool->lock as this path is taken only for unbound workqueues and
1060 * the release work item is scheduled on a per-cpu workqueue. To
1061 * avoid lockdep warning, unbound pool->locks are given lockdep
1062 * subclass of 1 in get_unbound_pool().
1064 schedule_work(&pwq->unbound_release_work);
1068 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1069 * @pwq: pool_workqueue to put (can be %NULL)
1071 * put_pwq() with locking. This function also allows %NULL @pwq.
1073 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1077 * As both pwqs and pools are sched-RCU protected, the
1078 * following lock operations are safe.
1080 spin_lock_irq(&pwq->pool->lock);
1082 spin_unlock_irq(&pwq->pool->lock);
1086 static void pwq_activate_delayed_work(struct work_struct *work)
1088 struct pool_workqueue *pwq = get_work_pwq(work);
1090 trace_workqueue_activate_work(work);
1091 move_linked_works(work, &pwq->pool->worklist, NULL);
1092 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1096 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1098 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1099 struct work_struct, entry);
1101 pwq_activate_delayed_work(work);
1105 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1106 * @pwq: pwq of interest
1107 * @color: color of work which left the queue
1109 * A work either has completed or is removed from pending queue,
1110 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1113 * spin_lock_irq(pool->lock).
1115 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1117 /* uncolored work items don't participate in flushing or nr_active */
1118 if (color == WORK_NO_COLOR)
1121 pwq->nr_in_flight[color]--;
1124 if (!list_empty(&pwq->delayed_works)) {
1125 /* one down, submit a delayed one */
1126 if (pwq->nr_active < pwq->max_active)
1127 pwq_activate_first_delayed(pwq);
1130 /* is flush in progress and are we at the flushing tip? */
1131 if (likely(pwq->flush_color != color))
1134 /* are there still in-flight works? */
1135 if (pwq->nr_in_flight[color])
1138 /* this pwq is done, clear flush_color */
1139 pwq->flush_color = -1;
1142 * If this was the last pwq, wake up the first flusher. It
1143 * will handle the rest.
1145 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1146 complete(&pwq->wq->first_flusher->done);
1152 * try_to_grab_pending - steal work item from worklist and disable irq
1153 * @work: work item to steal
1154 * @is_dwork: @work is a delayed_work
1155 * @flags: place to store irq state
1157 * Try to grab PENDING bit of @work. This function can handle @work in any
1158 * stable state - idle, on timer or on worklist.
1161 * 1 if @work was pending and we successfully stole PENDING
1162 * 0 if @work was idle and we claimed PENDING
1163 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1164 * -ENOENT if someone else is canceling @work, this state may persist
1165 * for arbitrarily long
1168 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1169 * interrupted while holding PENDING and @work off queue, irq must be
1170 * disabled on entry. This, combined with delayed_work->timer being
1171 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1173 * On successful return, >= 0, irq is disabled and the caller is
1174 * responsible for releasing it using local_irq_restore(*@flags).
1176 * This function is safe to call from any context including IRQ handler.
1178 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1179 unsigned long *flags)
1181 struct worker_pool *pool;
1182 struct pool_workqueue *pwq;
1184 local_irq_save(*flags);
1186 /* try to steal the timer if it exists */
1188 struct delayed_work *dwork = to_delayed_work(work);
1191 * dwork->timer is irqsafe. If del_timer() fails, it's
1192 * guaranteed that the timer is not queued anywhere and not
1193 * running on the local CPU.
1195 if (likely(del_timer(&dwork->timer)))
1199 /* try to claim PENDING the normal way */
1200 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1204 * The queueing is in progress, or it is already queued. Try to
1205 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1207 pool = get_work_pool(work);
1211 spin_lock(&pool->lock);
1213 * work->data is guaranteed to point to pwq only while the work
1214 * item is queued on pwq->wq, and both updating work->data to point
1215 * to pwq on queueing and to pool on dequeueing are done under
1216 * pwq->pool->lock. This in turn guarantees that, if work->data
1217 * points to pwq which is associated with a locked pool, the work
1218 * item is currently queued on that pool.
1220 pwq = get_work_pwq(work);
1221 if (pwq && pwq->pool == pool) {
1222 debug_work_deactivate(work);
1225 * A delayed work item cannot be grabbed directly because
1226 * it might have linked NO_COLOR work items which, if left
1227 * on the delayed_list, will confuse pwq->nr_active
1228 * management later on and cause stall. Make sure the work
1229 * item is activated before grabbing.
1231 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1232 pwq_activate_delayed_work(work);
1234 list_del_init(&work->entry);
1235 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1237 /* work->data points to pwq iff queued, point to pool */
1238 set_work_pool_and_keep_pending(work, pool->id);
1240 spin_unlock(&pool->lock);
1243 spin_unlock(&pool->lock);
1245 local_irq_restore(*flags);
1246 if (work_is_canceling(work))
1253 * insert_work - insert a work into a pool
1254 * @pwq: pwq @work belongs to
1255 * @work: work to insert
1256 * @head: insertion point
1257 * @extra_flags: extra WORK_STRUCT_* flags to set
1259 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1260 * work_struct flags.
1263 * spin_lock_irq(pool->lock).
1265 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1266 struct list_head *head, unsigned int extra_flags)
1268 struct worker_pool *pool = pwq->pool;
1270 /* we own @work, set data and link */
1271 set_work_pwq(work, pwq, extra_flags);
1272 list_add_tail(&work->entry, head);
1276 * Ensure either wq_worker_sleeping() sees the above
1277 * list_add_tail() or we see zero nr_running to avoid workers lying
1278 * around lazily while there are works to be processed.
1282 if (__need_more_worker(pool))
1283 wake_up_worker(pool);
1287 * Test whether @work is being queued from another work executing on the
1290 static bool is_chained_work(struct workqueue_struct *wq)
1292 struct worker *worker;
1294 worker = current_wq_worker();
1296 * Return %true iff I'm a worker execuing a work item on @wq. If
1297 * I'm @worker, it's safe to dereference it without locking.
1299 return worker && worker->current_pwq->wq == wq;
1302 static void __queue_work(int cpu, struct workqueue_struct *wq,
1303 struct work_struct *work)
1305 struct pool_workqueue *pwq;
1306 struct worker_pool *last_pool;
1307 struct list_head *worklist;
1308 unsigned int work_flags;
1309 unsigned int req_cpu = cpu;
1312 * While a work item is PENDING && off queue, a task trying to
1313 * steal the PENDING will busy-loop waiting for it to either get
1314 * queued or lose PENDING. Grabbing PENDING and queueing should
1315 * happen with IRQ disabled.
1317 WARN_ON_ONCE(!irqs_disabled());
1319 debug_work_activate(work);
1321 /* if draining, only works from the same workqueue are allowed */
1322 if (unlikely(wq->flags & __WQ_DRAINING) &&
1323 WARN_ON_ONCE(!is_chained_work(wq)))
1326 if (req_cpu == WORK_CPU_UNBOUND)
1327 cpu = raw_smp_processor_id();
1329 /* pwq which will be used unless @work is executing elsewhere */
1330 if (!(wq->flags & WQ_UNBOUND))
1331 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1333 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1336 * If @work was previously on a different pool, it might still be
1337 * running there, in which case the work needs to be queued on that
1338 * pool to guarantee non-reentrancy.
1340 last_pool = get_work_pool(work);
1341 if (last_pool && last_pool != pwq->pool) {
1342 struct worker *worker;
1344 spin_lock(&last_pool->lock);
1346 worker = find_worker_executing_work(last_pool, work);
1348 if (worker && worker->current_pwq->wq == wq) {
1349 pwq = worker->current_pwq;
1351 /* meh... not running there, queue here */
1352 spin_unlock(&last_pool->lock);
1353 spin_lock(&pwq->pool->lock);
1356 spin_lock(&pwq->pool->lock);
1360 * pwq is determined and locked. For unbound pools, we could have
1361 * raced with pwq release and it could already be dead. If its
1362 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1363 * without another pwq replacing it in the numa_pwq_tbl or while
1364 * work items are executing on it, so the retrying is guaranteed to
1365 * make forward-progress.
1367 if (unlikely(!pwq->refcnt)) {
1368 if (wq->flags & WQ_UNBOUND) {
1369 spin_unlock(&pwq->pool->lock);
1374 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1378 /* pwq determined, queue */
1379 trace_workqueue_queue_work(req_cpu, pwq, work);
1381 if (WARN_ON(!list_empty(&work->entry))) {
1382 spin_unlock(&pwq->pool->lock);
1386 pwq->nr_in_flight[pwq->work_color]++;
1387 work_flags = work_color_to_flags(pwq->work_color);
1389 if (likely(pwq->nr_active < pwq->max_active)) {
1390 trace_workqueue_activate_work(work);
1392 worklist = &pwq->pool->worklist;
1394 work_flags |= WORK_STRUCT_DELAYED;
1395 worklist = &pwq->delayed_works;
1398 insert_work(pwq, work, worklist, work_flags);
1400 spin_unlock(&pwq->pool->lock);
1404 * queue_work_on - queue work on specific cpu
1405 * @cpu: CPU number to execute work on
1406 * @wq: workqueue to use
1407 * @work: work to queue
1409 * We queue the work to a specific CPU, the caller must ensure it
1412 * Return: %false if @work was already on a queue, %true otherwise.
1414 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1415 struct work_struct *work)
1418 unsigned long flags;
1420 local_irq_save(flags);
1422 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1423 __queue_work(cpu, wq, work);
1427 local_irq_restore(flags);
1430 EXPORT_SYMBOL(queue_work_on);
1432 void delayed_work_timer_fn(unsigned long __data)
1434 struct delayed_work *dwork = (struct delayed_work *)__data;
1436 /* should have been called from irqsafe timer with irq already off */
1437 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1439 EXPORT_SYMBOL(delayed_work_timer_fn);
1441 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1442 struct delayed_work *dwork, unsigned long delay)
1444 struct timer_list *timer = &dwork->timer;
1445 struct work_struct *work = &dwork->work;
1447 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1448 timer->data != (unsigned long)dwork);
1449 WARN_ON_ONCE(timer_pending(timer));
1450 WARN_ON_ONCE(!list_empty(&work->entry));
1453 * If @delay is 0, queue @dwork->work immediately. This is for
1454 * both optimization and correctness. The earliest @timer can
1455 * expire is on the closest next tick and delayed_work users depend
1456 * on that there's no such delay when @delay is 0.
1459 __queue_work(cpu, wq, &dwork->work);
1463 timer_stats_timer_set_start_info(&dwork->timer);
1467 timer->expires = jiffies + delay;
1469 if (unlikely(cpu != WORK_CPU_UNBOUND))
1470 add_timer_on(timer, cpu);
1476 * queue_delayed_work_on - queue work on specific CPU after delay
1477 * @cpu: CPU number to execute work on
1478 * @wq: workqueue to use
1479 * @dwork: work to queue
1480 * @delay: number of jiffies to wait before queueing
1482 * Return: %false if @work was already on a queue, %true otherwise. If
1483 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1486 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1487 struct delayed_work *dwork, unsigned long delay)
1489 struct work_struct *work = &dwork->work;
1491 unsigned long flags;
1493 /* read the comment in __queue_work() */
1494 local_irq_save(flags);
1496 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1497 __queue_delayed_work(cpu, wq, dwork, delay);
1501 local_irq_restore(flags);
1504 EXPORT_SYMBOL(queue_delayed_work_on);
1507 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1508 * @cpu: CPU number to execute work on
1509 * @wq: workqueue to use
1510 * @dwork: work to queue
1511 * @delay: number of jiffies to wait before queueing
1513 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1514 * modify @dwork's timer so that it expires after @delay. If @delay is
1515 * zero, @work is guaranteed to be scheduled immediately regardless of its
1518 * Return: %false if @dwork was idle and queued, %true if @dwork was
1519 * pending and its timer was modified.
1521 * This function is safe to call from any context including IRQ handler.
1522 * See try_to_grab_pending() for details.
1524 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1525 struct delayed_work *dwork, unsigned long delay)
1527 unsigned long flags;
1531 ret = try_to_grab_pending(&dwork->work, true, &flags);
1532 } while (unlikely(ret == -EAGAIN));
1534 if (likely(ret >= 0)) {
1535 __queue_delayed_work(cpu, wq, dwork, delay);
1536 local_irq_restore(flags);
1539 /* -ENOENT from try_to_grab_pending() becomes %true */
1542 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1545 * worker_enter_idle - enter idle state
1546 * @worker: worker which is entering idle state
1548 * @worker is entering idle state. Update stats and idle timer if
1552 * spin_lock_irq(pool->lock).
1554 static void worker_enter_idle(struct worker *worker)
1556 struct worker_pool *pool = worker->pool;
1558 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1559 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1560 (worker->hentry.next || worker->hentry.pprev)))
1563 /* can't use worker_set_flags(), also called from create_worker() */
1564 worker->flags |= WORKER_IDLE;
1566 worker->last_active = jiffies;
1568 /* idle_list is LIFO */
1569 list_add(&worker->entry, &pool->idle_list);
1571 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1572 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1575 * Sanity check nr_running. Because wq_unbind_fn() releases
1576 * pool->lock between setting %WORKER_UNBOUND and zapping
1577 * nr_running, the warning may trigger spuriously. Check iff
1578 * unbind is not in progress.
1580 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1581 pool->nr_workers == pool->nr_idle &&
1582 atomic_read(&pool->nr_running));
1586 * worker_leave_idle - leave idle state
1587 * @worker: worker which is leaving idle state
1589 * @worker is leaving idle state. Update stats.
1592 * spin_lock_irq(pool->lock).
1594 static void worker_leave_idle(struct worker *worker)
1596 struct worker_pool *pool = worker->pool;
1598 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1600 worker_clr_flags(worker, WORKER_IDLE);
1602 list_del_init(&worker->entry);
1605 static struct worker *alloc_worker(int node)
1607 struct worker *worker;
1609 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1611 INIT_LIST_HEAD(&worker->entry);
1612 INIT_LIST_HEAD(&worker->scheduled);
1613 INIT_LIST_HEAD(&worker->node);
1614 /* on creation a worker is in !idle && prep state */
1615 worker->flags = WORKER_PREP;
1621 * worker_attach_to_pool() - attach a worker to a pool
1622 * @worker: worker to be attached
1623 * @pool: the target pool
1625 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1626 * cpu-binding of @worker are kept coordinated with the pool across
1629 static void worker_attach_to_pool(struct worker *worker,
1630 struct worker_pool *pool)
1632 mutex_lock(&pool->attach_mutex);
1635 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1636 * online CPUs. It'll be re-applied when any of the CPUs come up.
1638 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1641 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1642 * stable across this function. See the comments above the
1643 * flag definition for details.
1645 if (pool->flags & POOL_DISASSOCIATED)
1646 worker->flags |= WORKER_UNBOUND;
1648 list_add_tail(&worker->node, &pool->workers);
1650 mutex_unlock(&pool->attach_mutex);
1654 * worker_detach_from_pool() - detach a worker from its pool
1655 * @worker: worker which is attached to its pool
1656 * @pool: the pool @worker is attached to
1658 * Undo the attaching which had been done in worker_attach_to_pool(). The
1659 * caller worker shouldn't access to the pool after detached except it has
1660 * other reference to the pool.
1662 static void worker_detach_from_pool(struct worker *worker,
1663 struct worker_pool *pool)
1665 struct completion *detach_completion = NULL;
1667 mutex_lock(&pool->attach_mutex);
1668 list_del(&worker->node);
1669 if (list_empty(&pool->workers))
1670 detach_completion = pool->detach_completion;
1671 mutex_unlock(&pool->attach_mutex);
1673 /* clear leftover flags without pool->lock after it is detached */
1674 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1676 if (detach_completion)
1677 complete(detach_completion);
1681 * create_worker - create a new workqueue worker
1682 * @pool: pool the new worker will belong to
1684 * Create and start a new worker which is attached to @pool.
1687 * Might sleep. Does GFP_KERNEL allocations.
1690 * Pointer to the newly created worker.
1692 static struct worker *create_worker(struct worker_pool *pool)
1694 struct worker *worker = NULL;
1698 /* ID is needed to determine kthread name */
1699 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1703 worker = alloc_worker(pool->node);
1707 worker->pool = pool;
1711 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1712 pool->attrs->nice < 0 ? "H" : "");
1714 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1716 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1717 "kworker/%s", id_buf);
1718 if (IS_ERR(worker->task))
1721 set_user_nice(worker->task, pool->attrs->nice);
1723 /* prevent userland from meddling with cpumask of workqueue workers */
1724 worker->task->flags |= PF_NO_SETAFFINITY;
1726 /* successful, attach the worker to the pool */
1727 worker_attach_to_pool(worker, pool);
1729 /* start the newly created worker */
1730 spin_lock_irq(&pool->lock);
1731 worker->pool->nr_workers++;
1732 worker_enter_idle(worker);
1733 wake_up_process(worker->task);
1734 spin_unlock_irq(&pool->lock);
1740 ida_simple_remove(&pool->worker_ida, id);
1746 * destroy_worker - destroy a workqueue worker
1747 * @worker: worker to be destroyed
1749 * Destroy @worker and adjust @pool stats accordingly. The worker should
1753 * spin_lock_irq(pool->lock).
1755 static void destroy_worker(struct worker *worker)
1757 struct worker_pool *pool = worker->pool;
1759 lockdep_assert_held(&pool->lock);
1761 /* sanity check frenzy */
1762 if (WARN_ON(worker->current_work) ||
1763 WARN_ON(!list_empty(&worker->scheduled)) ||
1764 WARN_ON(!(worker->flags & WORKER_IDLE)))
1770 list_del_init(&worker->entry);
1771 worker->flags |= WORKER_DIE;
1772 wake_up_process(worker->task);
1775 static void idle_worker_timeout(unsigned long __pool)
1777 struct worker_pool *pool = (void *)__pool;
1779 spin_lock_irq(&pool->lock);
1781 while (too_many_workers(pool)) {
1782 struct worker *worker;
1783 unsigned long expires;
1785 /* idle_list is kept in LIFO order, check the last one */
1786 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1787 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1789 if (time_before(jiffies, expires)) {
1790 mod_timer(&pool->idle_timer, expires);
1794 destroy_worker(worker);
1797 spin_unlock_irq(&pool->lock);
1800 static void send_mayday(struct work_struct *work)
1802 struct pool_workqueue *pwq = get_work_pwq(work);
1803 struct workqueue_struct *wq = pwq->wq;
1805 lockdep_assert_held(&wq_mayday_lock);
1810 /* mayday mayday mayday */
1811 if (list_empty(&pwq->mayday_node)) {
1813 * If @pwq is for an unbound wq, its base ref may be put at
1814 * any time due to an attribute change. Pin @pwq until the
1815 * rescuer is done with it.
1818 list_add_tail(&pwq->mayday_node, &wq->maydays);
1819 wake_up_process(wq->rescuer->task);
1823 static void pool_mayday_timeout(unsigned long __pool)
1825 struct worker_pool *pool = (void *)__pool;
1826 struct work_struct *work;
1828 spin_lock_irq(&pool->lock);
1829 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1831 if (need_to_create_worker(pool)) {
1833 * We've been trying to create a new worker but
1834 * haven't been successful. We might be hitting an
1835 * allocation deadlock. Send distress signals to
1838 list_for_each_entry(work, &pool->worklist, entry)
1842 spin_unlock(&wq_mayday_lock);
1843 spin_unlock_irq(&pool->lock);
1845 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1849 * maybe_create_worker - create a new worker if necessary
1850 * @pool: pool to create a new worker for
1852 * Create a new worker for @pool if necessary. @pool is guaranteed to
1853 * have at least one idle worker on return from this function. If
1854 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1855 * sent to all rescuers with works scheduled on @pool to resolve
1856 * possible allocation deadlock.
1858 * On return, need_to_create_worker() is guaranteed to be %false and
1859 * may_start_working() %true.
1862 * spin_lock_irq(pool->lock) which may be released and regrabbed
1863 * multiple times. Does GFP_KERNEL allocations. Called only from
1866 static void maybe_create_worker(struct worker_pool *pool)
1867 __releases(&pool->lock)
1868 __acquires(&pool->lock)
1871 spin_unlock_irq(&pool->lock);
1873 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1874 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1877 if (create_worker(pool) || !need_to_create_worker(pool))
1880 schedule_timeout_interruptible(CREATE_COOLDOWN);
1882 if (!need_to_create_worker(pool))
1886 del_timer_sync(&pool->mayday_timer);
1887 spin_lock_irq(&pool->lock);
1889 * This is necessary even after a new worker was just successfully
1890 * created as @pool->lock was dropped and the new worker might have
1891 * already become busy.
1893 if (need_to_create_worker(pool))
1898 * manage_workers - manage worker pool
1901 * Assume the manager role and manage the worker pool @worker belongs
1902 * to. At any given time, there can be only zero or one manager per
1903 * pool. The exclusion is handled automatically by this function.
1905 * The caller can safely start processing works on false return. On
1906 * true return, it's guaranteed that need_to_create_worker() is false
1907 * and may_start_working() is true.
1910 * spin_lock_irq(pool->lock) which may be released and regrabbed
1911 * multiple times. Does GFP_KERNEL allocations.
1914 * %false if the pool doesn't need management and the caller can safely
1915 * start processing works, %true if management function was performed and
1916 * the conditions that the caller verified before calling the function may
1917 * no longer be true.
1919 static bool manage_workers(struct worker *worker)
1921 struct worker_pool *pool = worker->pool;
1924 * Anyone who successfully grabs manager_arb wins the arbitration
1925 * and becomes the manager. mutex_trylock() on pool->manager_arb
1926 * failure while holding pool->lock reliably indicates that someone
1927 * else is managing the pool and the worker which failed trylock
1928 * can proceed to executing work items. This means that anyone
1929 * grabbing manager_arb is responsible for actually performing
1930 * manager duties. If manager_arb is grabbed and released without
1931 * actual management, the pool may stall indefinitely.
1933 if (!mutex_trylock(&pool->manager_arb))
1935 pool->manager = worker;
1937 maybe_create_worker(pool);
1939 pool->manager = NULL;
1940 mutex_unlock(&pool->manager_arb);
1945 * process_one_work - process single work
1947 * @work: work to process
1949 * Process @work. This function contains all the logics necessary to
1950 * process a single work including synchronization against and
1951 * interaction with other workers on the same cpu, queueing and
1952 * flushing. As long as context requirement is met, any worker can
1953 * call this function to process a work.
1956 * spin_lock_irq(pool->lock) which is released and regrabbed.
1958 static void process_one_work(struct worker *worker, struct work_struct *work)
1959 __releases(&pool->lock)
1960 __acquires(&pool->lock)
1962 struct pool_workqueue *pwq = get_work_pwq(work);
1963 struct worker_pool *pool = worker->pool;
1964 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1966 struct worker *collision;
1967 #ifdef CONFIG_LOCKDEP
1969 * It is permissible to free the struct work_struct from
1970 * inside the function that is called from it, this we need to
1971 * take into account for lockdep too. To avoid bogus "held
1972 * lock freed" warnings as well as problems when looking into
1973 * work->lockdep_map, make a copy and use that here.
1975 struct lockdep_map lockdep_map;
1977 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1979 /* ensure we're on the correct CPU */
1980 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1981 raw_smp_processor_id() != pool->cpu);
1984 * A single work shouldn't be executed concurrently by
1985 * multiple workers on a single cpu. Check whether anyone is
1986 * already processing the work. If so, defer the work to the
1987 * currently executing one.
1989 collision = find_worker_executing_work(pool, work);
1990 if (unlikely(collision)) {
1991 move_linked_works(work, &collision->scheduled, NULL);
1995 /* claim and dequeue */
1996 debug_work_deactivate(work);
1997 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
1998 worker->current_work = work;
1999 worker->current_func = work->func;
2000 worker->current_pwq = pwq;
2001 work_color = get_work_color(work);
2003 list_del_init(&work->entry);
2006 * CPU intensive works don't participate in concurrency management.
2007 * They're the scheduler's responsibility. This takes @worker out
2008 * of concurrency management and the next code block will chain
2009 * execution of the pending work items.
2011 if (unlikely(cpu_intensive))
2012 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2015 * Wake up another worker if necessary. The condition is always
2016 * false for normal per-cpu workers since nr_running would always
2017 * be >= 1 at this point. This is used to chain execution of the
2018 * pending work items for WORKER_NOT_RUNNING workers such as the
2019 * UNBOUND and CPU_INTENSIVE ones.
2021 if (need_more_worker(pool))
2022 wake_up_worker(pool);
2025 * Record the last pool and clear PENDING which should be the last
2026 * update to @work. Also, do this inside @pool->lock so that
2027 * PENDING and queued state changes happen together while IRQ is
2030 set_work_pool_and_clear_pending(work, pool->id);
2032 spin_unlock_irq(&pool->lock);
2034 lock_map_acquire_read(&pwq->wq->lockdep_map);
2035 lock_map_acquire(&lockdep_map);
2036 trace_workqueue_execute_start(work);
2037 worker->current_func(work);
2039 * While we must be careful to not use "work" after this, the trace
2040 * point will only record its address.
2042 trace_workqueue_execute_end(work);
2043 lock_map_release(&lockdep_map);
2044 lock_map_release(&pwq->wq->lockdep_map);
2046 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2047 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2048 " last function: %pf\n",
2049 current->comm, preempt_count(), task_pid_nr(current),
2050 worker->current_func);
2051 debug_show_held_locks(current);
2056 * The following prevents a kworker from hogging CPU on !PREEMPT
2057 * kernels, where a requeueing work item waiting for something to
2058 * happen could deadlock with stop_machine as such work item could
2059 * indefinitely requeue itself while all other CPUs are trapped in
2060 * stop_machine. At the same time, report a quiescent RCU state so
2061 * the same condition doesn't freeze RCU.
2063 cond_resched_rcu_qs();
2065 spin_lock_irq(&pool->lock);
2067 /* clear cpu intensive status */
2068 if (unlikely(cpu_intensive))
2069 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2071 /* we're done with it, release */
2072 hash_del(&worker->hentry);
2073 worker->current_work = NULL;
2074 worker->current_func = NULL;
2075 worker->current_pwq = NULL;
2076 worker->desc_valid = false;
2077 pwq_dec_nr_in_flight(pwq, work_color);
2081 * process_scheduled_works - process scheduled works
2084 * Process all scheduled works. Please note that the scheduled list
2085 * may change while processing a work, so this function repeatedly
2086 * fetches a work from the top and executes it.
2089 * spin_lock_irq(pool->lock) which may be released and regrabbed
2092 static void process_scheduled_works(struct worker *worker)
2094 while (!list_empty(&worker->scheduled)) {
2095 struct work_struct *work = list_first_entry(&worker->scheduled,
2096 struct work_struct, entry);
2097 process_one_work(worker, work);
2102 * worker_thread - the worker thread function
2105 * The worker thread function. All workers belong to a worker_pool -
2106 * either a per-cpu one or dynamic unbound one. These workers process all
2107 * work items regardless of their specific target workqueue. The only
2108 * exception is work items which belong to workqueues with a rescuer which
2109 * will be explained in rescuer_thread().
2113 static int worker_thread(void *__worker)
2115 struct worker *worker = __worker;
2116 struct worker_pool *pool = worker->pool;
2118 /* tell the scheduler that this is a workqueue worker */
2119 worker->task->flags |= PF_WQ_WORKER;
2121 spin_lock_irq(&pool->lock);
2123 /* am I supposed to die? */
2124 if (unlikely(worker->flags & WORKER_DIE)) {
2125 spin_unlock_irq(&pool->lock);
2126 WARN_ON_ONCE(!list_empty(&worker->entry));
2127 worker->task->flags &= ~PF_WQ_WORKER;
2129 set_task_comm(worker->task, "kworker/dying");
2130 ida_simple_remove(&pool->worker_ida, worker->id);
2131 worker_detach_from_pool(worker, pool);
2136 worker_leave_idle(worker);
2138 /* no more worker necessary? */
2139 if (!need_more_worker(pool))
2142 /* do we need to manage? */
2143 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2147 * ->scheduled list can only be filled while a worker is
2148 * preparing to process a work or actually processing it.
2149 * Make sure nobody diddled with it while I was sleeping.
2151 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2154 * Finish PREP stage. We're guaranteed to have at least one idle
2155 * worker or that someone else has already assumed the manager
2156 * role. This is where @worker starts participating in concurrency
2157 * management if applicable and concurrency management is restored
2158 * after being rebound. See rebind_workers() for details.
2160 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2163 struct work_struct *work =
2164 list_first_entry(&pool->worklist,
2165 struct work_struct, entry);
2167 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2168 /* optimization path, not strictly necessary */
2169 process_one_work(worker, work);
2170 if (unlikely(!list_empty(&worker->scheduled)))
2171 process_scheduled_works(worker);
2173 move_linked_works(work, &worker->scheduled, NULL);
2174 process_scheduled_works(worker);
2176 } while (keep_working(pool));
2178 worker_set_flags(worker, WORKER_PREP);
2181 * pool->lock is held and there's no work to process and no need to
2182 * manage, sleep. Workers are woken up only while holding
2183 * pool->lock or from local cpu, so setting the current state
2184 * before releasing pool->lock is enough to prevent losing any
2187 worker_enter_idle(worker);
2188 __set_current_state(TASK_INTERRUPTIBLE);
2189 spin_unlock_irq(&pool->lock);
2195 * rescuer_thread - the rescuer thread function
2198 * Workqueue rescuer thread function. There's one rescuer for each
2199 * workqueue which has WQ_MEM_RECLAIM set.
2201 * Regular work processing on a pool may block trying to create a new
2202 * worker which uses GFP_KERNEL allocation which has slight chance of
2203 * developing into deadlock if some works currently on the same queue
2204 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2205 * the problem rescuer solves.
2207 * When such condition is possible, the pool summons rescuers of all
2208 * workqueues which have works queued on the pool and let them process
2209 * those works so that forward progress can be guaranteed.
2211 * This should happen rarely.
2215 static int rescuer_thread(void *__rescuer)
2217 struct worker *rescuer = __rescuer;
2218 struct workqueue_struct *wq = rescuer->rescue_wq;
2219 struct list_head *scheduled = &rescuer->scheduled;
2222 set_user_nice(current, RESCUER_NICE_LEVEL);
2225 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2226 * doesn't participate in concurrency management.
2228 rescuer->task->flags |= PF_WQ_WORKER;
2230 set_current_state(TASK_INTERRUPTIBLE);
2233 * By the time the rescuer is requested to stop, the workqueue
2234 * shouldn't have any work pending, but @wq->maydays may still have
2235 * pwq(s) queued. This can happen by non-rescuer workers consuming
2236 * all the work items before the rescuer got to them. Go through
2237 * @wq->maydays processing before acting on should_stop so that the
2238 * list is always empty on exit.
2240 should_stop = kthread_should_stop();
2242 /* see whether any pwq is asking for help */
2243 spin_lock_irq(&wq_mayday_lock);
2245 while (!list_empty(&wq->maydays)) {
2246 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2247 struct pool_workqueue, mayday_node);
2248 struct worker_pool *pool = pwq->pool;
2249 struct work_struct *work, *n;
2251 __set_current_state(TASK_RUNNING);
2252 list_del_init(&pwq->mayday_node);
2254 spin_unlock_irq(&wq_mayday_lock);
2256 worker_attach_to_pool(rescuer, pool);
2258 spin_lock_irq(&pool->lock);
2259 rescuer->pool = pool;
2262 * Slurp in all works issued via this workqueue and
2265 WARN_ON_ONCE(!list_empty(scheduled));
2266 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2267 if (get_work_pwq(work) == pwq)
2268 move_linked_works(work, scheduled, &n);
2270 if (!list_empty(scheduled)) {
2271 process_scheduled_works(rescuer);
2274 * The above execution of rescued work items could
2275 * have created more to rescue through
2276 * pwq_activate_first_delayed() or chained
2277 * queueing. Let's put @pwq back on mayday list so
2278 * that such back-to-back work items, which may be
2279 * being used to relieve memory pressure, don't
2280 * incur MAYDAY_INTERVAL delay inbetween.
2282 if (need_to_create_worker(pool)) {
2283 spin_lock(&wq_mayday_lock);
2285 list_move_tail(&pwq->mayday_node, &wq->maydays);
2286 spin_unlock(&wq_mayday_lock);
2291 * Put the reference grabbed by send_mayday(). @pool won't
2292 * go away while we're still attached to it.
2297 * Leave this pool. If need_more_worker() is %true, notify a
2298 * regular worker; otherwise, we end up with 0 concurrency
2299 * and stalling the execution.
2301 if (need_more_worker(pool))
2302 wake_up_worker(pool);
2304 rescuer->pool = NULL;
2305 spin_unlock_irq(&pool->lock);
2307 worker_detach_from_pool(rescuer, pool);
2309 spin_lock_irq(&wq_mayday_lock);
2312 spin_unlock_irq(&wq_mayday_lock);
2315 __set_current_state(TASK_RUNNING);
2316 rescuer->task->flags &= ~PF_WQ_WORKER;
2320 /* rescuers should never participate in concurrency management */
2321 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2327 struct work_struct work;
2328 struct completion done;
2329 struct task_struct *task; /* purely informational */
2332 static void wq_barrier_func(struct work_struct *work)
2334 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2335 complete(&barr->done);
2339 * insert_wq_barrier - insert a barrier work
2340 * @pwq: pwq to insert barrier into
2341 * @barr: wq_barrier to insert
2342 * @target: target work to attach @barr to
2343 * @worker: worker currently executing @target, NULL if @target is not executing
2345 * @barr is linked to @target such that @barr is completed only after
2346 * @target finishes execution. Please note that the ordering
2347 * guarantee is observed only with respect to @target and on the local
2350 * Currently, a queued barrier can't be canceled. This is because
2351 * try_to_grab_pending() can't determine whether the work to be
2352 * grabbed is at the head of the queue and thus can't clear LINKED
2353 * flag of the previous work while there must be a valid next work
2354 * after a work with LINKED flag set.
2356 * Note that when @worker is non-NULL, @target may be modified
2357 * underneath us, so we can't reliably determine pwq from @target.
2360 * spin_lock_irq(pool->lock).
2362 static void insert_wq_barrier(struct pool_workqueue *pwq,
2363 struct wq_barrier *barr,
2364 struct work_struct *target, struct worker *worker)
2366 struct list_head *head;
2367 unsigned int linked = 0;
2370 * debugobject calls are safe here even with pool->lock locked
2371 * as we know for sure that this will not trigger any of the
2372 * checks and call back into the fixup functions where we
2375 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2376 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2377 init_completion(&barr->done);
2378 barr->task = current;
2381 * If @target is currently being executed, schedule the
2382 * barrier to the worker; otherwise, put it after @target.
2385 head = worker->scheduled.next;
2387 unsigned long *bits = work_data_bits(target);
2389 head = target->entry.next;
2390 /* there can already be other linked works, inherit and set */
2391 linked = *bits & WORK_STRUCT_LINKED;
2392 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2395 debug_work_activate(&barr->work);
2396 insert_work(pwq, &barr->work, head,
2397 work_color_to_flags(WORK_NO_COLOR) | linked);
2401 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2402 * @wq: workqueue being flushed
2403 * @flush_color: new flush color, < 0 for no-op
2404 * @work_color: new work color, < 0 for no-op
2406 * Prepare pwqs for workqueue flushing.
2408 * If @flush_color is non-negative, flush_color on all pwqs should be
2409 * -1. If no pwq has in-flight commands at the specified color, all
2410 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2411 * has in flight commands, its pwq->flush_color is set to
2412 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2413 * wakeup logic is armed and %true is returned.
2415 * The caller should have initialized @wq->first_flusher prior to
2416 * calling this function with non-negative @flush_color. If
2417 * @flush_color is negative, no flush color update is done and %false
2420 * If @work_color is non-negative, all pwqs should have the same
2421 * work_color which is previous to @work_color and all will be
2422 * advanced to @work_color.
2425 * mutex_lock(wq->mutex).
2428 * %true if @flush_color >= 0 and there's something to flush. %false
2431 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2432 int flush_color, int work_color)
2435 struct pool_workqueue *pwq;
2437 if (flush_color >= 0) {
2438 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2439 atomic_set(&wq->nr_pwqs_to_flush, 1);
2442 for_each_pwq(pwq, wq) {
2443 struct worker_pool *pool = pwq->pool;
2445 spin_lock_irq(&pool->lock);
2447 if (flush_color >= 0) {
2448 WARN_ON_ONCE(pwq->flush_color != -1);
2450 if (pwq->nr_in_flight[flush_color]) {
2451 pwq->flush_color = flush_color;
2452 atomic_inc(&wq->nr_pwqs_to_flush);
2457 if (work_color >= 0) {
2458 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2459 pwq->work_color = work_color;
2462 spin_unlock_irq(&pool->lock);
2465 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2466 complete(&wq->first_flusher->done);
2472 * flush_workqueue - ensure that any scheduled work has run to completion.
2473 * @wq: workqueue to flush
2475 * This function sleeps until all work items which were queued on entry
2476 * have finished execution, but it is not livelocked by new incoming ones.
2478 void flush_workqueue(struct workqueue_struct *wq)
2480 struct wq_flusher this_flusher = {
2481 .list = LIST_HEAD_INIT(this_flusher.list),
2483 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2487 lock_map_acquire(&wq->lockdep_map);
2488 lock_map_release(&wq->lockdep_map);
2490 mutex_lock(&wq->mutex);
2493 * Start-to-wait phase
2495 next_color = work_next_color(wq->work_color);
2497 if (next_color != wq->flush_color) {
2499 * Color space is not full. The current work_color
2500 * becomes our flush_color and work_color is advanced
2503 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2504 this_flusher.flush_color = wq->work_color;
2505 wq->work_color = next_color;
2507 if (!wq->first_flusher) {
2508 /* no flush in progress, become the first flusher */
2509 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2511 wq->first_flusher = &this_flusher;
2513 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2515 /* nothing to flush, done */
2516 wq->flush_color = next_color;
2517 wq->first_flusher = NULL;
2522 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2523 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2524 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2528 * Oops, color space is full, wait on overflow queue.
2529 * The next flush completion will assign us
2530 * flush_color and transfer to flusher_queue.
2532 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2535 mutex_unlock(&wq->mutex);
2537 wait_for_completion(&this_flusher.done);
2540 * Wake-up-and-cascade phase
2542 * First flushers are responsible for cascading flushes and
2543 * handling overflow. Non-first flushers can simply return.
2545 if (wq->first_flusher != &this_flusher)
2548 mutex_lock(&wq->mutex);
2550 /* we might have raced, check again with mutex held */
2551 if (wq->first_flusher != &this_flusher)
2554 wq->first_flusher = NULL;
2556 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2557 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2560 struct wq_flusher *next, *tmp;
2562 /* complete all the flushers sharing the current flush color */
2563 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2564 if (next->flush_color != wq->flush_color)
2566 list_del_init(&next->list);
2567 complete(&next->done);
2570 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2571 wq->flush_color != work_next_color(wq->work_color));
2573 /* this flush_color is finished, advance by one */
2574 wq->flush_color = work_next_color(wq->flush_color);
2576 /* one color has been freed, handle overflow queue */
2577 if (!list_empty(&wq->flusher_overflow)) {
2579 * Assign the same color to all overflowed
2580 * flushers, advance work_color and append to
2581 * flusher_queue. This is the start-to-wait
2582 * phase for these overflowed flushers.
2584 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2585 tmp->flush_color = wq->work_color;
2587 wq->work_color = work_next_color(wq->work_color);
2589 list_splice_tail_init(&wq->flusher_overflow,
2590 &wq->flusher_queue);
2591 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2594 if (list_empty(&wq->flusher_queue)) {
2595 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2600 * Need to flush more colors. Make the next flusher
2601 * the new first flusher and arm pwqs.
2603 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2604 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2606 list_del_init(&next->list);
2607 wq->first_flusher = next;
2609 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2613 * Meh... this color is already done, clear first
2614 * flusher and repeat cascading.
2616 wq->first_flusher = NULL;
2620 mutex_unlock(&wq->mutex);
2622 EXPORT_SYMBOL_GPL(flush_workqueue);
2625 * drain_workqueue - drain a workqueue
2626 * @wq: workqueue to drain
2628 * Wait until the workqueue becomes empty. While draining is in progress,
2629 * only chain queueing is allowed. IOW, only currently pending or running
2630 * work items on @wq can queue further work items on it. @wq is flushed
2631 * repeatedly until it becomes empty. The number of flushing is determined
2632 * by the depth of chaining and should be relatively short. Whine if it
2635 void drain_workqueue(struct workqueue_struct *wq)
2637 unsigned int flush_cnt = 0;
2638 struct pool_workqueue *pwq;
2641 * __queue_work() needs to test whether there are drainers, is much
2642 * hotter than drain_workqueue() and already looks at @wq->flags.
2643 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2645 mutex_lock(&wq->mutex);
2646 if (!wq->nr_drainers++)
2647 wq->flags |= __WQ_DRAINING;
2648 mutex_unlock(&wq->mutex);
2650 flush_workqueue(wq);
2652 mutex_lock(&wq->mutex);
2654 for_each_pwq(pwq, wq) {
2657 spin_lock_irq(&pwq->pool->lock);
2658 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2659 spin_unlock_irq(&pwq->pool->lock);
2664 if (++flush_cnt == 10 ||
2665 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2666 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2667 wq->name, flush_cnt);
2669 mutex_unlock(&wq->mutex);
2673 if (!--wq->nr_drainers)
2674 wq->flags &= ~__WQ_DRAINING;
2675 mutex_unlock(&wq->mutex);
2677 EXPORT_SYMBOL_GPL(drain_workqueue);
2679 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2681 struct worker *worker = NULL;
2682 struct worker_pool *pool;
2683 struct pool_workqueue *pwq;
2687 local_irq_disable();
2688 pool = get_work_pool(work);
2694 spin_lock(&pool->lock);
2695 /* see the comment in try_to_grab_pending() with the same code */
2696 pwq = get_work_pwq(work);
2698 if (unlikely(pwq->pool != pool))
2701 worker = find_worker_executing_work(pool, work);
2704 pwq = worker->current_pwq;
2707 insert_wq_barrier(pwq, barr, work, worker);
2708 spin_unlock_irq(&pool->lock);
2711 * If @max_active is 1 or rescuer is in use, flushing another work
2712 * item on the same workqueue may lead to deadlock. Make sure the
2713 * flusher is not running on the same workqueue by verifying write
2716 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2717 lock_map_acquire(&pwq->wq->lockdep_map);
2719 lock_map_acquire_read(&pwq->wq->lockdep_map);
2720 lock_map_release(&pwq->wq->lockdep_map);
2724 spin_unlock_irq(&pool->lock);
2729 * flush_work - wait for a work to finish executing the last queueing instance
2730 * @work: the work to flush
2732 * Wait until @work has finished execution. @work is guaranteed to be idle
2733 * on return if it hasn't been requeued since flush started.
2736 * %true if flush_work() waited for the work to finish execution,
2737 * %false if it was already idle.
2739 bool flush_work(struct work_struct *work)
2741 struct wq_barrier barr;
2743 lock_map_acquire(&work->lockdep_map);
2744 lock_map_release(&work->lockdep_map);
2746 if (start_flush_work(work, &barr)) {
2747 wait_for_completion(&barr.done);
2748 destroy_work_on_stack(&barr.work);
2754 EXPORT_SYMBOL_GPL(flush_work);
2758 struct work_struct *work;
2761 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2763 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2765 if (cwait->work != key)
2767 return autoremove_wake_function(wait, mode, sync, key);
2770 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2772 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2773 unsigned long flags;
2777 ret = try_to_grab_pending(work, is_dwork, &flags);
2779 * If someone else is already canceling, wait for it to
2780 * finish. flush_work() doesn't work for PREEMPT_NONE
2781 * because we may get scheduled between @work's completion
2782 * and the other canceling task resuming and clearing
2783 * CANCELING - flush_work() will return false immediately
2784 * as @work is no longer busy, try_to_grab_pending() will
2785 * return -ENOENT as @work is still being canceled and the
2786 * other canceling task won't be able to clear CANCELING as
2787 * we're hogging the CPU.
2789 * Let's wait for completion using a waitqueue. As this
2790 * may lead to the thundering herd problem, use a custom
2791 * wake function which matches @work along with exclusive
2794 if (unlikely(ret == -ENOENT)) {
2795 struct cwt_wait cwait;
2797 init_wait(&cwait.wait);
2798 cwait.wait.func = cwt_wakefn;
2801 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2802 TASK_UNINTERRUPTIBLE);
2803 if (work_is_canceling(work))
2805 finish_wait(&cancel_waitq, &cwait.wait);
2807 } while (unlikely(ret < 0));
2809 /* tell other tasks trying to grab @work to back off */
2810 mark_work_canceling(work);
2811 local_irq_restore(flags);
2814 clear_work_data(work);
2817 * Paired with prepare_to_wait() above so that either
2818 * waitqueue_active() is visible here or !work_is_canceling() is
2822 if (waitqueue_active(&cancel_waitq))
2823 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2829 * cancel_work_sync - cancel a work and wait for it to finish
2830 * @work: the work to cancel
2832 * Cancel @work and wait for its execution to finish. This function
2833 * can be used even if the work re-queues itself or migrates to
2834 * another workqueue. On return from this function, @work is
2835 * guaranteed to be not pending or executing on any CPU.
2837 * cancel_work_sync(&delayed_work->work) must not be used for
2838 * delayed_work's. Use cancel_delayed_work_sync() instead.
2840 * The caller must ensure that the workqueue on which @work was last
2841 * queued can't be destroyed before this function returns.
2844 * %true if @work was pending, %false otherwise.
2846 bool cancel_work_sync(struct work_struct *work)
2848 return __cancel_work_timer(work, false);
2850 EXPORT_SYMBOL_GPL(cancel_work_sync);
2853 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2854 * @dwork: the delayed work to flush
2856 * Delayed timer is cancelled and the pending work is queued for
2857 * immediate execution. Like flush_work(), this function only
2858 * considers the last queueing instance of @dwork.
2861 * %true if flush_work() waited for the work to finish execution,
2862 * %false if it was already idle.
2864 bool flush_delayed_work(struct delayed_work *dwork)
2866 local_irq_disable();
2867 if (del_timer_sync(&dwork->timer))
2868 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2870 return flush_work(&dwork->work);
2872 EXPORT_SYMBOL(flush_delayed_work);
2875 * cancel_delayed_work - cancel a delayed work
2876 * @dwork: delayed_work to cancel
2878 * Kill off a pending delayed_work.
2880 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2884 * The work callback function may still be running on return, unless
2885 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2886 * use cancel_delayed_work_sync() to wait on it.
2888 * This function is safe to call from any context including IRQ handler.
2890 bool cancel_delayed_work(struct delayed_work *dwork)
2892 unsigned long flags;
2896 ret = try_to_grab_pending(&dwork->work, true, &flags);
2897 } while (unlikely(ret == -EAGAIN));
2899 if (unlikely(ret < 0))
2902 set_work_pool_and_clear_pending(&dwork->work,
2903 get_work_pool_id(&dwork->work));
2904 local_irq_restore(flags);
2907 EXPORT_SYMBOL(cancel_delayed_work);
2910 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2911 * @dwork: the delayed work cancel
2913 * This is cancel_work_sync() for delayed works.
2916 * %true if @dwork was pending, %false otherwise.
2918 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2920 return __cancel_work_timer(&dwork->work, true);
2922 EXPORT_SYMBOL(cancel_delayed_work_sync);
2925 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2926 * @func: the function to call
2928 * schedule_on_each_cpu() executes @func on each online CPU using the
2929 * system workqueue and blocks until all CPUs have completed.
2930 * schedule_on_each_cpu() is very slow.
2933 * 0 on success, -errno on failure.
2935 int schedule_on_each_cpu(work_func_t func)
2938 struct work_struct __percpu *works;
2940 works = alloc_percpu(struct work_struct);
2946 for_each_online_cpu(cpu) {
2947 struct work_struct *work = per_cpu_ptr(works, cpu);
2949 INIT_WORK(work, func);
2950 schedule_work_on(cpu, work);
2953 for_each_online_cpu(cpu)
2954 flush_work(per_cpu_ptr(works, cpu));
2962 * execute_in_process_context - reliably execute the routine with user context
2963 * @fn: the function to execute
2964 * @ew: guaranteed storage for the execute work structure (must
2965 * be available when the work executes)
2967 * Executes the function immediately if process context is available,
2968 * otherwise schedules the function for delayed execution.
2970 * Return: 0 - function was executed
2971 * 1 - function was scheduled for execution
2973 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2975 if (!in_interrupt()) {
2980 INIT_WORK(&ew->work, fn);
2981 schedule_work(&ew->work);
2985 EXPORT_SYMBOL_GPL(execute_in_process_context);
2988 * free_workqueue_attrs - free a workqueue_attrs
2989 * @attrs: workqueue_attrs to free
2991 * Undo alloc_workqueue_attrs().
2993 void free_workqueue_attrs(struct workqueue_attrs *attrs)
2996 free_cpumask_var(attrs->cpumask);
3002 * alloc_workqueue_attrs - allocate a workqueue_attrs
3003 * @gfp_mask: allocation mask to use
3005 * Allocate a new workqueue_attrs, initialize with default settings and
3008 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3010 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3012 struct workqueue_attrs *attrs;
3014 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3017 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3020 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3023 free_workqueue_attrs(attrs);
3027 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3028 const struct workqueue_attrs *from)
3030 to->nice = from->nice;
3031 cpumask_copy(to->cpumask, from->cpumask);
3033 * Unlike hash and equality test, this function doesn't ignore
3034 * ->no_numa as it is used for both pool and wq attrs. Instead,
3035 * get_unbound_pool() explicitly clears ->no_numa after copying.
3037 to->no_numa = from->no_numa;
3040 /* hash value of the content of @attr */
3041 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3045 hash = jhash_1word(attrs->nice, hash);
3046 hash = jhash(cpumask_bits(attrs->cpumask),
3047 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3051 /* content equality test */
3052 static bool wqattrs_equal(const struct workqueue_attrs *a,
3053 const struct workqueue_attrs *b)
3055 if (a->nice != b->nice)
3057 if (!cpumask_equal(a->cpumask, b->cpumask))
3063 * init_worker_pool - initialize a newly zalloc'd worker_pool
3064 * @pool: worker_pool to initialize
3066 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3068 * Return: 0 on success, -errno on failure. Even on failure, all fields
3069 * inside @pool proper are initialized and put_unbound_pool() can be called
3070 * on @pool safely to release it.
3072 static int init_worker_pool(struct worker_pool *pool)
3074 spin_lock_init(&pool->lock);
3077 pool->node = NUMA_NO_NODE;
3078 pool->flags |= POOL_DISASSOCIATED;
3079 INIT_LIST_HEAD(&pool->worklist);
3080 INIT_LIST_HEAD(&pool->idle_list);
3081 hash_init(pool->busy_hash);
3083 init_timer_deferrable(&pool->idle_timer);
3084 pool->idle_timer.function = idle_worker_timeout;
3085 pool->idle_timer.data = (unsigned long)pool;
3087 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3088 (unsigned long)pool);
3090 mutex_init(&pool->manager_arb);
3091 mutex_init(&pool->attach_mutex);
3092 INIT_LIST_HEAD(&pool->workers);
3094 ida_init(&pool->worker_ida);
3095 INIT_HLIST_NODE(&pool->hash_node);
3098 /* shouldn't fail above this point */
3099 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3105 static void rcu_free_wq(struct rcu_head *rcu)
3107 struct workqueue_struct *wq =
3108 container_of(rcu, struct workqueue_struct, rcu);
3110 if (!(wq->flags & WQ_UNBOUND))
3111 free_percpu(wq->cpu_pwqs);
3113 free_workqueue_attrs(wq->unbound_attrs);
3119 static void rcu_free_pool(struct rcu_head *rcu)
3121 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3123 ida_destroy(&pool->worker_ida);
3124 free_workqueue_attrs(pool->attrs);
3129 * put_unbound_pool - put a worker_pool
3130 * @pool: worker_pool to put
3132 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3133 * safe manner. get_unbound_pool() calls this function on its failure path
3134 * and this function should be able to release pools which went through,
3135 * successfully or not, init_worker_pool().
3137 * Should be called with wq_pool_mutex held.
3139 static void put_unbound_pool(struct worker_pool *pool)
3141 DECLARE_COMPLETION_ONSTACK(detach_completion);
3142 struct worker *worker;
3144 lockdep_assert_held(&wq_pool_mutex);
3150 if (WARN_ON(!(pool->cpu < 0)) ||
3151 WARN_ON(!list_empty(&pool->worklist)))
3154 /* release id and unhash */
3156 idr_remove(&worker_pool_idr, pool->id);
3157 hash_del(&pool->hash_node);
3160 * Become the manager and destroy all workers. Grabbing
3161 * manager_arb prevents @pool's workers from blocking on
3164 mutex_lock(&pool->manager_arb);
3166 spin_lock_irq(&pool->lock);
3167 while ((worker = first_idle_worker(pool)))
3168 destroy_worker(worker);
3169 WARN_ON(pool->nr_workers || pool->nr_idle);
3170 spin_unlock_irq(&pool->lock);
3172 mutex_lock(&pool->attach_mutex);
3173 if (!list_empty(&pool->workers))
3174 pool->detach_completion = &detach_completion;
3175 mutex_unlock(&pool->attach_mutex);
3177 if (pool->detach_completion)
3178 wait_for_completion(pool->detach_completion);
3180 mutex_unlock(&pool->manager_arb);
3182 /* shut down the timers */
3183 del_timer_sync(&pool->idle_timer);
3184 del_timer_sync(&pool->mayday_timer);
3186 /* sched-RCU protected to allow dereferences from get_work_pool() */
3187 call_rcu_sched(&pool->rcu, rcu_free_pool);
3191 * get_unbound_pool - get a worker_pool with the specified attributes
3192 * @attrs: the attributes of the worker_pool to get
3194 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3195 * reference count and return it. If there already is a matching
3196 * worker_pool, it will be used; otherwise, this function attempts to
3199 * Should be called with wq_pool_mutex held.
3201 * Return: On success, a worker_pool with the same attributes as @attrs.
3202 * On failure, %NULL.
3204 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3206 u32 hash = wqattrs_hash(attrs);
3207 struct worker_pool *pool;
3210 lockdep_assert_held(&wq_pool_mutex);
3212 /* do we already have a matching pool? */
3213 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3214 if (wqattrs_equal(pool->attrs, attrs)) {
3220 /* nope, create a new one */
3221 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3222 if (!pool || init_worker_pool(pool) < 0)
3225 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3226 copy_workqueue_attrs(pool->attrs, attrs);
3229 * no_numa isn't a worker_pool attribute, always clear it. See
3230 * 'struct workqueue_attrs' comments for detail.
3232 pool->attrs->no_numa = false;
3234 /* if cpumask is contained inside a NUMA node, we belong to that node */
3235 if (wq_numa_enabled) {
3236 for_each_node(node) {
3237 if (cpumask_subset(pool->attrs->cpumask,
3238 wq_numa_possible_cpumask[node])) {
3245 if (worker_pool_assign_id(pool) < 0)
3248 /* create and start the initial worker */
3249 if (!create_worker(pool))
3253 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3258 put_unbound_pool(pool);
3262 static void rcu_free_pwq(struct rcu_head *rcu)
3264 kmem_cache_free(pwq_cache,
3265 container_of(rcu, struct pool_workqueue, rcu));
3269 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3270 * and needs to be destroyed.
3272 static void pwq_unbound_release_workfn(struct work_struct *work)
3274 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3275 unbound_release_work);
3276 struct workqueue_struct *wq = pwq->wq;
3277 struct worker_pool *pool = pwq->pool;
3280 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3283 mutex_lock(&wq->mutex);
3284 list_del_rcu(&pwq->pwqs_node);
3285 is_last = list_empty(&wq->pwqs);
3286 mutex_unlock(&wq->mutex);
3288 mutex_lock(&wq_pool_mutex);
3289 put_unbound_pool(pool);
3290 mutex_unlock(&wq_pool_mutex);
3292 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3295 * If we're the last pwq going away, @wq is already dead and no one
3296 * is gonna access it anymore. Schedule RCU free.
3299 call_rcu_sched(&wq->rcu, rcu_free_wq);
3303 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3304 * @pwq: target pool_workqueue
3306 * If @pwq isn't freezing, set @pwq->max_active to the associated
3307 * workqueue's saved_max_active and activate delayed work items
3308 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3310 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3312 struct workqueue_struct *wq = pwq->wq;
3313 bool freezable = wq->flags & WQ_FREEZABLE;
3315 /* for @wq->saved_max_active */
3316 lockdep_assert_held(&wq->mutex);
3318 /* fast exit for non-freezable wqs */
3319 if (!freezable && pwq->max_active == wq->saved_max_active)
3322 spin_lock_irq(&pwq->pool->lock);
3325 * During [un]freezing, the caller is responsible for ensuring that
3326 * this function is called at least once after @workqueue_freezing
3327 * is updated and visible.
3329 if (!freezable || !workqueue_freezing) {
3330 pwq->max_active = wq->saved_max_active;
3332 while (!list_empty(&pwq->delayed_works) &&
3333 pwq->nr_active < pwq->max_active)
3334 pwq_activate_first_delayed(pwq);
3337 * Need to kick a worker after thawed or an unbound wq's
3338 * max_active is bumped. It's a slow path. Do it always.
3340 wake_up_worker(pwq->pool);
3342 pwq->max_active = 0;
3345 spin_unlock_irq(&pwq->pool->lock);
3348 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3349 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3350 struct worker_pool *pool)
3352 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3354 memset(pwq, 0, sizeof(*pwq));
3358 pwq->flush_color = -1;
3360 INIT_LIST_HEAD(&pwq->delayed_works);
3361 INIT_LIST_HEAD(&pwq->pwqs_node);
3362 INIT_LIST_HEAD(&pwq->mayday_node);
3363 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3366 /* sync @pwq with the current state of its associated wq and link it */
3367 static void link_pwq(struct pool_workqueue *pwq)
3369 struct workqueue_struct *wq = pwq->wq;
3371 lockdep_assert_held(&wq->mutex);
3373 /* may be called multiple times, ignore if already linked */
3374 if (!list_empty(&pwq->pwqs_node))
3377 /* set the matching work_color */
3378 pwq->work_color = wq->work_color;
3380 /* sync max_active to the current setting */
3381 pwq_adjust_max_active(pwq);
3384 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3387 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3388 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3389 const struct workqueue_attrs *attrs)
3391 struct worker_pool *pool;
3392 struct pool_workqueue *pwq;
3394 lockdep_assert_held(&wq_pool_mutex);
3396 pool = get_unbound_pool(attrs);
3400 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3402 put_unbound_pool(pool);
3406 init_pwq(pwq, wq, pool);
3411 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3412 * @attrs: the wq_attrs of the default pwq of the target workqueue
3413 * @node: the target NUMA node
3414 * @cpu_going_down: if >= 0, the CPU to consider as offline
3415 * @cpumask: outarg, the resulting cpumask
3417 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3418 * @cpu_going_down is >= 0, that cpu is considered offline during
3419 * calculation. The result is stored in @cpumask.
3421 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3422 * enabled and @node has online CPUs requested by @attrs, the returned
3423 * cpumask is the intersection of the possible CPUs of @node and
3426 * The caller is responsible for ensuring that the cpumask of @node stays
3429 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3432 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3433 int cpu_going_down, cpumask_t *cpumask)
3435 if (!wq_numa_enabled || attrs->no_numa)
3438 /* does @node have any online CPUs @attrs wants? */
3439 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3440 if (cpu_going_down >= 0)
3441 cpumask_clear_cpu(cpu_going_down, cpumask);
3443 if (cpumask_empty(cpumask))
3446 /* yeap, return possible CPUs in @node that @attrs wants */
3447 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3448 return !cpumask_equal(cpumask, attrs->cpumask);
3451 cpumask_copy(cpumask, attrs->cpumask);
3455 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3456 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3458 struct pool_workqueue *pwq)
3460 struct pool_workqueue *old_pwq;
3462 lockdep_assert_held(&wq_pool_mutex);
3463 lockdep_assert_held(&wq->mutex);
3465 /* link_pwq() can handle duplicate calls */
3468 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3469 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3473 /* context to store the prepared attrs & pwqs before applying */
3474 struct apply_wqattrs_ctx {
3475 struct workqueue_struct *wq; /* target workqueue */
3476 struct workqueue_attrs *attrs; /* attrs to apply */
3477 struct list_head list; /* queued for batching commit */
3478 struct pool_workqueue *dfl_pwq;
3479 struct pool_workqueue *pwq_tbl[];
3482 /* free the resources after success or abort */
3483 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3489 put_pwq_unlocked(ctx->pwq_tbl[node]);
3490 put_pwq_unlocked(ctx->dfl_pwq);
3492 free_workqueue_attrs(ctx->attrs);
3498 /* allocate the attrs and pwqs for later installation */
3499 static struct apply_wqattrs_ctx *
3500 apply_wqattrs_prepare(struct workqueue_struct *wq,
3501 const struct workqueue_attrs *attrs)
3503 struct apply_wqattrs_ctx *ctx;
3504 struct workqueue_attrs *new_attrs, *tmp_attrs;
3507 lockdep_assert_held(&wq_pool_mutex);
3509 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3512 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3513 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3514 if (!ctx || !new_attrs || !tmp_attrs)
3518 * Calculate the attrs of the default pwq.
3519 * If the user configured cpumask doesn't overlap with the
3520 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3522 copy_workqueue_attrs(new_attrs, attrs);
3523 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3524 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3525 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3528 * We may create multiple pwqs with differing cpumasks. Make a
3529 * copy of @new_attrs which will be modified and used to obtain
3532 copy_workqueue_attrs(tmp_attrs, new_attrs);
3535 * If something goes wrong during CPU up/down, we'll fall back to
3536 * the default pwq covering whole @attrs->cpumask. Always create
3537 * it even if we don't use it immediately.
3539 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3543 for_each_node(node) {
3544 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3545 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3546 if (!ctx->pwq_tbl[node])
3549 ctx->dfl_pwq->refcnt++;
3550 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3554 /* save the user configured attrs and sanitize it. */
3555 copy_workqueue_attrs(new_attrs, attrs);
3556 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3557 ctx->attrs = new_attrs;
3560 free_workqueue_attrs(tmp_attrs);
3564 free_workqueue_attrs(tmp_attrs);
3565 free_workqueue_attrs(new_attrs);
3566 apply_wqattrs_cleanup(ctx);
3570 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3571 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3575 /* all pwqs have been created successfully, let's install'em */
3576 mutex_lock(&ctx->wq->mutex);
3578 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3580 /* save the previous pwq and install the new one */
3582 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3583 ctx->pwq_tbl[node]);
3585 /* @dfl_pwq might not have been used, ensure it's linked */
3586 link_pwq(ctx->dfl_pwq);
3587 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3589 mutex_unlock(&ctx->wq->mutex);
3592 static void apply_wqattrs_lock(void)
3594 /* CPUs should stay stable across pwq creations and installations */
3596 mutex_lock(&wq_pool_mutex);
3599 static void apply_wqattrs_unlock(void)
3601 mutex_unlock(&wq_pool_mutex);
3605 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3606 const struct workqueue_attrs *attrs)
3608 struct apply_wqattrs_ctx *ctx;
3611 /* only unbound workqueues can change attributes */
3612 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3615 /* creating multiple pwqs breaks ordering guarantee */
3616 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3619 ctx = apply_wqattrs_prepare(wq, attrs);
3621 /* the ctx has been prepared successfully, let's commit it */
3623 apply_wqattrs_commit(ctx);
3627 apply_wqattrs_cleanup(ctx);
3633 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3634 * @wq: the target workqueue
3635 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3637 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3638 * machines, this function maps a separate pwq to each NUMA node with
3639 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3640 * NUMA node it was issued on. Older pwqs are released as in-flight work
3641 * items finish. Note that a work item which repeatedly requeues itself
3642 * back-to-back will stay on its current pwq.
3644 * Performs GFP_KERNEL allocations.
3646 * Return: 0 on success and -errno on failure.
3648 int apply_workqueue_attrs(struct workqueue_struct *wq,
3649 const struct workqueue_attrs *attrs)
3653 apply_wqattrs_lock();
3654 ret = apply_workqueue_attrs_locked(wq, attrs);
3655 apply_wqattrs_unlock();
3661 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3662 * @wq: the target workqueue
3663 * @cpu: the CPU coming up or going down
3664 * @online: whether @cpu is coming up or going down
3666 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3667 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3670 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3671 * falls back to @wq->dfl_pwq which may not be optimal but is always
3674 * Note that when the last allowed CPU of a NUMA node goes offline for a
3675 * workqueue with a cpumask spanning multiple nodes, the workers which were
3676 * already executing the work items for the workqueue will lose their CPU
3677 * affinity and may execute on any CPU. This is similar to how per-cpu
3678 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3679 * affinity, it's the user's responsibility to flush the work item from
3682 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3685 int node = cpu_to_node(cpu);
3686 int cpu_off = online ? -1 : cpu;
3687 struct pool_workqueue *old_pwq = NULL, *pwq;
3688 struct workqueue_attrs *target_attrs;
3691 lockdep_assert_held(&wq_pool_mutex);
3693 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3694 wq->unbound_attrs->no_numa)
3698 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3699 * Let's use a preallocated one. The following buf is protected by
3700 * CPU hotplug exclusion.
3702 target_attrs = wq_update_unbound_numa_attrs_buf;
3703 cpumask = target_attrs->cpumask;
3705 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3706 pwq = unbound_pwq_by_node(wq, node);
3709 * Let's determine what needs to be done. If the target cpumask is
3710 * different from the default pwq's, we need to compare it to @pwq's
3711 * and create a new one if they don't match. If the target cpumask
3712 * equals the default pwq's, the default pwq should be used.
3714 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3715 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3721 /* create a new pwq */
3722 pwq = alloc_unbound_pwq(wq, target_attrs);
3724 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3729 /* Install the new pwq. */
3730 mutex_lock(&wq->mutex);
3731 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3735 mutex_lock(&wq->mutex);
3736 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3737 get_pwq(wq->dfl_pwq);
3738 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3739 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3741 mutex_unlock(&wq->mutex);
3742 put_pwq_unlocked(old_pwq);
3745 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3747 bool highpri = wq->flags & WQ_HIGHPRI;
3750 if (!(wq->flags & WQ_UNBOUND)) {
3751 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3755 for_each_possible_cpu(cpu) {
3756 struct pool_workqueue *pwq =
3757 per_cpu_ptr(wq->cpu_pwqs, cpu);
3758 struct worker_pool *cpu_pools =
3759 per_cpu(cpu_worker_pools, cpu);
3761 init_pwq(pwq, wq, &cpu_pools[highpri]);
3763 mutex_lock(&wq->mutex);
3765 mutex_unlock(&wq->mutex);
3768 } else if (wq->flags & __WQ_ORDERED) {
3769 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3770 /* there should only be single pwq for ordering guarantee */
3771 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3772 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3773 "ordering guarantee broken for workqueue %s\n", wq->name);
3776 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3780 static int wq_clamp_max_active(int max_active, unsigned int flags,
3783 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3785 if (max_active < 1 || max_active > lim)
3786 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3787 max_active, name, 1, lim);
3789 return clamp_val(max_active, 1, lim);
3792 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3795 struct lock_class_key *key,
3796 const char *lock_name, ...)
3798 size_t tbl_size = 0;
3800 struct workqueue_struct *wq;
3801 struct pool_workqueue *pwq;
3803 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3804 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3805 flags |= WQ_UNBOUND;
3807 /* allocate wq and format name */
3808 if (flags & WQ_UNBOUND)
3809 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3811 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3815 if (flags & WQ_UNBOUND) {
3816 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3817 if (!wq->unbound_attrs)
3821 va_start(args, lock_name);
3822 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3825 max_active = max_active ?: WQ_DFL_ACTIVE;
3826 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3830 wq->saved_max_active = max_active;
3831 mutex_init(&wq->mutex);
3832 atomic_set(&wq->nr_pwqs_to_flush, 0);
3833 INIT_LIST_HEAD(&wq->pwqs);
3834 INIT_LIST_HEAD(&wq->flusher_queue);
3835 INIT_LIST_HEAD(&wq->flusher_overflow);
3836 INIT_LIST_HEAD(&wq->maydays);
3838 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3839 INIT_LIST_HEAD(&wq->list);
3841 if (alloc_and_link_pwqs(wq) < 0)
3845 * Workqueues which may be used during memory reclaim should
3846 * have a rescuer to guarantee forward progress.
3848 if (flags & WQ_MEM_RECLAIM) {
3849 struct worker *rescuer;
3851 rescuer = alloc_worker(NUMA_NO_NODE);
3855 rescuer->rescue_wq = wq;
3856 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3858 if (IS_ERR(rescuer->task)) {
3863 wq->rescuer = rescuer;
3864 rescuer->task->flags |= PF_NO_SETAFFINITY;
3865 wake_up_process(rescuer->task);
3868 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3872 * wq_pool_mutex protects global freeze state and workqueues list.
3873 * Grab it, adjust max_active and add the new @wq to workqueues
3876 mutex_lock(&wq_pool_mutex);
3878 mutex_lock(&wq->mutex);
3879 for_each_pwq(pwq, wq)
3880 pwq_adjust_max_active(pwq);
3881 mutex_unlock(&wq->mutex);
3883 list_add_tail_rcu(&wq->list, &workqueues);
3885 mutex_unlock(&wq_pool_mutex);
3890 free_workqueue_attrs(wq->unbound_attrs);
3894 destroy_workqueue(wq);
3897 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3900 * destroy_workqueue - safely terminate a workqueue
3901 * @wq: target workqueue
3903 * Safely destroy a workqueue. All work currently pending will be done first.
3905 void destroy_workqueue(struct workqueue_struct *wq)
3907 struct pool_workqueue *pwq;
3910 /* drain it before proceeding with destruction */
3911 drain_workqueue(wq);
3914 mutex_lock(&wq->mutex);
3915 for_each_pwq(pwq, wq) {
3918 for (i = 0; i < WORK_NR_COLORS; i++) {
3919 if (WARN_ON(pwq->nr_in_flight[i])) {
3920 mutex_unlock(&wq->mutex);
3925 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3926 WARN_ON(pwq->nr_active) ||
3927 WARN_ON(!list_empty(&pwq->delayed_works))) {
3928 mutex_unlock(&wq->mutex);
3932 mutex_unlock(&wq->mutex);
3935 * wq list is used to freeze wq, remove from list after
3936 * flushing is complete in case freeze races us.
3938 mutex_lock(&wq_pool_mutex);
3939 list_del_rcu(&wq->list);
3940 mutex_unlock(&wq_pool_mutex);
3942 workqueue_sysfs_unregister(wq);
3945 kthread_stop(wq->rescuer->task);
3947 if (!(wq->flags & WQ_UNBOUND)) {
3949 * The base ref is never dropped on per-cpu pwqs. Directly
3950 * schedule RCU free.
3952 call_rcu_sched(&wq->rcu, rcu_free_wq);
3955 * We're the sole accessor of @wq at this point. Directly
3956 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3957 * @wq will be freed when the last pwq is released.
3959 for_each_node(node) {
3960 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3961 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3962 put_pwq_unlocked(pwq);
3966 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
3967 * put. Don't access it afterwards.
3971 put_pwq_unlocked(pwq);
3974 EXPORT_SYMBOL_GPL(destroy_workqueue);
3977 * workqueue_set_max_active - adjust max_active of a workqueue
3978 * @wq: target workqueue
3979 * @max_active: new max_active value.
3981 * Set max_active of @wq to @max_active.
3984 * Don't call from IRQ context.
3986 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3988 struct pool_workqueue *pwq;
3990 /* disallow meddling with max_active for ordered workqueues */
3991 if (WARN_ON(wq->flags & __WQ_ORDERED))
3994 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3996 mutex_lock(&wq->mutex);
3998 wq->saved_max_active = max_active;
4000 for_each_pwq(pwq, wq)
4001 pwq_adjust_max_active(pwq);
4003 mutex_unlock(&wq->mutex);
4005 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4008 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4010 * Determine whether %current is a workqueue rescuer. Can be used from
4011 * work functions to determine whether it's being run off the rescuer task.
4013 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4015 bool current_is_workqueue_rescuer(void)
4017 struct worker *worker = current_wq_worker();
4019 return worker && worker->rescue_wq;
4023 * workqueue_congested - test whether a workqueue is congested
4024 * @cpu: CPU in question
4025 * @wq: target workqueue
4027 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4028 * no synchronization around this function and the test result is
4029 * unreliable and only useful as advisory hints or for debugging.
4031 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4032 * Note that both per-cpu and unbound workqueues may be associated with
4033 * multiple pool_workqueues which have separate congested states. A
4034 * workqueue being congested on one CPU doesn't mean the workqueue is also
4035 * contested on other CPUs / NUMA nodes.
4038 * %true if congested, %false otherwise.
4040 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4042 struct pool_workqueue *pwq;
4045 rcu_read_lock_sched();
4047 if (cpu == WORK_CPU_UNBOUND)
4048 cpu = smp_processor_id();
4050 if (!(wq->flags & WQ_UNBOUND))
4051 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4053 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4055 ret = !list_empty(&pwq->delayed_works);
4056 rcu_read_unlock_sched();
4060 EXPORT_SYMBOL_GPL(workqueue_congested);
4063 * work_busy - test whether a work is currently pending or running
4064 * @work: the work to be tested
4066 * Test whether @work is currently pending or running. There is no
4067 * synchronization around this function and the test result is
4068 * unreliable and only useful as advisory hints or for debugging.
4071 * OR'd bitmask of WORK_BUSY_* bits.
4073 unsigned int work_busy(struct work_struct *work)
4075 struct worker_pool *pool;
4076 unsigned long flags;
4077 unsigned int ret = 0;
4079 if (work_pending(work))
4080 ret |= WORK_BUSY_PENDING;
4082 local_irq_save(flags);
4083 pool = get_work_pool(work);
4085 spin_lock(&pool->lock);
4086 if (find_worker_executing_work(pool, work))
4087 ret |= WORK_BUSY_RUNNING;
4088 spin_unlock(&pool->lock);
4090 local_irq_restore(flags);
4094 EXPORT_SYMBOL_GPL(work_busy);
4097 * set_worker_desc - set description for the current work item
4098 * @fmt: printf-style format string
4099 * @...: arguments for the format string
4101 * This function can be called by a running work function to describe what
4102 * the work item is about. If the worker task gets dumped, this
4103 * information will be printed out together to help debugging. The
4104 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4106 void set_worker_desc(const char *fmt, ...)
4108 struct worker *worker = current_wq_worker();
4112 va_start(args, fmt);
4113 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4115 worker->desc_valid = true;
4120 * print_worker_info - print out worker information and description
4121 * @log_lvl: the log level to use when printing
4122 * @task: target task
4124 * If @task is a worker and currently executing a work item, print out the
4125 * name of the workqueue being serviced and worker description set with
4126 * set_worker_desc() by the currently executing work item.
4128 * This function can be safely called on any task as long as the
4129 * task_struct itself is accessible. While safe, this function isn't
4130 * synchronized and may print out mixups or garbages of limited length.
4132 void print_worker_info(const char *log_lvl, struct task_struct *task)
4134 work_func_t *fn = NULL;
4135 char name[WQ_NAME_LEN] = { };
4136 char desc[WORKER_DESC_LEN] = { };
4137 struct pool_workqueue *pwq = NULL;
4138 struct workqueue_struct *wq = NULL;
4139 bool desc_valid = false;
4140 struct worker *worker;
4142 if (!(task->flags & PF_WQ_WORKER))
4146 * This function is called without any synchronization and @task
4147 * could be in any state. Be careful with dereferences.
4149 worker = probe_kthread_data(task);
4152 * Carefully copy the associated workqueue's workfn and name. Keep
4153 * the original last '\0' in case the original contains garbage.
4155 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4156 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4157 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4158 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4160 /* copy worker description */
4161 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4163 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4165 if (fn || name[0] || desc[0]) {
4166 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4168 pr_cont(" (%s)", desc);
4173 static void pr_cont_pool_info(struct worker_pool *pool)
4175 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4176 if (pool->node != NUMA_NO_NODE)
4177 pr_cont(" node=%d", pool->node);
4178 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4181 static void pr_cont_work(bool comma, struct work_struct *work)
4183 if (work->func == wq_barrier_func) {
4184 struct wq_barrier *barr;
4186 barr = container_of(work, struct wq_barrier, work);
4188 pr_cont("%s BAR(%d)", comma ? "," : "",
4189 task_pid_nr(barr->task));
4191 pr_cont("%s %pf", comma ? "," : "", work->func);
4195 static void show_pwq(struct pool_workqueue *pwq)
4197 struct worker_pool *pool = pwq->pool;
4198 struct work_struct *work;
4199 struct worker *worker;
4200 bool has_in_flight = false, has_pending = false;
4203 pr_info(" pwq %d:", pool->id);
4204 pr_cont_pool_info(pool);
4206 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4207 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4209 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4210 if (worker->current_pwq == pwq) {
4211 has_in_flight = true;
4215 if (has_in_flight) {
4218 pr_info(" in-flight:");
4219 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4220 if (worker->current_pwq != pwq)
4223 pr_cont("%s %d%s:%pf", comma ? "," : "",
4224 task_pid_nr(worker->task),
4225 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4226 worker->current_func);
4227 list_for_each_entry(work, &worker->scheduled, entry)
4228 pr_cont_work(false, work);
4234 list_for_each_entry(work, &pool->worklist, entry) {
4235 if (get_work_pwq(work) == pwq) {
4243 pr_info(" pending:");
4244 list_for_each_entry(work, &pool->worklist, entry) {
4245 if (get_work_pwq(work) != pwq)
4248 pr_cont_work(comma, work);
4249 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4254 if (!list_empty(&pwq->delayed_works)) {
4257 pr_info(" delayed:");
4258 list_for_each_entry(work, &pwq->delayed_works, entry) {
4259 pr_cont_work(comma, work);
4260 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4267 * show_workqueue_state - dump workqueue state
4269 * Called from a sysrq handler and prints out all busy workqueues and
4272 void show_workqueue_state(void)
4274 struct workqueue_struct *wq;
4275 struct worker_pool *pool;
4276 unsigned long flags;
4279 rcu_read_lock_sched();
4281 pr_info("Showing busy workqueues and worker pools:\n");
4283 list_for_each_entry_rcu(wq, &workqueues, list) {
4284 struct pool_workqueue *pwq;
4287 for_each_pwq(pwq, wq) {
4288 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4296 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4298 for_each_pwq(pwq, wq) {
4299 spin_lock_irqsave(&pwq->pool->lock, flags);
4300 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4302 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4306 for_each_pool(pool, pi) {
4307 struct worker *worker;
4310 spin_lock_irqsave(&pool->lock, flags);
4311 if (pool->nr_workers == pool->nr_idle)
4314 pr_info("pool %d:", pool->id);
4315 pr_cont_pool_info(pool);
4316 pr_cont(" workers=%d", pool->nr_workers);
4318 pr_cont(" manager: %d",
4319 task_pid_nr(pool->manager->task));
4320 list_for_each_entry(worker, &pool->idle_list, entry) {
4321 pr_cont(" %s%d", first ? "idle: " : "",
4322 task_pid_nr(worker->task));
4327 spin_unlock_irqrestore(&pool->lock, flags);
4330 rcu_read_unlock_sched();
4336 * There are two challenges in supporting CPU hotplug. Firstly, there
4337 * are a lot of assumptions on strong associations among work, pwq and
4338 * pool which make migrating pending and scheduled works very
4339 * difficult to implement without impacting hot paths. Secondly,
4340 * worker pools serve mix of short, long and very long running works making
4341 * blocked draining impractical.
4343 * This is solved by allowing the pools to be disassociated from the CPU
4344 * running as an unbound one and allowing it to be reattached later if the
4345 * cpu comes back online.
4348 static void wq_unbind_fn(struct work_struct *work)
4350 int cpu = smp_processor_id();
4351 struct worker_pool *pool;
4352 struct worker *worker;
4354 for_each_cpu_worker_pool(pool, cpu) {
4355 mutex_lock(&pool->attach_mutex);
4356 spin_lock_irq(&pool->lock);
4359 * We've blocked all attach/detach operations. Make all workers
4360 * unbound and set DISASSOCIATED. Before this, all workers
4361 * except for the ones which are still executing works from
4362 * before the last CPU down must be on the cpu. After
4363 * this, they may become diasporas.
4365 for_each_pool_worker(worker, pool)
4366 worker->flags |= WORKER_UNBOUND;
4368 pool->flags |= POOL_DISASSOCIATED;
4370 spin_unlock_irq(&pool->lock);
4371 mutex_unlock(&pool->attach_mutex);
4374 * Call schedule() so that we cross rq->lock and thus can
4375 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4376 * This is necessary as scheduler callbacks may be invoked
4382 * Sched callbacks are disabled now. Zap nr_running.
4383 * After this, nr_running stays zero and need_more_worker()
4384 * and keep_working() are always true as long as the
4385 * worklist is not empty. This pool now behaves as an
4386 * unbound (in terms of concurrency management) pool which
4387 * are served by workers tied to the pool.
4389 atomic_set(&pool->nr_running, 0);
4392 * With concurrency management just turned off, a busy
4393 * worker blocking could lead to lengthy stalls. Kick off
4394 * unbound chain execution of currently pending work items.
4396 spin_lock_irq(&pool->lock);
4397 wake_up_worker(pool);
4398 spin_unlock_irq(&pool->lock);
4403 * rebind_workers - rebind all workers of a pool to the associated CPU
4404 * @pool: pool of interest
4406 * @pool->cpu is coming online. Rebind all workers to the CPU.
4408 static void rebind_workers(struct worker_pool *pool)
4410 struct worker *worker;
4412 lockdep_assert_held(&pool->attach_mutex);
4415 * Restore CPU affinity of all workers. As all idle workers should
4416 * be on the run-queue of the associated CPU before any local
4417 * wake-ups for concurrency management happen, restore CPU affinity
4418 * of all workers first and then clear UNBOUND. As we're called
4419 * from CPU_ONLINE, the following shouldn't fail.
4421 for_each_pool_worker(worker, pool)
4422 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4423 pool->attrs->cpumask) < 0);
4425 spin_lock_irq(&pool->lock);
4426 pool->flags &= ~POOL_DISASSOCIATED;
4428 for_each_pool_worker(worker, pool) {
4429 unsigned int worker_flags = worker->flags;
4432 * A bound idle worker should actually be on the runqueue
4433 * of the associated CPU for local wake-ups targeting it to
4434 * work. Kick all idle workers so that they migrate to the
4435 * associated CPU. Doing this in the same loop as
4436 * replacing UNBOUND with REBOUND is safe as no worker will
4437 * be bound before @pool->lock is released.
4439 if (worker_flags & WORKER_IDLE)
4440 wake_up_process(worker->task);
4443 * We want to clear UNBOUND but can't directly call
4444 * worker_clr_flags() or adjust nr_running. Atomically
4445 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4446 * @worker will clear REBOUND using worker_clr_flags() when
4447 * it initiates the next execution cycle thus restoring
4448 * concurrency management. Note that when or whether
4449 * @worker clears REBOUND doesn't affect correctness.
4451 * ACCESS_ONCE() is necessary because @worker->flags may be
4452 * tested without holding any lock in
4453 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4454 * fail incorrectly leading to premature concurrency
4455 * management operations.
4457 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4458 worker_flags |= WORKER_REBOUND;
4459 worker_flags &= ~WORKER_UNBOUND;
4460 ACCESS_ONCE(worker->flags) = worker_flags;
4463 spin_unlock_irq(&pool->lock);
4467 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4468 * @pool: unbound pool of interest
4469 * @cpu: the CPU which is coming up
4471 * An unbound pool may end up with a cpumask which doesn't have any online
4472 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4473 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4474 * online CPU before, cpus_allowed of all its workers should be restored.
4476 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4478 static cpumask_t cpumask;
4479 struct worker *worker;
4481 lockdep_assert_held(&pool->attach_mutex);
4483 /* is @cpu allowed for @pool? */
4484 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4487 /* is @cpu the only online CPU? */
4488 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4489 if (cpumask_weight(&cpumask) != 1)
4492 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4493 for_each_pool_worker(worker, pool)
4494 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4495 pool->attrs->cpumask) < 0);
4499 * Workqueues should be brought up before normal priority CPU notifiers.
4500 * This will be registered high priority CPU notifier.
4502 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4503 unsigned long action,
4506 int cpu = (unsigned long)hcpu;
4507 struct worker_pool *pool;
4508 struct workqueue_struct *wq;
4511 switch (action & ~CPU_TASKS_FROZEN) {
4512 case CPU_UP_PREPARE:
4513 for_each_cpu_worker_pool(pool, cpu) {
4514 if (pool->nr_workers)
4516 if (!create_worker(pool))
4521 case CPU_DOWN_FAILED:
4523 mutex_lock(&wq_pool_mutex);
4525 for_each_pool(pool, pi) {
4526 mutex_lock(&pool->attach_mutex);
4528 if (pool->cpu == cpu)
4529 rebind_workers(pool);
4530 else if (pool->cpu < 0)
4531 restore_unbound_workers_cpumask(pool, cpu);
4533 mutex_unlock(&pool->attach_mutex);
4536 /* update NUMA affinity of unbound workqueues */
4537 list_for_each_entry(wq, &workqueues, list)
4538 wq_update_unbound_numa(wq, cpu, true);
4540 mutex_unlock(&wq_pool_mutex);
4547 * Workqueues should be brought down after normal priority CPU notifiers.
4548 * This will be registered as low priority CPU notifier.
4550 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4551 unsigned long action,
4554 int cpu = (unsigned long)hcpu;
4555 struct work_struct unbind_work;
4556 struct workqueue_struct *wq;
4558 switch (action & ~CPU_TASKS_FROZEN) {
4559 case CPU_DOWN_PREPARE:
4560 /* unbinding per-cpu workers should happen on the local CPU */
4561 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4562 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4564 /* update NUMA affinity of unbound workqueues */
4565 mutex_lock(&wq_pool_mutex);
4566 list_for_each_entry(wq, &workqueues, list)
4567 wq_update_unbound_numa(wq, cpu, false);
4568 mutex_unlock(&wq_pool_mutex);
4570 /* wait for per-cpu unbinding to finish */
4571 flush_work(&unbind_work);
4572 destroy_work_on_stack(&unbind_work);
4580 struct work_for_cpu {
4581 struct work_struct work;
4587 static void work_for_cpu_fn(struct work_struct *work)
4589 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4591 wfc->ret = wfc->fn(wfc->arg);
4595 * work_on_cpu - run a function in user context on a particular cpu
4596 * @cpu: the cpu to run on
4597 * @fn: the function to run
4598 * @arg: the function arg
4600 * It is up to the caller to ensure that the cpu doesn't go offline.
4601 * The caller must not hold any locks which would prevent @fn from completing.
4603 * Return: The value @fn returns.
4605 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4607 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4609 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4610 schedule_work_on(cpu, &wfc.work);
4611 flush_work(&wfc.work);
4612 destroy_work_on_stack(&wfc.work);
4615 EXPORT_SYMBOL_GPL(work_on_cpu);
4616 #endif /* CONFIG_SMP */
4618 #ifdef CONFIG_FREEZER
4621 * freeze_workqueues_begin - begin freezing workqueues
4623 * Start freezing workqueues. After this function returns, all freezable
4624 * workqueues will queue new works to their delayed_works list instead of
4628 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4630 void freeze_workqueues_begin(void)
4632 struct workqueue_struct *wq;
4633 struct pool_workqueue *pwq;
4635 mutex_lock(&wq_pool_mutex);
4637 WARN_ON_ONCE(workqueue_freezing);
4638 workqueue_freezing = true;
4640 list_for_each_entry(wq, &workqueues, list) {
4641 mutex_lock(&wq->mutex);
4642 for_each_pwq(pwq, wq)
4643 pwq_adjust_max_active(pwq);
4644 mutex_unlock(&wq->mutex);
4647 mutex_unlock(&wq_pool_mutex);
4651 * freeze_workqueues_busy - are freezable workqueues still busy?
4653 * Check whether freezing is complete. This function must be called
4654 * between freeze_workqueues_begin() and thaw_workqueues().
4657 * Grabs and releases wq_pool_mutex.
4660 * %true if some freezable workqueues are still busy. %false if freezing
4663 bool freeze_workqueues_busy(void)
4666 struct workqueue_struct *wq;
4667 struct pool_workqueue *pwq;
4669 mutex_lock(&wq_pool_mutex);
4671 WARN_ON_ONCE(!workqueue_freezing);
4673 list_for_each_entry(wq, &workqueues, list) {
4674 if (!(wq->flags & WQ_FREEZABLE))
4677 * nr_active is monotonically decreasing. It's safe
4678 * to peek without lock.
4680 rcu_read_lock_sched();
4681 for_each_pwq(pwq, wq) {
4682 WARN_ON_ONCE(pwq->nr_active < 0);
4683 if (pwq->nr_active) {
4685 rcu_read_unlock_sched();
4689 rcu_read_unlock_sched();
4692 mutex_unlock(&wq_pool_mutex);
4697 * thaw_workqueues - thaw workqueues
4699 * Thaw workqueues. Normal queueing is restored and all collected
4700 * frozen works are transferred to their respective pool worklists.
4703 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4705 void thaw_workqueues(void)
4707 struct workqueue_struct *wq;
4708 struct pool_workqueue *pwq;
4710 mutex_lock(&wq_pool_mutex);
4712 if (!workqueue_freezing)
4715 workqueue_freezing = false;
4717 /* restore max_active and repopulate worklist */
4718 list_for_each_entry(wq, &workqueues, list) {
4719 mutex_lock(&wq->mutex);
4720 for_each_pwq(pwq, wq)
4721 pwq_adjust_max_active(pwq);
4722 mutex_unlock(&wq->mutex);
4726 mutex_unlock(&wq_pool_mutex);
4728 #endif /* CONFIG_FREEZER */
4730 static int workqueue_apply_unbound_cpumask(void)
4734 struct workqueue_struct *wq;
4735 struct apply_wqattrs_ctx *ctx, *n;
4737 lockdep_assert_held(&wq_pool_mutex);
4739 list_for_each_entry(wq, &workqueues, list) {
4740 if (!(wq->flags & WQ_UNBOUND))
4742 /* creating multiple pwqs breaks ordering guarantee */
4743 if (wq->flags & __WQ_ORDERED)
4746 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4752 list_add_tail(&ctx->list, &ctxs);
4755 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4757 apply_wqattrs_commit(ctx);
4758 apply_wqattrs_cleanup(ctx);
4765 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4766 * @cpumask: the cpumask to set
4768 * The low-level workqueues cpumask is a global cpumask that limits
4769 * the affinity of all unbound workqueues. This function check the @cpumask
4770 * and apply it to all unbound workqueues and updates all pwqs of them.
4772 * Retun: 0 - Success
4773 * -EINVAL - Invalid @cpumask
4774 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4776 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4779 cpumask_var_t saved_cpumask;
4781 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4784 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4785 if (!cpumask_empty(cpumask)) {
4786 apply_wqattrs_lock();
4788 /* save the old wq_unbound_cpumask. */
4789 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4791 /* update wq_unbound_cpumask at first and apply it to wqs. */
4792 cpumask_copy(wq_unbound_cpumask, cpumask);
4793 ret = workqueue_apply_unbound_cpumask();
4795 /* restore the wq_unbound_cpumask when failed. */
4797 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4799 apply_wqattrs_unlock();
4802 free_cpumask_var(saved_cpumask);
4808 * Workqueues with WQ_SYSFS flag set is visible to userland via
4809 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4810 * following attributes.
4812 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4813 * max_active RW int : maximum number of in-flight work items
4815 * Unbound workqueues have the following extra attributes.
4817 * id RO int : the associated pool ID
4818 * nice RW int : nice value of the workers
4819 * cpumask RW mask : bitmask of allowed CPUs for the workers
4822 struct workqueue_struct *wq;
4826 static struct workqueue_struct *dev_to_wq(struct device *dev)
4828 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4833 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4836 struct workqueue_struct *wq = dev_to_wq(dev);
4838 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4840 static DEVICE_ATTR_RO(per_cpu);
4842 static ssize_t max_active_show(struct device *dev,
4843 struct device_attribute *attr, char *buf)
4845 struct workqueue_struct *wq = dev_to_wq(dev);
4847 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4850 static ssize_t max_active_store(struct device *dev,
4851 struct device_attribute *attr, const char *buf,
4854 struct workqueue_struct *wq = dev_to_wq(dev);
4857 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4860 workqueue_set_max_active(wq, val);
4863 static DEVICE_ATTR_RW(max_active);
4865 static struct attribute *wq_sysfs_attrs[] = {
4866 &dev_attr_per_cpu.attr,
4867 &dev_attr_max_active.attr,
4870 ATTRIBUTE_GROUPS(wq_sysfs);
4872 static ssize_t wq_pool_ids_show(struct device *dev,
4873 struct device_attribute *attr, char *buf)
4875 struct workqueue_struct *wq = dev_to_wq(dev);
4876 const char *delim = "";
4877 int node, written = 0;
4879 rcu_read_lock_sched();
4880 for_each_node(node) {
4881 written += scnprintf(buf + written, PAGE_SIZE - written,
4882 "%s%d:%d", delim, node,
4883 unbound_pwq_by_node(wq, node)->pool->id);
4886 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4887 rcu_read_unlock_sched();
4892 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4895 struct workqueue_struct *wq = dev_to_wq(dev);
4898 mutex_lock(&wq->mutex);
4899 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4900 mutex_unlock(&wq->mutex);
4905 /* prepare workqueue_attrs for sysfs store operations */
4906 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4908 struct workqueue_attrs *attrs;
4910 lockdep_assert_held(&wq_pool_mutex);
4912 attrs = alloc_workqueue_attrs(GFP_KERNEL);
4916 copy_workqueue_attrs(attrs, wq->unbound_attrs);
4920 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4921 const char *buf, size_t count)
4923 struct workqueue_struct *wq = dev_to_wq(dev);
4924 struct workqueue_attrs *attrs;
4927 apply_wqattrs_lock();
4929 attrs = wq_sysfs_prep_attrs(wq);
4933 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4934 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4935 ret = apply_workqueue_attrs_locked(wq, attrs);
4940 apply_wqattrs_unlock();
4941 free_workqueue_attrs(attrs);
4942 return ret ?: count;
4945 static ssize_t wq_cpumask_show(struct device *dev,
4946 struct device_attribute *attr, char *buf)
4948 struct workqueue_struct *wq = dev_to_wq(dev);
4951 mutex_lock(&wq->mutex);
4952 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4953 cpumask_pr_args(wq->unbound_attrs->cpumask));
4954 mutex_unlock(&wq->mutex);
4958 static ssize_t wq_cpumask_store(struct device *dev,
4959 struct device_attribute *attr,
4960 const char *buf, size_t count)
4962 struct workqueue_struct *wq = dev_to_wq(dev);
4963 struct workqueue_attrs *attrs;
4966 apply_wqattrs_lock();
4968 attrs = wq_sysfs_prep_attrs(wq);
4972 ret = cpumask_parse(buf, attrs->cpumask);
4974 ret = apply_workqueue_attrs_locked(wq, attrs);
4977 apply_wqattrs_unlock();
4978 free_workqueue_attrs(attrs);
4979 return ret ?: count;
4982 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
4985 struct workqueue_struct *wq = dev_to_wq(dev);
4988 mutex_lock(&wq->mutex);
4989 written = scnprintf(buf, PAGE_SIZE, "%d\n",
4990 !wq->unbound_attrs->no_numa);
4991 mutex_unlock(&wq->mutex);
4996 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
4997 const char *buf, size_t count)
4999 struct workqueue_struct *wq = dev_to_wq(dev);
5000 struct workqueue_attrs *attrs;
5001 int v, ret = -ENOMEM;
5003 apply_wqattrs_lock();
5005 attrs = wq_sysfs_prep_attrs(wq);
5010 if (sscanf(buf, "%d", &v) == 1) {
5011 attrs->no_numa = !v;
5012 ret = apply_workqueue_attrs_locked(wq, attrs);
5016 apply_wqattrs_unlock();
5017 free_workqueue_attrs(attrs);
5018 return ret ?: count;
5021 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5022 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5023 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5024 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5025 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5029 static struct bus_type wq_subsys = {
5030 .name = "workqueue",
5031 .dev_groups = wq_sysfs_groups,
5034 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5035 struct device_attribute *attr, char *buf)
5039 mutex_lock(&wq_pool_mutex);
5040 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5041 cpumask_pr_args(wq_unbound_cpumask));
5042 mutex_unlock(&wq_pool_mutex);
5047 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5048 struct device_attribute *attr, const char *buf, size_t count)
5050 cpumask_var_t cpumask;
5053 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5056 ret = cpumask_parse(buf, cpumask);
5058 ret = workqueue_set_unbound_cpumask(cpumask);
5060 free_cpumask_var(cpumask);
5061 return ret ? ret : count;
5064 static struct device_attribute wq_sysfs_cpumask_attr =
5065 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5066 wq_unbound_cpumask_store);
5068 static int __init wq_sysfs_init(void)
5072 err = subsys_virtual_register(&wq_subsys, NULL);
5076 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5078 core_initcall(wq_sysfs_init);
5080 static void wq_device_release(struct device *dev)
5082 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5088 * workqueue_sysfs_register - make a workqueue visible in sysfs
5089 * @wq: the workqueue to register
5091 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5092 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5093 * which is the preferred method.
5095 * Workqueue user should use this function directly iff it wants to apply
5096 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5097 * apply_workqueue_attrs() may race against userland updating the
5100 * Return: 0 on success, -errno on failure.
5102 int workqueue_sysfs_register(struct workqueue_struct *wq)
5104 struct wq_device *wq_dev;
5108 * Adjusting max_active or creating new pwqs by applying
5109 * attributes breaks ordering guarantee. Disallow exposing ordered
5112 if (WARN_ON(wq->flags & __WQ_ORDERED))
5115 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5120 wq_dev->dev.bus = &wq_subsys;
5121 wq_dev->dev.init_name = wq->name;
5122 wq_dev->dev.release = wq_device_release;
5125 * unbound_attrs are created separately. Suppress uevent until
5126 * everything is ready.
5128 dev_set_uevent_suppress(&wq_dev->dev, true);
5130 ret = device_register(&wq_dev->dev);
5137 if (wq->flags & WQ_UNBOUND) {
5138 struct device_attribute *attr;
5140 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5141 ret = device_create_file(&wq_dev->dev, attr);
5143 device_unregister(&wq_dev->dev);
5150 dev_set_uevent_suppress(&wq_dev->dev, false);
5151 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5156 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5157 * @wq: the workqueue to unregister
5159 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5161 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5163 struct wq_device *wq_dev = wq->wq_dev;
5169 device_unregister(&wq_dev->dev);
5171 #else /* CONFIG_SYSFS */
5172 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5173 #endif /* CONFIG_SYSFS */
5175 static void __init wq_numa_init(void)
5180 if (num_possible_nodes() <= 1)
5183 if (wq_disable_numa) {
5184 pr_info("workqueue: NUMA affinity support disabled\n");
5188 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5189 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5192 * We want masks of possible CPUs of each node which isn't readily
5193 * available. Build one from cpu_to_node() which should have been
5194 * fully initialized by now.
5196 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5200 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5201 node_online(node) ? node : NUMA_NO_NODE));
5203 for_each_possible_cpu(cpu) {
5204 node = cpu_to_node(cpu);
5205 if (WARN_ON(node == NUMA_NO_NODE)) {
5206 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5207 /* happens iff arch is bonkers, let's just proceed */
5210 cpumask_set_cpu(cpu, tbl[node]);
5213 wq_numa_possible_cpumask = tbl;
5214 wq_numa_enabled = true;
5217 static int __init init_workqueues(void)
5219 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5222 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5224 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5225 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5227 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5229 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5230 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5234 /* initialize CPU pools */
5235 for_each_possible_cpu(cpu) {
5236 struct worker_pool *pool;
5239 for_each_cpu_worker_pool(pool, cpu) {
5240 BUG_ON(init_worker_pool(pool));
5242 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5243 pool->attrs->nice = std_nice[i++];
5244 pool->node = cpu_to_node(cpu);
5247 mutex_lock(&wq_pool_mutex);
5248 BUG_ON(worker_pool_assign_id(pool));
5249 mutex_unlock(&wq_pool_mutex);
5253 /* create the initial worker */
5254 for_each_online_cpu(cpu) {
5255 struct worker_pool *pool;
5257 for_each_cpu_worker_pool(pool, cpu) {
5258 pool->flags &= ~POOL_DISASSOCIATED;
5259 BUG_ON(!create_worker(pool));
5263 /* create default unbound and ordered wq attrs */
5264 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5265 struct workqueue_attrs *attrs;
5267 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5268 attrs->nice = std_nice[i];
5269 unbound_std_wq_attrs[i] = attrs;
5272 * An ordered wq should have only one pwq as ordering is
5273 * guaranteed by max_active which is enforced by pwqs.
5274 * Turn off NUMA so that dfl_pwq is used for all nodes.
5276 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5277 attrs->nice = std_nice[i];
5278 attrs->no_numa = true;
5279 ordered_wq_attrs[i] = attrs;
5282 system_wq = alloc_workqueue("events", 0, 0);
5283 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5284 system_long_wq = alloc_workqueue("events_long", 0, 0);
5285 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5286 WQ_UNBOUND_MAX_ACTIVE);
5287 system_freezable_wq = alloc_workqueue("events_freezable",
5289 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5290 WQ_POWER_EFFICIENT, 0);
5291 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5292 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5294 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5295 !system_unbound_wq || !system_freezable_wq ||
5296 !system_power_efficient_wq ||
5297 !system_freezable_power_efficient_wq);
5300 early_initcall(init_workqueues);