2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * 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 */
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
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,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * 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 */
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
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().
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;
215 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
218 * Structure used to wait for workqueue flush.
221 struct list_head list; /* WQ: list of flushers */
222 int flush_color; /* WQ: flush color waiting for */
223 struct completion done; /* flush completion */
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 */
254 struct wq_device *wq_dev; /* I: for sysfs interface */
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map;
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;
280 static bool wq_power_efficient;
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],
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 struct workqueue_struct *system_wq __read_mostly;
309 EXPORT_SYMBOL(system_wq);
310 struct workqueue_struct *system_highpri_wq __read_mostly;
311 EXPORT_SYMBOL_GPL(system_highpri_wq);
312 struct workqueue_struct *system_long_wq __read_mostly;
313 EXPORT_SYMBOL_GPL(system_long_wq);
314 struct workqueue_struct *system_unbound_wq __read_mostly;
315 EXPORT_SYMBOL_GPL(system_unbound_wq);
316 struct workqueue_struct *system_freezable_wq __read_mostly;
317 EXPORT_SYMBOL_GPL(system_freezable_wq);
318 struct workqueue_struct *system_power_efficient_wq __read_mostly;
319 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
320 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
321 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
323 static int worker_thread(void *__worker);
324 static void copy_workqueue_attrs(struct workqueue_attrs *to,
325 const struct workqueue_attrs *from);
327 #define CREATE_TRACE_POINTS
328 #include <trace/events/workqueue.h>
330 #define assert_rcu_or_pool_mutex() \
331 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
332 lockdep_is_held(&wq_pool_mutex), \
333 "sched RCU or wq_pool_mutex should be held")
335 #define assert_rcu_or_wq_mutex(wq) \
336 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
337 lockdep_is_held(&wq->mutex), \
338 "sched RCU or wq->mutex should be held")
340 #ifdef CONFIG_LOCKDEP
341 #define assert_manager_or_pool_lock(pool) \
342 WARN_ONCE(debug_locks && \
343 !lockdep_is_held(&(pool)->manager_mutex) && \
344 !lockdep_is_held(&(pool)->lock), \
345 "pool->manager_mutex or ->lock should be held")
347 #define assert_manager_or_pool_lock(pool) do { } while (0)
350 #define for_each_cpu_worker_pool(pool, cpu) \
351 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
352 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
356 * for_each_pool - iterate through all worker_pools in the system
357 * @pool: iteration cursor
358 * @pi: integer used for iteration
360 * This must be called either with wq_pool_mutex held or sched RCU read
361 * locked. If the pool needs to be used beyond the locking in effect, the
362 * caller is responsible for guaranteeing that the pool stays online.
364 * The if/else clause exists only for the lockdep assertion and can be
367 #define for_each_pool(pool, pi) \
368 idr_for_each_entry(&worker_pool_idr, pool, pi) \
369 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
373 * for_each_pool_worker - iterate through all workers of a worker_pool
374 * @worker: iteration cursor
375 * @wi: integer used for iteration
376 * @pool: worker_pool to iterate workers of
378 * This must be called with either @pool->manager_mutex or ->lock held.
380 * The if/else clause exists only for the lockdep assertion and can be
383 #define for_each_pool_worker(worker, wi, pool) \
384 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
385 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
389 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
390 * @pwq: iteration cursor
391 * @wq: the target workqueue
393 * This must be called either with wq->mutex held or sched RCU read locked.
394 * If the pwq needs to be used beyond the locking in effect, the caller is
395 * responsible for guaranteeing that the pwq stays online.
397 * The if/else clause exists only for the lockdep assertion and can be
400 #define for_each_pwq(pwq, wq) \
401 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
402 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
405 #ifdef CONFIG_DEBUG_OBJECTS_WORK
407 static struct debug_obj_descr work_debug_descr;
409 static void *work_debug_hint(void *addr)
411 return ((struct work_struct *) addr)->func;
415 * fixup_init is called when:
416 * - an active object is initialized
418 static int work_fixup_init(void *addr, enum debug_obj_state state)
420 struct work_struct *work = addr;
423 case ODEBUG_STATE_ACTIVE:
424 cancel_work_sync(work);
425 debug_object_init(work, &work_debug_descr);
433 * fixup_activate is called when:
434 * - an active object is activated
435 * - an unknown object is activated (might be a statically initialized object)
437 static int work_fixup_activate(void *addr, enum debug_obj_state state)
439 struct work_struct *work = addr;
443 case ODEBUG_STATE_NOTAVAILABLE:
445 * This is not really a fixup. The work struct was
446 * statically initialized. We just make sure that it
447 * is tracked in the object tracker.
449 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
450 debug_object_init(work, &work_debug_descr);
451 debug_object_activate(work, &work_debug_descr);
457 case ODEBUG_STATE_ACTIVE:
466 * fixup_free is called when:
467 * - an active object is freed
469 static int work_fixup_free(void *addr, enum debug_obj_state state)
471 struct work_struct *work = addr;
474 case ODEBUG_STATE_ACTIVE:
475 cancel_work_sync(work);
476 debug_object_free(work, &work_debug_descr);
483 static struct debug_obj_descr work_debug_descr = {
484 .name = "work_struct",
485 .debug_hint = work_debug_hint,
486 .fixup_init = work_fixup_init,
487 .fixup_activate = work_fixup_activate,
488 .fixup_free = work_fixup_free,
491 static inline void debug_work_activate(struct work_struct *work)
493 debug_object_activate(work, &work_debug_descr);
496 static inline void debug_work_deactivate(struct work_struct *work)
498 debug_object_deactivate(work, &work_debug_descr);
501 void __init_work(struct work_struct *work, int onstack)
504 debug_object_init_on_stack(work, &work_debug_descr);
506 debug_object_init(work, &work_debug_descr);
508 EXPORT_SYMBOL_GPL(__init_work);
510 void destroy_work_on_stack(struct work_struct *work)
512 debug_object_free(work, &work_debug_descr);
514 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
517 static inline void debug_work_activate(struct work_struct *work) { }
518 static inline void debug_work_deactivate(struct work_struct *work) { }
521 /* allocate ID and assign it to @pool */
522 static int worker_pool_assign_id(struct worker_pool *pool)
526 lockdep_assert_held(&wq_pool_mutex);
528 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
537 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
538 * @wq: the target workqueue
541 * This must be called either with pwq_lock held or sched RCU read locked.
542 * If the pwq needs to be used beyond the locking in effect, the caller is
543 * responsible for guaranteeing that the pwq stays online.
545 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
548 assert_rcu_or_wq_mutex(wq);
549 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
552 static unsigned int work_color_to_flags(int color)
554 return color << WORK_STRUCT_COLOR_SHIFT;
557 static int get_work_color(struct work_struct *work)
559 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
560 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
563 static int work_next_color(int color)
565 return (color + 1) % WORK_NR_COLORS;
569 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
570 * contain the pointer to the queued pwq. Once execution starts, the flag
571 * is cleared and the high bits contain OFFQ flags and pool ID.
573 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
574 * and clear_work_data() can be used to set the pwq, pool or clear
575 * work->data. These functions should only be called while the work is
576 * owned - ie. while the PENDING bit is set.
578 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
579 * corresponding to a work. Pool is available once the work has been
580 * queued anywhere after initialization until it is sync canceled. pwq is
581 * available only while the work item is queued.
583 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
584 * canceled. While being canceled, a work item may have its PENDING set
585 * but stay off timer and worklist for arbitrarily long and nobody should
586 * try to steal the PENDING bit.
588 static inline void set_work_data(struct work_struct *work, unsigned long data,
591 WARN_ON_ONCE(!work_pending(work));
592 atomic_long_set(&work->data, data | flags | work_static(work));
595 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
596 unsigned long extra_flags)
598 set_work_data(work, (unsigned long)pwq,
599 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
602 static void set_work_pool_and_keep_pending(struct work_struct *work,
605 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
606 WORK_STRUCT_PENDING);
609 static void set_work_pool_and_clear_pending(struct work_struct *work,
613 * The following wmb is paired with the implied mb in
614 * test_and_set_bit(PENDING) and ensures all updates to @work made
615 * here are visible to and precede any updates by the next PENDING
619 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
622 static void clear_work_data(struct work_struct *work)
624 smp_wmb(); /* see set_work_pool_and_clear_pending() */
625 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
628 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
630 unsigned long data = atomic_long_read(&work->data);
632 if (data & WORK_STRUCT_PWQ)
633 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
639 * get_work_pool - return the worker_pool a given work was associated with
640 * @work: the work item of interest
642 * Return the worker_pool @work was last associated with. %NULL if none.
644 * Pools are created and destroyed under wq_pool_mutex, and allows read
645 * access under sched-RCU read lock. As such, this function should be
646 * called under wq_pool_mutex or with preemption disabled.
648 * All fields of the returned pool are accessible as long as the above
649 * mentioned locking is in effect. If the returned pool needs to be used
650 * beyond the critical section, the caller is responsible for ensuring the
651 * returned pool is and stays online.
653 static struct worker_pool *get_work_pool(struct work_struct *work)
655 unsigned long data = atomic_long_read(&work->data);
658 assert_rcu_or_pool_mutex();
660 if (data & WORK_STRUCT_PWQ)
661 return ((struct pool_workqueue *)
662 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
664 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
665 if (pool_id == WORK_OFFQ_POOL_NONE)
668 return idr_find(&worker_pool_idr, pool_id);
672 * get_work_pool_id - return the worker pool ID a given work is associated with
673 * @work: the work item of interest
675 * Return the worker_pool ID @work was last associated with.
676 * %WORK_OFFQ_POOL_NONE if none.
678 static int get_work_pool_id(struct work_struct *work)
680 unsigned long data = atomic_long_read(&work->data);
682 if (data & WORK_STRUCT_PWQ)
683 return ((struct pool_workqueue *)
684 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
686 return data >> WORK_OFFQ_POOL_SHIFT;
689 static void mark_work_canceling(struct work_struct *work)
691 unsigned long pool_id = get_work_pool_id(work);
693 pool_id <<= WORK_OFFQ_POOL_SHIFT;
694 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
697 static bool work_is_canceling(struct work_struct *work)
699 unsigned long data = atomic_long_read(&work->data);
701 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
705 * Policy functions. These define the policies on how the global worker
706 * pools are managed. Unless noted otherwise, these functions assume that
707 * they're being called with pool->lock held.
710 static bool __need_more_worker(struct worker_pool *pool)
712 return !atomic_read(&pool->nr_running);
716 * Need to wake up a worker? Called from anything but currently
719 * Note that, because unbound workers never contribute to nr_running, this
720 * function will always return %true for unbound pools as long as the
721 * worklist isn't empty.
723 static bool need_more_worker(struct worker_pool *pool)
725 return !list_empty(&pool->worklist) && __need_more_worker(pool);
728 /* Can I start working? Called from busy but !running workers. */
729 static bool may_start_working(struct worker_pool *pool)
731 return pool->nr_idle;
734 /* Do I need to keep working? Called from currently running workers. */
735 static bool keep_working(struct worker_pool *pool)
737 return !list_empty(&pool->worklist) &&
738 atomic_read(&pool->nr_running) <= 1;
741 /* Do we need a new worker? Called from manager. */
742 static bool need_to_create_worker(struct worker_pool *pool)
744 return need_more_worker(pool) && !may_start_working(pool);
747 /* Do I need to be the manager? */
748 static bool need_to_manage_workers(struct worker_pool *pool)
750 return need_to_create_worker(pool) ||
751 (pool->flags & POOL_MANAGE_WORKERS);
754 /* Do we have too many workers and should some go away? */
755 static bool too_many_workers(struct worker_pool *pool)
757 bool managing = mutex_is_locked(&pool->manager_arb);
758 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
759 int nr_busy = pool->nr_workers - nr_idle;
762 * nr_idle and idle_list may disagree if idle rebinding is in
763 * progress. Never return %true if idle_list is empty.
765 if (list_empty(&pool->idle_list))
768 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
775 /* Return the first worker. Safe with preemption disabled */
776 static struct worker *first_worker(struct worker_pool *pool)
778 if (unlikely(list_empty(&pool->idle_list)))
781 return list_first_entry(&pool->idle_list, struct worker, entry);
785 * wake_up_worker - wake up an idle worker
786 * @pool: worker pool to wake worker from
788 * Wake up the first idle worker of @pool.
791 * spin_lock_irq(pool->lock).
793 static void wake_up_worker(struct worker_pool *pool)
795 struct worker *worker = first_worker(pool);
798 wake_up_process(worker->task);
802 * wq_worker_waking_up - a worker is waking up
803 * @task: task waking up
804 * @cpu: CPU @task is waking up to
806 * This function is called during try_to_wake_up() when a worker is
810 * spin_lock_irq(rq->lock)
812 void wq_worker_waking_up(struct task_struct *task, int cpu)
814 struct worker *worker = kthread_data(task);
816 if (!(worker->flags & WORKER_NOT_RUNNING)) {
817 WARN_ON_ONCE(worker->pool->cpu != cpu);
818 atomic_inc(&worker->pool->nr_running);
823 * wq_worker_sleeping - a worker is going to sleep
824 * @task: task going to sleep
825 * @cpu: CPU in question, must be the current CPU number
827 * This function is called during schedule() when a busy worker is
828 * going to sleep. Worker on the same cpu can be woken up by
829 * returning pointer to its task.
832 * spin_lock_irq(rq->lock)
835 * Worker task on @cpu to wake up, %NULL if none.
837 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
839 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
840 struct worker_pool *pool;
843 * Rescuers, which may not have all the fields set up like normal
844 * workers, also reach here, let's not access anything before
845 * checking NOT_RUNNING.
847 if (worker->flags & WORKER_NOT_RUNNING)
852 /* this can only happen on the local cpu */
853 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
857 * The counterpart of the following dec_and_test, implied mb,
858 * worklist not empty test sequence is in insert_work().
859 * Please read comment there.
861 * NOT_RUNNING is clear. This means that we're bound to and
862 * running on the local cpu w/ rq lock held and preemption
863 * disabled, which in turn means that none else could be
864 * manipulating idle_list, so dereferencing idle_list without pool
867 if (atomic_dec_and_test(&pool->nr_running) &&
868 !list_empty(&pool->worklist))
869 to_wakeup = first_worker(pool);
870 return to_wakeup ? to_wakeup->task : NULL;
874 * worker_set_flags - set worker flags and adjust nr_running accordingly
876 * @flags: flags to set
877 * @wakeup: wakeup an idle worker if necessary
879 * Set @flags in @worker->flags and adjust nr_running accordingly. If
880 * nr_running becomes zero and @wakeup is %true, an idle worker is
884 * spin_lock_irq(pool->lock)
886 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
889 struct worker_pool *pool = worker->pool;
891 WARN_ON_ONCE(worker->task != current);
894 * If transitioning into NOT_RUNNING, adjust nr_running and
895 * wake up an idle worker as necessary if requested by
898 if ((flags & WORKER_NOT_RUNNING) &&
899 !(worker->flags & WORKER_NOT_RUNNING)) {
901 if (atomic_dec_and_test(&pool->nr_running) &&
902 !list_empty(&pool->worklist))
903 wake_up_worker(pool);
905 atomic_dec(&pool->nr_running);
908 worker->flags |= flags;
912 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
914 * @flags: flags to clear
916 * Clear @flags in @worker->flags and adjust nr_running accordingly.
919 * spin_lock_irq(pool->lock)
921 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
923 struct worker_pool *pool = worker->pool;
924 unsigned int oflags = worker->flags;
926 WARN_ON_ONCE(worker->task != current);
928 worker->flags &= ~flags;
931 * If transitioning out of NOT_RUNNING, increment nr_running. Note
932 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
933 * of multiple flags, not a single flag.
935 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
936 if (!(worker->flags & WORKER_NOT_RUNNING))
937 atomic_inc(&pool->nr_running);
941 * find_worker_executing_work - find worker which is executing a work
942 * @pool: pool of interest
943 * @work: work to find worker for
945 * Find a worker which is executing @work on @pool by searching
946 * @pool->busy_hash which is keyed by the address of @work. For a worker
947 * to match, its current execution should match the address of @work and
948 * its work function. This is to avoid unwanted dependency between
949 * unrelated work executions through a work item being recycled while still
952 * This is a bit tricky. A work item may be freed once its execution
953 * starts and nothing prevents the freed area from being recycled for
954 * another work item. If the same work item address ends up being reused
955 * before the original execution finishes, workqueue will identify the
956 * recycled work item as currently executing and make it wait until the
957 * current execution finishes, introducing an unwanted dependency.
959 * This function checks the work item address and work function to avoid
960 * false positives. Note that this isn't complete as one may construct a
961 * work function which can introduce dependency onto itself through a
962 * recycled work item. Well, if somebody wants to shoot oneself in the
963 * foot that badly, there's only so much we can do, and if such deadlock
964 * actually occurs, it should be easy to locate the culprit work function.
967 * spin_lock_irq(pool->lock).
970 * Pointer to worker which is executing @work if found, NULL
973 static struct worker *find_worker_executing_work(struct worker_pool *pool,
974 struct work_struct *work)
976 struct worker *worker;
978 hash_for_each_possible(pool->busy_hash, worker, hentry,
980 if (worker->current_work == work &&
981 worker->current_func == work->func)
988 * move_linked_works - move linked works to a list
989 * @work: start of series of works to be scheduled
990 * @head: target list to append @work to
991 * @nextp: out paramter for nested worklist walking
993 * Schedule linked works starting from @work to @head. Work series to
994 * be scheduled starts at @work and includes any consecutive work with
995 * WORK_STRUCT_LINKED set in its predecessor.
997 * If @nextp is not NULL, it's updated to point to the next work of
998 * the last scheduled work. This allows move_linked_works() to be
999 * nested inside outer list_for_each_entry_safe().
1002 * spin_lock_irq(pool->lock).
1004 static void move_linked_works(struct work_struct *work, struct list_head *head,
1005 struct work_struct **nextp)
1007 struct work_struct *n;
1010 * Linked worklist will always end before the end of the list,
1011 * use NULL for list head.
1013 list_for_each_entry_safe_from(work, n, NULL, entry) {
1014 list_move_tail(&work->entry, head);
1015 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1020 * If we're already inside safe list traversal and have moved
1021 * multiple works to the scheduled queue, the next position
1022 * needs to be updated.
1029 * get_pwq - get an extra reference on the specified pool_workqueue
1030 * @pwq: pool_workqueue to get
1032 * Obtain an extra reference on @pwq. The caller should guarantee that
1033 * @pwq has positive refcnt and be holding the matching pool->lock.
1035 static void get_pwq(struct pool_workqueue *pwq)
1037 lockdep_assert_held(&pwq->pool->lock);
1038 WARN_ON_ONCE(pwq->refcnt <= 0);
1043 * put_pwq - put a pool_workqueue reference
1044 * @pwq: pool_workqueue to put
1046 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1047 * destruction. The caller should be holding the matching pool->lock.
1049 static void put_pwq(struct pool_workqueue *pwq)
1051 lockdep_assert_held(&pwq->pool->lock);
1052 if (likely(--pwq->refcnt))
1054 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1057 * @pwq can't be released under pool->lock, bounce to
1058 * pwq_unbound_release_workfn(). This never recurses on the same
1059 * pool->lock as this path is taken only for unbound workqueues and
1060 * the release work item is scheduled on a per-cpu workqueue. To
1061 * avoid lockdep warning, unbound pool->locks are given lockdep
1062 * subclass of 1 in get_unbound_pool().
1064 schedule_work(&pwq->unbound_release_work);
1068 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1069 * @pwq: pool_workqueue to put (can be %NULL)
1071 * put_pwq() with locking. This function also allows %NULL @pwq.
1073 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1077 * As both pwqs and pools are sched-RCU protected, the
1078 * following lock operations are safe.
1080 spin_lock_irq(&pwq->pool->lock);
1082 spin_unlock_irq(&pwq->pool->lock);
1086 static void pwq_activate_delayed_work(struct work_struct *work)
1088 struct pool_workqueue *pwq = get_work_pwq(work);
1090 trace_workqueue_activate_work(work);
1091 move_linked_works(work, &pwq->pool->worklist, NULL);
1092 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1096 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1098 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1099 struct work_struct, entry);
1101 pwq_activate_delayed_work(work);
1105 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1106 * @pwq: pwq of interest
1107 * @color: color of work which left the queue
1109 * A work either has completed or is removed from pending queue,
1110 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1113 * spin_lock_irq(pool->lock).
1115 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1117 /* uncolored work items don't participate in flushing or nr_active */
1118 if (color == WORK_NO_COLOR)
1121 pwq->nr_in_flight[color]--;
1124 if (!list_empty(&pwq->delayed_works)) {
1125 /* one down, submit a delayed one */
1126 if (pwq->nr_active < pwq->max_active)
1127 pwq_activate_first_delayed(pwq);
1130 /* is flush in progress and are we at the flushing tip? */
1131 if (likely(pwq->flush_color != color))
1134 /* are there still in-flight works? */
1135 if (pwq->nr_in_flight[color])
1138 /* this pwq is done, clear flush_color */
1139 pwq->flush_color = -1;
1142 * If this was the last pwq, wake up the first flusher. It
1143 * will handle the rest.
1145 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1146 complete(&pwq->wq->first_flusher->done);
1152 * try_to_grab_pending - steal work item from worklist and disable irq
1153 * @work: work item to steal
1154 * @is_dwork: @work is a delayed_work
1155 * @flags: place to store irq state
1157 * Try to grab PENDING bit of @work. This function can handle @work in any
1158 * stable state - idle, on timer or on worklist. Return values are
1160 * 1 if @work was pending and we successfully stole PENDING
1161 * 0 if @work was idle and we claimed PENDING
1162 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1163 * -ENOENT if someone else is canceling @work, this state may persist
1164 * for arbitrarily long
1166 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1167 * interrupted while holding PENDING and @work off queue, irq must be
1168 * disabled on entry. This, combined with delayed_work->timer being
1169 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1171 * On successful return, >= 0, irq is disabled and the caller is
1172 * responsible for releasing it using local_irq_restore(*@flags).
1174 * This function is safe to call from any context including IRQ handler.
1176 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1177 unsigned long *flags)
1179 struct worker_pool *pool;
1180 struct pool_workqueue *pwq;
1182 local_irq_save(*flags);
1184 /* try to steal the timer if it exists */
1186 struct delayed_work *dwork = to_delayed_work(work);
1189 * dwork->timer is irqsafe. If del_timer() fails, it's
1190 * guaranteed that the timer is not queued anywhere and not
1191 * running on the local CPU.
1193 if (likely(del_timer(&dwork->timer)))
1197 /* try to claim PENDING the normal way */
1198 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1202 * The queueing is in progress, or it is already queued. Try to
1203 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1205 pool = get_work_pool(work);
1209 spin_lock(&pool->lock);
1211 * work->data is guaranteed to point to pwq only while the work
1212 * item is queued on pwq->wq, and both updating work->data to point
1213 * to pwq on queueing and to pool on dequeueing are done under
1214 * pwq->pool->lock. This in turn guarantees that, if work->data
1215 * points to pwq which is associated with a locked pool, the work
1216 * item is currently queued on that pool.
1218 pwq = get_work_pwq(work);
1219 if (pwq && pwq->pool == pool) {
1220 debug_work_deactivate(work);
1223 * A delayed work item cannot be grabbed directly because
1224 * it might have linked NO_COLOR work items which, if left
1225 * on the delayed_list, will confuse pwq->nr_active
1226 * management later on and cause stall. Make sure the work
1227 * item is activated before grabbing.
1229 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1230 pwq_activate_delayed_work(work);
1232 list_del_init(&work->entry);
1233 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1235 /* work->data points to pwq iff queued, point to pool */
1236 set_work_pool_and_keep_pending(work, pool->id);
1238 spin_unlock(&pool->lock);
1241 spin_unlock(&pool->lock);
1243 local_irq_restore(*flags);
1244 if (work_is_canceling(work))
1251 * insert_work - insert a work into a pool
1252 * @pwq: pwq @work belongs to
1253 * @work: work to insert
1254 * @head: insertion point
1255 * @extra_flags: extra WORK_STRUCT_* flags to set
1257 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1258 * work_struct flags.
1261 * spin_lock_irq(pool->lock).
1263 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1264 struct list_head *head, unsigned int extra_flags)
1266 struct worker_pool *pool = pwq->pool;
1268 /* we own @work, set data and link */
1269 set_work_pwq(work, pwq, extra_flags);
1270 list_add_tail(&work->entry, head);
1274 * Ensure either wq_worker_sleeping() sees the above
1275 * list_add_tail() or we see zero nr_running to avoid workers lying
1276 * around lazily while there are works to be processed.
1280 if (__need_more_worker(pool))
1281 wake_up_worker(pool);
1285 * Test whether @work is being queued from another work executing on the
1288 static bool is_chained_work(struct workqueue_struct *wq)
1290 struct worker *worker;
1292 worker = current_wq_worker();
1294 * Return %true iff I'm a worker execuing a work item on @wq. If
1295 * I'm @worker, it's safe to dereference it without locking.
1297 return worker && worker->current_pwq->wq == wq;
1300 static void __queue_work(int cpu, struct workqueue_struct *wq,
1301 struct work_struct *work)
1303 struct pool_workqueue *pwq;
1304 struct worker_pool *last_pool;
1305 struct list_head *worklist;
1306 unsigned int work_flags;
1307 unsigned int req_cpu = cpu;
1310 * While a work item is PENDING && off queue, a task trying to
1311 * steal the PENDING will busy-loop waiting for it to either get
1312 * queued or lose PENDING. Grabbing PENDING and queueing should
1313 * happen with IRQ disabled.
1315 WARN_ON_ONCE(!irqs_disabled());
1317 debug_work_activate(work);
1319 /* if dying, only works from the same workqueue are allowed */
1320 if (unlikely(wq->flags & __WQ_DRAINING) &&
1321 WARN_ON_ONCE(!is_chained_work(wq)))
1324 if (req_cpu == WORK_CPU_UNBOUND)
1325 cpu = raw_smp_processor_id();
1327 /* pwq which will be used unless @work is executing elsewhere */
1328 if (!(wq->flags & WQ_UNBOUND))
1329 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1331 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1334 * If @work was previously on a different pool, it might still be
1335 * running there, in which case the work needs to be queued on that
1336 * pool to guarantee non-reentrancy.
1338 last_pool = get_work_pool(work);
1339 if (last_pool && last_pool != pwq->pool) {
1340 struct worker *worker;
1342 spin_lock(&last_pool->lock);
1344 worker = find_worker_executing_work(last_pool, work);
1346 if (worker && worker->current_pwq->wq == wq) {
1347 pwq = worker->current_pwq;
1349 /* meh... not running there, queue here */
1350 spin_unlock(&last_pool->lock);
1351 spin_lock(&pwq->pool->lock);
1354 spin_lock(&pwq->pool->lock);
1358 * pwq is determined and locked. For unbound pools, we could have
1359 * raced with pwq release and it could already be dead. If its
1360 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1361 * without another pwq replacing it in the numa_pwq_tbl or while
1362 * work items are executing on it, so the retrying is guaranteed to
1363 * make forward-progress.
1365 if (unlikely(!pwq->refcnt)) {
1366 if (wq->flags & WQ_UNBOUND) {
1367 spin_unlock(&pwq->pool->lock);
1372 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1376 /* pwq determined, queue */
1377 trace_workqueue_queue_work(req_cpu, pwq, work);
1379 if (WARN_ON(!list_empty(&work->entry))) {
1380 spin_unlock(&pwq->pool->lock);
1384 pwq->nr_in_flight[pwq->work_color]++;
1385 work_flags = work_color_to_flags(pwq->work_color);
1387 if (likely(pwq->nr_active < pwq->max_active)) {
1388 trace_workqueue_activate_work(work);
1390 worklist = &pwq->pool->worklist;
1392 work_flags |= WORK_STRUCT_DELAYED;
1393 worklist = &pwq->delayed_works;
1396 insert_work(pwq, work, worklist, work_flags);
1398 spin_unlock(&pwq->pool->lock);
1402 * queue_work_on - queue work on specific cpu
1403 * @cpu: CPU number to execute work on
1404 * @wq: workqueue to use
1405 * @work: work to queue
1407 * Returns %false if @work was already on a queue, %true otherwise.
1409 * We queue the work to a specific CPU, the caller must ensure it
1412 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1413 struct work_struct *work)
1416 unsigned long flags;
1418 local_irq_save(flags);
1420 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1421 __queue_work(cpu, wq, work);
1425 local_irq_restore(flags);
1428 EXPORT_SYMBOL(queue_work_on);
1430 void delayed_work_timer_fn(unsigned long __data)
1432 struct delayed_work *dwork = (struct delayed_work *)__data;
1434 /* should have been called from irqsafe timer with irq already off */
1435 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1437 EXPORT_SYMBOL(delayed_work_timer_fn);
1439 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1440 struct delayed_work *dwork, unsigned long delay)
1442 struct timer_list *timer = &dwork->timer;
1443 struct work_struct *work = &dwork->work;
1445 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1446 timer->data != (unsigned long)dwork);
1447 WARN_ON_ONCE(timer_pending(timer));
1448 WARN_ON_ONCE(!list_empty(&work->entry));
1451 * If @delay is 0, queue @dwork->work immediately. This is for
1452 * both optimization and correctness. The earliest @timer can
1453 * expire is on the closest next tick and delayed_work users depend
1454 * on that there's no such delay when @delay is 0.
1457 __queue_work(cpu, wq, &dwork->work);
1461 timer_stats_timer_set_start_info(&dwork->timer);
1465 timer->expires = jiffies + delay;
1467 if (unlikely(cpu != WORK_CPU_UNBOUND))
1468 add_timer_on(timer, cpu);
1474 * queue_delayed_work_on - queue work on specific CPU after delay
1475 * @cpu: CPU number to execute work on
1476 * @wq: workqueue to use
1477 * @dwork: work to queue
1478 * @delay: number of jiffies to wait before queueing
1480 * Returns %false if @work was already on a queue, %true otherwise. If
1481 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1484 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1485 struct delayed_work *dwork, unsigned long delay)
1487 struct work_struct *work = &dwork->work;
1489 unsigned long flags;
1491 /* read the comment in __queue_work() */
1492 local_irq_save(flags);
1494 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1495 __queue_delayed_work(cpu, wq, dwork, delay);
1499 local_irq_restore(flags);
1502 EXPORT_SYMBOL(queue_delayed_work_on);
1505 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1506 * @cpu: CPU number to execute work on
1507 * @wq: workqueue to use
1508 * @dwork: work to queue
1509 * @delay: number of jiffies to wait before queueing
1511 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1512 * modify @dwork's timer so that it expires after @delay. If @delay is
1513 * zero, @work is guaranteed to be scheduled immediately regardless of its
1516 * Returns %false if @dwork was idle and queued, %true if @dwork was
1517 * pending and its timer was modified.
1519 * This function is safe to call from any context including IRQ handler.
1520 * See try_to_grab_pending() for details.
1522 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1523 struct delayed_work *dwork, unsigned long delay)
1525 unsigned long flags;
1529 ret = try_to_grab_pending(&dwork->work, true, &flags);
1530 } while (unlikely(ret == -EAGAIN));
1532 if (likely(ret >= 0)) {
1533 __queue_delayed_work(cpu, wq, dwork, delay);
1534 local_irq_restore(flags);
1537 /* -ENOENT from try_to_grab_pending() becomes %true */
1540 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1543 * worker_enter_idle - enter idle state
1544 * @worker: worker which is entering idle state
1546 * @worker is entering idle state. Update stats and idle timer if
1550 * spin_lock_irq(pool->lock).
1552 static void worker_enter_idle(struct worker *worker)
1554 struct worker_pool *pool = worker->pool;
1556 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1557 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1558 (worker->hentry.next || worker->hentry.pprev)))
1561 /* can't use worker_set_flags(), also called from start_worker() */
1562 worker->flags |= WORKER_IDLE;
1564 worker->last_active = jiffies;
1566 /* idle_list is LIFO */
1567 list_add(&worker->entry, &pool->idle_list);
1569 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1570 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1573 * Sanity check nr_running. Because wq_unbind_fn() releases
1574 * pool->lock between setting %WORKER_UNBOUND and zapping
1575 * nr_running, the warning may trigger spuriously. Check iff
1576 * unbind is not in progress.
1578 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1579 pool->nr_workers == pool->nr_idle &&
1580 atomic_read(&pool->nr_running));
1584 * worker_leave_idle - leave idle state
1585 * @worker: worker which is leaving idle state
1587 * @worker is leaving idle state. Update stats.
1590 * spin_lock_irq(pool->lock).
1592 static void worker_leave_idle(struct worker *worker)
1594 struct worker_pool *pool = worker->pool;
1596 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1598 worker_clr_flags(worker, WORKER_IDLE);
1600 list_del_init(&worker->entry);
1604 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1605 * @pool: target worker_pool
1607 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1609 * Works which are scheduled while the cpu is online must at least be
1610 * scheduled to a worker which is bound to the cpu so that if they are
1611 * flushed from cpu callbacks while cpu is going down, they are
1612 * guaranteed to execute on the cpu.
1614 * This function is to be used by unbound workers and rescuers to bind
1615 * themselves to the target cpu and may race with cpu going down or
1616 * coming online. kthread_bind() can't be used because it may put the
1617 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1618 * verbatim as it's best effort and blocking and pool may be
1619 * [dis]associated in the meantime.
1621 * This function tries set_cpus_allowed() and locks pool and verifies the
1622 * binding against %POOL_DISASSOCIATED which is set during
1623 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1624 * enters idle state or fetches works without dropping lock, it can
1625 * guarantee the scheduling requirement described in the first paragraph.
1628 * Might sleep. Called without any lock but returns with pool->lock
1632 * %true if the associated pool is online (@worker is successfully
1633 * bound), %false if offline.
1635 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1636 __acquires(&pool->lock)
1640 * The following call may fail, succeed or succeed
1641 * without actually migrating the task to the cpu if
1642 * it races with cpu hotunplug operation. Verify
1643 * against POOL_DISASSOCIATED.
1645 if (!(pool->flags & POOL_DISASSOCIATED))
1646 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1648 spin_lock_irq(&pool->lock);
1649 if (pool->flags & POOL_DISASSOCIATED)
1651 if (task_cpu(current) == pool->cpu &&
1652 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1654 spin_unlock_irq(&pool->lock);
1657 * We've raced with CPU hot[un]plug. Give it a breather
1658 * and retry migration. cond_resched() is required here;
1659 * otherwise, we might deadlock against cpu_stop trying to
1660 * bring down the CPU on non-preemptive kernel.
1667 static struct worker *alloc_worker(void)
1669 struct worker *worker;
1671 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1673 INIT_LIST_HEAD(&worker->entry);
1674 INIT_LIST_HEAD(&worker->scheduled);
1675 /* on creation a worker is in !idle && prep state */
1676 worker->flags = WORKER_PREP;
1682 * create_worker - create a new workqueue worker
1683 * @pool: pool the new worker will belong to
1685 * Create a new worker which is bound to @pool. The returned worker
1686 * can be started by calling start_worker() or destroyed using
1690 * Might sleep. Does GFP_KERNEL allocations.
1693 * Pointer to the newly created worker.
1695 static struct worker *create_worker(struct worker_pool *pool)
1697 struct worker *worker = NULL;
1701 lockdep_assert_held(&pool->manager_mutex);
1704 * ID is needed to determine kthread name. Allocate ID first
1705 * without installing the pointer.
1707 idr_preload(GFP_KERNEL);
1708 spin_lock_irq(&pool->lock);
1710 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1712 spin_unlock_irq(&pool->lock);
1717 worker = alloc_worker();
1721 worker->pool = pool;
1725 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1726 pool->attrs->nice < 0 ? "H" : "");
1728 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1730 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1731 "kworker/%s", id_buf);
1732 if (IS_ERR(worker->task))
1736 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1737 * online CPUs. It'll be re-applied when any of the CPUs come up.
1739 set_user_nice(worker->task, pool->attrs->nice);
1740 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1742 /* prevent userland from meddling with cpumask of workqueue workers */
1743 worker->task->flags |= PF_NO_SETAFFINITY;
1746 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1747 * remains stable across this function. See the comments above the
1748 * flag definition for details.
1750 if (pool->flags & POOL_DISASSOCIATED)
1751 worker->flags |= WORKER_UNBOUND;
1753 /* successful, commit the pointer to idr */
1754 spin_lock_irq(&pool->lock);
1755 idr_replace(&pool->worker_idr, worker, worker->id);
1756 spin_unlock_irq(&pool->lock);
1762 spin_lock_irq(&pool->lock);
1763 idr_remove(&pool->worker_idr, id);
1764 spin_unlock_irq(&pool->lock);
1771 * start_worker - start a newly created worker
1772 * @worker: worker to start
1774 * Make the pool aware of @worker and start it.
1777 * spin_lock_irq(pool->lock).
1779 static void start_worker(struct worker *worker)
1781 worker->flags |= WORKER_STARTED;
1782 worker->pool->nr_workers++;
1783 worker_enter_idle(worker);
1784 wake_up_process(worker->task);
1788 * create_and_start_worker - create and start a worker for a pool
1789 * @pool: the target pool
1791 * Grab the managership of @pool and create and start a new worker for it.
1793 static int create_and_start_worker(struct worker_pool *pool)
1795 struct worker *worker;
1797 mutex_lock(&pool->manager_mutex);
1799 worker = create_worker(pool);
1801 spin_lock_irq(&pool->lock);
1802 start_worker(worker);
1803 spin_unlock_irq(&pool->lock);
1806 mutex_unlock(&pool->manager_mutex);
1808 return worker ? 0 : -ENOMEM;
1812 * destroy_worker - destroy a workqueue worker
1813 * @worker: worker to be destroyed
1815 * Destroy @worker and adjust @pool stats accordingly.
1818 * spin_lock_irq(pool->lock) which is released and regrabbed.
1820 static void destroy_worker(struct worker *worker)
1822 struct worker_pool *pool = worker->pool;
1824 lockdep_assert_held(&pool->manager_mutex);
1825 lockdep_assert_held(&pool->lock);
1827 /* sanity check frenzy */
1828 if (WARN_ON(worker->current_work) ||
1829 WARN_ON(!list_empty(&worker->scheduled)))
1832 if (worker->flags & WORKER_STARTED)
1834 if (worker->flags & WORKER_IDLE)
1837 list_del_init(&worker->entry);
1838 worker->flags |= WORKER_DIE;
1840 idr_remove(&pool->worker_idr, worker->id);
1842 spin_unlock_irq(&pool->lock);
1844 kthread_stop(worker->task);
1847 spin_lock_irq(&pool->lock);
1850 static void idle_worker_timeout(unsigned long __pool)
1852 struct worker_pool *pool = (void *)__pool;
1854 spin_lock_irq(&pool->lock);
1856 if (too_many_workers(pool)) {
1857 struct worker *worker;
1858 unsigned long expires;
1860 /* idle_list is kept in LIFO order, check the last one */
1861 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1862 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1864 if (time_before(jiffies, expires))
1865 mod_timer(&pool->idle_timer, expires);
1867 /* it's been idle for too long, wake up manager */
1868 pool->flags |= POOL_MANAGE_WORKERS;
1869 wake_up_worker(pool);
1873 spin_unlock_irq(&pool->lock);
1876 static void send_mayday(struct work_struct *work)
1878 struct pool_workqueue *pwq = get_work_pwq(work);
1879 struct workqueue_struct *wq = pwq->wq;
1881 lockdep_assert_held(&wq_mayday_lock);
1886 /* mayday mayday mayday */
1887 if (list_empty(&pwq->mayday_node)) {
1888 list_add_tail(&pwq->mayday_node, &wq->maydays);
1889 wake_up_process(wq->rescuer->task);
1893 static void pool_mayday_timeout(unsigned long __pool)
1895 struct worker_pool *pool = (void *)__pool;
1896 struct work_struct *work;
1898 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1899 spin_lock(&pool->lock);
1901 if (need_to_create_worker(pool)) {
1903 * We've been trying to create a new worker but
1904 * haven't been successful. We might be hitting an
1905 * allocation deadlock. Send distress signals to
1908 list_for_each_entry(work, &pool->worklist, entry)
1912 spin_unlock(&pool->lock);
1913 spin_unlock_irq(&wq_mayday_lock);
1915 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1919 * maybe_create_worker - create a new worker if necessary
1920 * @pool: pool to create a new worker for
1922 * Create a new worker for @pool if necessary. @pool is guaranteed to
1923 * have at least one idle worker on return from this function. If
1924 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1925 * sent to all rescuers with works scheduled on @pool to resolve
1926 * possible allocation deadlock.
1928 * On return, need_to_create_worker() is guaranteed to be %false and
1929 * may_start_working() %true.
1932 * spin_lock_irq(pool->lock) which may be released and regrabbed
1933 * multiple times. Does GFP_KERNEL allocations. Called only from
1937 * %false if no action was taken and pool->lock stayed locked, %true
1940 static bool maybe_create_worker(struct worker_pool *pool)
1941 __releases(&pool->lock)
1942 __acquires(&pool->lock)
1944 if (!need_to_create_worker(pool))
1947 spin_unlock_irq(&pool->lock);
1949 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1950 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1953 struct worker *worker;
1955 worker = create_worker(pool);
1957 del_timer_sync(&pool->mayday_timer);
1958 spin_lock_irq(&pool->lock);
1959 start_worker(worker);
1960 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1965 if (!need_to_create_worker(pool))
1968 __set_current_state(TASK_INTERRUPTIBLE);
1969 schedule_timeout(CREATE_COOLDOWN);
1971 if (!need_to_create_worker(pool))
1975 del_timer_sync(&pool->mayday_timer);
1976 spin_lock_irq(&pool->lock);
1977 if (need_to_create_worker(pool))
1983 * maybe_destroy_worker - destroy workers which have been idle for a while
1984 * @pool: pool to destroy workers for
1986 * Destroy @pool workers which have been idle for longer than
1987 * IDLE_WORKER_TIMEOUT.
1990 * spin_lock_irq(pool->lock) which may be released and regrabbed
1991 * multiple times. Called only from manager.
1994 * %false if no action was taken and pool->lock stayed locked, %true
1997 static bool maybe_destroy_workers(struct worker_pool *pool)
2001 while (too_many_workers(pool)) {
2002 struct worker *worker;
2003 unsigned long expires;
2005 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2006 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2008 if (time_before(jiffies, expires)) {
2009 mod_timer(&pool->idle_timer, expires);
2013 destroy_worker(worker);
2021 * manage_workers - manage worker pool
2024 * Assume the manager role and manage the worker pool @worker belongs
2025 * to. At any given time, there can be only zero or one manager per
2026 * pool. The exclusion is handled automatically by this function.
2028 * The caller can safely start processing works on false return. On
2029 * true return, it's guaranteed that need_to_create_worker() is false
2030 * and may_start_working() is true.
2033 * spin_lock_irq(pool->lock) which may be released and regrabbed
2034 * multiple times. Does GFP_KERNEL allocations.
2037 * %false if the pool don't need management and the caller can safely start
2038 * processing works, %true indicates that the function released pool->lock
2039 * and reacquired it to perform some management function and that the
2040 * conditions that the caller verified while holding the lock before
2041 * calling the function might no longer be true.
2043 static bool manage_workers(struct worker *worker)
2045 struct worker_pool *pool = worker->pool;
2049 * Managership is governed by two mutexes - manager_arb and
2050 * manager_mutex. manager_arb handles arbitration of manager role.
2051 * Anyone who successfully grabs manager_arb wins the arbitration
2052 * and becomes the manager. mutex_trylock() on pool->manager_arb
2053 * failure while holding pool->lock reliably indicates that someone
2054 * else is managing the pool and the worker which failed trylock
2055 * can proceed to executing work items. This means that anyone
2056 * grabbing manager_arb is responsible for actually performing
2057 * manager duties. If manager_arb is grabbed and released without
2058 * actual management, the pool may stall indefinitely.
2060 * manager_mutex is used for exclusion of actual management
2061 * operations. The holder of manager_mutex can be sure that none
2062 * of management operations, including creation and destruction of
2063 * workers, won't take place until the mutex is released. Because
2064 * manager_mutex doesn't interfere with manager role arbitration,
2065 * it is guaranteed that the pool's management, while may be
2066 * delayed, won't be disturbed by someone else grabbing
2069 if (!mutex_trylock(&pool->manager_arb))
2073 * With manager arbitration won, manager_mutex would be free in
2074 * most cases. trylock first without dropping @pool->lock.
2076 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2077 spin_unlock_irq(&pool->lock);
2078 mutex_lock(&pool->manager_mutex);
2079 spin_lock_irq(&pool->lock);
2083 pool->flags &= ~POOL_MANAGE_WORKERS;
2086 * Destroy and then create so that may_start_working() is true
2089 ret |= maybe_destroy_workers(pool);
2090 ret |= maybe_create_worker(pool);
2092 mutex_unlock(&pool->manager_mutex);
2093 mutex_unlock(&pool->manager_arb);
2098 * process_one_work - process single work
2100 * @work: work to process
2102 * Process @work. This function contains all the logics necessary to
2103 * process a single work including synchronization against and
2104 * interaction with other workers on the same cpu, queueing and
2105 * flushing. As long as context requirement is met, any worker can
2106 * call this function to process a work.
2109 * spin_lock_irq(pool->lock) which is released and regrabbed.
2111 static void process_one_work(struct worker *worker, struct work_struct *work)
2112 __releases(&pool->lock)
2113 __acquires(&pool->lock)
2115 struct pool_workqueue *pwq = get_work_pwq(work);
2116 struct worker_pool *pool = worker->pool;
2117 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2119 struct worker *collision;
2120 #ifdef CONFIG_LOCKDEP
2122 * It is permissible to free the struct work_struct from
2123 * inside the function that is called from it, this we need to
2124 * take into account for lockdep too. To avoid bogus "held
2125 * lock freed" warnings as well as problems when looking into
2126 * work->lockdep_map, make a copy and use that here.
2128 struct lockdep_map lockdep_map;
2130 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2133 * Ensure we're on the correct CPU. DISASSOCIATED test is
2134 * necessary to avoid spurious warnings from rescuers servicing the
2135 * unbound or a disassociated pool.
2137 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2138 !(pool->flags & POOL_DISASSOCIATED) &&
2139 raw_smp_processor_id() != pool->cpu);
2142 * A single work shouldn't be executed concurrently by
2143 * multiple workers on a single cpu. Check whether anyone is
2144 * already processing the work. If so, defer the work to the
2145 * currently executing one.
2147 collision = find_worker_executing_work(pool, work);
2148 if (unlikely(collision)) {
2149 move_linked_works(work, &collision->scheduled, NULL);
2153 /* claim and dequeue */
2154 debug_work_deactivate(work);
2155 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2156 worker->current_work = work;
2157 worker->current_func = work->func;
2158 worker->current_pwq = pwq;
2159 work_color = get_work_color(work);
2161 list_del_init(&work->entry);
2164 * CPU intensive works don't participate in concurrency
2165 * management. They're the scheduler's responsibility.
2167 if (unlikely(cpu_intensive))
2168 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2171 * Unbound pool isn't concurrency managed and work items should be
2172 * executed ASAP. Wake up another worker if necessary.
2174 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2175 wake_up_worker(pool);
2178 * Record the last pool and clear PENDING which should be the last
2179 * update to @work. Also, do this inside @pool->lock so that
2180 * PENDING and queued state changes happen together while IRQ is
2183 set_work_pool_and_clear_pending(work, pool->id);
2185 spin_unlock_irq(&pool->lock);
2187 lock_map_acquire_read(&pwq->wq->lockdep_map);
2188 lock_map_acquire(&lockdep_map);
2189 trace_workqueue_execute_start(work);
2190 worker->current_func(work);
2192 * While we must be careful to not use "work" after this, the trace
2193 * point will only record its address.
2195 trace_workqueue_execute_end(work);
2196 lock_map_release(&lockdep_map);
2197 lock_map_release(&pwq->wq->lockdep_map);
2199 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2200 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2201 " last function: %pf\n",
2202 current->comm, preempt_count(), task_pid_nr(current),
2203 worker->current_func);
2204 debug_show_held_locks(current);
2209 * The following prevents a kworker from hogging CPU on !PREEMPT
2210 * kernels, where a requeueing work item waiting for something to
2211 * happen could deadlock with stop_machine as such work item could
2212 * indefinitely requeue itself while all other CPUs are trapped in
2217 spin_lock_irq(&pool->lock);
2219 /* clear cpu intensive status */
2220 if (unlikely(cpu_intensive))
2221 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2223 /* we're done with it, release */
2224 hash_del(&worker->hentry);
2225 worker->current_work = NULL;
2226 worker->current_func = NULL;
2227 worker->current_pwq = NULL;
2228 worker->desc_valid = false;
2229 pwq_dec_nr_in_flight(pwq, work_color);
2233 * process_scheduled_works - process scheduled works
2236 * Process all scheduled works. Please note that the scheduled list
2237 * may change while processing a work, so this function repeatedly
2238 * fetches a work from the top and executes it.
2241 * spin_lock_irq(pool->lock) which may be released and regrabbed
2244 static void process_scheduled_works(struct worker *worker)
2246 while (!list_empty(&worker->scheduled)) {
2247 struct work_struct *work = list_first_entry(&worker->scheduled,
2248 struct work_struct, entry);
2249 process_one_work(worker, work);
2254 * worker_thread - the worker thread function
2257 * The worker thread function. All workers belong to a worker_pool -
2258 * either a per-cpu one or dynamic unbound one. These workers process all
2259 * work items regardless of their specific target workqueue. The only
2260 * exception is work items which belong to workqueues with a rescuer which
2261 * will be explained in rescuer_thread().
2263 static int worker_thread(void *__worker)
2265 struct worker *worker = __worker;
2266 struct worker_pool *pool = worker->pool;
2268 /* tell the scheduler that this is a workqueue worker */
2269 worker->task->flags |= PF_WQ_WORKER;
2271 spin_lock_irq(&pool->lock);
2273 /* am I supposed to die? */
2274 if (unlikely(worker->flags & WORKER_DIE)) {
2275 spin_unlock_irq(&pool->lock);
2276 WARN_ON_ONCE(!list_empty(&worker->entry));
2277 worker->task->flags &= ~PF_WQ_WORKER;
2281 worker_leave_idle(worker);
2283 /* no more worker necessary? */
2284 if (!need_more_worker(pool))
2287 /* do we need to manage? */
2288 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2292 * ->scheduled list can only be filled while a worker is
2293 * preparing to process a work or actually processing it.
2294 * Make sure nobody diddled with it while I was sleeping.
2296 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2299 * Finish PREP stage. We're guaranteed to have at least one idle
2300 * worker or that someone else has already assumed the manager
2301 * role. This is where @worker starts participating in concurrency
2302 * management if applicable and concurrency management is restored
2303 * after being rebound. See rebind_workers() for details.
2305 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2308 struct work_struct *work =
2309 list_first_entry(&pool->worklist,
2310 struct work_struct, entry);
2312 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2313 /* optimization path, not strictly necessary */
2314 process_one_work(worker, work);
2315 if (unlikely(!list_empty(&worker->scheduled)))
2316 process_scheduled_works(worker);
2318 move_linked_works(work, &worker->scheduled, NULL);
2319 process_scheduled_works(worker);
2321 } while (keep_working(pool));
2323 worker_set_flags(worker, WORKER_PREP, false);
2325 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2329 * pool->lock is held and there's no work to process and no need to
2330 * manage, sleep. Workers are woken up only while holding
2331 * pool->lock or from local cpu, so setting the current state
2332 * before releasing pool->lock is enough to prevent losing any
2335 worker_enter_idle(worker);
2336 __set_current_state(TASK_INTERRUPTIBLE);
2337 spin_unlock_irq(&pool->lock);
2343 * rescuer_thread - the rescuer thread function
2346 * Workqueue rescuer thread function. There's one rescuer for each
2347 * workqueue which has WQ_MEM_RECLAIM set.
2349 * Regular work processing on a pool may block trying to create a new
2350 * worker which uses GFP_KERNEL allocation which has slight chance of
2351 * developing into deadlock if some works currently on the same queue
2352 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2353 * the problem rescuer solves.
2355 * When such condition is possible, the pool summons rescuers of all
2356 * workqueues which have works queued on the pool and let them process
2357 * those works so that forward progress can be guaranteed.
2359 * This should happen rarely.
2361 static int rescuer_thread(void *__rescuer)
2363 struct worker *rescuer = __rescuer;
2364 struct workqueue_struct *wq = rescuer->rescue_wq;
2365 struct list_head *scheduled = &rescuer->scheduled;
2367 set_user_nice(current, RESCUER_NICE_LEVEL);
2370 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2371 * doesn't participate in concurrency management.
2373 rescuer->task->flags |= PF_WQ_WORKER;
2375 set_current_state(TASK_INTERRUPTIBLE);
2377 if (kthread_should_stop()) {
2378 __set_current_state(TASK_RUNNING);
2379 rescuer->task->flags &= ~PF_WQ_WORKER;
2383 /* see whether any pwq is asking for help */
2384 spin_lock_irq(&wq_mayday_lock);
2386 while (!list_empty(&wq->maydays)) {
2387 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2388 struct pool_workqueue, mayday_node);
2389 struct worker_pool *pool = pwq->pool;
2390 struct work_struct *work, *n;
2392 __set_current_state(TASK_RUNNING);
2393 list_del_init(&pwq->mayday_node);
2395 spin_unlock_irq(&wq_mayday_lock);
2397 /* migrate to the target cpu if possible */
2398 worker_maybe_bind_and_lock(pool);
2399 rescuer->pool = pool;
2402 * Slurp in all works issued via this workqueue and
2405 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2406 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2407 if (get_work_pwq(work) == pwq)
2408 move_linked_works(work, scheduled, &n);
2410 process_scheduled_works(rescuer);
2413 * Leave this pool. If keep_working() is %true, notify a
2414 * regular worker; otherwise, we end up with 0 concurrency
2415 * and stalling the execution.
2417 if (keep_working(pool))
2418 wake_up_worker(pool);
2420 rescuer->pool = NULL;
2421 spin_unlock(&pool->lock);
2422 spin_lock(&wq_mayday_lock);
2425 spin_unlock_irq(&wq_mayday_lock);
2427 /* rescuers should never participate in concurrency management */
2428 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2434 struct work_struct work;
2435 struct completion done;
2438 static void wq_barrier_func(struct work_struct *work)
2440 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2441 complete(&barr->done);
2445 * insert_wq_barrier - insert a barrier work
2446 * @pwq: pwq to insert barrier into
2447 * @barr: wq_barrier to insert
2448 * @target: target work to attach @barr to
2449 * @worker: worker currently executing @target, NULL if @target is not executing
2451 * @barr is linked to @target such that @barr is completed only after
2452 * @target finishes execution. Please note that the ordering
2453 * guarantee is observed only with respect to @target and on the local
2456 * Currently, a queued barrier can't be canceled. This is because
2457 * try_to_grab_pending() can't determine whether the work to be
2458 * grabbed is at the head of the queue and thus can't clear LINKED
2459 * flag of the previous work while there must be a valid next work
2460 * after a work with LINKED flag set.
2462 * Note that when @worker is non-NULL, @target may be modified
2463 * underneath us, so we can't reliably determine pwq from @target.
2466 * spin_lock_irq(pool->lock).
2468 static void insert_wq_barrier(struct pool_workqueue *pwq,
2469 struct wq_barrier *barr,
2470 struct work_struct *target, struct worker *worker)
2472 struct list_head *head;
2473 unsigned int linked = 0;
2476 * debugobject calls are safe here even with pool->lock locked
2477 * as we know for sure that this will not trigger any of the
2478 * checks and call back into the fixup functions where we
2481 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2482 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2483 init_completion(&barr->done);
2486 * If @target is currently being executed, schedule the
2487 * barrier to the worker; otherwise, put it after @target.
2490 head = worker->scheduled.next;
2492 unsigned long *bits = work_data_bits(target);
2494 head = target->entry.next;
2495 /* there can already be other linked works, inherit and set */
2496 linked = *bits & WORK_STRUCT_LINKED;
2497 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2500 debug_work_activate(&barr->work);
2501 insert_work(pwq, &barr->work, head,
2502 work_color_to_flags(WORK_NO_COLOR) | linked);
2506 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2507 * @wq: workqueue being flushed
2508 * @flush_color: new flush color, < 0 for no-op
2509 * @work_color: new work color, < 0 for no-op
2511 * Prepare pwqs for workqueue flushing.
2513 * If @flush_color is non-negative, flush_color on all pwqs should be
2514 * -1. If no pwq has in-flight commands at the specified color, all
2515 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2516 * has in flight commands, its pwq->flush_color is set to
2517 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2518 * wakeup logic is armed and %true is returned.
2520 * The caller should have initialized @wq->first_flusher prior to
2521 * calling this function with non-negative @flush_color. If
2522 * @flush_color is negative, no flush color update is done and %false
2525 * If @work_color is non-negative, all pwqs should have the same
2526 * work_color which is previous to @work_color and all will be
2527 * advanced to @work_color.
2530 * mutex_lock(wq->mutex).
2533 * %true if @flush_color >= 0 and there's something to flush. %false
2536 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2537 int flush_color, int work_color)
2540 struct pool_workqueue *pwq;
2542 if (flush_color >= 0) {
2543 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2544 atomic_set(&wq->nr_pwqs_to_flush, 1);
2547 for_each_pwq(pwq, wq) {
2548 struct worker_pool *pool = pwq->pool;
2550 spin_lock_irq(&pool->lock);
2552 if (flush_color >= 0) {
2553 WARN_ON_ONCE(pwq->flush_color != -1);
2555 if (pwq->nr_in_flight[flush_color]) {
2556 pwq->flush_color = flush_color;
2557 atomic_inc(&wq->nr_pwqs_to_flush);
2562 if (work_color >= 0) {
2563 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2564 pwq->work_color = work_color;
2567 spin_unlock_irq(&pool->lock);
2570 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2571 complete(&wq->first_flusher->done);
2577 * flush_workqueue - ensure that any scheduled work has run to completion.
2578 * @wq: workqueue to flush
2580 * This function sleeps until all work items which were queued on entry
2581 * have finished execution, but it is not livelocked by new incoming ones.
2583 void flush_workqueue(struct workqueue_struct *wq)
2585 struct wq_flusher this_flusher = {
2586 .list = LIST_HEAD_INIT(this_flusher.list),
2588 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2592 lock_map_acquire(&wq->lockdep_map);
2593 lock_map_release(&wq->lockdep_map);
2595 mutex_lock(&wq->mutex);
2598 * Start-to-wait phase
2600 next_color = work_next_color(wq->work_color);
2602 if (next_color != wq->flush_color) {
2604 * Color space is not full. The current work_color
2605 * becomes our flush_color and work_color is advanced
2608 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2609 this_flusher.flush_color = wq->work_color;
2610 wq->work_color = next_color;
2612 if (!wq->first_flusher) {
2613 /* no flush in progress, become the first flusher */
2614 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2616 wq->first_flusher = &this_flusher;
2618 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2620 /* nothing to flush, done */
2621 wq->flush_color = next_color;
2622 wq->first_flusher = NULL;
2627 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2628 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2629 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2633 * Oops, color space is full, wait on overflow queue.
2634 * The next flush completion will assign us
2635 * flush_color and transfer to flusher_queue.
2637 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2640 mutex_unlock(&wq->mutex);
2642 wait_for_completion(&this_flusher.done);
2645 * Wake-up-and-cascade phase
2647 * First flushers are responsible for cascading flushes and
2648 * handling overflow. Non-first flushers can simply return.
2650 if (wq->first_flusher != &this_flusher)
2653 mutex_lock(&wq->mutex);
2655 /* we might have raced, check again with mutex held */
2656 if (wq->first_flusher != &this_flusher)
2659 wq->first_flusher = NULL;
2661 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2662 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2665 struct wq_flusher *next, *tmp;
2667 /* complete all the flushers sharing the current flush color */
2668 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2669 if (next->flush_color != wq->flush_color)
2671 list_del_init(&next->list);
2672 complete(&next->done);
2675 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2676 wq->flush_color != work_next_color(wq->work_color));
2678 /* this flush_color is finished, advance by one */
2679 wq->flush_color = work_next_color(wq->flush_color);
2681 /* one color has been freed, handle overflow queue */
2682 if (!list_empty(&wq->flusher_overflow)) {
2684 * Assign the same color to all overflowed
2685 * flushers, advance work_color and append to
2686 * flusher_queue. This is the start-to-wait
2687 * phase for these overflowed flushers.
2689 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2690 tmp->flush_color = wq->work_color;
2692 wq->work_color = work_next_color(wq->work_color);
2694 list_splice_tail_init(&wq->flusher_overflow,
2695 &wq->flusher_queue);
2696 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2699 if (list_empty(&wq->flusher_queue)) {
2700 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2705 * Need to flush more colors. Make the next flusher
2706 * the new first flusher and arm pwqs.
2708 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2709 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2711 list_del_init(&next->list);
2712 wq->first_flusher = next;
2714 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2718 * Meh... this color is already done, clear first
2719 * flusher and repeat cascading.
2721 wq->first_flusher = NULL;
2725 mutex_unlock(&wq->mutex);
2727 EXPORT_SYMBOL_GPL(flush_workqueue);
2730 * drain_workqueue - drain a workqueue
2731 * @wq: workqueue to drain
2733 * Wait until the workqueue becomes empty. While draining is in progress,
2734 * only chain queueing is allowed. IOW, only currently pending or running
2735 * work items on @wq can queue further work items on it. @wq is flushed
2736 * repeatedly until it becomes empty. The number of flushing is detemined
2737 * by the depth of chaining and should be relatively short. Whine if it
2740 void drain_workqueue(struct workqueue_struct *wq)
2742 unsigned int flush_cnt = 0;
2743 struct pool_workqueue *pwq;
2746 * __queue_work() needs to test whether there are drainers, is much
2747 * hotter than drain_workqueue() and already looks at @wq->flags.
2748 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2750 mutex_lock(&wq->mutex);
2751 if (!wq->nr_drainers++)
2752 wq->flags |= __WQ_DRAINING;
2753 mutex_unlock(&wq->mutex);
2755 flush_workqueue(wq);
2757 mutex_lock(&wq->mutex);
2759 for_each_pwq(pwq, wq) {
2762 spin_lock_irq(&pwq->pool->lock);
2763 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2764 spin_unlock_irq(&pwq->pool->lock);
2769 if (++flush_cnt == 10 ||
2770 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2771 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2772 wq->name, flush_cnt);
2774 mutex_unlock(&wq->mutex);
2778 if (!--wq->nr_drainers)
2779 wq->flags &= ~__WQ_DRAINING;
2780 mutex_unlock(&wq->mutex);
2782 EXPORT_SYMBOL_GPL(drain_workqueue);
2784 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2786 struct worker *worker = NULL;
2787 struct worker_pool *pool;
2788 struct pool_workqueue *pwq;
2792 local_irq_disable();
2793 pool = get_work_pool(work);
2799 spin_lock(&pool->lock);
2800 /* see the comment in try_to_grab_pending() with the same code */
2801 pwq = get_work_pwq(work);
2803 if (unlikely(pwq->pool != pool))
2806 worker = find_worker_executing_work(pool, work);
2809 pwq = worker->current_pwq;
2812 insert_wq_barrier(pwq, barr, work, worker);
2813 spin_unlock_irq(&pool->lock);
2816 * If @max_active is 1 or rescuer is in use, flushing another work
2817 * item on the same workqueue may lead to deadlock. Make sure the
2818 * flusher is not running on the same workqueue by verifying write
2821 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2822 lock_map_acquire(&pwq->wq->lockdep_map);
2824 lock_map_acquire_read(&pwq->wq->lockdep_map);
2825 lock_map_release(&pwq->wq->lockdep_map);
2829 spin_unlock_irq(&pool->lock);
2833 static bool __flush_work(struct work_struct *work)
2835 struct wq_barrier barr;
2837 if (start_flush_work(work, &barr)) {
2838 wait_for_completion(&barr.done);
2839 destroy_work_on_stack(&barr.work);
2847 * flush_work - wait for a work to finish executing the last queueing instance
2848 * @work: the work to flush
2850 * Wait until @work has finished execution. @work is guaranteed to be idle
2851 * on return if it hasn't been requeued since flush started.
2854 * %true if flush_work() waited for the work to finish execution,
2855 * %false if it was already idle.
2857 bool flush_work(struct work_struct *work)
2859 lock_map_acquire(&work->lockdep_map);
2860 lock_map_release(&work->lockdep_map);
2862 return __flush_work(work);
2864 EXPORT_SYMBOL_GPL(flush_work);
2866 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2868 unsigned long flags;
2872 ret = try_to_grab_pending(work, is_dwork, &flags);
2874 * If someone else is canceling, wait for the same event it
2875 * would be waiting for before retrying.
2877 if (unlikely(ret == -ENOENT))
2879 } while (unlikely(ret < 0));
2881 /* tell other tasks trying to grab @work to back off */
2882 mark_work_canceling(work);
2883 local_irq_restore(flags);
2886 clear_work_data(work);
2891 * cancel_work_sync - cancel a work and wait for it to finish
2892 * @work: the work to cancel
2894 * Cancel @work and wait for its execution to finish. This function
2895 * can be used even if the work re-queues itself or migrates to
2896 * another workqueue. On return from this function, @work is
2897 * guaranteed to be not pending or executing on any CPU.
2899 * cancel_work_sync(&delayed_work->work) must not be used for
2900 * delayed_work's. Use cancel_delayed_work_sync() instead.
2902 * The caller must ensure that the workqueue on which @work was last
2903 * queued can't be destroyed before this function returns.
2906 * %true if @work was pending, %false otherwise.
2908 bool cancel_work_sync(struct work_struct *work)
2910 return __cancel_work_timer(work, false);
2912 EXPORT_SYMBOL_GPL(cancel_work_sync);
2915 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2916 * @dwork: the delayed work to flush
2918 * Delayed timer is cancelled and the pending work is queued for
2919 * immediate execution. Like flush_work(), this function only
2920 * considers the last queueing instance of @dwork.
2923 * %true if flush_work() waited for the work to finish execution,
2924 * %false if it was already idle.
2926 bool flush_delayed_work(struct delayed_work *dwork)
2928 local_irq_disable();
2929 if (del_timer_sync(&dwork->timer))
2930 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2932 return flush_work(&dwork->work);
2934 EXPORT_SYMBOL(flush_delayed_work);
2937 * cancel_delayed_work - cancel a delayed work
2938 * @dwork: delayed_work to cancel
2940 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2941 * and canceled; %false if wasn't pending. Note that the work callback
2942 * function may still be running on return, unless it returns %true and the
2943 * work doesn't re-arm itself. Explicitly flush or use
2944 * cancel_delayed_work_sync() to wait on it.
2946 * This function is safe to call from any context including IRQ handler.
2948 bool cancel_delayed_work(struct delayed_work *dwork)
2950 unsigned long flags;
2954 ret = try_to_grab_pending(&dwork->work, true, &flags);
2955 } while (unlikely(ret == -EAGAIN));
2957 if (unlikely(ret < 0))
2960 set_work_pool_and_clear_pending(&dwork->work,
2961 get_work_pool_id(&dwork->work));
2962 local_irq_restore(flags);
2965 EXPORT_SYMBOL(cancel_delayed_work);
2968 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2969 * @dwork: the delayed work cancel
2971 * This is cancel_work_sync() for delayed works.
2974 * %true if @dwork was pending, %false otherwise.
2976 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2978 return __cancel_work_timer(&dwork->work, true);
2980 EXPORT_SYMBOL(cancel_delayed_work_sync);
2983 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2984 * @func: the function to call
2986 * schedule_on_each_cpu() executes @func on each online CPU using the
2987 * system workqueue and blocks until all CPUs have completed.
2988 * schedule_on_each_cpu() is very slow.
2991 * 0 on success, -errno on failure.
2993 int schedule_on_each_cpu(work_func_t func)
2996 struct work_struct __percpu *works;
2998 works = alloc_percpu(struct work_struct);
3004 for_each_online_cpu(cpu) {
3005 struct work_struct *work = per_cpu_ptr(works, cpu);
3007 INIT_WORK(work, func);
3008 schedule_work_on(cpu, work);
3011 for_each_online_cpu(cpu)
3012 flush_work(per_cpu_ptr(works, cpu));
3020 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3022 * Forces execution of the kernel-global workqueue and blocks until its
3025 * Think twice before calling this function! It's very easy to get into
3026 * trouble if you don't take great care. Either of the following situations
3027 * will lead to deadlock:
3029 * One of the work items currently on the workqueue needs to acquire
3030 * a lock held by your code or its caller.
3032 * Your code is running in the context of a work routine.
3034 * They will be detected by lockdep when they occur, but the first might not
3035 * occur very often. It depends on what work items are on the workqueue and
3036 * what locks they need, which you have no control over.
3038 * In most situations flushing the entire workqueue is overkill; you merely
3039 * need to know that a particular work item isn't queued and isn't running.
3040 * In such cases you should use cancel_delayed_work_sync() or
3041 * cancel_work_sync() instead.
3043 void flush_scheduled_work(void)
3045 flush_workqueue(system_wq);
3047 EXPORT_SYMBOL(flush_scheduled_work);
3050 * execute_in_process_context - reliably execute the routine with user context
3051 * @fn: the function to execute
3052 * @ew: guaranteed storage for the execute work structure (must
3053 * be available when the work executes)
3055 * Executes the function immediately if process context is available,
3056 * otherwise schedules the function for delayed execution.
3058 * Returns: 0 - function was executed
3059 * 1 - function was scheduled for execution
3061 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3063 if (!in_interrupt()) {
3068 INIT_WORK(&ew->work, fn);
3069 schedule_work(&ew->work);
3073 EXPORT_SYMBOL_GPL(execute_in_process_context);
3077 * Workqueues with WQ_SYSFS flag set is visible to userland via
3078 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3079 * following attributes.
3081 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3082 * max_active RW int : maximum number of in-flight work items
3084 * Unbound workqueues have the following extra attributes.
3086 * id RO int : the associated pool ID
3087 * nice RW int : nice value of the workers
3088 * cpumask RW mask : bitmask of allowed CPUs for the workers
3091 struct workqueue_struct *wq;
3095 static struct workqueue_struct *dev_to_wq(struct device *dev)
3097 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3102 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3105 struct workqueue_struct *wq = dev_to_wq(dev);
3107 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3109 static DEVICE_ATTR_RO(per_cpu);
3111 static ssize_t max_active_show(struct device *dev,
3112 struct device_attribute *attr, char *buf)
3114 struct workqueue_struct *wq = dev_to_wq(dev);
3116 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3119 static ssize_t max_active_store(struct device *dev,
3120 struct device_attribute *attr, const char *buf,
3123 struct workqueue_struct *wq = dev_to_wq(dev);
3126 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3129 workqueue_set_max_active(wq, val);
3132 static DEVICE_ATTR_RW(max_active);
3134 static struct attribute *wq_sysfs_attrs[] = {
3135 &dev_attr_per_cpu.attr,
3136 &dev_attr_max_active.attr,
3139 ATTRIBUTE_GROUPS(wq_sysfs);
3141 static ssize_t wq_pool_ids_show(struct device *dev,
3142 struct device_attribute *attr, char *buf)
3144 struct workqueue_struct *wq = dev_to_wq(dev);
3145 const char *delim = "";
3146 int node, written = 0;
3148 rcu_read_lock_sched();
3149 for_each_node(node) {
3150 written += scnprintf(buf + written, PAGE_SIZE - written,
3151 "%s%d:%d", delim, node,
3152 unbound_pwq_by_node(wq, node)->pool->id);
3155 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3156 rcu_read_unlock_sched();
3161 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3164 struct workqueue_struct *wq = dev_to_wq(dev);
3167 mutex_lock(&wq->mutex);
3168 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3169 mutex_unlock(&wq->mutex);
3174 /* prepare workqueue_attrs for sysfs store operations */
3175 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3177 struct workqueue_attrs *attrs;
3179 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3183 mutex_lock(&wq->mutex);
3184 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3185 mutex_unlock(&wq->mutex);
3189 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3190 const char *buf, size_t count)
3192 struct workqueue_struct *wq = dev_to_wq(dev);
3193 struct workqueue_attrs *attrs;
3196 attrs = wq_sysfs_prep_attrs(wq);
3200 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3201 attrs->nice >= -20 && attrs->nice <= 19)
3202 ret = apply_workqueue_attrs(wq, attrs);
3206 free_workqueue_attrs(attrs);
3207 return ret ?: count;
3210 static ssize_t wq_cpumask_show(struct device *dev,
3211 struct device_attribute *attr, char *buf)
3213 struct workqueue_struct *wq = dev_to_wq(dev);
3216 mutex_lock(&wq->mutex);
3217 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3218 mutex_unlock(&wq->mutex);
3220 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3224 static ssize_t wq_cpumask_store(struct device *dev,
3225 struct device_attribute *attr,
3226 const char *buf, size_t count)
3228 struct workqueue_struct *wq = dev_to_wq(dev);
3229 struct workqueue_attrs *attrs;
3232 attrs = wq_sysfs_prep_attrs(wq);
3236 ret = cpumask_parse(buf, attrs->cpumask);
3238 ret = apply_workqueue_attrs(wq, attrs);
3240 free_workqueue_attrs(attrs);
3241 return ret ?: count;
3244 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3247 struct workqueue_struct *wq = dev_to_wq(dev);
3250 mutex_lock(&wq->mutex);
3251 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3252 !wq->unbound_attrs->no_numa);
3253 mutex_unlock(&wq->mutex);
3258 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3259 const char *buf, size_t count)
3261 struct workqueue_struct *wq = dev_to_wq(dev);
3262 struct workqueue_attrs *attrs;
3265 attrs = wq_sysfs_prep_attrs(wq);
3270 if (sscanf(buf, "%d", &v) == 1) {
3271 attrs->no_numa = !v;
3272 ret = apply_workqueue_attrs(wq, attrs);
3275 free_workqueue_attrs(attrs);
3276 return ret ?: count;
3279 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3280 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3281 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3282 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3283 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3287 static struct bus_type wq_subsys = {
3288 .name = "workqueue",
3289 .dev_groups = wq_sysfs_groups,
3292 static int __init wq_sysfs_init(void)
3294 return subsys_virtual_register(&wq_subsys, NULL);
3296 core_initcall(wq_sysfs_init);
3298 static void wq_device_release(struct device *dev)
3300 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3306 * workqueue_sysfs_register - make a workqueue visible in sysfs
3307 * @wq: the workqueue to register
3309 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3310 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3311 * which is the preferred method.
3313 * Workqueue user should use this function directly iff it wants to apply
3314 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3315 * apply_workqueue_attrs() may race against userland updating the
3318 * Returns 0 on success, -errno on failure.
3320 int workqueue_sysfs_register(struct workqueue_struct *wq)
3322 struct wq_device *wq_dev;
3326 * Adjusting max_active or creating new pwqs by applyting
3327 * attributes breaks ordering guarantee. Disallow exposing ordered
3330 if (WARN_ON(wq->flags & __WQ_ORDERED))
3333 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3338 wq_dev->dev.bus = &wq_subsys;
3339 wq_dev->dev.init_name = wq->name;
3340 wq_dev->dev.release = wq_device_release;
3343 * unbound_attrs are created separately. Suppress uevent until
3344 * everything is ready.
3346 dev_set_uevent_suppress(&wq_dev->dev, true);
3348 ret = device_register(&wq_dev->dev);
3355 if (wq->flags & WQ_UNBOUND) {
3356 struct device_attribute *attr;
3358 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3359 ret = device_create_file(&wq_dev->dev, attr);
3361 device_unregister(&wq_dev->dev);
3368 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3373 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3374 * @wq: the workqueue to unregister
3376 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3378 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3380 struct wq_device *wq_dev = wq->wq_dev;
3386 device_unregister(&wq_dev->dev);
3388 #else /* CONFIG_SYSFS */
3389 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3390 #endif /* CONFIG_SYSFS */
3393 * free_workqueue_attrs - free a workqueue_attrs
3394 * @attrs: workqueue_attrs to free
3396 * Undo alloc_workqueue_attrs().
3398 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3401 free_cpumask_var(attrs->cpumask);
3407 * alloc_workqueue_attrs - allocate a workqueue_attrs
3408 * @gfp_mask: allocation mask to use
3410 * Allocate a new workqueue_attrs, initialize with default settings and
3411 * return it. Returns NULL on failure.
3413 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3415 struct workqueue_attrs *attrs;
3417 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3420 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3423 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3426 free_workqueue_attrs(attrs);
3430 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3431 const struct workqueue_attrs *from)
3433 to->nice = from->nice;
3434 cpumask_copy(to->cpumask, from->cpumask);
3436 * Unlike hash and equality test, this function doesn't ignore
3437 * ->no_numa as it is used for both pool and wq attrs. Instead,
3438 * get_unbound_pool() explicitly clears ->no_numa after copying.
3440 to->no_numa = from->no_numa;
3443 /* hash value of the content of @attr */
3444 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3448 hash = jhash_1word(attrs->nice, hash);
3449 hash = jhash(cpumask_bits(attrs->cpumask),
3450 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3454 /* content equality test */
3455 static bool wqattrs_equal(const struct workqueue_attrs *a,
3456 const struct workqueue_attrs *b)
3458 if (a->nice != b->nice)
3460 if (!cpumask_equal(a->cpumask, b->cpumask))
3466 * init_worker_pool - initialize a newly zalloc'd worker_pool
3467 * @pool: worker_pool to initialize
3469 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3470 * Returns 0 on success, -errno on failure. Even on failure, all fields
3471 * inside @pool proper are initialized and put_unbound_pool() can be called
3472 * on @pool safely to release it.
3474 static int init_worker_pool(struct worker_pool *pool)
3476 spin_lock_init(&pool->lock);
3479 pool->node = NUMA_NO_NODE;
3480 pool->flags |= POOL_DISASSOCIATED;
3481 INIT_LIST_HEAD(&pool->worklist);
3482 INIT_LIST_HEAD(&pool->idle_list);
3483 hash_init(pool->busy_hash);
3485 init_timer_deferrable(&pool->idle_timer);
3486 pool->idle_timer.function = idle_worker_timeout;
3487 pool->idle_timer.data = (unsigned long)pool;
3489 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3490 (unsigned long)pool);
3492 mutex_init(&pool->manager_arb);
3493 mutex_init(&pool->manager_mutex);
3494 idr_init(&pool->worker_idr);
3496 INIT_HLIST_NODE(&pool->hash_node);
3499 /* shouldn't fail above this point */
3500 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3506 static void rcu_free_pool(struct rcu_head *rcu)
3508 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3510 idr_destroy(&pool->worker_idr);
3511 free_workqueue_attrs(pool->attrs);
3516 * put_unbound_pool - put a worker_pool
3517 * @pool: worker_pool to put
3519 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3520 * safe manner. get_unbound_pool() calls this function on its failure path
3521 * and this function should be able to release pools which went through,
3522 * successfully or not, init_worker_pool().
3524 * Should be called with wq_pool_mutex held.
3526 static void put_unbound_pool(struct worker_pool *pool)
3528 struct worker *worker;
3530 lockdep_assert_held(&wq_pool_mutex);
3536 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3537 WARN_ON(!list_empty(&pool->worklist)))
3540 /* release id and unhash */
3542 idr_remove(&worker_pool_idr, pool->id);
3543 hash_del(&pool->hash_node);
3546 * Become the manager and destroy all workers. Grabbing
3547 * manager_arb prevents @pool's workers from blocking on
3550 mutex_lock(&pool->manager_arb);
3551 mutex_lock(&pool->manager_mutex);
3552 spin_lock_irq(&pool->lock);
3554 while ((worker = first_worker(pool)))
3555 destroy_worker(worker);
3556 WARN_ON(pool->nr_workers || pool->nr_idle);
3558 spin_unlock_irq(&pool->lock);
3559 mutex_unlock(&pool->manager_mutex);
3560 mutex_unlock(&pool->manager_arb);
3562 /* shut down the timers */
3563 del_timer_sync(&pool->idle_timer);
3564 del_timer_sync(&pool->mayday_timer);
3566 /* sched-RCU protected to allow dereferences from get_work_pool() */
3567 call_rcu_sched(&pool->rcu, rcu_free_pool);
3571 * get_unbound_pool - get a worker_pool with the specified attributes
3572 * @attrs: the attributes of the worker_pool to get
3574 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3575 * reference count and return it. If there already is a matching
3576 * worker_pool, it will be used; otherwise, this function attempts to
3577 * create a new one. On failure, returns NULL.
3579 * Should be called with wq_pool_mutex held.
3581 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3583 u32 hash = wqattrs_hash(attrs);
3584 struct worker_pool *pool;
3587 lockdep_assert_held(&wq_pool_mutex);
3589 /* do we already have a matching pool? */
3590 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3591 if (wqattrs_equal(pool->attrs, attrs)) {
3597 /* nope, create a new one */
3598 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3599 if (!pool || init_worker_pool(pool) < 0)
3602 if (workqueue_freezing)
3603 pool->flags |= POOL_FREEZING;
3605 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3606 copy_workqueue_attrs(pool->attrs, attrs);
3609 * no_numa isn't a worker_pool attribute, always clear it. See
3610 * 'struct workqueue_attrs' comments for detail.
3612 pool->attrs->no_numa = false;
3614 /* if cpumask is contained inside a NUMA node, we belong to that node */
3615 if (wq_numa_enabled) {
3616 for_each_node(node) {
3617 if (cpumask_subset(pool->attrs->cpumask,
3618 wq_numa_possible_cpumask[node])) {
3625 if (worker_pool_assign_id(pool) < 0)
3628 /* create and start the initial worker */
3629 if (create_and_start_worker(pool) < 0)
3633 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3638 put_unbound_pool(pool);
3642 static void rcu_free_pwq(struct rcu_head *rcu)
3644 kmem_cache_free(pwq_cache,
3645 container_of(rcu, struct pool_workqueue, rcu));
3649 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3650 * and needs to be destroyed.
3652 static void pwq_unbound_release_workfn(struct work_struct *work)
3654 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3655 unbound_release_work);
3656 struct workqueue_struct *wq = pwq->wq;
3657 struct worker_pool *pool = pwq->pool;
3660 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3664 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3665 * necessary on release but do it anyway. It's easier to verify
3666 * and consistent with the linking path.
3668 mutex_lock(&wq->mutex);
3669 list_del_rcu(&pwq->pwqs_node);
3670 is_last = list_empty(&wq->pwqs);
3671 mutex_unlock(&wq->mutex);
3673 mutex_lock(&wq_pool_mutex);
3674 put_unbound_pool(pool);
3675 mutex_unlock(&wq_pool_mutex);
3677 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3680 * If we're the last pwq going away, @wq is already dead and no one
3681 * is gonna access it anymore. Free it.
3684 free_workqueue_attrs(wq->unbound_attrs);
3690 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3691 * @pwq: target pool_workqueue
3693 * If @pwq isn't freezing, set @pwq->max_active to the associated
3694 * workqueue's saved_max_active and activate delayed work items
3695 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3697 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3699 struct workqueue_struct *wq = pwq->wq;
3700 bool freezable = wq->flags & WQ_FREEZABLE;
3702 /* for @wq->saved_max_active */
3703 lockdep_assert_held(&wq->mutex);
3705 /* fast exit for non-freezable wqs */
3706 if (!freezable && pwq->max_active == wq->saved_max_active)
3709 spin_lock_irq(&pwq->pool->lock);
3711 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3712 pwq->max_active = wq->saved_max_active;
3714 while (!list_empty(&pwq->delayed_works) &&
3715 pwq->nr_active < pwq->max_active)
3716 pwq_activate_first_delayed(pwq);
3719 * Need to kick a worker after thawed or an unbound wq's
3720 * max_active is bumped. It's a slow path. Do it always.
3722 wake_up_worker(pwq->pool);
3724 pwq->max_active = 0;
3727 spin_unlock_irq(&pwq->pool->lock);
3730 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3731 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3732 struct worker_pool *pool)
3734 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3736 memset(pwq, 0, sizeof(*pwq));
3740 pwq->flush_color = -1;
3742 INIT_LIST_HEAD(&pwq->delayed_works);
3743 INIT_LIST_HEAD(&pwq->pwqs_node);
3744 INIT_LIST_HEAD(&pwq->mayday_node);
3745 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3748 /* sync @pwq with the current state of its associated wq and link it */
3749 static void link_pwq(struct pool_workqueue *pwq)
3751 struct workqueue_struct *wq = pwq->wq;
3753 lockdep_assert_held(&wq->mutex);
3755 /* may be called multiple times, ignore if already linked */
3756 if (!list_empty(&pwq->pwqs_node))
3760 * Set the matching work_color. This is synchronized with
3761 * wq->mutex to avoid confusing flush_workqueue().
3763 pwq->work_color = wq->work_color;
3765 /* sync max_active to the current setting */
3766 pwq_adjust_max_active(pwq);
3769 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3772 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3773 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3774 const struct workqueue_attrs *attrs)
3776 struct worker_pool *pool;
3777 struct pool_workqueue *pwq;
3779 lockdep_assert_held(&wq_pool_mutex);
3781 pool = get_unbound_pool(attrs);
3785 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3787 put_unbound_pool(pool);
3791 init_pwq(pwq, wq, pool);
3795 /* undo alloc_unbound_pwq(), used only in the error path */
3796 static void free_unbound_pwq(struct pool_workqueue *pwq)
3798 lockdep_assert_held(&wq_pool_mutex);
3801 put_unbound_pool(pwq->pool);
3802 kmem_cache_free(pwq_cache, pwq);
3807 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3808 * @attrs: the wq_attrs of interest
3809 * @node: the target NUMA node
3810 * @cpu_going_down: if >= 0, the CPU to consider as offline
3811 * @cpumask: outarg, the resulting cpumask
3813 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3814 * @cpu_going_down is >= 0, that cpu is considered offline during
3815 * calculation. The result is stored in @cpumask. This function returns
3816 * %true if the resulting @cpumask is different from @attrs->cpumask,
3819 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3820 * enabled and @node has online CPUs requested by @attrs, the returned
3821 * cpumask is the intersection of the possible CPUs of @node and
3824 * The caller is responsible for ensuring that the cpumask of @node stays
3827 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3828 int cpu_going_down, cpumask_t *cpumask)
3830 if (!wq_numa_enabled || attrs->no_numa)
3833 /* does @node have any online CPUs @attrs wants? */
3834 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3835 if (cpu_going_down >= 0)
3836 cpumask_clear_cpu(cpu_going_down, cpumask);
3838 if (cpumask_empty(cpumask))
3841 /* yeap, return possible CPUs in @node that @attrs wants */
3842 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3843 return !cpumask_equal(cpumask, attrs->cpumask);
3846 cpumask_copy(cpumask, attrs->cpumask);
3850 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3851 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3853 struct pool_workqueue *pwq)
3855 struct pool_workqueue *old_pwq;
3857 lockdep_assert_held(&wq->mutex);
3859 /* link_pwq() can handle duplicate calls */
3862 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3863 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3868 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3869 * @wq: the target workqueue
3870 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3872 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3873 * machines, this function maps a separate pwq to each NUMA node with
3874 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3875 * NUMA node it was issued on. Older pwqs are released as in-flight work
3876 * items finish. Note that a work item which repeatedly requeues itself
3877 * back-to-back will stay on its current pwq.
3879 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3882 int apply_workqueue_attrs(struct workqueue_struct *wq,
3883 const struct workqueue_attrs *attrs)
3885 struct workqueue_attrs *new_attrs, *tmp_attrs;
3886 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3889 /* only unbound workqueues can change attributes */
3890 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3893 /* creating multiple pwqs breaks ordering guarantee */
3894 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3897 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3898 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3899 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3900 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3903 /* make a copy of @attrs and sanitize it */
3904 copy_workqueue_attrs(new_attrs, attrs);
3905 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3908 * We may create multiple pwqs with differing cpumasks. Make a
3909 * copy of @new_attrs which will be modified and used to obtain
3912 copy_workqueue_attrs(tmp_attrs, new_attrs);
3915 * CPUs should stay stable across pwq creations and installations.
3916 * Pin CPUs, determine the target cpumask for each node and create
3921 mutex_lock(&wq_pool_mutex);
3924 * If something goes wrong during CPU up/down, we'll fall back to
3925 * the default pwq covering whole @attrs->cpumask. Always create
3926 * it even if we don't use it immediately.
3928 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3932 for_each_node(node) {
3933 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3934 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3939 pwq_tbl[node] = dfl_pwq;
3943 mutex_unlock(&wq_pool_mutex);
3945 /* all pwqs have been created successfully, let's install'em */
3946 mutex_lock(&wq->mutex);
3948 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3950 /* save the previous pwq and install the new one */
3952 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3954 /* @dfl_pwq might not have been used, ensure it's linked */
3956 swap(wq->dfl_pwq, dfl_pwq);
3958 mutex_unlock(&wq->mutex);
3960 /* put the old pwqs */
3962 put_pwq_unlocked(pwq_tbl[node]);
3963 put_pwq_unlocked(dfl_pwq);
3969 free_workqueue_attrs(tmp_attrs);
3970 free_workqueue_attrs(new_attrs);
3975 free_unbound_pwq(dfl_pwq);
3977 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3978 free_unbound_pwq(pwq_tbl[node]);
3979 mutex_unlock(&wq_pool_mutex);
3987 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3988 * @wq: the target workqueue
3989 * @cpu: the CPU coming up or going down
3990 * @online: whether @cpu is coming up or going down
3992 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3993 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3996 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3997 * falls back to @wq->dfl_pwq which may not be optimal but is always
4000 * Note that when the last allowed CPU of a NUMA node goes offline for a
4001 * workqueue with a cpumask spanning multiple nodes, the workers which were
4002 * already executing the work items for the workqueue will lose their CPU
4003 * affinity and may execute on any CPU. This is similar to how per-cpu
4004 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4005 * affinity, it's the user's responsibility to flush the work item from
4008 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4011 int node = cpu_to_node(cpu);
4012 int cpu_off = online ? -1 : cpu;
4013 struct pool_workqueue *old_pwq = NULL, *pwq;
4014 struct workqueue_attrs *target_attrs;
4017 lockdep_assert_held(&wq_pool_mutex);
4019 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4023 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4024 * Let's use a preallocated one. The following buf is protected by
4025 * CPU hotplug exclusion.
4027 target_attrs = wq_update_unbound_numa_attrs_buf;
4028 cpumask = target_attrs->cpumask;
4030 mutex_lock(&wq->mutex);
4031 if (wq->unbound_attrs->no_numa)
4034 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4035 pwq = unbound_pwq_by_node(wq, node);
4038 * Let's determine what needs to be done. If the target cpumask is
4039 * different from wq's, we need to compare it to @pwq's and create
4040 * a new one if they don't match. If the target cpumask equals
4041 * wq's, the default pwq should be used. If @pwq is already the
4042 * default one, nothing to do; otherwise, install the default one.
4044 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4045 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4048 if (pwq == wq->dfl_pwq)
4054 mutex_unlock(&wq->mutex);
4056 /* create a new pwq */
4057 pwq = alloc_unbound_pwq(wq, target_attrs);
4059 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4065 * Install the new pwq. As this function is called only from CPU
4066 * hotplug callbacks and applying a new attrs is wrapped with
4067 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4070 mutex_lock(&wq->mutex);
4071 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4075 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4076 get_pwq(wq->dfl_pwq);
4077 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4078 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4080 mutex_unlock(&wq->mutex);
4081 put_pwq_unlocked(old_pwq);
4084 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4086 bool highpri = wq->flags & WQ_HIGHPRI;
4089 if (!(wq->flags & WQ_UNBOUND)) {
4090 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4094 for_each_possible_cpu(cpu) {
4095 struct pool_workqueue *pwq =
4096 per_cpu_ptr(wq->cpu_pwqs, cpu);
4097 struct worker_pool *cpu_pools =
4098 per_cpu(cpu_worker_pools, cpu);
4100 init_pwq(pwq, wq, &cpu_pools[highpri]);
4102 mutex_lock(&wq->mutex);
4104 mutex_unlock(&wq->mutex);
4108 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4112 static int wq_clamp_max_active(int max_active, unsigned int flags,
4115 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4117 if (max_active < 1 || max_active > lim)
4118 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4119 max_active, name, 1, lim);
4121 return clamp_val(max_active, 1, lim);
4124 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4127 struct lock_class_key *key,
4128 const char *lock_name, ...)
4130 size_t tbl_size = 0;
4132 struct workqueue_struct *wq;
4133 struct pool_workqueue *pwq;
4135 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4136 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4137 flags |= WQ_UNBOUND;
4139 /* allocate wq and format name */
4140 if (flags & WQ_UNBOUND)
4141 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4143 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4147 if (flags & WQ_UNBOUND) {
4148 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4149 if (!wq->unbound_attrs)
4153 va_start(args, lock_name);
4154 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4157 max_active = max_active ?: WQ_DFL_ACTIVE;
4158 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4162 wq->saved_max_active = max_active;
4163 mutex_init(&wq->mutex);
4164 atomic_set(&wq->nr_pwqs_to_flush, 0);
4165 INIT_LIST_HEAD(&wq->pwqs);
4166 INIT_LIST_HEAD(&wq->flusher_queue);
4167 INIT_LIST_HEAD(&wq->flusher_overflow);
4168 INIT_LIST_HEAD(&wq->maydays);
4170 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4171 INIT_LIST_HEAD(&wq->list);
4173 if (alloc_and_link_pwqs(wq) < 0)
4177 * Workqueues which may be used during memory reclaim should
4178 * have a rescuer to guarantee forward progress.
4180 if (flags & WQ_MEM_RECLAIM) {
4181 struct worker *rescuer;
4183 rescuer = alloc_worker();
4187 rescuer->rescue_wq = wq;
4188 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4190 if (IS_ERR(rescuer->task)) {
4195 wq->rescuer = rescuer;
4196 rescuer->task->flags |= PF_NO_SETAFFINITY;
4197 wake_up_process(rescuer->task);
4200 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4204 * wq_pool_mutex protects global freeze state and workqueues list.
4205 * Grab it, adjust max_active and add the new @wq to workqueues
4208 mutex_lock(&wq_pool_mutex);
4210 mutex_lock(&wq->mutex);
4211 for_each_pwq(pwq, wq)
4212 pwq_adjust_max_active(pwq);
4213 mutex_unlock(&wq->mutex);
4215 list_add(&wq->list, &workqueues);
4217 mutex_unlock(&wq_pool_mutex);
4222 free_workqueue_attrs(wq->unbound_attrs);
4226 destroy_workqueue(wq);
4229 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4232 * destroy_workqueue - safely terminate a workqueue
4233 * @wq: target workqueue
4235 * Safely destroy a workqueue. All work currently pending will be done first.
4237 void destroy_workqueue(struct workqueue_struct *wq)
4239 struct pool_workqueue *pwq;
4242 /* drain it before proceeding with destruction */
4243 drain_workqueue(wq);
4246 mutex_lock(&wq->mutex);
4247 for_each_pwq(pwq, wq) {
4250 for (i = 0; i < WORK_NR_COLORS; i++) {
4251 if (WARN_ON(pwq->nr_in_flight[i])) {
4252 mutex_unlock(&wq->mutex);
4257 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4258 WARN_ON(pwq->nr_active) ||
4259 WARN_ON(!list_empty(&pwq->delayed_works))) {
4260 mutex_unlock(&wq->mutex);
4264 mutex_unlock(&wq->mutex);
4267 * wq list is used to freeze wq, remove from list after
4268 * flushing is complete in case freeze races us.
4270 mutex_lock(&wq_pool_mutex);
4271 list_del_init(&wq->list);
4272 mutex_unlock(&wq_pool_mutex);
4274 workqueue_sysfs_unregister(wq);
4277 kthread_stop(wq->rescuer->task);
4282 if (!(wq->flags & WQ_UNBOUND)) {
4284 * The base ref is never dropped on per-cpu pwqs. Directly
4285 * free the pwqs and wq.
4287 free_percpu(wq->cpu_pwqs);
4291 * We're the sole accessor of @wq at this point. Directly
4292 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4293 * @wq will be freed when the last pwq is released.
4295 for_each_node(node) {
4296 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4297 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4298 put_pwq_unlocked(pwq);
4302 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4303 * put. Don't access it afterwards.
4307 put_pwq_unlocked(pwq);
4310 EXPORT_SYMBOL_GPL(destroy_workqueue);
4313 * workqueue_set_max_active - adjust max_active of a workqueue
4314 * @wq: target workqueue
4315 * @max_active: new max_active value.
4317 * Set max_active of @wq to @max_active.
4320 * Don't call from IRQ context.
4322 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4324 struct pool_workqueue *pwq;
4326 /* disallow meddling with max_active for ordered workqueues */
4327 if (WARN_ON(wq->flags & __WQ_ORDERED))
4330 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4332 mutex_lock(&wq->mutex);
4334 wq->saved_max_active = max_active;
4336 for_each_pwq(pwq, wq)
4337 pwq_adjust_max_active(pwq);
4339 mutex_unlock(&wq->mutex);
4341 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4344 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4346 * Determine whether %current is a workqueue rescuer. Can be used from
4347 * work functions to determine whether it's being run off the rescuer task.
4349 bool current_is_workqueue_rescuer(void)
4351 struct worker *worker = current_wq_worker();
4353 return worker && worker->rescue_wq;
4357 * workqueue_congested - test whether a workqueue is congested
4358 * @cpu: CPU in question
4359 * @wq: target workqueue
4361 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4362 * no synchronization around this function and the test result is
4363 * unreliable and only useful as advisory hints or for debugging.
4365 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4366 * Note that both per-cpu and unbound workqueues may be associated with
4367 * multiple pool_workqueues which have separate congested states. A
4368 * workqueue being congested on one CPU doesn't mean the workqueue is also
4369 * contested on other CPUs / NUMA nodes.
4372 * %true if congested, %false otherwise.
4374 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4376 struct pool_workqueue *pwq;
4379 rcu_read_lock_sched();
4381 if (cpu == WORK_CPU_UNBOUND)
4382 cpu = smp_processor_id();
4384 if (!(wq->flags & WQ_UNBOUND))
4385 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4387 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4389 ret = !list_empty(&pwq->delayed_works);
4390 rcu_read_unlock_sched();
4394 EXPORT_SYMBOL_GPL(workqueue_congested);
4397 * work_busy - test whether a work is currently pending or running
4398 * @work: the work to be tested
4400 * Test whether @work is currently pending or running. There is no
4401 * synchronization around this function and the test result is
4402 * unreliable and only useful as advisory hints or for debugging.
4405 * OR'd bitmask of WORK_BUSY_* bits.
4407 unsigned int work_busy(struct work_struct *work)
4409 struct worker_pool *pool;
4410 unsigned long flags;
4411 unsigned int ret = 0;
4413 if (work_pending(work))
4414 ret |= WORK_BUSY_PENDING;
4416 local_irq_save(flags);
4417 pool = get_work_pool(work);
4419 spin_lock(&pool->lock);
4420 if (find_worker_executing_work(pool, work))
4421 ret |= WORK_BUSY_RUNNING;
4422 spin_unlock(&pool->lock);
4424 local_irq_restore(flags);
4428 EXPORT_SYMBOL_GPL(work_busy);
4431 * set_worker_desc - set description for the current work item
4432 * @fmt: printf-style format string
4433 * @...: arguments for the format string
4435 * This function can be called by a running work function to describe what
4436 * the work item is about. If the worker task gets dumped, this
4437 * information will be printed out together to help debugging. The
4438 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4440 void set_worker_desc(const char *fmt, ...)
4442 struct worker *worker = current_wq_worker();
4446 va_start(args, fmt);
4447 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4449 worker->desc_valid = true;
4454 * print_worker_info - print out worker information and description
4455 * @log_lvl: the log level to use when printing
4456 * @task: target task
4458 * If @task is a worker and currently executing a work item, print out the
4459 * name of the workqueue being serviced and worker description set with
4460 * set_worker_desc() by the currently executing work item.
4462 * This function can be safely called on any task as long as the
4463 * task_struct itself is accessible. While safe, this function isn't
4464 * synchronized and may print out mixups or garbages of limited length.
4466 void print_worker_info(const char *log_lvl, struct task_struct *task)
4468 work_func_t *fn = NULL;
4469 char name[WQ_NAME_LEN] = { };
4470 char desc[WORKER_DESC_LEN] = { };
4471 struct pool_workqueue *pwq = NULL;
4472 struct workqueue_struct *wq = NULL;
4473 bool desc_valid = false;
4474 struct worker *worker;
4476 if (!(task->flags & PF_WQ_WORKER))
4480 * This function is called without any synchronization and @task
4481 * could be in any state. Be careful with dereferences.
4483 worker = probe_kthread_data(task);
4486 * Carefully copy the associated workqueue's workfn and name. Keep
4487 * the original last '\0' in case the original contains garbage.
4489 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4490 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4491 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4492 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4494 /* copy worker description */
4495 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4497 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4499 if (fn || name[0] || desc[0]) {
4500 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4502 pr_cont(" (%s)", desc);
4510 * There are two challenges in supporting CPU hotplug. Firstly, there
4511 * are a lot of assumptions on strong associations among work, pwq and
4512 * pool which make migrating pending and scheduled works very
4513 * difficult to implement without impacting hot paths. Secondly,
4514 * worker pools serve mix of short, long and very long running works making
4515 * blocked draining impractical.
4517 * This is solved by allowing the pools to be disassociated from the CPU
4518 * running as an unbound one and allowing it to be reattached later if the
4519 * cpu comes back online.
4522 static void wq_unbind_fn(struct work_struct *work)
4524 int cpu = smp_processor_id();
4525 struct worker_pool *pool;
4526 struct worker *worker;
4529 for_each_cpu_worker_pool(pool, cpu) {
4530 WARN_ON_ONCE(cpu != smp_processor_id());
4532 mutex_lock(&pool->manager_mutex);
4533 spin_lock_irq(&pool->lock);
4536 * We've blocked all manager operations. Make all workers
4537 * unbound and set DISASSOCIATED. Before this, all workers
4538 * except for the ones which are still executing works from
4539 * before the last CPU down must be on the cpu. After
4540 * this, they may become diasporas.
4542 for_each_pool_worker(worker, wi, pool)
4543 worker->flags |= WORKER_UNBOUND;
4545 pool->flags |= POOL_DISASSOCIATED;
4547 spin_unlock_irq(&pool->lock);
4548 mutex_unlock(&pool->manager_mutex);
4551 * Call schedule() so that we cross rq->lock and thus can
4552 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4553 * This is necessary as scheduler callbacks may be invoked
4559 * Sched callbacks are disabled now. Zap nr_running.
4560 * After this, nr_running stays zero and need_more_worker()
4561 * and keep_working() are always true as long as the
4562 * worklist is not empty. This pool now behaves as an
4563 * unbound (in terms of concurrency management) pool which
4564 * are served by workers tied to the pool.
4566 atomic_set(&pool->nr_running, 0);
4569 * With concurrency management just turned off, a busy
4570 * worker blocking could lead to lengthy stalls. Kick off
4571 * unbound chain execution of currently pending work items.
4573 spin_lock_irq(&pool->lock);
4574 wake_up_worker(pool);
4575 spin_unlock_irq(&pool->lock);
4580 * rebind_workers - rebind all workers of a pool to the associated CPU
4581 * @pool: pool of interest
4583 * @pool->cpu is coming online. Rebind all workers to the CPU.
4585 static void rebind_workers(struct worker_pool *pool)
4587 struct worker *worker;
4590 lockdep_assert_held(&pool->manager_mutex);
4593 * Restore CPU affinity of all workers. As all idle workers should
4594 * be on the run-queue of the associated CPU before any local
4595 * wake-ups for concurrency management happen, restore CPU affinty
4596 * of all workers first and then clear UNBOUND. As we're called
4597 * from CPU_ONLINE, the following shouldn't fail.
4599 for_each_pool_worker(worker, wi, pool)
4600 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4601 pool->attrs->cpumask) < 0);
4603 spin_lock_irq(&pool->lock);
4605 for_each_pool_worker(worker, wi, pool) {
4606 unsigned int worker_flags = worker->flags;
4609 * A bound idle worker should actually be on the runqueue
4610 * of the associated CPU for local wake-ups targeting it to
4611 * work. Kick all idle workers so that they migrate to the
4612 * associated CPU. Doing this in the same loop as
4613 * replacing UNBOUND with REBOUND is safe as no worker will
4614 * be bound before @pool->lock is released.
4616 if (worker_flags & WORKER_IDLE)
4617 wake_up_process(worker->task);
4620 * We want to clear UNBOUND but can't directly call
4621 * worker_clr_flags() or adjust nr_running. Atomically
4622 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4623 * @worker will clear REBOUND using worker_clr_flags() when
4624 * it initiates the next execution cycle thus restoring
4625 * concurrency management. Note that when or whether
4626 * @worker clears REBOUND doesn't affect correctness.
4628 * ACCESS_ONCE() is necessary because @worker->flags may be
4629 * tested without holding any lock in
4630 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4631 * fail incorrectly leading to premature concurrency
4632 * management operations.
4634 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4635 worker_flags |= WORKER_REBOUND;
4636 worker_flags &= ~WORKER_UNBOUND;
4637 ACCESS_ONCE(worker->flags) = worker_flags;
4640 spin_unlock_irq(&pool->lock);
4644 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4645 * @pool: unbound pool of interest
4646 * @cpu: the CPU which is coming up
4648 * An unbound pool may end up with a cpumask which doesn't have any online
4649 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4650 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4651 * online CPU before, cpus_allowed of all its workers should be restored.
4653 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4655 static cpumask_t cpumask;
4656 struct worker *worker;
4659 lockdep_assert_held(&pool->manager_mutex);
4661 /* is @cpu allowed for @pool? */
4662 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4665 /* is @cpu the only online CPU? */
4666 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4667 if (cpumask_weight(&cpumask) != 1)
4670 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4671 for_each_pool_worker(worker, wi, pool)
4672 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4673 pool->attrs->cpumask) < 0);
4677 * Workqueues should be brought up before normal priority CPU notifiers.
4678 * This will be registered high priority CPU notifier.
4680 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4681 unsigned long action,
4684 int cpu = (unsigned long)hcpu;
4685 struct worker_pool *pool;
4686 struct workqueue_struct *wq;
4689 switch (action & ~CPU_TASKS_FROZEN) {
4690 case CPU_UP_PREPARE:
4691 for_each_cpu_worker_pool(pool, cpu) {
4692 if (pool->nr_workers)
4694 if (create_and_start_worker(pool) < 0)
4699 case CPU_DOWN_FAILED:
4701 mutex_lock(&wq_pool_mutex);
4703 for_each_pool(pool, pi) {
4704 mutex_lock(&pool->manager_mutex);
4706 if (pool->cpu == cpu) {
4707 spin_lock_irq(&pool->lock);
4708 pool->flags &= ~POOL_DISASSOCIATED;
4709 spin_unlock_irq(&pool->lock);
4711 rebind_workers(pool);
4712 } else if (pool->cpu < 0) {
4713 restore_unbound_workers_cpumask(pool, cpu);
4716 mutex_unlock(&pool->manager_mutex);
4719 /* update NUMA affinity of unbound workqueues */
4720 list_for_each_entry(wq, &workqueues, list)
4721 wq_update_unbound_numa(wq, cpu, true);
4723 mutex_unlock(&wq_pool_mutex);
4730 * Workqueues should be brought down after normal priority CPU notifiers.
4731 * This will be registered as low priority CPU notifier.
4733 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4734 unsigned long action,
4737 int cpu = (unsigned long)hcpu;
4738 struct work_struct unbind_work;
4739 struct workqueue_struct *wq;
4741 switch (action & ~CPU_TASKS_FROZEN) {
4742 case CPU_DOWN_PREPARE:
4743 /* unbinding per-cpu workers should happen on the local CPU */
4744 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4745 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4747 /* update NUMA affinity of unbound workqueues */
4748 mutex_lock(&wq_pool_mutex);
4749 list_for_each_entry(wq, &workqueues, list)
4750 wq_update_unbound_numa(wq, cpu, false);
4751 mutex_unlock(&wq_pool_mutex);
4753 /* wait for per-cpu unbinding to finish */
4754 flush_work(&unbind_work);
4762 struct work_for_cpu {
4763 struct work_struct work;
4769 static void work_for_cpu_fn(struct work_struct *work)
4771 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4773 wfc->ret = wfc->fn(wfc->arg);
4777 * work_on_cpu - run a function in user context on a particular cpu
4778 * @cpu: the cpu to run on
4779 * @fn: the function to run
4780 * @arg: the function arg
4782 * This will return the value @fn returns.
4783 * It is up to the caller to ensure that the cpu doesn't go offline.
4784 * The caller must not hold any locks which would prevent @fn from completing.
4786 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4788 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4790 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4791 schedule_work_on(cpu, &wfc.work);
4794 * The work item is on-stack and can't lead to deadlock through
4795 * flushing. Use __flush_work() to avoid spurious lockdep warnings
4796 * when work_on_cpu()s are nested.
4798 __flush_work(&wfc.work);
4802 EXPORT_SYMBOL_GPL(work_on_cpu);
4803 #endif /* CONFIG_SMP */
4805 #ifdef CONFIG_FREEZER
4808 * freeze_workqueues_begin - begin freezing workqueues
4810 * Start freezing workqueues. After this function returns, all freezable
4811 * workqueues will queue new works to their delayed_works list instead of
4815 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4817 void freeze_workqueues_begin(void)
4819 struct worker_pool *pool;
4820 struct workqueue_struct *wq;
4821 struct pool_workqueue *pwq;
4824 mutex_lock(&wq_pool_mutex);
4826 WARN_ON_ONCE(workqueue_freezing);
4827 workqueue_freezing = true;
4830 for_each_pool(pool, pi) {
4831 spin_lock_irq(&pool->lock);
4832 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4833 pool->flags |= POOL_FREEZING;
4834 spin_unlock_irq(&pool->lock);
4837 list_for_each_entry(wq, &workqueues, list) {
4838 mutex_lock(&wq->mutex);
4839 for_each_pwq(pwq, wq)
4840 pwq_adjust_max_active(pwq);
4841 mutex_unlock(&wq->mutex);
4844 mutex_unlock(&wq_pool_mutex);
4848 * freeze_workqueues_busy - are freezable workqueues still busy?
4850 * Check whether freezing is complete. This function must be called
4851 * between freeze_workqueues_begin() and thaw_workqueues().
4854 * Grabs and releases wq_pool_mutex.
4857 * %true if some freezable workqueues are still busy. %false if freezing
4860 bool freeze_workqueues_busy(void)
4863 struct workqueue_struct *wq;
4864 struct pool_workqueue *pwq;
4866 mutex_lock(&wq_pool_mutex);
4868 WARN_ON_ONCE(!workqueue_freezing);
4870 list_for_each_entry(wq, &workqueues, list) {
4871 if (!(wq->flags & WQ_FREEZABLE))
4874 * nr_active is monotonically decreasing. It's safe
4875 * to peek without lock.
4877 rcu_read_lock_sched();
4878 for_each_pwq(pwq, wq) {
4879 WARN_ON_ONCE(pwq->nr_active < 0);
4880 if (pwq->nr_active) {
4882 rcu_read_unlock_sched();
4886 rcu_read_unlock_sched();
4889 mutex_unlock(&wq_pool_mutex);
4894 * thaw_workqueues - thaw workqueues
4896 * Thaw workqueues. Normal queueing is restored and all collected
4897 * frozen works are transferred to their respective pool worklists.
4900 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4902 void thaw_workqueues(void)
4904 struct workqueue_struct *wq;
4905 struct pool_workqueue *pwq;
4906 struct worker_pool *pool;
4909 mutex_lock(&wq_pool_mutex);
4911 if (!workqueue_freezing)
4914 /* clear FREEZING */
4915 for_each_pool(pool, pi) {
4916 spin_lock_irq(&pool->lock);
4917 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4918 pool->flags &= ~POOL_FREEZING;
4919 spin_unlock_irq(&pool->lock);
4922 /* restore max_active and repopulate worklist */
4923 list_for_each_entry(wq, &workqueues, list) {
4924 mutex_lock(&wq->mutex);
4925 for_each_pwq(pwq, wq)
4926 pwq_adjust_max_active(pwq);
4927 mutex_unlock(&wq->mutex);
4930 workqueue_freezing = false;
4932 mutex_unlock(&wq_pool_mutex);
4934 #endif /* CONFIG_FREEZER */
4936 static void __init wq_numa_init(void)
4941 /* determine NUMA pwq table len - highest node id + 1 */
4943 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4945 if (num_possible_nodes() <= 1)
4948 if (wq_disable_numa) {
4949 pr_info("workqueue: NUMA affinity support disabled\n");
4953 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4954 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4957 * We want masks of possible CPUs of each node which isn't readily
4958 * available. Build one from cpu_to_node() which should have been
4959 * fully initialized by now.
4961 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4965 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4966 node_online(node) ? node : NUMA_NO_NODE));
4968 for_each_possible_cpu(cpu) {
4969 node = cpu_to_node(cpu);
4970 if (WARN_ON(node == NUMA_NO_NODE)) {
4971 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4972 /* happens iff arch is bonkers, let's just proceed */
4975 cpumask_set_cpu(cpu, tbl[node]);
4978 wq_numa_possible_cpumask = tbl;
4979 wq_numa_enabled = true;
4982 static int __init init_workqueues(void)
4984 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4987 /* make sure we have enough bits for OFFQ pool ID */
4988 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4989 WORK_CPU_END * NR_STD_WORKER_POOLS);
4991 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4993 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4995 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4996 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5000 /* initialize CPU pools */
5001 for_each_possible_cpu(cpu) {
5002 struct worker_pool *pool;
5005 for_each_cpu_worker_pool(pool, cpu) {
5006 BUG_ON(init_worker_pool(pool));
5008 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5009 pool->attrs->nice = std_nice[i++];
5010 pool->node = cpu_to_node(cpu);
5013 mutex_lock(&wq_pool_mutex);
5014 BUG_ON(worker_pool_assign_id(pool));
5015 mutex_unlock(&wq_pool_mutex);
5019 /* create the initial worker */
5020 for_each_online_cpu(cpu) {
5021 struct worker_pool *pool;
5023 for_each_cpu_worker_pool(pool, cpu) {
5024 pool->flags &= ~POOL_DISASSOCIATED;
5025 BUG_ON(create_and_start_worker(pool) < 0);
5029 /* create default unbound wq attrs */
5030 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5031 struct workqueue_attrs *attrs;
5033 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5034 attrs->nice = std_nice[i];
5035 unbound_std_wq_attrs[i] = attrs;
5038 system_wq = alloc_workqueue("events", 0, 0);
5039 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5040 system_long_wq = alloc_workqueue("events_long", 0, 0);
5041 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5042 WQ_UNBOUND_MAX_ACTIVE);
5043 system_freezable_wq = alloc_workqueue("events_freezable",
5045 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5046 WQ_POWER_EFFICIENT, 0);
5047 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5048 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5050 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5051 !system_unbound_wq || !system_freezable_wq ||
5052 !system_power_efficient_wq ||
5053 !system_freezable_power_efficient_wq);
5056 early_initcall(init_workqueues);