1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.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>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
53 #include <linux/kvm_para.h>
55 #include "workqueue_internal.h"
61 * A bound pool is either associated or disassociated with its CPU.
62 * While associated (!DISASSOCIATED), all workers are bound to the
63 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 * While DISASSOCIATED, the cpu may be offline and all workers have
67 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 * be executing on any CPU. The pool behaves as an unbound one.
70 * Note that DISASSOCIATED should be flipped only while holding
71 * wq_pool_attach_mutex to avoid changing binding state while
72 * worker_attach_to_pool() is in progress.
74 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
75 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
78 WORKER_DIE = 1 << 1, /* die die die */
79 WORKER_IDLE = 1 << 2, /* is idle */
80 WORKER_PREP = 1 << 3, /* preparing to run works */
81 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
82 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
83 WORKER_REBOUND = 1 << 8, /* worker was rebound */
85 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
86 WORKER_UNBOUND | WORKER_REBOUND,
88 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
90 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
91 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
93 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
94 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
96 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
97 /* call for help after 10ms
99 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
100 CREATE_COOLDOWN = HZ, /* time to breath after fail */
103 * Rescue workers are used only on emergencies and shared by
104 * all cpus. Give MIN_NICE.
106 RESCUER_NICE_LEVEL = MIN_NICE,
107 HIGHPRI_NICE_LEVEL = MIN_NICE,
113 * Structure fields follow one of the following exclusion rules.
115 * I: Modifiable by initialization/destruction paths and read-only for
118 * P: Preemption protected. Disabling preemption is enough and should
119 * only be modified and accessed from the local cpu.
121 * L: pool->lock protected. Access with pool->lock held.
123 * X: During normal operation, modification requires pool->lock and should
124 * be done only from local cpu. Either disabling preemption on local
125 * cpu or grabbing pool->lock is enough for read access. If
126 * POOL_DISASSOCIATED is set, it's identical to L.
128 * A: wq_pool_attach_mutex protected.
130 * PL: wq_pool_mutex protected.
132 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
134 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
136 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
139 * WQ: wq->mutex protected.
141 * WR: wq->mutex protected for writes. RCU protected for reads.
143 * MD: wq_mayday_lock protected.
146 /* struct worker is defined in workqueue_internal.h */
149 raw_spinlock_t lock; /* the pool lock */
150 int cpu; /* I: the associated cpu */
151 int node; /* I: the associated node ID */
152 int id; /* I: pool ID */
153 unsigned int flags; /* X: flags */
155 unsigned long watchdog_ts; /* L: watchdog timestamp */
157 /* The current concurrency level. */
160 struct list_head worklist; /* L: list of pending works */
162 int nr_workers; /* L: total number of workers */
163 int nr_idle; /* L: currently idle workers */
165 struct list_head idle_list; /* X: list of idle workers */
166 struct timer_list idle_timer; /* L: worker idle timeout */
167 struct timer_list mayday_timer; /* L: SOS timer for workers */
169 /* a workers is either on busy_hash or idle_list, or the manager */
170 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
171 /* L: hash of busy workers */
173 struct worker *manager; /* L: purely informational */
174 struct list_head workers; /* A: attached workers */
175 struct completion *detach_completion; /* all workers detached */
177 struct ida worker_ida; /* worker IDs for task name */
179 struct workqueue_attrs *attrs; /* I: worker attributes */
180 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
181 int refcnt; /* PL: refcnt for unbound pools */
184 * Destruction of pool is RCU protected to allow dereferences
185 * from get_work_pool().
191 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
192 * of work_struct->data are used for flags and the remaining high bits
193 * point to the pwq; thus, pwqs need to be aligned at two's power of the
194 * number of flag bits.
196 struct pool_workqueue {
197 struct worker_pool *pool; /* I: the associated pool */
198 struct workqueue_struct *wq; /* I: the owning workqueue */
199 int work_color; /* L: current color */
200 int flush_color; /* L: flushing color */
201 int refcnt; /* L: reference count */
202 int nr_in_flight[WORK_NR_COLORS];
203 /* L: nr of in_flight works */
206 * nr_active management and WORK_STRUCT_INACTIVE:
208 * When pwq->nr_active >= max_active, new work item is queued to
209 * pwq->inactive_works instead of pool->worklist and marked with
210 * WORK_STRUCT_INACTIVE.
212 * All work items marked with WORK_STRUCT_INACTIVE do not participate
213 * in pwq->nr_active and all work items in pwq->inactive_works are
214 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
215 * work items are in pwq->inactive_works. Some of them are ready to
216 * run in pool->worklist or worker->scheduled. Those work itmes are
217 * only struct wq_barrier which is used for flush_work() and should
218 * not participate in pwq->nr_active. For non-barrier work item, it
219 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
221 int nr_active; /* L: nr of active works */
222 int max_active; /* L: max active works */
223 struct list_head inactive_works; /* L: inactive works */
224 struct list_head pwqs_node; /* WR: node on wq->pwqs */
225 struct list_head mayday_node; /* MD: node on wq->maydays */
228 * Release of unbound pwq is punted to system_wq. See put_pwq()
229 * and pwq_unbound_release_workfn() for details. pool_workqueue
230 * itself is also RCU protected so that the first pwq can be
231 * determined without grabbing wq->mutex.
233 struct work_struct unbound_release_work;
235 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
238 * Structure used to wait for workqueue flush.
241 struct list_head list; /* WQ: list of flushers */
242 int flush_color; /* WQ: flush color waiting for */
243 struct completion done; /* flush completion */
249 * The externally visible workqueue. It relays the issued work items to
250 * the appropriate worker_pool through its pool_workqueues.
252 struct workqueue_struct {
253 struct list_head pwqs; /* WR: all pwqs of this wq */
254 struct list_head list; /* PR: list of all workqueues */
256 struct mutex mutex; /* protects this wq */
257 int work_color; /* WQ: current work color */
258 int flush_color; /* WQ: current flush color */
259 atomic_t nr_pwqs_to_flush; /* flush in progress */
260 struct wq_flusher *first_flusher; /* WQ: first flusher */
261 struct list_head flusher_queue; /* WQ: flush waiters */
262 struct list_head flusher_overflow; /* WQ: flush overflow list */
264 struct list_head maydays; /* MD: pwqs requesting rescue */
265 struct worker *rescuer; /* MD: rescue worker */
267 int nr_drainers; /* WQ: drain in progress */
268 int saved_max_active; /* WQ: saved pwq max_active */
270 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
271 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
274 struct wq_device *wq_dev; /* I: for sysfs interface */
276 #ifdef CONFIG_LOCKDEP
278 struct lock_class_key key;
279 struct lockdep_map lockdep_map;
281 char name[WQ_NAME_LEN]; /* I: workqueue name */
284 * Destruction of workqueue_struct is RCU protected to allow walking
285 * the workqueues list without grabbing wq_pool_mutex.
286 * This is used to dump all workqueues from sysrq.
290 /* hot fields used during command issue, aligned to cacheline */
291 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
292 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
293 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
296 static struct kmem_cache *pwq_cache;
298 static cpumask_var_t *wq_numa_possible_cpumask;
299 /* possible CPUs of each node */
301 static bool wq_disable_numa;
302 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
304 /* see the comment above the definition of WQ_POWER_EFFICIENT */
305 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
306 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
308 static bool wq_online; /* can kworkers be created yet? */
310 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
312 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
313 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
315 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
316 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
317 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
318 /* wait for manager to go away */
319 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
321 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
322 static bool workqueue_freezing; /* PL: have wqs started freezing? */
324 /* PL: allowable cpus for unbound wqs and work items */
325 static cpumask_var_t wq_unbound_cpumask;
327 /* CPU where unbound work was last round robin scheduled from this CPU */
328 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
331 * Local execution of unbound work items is no longer guaranteed. The
332 * following always forces round-robin CPU selection on unbound work items
333 * to uncover usages which depend on it.
335 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
336 static bool wq_debug_force_rr_cpu = true;
338 static bool wq_debug_force_rr_cpu = false;
340 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
342 /* the per-cpu worker pools */
343 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
345 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
347 /* PL: hash of all unbound pools keyed by pool->attrs */
348 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
350 /* I: attributes used when instantiating standard unbound pools on demand */
351 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
353 /* I: attributes used when instantiating ordered pools on demand */
354 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
356 struct workqueue_struct *system_wq __read_mostly;
357 EXPORT_SYMBOL(system_wq);
358 struct workqueue_struct *system_highpri_wq __read_mostly;
359 EXPORT_SYMBOL_GPL(system_highpri_wq);
360 struct workqueue_struct *system_long_wq __read_mostly;
361 EXPORT_SYMBOL_GPL(system_long_wq);
362 struct workqueue_struct *system_unbound_wq __read_mostly;
363 EXPORT_SYMBOL_GPL(system_unbound_wq);
364 struct workqueue_struct *system_freezable_wq __read_mostly;
365 EXPORT_SYMBOL_GPL(system_freezable_wq);
366 struct workqueue_struct *system_power_efficient_wq __read_mostly;
367 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
368 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
369 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
371 static int worker_thread(void *__worker);
372 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
373 static void show_pwq(struct pool_workqueue *pwq);
374 static void show_one_worker_pool(struct worker_pool *pool);
376 #define CREATE_TRACE_POINTS
377 #include <trace/events/workqueue.h>
379 #define assert_rcu_or_pool_mutex() \
380 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
381 !lockdep_is_held(&wq_pool_mutex), \
382 "RCU or wq_pool_mutex should be held")
384 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
385 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
386 !lockdep_is_held(&wq->mutex) && \
387 !lockdep_is_held(&wq_pool_mutex), \
388 "RCU, wq->mutex or wq_pool_mutex should be held")
390 #define for_each_cpu_worker_pool(pool, cpu) \
391 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
392 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
396 * for_each_pool - iterate through all worker_pools in the system
397 * @pool: iteration cursor
398 * @pi: integer used for iteration
400 * This must be called either with wq_pool_mutex held or RCU read
401 * locked. If the pool needs to be used beyond the locking in effect, the
402 * caller is responsible for guaranteeing that the pool stays online.
404 * The if/else clause exists only for the lockdep assertion and can be
407 #define for_each_pool(pool, pi) \
408 idr_for_each_entry(&worker_pool_idr, pool, pi) \
409 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
413 * for_each_pool_worker - iterate through all workers of a worker_pool
414 * @worker: iteration cursor
415 * @pool: worker_pool to iterate workers of
417 * This must be called with wq_pool_attach_mutex.
419 * The if/else clause exists only for the lockdep assertion and can be
422 #define for_each_pool_worker(worker, pool) \
423 list_for_each_entry((worker), &(pool)->workers, node) \
424 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
428 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
429 * @pwq: iteration cursor
430 * @wq: the target workqueue
432 * This must be called either with wq->mutex held or RCU read locked.
433 * If the pwq needs to be used beyond the locking in effect, the caller is
434 * responsible for guaranteeing that the pwq stays online.
436 * The if/else clause exists only for the lockdep assertion and can be
439 #define for_each_pwq(pwq, wq) \
440 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
441 lockdep_is_held(&(wq->mutex)))
443 #ifdef CONFIG_DEBUG_OBJECTS_WORK
445 static const struct debug_obj_descr work_debug_descr;
447 static void *work_debug_hint(void *addr)
449 return ((struct work_struct *) addr)->func;
452 static bool work_is_static_object(void *addr)
454 struct work_struct *work = addr;
456 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
460 * fixup_init is called when:
461 * - an active object is initialized
463 static bool work_fixup_init(void *addr, enum debug_obj_state state)
465 struct work_struct *work = addr;
468 case ODEBUG_STATE_ACTIVE:
469 cancel_work_sync(work);
470 debug_object_init(work, &work_debug_descr);
478 * fixup_free is called when:
479 * - an active object is freed
481 static bool work_fixup_free(void *addr, enum debug_obj_state state)
483 struct work_struct *work = addr;
486 case ODEBUG_STATE_ACTIVE:
487 cancel_work_sync(work);
488 debug_object_free(work, &work_debug_descr);
495 static const struct debug_obj_descr work_debug_descr = {
496 .name = "work_struct",
497 .debug_hint = work_debug_hint,
498 .is_static_object = work_is_static_object,
499 .fixup_init = work_fixup_init,
500 .fixup_free = work_fixup_free,
503 static inline void debug_work_activate(struct work_struct *work)
505 debug_object_activate(work, &work_debug_descr);
508 static inline void debug_work_deactivate(struct work_struct *work)
510 debug_object_deactivate(work, &work_debug_descr);
513 void __init_work(struct work_struct *work, int onstack)
516 debug_object_init_on_stack(work, &work_debug_descr);
518 debug_object_init(work, &work_debug_descr);
520 EXPORT_SYMBOL_GPL(__init_work);
522 void destroy_work_on_stack(struct work_struct *work)
524 debug_object_free(work, &work_debug_descr);
526 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
528 void destroy_delayed_work_on_stack(struct delayed_work *work)
530 destroy_timer_on_stack(&work->timer);
531 debug_object_free(&work->work, &work_debug_descr);
533 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
536 static inline void debug_work_activate(struct work_struct *work) { }
537 static inline void debug_work_deactivate(struct work_struct *work) { }
541 * worker_pool_assign_id - allocate ID and assign it to @pool
542 * @pool: the pool pointer of interest
544 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
545 * successfully, -errno on failure.
547 static int worker_pool_assign_id(struct worker_pool *pool)
551 lockdep_assert_held(&wq_pool_mutex);
553 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
563 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
564 * @wq: the target workqueue
567 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
569 * If the pwq needs to be used beyond the locking in effect, the caller is
570 * responsible for guaranteeing that the pwq stays online.
572 * Return: The unbound pool_workqueue for @node.
574 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
577 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
580 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
581 * delayed item is pending. The plan is to keep CPU -> NODE
582 * mapping valid and stable across CPU on/offlines. Once that
583 * happens, this workaround can be removed.
585 if (unlikely(node == NUMA_NO_NODE))
588 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
591 static unsigned int work_color_to_flags(int color)
593 return color << WORK_STRUCT_COLOR_SHIFT;
596 static int get_work_color(unsigned long work_data)
598 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
599 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
602 static int work_next_color(int color)
604 return (color + 1) % WORK_NR_COLORS;
608 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
609 * contain the pointer to the queued pwq. Once execution starts, the flag
610 * is cleared and the high bits contain OFFQ flags and pool ID.
612 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
613 * and clear_work_data() can be used to set the pwq, pool or clear
614 * work->data. These functions should only be called while the work is
615 * owned - ie. while the PENDING bit is set.
617 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
618 * corresponding to a work. Pool is available once the work has been
619 * queued anywhere after initialization until it is sync canceled. pwq is
620 * available only while the work item is queued.
622 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
623 * canceled. While being canceled, a work item may have its PENDING set
624 * but stay off timer and worklist for arbitrarily long and nobody should
625 * try to steal the PENDING bit.
627 static inline void set_work_data(struct work_struct *work, unsigned long data,
630 WARN_ON_ONCE(!work_pending(work));
631 atomic_long_set(&work->data, data | flags | work_static(work));
634 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
635 unsigned long extra_flags)
637 set_work_data(work, (unsigned long)pwq,
638 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
641 static void set_work_pool_and_keep_pending(struct work_struct *work,
644 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
645 WORK_STRUCT_PENDING);
648 static void set_work_pool_and_clear_pending(struct work_struct *work,
652 * The following wmb is paired with the implied mb in
653 * test_and_set_bit(PENDING) and ensures all updates to @work made
654 * here are visible to and precede any updates by the next PENDING
658 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
660 * The following mb guarantees that previous clear of a PENDING bit
661 * will not be reordered with any speculative LOADS or STORES from
662 * work->current_func, which is executed afterwards. This possible
663 * reordering can lead to a missed execution on attempt to queue
664 * the same @work. E.g. consider this case:
667 * ---------------------------- --------------------------------
669 * 1 STORE event_indicated
670 * 2 queue_work_on() {
671 * 3 test_and_set_bit(PENDING)
672 * 4 } set_..._and_clear_pending() {
673 * 5 set_work_data() # clear bit
675 * 7 work->current_func() {
676 * 8 LOAD event_indicated
679 * Without an explicit full barrier speculative LOAD on line 8 can
680 * be executed before CPU#0 does STORE on line 1. If that happens,
681 * CPU#0 observes the PENDING bit is still set and new execution of
682 * a @work is not queued in a hope, that CPU#1 will eventually
683 * finish the queued @work. Meanwhile CPU#1 does not see
684 * event_indicated is set, because speculative LOAD was executed
685 * before actual STORE.
690 static void clear_work_data(struct work_struct *work)
692 smp_wmb(); /* see set_work_pool_and_clear_pending() */
693 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
696 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
698 unsigned long data = atomic_long_read(&work->data);
700 if (data & WORK_STRUCT_PWQ)
701 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
707 * get_work_pool - return the worker_pool a given work was associated with
708 * @work: the work item of interest
710 * Pools are created and destroyed under wq_pool_mutex, and allows read
711 * access under RCU read lock. As such, this function should be
712 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
714 * All fields of the returned pool are accessible as long as the above
715 * mentioned locking is in effect. If the returned pool needs to be used
716 * beyond the critical section, the caller is responsible for ensuring the
717 * returned pool is and stays online.
719 * Return: The worker_pool @work was last associated with. %NULL if none.
721 static struct worker_pool *get_work_pool(struct work_struct *work)
723 unsigned long data = atomic_long_read(&work->data);
726 assert_rcu_or_pool_mutex();
728 if (data & WORK_STRUCT_PWQ)
729 return ((struct pool_workqueue *)
730 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
732 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
733 if (pool_id == WORK_OFFQ_POOL_NONE)
736 return idr_find(&worker_pool_idr, pool_id);
740 * get_work_pool_id - return the worker pool ID a given work is associated with
741 * @work: the work item of interest
743 * Return: The worker_pool ID @work was last associated with.
744 * %WORK_OFFQ_POOL_NONE if none.
746 static int get_work_pool_id(struct work_struct *work)
748 unsigned long data = atomic_long_read(&work->data);
750 if (data & WORK_STRUCT_PWQ)
751 return ((struct pool_workqueue *)
752 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
754 return data >> WORK_OFFQ_POOL_SHIFT;
757 static void mark_work_canceling(struct work_struct *work)
759 unsigned long pool_id = get_work_pool_id(work);
761 pool_id <<= WORK_OFFQ_POOL_SHIFT;
762 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
765 static bool work_is_canceling(struct work_struct *work)
767 unsigned long data = atomic_long_read(&work->data);
769 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
773 * Policy functions. These define the policies on how the global worker
774 * pools are managed. Unless noted otherwise, these functions assume that
775 * they're being called with pool->lock held.
778 static bool __need_more_worker(struct worker_pool *pool)
780 return !atomic_read(&pool->nr_running);
784 * Need to wake up a worker? Called from anything but currently
787 * Note that, because unbound workers never contribute to nr_running, this
788 * function will always return %true for unbound pools as long as the
789 * worklist isn't empty.
791 static bool need_more_worker(struct worker_pool *pool)
793 return !list_empty(&pool->worklist) && __need_more_worker(pool);
796 /* Can I start working? Called from busy but !running workers. */
797 static bool may_start_working(struct worker_pool *pool)
799 return pool->nr_idle;
802 /* Do I need to keep working? Called from currently running workers. */
803 static bool keep_working(struct worker_pool *pool)
805 return !list_empty(&pool->worklist) &&
806 atomic_read(&pool->nr_running) <= 1;
809 /* Do we need a new worker? Called from manager. */
810 static bool need_to_create_worker(struct worker_pool *pool)
812 return need_more_worker(pool) && !may_start_working(pool);
815 /* Do we have too many workers and should some go away? */
816 static bool too_many_workers(struct worker_pool *pool)
818 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
819 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
820 int nr_busy = pool->nr_workers - nr_idle;
822 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
829 /* Return the first idle worker. Safe with preemption disabled */
830 static struct worker *first_idle_worker(struct worker_pool *pool)
832 if (unlikely(list_empty(&pool->idle_list)))
835 return list_first_entry(&pool->idle_list, struct worker, entry);
839 * wake_up_worker - wake up an idle worker
840 * @pool: worker pool to wake worker from
842 * Wake up the first idle worker of @pool.
845 * raw_spin_lock_irq(pool->lock).
847 static void wake_up_worker(struct worker_pool *pool)
849 struct worker *worker = first_idle_worker(pool);
852 wake_up_process(worker->task);
856 * wq_worker_running - a worker is running again
857 * @task: task waking up
859 * This function is called when a worker returns from schedule()
861 void wq_worker_running(struct task_struct *task)
863 struct worker *worker = kthread_data(task);
865 if (!worker->sleeping)
869 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
870 * and the nr_running increment below, we may ruin the nr_running reset
871 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
872 * pool. Protect against such race.
875 if (!(worker->flags & WORKER_NOT_RUNNING))
876 atomic_inc(&worker->pool->nr_running);
878 worker->sleeping = 0;
882 * wq_worker_sleeping - a worker is going to sleep
883 * @task: task going to sleep
885 * This function is called from schedule() when a busy worker is
888 void wq_worker_sleeping(struct task_struct *task)
890 struct worker *next, *worker = kthread_data(task);
891 struct worker_pool *pool;
894 * Rescuers, which may not have all the fields set up like normal
895 * workers, also reach here, let's not access anything before
896 * checking NOT_RUNNING.
898 if (worker->flags & WORKER_NOT_RUNNING)
903 /* Return if preempted before wq_worker_running() was reached */
904 if (worker->sleeping)
907 worker->sleeping = 1;
908 raw_spin_lock_irq(&pool->lock);
911 * Recheck in case unbind_workers() preempted us. We don't
912 * want to decrement nr_running after the worker is unbound
913 * and nr_running has been reset.
915 if (worker->flags & WORKER_NOT_RUNNING) {
916 raw_spin_unlock_irq(&pool->lock);
921 * The counterpart of the following dec_and_test, implied mb,
922 * worklist not empty test sequence is in insert_work().
923 * Please read comment there.
925 * NOT_RUNNING is clear. This means that we're bound to and
926 * running on the local cpu w/ rq lock held and preemption
927 * disabled, which in turn means that none else could be
928 * manipulating idle_list, so dereferencing idle_list without pool
931 if (atomic_dec_and_test(&pool->nr_running) &&
932 !list_empty(&pool->worklist)) {
933 next = first_idle_worker(pool);
935 wake_up_process(next->task);
937 raw_spin_unlock_irq(&pool->lock);
941 * wq_worker_last_func - retrieve worker's last work function
942 * @task: Task to retrieve last work function of.
944 * Determine the last function a worker executed. This is called from
945 * the scheduler to get a worker's last known identity.
948 * raw_spin_lock_irq(rq->lock)
950 * This function is called during schedule() when a kworker is going
951 * to sleep. It's used by psi to identify aggregation workers during
952 * dequeuing, to allow periodic aggregation to shut-off when that
953 * worker is the last task in the system or cgroup to go to sleep.
955 * As this function doesn't involve any workqueue-related locking, it
956 * only returns stable values when called from inside the scheduler's
957 * queuing and dequeuing paths, when @task, which must be a kworker,
958 * is guaranteed to not be processing any works.
961 * The last work function %current executed as a worker, NULL if it
962 * hasn't executed any work yet.
964 work_func_t wq_worker_last_func(struct task_struct *task)
966 struct worker *worker = kthread_data(task);
968 return worker->last_func;
972 * worker_set_flags - set worker flags and adjust nr_running accordingly
974 * @flags: flags to set
976 * Set @flags in @worker->flags and adjust nr_running accordingly.
979 * raw_spin_lock_irq(pool->lock)
981 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
983 struct worker_pool *pool = worker->pool;
985 WARN_ON_ONCE(worker->task != current);
987 /* If transitioning into NOT_RUNNING, adjust nr_running. */
988 if ((flags & WORKER_NOT_RUNNING) &&
989 !(worker->flags & WORKER_NOT_RUNNING)) {
990 atomic_dec(&pool->nr_running);
993 worker->flags |= flags;
997 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
999 * @flags: flags to clear
1001 * Clear @flags in @worker->flags and adjust nr_running accordingly.
1004 * raw_spin_lock_irq(pool->lock)
1006 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
1008 struct worker_pool *pool = worker->pool;
1009 unsigned int oflags = worker->flags;
1011 WARN_ON_ONCE(worker->task != current);
1013 worker->flags &= ~flags;
1016 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1017 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1018 * of multiple flags, not a single flag.
1020 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1021 if (!(worker->flags & WORKER_NOT_RUNNING))
1022 atomic_inc(&pool->nr_running);
1026 * find_worker_executing_work - find worker which is executing a work
1027 * @pool: pool of interest
1028 * @work: work to find worker for
1030 * Find a worker which is executing @work on @pool by searching
1031 * @pool->busy_hash which is keyed by the address of @work. For a worker
1032 * to match, its current execution should match the address of @work and
1033 * its work function. This is to avoid unwanted dependency between
1034 * unrelated work executions through a work item being recycled while still
1037 * This is a bit tricky. A work item may be freed once its execution
1038 * starts and nothing prevents the freed area from being recycled for
1039 * another work item. If the same work item address ends up being reused
1040 * before the original execution finishes, workqueue will identify the
1041 * recycled work item as currently executing and make it wait until the
1042 * current execution finishes, introducing an unwanted dependency.
1044 * This function checks the work item address and work function to avoid
1045 * false positives. Note that this isn't complete as one may construct a
1046 * work function which can introduce dependency onto itself through a
1047 * recycled work item. Well, if somebody wants to shoot oneself in the
1048 * foot that badly, there's only so much we can do, and if such deadlock
1049 * actually occurs, it should be easy to locate the culprit work function.
1052 * raw_spin_lock_irq(pool->lock).
1055 * Pointer to worker which is executing @work if found, %NULL
1058 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1059 struct work_struct *work)
1061 struct worker *worker;
1063 hash_for_each_possible(pool->busy_hash, worker, hentry,
1064 (unsigned long)work)
1065 if (worker->current_work == work &&
1066 worker->current_func == work->func)
1073 * move_linked_works - move linked works to a list
1074 * @work: start of series of works to be scheduled
1075 * @head: target list to append @work to
1076 * @nextp: out parameter for nested worklist walking
1078 * Schedule linked works starting from @work to @head. Work series to
1079 * be scheduled starts at @work and includes any consecutive work with
1080 * WORK_STRUCT_LINKED set in its predecessor.
1082 * If @nextp is not NULL, it's updated to point to the next work of
1083 * the last scheduled work. This allows move_linked_works() to be
1084 * nested inside outer list_for_each_entry_safe().
1087 * raw_spin_lock_irq(pool->lock).
1089 static void move_linked_works(struct work_struct *work, struct list_head *head,
1090 struct work_struct **nextp)
1092 struct work_struct *n;
1095 * Linked worklist will always end before the end of the list,
1096 * use NULL for list head.
1098 list_for_each_entry_safe_from(work, n, NULL, entry) {
1099 list_move_tail(&work->entry, head);
1100 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1105 * If we're already inside safe list traversal and have moved
1106 * multiple works to the scheduled queue, the next position
1107 * needs to be updated.
1114 * get_pwq - get an extra reference on the specified pool_workqueue
1115 * @pwq: pool_workqueue to get
1117 * Obtain an extra reference on @pwq. The caller should guarantee that
1118 * @pwq has positive refcnt and be holding the matching pool->lock.
1120 static void get_pwq(struct pool_workqueue *pwq)
1122 lockdep_assert_held(&pwq->pool->lock);
1123 WARN_ON_ONCE(pwq->refcnt <= 0);
1128 * put_pwq - put a pool_workqueue reference
1129 * @pwq: pool_workqueue to put
1131 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1132 * destruction. The caller should be holding the matching pool->lock.
1134 static void put_pwq(struct pool_workqueue *pwq)
1136 lockdep_assert_held(&pwq->pool->lock);
1137 if (likely(--pwq->refcnt))
1139 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1142 * @pwq can't be released under pool->lock, bounce to
1143 * pwq_unbound_release_workfn(). This never recurses on the same
1144 * pool->lock as this path is taken only for unbound workqueues and
1145 * the release work item is scheduled on a per-cpu workqueue. To
1146 * avoid lockdep warning, unbound pool->locks are given lockdep
1147 * subclass of 1 in get_unbound_pool().
1149 schedule_work(&pwq->unbound_release_work);
1153 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1154 * @pwq: pool_workqueue to put (can be %NULL)
1156 * put_pwq() with locking. This function also allows %NULL @pwq.
1158 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1162 * As both pwqs and pools are RCU protected, the
1163 * following lock operations are safe.
1165 raw_spin_lock_irq(&pwq->pool->lock);
1167 raw_spin_unlock_irq(&pwq->pool->lock);
1171 static void pwq_activate_inactive_work(struct work_struct *work)
1173 struct pool_workqueue *pwq = get_work_pwq(work);
1175 trace_workqueue_activate_work(work);
1176 if (list_empty(&pwq->pool->worklist))
1177 pwq->pool->watchdog_ts = jiffies;
1178 move_linked_works(work, &pwq->pool->worklist, NULL);
1179 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1183 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1185 struct work_struct *work = list_first_entry(&pwq->inactive_works,
1186 struct work_struct, entry);
1188 pwq_activate_inactive_work(work);
1192 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1193 * @pwq: pwq of interest
1194 * @work_data: work_data of work which left the queue
1196 * A work either has completed or is removed from pending queue,
1197 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1200 * raw_spin_lock_irq(pool->lock).
1202 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1204 int color = get_work_color(work_data);
1206 if (!(work_data & WORK_STRUCT_INACTIVE)) {
1208 if (!list_empty(&pwq->inactive_works)) {
1209 /* one down, submit an inactive one */
1210 if (pwq->nr_active < pwq->max_active)
1211 pwq_activate_first_inactive(pwq);
1215 pwq->nr_in_flight[color]--;
1217 /* is flush in progress and are we at the flushing tip? */
1218 if (likely(pwq->flush_color != color))
1221 /* are there still in-flight works? */
1222 if (pwq->nr_in_flight[color])
1225 /* this pwq is done, clear flush_color */
1226 pwq->flush_color = -1;
1229 * If this was the last pwq, wake up the first flusher. It
1230 * will handle the rest.
1232 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1233 complete(&pwq->wq->first_flusher->done);
1239 * try_to_grab_pending - steal work item from worklist and disable irq
1240 * @work: work item to steal
1241 * @is_dwork: @work is a delayed_work
1242 * @flags: place to store irq state
1244 * Try to grab PENDING bit of @work. This function can handle @work in any
1245 * stable state - idle, on timer or on worklist.
1249 * ======== ================================================================
1250 * 1 if @work was pending and we successfully stole PENDING
1251 * 0 if @work was idle and we claimed PENDING
1252 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1253 * -ENOENT if someone else is canceling @work, this state may persist
1254 * for arbitrarily long
1255 * ======== ================================================================
1258 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1259 * interrupted while holding PENDING and @work off queue, irq must be
1260 * disabled on entry. This, combined with delayed_work->timer being
1261 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1263 * On successful return, >= 0, irq is disabled and the caller is
1264 * responsible for releasing it using local_irq_restore(*@flags).
1266 * This function is safe to call from any context including IRQ handler.
1268 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1269 unsigned long *flags)
1271 struct worker_pool *pool;
1272 struct pool_workqueue *pwq;
1274 local_irq_save(*flags);
1276 /* try to steal the timer if it exists */
1278 struct delayed_work *dwork = to_delayed_work(work);
1281 * dwork->timer is irqsafe. If del_timer() fails, it's
1282 * guaranteed that the timer is not queued anywhere and not
1283 * running on the local CPU.
1285 if (likely(del_timer(&dwork->timer)))
1289 /* try to claim PENDING the normal way */
1290 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1295 * The queueing is in progress, or it is already queued. Try to
1296 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1298 pool = get_work_pool(work);
1302 raw_spin_lock(&pool->lock);
1304 * work->data is guaranteed to point to pwq only while the work
1305 * item is queued on pwq->wq, and both updating work->data to point
1306 * to pwq on queueing and to pool on dequeueing are done under
1307 * pwq->pool->lock. This in turn guarantees that, if work->data
1308 * points to pwq which is associated with a locked pool, the work
1309 * item is currently queued on that pool.
1311 pwq = get_work_pwq(work);
1312 if (pwq && pwq->pool == pool) {
1313 debug_work_deactivate(work);
1316 * A cancelable inactive work item must be in the
1317 * pwq->inactive_works since a queued barrier can't be
1318 * canceled (see the comments in insert_wq_barrier()).
1320 * An inactive work item cannot be grabbed directly because
1321 * it might have linked barrier work items which, if left
1322 * on the inactive_works list, will confuse pwq->nr_active
1323 * management later on and cause stall. Make sure the work
1324 * item is activated before grabbing.
1326 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1327 pwq_activate_inactive_work(work);
1329 list_del_init(&work->entry);
1330 pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1332 /* work->data points to pwq iff queued, point to pool */
1333 set_work_pool_and_keep_pending(work, pool->id);
1335 raw_spin_unlock(&pool->lock);
1339 raw_spin_unlock(&pool->lock);
1342 local_irq_restore(*flags);
1343 if (work_is_canceling(work))
1350 * insert_work - insert a work into a pool
1351 * @pwq: pwq @work belongs to
1352 * @work: work to insert
1353 * @head: insertion point
1354 * @extra_flags: extra WORK_STRUCT_* flags to set
1356 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1357 * work_struct flags.
1360 * raw_spin_lock_irq(pool->lock).
1362 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1363 struct list_head *head, unsigned int extra_flags)
1365 struct worker_pool *pool = pwq->pool;
1367 /* record the work call stack in order to print it in KASAN reports */
1368 kasan_record_aux_stack_noalloc(work);
1370 /* we own @work, set data and link */
1371 set_work_pwq(work, pwq, extra_flags);
1372 list_add_tail(&work->entry, head);
1376 * Ensure either wq_worker_sleeping() sees the above
1377 * list_add_tail() or we see zero nr_running to avoid workers lying
1378 * around lazily while there are works to be processed.
1382 if (__need_more_worker(pool))
1383 wake_up_worker(pool);
1387 * Test whether @work is being queued from another work executing on the
1390 static bool is_chained_work(struct workqueue_struct *wq)
1392 struct worker *worker;
1394 worker = current_wq_worker();
1396 * Return %true iff I'm a worker executing a work item on @wq. If
1397 * I'm @worker, it's safe to dereference it without locking.
1399 return worker && worker->current_pwq->wq == wq;
1403 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1404 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1405 * avoid perturbing sensitive tasks.
1407 static int wq_select_unbound_cpu(int cpu)
1409 static bool printed_dbg_warning;
1412 if (likely(!wq_debug_force_rr_cpu)) {
1413 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1415 } else if (!printed_dbg_warning) {
1416 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1417 printed_dbg_warning = true;
1420 if (cpumask_empty(wq_unbound_cpumask))
1423 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1424 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1425 if (unlikely(new_cpu >= nr_cpu_ids)) {
1426 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1427 if (unlikely(new_cpu >= nr_cpu_ids))
1430 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1435 static void __queue_work(int cpu, struct workqueue_struct *wq,
1436 struct work_struct *work)
1438 struct pool_workqueue *pwq;
1439 struct worker_pool *last_pool;
1440 struct list_head *worklist;
1441 unsigned int work_flags;
1442 unsigned int req_cpu = cpu;
1445 * While a work item is PENDING && off queue, a task trying to
1446 * steal the PENDING will busy-loop waiting for it to either get
1447 * queued or lose PENDING. Grabbing PENDING and queueing should
1448 * happen with IRQ disabled.
1450 lockdep_assert_irqs_disabled();
1453 /* if draining, only works from the same workqueue are allowed */
1454 if (unlikely(wq->flags & __WQ_DRAINING) &&
1455 WARN_ON_ONCE(!is_chained_work(wq)))
1459 /* pwq which will be used unless @work is executing elsewhere */
1460 if (wq->flags & WQ_UNBOUND) {
1461 if (req_cpu == WORK_CPU_UNBOUND)
1462 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1463 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1465 if (req_cpu == WORK_CPU_UNBOUND)
1466 cpu = raw_smp_processor_id();
1467 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1471 * If @work was previously on a different pool, it might still be
1472 * running there, in which case the work needs to be queued on that
1473 * pool to guarantee non-reentrancy.
1475 last_pool = get_work_pool(work);
1476 if (last_pool && last_pool != pwq->pool) {
1477 struct worker *worker;
1479 raw_spin_lock(&last_pool->lock);
1481 worker = find_worker_executing_work(last_pool, work);
1483 if (worker && worker->current_pwq->wq == wq) {
1484 pwq = worker->current_pwq;
1486 /* meh... not running there, queue here */
1487 raw_spin_unlock(&last_pool->lock);
1488 raw_spin_lock(&pwq->pool->lock);
1491 raw_spin_lock(&pwq->pool->lock);
1495 * pwq is determined and locked. For unbound pools, we could have
1496 * raced with pwq release and it could already be dead. If its
1497 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1498 * without another pwq replacing it in the numa_pwq_tbl or while
1499 * work items are executing on it, so the retrying is guaranteed to
1500 * make forward-progress.
1502 if (unlikely(!pwq->refcnt)) {
1503 if (wq->flags & WQ_UNBOUND) {
1504 raw_spin_unlock(&pwq->pool->lock);
1509 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1513 /* pwq determined, queue */
1514 trace_workqueue_queue_work(req_cpu, pwq, work);
1516 if (WARN_ON(!list_empty(&work->entry)))
1519 pwq->nr_in_flight[pwq->work_color]++;
1520 work_flags = work_color_to_flags(pwq->work_color);
1522 if (likely(pwq->nr_active < pwq->max_active)) {
1523 trace_workqueue_activate_work(work);
1525 worklist = &pwq->pool->worklist;
1526 if (list_empty(worklist))
1527 pwq->pool->watchdog_ts = jiffies;
1529 work_flags |= WORK_STRUCT_INACTIVE;
1530 worklist = &pwq->inactive_works;
1533 debug_work_activate(work);
1534 insert_work(pwq, work, worklist, work_flags);
1537 raw_spin_unlock(&pwq->pool->lock);
1542 * queue_work_on - queue work on specific cpu
1543 * @cpu: CPU number to execute work on
1544 * @wq: workqueue to use
1545 * @work: work to queue
1547 * We queue the work to a specific CPU, the caller must ensure it
1548 * can't go away. Callers that fail to ensure that the specified
1549 * CPU cannot go away will execute on a randomly chosen CPU.
1551 * Return: %false if @work was already on a queue, %true otherwise.
1553 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1554 struct work_struct *work)
1557 unsigned long flags;
1559 local_irq_save(flags);
1561 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1562 __queue_work(cpu, wq, work);
1566 local_irq_restore(flags);
1569 EXPORT_SYMBOL(queue_work_on);
1572 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1573 * @node: NUMA node ID that we want to select a CPU from
1575 * This function will attempt to find a "random" cpu available on a given
1576 * node. If there are no CPUs available on the given node it will return
1577 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1578 * available CPU if we need to schedule this work.
1580 static int workqueue_select_cpu_near(int node)
1584 /* No point in doing this if NUMA isn't enabled for workqueues */
1585 if (!wq_numa_enabled)
1586 return WORK_CPU_UNBOUND;
1588 /* Delay binding to CPU if node is not valid or online */
1589 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1590 return WORK_CPU_UNBOUND;
1592 /* Use local node/cpu if we are already there */
1593 cpu = raw_smp_processor_id();
1594 if (node == cpu_to_node(cpu))
1597 /* Use "random" otherwise know as "first" online CPU of node */
1598 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1600 /* If CPU is valid return that, otherwise just defer */
1601 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1605 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1606 * @node: NUMA node that we are targeting the work for
1607 * @wq: workqueue to use
1608 * @work: work to queue
1610 * We queue the work to a "random" CPU within a given NUMA node. The basic
1611 * idea here is to provide a way to somehow associate work with a given
1614 * This function will only make a best effort attempt at getting this onto
1615 * the right NUMA node. If no node is requested or the requested node is
1616 * offline then we just fall back to standard queue_work behavior.
1618 * Currently the "random" CPU ends up being the first available CPU in the
1619 * intersection of cpu_online_mask and the cpumask of the node, unless we
1620 * are running on the node. In that case we just use the current CPU.
1622 * Return: %false if @work was already on a queue, %true otherwise.
1624 bool queue_work_node(int node, struct workqueue_struct *wq,
1625 struct work_struct *work)
1627 unsigned long flags;
1631 * This current implementation is specific to unbound workqueues.
1632 * Specifically we only return the first available CPU for a given
1633 * node instead of cycling through individual CPUs within the node.
1635 * If this is used with a per-cpu workqueue then the logic in
1636 * workqueue_select_cpu_near would need to be updated to allow for
1637 * some round robin type logic.
1639 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1641 local_irq_save(flags);
1643 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1644 int cpu = workqueue_select_cpu_near(node);
1646 __queue_work(cpu, wq, work);
1650 local_irq_restore(flags);
1653 EXPORT_SYMBOL_GPL(queue_work_node);
1655 void delayed_work_timer_fn(struct timer_list *t)
1657 struct delayed_work *dwork = from_timer(dwork, t, timer);
1659 /* should have been called from irqsafe timer with irq already off */
1660 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1662 EXPORT_SYMBOL(delayed_work_timer_fn);
1664 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1665 struct delayed_work *dwork, unsigned long delay)
1667 struct timer_list *timer = &dwork->timer;
1668 struct work_struct *work = &dwork->work;
1671 WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn);
1672 WARN_ON_ONCE(timer_pending(timer));
1673 WARN_ON_ONCE(!list_empty(&work->entry));
1676 * If @delay is 0, queue @dwork->work immediately. This is for
1677 * both optimization and correctness. The earliest @timer can
1678 * expire is on the closest next tick and delayed_work users depend
1679 * on that there's no such delay when @delay is 0.
1682 __queue_work(cpu, wq, &dwork->work);
1688 timer->expires = jiffies + delay;
1690 if (unlikely(cpu != WORK_CPU_UNBOUND))
1691 add_timer_on(timer, cpu);
1697 * queue_delayed_work_on - queue work on specific CPU after delay
1698 * @cpu: CPU number to execute work on
1699 * @wq: workqueue to use
1700 * @dwork: work to queue
1701 * @delay: number of jiffies to wait before queueing
1703 * Return: %false if @work was already on a queue, %true otherwise. If
1704 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1707 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1708 struct delayed_work *dwork, unsigned long delay)
1710 struct work_struct *work = &dwork->work;
1712 unsigned long flags;
1714 /* read the comment in __queue_work() */
1715 local_irq_save(flags);
1717 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1718 __queue_delayed_work(cpu, wq, dwork, delay);
1722 local_irq_restore(flags);
1725 EXPORT_SYMBOL(queue_delayed_work_on);
1728 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1729 * @cpu: CPU number to execute work on
1730 * @wq: workqueue to use
1731 * @dwork: work to queue
1732 * @delay: number of jiffies to wait before queueing
1734 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1735 * modify @dwork's timer so that it expires after @delay. If @delay is
1736 * zero, @work is guaranteed to be scheduled immediately regardless of its
1739 * Return: %false if @dwork was idle and queued, %true if @dwork was
1740 * pending and its timer was modified.
1742 * This function is safe to call from any context including IRQ handler.
1743 * See try_to_grab_pending() for details.
1745 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1746 struct delayed_work *dwork, unsigned long delay)
1748 unsigned long flags;
1752 ret = try_to_grab_pending(&dwork->work, true, &flags);
1753 } while (unlikely(ret == -EAGAIN));
1755 if (likely(ret >= 0)) {
1756 __queue_delayed_work(cpu, wq, dwork, delay);
1757 local_irq_restore(flags);
1760 /* -ENOENT from try_to_grab_pending() becomes %true */
1763 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1765 static void rcu_work_rcufn(struct rcu_head *rcu)
1767 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1769 /* read the comment in __queue_work() */
1770 local_irq_disable();
1771 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1776 * queue_rcu_work - queue work after a RCU grace period
1777 * @wq: workqueue to use
1778 * @rwork: work to queue
1780 * Return: %false if @rwork was already pending, %true otherwise. Note
1781 * that a full RCU grace period is guaranteed only after a %true return.
1782 * While @rwork is guaranteed to be executed after a %false return, the
1783 * execution may happen before a full RCU grace period has passed.
1785 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1787 struct work_struct *work = &rwork->work;
1789 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1791 call_rcu(&rwork->rcu, rcu_work_rcufn);
1797 EXPORT_SYMBOL(queue_rcu_work);
1800 * worker_enter_idle - enter idle state
1801 * @worker: worker which is entering idle state
1803 * @worker is entering idle state. Update stats and idle timer if
1807 * raw_spin_lock_irq(pool->lock).
1809 static void worker_enter_idle(struct worker *worker)
1811 struct worker_pool *pool = worker->pool;
1813 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1814 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1815 (worker->hentry.next || worker->hentry.pprev)))
1818 /* can't use worker_set_flags(), also called from create_worker() */
1819 worker->flags |= WORKER_IDLE;
1821 worker->last_active = jiffies;
1823 /* idle_list is LIFO */
1824 list_add(&worker->entry, &pool->idle_list);
1826 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1827 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1829 /* Sanity check nr_running. */
1830 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle &&
1831 atomic_read(&pool->nr_running));
1835 * worker_leave_idle - leave idle state
1836 * @worker: worker which is leaving idle state
1838 * @worker is leaving idle state. Update stats.
1841 * raw_spin_lock_irq(pool->lock).
1843 static void worker_leave_idle(struct worker *worker)
1845 struct worker_pool *pool = worker->pool;
1847 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1849 worker_clr_flags(worker, WORKER_IDLE);
1851 list_del_init(&worker->entry);
1854 static struct worker *alloc_worker(int node)
1856 struct worker *worker;
1858 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1860 INIT_LIST_HEAD(&worker->entry);
1861 INIT_LIST_HEAD(&worker->scheduled);
1862 INIT_LIST_HEAD(&worker->node);
1863 /* on creation a worker is in !idle && prep state */
1864 worker->flags = WORKER_PREP;
1870 * worker_attach_to_pool() - attach a worker to a pool
1871 * @worker: worker to be attached
1872 * @pool: the target pool
1874 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1875 * cpu-binding of @worker are kept coordinated with the pool across
1878 static void worker_attach_to_pool(struct worker *worker,
1879 struct worker_pool *pool)
1881 mutex_lock(&wq_pool_attach_mutex);
1884 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1885 * stable across this function. See the comments above the flag
1886 * definition for details.
1888 if (pool->flags & POOL_DISASSOCIATED)
1889 worker->flags |= WORKER_UNBOUND;
1891 kthread_set_per_cpu(worker->task, pool->cpu);
1893 if (worker->rescue_wq)
1894 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1896 list_add_tail(&worker->node, &pool->workers);
1897 worker->pool = pool;
1899 mutex_unlock(&wq_pool_attach_mutex);
1903 * worker_detach_from_pool() - detach a worker from its pool
1904 * @worker: worker which is attached to its pool
1906 * Undo the attaching which had been done in worker_attach_to_pool(). The
1907 * caller worker shouldn't access to the pool after detached except it has
1908 * other reference to the pool.
1910 static void worker_detach_from_pool(struct worker *worker)
1912 struct worker_pool *pool = worker->pool;
1913 struct completion *detach_completion = NULL;
1915 mutex_lock(&wq_pool_attach_mutex);
1917 kthread_set_per_cpu(worker->task, -1);
1918 list_del(&worker->node);
1919 worker->pool = NULL;
1921 if (list_empty(&pool->workers))
1922 detach_completion = pool->detach_completion;
1923 mutex_unlock(&wq_pool_attach_mutex);
1925 /* clear leftover flags without pool->lock after it is detached */
1926 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1928 if (detach_completion)
1929 complete(detach_completion);
1933 * create_worker - create a new workqueue worker
1934 * @pool: pool the new worker will belong to
1936 * Create and start a new worker which is attached to @pool.
1939 * Might sleep. Does GFP_KERNEL allocations.
1942 * Pointer to the newly created worker.
1944 static struct worker *create_worker(struct worker_pool *pool)
1946 struct worker *worker;
1950 /* ID is needed to determine kthread name */
1951 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
1955 worker = alloc_worker(pool->node);
1962 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1963 pool->attrs->nice < 0 ? "H" : "");
1965 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1967 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1968 "kworker/%s", id_buf);
1969 if (IS_ERR(worker->task))
1972 set_user_nice(worker->task, pool->attrs->nice);
1973 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1975 /* successful, attach the worker to the pool */
1976 worker_attach_to_pool(worker, pool);
1978 /* start the newly created worker */
1979 raw_spin_lock_irq(&pool->lock);
1980 worker->pool->nr_workers++;
1981 worker_enter_idle(worker);
1982 wake_up_process(worker->task);
1983 raw_spin_unlock_irq(&pool->lock);
1988 ida_free(&pool->worker_ida, id);
1994 * destroy_worker - destroy a workqueue worker
1995 * @worker: worker to be destroyed
1997 * Destroy @worker and adjust @pool stats accordingly. The worker should
2001 * raw_spin_lock_irq(pool->lock).
2003 static void destroy_worker(struct worker *worker)
2005 struct worker_pool *pool = worker->pool;
2007 lockdep_assert_held(&pool->lock);
2009 /* sanity check frenzy */
2010 if (WARN_ON(worker->current_work) ||
2011 WARN_ON(!list_empty(&worker->scheduled)) ||
2012 WARN_ON(!(worker->flags & WORKER_IDLE)))
2018 list_del_init(&worker->entry);
2019 worker->flags |= WORKER_DIE;
2020 wake_up_process(worker->task);
2023 static void idle_worker_timeout(struct timer_list *t)
2025 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2027 raw_spin_lock_irq(&pool->lock);
2029 while (too_many_workers(pool)) {
2030 struct worker *worker;
2031 unsigned long expires;
2033 /* idle_list is kept in LIFO order, check the last one */
2034 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2035 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2037 if (time_before(jiffies, expires)) {
2038 mod_timer(&pool->idle_timer, expires);
2042 destroy_worker(worker);
2045 raw_spin_unlock_irq(&pool->lock);
2048 static void send_mayday(struct work_struct *work)
2050 struct pool_workqueue *pwq = get_work_pwq(work);
2051 struct workqueue_struct *wq = pwq->wq;
2053 lockdep_assert_held(&wq_mayday_lock);
2058 /* mayday mayday mayday */
2059 if (list_empty(&pwq->mayday_node)) {
2061 * If @pwq is for an unbound wq, its base ref may be put at
2062 * any time due to an attribute change. Pin @pwq until the
2063 * rescuer is done with it.
2066 list_add_tail(&pwq->mayday_node, &wq->maydays);
2067 wake_up_process(wq->rescuer->task);
2071 static void pool_mayday_timeout(struct timer_list *t)
2073 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2074 struct work_struct *work;
2076 raw_spin_lock_irq(&pool->lock);
2077 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2079 if (need_to_create_worker(pool)) {
2081 * We've been trying to create a new worker but
2082 * haven't been successful. We might be hitting an
2083 * allocation deadlock. Send distress signals to
2086 list_for_each_entry(work, &pool->worklist, entry)
2090 raw_spin_unlock(&wq_mayday_lock);
2091 raw_spin_unlock_irq(&pool->lock);
2093 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2097 * maybe_create_worker - create a new worker if necessary
2098 * @pool: pool to create a new worker for
2100 * Create a new worker for @pool if necessary. @pool is guaranteed to
2101 * have at least one idle worker on return from this function. If
2102 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2103 * sent to all rescuers with works scheduled on @pool to resolve
2104 * possible allocation deadlock.
2106 * On return, need_to_create_worker() is guaranteed to be %false and
2107 * may_start_working() %true.
2110 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2111 * multiple times. Does GFP_KERNEL allocations. Called only from
2114 static void maybe_create_worker(struct worker_pool *pool)
2115 __releases(&pool->lock)
2116 __acquires(&pool->lock)
2119 raw_spin_unlock_irq(&pool->lock);
2121 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2122 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2125 if (create_worker(pool) || !need_to_create_worker(pool))
2128 schedule_timeout_interruptible(CREATE_COOLDOWN);
2130 if (!need_to_create_worker(pool))
2134 del_timer_sync(&pool->mayday_timer);
2135 raw_spin_lock_irq(&pool->lock);
2137 * This is necessary even after a new worker was just successfully
2138 * created as @pool->lock was dropped and the new worker might have
2139 * already become busy.
2141 if (need_to_create_worker(pool))
2146 * manage_workers - manage worker pool
2149 * Assume the manager role and manage the worker pool @worker belongs
2150 * to. At any given time, there can be only zero or one manager per
2151 * pool. The exclusion is handled automatically by this function.
2153 * The caller can safely start processing works on false return. On
2154 * true return, it's guaranteed that need_to_create_worker() is false
2155 * and may_start_working() is true.
2158 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2159 * multiple times. Does GFP_KERNEL allocations.
2162 * %false if the pool doesn't need management and the caller can safely
2163 * start processing works, %true if management function was performed and
2164 * the conditions that the caller verified before calling the function may
2165 * no longer be true.
2167 static bool manage_workers(struct worker *worker)
2169 struct worker_pool *pool = worker->pool;
2171 if (pool->flags & POOL_MANAGER_ACTIVE)
2174 pool->flags |= POOL_MANAGER_ACTIVE;
2175 pool->manager = worker;
2177 maybe_create_worker(pool);
2179 pool->manager = NULL;
2180 pool->flags &= ~POOL_MANAGER_ACTIVE;
2181 rcuwait_wake_up(&manager_wait);
2186 * process_one_work - process single work
2188 * @work: work to process
2190 * Process @work. This function contains all the logics necessary to
2191 * process a single work including synchronization against and
2192 * interaction with other workers on the same cpu, queueing and
2193 * flushing. As long as context requirement is met, any worker can
2194 * call this function to process a work.
2197 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2199 static void process_one_work(struct worker *worker, struct work_struct *work)
2200 __releases(&pool->lock)
2201 __acquires(&pool->lock)
2203 struct pool_workqueue *pwq = get_work_pwq(work);
2204 struct worker_pool *pool = worker->pool;
2205 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2206 unsigned long work_data;
2207 struct worker *collision;
2208 #ifdef CONFIG_LOCKDEP
2210 * It is permissible to free the struct work_struct from
2211 * inside the function that is called from it, this we need to
2212 * take into account for lockdep too. To avoid bogus "held
2213 * lock freed" warnings as well as problems when looking into
2214 * work->lockdep_map, make a copy and use that here.
2216 struct lockdep_map lockdep_map;
2218 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2220 /* ensure we're on the correct CPU */
2221 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2222 raw_smp_processor_id() != pool->cpu);
2225 * A single work shouldn't be executed concurrently by
2226 * multiple workers on a single cpu. Check whether anyone is
2227 * already processing the work. If so, defer the work to the
2228 * currently executing one.
2230 collision = find_worker_executing_work(pool, work);
2231 if (unlikely(collision)) {
2232 move_linked_works(work, &collision->scheduled, NULL);
2236 /* claim and dequeue */
2237 debug_work_deactivate(work);
2238 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2239 worker->current_work = work;
2240 worker->current_func = work->func;
2241 worker->current_pwq = pwq;
2242 work_data = *work_data_bits(work);
2243 worker->current_color = get_work_color(work_data);
2246 * Record wq name for cmdline and debug reporting, may get
2247 * overridden through set_worker_desc().
2249 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2251 list_del_init(&work->entry);
2254 * CPU intensive works don't participate in concurrency management.
2255 * They're the scheduler's responsibility. This takes @worker out
2256 * of concurrency management and the next code block will chain
2257 * execution of the pending work items.
2259 if (unlikely(cpu_intensive))
2260 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2263 * Wake up another worker if necessary. The condition is always
2264 * false for normal per-cpu workers since nr_running would always
2265 * be >= 1 at this point. This is used to chain execution of the
2266 * pending work items for WORKER_NOT_RUNNING workers such as the
2267 * UNBOUND and CPU_INTENSIVE ones.
2269 if (need_more_worker(pool))
2270 wake_up_worker(pool);
2273 * Record the last pool and clear PENDING which should be the last
2274 * update to @work. Also, do this inside @pool->lock so that
2275 * PENDING and queued state changes happen together while IRQ is
2278 set_work_pool_and_clear_pending(work, pool->id);
2280 raw_spin_unlock_irq(&pool->lock);
2282 lock_map_acquire(&pwq->wq->lockdep_map);
2283 lock_map_acquire(&lockdep_map);
2285 * Strictly speaking we should mark the invariant state without holding
2286 * any locks, that is, before these two lock_map_acquire()'s.
2288 * However, that would result in:
2295 * Which would create W1->C->W1 dependencies, even though there is no
2296 * actual deadlock possible. There are two solutions, using a
2297 * read-recursive acquire on the work(queue) 'locks', but this will then
2298 * hit the lockdep limitation on recursive locks, or simply discard
2301 * AFAICT there is no possible deadlock scenario between the
2302 * flush_work() and complete() primitives (except for single-threaded
2303 * workqueues), so hiding them isn't a problem.
2305 lockdep_invariant_state(true);
2306 trace_workqueue_execute_start(work);
2307 worker->current_func(work);
2309 * While we must be careful to not use "work" after this, the trace
2310 * point will only record its address.
2312 trace_workqueue_execute_end(work, worker->current_func);
2313 lock_map_release(&lockdep_map);
2314 lock_map_release(&pwq->wq->lockdep_map);
2316 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2317 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2318 " last function: %ps\n",
2319 current->comm, preempt_count(), task_pid_nr(current),
2320 worker->current_func);
2321 debug_show_held_locks(current);
2326 * The following prevents a kworker from hogging CPU on !PREEMPTION
2327 * kernels, where a requeueing work item waiting for something to
2328 * happen could deadlock with stop_machine as such work item could
2329 * indefinitely requeue itself while all other CPUs are trapped in
2330 * stop_machine. At the same time, report a quiescent RCU state so
2331 * the same condition doesn't freeze RCU.
2335 raw_spin_lock_irq(&pool->lock);
2337 /* clear cpu intensive status */
2338 if (unlikely(cpu_intensive))
2339 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2341 /* tag the worker for identification in schedule() */
2342 worker->last_func = worker->current_func;
2344 /* we're done with it, release */
2345 hash_del(&worker->hentry);
2346 worker->current_work = NULL;
2347 worker->current_func = NULL;
2348 worker->current_pwq = NULL;
2349 worker->current_color = INT_MAX;
2350 pwq_dec_nr_in_flight(pwq, work_data);
2354 * process_scheduled_works - process scheduled works
2357 * Process all scheduled works. Please note that the scheduled list
2358 * may change while processing a work, so this function repeatedly
2359 * fetches a work from the top and executes it.
2362 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2365 static void process_scheduled_works(struct worker *worker)
2367 while (!list_empty(&worker->scheduled)) {
2368 struct work_struct *work = list_first_entry(&worker->scheduled,
2369 struct work_struct, entry);
2370 process_one_work(worker, work);
2374 static void set_pf_worker(bool val)
2376 mutex_lock(&wq_pool_attach_mutex);
2378 current->flags |= PF_WQ_WORKER;
2380 current->flags &= ~PF_WQ_WORKER;
2381 mutex_unlock(&wq_pool_attach_mutex);
2385 * worker_thread - the worker thread function
2388 * The worker thread function. All workers belong to a worker_pool -
2389 * either a per-cpu one or dynamic unbound one. These workers process all
2390 * work items regardless of their specific target workqueue. The only
2391 * exception is work items which belong to workqueues with a rescuer which
2392 * will be explained in rescuer_thread().
2396 static int worker_thread(void *__worker)
2398 struct worker *worker = __worker;
2399 struct worker_pool *pool = worker->pool;
2401 /* tell the scheduler that this is a workqueue worker */
2402 set_pf_worker(true);
2404 raw_spin_lock_irq(&pool->lock);
2406 /* am I supposed to die? */
2407 if (unlikely(worker->flags & WORKER_DIE)) {
2408 raw_spin_unlock_irq(&pool->lock);
2409 WARN_ON_ONCE(!list_empty(&worker->entry));
2410 set_pf_worker(false);
2412 set_task_comm(worker->task, "kworker/dying");
2413 ida_free(&pool->worker_ida, worker->id);
2414 worker_detach_from_pool(worker);
2419 worker_leave_idle(worker);
2421 /* no more worker necessary? */
2422 if (!need_more_worker(pool))
2425 /* do we need to manage? */
2426 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2430 * ->scheduled list can only be filled while a worker is
2431 * preparing to process a work or actually processing it.
2432 * Make sure nobody diddled with it while I was sleeping.
2434 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2437 * Finish PREP stage. We're guaranteed to have at least one idle
2438 * worker or that someone else has already assumed the manager
2439 * role. This is where @worker starts participating in concurrency
2440 * management if applicable and concurrency management is restored
2441 * after being rebound. See rebind_workers() for details.
2443 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2446 struct work_struct *work =
2447 list_first_entry(&pool->worklist,
2448 struct work_struct, entry);
2450 pool->watchdog_ts = jiffies;
2452 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2453 /* optimization path, not strictly necessary */
2454 process_one_work(worker, work);
2455 if (unlikely(!list_empty(&worker->scheduled)))
2456 process_scheduled_works(worker);
2458 move_linked_works(work, &worker->scheduled, NULL);
2459 process_scheduled_works(worker);
2461 } while (keep_working(pool));
2463 worker_set_flags(worker, WORKER_PREP);
2466 * pool->lock is held and there's no work to process and no need to
2467 * manage, sleep. Workers are woken up only while holding
2468 * pool->lock or from local cpu, so setting the current state
2469 * before releasing pool->lock is enough to prevent losing any
2472 worker_enter_idle(worker);
2473 __set_current_state(TASK_IDLE);
2474 raw_spin_unlock_irq(&pool->lock);
2480 * rescuer_thread - the rescuer thread function
2483 * Workqueue rescuer thread function. There's one rescuer for each
2484 * workqueue which has WQ_MEM_RECLAIM set.
2486 * Regular work processing on a pool may block trying to create a new
2487 * worker which uses GFP_KERNEL allocation which has slight chance of
2488 * developing into deadlock if some works currently on the same queue
2489 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2490 * the problem rescuer solves.
2492 * When such condition is possible, the pool summons rescuers of all
2493 * workqueues which have works queued on the pool and let them process
2494 * those works so that forward progress can be guaranteed.
2496 * This should happen rarely.
2500 static int rescuer_thread(void *__rescuer)
2502 struct worker *rescuer = __rescuer;
2503 struct workqueue_struct *wq = rescuer->rescue_wq;
2504 struct list_head *scheduled = &rescuer->scheduled;
2507 set_user_nice(current, RESCUER_NICE_LEVEL);
2510 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2511 * doesn't participate in concurrency management.
2513 set_pf_worker(true);
2515 set_current_state(TASK_IDLE);
2518 * By the time the rescuer is requested to stop, the workqueue
2519 * shouldn't have any work pending, but @wq->maydays may still have
2520 * pwq(s) queued. This can happen by non-rescuer workers consuming
2521 * all the work items before the rescuer got to them. Go through
2522 * @wq->maydays processing before acting on should_stop so that the
2523 * list is always empty on exit.
2525 should_stop = kthread_should_stop();
2527 /* see whether any pwq is asking for help */
2528 raw_spin_lock_irq(&wq_mayday_lock);
2530 while (!list_empty(&wq->maydays)) {
2531 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2532 struct pool_workqueue, mayday_node);
2533 struct worker_pool *pool = pwq->pool;
2534 struct work_struct *work, *n;
2537 __set_current_state(TASK_RUNNING);
2538 list_del_init(&pwq->mayday_node);
2540 raw_spin_unlock_irq(&wq_mayday_lock);
2542 worker_attach_to_pool(rescuer, pool);
2544 raw_spin_lock_irq(&pool->lock);
2547 * Slurp in all works issued via this workqueue and
2550 WARN_ON_ONCE(!list_empty(scheduled));
2551 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2552 if (get_work_pwq(work) == pwq) {
2554 pool->watchdog_ts = jiffies;
2555 move_linked_works(work, scheduled, &n);
2560 if (!list_empty(scheduled)) {
2561 process_scheduled_works(rescuer);
2564 * The above execution of rescued work items could
2565 * have created more to rescue through
2566 * pwq_activate_first_inactive() or chained
2567 * queueing. Let's put @pwq back on mayday list so
2568 * that such back-to-back work items, which may be
2569 * being used to relieve memory pressure, don't
2570 * incur MAYDAY_INTERVAL delay inbetween.
2572 if (pwq->nr_active && need_to_create_worker(pool)) {
2573 raw_spin_lock(&wq_mayday_lock);
2575 * Queue iff we aren't racing destruction
2576 * and somebody else hasn't queued it already.
2578 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2580 list_add_tail(&pwq->mayday_node, &wq->maydays);
2582 raw_spin_unlock(&wq_mayday_lock);
2587 * Put the reference grabbed by send_mayday(). @pool won't
2588 * go away while we're still attached to it.
2593 * Leave this pool. If need_more_worker() is %true, notify a
2594 * regular worker; otherwise, we end up with 0 concurrency
2595 * and stalling the execution.
2597 if (need_more_worker(pool))
2598 wake_up_worker(pool);
2600 raw_spin_unlock_irq(&pool->lock);
2602 worker_detach_from_pool(rescuer);
2604 raw_spin_lock_irq(&wq_mayday_lock);
2607 raw_spin_unlock_irq(&wq_mayday_lock);
2610 __set_current_state(TASK_RUNNING);
2611 set_pf_worker(false);
2615 /* rescuers should never participate in concurrency management */
2616 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2622 * check_flush_dependency - check for flush dependency sanity
2623 * @target_wq: workqueue being flushed
2624 * @target_work: work item being flushed (NULL for workqueue flushes)
2626 * %current is trying to flush the whole @target_wq or @target_work on it.
2627 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2628 * reclaiming memory or running on a workqueue which doesn't have
2629 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2632 static void check_flush_dependency(struct workqueue_struct *target_wq,
2633 struct work_struct *target_work)
2635 work_func_t target_func = target_work ? target_work->func : NULL;
2636 struct worker *worker;
2638 if (target_wq->flags & WQ_MEM_RECLAIM)
2641 worker = current_wq_worker();
2643 WARN_ONCE(current->flags & PF_MEMALLOC,
2644 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2645 current->pid, current->comm, target_wq->name, target_func);
2646 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2647 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2648 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2649 worker->current_pwq->wq->name, worker->current_func,
2650 target_wq->name, target_func);
2654 struct work_struct work;
2655 struct completion done;
2656 struct task_struct *task; /* purely informational */
2659 static void wq_barrier_func(struct work_struct *work)
2661 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2662 complete(&barr->done);
2666 * insert_wq_barrier - insert a barrier work
2667 * @pwq: pwq to insert barrier into
2668 * @barr: wq_barrier to insert
2669 * @target: target work to attach @barr to
2670 * @worker: worker currently executing @target, NULL if @target is not executing
2672 * @barr is linked to @target such that @barr is completed only after
2673 * @target finishes execution. Please note that the ordering
2674 * guarantee is observed only with respect to @target and on the local
2677 * Currently, a queued barrier can't be canceled. This is because
2678 * try_to_grab_pending() can't determine whether the work to be
2679 * grabbed is at the head of the queue and thus can't clear LINKED
2680 * flag of the previous work while there must be a valid next work
2681 * after a work with LINKED flag set.
2683 * Note that when @worker is non-NULL, @target may be modified
2684 * underneath us, so we can't reliably determine pwq from @target.
2687 * raw_spin_lock_irq(pool->lock).
2689 static void insert_wq_barrier(struct pool_workqueue *pwq,
2690 struct wq_barrier *barr,
2691 struct work_struct *target, struct worker *worker)
2693 unsigned int work_flags = 0;
2694 unsigned int work_color;
2695 struct list_head *head;
2698 * debugobject calls are safe here even with pool->lock locked
2699 * as we know for sure that this will not trigger any of the
2700 * checks and call back into the fixup functions where we
2703 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2704 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2706 init_completion_map(&barr->done, &target->lockdep_map);
2708 barr->task = current;
2710 /* The barrier work item does not participate in pwq->nr_active. */
2711 work_flags |= WORK_STRUCT_INACTIVE;
2714 * If @target is currently being executed, schedule the
2715 * barrier to the worker; otherwise, put it after @target.
2718 head = worker->scheduled.next;
2719 work_color = worker->current_color;
2721 unsigned long *bits = work_data_bits(target);
2723 head = target->entry.next;
2724 /* there can already be other linked works, inherit and set */
2725 work_flags |= *bits & WORK_STRUCT_LINKED;
2726 work_color = get_work_color(*bits);
2727 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2730 pwq->nr_in_flight[work_color]++;
2731 work_flags |= work_color_to_flags(work_color);
2733 debug_work_activate(&barr->work);
2734 insert_work(pwq, &barr->work, head, work_flags);
2738 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2739 * @wq: workqueue being flushed
2740 * @flush_color: new flush color, < 0 for no-op
2741 * @work_color: new work color, < 0 for no-op
2743 * Prepare pwqs for workqueue flushing.
2745 * If @flush_color is non-negative, flush_color on all pwqs should be
2746 * -1. If no pwq has in-flight commands at the specified color, all
2747 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2748 * has in flight commands, its pwq->flush_color is set to
2749 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2750 * wakeup logic is armed and %true is returned.
2752 * The caller should have initialized @wq->first_flusher prior to
2753 * calling this function with non-negative @flush_color. If
2754 * @flush_color is negative, no flush color update is done and %false
2757 * If @work_color is non-negative, all pwqs should have the same
2758 * work_color which is previous to @work_color and all will be
2759 * advanced to @work_color.
2762 * mutex_lock(wq->mutex).
2765 * %true if @flush_color >= 0 and there's something to flush. %false
2768 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2769 int flush_color, int work_color)
2772 struct pool_workqueue *pwq;
2774 if (flush_color >= 0) {
2775 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2776 atomic_set(&wq->nr_pwqs_to_flush, 1);
2779 for_each_pwq(pwq, wq) {
2780 struct worker_pool *pool = pwq->pool;
2782 raw_spin_lock_irq(&pool->lock);
2784 if (flush_color >= 0) {
2785 WARN_ON_ONCE(pwq->flush_color != -1);
2787 if (pwq->nr_in_flight[flush_color]) {
2788 pwq->flush_color = flush_color;
2789 atomic_inc(&wq->nr_pwqs_to_flush);
2794 if (work_color >= 0) {
2795 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2796 pwq->work_color = work_color;
2799 raw_spin_unlock_irq(&pool->lock);
2802 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2803 complete(&wq->first_flusher->done);
2809 * flush_workqueue - ensure that any scheduled work has run to completion.
2810 * @wq: workqueue to flush
2812 * This function sleeps until all work items which were queued on entry
2813 * have finished execution, but it is not livelocked by new incoming ones.
2815 void flush_workqueue(struct workqueue_struct *wq)
2817 struct wq_flusher this_flusher = {
2818 .list = LIST_HEAD_INIT(this_flusher.list),
2820 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2824 if (WARN_ON(!wq_online))
2827 lock_map_acquire(&wq->lockdep_map);
2828 lock_map_release(&wq->lockdep_map);
2830 mutex_lock(&wq->mutex);
2833 * Start-to-wait phase
2835 next_color = work_next_color(wq->work_color);
2837 if (next_color != wq->flush_color) {
2839 * Color space is not full. The current work_color
2840 * becomes our flush_color and work_color is advanced
2843 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2844 this_flusher.flush_color = wq->work_color;
2845 wq->work_color = next_color;
2847 if (!wq->first_flusher) {
2848 /* no flush in progress, become the first flusher */
2849 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2851 wq->first_flusher = &this_flusher;
2853 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2855 /* nothing to flush, done */
2856 wq->flush_color = next_color;
2857 wq->first_flusher = NULL;
2862 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2863 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2864 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2868 * Oops, color space is full, wait on overflow queue.
2869 * The next flush completion will assign us
2870 * flush_color and transfer to flusher_queue.
2872 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2875 check_flush_dependency(wq, NULL);
2877 mutex_unlock(&wq->mutex);
2879 wait_for_completion(&this_flusher.done);
2882 * Wake-up-and-cascade phase
2884 * First flushers are responsible for cascading flushes and
2885 * handling overflow. Non-first flushers can simply return.
2887 if (READ_ONCE(wq->first_flusher) != &this_flusher)
2890 mutex_lock(&wq->mutex);
2892 /* we might have raced, check again with mutex held */
2893 if (wq->first_flusher != &this_flusher)
2896 WRITE_ONCE(wq->first_flusher, NULL);
2898 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2899 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2902 struct wq_flusher *next, *tmp;
2904 /* complete all the flushers sharing the current flush color */
2905 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2906 if (next->flush_color != wq->flush_color)
2908 list_del_init(&next->list);
2909 complete(&next->done);
2912 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2913 wq->flush_color != work_next_color(wq->work_color));
2915 /* this flush_color is finished, advance by one */
2916 wq->flush_color = work_next_color(wq->flush_color);
2918 /* one color has been freed, handle overflow queue */
2919 if (!list_empty(&wq->flusher_overflow)) {
2921 * Assign the same color to all overflowed
2922 * flushers, advance work_color and append to
2923 * flusher_queue. This is the start-to-wait
2924 * phase for these overflowed flushers.
2926 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2927 tmp->flush_color = wq->work_color;
2929 wq->work_color = work_next_color(wq->work_color);
2931 list_splice_tail_init(&wq->flusher_overflow,
2932 &wq->flusher_queue);
2933 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2936 if (list_empty(&wq->flusher_queue)) {
2937 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2942 * Need to flush more colors. Make the next flusher
2943 * the new first flusher and arm pwqs.
2945 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2946 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2948 list_del_init(&next->list);
2949 wq->first_flusher = next;
2951 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2955 * Meh... this color is already done, clear first
2956 * flusher and repeat cascading.
2958 wq->first_flusher = NULL;
2962 mutex_unlock(&wq->mutex);
2964 EXPORT_SYMBOL(flush_workqueue);
2967 * drain_workqueue - drain a workqueue
2968 * @wq: workqueue to drain
2970 * Wait until the workqueue becomes empty. While draining is in progress,
2971 * only chain queueing is allowed. IOW, only currently pending or running
2972 * work items on @wq can queue further work items on it. @wq is flushed
2973 * repeatedly until it becomes empty. The number of flushing is determined
2974 * by the depth of chaining and should be relatively short. Whine if it
2977 void drain_workqueue(struct workqueue_struct *wq)
2979 unsigned int flush_cnt = 0;
2980 struct pool_workqueue *pwq;
2983 * __queue_work() needs to test whether there are drainers, is much
2984 * hotter than drain_workqueue() and already looks at @wq->flags.
2985 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2987 mutex_lock(&wq->mutex);
2988 if (!wq->nr_drainers++)
2989 wq->flags |= __WQ_DRAINING;
2990 mutex_unlock(&wq->mutex);
2992 flush_workqueue(wq);
2994 mutex_lock(&wq->mutex);
2996 for_each_pwq(pwq, wq) {
2999 raw_spin_lock_irq(&pwq->pool->lock);
3000 drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
3001 raw_spin_unlock_irq(&pwq->pool->lock);
3006 if (++flush_cnt == 10 ||
3007 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3008 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3009 wq->name, __func__, flush_cnt);
3011 mutex_unlock(&wq->mutex);
3015 if (!--wq->nr_drainers)
3016 wq->flags &= ~__WQ_DRAINING;
3017 mutex_unlock(&wq->mutex);
3019 EXPORT_SYMBOL_GPL(drain_workqueue);
3021 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3024 struct worker *worker = NULL;
3025 struct worker_pool *pool;
3026 struct pool_workqueue *pwq;
3031 pool = get_work_pool(work);
3037 raw_spin_lock_irq(&pool->lock);
3038 /* see the comment in try_to_grab_pending() with the same code */
3039 pwq = get_work_pwq(work);
3041 if (unlikely(pwq->pool != pool))
3044 worker = find_worker_executing_work(pool, work);
3047 pwq = worker->current_pwq;
3050 check_flush_dependency(pwq->wq, work);
3052 insert_wq_barrier(pwq, barr, work, worker);
3053 raw_spin_unlock_irq(&pool->lock);
3056 * Force a lock recursion deadlock when using flush_work() inside a
3057 * single-threaded or rescuer equipped workqueue.
3059 * For single threaded workqueues the deadlock happens when the work
3060 * is after the work issuing the flush_work(). For rescuer equipped
3061 * workqueues the deadlock happens when the rescuer stalls, blocking
3065 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3066 lock_map_acquire(&pwq->wq->lockdep_map);
3067 lock_map_release(&pwq->wq->lockdep_map);
3072 raw_spin_unlock_irq(&pool->lock);
3077 static bool __flush_work(struct work_struct *work, bool from_cancel)
3079 struct wq_barrier barr;
3081 if (WARN_ON(!wq_online))
3084 if (WARN_ON(!work->func))
3088 lock_map_acquire(&work->lockdep_map);
3089 lock_map_release(&work->lockdep_map);
3092 if (start_flush_work(work, &barr, from_cancel)) {
3093 wait_for_completion(&barr.done);
3094 destroy_work_on_stack(&barr.work);
3102 * flush_work - wait for a work to finish executing the last queueing instance
3103 * @work: the work to flush
3105 * Wait until @work has finished execution. @work is guaranteed to be idle
3106 * on return if it hasn't been requeued since flush started.
3109 * %true if flush_work() waited for the work to finish execution,
3110 * %false if it was already idle.
3112 bool flush_work(struct work_struct *work)
3114 return __flush_work(work, false);
3116 EXPORT_SYMBOL_GPL(flush_work);
3119 wait_queue_entry_t wait;
3120 struct work_struct *work;
3123 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3125 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3127 if (cwait->work != key)
3129 return autoremove_wake_function(wait, mode, sync, key);
3132 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3134 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3135 unsigned long flags;
3139 ret = try_to_grab_pending(work, is_dwork, &flags);
3141 * If someone else is already canceling, wait for it to
3142 * finish. flush_work() doesn't work for PREEMPT_NONE
3143 * because we may get scheduled between @work's completion
3144 * and the other canceling task resuming and clearing
3145 * CANCELING - flush_work() will return false immediately
3146 * as @work is no longer busy, try_to_grab_pending() will
3147 * return -ENOENT as @work is still being canceled and the
3148 * other canceling task won't be able to clear CANCELING as
3149 * we're hogging the CPU.
3151 * Let's wait for completion using a waitqueue. As this
3152 * may lead to the thundering herd problem, use a custom
3153 * wake function which matches @work along with exclusive
3156 if (unlikely(ret == -ENOENT)) {
3157 struct cwt_wait cwait;
3159 init_wait(&cwait.wait);
3160 cwait.wait.func = cwt_wakefn;
3163 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3164 TASK_UNINTERRUPTIBLE);
3165 if (work_is_canceling(work))
3167 finish_wait(&cancel_waitq, &cwait.wait);
3169 } while (unlikely(ret < 0));
3171 /* tell other tasks trying to grab @work to back off */
3172 mark_work_canceling(work);
3173 local_irq_restore(flags);
3176 * This allows canceling during early boot. We know that @work
3180 __flush_work(work, true);
3182 clear_work_data(work);
3185 * Paired with prepare_to_wait() above so that either
3186 * waitqueue_active() is visible here or !work_is_canceling() is
3190 if (waitqueue_active(&cancel_waitq))
3191 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3197 * cancel_work_sync - cancel a work and wait for it to finish
3198 * @work: the work to cancel
3200 * Cancel @work and wait for its execution to finish. This function
3201 * can be used even if the work re-queues itself or migrates to
3202 * another workqueue. On return from this function, @work is
3203 * guaranteed to be not pending or executing on any CPU.
3205 * cancel_work_sync(&delayed_work->work) must not be used for
3206 * delayed_work's. Use cancel_delayed_work_sync() instead.
3208 * The caller must ensure that the workqueue on which @work was last
3209 * queued can't be destroyed before this function returns.
3212 * %true if @work was pending, %false otherwise.
3214 bool cancel_work_sync(struct work_struct *work)
3216 return __cancel_work_timer(work, false);
3218 EXPORT_SYMBOL_GPL(cancel_work_sync);
3221 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3222 * @dwork: the delayed work to flush
3224 * Delayed timer is cancelled and the pending work is queued for
3225 * immediate execution. Like flush_work(), this function only
3226 * considers the last queueing instance of @dwork.
3229 * %true if flush_work() waited for the work to finish execution,
3230 * %false if it was already idle.
3232 bool flush_delayed_work(struct delayed_work *dwork)
3234 local_irq_disable();
3235 if (del_timer_sync(&dwork->timer))
3236 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3238 return flush_work(&dwork->work);
3240 EXPORT_SYMBOL(flush_delayed_work);
3243 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3244 * @rwork: the rcu work to flush
3247 * %true if flush_rcu_work() waited for the work to finish execution,
3248 * %false if it was already idle.
3250 bool flush_rcu_work(struct rcu_work *rwork)
3252 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3254 flush_work(&rwork->work);
3257 return flush_work(&rwork->work);
3260 EXPORT_SYMBOL(flush_rcu_work);
3262 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3264 unsigned long flags;
3268 ret = try_to_grab_pending(work, is_dwork, &flags);
3269 } while (unlikely(ret == -EAGAIN));
3271 if (unlikely(ret < 0))
3274 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3275 local_irq_restore(flags);
3280 * cancel_delayed_work - cancel a delayed work
3281 * @dwork: delayed_work to cancel
3283 * Kill off a pending delayed_work.
3285 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3289 * The work callback function may still be running on return, unless
3290 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3291 * use cancel_delayed_work_sync() to wait on it.
3293 * This function is safe to call from any context including IRQ handler.
3295 bool cancel_delayed_work(struct delayed_work *dwork)
3297 return __cancel_work(&dwork->work, true);
3299 EXPORT_SYMBOL(cancel_delayed_work);
3302 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3303 * @dwork: the delayed work cancel
3305 * This is cancel_work_sync() for delayed works.
3308 * %true if @dwork was pending, %false otherwise.
3310 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3312 return __cancel_work_timer(&dwork->work, true);
3314 EXPORT_SYMBOL(cancel_delayed_work_sync);
3317 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3318 * @func: the function to call
3320 * schedule_on_each_cpu() executes @func on each online CPU using the
3321 * system workqueue and blocks until all CPUs have completed.
3322 * schedule_on_each_cpu() is very slow.
3325 * 0 on success, -errno on failure.
3327 int schedule_on_each_cpu(work_func_t func)
3330 struct work_struct __percpu *works;
3332 works = alloc_percpu(struct work_struct);
3338 for_each_online_cpu(cpu) {
3339 struct work_struct *work = per_cpu_ptr(works, cpu);
3341 INIT_WORK(work, func);
3342 schedule_work_on(cpu, work);
3345 for_each_online_cpu(cpu)
3346 flush_work(per_cpu_ptr(works, cpu));
3354 * execute_in_process_context - reliably execute the routine with user context
3355 * @fn: the function to execute
3356 * @ew: guaranteed storage for the execute work structure (must
3357 * be available when the work executes)
3359 * Executes the function immediately if process context is available,
3360 * otherwise schedules the function for delayed execution.
3362 * Return: 0 - function was executed
3363 * 1 - function was scheduled for execution
3365 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3367 if (!in_interrupt()) {
3372 INIT_WORK(&ew->work, fn);
3373 schedule_work(&ew->work);
3377 EXPORT_SYMBOL_GPL(execute_in_process_context);
3380 * free_workqueue_attrs - free a workqueue_attrs
3381 * @attrs: workqueue_attrs to free
3383 * Undo alloc_workqueue_attrs().
3385 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3388 free_cpumask_var(attrs->cpumask);
3394 * alloc_workqueue_attrs - allocate a workqueue_attrs
3396 * Allocate a new workqueue_attrs, initialize with default settings and
3399 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3401 struct workqueue_attrs *alloc_workqueue_attrs(void)
3403 struct workqueue_attrs *attrs;
3405 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3408 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3411 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3414 free_workqueue_attrs(attrs);
3418 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3419 const struct workqueue_attrs *from)
3421 to->nice = from->nice;
3422 cpumask_copy(to->cpumask, from->cpumask);
3424 * Unlike hash and equality test, this function doesn't ignore
3425 * ->no_numa as it is used for both pool and wq attrs. Instead,
3426 * get_unbound_pool() explicitly clears ->no_numa after copying.
3428 to->no_numa = from->no_numa;
3431 /* hash value of the content of @attr */
3432 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3436 hash = jhash_1word(attrs->nice, hash);
3437 hash = jhash(cpumask_bits(attrs->cpumask),
3438 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3442 /* content equality test */
3443 static bool wqattrs_equal(const struct workqueue_attrs *a,
3444 const struct workqueue_attrs *b)
3446 if (a->nice != b->nice)
3448 if (!cpumask_equal(a->cpumask, b->cpumask))
3454 * init_worker_pool - initialize a newly zalloc'd worker_pool
3455 * @pool: worker_pool to initialize
3457 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3459 * Return: 0 on success, -errno on failure. Even on failure, all fields
3460 * inside @pool proper are initialized and put_unbound_pool() can be called
3461 * on @pool safely to release it.
3463 static int init_worker_pool(struct worker_pool *pool)
3465 raw_spin_lock_init(&pool->lock);
3468 pool->node = NUMA_NO_NODE;
3469 pool->flags |= POOL_DISASSOCIATED;
3470 pool->watchdog_ts = jiffies;
3471 INIT_LIST_HEAD(&pool->worklist);
3472 INIT_LIST_HEAD(&pool->idle_list);
3473 hash_init(pool->busy_hash);
3475 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3477 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3479 INIT_LIST_HEAD(&pool->workers);
3481 ida_init(&pool->worker_ida);
3482 INIT_HLIST_NODE(&pool->hash_node);
3485 /* shouldn't fail above this point */
3486 pool->attrs = alloc_workqueue_attrs();
3492 #ifdef CONFIG_LOCKDEP
3493 static void wq_init_lockdep(struct workqueue_struct *wq)
3497 lockdep_register_key(&wq->key);
3498 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3500 lock_name = wq->name;
3502 wq->lock_name = lock_name;
3503 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3506 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3508 lockdep_unregister_key(&wq->key);
3511 static void wq_free_lockdep(struct workqueue_struct *wq)
3513 if (wq->lock_name != wq->name)
3514 kfree(wq->lock_name);
3517 static void wq_init_lockdep(struct workqueue_struct *wq)
3521 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3525 static void wq_free_lockdep(struct workqueue_struct *wq)
3530 static void rcu_free_wq(struct rcu_head *rcu)
3532 struct workqueue_struct *wq =
3533 container_of(rcu, struct workqueue_struct, rcu);
3535 wq_free_lockdep(wq);
3537 if (!(wq->flags & WQ_UNBOUND))
3538 free_percpu(wq->cpu_pwqs);
3540 free_workqueue_attrs(wq->unbound_attrs);
3545 static void rcu_free_pool(struct rcu_head *rcu)
3547 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3549 ida_destroy(&pool->worker_ida);
3550 free_workqueue_attrs(pool->attrs);
3554 /* This returns with the lock held on success (pool manager is inactive). */
3555 static bool wq_manager_inactive(struct worker_pool *pool)
3557 raw_spin_lock_irq(&pool->lock);
3559 if (pool->flags & POOL_MANAGER_ACTIVE) {
3560 raw_spin_unlock_irq(&pool->lock);
3567 * put_unbound_pool - put a worker_pool
3568 * @pool: worker_pool to put
3570 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3571 * safe manner. get_unbound_pool() calls this function on its failure path
3572 * and this function should be able to release pools which went through,
3573 * successfully or not, init_worker_pool().
3575 * Should be called with wq_pool_mutex held.
3577 static void put_unbound_pool(struct worker_pool *pool)
3579 DECLARE_COMPLETION_ONSTACK(detach_completion);
3580 struct worker *worker;
3582 lockdep_assert_held(&wq_pool_mutex);
3588 if (WARN_ON(!(pool->cpu < 0)) ||
3589 WARN_ON(!list_empty(&pool->worklist)))
3592 /* release id and unhash */
3594 idr_remove(&worker_pool_idr, pool->id);
3595 hash_del(&pool->hash_node);
3598 * Become the manager and destroy all workers. This prevents
3599 * @pool's workers from blocking on attach_mutex. We're the last
3600 * manager and @pool gets freed with the flag set.
3601 * Because of how wq_manager_inactive() works, we will hold the
3602 * spinlock after a successful wait.
3604 rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
3605 TASK_UNINTERRUPTIBLE);
3606 pool->flags |= POOL_MANAGER_ACTIVE;
3608 while ((worker = first_idle_worker(pool)))
3609 destroy_worker(worker);
3610 WARN_ON(pool->nr_workers || pool->nr_idle);
3611 raw_spin_unlock_irq(&pool->lock);
3613 mutex_lock(&wq_pool_attach_mutex);
3614 if (!list_empty(&pool->workers))
3615 pool->detach_completion = &detach_completion;
3616 mutex_unlock(&wq_pool_attach_mutex);
3618 if (pool->detach_completion)
3619 wait_for_completion(pool->detach_completion);
3621 /* shut down the timers */
3622 del_timer_sync(&pool->idle_timer);
3623 del_timer_sync(&pool->mayday_timer);
3625 /* RCU protected to allow dereferences from get_work_pool() */
3626 call_rcu(&pool->rcu, rcu_free_pool);
3630 * get_unbound_pool - get a worker_pool with the specified attributes
3631 * @attrs: the attributes of the worker_pool to get
3633 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3634 * reference count and return it. If there already is a matching
3635 * worker_pool, it will be used; otherwise, this function attempts to
3638 * Should be called with wq_pool_mutex held.
3640 * Return: On success, a worker_pool with the same attributes as @attrs.
3641 * On failure, %NULL.
3643 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3645 u32 hash = wqattrs_hash(attrs);
3646 struct worker_pool *pool;
3648 int target_node = NUMA_NO_NODE;
3650 lockdep_assert_held(&wq_pool_mutex);
3652 /* do we already have a matching pool? */
3653 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3654 if (wqattrs_equal(pool->attrs, attrs)) {
3660 /* if cpumask is contained inside a NUMA node, we belong to that node */
3661 if (wq_numa_enabled) {
3662 for_each_node(node) {
3663 if (cpumask_subset(attrs->cpumask,
3664 wq_numa_possible_cpumask[node])) {
3671 /* nope, create a new one */
3672 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3673 if (!pool || init_worker_pool(pool) < 0)
3676 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3677 copy_workqueue_attrs(pool->attrs, attrs);
3678 pool->node = target_node;
3681 * no_numa isn't a worker_pool attribute, always clear it. See
3682 * 'struct workqueue_attrs' comments for detail.
3684 pool->attrs->no_numa = false;
3686 if (worker_pool_assign_id(pool) < 0)
3689 /* create and start the initial worker */
3690 if (wq_online && !create_worker(pool))
3694 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3699 put_unbound_pool(pool);
3703 static void rcu_free_pwq(struct rcu_head *rcu)
3705 kmem_cache_free(pwq_cache,
3706 container_of(rcu, struct pool_workqueue, rcu));
3710 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3711 * and needs to be destroyed.
3713 static void pwq_unbound_release_workfn(struct work_struct *work)
3715 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3716 unbound_release_work);
3717 struct workqueue_struct *wq = pwq->wq;
3718 struct worker_pool *pool = pwq->pool;
3719 bool is_last = false;
3722 * when @pwq is not linked, it doesn't hold any reference to the
3723 * @wq, and @wq is invalid to access.
3725 if (!list_empty(&pwq->pwqs_node)) {
3726 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3729 mutex_lock(&wq->mutex);
3730 list_del_rcu(&pwq->pwqs_node);
3731 is_last = list_empty(&wq->pwqs);
3732 mutex_unlock(&wq->mutex);
3735 mutex_lock(&wq_pool_mutex);
3736 put_unbound_pool(pool);
3737 mutex_unlock(&wq_pool_mutex);
3739 call_rcu(&pwq->rcu, rcu_free_pwq);
3742 * If we're the last pwq going away, @wq is already dead and no one
3743 * is gonna access it anymore. Schedule RCU free.
3746 wq_unregister_lockdep(wq);
3747 call_rcu(&wq->rcu, rcu_free_wq);
3752 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3753 * @pwq: target pool_workqueue
3755 * If @pwq isn't freezing, set @pwq->max_active to the associated
3756 * workqueue's saved_max_active and activate inactive work items
3757 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3759 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3761 struct workqueue_struct *wq = pwq->wq;
3762 bool freezable = wq->flags & WQ_FREEZABLE;
3763 unsigned long flags;
3765 /* for @wq->saved_max_active */
3766 lockdep_assert_held(&wq->mutex);
3768 /* fast exit for non-freezable wqs */
3769 if (!freezable && pwq->max_active == wq->saved_max_active)
3772 /* this function can be called during early boot w/ irq disabled */
3773 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
3776 * During [un]freezing, the caller is responsible for ensuring that
3777 * this function is called at least once after @workqueue_freezing
3778 * is updated and visible.
3780 if (!freezable || !workqueue_freezing) {
3783 pwq->max_active = wq->saved_max_active;
3785 while (!list_empty(&pwq->inactive_works) &&
3786 pwq->nr_active < pwq->max_active) {
3787 pwq_activate_first_inactive(pwq);
3792 * Need to kick a worker after thawed or an unbound wq's
3793 * max_active is bumped. In realtime scenarios, always kicking a
3794 * worker will cause interference on the isolated cpu cores, so
3795 * let's kick iff work items were activated.
3798 wake_up_worker(pwq->pool);
3800 pwq->max_active = 0;
3803 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
3806 /* initialize newly allocated @pwq which is associated with @wq and @pool */
3807 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3808 struct worker_pool *pool)
3810 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3812 memset(pwq, 0, sizeof(*pwq));
3816 pwq->flush_color = -1;
3818 INIT_LIST_HEAD(&pwq->inactive_works);
3819 INIT_LIST_HEAD(&pwq->pwqs_node);
3820 INIT_LIST_HEAD(&pwq->mayday_node);
3821 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3824 /* sync @pwq with the current state of its associated wq and link it */
3825 static void link_pwq(struct pool_workqueue *pwq)
3827 struct workqueue_struct *wq = pwq->wq;
3829 lockdep_assert_held(&wq->mutex);
3831 /* may be called multiple times, ignore if already linked */
3832 if (!list_empty(&pwq->pwqs_node))
3835 /* set the matching work_color */
3836 pwq->work_color = wq->work_color;
3838 /* sync max_active to the current setting */
3839 pwq_adjust_max_active(pwq);
3842 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3845 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3846 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3847 const struct workqueue_attrs *attrs)
3849 struct worker_pool *pool;
3850 struct pool_workqueue *pwq;
3852 lockdep_assert_held(&wq_pool_mutex);
3854 pool = get_unbound_pool(attrs);
3858 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3860 put_unbound_pool(pool);
3864 init_pwq(pwq, wq, pool);
3869 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3870 * @attrs: the wq_attrs of the default pwq of the target workqueue
3871 * @node: the target NUMA node
3872 * @cpu_going_down: if >= 0, the CPU to consider as offline
3873 * @cpumask: outarg, the resulting cpumask
3875 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3876 * @cpu_going_down is >= 0, that cpu is considered offline during
3877 * calculation. The result is stored in @cpumask.
3879 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3880 * enabled and @node has online CPUs requested by @attrs, the returned
3881 * cpumask is the intersection of the possible CPUs of @node and
3884 * The caller is responsible for ensuring that the cpumask of @node stays
3887 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3890 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3891 int cpu_going_down, cpumask_t *cpumask)
3893 if (!wq_numa_enabled || attrs->no_numa)
3896 /* does @node have any online CPUs @attrs wants? */
3897 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3898 if (cpu_going_down >= 0)
3899 cpumask_clear_cpu(cpu_going_down, cpumask);
3901 if (cpumask_empty(cpumask))
3904 /* yeap, return possible CPUs in @node that @attrs wants */
3905 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3907 if (cpumask_empty(cpumask)) {
3908 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3909 "possible intersect\n");
3913 return !cpumask_equal(cpumask, attrs->cpumask);
3916 cpumask_copy(cpumask, attrs->cpumask);
3920 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3921 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3923 struct pool_workqueue *pwq)
3925 struct pool_workqueue *old_pwq;
3927 lockdep_assert_held(&wq_pool_mutex);
3928 lockdep_assert_held(&wq->mutex);
3930 /* link_pwq() can handle duplicate calls */
3933 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3934 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3938 /* context to store the prepared attrs & pwqs before applying */
3939 struct apply_wqattrs_ctx {
3940 struct workqueue_struct *wq; /* target workqueue */
3941 struct workqueue_attrs *attrs; /* attrs to apply */
3942 struct list_head list; /* queued for batching commit */
3943 struct pool_workqueue *dfl_pwq;
3944 struct pool_workqueue *pwq_tbl[];
3947 /* free the resources after success or abort */
3948 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3954 put_pwq_unlocked(ctx->pwq_tbl[node]);
3955 put_pwq_unlocked(ctx->dfl_pwq);
3957 free_workqueue_attrs(ctx->attrs);
3963 /* allocate the attrs and pwqs for later installation */
3964 static struct apply_wqattrs_ctx *
3965 apply_wqattrs_prepare(struct workqueue_struct *wq,
3966 const struct workqueue_attrs *attrs)
3968 struct apply_wqattrs_ctx *ctx;
3969 struct workqueue_attrs *new_attrs, *tmp_attrs;
3972 lockdep_assert_held(&wq_pool_mutex);
3974 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3976 new_attrs = alloc_workqueue_attrs();
3977 tmp_attrs = alloc_workqueue_attrs();
3978 if (!ctx || !new_attrs || !tmp_attrs)
3982 * Calculate the attrs of the default pwq.
3983 * If the user configured cpumask doesn't overlap with the
3984 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3986 copy_workqueue_attrs(new_attrs, attrs);
3987 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3988 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3989 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3992 * We may create multiple pwqs with differing cpumasks. Make a
3993 * copy of @new_attrs which will be modified and used to obtain
3996 copy_workqueue_attrs(tmp_attrs, new_attrs);
3999 * If something goes wrong during CPU up/down, we'll fall back to
4000 * the default pwq covering whole @attrs->cpumask. Always create
4001 * it even if we don't use it immediately.
4003 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4007 for_each_node(node) {
4008 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
4009 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
4010 if (!ctx->pwq_tbl[node])
4013 ctx->dfl_pwq->refcnt++;
4014 ctx->pwq_tbl[node] = ctx->dfl_pwq;
4018 /* save the user configured attrs and sanitize it. */
4019 copy_workqueue_attrs(new_attrs, attrs);
4020 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4021 ctx->attrs = new_attrs;
4024 free_workqueue_attrs(tmp_attrs);
4028 free_workqueue_attrs(tmp_attrs);
4029 free_workqueue_attrs(new_attrs);
4030 apply_wqattrs_cleanup(ctx);
4034 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4035 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4039 /* all pwqs have been created successfully, let's install'em */
4040 mutex_lock(&ctx->wq->mutex);
4042 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4044 /* save the previous pwq and install the new one */
4046 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
4047 ctx->pwq_tbl[node]);
4049 /* @dfl_pwq might not have been used, ensure it's linked */
4050 link_pwq(ctx->dfl_pwq);
4051 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4053 mutex_unlock(&ctx->wq->mutex);
4056 static void apply_wqattrs_lock(void)
4058 /* CPUs should stay stable across pwq creations and installations */
4060 mutex_lock(&wq_pool_mutex);
4063 static void apply_wqattrs_unlock(void)
4065 mutex_unlock(&wq_pool_mutex);
4069 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4070 const struct workqueue_attrs *attrs)
4072 struct apply_wqattrs_ctx *ctx;
4074 /* only unbound workqueues can change attributes */
4075 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4078 /* creating multiple pwqs breaks ordering guarantee */
4079 if (!list_empty(&wq->pwqs)) {
4080 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4083 wq->flags &= ~__WQ_ORDERED;
4086 ctx = apply_wqattrs_prepare(wq, attrs);
4090 /* the ctx has been prepared successfully, let's commit it */
4091 apply_wqattrs_commit(ctx);
4092 apply_wqattrs_cleanup(ctx);
4098 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4099 * @wq: the target workqueue
4100 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4102 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4103 * machines, this function maps a separate pwq to each NUMA node with
4104 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4105 * NUMA node it was issued on. Older pwqs are released as in-flight work
4106 * items finish. Note that a work item which repeatedly requeues itself
4107 * back-to-back will stay on its current pwq.
4109 * Performs GFP_KERNEL allocations.
4111 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4113 * Return: 0 on success and -errno on failure.
4115 int apply_workqueue_attrs(struct workqueue_struct *wq,
4116 const struct workqueue_attrs *attrs)
4120 lockdep_assert_cpus_held();
4122 mutex_lock(&wq_pool_mutex);
4123 ret = apply_workqueue_attrs_locked(wq, attrs);
4124 mutex_unlock(&wq_pool_mutex);
4130 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4131 * @wq: the target workqueue
4132 * @cpu: the CPU coming up or going down
4133 * @online: whether @cpu is coming up or going down
4135 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4136 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4139 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4140 * falls back to @wq->dfl_pwq which may not be optimal but is always
4143 * Note that when the last allowed CPU of a NUMA node goes offline for a
4144 * workqueue with a cpumask spanning multiple nodes, the workers which were
4145 * already executing the work items for the workqueue will lose their CPU
4146 * affinity and may execute on any CPU. This is similar to how per-cpu
4147 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4148 * affinity, it's the user's responsibility to flush the work item from
4151 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4154 int node = cpu_to_node(cpu);
4155 int cpu_off = online ? -1 : cpu;
4156 struct pool_workqueue *old_pwq = NULL, *pwq;
4157 struct workqueue_attrs *target_attrs;
4160 lockdep_assert_held(&wq_pool_mutex);
4162 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4163 wq->unbound_attrs->no_numa)
4167 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4168 * Let's use a preallocated one. The following buf is protected by
4169 * CPU hotplug exclusion.
4171 target_attrs = wq_update_unbound_numa_attrs_buf;
4172 cpumask = target_attrs->cpumask;
4174 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4175 pwq = unbound_pwq_by_node(wq, node);
4178 * Let's determine what needs to be done. If the target cpumask is
4179 * different from the default pwq's, we need to compare it to @pwq's
4180 * and create a new one if they don't match. If the target cpumask
4181 * equals the default pwq's, the default pwq should be used.
4183 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4184 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4190 /* create a new pwq */
4191 pwq = alloc_unbound_pwq(wq, target_attrs);
4193 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4198 /* Install the new pwq. */
4199 mutex_lock(&wq->mutex);
4200 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4204 mutex_lock(&wq->mutex);
4205 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4206 get_pwq(wq->dfl_pwq);
4207 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4208 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4210 mutex_unlock(&wq->mutex);
4211 put_pwq_unlocked(old_pwq);
4214 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4216 bool highpri = wq->flags & WQ_HIGHPRI;
4219 if (!(wq->flags & WQ_UNBOUND)) {
4220 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4224 for_each_possible_cpu(cpu) {
4225 struct pool_workqueue *pwq =
4226 per_cpu_ptr(wq->cpu_pwqs, cpu);
4227 struct worker_pool *cpu_pools =
4228 per_cpu(cpu_worker_pools, cpu);
4230 init_pwq(pwq, wq, &cpu_pools[highpri]);
4232 mutex_lock(&wq->mutex);
4234 mutex_unlock(&wq->mutex);
4240 if (wq->flags & __WQ_ORDERED) {
4241 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4242 /* there should only be single pwq for ordering guarantee */
4243 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4244 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4245 "ordering guarantee broken for workqueue %s\n", wq->name);
4247 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4254 static int wq_clamp_max_active(int max_active, unsigned int flags,
4257 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4259 if (max_active < 1 || max_active > lim)
4260 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4261 max_active, name, 1, lim);
4263 return clamp_val(max_active, 1, lim);
4267 * Workqueues which may be used during memory reclaim should have a rescuer
4268 * to guarantee forward progress.
4270 static int init_rescuer(struct workqueue_struct *wq)
4272 struct worker *rescuer;
4275 if (!(wq->flags & WQ_MEM_RECLAIM))
4278 rescuer = alloc_worker(NUMA_NO_NODE);
4282 rescuer->rescue_wq = wq;
4283 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4284 if (IS_ERR(rescuer->task)) {
4285 ret = PTR_ERR(rescuer->task);
4290 wq->rescuer = rescuer;
4291 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4292 wake_up_process(rescuer->task);
4298 struct workqueue_struct *alloc_workqueue(const char *fmt,
4300 int max_active, ...)
4302 size_t tbl_size = 0;
4304 struct workqueue_struct *wq;
4305 struct pool_workqueue *pwq;
4308 * Unbound && max_active == 1 used to imply ordered, which is no
4309 * longer the case on NUMA machines due to per-node pools. While
4310 * alloc_ordered_workqueue() is the right way to create an ordered
4311 * workqueue, keep the previous behavior to avoid subtle breakages
4314 if ((flags & WQ_UNBOUND) && max_active == 1)
4315 flags |= __WQ_ORDERED;
4317 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4318 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4319 flags |= WQ_UNBOUND;
4321 /* allocate wq and format name */
4322 if (flags & WQ_UNBOUND)
4323 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4325 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4329 if (flags & WQ_UNBOUND) {
4330 wq->unbound_attrs = alloc_workqueue_attrs();
4331 if (!wq->unbound_attrs)
4335 va_start(args, max_active);
4336 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4339 max_active = max_active ?: WQ_DFL_ACTIVE;
4340 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4344 wq->saved_max_active = max_active;
4345 mutex_init(&wq->mutex);
4346 atomic_set(&wq->nr_pwqs_to_flush, 0);
4347 INIT_LIST_HEAD(&wq->pwqs);
4348 INIT_LIST_HEAD(&wq->flusher_queue);
4349 INIT_LIST_HEAD(&wq->flusher_overflow);
4350 INIT_LIST_HEAD(&wq->maydays);
4352 wq_init_lockdep(wq);
4353 INIT_LIST_HEAD(&wq->list);
4355 if (alloc_and_link_pwqs(wq) < 0)
4356 goto err_unreg_lockdep;
4358 if (wq_online && init_rescuer(wq) < 0)
4361 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4365 * wq_pool_mutex protects global freeze state and workqueues list.
4366 * Grab it, adjust max_active and add the new @wq to workqueues
4369 mutex_lock(&wq_pool_mutex);
4371 mutex_lock(&wq->mutex);
4372 for_each_pwq(pwq, wq)
4373 pwq_adjust_max_active(pwq);
4374 mutex_unlock(&wq->mutex);
4376 list_add_tail_rcu(&wq->list, &workqueues);
4378 mutex_unlock(&wq_pool_mutex);
4383 wq_unregister_lockdep(wq);
4384 wq_free_lockdep(wq);
4386 free_workqueue_attrs(wq->unbound_attrs);
4390 destroy_workqueue(wq);
4393 EXPORT_SYMBOL_GPL(alloc_workqueue);
4395 static bool pwq_busy(struct pool_workqueue *pwq)
4399 for (i = 0; i < WORK_NR_COLORS; i++)
4400 if (pwq->nr_in_flight[i])
4403 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4405 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4412 * destroy_workqueue - safely terminate a workqueue
4413 * @wq: target workqueue
4415 * Safely destroy a workqueue. All work currently pending will be done first.
4417 void destroy_workqueue(struct workqueue_struct *wq)
4419 struct pool_workqueue *pwq;
4423 * Remove it from sysfs first so that sanity check failure doesn't
4424 * lead to sysfs name conflicts.
4426 workqueue_sysfs_unregister(wq);
4428 /* drain it before proceeding with destruction */
4429 drain_workqueue(wq);
4431 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4433 struct worker *rescuer = wq->rescuer;
4435 /* this prevents new queueing */
4436 raw_spin_lock_irq(&wq_mayday_lock);
4438 raw_spin_unlock_irq(&wq_mayday_lock);
4440 /* rescuer will empty maydays list before exiting */
4441 kthread_stop(rescuer->task);
4446 * Sanity checks - grab all the locks so that we wait for all
4447 * in-flight operations which may do put_pwq().
4449 mutex_lock(&wq_pool_mutex);
4450 mutex_lock(&wq->mutex);
4451 for_each_pwq(pwq, wq) {
4452 raw_spin_lock_irq(&pwq->pool->lock);
4453 if (WARN_ON(pwq_busy(pwq))) {
4454 pr_warn("%s: %s has the following busy pwq\n",
4455 __func__, wq->name);
4457 raw_spin_unlock_irq(&pwq->pool->lock);
4458 mutex_unlock(&wq->mutex);
4459 mutex_unlock(&wq_pool_mutex);
4460 show_one_workqueue(wq);
4463 raw_spin_unlock_irq(&pwq->pool->lock);
4465 mutex_unlock(&wq->mutex);
4468 * wq list is used to freeze wq, remove from list after
4469 * flushing is complete in case freeze races us.
4471 list_del_rcu(&wq->list);
4472 mutex_unlock(&wq_pool_mutex);
4474 if (!(wq->flags & WQ_UNBOUND)) {
4475 wq_unregister_lockdep(wq);
4477 * The base ref is never dropped on per-cpu pwqs. Directly
4478 * schedule RCU free.
4480 call_rcu(&wq->rcu, rcu_free_wq);
4483 * We're the sole accessor of @wq at this point. Directly
4484 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4485 * @wq will be freed when the last pwq is released.
4487 for_each_node(node) {
4488 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4489 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4490 put_pwq_unlocked(pwq);
4494 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4495 * put. Don't access it afterwards.
4499 put_pwq_unlocked(pwq);
4502 EXPORT_SYMBOL_GPL(destroy_workqueue);
4505 * workqueue_set_max_active - adjust max_active of a workqueue
4506 * @wq: target workqueue
4507 * @max_active: new max_active value.
4509 * Set max_active of @wq to @max_active.
4512 * Don't call from IRQ context.
4514 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4516 struct pool_workqueue *pwq;
4518 /* disallow meddling with max_active for ordered workqueues */
4519 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4522 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4524 mutex_lock(&wq->mutex);
4526 wq->flags &= ~__WQ_ORDERED;
4527 wq->saved_max_active = max_active;
4529 for_each_pwq(pwq, wq)
4530 pwq_adjust_max_active(pwq);
4532 mutex_unlock(&wq->mutex);
4534 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4537 * current_work - retrieve %current task's work struct
4539 * Determine if %current task is a workqueue worker and what it's working on.
4540 * Useful to find out the context that the %current task is running in.
4542 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4544 struct work_struct *current_work(void)
4546 struct worker *worker = current_wq_worker();
4548 return worker ? worker->current_work : NULL;
4550 EXPORT_SYMBOL(current_work);
4553 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4555 * Determine whether %current is a workqueue rescuer. Can be used from
4556 * work functions to determine whether it's being run off the rescuer task.
4558 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4560 bool current_is_workqueue_rescuer(void)
4562 struct worker *worker = current_wq_worker();
4564 return worker && worker->rescue_wq;
4568 * workqueue_congested - test whether a workqueue is congested
4569 * @cpu: CPU in question
4570 * @wq: target workqueue
4572 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4573 * no synchronization around this function and the test result is
4574 * unreliable and only useful as advisory hints or for debugging.
4576 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4577 * Note that both per-cpu and unbound workqueues may be associated with
4578 * multiple pool_workqueues which have separate congested states. A
4579 * workqueue being congested on one CPU doesn't mean the workqueue is also
4580 * contested on other CPUs / NUMA nodes.
4583 * %true if congested, %false otherwise.
4585 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4587 struct pool_workqueue *pwq;
4593 if (cpu == WORK_CPU_UNBOUND)
4594 cpu = smp_processor_id();
4596 if (!(wq->flags & WQ_UNBOUND))
4597 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4599 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4601 ret = !list_empty(&pwq->inactive_works);
4607 EXPORT_SYMBOL_GPL(workqueue_congested);
4610 * work_busy - test whether a work is currently pending or running
4611 * @work: the work to be tested
4613 * Test whether @work is currently pending or running. There is no
4614 * synchronization around this function and the test result is
4615 * unreliable and only useful as advisory hints or for debugging.
4618 * OR'd bitmask of WORK_BUSY_* bits.
4620 unsigned int work_busy(struct work_struct *work)
4622 struct worker_pool *pool;
4623 unsigned long flags;
4624 unsigned int ret = 0;
4626 if (work_pending(work))
4627 ret |= WORK_BUSY_PENDING;
4630 pool = get_work_pool(work);
4632 raw_spin_lock_irqsave(&pool->lock, flags);
4633 if (find_worker_executing_work(pool, work))
4634 ret |= WORK_BUSY_RUNNING;
4635 raw_spin_unlock_irqrestore(&pool->lock, flags);
4641 EXPORT_SYMBOL_GPL(work_busy);
4644 * set_worker_desc - set description for the current work item
4645 * @fmt: printf-style format string
4646 * @...: arguments for the format string
4648 * This function can be called by a running work function to describe what
4649 * the work item is about. If the worker task gets dumped, this
4650 * information will be printed out together to help debugging. The
4651 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4653 void set_worker_desc(const char *fmt, ...)
4655 struct worker *worker = current_wq_worker();
4659 va_start(args, fmt);
4660 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4664 EXPORT_SYMBOL_GPL(set_worker_desc);
4667 * print_worker_info - print out worker information and description
4668 * @log_lvl: the log level to use when printing
4669 * @task: target task
4671 * If @task is a worker and currently executing a work item, print out the
4672 * name of the workqueue being serviced and worker description set with
4673 * set_worker_desc() by the currently executing work item.
4675 * This function can be safely called on any task as long as the
4676 * task_struct itself is accessible. While safe, this function isn't
4677 * synchronized and may print out mixups or garbages of limited length.
4679 void print_worker_info(const char *log_lvl, struct task_struct *task)
4681 work_func_t *fn = NULL;
4682 char name[WQ_NAME_LEN] = { };
4683 char desc[WORKER_DESC_LEN] = { };
4684 struct pool_workqueue *pwq = NULL;
4685 struct workqueue_struct *wq = NULL;
4686 struct worker *worker;
4688 if (!(task->flags & PF_WQ_WORKER))
4692 * This function is called without any synchronization and @task
4693 * could be in any state. Be careful with dereferences.
4695 worker = kthread_probe_data(task);
4698 * Carefully copy the associated workqueue's workfn, name and desc.
4699 * Keep the original last '\0' in case the original is garbage.
4701 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
4702 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
4703 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
4704 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
4705 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
4707 if (fn || name[0] || desc[0]) {
4708 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4709 if (strcmp(name, desc))
4710 pr_cont(" (%s)", desc);
4715 static void pr_cont_pool_info(struct worker_pool *pool)
4717 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4718 if (pool->node != NUMA_NO_NODE)
4719 pr_cont(" node=%d", pool->node);
4720 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4723 static void pr_cont_work(bool comma, struct work_struct *work)
4725 if (work->func == wq_barrier_func) {
4726 struct wq_barrier *barr;
4728 barr = container_of(work, struct wq_barrier, work);
4730 pr_cont("%s BAR(%d)", comma ? "," : "",
4731 task_pid_nr(barr->task));
4733 pr_cont("%s %ps", comma ? "," : "", work->func);
4737 static void show_pwq(struct pool_workqueue *pwq)
4739 struct worker_pool *pool = pwq->pool;
4740 struct work_struct *work;
4741 struct worker *worker;
4742 bool has_in_flight = false, has_pending = false;
4745 pr_info(" pwq %d:", pool->id);
4746 pr_cont_pool_info(pool);
4748 pr_cont(" active=%d/%d refcnt=%d%s\n",
4749 pwq->nr_active, pwq->max_active, pwq->refcnt,
4750 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4752 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4753 if (worker->current_pwq == pwq) {
4754 has_in_flight = true;
4758 if (has_in_flight) {
4761 pr_info(" in-flight:");
4762 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4763 if (worker->current_pwq != pwq)
4766 pr_cont("%s %d%s:%ps", comma ? "," : "",
4767 task_pid_nr(worker->task),
4768 worker->rescue_wq ? "(RESCUER)" : "",
4769 worker->current_func);
4770 list_for_each_entry(work, &worker->scheduled, entry)
4771 pr_cont_work(false, work);
4777 list_for_each_entry(work, &pool->worklist, entry) {
4778 if (get_work_pwq(work) == pwq) {
4786 pr_info(" pending:");
4787 list_for_each_entry(work, &pool->worklist, entry) {
4788 if (get_work_pwq(work) != pwq)
4791 pr_cont_work(comma, work);
4792 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4797 if (!list_empty(&pwq->inactive_works)) {
4800 pr_info(" inactive:");
4801 list_for_each_entry(work, &pwq->inactive_works, entry) {
4802 pr_cont_work(comma, work);
4803 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4810 * show_one_workqueue - dump state of specified workqueue
4811 * @wq: workqueue whose state will be printed
4813 void show_one_workqueue(struct workqueue_struct *wq)
4815 struct pool_workqueue *pwq;
4817 unsigned long flags;
4819 for_each_pwq(pwq, wq) {
4820 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4825 if (idle) /* Nothing to print for idle workqueue */
4828 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4830 for_each_pwq(pwq, wq) {
4831 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4832 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4834 * Defer printing to avoid deadlocks in console
4835 * drivers that queue work while holding locks
4836 * also taken in their write paths.
4838 printk_deferred_enter();
4840 printk_deferred_exit();
4842 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4844 * We could be printing a lot from atomic context, e.g.
4845 * sysrq-t -> show_all_workqueues(). Avoid triggering
4848 touch_nmi_watchdog();
4854 * show_one_worker_pool - dump state of specified worker pool
4855 * @pool: worker pool whose state will be printed
4857 static void show_one_worker_pool(struct worker_pool *pool)
4859 struct worker *worker;
4861 unsigned long flags;
4863 raw_spin_lock_irqsave(&pool->lock, flags);
4864 if (pool->nr_workers == pool->nr_idle)
4867 * Defer printing to avoid deadlocks in console drivers that
4868 * queue work while holding locks also taken in their write
4871 printk_deferred_enter();
4872 pr_info("pool %d:", pool->id);
4873 pr_cont_pool_info(pool);
4874 pr_cont(" hung=%us workers=%d",
4875 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4878 pr_cont(" manager: %d",
4879 task_pid_nr(pool->manager->task));
4880 list_for_each_entry(worker, &pool->idle_list, entry) {
4881 pr_cont(" %s%d", first ? "idle: " : "",
4882 task_pid_nr(worker->task));
4886 printk_deferred_exit();
4888 raw_spin_unlock_irqrestore(&pool->lock, flags);
4890 * We could be printing a lot from atomic context, e.g.
4891 * sysrq-t -> show_all_workqueues(). Avoid triggering
4894 touch_nmi_watchdog();
4899 * show_all_workqueues - dump workqueue state
4901 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4902 * all busy workqueues and pools.
4904 void show_all_workqueues(void)
4906 struct workqueue_struct *wq;
4907 struct worker_pool *pool;
4912 pr_info("Showing busy workqueues and worker pools:\n");
4914 list_for_each_entry_rcu(wq, &workqueues, list)
4915 show_one_workqueue(wq);
4917 for_each_pool(pool, pi)
4918 show_one_worker_pool(pool);
4923 /* used to show worker information through /proc/PID/{comm,stat,status} */
4924 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4928 /* always show the actual comm */
4929 off = strscpy(buf, task->comm, size);
4933 /* stabilize PF_WQ_WORKER and worker pool association */
4934 mutex_lock(&wq_pool_attach_mutex);
4936 if (task->flags & PF_WQ_WORKER) {
4937 struct worker *worker = kthread_data(task);
4938 struct worker_pool *pool = worker->pool;
4941 raw_spin_lock_irq(&pool->lock);
4943 * ->desc tracks information (wq name or
4944 * set_worker_desc()) for the latest execution. If
4945 * current, prepend '+', otherwise '-'.
4947 if (worker->desc[0] != '\0') {
4948 if (worker->current_work)
4949 scnprintf(buf + off, size - off, "+%s",
4952 scnprintf(buf + off, size - off, "-%s",
4955 raw_spin_unlock_irq(&pool->lock);
4959 mutex_unlock(&wq_pool_attach_mutex);
4967 * There are two challenges in supporting CPU hotplug. Firstly, there
4968 * are a lot of assumptions on strong associations among work, pwq and
4969 * pool which make migrating pending and scheduled works very
4970 * difficult to implement without impacting hot paths. Secondly,
4971 * worker pools serve mix of short, long and very long running works making
4972 * blocked draining impractical.
4974 * This is solved by allowing the pools to be disassociated from the CPU
4975 * running as an unbound one and allowing it to be reattached later if the
4976 * cpu comes back online.
4979 static void unbind_workers(int cpu)
4981 struct worker_pool *pool;
4982 struct worker *worker;
4984 for_each_cpu_worker_pool(pool, cpu) {
4985 mutex_lock(&wq_pool_attach_mutex);
4986 raw_spin_lock_irq(&pool->lock);
4989 * We've blocked all attach/detach operations. Make all workers
4990 * unbound and set DISASSOCIATED. Before this, all workers
4991 * must be on the cpu. After this, they may become diasporas.
4992 * And the preemption disabled section in their sched callbacks
4993 * are guaranteed to see WORKER_UNBOUND since the code here
4994 * is on the same cpu.
4996 for_each_pool_worker(worker, pool)
4997 worker->flags |= WORKER_UNBOUND;
4999 pool->flags |= POOL_DISASSOCIATED;
5002 * The handling of nr_running in sched callbacks are disabled
5003 * now. Zap nr_running. After this, nr_running stays zero and
5004 * need_more_worker() and keep_working() are always true as
5005 * long as the worklist is not empty. This pool now behaves as
5006 * an unbound (in terms of concurrency management) pool which
5007 * are served by workers tied to the pool.
5009 atomic_set(&pool->nr_running, 0);
5012 * With concurrency management just turned off, a busy
5013 * worker blocking could lead to lengthy stalls. Kick off
5014 * unbound chain execution of currently pending work items.
5016 wake_up_worker(pool);
5018 raw_spin_unlock_irq(&pool->lock);
5020 for_each_pool_worker(worker, pool) {
5021 kthread_set_per_cpu(worker->task, -1);
5022 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
5025 mutex_unlock(&wq_pool_attach_mutex);
5030 * rebind_workers - rebind all workers of a pool to the associated CPU
5031 * @pool: pool of interest
5033 * @pool->cpu is coming online. Rebind all workers to the CPU.
5035 static void rebind_workers(struct worker_pool *pool)
5037 struct worker *worker;
5039 lockdep_assert_held(&wq_pool_attach_mutex);
5042 * Restore CPU affinity of all workers. As all idle workers should
5043 * be on the run-queue of the associated CPU before any local
5044 * wake-ups for concurrency management happen, restore CPU affinity
5045 * of all workers first and then clear UNBOUND. As we're called
5046 * from CPU_ONLINE, the following shouldn't fail.
5048 for_each_pool_worker(worker, pool) {
5049 kthread_set_per_cpu(worker->task, pool->cpu);
5050 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5051 pool->attrs->cpumask) < 0);
5054 raw_spin_lock_irq(&pool->lock);
5056 pool->flags &= ~POOL_DISASSOCIATED;
5058 for_each_pool_worker(worker, pool) {
5059 unsigned int worker_flags = worker->flags;
5062 * We want to clear UNBOUND but can't directly call
5063 * worker_clr_flags() or adjust nr_running. Atomically
5064 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5065 * @worker will clear REBOUND using worker_clr_flags() when
5066 * it initiates the next execution cycle thus restoring
5067 * concurrency management. Note that when or whether
5068 * @worker clears REBOUND doesn't affect correctness.
5070 * WRITE_ONCE() is necessary because @worker->flags may be
5071 * tested without holding any lock in
5072 * wq_worker_running(). Without it, NOT_RUNNING test may
5073 * fail incorrectly leading to premature concurrency
5074 * management operations.
5076 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5077 worker_flags |= WORKER_REBOUND;
5078 worker_flags &= ~WORKER_UNBOUND;
5079 WRITE_ONCE(worker->flags, worker_flags);
5082 raw_spin_unlock_irq(&pool->lock);
5086 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5087 * @pool: unbound pool of interest
5088 * @cpu: the CPU which is coming up
5090 * An unbound pool may end up with a cpumask which doesn't have any online
5091 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5092 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5093 * online CPU before, cpus_allowed of all its workers should be restored.
5095 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5097 static cpumask_t cpumask;
5098 struct worker *worker;
5100 lockdep_assert_held(&wq_pool_attach_mutex);
5102 /* is @cpu allowed for @pool? */
5103 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5106 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5108 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5109 for_each_pool_worker(worker, pool)
5110 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5113 int workqueue_prepare_cpu(unsigned int cpu)
5115 struct worker_pool *pool;
5117 for_each_cpu_worker_pool(pool, cpu) {
5118 if (pool->nr_workers)
5120 if (!create_worker(pool))
5126 int workqueue_online_cpu(unsigned int cpu)
5128 struct worker_pool *pool;
5129 struct workqueue_struct *wq;
5132 mutex_lock(&wq_pool_mutex);
5134 for_each_pool(pool, pi) {
5135 mutex_lock(&wq_pool_attach_mutex);
5137 if (pool->cpu == cpu)
5138 rebind_workers(pool);
5139 else if (pool->cpu < 0)
5140 restore_unbound_workers_cpumask(pool, cpu);
5142 mutex_unlock(&wq_pool_attach_mutex);
5145 /* update NUMA affinity of unbound workqueues */
5146 list_for_each_entry(wq, &workqueues, list)
5147 wq_update_unbound_numa(wq, cpu, true);
5149 mutex_unlock(&wq_pool_mutex);
5153 int workqueue_offline_cpu(unsigned int cpu)
5155 struct workqueue_struct *wq;
5157 /* unbinding per-cpu workers should happen on the local CPU */
5158 if (WARN_ON(cpu != smp_processor_id()))
5161 unbind_workers(cpu);
5163 /* update NUMA affinity of unbound workqueues */
5164 mutex_lock(&wq_pool_mutex);
5165 list_for_each_entry(wq, &workqueues, list)
5166 wq_update_unbound_numa(wq, cpu, false);
5167 mutex_unlock(&wq_pool_mutex);
5172 struct work_for_cpu {
5173 struct work_struct work;
5179 static void work_for_cpu_fn(struct work_struct *work)
5181 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5183 wfc->ret = wfc->fn(wfc->arg);
5187 * work_on_cpu - run a function in thread context on a particular cpu
5188 * @cpu: the cpu to run on
5189 * @fn: the function to run
5190 * @arg: the function arg
5192 * It is up to the caller to ensure that the cpu doesn't go offline.
5193 * The caller must not hold any locks which would prevent @fn from completing.
5195 * Return: The value @fn returns.
5197 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5199 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5201 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5202 schedule_work_on(cpu, &wfc.work);
5203 flush_work(&wfc.work);
5204 destroy_work_on_stack(&wfc.work);
5207 EXPORT_SYMBOL_GPL(work_on_cpu);
5210 * work_on_cpu_safe - run a function in thread context on a particular cpu
5211 * @cpu: the cpu to run on
5212 * @fn: the function to run
5213 * @arg: the function argument
5215 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5216 * any locks which would prevent @fn from completing.
5218 * Return: The value @fn returns.
5220 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5225 if (cpu_online(cpu))
5226 ret = work_on_cpu(cpu, fn, arg);
5230 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5231 #endif /* CONFIG_SMP */
5233 #ifdef CONFIG_FREEZER
5236 * freeze_workqueues_begin - begin freezing workqueues
5238 * Start freezing workqueues. After this function returns, all freezable
5239 * workqueues will queue new works to their inactive_works list instead of
5243 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5245 void freeze_workqueues_begin(void)
5247 struct workqueue_struct *wq;
5248 struct pool_workqueue *pwq;
5250 mutex_lock(&wq_pool_mutex);
5252 WARN_ON_ONCE(workqueue_freezing);
5253 workqueue_freezing = true;
5255 list_for_each_entry(wq, &workqueues, list) {
5256 mutex_lock(&wq->mutex);
5257 for_each_pwq(pwq, wq)
5258 pwq_adjust_max_active(pwq);
5259 mutex_unlock(&wq->mutex);
5262 mutex_unlock(&wq_pool_mutex);
5266 * freeze_workqueues_busy - are freezable workqueues still busy?
5268 * Check whether freezing is complete. This function must be called
5269 * between freeze_workqueues_begin() and thaw_workqueues().
5272 * Grabs and releases wq_pool_mutex.
5275 * %true if some freezable workqueues are still busy. %false if freezing
5278 bool freeze_workqueues_busy(void)
5281 struct workqueue_struct *wq;
5282 struct pool_workqueue *pwq;
5284 mutex_lock(&wq_pool_mutex);
5286 WARN_ON_ONCE(!workqueue_freezing);
5288 list_for_each_entry(wq, &workqueues, list) {
5289 if (!(wq->flags & WQ_FREEZABLE))
5292 * nr_active is monotonically decreasing. It's safe
5293 * to peek without lock.
5296 for_each_pwq(pwq, wq) {
5297 WARN_ON_ONCE(pwq->nr_active < 0);
5298 if (pwq->nr_active) {
5307 mutex_unlock(&wq_pool_mutex);
5312 * thaw_workqueues - thaw workqueues
5314 * Thaw workqueues. Normal queueing is restored and all collected
5315 * frozen works are transferred to their respective pool worklists.
5318 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5320 void thaw_workqueues(void)
5322 struct workqueue_struct *wq;
5323 struct pool_workqueue *pwq;
5325 mutex_lock(&wq_pool_mutex);
5327 if (!workqueue_freezing)
5330 workqueue_freezing = false;
5332 /* restore max_active and repopulate worklist */
5333 list_for_each_entry(wq, &workqueues, list) {
5334 mutex_lock(&wq->mutex);
5335 for_each_pwq(pwq, wq)
5336 pwq_adjust_max_active(pwq);
5337 mutex_unlock(&wq->mutex);
5341 mutex_unlock(&wq_pool_mutex);
5343 #endif /* CONFIG_FREEZER */
5345 static int workqueue_apply_unbound_cpumask(void)
5349 struct workqueue_struct *wq;
5350 struct apply_wqattrs_ctx *ctx, *n;
5352 lockdep_assert_held(&wq_pool_mutex);
5354 list_for_each_entry(wq, &workqueues, list) {
5355 if (!(wq->flags & WQ_UNBOUND))
5357 /* creating multiple pwqs breaks ordering guarantee */
5358 if (wq->flags & __WQ_ORDERED)
5361 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5367 list_add_tail(&ctx->list, &ctxs);
5370 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5372 apply_wqattrs_commit(ctx);
5373 apply_wqattrs_cleanup(ctx);
5380 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5381 * @cpumask: the cpumask to set
5383 * The low-level workqueues cpumask is a global cpumask that limits
5384 * the affinity of all unbound workqueues. This function check the @cpumask
5385 * and apply it to all unbound workqueues and updates all pwqs of them.
5387 * Return: 0 - Success
5388 * -EINVAL - Invalid @cpumask
5389 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5391 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5394 cpumask_var_t saved_cpumask;
5397 * Not excluding isolated cpus on purpose.
5398 * If the user wishes to include them, we allow that.
5400 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5401 if (!cpumask_empty(cpumask)) {
5402 apply_wqattrs_lock();
5403 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5408 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) {
5413 /* save the old wq_unbound_cpumask. */
5414 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5416 /* update wq_unbound_cpumask at first and apply it to wqs. */
5417 cpumask_copy(wq_unbound_cpumask, cpumask);
5418 ret = workqueue_apply_unbound_cpumask();
5420 /* restore the wq_unbound_cpumask when failed. */
5422 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5424 free_cpumask_var(saved_cpumask);
5426 apply_wqattrs_unlock();
5434 * Workqueues with WQ_SYSFS flag set is visible to userland via
5435 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5436 * following attributes.
5438 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5439 * max_active RW int : maximum number of in-flight work items
5441 * Unbound workqueues have the following extra attributes.
5443 * pool_ids RO int : the associated pool IDs for each node
5444 * nice RW int : nice value of the workers
5445 * cpumask RW mask : bitmask of allowed CPUs for the workers
5446 * numa RW bool : whether enable NUMA affinity
5449 struct workqueue_struct *wq;
5453 static struct workqueue_struct *dev_to_wq(struct device *dev)
5455 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5460 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5463 struct workqueue_struct *wq = dev_to_wq(dev);
5465 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5467 static DEVICE_ATTR_RO(per_cpu);
5469 static ssize_t max_active_show(struct device *dev,
5470 struct device_attribute *attr, char *buf)
5472 struct workqueue_struct *wq = dev_to_wq(dev);
5474 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5477 static ssize_t max_active_store(struct device *dev,
5478 struct device_attribute *attr, const char *buf,
5481 struct workqueue_struct *wq = dev_to_wq(dev);
5484 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5487 workqueue_set_max_active(wq, val);
5490 static DEVICE_ATTR_RW(max_active);
5492 static struct attribute *wq_sysfs_attrs[] = {
5493 &dev_attr_per_cpu.attr,
5494 &dev_attr_max_active.attr,
5497 ATTRIBUTE_GROUPS(wq_sysfs);
5499 static ssize_t wq_pool_ids_show(struct device *dev,
5500 struct device_attribute *attr, char *buf)
5502 struct workqueue_struct *wq = dev_to_wq(dev);
5503 const char *delim = "";
5504 int node, written = 0;
5508 for_each_node(node) {
5509 written += scnprintf(buf + written, PAGE_SIZE - written,
5510 "%s%d:%d", delim, node,
5511 unbound_pwq_by_node(wq, node)->pool->id);
5514 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5521 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5524 struct workqueue_struct *wq = dev_to_wq(dev);
5527 mutex_lock(&wq->mutex);
5528 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5529 mutex_unlock(&wq->mutex);
5534 /* prepare workqueue_attrs for sysfs store operations */
5535 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5537 struct workqueue_attrs *attrs;
5539 lockdep_assert_held(&wq_pool_mutex);
5541 attrs = alloc_workqueue_attrs();
5545 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5549 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5550 const char *buf, size_t count)
5552 struct workqueue_struct *wq = dev_to_wq(dev);
5553 struct workqueue_attrs *attrs;
5556 apply_wqattrs_lock();
5558 attrs = wq_sysfs_prep_attrs(wq);
5562 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5563 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5564 ret = apply_workqueue_attrs_locked(wq, attrs);
5569 apply_wqattrs_unlock();
5570 free_workqueue_attrs(attrs);
5571 return ret ?: count;
5574 static ssize_t wq_cpumask_show(struct device *dev,
5575 struct device_attribute *attr, char *buf)
5577 struct workqueue_struct *wq = dev_to_wq(dev);
5580 mutex_lock(&wq->mutex);
5581 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5582 cpumask_pr_args(wq->unbound_attrs->cpumask));
5583 mutex_unlock(&wq->mutex);
5587 static ssize_t wq_cpumask_store(struct device *dev,
5588 struct device_attribute *attr,
5589 const char *buf, size_t count)
5591 struct workqueue_struct *wq = dev_to_wq(dev);
5592 struct workqueue_attrs *attrs;
5595 apply_wqattrs_lock();
5597 attrs = wq_sysfs_prep_attrs(wq);
5601 ret = cpumask_parse(buf, attrs->cpumask);
5603 ret = apply_workqueue_attrs_locked(wq, attrs);
5606 apply_wqattrs_unlock();
5607 free_workqueue_attrs(attrs);
5608 return ret ?: count;
5611 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5614 struct workqueue_struct *wq = dev_to_wq(dev);
5617 mutex_lock(&wq->mutex);
5618 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5619 !wq->unbound_attrs->no_numa);
5620 mutex_unlock(&wq->mutex);
5625 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5626 const char *buf, size_t count)
5628 struct workqueue_struct *wq = dev_to_wq(dev);
5629 struct workqueue_attrs *attrs;
5630 int v, ret = -ENOMEM;
5632 apply_wqattrs_lock();
5634 attrs = wq_sysfs_prep_attrs(wq);
5639 if (sscanf(buf, "%d", &v) == 1) {
5640 attrs->no_numa = !v;
5641 ret = apply_workqueue_attrs_locked(wq, attrs);
5645 apply_wqattrs_unlock();
5646 free_workqueue_attrs(attrs);
5647 return ret ?: count;
5650 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5651 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5652 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5653 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5654 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5658 static struct bus_type wq_subsys = {
5659 .name = "workqueue",
5660 .dev_groups = wq_sysfs_groups,
5663 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5664 struct device_attribute *attr, char *buf)
5668 mutex_lock(&wq_pool_mutex);
5669 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5670 cpumask_pr_args(wq_unbound_cpumask));
5671 mutex_unlock(&wq_pool_mutex);
5676 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5677 struct device_attribute *attr, const char *buf, size_t count)
5679 cpumask_var_t cpumask;
5682 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5685 ret = cpumask_parse(buf, cpumask);
5687 ret = workqueue_set_unbound_cpumask(cpumask);
5689 free_cpumask_var(cpumask);
5690 return ret ? ret : count;
5693 static struct device_attribute wq_sysfs_cpumask_attr =
5694 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5695 wq_unbound_cpumask_store);
5697 static int __init wq_sysfs_init(void)
5701 err = subsys_virtual_register(&wq_subsys, NULL);
5705 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5707 core_initcall(wq_sysfs_init);
5709 static void wq_device_release(struct device *dev)
5711 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5717 * workqueue_sysfs_register - make a workqueue visible in sysfs
5718 * @wq: the workqueue to register
5720 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5721 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5722 * which is the preferred method.
5724 * Workqueue user should use this function directly iff it wants to apply
5725 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5726 * apply_workqueue_attrs() may race against userland updating the
5729 * Return: 0 on success, -errno on failure.
5731 int workqueue_sysfs_register(struct workqueue_struct *wq)
5733 struct wq_device *wq_dev;
5737 * Adjusting max_active or creating new pwqs by applying
5738 * attributes breaks ordering guarantee. Disallow exposing ordered
5741 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5744 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5749 wq_dev->dev.bus = &wq_subsys;
5750 wq_dev->dev.release = wq_device_release;
5751 dev_set_name(&wq_dev->dev, "%s", wq->name);
5754 * unbound_attrs are created separately. Suppress uevent until
5755 * everything is ready.
5757 dev_set_uevent_suppress(&wq_dev->dev, true);
5759 ret = device_register(&wq_dev->dev);
5761 put_device(&wq_dev->dev);
5766 if (wq->flags & WQ_UNBOUND) {
5767 struct device_attribute *attr;
5769 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5770 ret = device_create_file(&wq_dev->dev, attr);
5772 device_unregister(&wq_dev->dev);
5779 dev_set_uevent_suppress(&wq_dev->dev, false);
5780 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5785 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5786 * @wq: the workqueue to unregister
5788 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5790 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5792 struct wq_device *wq_dev = wq->wq_dev;
5798 device_unregister(&wq_dev->dev);
5800 #else /* CONFIG_SYSFS */
5801 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5802 #endif /* CONFIG_SYSFS */
5805 * Workqueue watchdog.
5807 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5808 * flush dependency, a concurrency managed work item which stays RUNNING
5809 * indefinitely. Workqueue stalls can be very difficult to debug as the
5810 * usual warning mechanisms don't trigger and internal workqueue state is
5813 * Workqueue watchdog monitors all worker pools periodically and dumps
5814 * state if some pools failed to make forward progress for a while where
5815 * forward progress is defined as the first item on ->worklist changing.
5817 * This mechanism is controlled through the kernel parameter
5818 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5819 * corresponding sysfs parameter file.
5821 #ifdef CONFIG_WQ_WATCHDOG
5823 static unsigned long wq_watchdog_thresh = 30;
5824 static struct timer_list wq_watchdog_timer;
5826 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5827 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5829 static void wq_watchdog_reset_touched(void)
5833 wq_watchdog_touched = jiffies;
5834 for_each_possible_cpu(cpu)
5835 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5838 static void wq_watchdog_timer_fn(struct timer_list *unused)
5840 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5841 bool lockup_detected = false;
5842 unsigned long now = jiffies;
5843 struct worker_pool *pool;
5851 for_each_pool(pool, pi) {
5852 unsigned long pool_ts, touched, ts;
5854 if (list_empty(&pool->worklist))
5858 * If a virtual machine is stopped by the host it can look to
5859 * the watchdog like a stall.
5861 kvm_check_and_clear_guest_paused();
5863 /* get the latest of pool and touched timestamps */
5865 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
5867 touched = READ_ONCE(wq_watchdog_touched);
5868 pool_ts = READ_ONCE(pool->watchdog_ts);
5870 if (time_after(pool_ts, touched))
5876 if (time_after(now, ts + thresh)) {
5877 lockup_detected = true;
5878 pr_emerg("BUG: workqueue lockup - pool");
5879 pr_cont_pool_info(pool);
5880 pr_cont(" stuck for %us!\n",
5881 jiffies_to_msecs(now - pool_ts) / 1000);
5887 if (lockup_detected)
5888 show_all_workqueues();
5890 wq_watchdog_reset_touched();
5891 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5894 notrace void wq_watchdog_touch(int cpu)
5897 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5899 wq_watchdog_touched = jiffies;
5902 static void wq_watchdog_set_thresh(unsigned long thresh)
5904 wq_watchdog_thresh = 0;
5905 del_timer_sync(&wq_watchdog_timer);
5908 wq_watchdog_thresh = thresh;
5909 wq_watchdog_reset_touched();
5910 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5914 static int wq_watchdog_param_set_thresh(const char *val,
5915 const struct kernel_param *kp)
5917 unsigned long thresh;
5920 ret = kstrtoul(val, 0, &thresh);
5925 wq_watchdog_set_thresh(thresh);
5927 wq_watchdog_thresh = thresh;
5932 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5933 .set = wq_watchdog_param_set_thresh,
5934 .get = param_get_ulong,
5937 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5940 static void wq_watchdog_init(void)
5942 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5943 wq_watchdog_set_thresh(wq_watchdog_thresh);
5946 #else /* CONFIG_WQ_WATCHDOG */
5948 static inline void wq_watchdog_init(void) { }
5950 #endif /* CONFIG_WQ_WATCHDOG */
5952 static void __init wq_numa_init(void)
5957 if (num_possible_nodes() <= 1)
5960 if (wq_disable_numa) {
5961 pr_info("workqueue: NUMA affinity support disabled\n");
5965 for_each_possible_cpu(cpu) {
5966 if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) {
5967 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5972 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5973 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5976 * We want masks of possible CPUs of each node which isn't readily
5977 * available. Build one from cpu_to_node() which should have been
5978 * fully initialized by now.
5980 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5984 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5985 node_online(node) ? node : NUMA_NO_NODE));
5987 for_each_possible_cpu(cpu) {
5988 node = cpu_to_node(cpu);
5989 cpumask_set_cpu(cpu, tbl[node]);
5992 wq_numa_possible_cpumask = tbl;
5993 wq_numa_enabled = true;
5997 * workqueue_init_early - early init for workqueue subsystem
5999 * This is the first half of two-staged workqueue subsystem initialization
6000 * and invoked as soon as the bare basics - memory allocation, cpumasks and
6001 * idr are up. It sets up all the data structures and system workqueues
6002 * and allows early boot code to create workqueues and queue/cancel work
6003 * items. Actual work item execution starts only after kthreads can be
6004 * created and scheduled right before early initcalls.
6006 void __init workqueue_init_early(void)
6008 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6011 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6013 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6014 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ));
6015 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN));
6017 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6019 /* initialize CPU pools */
6020 for_each_possible_cpu(cpu) {
6021 struct worker_pool *pool;
6024 for_each_cpu_worker_pool(pool, cpu) {
6025 BUG_ON(init_worker_pool(pool));
6027 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6028 pool->attrs->nice = std_nice[i++];
6029 pool->node = cpu_to_node(cpu);
6032 mutex_lock(&wq_pool_mutex);
6033 BUG_ON(worker_pool_assign_id(pool));
6034 mutex_unlock(&wq_pool_mutex);
6038 /* create default unbound and ordered wq attrs */
6039 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6040 struct workqueue_attrs *attrs;
6042 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6043 attrs->nice = std_nice[i];
6044 unbound_std_wq_attrs[i] = attrs;
6047 * An ordered wq should have only one pwq as ordering is
6048 * guaranteed by max_active which is enforced by pwqs.
6049 * Turn off NUMA so that dfl_pwq is used for all nodes.
6051 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6052 attrs->nice = std_nice[i];
6053 attrs->no_numa = true;
6054 ordered_wq_attrs[i] = attrs;
6057 system_wq = alloc_workqueue("events", 0, 0);
6058 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6059 system_long_wq = alloc_workqueue("events_long", 0, 0);
6060 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6061 WQ_UNBOUND_MAX_ACTIVE);
6062 system_freezable_wq = alloc_workqueue("events_freezable",
6064 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6065 WQ_POWER_EFFICIENT, 0);
6066 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6067 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6069 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6070 !system_unbound_wq || !system_freezable_wq ||
6071 !system_power_efficient_wq ||
6072 !system_freezable_power_efficient_wq);
6076 * workqueue_init - bring workqueue subsystem fully online
6078 * This is the latter half of two-staged workqueue subsystem initialization
6079 * and invoked as soon as kthreads can be created and scheduled.
6080 * Workqueues have been created and work items queued on them, but there
6081 * are no kworkers executing the work items yet. Populate the worker pools
6082 * with the initial workers and enable future kworker creations.
6084 void __init workqueue_init(void)
6086 struct workqueue_struct *wq;
6087 struct worker_pool *pool;
6091 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6092 * CPU to node mapping may not be available that early on some
6093 * archs such as power and arm64. As per-cpu pools created
6094 * previously could be missing node hint and unbound pools NUMA
6095 * affinity, fix them up.
6097 * Also, while iterating workqueues, create rescuers if requested.
6101 mutex_lock(&wq_pool_mutex);
6103 for_each_possible_cpu(cpu) {
6104 for_each_cpu_worker_pool(pool, cpu) {
6105 pool->node = cpu_to_node(cpu);
6109 list_for_each_entry(wq, &workqueues, list) {
6110 wq_update_unbound_numa(wq, smp_processor_id(), true);
6111 WARN(init_rescuer(wq),
6112 "workqueue: failed to create early rescuer for %s",
6116 mutex_unlock(&wq_pool_mutex);
6118 /* create the initial workers */
6119 for_each_online_cpu(cpu) {
6120 for_each_cpu_worker_pool(pool, cpu) {
6121 pool->flags &= ~POOL_DISASSOCIATED;
6122 BUG_ON(!create_worker(pool));
6126 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6127 BUG_ON(!create_worker(pool));