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 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
289 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
291 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
293 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
294 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
296 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
297 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
299 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
300 static bool workqueue_freezing; /* PL: have wqs started freezing? */
302 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
304 /* the per-cpu worker pools */
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
308 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
310 /* PL: hash of all unbound pools keyed by pool->attrs */
311 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
313 /* I: attributes used when instantiating standard unbound pools on demand */
314 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
316 /* I: attributes used when instantiating ordered pools on demand */
317 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
319 struct workqueue_struct *system_wq __read_mostly;
320 EXPORT_SYMBOL(system_wq);
321 struct workqueue_struct *system_highpri_wq __read_mostly;
322 EXPORT_SYMBOL_GPL(system_highpri_wq);
323 struct workqueue_struct *system_long_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_long_wq);
325 struct workqueue_struct *system_unbound_wq __read_mostly;
326 EXPORT_SYMBOL_GPL(system_unbound_wq);
327 struct workqueue_struct *system_freezable_wq __read_mostly;
328 EXPORT_SYMBOL_GPL(system_freezable_wq);
329 struct workqueue_struct *system_power_efficient_wq __read_mostly;
330 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
331 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
332 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
334 static int worker_thread(void *__worker);
335 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
337 #define CREATE_TRACE_POINTS
338 #include <trace/events/workqueue.h>
340 #define assert_rcu_or_pool_mutex() \
341 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
342 !lockdep_is_held(&wq_pool_mutex), \
343 "sched RCU or wq_pool_mutex should be held")
345 #define assert_rcu_or_wq_mutex(wq) \
346 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
347 !lockdep_is_held(&wq->mutex), \
348 "sched RCU or wq->mutex should be held")
350 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
351 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
352 !lockdep_is_held(&wq->mutex) && \
353 !lockdep_is_held(&wq_pool_mutex), \
354 "sched RCU, wq->mutex or wq_pool_mutex should be held")
356 #define for_each_cpu_worker_pool(pool, cpu) \
357 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
358 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
362 * for_each_pool - iterate through all worker_pools in the system
363 * @pool: iteration cursor
364 * @pi: integer used for iteration
366 * This must be called either with wq_pool_mutex held or sched RCU read
367 * locked. If the pool needs to be used beyond the locking in effect, the
368 * caller is responsible for guaranteeing that the pool stays online.
370 * The if/else clause exists only for the lockdep assertion and can be
373 #define for_each_pool(pool, pi) \
374 idr_for_each_entry(&worker_pool_idr, pool, pi) \
375 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
379 * for_each_pool_worker - iterate through all workers of a worker_pool
380 * @worker: iteration cursor
381 * @pool: worker_pool to iterate workers of
383 * This must be called with @pool->attach_mutex.
385 * The if/else clause exists only for the lockdep assertion and can be
388 #define for_each_pool_worker(worker, pool) \
389 list_for_each_entry((worker), &(pool)->workers, node) \
390 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
394 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
395 * @pwq: iteration cursor
396 * @wq: the target workqueue
398 * This must be called either with wq->mutex held or sched RCU read locked.
399 * If the pwq needs to be used beyond the locking in effect, the caller is
400 * responsible for guaranteeing that the pwq stays online.
402 * The if/else clause exists only for the lockdep assertion and can be
405 #define for_each_pwq(pwq, wq) \
406 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
407 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
410 #ifdef CONFIG_DEBUG_OBJECTS_WORK
412 static struct debug_obj_descr work_debug_descr;
414 static void *work_debug_hint(void *addr)
416 return ((struct work_struct *) addr)->func;
420 * fixup_init is called when:
421 * - an active object is initialized
423 static int work_fixup_init(void *addr, enum debug_obj_state state)
425 struct work_struct *work = addr;
428 case ODEBUG_STATE_ACTIVE:
429 cancel_work_sync(work);
430 debug_object_init(work, &work_debug_descr);
438 * fixup_activate is called when:
439 * - an active object is activated
440 * - an unknown object is activated (might be a statically initialized object)
442 static int work_fixup_activate(void *addr, enum debug_obj_state state)
444 struct work_struct *work = addr;
448 case ODEBUG_STATE_NOTAVAILABLE:
450 * This is not really a fixup. The work struct was
451 * statically initialized. We just make sure that it
452 * is tracked in the object tracker.
454 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
455 debug_object_init(work, &work_debug_descr);
456 debug_object_activate(work, &work_debug_descr);
462 case ODEBUG_STATE_ACTIVE:
471 * fixup_free is called when:
472 * - an active object is freed
474 static int work_fixup_free(void *addr, enum debug_obj_state state)
476 struct work_struct *work = addr;
479 case ODEBUG_STATE_ACTIVE:
480 cancel_work_sync(work);
481 debug_object_free(work, &work_debug_descr);
488 static struct debug_obj_descr work_debug_descr = {
489 .name = "work_struct",
490 .debug_hint = work_debug_hint,
491 .fixup_init = work_fixup_init,
492 .fixup_activate = work_fixup_activate,
493 .fixup_free = work_fixup_free,
496 static inline void debug_work_activate(struct work_struct *work)
498 debug_object_activate(work, &work_debug_descr);
501 static inline void debug_work_deactivate(struct work_struct *work)
503 debug_object_deactivate(work, &work_debug_descr);
506 void __init_work(struct work_struct *work, int onstack)
509 debug_object_init_on_stack(work, &work_debug_descr);
511 debug_object_init(work, &work_debug_descr);
513 EXPORT_SYMBOL_GPL(__init_work);
515 void destroy_work_on_stack(struct work_struct *work)
517 debug_object_free(work, &work_debug_descr);
519 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
521 void destroy_delayed_work_on_stack(struct delayed_work *work)
523 destroy_timer_on_stack(&work->timer);
524 debug_object_free(&work->work, &work_debug_descr);
526 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
529 static inline void debug_work_activate(struct work_struct *work) { }
530 static inline void debug_work_deactivate(struct work_struct *work) { }
534 * worker_pool_assign_id - allocate ID and assing it to @pool
535 * @pool: the pool pointer of interest
537 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
538 * successfully, -errno on failure.
540 static int worker_pool_assign_id(struct worker_pool *pool)
544 lockdep_assert_held(&wq_pool_mutex);
546 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
556 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
557 * @wq: the target workqueue
560 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
562 * If the pwq needs to be used beyond the locking in effect, the caller is
563 * responsible for guaranteeing that the pwq stays online.
565 * Return: The unbound pool_workqueue for @node.
567 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
570 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
573 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
574 * delayed item is pending. The plan is to keep CPU -> NODE
575 * mapping valid and stable across CPU on/offlines. Once that
576 * happens, this workaround can be removed.
578 if (unlikely(node == NUMA_NO_NODE))
581 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
584 static unsigned int work_color_to_flags(int color)
586 return color << WORK_STRUCT_COLOR_SHIFT;
589 static int get_work_color(struct work_struct *work)
591 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
592 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
595 static int work_next_color(int color)
597 return (color + 1) % WORK_NR_COLORS;
601 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
602 * contain the pointer to the queued pwq. Once execution starts, the flag
603 * is cleared and the high bits contain OFFQ flags and pool ID.
605 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
606 * and clear_work_data() can be used to set the pwq, pool or clear
607 * work->data. These functions should only be called while the work is
608 * owned - ie. while the PENDING bit is set.
610 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
611 * corresponding to a work. Pool is available once the work has been
612 * queued anywhere after initialization until it is sync canceled. pwq is
613 * available only while the work item is queued.
615 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
616 * canceled. While being canceled, a work item may have its PENDING set
617 * but stay off timer and worklist for arbitrarily long and nobody should
618 * try to steal the PENDING bit.
620 static inline void set_work_data(struct work_struct *work, unsigned long data,
623 WARN_ON_ONCE(!work_pending(work));
624 atomic_long_set(&work->data, data | flags | work_static(work));
627 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
628 unsigned long extra_flags)
630 set_work_data(work, (unsigned long)pwq,
631 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
634 static void set_work_pool_and_keep_pending(struct work_struct *work,
637 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
638 WORK_STRUCT_PENDING);
641 static void set_work_pool_and_clear_pending(struct work_struct *work,
645 * The following wmb is paired with the implied mb in
646 * test_and_set_bit(PENDING) and ensures all updates to @work made
647 * here are visible to and precede any updates by the next PENDING
651 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
654 static void clear_work_data(struct work_struct *work)
656 smp_wmb(); /* see set_work_pool_and_clear_pending() */
657 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
660 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
662 unsigned long data = atomic_long_read(&work->data);
664 if (data & WORK_STRUCT_PWQ)
665 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
671 * get_work_pool - return the worker_pool a given work was associated with
672 * @work: the work item of interest
674 * Pools are created and destroyed under wq_pool_mutex, and allows read
675 * access under sched-RCU read lock. As such, this function should be
676 * called under wq_pool_mutex or with preemption disabled.
678 * All fields of the returned pool are accessible as long as the above
679 * mentioned locking is in effect. If the returned pool needs to be used
680 * beyond the critical section, the caller is responsible for ensuring the
681 * returned pool is and stays online.
683 * Return: The worker_pool @work was last associated with. %NULL if none.
685 static struct worker_pool *get_work_pool(struct work_struct *work)
687 unsigned long data = atomic_long_read(&work->data);
690 assert_rcu_or_pool_mutex();
692 if (data & WORK_STRUCT_PWQ)
693 return ((struct pool_workqueue *)
694 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
696 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
697 if (pool_id == WORK_OFFQ_POOL_NONE)
700 return idr_find(&worker_pool_idr, pool_id);
704 * get_work_pool_id - return the worker pool ID a given work is associated with
705 * @work: the work item of interest
707 * Return: The worker_pool ID @work was last associated with.
708 * %WORK_OFFQ_POOL_NONE if none.
710 static int get_work_pool_id(struct work_struct *work)
712 unsigned long data = atomic_long_read(&work->data);
714 if (data & WORK_STRUCT_PWQ)
715 return ((struct pool_workqueue *)
716 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
718 return data >> WORK_OFFQ_POOL_SHIFT;
721 static void mark_work_canceling(struct work_struct *work)
723 unsigned long pool_id = get_work_pool_id(work);
725 pool_id <<= WORK_OFFQ_POOL_SHIFT;
726 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
729 static bool work_is_canceling(struct work_struct *work)
731 unsigned long data = atomic_long_read(&work->data);
733 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
737 * Policy functions. These define the policies on how the global worker
738 * pools are managed. Unless noted otherwise, these functions assume that
739 * they're being called with pool->lock held.
742 static bool __need_more_worker(struct worker_pool *pool)
744 return !atomic_read(&pool->nr_running);
748 * Need to wake up a worker? Called from anything but currently
751 * Note that, because unbound workers never contribute to nr_running, this
752 * function will always return %true for unbound pools as long as the
753 * worklist isn't empty.
755 static bool need_more_worker(struct worker_pool *pool)
757 return !list_empty(&pool->worklist) && __need_more_worker(pool);
760 /* Can I start working? Called from busy but !running workers. */
761 static bool may_start_working(struct worker_pool *pool)
763 return pool->nr_idle;
766 /* Do I need to keep working? Called from currently running workers. */
767 static bool keep_working(struct worker_pool *pool)
769 return !list_empty(&pool->worklist) &&
770 atomic_read(&pool->nr_running) <= 1;
773 /* Do we need a new worker? Called from manager. */
774 static bool need_to_create_worker(struct worker_pool *pool)
776 return need_more_worker(pool) && !may_start_working(pool);
779 /* Do we have too many workers and should some go away? */
780 static bool too_many_workers(struct worker_pool *pool)
782 bool managing = mutex_is_locked(&pool->manager_arb);
783 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
784 int nr_busy = pool->nr_workers - nr_idle;
786 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
793 /* Return the first idle worker. Safe with preemption disabled */
794 static struct worker *first_idle_worker(struct worker_pool *pool)
796 if (unlikely(list_empty(&pool->idle_list)))
799 return list_first_entry(&pool->idle_list, struct worker, entry);
803 * wake_up_worker - wake up an idle worker
804 * @pool: worker pool to wake worker from
806 * Wake up the first idle worker of @pool.
809 * spin_lock_irq(pool->lock).
811 static void wake_up_worker(struct worker_pool *pool)
813 struct worker *worker = first_idle_worker(pool);
816 wake_up_process(worker->task);
820 * wq_worker_waking_up - a worker is waking up
821 * @task: task waking up
822 * @cpu: CPU @task is waking up to
824 * This function is called during try_to_wake_up() when a worker is
828 * spin_lock_irq(rq->lock)
830 void wq_worker_waking_up(struct task_struct *task, int cpu)
832 struct worker *worker = kthread_data(task);
834 if (!(worker->flags & WORKER_NOT_RUNNING)) {
835 WARN_ON_ONCE(worker->pool->cpu != cpu);
836 atomic_inc(&worker->pool->nr_running);
841 * wq_worker_sleeping - a worker is going to sleep
842 * @task: task going to sleep
843 * @cpu: CPU in question, must be the current CPU number
845 * This function is called during schedule() when a busy worker is
846 * going to sleep. Worker on the same cpu can be woken up by
847 * returning pointer to its task.
850 * spin_lock_irq(rq->lock)
853 * Worker task on @cpu to wake up, %NULL if none.
855 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
857 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
858 struct worker_pool *pool;
861 * Rescuers, which may not have all the fields set up like normal
862 * workers, also reach here, let's not access anything before
863 * checking NOT_RUNNING.
865 if (worker->flags & WORKER_NOT_RUNNING)
870 /* this can only happen on the local cpu */
871 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
875 * The counterpart of the following dec_and_test, implied mb,
876 * worklist not empty test sequence is in insert_work().
877 * Please read comment there.
879 * NOT_RUNNING is clear. This means that we're bound to and
880 * running on the local cpu w/ rq lock held and preemption
881 * disabled, which in turn means that none else could be
882 * manipulating idle_list, so dereferencing idle_list without pool
885 if (atomic_dec_and_test(&pool->nr_running) &&
886 !list_empty(&pool->worklist))
887 to_wakeup = first_idle_worker(pool);
888 return to_wakeup ? to_wakeup->task : NULL;
892 * worker_set_flags - set worker flags and adjust nr_running accordingly
894 * @flags: flags to set
896 * Set @flags in @worker->flags and adjust nr_running accordingly.
899 * spin_lock_irq(pool->lock)
901 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
903 struct worker_pool *pool = worker->pool;
905 WARN_ON_ONCE(worker->task != current);
907 /* If transitioning into NOT_RUNNING, adjust nr_running. */
908 if ((flags & WORKER_NOT_RUNNING) &&
909 !(worker->flags & WORKER_NOT_RUNNING)) {
910 atomic_dec(&pool->nr_running);
913 worker->flags |= flags;
917 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
919 * @flags: flags to clear
921 * Clear @flags in @worker->flags and adjust nr_running accordingly.
924 * spin_lock_irq(pool->lock)
926 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
928 struct worker_pool *pool = worker->pool;
929 unsigned int oflags = worker->flags;
931 WARN_ON_ONCE(worker->task != current);
933 worker->flags &= ~flags;
936 * If transitioning out of NOT_RUNNING, increment nr_running. Note
937 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
938 * of multiple flags, not a single flag.
940 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
941 if (!(worker->flags & WORKER_NOT_RUNNING))
942 atomic_inc(&pool->nr_running);
946 * find_worker_executing_work - find worker which is executing a work
947 * @pool: pool of interest
948 * @work: work to find worker for
950 * Find a worker which is executing @work on @pool by searching
951 * @pool->busy_hash which is keyed by the address of @work. For a worker
952 * to match, its current execution should match the address of @work and
953 * its work function. This is to avoid unwanted dependency between
954 * unrelated work executions through a work item being recycled while still
957 * This is a bit tricky. A work item may be freed once its execution
958 * starts and nothing prevents the freed area from being recycled for
959 * another work item. If the same work item address ends up being reused
960 * before the original execution finishes, workqueue will identify the
961 * recycled work item as currently executing and make it wait until the
962 * current execution finishes, introducing an unwanted dependency.
964 * This function checks the work item address and work function to avoid
965 * false positives. Note that this isn't complete as one may construct a
966 * work function which can introduce dependency onto itself through a
967 * recycled work item. Well, if somebody wants to shoot oneself in the
968 * foot that badly, there's only so much we can do, and if such deadlock
969 * actually occurs, it should be easy to locate the culprit work function.
972 * spin_lock_irq(pool->lock).
975 * Pointer to worker which is executing @work if found, %NULL
978 static struct worker *find_worker_executing_work(struct worker_pool *pool,
979 struct work_struct *work)
981 struct worker *worker;
983 hash_for_each_possible(pool->busy_hash, worker, hentry,
985 if (worker->current_work == work &&
986 worker->current_func == work->func)
993 * move_linked_works - move linked works to a list
994 * @work: start of series of works to be scheduled
995 * @head: target list to append @work to
996 * @nextp: out parameter for nested worklist walking
998 * Schedule linked works starting from @work to @head. Work series to
999 * be scheduled starts at @work and includes any consecutive work with
1000 * WORK_STRUCT_LINKED set in its predecessor.
1002 * If @nextp is not NULL, it's updated to point to the next work of
1003 * the last scheduled work. This allows move_linked_works() to be
1004 * nested inside outer list_for_each_entry_safe().
1007 * spin_lock_irq(pool->lock).
1009 static void move_linked_works(struct work_struct *work, struct list_head *head,
1010 struct work_struct **nextp)
1012 struct work_struct *n;
1015 * Linked worklist will always end before the end of the list,
1016 * use NULL for list head.
1018 list_for_each_entry_safe_from(work, n, NULL, entry) {
1019 list_move_tail(&work->entry, head);
1020 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1025 * If we're already inside safe list traversal and have moved
1026 * multiple works to the scheduled queue, the next position
1027 * needs to be updated.
1034 * get_pwq - get an extra reference on the specified pool_workqueue
1035 * @pwq: pool_workqueue to get
1037 * Obtain an extra reference on @pwq. The caller should guarantee that
1038 * @pwq has positive refcnt and be holding the matching pool->lock.
1040 static void get_pwq(struct pool_workqueue *pwq)
1042 lockdep_assert_held(&pwq->pool->lock);
1043 WARN_ON_ONCE(pwq->refcnt <= 0);
1048 * put_pwq - put a pool_workqueue reference
1049 * @pwq: pool_workqueue to put
1051 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1052 * destruction. The caller should be holding the matching pool->lock.
1054 static void put_pwq(struct pool_workqueue *pwq)
1056 lockdep_assert_held(&pwq->pool->lock);
1057 if (likely(--pwq->refcnt))
1059 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1062 * @pwq can't be released under pool->lock, bounce to
1063 * pwq_unbound_release_workfn(). This never recurses on the same
1064 * pool->lock as this path is taken only for unbound workqueues and
1065 * the release work item is scheduled on a per-cpu workqueue. To
1066 * avoid lockdep warning, unbound pool->locks are given lockdep
1067 * subclass of 1 in get_unbound_pool().
1069 schedule_work(&pwq->unbound_release_work);
1073 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1074 * @pwq: pool_workqueue to put (can be %NULL)
1076 * put_pwq() with locking. This function also allows %NULL @pwq.
1078 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1082 * As both pwqs and pools are sched-RCU protected, the
1083 * following lock operations are safe.
1085 spin_lock_irq(&pwq->pool->lock);
1087 spin_unlock_irq(&pwq->pool->lock);
1091 static void pwq_activate_delayed_work(struct work_struct *work)
1093 struct pool_workqueue *pwq = get_work_pwq(work);
1095 trace_workqueue_activate_work(work);
1096 move_linked_works(work, &pwq->pool->worklist, NULL);
1097 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1101 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1103 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1104 struct work_struct, entry);
1106 pwq_activate_delayed_work(work);
1110 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1111 * @pwq: pwq of interest
1112 * @color: color of work which left the queue
1114 * A work either has completed or is removed from pending queue,
1115 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1118 * spin_lock_irq(pool->lock).
1120 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1122 /* uncolored work items don't participate in flushing or nr_active */
1123 if (color == WORK_NO_COLOR)
1126 pwq->nr_in_flight[color]--;
1129 if (!list_empty(&pwq->delayed_works)) {
1130 /* one down, submit a delayed one */
1131 if (pwq->nr_active < pwq->max_active)
1132 pwq_activate_first_delayed(pwq);
1135 /* is flush in progress and are we at the flushing tip? */
1136 if (likely(pwq->flush_color != color))
1139 /* are there still in-flight works? */
1140 if (pwq->nr_in_flight[color])
1143 /* this pwq is done, clear flush_color */
1144 pwq->flush_color = -1;
1147 * If this was the last pwq, wake up the first flusher. It
1148 * will handle the rest.
1150 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1151 complete(&pwq->wq->first_flusher->done);
1157 * try_to_grab_pending - steal work item from worklist and disable irq
1158 * @work: work item to steal
1159 * @is_dwork: @work is a delayed_work
1160 * @flags: place to store irq state
1162 * Try to grab PENDING bit of @work. This function can handle @work in any
1163 * stable state - idle, on timer or on worklist.
1166 * 1 if @work was pending and we successfully stole PENDING
1167 * 0 if @work was idle and we claimed PENDING
1168 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1169 * -ENOENT if someone else is canceling @work, this state may persist
1170 * for arbitrarily long
1173 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1174 * interrupted while holding PENDING and @work off queue, irq must be
1175 * disabled on entry. This, combined with delayed_work->timer being
1176 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1178 * On successful return, >= 0, irq is disabled and the caller is
1179 * responsible for releasing it using local_irq_restore(*@flags).
1181 * This function is safe to call from any context including IRQ handler.
1183 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1184 unsigned long *flags)
1186 struct worker_pool *pool;
1187 struct pool_workqueue *pwq;
1189 local_irq_save(*flags);
1191 /* try to steal the timer if it exists */
1193 struct delayed_work *dwork = to_delayed_work(work);
1196 * dwork->timer is irqsafe. If del_timer() fails, it's
1197 * guaranteed that the timer is not queued anywhere and not
1198 * running on the local CPU.
1200 if (likely(del_timer(&dwork->timer)))
1204 /* try to claim PENDING the normal way */
1205 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1209 * The queueing is in progress, or it is already queued. Try to
1210 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1212 pool = get_work_pool(work);
1216 spin_lock(&pool->lock);
1218 * work->data is guaranteed to point to pwq only while the work
1219 * item is queued on pwq->wq, and both updating work->data to point
1220 * to pwq on queueing and to pool on dequeueing are done under
1221 * pwq->pool->lock. This in turn guarantees that, if work->data
1222 * points to pwq which is associated with a locked pool, the work
1223 * item is currently queued on that pool.
1225 pwq = get_work_pwq(work);
1226 if (pwq && pwq->pool == pool) {
1227 debug_work_deactivate(work);
1230 * A delayed work item cannot be grabbed directly because
1231 * it might have linked NO_COLOR work items which, if left
1232 * on the delayed_list, will confuse pwq->nr_active
1233 * management later on and cause stall. Make sure the work
1234 * item is activated before grabbing.
1236 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1237 pwq_activate_delayed_work(work);
1239 list_del_init(&work->entry);
1240 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1242 /* work->data points to pwq iff queued, point to pool */
1243 set_work_pool_and_keep_pending(work, pool->id);
1245 spin_unlock(&pool->lock);
1248 spin_unlock(&pool->lock);
1250 local_irq_restore(*flags);
1251 if (work_is_canceling(work))
1258 * insert_work - insert a work into a pool
1259 * @pwq: pwq @work belongs to
1260 * @work: work to insert
1261 * @head: insertion point
1262 * @extra_flags: extra WORK_STRUCT_* flags to set
1264 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1265 * work_struct flags.
1268 * spin_lock_irq(pool->lock).
1270 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1271 struct list_head *head, unsigned int extra_flags)
1273 struct worker_pool *pool = pwq->pool;
1275 /* we own @work, set data and link */
1276 set_work_pwq(work, pwq, extra_flags);
1277 list_add_tail(&work->entry, head);
1281 * Ensure either wq_worker_sleeping() sees the above
1282 * list_add_tail() or we see zero nr_running to avoid workers lying
1283 * around lazily while there are works to be processed.
1287 if (__need_more_worker(pool))
1288 wake_up_worker(pool);
1292 * Test whether @work is being queued from another work executing on the
1295 static bool is_chained_work(struct workqueue_struct *wq)
1297 struct worker *worker;
1299 worker = current_wq_worker();
1301 * Return %true iff I'm a worker execuing a work item on @wq. If
1302 * I'm @worker, it's safe to dereference it without locking.
1304 return worker && worker->current_pwq->wq == wq;
1307 static void __queue_work(int cpu, struct workqueue_struct *wq,
1308 struct work_struct *work)
1310 struct pool_workqueue *pwq;
1311 struct worker_pool *last_pool;
1312 struct list_head *worklist;
1313 unsigned int work_flags;
1314 unsigned int req_cpu = cpu;
1317 * While a work item is PENDING && off queue, a task trying to
1318 * steal the PENDING will busy-loop waiting for it to either get
1319 * queued or lose PENDING. Grabbing PENDING and queueing should
1320 * happen with IRQ disabled.
1322 WARN_ON_ONCE(!irqs_disabled());
1324 debug_work_activate(work);
1326 /* if draining, only works from the same workqueue are allowed */
1327 if (unlikely(wq->flags & __WQ_DRAINING) &&
1328 WARN_ON_ONCE(!is_chained_work(wq)))
1331 if (req_cpu == WORK_CPU_UNBOUND)
1332 cpu = raw_smp_processor_id();
1334 /* pwq which will be used unless @work is executing elsewhere */
1335 if (!(wq->flags & WQ_UNBOUND))
1336 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1338 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1341 * If @work was previously on a different pool, it might still be
1342 * running there, in which case the work needs to be queued on that
1343 * pool to guarantee non-reentrancy.
1345 last_pool = get_work_pool(work);
1346 if (last_pool && last_pool != pwq->pool) {
1347 struct worker *worker;
1349 spin_lock(&last_pool->lock);
1351 worker = find_worker_executing_work(last_pool, work);
1353 if (worker && worker->current_pwq->wq == wq) {
1354 pwq = worker->current_pwq;
1356 /* meh... not running there, queue here */
1357 spin_unlock(&last_pool->lock);
1358 spin_lock(&pwq->pool->lock);
1361 spin_lock(&pwq->pool->lock);
1365 * pwq is determined and locked. For unbound pools, we could have
1366 * raced with pwq release and it could already be dead. If its
1367 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1368 * without another pwq replacing it in the numa_pwq_tbl or while
1369 * work items are executing on it, so the retrying is guaranteed to
1370 * make forward-progress.
1372 if (unlikely(!pwq->refcnt)) {
1373 if (wq->flags & WQ_UNBOUND) {
1374 spin_unlock(&pwq->pool->lock);
1379 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1383 /* pwq determined, queue */
1384 trace_workqueue_queue_work(req_cpu, pwq, work);
1386 if (WARN_ON(!list_empty(&work->entry))) {
1387 spin_unlock(&pwq->pool->lock);
1391 pwq->nr_in_flight[pwq->work_color]++;
1392 work_flags = work_color_to_flags(pwq->work_color);
1394 if (likely(pwq->nr_active < pwq->max_active)) {
1395 trace_workqueue_activate_work(work);
1397 worklist = &pwq->pool->worklist;
1399 work_flags |= WORK_STRUCT_DELAYED;
1400 worklist = &pwq->delayed_works;
1403 insert_work(pwq, work, worklist, work_flags);
1405 spin_unlock(&pwq->pool->lock);
1409 * queue_work_on - queue work on specific cpu
1410 * @cpu: CPU number to execute work on
1411 * @wq: workqueue to use
1412 * @work: work to queue
1414 * We queue the work to a specific CPU, the caller must ensure it
1417 * Return: %false if @work was already on a queue, %true otherwise.
1419 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1420 struct work_struct *work)
1423 unsigned long flags;
1425 local_irq_save(flags);
1427 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1428 __queue_work(cpu, wq, work);
1432 local_irq_restore(flags);
1435 EXPORT_SYMBOL(queue_work_on);
1437 void delayed_work_timer_fn(unsigned long __data)
1439 struct delayed_work *dwork = (struct delayed_work *)__data;
1441 /* should have been called from irqsafe timer with irq already off */
1442 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1444 EXPORT_SYMBOL(delayed_work_timer_fn);
1446 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1447 struct delayed_work *dwork, unsigned long delay)
1449 struct timer_list *timer = &dwork->timer;
1450 struct work_struct *work = &dwork->work;
1452 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1453 timer->data != (unsigned long)dwork);
1454 WARN_ON_ONCE(timer_pending(timer));
1455 WARN_ON_ONCE(!list_empty(&work->entry));
1458 * If @delay is 0, queue @dwork->work immediately. This is for
1459 * both optimization and correctness. The earliest @timer can
1460 * expire is on the closest next tick and delayed_work users depend
1461 * on that there's no such delay when @delay is 0.
1464 __queue_work(cpu, wq, &dwork->work);
1468 timer_stats_timer_set_start_info(&dwork->timer);
1472 timer->expires = jiffies + delay;
1474 if (unlikely(cpu != WORK_CPU_UNBOUND))
1475 add_timer_on(timer, cpu);
1481 * queue_delayed_work_on - queue work on specific CPU after delay
1482 * @cpu: CPU number to execute work on
1483 * @wq: workqueue to use
1484 * @dwork: work to queue
1485 * @delay: number of jiffies to wait before queueing
1487 * Return: %false if @work was already on a queue, %true otherwise. If
1488 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1491 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1492 struct delayed_work *dwork, unsigned long delay)
1494 struct work_struct *work = &dwork->work;
1496 unsigned long flags;
1498 /* read the comment in __queue_work() */
1499 local_irq_save(flags);
1501 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1502 __queue_delayed_work(cpu, wq, dwork, delay);
1506 local_irq_restore(flags);
1509 EXPORT_SYMBOL(queue_delayed_work_on);
1512 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1513 * @cpu: CPU number to execute work on
1514 * @wq: workqueue to use
1515 * @dwork: work to queue
1516 * @delay: number of jiffies to wait before queueing
1518 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1519 * modify @dwork's timer so that it expires after @delay. If @delay is
1520 * zero, @work is guaranteed to be scheduled immediately regardless of its
1523 * Return: %false if @dwork was idle and queued, %true if @dwork was
1524 * pending and its timer was modified.
1526 * This function is safe to call from any context including IRQ handler.
1527 * See try_to_grab_pending() for details.
1529 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1530 struct delayed_work *dwork, unsigned long delay)
1532 unsigned long flags;
1536 ret = try_to_grab_pending(&dwork->work, true, &flags);
1537 } while (unlikely(ret == -EAGAIN));
1539 if (likely(ret >= 0)) {
1540 __queue_delayed_work(cpu, wq, dwork, delay);
1541 local_irq_restore(flags);
1544 /* -ENOENT from try_to_grab_pending() becomes %true */
1547 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1550 * worker_enter_idle - enter idle state
1551 * @worker: worker which is entering idle state
1553 * @worker is entering idle state. Update stats and idle timer if
1557 * spin_lock_irq(pool->lock).
1559 static void worker_enter_idle(struct worker *worker)
1561 struct worker_pool *pool = worker->pool;
1563 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1564 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1565 (worker->hentry.next || worker->hentry.pprev)))
1568 /* can't use worker_set_flags(), also called from create_worker() */
1569 worker->flags |= WORKER_IDLE;
1571 worker->last_active = jiffies;
1573 /* idle_list is LIFO */
1574 list_add(&worker->entry, &pool->idle_list);
1576 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1577 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1580 * Sanity check nr_running. Because wq_unbind_fn() releases
1581 * pool->lock between setting %WORKER_UNBOUND and zapping
1582 * nr_running, the warning may trigger spuriously. Check iff
1583 * unbind is not in progress.
1585 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1586 pool->nr_workers == pool->nr_idle &&
1587 atomic_read(&pool->nr_running));
1591 * worker_leave_idle - leave idle state
1592 * @worker: worker which is leaving idle state
1594 * @worker is leaving idle state. Update stats.
1597 * spin_lock_irq(pool->lock).
1599 static void worker_leave_idle(struct worker *worker)
1601 struct worker_pool *pool = worker->pool;
1603 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1605 worker_clr_flags(worker, WORKER_IDLE);
1607 list_del_init(&worker->entry);
1610 static struct worker *alloc_worker(int node)
1612 struct worker *worker;
1614 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1616 INIT_LIST_HEAD(&worker->entry);
1617 INIT_LIST_HEAD(&worker->scheduled);
1618 INIT_LIST_HEAD(&worker->node);
1619 /* on creation a worker is in !idle && prep state */
1620 worker->flags = WORKER_PREP;
1626 * worker_attach_to_pool() - attach a worker to a pool
1627 * @worker: worker to be attached
1628 * @pool: the target pool
1630 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1631 * cpu-binding of @worker are kept coordinated with the pool across
1634 static void worker_attach_to_pool(struct worker *worker,
1635 struct worker_pool *pool)
1637 mutex_lock(&pool->attach_mutex);
1640 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1641 * online CPUs. It'll be re-applied when any of the CPUs come up.
1643 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1646 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1647 * stable across this function. See the comments above the
1648 * flag definition for details.
1650 if (pool->flags & POOL_DISASSOCIATED)
1651 worker->flags |= WORKER_UNBOUND;
1653 list_add_tail(&worker->node, &pool->workers);
1655 mutex_unlock(&pool->attach_mutex);
1659 * worker_detach_from_pool() - detach a worker from its pool
1660 * @worker: worker which is attached to its pool
1661 * @pool: the pool @worker is attached to
1663 * Undo the attaching which had been done in worker_attach_to_pool(). The
1664 * caller worker shouldn't access to the pool after detached except it has
1665 * other reference to the pool.
1667 static void worker_detach_from_pool(struct worker *worker,
1668 struct worker_pool *pool)
1670 struct completion *detach_completion = NULL;
1672 mutex_lock(&pool->attach_mutex);
1673 list_del(&worker->node);
1674 if (list_empty(&pool->workers))
1675 detach_completion = pool->detach_completion;
1676 mutex_unlock(&pool->attach_mutex);
1678 /* clear leftover flags without pool->lock after it is detached */
1679 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1681 if (detach_completion)
1682 complete(detach_completion);
1686 * create_worker - create a new workqueue worker
1687 * @pool: pool the new worker will belong to
1689 * Create and start a new worker which is attached to @pool.
1692 * Might sleep. Does GFP_KERNEL allocations.
1695 * Pointer to the newly created worker.
1697 static struct worker *create_worker(struct worker_pool *pool)
1699 struct worker *worker = NULL;
1703 /* ID is needed to determine kthread name */
1704 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1708 worker = alloc_worker(pool->node);
1712 worker->pool = pool;
1716 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1717 pool->attrs->nice < 0 ? "H" : "");
1719 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1721 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1722 "kworker/%s", id_buf);
1723 if (IS_ERR(worker->task))
1726 set_user_nice(worker->task, pool->attrs->nice);
1727 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1729 /* successful, attach the worker to the pool */
1730 worker_attach_to_pool(worker, pool);
1732 /* start the newly created worker */
1733 spin_lock_irq(&pool->lock);
1734 worker->pool->nr_workers++;
1735 worker_enter_idle(worker);
1736 wake_up_process(worker->task);
1737 spin_unlock_irq(&pool->lock);
1743 ida_simple_remove(&pool->worker_ida, id);
1749 * destroy_worker - destroy a workqueue worker
1750 * @worker: worker to be destroyed
1752 * Destroy @worker and adjust @pool stats accordingly. The worker should
1756 * spin_lock_irq(pool->lock).
1758 static void destroy_worker(struct worker *worker)
1760 struct worker_pool *pool = worker->pool;
1762 lockdep_assert_held(&pool->lock);
1764 /* sanity check frenzy */
1765 if (WARN_ON(worker->current_work) ||
1766 WARN_ON(!list_empty(&worker->scheduled)) ||
1767 WARN_ON(!(worker->flags & WORKER_IDLE)))
1773 list_del_init(&worker->entry);
1774 worker->flags |= WORKER_DIE;
1775 wake_up_process(worker->task);
1778 static void idle_worker_timeout(unsigned long __pool)
1780 struct worker_pool *pool = (void *)__pool;
1782 spin_lock_irq(&pool->lock);
1784 while (too_many_workers(pool)) {
1785 struct worker *worker;
1786 unsigned long expires;
1788 /* idle_list is kept in LIFO order, check the last one */
1789 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1790 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1792 if (time_before(jiffies, expires)) {
1793 mod_timer(&pool->idle_timer, expires);
1797 destroy_worker(worker);
1800 spin_unlock_irq(&pool->lock);
1803 static void send_mayday(struct work_struct *work)
1805 struct pool_workqueue *pwq = get_work_pwq(work);
1806 struct workqueue_struct *wq = pwq->wq;
1808 lockdep_assert_held(&wq_mayday_lock);
1813 /* mayday mayday mayday */
1814 if (list_empty(&pwq->mayday_node)) {
1816 * If @pwq is for an unbound wq, its base ref may be put at
1817 * any time due to an attribute change. Pin @pwq until the
1818 * rescuer is done with it.
1821 list_add_tail(&pwq->mayday_node, &wq->maydays);
1822 wake_up_process(wq->rescuer->task);
1826 static void pool_mayday_timeout(unsigned long __pool)
1828 struct worker_pool *pool = (void *)__pool;
1829 struct work_struct *work;
1831 spin_lock_irq(&pool->lock);
1832 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1834 if (need_to_create_worker(pool)) {
1836 * We've been trying to create a new worker but
1837 * haven't been successful. We might be hitting an
1838 * allocation deadlock. Send distress signals to
1841 list_for_each_entry(work, &pool->worklist, entry)
1845 spin_unlock(&wq_mayday_lock);
1846 spin_unlock_irq(&pool->lock);
1848 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1852 * maybe_create_worker - create a new worker if necessary
1853 * @pool: pool to create a new worker for
1855 * Create a new worker for @pool if necessary. @pool is guaranteed to
1856 * have at least one idle worker on return from this function. If
1857 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1858 * sent to all rescuers with works scheduled on @pool to resolve
1859 * possible allocation deadlock.
1861 * On return, need_to_create_worker() is guaranteed to be %false and
1862 * may_start_working() %true.
1865 * spin_lock_irq(pool->lock) which may be released and regrabbed
1866 * multiple times. Does GFP_KERNEL allocations. Called only from
1869 static void maybe_create_worker(struct worker_pool *pool)
1870 __releases(&pool->lock)
1871 __acquires(&pool->lock)
1874 spin_unlock_irq(&pool->lock);
1876 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1877 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1880 if (create_worker(pool) || !need_to_create_worker(pool))
1883 schedule_timeout_interruptible(CREATE_COOLDOWN);
1885 if (!need_to_create_worker(pool))
1889 del_timer_sync(&pool->mayday_timer);
1890 spin_lock_irq(&pool->lock);
1892 * This is necessary even after a new worker was just successfully
1893 * created as @pool->lock was dropped and the new worker might have
1894 * already become busy.
1896 if (need_to_create_worker(pool))
1901 * manage_workers - manage worker pool
1904 * Assume the manager role and manage the worker pool @worker belongs
1905 * to. At any given time, there can be only zero or one manager per
1906 * pool. The exclusion is handled automatically by this function.
1908 * The caller can safely start processing works on false return. On
1909 * true return, it's guaranteed that need_to_create_worker() is false
1910 * and may_start_working() is true.
1913 * spin_lock_irq(pool->lock) which may be released and regrabbed
1914 * multiple times. Does GFP_KERNEL allocations.
1917 * %false if the pool doesn't need management and the caller can safely
1918 * start processing works, %true if management function was performed and
1919 * the conditions that the caller verified before calling the function may
1920 * no longer be true.
1922 static bool manage_workers(struct worker *worker)
1924 struct worker_pool *pool = worker->pool;
1927 * Anyone who successfully grabs manager_arb wins the arbitration
1928 * and becomes the manager. mutex_trylock() on pool->manager_arb
1929 * failure while holding pool->lock reliably indicates that someone
1930 * else is managing the pool and the worker which failed trylock
1931 * can proceed to executing work items. This means that anyone
1932 * grabbing manager_arb is responsible for actually performing
1933 * manager duties. If manager_arb is grabbed and released without
1934 * actual management, the pool may stall indefinitely.
1936 if (!mutex_trylock(&pool->manager_arb))
1938 pool->manager = worker;
1940 maybe_create_worker(pool);
1942 pool->manager = NULL;
1943 mutex_unlock(&pool->manager_arb);
1948 * process_one_work - process single work
1950 * @work: work to process
1952 * Process @work. This function contains all the logics necessary to
1953 * process a single work including synchronization against and
1954 * interaction with other workers on the same cpu, queueing and
1955 * flushing. As long as context requirement is met, any worker can
1956 * call this function to process a work.
1959 * spin_lock_irq(pool->lock) which is released and regrabbed.
1961 static void process_one_work(struct worker *worker, struct work_struct *work)
1962 __releases(&pool->lock)
1963 __acquires(&pool->lock)
1965 struct pool_workqueue *pwq = get_work_pwq(work);
1966 struct worker_pool *pool = worker->pool;
1967 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1969 struct worker *collision;
1970 #ifdef CONFIG_LOCKDEP
1972 * It is permissible to free the struct work_struct from
1973 * inside the function that is called from it, this we need to
1974 * take into account for lockdep too. To avoid bogus "held
1975 * lock freed" warnings as well as problems when looking into
1976 * work->lockdep_map, make a copy and use that here.
1978 struct lockdep_map lockdep_map;
1980 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1982 /* ensure we're on the correct CPU */
1983 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1984 raw_smp_processor_id() != pool->cpu);
1987 * A single work shouldn't be executed concurrently by
1988 * multiple workers on a single cpu. Check whether anyone is
1989 * already processing the work. If so, defer the work to the
1990 * currently executing one.
1992 collision = find_worker_executing_work(pool, work);
1993 if (unlikely(collision)) {
1994 move_linked_works(work, &collision->scheduled, NULL);
1998 /* claim and dequeue */
1999 debug_work_deactivate(work);
2000 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2001 worker->current_work = work;
2002 worker->current_func = work->func;
2003 worker->current_pwq = pwq;
2004 work_color = get_work_color(work);
2006 list_del_init(&work->entry);
2009 * CPU intensive works don't participate in concurrency management.
2010 * They're the scheduler's responsibility. This takes @worker out
2011 * of concurrency management and the next code block will chain
2012 * execution of the pending work items.
2014 if (unlikely(cpu_intensive))
2015 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2018 * Wake up another worker if necessary. The condition is always
2019 * false for normal per-cpu workers since nr_running would always
2020 * be >= 1 at this point. This is used to chain execution of the
2021 * pending work items for WORKER_NOT_RUNNING workers such as the
2022 * UNBOUND and CPU_INTENSIVE ones.
2024 if (need_more_worker(pool))
2025 wake_up_worker(pool);
2028 * Record the last pool and clear PENDING which should be the last
2029 * update to @work. Also, do this inside @pool->lock so that
2030 * PENDING and queued state changes happen together while IRQ is
2033 set_work_pool_and_clear_pending(work, pool->id);
2035 spin_unlock_irq(&pool->lock);
2037 lock_map_acquire_read(&pwq->wq->lockdep_map);
2038 lock_map_acquire(&lockdep_map);
2039 trace_workqueue_execute_start(work);
2040 worker->current_func(work);
2042 * While we must be careful to not use "work" after this, the trace
2043 * point will only record its address.
2045 trace_workqueue_execute_end(work);
2046 lock_map_release(&lockdep_map);
2047 lock_map_release(&pwq->wq->lockdep_map);
2049 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2050 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2051 " last function: %pf\n",
2052 current->comm, preempt_count(), task_pid_nr(current),
2053 worker->current_func);
2054 debug_show_held_locks(current);
2059 * The following prevents a kworker from hogging CPU on !PREEMPT
2060 * kernels, where a requeueing work item waiting for something to
2061 * happen could deadlock with stop_machine as such work item could
2062 * indefinitely requeue itself while all other CPUs are trapped in
2063 * stop_machine. At the same time, report a quiescent RCU state so
2064 * the same condition doesn't freeze RCU.
2066 cond_resched_rcu_qs();
2068 spin_lock_irq(&pool->lock);
2070 /* clear cpu intensive status */
2071 if (unlikely(cpu_intensive))
2072 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2074 /* we're done with it, release */
2075 hash_del(&worker->hentry);
2076 worker->current_work = NULL;
2077 worker->current_func = NULL;
2078 worker->current_pwq = NULL;
2079 worker->desc_valid = false;
2080 pwq_dec_nr_in_flight(pwq, work_color);
2084 * process_scheduled_works - process scheduled works
2087 * Process all scheduled works. Please note that the scheduled list
2088 * may change while processing a work, so this function repeatedly
2089 * fetches a work from the top and executes it.
2092 * spin_lock_irq(pool->lock) which may be released and regrabbed
2095 static void process_scheduled_works(struct worker *worker)
2097 while (!list_empty(&worker->scheduled)) {
2098 struct work_struct *work = list_first_entry(&worker->scheduled,
2099 struct work_struct, entry);
2100 process_one_work(worker, work);
2105 * worker_thread - the worker thread function
2108 * The worker thread function. All workers belong to a worker_pool -
2109 * either a per-cpu one or dynamic unbound one. These workers process all
2110 * work items regardless of their specific target workqueue. The only
2111 * exception is work items which belong to workqueues with a rescuer which
2112 * will be explained in rescuer_thread().
2116 static int worker_thread(void *__worker)
2118 struct worker *worker = __worker;
2119 struct worker_pool *pool = worker->pool;
2121 /* tell the scheduler that this is a workqueue worker */
2122 worker->task->flags |= PF_WQ_WORKER;
2124 spin_lock_irq(&pool->lock);
2126 /* am I supposed to die? */
2127 if (unlikely(worker->flags & WORKER_DIE)) {
2128 spin_unlock_irq(&pool->lock);
2129 WARN_ON_ONCE(!list_empty(&worker->entry));
2130 worker->task->flags &= ~PF_WQ_WORKER;
2132 set_task_comm(worker->task, "kworker/dying");
2133 ida_simple_remove(&pool->worker_ida, worker->id);
2134 worker_detach_from_pool(worker, pool);
2139 worker_leave_idle(worker);
2141 /* no more worker necessary? */
2142 if (!need_more_worker(pool))
2145 /* do we need to manage? */
2146 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2150 * ->scheduled list can only be filled while a worker is
2151 * preparing to process a work or actually processing it.
2152 * Make sure nobody diddled with it while I was sleeping.
2154 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2157 * Finish PREP stage. We're guaranteed to have at least one idle
2158 * worker or that someone else has already assumed the manager
2159 * role. This is where @worker starts participating in concurrency
2160 * management if applicable and concurrency management is restored
2161 * after being rebound. See rebind_workers() for details.
2163 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2166 struct work_struct *work =
2167 list_first_entry(&pool->worklist,
2168 struct work_struct, entry);
2170 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2171 /* optimization path, not strictly necessary */
2172 process_one_work(worker, work);
2173 if (unlikely(!list_empty(&worker->scheduled)))
2174 process_scheduled_works(worker);
2176 move_linked_works(work, &worker->scheduled, NULL);
2177 process_scheduled_works(worker);
2179 } while (keep_working(pool));
2181 worker_set_flags(worker, WORKER_PREP);
2184 * pool->lock is held and there's no work to process and no need to
2185 * manage, sleep. Workers are woken up only while holding
2186 * pool->lock or from local cpu, so setting the current state
2187 * before releasing pool->lock is enough to prevent losing any
2190 worker_enter_idle(worker);
2191 __set_current_state(TASK_INTERRUPTIBLE);
2192 spin_unlock_irq(&pool->lock);
2198 * rescuer_thread - the rescuer thread function
2201 * Workqueue rescuer thread function. There's one rescuer for each
2202 * workqueue which has WQ_MEM_RECLAIM set.
2204 * Regular work processing on a pool may block trying to create a new
2205 * worker which uses GFP_KERNEL allocation which has slight chance of
2206 * developing into deadlock if some works currently on the same queue
2207 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2208 * the problem rescuer solves.
2210 * When such condition is possible, the pool summons rescuers of all
2211 * workqueues which have works queued on the pool and let them process
2212 * those works so that forward progress can be guaranteed.
2214 * This should happen rarely.
2218 static int rescuer_thread(void *__rescuer)
2220 struct worker *rescuer = __rescuer;
2221 struct workqueue_struct *wq = rescuer->rescue_wq;
2222 struct list_head *scheduled = &rescuer->scheduled;
2225 set_user_nice(current, RESCUER_NICE_LEVEL);
2228 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2229 * doesn't participate in concurrency management.
2231 rescuer->task->flags |= PF_WQ_WORKER;
2233 set_current_state(TASK_INTERRUPTIBLE);
2236 * By the time the rescuer is requested to stop, the workqueue
2237 * shouldn't have any work pending, but @wq->maydays may still have
2238 * pwq(s) queued. This can happen by non-rescuer workers consuming
2239 * all the work items before the rescuer got to them. Go through
2240 * @wq->maydays processing before acting on should_stop so that the
2241 * list is always empty on exit.
2243 should_stop = kthread_should_stop();
2245 /* see whether any pwq is asking for help */
2246 spin_lock_irq(&wq_mayday_lock);
2248 while (!list_empty(&wq->maydays)) {
2249 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2250 struct pool_workqueue, mayday_node);
2251 struct worker_pool *pool = pwq->pool;
2252 struct work_struct *work, *n;
2254 __set_current_state(TASK_RUNNING);
2255 list_del_init(&pwq->mayday_node);
2257 spin_unlock_irq(&wq_mayday_lock);
2259 worker_attach_to_pool(rescuer, pool);
2261 spin_lock_irq(&pool->lock);
2262 rescuer->pool = pool;
2265 * Slurp in all works issued via this workqueue and
2268 WARN_ON_ONCE(!list_empty(scheduled));
2269 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2270 if (get_work_pwq(work) == pwq)
2271 move_linked_works(work, scheduled, &n);
2273 if (!list_empty(scheduled)) {
2274 process_scheduled_works(rescuer);
2277 * The above execution of rescued work items could
2278 * have created more to rescue through
2279 * pwq_activate_first_delayed() or chained
2280 * queueing. Let's put @pwq back on mayday list so
2281 * that such back-to-back work items, which may be
2282 * being used to relieve memory pressure, don't
2283 * incur MAYDAY_INTERVAL delay inbetween.
2285 if (need_to_create_worker(pool)) {
2286 spin_lock(&wq_mayday_lock);
2288 list_move_tail(&pwq->mayday_node, &wq->maydays);
2289 spin_unlock(&wq_mayday_lock);
2294 * Put the reference grabbed by send_mayday(). @pool won't
2295 * go away while we're still attached to it.
2300 * Leave this pool. If need_more_worker() is %true, notify a
2301 * regular worker; otherwise, we end up with 0 concurrency
2302 * and stalling the execution.
2304 if (need_more_worker(pool))
2305 wake_up_worker(pool);
2307 rescuer->pool = NULL;
2308 spin_unlock_irq(&pool->lock);
2310 worker_detach_from_pool(rescuer, pool);
2312 spin_lock_irq(&wq_mayday_lock);
2315 spin_unlock_irq(&wq_mayday_lock);
2318 __set_current_state(TASK_RUNNING);
2319 rescuer->task->flags &= ~PF_WQ_WORKER;
2323 /* rescuers should never participate in concurrency management */
2324 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2330 struct work_struct work;
2331 struct completion done;
2332 struct task_struct *task; /* purely informational */
2335 static void wq_barrier_func(struct work_struct *work)
2337 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2338 complete(&barr->done);
2342 * insert_wq_barrier - insert a barrier work
2343 * @pwq: pwq to insert barrier into
2344 * @barr: wq_barrier to insert
2345 * @target: target work to attach @barr to
2346 * @worker: worker currently executing @target, NULL if @target is not executing
2348 * @barr is linked to @target such that @barr is completed only after
2349 * @target finishes execution. Please note that the ordering
2350 * guarantee is observed only with respect to @target and on the local
2353 * Currently, a queued barrier can't be canceled. This is because
2354 * try_to_grab_pending() can't determine whether the work to be
2355 * grabbed is at the head of the queue and thus can't clear LINKED
2356 * flag of the previous work while there must be a valid next work
2357 * after a work with LINKED flag set.
2359 * Note that when @worker is non-NULL, @target may be modified
2360 * underneath us, so we can't reliably determine pwq from @target.
2363 * spin_lock_irq(pool->lock).
2365 static void insert_wq_barrier(struct pool_workqueue *pwq,
2366 struct wq_barrier *barr,
2367 struct work_struct *target, struct worker *worker)
2369 struct list_head *head;
2370 unsigned int linked = 0;
2373 * debugobject calls are safe here even with pool->lock locked
2374 * as we know for sure that this will not trigger any of the
2375 * checks and call back into the fixup functions where we
2378 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2379 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2380 init_completion(&barr->done);
2381 barr->task = current;
2384 * If @target is currently being executed, schedule the
2385 * barrier to the worker; otherwise, put it after @target.
2388 head = worker->scheduled.next;
2390 unsigned long *bits = work_data_bits(target);
2392 head = target->entry.next;
2393 /* there can already be other linked works, inherit and set */
2394 linked = *bits & WORK_STRUCT_LINKED;
2395 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2398 debug_work_activate(&barr->work);
2399 insert_work(pwq, &barr->work, head,
2400 work_color_to_flags(WORK_NO_COLOR) | linked);
2404 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2405 * @wq: workqueue being flushed
2406 * @flush_color: new flush color, < 0 for no-op
2407 * @work_color: new work color, < 0 for no-op
2409 * Prepare pwqs for workqueue flushing.
2411 * If @flush_color is non-negative, flush_color on all pwqs should be
2412 * -1. If no pwq has in-flight commands at the specified color, all
2413 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2414 * has in flight commands, its pwq->flush_color is set to
2415 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2416 * wakeup logic is armed and %true is returned.
2418 * The caller should have initialized @wq->first_flusher prior to
2419 * calling this function with non-negative @flush_color. If
2420 * @flush_color is negative, no flush color update is done and %false
2423 * If @work_color is non-negative, all pwqs should have the same
2424 * work_color which is previous to @work_color and all will be
2425 * advanced to @work_color.
2428 * mutex_lock(wq->mutex).
2431 * %true if @flush_color >= 0 and there's something to flush. %false
2434 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2435 int flush_color, int work_color)
2438 struct pool_workqueue *pwq;
2440 if (flush_color >= 0) {
2441 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2442 atomic_set(&wq->nr_pwqs_to_flush, 1);
2445 for_each_pwq(pwq, wq) {
2446 struct worker_pool *pool = pwq->pool;
2448 spin_lock_irq(&pool->lock);
2450 if (flush_color >= 0) {
2451 WARN_ON_ONCE(pwq->flush_color != -1);
2453 if (pwq->nr_in_flight[flush_color]) {
2454 pwq->flush_color = flush_color;
2455 atomic_inc(&wq->nr_pwqs_to_flush);
2460 if (work_color >= 0) {
2461 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2462 pwq->work_color = work_color;
2465 spin_unlock_irq(&pool->lock);
2468 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2469 complete(&wq->first_flusher->done);
2475 * flush_workqueue - ensure that any scheduled work has run to completion.
2476 * @wq: workqueue to flush
2478 * This function sleeps until all work items which were queued on entry
2479 * have finished execution, but it is not livelocked by new incoming ones.
2481 void flush_workqueue(struct workqueue_struct *wq)
2483 struct wq_flusher this_flusher = {
2484 .list = LIST_HEAD_INIT(this_flusher.list),
2486 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2490 lock_map_acquire(&wq->lockdep_map);
2491 lock_map_release(&wq->lockdep_map);
2493 mutex_lock(&wq->mutex);
2496 * Start-to-wait phase
2498 next_color = work_next_color(wq->work_color);
2500 if (next_color != wq->flush_color) {
2502 * Color space is not full. The current work_color
2503 * becomes our flush_color and work_color is advanced
2506 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2507 this_flusher.flush_color = wq->work_color;
2508 wq->work_color = next_color;
2510 if (!wq->first_flusher) {
2511 /* no flush in progress, become the first flusher */
2512 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2514 wq->first_flusher = &this_flusher;
2516 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2518 /* nothing to flush, done */
2519 wq->flush_color = next_color;
2520 wq->first_flusher = NULL;
2525 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2526 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2527 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2531 * Oops, color space is full, wait on overflow queue.
2532 * The next flush completion will assign us
2533 * flush_color and transfer to flusher_queue.
2535 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2538 mutex_unlock(&wq->mutex);
2540 wait_for_completion(&this_flusher.done);
2543 * Wake-up-and-cascade phase
2545 * First flushers are responsible for cascading flushes and
2546 * handling overflow. Non-first flushers can simply return.
2548 if (wq->first_flusher != &this_flusher)
2551 mutex_lock(&wq->mutex);
2553 /* we might have raced, check again with mutex held */
2554 if (wq->first_flusher != &this_flusher)
2557 wq->first_flusher = NULL;
2559 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2560 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2563 struct wq_flusher *next, *tmp;
2565 /* complete all the flushers sharing the current flush color */
2566 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2567 if (next->flush_color != wq->flush_color)
2569 list_del_init(&next->list);
2570 complete(&next->done);
2573 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2574 wq->flush_color != work_next_color(wq->work_color));
2576 /* this flush_color is finished, advance by one */
2577 wq->flush_color = work_next_color(wq->flush_color);
2579 /* one color has been freed, handle overflow queue */
2580 if (!list_empty(&wq->flusher_overflow)) {
2582 * Assign the same color to all overflowed
2583 * flushers, advance work_color and append to
2584 * flusher_queue. This is the start-to-wait
2585 * phase for these overflowed flushers.
2587 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2588 tmp->flush_color = wq->work_color;
2590 wq->work_color = work_next_color(wq->work_color);
2592 list_splice_tail_init(&wq->flusher_overflow,
2593 &wq->flusher_queue);
2594 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2597 if (list_empty(&wq->flusher_queue)) {
2598 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2603 * Need to flush more colors. Make the next flusher
2604 * the new first flusher and arm pwqs.
2606 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2607 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2609 list_del_init(&next->list);
2610 wq->first_flusher = next;
2612 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2616 * Meh... this color is already done, clear first
2617 * flusher and repeat cascading.
2619 wq->first_flusher = NULL;
2623 mutex_unlock(&wq->mutex);
2625 EXPORT_SYMBOL(flush_workqueue);
2628 * drain_workqueue - drain a workqueue
2629 * @wq: workqueue to drain
2631 * Wait until the workqueue becomes empty. While draining is in progress,
2632 * only chain queueing is allowed. IOW, only currently pending or running
2633 * work items on @wq can queue further work items on it. @wq is flushed
2634 * repeatedly until it becomes empty. The number of flushing is determined
2635 * by the depth of chaining and should be relatively short. Whine if it
2638 void drain_workqueue(struct workqueue_struct *wq)
2640 unsigned int flush_cnt = 0;
2641 struct pool_workqueue *pwq;
2644 * __queue_work() needs to test whether there are drainers, is much
2645 * hotter than drain_workqueue() and already looks at @wq->flags.
2646 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2648 mutex_lock(&wq->mutex);
2649 if (!wq->nr_drainers++)
2650 wq->flags |= __WQ_DRAINING;
2651 mutex_unlock(&wq->mutex);
2653 flush_workqueue(wq);
2655 mutex_lock(&wq->mutex);
2657 for_each_pwq(pwq, wq) {
2660 spin_lock_irq(&pwq->pool->lock);
2661 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2662 spin_unlock_irq(&pwq->pool->lock);
2667 if (++flush_cnt == 10 ||
2668 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2669 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2670 wq->name, flush_cnt);
2672 mutex_unlock(&wq->mutex);
2676 if (!--wq->nr_drainers)
2677 wq->flags &= ~__WQ_DRAINING;
2678 mutex_unlock(&wq->mutex);
2680 EXPORT_SYMBOL_GPL(drain_workqueue);
2682 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2684 struct worker *worker = NULL;
2685 struct worker_pool *pool;
2686 struct pool_workqueue *pwq;
2690 local_irq_disable();
2691 pool = get_work_pool(work);
2697 spin_lock(&pool->lock);
2698 /* see the comment in try_to_grab_pending() with the same code */
2699 pwq = get_work_pwq(work);
2701 if (unlikely(pwq->pool != pool))
2704 worker = find_worker_executing_work(pool, work);
2707 pwq = worker->current_pwq;
2710 insert_wq_barrier(pwq, barr, work, worker);
2711 spin_unlock_irq(&pool->lock);
2714 * If @max_active is 1 or rescuer is in use, flushing another work
2715 * item on the same workqueue may lead to deadlock. Make sure the
2716 * flusher is not running on the same workqueue by verifying write
2719 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2720 lock_map_acquire(&pwq->wq->lockdep_map);
2722 lock_map_acquire_read(&pwq->wq->lockdep_map);
2723 lock_map_release(&pwq->wq->lockdep_map);
2727 spin_unlock_irq(&pool->lock);
2732 * flush_work - wait for a work to finish executing the last queueing instance
2733 * @work: the work to flush
2735 * Wait until @work has finished execution. @work is guaranteed to be idle
2736 * on return if it hasn't been requeued since flush started.
2739 * %true if flush_work() waited for the work to finish execution,
2740 * %false if it was already idle.
2742 bool flush_work(struct work_struct *work)
2744 struct wq_barrier barr;
2746 lock_map_acquire(&work->lockdep_map);
2747 lock_map_release(&work->lockdep_map);
2749 if (start_flush_work(work, &barr)) {
2750 wait_for_completion(&barr.done);
2751 destroy_work_on_stack(&barr.work);
2757 EXPORT_SYMBOL_GPL(flush_work);
2761 struct work_struct *work;
2764 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2766 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2768 if (cwait->work != key)
2770 return autoremove_wake_function(wait, mode, sync, key);
2773 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2775 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2776 unsigned long flags;
2780 ret = try_to_grab_pending(work, is_dwork, &flags);
2782 * If someone else is already canceling, wait for it to
2783 * finish. flush_work() doesn't work for PREEMPT_NONE
2784 * because we may get scheduled between @work's completion
2785 * and the other canceling task resuming and clearing
2786 * CANCELING - flush_work() will return false immediately
2787 * as @work is no longer busy, try_to_grab_pending() will
2788 * return -ENOENT as @work is still being canceled and the
2789 * other canceling task won't be able to clear CANCELING as
2790 * we're hogging the CPU.
2792 * Let's wait for completion using a waitqueue. As this
2793 * may lead to the thundering herd problem, use a custom
2794 * wake function which matches @work along with exclusive
2797 if (unlikely(ret == -ENOENT)) {
2798 struct cwt_wait cwait;
2800 init_wait(&cwait.wait);
2801 cwait.wait.func = cwt_wakefn;
2804 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2805 TASK_UNINTERRUPTIBLE);
2806 if (work_is_canceling(work))
2808 finish_wait(&cancel_waitq, &cwait.wait);
2810 } while (unlikely(ret < 0));
2812 /* tell other tasks trying to grab @work to back off */
2813 mark_work_canceling(work);
2814 local_irq_restore(flags);
2817 clear_work_data(work);
2820 * Paired with prepare_to_wait() above so that either
2821 * waitqueue_active() is visible here or !work_is_canceling() is
2825 if (waitqueue_active(&cancel_waitq))
2826 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2832 * cancel_work_sync - cancel a work and wait for it to finish
2833 * @work: the work to cancel
2835 * Cancel @work and wait for its execution to finish. This function
2836 * can be used even if the work re-queues itself or migrates to
2837 * another workqueue. On return from this function, @work is
2838 * guaranteed to be not pending or executing on any CPU.
2840 * cancel_work_sync(&delayed_work->work) must not be used for
2841 * delayed_work's. Use cancel_delayed_work_sync() instead.
2843 * The caller must ensure that the workqueue on which @work was last
2844 * queued can't be destroyed before this function returns.
2847 * %true if @work was pending, %false otherwise.
2849 bool cancel_work_sync(struct work_struct *work)
2851 return __cancel_work_timer(work, false);
2853 EXPORT_SYMBOL_GPL(cancel_work_sync);
2856 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2857 * @dwork: the delayed work to flush
2859 * Delayed timer is cancelled and the pending work is queued for
2860 * immediate execution. Like flush_work(), this function only
2861 * considers the last queueing instance of @dwork.
2864 * %true if flush_work() waited for the work to finish execution,
2865 * %false if it was already idle.
2867 bool flush_delayed_work(struct delayed_work *dwork)
2869 local_irq_disable();
2870 if (del_timer_sync(&dwork->timer))
2871 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2873 return flush_work(&dwork->work);
2875 EXPORT_SYMBOL(flush_delayed_work);
2878 * cancel_delayed_work - cancel a delayed work
2879 * @dwork: delayed_work to cancel
2881 * Kill off a pending delayed_work.
2883 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2887 * The work callback function may still be running on return, unless
2888 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2889 * use cancel_delayed_work_sync() to wait on it.
2891 * This function is safe to call from any context including IRQ handler.
2893 bool cancel_delayed_work(struct delayed_work *dwork)
2895 unsigned long flags;
2899 ret = try_to_grab_pending(&dwork->work, true, &flags);
2900 } while (unlikely(ret == -EAGAIN));
2902 if (unlikely(ret < 0))
2905 set_work_pool_and_clear_pending(&dwork->work,
2906 get_work_pool_id(&dwork->work));
2907 local_irq_restore(flags);
2910 EXPORT_SYMBOL(cancel_delayed_work);
2913 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2914 * @dwork: the delayed work cancel
2916 * This is cancel_work_sync() for delayed works.
2919 * %true if @dwork was pending, %false otherwise.
2921 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2923 return __cancel_work_timer(&dwork->work, true);
2925 EXPORT_SYMBOL(cancel_delayed_work_sync);
2928 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2929 * @func: the function to call
2931 * schedule_on_each_cpu() executes @func on each online CPU using the
2932 * system workqueue and blocks until all CPUs have completed.
2933 * schedule_on_each_cpu() is very slow.
2936 * 0 on success, -errno on failure.
2938 int schedule_on_each_cpu(work_func_t func)
2941 struct work_struct __percpu *works;
2943 works = alloc_percpu(struct work_struct);
2949 for_each_online_cpu(cpu) {
2950 struct work_struct *work = per_cpu_ptr(works, cpu);
2952 INIT_WORK(work, func);
2953 schedule_work_on(cpu, work);
2956 for_each_online_cpu(cpu)
2957 flush_work(per_cpu_ptr(works, cpu));
2965 * execute_in_process_context - reliably execute the routine with user context
2966 * @fn: the function to execute
2967 * @ew: guaranteed storage for the execute work structure (must
2968 * be available when the work executes)
2970 * Executes the function immediately if process context is available,
2971 * otherwise schedules the function for delayed execution.
2973 * Return: 0 - function was executed
2974 * 1 - function was scheduled for execution
2976 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2978 if (!in_interrupt()) {
2983 INIT_WORK(&ew->work, fn);
2984 schedule_work(&ew->work);
2988 EXPORT_SYMBOL_GPL(execute_in_process_context);
2991 * free_workqueue_attrs - free a workqueue_attrs
2992 * @attrs: workqueue_attrs to free
2994 * Undo alloc_workqueue_attrs().
2996 void free_workqueue_attrs(struct workqueue_attrs *attrs)
2999 free_cpumask_var(attrs->cpumask);
3005 * alloc_workqueue_attrs - allocate a workqueue_attrs
3006 * @gfp_mask: allocation mask to use
3008 * Allocate a new workqueue_attrs, initialize with default settings and
3011 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3013 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3015 struct workqueue_attrs *attrs;
3017 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3020 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3023 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3026 free_workqueue_attrs(attrs);
3030 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3031 const struct workqueue_attrs *from)
3033 to->nice = from->nice;
3034 cpumask_copy(to->cpumask, from->cpumask);
3036 * Unlike hash and equality test, this function doesn't ignore
3037 * ->no_numa as it is used for both pool and wq attrs. Instead,
3038 * get_unbound_pool() explicitly clears ->no_numa after copying.
3040 to->no_numa = from->no_numa;
3043 /* hash value of the content of @attr */
3044 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3048 hash = jhash_1word(attrs->nice, hash);
3049 hash = jhash(cpumask_bits(attrs->cpumask),
3050 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3054 /* content equality test */
3055 static bool wqattrs_equal(const struct workqueue_attrs *a,
3056 const struct workqueue_attrs *b)
3058 if (a->nice != b->nice)
3060 if (!cpumask_equal(a->cpumask, b->cpumask))
3066 * init_worker_pool - initialize a newly zalloc'd worker_pool
3067 * @pool: worker_pool to initialize
3069 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3071 * Return: 0 on success, -errno on failure. Even on failure, all fields
3072 * inside @pool proper are initialized and put_unbound_pool() can be called
3073 * on @pool safely to release it.
3075 static int init_worker_pool(struct worker_pool *pool)
3077 spin_lock_init(&pool->lock);
3080 pool->node = NUMA_NO_NODE;
3081 pool->flags |= POOL_DISASSOCIATED;
3082 INIT_LIST_HEAD(&pool->worklist);
3083 INIT_LIST_HEAD(&pool->idle_list);
3084 hash_init(pool->busy_hash);
3086 init_timer_deferrable(&pool->idle_timer);
3087 pool->idle_timer.function = idle_worker_timeout;
3088 pool->idle_timer.data = (unsigned long)pool;
3090 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3091 (unsigned long)pool);
3093 mutex_init(&pool->manager_arb);
3094 mutex_init(&pool->attach_mutex);
3095 INIT_LIST_HEAD(&pool->workers);
3097 ida_init(&pool->worker_ida);
3098 INIT_HLIST_NODE(&pool->hash_node);
3101 /* shouldn't fail above this point */
3102 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3108 static void rcu_free_wq(struct rcu_head *rcu)
3110 struct workqueue_struct *wq =
3111 container_of(rcu, struct workqueue_struct, rcu);
3113 if (!(wq->flags & WQ_UNBOUND))
3114 free_percpu(wq->cpu_pwqs);
3116 free_workqueue_attrs(wq->unbound_attrs);
3122 static void rcu_free_pool(struct rcu_head *rcu)
3124 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3126 ida_destroy(&pool->worker_ida);
3127 free_workqueue_attrs(pool->attrs);
3132 * put_unbound_pool - put a worker_pool
3133 * @pool: worker_pool to put
3135 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3136 * safe manner. get_unbound_pool() calls this function on its failure path
3137 * and this function should be able to release pools which went through,
3138 * successfully or not, init_worker_pool().
3140 * Should be called with wq_pool_mutex held.
3142 static void put_unbound_pool(struct worker_pool *pool)
3144 DECLARE_COMPLETION_ONSTACK(detach_completion);
3145 struct worker *worker;
3147 lockdep_assert_held(&wq_pool_mutex);
3153 if (WARN_ON(!(pool->cpu < 0)) ||
3154 WARN_ON(!list_empty(&pool->worklist)))
3157 /* release id and unhash */
3159 idr_remove(&worker_pool_idr, pool->id);
3160 hash_del(&pool->hash_node);
3163 * Become the manager and destroy all workers. Grabbing
3164 * manager_arb prevents @pool's workers from blocking on
3167 mutex_lock(&pool->manager_arb);
3169 spin_lock_irq(&pool->lock);
3170 while ((worker = first_idle_worker(pool)))
3171 destroy_worker(worker);
3172 WARN_ON(pool->nr_workers || pool->nr_idle);
3173 spin_unlock_irq(&pool->lock);
3175 mutex_lock(&pool->attach_mutex);
3176 if (!list_empty(&pool->workers))
3177 pool->detach_completion = &detach_completion;
3178 mutex_unlock(&pool->attach_mutex);
3180 if (pool->detach_completion)
3181 wait_for_completion(pool->detach_completion);
3183 mutex_unlock(&pool->manager_arb);
3185 /* shut down the timers */
3186 del_timer_sync(&pool->idle_timer);
3187 del_timer_sync(&pool->mayday_timer);
3189 /* sched-RCU protected to allow dereferences from get_work_pool() */
3190 call_rcu_sched(&pool->rcu, rcu_free_pool);
3194 * get_unbound_pool - get a worker_pool with the specified attributes
3195 * @attrs: the attributes of the worker_pool to get
3197 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3198 * reference count and return it. If there already is a matching
3199 * worker_pool, it will be used; otherwise, this function attempts to
3202 * Should be called with wq_pool_mutex held.
3204 * Return: On success, a worker_pool with the same attributes as @attrs.
3205 * On failure, %NULL.
3207 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3209 u32 hash = wqattrs_hash(attrs);
3210 struct worker_pool *pool;
3212 int target_node = NUMA_NO_NODE;
3214 lockdep_assert_held(&wq_pool_mutex);
3216 /* do we already have a matching pool? */
3217 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3218 if (wqattrs_equal(pool->attrs, attrs)) {
3224 /* if cpumask is contained inside a NUMA node, we belong to that node */
3225 if (wq_numa_enabled) {
3226 for_each_node(node) {
3227 if (cpumask_subset(attrs->cpumask,
3228 wq_numa_possible_cpumask[node])) {
3235 /* nope, create a new one */
3236 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3237 if (!pool || init_worker_pool(pool) < 0)
3240 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3241 copy_workqueue_attrs(pool->attrs, attrs);
3242 pool->node = target_node;
3245 * no_numa isn't a worker_pool attribute, always clear it. See
3246 * 'struct workqueue_attrs' comments for detail.
3248 pool->attrs->no_numa = false;
3250 if (worker_pool_assign_id(pool) < 0)
3253 /* create and start the initial worker */
3254 if (!create_worker(pool))
3258 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3263 put_unbound_pool(pool);
3267 static void rcu_free_pwq(struct rcu_head *rcu)
3269 kmem_cache_free(pwq_cache,
3270 container_of(rcu, struct pool_workqueue, rcu));
3274 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3275 * and needs to be destroyed.
3277 static void pwq_unbound_release_workfn(struct work_struct *work)
3279 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3280 unbound_release_work);
3281 struct workqueue_struct *wq = pwq->wq;
3282 struct worker_pool *pool = pwq->pool;
3285 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3288 mutex_lock(&wq->mutex);
3289 list_del_rcu(&pwq->pwqs_node);
3290 is_last = list_empty(&wq->pwqs);
3291 mutex_unlock(&wq->mutex);
3293 mutex_lock(&wq_pool_mutex);
3294 put_unbound_pool(pool);
3295 mutex_unlock(&wq_pool_mutex);
3297 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3300 * If we're the last pwq going away, @wq is already dead and no one
3301 * is gonna access it anymore. Schedule RCU free.
3304 call_rcu_sched(&wq->rcu, rcu_free_wq);
3308 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3309 * @pwq: target pool_workqueue
3311 * If @pwq isn't freezing, set @pwq->max_active to the associated
3312 * workqueue's saved_max_active and activate delayed work items
3313 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3315 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3317 struct workqueue_struct *wq = pwq->wq;
3318 bool freezable = wq->flags & WQ_FREEZABLE;
3320 /* for @wq->saved_max_active */
3321 lockdep_assert_held(&wq->mutex);
3323 /* fast exit for non-freezable wqs */
3324 if (!freezable && pwq->max_active == wq->saved_max_active)
3327 spin_lock_irq(&pwq->pool->lock);
3330 * During [un]freezing, the caller is responsible for ensuring that
3331 * this function is called at least once after @workqueue_freezing
3332 * is updated and visible.
3334 if (!freezable || !workqueue_freezing) {
3335 pwq->max_active = wq->saved_max_active;
3337 while (!list_empty(&pwq->delayed_works) &&
3338 pwq->nr_active < pwq->max_active)
3339 pwq_activate_first_delayed(pwq);
3342 * Need to kick a worker after thawed or an unbound wq's
3343 * max_active is bumped. It's a slow path. Do it always.
3345 wake_up_worker(pwq->pool);
3347 pwq->max_active = 0;
3350 spin_unlock_irq(&pwq->pool->lock);
3353 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3354 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3355 struct worker_pool *pool)
3357 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3359 memset(pwq, 0, sizeof(*pwq));
3363 pwq->flush_color = -1;
3365 INIT_LIST_HEAD(&pwq->delayed_works);
3366 INIT_LIST_HEAD(&pwq->pwqs_node);
3367 INIT_LIST_HEAD(&pwq->mayday_node);
3368 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3371 /* sync @pwq with the current state of its associated wq and link it */
3372 static void link_pwq(struct pool_workqueue *pwq)
3374 struct workqueue_struct *wq = pwq->wq;
3376 lockdep_assert_held(&wq->mutex);
3378 /* may be called multiple times, ignore if already linked */
3379 if (!list_empty(&pwq->pwqs_node))
3382 /* set the matching work_color */
3383 pwq->work_color = wq->work_color;
3385 /* sync max_active to the current setting */
3386 pwq_adjust_max_active(pwq);
3389 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3392 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3393 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3394 const struct workqueue_attrs *attrs)
3396 struct worker_pool *pool;
3397 struct pool_workqueue *pwq;
3399 lockdep_assert_held(&wq_pool_mutex);
3401 pool = get_unbound_pool(attrs);
3405 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3407 put_unbound_pool(pool);
3411 init_pwq(pwq, wq, pool);
3416 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3417 * @attrs: the wq_attrs of the default pwq of the target workqueue
3418 * @node: the target NUMA node
3419 * @cpu_going_down: if >= 0, the CPU to consider as offline
3420 * @cpumask: outarg, the resulting cpumask
3422 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3423 * @cpu_going_down is >= 0, that cpu is considered offline during
3424 * calculation. The result is stored in @cpumask.
3426 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3427 * enabled and @node has online CPUs requested by @attrs, the returned
3428 * cpumask is the intersection of the possible CPUs of @node and
3431 * The caller is responsible for ensuring that the cpumask of @node stays
3434 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3437 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3438 int cpu_going_down, cpumask_t *cpumask)
3440 if (!wq_numa_enabled || attrs->no_numa)
3443 /* does @node have any online CPUs @attrs wants? */
3444 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3445 if (cpu_going_down >= 0)
3446 cpumask_clear_cpu(cpu_going_down, cpumask);
3448 if (cpumask_empty(cpumask))
3451 /* yeap, return possible CPUs in @node that @attrs wants */
3452 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3453 return !cpumask_equal(cpumask, attrs->cpumask);
3456 cpumask_copy(cpumask, attrs->cpumask);
3460 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3461 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3463 struct pool_workqueue *pwq)
3465 struct pool_workqueue *old_pwq;
3467 lockdep_assert_held(&wq_pool_mutex);
3468 lockdep_assert_held(&wq->mutex);
3470 /* link_pwq() can handle duplicate calls */
3473 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3474 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3478 /* context to store the prepared attrs & pwqs before applying */
3479 struct apply_wqattrs_ctx {
3480 struct workqueue_struct *wq; /* target workqueue */
3481 struct workqueue_attrs *attrs; /* attrs to apply */
3482 struct list_head list; /* queued for batching commit */
3483 struct pool_workqueue *dfl_pwq;
3484 struct pool_workqueue *pwq_tbl[];
3487 /* free the resources after success or abort */
3488 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3494 put_pwq_unlocked(ctx->pwq_tbl[node]);
3495 put_pwq_unlocked(ctx->dfl_pwq);
3497 free_workqueue_attrs(ctx->attrs);
3503 /* allocate the attrs and pwqs for later installation */
3504 static struct apply_wqattrs_ctx *
3505 apply_wqattrs_prepare(struct workqueue_struct *wq,
3506 const struct workqueue_attrs *attrs)
3508 struct apply_wqattrs_ctx *ctx;
3509 struct workqueue_attrs *new_attrs, *tmp_attrs;
3512 lockdep_assert_held(&wq_pool_mutex);
3514 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3517 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3518 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3519 if (!ctx || !new_attrs || !tmp_attrs)
3523 * Calculate the attrs of the default pwq.
3524 * If the user configured cpumask doesn't overlap with the
3525 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3527 copy_workqueue_attrs(new_attrs, attrs);
3528 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3529 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3530 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3533 * We may create multiple pwqs with differing cpumasks. Make a
3534 * copy of @new_attrs which will be modified and used to obtain
3537 copy_workqueue_attrs(tmp_attrs, new_attrs);
3540 * If something goes wrong during CPU up/down, we'll fall back to
3541 * the default pwq covering whole @attrs->cpumask. Always create
3542 * it even if we don't use it immediately.
3544 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3548 for_each_node(node) {
3549 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3550 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3551 if (!ctx->pwq_tbl[node])
3554 ctx->dfl_pwq->refcnt++;
3555 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3559 /* save the user configured attrs and sanitize it. */
3560 copy_workqueue_attrs(new_attrs, attrs);
3561 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3562 ctx->attrs = new_attrs;
3565 free_workqueue_attrs(tmp_attrs);
3569 free_workqueue_attrs(tmp_attrs);
3570 free_workqueue_attrs(new_attrs);
3571 apply_wqattrs_cleanup(ctx);
3575 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3576 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3580 /* all pwqs have been created successfully, let's install'em */
3581 mutex_lock(&ctx->wq->mutex);
3583 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3585 /* save the previous pwq and install the new one */
3587 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3588 ctx->pwq_tbl[node]);
3590 /* @dfl_pwq might not have been used, ensure it's linked */
3591 link_pwq(ctx->dfl_pwq);
3592 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3594 mutex_unlock(&ctx->wq->mutex);
3597 static void apply_wqattrs_lock(void)
3599 /* CPUs should stay stable across pwq creations and installations */
3601 mutex_lock(&wq_pool_mutex);
3604 static void apply_wqattrs_unlock(void)
3606 mutex_unlock(&wq_pool_mutex);
3610 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3611 const struct workqueue_attrs *attrs)
3613 struct apply_wqattrs_ctx *ctx;
3616 /* only unbound workqueues can change attributes */
3617 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3620 /* creating multiple pwqs breaks ordering guarantee */
3621 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3624 ctx = apply_wqattrs_prepare(wq, attrs);
3626 /* the ctx has been prepared successfully, let's commit it */
3628 apply_wqattrs_commit(ctx);
3632 apply_wqattrs_cleanup(ctx);
3638 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3639 * @wq: the target workqueue
3640 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3642 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3643 * machines, this function maps a separate pwq to each NUMA node with
3644 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3645 * NUMA node it was issued on. Older pwqs are released as in-flight work
3646 * items finish. Note that a work item which repeatedly requeues itself
3647 * back-to-back will stay on its current pwq.
3649 * Performs GFP_KERNEL allocations.
3651 * Return: 0 on success and -errno on failure.
3653 int apply_workqueue_attrs(struct workqueue_struct *wq,
3654 const struct workqueue_attrs *attrs)
3658 apply_wqattrs_lock();
3659 ret = apply_workqueue_attrs_locked(wq, attrs);
3660 apply_wqattrs_unlock();
3666 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3667 * @wq: the target workqueue
3668 * @cpu: the CPU coming up or going down
3669 * @online: whether @cpu is coming up or going down
3671 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3672 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3675 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3676 * falls back to @wq->dfl_pwq which may not be optimal but is always
3679 * Note that when the last allowed CPU of a NUMA node goes offline for a
3680 * workqueue with a cpumask spanning multiple nodes, the workers which were
3681 * already executing the work items for the workqueue will lose their CPU
3682 * affinity and may execute on any CPU. This is similar to how per-cpu
3683 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3684 * affinity, it's the user's responsibility to flush the work item from
3687 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3690 int node = cpu_to_node(cpu);
3691 int cpu_off = online ? -1 : cpu;
3692 struct pool_workqueue *old_pwq = NULL, *pwq;
3693 struct workqueue_attrs *target_attrs;
3696 lockdep_assert_held(&wq_pool_mutex);
3698 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3699 wq->unbound_attrs->no_numa)
3703 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3704 * Let's use a preallocated one. The following buf is protected by
3705 * CPU hotplug exclusion.
3707 target_attrs = wq_update_unbound_numa_attrs_buf;
3708 cpumask = target_attrs->cpumask;
3710 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3711 pwq = unbound_pwq_by_node(wq, node);
3714 * Let's determine what needs to be done. If the target cpumask is
3715 * different from the default pwq's, we need to compare it to @pwq's
3716 * and create a new one if they don't match. If the target cpumask
3717 * equals the default pwq's, the default pwq should be used.
3719 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3720 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3726 /* create a new pwq */
3727 pwq = alloc_unbound_pwq(wq, target_attrs);
3729 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3734 /* Install the new pwq. */
3735 mutex_lock(&wq->mutex);
3736 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3740 mutex_lock(&wq->mutex);
3741 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3742 get_pwq(wq->dfl_pwq);
3743 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3744 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3746 mutex_unlock(&wq->mutex);
3747 put_pwq_unlocked(old_pwq);
3750 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3752 bool highpri = wq->flags & WQ_HIGHPRI;
3755 if (!(wq->flags & WQ_UNBOUND)) {
3756 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3760 for_each_possible_cpu(cpu) {
3761 struct pool_workqueue *pwq =
3762 per_cpu_ptr(wq->cpu_pwqs, cpu);
3763 struct worker_pool *cpu_pools =
3764 per_cpu(cpu_worker_pools, cpu);
3766 init_pwq(pwq, wq, &cpu_pools[highpri]);
3768 mutex_lock(&wq->mutex);
3770 mutex_unlock(&wq->mutex);
3773 } else if (wq->flags & __WQ_ORDERED) {
3774 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3775 /* there should only be single pwq for ordering guarantee */
3776 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3777 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3778 "ordering guarantee broken for workqueue %s\n", wq->name);
3781 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3785 static int wq_clamp_max_active(int max_active, unsigned int flags,
3788 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3790 if (max_active < 1 || max_active > lim)
3791 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3792 max_active, name, 1, lim);
3794 return clamp_val(max_active, 1, lim);
3797 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3800 struct lock_class_key *key,
3801 const char *lock_name, ...)
3803 size_t tbl_size = 0;
3805 struct workqueue_struct *wq;
3806 struct pool_workqueue *pwq;
3808 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3809 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3810 flags |= WQ_UNBOUND;
3812 /* allocate wq and format name */
3813 if (flags & WQ_UNBOUND)
3814 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3816 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3820 if (flags & WQ_UNBOUND) {
3821 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3822 if (!wq->unbound_attrs)
3826 va_start(args, lock_name);
3827 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3830 max_active = max_active ?: WQ_DFL_ACTIVE;
3831 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3835 wq->saved_max_active = max_active;
3836 mutex_init(&wq->mutex);
3837 atomic_set(&wq->nr_pwqs_to_flush, 0);
3838 INIT_LIST_HEAD(&wq->pwqs);
3839 INIT_LIST_HEAD(&wq->flusher_queue);
3840 INIT_LIST_HEAD(&wq->flusher_overflow);
3841 INIT_LIST_HEAD(&wq->maydays);
3843 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3844 INIT_LIST_HEAD(&wq->list);
3846 if (alloc_and_link_pwqs(wq) < 0)
3850 * Workqueues which may be used during memory reclaim should
3851 * have a rescuer to guarantee forward progress.
3853 if (flags & WQ_MEM_RECLAIM) {
3854 struct worker *rescuer;
3856 rescuer = alloc_worker(NUMA_NO_NODE);
3860 rescuer->rescue_wq = wq;
3861 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3863 if (IS_ERR(rescuer->task)) {
3868 wq->rescuer = rescuer;
3869 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3870 wake_up_process(rescuer->task);
3873 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3877 * wq_pool_mutex protects global freeze state and workqueues list.
3878 * Grab it, adjust max_active and add the new @wq to workqueues
3881 mutex_lock(&wq_pool_mutex);
3883 mutex_lock(&wq->mutex);
3884 for_each_pwq(pwq, wq)
3885 pwq_adjust_max_active(pwq);
3886 mutex_unlock(&wq->mutex);
3888 list_add_tail_rcu(&wq->list, &workqueues);
3890 mutex_unlock(&wq_pool_mutex);
3895 free_workqueue_attrs(wq->unbound_attrs);
3899 destroy_workqueue(wq);
3902 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3905 * destroy_workqueue - safely terminate a workqueue
3906 * @wq: target workqueue
3908 * Safely destroy a workqueue. All work currently pending will be done first.
3910 void destroy_workqueue(struct workqueue_struct *wq)
3912 struct pool_workqueue *pwq;
3915 /* drain it before proceeding with destruction */
3916 drain_workqueue(wq);
3919 mutex_lock(&wq->mutex);
3920 for_each_pwq(pwq, wq) {
3923 for (i = 0; i < WORK_NR_COLORS; i++) {
3924 if (WARN_ON(pwq->nr_in_flight[i])) {
3925 mutex_unlock(&wq->mutex);
3930 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3931 WARN_ON(pwq->nr_active) ||
3932 WARN_ON(!list_empty(&pwq->delayed_works))) {
3933 mutex_unlock(&wq->mutex);
3937 mutex_unlock(&wq->mutex);
3940 * wq list is used to freeze wq, remove from list after
3941 * flushing is complete in case freeze races us.
3943 mutex_lock(&wq_pool_mutex);
3944 list_del_rcu(&wq->list);
3945 mutex_unlock(&wq_pool_mutex);
3947 workqueue_sysfs_unregister(wq);
3950 kthread_stop(wq->rescuer->task);
3952 if (!(wq->flags & WQ_UNBOUND)) {
3954 * The base ref is never dropped on per-cpu pwqs. Directly
3955 * schedule RCU free.
3957 call_rcu_sched(&wq->rcu, rcu_free_wq);
3960 * We're the sole accessor of @wq at this point. Directly
3961 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3962 * @wq will be freed when the last pwq is released.
3964 for_each_node(node) {
3965 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3966 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3967 put_pwq_unlocked(pwq);
3971 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
3972 * put. Don't access it afterwards.
3976 put_pwq_unlocked(pwq);
3979 EXPORT_SYMBOL_GPL(destroy_workqueue);
3982 * workqueue_set_max_active - adjust max_active of a workqueue
3983 * @wq: target workqueue
3984 * @max_active: new max_active value.
3986 * Set max_active of @wq to @max_active.
3989 * Don't call from IRQ context.
3991 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3993 struct pool_workqueue *pwq;
3995 /* disallow meddling with max_active for ordered workqueues */
3996 if (WARN_ON(wq->flags & __WQ_ORDERED))
3999 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4001 mutex_lock(&wq->mutex);
4003 wq->saved_max_active = max_active;
4005 for_each_pwq(pwq, wq)
4006 pwq_adjust_max_active(pwq);
4008 mutex_unlock(&wq->mutex);
4010 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4013 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4015 * Determine whether %current is a workqueue rescuer. Can be used from
4016 * work functions to determine whether it's being run off the rescuer task.
4018 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4020 bool current_is_workqueue_rescuer(void)
4022 struct worker *worker = current_wq_worker();
4024 return worker && worker->rescue_wq;
4028 * workqueue_congested - test whether a workqueue is congested
4029 * @cpu: CPU in question
4030 * @wq: target workqueue
4032 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4033 * no synchronization around this function and the test result is
4034 * unreliable and only useful as advisory hints or for debugging.
4036 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4037 * Note that both per-cpu and unbound workqueues may be associated with
4038 * multiple pool_workqueues which have separate congested states. A
4039 * workqueue being congested on one CPU doesn't mean the workqueue is also
4040 * contested on other CPUs / NUMA nodes.
4043 * %true if congested, %false otherwise.
4045 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4047 struct pool_workqueue *pwq;
4050 rcu_read_lock_sched();
4052 if (cpu == WORK_CPU_UNBOUND)
4053 cpu = smp_processor_id();
4055 if (!(wq->flags & WQ_UNBOUND))
4056 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4058 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4060 ret = !list_empty(&pwq->delayed_works);
4061 rcu_read_unlock_sched();
4065 EXPORT_SYMBOL_GPL(workqueue_congested);
4068 * work_busy - test whether a work is currently pending or running
4069 * @work: the work to be tested
4071 * Test whether @work is currently pending or running. There is no
4072 * synchronization around this function and the test result is
4073 * unreliable and only useful as advisory hints or for debugging.
4076 * OR'd bitmask of WORK_BUSY_* bits.
4078 unsigned int work_busy(struct work_struct *work)
4080 struct worker_pool *pool;
4081 unsigned long flags;
4082 unsigned int ret = 0;
4084 if (work_pending(work))
4085 ret |= WORK_BUSY_PENDING;
4087 local_irq_save(flags);
4088 pool = get_work_pool(work);
4090 spin_lock(&pool->lock);
4091 if (find_worker_executing_work(pool, work))
4092 ret |= WORK_BUSY_RUNNING;
4093 spin_unlock(&pool->lock);
4095 local_irq_restore(flags);
4099 EXPORT_SYMBOL_GPL(work_busy);
4102 * set_worker_desc - set description for the current work item
4103 * @fmt: printf-style format string
4104 * @...: arguments for the format string
4106 * This function can be called by a running work function to describe what
4107 * the work item is about. If the worker task gets dumped, this
4108 * information will be printed out together to help debugging. The
4109 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4111 void set_worker_desc(const char *fmt, ...)
4113 struct worker *worker = current_wq_worker();
4117 va_start(args, fmt);
4118 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4120 worker->desc_valid = true;
4125 * print_worker_info - print out worker information and description
4126 * @log_lvl: the log level to use when printing
4127 * @task: target task
4129 * If @task is a worker and currently executing a work item, print out the
4130 * name of the workqueue being serviced and worker description set with
4131 * set_worker_desc() by the currently executing work item.
4133 * This function can be safely called on any task as long as the
4134 * task_struct itself is accessible. While safe, this function isn't
4135 * synchronized and may print out mixups or garbages of limited length.
4137 void print_worker_info(const char *log_lvl, struct task_struct *task)
4139 work_func_t *fn = NULL;
4140 char name[WQ_NAME_LEN] = { };
4141 char desc[WORKER_DESC_LEN] = { };
4142 struct pool_workqueue *pwq = NULL;
4143 struct workqueue_struct *wq = NULL;
4144 bool desc_valid = false;
4145 struct worker *worker;
4147 if (!(task->flags & PF_WQ_WORKER))
4151 * This function is called without any synchronization and @task
4152 * could be in any state. Be careful with dereferences.
4154 worker = probe_kthread_data(task);
4157 * Carefully copy the associated workqueue's workfn and name. Keep
4158 * the original last '\0' in case the original contains garbage.
4160 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4161 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4162 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4163 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4165 /* copy worker description */
4166 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4168 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4170 if (fn || name[0] || desc[0]) {
4171 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4173 pr_cont(" (%s)", desc);
4178 static void pr_cont_pool_info(struct worker_pool *pool)
4180 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4181 if (pool->node != NUMA_NO_NODE)
4182 pr_cont(" node=%d", pool->node);
4183 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4186 static void pr_cont_work(bool comma, struct work_struct *work)
4188 if (work->func == wq_barrier_func) {
4189 struct wq_barrier *barr;
4191 barr = container_of(work, struct wq_barrier, work);
4193 pr_cont("%s BAR(%d)", comma ? "," : "",
4194 task_pid_nr(barr->task));
4196 pr_cont("%s %pf", comma ? "," : "", work->func);
4200 static void show_pwq(struct pool_workqueue *pwq)
4202 struct worker_pool *pool = pwq->pool;
4203 struct work_struct *work;
4204 struct worker *worker;
4205 bool has_in_flight = false, has_pending = false;
4208 pr_info(" pwq %d:", pool->id);
4209 pr_cont_pool_info(pool);
4211 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4212 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4214 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4215 if (worker->current_pwq == pwq) {
4216 has_in_flight = true;
4220 if (has_in_flight) {
4223 pr_info(" in-flight:");
4224 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4225 if (worker->current_pwq != pwq)
4228 pr_cont("%s %d%s:%pf", comma ? "," : "",
4229 task_pid_nr(worker->task),
4230 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4231 worker->current_func);
4232 list_for_each_entry(work, &worker->scheduled, entry)
4233 pr_cont_work(false, work);
4239 list_for_each_entry(work, &pool->worklist, entry) {
4240 if (get_work_pwq(work) == pwq) {
4248 pr_info(" pending:");
4249 list_for_each_entry(work, &pool->worklist, entry) {
4250 if (get_work_pwq(work) != pwq)
4253 pr_cont_work(comma, work);
4254 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4259 if (!list_empty(&pwq->delayed_works)) {
4262 pr_info(" delayed:");
4263 list_for_each_entry(work, &pwq->delayed_works, entry) {
4264 pr_cont_work(comma, work);
4265 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4272 * show_workqueue_state - dump workqueue state
4274 * Called from a sysrq handler and prints out all busy workqueues and
4277 void show_workqueue_state(void)
4279 struct workqueue_struct *wq;
4280 struct worker_pool *pool;
4281 unsigned long flags;
4284 rcu_read_lock_sched();
4286 pr_info("Showing busy workqueues and worker pools:\n");
4288 list_for_each_entry_rcu(wq, &workqueues, list) {
4289 struct pool_workqueue *pwq;
4292 for_each_pwq(pwq, wq) {
4293 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4301 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4303 for_each_pwq(pwq, wq) {
4304 spin_lock_irqsave(&pwq->pool->lock, flags);
4305 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4307 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4311 for_each_pool(pool, pi) {
4312 struct worker *worker;
4315 spin_lock_irqsave(&pool->lock, flags);
4316 if (pool->nr_workers == pool->nr_idle)
4319 pr_info("pool %d:", pool->id);
4320 pr_cont_pool_info(pool);
4321 pr_cont(" workers=%d", pool->nr_workers);
4323 pr_cont(" manager: %d",
4324 task_pid_nr(pool->manager->task));
4325 list_for_each_entry(worker, &pool->idle_list, entry) {
4326 pr_cont(" %s%d", first ? "idle: " : "",
4327 task_pid_nr(worker->task));
4332 spin_unlock_irqrestore(&pool->lock, flags);
4335 rcu_read_unlock_sched();
4341 * There are two challenges in supporting CPU hotplug. Firstly, there
4342 * are a lot of assumptions on strong associations among work, pwq and
4343 * pool which make migrating pending and scheduled works very
4344 * difficult to implement without impacting hot paths. Secondly,
4345 * worker pools serve mix of short, long and very long running works making
4346 * blocked draining impractical.
4348 * This is solved by allowing the pools to be disassociated from the CPU
4349 * running as an unbound one and allowing it to be reattached later if the
4350 * cpu comes back online.
4353 static void wq_unbind_fn(struct work_struct *work)
4355 int cpu = smp_processor_id();
4356 struct worker_pool *pool;
4357 struct worker *worker;
4359 for_each_cpu_worker_pool(pool, cpu) {
4360 mutex_lock(&pool->attach_mutex);
4361 spin_lock_irq(&pool->lock);
4364 * We've blocked all attach/detach operations. Make all workers
4365 * unbound and set DISASSOCIATED. Before this, all workers
4366 * except for the ones which are still executing works from
4367 * before the last CPU down must be on the cpu. After
4368 * this, they may become diasporas.
4370 for_each_pool_worker(worker, pool)
4371 worker->flags |= WORKER_UNBOUND;
4373 pool->flags |= POOL_DISASSOCIATED;
4375 spin_unlock_irq(&pool->lock);
4376 mutex_unlock(&pool->attach_mutex);
4379 * Call schedule() so that we cross rq->lock and thus can
4380 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4381 * This is necessary as scheduler callbacks may be invoked
4387 * Sched callbacks are disabled now. Zap nr_running.
4388 * After this, nr_running stays zero and need_more_worker()
4389 * and keep_working() are always true as long as the
4390 * worklist is not empty. This pool now behaves as an
4391 * unbound (in terms of concurrency management) pool which
4392 * are served by workers tied to the pool.
4394 atomic_set(&pool->nr_running, 0);
4397 * With concurrency management just turned off, a busy
4398 * worker blocking could lead to lengthy stalls. Kick off
4399 * unbound chain execution of currently pending work items.
4401 spin_lock_irq(&pool->lock);
4402 wake_up_worker(pool);
4403 spin_unlock_irq(&pool->lock);
4408 * rebind_workers - rebind all workers of a pool to the associated CPU
4409 * @pool: pool of interest
4411 * @pool->cpu is coming online. Rebind all workers to the CPU.
4413 static void rebind_workers(struct worker_pool *pool)
4415 struct worker *worker;
4417 lockdep_assert_held(&pool->attach_mutex);
4420 * Restore CPU affinity of all workers. As all idle workers should
4421 * be on the run-queue of the associated CPU before any local
4422 * wake-ups for concurrency management happen, restore CPU affinity
4423 * of all workers first and then clear UNBOUND. As we're called
4424 * from CPU_ONLINE, the following shouldn't fail.
4426 for_each_pool_worker(worker, pool)
4427 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4428 pool->attrs->cpumask) < 0);
4430 spin_lock_irq(&pool->lock);
4431 pool->flags &= ~POOL_DISASSOCIATED;
4433 for_each_pool_worker(worker, pool) {
4434 unsigned int worker_flags = worker->flags;
4437 * A bound idle worker should actually be on the runqueue
4438 * of the associated CPU for local wake-ups targeting it to
4439 * work. Kick all idle workers so that they migrate to the
4440 * associated CPU. Doing this in the same loop as
4441 * replacing UNBOUND with REBOUND is safe as no worker will
4442 * be bound before @pool->lock is released.
4444 if (worker_flags & WORKER_IDLE)
4445 wake_up_process(worker->task);
4448 * We want to clear UNBOUND but can't directly call
4449 * worker_clr_flags() or adjust nr_running. Atomically
4450 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4451 * @worker will clear REBOUND using worker_clr_flags() when
4452 * it initiates the next execution cycle thus restoring
4453 * concurrency management. Note that when or whether
4454 * @worker clears REBOUND doesn't affect correctness.
4456 * ACCESS_ONCE() is necessary because @worker->flags may be
4457 * tested without holding any lock in
4458 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4459 * fail incorrectly leading to premature concurrency
4460 * management operations.
4462 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4463 worker_flags |= WORKER_REBOUND;
4464 worker_flags &= ~WORKER_UNBOUND;
4465 ACCESS_ONCE(worker->flags) = worker_flags;
4468 spin_unlock_irq(&pool->lock);
4472 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4473 * @pool: unbound pool of interest
4474 * @cpu: the CPU which is coming up
4476 * An unbound pool may end up with a cpumask which doesn't have any online
4477 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4478 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4479 * online CPU before, cpus_allowed of all its workers should be restored.
4481 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4483 static cpumask_t cpumask;
4484 struct worker *worker;
4486 lockdep_assert_held(&pool->attach_mutex);
4488 /* is @cpu allowed for @pool? */
4489 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4492 /* is @cpu the only online CPU? */
4493 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4494 if (cpumask_weight(&cpumask) != 1)
4497 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4498 for_each_pool_worker(worker, pool)
4499 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4500 pool->attrs->cpumask) < 0);
4504 * Workqueues should be brought up before normal priority CPU notifiers.
4505 * This will be registered high priority CPU notifier.
4507 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4508 unsigned long action,
4511 int cpu = (unsigned long)hcpu;
4512 struct worker_pool *pool;
4513 struct workqueue_struct *wq;
4516 switch (action & ~CPU_TASKS_FROZEN) {
4517 case CPU_UP_PREPARE:
4518 for_each_cpu_worker_pool(pool, cpu) {
4519 if (pool->nr_workers)
4521 if (!create_worker(pool))
4526 case CPU_DOWN_FAILED:
4528 mutex_lock(&wq_pool_mutex);
4530 for_each_pool(pool, pi) {
4531 mutex_lock(&pool->attach_mutex);
4533 if (pool->cpu == cpu)
4534 rebind_workers(pool);
4535 else if (pool->cpu < 0)
4536 restore_unbound_workers_cpumask(pool, cpu);
4538 mutex_unlock(&pool->attach_mutex);
4541 /* update NUMA affinity of unbound workqueues */
4542 list_for_each_entry(wq, &workqueues, list)
4543 wq_update_unbound_numa(wq, cpu, true);
4545 mutex_unlock(&wq_pool_mutex);
4552 * Workqueues should be brought down after normal priority CPU notifiers.
4553 * This will be registered as low priority CPU notifier.
4555 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4556 unsigned long action,
4559 int cpu = (unsigned long)hcpu;
4560 struct work_struct unbind_work;
4561 struct workqueue_struct *wq;
4563 switch (action & ~CPU_TASKS_FROZEN) {
4564 case CPU_DOWN_PREPARE:
4565 /* unbinding per-cpu workers should happen on the local CPU */
4566 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4567 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4569 /* update NUMA affinity of unbound workqueues */
4570 mutex_lock(&wq_pool_mutex);
4571 list_for_each_entry(wq, &workqueues, list)
4572 wq_update_unbound_numa(wq, cpu, false);
4573 mutex_unlock(&wq_pool_mutex);
4575 /* wait for per-cpu unbinding to finish */
4576 flush_work(&unbind_work);
4577 destroy_work_on_stack(&unbind_work);
4585 struct work_for_cpu {
4586 struct work_struct work;
4592 static void work_for_cpu_fn(struct work_struct *work)
4594 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4596 wfc->ret = wfc->fn(wfc->arg);
4600 * work_on_cpu - run a function in user context on a particular cpu
4601 * @cpu: the cpu to run on
4602 * @fn: the function to run
4603 * @arg: the function arg
4605 * It is up to the caller to ensure that the cpu doesn't go offline.
4606 * The caller must not hold any locks which would prevent @fn from completing.
4608 * Return: The value @fn returns.
4610 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4612 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4614 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4615 schedule_work_on(cpu, &wfc.work);
4616 flush_work(&wfc.work);
4617 destroy_work_on_stack(&wfc.work);
4620 EXPORT_SYMBOL_GPL(work_on_cpu);
4621 #endif /* CONFIG_SMP */
4623 #ifdef CONFIG_FREEZER
4626 * freeze_workqueues_begin - begin freezing workqueues
4628 * Start freezing workqueues. After this function returns, all freezable
4629 * workqueues will queue new works to their delayed_works list instead of
4633 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4635 void freeze_workqueues_begin(void)
4637 struct workqueue_struct *wq;
4638 struct pool_workqueue *pwq;
4640 mutex_lock(&wq_pool_mutex);
4642 WARN_ON_ONCE(workqueue_freezing);
4643 workqueue_freezing = true;
4645 list_for_each_entry(wq, &workqueues, list) {
4646 mutex_lock(&wq->mutex);
4647 for_each_pwq(pwq, wq)
4648 pwq_adjust_max_active(pwq);
4649 mutex_unlock(&wq->mutex);
4652 mutex_unlock(&wq_pool_mutex);
4656 * freeze_workqueues_busy - are freezable workqueues still busy?
4658 * Check whether freezing is complete. This function must be called
4659 * between freeze_workqueues_begin() and thaw_workqueues().
4662 * Grabs and releases wq_pool_mutex.
4665 * %true if some freezable workqueues are still busy. %false if freezing
4668 bool freeze_workqueues_busy(void)
4671 struct workqueue_struct *wq;
4672 struct pool_workqueue *pwq;
4674 mutex_lock(&wq_pool_mutex);
4676 WARN_ON_ONCE(!workqueue_freezing);
4678 list_for_each_entry(wq, &workqueues, list) {
4679 if (!(wq->flags & WQ_FREEZABLE))
4682 * nr_active is monotonically decreasing. It's safe
4683 * to peek without lock.
4685 rcu_read_lock_sched();
4686 for_each_pwq(pwq, wq) {
4687 WARN_ON_ONCE(pwq->nr_active < 0);
4688 if (pwq->nr_active) {
4690 rcu_read_unlock_sched();
4694 rcu_read_unlock_sched();
4697 mutex_unlock(&wq_pool_mutex);
4702 * thaw_workqueues - thaw workqueues
4704 * Thaw workqueues. Normal queueing is restored and all collected
4705 * frozen works are transferred to their respective pool worklists.
4708 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4710 void thaw_workqueues(void)
4712 struct workqueue_struct *wq;
4713 struct pool_workqueue *pwq;
4715 mutex_lock(&wq_pool_mutex);
4717 if (!workqueue_freezing)
4720 workqueue_freezing = false;
4722 /* restore max_active and repopulate worklist */
4723 list_for_each_entry(wq, &workqueues, list) {
4724 mutex_lock(&wq->mutex);
4725 for_each_pwq(pwq, wq)
4726 pwq_adjust_max_active(pwq);
4727 mutex_unlock(&wq->mutex);
4731 mutex_unlock(&wq_pool_mutex);
4733 #endif /* CONFIG_FREEZER */
4735 static int workqueue_apply_unbound_cpumask(void)
4739 struct workqueue_struct *wq;
4740 struct apply_wqattrs_ctx *ctx, *n;
4742 lockdep_assert_held(&wq_pool_mutex);
4744 list_for_each_entry(wq, &workqueues, list) {
4745 if (!(wq->flags & WQ_UNBOUND))
4747 /* creating multiple pwqs breaks ordering guarantee */
4748 if (wq->flags & __WQ_ORDERED)
4751 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4757 list_add_tail(&ctx->list, &ctxs);
4760 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4762 apply_wqattrs_commit(ctx);
4763 apply_wqattrs_cleanup(ctx);
4770 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4771 * @cpumask: the cpumask to set
4773 * The low-level workqueues cpumask is a global cpumask that limits
4774 * the affinity of all unbound workqueues. This function check the @cpumask
4775 * and apply it to all unbound workqueues and updates all pwqs of them.
4777 * Retun: 0 - Success
4778 * -EINVAL - Invalid @cpumask
4779 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4781 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4784 cpumask_var_t saved_cpumask;
4786 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4789 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4790 if (!cpumask_empty(cpumask)) {
4791 apply_wqattrs_lock();
4793 /* save the old wq_unbound_cpumask. */
4794 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4796 /* update wq_unbound_cpumask at first and apply it to wqs. */
4797 cpumask_copy(wq_unbound_cpumask, cpumask);
4798 ret = workqueue_apply_unbound_cpumask();
4800 /* restore the wq_unbound_cpumask when failed. */
4802 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4804 apply_wqattrs_unlock();
4807 free_cpumask_var(saved_cpumask);
4813 * Workqueues with WQ_SYSFS flag set is visible to userland via
4814 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4815 * following attributes.
4817 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4818 * max_active RW int : maximum number of in-flight work items
4820 * Unbound workqueues have the following extra attributes.
4822 * id RO int : the associated pool ID
4823 * nice RW int : nice value of the workers
4824 * cpumask RW mask : bitmask of allowed CPUs for the workers
4827 struct workqueue_struct *wq;
4831 static struct workqueue_struct *dev_to_wq(struct device *dev)
4833 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4838 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4841 struct workqueue_struct *wq = dev_to_wq(dev);
4843 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4845 static DEVICE_ATTR_RO(per_cpu);
4847 static ssize_t max_active_show(struct device *dev,
4848 struct device_attribute *attr, char *buf)
4850 struct workqueue_struct *wq = dev_to_wq(dev);
4852 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4855 static ssize_t max_active_store(struct device *dev,
4856 struct device_attribute *attr, const char *buf,
4859 struct workqueue_struct *wq = dev_to_wq(dev);
4862 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4865 workqueue_set_max_active(wq, val);
4868 static DEVICE_ATTR_RW(max_active);
4870 static struct attribute *wq_sysfs_attrs[] = {
4871 &dev_attr_per_cpu.attr,
4872 &dev_attr_max_active.attr,
4875 ATTRIBUTE_GROUPS(wq_sysfs);
4877 static ssize_t wq_pool_ids_show(struct device *dev,
4878 struct device_attribute *attr, char *buf)
4880 struct workqueue_struct *wq = dev_to_wq(dev);
4881 const char *delim = "";
4882 int node, written = 0;
4884 rcu_read_lock_sched();
4885 for_each_node(node) {
4886 written += scnprintf(buf + written, PAGE_SIZE - written,
4887 "%s%d:%d", delim, node,
4888 unbound_pwq_by_node(wq, node)->pool->id);
4891 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4892 rcu_read_unlock_sched();
4897 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4900 struct workqueue_struct *wq = dev_to_wq(dev);
4903 mutex_lock(&wq->mutex);
4904 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4905 mutex_unlock(&wq->mutex);
4910 /* prepare workqueue_attrs for sysfs store operations */
4911 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4913 struct workqueue_attrs *attrs;
4915 lockdep_assert_held(&wq_pool_mutex);
4917 attrs = alloc_workqueue_attrs(GFP_KERNEL);
4921 copy_workqueue_attrs(attrs, wq->unbound_attrs);
4925 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4926 const char *buf, size_t count)
4928 struct workqueue_struct *wq = dev_to_wq(dev);
4929 struct workqueue_attrs *attrs;
4932 apply_wqattrs_lock();
4934 attrs = wq_sysfs_prep_attrs(wq);
4938 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4939 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4940 ret = apply_workqueue_attrs_locked(wq, attrs);
4945 apply_wqattrs_unlock();
4946 free_workqueue_attrs(attrs);
4947 return ret ?: count;
4950 static ssize_t wq_cpumask_show(struct device *dev,
4951 struct device_attribute *attr, char *buf)
4953 struct workqueue_struct *wq = dev_to_wq(dev);
4956 mutex_lock(&wq->mutex);
4957 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4958 cpumask_pr_args(wq->unbound_attrs->cpumask));
4959 mutex_unlock(&wq->mutex);
4963 static ssize_t wq_cpumask_store(struct device *dev,
4964 struct device_attribute *attr,
4965 const char *buf, size_t count)
4967 struct workqueue_struct *wq = dev_to_wq(dev);
4968 struct workqueue_attrs *attrs;
4971 apply_wqattrs_lock();
4973 attrs = wq_sysfs_prep_attrs(wq);
4977 ret = cpumask_parse(buf, attrs->cpumask);
4979 ret = apply_workqueue_attrs_locked(wq, attrs);
4982 apply_wqattrs_unlock();
4983 free_workqueue_attrs(attrs);
4984 return ret ?: count;
4987 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
4990 struct workqueue_struct *wq = dev_to_wq(dev);
4993 mutex_lock(&wq->mutex);
4994 written = scnprintf(buf, PAGE_SIZE, "%d\n",
4995 !wq->unbound_attrs->no_numa);
4996 mutex_unlock(&wq->mutex);
5001 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5002 const char *buf, size_t count)
5004 struct workqueue_struct *wq = dev_to_wq(dev);
5005 struct workqueue_attrs *attrs;
5006 int v, ret = -ENOMEM;
5008 apply_wqattrs_lock();
5010 attrs = wq_sysfs_prep_attrs(wq);
5015 if (sscanf(buf, "%d", &v) == 1) {
5016 attrs->no_numa = !v;
5017 ret = apply_workqueue_attrs_locked(wq, attrs);
5021 apply_wqattrs_unlock();
5022 free_workqueue_attrs(attrs);
5023 return ret ?: count;
5026 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5027 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5028 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5029 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5030 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5034 static struct bus_type wq_subsys = {
5035 .name = "workqueue",
5036 .dev_groups = wq_sysfs_groups,
5039 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5040 struct device_attribute *attr, char *buf)
5044 mutex_lock(&wq_pool_mutex);
5045 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5046 cpumask_pr_args(wq_unbound_cpumask));
5047 mutex_unlock(&wq_pool_mutex);
5052 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5053 struct device_attribute *attr, const char *buf, size_t count)
5055 cpumask_var_t cpumask;
5058 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5061 ret = cpumask_parse(buf, cpumask);
5063 ret = workqueue_set_unbound_cpumask(cpumask);
5065 free_cpumask_var(cpumask);
5066 return ret ? ret : count;
5069 static struct device_attribute wq_sysfs_cpumask_attr =
5070 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5071 wq_unbound_cpumask_store);
5073 static int __init wq_sysfs_init(void)
5077 err = subsys_virtual_register(&wq_subsys, NULL);
5081 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5083 core_initcall(wq_sysfs_init);
5085 static void wq_device_release(struct device *dev)
5087 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5093 * workqueue_sysfs_register - make a workqueue visible in sysfs
5094 * @wq: the workqueue to register
5096 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5097 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5098 * which is the preferred method.
5100 * Workqueue user should use this function directly iff it wants to apply
5101 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5102 * apply_workqueue_attrs() may race against userland updating the
5105 * Return: 0 on success, -errno on failure.
5107 int workqueue_sysfs_register(struct workqueue_struct *wq)
5109 struct wq_device *wq_dev;
5113 * Adjusting max_active or creating new pwqs by applying
5114 * attributes breaks ordering guarantee. Disallow exposing ordered
5117 if (WARN_ON(wq->flags & __WQ_ORDERED))
5120 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5125 wq_dev->dev.bus = &wq_subsys;
5126 wq_dev->dev.init_name = wq->name;
5127 wq_dev->dev.release = wq_device_release;
5130 * unbound_attrs are created separately. Suppress uevent until
5131 * everything is ready.
5133 dev_set_uevent_suppress(&wq_dev->dev, true);
5135 ret = device_register(&wq_dev->dev);
5142 if (wq->flags & WQ_UNBOUND) {
5143 struct device_attribute *attr;
5145 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5146 ret = device_create_file(&wq_dev->dev, attr);
5148 device_unregister(&wq_dev->dev);
5155 dev_set_uevent_suppress(&wq_dev->dev, false);
5156 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5161 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5162 * @wq: the workqueue to unregister
5164 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5166 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5168 struct wq_device *wq_dev = wq->wq_dev;
5174 device_unregister(&wq_dev->dev);
5176 #else /* CONFIG_SYSFS */
5177 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5178 #endif /* CONFIG_SYSFS */
5180 static void __init wq_numa_init(void)
5185 if (num_possible_nodes() <= 1)
5188 if (wq_disable_numa) {
5189 pr_info("workqueue: NUMA affinity support disabled\n");
5193 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5194 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5197 * We want masks of possible CPUs of each node which isn't readily
5198 * available. Build one from cpu_to_node() which should have been
5199 * fully initialized by now.
5201 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5205 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5206 node_online(node) ? node : NUMA_NO_NODE));
5208 for_each_possible_cpu(cpu) {
5209 node = cpu_to_node(cpu);
5210 if (WARN_ON(node == NUMA_NO_NODE)) {
5211 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5212 /* happens iff arch is bonkers, let's just proceed */
5215 cpumask_set_cpu(cpu, tbl[node]);
5218 wq_numa_possible_cpumask = tbl;
5219 wq_numa_enabled = true;
5222 static int __init init_workqueues(void)
5224 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5227 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5229 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5230 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5232 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5234 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5235 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5239 /* initialize CPU pools */
5240 for_each_possible_cpu(cpu) {
5241 struct worker_pool *pool;
5244 for_each_cpu_worker_pool(pool, cpu) {
5245 BUG_ON(init_worker_pool(pool));
5247 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5248 pool->attrs->nice = std_nice[i++];
5249 pool->node = cpu_to_node(cpu);
5252 mutex_lock(&wq_pool_mutex);
5253 BUG_ON(worker_pool_assign_id(pool));
5254 mutex_unlock(&wq_pool_mutex);
5258 /* create the initial worker */
5259 for_each_online_cpu(cpu) {
5260 struct worker_pool *pool;
5262 for_each_cpu_worker_pool(pool, cpu) {
5263 pool->flags &= ~POOL_DISASSOCIATED;
5264 BUG_ON(!create_worker(pool));
5268 /* create default unbound and ordered wq attrs */
5269 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5270 struct workqueue_attrs *attrs;
5272 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5273 attrs->nice = std_nice[i];
5274 unbound_std_wq_attrs[i] = attrs;
5277 * An ordered wq should have only one pwq as ordering is
5278 * guaranteed by max_active which is enforced by pwqs.
5279 * Turn off NUMA so that dfl_pwq is used for all nodes.
5281 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5282 attrs->nice = std_nice[i];
5283 attrs->no_numa = true;
5284 ordered_wq_attrs[i] = attrs;
5287 system_wq = alloc_workqueue("events", 0, 0);
5288 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5289 system_long_wq = alloc_workqueue("events_long", 0, 0);
5290 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5291 WQ_UNBOUND_MAX_ACTIVE);
5292 system_freezable_wq = alloc_workqueue("events_freezable",
5294 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5295 WQ_POWER_EFFICIENT, 0);
5296 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5297 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5299 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5300 !system_unbound_wq || !system_freezable_wq ||
5301 !system_power_efficient_wq ||
5302 !system_freezable_power_efficient_wq);
5305 early_initcall(init_workqueues);