x86/xen: resume timer irqs early
[linux/fpc-iii.git] / kernel / workqueue.c
blob3fafbbb319273dde3f460fcd46f02324560ba6bf
1 /*
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>
8 * Andrew Morton
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"
54 enum {
56 * worker_pool flags
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
61 * is in effect.
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 * manager_mutex to avoid changing binding state while
69 * create_worker() is in progress.
71 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
72 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
73 POOL_FREEZING = 1 << 3, /* freeze in progress */
75 /* worker flags */
76 WORKER_STARTED = 1 << 0, /* started */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
97 (min two ticks) */
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give -20.
105 RESCUER_NICE_LEVEL = -20,
106 HIGHPRI_NICE_LEVEL = -20,
108 WQ_NAME_LEN = 24,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
115 * everyone else.
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * MG: pool->manager_mutex and pool->lock protected. Writes require both
128 * locks. Reads can happen under either lock.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
134 * WQ: wq->mutex protected.
136 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138 * MD: wq_mayday_lock protected.
141 /* struct worker is defined in workqueue_internal.h */
143 struct worker_pool {
144 spinlock_t lock; /* the pool lock */
145 int cpu; /* I: the associated cpu */
146 int node; /* I: the associated node ID */
147 int id; /* I: pool ID */
148 unsigned int flags; /* X: flags */
150 struct list_head worklist; /* L: list of pending works */
151 int nr_workers; /* L: total number of workers */
153 /* nr_idle includes the ones off idle_list for rebinding */
154 int nr_idle; /* L: currently idle ones */
156 struct list_head idle_list; /* X: list of idle workers */
157 struct timer_list idle_timer; /* L: worker idle timeout */
158 struct timer_list mayday_timer; /* L: SOS timer for workers */
160 /* a workers is either on busy_hash or idle_list, or the manager */
161 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
162 /* L: hash of busy workers */
164 /* see manage_workers() for details on the two manager mutexes */
165 struct mutex manager_arb; /* manager arbitration */
166 struct mutex manager_mutex; /* manager exclusion */
167 struct idr worker_idr; /* MG: worker IDs and iteration */
169 struct workqueue_attrs *attrs; /* I: worker attributes */
170 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
171 int refcnt; /* PL: refcnt for unbound pools */
174 * The current concurrency level. As it's likely to be accessed
175 * from other CPUs during try_to_wake_up(), put it in a separate
176 * cacheline.
178 atomic_t nr_running ____cacheline_aligned_in_smp;
181 * Destruction of pool is sched-RCU protected to allow dereferences
182 * from get_work_pool().
184 struct rcu_head rcu;
185 } ____cacheline_aligned_in_smp;
188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
193 struct pool_workqueue {
194 struct worker_pool *pool; /* I: the associated pool */
195 struct workqueue_struct *wq; /* I: the owning workqueue */
196 int work_color; /* L: current color */
197 int flush_color; /* L: flushing color */
198 int refcnt; /* L: reference count */
199 int nr_in_flight[WORK_NR_COLORS];
200 /* L: nr of in_flight works */
201 int nr_active; /* L: nr of active works */
202 int max_active; /* L: max active works */
203 struct list_head delayed_works; /* L: delayed works */
204 struct list_head pwqs_node; /* WR: node on wq->pwqs */
205 struct list_head mayday_node; /* MD: node on wq->maydays */
208 * Release of unbound pwq is punted to system_wq. See put_pwq()
209 * and pwq_unbound_release_workfn() for details. pool_workqueue
210 * itself is also sched-RCU protected so that the first pwq can be
211 * determined without grabbing wq->mutex.
213 struct work_struct unbound_release_work;
214 struct rcu_head rcu;
215 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
218 * Structure used to wait for workqueue flush.
220 struct wq_flusher {
221 struct list_head list; /* WQ: list of flushers */
222 int flush_color; /* WQ: flush color waiting for */
223 struct completion done; /* flush completion */
226 struct wq_device;
229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
232 struct workqueue_struct {
233 struct list_head pwqs; /* WR: all pwqs of this wq */
234 struct list_head list; /* PL: list of all workqueues */
236 struct mutex mutex; /* protects this wq */
237 int work_color; /* WQ: current work color */
238 int flush_color; /* WQ: current flush color */
239 atomic_t nr_pwqs_to_flush; /* flush in progress */
240 struct wq_flusher *first_flusher; /* WQ: first flusher */
241 struct list_head flusher_queue; /* WQ: flush waiters */
242 struct list_head flusher_overflow; /* WQ: flush overflow list */
244 struct list_head maydays; /* MD: pwqs requesting rescue */
245 struct worker *rescuer; /* I: rescue worker */
247 int nr_drainers; /* WQ: drain in progress */
248 int saved_max_active; /* WQ: saved pwq max_active */
250 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
251 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
253 #ifdef CONFIG_SYSFS
254 struct wq_device *wq_dev; /* I: for sysfs interface */
255 #endif
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map;
258 #endif
259 char name[WQ_NAME_LEN]; /* I: workqueue name */
261 /* hot fields used during command issue, aligned to cacheline */
262 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
263 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
264 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
267 static struct kmem_cache *pwq_cache;
269 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
270 static cpumask_var_t *wq_numa_possible_cpumask;
271 /* possible CPUs of each node */
273 static bool wq_disable_numa;
274 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
276 /* see the comment above the definition of WQ_POWER_EFFICIENT */
277 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
278 static bool wq_power_efficient = true;
279 #else
280 static bool wq_power_efficient;
281 #endif
283 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
285 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
287 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
288 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
290 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
291 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
293 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
294 static bool workqueue_freezing; /* PL: have wqs started freezing? */
296 /* the per-cpu worker pools */
297 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
298 cpu_worker_pools);
300 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
302 /* PL: hash of all unbound pools keyed by pool->attrs */
303 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
305 /* I: attributes used when instantiating standard unbound pools on demand */
306 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
308 /* I: attributes used when instantiating ordered pools on demand */
309 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
311 struct workqueue_struct *system_wq __read_mostly;
312 EXPORT_SYMBOL(system_wq);
313 struct workqueue_struct *system_highpri_wq __read_mostly;
314 EXPORT_SYMBOL_GPL(system_highpri_wq);
315 struct workqueue_struct *system_long_wq __read_mostly;
316 EXPORT_SYMBOL_GPL(system_long_wq);
317 struct workqueue_struct *system_unbound_wq __read_mostly;
318 EXPORT_SYMBOL_GPL(system_unbound_wq);
319 struct workqueue_struct *system_freezable_wq __read_mostly;
320 EXPORT_SYMBOL_GPL(system_freezable_wq);
321 struct workqueue_struct *system_power_efficient_wq __read_mostly;
322 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
323 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
326 static int worker_thread(void *__worker);
327 static void copy_workqueue_attrs(struct workqueue_attrs *to,
328 const struct workqueue_attrs *from);
330 #define CREATE_TRACE_POINTS
331 #include <trace/events/workqueue.h>
333 #define assert_rcu_or_pool_mutex() \
334 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
335 lockdep_is_held(&wq_pool_mutex), \
336 "sched RCU or wq_pool_mutex should be held")
338 #define assert_rcu_or_wq_mutex(wq) \
339 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
340 lockdep_is_held(&wq->mutex), \
341 "sched RCU or wq->mutex should be held")
343 #ifdef CONFIG_LOCKDEP
344 #define assert_manager_or_pool_lock(pool) \
345 WARN_ONCE(debug_locks && \
346 !lockdep_is_held(&(pool)->manager_mutex) && \
347 !lockdep_is_held(&(pool)->lock), \
348 "pool->manager_mutex or ->lock should be held")
349 #else
350 #define assert_manager_or_pool_lock(pool) do { } while (0)
351 #endif
353 #define for_each_cpu_worker_pool(pool, cpu) \
354 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
355 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
356 (pool)++)
359 * for_each_pool - iterate through all worker_pools in the system
360 * @pool: iteration cursor
361 * @pi: integer used for iteration
363 * This must be called either with wq_pool_mutex held or sched RCU read
364 * locked. If the pool needs to be used beyond the locking in effect, the
365 * caller is responsible for guaranteeing that the pool stays online.
367 * The if/else clause exists only for the lockdep assertion and can be
368 * ignored.
370 #define for_each_pool(pool, pi) \
371 idr_for_each_entry(&worker_pool_idr, pool, pi) \
372 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
373 else
376 * for_each_pool_worker - iterate through all workers of a worker_pool
377 * @worker: iteration cursor
378 * @wi: integer used for iteration
379 * @pool: worker_pool to iterate workers of
381 * This must be called with either @pool->manager_mutex or ->lock held.
383 * The if/else clause exists only for the lockdep assertion and can be
384 * ignored.
386 #define for_each_pool_worker(worker, wi, pool) \
387 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
388 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
389 else
392 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
393 * @pwq: iteration cursor
394 * @wq: the target workqueue
396 * This must be called either with wq->mutex held or sched RCU read locked.
397 * If the pwq needs to be used beyond the locking in effect, the caller is
398 * responsible for guaranteeing that the pwq stays online.
400 * The if/else clause exists only for the lockdep assertion and can be
401 * ignored.
403 #define for_each_pwq(pwq, wq) \
404 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
405 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
406 else
408 #ifdef CONFIG_DEBUG_OBJECTS_WORK
410 static struct debug_obj_descr work_debug_descr;
412 static void *work_debug_hint(void *addr)
414 return ((struct work_struct *) addr)->func;
418 * fixup_init is called when:
419 * - an active object is initialized
421 static int work_fixup_init(void *addr, enum debug_obj_state state)
423 struct work_struct *work = addr;
425 switch (state) {
426 case ODEBUG_STATE_ACTIVE:
427 cancel_work_sync(work);
428 debug_object_init(work, &work_debug_descr);
429 return 1;
430 default:
431 return 0;
436 * fixup_activate is called when:
437 * - an active object is activated
438 * - an unknown object is activated (might be a statically initialized object)
440 static int work_fixup_activate(void *addr, enum debug_obj_state state)
442 struct work_struct *work = addr;
444 switch (state) {
446 case ODEBUG_STATE_NOTAVAILABLE:
448 * This is not really a fixup. The work struct was
449 * statically initialized. We just make sure that it
450 * is tracked in the object tracker.
452 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
453 debug_object_init(work, &work_debug_descr);
454 debug_object_activate(work, &work_debug_descr);
455 return 0;
457 WARN_ON_ONCE(1);
458 return 0;
460 case ODEBUG_STATE_ACTIVE:
461 WARN_ON(1);
463 default:
464 return 0;
469 * fixup_free is called when:
470 * - an active object is freed
472 static int work_fixup_free(void *addr, enum debug_obj_state state)
474 struct work_struct *work = addr;
476 switch (state) {
477 case ODEBUG_STATE_ACTIVE:
478 cancel_work_sync(work);
479 debug_object_free(work, &work_debug_descr);
480 return 1;
481 default:
482 return 0;
486 static struct debug_obj_descr work_debug_descr = {
487 .name = "work_struct",
488 .debug_hint = work_debug_hint,
489 .fixup_init = work_fixup_init,
490 .fixup_activate = work_fixup_activate,
491 .fixup_free = work_fixup_free,
494 static inline void debug_work_activate(struct work_struct *work)
496 debug_object_activate(work, &work_debug_descr);
499 static inline void debug_work_deactivate(struct work_struct *work)
501 debug_object_deactivate(work, &work_debug_descr);
504 void __init_work(struct work_struct *work, int onstack)
506 if (onstack)
507 debug_object_init_on_stack(work, &work_debug_descr);
508 else
509 debug_object_init(work, &work_debug_descr);
511 EXPORT_SYMBOL_GPL(__init_work);
513 void destroy_work_on_stack(struct work_struct *work)
515 debug_object_free(work, &work_debug_descr);
517 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
519 #else
520 static inline void debug_work_activate(struct work_struct *work) { }
521 static inline void debug_work_deactivate(struct work_struct *work) { }
522 #endif
524 /* allocate ID and assign it to @pool */
525 static int worker_pool_assign_id(struct worker_pool *pool)
527 int ret;
529 lockdep_assert_held(&wq_pool_mutex);
531 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
532 if (ret >= 0) {
533 pool->id = ret;
534 return 0;
536 return ret;
540 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
541 * @wq: the target workqueue
542 * @node: the node ID
544 * This must be called either with pwq_lock held or sched RCU read locked.
545 * If the pwq needs to be used beyond the locking in effect, the caller is
546 * responsible for guaranteeing that the pwq stays online.
548 * Return: The unbound pool_workqueue for @node.
550 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
551 int node)
553 assert_rcu_or_wq_mutex(wq);
554 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
557 static unsigned int work_color_to_flags(int color)
559 return color << WORK_STRUCT_COLOR_SHIFT;
562 static int get_work_color(struct work_struct *work)
564 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
565 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
568 static int work_next_color(int color)
570 return (color + 1) % WORK_NR_COLORS;
574 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
575 * contain the pointer to the queued pwq. Once execution starts, the flag
576 * is cleared and the high bits contain OFFQ flags and pool ID.
578 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
579 * and clear_work_data() can be used to set the pwq, pool or clear
580 * work->data. These functions should only be called while the work is
581 * owned - ie. while the PENDING bit is set.
583 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
584 * corresponding to a work. Pool is available once the work has been
585 * queued anywhere after initialization until it is sync canceled. pwq is
586 * available only while the work item is queued.
588 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
589 * canceled. While being canceled, a work item may have its PENDING set
590 * but stay off timer and worklist for arbitrarily long and nobody should
591 * try to steal the PENDING bit.
593 static inline void set_work_data(struct work_struct *work, unsigned long data,
594 unsigned long flags)
596 WARN_ON_ONCE(!work_pending(work));
597 atomic_long_set(&work->data, data | flags | work_static(work));
600 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
601 unsigned long extra_flags)
603 set_work_data(work, (unsigned long)pwq,
604 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
607 static void set_work_pool_and_keep_pending(struct work_struct *work,
608 int pool_id)
610 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
611 WORK_STRUCT_PENDING);
614 static void set_work_pool_and_clear_pending(struct work_struct *work,
615 int pool_id)
618 * The following wmb is paired with the implied mb in
619 * test_and_set_bit(PENDING) and ensures all updates to @work made
620 * here are visible to and precede any updates by the next PENDING
621 * owner.
623 smp_wmb();
624 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
627 static void clear_work_data(struct work_struct *work)
629 smp_wmb(); /* see set_work_pool_and_clear_pending() */
630 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
633 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
635 unsigned long data = atomic_long_read(&work->data);
637 if (data & WORK_STRUCT_PWQ)
638 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
639 else
640 return NULL;
644 * get_work_pool - return the worker_pool a given work was associated with
645 * @work: the work item of interest
647 * Pools are created and destroyed under wq_pool_mutex, and allows read
648 * access under sched-RCU read lock. As such, this function should be
649 * called under wq_pool_mutex or with preemption disabled.
651 * All fields of the returned pool are accessible as long as the above
652 * mentioned locking is in effect. If the returned pool needs to be used
653 * beyond the critical section, the caller is responsible for ensuring the
654 * returned pool is and stays online.
656 * Return: The worker_pool @work was last associated with. %NULL if none.
658 static struct worker_pool *get_work_pool(struct work_struct *work)
660 unsigned long data = atomic_long_read(&work->data);
661 int pool_id;
663 assert_rcu_or_pool_mutex();
665 if (data & WORK_STRUCT_PWQ)
666 return ((struct pool_workqueue *)
667 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
669 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
670 if (pool_id == WORK_OFFQ_POOL_NONE)
671 return NULL;
673 return idr_find(&worker_pool_idr, pool_id);
677 * get_work_pool_id - return the worker pool ID a given work is associated with
678 * @work: the work item of interest
680 * Return: The worker_pool ID @work was last associated with.
681 * %WORK_OFFQ_POOL_NONE if none.
683 static int get_work_pool_id(struct work_struct *work)
685 unsigned long data = atomic_long_read(&work->data);
687 if (data & WORK_STRUCT_PWQ)
688 return ((struct pool_workqueue *)
689 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
691 return data >> WORK_OFFQ_POOL_SHIFT;
694 static void mark_work_canceling(struct work_struct *work)
696 unsigned long pool_id = get_work_pool_id(work);
698 pool_id <<= WORK_OFFQ_POOL_SHIFT;
699 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
702 static bool work_is_canceling(struct work_struct *work)
704 unsigned long data = atomic_long_read(&work->data);
706 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
710 * Policy functions. These define the policies on how the global worker
711 * pools are managed. Unless noted otherwise, these functions assume that
712 * they're being called with pool->lock held.
715 static bool __need_more_worker(struct worker_pool *pool)
717 return !atomic_read(&pool->nr_running);
721 * Need to wake up a worker? Called from anything but currently
722 * running workers.
724 * Note that, because unbound workers never contribute to nr_running, this
725 * function will always return %true for unbound pools as long as the
726 * worklist isn't empty.
728 static bool need_more_worker(struct worker_pool *pool)
730 return !list_empty(&pool->worklist) && __need_more_worker(pool);
733 /* Can I start working? Called from busy but !running workers. */
734 static bool may_start_working(struct worker_pool *pool)
736 return pool->nr_idle;
739 /* Do I need to keep working? Called from currently running workers. */
740 static bool keep_working(struct worker_pool *pool)
742 return !list_empty(&pool->worklist) &&
743 atomic_read(&pool->nr_running) <= 1;
746 /* Do we need a new worker? Called from manager. */
747 static bool need_to_create_worker(struct worker_pool *pool)
749 return need_more_worker(pool) && !may_start_working(pool);
752 /* Do I need to be the manager? */
753 static bool need_to_manage_workers(struct worker_pool *pool)
755 return need_to_create_worker(pool) ||
756 (pool->flags & POOL_MANAGE_WORKERS);
759 /* Do we have too many workers and should some go away? */
760 static bool too_many_workers(struct worker_pool *pool)
762 bool managing = mutex_is_locked(&pool->manager_arb);
763 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
764 int nr_busy = pool->nr_workers - nr_idle;
767 * nr_idle and idle_list may disagree if idle rebinding is in
768 * progress. Never return %true if idle_list is empty.
770 if (list_empty(&pool->idle_list))
771 return false;
773 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
777 * Wake up functions.
780 /* Return the first worker. Safe with preemption disabled */
781 static struct worker *first_worker(struct worker_pool *pool)
783 if (unlikely(list_empty(&pool->idle_list)))
784 return NULL;
786 return list_first_entry(&pool->idle_list, struct worker, entry);
790 * wake_up_worker - wake up an idle worker
791 * @pool: worker pool to wake worker from
793 * Wake up the first idle worker of @pool.
795 * CONTEXT:
796 * spin_lock_irq(pool->lock).
798 static void wake_up_worker(struct worker_pool *pool)
800 struct worker *worker = first_worker(pool);
802 if (likely(worker))
803 wake_up_process(worker->task);
807 * wq_worker_waking_up - a worker is waking up
808 * @task: task waking up
809 * @cpu: CPU @task is waking up to
811 * This function is called during try_to_wake_up() when a worker is
812 * being awoken.
814 * CONTEXT:
815 * spin_lock_irq(rq->lock)
817 void wq_worker_waking_up(struct task_struct *task, int cpu)
819 struct worker *worker = kthread_data(task);
821 if (!(worker->flags & WORKER_NOT_RUNNING)) {
822 WARN_ON_ONCE(worker->pool->cpu != cpu);
823 atomic_inc(&worker->pool->nr_running);
828 * wq_worker_sleeping - a worker is going to sleep
829 * @task: task going to sleep
830 * @cpu: CPU in question, must be the current CPU number
832 * This function is called during schedule() when a busy worker is
833 * going to sleep. Worker on the same cpu can be woken up by
834 * returning pointer to its task.
836 * CONTEXT:
837 * spin_lock_irq(rq->lock)
839 * Return:
840 * Worker task on @cpu to wake up, %NULL if none.
842 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
844 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
845 struct worker_pool *pool;
848 * Rescuers, which may not have all the fields set up like normal
849 * workers, also reach here, let's not access anything before
850 * checking NOT_RUNNING.
852 if (worker->flags & WORKER_NOT_RUNNING)
853 return NULL;
855 pool = worker->pool;
857 /* this can only happen on the local cpu */
858 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
859 return NULL;
862 * The counterpart of the following dec_and_test, implied mb,
863 * worklist not empty test sequence is in insert_work().
864 * Please read comment there.
866 * NOT_RUNNING is clear. This means that we're bound to and
867 * running on the local cpu w/ rq lock held and preemption
868 * disabled, which in turn means that none else could be
869 * manipulating idle_list, so dereferencing idle_list without pool
870 * lock is safe.
872 if (atomic_dec_and_test(&pool->nr_running) &&
873 !list_empty(&pool->worklist))
874 to_wakeup = first_worker(pool);
875 return to_wakeup ? to_wakeup->task : NULL;
879 * worker_set_flags - set worker flags and adjust nr_running accordingly
880 * @worker: self
881 * @flags: flags to set
882 * @wakeup: wakeup an idle worker if necessary
884 * Set @flags in @worker->flags and adjust nr_running accordingly. If
885 * nr_running becomes zero and @wakeup is %true, an idle worker is
886 * woken up.
888 * CONTEXT:
889 * spin_lock_irq(pool->lock)
891 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
892 bool wakeup)
894 struct worker_pool *pool = worker->pool;
896 WARN_ON_ONCE(worker->task != current);
899 * If transitioning into NOT_RUNNING, adjust nr_running and
900 * wake up an idle worker as necessary if requested by
901 * @wakeup.
903 if ((flags & WORKER_NOT_RUNNING) &&
904 !(worker->flags & WORKER_NOT_RUNNING)) {
905 if (wakeup) {
906 if (atomic_dec_and_test(&pool->nr_running) &&
907 !list_empty(&pool->worklist))
908 wake_up_worker(pool);
909 } else
910 atomic_dec(&pool->nr_running);
913 worker->flags |= flags;
917 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
918 * @worker: self
919 * @flags: flags to clear
921 * Clear @flags in @worker->flags and adjust nr_running accordingly.
923 * CONTEXT:
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
955 * being executed.
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.
971 * CONTEXT:
972 * spin_lock_irq(pool->lock).
974 * Return:
975 * Pointer to worker which is executing @work if found, %NULL
976 * otherwise.
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,
984 (unsigned long)work)
985 if (worker->current_work == work &&
986 worker->current_func == work->func)
987 return worker;
989 return NULL;
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 paramter 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().
1006 * CONTEXT:
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))
1021 break;
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.
1029 if (nextp)
1030 *nextp = n;
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);
1044 pwq->refcnt++;
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))
1058 return;
1059 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1060 return;
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)
1080 if (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);
1086 put_pwq(pwq);
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));
1098 pwq->nr_active++;
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.
1117 * CONTEXT:
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)
1124 goto out_put;
1126 pwq->nr_in_flight[color]--;
1128 pwq->nr_active--;
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))
1137 goto out_put;
1139 /* are there still in-flight works? */
1140 if (pwq->nr_in_flight[color])
1141 goto out_put;
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);
1152 out_put:
1153 put_pwq(pwq);
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.
1165 * Return:
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
1172 * Note:
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 */
1192 if (is_dwork) {
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)))
1201 return 1;
1204 /* try to claim PENDING the normal way */
1205 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1206 return 0;
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);
1213 if (!pool)
1214 goto fail;
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(get_work_pwq(work), 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);
1246 return 1;
1248 spin_unlock(&pool->lock);
1249 fail:
1250 local_irq_restore(*flags);
1251 if (work_is_canceling(work))
1252 return -ENOENT;
1253 cpu_relax();
1254 return -EAGAIN;
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.
1267 * CONTEXT:
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);
1278 get_pwq(pwq);
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.
1285 smp_mb();
1287 if (__need_more_worker(pool))
1288 wake_up_worker(pool);
1292 * Test whether @work is being queued from another work executing on the
1293 * same workqueue.
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 dying, only works from the same workqueue are allowed */
1327 if (unlikely(wq->flags & __WQ_DRAINING) &&
1328 WARN_ON_ONCE(!is_chained_work(wq)))
1329 return;
1330 retry:
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);
1337 else
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;
1355 } else {
1356 /* meh... not running there, queue here */
1357 spin_unlock(&last_pool->lock);
1358 spin_lock(&pwq->pool->lock);
1360 } else {
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);
1375 cpu_relax();
1376 goto retry;
1378 /* oops */
1379 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1380 wq->name, cpu);
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);
1388 return;
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);
1396 pwq->nr_active++;
1397 worklist = &pwq->pool->worklist;
1398 } else {
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
1415 * can't go away.
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)
1422 bool ret = false;
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);
1429 ret = true;
1432 local_irq_restore(flags);
1433 return ret;
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.
1463 if (!delay) {
1464 __queue_work(cpu, wq, &dwork->work);
1465 return;
1468 timer_stats_timer_set_start_info(&dwork->timer);
1470 dwork->wq = wq;
1471 dwork->cpu = cpu;
1472 timer->expires = jiffies + delay;
1474 if (unlikely(cpu != WORK_CPU_UNBOUND))
1475 add_timer_on(timer, cpu);
1476 else
1477 add_timer(timer);
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
1489 * execution.
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;
1495 bool ret = false;
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);
1503 ret = true;
1506 local_irq_restore(flags);
1507 return ret;
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
1521 * current state.
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;
1533 int ret;
1535 do {
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 */
1545 return ret;
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
1554 * necessary.
1556 * LOCKING:
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)))
1566 return;
1568 /* can't use worker_set_flags(), also called from start_worker() */
1569 worker->flags |= WORKER_IDLE;
1570 pool->nr_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.
1596 * LOCKING:
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)))
1604 return;
1605 worker_clr_flags(worker, WORKER_IDLE);
1606 pool->nr_idle--;
1607 list_del_init(&worker->entry);
1611 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1612 * @pool: target worker_pool
1614 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1616 * Works which are scheduled while the cpu is online must at least be
1617 * scheduled to a worker which is bound to the cpu so that if they are
1618 * flushed from cpu callbacks while cpu is going down, they are
1619 * guaranteed to execute on the cpu.
1621 * This function is to be used by unbound workers and rescuers to bind
1622 * themselves to the target cpu and may race with cpu going down or
1623 * coming online. kthread_bind() can't be used because it may put the
1624 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1625 * verbatim as it's best effort and blocking and pool may be
1626 * [dis]associated in the meantime.
1628 * This function tries set_cpus_allowed() and locks pool and verifies the
1629 * binding against %POOL_DISASSOCIATED which is set during
1630 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1631 * enters idle state or fetches works without dropping lock, it can
1632 * guarantee the scheduling requirement described in the first paragraph.
1634 * CONTEXT:
1635 * Might sleep. Called without any lock but returns with pool->lock
1636 * held.
1638 * Return:
1639 * %true if the associated pool is online (@worker is successfully
1640 * bound), %false if offline.
1642 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1643 __acquires(&pool->lock)
1645 while (true) {
1647 * The following call may fail, succeed or succeed
1648 * without actually migrating the task to the cpu if
1649 * it races with cpu hotunplug operation. Verify
1650 * against POOL_DISASSOCIATED.
1652 if (!(pool->flags & POOL_DISASSOCIATED))
1653 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1655 spin_lock_irq(&pool->lock);
1656 if (pool->flags & POOL_DISASSOCIATED)
1657 return false;
1658 if (task_cpu(current) == pool->cpu &&
1659 cpumask_equal(&current->cpus_allowed, pool->attrs->cpumask))
1660 return true;
1661 spin_unlock_irq(&pool->lock);
1664 * We've raced with CPU hot[un]plug. Give it a breather
1665 * and retry migration. cond_resched() is required here;
1666 * otherwise, we might deadlock against cpu_stop trying to
1667 * bring down the CPU on non-preemptive kernel.
1669 cpu_relax();
1670 cond_resched();
1674 static struct worker *alloc_worker(void)
1676 struct worker *worker;
1678 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1679 if (worker) {
1680 INIT_LIST_HEAD(&worker->entry);
1681 INIT_LIST_HEAD(&worker->scheduled);
1682 /* on creation a worker is in !idle && prep state */
1683 worker->flags = WORKER_PREP;
1685 return worker;
1689 * create_worker - create a new workqueue worker
1690 * @pool: pool the new worker will belong to
1692 * Create a new worker which is bound to @pool. The returned worker
1693 * can be started by calling start_worker() or destroyed using
1694 * destroy_worker().
1696 * CONTEXT:
1697 * Might sleep. Does GFP_KERNEL allocations.
1699 * Return:
1700 * Pointer to the newly created worker.
1702 static struct worker *create_worker(struct worker_pool *pool)
1704 struct worker *worker = NULL;
1705 int id = -1;
1706 char id_buf[16];
1708 lockdep_assert_held(&pool->manager_mutex);
1711 * ID is needed to determine kthread name. Allocate ID first
1712 * without installing the pointer.
1714 idr_preload(GFP_KERNEL);
1715 spin_lock_irq(&pool->lock);
1717 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1719 spin_unlock_irq(&pool->lock);
1720 idr_preload_end();
1721 if (id < 0)
1722 goto fail;
1724 worker = alloc_worker();
1725 if (!worker)
1726 goto fail;
1728 worker->pool = pool;
1729 worker->id = id;
1731 if (pool->cpu >= 0)
1732 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1733 pool->attrs->nice < 0 ? "H" : "");
1734 else
1735 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1737 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1738 "kworker/%s", id_buf);
1739 if (IS_ERR(worker->task))
1740 goto fail;
1743 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1744 * online CPUs. It'll be re-applied when any of the CPUs come up.
1746 set_user_nice(worker->task, pool->attrs->nice);
1747 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1749 /* prevent userland from meddling with cpumask of workqueue workers */
1750 worker->task->flags |= PF_NO_SETAFFINITY;
1753 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1754 * remains stable across this function. See the comments above the
1755 * flag definition for details.
1757 if (pool->flags & POOL_DISASSOCIATED)
1758 worker->flags |= WORKER_UNBOUND;
1760 /* successful, commit the pointer to idr */
1761 spin_lock_irq(&pool->lock);
1762 idr_replace(&pool->worker_idr, worker, worker->id);
1763 spin_unlock_irq(&pool->lock);
1765 return worker;
1767 fail:
1768 if (id >= 0) {
1769 spin_lock_irq(&pool->lock);
1770 idr_remove(&pool->worker_idr, id);
1771 spin_unlock_irq(&pool->lock);
1773 kfree(worker);
1774 return NULL;
1778 * start_worker - start a newly created worker
1779 * @worker: worker to start
1781 * Make the pool aware of @worker and start it.
1783 * CONTEXT:
1784 * spin_lock_irq(pool->lock).
1786 static void start_worker(struct worker *worker)
1788 worker->flags |= WORKER_STARTED;
1789 worker->pool->nr_workers++;
1790 worker_enter_idle(worker);
1791 wake_up_process(worker->task);
1795 * create_and_start_worker - create and start a worker for a pool
1796 * @pool: the target pool
1798 * Grab the managership of @pool and create and start a new worker for it.
1800 * Return: 0 on success. A negative error code otherwise.
1802 static int create_and_start_worker(struct worker_pool *pool)
1804 struct worker *worker;
1806 mutex_lock(&pool->manager_mutex);
1808 worker = create_worker(pool);
1809 if (worker) {
1810 spin_lock_irq(&pool->lock);
1811 start_worker(worker);
1812 spin_unlock_irq(&pool->lock);
1815 mutex_unlock(&pool->manager_mutex);
1817 return worker ? 0 : -ENOMEM;
1821 * destroy_worker - destroy a workqueue worker
1822 * @worker: worker to be destroyed
1824 * Destroy @worker and adjust @pool stats accordingly.
1826 * CONTEXT:
1827 * spin_lock_irq(pool->lock) which is released and regrabbed.
1829 static void destroy_worker(struct worker *worker)
1831 struct worker_pool *pool = worker->pool;
1833 lockdep_assert_held(&pool->manager_mutex);
1834 lockdep_assert_held(&pool->lock);
1836 /* sanity check frenzy */
1837 if (WARN_ON(worker->current_work) ||
1838 WARN_ON(!list_empty(&worker->scheduled)))
1839 return;
1841 if (worker->flags & WORKER_STARTED)
1842 pool->nr_workers--;
1843 if (worker->flags & WORKER_IDLE)
1844 pool->nr_idle--;
1847 * Once WORKER_DIE is set, the kworker may destroy itself at any
1848 * point. Pin to ensure the task stays until we're done with it.
1850 get_task_struct(worker->task);
1852 list_del_init(&worker->entry);
1853 worker->flags |= WORKER_DIE;
1855 idr_remove(&pool->worker_idr, worker->id);
1857 spin_unlock_irq(&pool->lock);
1859 kthread_stop(worker->task);
1860 put_task_struct(worker->task);
1861 kfree(worker);
1863 spin_lock_irq(&pool->lock);
1866 static void idle_worker_timeout(unsigned long __pool)
1868 struct worker_pool *pool = (void *)__pool;
1870 spin_lock_irq(&pool->lock);
1872 if (too_many_workers(pool)) {
1873 struct worker *worker;
1874 unsigned long expires;
1876 /* idle_list is kept in LIFO order, check the last one */
1877 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1878 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1880 if (time_before(jiffies, expires))
1881 mod_timer(&pool->idle_timer, expires);
1882 else {
1883 /* it's been idle for too long, wake up manager */
1884 pool->flags |= POOL_MANAGE_WORKERS;
1885 wake_up_worker(pool);
1889 spin_unlock_irq(&pool->lock);
1892 static void send_mayday(struct work_struct *work)
1894 struct pool_workqueue *pwq = get_work_pwq(work);
1895 struct workqueue_struct *wq = pwq->wq;
1897 lockdep_assert_held(&wq_mayday_lock);
1899 if (!wq->rescuer)
1900 return;
1902 /* mayday mayday mayday */
1903 if (list_empty(&pwq->mayday_node)) {
1905 * If @pwq is for an unbound wq, its base ref may be put at
1906 * any time due to an attribute change. Pin @pwq until the
1907 * rescuer is done with it.
1909 get_pwq(pwq);
1910 list_add_tail(&pwq->mayday_node, &wq->maydays);
1911 wake_up_process(wq->rescuer->task);
1915 static void pool_mayday_timeout(unsigned long __pool)
1917 struct worker_pool *pool = (void *)__pool;
1918 struct work_struct *work;
1920 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1921 spin_lock(&pool->lock);
1923 if (need_to_create_worker(pool)) {
1925 * We've been trying to create a new worker but
1926 * haven't been successful. We might be hitting an
1927 * allocation deadlock. Send distress signals to
1928 * rescuers.
1930 list_for_each_entry(work, &pool->worklist, entry)
1931 send_mayday(work);
1934 spin_unlock(&pool->lock);
1935 spin_unlock_irq(&wq_mayday_lock);
1937 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1941 * maybe_create_worker - create a new worker if necessary
1942 * @pool: pool to create a new worker for
1944 * Create a new worker for @pool if necessary. @pool is guaranteed to
1945 * have at least one idle worker on return from this function. If
1946 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1947 * sent to all rescuers with works scheduled on @pool to resolve
1948 * possible allocation deadlock.
1950 * On return, need_to_create_worker() is guaranteed to be %false and
1951 * may_start_working() %true.
1953 * LOCKING:
1954 * spin_lock_irq(pool->lock) which may be released and regrabbed
1955 * multiple times. Does GFP_KERNEL allocations. Called only from
1956 * manager.
1958 * Return:
1959 * %false if no action was taken and pool->lock stayed locked, %true
1960 * otherwise.
1962 static bool maybe_create_worker(struct worker_pool *pool)
1963 __releases(&pool->lock)
1964 __acquires(&pool->lock)
1966 if (!need_to_create_worker(pool))
1967 return false;
1968 restart:
1969 spin_unlock_irq(&pool->lock);
1971 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1972 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1974 while (true) {
1975 struct worker *worker;
1977 worker = create_worker(pool);
1978 if (worker) {
1979 del_timer_sync(&pool->mayday_timer);
1980 spin_lock_irq(&pool->lock);
1981 start_worker(worker);
1982 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1983 goto restart;
1984 return true;
1987 if (!need_to_create_worker(pool))
1988 break;
1990 __set_current_state(TASK_INTERRUPTIBLE);
1991 schedule_timeout(CREATE_COOLDOWN);
1993 if (!need_to_create_worker(pool))
1994 break;
1997 del_timer_sync(&pool->mayday_timer);
1998 spin_lock_irq(&pool->lock);
1999 if (need_to_create_worker(pool))
2000 goto restart;
2001 return true;
2005 * maybe_destroy_worker - destroy workers which have been idle for a while
2006 * @pool: pool to destroy workers for
2008 * Destroy @pool workers which have been idle for longer than
2009 * IDLE_WORKER_TIMEOUT.
2011 * LOCKING:
2012 * spin_lock_irq(pool->lock) which may be released and regrabbed
2013 * multiple times. Called only from manager.
2015 * Return:
2016 * %false if no action was taken and pool->lock stayed locked, %true
2017 * otherwise.
2019 static bool maybe_destroy_workers(struct worker_pool *pool)
2021 bool ret = false;
2023 while (too_many_workers(pool)) {
2024 struct worker *worker;
2025 unsigned long expires;
2027 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2028 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2030 if (time_before(jiffies, expires)) {
2031 mod_timer(&pool->idle_timer, expires);
2032 break;
2035 destroy_worker(worker);
2036 ret = true;
2039 return ret;
2043 * manage_workers - manage worker pool
2044 * @worker: self
2046 * Assume the manager role and manage the worker pool @worker belongs
2047 * to. At any given time, there can be only zero or one manager per
2048 * pool. The exclusion is handled automatically by this function.
2050 * The caller can safely start processing works on false return. On
2051 * true return, it's guaranteed that need_to_create_worker() is false
2052 * and may_start_working() is true.
2054 * CONTEXT:
2055 * spin_lock_irq(pool->lock) which may be released and regrabbed
2056 * multiple times. Does GFP_KERNEL allocations.
2058 * Return:
2059 * %false if the pool don't need management and the caller can safely start
2060 * processing works, %true indicates that the function released pool->lock
2061 * and reacquired it to perform some management function and that the
2062 * conditions that the caller verified while holding the lock before
2063 * calling the function might no longer be true.
2065 static bool manage_workers(struct worker *worker)
2067 struct worker_pool *pool = worker->pool;
2068 bool ret = false;
2071 * Managership is governed by two mutexes - manager_arb and
2072 * manager_mutex. manager_arb handles arbitration of manager role.
2073 * Anyone who successfully grabs manager_arb wins the arbitration
2074 * and becomes the manager. mutex_trylock() on pool->manager_arb
2075 * failure while holding pool->lock reliably indicates that someone
2076 * else is managing the pool and the worker which failed trylock
2077 * can proceed to executing work items. This means that anyone
2078 * grabbing manager_arb is responsible for actually performing
2079 * manager duties. If manager_arb is grabbed and released without
2080 * actual management, the pool may stall indefinitely.
2082 * manager_mutex is used for exclusion of actual management
2083 * operations. The holder of manager_mutex can be sure that none
2084 * of management operations, including creation and destruction of
2085 * workers, won't take place until the mutex is released. Because
2086 * manager_mutex doesn't interfere with manager role arbitration,
2087 * it is guaranteed that the pool's management, while may be
2088 * delayed, won't be disturbed by someone else grabbing
2089 * manager_mutex.
2091 if (!mutex_trylock(&pool->manager_arb))
2092 return ret;
2095 * With manager arbitration won, manager_mutex would be free in
2096 * most cases. trylock first without dropping @pool->lock.
2098 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2099 spin_unlock_irq(&pool->lock);
2100 mutex_lock(&pool->manager_mutex);
2101 spin_lock_irq(&pool->lock);
2102 ret = true;
2105 pool->flags &= ~POOL_MANAGE_WORKERS;
2108 * Destroy and then create so that may_start_working() is true
2109 * on return.
2111 ret |= maybe_destroy_workers(pool);
2112 ret |= maybe_create_worker(pool);
2114 mutex_unlock(&pool->manager_mutex);
2115 mutex_unlock(&pool->manager_arb);
2116 return ret;
2120 * process_one_work - process single work
2121 * @worker: self
2122 * @work: work to process
2124 * Process @work. This function contains all the logics necessary to
2125 * process a single work including synchronization against and
2126 * interaction with other workers on the same cpu, queueing and
2127 * flushing. As long as context requirement is met, any worker can
2128 * call this function to process a work.
2130 * CONTEXT:
2131 * spin_lock_irq(pool->lock) which is released and regrabbed.
2133 static void process_one_work(struct worker *worker, struct work_struct *work)
2134 __releases(&pool->lock)
2135 __acquires(&pool->lock)
2137 struct pool_workqueue *pwq = get_work_pwq(work);
2138 struct worker_pool *pool = worker->pool;
2139 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2140 int work_color;
2141 struct worker *collision;
2142 #ifdef CONFIG_LOCKDEP
2144 * It is permissible to free the struct work_struct from
2145 * inside the function that is called from it, this we need to
2146 * take into account for lockdep too. To avoid bogus "held
2147 * lock freed" warnings as well as problems when looking into
2148 * work->lockdep_map, make a copy and use that here.
2150 struct lockdep_map lockdep_map;
2152 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2153 #endif
2155 * Ensure we're on the correct CPU. DISASSOCIATED test is
2156 * necessary to avoid spurious warnings from rescuers servicing the
2157 * unbound or a disassociated pool.
2159 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2160 !(pool->flags & POOL_DISASSOCIATED) &&
2161 raw_smp_processor_id() != pool->cpu);
2164 * A single work shouldn't be executed concurrently by
2165 * multiple workers on a single cpu. Check whether anyone is
2166 * already processing the work. If so, defer the work to the
2167 * currently executing one.
2169 collision = find_worker_executing_work(pool, work);
2170 if (unlikely(collision)) {
2171 move_linked_works(work, &collision->scheduled, NULL);
2172 return;
2175 /* claim and dequeue */
2176 debug_work_deactivate(work);
2177 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2178 worker->current_work = work;
2179 worker->current_func = work->func;
2180 worker->current_pwq = pwq;
2181 work_color = get_work_color(work);
2183 list_del_init(&work->entry);
2186 * CPU intensive works don't participate in concurrency
2187 * management. They're the scheduler's responsibility.
2189 if (unlikely(cpu_intensive))
2190 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2193 * Unbound pool isn't concurrency managed and work items should be
2194 * executed ASAP. Wake up another worker if necessary.
2196 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2197 wake_up_worker(pool);
2200 * Record the last pool and clear PENDING which should be the last
2201 * update to @work. Also, do this inside @pool->lock so that
2202 * PENDING and queued state changes happen together while IRQ is
2203 * disabled.
2205 set_work_pool_and_clear_pending(work, pool->id);
2207 spin_unlock_irq(&pool->lock);
2209 lock_map_acquire_read(&pwq->wq->lockdep_map);
2210 lock_map_acquire(&lockdep_map);
2211 trace_workqueue_execute_start(work);
2212 worker->current_func(work);
2214 * While we must be careful to not use "work" after this, the trace
2215 * point will only record its address.
2217 trace_workqueue_execute_end(work);
2218 lock_map_release(&lockdep_map);
2219 lock_map_release(&pwq->wq->lockdep_map);
2221 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2222 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2223 " last function: %pf\n",
2224 current->comm, preempt_count(), task_pid_nr(current),
2225 worker->current_func);
2226 debug_show_held_locks(current);
2227 dump_stack();
2231 * The following prevents a kworker from hogging CPU on !PREEMPT
2232 * kernels, where a requeueing work item waiting for something to
2233 * happen could deadlock with stop_machine as such work item could
2234 * indefinitely requeue itself while all other CPUs are trapped in
2235 * stop_machine.
2237 cond_resched();
2239 spin_lock_irq(&pool->lock);
2241 /* clear cpu intensive status */
2242 if (unlikely(cpu_intensive))
2243 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2245 /* we're done with it, release */
2246 hash_del(&worker->hentry);
2247 worker->current_work = NULL;
2248 worker->current_func = NULL;
2249 worker->current_pwq = NULL;
2250 worker->desc_valid = false;
2251 pwq_dec_nr_in_flight(pwq, work_color);
2255 * process_scheduled_works - process scheduled works
2256 * @worker: self
2258 * Process all scheduled works. Please note that the scheduled list
2259 * may change while processing a work, so this function repeatedly
2260 * fetches a work from the top and executes it.
2262 * CONTEXT:
2263 * spin_lock_irq(pool->lock) which may be released and regrabbed
2264 * multiple times.
2266 static void process_scheduled_works(struct worker *worker)
2268 while (!list_empty(&worker->scheduled)) {
2269 struct work_struct *work = list_first_entry(&worker->scheduled,
2270 struct work_struct, entry);
2271 process_one_work(worker, work);
2276 * worker_thread - the worker thread function
2277 * @__worker: self
2279 * The worker thread function. All workers belong to a worker_pool -
2280 * either a per-cpu one or dynamic unbound one. These workers process all
2281 * work items regardless of their specific target workqueue. The only
2282 * exception is work items which belong to workqueues with a rescuer which
2283 * will be explained in rescuer_thread().
2285 * Return: 0
2287 static int worker_thread(void *__worker)
2289 struct worker *worker = __worker;
2290 struct worker_pool *pool = worker->pool;
2292 /* tell the scheduler that this is a workqueue worker */
2293 worker->task->flags |= PF_WQ_WORKER;
2294 woke_up:
2295 spin_lock_irq(&pool->lock);
2297 /* am I supposed to die? */
2298 if (unlikely(worker->flags & WORKER_DIE)) {
2299 spin_unlock_irq(&pool->lock);
2300 WARN_ON_ONCE(!list_empty(&worker->entry));
2301 worker->task->flags &= ~PF_WQ_WORKER;
2302 return 0;
2305 worker_leave_idle(worker);
2306 recheck:
2307 /* no more worker necessary? */
2308 if (!need_more_worker(pool))
2309 goto sleep;
2311 /* do we need to manage? */
2312 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2313 goto recheck;
2316 * ->scheduled list can only be filled while a worker is
2317 * preparing to process a work or actually processing it.
2318 * Make sure nobody diddled with it while I was sleeping.
2320 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2323 * Finish PREP stage. We're guaranteed to have at least one idle
2324 * worker or that someone else has already assumed the manager
2325 * role. This is where @worker starts participating in concurrency
2326 * management if applicable and concurrency management is restored
2327 * after being rebound. See rebind_workers() for details.
2329 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2331 do {
2332 struct work_struct *work =
2333 list_first_entry(&pool->worklist,
2334 struct work_struct, entry);
2336 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2337 /* optimization path, not strictly necessary */
2338 process_one_work(worker, work);
2339 if (unlikely(!list_empty(&worker->scheduled)))
2340 process_scheduled_works(worker);
2341 } else {
2342 move_linked_works(work, &worker->scheduled, NULL);
2343 process_scheduled_works(worker);
2345 } while (keep_working(pool));
2347 worker_set_flags(worker, WORKER_PREP, false);
2348 sleep:
2349 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2350 goto recheck;
2353 * pool->lock is held and there's no work to process and no need to
2354 * manage, sleep. Workers are woken up only while holding
2355 * pool->lock or from local cpu, so setting the current state
2356 * before releasing pool->lock is enough to prevent losing any
2357 * event.
2359 worker_enter_idle(worker);
2360 __set_current_state(TASK_INTERRUPTIBLE);
2361 spin_unlock_irq(&pool->lock);
2362 schedule();
2363 goto woke_up;
2367 * rescuer_thread - the rescuer thread function
2368 * @__rescuer: self
2370 * Workqueue rescuer thread function. There's one rescuer for each
2371 * workqueue which has WQ_MEM_RECLAIM set.
2373 * Regular work processing on a pool may block trying to create a new
2374 * worker which uses GFP_KERNEL allocation which has slight chance of
2375 * developing into deadlock if some works currently on the same queue
2376 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2377 * the problem rescuer solves.
2379 * When such condition is possible, the pool summons rescuers of all
2380 * workqueues which have works queued on the pool and let them process
2381 * those works so that forward progress can be guaranteed.
2383 * This should happen rarely.
2385 * Return: 0
2387 static int rescuer_thread(void *__rescuer)
2389 struct worker *rescuer = __rescuer;
2390 struct workqueue_struct *wq = rescuer->rescue_wq;
2391 struct list_head *scheduled = &rescuer->scheduled;
2392 bool should_stop;
2394 set_user_nice(current, RESCUER_NICE_LEVEL);
2397 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2398 * doesn't participate in concurrency management.
2400 rescuer->task->flags |= PF_WQ_WORKER;
2401 repeat:
2402 set_current_state(TASK_INTERRUPTIBLE);
2405 * By the time the rescuer is requested to stop, the workqueue
2406 * shouldn't have any work pending, but @wq->maydays may still have
2407 * pwq(s) queued. This can happen by non-rescuer workers consuming
2408 * all the work items before the rescuer got to them. Go through
2409 * @wq->maydays processing before acting on should_stop so that the
2410 * list is always empty on exit.
2412 should_stop = kthread_should_stop();
2414 /* see whether any pwq is asking for help */
2415 spin_lock_irq(&wq_mayday_lock);
2417 while (!list_empty(&wq->maydays)) {
2418 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2419 struct pool_workqueue, mayday_node);
2420 struct worker_pool *pool = pwq->pool;
2421 struct work_struct *work, *n;
2423 __set_current_state(TASK_RUNNING);
2424 list_del_init(&pwq->mayday_node);
2426 spin_unlock_irq(&wq_mayday_lock);
2428 /* migrate to the target cpu if possible */
2429 worker_maybe_bind_and_lock(pool);
2430 rescuer->pool = pool;
2433 * Slurp in all works issued via this workqueue and
2434 * process'em.
2436 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2437 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2438 if (get_work_pwq(work) == pwq)
2439 move_linked_works(work, scheduled, &n);
2441 process_scheduled_works(rescuer);
2444 * Put the reference grabbed by send_mayday(). @pool won't
2445 * go away while we're holding its lock.
2447 put_pwq(pwq);
2450 * Leave this pool. If keep_working() is %true, notify a
2451 * regular worker; otherwise, we end up with 0 concurrency
2452 * and stalling the execution.
2454 if (keep_working(pool))
2455 wake_up_worker(pool);
2457 rescuer->pool = NULL;
2458 spin_unlock(&pool->lock);
2459 spin_lock(&wq_mayday_lock);
2462 spin_unlock_irq(&wq_mayday_lock);
2464 if (should_stop) {
2465 __set_current_state(TASK_RUNNING);
2466 rescuer->task->flags &= ~PF_WQ_WORKER;
2467 return 0;
2470 /* rescuers should never participate in concurrency management */
2471 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2472 schedule();
2473 goto repeat;
2476 struct wq_barrier {
2477 struct work_struct work;
2478 struct completion done;
2481 static void wq_barrier_func(struct work_struct *work)
2483 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2484 complete(&barr->done);
2488 * insert_wq_barrier - insert a barrier work
2489 * @pwq: pwq to insert barrier into
2490 * @barr: wq_barrier to insert
2491 * @target: target work to attach @barr to
2492 * @worker: worker currently executing @target, NULL if @target is not executing
2494 * @barr is linked to @target such that @barr is completed only after
2495 * @target finishes execution. Please note that the ordering
2496 * guarantee is observed only with respect to @target and on the local
2497 * cpu.
2499 * Currently, a queued barrier can't be canceled. This is because
2500 * try_to_grab_pending() can't determine whether the work to be
2501 * grabbed is at the head of the queue and thus can't clear LINKED
2502 * flag of the previous work while there must be a valid next work
2503 * after a work with LINKED flag set.
2505 * Note that when @worker is non-NULL, @target may be modified
2506 * underneath us, so we can't reliably determine pwq from @target.
2508 * CONTEXT:
2509 * spin_lock_irq(pool->lock).
2511 static void insert_wq_barrier(struct pool_workqueue *pwq,
2512 struct wq_barrier *barr,
2513 struct work_struct *target, struct worker *worker)
2515 struct list_head *head;
2516 unsigned int linked = 0;
2519 * debugobject calls are safe here even with pool->lock locked
2520 * as we know for sure that this will not trigger any of the
2521 * checks and call back into the fixup functions where we
2522 * might deadlock.
2524 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2525 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2526 init_completion(&barr->done);
2529 * If @target is currently being executed, schedule the
2530 * barrier to the worker; otherwise, put it after @target.
2532 if (worker)
2533 head = worker->scheduled.next;
2534 else {
2535 unsigned long *bits = work_data_bits(target);
2537 head = target->entry.next;
2538 /* there can already be other linked works, inherit and set */
2539 linked = *bits & WORK_STRUCT_LINKED;
2540 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2543 debug_work_activate(&barr->work);
2544 insert_work(pwq, &barr->work, head,
2545 work_color_to_flags(WORK_NO_COLOR) | linked);
2549 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2550 * @wq: workqueue being flushed
2551 * @flush_color: new flush color, < 0 for no-op
2552 * @work_color: new work color, < 0 for no-op
2554 * Prepare pwqs for workqueue flushing.
2556 * If @flush_color is non-negative, flush_color on all pwqs should be
2557 * -1. If no pwq has in-flight commands at the specified color, all
2558 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2559 * has in flight commands, its pwq->flush_color is set to
2560 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2561 * wakeup logic is armed and %true is returned.
2563 * The caller should have initialized @wq->first_flusher prior to
2564 * calling this function with non-negative @flush_color. If
2565 * @flush_color is negative, no flush color update is done and %false
2566 * is returned.
2568 * If @work_color is non-negative, all pwqs should have the same
2569 * work_color which is previous to @work_color and all will be
2570 * advanced to @work_color.
2572 * CONTEXT:
2573 * mutex_lock(wq->mutex).
2575 * Return:
2576 * %true if @flush_color >= 0 and there's something to flush. %false
2577 * otherwise.
2579 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2580 int flush_color, int work_color)
2582 bool wait = false;
2583 struct pool_workqueue *pwq;
2585 if (flush_color >= 0) {
2586 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2587 atomic_set(&wq->nr_pwqs_to_flush, 1);
2590 for_each_pwq(pwq, wq) {
2591 struct worker_pool *pool = pwq->pool;
2593 spin_lock_irq(&pool->lock);
2595 if (flush_color >= 0) {
2596 WARN_ON_ONCE(pwq->flush_color != -1);
2598 if (pwq->nr_in_flight[flush_color]) {
2599 pwq->flush_color = flush_color;
2600 atomic_inc(&wq->nr_pwqs_to_flush);
2601 wait = true;
2605 if (work_color >= 0) {
2606 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2607 pwq->work_color = work_color;
2610 spin_unlock_irq(&pool->lock);
2613 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2614 complete(&wq->first_flusher->done);
2616 return wait;
2620 * flush_workqueue - ensure that any scheduled work has run to completion.
2621 * @wq: workqueue to flush
2623 * This function sleeps until all work items which were queued on entry
2624 * have finished execution, but it is not livelocked by new incoming ones.
2626 void flush_workqueue(struct workqueue_struct *wq)
2628 struct wq_flusher this_flusher = {
2629 .list = LIST_HEAD_INIT(this_flusher.list),
2630 .flush_color = -1,
2631 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2633 int next_color;
2635 lock_map_acquire(&wq->lockdep_map);
2636 lock_map_release(&wq->lockdep_map);
2638 mutex_lock(&wq->mutex);
2641 * Start-to-wait phase
2643 next_color = work_next_color(wq->work_color);
2645 if (next_color != wq->flush_color) {
2647 * Color space is not full. The current work_color
2648 * becomes our flush_color and work_color is advanced
2649 * by one.
2651 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2652 this_flusher.flush_color = wq->work_color;
2653 wq->work_color = next_color;
2655 if (!wq->first_flusher) {
2656 /* no flush in progress, become the first flusher */
2657 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2659 wq->first_flusher = &this_flusher;
2661 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2662 wq->work_color)) {
2663 /* nothing to flush, done */
2664 wq->flush_color = next_color;
2665 wq->first_flusher = NULL;
2666 goto out_unlock;
2668 } else {
2669 /* wait in queue */
2670 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2671 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2672 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2674 } else {
2676 * Oops, color space is full, wait on overflow queue.
2677 * The next flush completion will assign us
2678 * flush_color and transfer to flusher_queue.
2680 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2683 mutex_unlock(&wq->mutex);
2685 wait_for_completion(&this_flusher.done);
2688 * Wake-up-and-cascade phase
2690 * First flushers are responsible for cascading flushes and
2691 * handling overflow. Non-first flushers can simply return.
2693 if (wq->first_flusher != &this_flusher)
2694 return;
2696 mutex_lock(&wq->mutex);
2698 /* we might have raced, check again with mutex held */
2699 if (wq->first_flusher != &this_flusher)
2700 goto out_unlock;
2702 wq->first_flusher = NULL;
2704 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2705 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2707 while (true) {
2708 struct wq_flusher *next, *tmp;
2710 /* complete all the flushers sharing the current flush color */
2711 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2712 if (next->flush_color != wq->flush_color)
2713 break;
2714 list_del_init(&next->list);
2715 complete(&next->done);
2718 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2719 wq->flush_color != work_next_color(wq->work_color));
2721 /* this flush_color is finished, advance by one */
2722 wq->flush_color = work_next_color(wq->flush_color);
2724 /* one color has been freed, handle overflow queue */
2725 if (!list_empty(&wq->flusher_overflow)) {
2727 * Assign the same color to all overflowed
2728 * flushers, advance work_color and append to
2729 * flusher_queue. This is the start-to-wait
2730 * phase for these overflowed flushers.
2732 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2733 tmp->flush_color = wq->work_color;
2735 wq->work_color = work_next_color(wq->work_color);
2737 list_splice_tail_init(&wq->flusher_overflow,
2738 &wq->flusher_queue);
2739 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2742 if (list_empty(&wq->flusher_queue)) {
2743 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2744 break;
2748 * Need to flush more colors. Make the next flusher
2749 * the new first flusher and arm pwqs.
2751 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2752 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2754 list_del_init(&next->list);
2755 wq->first_flusher = next;
2757 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2758 break;
2761 * Meh... this color is already done, clear first
2762 * flusher and repeat cascading.
2764 wq->first_flusher = NULL;
2767 out_unlock:
2768 mutex_unlock(&wq->mutex);
2770 EXPORT_SYMBOL_GPL(flush_workqueue);
2773 * drain_workqueue - drain a workqueue
2774 * @wq: workqueue to drain
2776 * Wait until the workqueue becomes empty. While draining is in progress,
2777 * only chain queueing is allowed. IOW, only currently pending or running
2778 * work items on @wq can queue further work items on it. @wq is flushed
2779 * repeatedly until it becomes empty. The number of flushing is detemined
2780 * by the depth of chaining and should be relatively short. Whine if it
2781 * takes too long.
2783 void drain_workqueue(struct workqueue_struct *wq)
2785 unsigned int flush_cnt = 0;
2786 struct pool_workqueue *pwq;
2789 * __queue_work() needs to test whether there are drainers, is much
2790 * hotter than drain_workqueue() and already looks at @wq->flags.
2791 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2793 mutex_lock(&wq->mutex);
2794 if (!wq->nr_drainers++)
2795 wq->flags |= __WQ_DRAINING;
2796 mutex_unlock(&wq->mutex);
2797 reflush:
2798 flush_workqueue(wq);
2800 mutex_lock(&wq->mutex);
2802 for_each_pwq(pwq, wq) {
2803 bool drained;
2805 spin_lock_irq(&pwq->pool->lock);
2806 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2807 spin_unlock_irq(&pwq->pool->lock);
2809 if (drained)
2810 continue;
2812 if (++flush_cnt == 10 ||
2813 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2814 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2815 wq->name, flush_cnt);
2817 mutex_unlock(&wq->mutex);
2818 goto reflush;
2821 if (!--wq->nr_drainers)
2822 wq->flags &= ~__WQ_DRAINING;
2823 mutex_unlock(&wq->mutex);
2825 EXPORT_SYMBOL_GPL(drain_workqueue);
2827 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2829 struct worker *worker = NULL;
2830 struct worker_pool *pool;
2831 struct pool_workqueue *pwq;
2833 might_sleep();
2835 local_irq_disable();
2836 pool = get_work_pool(work);
2837 if (!pool) {
2838 local_irq_enable();
2839 return false;
2842 spin_lock(&pool->lock);
2843 /* see the comment in try_to_grab_pending() with the same code */
2844 pwq = get_work_pwq(work);
2845 if (pwq) {
2846 if (unlikely(pwq->pool != pool))
2847 goto already_gone;
2848 } else {
2849 worker = find_worker_executing_work(pool, work);
2850 if (!worker)
2851 goto already_gone;
2852 pwq = worker->current_pwq;
2855 insert_wq_barrier(pwq, barr, work, worker);
2856 spin_unlock_irq(&pool->lock);
2859 * If @max_active is 1 or rescuer is in use, flushing another work
2860 * item on the same workqueue may lead to deadlock. Make sure the
2861 * flusher is not running on the same workqueue by verifying write
2862 * access.
2864 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2865 lock_map_acquire(&pwq->wq->lockdep_map);
2866 else
2867 lock_map_acquire_read(&pwq->wq->lockdep_map);
2868 lock_map_release(&pwq->wq->lockdep_map);
2870 return true;
2871 already_gone:
2872 spin_unlock_irq(&pool->lock);
2873 return false;
2876 static bool __flush_work(struct work_struct *work)
2878 struct wq_barrier barr;
2880 if (start_flush_work(work, &barr)) {
2881 wait_for_completion(&barr.done);
2882 destroy_work_on_stack(&barr.work);
2883 return true;
2884 } else {
2885 return false;
2890 * flush_work - wait for a work to finish executing the last queueing instance
2891 * @work: the work to flush
2893 * Wait until @work has finished execution. @work is guaranteed to be idle
2894 * on return if it hasn't been requeued since flush started.
2896 * Return:
2897 * %true if flush_work() waited for the work to finish execution,
2898 * %false if it was already idle.
2900 bool flush_work(struct work_struct *work)
2902 lock_map_acquire(&work->lockdep_map);
2903 lock_map_release(&work->lockdep_map);
2905 return __flush_work(work);
2907 EXPORT_SYMBOL_GPL(flush_work);
2909 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2911 unsigned long flags;
2912 int ret;
2914 do {
2915 ret = try_to_grab_pending(work, is_dwork, &flags);
2917 * If someone else is canceling, wait for the same event it
2918 * would be waiting for before retrying.
2920 if (unlikely(ret == -ENOENT))
2921 flush_work(work);
2922 } while (unlikely(ret < 0));
2924 /* tell other tasks trying to grab @work to back off */
2925 mark_work_canceling(work);
2926 local_irq_restore(flags);
2928 flush_work(work);
2929 clear_work_data(work);
2930 return ret;
2934 * cancel_work_sync - cancel a work and wait for it to finish
2935 * @work: the work to cancel
2937 * Cancel @work and wait for its execution to finish. This function
2938 * can be used even if the work re-queues itself or migrates to
2939 * another workqueue. On return from this function, @work is
2940 * guaranteed to be not pending or executing on any CPU.
2942 * cancel_work_sync(&delayed_work->work) must not be used for
2943 * delayed_work's. Use cancel_delayed_work_sync() instead.
2945 * The caller must ensure that the workqueue on which @work was last
2946 * queued can't be destroyed before this function returns.
2948 * Return:
2949 * %true if @work was pending, %false otherwise.
2951 bool cancel_work_sync(struct work_struct *work)
2953 return __cancel_work_timer(work, false);
2955 EXPORT_SYMBOL_GPL(cancel_work_sync);
2958 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2959 * @dwork: the delayed work to flush
2961 * Delayed timer is cancelled and the pending work is queued for
2962 * immediate execution. Like flush_work(), this function only
2963 * considers the last queueing instance of @dwork.
2965 * Return:
2966 * %true if flush_work() waited for the work to finish execution,
2967 * %false if it was already idle.
2969 bool flush_delayed_work(struct delayed_work *dwork)
2971 local_irq_disable();
2972 if (del_timer_sync(&dwork->timer))
2973 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2974 local_irq_enable();
2975 return flush_work(&dwork->work);
2977 EXPORT_SYMBOL(flush_delayed_work);
2980 * cancel_delayed_work - cancel a delayed work
2981 * @dwork: delayed_work to cancel
2983 * Kill off a pending delayed_work.
2985 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2986 * pending.
2988 * Note:
2989 * The work callback function may still be running on return, unless
2990 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2991 * use cancel_delayed_work_sync() to wait on it.
2993 * This function is safe to call from any context including IRQ handler.
2995 bool cancel_delayed_work(struct delayed_work *dwork)
2997 unsigned long flags;
2998 int ret;
3000 do {
3001 ret = try_to_grab_pending(&dwork->work, true, &flags);
3002 } while (unlikely(ret == -EAGAIN));
3004 if (unlikely(ret < 0))
3005 return false;
3007 set_work_pool_and_clear_pending(&dwork->work,
3008 get_work_pool_id(&dwork->work));
3009 local_irq_restore(flags);
3010 return ret;
3012 EXPORT_SYMBOL(cancel_delayed_work);
3015 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3016 * @dwork: the delayed work cancel
3018 * This is cancel_work_sync() for delayed works.
3020 * Return:
3021 * %true if @dwork was pending, %false otherwise.
3023 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3025 return __cancel_work_timer(&dwork->work, true);
3027 EXPORT_SYMBOL(cancel_delayed_work_sync);
3030 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3031 * @func: the function to call
3033 * schedule_on_each_cpu() executes @func on each online CPU using the
3034 * system workqueue and blocks until all CPUs have completed.
3035 * schedule_on_each_cpu() is very slow.
3037 * Return:
3038 * 0 on success, -errno on failure.
3040 int schedule_on_each_cpu(work_func_t func)
3042 int cpu;
3043 struct work_struct __percpu *works;
3045 works = alloc_percpu(struct work_struct);
3046 if (!works)
3047 return -ENOMEM;
3049 get_online_cpus();
3051 for_each_online_cpu(cpu) {
3052 struct work_struct *work = per_cpu_ptr(works, cpu);
3054 INIT_WORK(work, func);
3055 schedule_work_on(cpu, work);
3058 for_each_online_cpu(cpu)
3059 flush_work(per_cpu_ptr(works, cpu));
3061 put_online_cpus();
3062 free_percpu(works);
3063 return 0;
3067 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3069 * Forces execution of the kernel-global workqueue and blocks until its
3070 * completion.
3072 * Think twice before calling this function! It's very easy to get into
3073 * trouble if you don't take great care. Either of the following situations
3074 * will lead to deadlock:
3076 * One of the work items currently on the workqueue needs to acquire
3077 * a lock held by your code or its caller.
3079 * Your code is running in the context of a work routine.
3081 * They will be detected by lockdep when they occur, but the first might not
3082 * occur very often. It depends on what work items are on the workqueue and
3083 * what locks they need, which you have no control over.
3085 * In most situations flushing the entire workqueue is overkill; you merely
3086 * need to know that a particular work item isn't queued and isn't running.
3087 * In such cases you should use cancel_delayed_work_sync() or
3088 * cancel_work_sync() instead.
3090 void flush_scheduled_work(void)
3092 flush_workqueue(system_wq);
3094 EXPORT_SYMBOL(flush_scheduled_work);
3097 * execute_in_process_context - reliably execute the routine with user context
3098 * @fn: the function to execute
3099 * @ew: guaranteed storage for the execute work structure (must
3100 * be available when the work executes)
3102 * Executes the function immediately if process context is available,
3103 * otherwise schedules the function for delayed execution.
3105 * Return: 0 - function was executed
3106 * 1 - function was scheduled for execution
3108 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3110 if (!in_interrupt()) {
3111 fn(&ew->work);
3112 return 0;
3115 INIT_WORK(&ew->work, fn);
3116 schedule_work(&ew->work);
3118 return 1;
3120 EXPORT_SYMBOL_GPL(execute_in_process_context);
3122 #ifdef CONFIG_SYSFS
3124 * Workqueues with WQ_SYSFS flag set is visible to userland via
3125 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3126 * following attributes.
3128 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3129 * max_active RW int : maximum number of in-flight work items
3131 * Unbound workqueues have the following extra attributes.
3133 * id RO int : the associated pool ID
3134 * nice RW int : nice value of the workers
3135 * cpumask RW mask : bitmask of allowed CPUs for the workers
3137 struct wq_device {
3138 struct workqueue_struct *wq;
3139 struct device dev;
3142 static struct workqueue_struct *dev_to_wq(struct device *dev)
3144 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3146 return wq_dev->wq;
3149 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3150 char *buf)
3152 struct workqueue_struct *wq = dev_to_wq(dev);
3154 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3156 static DEVICE_ATTR_RO(per_cpu);
3158 static ssize_t max_active_show(struct device *dev,
3159 struct device_attribute *attr, char *buf)
3161 struct workqueue_struct *wq = dev_to_wq(dev);
3163 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3166 static ssize_t max_active_store(struct device *dev,
3167 struct device_attribute *attr, const char *buf,
3168 size_t count)
3170 struct workqueue_struct *wq = dev_to_wq(dev);
3171 int val;
3173 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3174 return -EINVAL;
3176 workqueue_set_max_active(wq, val);
3177 return count;
3179 static DEVICE_ATTR_RW(max_active);
3181 static struct attribute *wq_sysfs_attrs[] = {
3182 &dev_attr_per_cpu.attr,
3183 &dev_attr_max_active.attr,
3184 NULL,
3186 ATTRIBUTE_GROUPS(wq_sysfs);
3188 static ssize_t wq_pool_ids_show(struct device *dev,
3189 struct device_attribute *attr, char *buf)
3191 struct workqueue_struct *wq = dev_to_wq(dev);
3192 const char *delim = "";
3193 int node, written = 0;
3195 rcu_read_lock_sched();
3196 for_each_node(node) {
3197 written += scnprintf(buf + written, PAGE_SIZE - written,
3198 "%s%d:%d", delim, node,
3199 unbound_pwq_by_node(wq, node)->pool->id);
3200 delim = " ";
3202 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3203 rcu_read_unlock_sched();
3205 return written;
3208 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3209 char *buf)
3211 struct workqueue_struct *wq = dev_to_wq(dev);
3212 int written;
3214 mutex_lock(&wq->mutex);
3215 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3216 mutex_unlock(&wq->mutex);
3218 return written;
3221 /* prepare workqueue_attrs for sysfs store operations */
3222 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3224 struct workqueue_attrs *attrs;
3226 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3227 if (!attrs)
3228 return NULL;
3230 mutex_lock(&wq->mutex);
3231 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3232 mutex_unlock(&wq->mutex);
3233 return attrs;
3236 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3237 const char *buf, size_t count)
3239 struct workqueue_struct *wq = dev_to_wq(dev);
3240 struct workqueue_attrs *attrs;
3241 int ret;
3243 attrs = wq_sysfs_prep_attrs(wq);
3244 if (!attrs)
3245 return -ENOMEM;
3247 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3248 attrs->nice >= -20 && attrs->nice <= 19)
3249 ret = apply_workqueue_attrs(wq, attrs);
3250 else
3251 ret = -EINVAL;
3253 free_workqueue_attrs(attrs);
3254 return ret ?: count;
3257 static ssize_t wq_cpumask_show(struct device *dev,
3258 struct device_attribute *attr, char *buf)
3260 struct workqueue_struct *wq = dev_to_wq(dev);
3261 int written;
3263 mutex_lock(&wq->mutex);
3264 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3265 mutex_unlock(&wq->mutex);
3267 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3268 return written;
3271 static ssize_t wq_cpumask_store(struct device *dev,
3272 struct device_attribute *attr,
3273 const char *buf, size_t count)
3275 struct workqueue_struct *wq = dev_to_wq(dev);
3276 struct workqueue_attrs *attrs;
3277 int ret;
3279 attrs = wq_sysfs_prep_attrs(wq);
3280 if (!attrs)
3281 return -ENOMEM;
3283 ret = cpumask_parse(buf, attrs->cpumask);
3284 if (!ret)
3285 ret = apply_workqueue_attrs(wq, attrs);
3287 free_workqueue_attrs(attrs);
3288 return ret ?: count;
3291 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3292 char *buf)
3294 struct workqueue_struct *wq = dev_to_wq(dev);
3295 int written;
3297 mutex_lock(&wq->mutex);
3298 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3299 !wq->unbound_attrs->no_numa);
3300 mutex_unlock(&wq->mutex);
3302 return written;
3305 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3306 const char *buf, size_t count)
3308 struct workqueue_struct *wq = dev_to_wq(dev);
3309 struct workqueue_attrs *attrs;
3310 int v, ret;
3312 attrs = wq_sysfs_prep_attrs(wq);
3313 if (!attrs)
3314 return -ENOMEM;
3316 ret = -EINVAL;
3317 if (sscanf(buf, "%d", &v) == 1) {
3318 attrs->no_numa = !v;
3319 ret = apply_workqueue_attrs(wq, attrs);
3322 free_workqueue_attrs(attrs);
3323 return ret ?: count;
3326 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3327 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3328 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3329 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3330 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3331 __ATTR_NULL,
3334 static struct bus_type wq_subsys = {
3335 .name = "workqueue",
3336 .dev_groups = wq_sysfs_groups,
3339 static int __init wq_sysfs_init(void)
3341 return subsys_virtual_register(&wq_subsys, NULL);
3343 core_initcall(wq_sysfs_init);
3345 static void wq_device_release(struct device *dev)
3347 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3349 kfree(wq_dev);
3353 * workqueue_sysfs_register - make a workqueue visible in sysfs
3354 * @wq: the workqueue to register
3356 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3357 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3358 * which is the preferred method.
3360 * Workqueue user should use this function directly iff it wants to apply
3361 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3362 * apply_workqueue_attrs() may race against userland updating the
3363 * attributes.
3365 * Return: 0 on success, -errno on failure.
3367 int workqueue_sysfs_register(struct workqueue_struct *wq)
3369 struct wq_device *wq_dev;
3370 int ret;
3373 * Adjusting max_active or creating new pwqs by applyting
3374 * attributes breaks ordering guarantee. Disallow exposing ordered
3375 * workqueues.
3377 if (WARN_ON(wq->flags & __WQ_ORDERED))
3378 return -EINVAL;
3380 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3381 if (!wq_dev)
3382 return -ENOMEM;
3384 wq_dev->wq = wq;
3385 wq_dev->dev.bus = &wq_subsys;
3386 wq_dev->dev.init_name = wq->name;
3387 wq_dev->dev.release = wq_device_release;
3390 * unbound_attrs are created separately. Suppress uevent until
3391 * everything is ready.
3393 dev_set_uevent_suppress(&wq_dev->dev, true);
3395 ret = device_register(&wq_dev->dev);
3396 if (ret) {
3397 kfree(wq_dev);
3398 wq->wq_dev = NULL;
3399 return ret;
3402 if (wq->flags & WQ_UNBOUND) {
3403 struct device_attribute *attr;
3405 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3406 ret = device_create_file(&wq_dev->dev, attr);
3407 if (ret) {
3408 device_unregister(&wq_dev->dev);
3409 wq->wq_dev = NULL;
3410 return ret;
3415 dev_set_uevent_suppress(&wq_dev->dev, false);
3416 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3417 return 0;
3421 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3422 * @wq: the workqueue to unregister
3424 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3426 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3428 struct wq_device *wq_dev = wq->wq_dev;
3430 if (!wq->wq_dev)
3431 return;
3433 wq->wq_dev = NULL;
3434 device_unregister(&wq_dev->dev);
3436 #else /* CONFIG_SYSFS */
3437 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3438 #endif /* CONFIG_SYSFS */
3441 * free_workqueue_attrs - free a workqueue_attrs
3442 * @attrs: workqueue_attrs to free
3444 * Undo alloc_workqueue_attrs().
3446 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3448 if (attrs) {
3449 free_cpumask_var(attrs->cpumask);
3450 kfree(attrs);
3455 * alloc_workqueue_attrs - allocate a workqueue_attrs
3456 * @gfp_mask: allocation mask to use
3458 * Allocate a new workqueue_attrs, initialize with default settings and
3459 * return it.
3461 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3463 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3465 struct workqueue_attrs *attrs;
3467 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3468 if (!attrs)
3469 goto fail;
3470 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3471 goto fail;
3473 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3474 return attrs;
3475 fail:
3476 free_workqueue_attrs(attrs);
3477 return NULL;
3480 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3481 const struct workqueue_attrs *from)
3483 to->nice = from->nice;
3484 cpumask_copy(to->cpumask, from->cpumask);
3486 * Unlike hash and equality test, this function doesn't ignore
3487 * ->no_numa as it is used for both pool and wq attrs. Instead,
3488 * get_unbound_pool() explicitly clears ->no_numa after copying.
3490 to->no_numa = from->no_numa;
3493 /* hash value of the content of @attr */
3494 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3496 u32 hash = 0;
3498 hash = jhash_1word(attrs->nice, hash);
3499 hash = jhash(cpumask_bits(attrs->cpumask),
3500 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3501 return hash;
3504 /* content equality test */
3505 static bool wqattrs_equal(const struct workqueue_attrs *a,
3506 const struct workqueue_attrs *b)
3508 if (a->nice != b->nice)
3509 return false;
3510 if (!cpumask_equal(a->cpumask, b->cpumask))
3511 return false;
3512 return true;
3516 * init_worker_pool - initialize a newly zalloc'd worker_pool
3517 * @pool: worker_pool to initialize
3519 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3521 * Return: 0 on success, -errno on failure. Even on failure, all fields
3522 * inside @pool proper are initialized and put_unbound_pool() can be called
3523 * on @pool safely to release it.
3525 static int init_worker_pool(struct worker_pool *pool)
3527 spin_lock_init(&pool->lock);
3528 pool->id = -1;
3529 pool->cpu = -1;
3530 pool->node = NUMA_NO_NODE;
3531 pool->flags |= POOL_DISASSOCIATED;
3532 INIT_LIST_HEAD(&pool->worklist);
3533 INIT_LIST_HEAD(&pool->idle_list);
3534 hash_init(pool->busy_hash);
3536 init_timer_deferrable(&pool->idle_timer);
3537 pool->idle_timer.function = idle_worker_timeout;
3538 pool->idle_timer.data = (unsigned long)pool;
3540 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3541 (unsigned long)pool);
3543 mutex_init(&pool->manager_arb);
3544 mutex_init(&pool->manager_mutex);
3545 idr_init(&pool->worker_idr);
3547 INIT_HLIST_NODE(&pool->hash_node);
3548 pool->refcnt = 1;
3550 /* shouldn't fail above this point */
3551 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3552 if (!pool->attrs)
3553 return -ENOMEM;
3554 return 0;
3557 static void rcu_free_pool(struct rcu_head *rcu)
3559 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3561 idr_destroy(&pool->worker_idr);
3562 free_workqueue_attrs(pool->attrs);
3563 kfree(pool);
3567 * put_unbound_pool - put a worker_pool
3568 * @pool: worker_pool to put
3570 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3571 * safe manner. get_unbound_pool() calls this function on its failure path
3572 * and this function should be able to release pools which went through,
3573 * successfully or not, init_worker_pool().
3575 * Should be called with wq_pool_mutex held.
3577 static void put_unbound_pool(struct worker_pool *pool)
3579 struct worker *worker;
3581 lockdep_assert_held(&wq_pool_mutex);
3583 if (--pool->refcnt)
3584 return;
3586 /* sanity checks */
3587 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3588 WARN_ON(!list_empty(&pool->worklist)))
3589 return;
3591 /* release id and unhash */
3592 if (pool->id >= 0)
3593 idr_remove(&worker_pool_idr, pool->id);
3594 hash_del(&pool->hash_node);
3597 * Become the manager and destroy all workers. Grabbing
3598 * manager_arb prevents @pool's workers from blocking on
3599 * manager_mutex.
3601 mutex_lock(&pool->manager_arb);
3602 mutex_lock(&pool->manager_mutex);
3603 spin_lock_irq(&pool->lock);
3605 while ((worker = first_worker(pool)))
3606 destroy_worker(worker);
3607 WARN_ON(pool->nr_workers || pool->nr_idle);
3609 spin_unlock_irq(&pool->lock);
3610 mutex_unlock(&pool->manager_mutex);
3611 mutex_unlock(&pool->manager_arb);
3613 /* shut down the timers */
3614 del_timer_sync(&pool->idle_timer);
3615 del_timer_sync(&pool->mayday_timer);
3617 /* sched-RCU protected to allow dereferences from get_work_pool() */
3618 call_rcu_sched(&pool->rcu, rcu_free_pool);
3622 * get_unbound_pool - get a worker_pool with the specified attributes
3623 * @attrs: the attributes of the worker_pool to get
3625 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3626 * reference count and return it. If there already is a matching
3627 * worker_pool, it will be used; otherwise, this function attempts to
3628 * create a new one.
3630 * Should be called with wq_pool_mutex held.
3632 * Return: On success, a worker_pool with the same attributes as @attrs.
3633 * On failure, %NULL.
3635 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3637 u32 hash = wqattrs_hash(attrs);
3638 struct worker_pool *pool;
3639 int node;
3641 lockdep_assert_held(&wq_pool_mutex);
3643 /* do we already have a matching pool? */
3644 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3645 if (wqattrs_equal(pool->attrs, attrs)) {
3646 pool->refcnt++;
3647 goto out_unlock;
3651 /* nope, create a new one */
3652 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3653 if (!pool || init_worker_pool(pool) < 0)
3654 goto fail;
3656 if (workqueue_freezing)
3657 pool->flags |= POOL_FREEZING;
3659 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3660 copy_workqueue_attrs(pool->attrs, attrs);
3663 * no_numa isn't a worker_pool attribute, always clear it. See
3664 * 'struct workqueue_attrs' comments for detail.
3666 pool->attrs->no_numa = false;
3668 /* if cpumask is contained inside a NUMA node, we belong to that node */
3669 if (wq_numa_enabled) {
3670 for_each_node(node) {
3671 if (cpumask_subset(pool->attrs->cpumask,
3672 wq_numa_possible_cpumask[node])) {
3673 pool->node = node;
3674 break;
3679 if (worker_pool_assign_id(pool) < 0)
3680 goto fail;
3682 /* create and start the initial worker */
3683 if (create_and_start_worker(pool) < 0)
3684 goto fail;
3686 /* install */
3687 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3688 out_unlock:
3689 return pool;
3690 fail:
3691 if (pool)
3692 put_unbound_pool(pool);
3693 return NULL;
3696 static void rcu_free_pwq(struct rcu_head *rcu)
3698 kmem_cache_free(pwq_cache,
3699 container_of(rcu, struct pool_workqueue, rcu));
3703 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3704 * and needs to be destroyed.
3706 static void pwq_unbound_release_workfn(struct work_struct *work)
3708 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3709 unbound_release_work);
3710 struct workqueue_struct *wq = pwq->wq;
3711 struct worker_pool *pool = pwq->pool;
3712 bool is_last;
3714 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3715 return;
3718 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3719 * necessary on release but do it anyway. It's easier to verify
3720 * and consistent with the linking path.
3722 mutex_lock(&wq->mutex);
3723 list_del_rcu(&pwq->pwqs_node);
3724 is_last = list_empty(&wq->pwqs);
3725 mutex_unlock(&wq->mutex);
3727 mutex_lock(&wq_pool_mutex);
3728 put_unbound_pool(pool);
3729 mutex_unlock(&wq_pool_mutex);
3731 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3734 * If we're the last pwq going away, @wq is already dead and no one
3735 * is gonna access it anymore. Free it.
3737 if (is_last) {
3738 free_workqueue_attrs(wq->unbound_attrs);
3739 kfree(wq);
3744 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3745 * @pwq: target pool_workqueue
3747 * If @pwq isn't freezing, set @pwq->max_active to the associated
3748 * workqueue's saved_max_active and activate delayed work items
3749 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3751 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3753 struct workqueue_struct *wq = pwq->wq;
3754 bool freezable = wq->flags & WQ_FREEZABLE;
3756 /* for @wq->saved_max_active */
3757 lockdep_assert_held(&wq->mutex);
3759 /* fast exit for non-freezable wqs */
3760 if (!freezable && pwq->max_active == wq->saved_max_active)
3761 return;
3763 spin_lock_irq(&pwq->pool->lock);
3765 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3766 pwq->max_active = wq->saved_max_active;
3768 while (!list_empty(&pwq->delayed_works) &&
3769 pwq->nr_active < pwq->max_active)
3770 pwq_activate_first_delayed(pwq);
3773 * Need to kick a worker after thawed or an unbound wq's
3774 * max_active is bumped. It's a slow path. Do it always.
3776 wake_up_worker(pwq->pool);
3777 } else {
3778 pwq->max_active = 0;
3781 spin_unlock_irq(&pwq->pool->lock);
3784 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3785 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3786 struct worker_pool *pool)
3788 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3790 memset(pwq, 0, sizeof(*pwq));
3792 pwq->pool = pool;
3793 pwq->wq = wq;
3794 pwq->flush_color = -1;
3795 pwq->refcnt = 1;
3796 INIT_LIST_HEAD(&pwq->delayed_works);
3797 INIT_LIST_HEAD(&pwq->pwqs_node);
3798 INIT_LIST_HEAD(&pwq->mayday_node);
3799 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3802 /* sync @pwq with the current state of its associated wq and link it */
3803 static void link_pwq(struct pool_workqueue *pwq)
3805 struct workqueue_struct *wq = pwq->wq;
3807 lockdep_assert_held(&wq->mutex);
3809 /* may be called multiple times, ignore if already linked */
3810 if (!list_empty(&pwq->pwqs_node))
3811 return;
3814 * Set the matching work_color. This is synchronized with
3815 * wq->mutex to avoid confusing flush_workqueue().
3817 pwq->work_color = wq->work_color;
3819 /* sync max_active to the current setting */
3820 pwq_adjust_max_active(pwq);
3822 /* link in @pwq */
3823 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3826 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3827 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3828 const struct workqueue_attrs *attrs)
3830 struct worker_pool *pool;
3831 struct pool_workqueue *pwq;
3833 lockdep_assert_held(&wq_pool_mutex);
3835 pool = get_unbound_pool(attrs);
3836 if (!pool)
3837 return NULL;
3839 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3840 if (!pwq) {
3841 put_unbound_pool(pool);
3842 return NULL;
3845 init_pwq(pwq, wq, pool);
3846 return pwq;
3849 /* undo alloc_unbound_pwq(), used only in the error path */
3850 static void free_unbound_pwq(struct pool_workqueue *pwq)
3852 lockdep_assert_held(&wq_pool_mutex);
3854 if (pwq) {
3855 put_unbound_pool(pwq->pool);
3856 kmem_cache_free(pwq_cache, pwq);
3861 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3862 * @attrs: the wq_attrs of interest
3863 * @node: the target NUMA node
3864 * @cpu_going_down: if >= 0, the CPU to consider as offline
3865 * @cpumask: outarg, the resulting cpumask
3867 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3868 * @cpu_going_down is >= 0, that cpu is considered offline during
3869 * calculation. The result is stored in @cpumask.
3871 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3872 * enabled and @node has online CPUs requested by @attrs, the returned
3873 * cpumask is the intersection of the possible CPUs of @node and
3874 * @attrs->cpumask.
3876 * The caller is responsible for ensuring that the cpumask of @node stays
3877 * stable.
3879 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3880 * %false if equal.
3882 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3883 int cpu_going_down, cpumask_t *cpumask)
3885 if (!wq_numa_enabled || attrs->no_numa)
3886 goto use_dfl;
3888 /* does @node have any online CPUs @attrs wants? */
3889 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3890 if (cpu_going_down >= 0)
3891 cpumask_clear_cpu(cpu_going_down, cpumask);
3893 if (cpumask_empty(cpumask))
3894 goto use_dfl;
3896 /* yeap, return possible CPUs in @node that @attrs wants */
3897 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3898 return !cpumask_equal(cpumask, attrs->cpumask);
3900 use_dfl:
3901 cpumask_copy(cpumask, attrs->cpumask);
3902 return false;
3905 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3906 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3907 int node,
3908 struct pool_workqueue *pwq)
3910 struct pool_workqueue *old_pwq;
3912 lockdep_assert_held(&wq->mutex);
3914 /* link_pwq() can handle duplicate calls */
3915 link_pwq(pwq);
3917 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3918 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3919 return old_pwq;
3923 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3924 * @wq: the target workqueue
3925 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3927 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3928 * machines, this function maps a separate pwq to each NUMA node with
3929 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3930 * NUMA node it was issued on. Older pwqs are released as in-flight work
3931 * items finish. Note that a work item which repeatedly requeues itself
3932 * back-to-back will stay on its current pwq.
3934 * Performs GFP_KERNEL allocations.
3936 * Return: 0 on success and -errno on failure.
3938 int apply_workqueue_attrs(struct workqueue_struct *wq,
3939 const struct workqueue_attrs *attrs)
3941 struct workqueue_attrs *new_attrs, *tmp_attrs;
3942 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3943 int node, ret;
3945 /* only unbound workqueues can change attributes */
3946 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3947 return -EINVAL;
3949 /* creating multiple pwqs breaks ordering guarantee */
3950 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3951 return -EINVAL;
3953 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3954 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3955 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3956 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3957 goto enomem;
3959 /* make a copy of @attrs and sanitize it */
3960 copy_workqueue_attrs(new_attrs, attrs);
3961 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3964 * We may create multiple pwqs with differing cpumasks. Make a
3965 * copy of @new_attrs which will be modified and used to obtain
3966 * pools.
3968 copy_workqueue_attrs(tmp_attrs, new_attrs);
3971 * CPUs should stay stable across pwq creations and installations.
3972 * Pin CPUs, determine the target cpumask for each node and create
3973 * pwqs accordingly.
3975 get_online_cpus();
3977 mutex_lock(&wq_pool_mutex);
3980 * If something goes wrong during CPU up/down, we'll fall back to
3981 * the default pwq covering whole @attrs->cpumask. Always create
3982 * it even if we don't use it immediately.
3984 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3985 if (!dfl_pwq)
3986 goto enomem_pwq;
3988 for_each_node(node) {
3989 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3990 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3991 if (!pwq_tbl[node])
3992 goto enomem_pwq;
3993 } else {
3994 dfl_pwq->refcnt++;
3995 pwq_tbl[node] = dfl_pwq;
3999 mutex_unlock(&wq_pool_mutex);
4001 /* all pwqs have been created successfully, let's install'em */
4002 mutex_lock(&wq->mutex);
4004 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
4006 /* save the previous pwq and install the new one */
4007 for_each_node(node)
4008 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
4010 /* @dfl_pwq might not have been used, ensure it's linked */
4011 link_pwq(dfl_pwq);
4012 swap(wq->dfl_pwq, dfl_pwq);
4014 mutex_unlock(&wq->mutex);
4016 /* put the old pwqs */
4017 for_each_node(node)
4018 put_pwq_unlocked(pwq_tbl[node]);
4019 put_pwq_unlocked(dfl_pwq);
4021 put_online_cpus();
4022 ret = 0;
4023 /* fall through */
4024 out_free:
4025 free_workqueue_attrs(tmp_attrs);
4026 free_workqueue_attrs(new_attrs);
4027 kfree(pwq_tbl);
4028 return ret;
4030 enomem_pwq:
4031 free_unbound_pwq(dfl_pwq);
4032 for_each_node(node)
4033 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
4034 free_unbound_pwq(pwq_tbl[node]);
4035 mutex_unlock(&wq_pool_mutex);
4036 put_online_cpus();
4037 enomem:
4038 ret = -ENOMEM;
4039 goto out_free;
4043 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4044 * @wq: the target workqueue
4045 * @cpu: the CPU coming up or going down
4046 * @online: whether @cpu is coming up or going down
4048 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4049 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4050 * @wq accordingly.
4052 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4053 * falls back to @wq->dfl_pwq which may not be optimal but is always
4054 * correct.
4056 * Note that when the last allowed CPU of a NUMA node goes offline for a
4057 * workqueue with a cpumask spanning multiple nodes, the workers which were
4058 * already executing the work items for the workqueue will lose their CPU
4059 * affinity and may execute on any CPU. This is similar to how per-cpu
4060 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4061 * affinity, it's the user's responsibility to flush the work item from
4062 * CPU_DOWN_PREPARE.
4064 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4065 bool online)
4067 int node = cpu_to_node(cpu);
4068 int cpu_off = online ? -1 : cpu;
4069 struct pool_workqueue *old_pwq = NULL, *pwq;
4070 struct workqueue_attrs *target_attrs;
4071 cpumask_t *cpumask;
4073 lockdep_assert_held(&wq_pool_mutex);
4075 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4076 return;
4079 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4080 * Let's use a preallocated one. The following buf is protected by
4081 * CPU hotplug exclusion.
4083 target_attrs = wq_update_unbound_numa_attrs_buf;
4084 cpumask = target_attrs->cpumask;
4086 mutex_lock(&wq->mutex);
4087 if (wq->unbound_attrs->no_numa)
4088 goto out_unlock;
4090 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4091 pwq = unbound_pwq_by_node(wq, node);
4094 * Let's determine what needs to be done. If the target cpumask is
4095 * different from wq's, we need to compare it to @pwq's and create
4096 * a new one if they don't match. If the target cpumask equals
4097 * wq's, the default pwq should be used. If @pwq is already the
4098 * default one, nothing to do; otherwise, install the default one.
4100 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4101 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4102 goto out_unlock;
4103 } else {
4104 if (pwq == wq->dfl_pwq)
4105 goto out_unlock;
4106 else
4107 goto use_dfl_pwq;
4110 mutex_unlock(&wq->mutex);
4112 /* create a new pwq */
4113 pwq = alloc_unbound_pwq(wq, target_attrs);
4114 if (!pwq) {
4115 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4116 wq->name);
4117 mutex_lock(&wq->mutex);
4118 goto use_dfl_pwq;
4122 * Install the new pwq. As this function is called only from CPU
4123 * hotplug callbacks and applying a new attrs is wrapped with
4124 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4125 * inbetween.
4127 mutex_lock(&wq->mutex);
4128 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4129 goto out_unlock;
4131 use_dfl_pwq:
4132 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4133 get_pwq(wq->dfl_pwq);
4134 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4135 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4136 out_unlock:
4137 mutex_unlock(&wq->mutex);
4138 put_pwq_unlocked(old_pwq);
4141 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4143 bool highpri = wq->flags & WQ_HIGHPRI;
4144 int cpu, ret;
4146 if (!(wq->flags & WQ_UNBOUND)) {
4147 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4148 if (!wq->cpu_pwqs)
4149 return -ENOMEM;
4151 for_each_possible_cpu(cpu) {
4152 struct pool_workqueue *pwq =
4153 per_cpu_ptr(wq->cpu_pwqs, cpu);
4154 struct worker_pool *cpu_pools =
4155 per_cpu(cpu_worker_pools, cpu);
4157 init_pwq(pwq, wq, &cpu_pools[highpri]);
4159 mutex_lock(&wq->mutex);
4160 link_pwq(pwq);
4161 mutex_unlock(&wq->mutex);
4163 return 0;
4164 } else if (wq->flags & __WQ_ORDERED) {
4165 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4166 /* there should only be single pwq for ordering guarantee */
4167 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4168 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4169 "ordering guarantee broken for workqueue %s\n", wq->name);
4170 return ret;
4171 } else {
4172 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4176 static int wq_clamp_max_active(int max_active, unsigned int flags,
4177 const char *name)
4179 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4181 if (max_active < 1 || max_active > lim)
4182 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4183 max_active, name, 1, lim);
4185 return clamp_val(max_active, 1, lim);
4188 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4189 unsigned int flags,
4190 int max_active,
4191 struct lock_class_key *key,
4192 const char *lock_name, ...)
4194 size_t tbl_size = 0;
4195 va_list args;
4196 struct workqueue_struct *wq;
4197 struct pool_workqueue *pwq;
4199 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4200 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4201 flags |= WQ_UNBOUND;
4203 /* allocate wq and format name */
4204 if (flags & WQ_UNBOUND)
4205 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4207 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4208 if (!wq)
4209 return NULL;
4211 if (flags & WQ_UNBOUND) {
4212 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4213 if (!wq->unbound_attrs)
4214 goto err_free_wq;
4217 va_start(args, lock_name);
4218 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4219 va_end(args);
4221 max_active = max_active ?: WQ_DFL_ACTIVE;
4222 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4224 /* init wq */
4225 wq->flags = flags;
4226 wq->saved_max_active = max_active;
4227 mutex_init(&wq->mutex);
4228 atomic_set(&wq->nr_pwqs_to_flush, 0);
4229 INIT_LIST_HEAD(&wq->pwqs);
4230 INIT_LIST_HEAD(&wq->flusher_queue);
4231 INIT_LIST_HEAD(&wq->flusher_overflow);
4232 INIT_LIST_HEAD(&wq->maydays);
4234 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4235 INIT_LIST_HEAD(&wq->list);
4237 if (alloc_and_link_pwqs(wq) < 0)
4238 goto err_free_wq;
4241 * Workqueues which may be used during memory reclaim should
4242 * have a rescuer to guarantee forward progress.
4244 if (flags & WQ_MEM_RECLAIM) {
4245 struct worker *rescuer;
4247 rescuer = alloc_worker();
4248 if (!rescuer)
4249 goto err_destroy;
4251 rescuer->rescue_wq = wq;
4252 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4253 wq->name);
4254 if (IS_ERR(rescuer->task)) {
4255 kfree(rescuer);
4256 goto err_destroy;
4259 wq->rescuer = rescuer;
4260 rescuer->task->flags |= PF_NO_SETAFFINITY;
4261 wake_up_process(rescuer->task);
4264 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4265 goto err_destroy;
4268 * wq_pool_mutex protects global freeze state and workqueues list.
4269 * Grab it, adjust max_active and add the new @wq to workqueues
4270 * list.
4272 mutex_lock(&wq_pool_mutex);
4274 mutex_lock(&wq->mutex);
4275 for_each_pwq(pwq, wq)
4276 pwq_adjust_max_active(pwq);
4277 mutex_unlock(&wq->mutex);
4279 list_add(&wq->list, &workqueues);
4281 mutex_unlock(&wq_pool_mutex);
4283 return wq;
4285 err_free_wq:
4286 free_workqueue_attrs(wq->unbound_attrs);
4287 kfree(wq);
4288 return NULL;
4289 err_destroy:
4290 destroy_workqueue(wq);
4291 return NULL;
4293 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4296 * destroy_workqueue - safely terminate a workqueue
4297 * @wq: target workqueue
4299 * Safely destroy a workqueue. All work currently pending will be done first.
4301 void destroy_workqueue(struct workqueue_struct *wq)
4303 struct pool_workqueue *pwq;
4304 int node;
4306 /* drain it before proceeding with destruction */
4307 drain_workqueue(wq);
4309 /* sanity checks */
4310 mutex_lock(&wq->mutex);
4311 for_each_pwq(pwq, wq) {
4312 int i;
4314 for (i = 0; i < WORK_NR_COLORS; i++) {
4315 if (WARN_ON(pwq->nr_in_flight[i])) {
4316 mutex_unlock(&wq->mutex);
4317 return;
4321 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4322 WARN_ON(pwq->nr_active) ||
4323 WARN_ON(!list_empty(&pwq->delayed_works))) {
4324 mutex_unlock(&wq->mutex);
4325 return;
4328 mutex_unlock(&wq->mutex);
4331 * wq list is used to freeze wq, remove from list after
4332 * flushing is complete in case freeze races us.
4334 mutex_lock(&wq_pool_mutex);
4335 list_del_init(&wq->list);
4336 mutex_unlock(&wq_pool_mutex);
4338 workqueue_sysfs_unregister(wq);
4340 if (wq->rescuer) {
4341 kthread_stop(wq->rescuer->task);
4342 kfree(wq->rescuer);
4343 wq->rescuer = NULL;
4346 if (!(wq->flags & WQ_UNBOUND)) {
4348 * The base ref is never dropped on per-cpu pwqs. Directly
4349 * free the pwqs and wq.
4351 free_percpu(wq->cpu_pwqs);
4352 kfree(wq);
4353 } else {
4355 * We're the sole accessor of @wq at this point. Directly
4356 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4357 * @wq will be freed when the last pwq is released.
4359 for_each_node(node) {
4360 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4361 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4362 put_pwq_unlocked(pwq);
4366 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4367 * put. Don't access it afterwards.
4369 pwq = wq->dfl_pwq;
4370 wq->dfl_pwq = NULL;
4371 put_pwq_unlocked(pwq);
4374 EXPORT_SYMBOL_GPL(destroy_workqueue);
4377 * workqueue_set_max_active - adjust max_active of a workqueue
4378 * @wq: target workqueue
4379 * @max_active: new max_active value.
4381 * Set max_active of @wq to @max_active.
4383 * CONTEXT:
4384 * Don't call from IRQ context.
4386 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4388 struct pool_workqueue *pwq;
4390 /* disallow meddling with max_active for ordered workqueues */
4391 if (WARN_ON(wq->flags & __WQ_ORDERED))
4392 return;
4394 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4396 mutex_lock(&wq->mutex);
4398 wq->saved_max_active = max_active;
4400 for_each_pwq(pwq, wq)
4401 pwq_adjust_max_active(pwq);
4403 mutex_unlock(&wq->mutex);
4405 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4408 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4410 * Determine whether %current is a workqueue rescuer. Can be used from
4411 * work functions to determine whether it's being run off the rescuer task.
4413 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4415 bool current_is_workqueue_rescuer(void)
4417 struct worker *worker = current_wq_worker();
4419 return worker && worker->rescue_wq;
4423 * workqueue_congested - test whether a workqueue is congested
4424 * @cpu: CPU in question
4425 * @wq: target workqueue
4427 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4428 * no synchronization around this function and the test result is
4429 * unreliable and only useful as advisory hints or for debugging.
4431 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4432 * Note that both per-cpu and unbound workqueues may be associated with
4433 * multiple pool_workqueues which have separate congested states. A
4434 * workqueue being congested on one CPU doesn't mean the workqueue is also
4435 * contested on other CPUs / NUMA nodes.
4437 * Return:
4438 * %true if congested, %false otherwise.
4440 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4442 struct pool_workqueue *pwq;
4443 bool ret;
4445 rcu_read_lock_sched();
4447 if (cpu == WORK_CPU_UNBOUND)
4448 cpu = smp_processor_id();
4450 if (!(wq->flags & WQ_UNBOUND))
4451 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4452 else
4453 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4455 ret = !list_empty(&pwq->delayed_works);
4456 rcu_read_unlock_sched();
4458 return ret;
4460 EXPORT_SYMBOL_GPL(workqueue_congested);
4463 * work_busy - test whether a work is currently pending or running
4464 * @work: the work to be tested
4466 * Test whether @work is currently pending or running. There is no
4467 * synchronization around this function and the test result is
4468 * unreliable and only useful as advisory hints or for debugging.
4470 * Return:
4471 * OR'd bitmask of WORK_BUSY_* bits.
4473 unsigned int work_busy(struct work_struct *work)
4475 struct worker_pool *pool;
4476 unsigned long flags;
4477 unsigned int ret = 0;
4479 if (work_pending(work))
4480 ret |= WORK_BUSY_PENDING;
4482 local_irq_save(flags);
4483 pool = get_work_pool(work);
4484 if (pool) {
4485 spin_lock(&pool->lock);
4486 if (find_worker_executing_work(pool, work))
4487 ret |= WORK_BUSY_RUNNING;
4488 spin_unlock(&pool->lock);
4490 local_irq_restore(flags);
4492 return ret;
4494 EXPORT_SYMBOL_GPL(work_busy);
4497 * set_worker_desc - set description for the current work item
4498 * @fmt: printf-style format string
4499 * @...: arguments for the format string
4501 * This function can be called by a running work function to describe what
4502 * the work item is about. If the worker task gets dumped, this
4503 * information will be printed out together to help debugging. The
4504 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4506 void set_worker_desc(const char *fmt, ...)
4508 struct worker *worker = current_wq_worker();
4509 va_list args;
4511 if (worker) {
4512 va_start(args, fmt);
4513 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4514 va_end(args);
4515 worker->desc_valid = true;
4520 * print_worker_info - print out worker information and description
4521 * @log_lvl: the log level to use when printing
4522 * @task: target task
4524 * If @task is a worker and currently executing a work item, print out the
4525 * name of the workqueue being serviced and worker description set with
4526 * set_worker_desc() by the currently executing work item.
4528 * This function can be safely called on any task as long as the
4529 * task_struct itself is accessible. While safe, this function isn't
4530 * synchronized and may print out mixups or garbages of limited length.
4532 void print_worker_info(const char *log_lvl, struct task_struct *task)
4534 work_func_t *fn = NULL;
4535 char name[WQ_NAME_LEN] = { };
4536 char desc[WORKER_DESC_LEN] = { };
4537 struct pool_workqueue *pwq = NULL;
4538 struct workqueue_struct *wq = NULL;
4539 bool desc_valid = false;
4540 struct worker *worker;
4542 if (!(task->flags & PF_WQ_WORKER))
4543 return;
4546 * This function is called without any synchronization and @task
4547 * could be in any state. Be careful with dereferences.
4549 worker = probe_kthread_data(task);
4552 * Carefully copy the associated workqueue's workfn and name. Keep
4553 * the original last '\0' in case the original contains garbage.
4555 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4556 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4557 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4558 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4560 /* copy worker description */
4561 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4562 if (desc_valid)
4563 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4565 if (fn || name[0] || desc[0]) {
4566 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4567 if (desc[0])
4568 pr_cont(" (%s)", desc);
4569 pr_cont("\n");
4574 * CPU hotplug.
4576 * There are two challenges in supporting CPU hotplug. Firstly, there
4577 * are a lot of assumptions on strong associations among work, pwq and
4578 * pool which make migrating pending and scheduled works very
4579 * difficult to implement without impacting hot paths. Secondly,
4580 * worker pools serve mix of short, long and very long running works making
4581 * blocked draining impractical.
4583 * This is solved by allowing the pools to be disassociated from the CPU
4584 * running as an unbound one and allowing it to be reattached later if the
4585 * cpu comes back online.
4588 static void wq_unbind_fn(struct work_struct *work)
4590 int cpu = smp_processor_id();
4591 struct worker_pool *pool;
4592 struct worker *worker;
4593 int wi;
4595 for_each_cpu_worker_pool(pool, cpu) {
4596 WARN_ON_ONCE(cpu != smp_processor_id());
4598 mutex_lock(&pool->manager_mutex);
4599 spin_lock_irq(&pool->lock);
4602 * We've blocked all manager operations. Make all workers
4603 * unbound and set DISASSOCIATED. Before this, all workers
4604 * except for the ones which are still executing works from
4605 * before the last CPU down must be on the cpu. After
4606 * this, they may become diasporas.
4608 for_each_pool_worker(worker, wi, pool)
4609 worker->flags |= WORKER_UNBOUND;
4611 pool->flags |= POOL_DISASSOCIATED;
4613 spin_unlock_irq(&pool->lock);
4614 mutex_unlock(&pool->manager_mutex);
4617 * Call schedule() so that we cross rq->lock and thus can
4618 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4619 * This is necessary as scheduler callbacks may be invoked
4620 * from other cpus.
4622 schedule();
4625 * Sched callbacks are disabled now. Zap nr_running.
4626 * After this, nr_running stays zero and need_more_worker()
4627 * and keep_working() are always true as long as the
4628 * worklist is not empty. This pool now behaves as an
4629 * unbound (in terms of concurrency management) pool which
4630 * are served by workers tied to the pool.
4632 atomic_set(&pool->nr_running, 0);
4635 * With concurrency management just turned off, a busy
4636 * worker blocking could lead to lengthy stalls. Kick off
4637 * unbound chain execution of currently pending work items.
4639 spin_lock_irq(&pool->lock);
4640 wake_up_worker(pool);
4641 spin_unlock_irq(&pool->lock);
4646 * rebind_workers - rebind all workers of a pool to the associated CPU
4647 * @pool: pool of interest
4649 * @pool->cpu is coming online. Rebind all workers to the CPU.
4651 static void rebind_workers(struct worker_pool *pool)
4653 struct worker *worker;
4654 int wi;
4656 lockdep_assert_held(&pool->manager_mutex);
4659 * Restore CPU affinity of all workers. As all idle workers should
4660 * be on the run-queue of the associated CPU before any local
4661 * wake-ups for concurrency management happen, restore CPU affinty
4662 * of all workers first and then clear UNBOUND. As we're called
4663 * from CPU_ONLINE, the following shouldn't fail.
4665 for_each_pool_worker(worker, wi, pool)
4666 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4667 pool->attrs->cpumask) < 0);
4669 spin_lock_irq(&pool->lock);
4671 for_each_pool_worker(worker, wi, pool) {
4672 unsigned int worker_flags = worker->flags;
4675 * A bound idle worker should actually be on the runqueue
4676 * of the associated CPU for local wake-ups targeting it to
4677 * work. Kick all idle workers so that they migrate to the
4678 * associated CPU. Doing this in the same loop as
4679 * replacing UNBOUND with REBOUND is safe as no worker will
4680 * be bound before @pool->lock is released.
4682 if (worker_flags & WORKER_IDLE)
4683 wake_up_process(worker->task);
4686 * We want to clear UNBOUND but can't directly call
4687 * worker_clr_flags() or adjust nr_running. Atomically
4688 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4689 * @worker will clear REBOUND using worker_clr_flags() when
4690 * it initiates the next execution cycle thus restoring
4691 * concurrency management. Note that when or whether
4692 * @worker clears REBOUND doesn't affect correctness.
4694 * ACCESS_ONCE() is necessary because @worker->flags may be
4695 * tested without holding any lock in
4696 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4697 * fail incorrectly leading to premature concurrency
4698 * management operations.
4700 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4701 worker_flags |= WORKER_REBOUND;
4702 worker_flags &= ~WORKER_UNBOUND;
4703 ACCESS_ONCE(worker->flags) = worker_flags;
4706 spin_unlock_irq(&pool->lock);
4710 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4711 * @pool: unbound pool of interest
4712 * @cpu: the CPU which is coming up
4714 * An unbound pool may end up with a cpumask which doesn't have any online
4715 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4716 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4717 * online CPU before, cpus_allowed of all its workers should be restored.
4719 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4721 static cpumask_t cpumask;
4722 struct worker *worker;
4723 int wi;
4725 lockdep_assert_held(&pool->manager_mutex);
4727 /* is @cpu allowed for @pool? */
4728 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4729 return;
4731 /* is @cpu the only online CPU? */
4732 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4733 if (cpumask_weight(&cpumask) != 1)
4734 return;
4736 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4737 for_each_pool_worker(worker, wi, pool)
4738 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4739 pool->attrs->cpumask) < 0);
4743 * Workqueues should be brought up before normal priority CPU notifiers.
4744 * This will be registered high priority CPU notifier.
4746 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4747 unsigned long action,
4748 void *hcpu)
4750 int cpu = (unsigned long)hcpu;
4751 struct worker_pool *pool;
4752 struct workqueue_struct *wq;
4753 int pi;
4755 switch (action & ~CPU_TASKS_FROZEN) {
4756 case CPU_UP_PREPARE:
4757 for_each_cpu_worker_pool(pool, cpu) {
4758 if (pool->nr_workers)
4759 continue;
4760 if (create_and_start_worker(pool) < 0)
4761 return NOTIFY_BAD;
4763 break;
4765 case CPU_DOWN_FAILED:
4766 case CPU_ONLINE:
4767 mutex_lock(&wq_pool_mutex);
4769 for_each_pool(pool, pi) {
4770 mutex_lock(&pool->manager_mutex);
4772 if (pool->cpu == cpu) {
4773 spin_lock_irq(&pool->lock);
4774 pool->flags &= ~POOL_DISASSOCIATED;
4775 spin_unlock_irq(&pool->lock);
4777 rebind_workers(pool);
4778 } else if (pool->cpu < 0) {
4779 restore_unbound_workers_cpumask(pool, cpu);
4782 mutex_unlock(&pool->manager_mutex);
4785 /* update NUMA affinity of unbound workqueues */
4786 list_for_each_entry(wq, &workqueues, list)
4787 wq_update_unbound_numa(wq, cpu, true);
4789 mutex_unlock(&wq_pool_mutex);
4790 break;
4792 return NOTIFY_OK;
4796 * Workqueues should be brought down after normal priority CPU notifiers.
4797 * This will be registered as low priority CPU notifier.
4799 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4800 unsigned long action,
4801 void *hcpu)
4803 int cpu = (unsigned long)hcpu;
4804 struct work_struct unbind_work;
4805 struct workqueue_struct *wq;
4807 switch (action & ~CPU_TASKS_FROZEN) {
4808 case CPU_DOWN_PREPARE:
4809 /* unbinding per-cpu workers should happen on the local CPU */
4810 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4811 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4813 /* update NUMA affinity of unbound workqueues */
4814 mutex_lock(&wq_pool_mutex);
4815 list_for_each_entry(wq, &workqueues, list)
4816 wq_update_unbound_numa(wq, cpu, false);
4817 mutex_unlock(&wq_pool_mutex);
4819 /* wait for per-cpu unbinding to finish */
4820 flush_work(&unbind_work);
4821 break;
4823 return NOTIFY_OK;
4826 #ifdef CONFIG_SMP
4828 struct work_for_cpu {
4829 struct work_struct work;
4830 long (*fn)(void *);
4831 void *arg;
4832 long ret;
4835 static void work_for_cpu_fn(struct work_struct *work)
4837 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4839 wfc->ret = wfc->fn(wfc->arg);
4843 * work_on_cpu - run a function in user context on a particular cpu
4844 * @cpu: the cpu to run on
4845 * @fn: the function to run
4846 * @arg: the function arg
4848 * It is up to the caller to ensure that the cpu doesn't go offline.
4849 * The caller must not hold any locks which would prevent @fn from completing.
4851 * Return: The value @fn returns.
4853 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4855 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4857 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4858 schedule_work_on(cpu, &wfc.work);
4861 * The work item is on-stack and can't lead to deadlock through
4862 * flushing. Use __flush_work() to avoid spurious lockdep warnings
4863 * when work_on_cpu()s are nested.
4865 __flush_work(&wfc.work);
4867 return wfc.ret;
4869 EXPORT_SYMBOL_GPL(work_on_cpu);
4870 #endif /* CONFIG_SMP */
4872 #ifdef CONFIG_FREEZER
4875 * freeze_workqueues_begin - begin freezing workqueues
4877 * Start freezing workqueues. After this function returns, all freezable
4878 * workqueues will queue new works to their delayed_works list instead of
4879 * pool->worklist.
4881 * CONTEXT:
4882 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4884 void freeze_workqueues_begin(void)
4886 struct worker_pool *pool;
4887 struct workqueue_struct *wq;
4888 struct pool_workqueue *pwq;
4889 int pi;
4891 mutex_lock(&wq_pool_mutex);
4893 WARN_ON_ONCE(workqueue_freezing);
4894 workqueue_freezing = true;
4896 /* set FREEZING */
4897 for_each_pool(pool, pi) {
4898 spin_lock_irq(&pool->lock);
4899 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4900 pool->flags |= POOL_FREEZING;
4901 spin_unlock_irq(&pool->lock);
4904 list_for_each_entry(wq, &workqueues, list) {
4905 mutex_lock(&wq->mutex);
4906 for_each_pwq(pwq, wq)
4907 pwq_adjust_max_active(pwq);
4908 mutex_unlock(&wq->mutex);
4911 mutex_unlock(&wq_pool_mutex);
4915 * freeze_workqueues_busy - are freezable workqueues still busy?
4917 * Check whether freezing is complete. This function must be called
4918 * between freeze_workqueues_begin() and thaw_workqueues().
4920 * CONTEXT:
4921 * Grabs and releases wq_pool_mutex.
4923 * Return:
4924 * %true if some freezable workqueues are still busy. %false if freezing
4925 * is complete.
4927 bool freeze_workqueues_busy(void)
4929 bool busy = false;
4930 struct workqueue_struct *wq;
4931 struct pool_workqueue *pwq;
4933 mutex_lock(&wq_pool_mutex);
4935 WARN_ON_ONCE(!workqueue_freezing);
4937 list_for_each_entry(wq, &workqueues, list) {
4938 if (!(wq->flags & WQ_FREEZABLE))
4939 continue;
4941 * nr_active is monotonically decreasing. It's safe
4942 * to peek without lock.
4944 rcu_read_lock_sched();
4945 for_each_pwq(pwq, wq) {
4946 WARN_ON_ONCE(pwq->nr_active < 0);
4947 if (pwq->nr_active) {
4948 busy = true;
4949 rcu_read_unlock_sched();
4950 goto out_unlock;
4953 rcu_read_unlock_sched();
4955 out_unlock:
4956 mutex_unlock(&wq_pool_mutex);
4957 return busy;
4961 * thaw_workqueues - thaw workqueues
4963 * Thaw workqueues. Normal queueing is restored and all collected
4964 * frozen works are transferred to their respective pool worklists.
4966 * CONTEXT:
4967 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4969 void thaw_workqueues(void)
4971 struct workqueue_struct *wq;
4972 struct pool_workqueue *pwq;
4973 struct worker_pool *pool;
4974 int pi;
4976 mutex_lock(&wq_pool_mutex);
4978 if (!workqueue_freezing)
4979 goto out_unlock;
4981 /* clear FREEZING */
4982 for_each_pool(pool, pi) {
4983 spin_lock_irq(&pool->lock);
4984 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4985 pool->flags &= ~POOL_FREEZING;
4986 spin_unlock_irq(&pool->lock);
4989 /* restore max_active and repopulate worklist */
4990 list_for_each_entry(wq, &workqueues, list) {
4991 mutex_lock(&wq->mutex);
4992 for_each_pwq(pwq, wq)
4993 pwq_adjust_max_active(pwq);
4994 mutex_unlock(&wq->mutex);
4997 workqueue_freezing = false;
4998 out_unlock:
4999 mutex_unlock(&wq_pool_mutex);
5001 #endif /* CONFIG_FREEZER */
5003 static void __init wq_numa_init(void)
5005 cpumask_var_t *tbl;
5006 int node, cpu;
5008 /* determine NUMA pwq table len - highest node id + 1 */
5009 for_each_node(node)
5010 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
5012 if (num_possible_nodes() <= 1)
5013 return;
5015 if (wq_disable_numa) {
5016 pr_info("workqueue: NUMA affinity support disabled\n");
5017 return;
5020 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5021 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5024 * We want masks of possible CPUs of each node which isn't readily
5025 * available. Build one from cpu_to_node() which should have been
5026 * fully initialized by now.
5028 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
5029 BUG_ON(!tbl);
5031 for_each_node(node)
5032 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5033 node_online(node) ? node : NUMA_NO_NODE));
5035 for_each_possible_cpu(cpu) {
5036 node = cpu_to_node(cpu);
5037 if (WARN_ON(node == NUMA_NO_NODE)) {
5038 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5039 /* happens iff arch is bonkers, let's just proceed */
5040 return;
5042 cpumask_set_cpu(cpu, tbl[node]);
5045 wq_numa_possible_cpumask = tbl;
5046 wq_numa_enabled = true;
5049 static int __init init_workqueues(void)
5051 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5052 int i, cpu;
5054 /* make sure we have enough bits for OFFQ pool ID */
5055 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
5056 WORK_CPU_END * NR_STD_WORKER_POOLS);
5058 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5060 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5062 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5063 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5065 wq_numa_init();
5067 /* initialize CPU pools */
5068 for_each_possible_cpu(cpu) {
5069 struct worker_pool *pool;
5071 i = 0;
5072 for_each_cpu_worker_pool(pool, cpu) {
5073 BUG_ON(init_worker_pool(pool));
5074 pool->cpu = cpu;
5075 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5076 pool->attrs->nice = std_nice[i++];
5077 pool->node = cpu_to_node(cpu);
5079 /* alloc pool ID */
5080 mutex_lock(&wq_pool_mutex);
5081 BUG_ON(worker_pool_assign_id(pool));
5082 mutex_unlock(&wq_pool_mutex);
5086 /* create the initial worker */
5087 for_each_online_cpu(cpu) {
5088 struct worker_pool *pool;
5090 for_each_cpu_worker_pool(pool, cpu) {
5091 pool->flags &= ~POOL_DISASSOCIATED;
5092 BUG_ON(create_and_start_worker(pool) < 0);
5096 /* create default unbound and ordered wq attrs */
5097 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5098 struct workqueue_attrs *attrs;
5100 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5101 attrs->nice = std_nice[i];
5102 unbound_std_wq_attrs[i] = attrs;
5105 * An ordered wq should have only one pwq as ordering is
5106 * guaranteed by max_active which is enforced by pwqs.
5107 * Turn off NUMA so that dfl_pwq is used for all nodes.
5109 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5110 attrs->nice = std_nice[i];
5111 attrs->no_numa = true;
5112 ordered_wq_attrs[i] = attrs;
5115 system_wq = alloc_workqueue("events", 0, 0);
5116 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5117 system_long_wq = alloc_workqueue("events_long", 0, 0);
5118 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5119 WQ_UNBOUND_MAX_ACTIVE);
5120 system_freezable_wq = alloc_workqueue("events_freezable",
5121 WQ_FREEZABLE, 0);
5122 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5123 WQ_POWER_EFFICIENT, 0);
5124 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5125 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5127 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5128 !system_unbound_wq || !system_freezable_wq ||
5129 !system_power_efficient_wq ||
5130 !system_freezable_power_efficient_wq);
5131 return 0;
5133 early_initcall(init_workqueues);