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 is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
51 #include "workqueue_internal.h"
57 * A bound pool is either associated or disassociated with its CPU.
58 * While associated (!DISASSOCIATED), all workers are bound to the
59 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * While DISASSOCIATED, the cpu may be offline and all workers have
63 * %WORKER_UNBOUND set and concurrency management disabled, and may
64 * be executing on any CPU. The pool behaves as an unbound one.
66 * Note that DISASSOCIATED should be flipped only while holding
67 * manager_mutex to avoid changing binding state while
68 * create_worker() is in progress.
70 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72 POOL_FREEZING = 1 << 3, /* freeze in progress */
75 WORKER_STARTED = 1 << 0, /* started */
76 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
78 WORKER_PREP = 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give -20.
104 RESCUER_NICE_LEVEL = -20,
105 HIGHPRI_NICE_LEVEL = -20,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * MG: pool->manager_mutex and pool->lock protected. Writes require both
127 * locks. Reads can happen under either lock.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * WQ: wq->mutex protected.
135 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
137 * MD: wq_mayday_lock protected.
140 /* struct worker is defined in workqueue_internal.h */
143 spinlock_t lock; /* the pool lock */
144 int cpu; /* I: the associated cpu */
145 int node; /* I: the associated node ID */
146 int id; /* I: pool ID */
147 unsigned int flags; /* X: flags */
149 struct list_head worklist; /* L: list of pending works */
150 int nr_workers; /* L: total number of workers */
152 /* nr_idle includes the ones off idle_list for rebinding */
153 int nr_idle; /* L: currently idle ones */
155 struct list_head idle_list; /* X: list of idle workers */
156 struct timer_list idle_timer; /* L: worker idle timeout */
157 struct timer_list mayday_timer; /* L: SOS timer for workers */
159 /* a workers is either on busy_hash or idle_list, or the manager */
160 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
161 /* L: hash of busy workers */
163 /* see manage_workers() for details on the two manager mutexes */
164 struct mutex manager_arb; /* manager arbitration */
165 struct mutex manager_mutex; /* manager exclusion */
166 struct idr worker_idr; /* MG: worker IDs and iteration */
168 struct workqueue_attrs *attrs; /* I: worker attributes */
169 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
170 int refcnt; /* PL: refcnt for unbound pools */
173 * The current concurrency level. As it's likely to be accessed
174 * from other CPUs during try_to_wake_up(), put it in a separate
177 atomic_t nr_running ____cacheline_aligned_in_smp;
180 * Destruction of pool is sched-RCU protected to allow dereferences
181 * from get_work_pool().
184 } ____cacheline_aligned_in_smp;
187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 * of work_struct->data are used for flags and the remaining high bits
189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 * number of flag bits.
192 struct pool_workqueue {
193 struct worker_pool *pool; /* I: the associated pool */
194 struct workqueue_struct *wq; /* I: the owning workqueue */
195 int work_color; /* L: current color */
196 int flush_color; /* L: flushing color */
197 int refcnt; /* L: reference count */
198 int nr_in_flight[WORK_NR_COLORS];
199 /* L: nr of in_flight works */
200 int nr_active; /* L: nr of active works */
201 int max_active; /* L: max active works */
202 struct list_head delayed_works; /* L: delayed works */
203 struct list_head pwqs_node; /* WR: node on wq->pwqs */
204 struct list_head mayday_node; /* MD: node on wq->maydays */
207 * Release of unbound pwq is punted to system_wq. See put_pwq()
208 * and pwq_unbound_release_workfn() for details. pool_workqueue
209 * itself is also sched-RCU protected so that the first pwq can be
210 * determined without grabbing wq->mutex.
212 struct work_struct unbound_release_work;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
217 * Structure used to wait for workqueue flush.
220 struct list_head list; /* WQ: list of flushers */
221 int flush_color; /* WQ: flush color waiting for */
222 struct completion done; /* flush completion */
228 * The externally visible workqueue. It relays the issued work items to
229 * the appropriate worker_pool through its pool_workqueues.
231 struct workqueue_struct {
232 struct list_head pwqs; /* WR: all pwqs of this wq */
233 struct list_head list; /* PL: list of all workqueues */
235 struct mutex mutex; /* protects this wq */
236 int work_color; /* WQ: current work color */
237 int flush_color; /* WQ: current flush color */
238 atomic_t nr_pwqs_to_flush; /* flush in progress */
239 struct wq_flusher *first_flusher; /* WQ: first flusher */
240 struct list_head flusher_queue; /* WQ: flush waiters */
241 struct list_head flusher_overflow; /* WQ: flush overflow list */
243 struct list_head maydays; /* MD: pwqs requesting rescue */
244 struct worker *rescuer; /* I: rescue worker */
246 int nr_drainers; /* WQ: drain in progress */
247 int saved_max_active; /* WQ: saved pwq max_active */
249 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
253 struct wq_device *wq_dev; /* I: for sysfs interface */
255 #ifdef CONFIG_LOCKDEP
256 struct lockdep_map lockdep_map;
258 char name[WQ_NAME_LEN]; /* I: workqueue name */
260 /* hot fields used during command issue, aligned to cacheline */
261 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
262 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
263 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
266 static struct kmem_cache *pwq_cache;
268 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
269 static cpumask_var_t *wq_numa_possible_cpumask;
270 /* possible CPUs of each node */
272 static bool wq_disable_numa;
273 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
275 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
277 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
278 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
280 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
281 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
283 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
284 static bool workqueue_freezing; /* PL: have wqs started freezing? */
286 /* the per-cpu worker pools */
287 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
290 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
292 /* PL: hash of all unbound pools keyed by pool->attrs */
293 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
295 /* I: attributes used when instantiating standard unbound pools on demand */
296 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
298 /* I: attributes used when instantiating ordered pools on demand */
299 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
301 struct workqueue_struct *system_wq __read_mostly;
302 EXPORT_SYMBOL(system_wq);
303 struct workqueue_struct *system_highpri_wq __read_mostly;
304 EXPORT_SYMBOL_GPL(system_highpri_wq);
305 struct workqueue_struct *system_long_wq __read_mostly;
306 EXPORT_SYMBOL_GPL(system_long_wq);
307 struct workqueue_struct *system_unbound_wq __read_mostly;
308 EXPORT_SYMBOL_GPL(system_unbound_wq);
309 struct workqueue_struct *system_freezable_wq __read_mostly;
310 EXPORT_SYMBOL_GPL(system_freezable_wq);
312 static int worker_thread(void *__worker);
313 static void copy_workqueue_attrs(struct workqueue_attrs *to,
314 const struct workqueue_attrs *from);
316 #define CREATE_TRACE_POINTS
317 #include <trace/events/workqueue.h>
319 #define assert_rcu_or_pool_mutex() \
320 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
321 lockdep_is_held(&wq_pool_mutex), \
322 "sched RCU or wq_pool_mutex should be held")
324 #define assert_rcu_or_wq_mutex(wq) \
325 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
326 lockdep_is_held(&wq->mutex), \
327 "sched RCU or wq->mutex should be held")
329 #ifdef CONFIG_LOCKDEP
330 #define assert_manager_or_pool_lock(pool) \
331 WARN_ONCE(debug_locks && \
332 !lockdep_is_held(&(pool)->manager_mutex) && \
333 !lockdep_is_held(&(pool)->lock), \
334 "pool->manager_mutex or ->lock should be held")
336 #define assert_manager_or_pool_lock(pool) do { } while (0)
339 #define for_each_cpu_worker_pool(pool, cpu) \
340 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
341 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
345 * for_each_pool - iterate through all worker_pools in the system
346 * @pool: iteration cursor
347 * @pi: integer used for iteration
349 * This must be called either with wq_pool_mutex held or sched RCU read
350 * locked. If the pool needs to be used beyond the locking in effect, the
351 * caller is responsible for guaranteeing that the pool stays online.
353 * The if/else clause exists only for the lockdep assertion and can be
356 #define for_each_pool(pool, pi) \
357 idr_for_each_entry(&worker_pool_idr, pool, pi) \
358 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
362 * for_each_pool_worker - iterate through all workers of a worker_pool
363 * @worker: iteration cursor
364 * @wi: integer used for iteration
365 * @pool: worker_pool to iterate workers of
367 * This must be called with either @pool->manager_mutex or ->lock held.
369 * The if/else clause exists only for the lockdep assertion and can be
372 #define for_each_pool_worker(worker, wi, pool) \
373 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
374 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
378 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
379 * @pwq: iteration cursor
380 * @wq: the target workqueue
382 * This must be called either with wq->mutex held or sched RCU read locked.
383 * If the pwq needs to be used beyond the locking in effect, the caller is
384 * responsible for guaranteeing that the pwq stays online.
386 * The if/else clause exists only for the lockdep assertion and can be
389 #define for_each_pwq(pwq, wq) \
390 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
391 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
394 #ifdef CONFIG_DEBUG_OBJECTS_WORK
396 static struct debug_obj_descr work_debug_descr;
398 static void *work_debug_hint(void *addr)
400 return ((struct work_struct *) addr)->func;
404 * fixup_init is called when:
405 * - an active object is initialized
407 static int work_fixup_init(void *addr, enum debug_obj_state state)
409 struct work_struct *work = addr;
412 case ODEBUG_STATE_ACTIVE:
413 cancel_work_sync(work);
414 debug_object_init(work, &work_debug_descr);
422 * fixup_activate is called when:
423 * - an active object is activated
424 * - an unknown object is activated (might be a statically initialized object)
426 static int work_fixup_activate(void *addr, enum debug_obj_state state)
428 struct work_struct *work = addr;
432 case ODEBUG_STATE_NOTAVAILABLE:
434 * This is not really a fixup. The work struct was
435 * statically initialized. We just make sure that it
436 * is tracked in the object tracker.
438 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
439 debug_object_init(work, &work_debug_descr);
440 debug_object_activate(work, &work_debug_descr);
446 case ODEBUG_STATE_ACTIVE:
455 * fixup_free is called when:
456 * - an active object is freed
458 static int work_fixup_free(void *addr, enum debug_obj_state state)
460 struct work_struct *work = addr;
463 case ODEBUG_STATE_ACTIVE:
464 cancel_work_sync(work);
465 debug_object_free(work, &work_debug_descr);
472 static struct debug_obj_descr work_debug_descr = {
473 .name = "work_struct",
474 .debug_hint = work_debug_hint,
475 .fixup_init = work_fixup_init,
476 .fixup_activate = work_fixup_activate,
477 .fixup_free = work_fixup_free,
480 static inline void debug_work_activate(struct work_struct *work)
482 debug_object_activate(work, &work_debug_descr);
485 static inline void debug_work_deactivate(struct work_struct *work)
487 debug_object_deactivate(work, &work_debug_descr);
490 void __init_work(struct work_struct *work, int onstack)
493 debug_object_init_on_stack(work, &work_debug_descr);
495 debug_object_init(work, &work_debug_descr);
497 EXPORT_SYMBOL_GPL(__init_work);
499 void destroy_work_on_stack(struct work_struct *work)
501 debug_object_free(work, &work_debug_descr);
503 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
506 static inline void debug_work_activate(struct work_struct *work) { }
507 static inline void debug_work_deactivate(struct work_struct *work) { }
510 /* allocate ID and assign it to @pool */
511 static int worker_pool_assign_id(struct worker_pool *pool)
515 lockdep_assert_held(&wq_pool_mutex);
517 ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
526 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
527 * @wq: the target workqueue
530 * This must be called either with pwq_lock held or sched RCU read locked.
531 * If the pwq needs to be used beyond the locking in effect, the caller is
532 * responsible for guaranteeing that the pwq stays online.
534 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
537 assert_rcu_or_wq_mutex(wq);
538 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
541 static unsigned int work_color_to_flags(int color)
543 return color << WORK_STRUCT_COLOR_SHIFT;
546 static int get_work_color(struct work_struct *work)
548 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
549 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
552 static int work_next_color(int color)
554 return (color + 1) % WORK_NR_COLORS;
558 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
559 * contain the pointer to the queued pwq. Once execution starts, the flag
560 * is cleared and the high bits contain OFFQ flags and pool ID.
562 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
563 * and clear_work_data() can be used to set the pwq, pool or clear
564 * work->data. These functions should only be called while the work is
565 * owned - ie. while the PENDING bit is set.
567 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
568 * corresponding to a work. Pool is available once the work has been
569 * queued anywhere after initialization until it is sync canceled. pwq is
570 * available only while the work item is queued.
572 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
573 * canceled. While being canceled, a work item may have its PENDING set
574 * but stay off timer and worklist for arbitrarily long and nobody should
575 * try to steal the PENDING bit.
577 static inline void set_work_data(struct work_struct *work, unsigned long data,
580 WARN_ON_ONCE(!work_pending(work));
581 atomic_long_set(&work->data, data | flags | work_static(work));
584 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
585 unsigned long extra_flags)
587 set_work_data(work, (unsigned long)pwq,
588 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
591 static void set_work_pool_and_keep_pending(struct work_struct *work,
594 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
595 WORK_STRUCT_PENDING);
598 static void set_work_pool_and_clear_pending(struct work_struct *work,
602 * The following wmb is paired with the implied mb in
603 * test_and_set_bit(PENDING) and ensures all updates to @work made
604 * here are visible to and precede any updates by the next PENDING
608 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
611 static void clear_work_data(struct work_struct *work)
613 smp_wmb(); /* see set_work_pool_and_clear_pending() */
614 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
617 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
619 unsigned long data = atomic_long_read(&work->data);
621 if (data & WORK_STRUCT_PWQ)
622 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
628 * get_work_pool - return the worker_pool a given work was associated with
629 * @work: the work item of interest
631 * Return the worker_pool @work was last associated with. %NULL if none.
633 * Pools are created and destroyed under wq_pool_mutex, and allows read
634 * access under sched-RCU read lock. As such, this function should be
635 * called under wq_pool_mutex or with preemption disabled.
637 * All fields of the returned pool are accessible as long as the above
638 * mentioned locking is in effect. If the returned pool needs to be used
639 * beyond the critical section, the caller is responsible for ensuring the
640 * returned pool is and stays online.
642 static struct worker_pool *get_work_pool(struct work_struct *work)
644 unsigned long data = atomic_long_read(&work->data);
647 assert_rcu_or_pool_mutex();
649 if (data & WORK_STRUCT_PWQ)
650 return ((struct pool_workqueue *)
651 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
653 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
654 if (pool_id == WORK_OFFQ_POOL_NONE)
657 return idr_find(&worker_pool_idr, pool_id);
661 * get_work_pool_id - return the worker pool ID a given work is associated with
662 * @work: the work item of interest
664 * Return the worker_pool ID @work was last associated with.
665 * %WORK_OFFQ_POOL_NONE if none.
667 static int get_work_pool_id(struct work_struct *work)
669 unsigned long data = atomic_long_read(&work->data);
671 if (data & WORK_STRUCT_PWQ)
672 return ((struct pool_workqueue *)
673 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
675 return data >> WORK_OFFQ_POOL_SHIFT;
678 static void mark_work_canceling(struct work_struct *work)
680 unsigned long pool_id = get_work_pool_id(work);
682 pool_id <<= WORK_OFFQ_POOL_SHIFT;
683 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
686 static bool work_is_canceling(struct work_struct *work)
688 unsigned long data = atomic_long_read(&work->data);
690 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
694 * Policy functions. These define the policies on how the global worker
695 * pools are managed. Unless noted otherwise, these functions assume that
696 * they're being called with pool->lock held.
699 static bool __need_more_worker(struct worker_pool *pool)
701 return !atomic_read(&pool->nr_running);
705 * Need to wake up a worker? Called from anything but currently
708 * Note that, because unbound workers never contribute to nr_running, this
709 * function will always return %true for unbound pools as long as the
710 * worklist isn't empty.
712 static bool need_more_worker(struct worker_pool *pool)
714 return !list_empty(&pool->worklist) && __need_more_worker(pool);
717 /* Can I start working? Called from busy but !running workers. */
718 static bool may_start_working(struct worker_pool *pool)
720 return pool->nr_idle;
723 /* Do I need to keep working? Called from currently running workers. */
724 static bool keep_working(struct worker_pool *pool)
726 return !list_empty(&pool->worklist) &&
727 atomic_read(&pool->nr_running) <= 1;
730 /* Do we need a new worker? Called from manager. */
731 static bool need_to_create_worker(struct worker_pool *pool)
733 return need_more_worker(pool) && !may_start_working(pool);
736 /* Do I need to be the manager? */
737 static bool need_to_manage_workers(struct worker_pool *pool)
739 return need_to_create_worker(pool) ||
740 (pool->flags & POOL_MANAGE_WORKERS);
743 /* Do we have too many workers and should some go away? */
744 static bool too_many_workers(struct worker_pool *pool)
746 bool managing = mutex_is_locked(&pool->manager_arb);
747 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
748 int nr_busy = pool->nr_workers - nr_idle;
751 * nr_idle and idle_list may disagree if idle rebinding is in
752 * progress. Never return %true if idle_list is empty.
754 if (list_empty(&pool->idle_list))
757 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
764 /* Return the first worker. Safe with preemption disabled */
765 static struct worker *first_worker(struct worker_pool *pool)
767 if (unlikely(list_empty(&pool->idle_list)))
770 return list_first_entry(&pool->idle_list, struct worker, entry);
774 * wake_up_worker - wake up an idle worker
775 * @pool: worker pool to wake worker from
777 * Wake up the first idle worker of @pool.
780 * spin_lock_irq(pool->lock).
782 static void wake_up_worker(struct worker_pool *pool)
784 struct worker *worker = first_worker(pool);
787 wake_up_process(worker->task);
791 * wq_worker_waking_up - a worker is waking up
792 * @task: task waking up
793 * @cpu: CPU @task is waking up to
795 * This function is called during try_to_wake_up() when a worker is
799 * spin_lock_irq(rq->lock)
801 void wq_worker_waking_up(struct task_struct *task, int cpu)
803 struct worker *worker = kthread_data(task);
805 if (!(worker->flags & WORKER_NOT_RUNNING)) {
806 WARN_ON_ONCE(worker->pool->cpu != cpu);
807 atomic_inc(&worker->pool->nr_running);
812 * wq_worker_sleeping - a worker is going to sleep
813 * @task: task going to sleep
814 * @cpu: CPU in question, must be the current CPU number
816 * This function is called during schedule() when a busy worker is
817 * going to sleep. Worker on the same cpu can be woken up by
818 * returning pointer to its task.
821 * spin_lock_irq(rq->lock)
824 * Worker task on @cpu to wake up, %NULL if none.
826 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
828 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
829 struct worker_pool *pool;
832 * Rescuers, which may not have all the fields set up like normal
833 * workers, also reach here, let's not access anything before
834 * checking NOT_RUNNING.
836 if (worker->flags & WORKER_NOT_RUNNING)
841 /* this can only happen on the local cpu */
842 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
846 * The counterpart of the following dec_and_test, implied mb,
847 * worklist not empty test sequence is in insert_work().
848 * Please read comment there.
850 * NOT_RUNNING is clear. This means that we're bound to and
851 * running on the local cpu w/ rq lock held and preemption
852 * disabled, which in turn means that none else could be
853 * manipulating idle_list, so dereferencing idle_list without pool
856 if (atomic_dec_and_test(&pool->nr_running) &&
857 !list_empty(&pool->worklist))
858 to_wakeup = first_worker(pool);
859 return to_wakeup ? to_wakeup->task : NULL;
863 * worker_set_flags - set worker flags and adjust nr_running accordingly
865 * @flags: flags to set
866 * @wakeup: wakeup an idle worker if necessary
868 * Set @flags in @worker->flags and adjust nr_running accordingly. If
869 * nr_running becomes zero and @wakeup is %true, an idle worker is
873 * spin_lock_irq(pool->lock)
875 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
878 struct worker_pool *pool = worker->pool;
880 WARN_ON_ONCE(worker->task != current);
883 * If transitioning into NOT_RUNNING, adjust nr_running and
884 * wake up an idle worker as necessary if requested by
887 if ((flags & WORKER_NOT_RUNNING) &&
888 !(worker->flags & WORKER_NOT_RUNNING)) {
890 if (atomic_dec_and_test(&pool->nr_running) &&
891 !list_empty(&pool->worklist))
892 wake_up_worker(pool);
894 atomic_dec(&pool->nr_running);
897 worker->flags |= flags;
901 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
903 * @flags: flags to clear
905 * Clear @flags in @worker->flags and adjust nr_running accordingly.
908 * spin_lock_irq(pool->lock)
910 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
912 struct worker_pool *pool = worker->pool;
913 unsigned int oflags = worker->flags;
915 WARN_ON_ONCE(worker->task != current);
917 worker->flags &= ~flags;
920 * If transitioning out of NOT_RUNNING, increment nr_running. Note
921 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
922 * of multiple flags, not a single flag.
924 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
925 if (!(worker->flags & WORKER_NOT_RUNNING))
926 atomic_inc(&pool->nr_running);
930 * find_worker_executing_work - find worker which is executing a work
931 * @pool: pool of interest
932 * @work: work to find worker for
934 * Find a worker which is executing @work on @pool by searching
935 * @pool->busy_hash which is keyed by the address of @work. For a worker
936 * to match, its current execution should match the address of @work and
937 * its work function. This is to avoid unwanted dependency between
938 * unrelated work executions through a work item being recycled while still
941 * This is a bit tricky. A work item may be freed once its execution
942 * starts and nothing prevents the freed area from being recycled for
943 * another work item. If the same work item address ends up being reused
944 * before the original execution finishes, workqueue will identify the
945 * recycled work item as currently executing and make it wait until the
946 * current execution finishes, introducing an unwanted dependency.
948 * This function checks the work item address and work function to avoid
949 * false positives. Note that this isn't complete as one may construct a
950 * work function which can introduce dependency onto itself through a
951 * recycled work item. Well, if somebody wants to shoot oneself in the
952 * foot that badly, there's only so much we can do, and if such deadlock
953 * actually occurs, it should be easy to locate the culprit work function.
956 * spin_lock_irq(pool->lock).
959 * Pointer to worker which is executing @work if found, NULL
962 static struct worker *find_worker_executing_work(struct worker_pool *pool,
963 struct work_struct *work)
965 struct worker *worker;
967 hash_for_each_possible(pool->busy_hash, worker, hentry,
969 if (worker->current_work == work &&
970 worker->current_func == work->func)
977 * move_linked_works - move linked works to a list
978 * @work: start of series of works to be scheduled
979 * @head: target list to append @work to
980 * @nextp: out paramter for nested worklist walking
982 * Schedule linked works starting from @work to @head. Work series to
983 * be scheduled starts at @work and includes any consecutive work with
984 * WORK_STRUCT_LINKED set in its predecessor.
986 * If @nextp is not NULL, it's updated to point to the next work of
987 * the last scheduled work. This allows move_linked_works() to be
988 * nested inside outer list_for_each_entry_safe().
991 * spin_lock_irq(pool->lock).
993 static void move_linked_works(struct work_struct *work, struct list_head *head,
994 struct work_struct **nextp)
996 struct work_struct *n;
999 * Linked worklist will always end before the end of the list,
1000 * use NULL for list head.
1002 list_for_each_entry_safe_from(work, n, NULL, entry) {
1003 list_move_tail(&work->entry, head);
1004 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1009 * If we're already inside safe list traversal and have moved
1010 * multiple works to the scheduled queue, the next position
1011 * needs to be updated.
1018 * get_pwq - get an extra reference on the specified pool_workqueue
1019 * @pwq: pool_workqueue to get
1021 * Obtain an extra reference on @pwq. The caller should guarantee that
1022 * @pwq has positive refcnt and be holding the matching pool->lock.
1024 static void get_pwq(struct pool_workqueue *pwq)
1026 lockdep_assert_held(&pwq->pool->lock);
1027 WARN_ON_ONCE(pwq->refcnt <= 0);
1032 * put_pwq - put a pool_workqueue reference
1033 * @pwq: pool_workqueue to put
1035 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1036 * destruction. The caller should be holding the matching pool->lock.
1038 static void put_pwq(struct pool_workqueue *pwq)
1040 lockdep_assert_held(&pwq->pool->lock);
1041 if (likely(--pwq->refcnt))
1043 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1046 * @pwq can't be released under pool->lock, bounce to
1047 * pwq_unbound_release_workfn(). This never recurses on the same
1048 * pool->lock as this path is taken only for unbound workqueues and
1049 * the release work item is scheduled on a per-cpu workqueue. To
1050 * avoid lockdep warning, unbound pool->locks are given lockdep
1051 * subclass of 1 in get_unbound_pool().
1053 schedule_work(&pwq->unbound_release_work);
1057 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1058 * @pwq: pool_workqueue to put (can be %NULL)
1060 * put_pwq() with locking. This function also allows %NULL @pwq.
1062 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1066 * As both pwqs and pools are sched-RCU protected, the
1067 * following lock operations are safe.
1069 spin_lock_irq(&pwq->pool->lock);
1071 spin_unlock_irq(&pwq->pool->lock);
1075 static void pwq_activate_delayed_work(struct work_struct *work)
1077 struct pool_workqueue *pwq = get_work_pwq(work);
1079 trace_workqueue_activate_work(work);
1080 move_linked_works(work, &pwq->pool->worklist, NULL);
1081 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1085 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1087 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1088 struct work_struct, entry);
1090 pwq_activate_delayed_work(work);
1094 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1095 * @pwq: pwq of interest
1096 * @color: color of work which left the queue
1098 * A work either has completed or is removed from pending queue,
1099 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1102 * spin_lock_irq(pool->lock).
1104 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1106 /* uncolored work items don't participate in flushing or nr_active */
1107 if (color == WORK_NO_COLOR)
1110 pwq->nr_in_flight[color]--;
1113 if (!list_empty(&pwq->delayed_works)) {
1114 /* one down, submit a delayed one */
1115 if (pwq->nr_active < pwq->max_active)
1116 pwq_activate_first_delayed(pwq);
1119 /* is flush in progress and are we at the flushing tip? */
1120 if (likely(pwq->flush_color != color))
1123 /* are there still in-flight works? */
1124 if (pwq->nr_in_flight[color])
1127 /* this pwq is done, clear flush_color */
1128 pwq->flush_color = -1;
1131 * If this was the last pwq, wake up the first flusher. It
1132 * will handle the rest.
1134 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1135 complete(&pwq->wq->first_flusher->done);
1141 * try_to_grab_pending - steal work item from worklist and disable irq
1142 * @work: work item to steal
1143 * @is_dwork: @work is a delayed_work
1144 * @flags: place to store irq state
1146 * Try to grab PENDING bit of @work. This function can handle @work in any
1147 * stable state - idle, on timer or on worklist. Return values are
1149 * 1 if @work was pending and we successfully stole PENDING
1150 * 0 if @work was idle and we claimed PENDING
1151 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1152 * -ENOENT if someone else is canceling @work, this state may persist
1153 * for arbitrarily long
1155 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1156 * interrupted while holding PENDING and @work off queue, irq must be
1157 * disabled on entry. This, combined with delayed_work->timer being
1158 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1160 * On successful return, >= 0, irq is disabled and the caller is
1161 * responsible for releasing it using local_irq_restore(*@flags).
1163 * This function is safe to call from any context including IRQ handler.
1165 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1166 unsigned long *flags)
1168 struct worker_pool *pool;
1169 struct pool_workqueue *pwq;
1171 local_irq_save(*flags);
1173 /* try to steal the timer if it exists */
1175 struct delayed_work *dwork = to_delayed_work(work);
1178 * dwork->timer is irqsafe. If del_timer() fails, it's
1179 * guaranteed that the timer is not queued anywhere and not
1180 * running on the local CPU.
1182 if (likely(del_timer(&dwork->timer)))
1186 /* try to claim PENDING the normal way */
1187 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1191 * The queueing is in progress, or it is already queued. Try to
1192 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1194 pool = get_work_pool(work);
1198 spin_lock(&pool->lock);
1200 * work->data is guaranteed to point to pwq only while the work
1201 * item is queued on pwq->wq, and both updating work->data to point
1202 * to pwq on queueing and to pool on dequeueing are done under
1203 * pwq->pool->lock. This in turn guarantees that, if work->data
1204 * points to pwq which is associated with a locked pool, the work
1205 * item is currently queued on that pool.
1207 pwq = get_work_pwq(work);
1208 if (pwq && pwq->pool == pool) {
1209 debug_work_deactivate(work);
1212 * A delayed work item cannot be grabbed directly because
1213 * it might have linked NO_COLOR work items which, if left
1214 * on the delayed_list, will confuse pwq->nr_active
1215 * management later on and cause stall. Make sure the work
1216 * item is activated before grabbing.
1218 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1219 pwq_activate_delayed_work(work);
1221 list_del_init(&work->entry);
1222 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1224 /* work->data points to pwq iff queued, point to pool */
1225 set_work_pool_and_keep_pending(work, pool->id);
1227 spin_unlock(&pool->lock);
1230 spin_unlock(&pool->lock);
1232 local_irq_restore(*flags);
1233 if (work_is_canceling(work))
1240 * insert_work - insert a work into a pool
1241 * @pwq: pwq @work belongs to
1242 * @work: work to insert
1243 * @head: insertion point
1244 * @extra_flags: extra WORK_STRUCT_* flags to set
1246 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1247 * work_struct flags.
1250 * spin_lock_irq(pool->lock).
1252 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1253 struct list_head *head, unsigned int extra_flags)
1255 struct worker_pool *pool = pwq->pool;
1257 /* we own @work, set data and link */
1258 set_work_pwq(work, pwq, extra_flags);
1259 list_add_tail(&work->entry, head);
1263 * Ensure either wq_worker_sleeping() sees the above
1264 * list_add_tail() or we see zero nr_running to avoid workers lying
1265 * around lazily while there are works to be processed.
1269 if (__need_more_worker(pool))
1270 wake_up_worker(pool);
1274 * Test whether @work is being queued from another work executing on the
1277 static bool is_chained_work(struct workqueue_struct *wq)
1279 struct worker *worker;
1281 worker = current_wq_worker();
1283 * Return %true iff I'm a worker execuing a work item on @wq. If
1284 * I'm @worker, it's safe to dereference it without locking.
1286 return worker && worker->current_pwq->wq == wq;
1289 static void __queue_work(int cpu, struct workqueue_struct *wq,
1290 struct work_struct *work)
1292 struct pool_workqueue *pwq;
1293 struct worker_pool *last_pool;
1294 struct list_head *worklist;
1295 unsigned int work_flags;
1296 unsigned int req_cpu = cpu;
1299 * While a work item is PENDING && off queue, a task trying to
1300 * steal the PENDING will busy-loop waiting for it to either get
1301 * queued or lose PENDING. Grabbing PENDING and queueing should
1302 * happen with IRQ disabled.
1304 WARN_ON_ONCE(!irqs_disabled());
1306 debug_work_activate(work);
1308 /* if dying, only works from the same workqueue are allowed */
1309 if (unlikely(wq->flags & __WQ_DRAINING) &&
1310 WARN_ON_ONCE(!is_chained_work(wq)))
1313 if (req_cpu == WORK_CPU_UNBOUND)
1314 cpu = raw_smp_processor_id();
1316 /* pwq which will be used unless @work is executing elsewhere */
1317 if (!(wq->flags & WQ_UNBOUND))
1318 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1320 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1323 * If @work was previously on a different pool, it might still be
1324 * running there, in which case the work needs to be queued on that
1325 * pool to guarantee non-reentrancy.
1327 last_pool = get_work_pool(work);
1328 if (last_pool && last_pool != pwq->pool) {
1329 struct worker *worker;
1331 spin_lock(&last_pool->lock);
1333 worker = find_worker_executing_work(last_pool, work);
1335 if (worker && worker->current_pwq->wq == wq) {
1336 pwq = worker->current_pwq;
1338 /* meh... not running there, queue here */
1339 spin_unlock(&last_pool->lock);
1340 spin_lock(&pwq->pool->lock);
1343 spin_lock(&pwq->pool->lock);
1347 * pwq is determined and locked. For unbound pools, we could have
1348 * raced with pwq release and it could already be dead. If its
1349 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1350 * without another pwq replacing it in the numa_pwq_tbl or while
1351 * work items are executing on it, so the retrying is guaranteed to
1352 * make forward-progress.
1354 if (unlikely(!pwq->refcnt)) {
1355 if (wq->flags & WQ_UNBOUND) {
1356 spin_unlock(&pwq->pool->lock);
1361 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1365 /* pwq determined, queue */
1366 trace_workqueue_queue_work(req_cpu, pwq, work);
1368 if (WARN_ON(!list_empty(&work->entry))) {
1369 spin_unlock(&pwq->pool->lock);
1373 pwq->nr_in_flight[pwq->work_color]++;
1374 work_flags = work_color_to_flags(pwq->work_color);
1376 if (likely(pwq->nr_active < pwq->max_active)) {
1377 trace_workqueue_activate_work(work);
1379 worklist = &pwq->pool->worklist;
1381 work_flags |= WORK_STRUCT_DELAYED;
1382 worklist = &pwq->delayed_works;
1385 insert_work(pwq, work, worklist, work_flags);
1387 spin_unlock(&pwq->pool->lock);
1391 * queue_work_on - queue work on specific cpu
1392 * @cpu: CPU number to execute work on
1393 * @wq: workqueue to use
1394 * @work: work to queue
1396 * Returns %false if @work was already on a queue, %true otherwise.
1398 * We queue the work to a specific CPU, the caller must ensure it
1401 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1402 struct work_struct *work)
1405 unsigned long flags;
1407 local_irq_save(flags);
1409 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1410 __queue_work(cpu, wq, work);
1414 local_irq_restore(flags);
1417 EXPORT_SYMBOL(queue_work_on);
1419 void delayed_work_timer_fn(unsigned long __data)
1421 struct delayed_work *dwork = (struct delayed_work *)__data;
1423 /* should have been called from irqsafe timer with irq already off */
1424 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1426 EXPORT_SYMBOL(delayed_work_timer_fn);
1428 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1429 struct delayed_work *dwork, unsigned long delay)
1431 struct timer_list *timer = &dwork->timer;
1432 struct work_struct *work = &dwork->work;
1434 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1435 timer->data != (unsigned long)dwork);
1436 WARN_ON_ONCE(timer_pending(timer));
1437 WARN_ON_ONCE(!list_empty(&work->entry));
1440 * If @delay is 0, queue @dwork->work immediately. This is for
1441 * both optimization and correctness. The earliest @timer can
1442 * expire is on the closest next tick and delayed_work users depend
1443 * on that there's no such delay when @delay is 0.
1446 __queue_work(cpu, wq, &dwork->work);
1450 timer_stats_timer_set_start_info(&dwork->timer);
1454 timer->expires = jiffies + delay;
1456 if (unlikely(cpu != WORK_CPU_UNBOUND))
1457 add_timer_on(timer, cpu);
1463 * queue_delayed_work_on - queue work on specific CPU after delay
1464 * @cpu: CPU number to execute work on
1465 * @wq: workqueue to use
1466 * @dwork: work to queue
1467 * @delay: number of jiffies to wait before queueing
1469 * Returns %false if @work was already on a queue, %true otherwise. If
1470 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1473 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1474 struct delayed_work *dwork, unsigned long delay)
1476 struct work_struct *work = &dwork->work;
1478 unsigned long flags;
1480 /* read the comment in __queue_work() */
1481 local_irq_save(flags);
1483 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1484 __queue_delayed_work(cpu, wq, dwork, delay);
1488 local_irq_restore(flags);
1491 EXPORT_SYMBOL(queue_delayed_work_on);
1494 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1495 * @cpu: CPU number to execute work on
1496 * @wq: workqueue to use
1497 * @dwork: work to queue
1498 * @delay: number of jiffies to wait before queueing
1500 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1501 * modify @dwork's timer so that it expires after @delay. If @delay is
1502 * zero, @work is guaranteed to be scheduled immediately regardless of its
1505 * Returns %false if @dwork was idle and queued, %true if @dwork was
1506 * pending and its timer was modified.
1508 * This function is safe to call from any context including IRQ handler.
1509 * See try_to_grab_pending() for details.
1511 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1512 struct delayed_work *dwork, unsigned long delay)
1514 unsigned long flags;
1518 ret = try_to_grab_pending(&dwork->work, true, &flags);
1519 } while (unlikely(ret == -EAGAIN));
1521 if (likely(ret >= 0)) {
1522 __queue_delayed_work(cpu, wq, dwork, delay);
1523 local_irq_restore(flags);
1526 /* -ENOENT from try_to_grab_pending() becomes %true */
1529 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1532 * worker_enter_idle - enter idle state
1533 * @worker: worker which is entering idle state
1535 * @worker is entering idle state. Update stats and idle timer if
1539 * spin_lock_irq(pool->lock).
1541 static void worker_enter_idle(struct worker *worker)
1543 struct worker_pool *pool = worker->pool;
1545 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1546 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1547 (worker->hentry.next || worker->hentry.pprev)))
1550 /* can't use worker_set_flags(), also called from start_worker() */
1551 worker->flags |= WORKER_IDLE;
1553 worker->last_active = jiffies;
1555 /* idle_list is LIFO */
1556 list_add(&worker->entry, &pool->idle_list);
1558 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1559 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1562 * Sanity check nr_running. Because wq_unbind_fn() releases
1563 * pool->lock between setting %WORKER_UNBOUND and zapping
1564 * nr_running, the warning may trigger spuriously. Check iff
1565 * unbind is not in progress.
1567 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1568 pool->nr_workers == pool->nr_idle &&
1569 atomic_read(&pool->nr_running));
1573 * worker_leave_idle - leave idle state
1574 * @worker: worker which is leaving idle state
1576 * @worker is leaving idle state. Update stats.
1579 * spin_lock_irq(pool->lock).
1581 static void worker_leave_idle(struct worker *worker)
1583 struct worker_pool *pool = worker->pool;
1585 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1587 worker_clr_flags(worker, WORKER_IDLE);
1589 list_del_init(&worker->entry);
1593 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1594 * @pool: target worker_pool
1596 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1598 * Works which are scheduled while the cpu is online must at least be
1599 * scheduled to a worker which is bound to the cpu so that if they are
1600 * flushed from cpu callbacks while cpu is going down, they are
1601 * guaranteed to execute on the cpu.
1603 * This function is to be used by unbound workers and rescuers to bind
1604 * themselves to the target cpu and may race with cpu going down or
1605 * coming online. kthread_bind() can't be used because it may put the
1606 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1607 * verbatim as it's best effort and blocking and pool may be
1608 * [dis]associated in the meantime.
1610 * This function tries set_cpus_allowed() and locks pool and verifies the
1611 * binding against %POOL_DISASSOCIATED which is set during
1612 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1613 * enters idle state or fetches works without dropping lock, it can
1614 * guarantee the scheduling requirement described in the first paragraph.
1617 * Might sleep. Called without any lock but returns with pool->lock
1621 * %true if the associated pool is online (@worker is successfully
1622 * bound), %false if offline.
1624 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1625 __acquires(&pool->lock)
1629 * The following call may fail, succeed or succeed
1630 * without actually migrating the task to the cpu if
1631 * it races with cpu hotunplug operation. Verify
1632 * against POOL_DISASSOCIATED.
1634 if (!(pool->flags & POOL_DISASSOCIATED))
1635 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1637 spin_lock_irq(&pool->lock);
1638 if (pool->flags & POOL_DISASSOCIATED)
1640 if (task_cpu(current) == pool->cpu &&
1641 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1643 spin_unlock_irq(&pool->lock);
1646 * We've raced with CPU hot[un]plug. Give it a breather
1647 * and retry migration. cond_resched() is required here;
1648 * otherwise, we might deadlock against cpu_stop trying to
1649 * bring down the CPU on non-preemptive kernel.
1656 static struct worker *alloc_worker(void)
1658 struct worker *worker;
1660 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1662 INIT_LIST_HEAD(&worker->entry);
1663 INIT_LIST_HEAD(&worker->scheduled);
1664 /* on creation a worker is in !idle && prep state */
1665 worker->flags = WORKER_PREP;
1671 * create_worker - create a new workqueue worker
1672 * @pool: pool the new worker will belong to
1674 * Create a new worker which is bound to @pool. The returned worker
1675 * can be started by calling start_worker() or destroyed using
1679 * Might sleep. Does GFP_KERNEL allocations.
1682 * Pointer to the newly created worker.
1684 static struct worker *create_worker(struct worker_pool *pool)
1686 struct worker *worker = NULL;
1690 lockdep_assert_held(&pool->manager_mutex);
1693 * ID is needed to determine kthread name. Allocate ID first
1694 * without installing the pointer.
1696 idr_preload(GFP_KERNEL);
1697 spin_lock_irq(&pool->lock);
1699 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1701 spin_unlock_irq(&pool->lock);
1706 worker = alloc_worker();
1710 worker->pool = pool;
1714 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1715 pool->attrs->nice < 0 ? "H" : "");
1717 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1719 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1720 "kworker/%s", id_buf);
1721 if (IS_ERR(worker->task))
1725 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1726 * online CPUs. It'll be re-applied when any of the CPUs come up.
1728 set_user_nice(worker->task, pool->attrs->nice);
1729 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1731 /* prevent userland from meddling with cpumask of workqueue workers */
1732 worker->task->flags |= PF_NO_SETAFFINITY;
1735 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1736 * remains stable across this function. See the comments above the
1737 * flag definition for details.
1739 if (pool->flags & POOL_DISASSOCIATED)
1740 worker->flags |= WORKER_UNBOUND;
1742 /* successful, commit the pointer to idr */
1743 spin_lock_irq(&pool->lock);
1744 idr_replace(&pool->worker_idr, worker, worker->id);
1745 spin_unlock_irq(&pool->lock);
1751 spin_lock_irq(&pool->lock);
1752 idr_remove(&pool->worker_idr, id);
1753 spin_unlock_irq(&pool->lock);
1760 * start_worker - start a newly created worker
1761 * @worker: worker to start
1763 * Make the pool aware of @worker and start it.
1766 * spin_lock_irq(pool->lock).
1768 static void start_worker(struct worker *worker)
1770 worker->flags |= WORKER_STARTED;
1771 worker->pool->nr_workers++;
1772 worker_enter_idle(worker);
1773 wake_up_process(worker->task);
1777 * create_and_start_worker - create and start a worker for a pool
1778 * @pool: the target pool
1780 * Grab the managership of @pool and create and start a new worker for it.
1782 static int create_and_start_worker(struct worker_pool *pool)
1784 struct worker *worker;
1786 mutex_lock(&pool->manager_mutex);
1788 worker = create_worker(pool);
1790 spin_lock_irq(&pool->lock);
1791 start_worker(worker);
1792 spin_unlock_irq(&pool->lock);
1795 mutex_unlock(&pool->manager_mutex);
1797 return worker ? 0 : -ENOMEM;
1801 * destroy_worker - destroy a workqueue worker
1802 * @worker: worker to be destroyed
1804 * Destroy @worker and adjust @pool stats accordingly.
1807 * spin_lock_irq(pool->lock) which is released and regrabbed.
1809 static void destroy_worker(struct worker *worker)
1811 struct worker_pool *pool = worker->pool;
1813 lockdep_assert_held(&pool->manager_mutex);
1814 lockdep_assert_held(&pool->lock);
1816 /* sanity check frenzy */
1817 if (WARN_ON(worker->current_work) ||
1818 WARN_ON(!list_empty(&worker->scheduled)))
1821 if (worker->flags & WORKER_STARTED)
1823 if (worker->flags & WORKER_IDLE)
1827 * Once WORKER_DIE is set, the kworker may destroy itself at any
1828 * point. Pin to ensure the task stays until we're done with it.
1830 get_task_struct(worker->task);
1832 list_del_init(&worker->entry);
1833 worker->flags |= WORKER_DIE;
1835 idr_remove(&pool->worker_idr, worker->id);
1837 spin_unlock_irq(&pool->lock);
1839 kthread_stop(worker->task);
1840 put_task_struct(worker->task);
1843 spin_lock_irq(&pool->lock);
1846 static void idle_worker_timeout(unsigned long __pool)
1848 struct worker_pool *pool = (void *)__pool;
1850 spin_lock_irq(&pool->lock);
1852 if (too_many_workers(pool)) {
1853 struct worker *worker;
1854 unsigned long expires;
1856 /* idle_list is kept in LIFO order, check the last one */
1857 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1858 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1860 if (time_before(jiffies, expires))
1861 mod_timer(&pool->idle_timer, expires);
1863 /* it's been idle for too long, wake up manager */
1864 pool->flags |= POOL_MANAGE_WORKERS;
1865 wake_up_worker(pool);
1869 spin_unlock_irq(&pool->lock);
1872 static void send_mayday(struct work_struct *work)
1874 struct pool_workqueue *pwq = get_work_pwq(work);
1875 struct workqueue_struct *wq = pwq->wq;
1877 lockdep_assert_held(&wq_mayday_lock);
1882 /* mayday mayday mayday */
1883 if (list_empty(&pwq->mayday_node)) {
1885 * If @pwq is for an unbound wq, its base ref may be put at
1886 * any time due to an attribute change. Pin @pwq until the
1887 * rescuer is done with it.
1890 list_add_tail(&pwq->mayday_node, &wq->maydays);
1891 wake_up_process(wq->rescuer->task);
1895 static void pool_mayday_timeout(unsigned long __pool)
1897 struct worker_pool *pool = (void *)__pool;
1898 struct work_struct *work;
1900 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1901 spin_lock(&pool->lock);
1903 if (need_to_create_worker(pool)) {
1905 * We've been trying to create a new worker but
1906 * haven't been successful. We might be hitting an
1907 * allocation deadlock. Send distress signals to
1910 list_for_each_entry(work, &pool->worklist, entry)
1914 spin_unlock(&pool->lock);
1915 spin_unlock_irq(&wq_mayday_lock);
1917 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1921 * maybe_create_worker - create a new worker if necessary
1922 * @pool: pool to create a new worker for
1924 * Create a new worker for @pool if necessary. @pool is guaranteed to
1925 * have at least one idle worker on return from this function. If
1926 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1927 * sent to all rescuers with works scheduled on @pool to resolve
1928 * possible allocation deadlock.
1930 * On return, need_to_create_worker() is guaranteed to be %false and
1931 * may_start_working() %true.
1934 * spin_lock_irq(pool->lock) which may be released and regrabbed
1935 * multiple times. Does GFP_KERNEL allocations. Called only from
1938 static void maybe_create_worker(struct worker_pool *pool)
1939 __releases(&pool->lock)
1940 __acquires(&pool->lock)
1942 if (!need_to_create_worker(pool))
1945 spin_unlock_irq(&pool->lock);
1947 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1948 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1951 struct worker *worker;
1953 worker = create_worker(pool);
1955 del_timer_sync(&pool->mayday_timer);
1956 spin_lock_irq(&pool->lock);
1957 start_worker(worker);
1958 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1963 if (!need_to_create_worker(pool))
1966 __set_current_state(TASK_INTERRUPTIBLE);
1967 schedule_timeout(CREATE_COOLDOWN);
1969 if (!need_to_create_worker(pool))
1973 del_timer_sync(&pool->mayday_timer);
1974 spin_lock_irq(&pool->lock);
1975 if (need_to_create_worker(pool))
1981 * maybe_destroy_worker - destroy workers which have been idle for a while
1982 * @pool: pool to destroy workers for
1984 * Destroy @pool workers which have been idle for longer than
1985 * IDLE_WORKER_TIMEOUT.
1988 * spin_lock_irq(pool->lock) which may be released and regrabbed
1989 * multiple times. Called only from manager.
1991 static void maybe_destroy_workers(struct worker_pool *pool)
1993 while (too_many_workers(pool)) {
1994 struct worker *worker;
1995 unsigned long expires;
1997 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1998 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2000 if (time_before(jiffies, expires)) {
2001 mod_timer(&pool->idle_timer, expires);
2005 destroy_worker(worker);
2010 * manage_workers - manage worker pool
2013 * Assume the manager role and manage the worker pool @worker belongs
2014 * to. At any given time, there can be only zero or one manager per
2015 * pool. The exclusion is handled automatically by this function.
2017 * The caller can safely start processing works on false return. On
2018 * true return, it's guaranteed that need_to_create_worker() is false
2019 * and may_start_working() is true.
2022 * spin_lock_irq(pool->lock) which may be released and regrabbed
2023 * multiple times. Does GFP_KERNEL allocations.
2026 * %false if the pool doesn't need management and the caller can safely
2027 * start processing works, %true if management function was performed and
2028 * the conditions that the caller verified before calling the function may
2029 * no longer be true.
2031 static bool manage_workers(struct worker *worker)
2033 struct worker_pool *pool = worker->pool;
2036 * Managership is governed by two mutexes - manager_arb and
2037 * manager_mutex. manager_arb handles arbitration of manager role.
2038 * Anyone who successfully grabs manager_arb wins the arbitration
2039 * and becomes the manager. mutex_trylock() on pool->manager_arb
2040 * failure while holding pool->lock reliably indicates that someone
2041 * else is managing the pool and the worker which failed trylock
2042 * can proceed to executing work items. This means that anyone
2043 * grabbing manager_arb is responsible for actually performing
2044 * manager duties. If manager_arb is grabbed and released without
2045 * actual management, the pool may stall indefinitely.
2047 * manager_mutex is used for exclusion of actual management
2048 * operations. The holder of manager_mutex can be sure that none
2049 * of management operations, including creation and destruction of
2050 * workers, won't take place until the mutex is released. Because
2051 * manager_mutex doesn't interfere with manager role arbitration,
2052 * it is guaranteed that the pool's management, while may be
2053 * delayed, won't be disturbed by someone else grabbing
2056 if (!mutex_trylock(&pool->manager_arb))
2060 * With manager arbitration won, manager_mutex would be free in
2061 * most cases. trylock first without dropping @pool->lock.
2063 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2064 spin_unlock_irq(&pool->lock);
2065 mutex_lock(&pool->manager_mutex);
2066 spin_lock_irq(&pool->lock);
2069 pool->flags &= ~POOL_MANAGE_WORKERS;
2072 * Destroy and then create so that may_start_working() is true
2075 maybe_destroy_workers(pool);
2076 maybe_create_worker(pool);
2078 mutex_unlock(&pool->manager_mutex);
2079 mutex_unlock(&pool->manager_arb);
2084 * process_one_work - process single work
2086 * @work: work to process
2088 * Process @work. This function contains all the logics necessary to
2089 * process a single work including synchronization against and
2090 * interaction with other workers on the same cpu, queueing and
2091 * flushing. As long as context requirement is met, any worker can
2092 * call this function to process a work.
2095 * spin_lock_irq(pool->lock) which is released and regrabbed.
2097 static void process_one_work(struct worker *worker, struct work_struct *work)
2098 __releases(&pool->lock)
2099 __acquires(&pool->lock)
2101 struct pool_workqueue *pwq = get_work_pwq(work);
2102 struct worker_pool *pool = worker->pool;
2103 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2105 struct worker *collision;
2106 #ifdef CONFIG_LOCKDEP
2108 * It is permissible to free the struct work_struct from
2109 * inside the function that is called from it, this we need to
2110 * take into account for lockdep too. To avoid bogus "held
2111 * lock freed" warnings as well as problems when looking into
2112 * work->lockdep_map, make a copy and use that here.
2114 struct lockdep_map lockdep_map;
2116 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2119 * Ensure we're on the correct CPU. DISASSOCIATED test is
2120 * necessary to avoid spurious warnings from rescuers servicing the
2121 * unbound or a disassociated pool.
2123 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2124 !(pool->flags & POOL_DISASSOCIATED) &&
2125 raw_smp_processor_id() != pool->cpu);
2128 * A single work shouldn't be executed concurrently by
2129 * multiple workers on a single cpu. Check whether anyone is
2130 * already processing the work. If so, defer the work to the
2131 * currently executing one.
2133 collision = find_worker_executing_work(pool, work);
2134 if (unlikely(collision)) {
2135 move_linked_works(work, &collision->scheduled, NULL);
2139 /* claim and dequeue */
2140 debug_work_deactivate(work);
2141 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2142 worker->current_work = work;
2143 worker->current_func = work->func;
2144 worker->current_pwq = pwq;
2145 work_color = get_work_color(work);
2147 list_del_init(&work->entry);
2150 * CPU intensive works don't participate in concurrency
2151 * management. They're the scheduler's responsibility.
2153 if (unlikely(cpu_intensive))
2154 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2157 * Unbound pool isn't concurrency managed and work items should be
2158 * executed ASAP. Wake up another worker if necessary.
2160 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2161 wake_up_worker(pool);
2164 * Record the last pool and clear PENDING which should be the last
2165 * update to @work. Also, do this inside @pool->lock so that
2166 * PENDING and queued state changes happen together while IRQ is
2169 set_work_pool_and_clear_pending(work, pool->id);
2171 spin_unlock_irq(&pool->lock);
2173 lock_map_acquire_read(&pwq->wq->lockdep_map);
2174 lock_map_acquire(&lockdep_map);
2175 trace_workqueue_execute_start(work);
2176 worker->current_func(work);
2178 * While we must be careful to not use "work" after this, the trace
2179 * point will only record its address.
2181 trace_workqueue_execute_end(work);
2182 lock_map_release(&lockdep_map);
2183 lock_map_release(&pwq->wq->lockdep_map);
2185 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2186 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2187 " last function: %pf\n",
2188 current->comm, preempt_count(), task_pid_nr(current),
2189 worker->current_func);
2190 debug_show_held_locks(current);
2195 * The following prevents a kworker from hogging CPU on !PREEMPT
2196 * kernels, where a requeueing work item waiting for something to
2197 * happen could deadlock with stop_machine as such work item could
2198 * indefinitely requeue itself while all other CPUs are trapped in
2203 spin_lock_irq(&pool->lock);
2205 /* clear cpu intensive status */
2206 if (unlikely(cpu_intensive))
2207 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2209 /* we're done with it, release */
2210 hash_del(&worker->hentry);
2211 worker->current_work = NULL;
2212 worker->current_func = NULL;
2213 worker->current_pwq = NULL;
2214 worker->desc_valid = false;
2215 pwq_dec_nr_in_flight(pwq, work_color);
2219 * process_scheduled_works - process scheduled works
2222 * Process all scheduled works. Please note that the scheduled list
2223 * may change while processing a work, so this function repeatedly
2224 * fetches a work from the top and executes it.
2227 * spin_lock_irq(pool->lock) which may be released and regrabbed
2230 static void process_scheduled_works(struct worker *worker)
2232 while (!list_empty(&worker->scheduled)) {
2233 struct work_struct *work = list_first_entry(&worker->scheduled,
2234 struct work_struct, entry);
2235 process_one_work(worker, work);
2240 * worker_thread - the worker thread function
2243 * The worker thread function. All workers belong to a worker_pool -
2244 * either a per-cpu one or dynamic unbound one. These workers process all
2245 * work items regardless of their specific target workqueue. The only
2246 * exception is work items which belong to workqueues with a rescuer which
2247 * will be explained in rescuer_thread().
2249 static int worker_thread(void *__worker)
2251 struct worker *worker = __worker;
2252 struct worker_pool *pool = worker->pool;
2254 /* tell the scheduler that this is a workqueue worker */
2255 worker->task->flags |= PF_WQ_WORKER;
2257 spin_lock_irq(&pool->lock);
2259 /* am I supposed to die? */
2260 if (unlikely(worker->flags & WORKER_DIE)) {
2261 spin_unlock_irq(&pool->lock);
2262 WARN_ON_ONCE(!list_empty(&worker->entry));
2263 worker->task->flags &= ~PF_WQ_WORKER;
2267 worker_leave_idle(worker);
2269 /* no more worker necessary? */
2270 if (!need_more_worker(pool))
2273 /* do we need to manage? */
2274 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2278 * ->scheduled list can only be filled while a worker is
2279 * preparing to process a work or actually processing it.
2280 * Make sure nobody diddled with it while I was sleeping.
2282 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2285 * Finish PREP stage. We're guaranteed to have at least one idle
2286 * worker or that someone else has already assumed the manager
2287 * role. This is where @worker starts participating in concurrency
2288 * management if applicable and concurrency management is restored
2289 * after being rebound. See rebind_workers() for details.
2291 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2294 struct work_struct *work =
2295 list_first_entry(&pool->worklist,
2296 struct work_struct, entry);
2298 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2299 /* optimization path, not strictly necessary */
2300 process_one_work(worker, work);
2301 if (unlikely(!list_empty(&worker->scheduled)))
2302 process_scheduled_works(worker);
2304 move_linked_works(work, &worker->scheduled, NULL);
2305 process_scheduled_works(worker);
2307 } while (keep_working(pool));
2309 worker_set_flags(worker, WORKER_PREP, false);
2311 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2315 * pool->lock is held and there's no work to process and no need to
2316 * manage, sleep. Workers are woken up only while holding
2317 * pool->lock or from local cpu, so setting the current state
2318 * before releasing pool->lock is enough to prevent losing any
2321 worker_enter_idle(worker);
2322 __set_current_state(TASK_INTERRUPTIBLE);
2323 spin_unlock_irq(&pool->lock);
2329 * rescuer_thread - the rescuer thread function
2332 * Workqueue rescuer thread function. There's one rescuer for each
2333 * workqueue which has WQ_MEM_RECLAIM set.
2335 * Regular work processing on a pool may block trying to create a new
2336 * worker which uses GFP_KERNEL allocation which has slight chance of
2337 * developing into deadlock if some works currently on the same queue
2338 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2339 * the problem rescuer solves.
2341 * When such condition is possible, the pool summons rescuers of all
2342 * workqueues which have works queued on the pool and let them process
2343 * those works so that forward progress can be guaranteed.
2345 * This should happen rarely.
2347 static int rescuer_thread(void *__rescuer)
2349 struct worker *rescuer = __rescuer;
2350 struct workqueue_struct *wq = rescuer->rescue_wq;
2351 struct list_head *scheduled = &rescuer->scheduled;
2354 set_user_nice(current, RESCUER_NICE_LEVEL);
2357 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2358 * doesn't participate in concurrency management.
2360 rescuer->task->flags |= PF_WQ_WORKER;
2362 set_current_state(TASK_INTERRUPTIBLE);
2365 * By the time the rescuer is requested to stop, the workqueue
2366 * shouldn't have any work pending, but @wq->maydays may still have
2367 * pwq(s) queued. This can happen by non-rescuer workers consuming
2368 * all the work items before the rescuer got to them. Go through
2369 * @wq->maydays processing before acting on should_stop so that the
2370 * list is always empty on exit.
2372 should_stop = kthread_should_stop();
2374 /* see whether any pwq is asking for help */
2375 spin_lock_irq(&wq_mayday_lock);
2377 while (!list_empty(&wq->maydays)) {
2378 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2379 struct pool_workqueue, mayday_node);
2380 struct worker_pool *pool = pwq->pool;
2381 struct work_struct *work, *n;
2383 __set_current_state(TASK_RUNNING);
2384 list_del_init(&pwq->mayday_node);
2386 spin_unlock_irq(&wq_mayday_lock);
2388 /* migrate to the target cpu if possible */
2389 worker_maybe_bind_and_lock(pool);
2390 rescuer->pool = pool;
2393 * Slurp in all works issued via this workqueue and
2396 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2397 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2398 if (get_work_pwq(work) == pwq)
2399 move_linked_works(work, scheduled, &n);
2401 process_scheduled_works(rescuer);
2404 * Put the reference grabbed by send_mayday(). @pool won't
2405 * go away while we're holding its lock.
2410 * Leave this pool. If keep_working() is %true, notify a
2411 * regular worker; otherwise, we end up with 0 concurrency
2412 * and stalling the execution.
2414 if (keep_working(pool))
2415 wake_up_worker(pool);
2417 rescuer->pool = NULL;
2418 spin_unlock(&pool->lock);
2419 spin_lock(&wq_mayday_lock);
2422 spin_unlock_irq(&wq_mayday_lock);
2425 __set_current_state(TASK_RUNNING);
2426 rescuer->task->flags &= ~PF_WQ_WORKER;
2430 /* rescuers should never participate in concurrency management */
2431 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2437 struct work_struct work;
2438 struct completion done;
2441 static void wq_barrier_func(struct work_struct *work)
2443 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2444 complete(&barr->done);
2448 * insert_wq_barrier - insert a barrier work
2449 * @pwq: pwq to insert barrier into
2450 * @barr: wq_barrier to insert
2451 * @target: target work to attach @barr to
2452 * @worker: worker currently executing @target, NULL if @target is not executing
2454 * @barr is linked to @target such that @barr is completed only after
2455 * @target finishes execution. Please note that the ordering
2456 * guarantee is observed only with respect to @target and on the local
2459 * Currently, a queued barrier can't be canceled. This is because
2460 * try_to_grab_pending() can't determine whether the work to be
2461 * grabbed is at the head of the queue and thus can't clear LINKED
2462 * flag of the previous work while there must be a valid next work
2463 * after a work with LINKED flag set.
2465 * Note that when @worker is non-NULL, @target may be modified
2466 * underneath us, so we can't reliably determine pwq from @target.
2469 * spin_lock_irq(pool->lock).
2471 static void insert_wq_barrier(struct pool_workqueue *pwq,
2472 struct wq_barrier *barr,
2473 struct work_struct *target, struct worker *worker)
2475 struct list_head *head;
2476 unsigned int linked = 0;
2479 * debugobject calls are safe here even with pool->lock locked
2480 * as we know for sure that this will not trigger any of the
2481 * checks and call back into the fixup functions where we
2484 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2485 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2486 init_completion(&barr->done);
2489 * If @target is currently being executed, schedule the
2490 * barrier to the worker; otherwise, put it after @target.
2493 head = worker->scheduled.next;
2495 unsigned long *bits = work_data_bits(target);
2497 head = target->entry.next;
2498 /* there can already be other linked works, inherit and set */
2499 linked = *bits & WORK_STRUCT_LINKED;
2500 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2503 debug_work_activate(&barr->work);
2504 insert_work(pwq, &barr->work, head,
2505 work_color_to_flags(WORK_NO_COLOR) | linked);
2509 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2510 * @wq: workqueue being flushed
2511 * @flush_color: new flush color, < 0 for no-op
2512 * @work_color: new work color, < 0 for no-op
2514 * Prepare pwqs for workqueue flushing.
2516 * If @flush_color is non-negative, flush_color on all pwqs should be
2517 * -1. If no pwq has in-flight commands at the specified color, all
2518 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2519 * has in flight commands, its pwq->flush_color is set to
2520 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2521 * wakeup logic is armed and %true is returned.
2523 * The caller should have initialized @wq->first_flusher prior to
2524 * calling this function with non-negative @flush_color. If
2525 * @flush_color is negative, no flush color update is done and %false
2528 * If @work_color is non-negative, all pwqs should have the same
2529 * work_color which is previous to @work_color and all will be
2530 * advanced to @work_color.
2533 * mutex_lock(wq->mutex).
2536 * %true if @flush_color >= 0 and there's something to flush. %false
2539 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2540 int flush_color, int work_color)
2543 struct pool_workqueue *pwq;
2545 if (flush_color >= 0) {
2546 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2547 atomic_set(&wq->nr_pwqs_to_flush, 1);
2550 for_each_pwq(pwq, wq) {
2551 struct worker_pool *pool = pwq->pool;
2553 spin_lock_irq(&pool->lock);
2555 if (flush_color >= 0) {
2556 WARN_ON_ONCE(pwq->flush_color != -1);
2558 if (pwq->nr_in_flight[flush_color]) {
2559 pwq->flush_color = flush_color;
2560 atomic_inc(&wq->nr_pwqs_to_flush);
2565 if (work_color >= 0) {
2566 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2567 pwq->work_color = work_color;
2570 spin_unlock_irq(&pool->lock);
2573 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2574 complete(&wq->first_flusher->done);
2580 * flush_workqueue - ensure that any scheduled work has run to completion.
2581 * @wq: workqueue to flush
2583 * This function sleeps until all work items which were queued on entry
2584 * have finished execution, but it is not livelocked by new incoming ones.
2586 void flush_workqueue(struct workqueue_struct *wq)
2588 struct wq_flusher this_flusher = {
2589 .list = LIST_HEAD_INIT(this_flusher.list),
2591 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2595 lock_map_acquire(&wq->lockdep_map);
2596 lock_map_release(&wq->lockdep_map);
2598 mutex_lock(&wq->mutex);
2601 * Start-to-wait phase
2603 next_color = work_next_color(wq->work_color);
2605 if (next_color != wq->flush_color) {
2607 * Color space is not full. The current work_color
2608 * becomes our flush_color and work_color is advanced
2611 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2612 this_flusher.flush_color = wq->work_color;
2613 wq->work_color = next_color;
2615 if (!wq->first_flusher) {
2616 /* no flush in progress, become the first flusher */
2617 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2619 wq->first_flusher = &this_flusher;
2621 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2623 /* nothing to flush, done */
2624 wq->flush_color = next_color;
2625 wq->first_flusher = NULL;
2630 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2631 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2632 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2636 * Oops, color space is full, wait on overflow queue.
2637 * The next flush completion will assign us
2638 * flush_color and transfer to flusher_queue.
2640 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2643 mutex_unlock(&wq->mutex);
2645 wait_for_completion(&this_flusher.done);
2648 * Wake-up-and-cascade phase
2650 * First flushers are responsible for cascading flushes and
2651 * handling overflow. Non-first flushers can simply return.
2653 if (wq->first_flusher != &this_flusher)
2656 mutex_lock(&wq->mutex);
2658 /* we might have raced, check again with mutex held */
2659 if (wq->first_flusher != &this_flusher)
2662 wq->first_flusher = NULL;
2664 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2665 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2668 struct wq_flusher *next, *tmp;
2670 /* complete all the flushers sharing the current flush color */
2671 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2672 if (next->flush_color != wq->flush_color)
2674 list_del_init(&next->list);
2675 complete(&next->done);
2678 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2679 wq->flush_color != work_next_color(wq->work_color));
2681 /* this flush_color is finished, advance by one */
2682 wq->flush_color = work_next_color(wq->flush_color);
2684 /* one color has been freed, handle overflow queue */
2685 if (!list_empty(&wq->flusher_overflow)) {
2687 * Assign the same color to all overflowed
2688 * flushers, advance work_color and append to
2689 * flusher_queue. This is the start-to-wait
2690 * phase for these overflowed flushers.
2692 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2693 tmp->flush_color = wq->work_color;
2695 wq->work_color = work_next_color(wq->work_color);
2697 list_splice_tail_init(&wq->flusher_overflow,
2698 &wq->flusher_queue);
2699 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2702 if (list_empty(&wq->flusher_queue)) {
2703 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2708 * Need to flush more colors. Make the next flusher
2709 * the new first flusher and arm pwqs.
2711 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2712 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2714 list_del_init(&next->list);
2715 wq->first_flusher = next;
2717 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2721 * Meh... this color is already done, clear first
2722 * flusher and repeat cascading.
2724 wq->first_flusher = NULL;
2728 mutex_unlock(&wq->mutex);
2730 EXPORT_SYMBOL_GPL(flush_workqueue);
2733 * drain_workqueue - drain a workqueue
2734 * @wq: workqueue to drain
2736 * Wait until the workqueue becomes empty. While draining is in progress,
2737 * only chain queueing is allowed. IOW, only currently pending or running
2738 * work items on @wq can queue further work items on it. @wq is flushed
2739 * repeatedly until it becomes empty. The number of flushing is detemined
2740 * by the depth of chaining and should be relatively short. Whine if it
2743 void drain_workqueue(struct workqueue_struct *wq)
2745 unsigned int flush_cnt = 0;
2746 struct pool_workqueue *pwq;
2749 * __queue_work() needs to test whether there are drainers, is much
2750 * hotter than drain_workqueue() and already looks at @wq->flags.
2751 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2753 mutex_lock(&wq->mutex);
2754 if (!wq->nr_drainers++)
2755 wq->flags |= __WQ_DRAINING;
2756 mutex_unlock(&wq->mutex);
2758 flush_workqueue(wq);
2760 mutex_lock(&wq->mutex);
2762 for_each_pwq(pwq, wq) {
2765 spin_lock_irq(&pwq->pool->lock);
2766 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2767 spin_unlock_irq(&pwq->pool->lock);
2772 if (++flush_cnt == 10 ||
2773 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2774 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2775 wq->name, flush_cnt);
2777 mutex_unlock(&wq->mutex);
2781 if (!--wq->nr_drainers)
2782 wq->flags &= ~__WQ_DRAINING;
2783 mutex_unlock(&wq->mutex);
2785 EXPORT_SYMBOL_GPL(drain_workqueue);
2787 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2789 struct worker *worker = NULL;
2790 struct worker_pool *pool;
2791 struct pool_workqueue *pwq;
2795 local_irq_disable();
2796 pool = get_work_pool(work);
2802 spin_lock(&pool->lock);
2803 /* see the comment in try_to_grab_pending() with the same code */
2804 pwq = get_work_pwq(work);
2806 if (unlikely(pwq->pool != pool))
2809 worker = find_worker_executing_work(pool, work);
2812 pwq = worker->current_pwq;
2815 insert_wq_barrier(pwq, barr, work, worker);
2816 spin_unlock_irq(&pool->lock);
2819 * If @max_active is 1 or rescuer is in use, flushing another work
2820 * item on the same workqueue may lead to deadlock. Make sure the
2821 * flusher is not running on the same workqueue by verifying write
2824 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2825 lock_map_acquire(&pwq->wq->lockdep_map);
2827 lock_map_acquire_read(&pwq->wq->lockdep_map);
2828 lock_map_release(&pwq->wq->lockdep_map);
2832 spin_unlock_irq(&pool->lock);
2837 * flush_work - wait for a work to finish executing the last queueing instance
2838 * @work: the work to flush
2840 * Wait until @work has finished execution. @work is guaranteed to be idle
2841 * on return if it hasn't been requeued since flush started.
2844 * %true if flush_work() waited for the work to finish execution,
2845 * %false if it was already idle.
2847 bool flush_work(struct work_struct *work)
2849 struct wq_barrier barr;
2851 lock_map_acquire(&work->lockdep_map);
2852 lock_map_release(&work->lockdep_map);
2854 if (start_flush_work(work, &barr)) {
2855 wait_for_completion(&barr.done);
2856 destroy_work_on_stack(&barr.work);
2862 EXPORT_SYMBOL_GPL(flush_work);
2864 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2866 unsigned long flags;
2870 ret = try_to_grab_pending(work, is_dwork, &flags);
2872 * If someone else is canceling, wait for the same event it
2873 * would be waiting for before retrying.
2875 if (unlikely(ret == -ENOENT))
2877 } while (unlikely(ret < 0));
2879 /* tell other tasks trying to grab @work to back off */
2880 mark_work_canceling(work);
2881 local_irq_restore(flags);
2884 clear_work_data(work);
2889 * cancel_work_sync - cancel a work and wait for it to finish
2890 * @work: the work to cancel
2892 * Cancel @work and wait for its execution to finish. This function
2893 * can be used even if the work re-queues itself or migrates to
2894 * another workqueue. On return from this function, @work is
2895 * guaranteed to be not pending or executing on any CPU.
2897 * cancel_work_sync(&delayed_work->work) must not be used for
2898 * delayed_work's. Use cancel_delayed_work_sync() instead.
2900 * The caller must ensure that the workqueue on which @work was last
2901 * queued can't be destroyed before this function returns.
2904 * %true if @work was pending, %false otherwise.
2906 bool cancel_work_sync(struct work_struct *work)
2908 return __cancel_work_timer(work, false);
2910 EXPORT_SYMBOL_GPL(cancel_work_sync);
2913 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2914 * @dwork: the delayed work to flush
2916 * Delayed timer is cancelled and the pending work is queued for
2917 * immediate execution. Like flush_work(), this function only
2918 * considers the last queueing instance of @dwork.
2921 * %true if flush_work() waited for the work to finish execution,
2922 * %false if it was already idle.
2924 bool flush_delayed_work(struct delayed_work *dwork)
2926 local_irq_disable();
2927 if (del_timer_sync(&dwork->timer))
2928 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2930 return flush_work(&dwork->work);
2932 EXPORT_SYMBOL(flush_delayed_work);
2935 * cancel_delayed_work - cancel a delayed work
2936 * @dwork: delayed_work to cancel
2938 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2939 * and canceled; %false if wasn't pending. Note that the work callback
2940 * function may still be running on return, unless it returns %true and the
2941 * work doesn't re-arm itself. Explicitly flush or use
2942 * cancel_delayed_work_sync() to wait on it.
2944 * This function is safe to call from any context including IRQ handler.
2946 bool cancel_delayed_work(struct delayed_work *dwork)
2948 unsigned long flags;
2952 ret = try_to_grab_pending(&dwork->work, true, &flags);
2953 } while (unlikely(ret == -EAGAIN));
2955 if (unlikely(ret < 0))
2958 set_work_pool_and_clear_pending(&dwork->work,
2959 get_work_pool_id(&dwork->work));
2960 local_irq_restore(flags);
2963 EXPORT_SYMBOL(cancel_delayed_work);
2966 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2967 * @dwork: the delayed work cancel
2969 * This is cancel_work_sync() for delayed works.
2972 * %true if @dwork was pending, %false otherwise.
2974 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2976 return __cancel_work_timer(&dwork->work, true);
2978 EXPORT_SYMBOL(cancel_delayed_work_sync);
2981 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2982 * @func: the function to call
2984 * schedule_on_each_cpu() executes @func on each online CPU using the
2985 * system workqueue and blocks until all CPUs have completed.
2986 * schedule_on_each_cpu() is very slow.
2989 * 0 on success, -errno on failure.
2991 int schedule_on_each_cpu(work_func_t func)
2994 struct work_struct __percpu *works;
2996 works = alloc_percpu(struct work_struct);
3002 for_each_online_cpu(cpu) {
3003 struct work_struct *work = per_cpu_ptr(works, cpu);
3005 INIT_WORK(work, func);
3006 schedule_work_on(cpu, work);
3009 for_each_online_cpu(cpu)
3010 flush_work(per_cpu_ptr(works, cpu));
3018 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3020 * Forces execution of the kernel-global workqueue and blocks until its
3023 * Think twice before calling this function! It's very easy to get into
3024 * trouble if you don't take great care. Either of the following situations
3025 * will lead to deadlock:
3027 * One of the work items currently on the workqueue needs to acquire
3028 * a lock held by your code or its caller.
3030 * Your code is running in the context of a work routine.
3032 * They will be detected by lockdep when they occur, but the first might not
3033 * occur very often. It depends on what work items are on the workqueue and
3034 * what locks they need, which you have no control over.
3036 * In most situations flushing the entire workqueue is overkill; you merely
3037 * need to know that a particular work item isn't queued and isn't running.
3038 * In such cases you should use cancel_delayed_work_sync() or
3039 * cancel_work_sync() instead.
3041 void flush_scheduled_work(void)
3043 flush_workqueue(system_wq);
3045 EXPORT_SYMBOL(flush_scheduled_work);
3048 * execute_in_process_context - reliably execute the routine with user context
3049 * @fn: the function to execute
3050 * @ew: guaranteed storage for the execute work structure (must
3051 * be available when the work executes)
3053 * Executes the function immediately if process context is available,
3054 * otherwise schedules the function for delayed execution.
3056 * Returns: 0 - function was executed
3057 * 1 - function was scheduled for execution
3059 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3061 if (!in_interrupt()) {
3066 INIT_WORK(&ew->work, fn);
3067 schedule_work(&ew->work);
3071 EXPORT_SYMBOL_GPL(execute_in_process_context);
3075 * Workqueues with WQ_SYSFS flag set is visible to userland via
3076 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3077 * following attributes.
3079 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3080 * max_active RW int : maximum number of in-flight work items
3082 * Unbound workqueues have the following extra attributes.
3084 * id RO int : the associated pool ID
3085 * nice RW int : nice value of the workers
3086 * cpumask RW mask : bitmask of allowed CPUs for the workers
3089 struct workqueue_struct *wq;
3093 static struct workqueue_struct *dev_to_wq(struct device *dev)
3095 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3100 static ssize_t wq_per_cpu_show(struct device *dev,
3101 struct device_attribute *attr, char *buf)
3103 struct workqueue_struct *wq = dev_to_wq(dev);
3105 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3108 static ssize_t wq_max_active_show(struct device *dev,
3109 struct device_attribute *attr, char *buf)
3111 struct workqueue_struct *wq = dev_to_wq(dev);
3113 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3116 static ssize_t wq_max_active_store(struct device *dev,
3117 struct device_attribute *attr,
3118 const char *buf, size_t count)
3120 struct workqueue_struct *wq = dev_to_wq(dev);
3123 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3126 workqueue_set_max_active(wq, val);
3130 static struct device_attribute wq_sysfs_attrs[] = {
3131 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3132 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3136 static ssize_t wq_pool_ids_show(struct device *dev,
3137 struct device_attribute *attr, char *buf)
3139 struct workqueue_struct *wq = dev_to_wq(dev);
3140 const char *delim = "";
3141 int node, written = 0;
3143 rcu_read_lock_sched();
3144 for_each_node(node) {
3145 written += scnprintf(buf + written, PAGE_SIZE - written,
3146 "%s%d:%d", delim, node,
3147 unbound_pwq_by_node(wq, node)->pool->id);
3150 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3151 rcu_read_unlock_sched();
3156 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3159 struct workqueue_struct *wq = dev_to_wq(dev);
3162 mutex_lock(&wq->mutex);
3163 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3164 mutex_unlock(&wq->mutex);
3169 /* prepare workqueue_attrs for sysfs store operations */
3170 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3172 struct workqueue_attrs *attrs;
3174 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3178 mutex_lock(&wq->mutex);
3179 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3180 mutex_unlock(&wq->mutex);
3184 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3185 const char *buf, size_t count)
3187 struct workqueue_struct *wq = dev_to_wq(dev);
3188 struct workqueue_attrs *attrs;
3191 attrs = wq_sysfs_prep_attrs(wq);
3195 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3196 attrs->nice >= -20 && attrs->nice <= 19)
3197 ret = apply_workqueue_attrs(wq, attrs);
3201 free_workqueue_attrs(attrs);
3202 return ret ?: count;
3205 static ssize_t wq_cpumask_show(struct device *dev,
3206 struct device_attribute *attr, char *buf)
3208 struct workqueue_struct *wq = dev_to_wq(dev);
3211 mutex_lock(&wq->mutex);
3212 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3213 mutex_unlock(&wq->mutex);
3215 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3219 static ssize_t wq_cpumask_store(struct device *dev,
3220 struct device_attribute *attr,
3221 const char *buf, size_t count)
3223 struct workqueue_struct *wq = dev_to_wq(dev);
3224 struct workqueue_attrs *attrs;
3227 attrs = wq_sysfs_prep_attrs(wq);
3231 ret = cpumask_parse(buf, attrs->cpumask);
3233 ret = apply_workqueue_attrs(wq, attrs);
3235 free_workqueue_attrs(attrs);
3236 return ret ?: count;
3239 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3242 struct workqueue_struct *wq = dev_to_wq(dev);
3245 mutex_lock(&wq->mutex);
3246 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3247 !wq->unbound_attrs->no_numa);
3248 mutex_unlock(&wq->mutex);
3253 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3254 const char *buf, size_t count)
3256 struct workqueue_struct *wq = dev_to_wq(dev);
3257 struct workqueue_attrs *attrs;
3260 attrs = wq_sysfs_prep_attrs(wq);
3265 if (sscanf(buf, "%d", &v) == 1) {
3266 attrs->no_numa = !v;
3267 ret = apply_workqueue_attrs(wq, attrs);
3270 free_workqueue_attrs(attrs);
3271 return ret ?: count;
3274 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3275 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3276 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3277 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3278 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3282 static struct bus_type wq_subsys = {
3283 .name = "workqueue",
3284 .dev_attrs = wq_sysfs_attrs,
3287 static int __init wq_sysfs_init(void)
3289 return subsys_virtual_register(&wq_subsys, NULL);
3291 core_initcall(wq_sysfs_init);
3293 static void wq_device_release(struct device *dev)
3295 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3301 * workqueue_sysfs_register - make a workqueue visible in sysfs
3302 * @wq: the workqueue to register
3304 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3305 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3306 * which is the preferred method.
3308 * Workqueue user should use this function directly iff it wants to apply
3309 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3310 * apply_workqueue_attrs() may race against userland updating the
3313 * Returns 0 on success, -errno on failure.
3315 int workqueue_sysfs_register(struct workqueue_struct *wq)
3317 struct wq_device *wq_dev;
3321 * Adjusting max_active or creating new pwqs by applyting
3322 * attributes breaks ordering guarantee. Disallow exposing ordered
3325 if (WARN_ON(wq->flags & __WQ_ORDERED))
3328 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3333 wq_dev->dev.bus = &wq_subsys;
3334 wq_dev->dev.init_name = wq->name;
3335 wq_dev->dev.release = wq_device_release;
3338 * unbound_attrs are created separately. Suppress uevent until
3339 * everything is ready.
3341 dev_set_uevent_suppress(&wq_dev->dev, true);
3343 ret = device_register(&wq_dev->dev);
3350 if (wq->flags & WQ_UNBOUND) {
3351 struct device_attribute *attr;
3353 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3354 ret = device_create_file(&wq_dev->dev, attr);
3356 device_unregister(&wq_dev->dev);
3363 dev_set_uevent_suppress(&wq_dev->dev, false);
3364 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3369 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3370 * @wq: the workqueue to unregister
3372 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3374 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3376 struct wq_device *wq_dev = wq->wq_dev;
3382 device_unregister(&wq_dev->dev);
3384 #else /* CONFIG_SYSFS */
3385 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3386 #endif /* CONFIG_SYSFS */
3389 * free_workqueue_attrs - free a workqueue_attrs
3390 * @attrs: workqueue_attrs to free
3392 * Undo alloc_workqueue_attrs().
3394 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3397 free_cpumask_var(attrs->cpumask);
3403 * alloc_workqueue_attrs - allocate a workqueue_attrs
3404 * @gfp_mask: allocation mask to use
3406 * Allocate a new workqueue_attrs, initialize with default settings and
3407 * return it. Returns NULL on failure.
3409 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3411 struct workqueue_attrs *attrs;
3413 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3416 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3419 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3422 free_workqueue_attrs(attrs);
3426 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3427 const struct workqueue_attrs *from)
3429 to->nice = from->nice;
3430 cpumask_copy(to->cpumask, from->cpumask);
3432 * Unlike hash and equality test, this function doesn't ignore
3433 * ->no_numa as it is used for both pool and wq attrs. Instead,
3434 * get_unbound_pool() explicitly clears ->no_numa after copying.
3436 to->no_numa = from->no_numa;
3439 /* hash value of the content of @attr */
3440 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3444 hash = jhash_1word(attrs->nice, hash);
3445 hash = jhash(cpumask_bits(attrs->cpumask),
3446 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3450 /* content equality test */
3451 static bool wqattrs_equal(const struct workqueue_attrs *a,
3452 const struct workqueue_attrs *b)
3454 if (a->nice != b->nice)
3456 if (!cpumask_equal(a->cpumask, b->cpumask))
3462 * init_worker_pool - initialize a newly zalloc'd worker_pool
3463 * @pool: worker_pool to initialize
3465 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3466 * Returns 0 on success, -errno on failure. Even on failure, all fields
3467 * inside @pool proper are initialized and put_unbound_pool() can be called
3468 * on @pool safely to release it.
3470 static int init_worker_pool(struct worker_pool *pool)
3472 spin_lock_init(&pool->lock);
3475 pool->node = NUMA_NO_NODE;
3476 pool->flags |= POOL_DISASSOCIATED;
3477 INIT_LIST_HEAD(&pool->worklist);
3478 INIT_LIST_HEAD(&pool->idle_list);
3479 hash_init(pool->busy_hash);
3481 init_timer_deferrable(&pool->idle_timer);
3482 pool->idle_timer.function = idle_worker_timeout;
3483 pool->idle_timer.data = (unsigned long)pool;
3485 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3486 (unsigned long)pool);
3488 mutex_init(&pool->manager_arb);
3489 mutex_init(&pool->manager_mutex);
3490 idr_init(&pool->worker_idr);
3492 INIT_HLIST_NODE(&pool->hash_node);
3495 /* shouldn't fail above this point */
3496 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3502 static void rcu_free_pool(struct rcu_head *rcu)
3504 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3506 idr_destroy(&pool->worker_idr);
3507 free_workqueue_attrs(pool->attrs);
3512 * put_unbound_pool - put a worker_pool
3513 * @pool: worker_pool to put
3515 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3516 * safe manner. get_unbound_pool() calls this function on its failure path
3517 * and this function should be able to release pools which went through,
3518 * successfully or not, init_worker_pool().
3520 * Should be called with wq_pool_mutex held.
3522 static void put_unbound_pool(struct worker_pool *pool)
3524 struct worker *worker;
3526 lockdep_assert_held(&wq_pool_mutex);
3532 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3533 WARN_ON(!list_empty(&pool->worklist)))
3536 /* release id and unhash */
3538 idr_remove(&worker_pool_idr, pool->id);
3539 hash_del(&pool->hash_node);
3542 * Become the manager and destroy all workers. Grabbing
3543 * manager_arb prevents @pool's workers from blocking on
3546 mutex_lock(&pool->manager_arb);
3547 mutex_lock(&pool->manager_mutex);
3548 spin_lock_irq(&pool->lock);
3550 while ((worker = first_worker(pool)))
3551 destroy_worker(worker);
3552 WARN_ON(pool->nr_workers || pool->nr_idle);
3554 spin_unlock_irq(&pool->lock);
3555 mutex_unlock(&pool->manager_mutex);
3556 mutex_unlock(&pool->manager_arb);
3558 /* shut down the timers */
3559 del_timer_sync(&pool->idle_timer);
3560 del_timer_sync(&pool->mayday_timer);
3562 /* sched-RCU protected to allow dereferences from get_work_pool() */
3563 call_rcu_sched(&pool->rcu, rcu_free_pool);
3567 * get_unbound_pool - get a worker_pool with the specified attributes
3568 * @attrs: the attributes of the worker_pool to get
3570 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3571 * reference count and return it. If there already is a matching
3572 * worker_pool, it will be used; otherwise, this function attempts to
3573 * create a new one. On failure, returns NULL.
3575 * Should be called with wq_pool_mutex held.
3577 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3579 u32 hash = wqattrs_hash(attrs);
3580 struct worker_pool *pool;
3583 lockdep_assert_held(&wq_pool_mutex);
3585 /* do we already have a matching pool? */
3586 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3587 if (wqattrs_equal(pool->attrs, attrs)) {
3593 /* nope, create a new one */
3594 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3595 if (!pool || init_worker_pool(pool) < 0)
3598 if (workqueue_freezing)
3599 pool->flags |= POOL_FREEZING;
3601 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3602 copy_workqueue_attrs(pool->attrs, attrs);
3605 * no_numa isn't a worker_pool attribute, always clear it. See
3606 * 'struct workqueue_attrs' comments for detail.
3608 pool->attrs->no_numa = false;
3610 /* if cpumask is contained inside a NUMA node, we belong to that node */
3611 if (wq_numa_enabled) {
3612 for_each_node(node) {
3613 if (cpumask_subset(pool->attrs->cpumask,
3614 wq_numa_possible_cpumask[node])) {
3621 if (worker_pool_assign_id(pool) < 0)
3624 /* create and start the initial worker */
3625 if (create_and_start_worker(pool) < 0)
3629 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3634 put_unbound_pool(pool);
3638 static void rcu_free_pwq(struct rcu_head *rcu)
3640 kmem_cache_free(pwq_cache,
3641 container_of(rcu, struct pool_workqueue, rcu));
3645 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3646 * and needs to be destroyed.
3648 static void pwq_unbound_release_workfn(struct work_struct *work)
3650 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3651 unbound_release_work);
3652 struct workqueue_struct *wq = pwq->wq;
3653 struct worker_pool *pool = pwq->pool;
3656 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3660 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3661 * necessary on release but do it anyway. It's easier to verify
3662 * and consistent with the linking path.
3664 mutex_lock(&wq->mutex);
3665 list_del_rcu(&pwq->pwqs_node);
3666 is_last = list_empty(&wq->pwqs);
3667 mutex_unlock(&wq->mutex);
3669 mutex_lock(&wq_pool_mutex);
3670 put_unbound_pool(pool);
3671 mutex_unlock(&wq_pool_mutex);
3673 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3676 * If we're the last pwq going away, @wq is already dead and no one
3677 * is gonna access it anymore. Free it.
3680 free_workqueue_attrs(wq->unbound_attrs);
3686 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3687 * @pwq: target pool_workqueue
3689 * If @pwq isn't freezing, set @pwq->max_active to the associated
3690 * workqueue's saved_max_active and activate delayed work items
3691 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3693 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3695 struct workqueue_struct *wq = pwq->wq;
3696 bool freezable = wq->flags & WQ_FREEZABLE;
3698 /* for @wq->saved_max_active */
3699 lockdep_assert_held(&wq->mutex);
3701 /* fast exit for non-freezable wqs */
3702 if (!freezable && pwq->max_active == wq->saved_max_active)
3705 spin_lock_irq(&pwq->pool->lock);
3707 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3708 pwq->max_active = wq->saved_max_active;
3710 while (!list_empty(&pwq->delayed_works) &&
3711 pwq->nr_active < pwq->max_active)
3712 pwq_activate_first_delayed(pwq);
3715 * Need to kick a worker after thawed or an unbound wq's
3716 * max_active is bumped. It's a slow path. Do it always.
3718 wake_up_worker(pwq->pool);
3720 pwq->max_active = 0;
3723 spin_unlock_irq(&pwq->pool->lock);
3726 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3727 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3728 struct worker_pool *pool)
3730 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3732 memset(pwq, 0, sizeof(*pwq));
3736 pwq->flush_color = -1;
3738 INIT_LIST_HEAD(&pwq->delayed_works);
3739 INIT_LIST_HEAD(&pwq->pwqs_node);
3740 INIT_LIST_HEAD(&pwq->mayday_node);
3741 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3744 /* sync @pwq with the current state of its associated wq and link it */
3745 static void link_pwq(struct pool_workqueue *pwq)
3747 struct workqueue_struct *wq = pwq->wq;
3749 lockdep_assert_held(&wq->mutex);
3751 /* may be called multiple times, ignore if already linked */
3752 if (!list_empty(&pwq->pwqs_node))
3756 * Set the matching work_color. This is synchronized with
3757 * wq->mutex to avoid confusing flush_workqueue().
3759 pwq->work_color = wq->work_color;
3761 /* sync max_active to the current setting */
3762 pwq_adjust_max_active(pwq);
3765 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3768 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3769 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3770 const struct workqueue_attrs *attrs)
3772 struct worker_pool *pool;
3773 struct pool_workqueue *pwq;
3775 lockdep_assert_held(&wq_pool_mutex);
3777 pool = get_unbound_pool(attrs);
3781 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3783 put_unbound_pool(pool);
3787 init_pwq(pwq, wq, pool);
3791 /* undo alloc_unbound_pwq(), used only in the error path */
3792 static void free_unbound_pwq(struct pool_workqueue *pwq)
3794 lockdep_assert_held(&wq_pool_mutex);
3797 put_unbound_pool(pwq->pool);
3798 kmem_cache_free(pwq_cache, pwq);
3803 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3804 * @attrs: the wq_attrs of interest
3805 * @node: the target NUMA node
3806 * @cpu_going_down: if >= 0, the CPU to consider as offline
3807 * @cpumask: outarg, the resulting cpumask
3809 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3810 * @cpu_going_down is >= 0, that cpu is considered offline during
3811 * calculation. The result is stored in @cpumask. This function returns
3812 * %true if the resulting @cpumask is different from @attrs->cpumask,
3815 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3816 * enabled and @node has online CPUs requested by @attrs, the returned
3817 * cpumask is the intersection of the possible CPUs of @node and
3820 * The caller is responsible for ensuring that the cpumask of @node stays
3823 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3824 int cpu_going_down, cpumask_t *cpumask)
3826 if (!wq_numa_enabled || attrs->no_numa)
3829 /* does @node have any online CPUs @attrs wants? */
3830 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3831 if (cpu_going_down >= 0)
3832 cpumask_clear_cpu(cpu_going_down, cpumask);
3834 if (cpumask_empty(cpumask))
3837 /* yeap, return possible CPUs in @node that @attrs wants */
3838 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3839 return !cpumask_equal(cpumask, attrs->cpumask);
3842 cpumask_copy(cpumask, attrs->cpumask);
3846 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3847 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3849 struct pool_workqueue *pwq)
3851 struct pool_workqueue *old_pwq;
3853 lockdep_assert_held(&wq->mutex);
3855 /* link_pwq() can handle duplicate calls */
3858 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3859 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3864 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3865 * @wq: the target workqueue
3866 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3868 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3869 * machines, this function maps a separate pwq to each NUMA node with
3870 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3871 * NUMA node it was issued on. Older pwqs are released as in-flight work
3872 * items finish. Note that a work item which repeatedly requeues itself
3873 * back-to-back will stay on its current pwq.
3875 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3878 int apply_workqueue_attrs(struct workqueue_struct *wq,
3879 const struct workqueue_attrs *attrs)
3881 struct workqueue_attrs *new_attrs, *tmp_attrs;
3882 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3885 /* only unbound workqueues can change attributes */
3886 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3889 /* creating multiple pwqs breaks ordering guarantee */
3890 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3893 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3894 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3895 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3896 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3899 /* make a copy of @attrs and sanitize it */
3900 copy_workqueue_attrs(new_attrs, attrs);
3901 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3904 * We may create multiple pwqs with differing cpumasks. Make a
3905 * copy of @new_attrs which will be modified and used to obtain
3908 copy_workqueue_attrs(tmp_attrs, new_attrs);
3911 * CPUs should stay stable across pwq creations and installations.
3912 * Pin CPUs, determine the target cpumask for each node and create
3917 mutex_lock(&wq_pool_mutex);
3920 * If something goes wrong during CPU up/down, we'll fall back to
3921 * the default pwq covering whole @attrs->cpumask. Always create
3922 * it even if we don't use it immediately.
3924 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3928 for_each_node(node) {
3929 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3930 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3935 pwq_tbl[node] = dfl_pwq;
3939 mutex_unlock(&wq_pool_mutex);
3941 /* all pwqs have been created successfully, let's install'em */
3942 mutex_lock(&wq->mutex);
3944 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3946 /* save the previous pwq and install the new one */
3948 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3950 /* @dfl_pwq might not have been used, ensure it's linked */
3952 swap(wq->dfl_pwq, dfl_pwq);
3954 mutex_unlock(&wq->mutex);
3956 /* put the old pwqs */
3958 put_pwq_unlocked(pwq_tbl[node]);
3959 put_pwq_unlocked(dfl_pwq);
3965 free_workqueue_attrs(tmp_attrs);
3966 free_workqueue_attrs(new_attrs);
3971 free_unbound_pwq(dfl_pwq);
3973 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3974 free_unbound_pwq(pwq_tbl[node]);
3975 mutex_unlock(&wq_pool_mutex);
3983 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3984 * @wq: the target workqueue
3985 * @cpu: the CPU coming up or going down
3986 * @online: whether @cpu is coming up or going down
3988 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3989 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3992 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3993 * falls back to @wq->dfl_pwq which may not be optimal but is always
3996 * Note that when the last allowed CPU of a NUMA node goes offline for a
3997 * workqueue with a cpumask spanning multiple nodes, the workers which were
3998 * already executing the work items for the workqueue will lose their CPU
3999 * affinity and may execute on any CPU. This is similar to how per-cpu
4000 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4001 * affinity, it's the user's responsibility to flush the work item from
4004 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4007 int node = cpu_to_node(cpu);
4008 int cpu_off = online ? -1 : cpu;
4009 struct pool_workqueue *old_pwq = NULL, *pwq;
4010 struct workqueue_attrs *target_attrs;
4013 lockdep_assert_held(&wq_pool_mutex);
4015 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4019 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4020 * Let's use a preallocated one. The following buf is protected by
4021 * CPU hotplug exclusion.
4023 target_attrs = wq_update_unbound_numa_attrs_buf;
4024 cpumask = target_attrs->cpumask;
4026 mutex_lock(&wq->mutex);
4027 if (wq->unbound_attrs->no_numa)
4030 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4031 pwq = unbound_pwq_by_node(wq, node);
4034 * Let's determine what needs to be done. If the target cpumask is
4035 * different from wq's, we need to compare it to @pwq's and create
4036 * a new one if they don't match. If the target cpumask equals
4037 * wq's, the default pwq should be used. If @pwq is already the
4038 * default one, nothing to do; otherwise, install the default one.
4040 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4041 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4044 if (pwq == wq->dfl_pwq)
4050 mutex_unlock(&wq->mutex);
4052 /* create a new pwq */
4053 pwq = alloc_unbound_pwq(wq, target_attrs);
4055 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4057 mutex_lock(&wq->mutex);
4062 * Install the new pwq. As this function is called only from CPU
4063 * hotplug callbacks and applying a new attrs is wrapped with
4064 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4067 mutex_lock(&wq->mutex);
4068 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4072 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4073 get_pwq(wq->dfl_pwq);
4074 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4075 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4077 mutex_unlock(&wq->mutex);
4078 put_pwq_unlocked(old_pwq);
4081 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4083 bool highpri = wq->flags & WQ_HIGHPRI;
4086 if (!(wq->flags & WQ_UNBOUND)) {
4087 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4091 for_each_possible_cpu(cpu) {
4092 struct pool_workqueue *pwq =
4093 per_cpu_ptr(wq->cpu_pwqs, cpu);
4094 struct worker_pool *cpu_pools =
4095 per_cpu(cpu_worker_pools, cpu);
4097 init_pwq(pwq, wq, &cpu_pools[highpri]);
4099 mutex_lock(&wq->mutex);
4101 mutex_unlock(&wq->mutex);
4104 } else if (wq->flags & __WQ_ORDERED) {
4105 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4106 /* there should only be single pwq for ordering guarantee */
4107 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4108 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4109 "ordering guarantee broken for workqueue %s\n", wq->name);
4112 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4116 static int wq_clamp_max_active(int max_active, unsigned int flags,
4119 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4121 if (max_active < 1 || max_active > lim)
4122 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4123 max_active, name, 1, lim);
4125 return clamp_val(max_active, 1, lim);
4128 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4131 struct lock_class_key *key,
4132 const char *lock_name, ...)
4134 size_t tbl_size = 0;
4136 struct workqueue_struct *wq;
4137 struct pool_workqueue *pwq;
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 __cpuinit 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 __cpuinit 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);
4792 flush_work(&wfc.work);
4795 EXPORT_SYMBOL_GPL(work_on_cpu);
4796 #endif /* CONFIG_SMP */
4798 #ifdef CONFIG_FREEZER
4801 * freeze_workqueues_begin - begin freezing workqueues
4803 * Start freezing workqueues. After this function returns, all freezable
4804 * workqueues will queue new works to their delayed_works list instead of
4808 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4810 void freeze_workqueues_begin(void)
4812 struct worker_pool *pool;
4813 struct workqueue_struct *wq;
4814 struct pool_workqueue *pwq;
4817 mutex_lock(&wq_pool_mutex);
4819 WARN_ON_ONCE(workqueue_freezing);
4820 workqueue_freezing = true;
4823 for_each_pool(pool, pi) {
4824 spin_lock_irq(&pool->lock);
4825 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4826 pool->flags |= POOL_FREEZING;
4827 spin_unlock_irq(&pool->lock);
4830 list_for_each_entry(wq, &workqueues, list) {
4831 mutex_lock(&wq->mutex);
4832 for_each_pwq(pwq, wq)
4833 pwq_adjust_max_active(pwq);
4834 mutex_unlock(&wq->mutex);
4837 mutex_unlock(&wq_pool_mutex);
4841 * freeze_workqueues_busy - are freezable workqueues still busy?
4843 * Check whether freezing is complete. This function must be called
4844 * between freeze_workqueues_begin() and thaw_workqueues().
4847 * Grabs and releases wq_pool_mutex.
4850 * %true if some freezable workqueues are still busy. %false if freezing
4853 bool freeze_workqueues_busy(void)
4856 struct workqueue_struct *wq;
4857 struct pool_workqueue *pwq;
4859 mutex_lock(&wq_pool_mutex);
4861 WARN_ON_ONCE(!workqueue_freezing);
4863 list_for_each_entry(wq, &workqueues, list) {
4864 if (!(wq->flags & WQ_FREEZABLE))
4867 * nr_active is monotonically decreasing. It's safe
4868 * to peek without lock.
4870 rcu_read_lock_sched();
4871 for_each_pwq(pwq, wq) {
4872 WARN_ON_ONCE(pwq->nr_active < 0);
4873 if (pwq->nr_active) {
4875 rcu_read_unlock_sched();
4879 rcu_read_unlock_sched();
4882 mutex_unlock(&wq_pool_mutex);
4887 * thaw_workqueues - thaw workqueues
4889 * Thaw workqueues. Normal queueing is restored and all collected
4890 * frozen works are transferred to their respective pool worklists.
4893 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4895 void thaw_workqueues(void)
4897 struct workqueue_struct *wq;
4898 struct pool_workqueue *pwq;
4899 struct worker_pool *pool;
4902 mutex_lock(&wq_pool_mutex);
4904 if (!workqueue_freezing)
4907 /* clear FREEZING */
4908 for_each_pool(pool, pi) {
4909 spin_lock_irq(&pool->lock);
4910 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4911 pool->flags &= ~POOL_FREEZING;
4912 spin_unlock_irq(&pool->lock);
4915 /* restore max_active and repopulate worklist */
4916 list_for_each_entry(wq, &workqueues, list) {
4917 mutex_lock(&wq->mutex);
4918 for_each_pwq(pwq, wq)
4919 pwq_adjust_max_active(pwq);
4920 mutex_unlock(&wq->mutex);
4923 workqueue_freezing = false;
4925 mutex_unlock(&wq_pool_mutex);
4927 #endif /* CONFIG_FREEZER */
4929 static void __init wq_numa_init(void)
4934 /* determine NUMA pwq table len - highest node id + 1 */
4936 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4938 if (num_possible_nodes() <= 1)
4941 if (wq_disable_numa) {
4942 pr_info("workqueue: NUMA affinity support disabled\n");
4946 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4947 BUG_ON(!wq_update_unbound_numa_attrs_buf);
4950 * We want masks of possible CPUs of each node which isn't readily
4951 * available. Build one from cpu_to_node() which should have been
4952 * fully initialized by now.
4954 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4958 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4959 node_online(node) ? node : NUMA_NO_NODE));
4961 for_each_possible_cpu(cpu) {
4962 node = cpu_to_node(cpu);
4963 if (WARN_ON(node == NUMA_NO_NODE)) {
4964 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4965 /* happens iff arch is bonkers, let's just proceed */
4968 cpumask_set_cpu(cpu, tbl[node]);
4971 wq_numa_possible_cpumask = tbl;
4972 wq_numa_enabled = true;
4975 static int __init init_workqueues(void)
4977 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4980 /* make sure we have enough bits for OFFQ pool ID */
4981 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4982 WORK_CPU_END * NR_STD_WORKER_POOLS);
4984 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4986 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4988 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4989 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4993 /* initialize CPU pools */
4994 for_each_possible_cpu(cpu) {
4995 struct worker_pool *pool;
4998 for_each_cpu_worker_pool(pool, cpu) {
4999 BUG_ON(init_worker_pool(pool));
5001 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5002 pool->attrs->nice = std_nice[i++];
5003 pool->node = cpu_to_node(cpu);
5006 mutex_lock(&wq_pool_mutex);
5007 BUG_ON(worker_pool_assign_id(pool));
5008 mutex_unlock(&wq_pool_mutex);
5012 /* create the initial worker */
5013 for_each_online_cpu(cpu) {
5014 struct worker_pool *pool;
5016 for_each_cpu_worker_pool(pool, cpu) {
5017 pool->flags &= ~POOL_DISASSOCIATED;
5018 BUG_ON(create_and_start_worker(pool) < 0);
5022 /* create default unbound and ordered wq attrs */
5023 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5024 struct workqueue_attrs *attrs;
5026 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5027 attrs->nice = std_nice[i];
5028 unbound_std_wq_attrs[i] = attrs;
5031 * An ordered wq should have only one pwq as ordering is
5032 * guaranteed by max_active which is enforced by pwqs.
5033 * Turn off NUMA so that dfl_pwq is used for all nodes.
5035 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5036 attrs->nice = std_nice[i];
5037 attrs->no_numa = true;
5038 ordered_wq_attrs[i] = attrs;
5041 system_wq = alloc_workqueue("events", 0, 0);
5042 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5043 system_long_wq = alloc_workqueue("events_long", 0, 0);
5044 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5045 WQ_UNBOUND_MAX_ACTIVE);
5046 system_freezable_wq = alloc_workqueue("events_freezable",
5048 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5049 !system_unbound_wq || !system_freezable_wq);
5052 early_initcall(init_workqueues);