2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * manager_mutex to avoid changing binding state while
69 * create_worker() is in progress.
71 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
72 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
73 POOL_FREEZING = 1 << 3, /* freeze in progress */
76 WORKER_STARTED = 1 << 0, /* started */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give -20.
105 RESCUER_NICE_LEVEL = -20,
106 HIGHPRI_NICE_LEVEL = -20,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * MG: pool->manager_mutex and pool->lock protected. Writes require both
128 * locks. Reads can happen under either lock.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
134 * WQ: wq->mutex protected.
136 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138 * MD: wq_mayday_lock protected.
141 /* struct worker is defined in workqueue_internal.h */
144 spinlock_t lock; /* the pool lock */
145 int cpu; /* I: the associated cpu */
146 int node; /* I: the associated node ID */
147 int id; /* I: pool ID */
148 unsigned int flags; /* X: flags */
150 struct list_head worklist; /* L: list of pending works */
151 int nr_workers; /* L: total number of workers */
153 /* nr_idle includes the ones off idle_list for rebinding */
154 int nr_idle; /* L: currently idle ones */
156 struct list_head idle_list; /* X: list of idle workers */
157 struct timer_list idle_timer; /* L: worker idle timeout */
158 struct timer_list mayday_timer; /* L: SOS timer for workers */
160 /* a workers is either on busy_hash or idle_list, or the manager */
161 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
162 /* L: hash of busy workers */
164 /* see manage_workers() for details on the two manager mutexes */
165 struct mutex manager_arb; /* manager arbitration */
166 struct mutex manager_mutex; /* manager exclusion */
167 struct idr worker_idr; /* MG: worker IDs and iteration */
169 struct workqueue_attrs *attrs; /* I: worker attributes */
170 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
171 int refcnt; /* PL: refcnt for unbound pools */
174 * The current concurrency level. As it's likely to be accessed
175 * from other CPUs during try_to_wake_up(), put it in a separate
178 atomic_t nr_running ____cacheline_aligned_in_smp;
181 * Destruction of pool is sched-RCU protected to allow dereferences
182 * from get_work_pool().
185 } ____cacheline_aligned_in_smp;
188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
193 struct pool_workqueue {
194 struct worker_pool *pool; /* I: the associated pool */
195 struct workqueue_struct *wq; /* I: the owning workqueue */
196 int work_color; /* L: current color */
197 int flush_color; /* L: flushing color */
198 int refcnt; /* L: reference count */
199 int nr_in_flight[WORK_NR_COLORS];
200 /* L: nr of in_flight works */
201 int nr_active; /* L: nr of active works */
202 int max_active; /* L: max active works */
203 struct list_head delayed_works; /* L: delayed works */
204 struct list_head pwqs_node; /* WR: node on wq->pwqs */
205 struct list_head mayday_node; /* MD: node on wq->maydays */
208 * Release of unbound pwq is punted to system_wq. See put_pwq()
209 * and pwq_unbound_release_workfn() for details. pool_workqueue
210 * itself is also sched-RCU protected so that the first pwq can be
211 * determined without grabbing wq->mutex.
213 struct work_struct unbound_release_work;
215 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
218 * Structure used to wait for workqueue flush.
221 struct list_head list; /* WQ: list of flushers */
222 int flush_color; /* WQ: flush color waiting for */
223 struct completion done; /* flush completion */
229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
232 struct workqueue_struct {
233 struct list_head pwqs; /* WR: all pwqs of this wq */
234 struct list_head list; /* PL: list of all workqueues */
236 struct mutex mutex; /* protects this wq */
237 int work_color; /* WQ: current work color */
238 int flush_color; /* WQ: current flush color */
239 atomic_t nr_pwqs_to_flush; /* flush in progress */
240 struct wq_flusher *first_flusher; /* WQ: first flusher */
241 struct list_head flusher_queue; /* WQ: flush waiters */
242 struct list_head flusher_overflow; /* WQ: flush overflow list */
244 struct list_head maydays; /* MD: pwqs requesting rescue */
245 struct worker *rescuer; /* I: rescue worker */
247 int nr_drainers; /* WQ: drain in progress */
248 int saved_max_active; /* WQ: saved pwq max_active */
250 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
251 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
254 struct wq_device *wq_dev; /* I: for sysfs interface */
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map;
259 char name[WQ_NAME_LEN]; /* I: workqueue name */
261 /* hot fields used during command issue, aligned to cacheline */
262 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
263 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
264 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
267 static struct kmem_cache *pwq_cache;
269 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
270 static cpumask_var_t *wq_numa_possible_cpumask;
271 /* possible CPUs of each node */
273 static bool wq_disable_numa;
274 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
276 /* see the comment above the definition of WQ_POWER_EFFICIENT */
277 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
278 static bool wq_power_efficient = true;
280 static bool wq_power_efficient;
283 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
285 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
287 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
288 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
290 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
291 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
293 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
294 static bool workqueue_freezing; /* PL: have wqs started freezing? */
296 /* the per-cpu worker pools */
297 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
300 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
302 /* PL: hash of all unbound pools keyed by pool->attrs */
303 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
305 /* I: attributes used when instantiating standard unbound pools on demand */
306 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
308 /* I: attributes used when instantiating ordered pools on demand */
309 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
311 struct workqueue_struct *system_wq __read_mostly;
312 EXPORT_SYMBOL(system_wq);
313 struct workqueue_struct *system_highpri_wq __read_mostly;
314 EXPORT_SYMBOL_GPL(system_highpri_wq);
315 struct workqueue_struct *system_long_wq __read_mostly;
316 EXPORT_SYMBOL_GPL(system_long_wq);
317 struct workqueue_struct *system_unbound_wq __read_mostly;
318 EXPORT_SYMBOL_GPL(system_unbound_wq);
319 struct workqueue_struct *system_freezable_wq __read_mostly;
320 EXPORT_SYMBOL_GPL(system_freezable_wq);
321 struct workqueue_struct *system_power_efficient_wq __read_mostly;
322 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
323 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
326 static int worker_thread(void *__worker);
327 static void copy_workqueue_attrs(struct workqueue_attrs *to,
328 const struct workqueue_attrs *from);
330 #define CREATE_TRACE_POINTS
331 #include <trace/events/workqueue.h>
333 #define assert_rcu_or_pool_mutex() \
334 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
335 lockdep_is_held(&wq_pool_mutex), \
336 "sched RCU or wq_pool_mutex should be held")
338 #define assert_rcu_or_wq_mutex(wq) \
339 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
340 lockdep_is_held(&wq->mutex), \
341 "sched RCU or wq->mutex should be held")
343 #ifdef CONFIG_LOCKDEP
344 #define assert_manager_or_pool_lock(pool) \
345 WARN_ONCE(debug_locks && \
346 !lockdep_is_held(&(pool)->manager_mutex) && \
347 !lockdep_is_held(&(pool)->lock), \
348 "pool->manager_mutex or ->lock should be held")
350 #define assert_manager_or_pool_lock(pool) do { } while (0)
353 #define for_each_cpu_worker_pool(pool, cpu) \
354 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
355 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
359 * for_each_pool - iterate through all worker_pools in the system
360 * @pool: iteration cursor
361 * @pi: integer used for iteration
363 * This must be called either with wq_pool_mutex held or sched RCU read
364 * locked. If the pool needs to be used beyond the locking in effect, the
365 * caller is responsible for guaranteeing that the pool stays online.
367 * The if/else clause exists only for the lockdep assertion and can be
370 #define for_each_pool(pool, pi) \
371 idr_for_each_entry(&worker_pool_idr, pool, pi) \
372 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
376 * for_each_pool_worker - iterate through all workers of a worker_pool
377 * @worker: iteration cursor
378 * @wi: integer used for iteration
379 * @pool: worker_pool to iterate workers of
381 * This must be called with either @pool->manager_mutex or ->lock held.
383 * The if/else clause exists only for the lockdep assertion and can be
386 #define for_each_pool_worker(worker, wi, pool) \
387 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
388 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
392 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
393 * @pwq: iteration cursor
394 * @wq: the target workqueue
396 * This must be called either with wq->mutex held or sched RCU read locked.
397 * If the pwq needs to be used beyond the locking in effect, the caller is
398 * responsible for guaranteeing that the pwq stays online.
400 * The if/else clause exists only for the lockdep assertion and can be
403 #define for_each_pwq(pwq, wq) \
404 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
405 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
408 #ifdef CONFIG_DEBUG_OBJECTS_WORK
410 static struct debug_obj_descr work_debug_descr;
412 static void *work_debug_hint(void *addr)
414 return ((struct work_struct *) addr)->func;
418 * fixup_init is called when:
419 * - an active object is initialized
421 static int work_fixup_init(void *addr, enum debug_obj_state state)
423 struct work_struct *work = addr;
426 case ODEBUG_STATE_ACTIVE:
427 cancel_work_sync(work);
428 debug_object_init(work, &work_debug_descr);
436 * fixup_activate is called when:
437 * - an active object is activated
438 * - an unknown object is activated (might be a statically initialized object)
440 static int work_fixup_activate(void *addr, enum debug_obj_state state)
442 struct work_struct *work = addr;
446 case ODEBUG_STATE_NOTAVAILABLE:
448 * This is not really a fixup. The work struct was
449 * statically initialized. We just make sure that it
450 * is tracked in the object tracker.
452 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
453 debug_object_init(work, &work_debug_descr);
454 debug_object_activate(work, &work_debug_descr);
460 case ODEBUG_STATE_ACTIVE:
469 * fixup_free is called when:
470 * - an active object is freed
472 static int work_fixup_free(void *addr, enum debug_obj_state state)
474 struct work_struct *work = addr;
477 case ODEBUG_STATE_ACTIVE:
478 cancel_work_sync(work);
479 debug_object_free(work, &work_debug_descr);
486 static struct debug_obj_descr work_debug_descr = {
487 .name = "work_struct",
488 .debug_hint = work_debug_hint,
489 .fixup_init = work_fixup_init,
490 .fixup_activate = work_fixup_activate,
491 .fixup_free = work_fixup_free,
494 static inline void debug_work_activate(struct work_struct *work)
496 debug_object_activate(work, &work_debug_descr);
499 static inline void debug_work_deactivate(struct work_struct *work)
501 debug_object_deactivate(work, &work_debug_descr);
504 void __init_work(struct work_struct *work, int onstack)
507 debug_object_init_on_stack(work, &work_debug_descr);
509 debug_object_init(work, &work_debug_descr);
511 EXPORT_SYMBOL_GPL(__init_work);
513 void destroy_work_on_stack(struct work_struct *work)
515 debug_object_free(work, &work_debug_descr);
517 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
519 void destroy_delayed_work_on_stack(struct delayed_work *work)
521 destroy_timer_on_stack(&work->timer);
522 debug_object_free(&work->work, &work_debug_descr);
524 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
527 static inline void debug_work_activate(struct work_struct *work) { }
528 static inline void debug_work_deactivate(struct work_struct *work) { }
532 * worker_pool_assign_id - allocate ID and assing it to @pool
533 * @pool: the pool pointer of interest
535 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
536 * successfully, -errno on failure.
538 static int worker_pool_assign_id(struct worker_pool *pool)
542 lockdep_assert_held(&wq_pool_mutex);
544 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
554 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
555 * @wq: the target workqueue
558 * This must be called either with pwq_lock held or sched RCU read locked.
559 * If the pwq needs to be used beyond the locking in effect, the caller is
560 * responsible for guaranteeing that the pwq stays online.
562 * Return: The unbound pool_workqueue for @node.
564 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
567 assert_rcu_or_wq_mutex(wq);
568 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
571 static unsigned int work_color_to_flags(int color)
573 return color << WORK_STRUCT_COLOR_SHIFT;
576 static int get_work_color(struct work_struct *work)
578 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
579 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
582 static int work_next_color(int color)
584 return (color + 1) % WORK_NR_COLORS;
588 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
589 * contain the pointer to the queued pwq. Once execution starts, the flag
590 * is cleared and the high bits contain OFFQ flags and pool ID.
592 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
593 * and clear_work_data() can be used to set the pwq, pool or clear
594 * work->data. These functions should only be called while the work is
595 * owned - ie. while the PENDING bit is set.
597 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
598 * corresponding to a work. Pool is available once the work has been
599 * queued anywhere after initialization until it is sync canceled. pwq is
600 * available only while the work item is queued.
602 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
603 * canceled. While being canceled, a work item may have its PENDING set
604 * but stay off timer and worklist for arbitrarily long and nobody should
605 * try to steal the PENDING bit.
607 static inline void set_work_data(struct work_struct *work, unsigned long data,
610 WARN_ON_ONCE(!work_pending(work));
611 atomic_long_set(&work->data, data | flags | work_static(work));
614 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
615 unsigned long extra_flags)
617 set_work_data(work, (unsigned long)pwq,
618 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
621 static void set_work_pool_and_keep_pending(struct work_struct *work,
624 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
625 WORK_STRUCT_PENDING);
628 static void set_work_pool_and_clear_pending(struct work_struct *work,
632 * The following wmb is paired with the implied mb in
633 * test_and_set_bit(PENDING) and ensures all updates to @work made
634 * here are visible to and precede any updates by the next PENDING
638 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
641 static void clear_work_data(struct work_struct *work)
643 smp_wmb(); /* see set_work_pool_and_clear_pending() */
644 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
647 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
649 unsigned long data = atomic_long_read(&work->data);
651 if (data & WORK_STRUCT_PWQ)
652 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
658 * get_work_pool - return the worker_pool a given work was associated with
659 * @work: the work item of interest
661 * Pools are created and destroyed under wq_pool_mutex, and allows read
662 * access under sched-RCU read lock. As such, this function should be
663 * called under wq_pool_mutex or with preemption disabled.
665 * All fields of the returned pool are accessible as long as the above
666 * mentioned locking is in effect. If the returned pool needs to be used
667 * beyond the critical section, the caller is responsible for ensuring the
668 * returned pool is and stays online.
670 * Return: The worker_pool @work was last associated with. %NULL if none.
672 static struct worker_pool *get_work_pool(struct work_struct *work)
674 unsigned long data = atomic_long_read(&work->data);
677 assert_rcu_or_pool_mutex();
679 if (data & WORK_STRUCT_PWQ)
680 return ((struct pool_workqueue *)
681 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
683 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
684 if (pool_id == WORK_OFFQ_POOL_NONE)
687 return idr_find(&worker_pool_idr, pool_id);
691 * get_work_pool_id - return the worker pool ID a given work is associated with
692 * @work: the work item of interest
694 * Return: The worker_pool ID @work was last associated with.
695 * %WORK_OFFQ_POOL_NONE if none.
697 static int get_work_pool_id(struct work_struct *work)
699 unsigned long data = atomic_long_read(&work->data);
701 if (data & WORK_STRUCT_PWQ)
702 return ((struct pool_workqueue *)
703 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
705 return data >> WORK_OFFQ_POOL_SHIFT;
708 static void mark_work_canceling(struct work_struct *work)
710 unsigned long pool_id = get_work_pool_id(work);
712 pool_id <<= WORK_OFFQ_POOL_SHIFT;
713 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
716 static bool work_is_canceling(struct work_struct *work)
718 unsigned long data = atomic_long_read(&work->data);
720 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
724 * Policy functions. These define the policies on how the global worker
725 * pools are managed. Unless noted otherwise, these functions assume that
726 * they're being called with pool->lock held.
729 static bool __need_more_worker(struct worker_pool *pool)
731 return !atomic_read(&pool->nr_running);
735 * Need to wake up a worker? Called from anything but currently
738 * Note that, because unbound workers never contribute to nr_running, this
739 * function will always return %true for unbound pools as long as the
740 * worklist isn't empty.
742 static bool need_more_worker(struct worker_pool *pool)
744 return !list_empty(&pool->worklist) && __need_more_worker(pool);
747 /* Can I start working? Called from busy but !running workers. */
748 static bool may_start_working(struct worker_pool *pool)
750 return pool->nr_idle;
753 /* Do I need to keep working? Called from currently running workers. */
754 static bool keep_working(struct worker_pool *pool)
756 return !list_empty(&pool->worklist) &&
757 atomic_read(&pool->nr_running) <= 1;
760 /* Do we need a new worker? Called from manager. */
761 static bool need_to_create_worker(struct worker_pool *pool)
763 return need_more_worker(pool) && !may_start_working(pool);
766 /* Do I need to be the manager? */
767 static bool need_to_manage_workers(struct worker_pool *pool)
769 return need_to_create_worker(pool) ||
770 (pool->flags & POOL_MANAGE_WORKERS);
773 /* Do we have too many workers and should some go away? */
774 static bool too_many_workers(struct worker_pool *pool)
776 bool managing = mutex_is_locked(&pool->manager_arb);
777 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
778 int nr_busy = pool->nr_workers - nr_idle;
781 * nr_idle and idle_list may disagree if idle rebinding is in
782 * progress. Never return %true if idle_list is empty.
784 if (list_empty(&pool->idle_list))
787 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
794 /* Return the first worker. Safe with preemption disabled */
795 static struct worker *first_worker(struct worker_pool *pool)
797 if (unlikely(list_empty(&pool->idle_list)))
800 return list_first_entry(&pool->idle_list, struct worker, entry);
804 * wake_up_worker - wake up an idle worker
805 * @pool: worker pool to wake worker from
807 * Wake up the first idle worker of @pool.
810 * spin_lock_irq(pool->lock).
812 static void wake_up_worker(struct worker_pool *pool)
814 struct worker *worker = first_worker(pool);
817 wake_up_process(worker->task);
821 * wq_worker_waking_up - a worker is waking up
822 * @task: task waking up
823 * @cpu: CPU @task is waking up to
825 * This function is called during try_to_wake_up() when a worker is
829 * spin_lock_irq(rq->lock)
831 void wq_worker_waking_up(struct task_struct *task, int cpu)
833 struct worker *worker = kthread_data(task);
835 if (!(worker->flags & WORKER_NOT_RUNNING)) {
836 WARN_ON_ONCE(worker->pool->cpu != cpu);
837 atomic_inc(&worker->pool->nr_running);
842 * wq_worker_sleeping - a worker is going to sleep
843 * @task: task going to sleep
844 * @cpu: CPU in question, must be the current CPU number
846 * This function is called during schedule() when a busy worker is
847 * going to sleep. Worker on the same cpu can be woken up by
848 * returning pointer to its task.
851 * spin_lock_irq(rq->lock)
854 * Worker task on @cpu to wake up, %NULL if none.
856 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
858 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
859 struct worker_pool *pool;
862 * Rescuers, which may not have all the fields set up like normal
863 * workers, also reach here, let's not access anything before
864 * checking NOT_RUNNING.
866 if (worker->flags & WORKER_NOT_RUNNING)
871 /* this can only happen on the local cpu */
872 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
876 * The counterpart of the following dec_and_test, implied mb,
877 * worklist not empty test sequence is in insert_work().
878 * Please read comment there.
880 * NOT_RUNNING is clear. This means that we're bound to and
881 * running on the local cpu w/ rq lock held and preemption
882 * disabled, which in turn means that none else could be
883 * manipulating idle_list, so dereferencing idle_list without pool
886 if (atomic_dec_and_test(&pool->nr_running) &&
887 !list_empty(&pool->worklist))
888 to_wakeup = first_worker(pool);
889 return to_wakeup ? to_wakeup->task : NULL;
893 * worker_set_flags - set worker flags and adjust nr_running accordingly
895 * @flags: flags to set
896 * @wakeup: wakeup an idle worker if necessary
898 * Set @flags in @worker->flags and adjust nr_running accordingly. If
899 * nr_running becomes zero and @wakeup is %true, an idle worker is
903 * spin_lock_irq(pool->lock)
905 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
908 struct worker_pool *pool = worker->pool;
910 WARN_ON_ONCE(worker->task != current);
913 * If transitioning into NOT_RUNNING, adjust nr_running and
914 * wake up an idle worker as necessary if requested by
917 if ((flags & WORKER_NOT_RUNNING) &&
918 !(worker->flags & WORKER_NOT_RUNNING)) {
920 if (atomic_dec_and_test(&pool->nr_running) &&
921 !list_empty(&pool->worklist))
922 wake_up_worker(pool);
924 atomic_dec(&pool->nr_running);
927 worker->flags |= flags;
931 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
933 * @flags: flags to clear
935 * Clear @flags in @worker->flags and adjust nr_running accordingly.
938 * spin_lock_irq(pool->lock)
940 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
942 struct worker_pool *pool = worker->pool;
943 unsigned int oflags = worker->flags;
945 WARN_ON_ONCE(worker->task != current);
947 worker->flags &= ~flags;
950 * If transitioning out of NOT_RUNNING, increment nr_running. Note
951 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
952 * of multiple flags, not a single flag.
954 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
955 if (!(worker->flags & WORKER_NOT_RUNNING))
956 atomic_inc(&pool->nr_running);
960 * find_worker_executing_work - find worker which is executing a work
961 * @pool: pool of interest
962 * @work: work to find worker for
964 * Find a worker which is executing @work on @pool by searching
965 * @pool->busy_hash which is keyed by the address of @work. For a worker
966 * to match, its current execution should match the address of @work and
967 * its work function. This is to avoid unwanted dependency between
968 * unrelated work executions through a work item being recycled while still
971 * This is a bit tricky. A work item may be freed once its execution
972 * starts and nothing prevents the freed area from being recycled for
973 * another work item. If the same work item address ends up being reused
974 * before the original execution finishes, workqueue will identify the
975 * recycled work item as currently executing and make it wait until the
976 * current execution finishes, introducing an unwanted dependency.
978 * This function checks the work item address and work function to avoid
979 * false positives. Note that this isn't complete as one may construct a
980 * work function which can introduce dependency onto itself through a
981 * recycled work item. Well, if somebody wants to shoot oneself in the
982 * foot that badly, there's only so much we can do, and if such deadlock
983 * actually occurs, it should be easy to locate the culprit work function.
986 * spin_lock_irq(pool->lock).
989 * Pointer to worker which is executing @work if found, %NULL
992 static struct worker *find_worker_executing_work(struct worker_pool *pool,
993 struct work_struct *work)
995 struct worker *worker;
997 hash_for_each_possible(pool->busy_hash, worker, hentry,
999 if (worker->current_work == work &&
1000 worker->current_func == work->func)
1007 * move_linked_works - move linked works to a list
1008 * @work: start of series of works to be scheduled
1009 * @head: target list to append @work to
1010 * @nextp: out paramter for nested worklist walking
1012 * Schedule linked works starting from @work to @head. Work series to
1013 * be scheduled starts at @work and includes any consecutive work with
1014 * WORK_STRUCT_LINKED set in its predecessor.
1016 * If @nextp is not NULL, it's updated to point to the next work of
1017 * the last scheduled work. This allows move_linked_works() to be
1018 * nested inside outer list_for_each_entry_safe().
1021 * spin_lock_irq(pool->lock).
1023 static void move_linked_works(struct work_struct *work, struct list_head *head,
1024 struct work_struct **nextp)
1026 struct work_struct *n;
1029 * Linked worklist will always end before the end of the list,
1030 * use NULL for list head.
1032 list_for_each_entry_safe_from(work, n, NULL, entry) {
1033 list_move_tail(&work->entry, head);
1034 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1039 * If we're already inside safe list traversal and have moved
1040 * multiple works to the scheduled queue, the next position
1041 * needs to be updated.
1048 * get_pwq - get an extra reference on the specified pool_workqueue
1049 * @pwq: pool_workqueue to get
1051 * Obtain an extra reference on @pwq. The caller should guarantee that
1052 * @pwq has positive refcnt and be holding the matching pool->lock.
1054 static void get_pwq(struct pool_workqueue *pwq)
1056 lockdep_assert_held(&pwq->pool->lock);
1057 WARN_ON_ONCE(pwq->refcnt <= 0);
1062 * put_pwq - put a pool_workqueue reference
1063 * @pwq: pool_workqueue to put
1065 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1066 * destruction. The caller should be holding the matching pool->lock.
1068 static void put_pwq(struct pool_workqueue *pwq)
1070 lockdep_assert_held(&pwq->pool->lock);
1071 if (likely(--pwq->refcnt))
1073 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1076 * @pwq can't be released under pool->lock, bounce to
1077 * pwq_unbound_release_workfn(). This never recurses on the same
1078 * pool->lock as this path is taken only for unbound workqueues and
1079 * the release work item is scheduled on a per-cpu workqueue. To
1080 * avoid lockdep warning, unbound pool->locks are given lockdep
1081 * subclass of 1 in get_unbound_pool().
1083 schedule_work(&pwq->unbound_release_work);
1087 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1088 * @pwq: pool_workqueue to put (can be %NULL)
1090 * put_pwq() with locking. This function also allows %NULL @pwq.
1092 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1096 * As both pwqs and pools are sched-RCU protected, the
1097 * following lock operations are safe.
1099 spin_lock_irq(&pwq->pool->lock);
1101 spin_unlock_irq(&pwq->pool->lock);
1105 static void pwq_activate_delayed_work(struct work_struct *work)
1107 struct pool_workqueue *pwq = get_work_pwq(work);
1109 trace_workqueue_activate_work(work);
1110 move_linked_works(work, &pwq->pool->worklist, NULL);
1111 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1115 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1117 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1118 struct work_struct, entry);
1120 pwq_activate_delayed_work(work);
1124 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1125 * @pwq: pwq of interest
1126 * @color: color of work which left the queue
1128 * A work either has completed or is removed from pending queue,
1129 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1132 * spin_lock_irq(pool->lock).
1134 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1136 /* uncolored work items don't participate in flushing or nr_active */
1137 if (color == WORK_NO_COLOR)
1140 pwq->nr_in_flight[color]--;
1143 if (!list_empty(&pwq->delayed_works)) {
1144 /* one down, submit a delayed one */
1145 if (pwq->nr_active < pwq->max_active)
1146 pwq_activate_first_delayed(pwq);
1149 /* is flush in progress and are we at the flushing tip? */
1150 if (likely(pwq->flush_color != color))
1153 /* are there still in-flight works? */
1154 if (pwq->nr_in_flight[color])
1157 /* this pwq is done, clear flush_color */
1158 pwq->flush_color = -1;
1161 * If this was the last pwq, wake up the first flusher. It
1162 * will handle the rest.
1164 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1165 complete(&pwq->wq->first_flusher->done);
1171 * try_to_grab_pending - steal work item from worklist and disable irq
1172 * @work: work item to steal
1173 * @is_dwork: @work is a delayed_work
1174 * @flags: place to store irq state
1176 * Try to grab PENDING bit of @work. This function can handle @work in any
1177 * stable state - idle, on timer or on worklist.
1180 * 1 if @work was pending and we successfully stole PENDING
1181 * 0 if @work was idle and we claimed PENDING
1182 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1183 * -ENOENT if someone else is canceling @work, this state may persist
1184 * for arbitrarily long
1187 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1188 * interrupted while holding PENDING and @work off queue, irq must be
1189 * disabled on entry. This, combined with delayed_work->timer being
1190 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1192 * On successful return, >= 0, irq is disabled and the caller is
1193 * responsible for releasing it using local_irq_restore(*@flags).
1195 * This function is safe to call from any context including IRQ handler.
1197 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1198 unsigned long *flags)
1200 struct worker_pool *pool;
1201 struct pool_workqueue *pwq;
1203 local_irq_save(*flags);
1205 /* try to steal the timer if it exists */
1207 struct delayed_work *dwork = to_delayed_work(work);
1210 * dwork->timer is irqsafe. If del_timer() fails, it's
1211 * guaranteed that the timer is not queued anywhere and not
1212 * running on the local CPU.
1214 if (likely(del_timer(&dwork->timer)))
1218 /* try to claim PENDING the normal way */
1219 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1223 * The queueing is in progress, or it is already queued. Try to
1224 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1226 pool = get_work_pool(work);
1230 spin_lock(&pool->lock);
1232 * work->data is guaranteed to point to pwq only while the work
1233 * item is queued on pwq->wq, and both updating work->data to point
1234 * to pwq on queueing and to pool on dequeueing are done under
1235 * pwq->pool->lock. This in turn guarantees that, if work->data
1236 * points to pwq which is associated with a locked pool, the work
1237 * item is currently queued on that pool.
1239 pwq = get_work_pwq(work);
1240 if (pwq && pwq->pool == pool) {
1241 debug_work_deactivate(work);
1244 * A delayed work item cannot be grabbed directly because
1245 * it might have linked NO_COLOR work items which, if left
1246 * on the delayed_list, will confuse pwq->nr_active
1247 * management later on and cause stall. Make sure the work
1248 * item is activated before grabbing.
1250 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1251 pwq_activate_delayed_work(work);
1253 list_del_init(&work->entry);
1254 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1256 /* work->data points to pwq iff queued, point to pool */
1257 set_work_pool_and_keep_pending(work, pool->id);
1259 spin_unlock(&pool->lock);
1262 spin_unlock(&pool->lock);
1264 local_irq_restore(*flags);
1265 if (work_is_canceling(work))
1272 * insert_work - insert a work into a pool
1273 * @pwq: pwq @work belongs to
1274 * @work: work to insert
1275 * @head: insertion point
1276 * @extra_flags: extra WORK_STRUCT_* flags to set
1278 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1279 * work_struct flags.
1282 * spin_lock_irq(pool->lock).
1284 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1285 struct list_head *head, unsigned int extra_flags)
1287 struct worker_pool *pool = pwq->pool;
1289 /* we own @work, set data and link */
1290 set_work_pwq(work, pwq, extra_flags);
1291 list_add_tail(&work->entry, head);
1295 * Ensure either wq_worker_sleeping() sees the above
1296 * list_add_tail() or we see zero nr_running to avoid workers lying
1297 * around lazily while there are works to be processed.
1301 if (__need_more_worker(pool))
1302 wake_up_worker(pool);
1306 * Test whether @work is being queued from another work executing on the
1309 static bool is_chained_work(struct workqueue_struct *wq)
1311 struct worker *worker;
1313 worker = current_wq_worker();
1315 * Return %true iff I'm a worker execuing a work item on @wq. If
1316 * I'm @worker, it's safe to dereference it without locking.
1318 return worker && worker->current_pwq->wq == wq;
1321 static void __queue_work(int cpu, struct workqueue_struct *wq,
1322 struct work_struct *work)
1324 struct pool_workqueue *pwq;
1325 struct worker_pool *last_pool;
1326 struct list_head *worklist;
1327 unsigned int work_flags;
1328 unsigned int req_cpu = cpu;
1331 * While a work item is PENDING && off queue, a task trying to
1332 * steal the PENDING will busy-loop waiting for it to either get
1333 * queued or lose PENDING. Grabbing PENDING and queueing should
1334 * happen with IRQ disabled.
1336 WARN_ON_ONCE(!irqs_disabled());
1338 debug_work_activate(work);
1340 /* if draining, only works from the same workqueue are allowed */
1341 if (unlikely(wq->flags & __WQ_DRAINING) &&
1342 WARN_ON_ONCE(!is_chained_work(wq)))
1345 if (req_cpu == WORK_CPU_UNBOUND)
1346 cpu = raw_smp_processor_id();
1348 /* pwq which will be used unless @work is executing elsewhere */
1349 if (!(wq->flags & WQ_UNBOUND))
1350 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1352 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1355 * If @work was previously on a different pool, it might still be
1356 * running there, in which case the work needs to be queued on that
1357 * pool to guarantee non-reentrancy.
1359 last_pool = get_work_pool(work);
1360 if (last_pool && last_pool != pwq->pool) {
1361 struct worker *worker;
1363 spin_lock(&last_pool->lock);
1365 worker = find_worker_executing_work(last_pool, work);
1367 if (worker && worker->current_pwq->wq == wq) {
1368 pwq = worker->current_pwq;
1370 /* meh... not running there, queue here */
1371 spin_unlock(&last_pool->lock);
1372 spin_lock(&pwq->pool->lock);
1375 spin_lock(&pwq->pool->lock);
1379 * pwq is determined and locked. For unbound pools, we could have
1380 * raced with pwq release and it could already be dead. If its
1381 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1382 * without another pwq replacing it in the numa_pwq_tbl or while
1383 * work items are executing on it, so the retrying is guaranteed to
1384 * make forward-progress.
1386 if (unlikely(!pwq->refcnt)) {
1387 if (wq->flags & WQ_UNBOUND) {
1388 spin_unlock(&pwq->pool->lock);
1393 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1397 /* pwq determined, queue */
1398 trace_workqueue_queue_work(req_cpu, pwq, work);
1400 if (WARN_ON(!list_empty(&work->entry))) {
1401 spin_unlock(&pwq->pool->lock);
1405 pwq->nr_in_flight[pwq->work_color]++;
1406 work_flags = work_color_to_flags(pwq->work_color);
1408 if (likely(pwq->nr_active < pwq->max_active)) {
1409 trace_workqueue_activate_work(work);
1411 worklist = &pwq->pool->worklist;
1413 work_flags |= WORK_STRUCT_DELAYED;
1414 worklist = &pwq->delayed_works;
1417 insert_work(pwq, work, worklist, work_flags);
1419 spin_unlock(&pwq->pool->lock);
1423 * queue_work_on - queue work on specific cpu
1424 * @cpu: CPU number to execute work on
1425 * @wq: workqueue to use
1426 * @work: work to queue
1428 * We queue the work to a specific CPU, the caller must ensure it
1431 * Return: %false if @work was already on a queue, %true otherwise.
1433 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1434 struct work_struct *work)
1437 unsigned long flags;
1439 local_irq_save(flags);
1441 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1442 __queue_work(cpu, wq, work);
1446 local_irq_restore(flags);
1449 EXPORT_SYMBOL(queue_work_on);
1451 void delayed_work_timer_fn(unsigned long __data)
1453 struct delayed_work *dwork = (struct delayed_work *)__data;
1455 /* should have been called from irqsafe timer with irq already off */
1456 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1458 EXPORT_SYMBOL(delayed_work_timer_fn);
1460 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1461 struct delayed_work *dwork, unsigned long delay)
1463 struct timer_list *timer = &dwork->timer;
1464 struct work_struct *work = &dwork->work;
1466 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1467 timer->data != (unsigned long)dwork);
1468 WARN_ON_ONCE(timer_pending(timer));
1469 WARN_ON_ONCE(!list_empty(&work->entry));
1472 * If @delay is 0, queue @dwork->work immediately. This is for
1473 * both optimization and correctness. The earliest @timer can
1474 * expire is on the closest next tick and delayed_work users depend
1475 * on that there's no such delay when @delay is 0.
1478 __queue_work(cpu, wq, &dwork->work);
1482 timer_stats_timer_set_start_info(&dwork->timer);
1486 timer->expires = jiffies + delay;
1488 if (unlikely(cpu != WORK_CPU_UNBOUND))
1489 add_timer_on(timer, cpu);
1495 * queue_delayed_work_on - queue work on specific CPU after delay
1496 * @cpu: CPU number to execute work on
1497 * @wq: workqueue to use
1498 * @dwork: work to queue
1499 * @delay: number of jiffies to wait before queueing
1501 * Return: %false if @work was already on a queue, %true otherwise. If
1502 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1505 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1506 struct delayed_work *dwork, unsigned long delay)
1508 struct work_struct *work = &dwork->work;
1510 unsigned long flags;
1512 /* read the comment in __queue_work() */
1513 local_irq_save(flags);
1515 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1516 __queue_delayed_work(cpu, wq, dwork, delay);
1520 local_irq_restore(flags);
1523 EXPORT_SYMBOL(queue_delayed_work_on);
1526 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1527 * @cpu: CPU number to execute work on
1528 * @wq: workqueue to use
1529 * @dwork: work to queue
1530 * @delay: number of jiffies to wait before queueing
1532 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1533 * modify @dwork's timer so that it expires after @delay. If @delay is
1534 * zero, @work is guaranteed to be scheduled immediately regardless of its
1537 * Return: %false if @dwork was idle and queued, %true if @dwork was
1538 * pending and its timer was modified.
1540 * This function is safe to call from any context including IRQ handler.
1541 * See try_to_grab_pending() for details.
1543 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1544 struct delayed_work *dwork, unsigned long delay)
1546 unsigned long flags;
1550 ret = try_to_grab_pending(&dwork->work, true, &flags);
1551 } while (unlikely(ret == -EAGAIN));
1553 if (likely(ret >= 0)) {
1554 __queue_delayed_work(cpu, wq, dwork, delay);
1555 local_irq_restore(flags);
1558 /* -ENOENT from try_to_grab_pending() becomes %true */
1561 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1564 * worker_enter_idle - enter idle state
1565 * @worker: worker which is entering idle state
1567 * @worker is entering idle state. Update stats and idle timer if
1571 * spin_lock_irq(pool->lock).
1573 static void worker_enter_idle(struct worker *worker)
1575 struct worker_pool *pool = worker->pool;
1577 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1578 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1579 (worker->hentry.next || worker->hentry.pprev)))
1582 /* can't use worker_set_flags(), also called from start_worker() */
1583 worker->flags |= WORKER_IDLE;
1585 worker->last_active = jiffies;
1587 /* idle_list is LIFO */
1588 list_add(&worker->entry, &pool->idle_list);
1590 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1591 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1594 * Sanity check nr_running. Because wq_unbind_fn() releases
1595 * pool->lock between setting %WORKER_UNBOUND and zapping
1596 * nr_running, the warning may trigger spuriously. Check iff
1597 * unbind is not in progress.
1599 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1600 pool->nr_workers == pool->nr_idle &&
1601 atomic_read(&pool->nr_running));
1605 * worker_leave_idle - leave idle state
1606 * @worker: worker which is leaving idle state
1608 * @worker is leaving idle state. Update stats.
1611 * spin_lock_irq(pool->lock).
1613 static void worker_leave_idle(struct worker *worker)
1615 struct worker_pool *pool = worker->pool;
1617 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1619 worker_clr_flags(worker, WORKER_IDLE);
1621 list_del_init(&worker->entry);
1625 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1626 * @pool: target worker_pool
1628 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1630 * Works which are scheduled while the cpu is online must at least be
1631 * scheduled to a worker which is bound to the cpu so that if they are
1632 * flushed from cpu callbacks while cpu is going down, they are
1633 * guaranteed to execute on the cpu.
1635 * This function is to be used by unbound workers and rescuers to bind
1636 * themselves to the target cpu and may race with cpu going down or
1637 * coming online. kthread_bind() can't be used because it may put the
1638 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1639 * verbatim as it's best effort and blocking and pool may be
1640 * [dis]associated in the meantime.
1642 * This function tries set_cpus_allowed() and locks pool and verifies the
1643 * binding against %POOL_DISASSOCIATED which is set during
1644 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1645 * enters idle state or fetches works without dropping lock, it can
1646 * guarantee the scheduling requirement described in the first paragraph.
1649 * Might sleep. Called without any lock but returns with pool->lock
1653 * %true if the associated pool is online (@worker is successfully
1654 * bound), %false if offline.
1656 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1657 __acquires(&pool->lock)
1661 * The following call may fail, succeed or succeed
1662 * without actually migrating the task to the cpu if
1663 * it races with cpu hotunplug operation. Verify
1664 * against POOL_DISASSOCIATED.
1666 if (!(pool->flags & POOL_DISASSOCIATED))
1667 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1669 spin_lock_irq(&pool->lock);
1670 if (pool->flags & POOL_DISASSOCIATED)
1672 if (task_cpu(current) == pool->cpu &&
1673 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1675 spin_unlock_irq(&pool->lock);
1678 * We've raced with CPU hot[un]plug. Give it a breather
1679 * and retry migration. cond_resched() is required here;
1680 * otherwise, we might deadlock against cpu_stop trying to
1681 * bring down the CPU on non-preemptive kernel.
1688 static struct worker *alloc_worker(void)
1690 struct worker *worker;
1692 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1694 INIT_LIST_HEAD(&worker->entry);
1695 INIT_LIST_HEAD(&worker->scheduled);
1696 /* on creation a worker is in !idle && prep state */
1697 worker->flags = WORKER_PREP;
1703 * create_worker - create a new workqueue worker
1704 * @pool: pool the new worker will belong to
1706 * Create a new worker which is bound to @pool. The returned worker
1707 * can be started by calling start_worker() or destroyed using
1711 * Might sleep. Does GFP_KERNEL allocations.
1714 * Pointer to the newly created worker.
1716 static struct worker *create_worker(struct worker_pool *pool)
1718 struct worker *worker = NULL;
1722 lockdep_assert_held(&pool->manager_mutex);
1725 * ID is needed to determine kthread name. Allocate ID first
1726 * without installing the pointer.
1728 idr_preload(GFP_KERNEL);
1729 spin_lock_irq(&pool->lock);
1731 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1733 spin_unlock_irq(&pool->lock);
1738 worker = alloc_worker();
1742 worker->pool = pool;
1746 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1747 pool->attrs->nice < 0 ? "H" : "");
1749 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1751 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1752 "kworker/%s", id_buf);
1753 if (IS_ERR(worker->task))
1756 set_user_nice(worker->task, pool->attrs->nice);
1758 /* prevent userland from meddling with cpumask of workqueue workers */
1759 worker->task->flags |= PF_NO_SETAFFINITY;
1762 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1763 * online CPUs. It'll be re-applied when any of the CPUs come up.
1765 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1768 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1769 * remains stable across this function. See the comments above the
1770 * flag definition for details.
1772 if (pool->flags & POOL_DISASSOCIATED)
1773 worker->flags |= WORKER_UNBOUND;
1775 /* successful, commit the pointer to idr */
1776 spin_lock_irq(&pool->lock);
1777 idr_replace(&pool->worker_idr, worker, worker->id);
1778 spin_unlock_irq(&pool->lock);
1784 spin_lock_irq(&pool->lock);
1785 idr_remove(&pool->worker_idr, id);
1786 spin_unlock_irq(&pool->lock);
1793 * start_worker - start a newly created worker
1794 * @worker: worker to start
1796 * Make the pool aware of @worker and start it.
1799 * spin_lock_irq(pool->lock).
1801 static void start_worker(struct worker *worker)
1803 worker->flags |= WORKER_STARTED;
1804 worker->pool->nr_workers++;
1805 worker_enter_idle(worker);
1806 wake_up_process(worker->task);
1810 * create_and_start_worker - create and start a worker for a pool
1811 * @pool: the target pool
1813 * Grab the managership of @pool and create and start a new worker for it.
1815 * Return: 0 on success. A negative error code otherwise.
1817 static int create_and_start_worker(struct worker_pool *pool)
1819 struct worker *worker;
1821 mutex_lock(&pool->manager_mutex);
1823 worker = create_worker(pool);
1825 spin_lock_irq(&pool->lock);
1826 start_worker(worker);
1827 spin_unlock_irq(&pool->lock);
1830 mutex_unlock(&pool->manager_mutex);
1832 return worker ? 0 : -ENOMEM;
1836 * destroy_worker - destroy a workqueue worker
1837 * @worker: worker to be destroyed
1839 * Destroy @worker and adjust @pool stats accordingly.
1842 * spin_lock_irq(pool->lock) which is released and regrabbed.
1844 static void destroy_worker(struct worker *worker)
1846 struct worker_pool *pool = worker->pool;
1848 lockdep_assert_held(&pool->manager_mutex);
1849 lockdep_assert_held(&pool->lock);
1851 /* sanity check frenzy */
1852 if (WARN_ON(worker->current_work) ||
1853 WARN_ON(!list_empty(&worker->scheduled)))
1856 if (worker->flags & WORKER_STARTED)
1858 if (worker->flags & WORKER_IDLE)
1862 * Once WORKER_DIE is set, the kworker may destroy itself at any
1863 * point. Pin to ensure the task stays until we're done with it.
1865 get_task_struct(worker->task);
1867 list_del_init(&worker->entry);
1868 worker->flags |= WORKER_DIE;
1870 idr_remove(&pool->worker_idr, worker->id);
1872 spin_unlock_irq(&pool->lock);
1874 kthread_stop(worker->task);
1875 put_task_struct(worker->task);
1878 spin_lock_irq(&pool->lock);
1881 static void idle_worker_timeout(unsigned long __pool)
1883 struct worker_pool *pool = (void *)__pool;
1885 spin_lock_irq(&pool->lock);
1887 if (too_many_workers(pool)) {
1888 struct worker *worker;
1889 unsigned long expires;
1891 /* idle_list is kept in LIFO order, check the last one */
1892 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1893 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1895 if (time_before(jiffies, expires))
1896 mod_timer(&pool->idle_timer, expires);
1898 /* it's been idle for too long, wake up manager */
1899 pool->flags |= POOL_MANAGE_WORKERS;
1900 wake_up_worker(pool);
1904 spin_unlock_irq(&pool->lock);
1907 static void send_mayday(struct work_struct *work)
1909 struct pool_workqueue *pwq = get_work_pwq(work);
1910 struct workqueue_struct *wq = pwq->wq;
1912 lockdep_assert_held(&wq_mayday_lock);
1917 /* mayday mayday mayday */
1918 if (list_empty(&pwq->mayday_node)) {
1919 list_add_tail(&pwq->mayday_node, &wq->maydays);
1920 wake_up_process(wq->rescuer->task);
1924 static void pool_mayday_timeout(unsigned long __pool)
1926 struct worker_pool *pool = (void *)__pool;
1927 struct work_struct *work;
1929 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1930 spin_lock(&pool->lock);
1932 if (need_to_create_worker(pool)) {
1934 * We've been trying to create a new worker but
1935 * haven't been successful. We might be hitting an
1936 * allocation deadlock. Send distress signals to
1939 list_for_each_entry(work, &pool->worklist, entry)
1943 spin_unlock(&pool->lock);
1944 spin_unlock_irq(&wq_mayday_lock);
1946 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1950 * maybe_create_worker - create a new worker if necessary
1951 * @pool: pool to create a new worker for
1953 * Create a new worker for @pool if necessary. @pool is guaranteed to
1954 * have at least one idle worker on return from this function. If
1955 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1956 * sent to all rescuers with works scheduled on @pool to resolve
1957 * possible allocation deadlock.
1959 * On return, need_to_create_worker() is guaranteed to be %false and
1960 * may_start_working() %true.
1963 * spin_lock_irq(pool->lock) which may be released and regrabbed
1964 * multiple times. Does GFP_KERNEL allocations. Called only from
1968 * %false if no action was taken and pool->lock stayed locked, %true
1971 static bool maybe_create_worker(struct worker_pool *pool)
1972 __releases(&pool->lock)
1973 __acquires(&pool->lock)
1975 if (!need_to_create_worker(pool))
1978 spin_unlock_irq(&pool->lock);
1980 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1981 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1984 struct worker *worker;
1986 worker = create_worker(pool);
1988 del_timer_sync(&pool->mayday_timer);
1989 spin_lock_irq(&pool->lock);
1990 start_worker(worker);
1991 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1996 if (!need_to_create_worker(pool))
1999 __set_current_state(TASK_INTERRUPTIBLE);
2000 schedule_timeout(CREATE_COOLDOWN);
2002 if (!need_to_create_worker(pool))
2006 del_timer_sync(&pool->mayday_timer);
2007 spin_lock_irq(&pool->lock);
2008 if (need_to_create_worker(pool))
2014 * maybe_destroy_worker - destroy workers which have been idle for a while
2015 * @pool: pool to destroy workers for
2017 * Destroy @pool workers which have been idle for longer than
2018 * IDLE_WORKER_TIMEOUT.
2021 * spin_lock_irq(pool->lock) which may be released and regrabbed
2022 * multiple times. Called only from manager.
2025 * %false if no action was taken and pool->lock stayed locked, %true
2028 static bool maybe_destroy_workers(struct worker_pool *pool)
2032 while (too_many_workers(pool)) {
2033 struct worker *worker;
2034 unsigned long expires;
2036 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2037 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2039 if (time_before(jiffies, expires)) {
2040 mod_timer(&pool->idle_timer, expires);
2044 destroy_worker(worker);
2052 * manage_workers - manage worker pool
2055 * Assume the manager role and manage the worker pool @worker belongs
2056 * to. At any given time, there can be only zero or one manager per
2057 * pool. The exclusion is handled automatically by this function.
2059 * The caller can safely start processing works on false return. On
2060 * true return, it's guaranteed that need_to_create_worker() is false
2061 * and may_start_working() is true.
2064 * spin_lock_irq(pool->lock) which may be released and regrabbed
2065 * multiple times. Does GFP_KERNEL allocations.
2068 * %false if the pool don't need management and the caller can safely start
2069 * processing works, %true indicates that the function released pool->lock
2070 * and reacquired it to perform some management function and that the
2071 * conditions that the caller verified while holding the lock before
2072 * calling the function might no longer be true.
2074 static bool manage_workers(struct worker *worker)
2076 struct worker_pool *pool = worker->pool;
2080 * Managership is governed by two mutexes - manager_arb and
2081 * manager_mutex. manager_arb handles arbitration of manager role.
2082 * Anyone who successfully grabs manager_arb wins the arbitration
2083 * and becomes the manager. mutex_trylock() on pool->manager_arb
2084 * failure while holding pool->lock reliably indicates that someone
2085 * else is managing the pool and the worker which failed trylock
2086 * can proceed to executing work items. This means that anyone
2087 * grabbing manager_arb is responsible for actually performing
2088 * manager duties. If manager_arb is grabbed and released without
2089 * actual management, the pool may stall indefinitely.
2091 * manager_mutex is used for exclusion of actual management
2092 * operations. The holder of manager_mutex can be sure that none
2093 * of management operations, including creation and destruction of
2094 * workers, won't take place until the mutex is released. Because
2095 * manager_mutex doesn't interfere with manager role arbitration,
2096 * it is guaranteed that the pool's management, while may be
2097 * delayed, won't be disturbed by someone else grabbing
2100 if (!mutex_trylock(&pool->manager_arb))
2104 * With manager arbitration won, manager_mutex would be free in
2105 * most cases. trylock first without dropping @pool->lock.
2107 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2108 spin_unlock_irq(&pool->lock);
2109 mutex_lock(&pool->manager_mutex);
2110 spin_lock_irq(&pool->lock);
2114 pool->flags &= ~POOL_MANAGE_WORKERS;
2117 * Destroy and then create so that may_start_working() is true
2120 ret |= maybe_destroy_workers(pool);
2121 ret |= maybe_create_worker(pool);
2123 mutex_unlock(&pool->manager_mutex);
2124 mutex_unlock(&pool->manager_arb);
2129 * process_one_work - process single work
2131 * @work: work to process
2133 * Process @work. This function contains all the logics necessary to
2134 * process a single work including synchronization against and
2135 * interaction with other workers on the same cpu, queueing and
2136 * flushing. As long as context requirement is met, any worker can
2137 * call this function to process a work.
2140 * spin_lock_irq(pool->lock) which is released and regrabbed.
2142 static void process_one_work(struct worker *worker, struct work_struct *work)
2143 __releases(&pool->lock)
2144 __acquires(&pool->lock)
2146 struct pool_workqueue *pwq = get_work_pwq(work);
2147 struct worker_pool *pool = worker->pool;
2148 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2150 struct worker *collision;
2151 #ifdef CONFIG_LOCKDEP
2153 * It is permissible to free the struct work_struct from
2154 * inside the function that is called from it, this we need to
2155 * take into account for lockdep too. To avoid bogus "held
2156 * lock freed" warnings as well as problems when looking into
2157 * work->lockdep_map, make a copy and use that here.
2159 struct lockdep_map lockdep_map;
2161 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2164 * Ensure we're on the correct CPU. DISASSOCIATED test is
2165 * necessary to avoid spurious warnings from rescuers servicing the
2166 * unbound or a disassociated pool.
2168 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2169 !(pool->flags & POOL_DISASSOCIATED) &&
2170 raw_smp_processor_id() != pool->cpu);
2173 * A single work shouldn't be executed concurrently by
2174 * multiple workers on a single cpu. Check whether anyone is
2175 * already processing the work. If so, defer the work to the
2176 * currently executing one.
2178 collision = find_worker_executing_work(pool, work);
2179 if (unlikely(collision)) {
2180 move_linked_works(work, &collision->scheduled, NULL);
2184 /* claim and dequeue */
2185 debug_work_deactivate(work);
2186 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2187 worker->current_work = work;
2188 worker->current_func = work->func;
2189 worker->current_pwq = pwq;
2190 work_color = get_work_color(work);
2192 list_del_init(&work->entry);
2195 * CPU intensive works don't participate in concurrency
2196 * management. They're the scheduler's responsibility.
2198 if (unlikely(cpu_intensive))
2199 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2202 * Unbound pool isn't concurrency managed and work items should be
2203 * executed ASAP. Wake up another worker if necessary.
2205 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2206 wake_up_worker(pool);
2209 * Record the last pool and clear PENDING which should be the last
2210 * update to @work. Also, do this inside @pool->lock so that
2211 * PENDING and queued state changes happen together while IRQ is
2214 set_work_pool_and_clear_pending(work, pool->id);
2216 spin_unlock_irq(&pool->lock);
2218 lock_map_acquire_read(&pwq->wq->lockdep_map);
2219 lock_map_acquire(&lockdep_map);
2220 trace_workqueue_execute_start(work);
2221 worker->current_func(work);
2223 * While we must be careful to not use "work" after this, the trace
2224 * point will only record its address.
2226 trace_workqueue_execute_end(work);
2227 lock_map_release(&lockdep_map);
2228 lock_map_release(&pwq->wq->lockdep_map);
2230 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2231 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2232 " last function: %pf\n",
2233 current->comm, preempt_count(), task_pid_nr(current),
2234 worker->current_func);
2235 debug_show_held_locks(current);
2240 * The following prevents a kworker from hogging CPU on !PREEMPT
2241 * kernels, where a requeueing work item waiting for something to
2242 * happen could deadlock with stop_machine as such work item could
2243 * indefinitely requeue itself while all other CPUs are trapped in
2248 spin_lock_irq(&pool->lock);
2250 /* clear cpu intensive status */
2251 if (unlikely(cpu_intensive))
2252 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2254 /* we're done with it, release */
2255 hash_del(&worker->hentry);
2256 worker->current_work = NULL;
2257 worker->current_func = NULL;
2258 worker->current_pwq = NULL;
2259 worker->desc_valid = false;
2260 pwq_dec_nr_in_flight(pwq, work_color);
2264 * process_scheduled_works - process scheduled works
2267 * Process all scheduled works. Please note that the scheduled list
2268 * may change while processing a work, so this function repeatedly
2269 * fetches a work from the top and executes it.
2272 * spin_lock_irq(pool->lock) which may be released and regrabbed
2275 static void process_scheduled_works(struct worker *worker)
2277 while (!list_empty(&worker->scheduled)) {
2278 struct work_struct *work = list_first_entry(&worker->scheduled,
2279 struct work_struct, entry);
2280 process_one_work(worker, work);
2285 * worker_thread - the worker thread function
2288 * The worker thread function. All workers belong to a worker_pool -
2289 * either a per-cpu one or dynamic unbound one. These workers process all
2290 * work items regardless of their specific target workqueue. The only
2291 * exception is work items which belong to workqueues with a rescuer which
2292 * will be explained in rescuer_thread().
2296 static int worker_thread(void *__worker)
2298 struct worker *worker = __worker;
2299 struct worker_pool *pool = worker->pool;
2301 /* tell the scheduler that this is a workqueue worker */
2302 worker->task->flags |= PF_WQ_WORKER;
2304 spin_lock_irq(&pool->lock);
2306 /* am I supposed to die? */
2307 if (unlikely(worker->flags & WORKER_DIE)) {
2308 spin_unlock_irq(&pool->lock);
2309 WARN_ON_ONCE(!list_empty(&worker->entry));
2310 worker->task->flags &= ~PF_WQ_WORKER;
2314 worker_leave_idle(worker);
2316 /* no more worker necessary? */
2317 if (!need_more_worker(pool))
2320 /* do we need to manage? */
2321 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2325 * ->scheduled list can only be filled while a worker is
2326 * preparing to process a work or actually processing it.
2327 * Make sure nobody diddled with it while I was sleeping.
2329 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2332 * Finish PREP stage. We're guaranteed to have at least one idle
2333 * worker or that someone else has already assumed the manager
2334 * role. This is where @worker starts participating in concurrency
2335 * management if applicable and concurrency management is restored
2336 * after being rebound. See rebind_workers() for details.
2338 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2341 struct work_struct *work =
2342 list_first_entry(&pool->worklist,
2343 struct work_struct, entry);
2345 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2346 /* optimization path, not strictly necessary */
2347 process_one_work(worker, work);
2348 if (unlikely(!list_empty(&worker->scheduled)))
2349 process_scheduled_works(worker);
2351 move_linked_works(work, &worker->scheduled, NULL);
2352 process_scheduled_works(worker);
2354 } while (keep_working(pool));
2356 worker_set_flags(worker, WORKER_PREP, false);
2358 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2362 * pool->lock is held and there's no work to process and no need to
2363 * manage, sleep. Workers are woken up only while holding
2364 * pool->lock or from local cpu, so setting the current state
2365 * before releasing pool->lock is enough to prevent losing any
2368 worker_enter_idle(worker);
2369 __set_current_state(TASK_INTERRUPTIBLE);
2370 spin_unlock_irq(&pool->lock);
2376 * rescuer_thread - the rescuer thread function
2379 * Workqueue rescuer thread function. There's one rescuer for each
2380 * workqueue which has WQ_MEM_RECLAIM set.
2382 * Regular work processing on a pool may block trying to create a new
2383 * worker which uses GFP_KERNEL allocation which has slight chance of
2384 * developing into deadlock if some works currently on the same queue
2385 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2386 * the problem rescuer solves.
2388 * When such condition is possible, the pool summons rescuers of all
2389 * workqueues which have works queued on the pool and let them process
2390 * those works so that forward progress can be guaranteed.
2392 * This should happen rarely.
2396 static int rescuer_thread(void *__rescuer)
2398 struct worker *rescuer = __rescuer;
2399 struct workqueue_struct *wq = rescuer->rescue_wq;
2400 struct list_head *scheduled = &rescuer->scheduled;
2403 set_user_nice(current, RESCUER_NICE_LEVEL);
2406 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2407 * doesn't participate in concurrency management.
2409 rescuer->task->flags |= PF_WQ_WORKER;
2411 set_current_state(TASK_INTERRUPTIBLE);
2414 * By the time the rescuer is requested to stop, the workqueue
2415 * shouldn't have any work pending, but @wq->maydays may still have
2416 * pwq(s) queued. This can happen by non-rescuer workers consuming
2417 * all the work items before the rescuer got to them. Go through
2418 * @wq->maydays processing before acting on should_stop so that the
2419 * list is always empty on exit.
2421 should_stop = kthread_should_stop();
2423 /* see whether any pwq is asking for help */
2424 spin_lock_irq(&wq_mayday_lock);
2426 while (!list_empty(&wq->maydays)) {
2427 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2428 struct pool_workqueue, mayday_node);
2429 struct worker_pool *pool = pwq->pool;
2430 struct work_struct *work, *n;
2432 __set_current_state(TASK_RUNNING);
2433 list_del_init(&pwq->mayday_node);
2435 spin_unlock_irq(&wq_mayday_lock);
2437 /* migrate to the target cpu if possible */
2438 worker_maybe_bind_and_lock(pool);
2439 rescuer->pool = pool;
2442 * Slurp in all works issued via this workqueue and
2445 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2446 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2447 if (get_work_pwq(work) == pwq)
2448 move_linked_works(work, scheduled, &n);
2450 process_scheduled_works(rescuer);
2453 * Leave this pool. If keep_working() is %true, notify a
2454 * regular worker; otherwise, we end up with 0 concurrency
2455 * and stalling the execution.
2457 if (keep_working(pool))
2458 wake_up_worker(pool);
2460 rescuer->pool = NULL;
2461 spin_unlock(&pool->lock);
2462 spin_lock(&wq_mayday_lock);
2465 spin_unlock_irq(&wq_mayday_lock);
2468 __set_current_state(TASK_RUNNING);
2469 rescuer->task->flags &= ~PF_WQ_WORKER;
2473 /* rescuers should never participate in concurrency management */
2474 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2480 struct work_struct work;
2481 struct completion done;
2484 static void wq_barrier_func(struct work_struct *work)
2486 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2487 complete(&barr->done);
2491 * insert_wq_barrier - insert a barrier work
2492 * @pwq: pwq to insert barrier into
2493 * @barr: wq_barrier to insert
2494 * @target: target work to attach @barr to
2495 * @worker: worker currently executing @target, NULL if @target is not executing
2497 * @barr is linked to @target such that @barr is completed only after
2498 * @target finishes execution. Please note that the ordering
2499 * guarantee is observed only with respect to @target and on the local
2502 * Currently, a queued barrier can't be canceled. This is because
2503 * try_to_grab_pending() can't determine whether the work to be
2504 * grabbed is at the head of the queue and thus can't clear LINKED
2505 * flag of the previous work while there must be a valid next work
2506 * after a work with LINKED flag set.
2508 * Note that when @worker is non-NULL, @target may be modified
2509 * underneath us, so we can't reliably determine pwq from @target.
2512 * spin_lock_irq(pool->lock).
2514 static void insert_wq_barrier(struct pool_workqueue *pwq,
2515 struct wq_barrier *barr,
2516 struct work_struct *target, struct worker *worker)
2518 struct list_head *head;
2519 unsigned int linked = 0;
2522 * debugobject calls are safe here even with pool->lock locked
2523 * as we know for sure that this will not trigger any of the
2524 * checks and call back into the fixup functions where we
2527 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2528 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2529 init_completion(&barr->done);
2532 * If @target is currently being executed, schedule the
2533 * barrier to the worker; otherwise, put it after @target.
2536 head = worker->scheduled.next;
2538 unsigned long *bits = work_data_bits(target);
2540 head = target->entry.next;
2541 /* there can already be other linked works, inherit and set */
2542 linked = *bits & WORK_STRUCT_LINKED;
2543 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2546 debug_work_activate(&barr->work);
2547 insert_work(pwq, &barr->work, head,
2548 work_color_to_flags(WORK_NO_COLOR) | linked);
2552 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2553 * @wq: workqueue being flushed
2554 * @flush_color: new flush color, < 0 for no-op
2555 * @work_color: new work color, < 0 for no-op
2557 * Prepare pwqs for workqueue flushing.
2559 * If @flush_color is non-negative, flush_color on all pwqs should be
2560 * -1. If no pwq has in-flight commands at the specified color, all
2561 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2562 * has in flight commands, its pwq->flush_color is set to
2563 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2564 * wakeup logic is armed and %true is returned.
2566 * The caller should have initialized @wq->first_flusher prior to
2567 * calling this function with non-negative @flush_color. If
2568 * @flush_color is negative, no flush color update is done and %false
2571 * If @work_color is non-negative, all pwqs should have the same
2572 * work_color which is previous to @work_color and all will be
2573 * advanced to @work_color.
2576 * mutex_lock(wq->mutex).
2579 * %true if @flush_color >= 0 and there's something to flush. %false
2582 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2583 int flush_color, int work_color)
2586 struct pool_workqueue *pwq;
2588 if (flush_color >= 0) {
2589 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2590 atomic_set(&wq->nr_pwqs_to_flush, 1);
2593 for_each_pwq(pwq, wq) {
2594 struct worker_pool *pool = pwq->pool;
2596 spin_lock_irq(&pool->lock);
2598 if (flush_color >= 0) {
2599 WARN_ON_ONCE(pwq->flush_color != -1);
2601 if (pwq->nr_in_flight[flush_color]) {
2602 pwq->flush_color = flush_color;
2603 atomic_inc(&wq->nr_pwqs_to_flush);
2608 if (work_color >= 0) {
2609 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2610 pwq->work_color = work_color;
2613 spin_unlock_irq(&pool->lock);
2616 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2617 complete(&wq->first_flusher->done);
2623 * flush_workqueue - ensure that any scheduled work has run to completion.
2624 * @wq: workqueue to flush
2626 * This function sleeps until all work items which were queued on entry
2627 * have finished execution, but it is not livelocked by new incoming ones.
2629 void flush_workqueue(struct workqueue_struct *wq)
2631 struct wq_flusher this_flusher = {
2632 .list = LIST_HEAD_INIT(this_flusher.list),
2634 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2638 lock_map_acquire(&wq->lockdep_map);
2639 lock_map_release(&wq->lockdep_map);
2641 mutex_lock(&wq->mutex);
2644 * Start-to-wait phase
2646 next_color = work_next_color(wq->work_color);
2648 if (next_color != wq->flush_color) {
2650 * Color space is not full. The current work_color
2651 * becomes our flush_color and work_color is advanced
2654 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2655 this_flusher.flush_color = wq->work_color;
2656 wq->work_color = next_color;
2658 if (!wq->first_flusher) {
2659 /* no flush in progress, become the first flusher */
2660 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2662 wq->first_flusher = &this_flusher;
2664 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2666 /* nothing to flush, done */
2667 wq->flush_color = next_color;
2668 wq->first_flusher = NULL;
2673 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2674 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2675 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2679 * Oops, color space is full, wait on overflow queue.
2680 * The next flush completion will assign us
2681 * flush_color and transfer to flusher_queue.
2683 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2686 mutex_unlock(&wq->mutex);
2688 wait_for_completion(&this_flusher.done);
2691 * Wake-up-and-cascade phase
2693 * First flushers are responsible for cascading flushes and
2694 * handling overflow. Non-first flushers can simply return.
2696 if (wq->first_flusher != &this_flusher)
2699 mutex_lock(&wq->mutex);
2701 /* we might have raced, check again with mutex held */
2702 if (wq->first_flusher != &this_flusher)
2705 wq->first_flusher = NULL;
2707 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2708 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2711 struct wq_flusher *next, *tmp;
2713 /* complete all the flushers sharing the current flush color */
2714 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2715 if (next->flush_color != wq->flush_color)
2717 list_del_init(&next->list);
2718 complete(&next->done);
2721 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2722 wq->flush_color != work_next_color(wq->work_color));
2724 /* this flush_color is finished, advance by one */
2725 wq->flush_color = work_next_color(wq->flush_color);
2727 /* one color has been freed, handle overflow queue */
2728 if (!list_empty(&wq->flusher_overflow)) {
2730 * Assign the same color to all overflowed
2731 * flushers, advance work_color and append to
2732 * flusher_queue. This is the start-to-wait
2733 * phase for these overflowed flushers.
2735 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2736 tmp->flush_color = wq->work_color;
2738 wq->work_color = work_next_color(wq->work_color);
2740 list_splice_tail_init(&wq->flusher_overflow,
2741 &wq->flusher_queue);
2742 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2745 if (list_empty(&wq->flusher_queue)) {
2746 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2751 * Need to flush more colors. Make the next flusher
2752 * the new first flusher and arm pwqs.
2754 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2755 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2757 list_del_init(&next->list);
2758 wq->first_flusher = next;
2760 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2764 * Meh... this color is already done, clear first
2765 * flusher and repeat cascading.
2767 wq->first_flusher = NULL;
2771 mutex_unlock(&wq->mutex);
2773 EXPORT_SYMBOL_GPL(flush_workqueue);
2776 * drain_workqueue - drain a workqueue
2777 * @wq: workqueue to drain
2779 * Wait until the workqueue becomes empty. While draining is in progress,
2780 * only chain queueing is allowed. IOW, only currently pending or running
2781 * work items on @wq can queue further work items on it. @wq is flushed
2782 * repeatedly until it becomes empty. The number of flushing is detemined
2783 * by the depth of chaining and should be relatively short. Whine if it
2786 void drain_workqueue(struct workqueue_struct *wq)
2788 unsigned int flush_cnt = 0;
2789 struct pool_workqueue *pwq;
2792 * __queue_work() needs to test whether there are drainers, is much
2793 * hotter than drain_workqueue() and already looks at @wq->flags.
2794 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2796 mutex_lock(&wq->mutex);
2797 if (!wq->nr_drainers++)
2798 wq->flags |= __WQ_DRAINING;
2799 mutex_unlock(&wq->mutex);
2801 flush_workqueue(wq);
2803 mutex_lock(&wq->mutex);
2805 for_each_pwq(pwq, wq) {
2808 spin_lock_irq(&pwq->pool->lock);
2809 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2810 spin_unlock_irq(&pwq->pool->lock);
2815 if (++flush_cnt == 10 ||
2816 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2817 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2818 wq->name, flush_cnt);
2820 mutex_unlock(&wq->mutex);
2824 if (!--wq->nr_drainers)
2825 wq->flags &= ~__WQ_DRAINING;
2826 mutex_unlock(&wq->mutex);
2828 EXPORT_SYMBOL_GPL(drain_workqueue);
2830 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2832 struct worker *worker = NULL;
2833 struct worker_pool *pool;
2834 struct pool_workqueue *pwq;
2838 local_irq_disable();
2839 pool = get_work_pool(work);
2845 spin_lock(&pool->lock);
2846 /* see the comment in try_to_grab_pending() with the same code */
2847 pwq = get_work_pwq(work);
2849 if (unlikely(pwq->pool != pool))
2852 worker = find_worker_executing_work(pool, work);
2855 pwq = worker->current_pwq;
2858 insert_wq_barrier(pwq, barr, work, worker);
2859 spin_unlock_irq(&pool->lock);
2862 * If @max_active is 1 or rescuer is in use, flushing another work
2863 * item on the same workqueue may lead to deadlock. Make sure the
2864 * flusher is not running on the same workqueue by verifying write
2867 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2868 lock_map_acquire(&pwq->wq->lockdep_map);
2870 lock_map_acquire_read(&pwq->wq->lockdep_map);
2871 lock_map_release(&pwq->wq->lockdep_map);
2875 spin_unlock_irq(&pool->lock);
2880 * flush_work - wait for a work to finish executing the last queueing instance
2881 * @work: the work to flush
2883 * Wait until @work has finished execution. @work is guaranteed to be idle
2884 * on return if it hasn't been requeued since flush started.
2887 * %true if flush_work() waited for the work to finish execution,
2888 * %false if it was already idle.
2890 bool flush_work(struct work_struct *work)
2892 struct wq_barrier barr;
2894 lock_map_acquire(&work->lockdep_map);
2895 lock_map_release(&work->lockdep_map);
2897 if (start_flush_work(work, &barr)) {
2898 wait_for_completion(&barr.done);
2899 destroy_work_on_stack(&barr.work);
2905 EXPORT_SYMBOL_GPL(flush_work);
2907 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2909 unsigned long flags;
2913 ret = try_to_grab_pending(work, is_dwork, &flags);
2915 * If someone else is canceling, wait for the same event it
2916 * would be waiting for before retrying.
2918 if (unlikely(ret == -ENOENT))
2920 } while (unlikely(ret < 0));
2922 /* tell other tasks trying to grab @work to back off */
2923 mark_work_canceling(work);
2924 local_irq_restore(flags);
2927 clear_work_data(work);
2932 * cancel_work_sync - cancel a work and wait for it to finish
2933 * @work: the work to cancel
2935 * Cancel @work and wait for its execution to finish. This function
2936 * can be used even if the work re-queues itself or migrates to
2937 * another workqueue. On return from this function, @work is
2938 * guaranteed to be not pending or executing on any CPU.
2940 * cancel_work_sync(&delayed_work->work) must not be used for
2941 * delayed_work's. Use cancel_delayed_work_sync() instead.
2943 * The caller must ensure that the workqueue on which @work was last
2944 * queued can't be destroyed before this function returns.
2947 * %true if @work was pending, %false otherwise.
2949 bool cancel_work_sync(struct work_struct *work)
2951 return __cancel_work_timer(work, false);
2953 EXPORT_SYMBOL_GPL(cancel_work_sync);
2956 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2957 * @dwork: the delayed work to flush
2959 * Delayed timer is cancelled and the pending work is queued for
2960 * immediate execution. Like flush_work(), this function only
2961 * considers the last queueing instance of @dwork.
2964 * %true if flush_work() waited for the work to finish execution,
2965 * %false if it was already idle.
2967 bool flush_delayed_work(struct delayed_work *dwork)
2969 local_irq_disable();
2970 if (del_timer_sync(&dwork->timer))
2971 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2973 return flush_work(&dwork->work);
2975 EXPORT_SYMBOL(flush_delayed_work);
2978 * cancel_delayed_work - cancel a delayed work
2979 * @dwork: delayed_work to cancel
2981 * Kill off a pending delayed_work.
2983 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2987 * The work callback function may still be running on return, unless
2988 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2989 * use cancel_delayed_work_sync() to wait on it.
2991 * This function is safe to call from any context including IRQ handler.
2993 bool cancel_delayed_work(struct delayed_work *dwork)
2995 unsigned long flags;
2999 ret = try_to_grab_pending(&dwork->work, true, &flags);
3000 } while (unlikely(ret == -EAGAIN));
3002 if (unlikely(ret < 0))
3005 set_work_pool_and_clear_pending(&dwork->work,
3006 get_work_pool_id(&dwork->work));
3007 local_irq_restore(flags);
3010 EXPORT_SYMBOL(cancel_delayed_work);
3013 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3014 * @dwork: the delayed work cancel
3016 * This is cancel_work_sync() for delayed works.
3019 * %true if @dwork was pending, %false otherwise.
3021 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3023 return __cancel_work_timer(&dwork->work, true);
3025 EXPORT_SYMBOL(cancel_delayed_work_sync);
3028 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3029 * @func: the function to call
3031 * schedule_on_each_cpu() executes @func on each online CPU using the
3032 * system workqueue and blocks until all CPUs have completed.
3033 * schedule_on_each_cpu() is very slow.
3036 * 0 on success, -errno on failure.
3038 int schedule_on_each_cpu(work_func_t func)
3041 struct work_struct __percpu *works;
3043 works = alloc_percpu(struct work_struct);
3049 for_each_online_cpu(cpu) {
3050 struct work_struct *work = per_cpu_ptr(works, cpu);
3052 INIT_WORK(work, func);
3053 schedule_work_on(cpu, work);
3056 for_each_online_cpu(cpu)
3057 flush_work(per_cpu_ptr(works, cpu));
3065 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3067 * Forces execution of the kernel-global workqueue and blocks until its
3070 * Think twice before calling this function! It's very easy to get into
3071 * trouble if you don't take great care. Either of the following situations
3072 * will lead to deadlock:
3074 * One of the work items currently on the workqueue needs to acquire
3075 * a lock held by your code or its caller.
3077 * Your code is running in the context of a work routine.
3079 * They will be detected by lockdep when they occur, but the first might not
3080 * occur very often. It depends on what work items are on the workqueue and
3081 * what locks they need, which you have no control over.
3083 * In most situations flushing the entire workqueue is overkill; you merely
3084 * need to know that a particular work item isn't queued and isn't running.
3085 * In such cases you should use cancel_delayed_work_sync() or
3086 * cancel_work_sync() instead.
3088 void flush_scheduled_work(void)
3090 flush_workqueue(system_wq);
3092 EXPORT_SYMBOL(flush_scheduled_work);
3095 * execute_in_process_context - reliably execute the routine with user context
3096 * @fn: the function to execute
3097 * @ew: guaranteed storage for the execute work structure (must
3098 * be available when the work executes)
3100 * Executes the function immediately if process context is available,
3101 * otherwise schedules the function for delayed execution.
3103 * Return: 0 - function was executed
3104 * 1 - function was scheduled for execution
3106 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3108 if (!in_interrupt()) {
3113 INIT_WORK(&ew->work, fn);
3114 schedule_work(&ew->work);
3118 EXPORT_SYMBOL_GPL(execute_in_process_context);
3122 * Workqueues with WQ_SYSFS flag set is visible to userland via
3123 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3124 * following attributes.
3126 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3127 * max_active RW int : maximum number of in-flight work items
3129 * Unbound workqueues have the following extra attributes.
3131 * id RO int : the associated pool ID
3132 * nice RW int : nice value of the workers
3133 * cpumask RW mask : bitmask of allowed CPUs for the workers
3136 struct workqueue_struct *wq;
3140 static struct workqueue_struct *dev_to_wq(struct device *dev)
3142 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3147 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3150 struct workqueue_struct *wq = dev_to_wq(dev);
3152 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3154 static DEVICE_ATTR_RO(per_cpu);
3156 static ssize_t max_active_show(struct device *dev,
3157 struct device_attribute *attr, char *buf)
3159 struct workqueue_struct *wq = dev_to_wq(dev);
3161 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3164 static ssize_t max_active_store(struct device *dev,
3165 struct device_attribute *attr, const char *buf,
3168 struct workqueue_struct *wq = dev_to_wq(dev);
3171 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3174 workqueue_set_max_active(wq, val);
3177 static DEVICE_ATTR_RW(max_active);
3179 static struct attribute *wq_sysfs_attrs[] = {
3180 &dev_attr_per_cpu.attr,
3181 &dev_attr_max_active.attr,
3184 ATTRIBUTE_GROUPS(wq_sysfs);
3186 static ssize_t wq_pool_ids_show(struct device *dev,
3187 struct device_attribute *attr, char *buf)
3189 struct workqueue_struct *wq = dev_to_wq(dev);
3190 const char *delim = "";
3191 int node, written = 0;
3193 rcu_read_lock_sched();
3194 for_each_node(node) {
3195 written += scnprintf(buf + written, PAGE_SIZE - written,
3196 "%s%d:%d", delim, node,
3197 unbound_pwq_by_node(wq, node)->pool->id);
3200 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3201 rcu_read_unlock_sched();
3206 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3209 struct workqueue_struct *wq = dev_to_wq(dev);
3212 mutex_lock(&wq->mutex);
3213 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3214 mutex_unlock(&wq->mutex);
3219 /* prepare workqueue_attrs for sysfs store operations */
3220 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3222 struct workqueue_attrs *attrs;
3224 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3228 mutex_lock(&wq->mutex);
3229 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3230 mutex_unlock(&wq->mutex);
3234 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3235 const char *buf, size_t count)
3237 struct workqueue_struct *wq = dev_to_wq(dev);
3238 struct workqueue_attrs *attrs;
3241 attrs = wq_sysfs_prep_attrs(wq);
3245 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3246 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
3247 ret = apply_workqueue_attrs(wq, attrs);
3251 free_workqueue_attrs(attrs);
3252 return ret ?: count;
3255 static ssize_t wq_cpumask_show(struct device *dev,
3256 struct device_attribute *attr, char *buf)
3258 struct workqueue_struct *wq = dev_to_wq(dev);
3261 mutex_lock(&wq->mutex);
3262 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3263 mutex_unlock(&wq->mutex);
3265 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3269 static ssize_t wq_cpumask_store(struct device *dev,
3270 struct device_attribute *attr,
3271 const char *buf, size_t count)
3273 struct workqueue_struct *wq = dev_to_wq(dev);
3274 struct workqueue_attrs *attrs;
3277 attrs = wq_sysfs_prep_attrs(wq);
3281 ret = cpumask_parse(buf, attrs->cpumask);
3283 ret = apply_workqueue_attrs(wq, attrs);
3285 free_workqueue_attrs(attrs);
3286 return ret ?: count;
3289 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3292 struct workqueue_struct *wq = dev_to_wq(dev);
3295 mutex_lock(&wq->mutex);
3296 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3297 !wq->unbound_attrs->no_numa);
3298 mutex_unlock(&wq->mutex);
3303 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3304 const char *buf, size_t count)
3306 struct workqueue_struct *wq = dev_to_wq(dev);
3307 struct workqueue_attrs *attrs;
3310 attrs = wq_sysfs_prep_attrs(wq);
3315 if (sscanf(buf, "%d", &v) == 1) {
3316 attrs->no_numa = !v;
3317 ret = apply_workqueue_attrs(wq, attrs);
3320 free_workqueue_attrs(attrs);
3321 return ret ?: count;
3324 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3325 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3326 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3327 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3328 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3332 static struct bus_type wq_subsys = {
3333 .name = "workqueue",
3334 .dev_groups = wq_sysfs_groups,
3337 static int __init wq_sysfs_init(void)
3339 return subsys_virtual_register(&wq_subsys, NULL);
3341 core_initcall(wq_sysfs_init);
3343 static void wq_device_release(struct device *dev)
3345 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3351 * workqueue_sysfs_register - make a workqueue visible in sysfs
3352 * @wq: the workqueue to register
3354 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3355 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3356 * which is the preferred method.
3358 * Workqueue user should use this function directly iff it wants to apply
3359 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3360 * apply_workqueue_attrs() may race against userland updating the
3363 * Return: 0 on success, -errno on failure.
3365 int workqueue_sysfs_register(struct workqueue_struct *wq)
3367 struct wq_device *wq_dev;
3371 * Adjusting max_active or creating new pwqs by applyting
3372 * attributes breaks ordering guarantee. Disallow exposing ordered
3375 if (WARN_ON(wq->flags & __WQ_ORDERED))
3378 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3383 wq_dev->dev.bus = &wq_subsys;
3384 wq_dev->dev.init_name = wq->name;
3385 wq_dev->dev.release = wq_device_release;
3388 * unbound_attrs are created separately. Suppress uevent until
3389 * everything is ready.
3391 dev_set_uevent_suppress(&wq_dev->dev, true);
3393 ret = device_register(&wq_dev->dev);
3400 if (wq->flags & WQ_UNBOUND) {
3401 struct device_attribute *attr;
3403 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3404 ret = device_create_file(&wq_dev->dev, attr);
3406 device_unregister(&wq_dev->dev);
3413 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3418 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3419 * @wq: the workqueue to unregister
3421 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3423 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3425 struct wq_device *wq_dev = wq->wq_dev;
3431 device_unregister(&wq_dev->dev);
3433 #else /* CONFIG_SYSFS */
3434 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3435 #endif /* CONFIG_SYSFS */
3438 * free_workqueue_attrs - free a workqueue_attrs
3439 * @attrs: workqueue_attrs to free
3441 * Undo alloc_workqueue_attrs().
3443 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3446 free_cpumask_var(attrs->cpumask);
3452 * alloc_workqueue_attrs - allocate a workqueue_attrs
3453 * @gfp_mask: allocation mask to use
3455 * Allocate a new workqueue_attrs, initialize with default settings and
3458 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3460 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3462 struct workqueue_attrs *attrs;
3464 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3467 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3470 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3473 free_workqueue_attrs(attrs);
3477 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3478 const struct workqueue_attrs *from)
3480 to->nice = from->nice;
3481 cpumask_copy(to->cpumask, from->cpumask);
3483 * Unlike hash and equality test, this function doesn't ignore
3484 * ->no_numa as it is used for both pool and wq attrs. Instead,
3485 * get_unbound_pool() explicitly clears ->no_numa after copying.
3487 to->no_numa = from->no_numa;
3490 /* hash value of the content of @attr */
3491 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3495 hash = jhash_1word(attrs->nice, hash);
3496 hash = jhash(cpumask_bits(attrs->cpumask),
3497 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3501 /* content equality test */
3502 static bool wqattrs_equal(const struct workqueue_attrs *a,
3503 const struct workqueue_attrs *b)
3505 if (a->nice != b->nice)
3507 if (!cpumask_equal(a->cpumask, b->cpumask))
3513 * init_worker_pool - initialize a newly zalloc'd worker_pool
3514 * @pool: worker_pool to initialize
3516 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3518 * Return: 0 on success, -errno on failure. Even on failure, all fields
3519 * inside @pool proper are initialized and put_unbound_pool() can be called
3520 * on @pool safely to release it.
3522 static int init_worker_pool(struct worker_pool *pool)
3524 spin_lock_init(&pool->lock);
3527 pool->node = NUMA_NO_NODE;
3528 pool->flags |= POOL_DISASSOCIATED;
3529 INIT_LIST_HEAD(&pool->worklist);
3530 INIT_LIST_HEAD(&pool->idle_list);
3531 hash_init(pool->busy_hash);
3533 init_timer_deferrable(&pool->idle_timer);
3534 pool->idle_timer.function = idle_worker_timeout;
3535 pool->idle_timer.data = (unsigned long)pool;
3537 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3538 (unsigned long)pool);
3540 mutex_init(&pool->manager_arb);
3541 mutex_init(&pool->manager_mutex);
3542 idr_init(&pool->worker_idr);
3544 INIT_HLIST_NODE(&pool->hash_node);
3547 /* shouldn't fail above this point */
3548 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3554 static void rcu_free_pool(struct rcu_head *rcu)
3556 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3558 idr_destroy(&pool->worker_idr);
3559 free_workqueue_attrs(pool->attrs);
3564 * put_unbound_pool - put a worker_pool
3565 * @pool: worker_pool to put
3567 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3568 * safe manner. get_unbound_pool() calls this function on its failure path
3569 * and this function should be able to release pools which went through,
3570 * successfully or not, init_worker_pool().
3572 * Should be called with wq_pool_mutex held.
3574 static void put_unbound_pool(struct worker_pool *pool)
3576 struct worker *worker;
3578 lockdep_assert_held(&wq_pool_mutex);
3584 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3585 WARN_ON(!list_empty(&pool->worklist)))
3588 /* release id and unhash */
3590 idr_remove(&worker_pool_idr, pool->id);
3591 hash_del(&pool->hash_node);
3594 * Become the manager and destroy all workers. Grabbing
3595 * manager_arb prevents @pool's workers from blocking on
3598 mutex_lock(&pool->manager_arb);
3599 mutex_lock(&pool->manager_mutex);
3600 spin_lock_irq(&pool->lock);
3602 while ((worker = first_worker(pool)))
3603 destroy_worker(worker);
3604 WARN_ON(pool->nr_workers || pool->nr_idle);
3606 spin_unlock_irq(&pool->lock);
3607 mutex_unlock(&pool->manager_mutex);
3608 mutex_unlock(&pool->manager_arb);
3610 /* shut down the timers */
3611 del_timer_sync(&pool->idle_timer);
3612 del_timer_sync(&pool->mayday_timer);
3614 /* sched-RCU protected to allow dereferences from get_work_pool() */
3615 call_rcu_sched(&pool->rcu, rcu_free_pool);
3619 * get_unbound_pool - get a worker_pool with the specified attributes
3620 * @attrs: the attributes of the worker_pool to get
3622 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3623 * reference count and return it. If there already is a matching
3624 * worker_pool, it will be used; otherwise, this function attempts to
3627 * Should be called with wq_pool_mutex held.
3629 * Return: On success, a worker_pool with the same attributes as @attrs.
3630 * On failure, %NULL.
3632 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3634 u32 hash = wqattrs_hash(attrs);
3635 struct worker_pool *pool;
3638 lockdep_assert_held(&wq_pool_mutex);
3640 /* do we already have a matching pool? */
3641 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3642 if (wqattrs_equal(pool->attrs, attrs)) {
3648 /* nope, create a new one */
3649 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3650 if (!pool || init_worker_pool(pool) < 0)
3653 if (workqueue_freezing)
3654 pool->flags |= POOL_FREEZING;
3656 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3657 copy_workqueue_attrs(pool->attrs, attrs);
3660 * no_numa isn't a worker_pool attribute, always clear it. See
3661 * 'struct workqueue_attrs' comments for detail.
3663 pool->attrs->no_numa = false;
3665 /* if cpumask is contained inside a NUMA node, we belong to that node */
3666 if (wq_numa_enabled) {
3667 for_each_node(node) {
3668 if (cpumask_subset(pool->attrs->cpumask,
3669 wq_numa_possible_cpumask[node])) {
3676 if (worker_pool_assign_id(pool) < 0)
3679 /* create and start the initial worker */
3680 if (create_and_start_worker(pool) < 0)
3684 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3689 put_unbound_pool(pool);
3693 static void rcu_free_pwq(struct rcu_head *rcu)
3695 kmem_cache_free(pwq_cache,
3696 container_of(rcu, struct pool_workqueue, rcu));
3700 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3701 * and needs to be destroyed.
3703 static void pwq_unbound_release_workfn(struct work_struct *work)
3705 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3706 unbound_release_work);
3707 struct workqueue_struct *wq = pwq->wq;
3708 struct worker_pool *pool = pwq->pool;
3711 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3715 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3716 * necessary on release but do it anyway. It's easier to verify
3717 * and consistent with the linking path.
3719 mutex_lock(&wq->mutex);
3720 list_del_rcu(&pwq->pwqs_node);
3721 is_last = list_empty(&wq->pwqs);
3722 mutex_unlock(&wq->mutex);
3724 mutex_lock(&wq_pool_mutex);
3725 put_unbound_pool(pool);
3726 mutex_unlock(&wq_pool_mutex);
3728 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3731 * If we're the last pwq going away, @wq is already dead and no one
3732 * is gonna access it anymore. Free it.
3735 free_workqueue_attrs(wq->unbound_attrs);
3741 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3742 * @pwq: target pool_workqueue
3744 * If @pwq isn't freezing, set @pwq->max_active to the associated
3745 * workqueue's saved_max_active and activate delayed work items
3746 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3748 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3750 struct workqueue_struct *wq = pwq->wq;
3751 bool freezable = wq->flags & WQ_FREEZABLE;
3753 /* for @wq->saved_max_active */
3754 lockdep_assert_held(&wq->mutex);
3756 /* fast exit for non-freezable wqs */
3757 if (!freezable && pwq->max_active == wq->saved_max_active)
3760 spin_lock_irq(&pwq->pool->lock);
3762 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3763 pwq->max_active = wq->saved_max_active;
3765 while (!list_empty(&pwq->delayed_works) &&
3766 pwq->nr_active < pwq->max_active)
3767 pwq_activate_first_delayed(pwq);
3770 * Need to kick a worker after thawed or an unbound wq's
3771 * max_active is bumped. It's a slow path. Do it always.
3773 wake_up_worker(pwq->pool);
3775 pwq->max_active = 0;
3778 spin_unlock_irq(&pwq->pool->lock);
3781 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3782 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3783 struct worker_pool *pool)
3785 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3787 memset(pwq, 0, sizeof(*pwq));
3791 pwq->flush_color = -1;
3793 INIT_LIST_HEAD(&pwq->delayed_works);
3794 INIT_LIST_HEAD(&pwq->pwqs_node);
3795 INIT_LIST_HEAD(&pwq->mayday_node);
3796 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3799 /* sync @pwq with the current state of its associated wq and link it */
3800 static void link_pwq(struct pool_workqueue *pwq)
3802 struct workqueue_struct *wq = pwq->wq;
3804 lockdep_assert_held(&wq->mutex);
3806 /* may be called multiple times, ignore if already linked */
3807 if (!list_empty(&pwq->pwqs_node))
3811 * Set the matching work_color. This is synchronized with
3812 * wq->mutex to avoid confusing flush_workqueue().
3814 pwq->work_color = wq->work_color;
3816 /* sync max_active to the current setting */
3817 pwq_adjust_max_active(pwq);
3820 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3823 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3824 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3825 const struct workqueue_attrs *attrs)
3827 struct worker_pool *pool;
3828 struct pool_workqueue *pwq;
3830 lockdep_assert_held(&wq_pool_mutex);
3832 pool = get_unbound_pool(attrs);
3836 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3838 put_unbound_pool(pool);
3842 init_pwq(pwq, wq, pool);
3846 /* undo alloc_unbound_pwq(), used only in the error path */
3847 static void free_unbound_pwq(struct pool_workqueue *pwq)
3849 lockdep_assert_held(&wq_pool_mutex);
3852 put_unbound_pool(pwq->pool);
3853 kmem_cache_free(pwq_cache, pwq);
3858 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3859 * @attrs: the wq_attrs of interest
3860 * @node: the target NUMA node
3861 * @cpu_going_down: if >= 0, the CPU to consider as offline
3862 * @cpumask: outarg, the resulting cpumask
3864 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3865 * @cpu_going_down is >= 0, that cpu is considered offline during
3866 * calculation. The result is stored in @cpumask.
3868 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3869 * enabled and @node has online CPUs requested by @attrs, the returned
3870 * cpumask is the intersection of the possible CPUs of @node and
3873 * The caller is responsible for ensuring that the cpumask of @node stays
3876 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3879 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3880 int cpu_going_down, cpumask_t *cpumask)
3882 if (!wq_numa_enabled || attrs->no_numa)
3885 /* does @node have any online CPUs @attrs wants? */
3886 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3887 if (cpu_going_down >= 0)
3888 cpumask_clear_cpu(cpu_going_down, cpumask);
3890 if (cpumask_empty(cpumask))
3893 /* yeap, return possible CPUs in @node that @attrs wants */
3894 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3895 return !cpumask_equal(cpumask, attrs->cpumask);
3898 cpumask_copy(cpumask, attrs->cpumask);
3902 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3903 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3905 struct pool_workqueue *pwq)
3907 struct pool_workqueue *old_pwq;
3909 lockdep_assert_held(&wq->mutex);
3911 /* link_pwq() can handle duplicate calls */
3914 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3915 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3920 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3921 * @wq: the target workqueue
3922 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3924 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3925 * machines, this function maps a separate pwq to each NUMA node with
3926 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3927 * NUMA node it was issued on. Older pwqs are released as in-flight work
3928 * items finish. Note that a work item which repeatedly requeues itself
3929 * back-to-back will stay on its current pwq.
3931 * Performs GFP_KERNEL allocations.
3933 * Return: 0 on success and -errno on failure.
3935 int apply_workqueue_attrs(struct workqueue_struct *wq,
3936 const struct workqueue_attrs *attrs)
3938 struct workqueue_attrs *new_attrs, *tmp_attrs;
3939 struct pool_workqueue **pwq_tbl, *dfl_pwq;
3942 /* only unbound workqueues can change attributes */
3943 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3946 /* creating multiple pwqs breaks ordering guarantee */
3947 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3950 pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3951 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3952 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3953 if (!pwq_tbl || !new_attrs || !tmp_attrs)
3956 /* make a copy of @attrs and sanitize it */
3957 copy_workqueue_attrs(new_attrs, attrs);
3958 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3961 * We may create multiple pwqs with differing cpumasks. Make a
3962 * copy of @new_attrs which will be modified and used to obtain
3965 copy_workqueue_attrs(tmp_attrs, new_attrs);
3968 * CPUs should stay stable across pwq creations and installations.
3969 * Pin CPUs, determine the target cpumask for each node and create
3974 mutex_lock(&wq_pool_mutex);
3977 * If something goes wrong during CPU up/down, we'll fall back to
3978 * the default pwq covering whole @attrs->cpumask. Always create
3979 * it even if we don't use it immediately.
3981 dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3985 for_each_node(node) {
3986 if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3987 pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3992 pwq_tbl[node] = dfl_pwq;
3996 mutex_unlock(&wq_pool_mutex);
3998 /* all pwqs have been created successfully, let's install'em */
3999 mutex_lock(&wq->mutex);
4001 copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
4003 /* save the previous pwq and install the new one */
4005 pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
4007 /* @dfl_pwq might not have been used, ensure it's linked */
4009 swap(wq->dfl_pwq, dfl_pwq);
4011 mutex_unlock(&wq->mutex);
4013 /* put the old pwqs */
4015 put_pwq_unlocked(pwq_tbl[node]);
4016 put_pwq_unlocked(dfl_pwq);
4022 free_workqueue_attrs(tmp_attrs);
4023 free_workqueue_attrs(new_attrs);
4028 free_unbound_pwq(dfl_pwq);
4030 if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
4031 free_unbound_pwq(pwq_tbl[node]);
4032 mutex_unlock(&wq_pool_mutex);
4040 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4041 * @wq: the target workqueue
4042 * @cpu: the CPU coming up or going down
4043 * @online: whether @cpu is coming up or going down
4045 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4046 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4049 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4050 * falls back to @wq->dfl_pwq which may not be optimal but is always
4053 * Note that when the last allowed CPU of a NUMA node goes offline for a
4054 * workqueue with a cpumask spanning multiple nodes, the workers which were
4055 * already executing the work items for the workqueue will lose their CPU
4056 * affinity and may execute on any CPU. This is similar to how per-cpu
4057 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4058 * affinity, it's the user's responsibility to flush the work item from
4061 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4064 int node = cpu_to_node(cpu);
4065 int cpu_off = online ? -1 : cpu;
4066 struct pool_workqueue *old_pwq = NULL, *pwq;
4067 struct workqueue_attrs *target_attrs;
4070 lockdep_assert_held(&wq_pool_mutex);
4072 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4076 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4077 * Let's use a preallocated one. The following buf is protected by
4078 * CPU hotplug exclusion.
4080 target_attrs = wq_update_unbound_numa_attrs_buf;
4081 cpumask = target_attrs->cpumask;
4083 mutex_lock(&wq->mutex);
4084 if (wq->unbound_attrs->no_numa)
4087 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4088 pwq = unbound_pwq_by_node(wq, node);
4091 * Let's determine what needs to be done. If the target cpumask is
4092 * different from wq's, we need to compare it to @pwq's and create
4093 * a new one if they don't match. If the target cpumask equals
4094 * wq's, the default pwq should be used. If @pwq is already the
4095 * default one, nothing to do; otherwise, install the default one.
4097 if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4098 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4101 if (pwq == wq->dfl_pwq)
4107 mutex_unlock(&wq->mutex);
4109 /* create a new pwq */
4110 pwq = alloc_unbound_pwq(wq, target_attrs);
4112 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4114 mutex_lock(&wq->mutex);
4119 * Install the new pwq. As this function is called only from CPU
4120 * hotplug callbacks and applying a new attrs is wrapped with
4121 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4124 mutex_lock(&wq->mutex);
4125 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4129 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4130 get_pwq(wq->dfl_pwq);
4131 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4132 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4134 mutex_unlock(&wq->mutex);
4135 put_pwq_unlocked(old_pwq);
4138 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4140 bool highpri = wq->flags & WQ_HIGHPRI;
4143 if (!(wq->flags & WQ_UNBOUND)) {
4144 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4148 for_each_possible_cpu(cpu) {
4149 struct pool_workqueue *pwq =
4150 per_cpu_ptr(wq->cpu_pwqs, cpu);
4151 struct worker_pool *cpu_pools =
4152 per_cpu(cpu_worker_pools, cpu);
4154 init_pwq(pwq, wq, &cpu_pools[highpri]);
4156 mutex_lock(&wq->mutex);
4158 mutex_unlock(&wq->mutex);
4161 } else if (wq->flags & __WQ_ORDERED) {
4162 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4163 /* there should only be single pwq for ordering guarantee */
4164 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4165 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4166 "ordering guarantee broken for workqueue %s\n", wq->name);
4169 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4173 static int wq_clamp_max_active(int max_active, unsigned int flags,
4176 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4178 if (max_active < 1 || max_active > lim)
4179 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4180 max_active, name, 1, lim);
4182 return clamp_val(max_active, 1, lim);
4185 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4188 struct lock_class_key *key,
4189 const char *lock_name, ...)
4191 size_t tbl_size = 0;
4193 struct workqueue_struct *wq;
4194 struct pool_workqueue *pwq;
4196 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4197 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4198 flags |= WQ_UNBOUND;
4200 /* allocate wq and format name */
4201 if (flags & WQ_UNBOUND)
4202 tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4204 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4208 if (flags & WQ_UNBOUND) {
4209 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4210 if (!wq->unbound_attrs)
4214 va_start(args, lock_name);
4215 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4218 max_active = max_active ?: WQ_DFL_ACTIVE;
4219 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4223 wq->saved_max_active = max_active;
4224 mutex_init(&wq->mutex);
4225 atomic_set(&wq->nr_pwqs_to_flush, 0);
4226 INIT_LIST_HEAD(&wq->pwqs);
4227 INIT_LIST_HEAD(&wq->flusher_queue);
4228 INIT_LIST_HEAD(&wq->flusher_overflow);
4229 INIT_LIST_HEAD(&wq->maydays);
4231 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4232 INIT_LIST_HEAD(&wq->list);
4234 if (alloc_and_link_pwqs(wq) < 0)
4238 * Workqueues which may be used during memory reclaim should
4239 * have a rescuer to guarantee forward progress.
4241 if (flags & WQ_MEM_RECLAIM) {
4242 struct worker *rescuer;
4244 rescuer = alloc_worker();
4248 rescuer->rescue_wq = wq;
4249 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4251 if (IS_ERR(rescuer->task)) {
4256 wq->rescuer = rescuer;
4257 rescuer->task->flags |= PF_NO_SETAFFINITY;
4258 wake_up_process(rescuer->task);
4261 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4265 * wq_pool_mutex protects global freeze state and workqueues list.
4266 * Grab it, adjust max_active and add the new @wq to workqueues
4269 mutex_lock(&wq_pool_mutex);
4271 mutex_lock(&wq->mutex);
4272 for_each_pwq(pwq, wq)
4273 pwq_adjust_max_active(pwq);
4274 mutex_unlock(&wq->mutex);
4276 list_add(&wq->list, &workqueues);
4278 mutex_unlock(&wq_pool_mutex);
4283 free_workqueue_attrs(wq->unbound_attrs);
4287 destroy_workqueue(wq);
4290 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4293 * destroy_workqueue - safely terminate a workqueue
4294 * @wq: target workqueue
4296 * Safely destroy a workqueue. All work currently pending will be done first.
4298 void destroy_workqueue(struct workqueue_struct *wq)
4300 struct pool_workqueue *pwq;
4303 /* drain it before proceeding with destruction */
4304 drain_workqueue(wq);
4307 mutex_lock(&wq->mutex);
4308 for_each_pwq(pwq, wq) {
4311 for (i = 0; i < WORK_NR_COLORS; i++) {
4312 if (WARN_ON(pwq->nr_in_flight[i])) {
4313 mutex_unlock(&wq->mutex);
4318 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4319 WARN_ON(pwq->nr_active) ||
4320 WARN_ON(!list_empty(&pwq->delayed_works))) {
4321 mutex_unlock(&wq->mutex);
4325 mutex_unlock(&wq->mutex);
4328 * wq list is used to freeze wq, remove from list after
4329 * flushing is complete in case freeze races us.
4331 mutex_lock(&wq_pool_mutex);
4332 list_del_init(&wq->list);
4333 mutex_unlock(&wq_pool_mutex);
4335 workqueue_sysfs_unregister(wq);
4338 kthread_stop(wq->rescuer->task);
4343 if (!(wq->flags & WQ_UNBOUND)) {
4345 * The base ref is never dropped on per-cpu pwqs. Directly
4346 * free the pwqs and wq.
4348 free_percpu(wq->cpu_pwqs);
4352 * We're the sole accessor of @wq at this point. Directly
4353 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4354 * @wq will be freed when the last pwq is released.
4356 for_each_node(node) {
4357 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4358 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4359 put_pwq_unlocked(pwq);
4363 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4364 * put. Don't access it afterwards.
4368 put_pwq_unlocked(pwq);
4371 EXPORT_SYMBOL_GPL(destroy_workqueue);
4374 * workqueue_set_max_active - adjust max_active of a workqueue
4375 * @wq: target workqueue
4376 * @max_active: new max_active value.
4378 * Set max_active of @wq to @max_active.
4381 * Don't call from IRQ context.
4383 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4385 struct pool_workqueue *pwq;
4387 /* disallow meddling with max_active for ordered workqueues */
4388 if (WARN_ON(wq->flags & __WQ_ORDERED))
4391 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4393 mutex_lock(&wq->mutex);
4395 wq->saved_max_active = max_active;
4397 for_each_pwq(pwq, wq)
4398 pwq_adjust_max_active(pwq);
4400 mutex_unlock(&wq->mutex);
4402 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4405 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4407 * Determine whether %current is a workqueue rescuer. Can be used from
4408 * work functions to determine whether it's being run off the rescuer task.
4410 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4412 bool current_is_workqueue_rescuer(void)
4414 struct worker *worker = current_wq_worker();
4416 return worker && worker->rescue_wq;
4420 * workqueue_congested - test whether a workqueue is congested
4421 * @cpu: CPU in question
4422 * @wq: target workqueue
4424 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4425 * no synchronization around this function and the test result is
4426 * unreliable and only useful as advisory hints or for debugging.
4428 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4429 * Note that both per-cpu and unbound workqueues may be associated with
4430 * multiple pool_workqueues which have separate congested states. A
4431 * workqueue being congested on one CPU doesn't mean the workqueue is also
4432 * contested on other CPUs / NUMA nodes.
4435 * %true if congested, %false otherwise.
4437 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4439 struct pool_workqueue *pwq;
4442 rcu_read_lock_sched();
4444 if (cpu == WORK_CPU_UNBOUND)
4445 cpu = smp_processor_id();
4447 if (!(wq->flags & WQ_UNBOUND))
4448 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4450 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4452 ret = !list_empty(&pwq->delayed_works);
4453 rcu_read_unlock_sched();
4457 EXPORT_SYMBOL_GPL(workqueue_congested);
4460 * work_busy - test whether a work is currently pending or running
4461 * @work: the work to be tested
4463 * Test whether @work is currently pending or running. There is no
4464 * synchronization around this function and the test result is
4465 * unreliable and only useful as advisory hints or for debugging.
4468 * OR'd bitmask of WORK_BUSY_* bits.
4470 unsigned int work_busy(struct work_struct *work)
4472 struct worker_pool *pool;
4473 unsigned long flags;
4474 unsigned int ret = 0;
4476 if (work_pending(work))
4477 ret |= WORK_BUSY_PENDING;
4479 local_irq_save(flags);
4480 pool = get_work_pool(work);
4482 spin_lock(&pool->lock);
4483 if (find_worker_executing_work(pool, work))
4484 ret |= WORK_BUSY_RUNNING;
4485 spin_unlock(&pool->lock);
4487 local_irq_restore(flags);
4491 EXPORT_SYMBOL_GPL(work_busy);
4494 * set_worker_desc - set description for the current work item
4495 * @fmt: printf-style format string
4496 * @...: arguments for the format string
4498 * This function can be called by a running work function to describe what
4499 * the work item is about. If the worker task gets dumped, this
4500 * information will be printed out together to help debugging. The
4501 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4503 void set_worker_desc(const char *fmt, ...)
4505 struct worker *worker = current_wq_worker();
4509 va_start(args, fmt);
4510 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4512 worker->desc_valid = true;
4517 * print_worker_info - print out worker information and description
4518 * @log_lvl: the log level to use when printing
4519 * @task: target task
4521 * If @task is a worker and currently executing a work item, print out the
4522 * name of the workqueue being serviced and worker description set with
4523 * set_worker_desc() by the currently executing work item.
4525 * This function can be safely called on any task as long as the
4526 * task_struct itself is accessible. While safe, this function isn't
4527 * synchronized and may print out mixups or garbages of limited length.
4529 void print_worker_info(const char *log_lvl, struct task_struct *task)
4531 work_func_t *fn = NULL;
4532 char name[WQ_NAME_LEN] = { };
4533 char desc[WORKER_DESC_LEN] = { };
4534 struct pool_workqueue *pwq = NULL;
4535 struct workqueue_struct *wq = NULL;
4536 bool desc_valid = false;
4537 struct worker *worker;
4539 if (!(task->flags & PF_WQ_WORKER))
4543 * This function is called without any synchronization and @task
4544 * could be in any state. Be careful with dereferences.
4546 worker = probe_kthread_data(task);
4549 * Carefully copy the associated workqueue's workfn and name. Keep
4550 * the original last '\0' in case the original contains garbage.
4552 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4553 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4554 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4555 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4557 /* copy worker description */
4558 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4560 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4562 if (fn || name[0] || desc[0]) {
4563 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4565 pr_cont(" (%s)", desc);
4573 * There are two challenges in supporting CPU hotplug. Firstly, there
4574 * are a lot of assumptions on strong associations among work, pwq and
4575 * pool which make migrating pending and scheduled works very
4576 * difficult to implement without impacting hot paths. Secondly,
4577 * worker pools serve mix of short, long and very long running works making
4578 * blocked draining impractical.
4580 * This is solved by allowing the pools to be disassociated from the CPU
4581 * running as an unbound one and allowing it to be reattached later if the
4582 * cpu comes back online.
4585 static void wq_unbind_fn(struct work_struct *work)
4587 int cpu = smp_processor_id();
4588 struct worker_pool *pool;
4589 struct worker *worker;
4592 for_each_cpu_worker_pool(pool, cpu) {
4593 WARN_ON_ONCE(cpu != smp_processor_id());
4595 mutex_lock(&pool->manager_mutex);
4596 spin_lock_irq(&pool->lock);
4599 * We've blocked all manager operations. Make all workers
4600 * unbound and set DISASSOCIATED. Before this, all workers
4601 * except for the ones which are still executing works from
4602 * before the last CPU down must be on the cpu. After
4603 * this, they may become diasporas.
4605 for_each_pool_worker(worker, wi, pool)
4606 worker->flags |= WORKER_UNBOUND;
4608 pool->flags |= POOL_DISASSOCIATED;
4610 spin_unlock_irq(&pool->lock);
4611 mutex_unlock(&pool->manager_mutex);
4614 * Call schedule() so that we cross rq->lock and thus can
4615 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4616 * This is necessary as scheduler callbacks may be invoked
4622 * Sched callbacks are disabled now. Zap nr_running.
4623 * After this, nr_running stays zero and need_more_worker()
4624 * and keep_working() are always true as long as the
4625 * worklist is not empty. This pool now behaves as an
4626 * unbound (in terms of concurrency management) pool which
4627 * are served by workers tied to the pool.
4629 atomic_set(&pool->nr_running, 0);
4632 * With concurrency management just turned off, a busy
4633 * worker blocking could lead to lengthy stalls. Kick off
4634 * unbound chain execution of currently pending work items.
4636 spin_lock_irq(&pool->lock);
4637 wake_up_worker(pool);
4638 spin_unlock_irq(&pool->lock);
4643 * rebind_workers - rebind all workers of a pool to the associated CPU
4644 * @pool: pool of interest
4646 * @pool->cpu is coming online. Rebind all workers to the CPU.
4648 static void rebind_workers(struct worker_pool *pool)
4650 struct worker *worker;
4653 lockdep_assert_held(&pool->manager_mutex);
4656 * Restore CPU affinity of all workers. As all idle workers should
4657 * be on the run-queue of the associated CPU before any local
4658 * wake-ups for concurrency management happen, restore CPU affinty
4659 * of all workers first and then clear UNBOUND. As we're called
4660 * from CPU_ONLINE, the following shouldn't fail.
4662 for_each_pool_worker(worker, wi, pool)
4663 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4664 pool->attrs->cpumask) < 0);
4666 spin_lock_irq(&pool->lock);
4668 for_each_pool_worker(worker, wi, pool) {
4669 unsigned int worker_flags = worker->flags;
4672 * A bound idle worker should actually be on the runqueue
4673 * of the associated CPU for local wake-ups targeting it to
4674 * work. Kick all idle workers so that they migrate to the
4675 * associated CPU. Doing this in the same loop as
4676 * replacing UNBOUND with REBOUND is safe as no worker will
4677 * be bound before @pool->lock is released.
4679 if (worker_flags & WORKER_IDLE)
4680 wake_up_process(worker->task);
4683 * We want to clear UNBOUND but can't directly call
4684 * worker_clr_flags() or adjust nr_running. Atomically
4685 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4686 * @worker will clear REBOUND using worker_clr_flags() when
4687 * it initiates the next execution cycle thus restoring
4688 * concurrency management. Note that when or whether
4689 * @worker clears REBOUND doesn't affect correctness.
4691 * ACCESS_ONCE() is necessary because @worker->flags may be
4692 * tested without holding any lock in
4693 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4694 * fail incorrectly leading to premature concurrency
4695 * management operations.
4697 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4698 worker_flags |= WORKER_REBOUND;
4699 worker_flags &= ~WORKER_UNBOUND;
4700 ACCESS_ONCE(worker->flags) = worker_flags;
4703 spin_unlock_irq(&pool->lock);
4707 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4708 * @pool: unbound pool of interest
4709 * @cpu: the CPU which is coming up
4711 * An unbound pool may end up with a cpumask which doesn't have any online
4712 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4713 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4714 * online CPU before, cpus_allowed of all its workers should be restored.
4716 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4718 static cpumask_t cpumask;
4719 struct worker *worker;
4722 lockdep_assert_held(&pool->manager_mutex);
4724 /* is @cpu allowed for @pool? */
4725 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4728 /* is @cpu the only online CPU? */
4729 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4730 if (cpumask_weight(&cpumask) != 1)
4733 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4734 for_each_pool_worker(worker, wi, pool)
4735 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4736 pool->attrs->cpumask) < 0);
4740 * Workqueues should be brought up before normal priority CPU notifiers.
4741 * This will be registered high priority CPU notifier.
4743 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4744 unsigned long action,
4747 int cpu = (unsigned long)hcpu;
4748 struct worker_pool *pool;
4749 struct workqueue_struct *wq;
4752 switch (action & ~CPU_TASKS_FROZEN) {
4753 case CPU_UP_PREPARE:
4754 for_each_cpu_worker_pool(pool, cpu) {
4755 if (pool->nr_workers)
4757 if (create_and_start_worker(pool) < 0)
4762 case CPU_DOWN_FAILED:
4764 mutex_lock(&wq_pool_mutex);
4766 for_each_pool(pool, pi) {
4767 mutex_lock(&pool->manager_mutex);
4769 if (pool->cpu == cpu) {
4770 spin_lock_irq(&pool->lock);
4771 pool->flags &= ~POOL_DISASSOCIATED;
4772 spin_unlock_irq(&pool->lock);
4774 rebind_workers(pool);
4775 } else if (pool->cpu < 0) {
4776 restore_unbound_workers_cpumask(pool, cpu);
4779 mutex_unlock(&pool->manager_mutex);
4782 /* update NUMA affinity of unbound workqueues */
4783 list_for_each_entry(wq, &workqueues, list)
4784 wq_update_unbound_numa(wq, cpu, true);
4786 mutex_unlock(&wq_pool_mutex);
4793 * Workqueues should be brought down after normal priority CPU notifiers.
4794 * This will be registered as low priority CPU notifier.
4796 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4797 unsigned long action,
4800 int cpu = (unsigned long)hcpu;
4801 struct work_struct unbind_work;
4802 struct workqueue_struct *wq;
4804 switch (action & ~CPU_TASKS_FROZEN) {
4805 case CPU_DOWN_PREPARE:
4806 /* unbinding per-cpu workers should happen on the local CPU */
4807 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4808 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4810 /* update NUMA affinity of unbound workqueues */
4811 mutex_lock(&wq_pool_mutex);
4812 list_for_each_entry(wq, &workqueues, list)
4813 wq_update_unbound_numa(wq, cpu, false);
4814 mutex_unlock(&wq_pool_mutex);
4816 /* wait for per-cpu unbinding to finish */
4817 flush_work(&unbind_work);
4818 destroy_work_on_stack(&unbind_work);
4826 struct work_for_cpu {
4827 struct work_struct work;
4833 static void work_for_cpu_fn(struct work_struct *work)
4835 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4837 wfc->ret = wfc->fn(wfc->arg);
4841 * work_on_cpu - run a function in user context on a particular cpu
4842 * @cpu: the cpu to run on
4843 * @fn: the function to run
4844 * @arg: the function arg
4846 * It is up to the caller to ensure that the cpu doesn't go offline.
4847 * The caller must not hold any locks which would prevent @fn from completing.
4849 * Return: The value @fn returns.
4851 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4853 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4855 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4856 schedule_work_on(cpu, &wfc.work);
4857 flush_work(&wfc.work);
4858 destroy_work_on_stack(&wfc.work);
4861 EXPORT_SYMBOL_GPL(work_on_cpu);
4862 #endif /* CONFIG_SMP */
4864 #ifdef CONFIG_FREEZER
4867 * freeze_workqueues_begin - begin freezing workqueues
4869 * Start freezing workqueues. After this function returns, all freezable
4870 * workqueues will queue new works to their delayed_works list instead of
4874 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4876 void freeze_workqueues_begin(void)
4878 struct worker_pool *pool;
4879 struct workqueue_struct *wq;
4880 struct pool_workqueue *pwq;
4883 mutex_lock(&wq_pool_mutex);
4885 WARN_ON_ONCE(workqueue_freezing);
4886 workqueue_freezing = true;
4889 for_each_pool(pool, pi) {
4890 spin_lock_irq(&pool->lock);
4891 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4892 pool->flags |= POOL_FREEZING;
4893 spin_unlock_irq(&pool->lock);
4896 list_for_each_entry(wq, &workqueues, list) {
4897 mutex_lock(&wq->mutex);
4898 for_each_pwq(pwq, wq)
4899 pwq_adjust_max_active(pwq);
4900 mutex_unlock(&wq->mutex);
4903 mutex_unlock(&wq_pool_mutex);
4907 * freeze_workqueues_busy - are freezable workqueues still busy?
4909 * Check whether freezing is complete. This function must be called
4910 * between freeze_workqueues_begin() and thaw_workqueues().
4913 * Grabs and releases wq_pool_mutex.
4916 * %true if some freezable workqueues are still busy. %false if freezing
4919 bool freeze_workqueues_busy(void)
4922 struct workqueue_struct *wq;
4923 struct pool_workqueue *pwq;
4925 mutex_lock(&wq_pool_mutex);
4927 WARN_ON_ONCE(!workqueue_freezing);
4929 list_for_each_entry(wq, &workqueues, list) {
4930 if (!(wq->flags & WQ_FREEZABLE))
4933 * nr_active is monotonically decreasing. It's safe
4934 * to peek without lock.
4936 rcu_read_lock_sched();
4937 for_each_pwq(pwq, wq) {
4938 WARN_ON_ONCE(pwq->nr_active < 0);
4939 if (pwq->nr_active) {
4941 rcu_read_unlock_sched();
4945 rcu_read_unlock_sched();
4948 mutex_unlock(&wq_pool_mutex);
4953 * thaw_workqueues - thaw workqueues
4955 * Thaw workqueues. Normal queueing is restored and all collected
4956 * frozen works are transferred to their respective pool worklists.
4959 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4961 void thaw_workqueues(void)
4963 struct workqueue_struct *wq;
4964 struct pool_workqueue *pwq;
4965 struct worker_pool *pool;
4968 mutex_lock(&wq_pool_mutex);
4970 if (!workqueue_freezing)
4973 /* clear FREEZING */
4974 for_each_pool(pool, pi) {
4975 spin_lock_irq(&pool->lock);
4976 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4977 pool->flags &= ~POOL_FREEZING;
4978 spin_unlock_irq(&pool->lock);
4981 /* restore max_active and repopulate worklist */
4982 list_for_each_entry(wq, &workqueues, list) {
4983 mutex_lock(&wq->mutex);
4984 for_each_pwq(pwq, wq)
4985 pwq_adjust_max_active(pwq);
4986 mutex_unlock(&wq->mutex);
4989 workqueue_freezing = false;
4991 mutex_unlock(&wq_pool_mutex);
4993 #endif /* CONFIG_FREEZER */
4995 static void __init wq_numa_init(void)
5000 /* determine NUMA pwq table len - highest node id + 1 */
5002 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
5004 if (num_possible_nodes() <= 1)
5007 if (wq_disable_numa) {
5008 pr_info("workqueue: NUMA affinity support disabled\n");
5012 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5013 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5016 * We want masks of possible CPUs of each node which isn't readily
5017 * available. Build one from cpu_to_node() which should have been
5018 * fully initialized by now.
5020 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
5024 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5025 node_online(node) ? node : NUMA_NO_NODE));
5027 for_each_possible_cpu(cpu) {
5028 node = cpu_to_node(cpu);
5029 if (WARN_ON(node == NUMA_NO_NODE)) {
5030 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5031 /* happens iff arch is bonkers, let's just proceed */
5034 cpumask_set_cpu(cpu, tbl[node]);
5037 wq_numa_possible_cpumask = tbl;
5038 wq_numa_enabled = true;
5041 static int __init init_workqueues(void)
5043 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5046 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5048 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5050 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5051 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5055 /* initialize CPU pools */
5056 for_each_possible_cpu(cpu) {
5057 struct worker_pool *pool;
5060 for_each_cpu_worker_pool(pool, cpu) {
5061 BUG_ON(init_worker_pool(pool));
5063 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5064 pool->attrs->nice = std_nice[i++];
5065 pool->node = cpu_to_node(cpu);
5068 mutex_lock(&wq_pool_mutex);
5069 BUG_ON(worker_pool_assign_id(pool));
5070 mutex_unlock(&wq_pool_mutex);
5074 /* create the initial worker */
5075 for_each_online_cpu(cpu) {
5076 struct worker_pool *pool;
5078 for_each_cpu_worker_pool(pool, cpu) {
5079 pool->flags &= ~POOL_DISASSOCIATED;
5080 BUG_ON(create_and_start_worker(pool) < 0);
5084 /* create default unbound and ordered wq attrs */
5085 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5086 struct workqueue_attrs *attrs;
5088 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5089 attrs->nice = std_nice[i];
5090 unbound_std_wq_attrs[i] = attrs;
5093 * An ordered wq should have only one pwq as ordering is
5094 * guaranteed by max_active which is enforced by pwqs.
5095 * Turn off NUMA so that dfl_pwq is used for all nodes.
5097 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5098 attrs->nice = std_nice[i];
5099 attrs->no_numa = true;
5100 ordered_wq_attrs[i] = attrs;
5103 system_wq = alloc_workqueue("events", 0, 0);
5104 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5105 system_long_wq = alloc_workqueue("events_long", 0, 0);
5106 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5107 WQ_UNBOUND_MAX_ACTIVE);
5108 system_freezable_wq = alloc_workqueue("events_freezable",
5110 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5111 WQ_POWER_EFFICIENT, 0);
5112 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5113 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5115 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5116 !system_unbound_wq || !system_freezable_wq ||
5117 !system_power_efficient_wq ||
5118 !system_freezable_power_efficient_wq);
5121 early_initcall(init_workqueues);