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/sched/debug.h>
53 #include <linux/nmi.h>
54 #include <linux/kvm_para.h>
55 #include <linux/delay.h>
57 #include "workqueue_internal.h"
59 enum worker_pool_flags {
63 * A bound pool is either associated or disassociated with its CPU.
64 * While associated (!DISASSOCIATED), all workers are bound to the
65 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * While DISASSOCIATED, the cpu may be offline and all workers have
69 * %WORKER_UNBOUND set and concurrency management disabled, and may
70 * be executing on any CPU. The pool behaves as an unbound one.
72 * Note that DISASSOCIATED should be flipped only while holding
73 * wq_pool_attach_mutex to avoid changing binding state while
74 * worker_attach_to_pool() is in progress.
76 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
77 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
82 WORKER_DIE = 1 << 1, /* die die die */
83 WORKER_IDLE = 1 << 2, /* is idle */
84 WORKER_PREP = 1 << 3, /* preparing to run works */
85 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
86 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
87 WORKER_REBOUND = 1 << 8, /* worker was rebound */
89 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
90 WORKER_UNBOUND | WORKER_REBOUND,
93 enum wq_internal_consts {
94 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
96 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
97 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
99 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
100 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
102 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
103 /* call for help after 10ms
105 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
106 CREATE_COOLDOWN = HZ, /* time to breath after fail */
109 * Rescue workers are used only on emergencies and shared by
110 * all cpus. Give MIN_NICE.
112 RESCUER_NICE_LEVEL = MIN_NICE,
113 HIGHPRI_NICE_LEVEL = MIN_NICE,
119 * Structure fields follow one of the following exclusion rules.
121 * I: Modifiable by initialization/destruction paths and read-only for
124 * P: Preemption protected. Disabling preemption is enough and should
125 * only be modified and accessed from the local cpu.
127 * L: pool->lock protected. Access with pool->lock held.
129 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
132 * K: Only modified by worker while holding pool->lock. Can be safely read by
133 * self, while holding pool->lock or from IRQ context if %current is the
136 * S: Only modified by worker self.
138 * A: wq_pool_attach_mutex protected.
140 * PL: wq_pool_mutex protected.
142 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
144 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
146 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
149 * WQ: wq->mutex protected.
151 * WR: wq->mutex protected for writes. RCU protected for reads.
153 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
154 * with READ_ONCE() without locking.
156 * MD: wq_mayday_lock protected.
158 * WD: Used internally by the watchdog.
161 /* struct worker is defined in workqueue_internal.h */
164 raw_spinlock_t lock; /* the pool lock */
165 int cpu; /* I: the associated cpu */
166 int node; /* I: the associated node ID */
167 int id; /* I: pool ID */
168 unsigned int flags; /* L: flags */
170 unsigned long watchdog_ts; /* L: watchdog timestamp */
171 bool cpu_stall; /* WD: stalled cpu bound pool */
174 * The counter is incremented in a process context on the associated CPU
175 * w/ preemption disabled, and decremented or reset in the same context
176 * but w/ pool->lock held. The readers grab pool->lock and are
177 * guaranteed to see if the counter reached zero.
181 struct list_head worklist; /* L: list of pending works */
183 int nr_workers; /* L: total number of workers */
184 int nr_idle; /* L: currently idle workers */
186 struct list_head idle_list; /* L: list of idle workers */
187 struct timer_list idle_timer; /* L: worker idle timeout */
188 struct work_struct idle_cull_work; /* L: worker idle cleanup */
190 struct timer_list mayday_timer; /* L: SOS timer for workers */
192 /* a workers is either on busy_hash or idle_list, or the manager */
193 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
194 /* L: hash of busy workers */
196 struct worker *manager; /* L: purely informational */
197 struct list_head workers; /* A: attached workers */
198 struct list_head dying_workers; /* A: workers about to die */
199 struct completion *detach_completion; /* all workers detached */
201 struct ida worker_ida; /* worker IDs for task name */
203 struct workqueue_attrs *attrs; /* I: worker attributes */
204 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
205 int refcnt; /* PL: refcnt for unbound pools */
208 * Destruction of pool is RCU protected to allow dereferences
209 * from get_work_pool().
215 * Per-pool_workqueue statistics. These can be monitored using
216 * tools/workqueue/wq_monitor.py.
218 enum pool_workqueue_stats {
219 PWQ_STAT_STARTED, /* work items started execution */
220 PWQ_STAT_COMPLETED, /* work items completed execution */
221 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
222 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
223 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
224 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
225 PWQ_STAT_MAYDAY, /* maydays to rescuer */
226 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
232 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
233 * of work_struct->data are used for flags and the remaining high bits
234 * point to the pwq; thus, pwqs need to be aligned at two's power of the
235 * number of flag bits.
237 struct pool_workqueue {
238 struct worker_pool *pool; /* I: the associated pool */
239 struct workqueue_struct *wq; /* I: the owning workqueue */
240 int work_color; /* L: current color */
241 int flush_color; /* L: flushing color */
242 int refcnt; /* L: reference count */
243 int nr_in_flight[WORK_NR_COLORS];
244 /* L: nr of in_flight works */
247 * nr_active management and WORK_STRUCT_INACTIVE:
249 * When pwq->nr_active >= max_active, new work item is queued to
250 * pwq->inactive_works instead of pool->worklist and marked with
251 * WORK_STRUCT_INACTIVE.
253 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
254 * nr_active and all work items in pwq->inactive_works are marked with
255 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
256 * in pwq->inactive_works. Some of them are ready to run in
257 * pool->worklist or worker->scheduled. Those work itmes are only struct
258 * wq_barrier which is used for flush_work() and should not participate
259 * in nr_active. For non-barrier work item, it is marked with
260 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
262 int nr_active; /* L: nr of active works */
263 struct list_head inactive_works; /* L: inactive works */
264 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
265 struct list_head pwqs_node; /* WR: node on wq->pwqs */
266 struct list_head mayday_node; /* MD: node on wq->maydays */
268 u64 stats[PWQ_NR_STATS];
271 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
272 * and pwq_release_workfn() for details. pool_workqueue itself is also
273 * RCU protected so that the first pwq can be determined without
274 * grabbing wq->mutex.
276 struct kthread_work release_work;
278 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
281 * Structure used to wait for workqueue flush.
284 struct list_head list; /* WQ: list of flushers */
285 int flush_color; /* WQ: flush color waiting for */
286 struct completion done; /* flush completion */
292 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
293 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
294 * As sharing a single nr_active across multiple sockets can be very expensive,
295 * the counting and enforcement is per NUMA node.
297 * The following struct is used to enforce per-node max_active. When a pwq wants
298 * to start executing a work item, it should increment ->nr using
299 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
300 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
301 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
304 struct wq_node_nr_active {
305 int max; /* per-node max_active */
306 atomic_t nr; /* per-node nr_active */
307 raw_spinlock_t lock; /* nests inside pool locks */
308 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
312 * The externally visible workqueue. It relays the issued work items to
313 * the appropriate worker_pool through its pool_workqueues.
315 struct workqueue_struct {
316 struct list_head pwqs; /* WR: all pwqs of this wq */
317 struct list_head list; /* PR: list of all workqueues */
319 struct mutex mutex; /* protects this wq */
320 int work_color; /* WQ: current work color */
321 int flush_color; /* WQ: current flush color */
322 atomic_t nr_pwqs_to_flush; /* flush in progress */
323 struct wq_flusher *first_flusher; /* WQ: first flusher */
324 struct list_head flusher_queue; /* WQ: flush waiters */
325 struct list_head flusher_overflow; /* WQ: flush overflow list */
327 struct list_head maydays; /* MD: pwqs requesting rescue */
328 struct worker *rescuer; /* MD: rescue worker */
330 int nr_drainers; /* WQ: drain in progress */
332 /* See alloc_workqueue() function comment for info on min/max_active */
333 int max_active; /* WO: max active works */
334 int min_active; /* WO: min active works */
335 int saved_max_active; /* WQ: saved max_active */
336 int saved_min_active; /* WQ: saved min_active */
338 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
339 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
342 struct wq_device *wq_dev; /* I: for sysfs interface */
344 #ifdef CONFIG_LOCKDEP
346 struct lock_class_key key;
347 struct lockdep_map lockdep_map;
349 char name[WQ_NAME_LEN]; /* I: workqueue name */
352 * Destruction of workqueue_struct is RCU protected to allow walking
353 * the workqueues list without grabbing wq_pool_mutex.
354 * This is used to dump all workqueues from sysrq.
358 /* hot fields used during command issue, aligned to cacheline */
359 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
360 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
361 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
364 static struct kmem_cache *pwq_cache;
367 * Each pod type describes how CPUs should be grouped for unbound workqueues.
368 * See the comment above workqueue_attrs->affn_scope.
371 int nr_pods; /* number of pods */
372 cpumask_var_t *pod_cpus; /* pod -> cpus */
373 int *pod_node; /* pod -> node */
374 int *cpu_pod; /* cpu -> pod */
377 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
378 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
380 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
381 [WQ_AFFN_DFL] = "default",
382 [WQ_AFFN_CPU] = "cpu",
383 [WQ_AFFN_SMT] = "smt",
384 [WQ_AFFN_CACHE] = "cache",
385 [WQ_AFFN_NUMA] = "numa",
386 [WQ_AFFN_SYSTEM] = "system",
390 * Per-cpu work items which run for longer than the following threshold are
391 * automatically considered CPU intensive and excluded from concurrency
392 * management to prevent them from noticeably delaying other per-cpu work items.
393 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
394 * The actual value is initialized in wq_cpu_intensive_thresh_init().
396 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
397 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
399 /* see the comment above the definition of WQ_POWER_EFFICIENT */
400 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
401 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
403 static bool wq_online; /* can kworkers be created yet? */
405 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
406 static struct workqueue_attrs *wq_update_pod_attrs_buf;
408 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
409 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
410 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
411 /* wait for manager to go away */
412 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
414 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
415 static bool workqueue_freezing; /* PL: have wqs started freezing? */
417 /* PL&A: allowable cpus for unbound wqs and work items */
418 static cpumask_var_t wq_unbound_cpumask;
420 /* PL: user requested unbound cpumask via sysfs */
421 static cpumask_var_t wq_requested_unbound_cpumask;
423 /* PL: isolated cpumask to be excluded from unbound cpumask */
424 static cpumask_var_t wq_isolated_cpumask;
426 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
427 static struct cpumask wq_cmdline_cpumask __initdata;
429 /* CPU where unbound work was last round robin scheduled from this CPU */
430 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
433 * Local execution of unbound work items is no longer guaranteed. The
434 * following always forces round-robin CPU selection on unbound work items
435 * to uncover usages which depend on it.
437 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
438 static bool wq_debug_force_rr_cpu = true;
440 static bool wq_debug_force_rr_cpu = false;
442 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
444 /* the per-cpu worker pools */
445 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
447 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
449 /* PL: hash of all unbound pools keyed by pool->attrs */
450 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
452 /* I: attributes used when instantiating standard unbound pools on demand */
453 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
455 /* I: attributes used when instantiating ordered pools on demand */
456 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
459 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
460 * process context while holding a pool lock. Bounce to a dedicated kthread
461 * worker to avoid A-A deadlocks.
463 static struct kthread_worker *pwq_release_worker __ro_after_init;
465 struct workqueue_struct *system_wq __ro_after_init;
466 EXPORT_SYMBOL(system_wq);
467 struct workqueue_struct *system_highpri_wq __ro_after_init;
468 EXPORT_SYMBOL_GPL(system_highpri_wq);
469 struct workqueue_struct *system_long_wq __ro_after_init;
470 EXPORT_SYMBOL_GPL(system_long_wq);
471 struct workqueue_struct *system_unbound_wq __ro_after_init;
472 EXPORT_SYMBOL_GPL(system_unbound_wq);
473 struct workqueue_struct *system_freezable_wq __ro_after_init;
474 EXPORT_SYMBOL_GPL(system_freezable_wq);
475 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
476 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
477 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
478 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
480 static int worker_thread(void *__worker);
481 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
482 static void show_pwq(struct pool_workqueue *pwq);
483 static void show_one_worker_pool(struct worker_pool *pool);
485 #define CREATE_TRACE_POINTS
486 #include <trace/events/workqueue.h>
488 #define assert_rcu_or_pool_mutex() \
489 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
490 !lockdep_is_held(&wq_pool_mutex), \
491 "RCU or wq_pool_mutex should be held")
493 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
494 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
495 !lockdep_is_held(&wq->mutex) && \
496 !lockdep_is_held(&wq_pool_mutex), \
497 "RCU, wq->mutex or wq_pool_mutex should be held")
499 #define for_each_cpu_worker_pool(pool, cpu) \
500 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
501 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
505 * for_each_pool - iterate through all worker_pools in the system
506 * @pool: iteration cursor
507 * @pi: integer used for iteration
509 * This must be called either with wq_pool_mutex held or RCU read
510 * locked. If the pool needs to be used beyond the locking in effect, the
511 * caller is responsible for guaranteeing that the pool stays online.
513 * The if/else clause exists only for the lockdep assertion and can be
516 #define for_each_pool(pool, pi) \
517 idr_for_each_entry(&worker_pool_idr, pool, pi) \
518 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
522 * for_each_pool_worker - iterate through all workers of a worker_pool
523 * @worker: iteration cursor
524 * @pool: worker_pool to iterate workers of
526 * This must be called with wq_pool_attach_mutex.
528 * The if/else clause exists only for the lockdep assertion and can be
531 #define for_each_pool_worker(worker, pool) \
532 list_for_each_entry((worker), &(pool)->workers, node) \
533 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
537 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
538 * @pwq: iteration cursor
539 * @wq: the target workqueue
541 * This must be called either with wq->mutex held or RCU read locked.
542 * If the pwq needs to be used beyond the locking in effect, the caller is
543 * responsible for guaranteeing that the pwq stays online.
545 * The if/else clause exists only for the lockdep assertion and can be
548 #define for_each_pwq(pwq, wq) \
549 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
550 lockdep_is_held(&(wq->mutex)))
552 #ifdef CONFIG_DEBUG_OBJECTS_WORK
554 static const struct debug_obj_descr work_debug_descr;
556 static void *work_debug_hint(void *addr)
558 return ((struct work_struct *) addr)->func;
561 static bool work_is_static_object(void *addr)
563 struct work_struct *work = addr;
565 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
569 * fixup_init is called when:
570 * - an active object is initialized
572 static bool work_fixup_init(void *addr, enum debug_obj_state state)
574 struct work_struct *work = addr;
577 case ODEBUG_STATE_ACTIVE:
578 cancel_work_sync(work);
579 debug_object_init(work, &work_debug_descr);
587 * fixup_free is called when:
588 * - an active object is freed
590 static bool work_fixup_free(void *addr, enum debug_obj_state state)
592 struct work_struct *work = addr;
595 case ODEBUG_STATE_ACTIVE:
596 cancel_work_sync(work);
597 debug_object_free(work, &work_debug_descr);
604 static const struct debug_obj_descr work_debug_descr = {
605 .name = "work_struct",
606 .debug_hint = work_debug_hint,
607 .is_static_object = work_is_static_object,
608 .fixup_init = work_fixup_init,
609 .fixup_free = work_fixup_free,
612 static inline void debug_work_activate(struct work_struct *work)
614 debug_object_activate(work, &work_debug_descr);
617 static inline void debug_work_deactivate(struct work_struct *work)
619 debug_object_deactivate(work, &work_debug_descr);
622 void __init_work(struct work_struct *work, int onstack)
625 debug_object_init_on_stack(work, &work_debug_descr);
627 debug_object_init(work, &work_debug_descr);
629 EXPORT_SYMBOL_GPL(__init_work);
631 void destroy_work_on_stack(struct work_struct *work)
633 debug_object_free(work, &work_debug_descr);
635 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
637 void destroy_delayed_work_on_stack(struct delayed_work *work)
639 destroy_timer_on_stack(&work->timer);
640 debug_object_free(&work->work, &work_debug_descr);
642 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
645 static inline void debug_work_activate(struct work_struct *work) { }
646 static inline void debug_work_deactivate(struct work_struct *work) { }
650 * worker_pool_assign_id - allocate ID and assign it to @pool
651 * @pool: the pool pointer of interest
653 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
654 * successfully, -errno on failure.
656 static int worker_pool_assign_id(struct worker_pool *pool)
660 lockdep_assert_held(&wq_pool_mutex);
662 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
671 static struct pool_workqueue __rcu **
672 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
675 return per_cpu_ptr(wq->cpu_pwq, cpu);
680 /* @cpu < 0 for dfl_pwq */
681 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
683 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
684 lockdep_is_held(&wq_pool_mutex) ||
685 lockdep_is_held(&wq->mutex));
689 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
690 * @wq: workqueue of interest
692 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
693 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
694 * default pwq is always mapped to the pool with the current effective cpumask.
696 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
698 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
701 static unsigned int work_color_to_flags(int color)
703 return color << WORK_STRUCT_COLOR_SHIFT;
706 static int get_work_color(unsigned long work_data)
708 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
709 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
712 static int work_next_color(int color)
714 return (color + 1) % WORK_NR_COLORS;
718 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
719 * contain the pointer to the queued pwq. Once execution starts, the flag
720 * is cleared and the high bits contain OFFQ flags and pool ID.
722 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
723 * and clear_work_data() can be used to set the pwq, pool or clear
724 * work->data. These functions should only be called while the work is
725 * owned - ie. while the PENDING bit is set.
727 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
728 * corresponding to a work. Pool is available once the work has been
729 * queued anywhere after initialization until it is sync canceled. pwq is
730 * available only while the work item is queued.
732 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
733 * canceled. While being canceled, a work item may have its PENDING set
734 * but stay off timer and worklist for arbitrarily long and nobody should
735 * try to steal the PENDING bit.
737 static inline void set_work_data(struct work_struct *work, unsigned long data,
740 WARN_ON_ONCE(!work_pending(work));
741 atomic_long_set(&work->data, data | flags | work_static(work));
744 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
745 unsigned long extra_flags)
747 set_work_data(work, (unsigned long)pwq,
748 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
751 static void set_work_pool_and_keep_pending(struct work_struct *work,
754 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
755 WORK_STRUCT_PENDING);
758 static void set_work_pool_and_clear_pending(struct work_struct *work,
762 * The following wmb is paired with the implied mb in
763 * test_and_set_bit(PENDING) and ensures all updates to @work made
764 * here are visible to and precede any updates by the next PENDING
768 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
770 * The following mb guarantees that previous clear of a PENDING bit
771 * will not be reordered with any speculative LOADS or STORES from
772 * work->current_func, which is executed afterwards. This possible
773 * reordering can lead to a missed execution on attempt to queue
774 * the same @work. E.g. consider this case:
777 * ---------------------------- --------------------------------
779 * 1 STORE event_indicated
780 * 2 queue_work_on() {
781 * 3 test_and_set_bit(PENDING)
782 * 4 } set_..._and_clear_pending() {
783 * 5 set_work_data() # clear bit
785 * 7 work->current_func() {
786 * 8 LOAD event_indicated
789 * Without an explicit full barrier speculative LOAD on line 8 can
790 * be executed before CPU#0 does STORE on line 1. If that happens,
791 * CPU#0 observes the PENDING bit is still set and new execution of
792 * a @work is not queued in a hope, that CPU#1 will eventually
793 * finish the queued @work. Meanwhile CPU#1 does not see
794 * event_indicated is set, because speculative LOAD was executed
795 * before actual STORE.
800 static void clear_work_data(struct work_struct *work)
802 smp_wmb(); /* see set_work_pool_and_clear_pending() */
803 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
806 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
808 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
811 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
813 unsigned long data = atomic_long_read(&work->data);
815 if (data & WORK_STRUCT_PWQ)
816 return work_struct_pwq(data);
822 * get_work_pool - return the worker_pool a given work was associated with
823 * @work: the work item of interest
825 * Pools are created and destroyed under wq_pool_mutex, and allows read
826 * access under RCU read lock. As such, this function should be
827 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
829 * All fields of the returned pool are accessible as long as the above
830 * mentioned locking is in effect. If the returned pool needs to be used
831 * beyond the critical section, the caller is responsible for ensuring the
832 * returned pool is and stays online.
834 * Return: The worker_pool @work was last associated with. %NULL if none.
836 static struct worker_pool *get_work_pool(struct work_struct *work)
838 unsigned long data = atomic_long_read(&work->data);
841 assert_rcu_or_pool_mutex();
843 if (data & WORK_STRUCT_PWQ)
844 return work_struct_pwq(data)->pool;
846 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
847 if (pool_id == WORK_OFFQ_POOL_NONE)
850 return idr_find(&worker_pool_idr, pool_id);
854 * get_work_pool_id - return the worker pool ID a given work is associated with
855 * @work: the work item of interest
857 * Return: The worker_pool ID @work was last associated with.
858 * %WORK_OFFQ_POOL_NONE if none.
860 static int get_work_pool_id(struct work_struct *work)
862 unsigned long data = atomic_long_read(&work->data);
864 if (data & WORK_STRUCT_PWQ)
865 return work_struct_pwq(data)->pool->id;
867 return data >> WORK_OFFQ_POOL_SHIFT;
870 static void mark_work_canceling(struct work_struct *work)
872 unsigned long pool_id = get_work_pool_id(work);
874 pool_id <<= WORK_OFFQ_POOL_SHIFT;
875 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
878 static bool work_is_canceling(struct work_struct *work)
880 unsigned long data = atomic_long_read(&work->data);
882 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
886 * Policy functions. These define the policies on how the global worker
887 * pools are managed. Unless noted otherwise, these functions assume that
888 * they're being called with pool->lock held.
892 * Need to wake up a worker? Called from anything but currently
895 * Note that, because unbound workers never contribute to nr_running, this
896 * function will always return %true for unbound pools as long as the
897 * worklist isn't empty.
899 static bool need_more_worker(struct worker_pool *pool)
901 return !list_empty(&pool->worklist) && !pool->nr_running;
904 /* Can I start working? Called from busy but !running workers. */
905 static bool may_start_working(struct worker_pool *pool)
907 return pool->nr_idle;
910 /* Do I need to keep working? Called from currently running workers. */
911 static bool keep_working(struct worker_pool *pool)
913 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
916 /* Do we need a new worker? Called from manager. */
917 static bool need_to_create_worker(struct worker_pool *pool)
919 return need_more_worker(pool) && !may_start_working(pool);
922 /* Do we have too many workers and should some go away? */
923 static bool too_many_workers(struct worker_pool *pool)
925 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
926 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
927 int nr_busy = pool->nr_workers - nr_idle;
929 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
933 * worker_set_flags - set worker flags and adjust nr_running accordingly
935 * @flags: flags to set
937 * Set @flags in @worker->flags and adjust nr_running accordingly.
939 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
941 struct worker_pool *pool = worker->pool;
943 lockdep_assert_held(&pool->lock);
945 /* If transitioning into NOT_RUNNING, adjust nr_running. */
946 if ((flags & WORKER_NOT_RUNNING) &&
947 !(worker->flags & WORKER_NOT_RUNNING)) {
951 worker->flags |= flags;
955 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
957 * @flags: flags to clear
959 * Clear @flags in @worker->flags and adjust nr_running accordingly.
961 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
963 struct worker_pool *pool = worker->pool;
964 unsigned int oflags = worker->flags;
966 lockdep_assert_held(&pool->lock);
968 worker->flags &= ~flags;
971 * If transitioning out of NOT_RUNNING, increment nr_running. Note
972 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
973 * of multiple flags, not a single flag.
975 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
976 if (!(worker->flags & WORKER_NOT_RUNNING))
980 /* Return the first idle worker. Called with pool->lock held. */
981 static struct worker *first_idle_worker(struct worker_pool *pool)
983 if (unlikely(list_empty(&pool->idle_list)))
986 return list_first_entry(&pool->idle_list, struct worker, entry);
990 * worker_enter_idle - enter idle state
991 * @worker: worker which is entering idle state
993 * @worker is entering idle state. Update stats and idle timer if
997 * raw_spin_lock_irq(pool->lock).
999 static void worker_enter_idle(struct worker *worker)
1001 struct worker_pool *pool = worker->pool;
1003 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1004 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1005 (worker->hentry.next || worker->hentry.pprev)))
1008 /* can't use worker_set_flags(), also called from create_worker() */
1009 worker->flags |= WORKER_IDLE;
1011 worker->last_active = jiffies;
1013 /* idle_list is LIFO */
1014 list_add(&worker->entry, &pool->idle_list);
1016 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1017 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1019 /* Sanity check nr_running. */
1020 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1024 * worker_leave_idle - leave idle state
1025 * @worker: worker which is leaving idle state
1027 * @worker is leaving idle state. Update stats.
1030 * raw_spin_lock_irq(pool->lock).
1032 static void worker_leave_idle(struct worker *worker)
1034 struct worker_pool *pool = worker->pool;
1036 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1038 worker_clr_flags(worker, WORKER_IDLE);
1040 list_del_init(&worker->entry);
1044 * find_worker_executing_work - find worker which is executing a work
1045 * @pool: pool of interest
1046 * @work: work to find worker for
1048 * Find a worker which is executing @work on @pool by searching
1049 * @pool->busy_hash which is keyed by the address of @work. For a worker
1050 * to match, its current execution should match the address of @work and
1051 * its work function. This is to avoid unwanted dependency between
1052 * unrelated work executions through a work item being recycled while still
1055 * This is a bit tricky. A work item may be freed once its execution
1056 * starts and nothing prevents the freed area from being recycled for
1057 * another work item. If the same work item address ends up being reused
1058 * before the original execution finishes, workqueue will identify the
1059 * recycled work item as currently executing and make it wait until the
1060 * current execution finishes, introducing an unwanted dependency.
1062 * This function checks the work item address and work function to avoid
1063 * false positives. Note that this isn't complete as one may construct a
1064 * work function which can introduce dependency onto itself through a
1065 * recycled work item. Well, if somebody wants to shoot oneself in the
1066 * foot that badly, there's only so much we can do, and if such deadlock
1067 * actually occurs, it should be easy to locate the culprit work function.
1070 * raw_spin_lock_irq(pool->lock).
1073 * Pointer to worker which is executing @work if found, %NULL
1076 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1077 struct work_struct *work)
1079 struct worker *worker;
1081 hash_for_each_possible(pool->busy_hash, worker, hentry,
1082 (unsigned long)work)
1083 if (worker->current_work == work &&
1084 worker->current_func == work->func)
1091 * move_linked_works - move linked works to a list
1092 * @work: start of series of works to be scheduled
1093 * @head: target list to append @work to
1094 * @nextp: out parameter for nested worklist walking
1096 * Schedule linked works starting from @work to @head. Work series to be
1097 * scheduled starts at @work and includes any consecutive work with
1098 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1102 * raw_spin_lock_irq(pool->lock).
1104 static void move_linked_works(struct work_struct *work, struct list_head *head,
1105 struct work_struct **nextp)
1107 struct work_struct *n;
1110 * Linked worklist will always end before the end of the list,
1111 * use NULL for list head.
1113 list_for_each_entry_safe_from(work, n, NULL, entry) {
1114 list_move_tail(&work->entry, head);
1115 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1120 * If we're already inside safe list traversal and have moved
1121 * multiple works to the scheduled queue, the next position
1122 * needs to be updated.
1129 * assign_work - assign a work item and its linked work items to a worker
1130 * @work: work to assign
1131 * @worker: worker to assign to
1132 * @nextp: out parameter for nested worklist walking
1134 * Assign @work and its linked work items to @worker. If @work is already being
1135 * executed by another worker in the same pool, it'll be punted there.
1137 * If @nextp is not NULL, it's updated to point to the next work of the last
1138 * scheduled work. This allows assign_work() to be nested inside
1139 * list_for_each_entry_safe().
1141 * Returns %true if @work was successfully assigned to @worker. %false if @work
1142 * was punted to another worker already executing it.
1144 static bool assign_work(struct work_struct *work, struct worker *worker,
1145 struct work_struct **nextp)
1147 struct worker_pool *pool = worker->pool;
1148 struct worker *collision;
1150 lockdep_assert_held(&pool->lock);
1153 * A single work shouldn't be executed concurrently by multiple workers.
1154 * __queue_work() ensures that @work doesn't jump to a different pool
1155 * while still running in the previous pool. Here, we should ensure that
1156 * @work is not executed concurrently by multiple workers from the same
1157 * pool. Check whether anyone is already processing the work. If so,
1158 * defer the work to the currently executing one.
1160 collision = find_worker_executing_work(pool, work);
1161 if (unlikely(collision)) {
1162 move_linked_works(work, &collision->scheduled, nextp);
1166 move_linked_works(work, &worker->scheduled, nextp);
1171 * kick_pool - wake up an idle worker if necessary
1172 * @pool: pool to kick
1174 * @pool may have pending work items. Wake up worker if necessary. Returns
1175 * whether a worker was woken up.
1177 static bool kick_pool(struct worker_pool *pool)
1179 struct worker *worker = first_idle_worker(pool);
1180 struct task_struct *p;
1182 lockdep_assert_held(&pool->lock);
1184 if (!need_more_worker(pool) || !worker)
1191 * Idle @worker is about to execute @work and waking up provides an
1192 * opportunity to migrate @worker at a lower cost by setting the task's
1193 * wake_cpu field. Let's see if we want to move @worker to improve
1194 * execution locality.
1196 * We're waking the worker that went idle the latest and there's some
1197 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1198 * so, setting the wake_cpu won't do anything. As this is a best-effort
1199 * optimization and the race window is narrow, let's leave as-is for
1200 * now. If this becomes pronounced, we can skip over workers which are
1201 * still on cpu when picking an idle worker.
1203 * If @pool has non-strict affinity, @worker might have ended up outside
1204 * its affinity scope. Repatriate.
1206 if (!pool->attrs->affn_strict &&
1207 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1208 struct work_struct *work = list_first_entry(&pool->worklist,
1209 struct work_struct, entry);
1210 p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1211 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1218 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1221 * Concurrency-managed per-cpu work items that hog CPU for longer than
1222 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1223 * which prevents them from stalling other concurrency-managed work items. If a
1224 * work function keeps triggering this mechanism, it's likely that the work item
1225 * should be using an unbound workqueue instead.
1227 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1228 * and report them so that they can be examined and converted to use unbound
1229 * workqueues as appropriate. To avoid flooding the console, each violating work
1230 * function is tracked and reported with exponential backoff.
1232 #define WCI_MAX_ENTS 128
1237 struct hlist_node hash_node;
1240 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1241 static int wci_nr_ents;
1242 static DEFINE_RAW_SPINLOCK(wci_lock);
1243 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1245 static struct wci_ent *wci_find_ent(work_func_t func)
1247 struct wci_ent *ent;
1249 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1250 (unsigned long)func) {
1251 if (ent->func == func)
1257 static void wq_cpu_intensive_report(work_func_t func)
1259 struct wci_ent *ent;
1262 ent = wci_find_ent(func);
1267 * Start reporting from the fourth time and back off
1270 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1271 if (cnt >= 4 && is_power_of_2(cnt))
1272 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1273 ent->func, wq_cpu_intensive_thresh_us,
1274 atomic64_read(&ent->cnt));
1279 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1280 * is exhausted, something went really wrong and we probably made enough
1283 if (wci_nr_ents >= WCI_MAX_ENTS)
1286 raw_spin_lock(&wci_lock);
1288 if (wci_nr_ents >= WCI_MAX_ENTS) {
1289 raw_spin_unlock(&wci_lock);
1293 if (wci_find_ent(func)) {
1294 raw_spin_unlock(&wci_lock);
1298 ent = &wci_ents[wci_nr_ents++];
1300 atomic64_set(&ent->cnt, 1);
1301 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1303 raw_spin_unlock(&wci_lock);
1306 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1307 static void wq_cpu_intensive_report(work_func_t func) {}
1308 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1311 * wq_worker_running - a worker is running again
1312 * @task: task waking up
1314 * This function is called when a worker returns from schedule()
1316 void wq_worker_running(struct task_struct *task)
1318 struct worker *worker = kthread_data(task);
1320 if (!READ_ONCE(worker->sleeping))
1324 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1325 * and the nr_running increment below, we may ruin the nr_running reset
1326 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1327 * pool. Protect against such race.
1330 if (!(worker->flags & WORKER_NOT_RUNNING))
1331 worker->pool->nr_running++;
1335 * CPU intensive auto-detection cares about how long a work item hogged
1336 * CPU without sleeping. Reset the starting timestamp on wakeup.
1338 worker->current_at = worker->task->se.sum_exec_runtime;
1340 WRITE_ONCE(worker->sleeping, 0);
1344 * wq_worker_sleeping - a worker is going to sleep
1345 * @task: task going to sleep
1347 * This function is called from schedule() when a busy worker is
1350 void wq_worker_sleeping(struct task_struct *task)
1352 struct worker *worker = kthread_data(task);
1353 struct worker_pool *pool;
1356 * Rescuers, which may not have all the fields set up like normal
1357 * workers, also reach here, let's not access anything before
1358 * checking NOT_RUNNING.
1360 if (worker->flags & WORKER_NOT_RUNNING)
1363 pool = worker->pool;
1365 /* Return if preempted before wq_worker_running() was reached */
1366 if (READ_ONCE(worker->sleeping))
1369 WRITE_ONCE(worker->sleeping, 1);
1370 raw_spin_lock_irq(&pool->lock);
1373 * Recheck in case unbind_workers() preempted us. We don't
1374 * want to decrement nr_running after the worker is unbound
1375 * and nr_running has been reset.
1377 if (worker->flags & WORKER_NOT_RUNNING) {
1378 raw_spin_unlock_irq(&pool->lock);
1383 if (kick_pool(pool))
1384 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1386 raw_spin_unlock_irq(&pool->lock);
1390 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1391 * @task: task currently running
1393 * Called from scheduler_tick(). We're in the IRQ context and the current
1394 * worker's fields which follow the 'K' locking rule can be accessed safely.
1396 void wq_worker_tick(struct task_struct *task)
1398 struct worker *worker = kthread_data(task);
1399 struct pool_workqueue *pwq = worker->current_pwq;
1400 struct worker_pool *pool = worker->pool;
1405 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1407 if (!wq_cpu_intensive_thresh_us)
1411 * If the current worker is concurrency managed and hogged the CPU for
1412 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1413 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1415 * Set @worker->sleeping means that @worker is in the process of
1416 * switching out voluntarily and won't be contributing to
1417 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1418 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1419 * double decrements. The task is releasing the CPU anyway. Let's skip.
1420 * We probably want to make this prettier in the future.
1422 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1423 worker->task->se.sum_exec_runtime - worker->current_at <
1424 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1427 raw_spin_lock(&pool->lock);
1429 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1430 wq_cpu_intensive_report(worker->current_func);
1431 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1433 if (kick_pool(pool))
1434 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1436 raw_spin_unlock(&pool->lock);
1440 * wq_worker_last_func - retrieve worker's last work function
1441 * @task: Task to retrieve last work function of.
1443 * Determine the last function a worker executed. This is called from
1444 * the scheduler to get a worker's last known identity.
1447 * raw_spin_lock_irq(rq->lock)
1449 * This function is called during schedule() when a kworker is going
1450 * to sleep. It's used by psi to identify aggregation workers during
1451 * dequeuing, to allow periodic aggregation to shut-off when that
1452 * worker is the last task in the system or cgroup to go to sleep.
1454 * As this function doesn't involve any workqueue-related locking, it
1455 * only returns stable values when called from inside the scheduler's
1456 * queuing and dequeuing paths, when @task, which must be a kworker,
1457 * is guaranteed to not be processing any works.
1460 * The last work function %current executed as a worker, NULL if it
1461 * hasn't executed any work yet.
1463 work_func_t wq_worker_last_func(struct task_struct *task)
1465 struct worker *worker = kthread_data(task);
1467 return worker->last_func;
1471 * wq_node_nr_active - Determine wq_node_nr_active to use
1472 * @wq: workqueue of interest
1473 * @node: NUMA node, can be %NUMA_NO_NODE
1475 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1477 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1479 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1481 * - Otherwise, node_nr_active[@node].
1483 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1486 if (!(wq->flags & WQ_UNBOUND))
1489 if (node == NUMA_NO_NODE)
1492 return wq->node_nr_active[node];
1496 * wq_update_node_max_active - Update per-node max_actives to use
1497 * @wq: workqueue to update
1498 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1500 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1501 * distributed among nodes according to the proportions of numbers of online
1502 * cpus. The result is always between @wq->min_active and max_active.
1504 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1506 struct cpumask *effective = unbound_effective_cpumask(wq);
1507 int min_active = READ_ONCE(wq->min_active);
1508 int max_active = READ_ONCE(wq->max_active);
1509 int total_cpus, node;
1511 lockdep_assert_held(&wq->mutex);
1513 if (!cpumask_test_cpu(off_cpu, effective))
1516 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1520 for_each_node(node) {
1523 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1524 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1527 wq_node_nr_active(wq, node)->max =
1528 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1529 min_active, max_active);
1532 wq_node_nr_active(wq, NUMA_NO_NODE)->max = min_active;
1536 * get_pwq - get an extra reference on the specified pool_workqueue
1537 * @pwq: pool_workqueue to get
1539 * Obtain an extra reference on @pwq. The caller should guarantee that
1540 * @pwq has positive refcnt and be holding the matching pool->lock.
1542 static void get_pwq(struct pool_workqueue *pwq)
1544 lockdep_assert_held(&pwq->pool->lock);
1545 WARN_ON_ONCE(pwq->refcnt <= 0);
1550 * put_pwq - put a pool_workqueue reference
1551 * @pwq: pool_workqueue to put
1553 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1554 * destruction. The caller should be holding the matching pool->lock.
1556 static void put_pwq(struct pool_workqueue *pwq)
1558 lockdep_assert_held(&pwq->pool->lock);
1559 if (likely(--pwq->refcnt))
1562 * @pwq can't be released under pool->lock, bounce to a dedicated
1563 * kthread_worker to avoid A-A deadlocks.
1565 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1569 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1570 * @pwq: pool_workqueue to put (can be %NULL)
1572 * put_pwq() with locking. This function also allows %NULL @pwq.
1574 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1578 * As both pwqs and pools are RCU protected, the
1579 * following lock operations are safe.
1581 raw_spin_lock_irq(&pwq->pool->lock);
1583 raw_spin_unlock_irq(&pwq->pool->lock);
1587 static bool pwq_is_empty(struct pool_workqueue *pwq)
1589 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1592 static void __pwq_activate_work(struct pool_workqueue *pwq,
1593 struct work_struct *work)
1595 unsigned long *wdb = work_data_bits(work);
1597 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1598 trace_workqueue_activate_work(work);
1599 if (list_empty(&pwq->pool->worklist))
1600 pwq->pool->watchdog_ts = jiffies;
1601 move_linked_works(work, &pwq->pool->worklist, NULL);
1602 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1606 * pwq_activate_work - Activate a work item if inactive
1607 * @pwq: pool_workqueue @work belongs to
1608 * @work: work item to activate
1610 * Returns %true if activated. %false if already active.
1612 static bool pwq_activate_work(struct pool_workqueue *pwq,
1613 struct work_struct *work)
1615 struct worker_pool *pool = pwq->pool;
1616 struct wq_node_nr_active *nna;
1618 lockdep_assert_held(&pool->lock);
1620 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE))
1623 nna = wq_node_nr_active(pwq->wq, pool->node);
1625 atomic_inc(&nna->nr);
1628 __pwq_activate_work(pwq, work);
1632 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1634 int max = READ_ONCE(nna->max);
1639 old = atomic_read(&nna->nr);
1642 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1649 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1650 * @pwq: pool_workqueue of interest
1651 * @fill: max_active may have increased, try to increase concurrency level
1653 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1654 * successfully obtained. %false otherwise.
1656 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1658 struct workqueue_struct *wq = pwq->wq;
1659 struct worker_pool *pool = pwq->pool;
1660 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1661 bool obtained = false;
1663 lockdep_assert_held(&pool->lock);
1666 /* per-cpu workqueue, pwq->nr_active is sufficient */
1667 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1672 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1673 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1674 * concurrency level. Don't jump the line.
1676 * We need to ignore the pending test after max_active has increased as
1677 * pwq_dec_nr_active() can only maintain the concurrency level but not
1678 * increase it. This is indicated by @fill.
1680 if (!list_empty(&pwq->pending_node) && likely(!fill))
1683 obtained = tryinc_node_nr_active(nna);
1688 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1689 * and try again. The smp_mb() is paired with the implied memory barrier
1690 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1691 * we see the decremented $nna->nr or they see non-empty
1692 * $nna->pending_pwqs.
1694 raw_spin_lock(&nna->lock);
1696 if (list_empty(&pwq->pending_node))
1697 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1698 else if (likely(!fill))
1703 obtained = tryinc_node_nr_active(nna);
1706 * If @fill, @pwq might have already been pending. Being spuriously
1707 * pending in cold paths doesn't affect anything. Let's leave it be.
1709 if (obtained && likely(!fill))
1710 list_del_init(&pwq->pending_node);
1713 raw_spin_unlock(&nna->lock);
1721 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1722 * @pwq: pool_workqueue of interest
1723 * @fill: max_active may have increased, try to increase concurrency level
1725 * Activate the first inactive work item of @pwq if available and allowed by
1728 * Returns %true if an inactive work item has been activated. %false if no
1729 * inactive work item is found or max_active limit is reached.
1731 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1733 struct work_struct *work =
1734 list_first_entry_or_null(&pwq->inactive_works,
1735 struct work_struct, entry);
1737 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1738 __pwq_activate_work(pwq, work);
1746 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1747 * @nna: wq_node_nr_active to activate a pending pwq for
1748 * @caller_pool: worker_pool the caller is locking
1750 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1751 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1753 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1754 struct worker_pool *caller_pool)
1756 struct worker_pool *locked_pool = caller_pool;
1757 struct pool_workqueue *pwq;
1758 struct work_struct *work;
1760 lockdep_assert_held(&caller_pool->lock);
1762 raw_spin_lock(&nna->lock);
1764 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1765 struct pool_workqueue, pending_node);
1770 * If @pwq is for a different pool than @locked_pool, we need to lock
1771 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1772 * / lock dance. For that, we also need to release @nna->lock as it's
1773 * nested inside pool locks.
1775 if (pwq->pool != locked_pool) {
1776 raw_spin_unlock(&locked_pool->lock);
1777 locked_pool = pwq->pool;
1778 if (!raw_spin_trylock(&locked_pool->lock)) {
1779 raw_spin_unlock(&nna->lock);
1780 raw_spin_lock(&locked_pool->lock);
1781 raw_spin_lock(&nna->lock);
1787 * $pwq may not have any inactive work items due to e.g. cancellations.
1788 * Drop it from pending_pwqs and see if there's another one.
1790 work = list_first_entry_or_null(&pwq->inactive_works,
1791 struct work_struct, entry);
1793 list_del_init(&pwq->pending_node);
1798 * Acquire an nr_active count and activate the inactive work item. If
1799 * $pwq still has inactive work items, rotate it to the end of the
1800 * pending_pwqs so that we round-robin through them. This means that
1801 * inactive work items are not activated in queueing order which is fine
1802 * given that there has never been any ordering across different pwqs.
1804 if (likely(tryinc_node_nr_active(nna))) {
1806 __pwq_activate_work(pwq, work);
1808 if (list_empty(&pwq->inactive_works))
1809 list_del_init(&pwq->pending_node);
1811 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1813 /* if activating a foreign pool, make sure it's running */
1814 if (pwq->pool != caller_pool)
1815 kick_pool(pwq->pool);
1819 raw_spin_unlock(&nna->lock);
1820 if (locked_pool != caller_pool) {
1821 raw_spin_unlock(&locked_pool->lock);
1822 raw_spin_lock(&caller_pool->lock);
1827 * pwq_dec_nr_active - Retire an active count
1828 * @pwq: pool_workqueue of interest
1830 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1831 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1833 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1835 struct worker_pool *pool = pwq->pool;
1836 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1838 lockdep_assert_held(&pool->lock);
1841 * @pwq->nr_active should be decremented for both percpu and unbound
1847 * For a percpu workqueue, it's simple. Just need to kick the first
1848 * inactive work item on @pwq itself.
1851 pwq_activate_first_inactive(pwq, false);
1856 * If @pwq is for an unbound workqueue, it's more complicated because
1857 * multiple pwqs and pools may be sharing the nr_active count. When a
1858 * pwq needs to wait for an nr_active count, it puts itself on
1859 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1860 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1861 * guarantee that either we see non-empty pending_pwqs or they see
1862 * decremented $nna->nr.
1864 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1865 * max_active gets updated. However, it is guaranteed to be equal to or
1866 * larger than @pwq->wq->min_active which is above zero unless freezing.
1867 * This maintains the forward progress guarantee.
1869 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1872 if (!list_empty(&nna->pending_pwqs))
1873 node_activate_pending_pwq(nna, pool);
1877 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1878 * @pwq: pwq of interest
1879 * @work_data: work_data of work which left the queue
1881 * A work either has completed or is removed from pending queue,
1882 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1885 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1886 * and thus should be called after all other state updates for the in-flight
1887 * work item is complete.
1890 * raw_spin_lock_irq(pool->lock).
1892 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1894 int color = get_work_color(work_data);
1896 if (!(work_data & WORK_STRUCT_INACTIVE))
1897 pwq_dec_nr_active(pwq);
1899 pwq->nr_in_flight[color]--;
1901 /* is flush in progress and are we at the flushing tip? */
1902 if (likely(pwq->flush_color != color))
1905 /* are there still in-flight works? */
1906 if (pwq->nr_in_flight[color])
1909 /* this pwq is done, clear flush_color */
1910 pwq->flush_color = -1;
1913 * If this was the last pwq, wake up the first flusher. It
1914 * will handle the rest.
1916 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1917 complete(&pwq->wq->first_flusher->done);
1923 * try_to_grab_pending - steal work item from worklist and disable irq
1924 * @work: work item to steal
1925 * @is_dwork: @work is a delayed_work
1926 * @flags: place to store irq state
1928 * Try to grab PENDING bit of @work. This function can handle @work in any
1929 * stable state - idle, on timer or on worklist.
1933 * ======== ================================================================
1934 * 1 if @work was pending and we successfully stole PENDING
1935 * 0 if @work was idle and we claimed PENDING
1936 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1937 * -ENOENT if someone else is canceling @work, this state may persist
1938 * for arbitrarily long
1939 * ======== ================================================================
1942 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1943 * interrupted while holding PENDING and @work off queue, irq must be
1944 * disabled on entry. This, combined with delayed_work->timer being
1945 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1947 * On successful return, >= 0, irq is disabled and the caller is
1948 * responsible for releasing it using local_irq_restore(*@flags).
1950 * This function is safe to call from any context including IRQ handler.
1952 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1953 unsigned long *flags)
1955 struct worker_pool *pool;
1956 struct pool_workqueue *pwq;
1958 local_irq_save(*flags);
1960 /* try to steal the timer if it exists */
1962 struct delayed_work *dwork = to_delayed_work(work);
1965 * dwork->timer is irqsafe. If del_timer() fails, it's
1966 * guaranteed that the timer is not queued anywhere and not
1967 * running on the local CPU.
1969 if (likely(del_timer(&dwork->timer)))
1973 /* try to claim PENDING the normal way */
1974 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1979 * The queueing is in progress, or it is already queued. Try to
1980 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1982 pool = get_work_pool(work);
1986 raw_spin_lock(&pool->lock);
1988 * work->data is guaranteed to point to pwq only while the work
1989 * item is queued on pwq->wq, and both updating work->data to point
1990 * to pwq on queueing and to pool on dequeueing are done under
1991 * pwq->pool->lock. This in turn guarantees that, if work->data
1992 * points to pwq which is associated with a locked pool, the work
1993 * item is currently queued on that pool.
1995 pwq = get_work_pwq(work);
1996 if (pwq && pwq->pool == pool) {
1997 debug_work_deactivate(work);
2000 * A cancelable inactive work item must be in the
2001 * pwq->inactive_works since a queued barrier can't be
2002 * canceled (see the comments in insert_wq_barrier()).
2004 * An inactive work item cannot be grabbed directly because
2005 * it might have linked barrier work items which, if left
2006 * on the inactive_works list, will confuse pwq->nr_active
2007 * management later on and cause stall. Make sure the work
2008 * item is activated before grabbing.
2010 pwq_activate_work(pwq, work);
2012 list_del_init(&work->entry);
2014 /* work->data points to pwq iff queued, point to pool */
2015 set_work_pool_and_keep_pending(work, pool->id);
2017 /* must be the last step, see the function comment */
2018 pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
2020 raw_spin_unlock(&pool->lock);
2024 raw_spin_unlock(&pool->lock);
2027 local_irq_restore(*flags);
2028 if (work_is_canceling(work))
2035 * insert_work - insert a work into a pool
2036 * @pwq: pwq @work belongs to
2037 * @work: work to insert
2038 * @head: insertion point
2039 * @extra_flags: extra WORK_STRUCT_* flags to set
2041 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2042 * work_struct flags.
2045 * raw_spin_lock_irq(pool->lock).
2047 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2048 struct list_head *head, unsigned int extra_flags)
2050 debug_work_activate(work);
2052 /* record the work call stack in order to print it in KASAN reports */
2053 kasan_record_aux_stack_noalloc(work);
2055 /* we own @work, set data and link */
2056 set_work_pwq(work, pwq, extra_flags);
2057 list_add_tail(&work->entry, head);
2062 * Test whether @work is being queued from another work executing on the
2065 static bool is_chained_work(struct workqueue_struct *wq)
2067 struct worker *worker;
2069 worker = current_wq_worker();
2071 * Return %true iff I'm a worker executing a work item on @wq. If
2072 * I'm @worker, it's safe to dereference it without locking.
2074 return worker && worker->current_pwq->wq == wq;
2078 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2079 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2080 * avoid perturbing sensitive tasks.
2082 static int wq_select_unbound_cpu(int cpu)
2086 if (likely(!wq_debug_force_rr_cpu)) {
2087 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2090 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2093 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2094 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2095 if (unlikely(new_cpu >= nr_cpu_ids)) {
2096 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2097 if (unlikely(new_cpu >= nr_cpu_ids))
2100 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2105 static void __queue_work(int cpu, struct workqueue_struct *wq,
2106 struct work_struct *work)
2108 struct pool_workqueue *pwq;
2109 struct worker_pool *last_pool, *pool;
2110 unsigned int work_flags;
2111 unsigned int req_cpu = cpu;
2114 * While a work item is PENDING && off queue, a task trying to
2115 * steal the PENDING will busy-loop waiting for it to either get
2116 * queued or lose PENDING. Grabbing PENDING and queueing should
2117 * happen with IRQ disabled.
2119 lockdep_assert_irqs_disabled();
2123 * For a draining wq, only works from the same workqueue are
2124 * allowed. The __WQ_DESTROYING helps to spot the issue that
2125 * queues a new work item to a wq after destroy_workqueue(wq).
2127 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2128 WARN_ON_ONCE(!is_chained_work(wq))))
2132 /* pwq which will be used unless @work is executing elsewhere */
2133 if (req_cpu == WORK_CPU_UNBOUND) {
2134 if (wq->flags & WQ_UNBOUND)
2135 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2137 cpu = raw_smp_processor_id();
2140 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2144 * If @work was previously on a different pool, it might still be
2145 * running there, in which case the work needs to be queued on that
2146 * pool to guarantee non-reentrancy.
2148 last_pool = get_work_pool(work);
2149 if (last_pool && last_pool != pool) {
2150 struct worker *worker;
2152 raw_spin_lock(&last_pool->lock);
2154 worker = find_worker_executing_work(last_pool, work);
2156 if (worker && worker->current_pwq->wq == wq) {
2157 pwq = worker->current_pwq;
2159 WARN_ON_ONCE(pool != last_pool);
2161 /* meh... not running there, queue here */
2162 raw_spin_unlock(&last_pool->lock);
2163 raw_spin_lock(&pool->lock);
2166 raw_spin_lock(&pool->lock);
2170 * pwq is determined and locked. For unbound pools, we could have raced
2171 * with pwq release and it could already be dead. If its refcnt is zero,
2172 * repeat pwq selection. Note that unbound pwqs never die without
2173 * another pwq replacing it in cpu_pwq or while work items are executing
2174 * on it, so the retrying is guaranteed to make forward-progress.
2176 if (unlikely(!pwq->refcnt)) {
2177 if (wq->flags & WQ_UNBOUND) {
2178 raw_spin_unlock(&pool->lock);
2183 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2187 /* pwq determined, queue */
2188 trace_workqueue_queue_work(req_cpu, pwq, work);
2190 if (WARN_ON(!list_empty(&work->entry)))
2193 pwq->nr_in_flight[pwq->work_color]++;
2194 work_flags = work_color_to_flags(pwq->work_color);
2197 * Limit the number of concurrently active work items to max_active.
2198 * @work must also queue behind existing inactive work items to maintain
2199 * ordering when max_active changes. See wq_adjust_max_active().
2201 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2202 if (list_empty(&pool->worklist))
2203 pool->watchdog_ts = jiffies;
2205 trace_workqueue_activate_work(work);
2206 insert_work(pwq, work, &pool->worklist, work_flags);
2209 work_flags |= WORK_STRUCT_INACTIVE;
2210 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2214 raw_spin_unlock(&pool->lock);
2219 * queue_work_on - queue work on specific cpu
2220 * @cpu: CPU number to execute work on
2221 * @wq: workqueue to use
2222 * @work: work to queue
2224 * We queue the work to a specific CPU, the caller must ensure it
2225 * can't go away. Callers that fail to ensure that the specified
2226 * CPU cannot go away will execute on a randomly chosen CPU.
2227 * But note well that callers specifying a CPU that never has been
2228 * online will get a splat.
2230 * Return: %false if @work was already on a queue, %true otherwise.
2232 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2233 struct work_struct *work)
2236 unsigned long flags;
2238 local_irq_save(flags);
2240 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2241 __queue_work(cpu, wq, work);
2245 local_irq_restore(flags);
2248 EXPORT_SYMBOL(queue_work_on);
2251 * select_numa_node_cpu - Select a CPU based on NUMA node
2252 * @node: NUMA node ID that we want to select a CPU from
2254 * This function will attempt to find a "random" cpu available on a given
2255 * node. If there are no CPUs available on the given node it will return
2256 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2257 * available CPU if we need to schedule this work.
2259 static int select_numa_node_cpu(int node)
2263 /* Delay binding to CPU if node is not valid or online */
2264 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2265 return WORK_CPU_UNBOUND;
2267 /* Use local node/cpu if we are already there */
2268 cpu = raw_smp_processor_id();
2269 if (node == cpu_to_node(cpu))
2272 /* Use "random" otherwise know as "first" online CPU of node */
2273 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2275 /* If CPU is valid return that, otherwise just defer */
2276 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2280 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2281 * @node: NUMA node that we are targeting the work for
2282 * @wq: workqueue to use
2283 * @work: work to queue
2285 * We queue the work to a "random" CPU within a given NUMA node. The basic
2286 * idea here is to provide a way to somehow associate work with a given
2289 * This function will only make a best effort attempt at getting this onto
2290 * the right NUMA node. If no node is requested or the requested node is
2291 * offline then we just fall back to standard queue_work behavior.
2293 * Currently the "random" CPU ends up being the first available CPU in the
2294 * intersection of cpu_online_mask and the cpumask of the node, unless we
2295 * are running on the node. In that case we just use the current CPU.
2297 * Return: %false if @work was already on a queue, %true otherwise.
2299 bool queue_work_node(int node, struct workqueue_struct *wq,
2300 struct work_struct *work)
2302 unsigned long flags;
2306 * This current implementation is specific to unbound workqueues.
2307 * Specifically we only return the first available CPU for a given
2308 * node instead of cycling through individual CPUs within the node.
2310 * If this is used with a per-cpu workqueue then the logic in
2311 * workqueue_select_cpu_near would need to be updated to allow for
2312 * some round robin type logic.
2314 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2316 local_irq_save(flags);
2318 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2319 int cpu = select_numa_node_cpu(node);
2321 __queue_work(cpu, wq, work);
2325 local_irq_restore(flags);
2328 EXPORT_SYMBOL_GPL(queue_work_node);
2330 void delayed_work_timer_fn(struct timer_list *t)
2332 struct delayed_work *dwork = from_timer(dwork, t, timer);
2334 /* should have been called from irqsafe timer with irq already off */
2335 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2337 EXPORT_SYMBOL(delayed_work_timer_fn);
2339 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2340 struct delayed_work *dwork, unsigned long delay)
2342 struct timer_list *timer = &dwork->timer;
2343 struct work_struct *work = &dwork->work;
2346 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2347 WARN_ON_ONCE(timer_pending(timer));
2348 WARN_ON_ONCE(!list_empty(&work->entry));
2351 * If @delay is 0, queue @dwork->work immediately. This is for
2352 * both optimization and correctness. The earliest @timer can
2353 * expire is on the closest next tick and delayed_work users depend
2354 * on that there's no such delay when @delay is 0.
2357 __queue_work(cpu, wq, &dwork->work);
2363 timer->expires = jiffies + delay;
2365 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2366 /* If the current cpu is a housekeeping cpu, use it. */
2367 cpu = smp_processor_id();
2368 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2369 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2370 add_timer_on(timer, cpu);
2372 if (likely(cpu == WORK_CPU_UNBOUND))
2375 add_timer_on(timer, cpu);
2380 * queue_delayed_work_on - queue work on specific CPU after delay
2381 * @cpu: CPU number to execute work on
2382 * @wq: workqueue to use
2383 * @dwork: work to queue
2384 * @delay: number of jiffies to wait before queueing
2386 * Return: %false if @work was already on a queue, %true otherwise. If
2387 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2390 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2391 struct delayed_work *dwork, unsigned long delay)
2393 struct work_struct *work = &dwork->work;
2395 unsigned long flags;
2397 /* read the comment in __queue_work() */
2398 local_irq_save(flags);
2400 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2401 __queue_delayed_work(cpu, wq, dwork, delay);
2405 local_irq_restore(flags);
2408 EXPORT_SYMBOL(queue_delayed_work_on);
2411 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2412 * @cpu: CPU number to execute work on
2413 * @wq: workqueue to use
2414 * @dwork: work to queue
2415 * @delay: number of jiffies to wait before queueing
2417 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2418 * modify @dwork's timer so that it expires after @delay. If @delay is
2419 * zero, @work is guaranteed to be scheduled immediately regardless of its
2422 * Return: %false if @dwork was idle and queued, %true if @dwork was
2423 * pending and its timer was modified.
2425 * This function is safe to call from any context including IRQ handler.
2426 * See try_to_grab_pending() for details.
2428 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2429 struct delayed_work *dwork, unsigned long delay)
2431 unsigned long flags;
2435 ret = try_to_grab_pending(&dwork->work, true, &flags);
2436 } while (unlikely(ret == -EAGAIN));
2438 if (likely(ret >= 0)) {
2439 __queue_delayed_work(cpu, wq, dwork, delay);
2440 local_irq_restore(flags);
2443 /* -ENOENT from try_to_grab_pending() becomes %true */
2446 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2448 static void rcu_work_rcufn(struct rcu_head *rcu)
2450 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2452 /* read the comment in __queue_work() */
2453 local_irq_disable();
2454 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2459 * queue_rcu_work - queue work after a RCU grace period
2460 * @wq: workqueue to use
2461 * @rwork: work to queue
2463 * Return: %false if @rwork was already pending, %true otherwise. Note
2464 * that a full RCU grace period is guaranteed only after a %true return.
2465 * While @rwork is guaranteed to be executed after a %false return, the
2466 * execution may happen before a full RCU grace period has passed.
2468 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2470 struct work_struct *work = &rwork->work;
2472 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2474 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2480 EXPORT_SYMBOL(queue_rcu_work);
2482 static struct worker *alloc_worker(int node)
2484 struct worker *worker;
2486 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2488 INIT_LIST_HEAD(&worker->entry);
2489 INIT_LIST_HEAD(&worker->scheduled);
2490 INIT_LIST_HEAD(&worker->node);
2491 /* on creation a worker is in !idle && prep state */
2492 worker->flags = WORKER_PREP;
2497 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2499 if (pool->cpu < 0 && pool->attrs->affn_strict)
2500 return pool->attrs->__pod_cpumask;
2502 return pool->attrs->cpumask;
2506 * worker_attach_to_pool() - attach a worker to a pool
2507 * @worker: worker to be attached
2508 * @pool: the target pool
2510 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2511 * cpu-binding of @worker are kept coordinated with the pool across
2514 static void worker_attach_to_pool(struct worker *worker,
2515 struct worker_pool *pool)
2517 mutex_lock(&wq_pool_attach_mutex);
2520 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2521 * stable across this function. See the comments above the flag
2522 * definition for details.
2524 if (pool->flags & POOL_DISASSOCIATED)
2525 worker->flags |= WORKER_UNBOUND;
2527 kthread_set_per_cpu(worker->task, pool->cpu);
2529 if (worker->rescue_wq)
2530 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2532 list_add_tail(&worker->node, &pool->workers);
2533 worker->pool = pool;
2535 mutex_unlock(&wq_pool_attach_mutex);
2539 * worker_detach_from_pool() - detach a worker from its pool
2540 * @worker: worker which is attached to its pool
2542 * Undo the attaching which had been done in worker_attach_to_pool(). The
2543 * caller worker shouldn't access to the pool after detached except it has
2544 * other reference to the pool.
2546 static void worker_detach_from_pool(struct worker *worker)
2548 struct worker_pool *pool = worker->pool;
2549 struct completion *detach_completion = NULL;
2551 mutex_lock(&wq_pool_attach_mutex);
2553 kthread_set_per_cpu(worker->task, -1);
2554 list_del(&worker->node);
2555 worker->pool = NULL;
2557 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2558 detach_completion = pool->detach_completion;
2559 mutex_unlock(&wq_pool_attach_mutex);
2561 /* clear leftover flags without pool->lock after it is detached */
2562 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2564 if (detach_completion)
2565 complete(detach_completion);
2569 * create_worker - create a new workqueue worker
2570 * @pool: pool the new worker will belong to
2572 * Create and start a new worker which is attached to @pool.
2575 * Might sleep. Does GFP_KERNEL allocations.
2578 * Pointer to the newly created worker.
2580 static struct worker *create_worker(struct worker_pool *pool)
2582 struct worker *worker;
2586 /* ID is needed to determine kthread name */
2587 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2589 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2594 worker = alloc_worker(pool->node);
2596 pr_err_once("workqueue: Failed to allocate a worker\n");
2603 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2604 pool->attrs->nice < 0 ? "H" : "");
2606 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2608 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2609 "kworker/%s", id_buf);
2610 if (IS_ERR(worker->task)) {
2611 if (PTR_ERR(worker->task) == -EINTR) {
2612 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2615 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2621 set_user_nice(worker->task, pool->attrs->nice);
2622 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2624 /* successful, attach the worker to the pool */
2625 worker_attach_to_pool(worker, pool);
2627 /* start the newly created worker */
2628 raw_spin_lock_irq(&pool->lock);
2630 worker->pool->nr_workers++;
2631 worker_enter_idle(worker);
2634 * @worker is waiting on a completion in kthread() and will trigger hung
2635 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2636 * wake it up explicitly.
2638 wake_up_process(worker->task);
2640 raw_spin_unlock_irq(&pool->lock);
2645 ida_free(&pool->worker_ida, id);
2650 static void unbind_worker(struct worker *worker)
2652 lockdep_assert_held(&wq_pool_attach_mutex);
2654 kthread_set_per_cpu(worker->task, -1);
2655 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2656 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2658 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2661 static void wake_dying_workers(struct list_head *cull_list)
2663 struct worker *worker, *tmp;
2665 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2666 list_del_init(&worker->entry);
2667 unbind_worker(worker);
2669 * If the worker was somehow already running, then it had to be
2670 * in pool->idle_list when set_worker_dying() happened or we
2671 * wouldn't have gotten here.
2673 * Thus, the worker must either have observed the WORKER_DIE
2674 * flag, or have set its state to TASK_IDLE. Either way, the
2675 * below will be observed by the worker and is safe to do
2676 * outside of pool->lock.
2678 wake_up_process(worker->task);
2683 * set_worker_dying - Tag a worker for destruction
2684 * @worker: worker to be destroyed
2685 * @list: transfer worker away from its pool->idle_list and into list
2687 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2691 * raw_spin_lock_irq(pool->lock).
2693 static void set_worker_dying(struct worker *worker, struct list_head *list)
2695 struct worker_pool *pool = worker->pool;
2697 lockdep_assert_held(&pool->lock);
2698 lockdep_assert_held(&wq_pool_attach_mutex);
2700 /* sanity check frenzy */
2701 if (WARN_ON(worker->current_work) ||
2702 WARN_ON(!list_empty(&worker->scheduled)) ||
2703 WARN_ON(!(worker->flags & WORKER_IDLE)))
2709 worker->flags |= WORKER_DIE;
2711 list_move(&worker->entry, list);
2712 list_move(&worker->node, &pool->dying_workers);
2716 * idle_worker_timeout - check if some idle workers can now be deleted.
2717 * @t: The pool's idle_timer that just expired
2719 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2720 * worker_leave_idle(), as a worker flicking between idle and active while its
2721 * pool is at the too_many_workers() tipping point would cause too much timer
2722 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2723 * it expire and re-evaluate things from there.
2725 static void idle_worker_timeout(struct timer_list *t)
2727 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2728 bool do_cull = false;
2730 if (work_pending(&pool->idle_cull_work))
2733 raw_spin_lock_irq(&pool->lock);
2735 if (too_many_workers(pool)) {
2736 struct worker *worker;
2737 unsigned long expires;
2739 /* idle_list is kept in LIFO order, check the last one */
2740 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2741 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2742 do_cull = !time_before(jiffies, expires);
2745 mod_timer(&pool->idle_timer, expires);
2747 raw_spin_unlock_irq(&pool->lock);
2750 queue_work(system_unbound_wq, &pool->idle_cull_work);
2754 * idle_cull_fn - cull workers that have been idle for too long.
2755 * @work: the pool's work for handling these idle workers
2757 * This goes through a pool's idle workers and gets rid of those that have been
2758 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2760 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2761 * culled, so this also resets worker affinity. This requires a sleepable
2762 * context, hence the split between timer callback and work item.
2764 static void idle_cull_fn(struct work_struct *work)
2766 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2767 LIST_HEAD(cull_list);
2770 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2771 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2772 * path. This is required as a previously-preempted worker could run after
2773 * set_worker_dying() has happened but before wake_dying_workers() did.
2775 mutex_lock(&wq_pool_attach_mutex);
2776 raw_spin_lock_irq(&pool->lock);
2778 while (too_many_workers(pool)) {
2779 struct worker *worker;
2780 unsigned long expires;
2782 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2783 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2785 if (time_before(jiffies, expires)) {
2786 mod_timer(&pool->idle_timer, expires);
2790 set_worker_dying(worker, &cull_list);
2793 raw_spin_unlock_irq(&pool->lock);
2794 wake_dying_workers(&cull_list);
2795 mutex_unlock(&wq_pool_attach_mutex);
2798 static void send_mayday(struct work_struct *work)
2800 struct pool_workqueue *pwq = get_work_pwq(work);
2801 struct workqueue_struct *wq = pwq->wq;
2803 lockdep_assert_held(&wq_mayday_lock);
2808 /* mayday mayday mayday */
2809 if (list_empty(&pwq->mayday_node)) {
2811 * If @pwq is for an unbound wq, its base ref may be put at
2812 * any time due to an attribute change. Pin @pwq until the
2813 * rescuer is done with it.
2816 list_add_tail(&pwq->mayday_node, &wq->maydays);
2817 wake_up_process(wq->rescuer->task);
2818 pwq->stats[PWQ_STAT_MAYDAY]++;
2822 static void pool_mayday_timeout(struct timer_list *t)
2824 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2825 struct work_struct *work;
2827 raw_spin_lock_irq(&pool->lock);
2828 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2830 if (need_to_create_worker(pool)) {
2832 * We've been trying to create a new worker but
2833 * haven't been successful. We might be hitting an
2834 * allocation deadlock. Send distress signals to
2837 list_for_each_entry(work, &pool->worklist, entry)
2841 raw_spin_unlock(&wq_mayday_lock);
2842 raw_spin_unlock_irq(&pool->lock);
2844 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2848 * maybe_create_worker - create a new worker if necessary
2849 * @pool: pool to create a new worker for
2851 * Create a new worker for @pool if necessary. @pool is guaranteed to
2852 * have at least one idle worker on return from this function. If
2853 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2854 * sent to all rescuers with works scheduled on @pool to resolve
2855 * possible allocation deadlock.
2857 * On return, need_to_create_worker() is guaranteed to be %false and
2858 * may_start_working() %true.
2861 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2862 * multiple times. Does GFP_KERNEL allocations. Called only from
2865 static void maybe_create_worker(struct worker_pool *pool)
2866 __releases(&pool->lock)
2867 __acquires(&pool->lock)
2870 raw_spin_unlock_irq(&pool->lock);
2872 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2873 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2876 if (create_worker(pool) || !need_to_create_worker(pool))
2879 schedule_timeout_interruptible(CREATE_COOLDOWN);
2881 if (!need_to_create_worker(pool))
2885 del_timer_sync(&pool->mayday_timer);
2886 raw_spin_lock_irq(&pool->lock);
2888 * This is necessary even after a new worker was just successfully
2889 * created as @pool->lock was dropped and the new worker might have
2890 * already become busy.
2892 if (need_to_create_worker(pool))
2897 * manage_workers - manage worker pool
2900 * Assume the manager role and manage the worker pool @worker belongs
2901 * to. At any given time, there can be only zero or one manager per
2902 * pool. The exclusion is handled automatically by this function.
2904 * The caller can safely start processing works on false return. On
2905 * true return, it's guaranteed that need_to_create_worker() is false
2906 * and may_start_working() is true.
2909 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2910 * multiple times. Does GFP_KERNEL allocations.
2913 * %false if the pool doesn't need management and the caller can safely
2914 * start processing works, %true if management function was performed and
2915 * the conditions that the caller verified before calling the function may
2916 * no longer be true.
2918 static bool manage_workers(struct worker *worker)
2920 struct worker_pool *pool = worker->pool;
2922 if (pool->flags & POOL_MANAGER_ACTIVE)
2925 pool->flags |= POOL_MANAGER_ACTIVE;
2926 pool->manager = worker;
2928 maybe_create_worker(pool);
2930 pool->manager = NULL;
2931 pool->flags &= ~POOL_MANAGER_ACTIVE;
2932 rcuwait_wake_up(&manager_wait);
2937 * process_one_work - process single work
2939 * @work: work to process
2941 * Process @work. This function contains all the logics necessary to
2942 * process a single work including synchronization against and
2943 * interaction with other workers on the same cpu, queueing and
2944 * flushing. As long as context requirement is met, any worker can
2945 * call this function to process a work.
2948 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2950 static void process_one_work(struct worker *worker, struct work_struct *work)
2951 __releases(&pool->lock)
2952 __acquires(&pool->lock)
2954 struct pool_workqueue *pwq = get_work_pwq(work);
2955 struct worker_pool *pool = worker->pool;
2956 unsigned long work_data;
2957 #ifdef CONFIG_LOCKDEP
2959 * It is permissible to free the struct work_struct from
2960 * inside the function that is called from it, this we need to
2961 * take into account for lockdep too. To avoid bogus "held
2962 * lock freed" warnings as well as problems when looking into
2963 * work->lockdep_map, make a copy and use that here.
2965 struct lockdep_map lockdep_map;
2967 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2969 /* ensure we're on the correct CPU */
2970 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2971 raw_smp_processor_id() != pool->cpu);
2973 /* claim and dequeue */
2974 debug_work_deactivate(work);
2975 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2976 worker->current_work = work;
2977 worker->current_func = work->func;
2978 worker->current_pwq = pwq;
2979 worker->current_at = worker->task->se.sum_exec_runtime;
2980 work_data = *work_data_bits(work);
2981 worker->current_color = get_work_color(work_data);
2984 * Record wq name for cmdline and debug reporting, may get
2985 * overridden through set_worker_desc().
2987 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2989 list_del_init(&work->entry);
2992 * CPU intensive works don't participate in concurrency management.
2993 * They're the scheduler's responsibility. This takes @worker out
2994 * of concurrency management and the next code block will chain
2995 * execution of the pending work items.
2997 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
2998 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3001 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3002 * since nr_running would always be >= 1 at this point. This is used to
3003 * chain execution of the pending work items for WORKER_NOT_RUNNING
3004 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3009 * Record the last pool and clear PENDING which should be the last
3010 * update to @work. Also, do this inside @pool->lock so that
3011 * PENDING and queued state changes happen together while IRQ is
3014 set_work_pool_and_clear_pending(work, pool->id);
3016 pwq->stats[PWQ_STAT_STARTED]++;
3017 raw_spin_unlock_irq(&pool->lock);
3019 lock_map_acquire(&pwq->wq->lockdep_map);
3020 lock_map_acquire(&lockdep_map);
3022 * Strictly speaking we should mark the invariant state without holding
3023 * any locks, that is, before these two lock_map_acquire()'s.
3025 * However, that would result in:
3032 * Which would create W1->C->W1 dependencies, even though there is no
3033 * actual deadlock possible. There are two solutions, using a
3034 * read-recursive acquire on the work(queue) 'locks', but this will then
3035 * hit the lockdep limitation on recursive locks, or simply discard
3038 * AFAICT there is no possible deadlock scenario between the
3039 * flush_work() and complete() primitives (except for single-threaded
3040 * workqueues), so hiding them isn't a problem.
3042 lockdep_invariant_state(true);
3043 trace_workqueue_execute_start(work);
3044 worker->current_func(work);
3046 * While we must be careful to not use "work" after this, the trace
3047 * point will only record its address.
3049 trace_workqueue_execute_end(work, worker->current_func);
3050 pwq->stats[PWQ_STAT_COMPLETED]++;
3051 lock_map_release(&lockdep_map);
3052 lock_map_release(&pwq->wq->lockdep_map);
3054 if (unlikely(in_atomic() || lockdep_depth(current) > 0 ||
3055 rcu_preempt_depth() > 0)) {
3056 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d/%d\n"
3057 " last function: %ps\n",
3058 current->comm, preempt_count(), rcu_preempt_depth(),
3059 task_pid_nr(current), worker->current_func);
3060 debug_show_held_locks(current);
3065 * The following prevents a kworker from hogging CPU on !PREEMPTION
3066 * kernels, where a requeueing work item waiting for something to
3067 * happen could deadlock with stop_machine as such work item could
3068 * indefinitely requeue itself while all other CPUs are trapped in
3069 * stop_machine. At the same time, report a quiescent RCU state so
3070 * the same condition doesn't freeze RCU.
3074 raw_spin_lock_irq(&pool->lock);
3077 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3078 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3079 * wq_cpu_intensive_thresh_us. Clear it.
3081 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3083 /* tag the worker for identification in schedule() */
3084 worker->last_func = worker->current_func;
3086 /* we're done with it, release */
3087 hash_del(&worker->hentry);
3088 worker->current_work = NULL;
3089 worker->current_func = NULL;
3090 worker->current_pwq = NULL;
3091 worker->current_color = INT_MAX;
3093 /* must be the last step, see the function comment */
3094 pwq_dec_nr_in_flight(pwq, work_data);
3098 * process_scheduled_works - process scheduled works
3101 * Process all scheduled works. Please note that the scheduled list
3102 * may change while processing a work, so this function repeatedly
3103 * fetches a work from the top and executes it.
3106 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3109 static void process_scheduled_works(struct worker *worker)
3111 struct work_struct *work;
3114 while ((work = list_first_entry_or_null(&worker->scheduled,
3115 struct work_struct, entry))) {
3117 worker->pool->watchdog_ts = jiffies;
3120 process_one_work(worker, work);
3124 static void set_pf_worker(bool val)
3126 mutex_lock(&wq_pool_attach_mutex);
3128 current->flags |= PF_WQ_WORKER;
3130 current->flags &= ~PF_WQ_WORKER;
3131 mutex_unlock(&wq_pool_attach_mutex);
3135 * worker_thread - the worker thread function
3138 * The worker thread function. All workers belong to a worker_pool -
3139 * either a per-cpu one or dynamic unbound one. These workers process all
3140 * work items regardless of their specific target workqueue. The only
3141 * exception is work items which belong to workqueues with a rescuer which
3142 * will be explained in rescuer_thread().
3146 static int worker_thread(void *__worker)
3148 struct worker *worker = __worker;
3149 struct worker_pool *pool = worker->pool;
3151 /* tell the scheduler that this is a workqueue worker */
3152 set_pf_worker(true);
3154 raw_spin_lock_irq(&pool->lock);
3156 /* am I supposed to die? */
3157 if (unlikely(worker->flags & WORKER_DIE)) {
3158 raw_spin_unlock_irq(&pool->lock);
3159 set_pf_worker(false);
3161 set_task_comm(worker->task, "kworker/dying");
3162 ida_free(&pool->worker_ida, worker->id);
3163 worker_detach_from_pool(worker);
3164 WARN_ON_ONCE(!list_empty(&worker->entry));
3169 worker_leave_idle(worker);
3171 /* no more worker necessary? */
3172 if (!need_more_worker(pool))
3175 /* do we need to manage? */
3176 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3180 * ->scheduled list can only be filled while a worker is
3181 * preparing to process a work or actually processing it.
3182 * Make sure nobody diddled with it while I was sleeping.
3184 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3187 * Finish PREP stage. We're guaranteed to have at least one idle
3188 * worker or that someone else has already assumed the manager
3189 * role. This is where @worker starts participating in concurrency
3190 * management if applicable and concurrency management is restored
3191 * after being rebound. See rebind_workers() for details.
3193 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3196 struct work_struct *work =
3197 list_first_entry(&pool->worklist,
3198 struct work_struct, entry);
3200 if (assign_work(work, worker, NULL))
3201 process_scheduled_works(worker);
3202 } while (keep_working(pool));
3204 worker_set_flags(worker, WORKER_PREP);
3207 * pool->lock is held and there's no work to process and no need to
3208 * manage, sleep. Workers are woken up only while holding
3209 * pool->lock or from local cpu, so setting the current state
3210 * before releasing pool->lock is enough to prevent losing any
3213 worker_enter_idle(worker);
3214 __set_current_state(TASK_IDLE);
3215 raw_spin_unlock_irq(&pool->lock);
3221 * rescuer_thread - the rescuer thread function
3224 * Workqueue rescuer thread function. There's one rescuer for each
3225 * workqueue which has WQ_MEM_RECLAIM set.
3227 * Regular work processing on a pool may block trying to create a new
3228 * worker which uses GFP_KERNEL allocation which has slight chance of
3229 * developing into deadlock if some works currently on the same queue
3230 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3231 * the problem rescuer solves.
3233 * When such condition is possible, the pool summons rescuers of all
3234 * workqueues which have works queued on the pool and let them process
3235 * those works so that forward progress can be guaranteed.
3237 * This should happen rarely.
3241 static int rescuer_thread(void *__rescuer)
3243 struct worker *rescuer = __rescuer;
3244 struct workqueue_struct *wq = rescuer->rescue_wq;
3247 set_user_nice(current, RESCUER_NICE_LEVEL);
3250 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3251 * doesn't participate in concurrency management.
3253 set_pf_worker(true);
3255 set_current_state(TASK_IDLE);
3258 * By the time the rescuer is requested to stop, the workqueue
3259 * shouldn't have any work pending, but @wq->maydays may still have
3260 * pwq(s) queued. This can happen by non-rescuer workers consuming
3261 * all the work items before the rescuer got to them. Go through
3262 * @wq->maydays processing before acting on should_stop so that the
3263 * list is always empty on exit.
3265 should_stop = kthread_should_stop();
3267 /* see whether any pwq is asking for help */
3268 raw_spin_lock_irq(&wq_mayday_lock);
3270 while (!list_empty(&wq->maydays)) {
3271 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3272 struct pool_workqueue, mayday_node);
3273 struct worker_pool *pool = pwq->pool;
3274 struct work_struct *work, *n;
3276 __set_current_state(TASK_RUNNING);
3277 list_del_init(&pwq->mayday_node);
3279 raw_spin_unlock_irq(&wq_mayday_lock);
3281 worker_attach_to_pool(rescuer, pool);
3283 raw_spin_lock_irq(&pool->lock);
3286 * Slurp in all works issued via this workqueue and
3289 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3290 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3291 if (get_work_pwq(work) == pwq &&
3292 assign_work(work, rescuer, &n))
3293 pwq->stats[PWQ_STAT_RESCUED]++;
3296 if (!list_empty(&rescuer->scheduled)) {
3297 process_scheduled_works(rescuer);
3300 * The above execution of rescued work items could
3301 * have created more to rescue through
3302 * pwq_activate_first_inactive() or chained
3303 * queueing. Let's put @pwq back on mayday list so
3304 * that such back-to-back work items, which may be
3305 * being used to relieve memory pressure, don't
3306 * incur MAYDAY_INTERVAL delay inbetween.
3308 if (pwq->nr_active && need_to_create_worker(pool)) {
3309 raw_spin_lock(&wq_mayday_lock);
3311 * Queue iff we aren't racing destruction
3312 * and somebody else hasn't queued it already.
3314 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3316 list_add_tail(&pwq->mayday_node, &wq->maydays);
3318 raw_spin_unlock(&wq_mayday_lock);
3323 * Put the reference grabbed by send_mayday(). @pool won't
3324 * go away while we're still attached to it.
3329 * Leave this pool. Notify regular workers; otherwise, we end up
3330 * with 0 concurrency and stalling the execution.
3334 raw_spin_unlock_irq(&pool->lock);
3336 worker_detach_from_pool(rescuer);
3338 raw_spin_lock_irq(&wq_mayday_lock);
3341 raw_spin_unlock_irq(&wq_mayday_lock);
3344 __set_current_state(TASK_RUNNING);
3345 set_pf_worker(false);
3349 /* rescuers should never participate in concurrency management */
3350 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3356 * check_flush_dependency - check for flush dependency sanity
3357 * @target_wq: workqueue being flushed
3358 * @target_work: work item being flushed (NULL for workqueue flushes)
3360 * %current is trying to flush the whole @target_wq or @target_work on it.
3361 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3362 * reclaiming memory or running on a workqueue which doesn't have
3363 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3366 static void check_flush_dependency(struct workqueue_struct *target_wq,
3367 struct work_struct *target_work)
3369 work_func_t target_func = target_work ? target_work->func : NULL;
3370 struct worker *worker;
3372 if (target_wq->flags & WQ_MEM_RECLAIM)
3375 worker = current_wq_worker();
3377 WARN_ONCE(current->flags & PF_MEMALLOC,
3378 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3379 current->pid, current->comm, target_wq->name, target_func);
3380 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3381 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3382 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3383 worker->current_pwq->wq->name, worker->current_func,
3384 target_wq->name, target_func);
3388 struct work_struct work;
3389 struct completion done;
3390 struct task_struct *task; /* purely informational */
3393 static void wq_barrier_func(struct work_struct *work)
3395 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3396 complete(&barr->done);
3400 * insert_wq_barrier - insert a barrier work
3401 * @pwq: pwq to insert barrier into
3402 * @barr: wq_barrier to insert
3403 * @target: target work to attach @barr to
3404 * @worker: worker currently executing @target, NULL if @target is not executing
3406 * @barr is linked to @target such that @barr is completed only after
3407 * @target finishes execution. Please note that the ordering
3408 * guarantee is observed only with respect to @target and on the local
3411 * Currently, a queued barrier can't be canceled. This is because
3412 * try_to_grab_pending() can't determine whether the work to be
3413 * grabbed is at the head of the queue and thus can't clear LINKED
3414 * flag of the previous work while there must be a valid next work
3415 * after a work with LINKED flag set.
3417 * Note that when @worker is non-NULL, @target may be modified
3418 * underneath us, so we can't reliably determine pwq from @target.
3421 * raw_spin_lock_irq(pool->lock).
3423 static void insert_wq_barrier(struct pool_workqueue *pwq,
3424 struct wq_barrier *barr,
3425 struct work_struct *target, struct worker *worker)
3427 unsigned int work_flags = 0;
3428 unsigned int work_color;
3429 struct list_head *head;
3432 * debugobject calls are safe here even with pool->lock locked
3433 * as we know for sure that this will not trigger any of the
3434 * checks and call back into the fixup functions where we
3437 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3438 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3440 init_completion_map(&barr->done, &target->lockdep_map);
3442 barr->task = current;
3444 /* The barrier work item does not participate in nr_active. */
3445 work_flags |= WORK_STRUCT_INACTIVE;
3448 * If @target is currently being executed, schedule the
3449 * barrier to the worker; otherwise, put it after @target.
3452 head = worker->scheduled.next;
3453 work_color = worker->current_color;
3455 unsigned long *bits = work_data_bits(target);
3457 head = target->entry.next;
3458 /* there can already be other linked works, inherit and set */
3459 work_flags |= *bits & WORK_STRUCT_LINKED;
3460 work_color = get_work_color(*bits);
3461 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3464 pwq->nr_in_flight[work_color]++;
3465 work_flags |= work_color_to_flags(work_color);
3467 insert_work(pwq, &barr->work, head, work_flags);
3471 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3472 * @wq: workqueue being flushed
3473 * @flush_color: new flush color, < 0 for no-op
3474 * @work_color: new work color, < 0 for no-op
3476 * Prepare pwqs for workqueue flushing.
3478 * If @flush_color is non-negative, flush_color on all pwqs should be
3479 * -1. If no pwq has in-flight commands at the specified color, all
3480 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3481 * has in flight commands, its pwq->flush_color is set to
3482 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3483 * wakeup logic is armed and %true is returned.
3485 * The caller should have initialized @wq->first_flusher prior to
3486 * calling this function with non-negative @flush_color. If
3487 * @flush_color is negative, no flush color update is done and %false
3490 * If @work_color is non-negative, all pwqs should have the same
3491 * work_color which is previous to @work_color and all will be
3492 * advanced to @work_color.
3495 * mutex_lock(wq->mutex).
3498 * %true if @flush_color >= 0 and there's something to flush. %false
3501 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3502 int flush_color, int work_color)
3505 struct pool_workqueue *pwq;
3507 if (flush_color >= 0) {
3508 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3509 atomic_set(&wq->nr_pwqs_to_flush, 1);
3512 for_each_pwq(pwq, wq) {
3513 struct worker_pool *pool = pwq->pool;
3515 raw_spin_lock_irq(&pool->lock);
3517 if (flush_color >= 0) {
3518 WARN_ON_ONCE(pwq->flush_color != -1);
3520 if (pwq->nr_in_flight[flush_color]) {
3521 pwq->flush_color = flush_color;
3522 atomic_inc(&wq->nr_pwqs_to_flush);
3527 if (work_color >= 0) {
3528 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3529 pwq->work_color = work_color;
3532 raw_spin_unlock_irq(&pool->lock);
3535 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3536 complete(&wq->first_flusher->done);
3542 * __flush_workqueue - ensure that any scheduled work has run to completion.
3543 * @wq: workqueue to flush
3545 * This function sleeps until all work items which were queued on entry
3546 * have finished execution, but it is not livelocked by new incoming ones.
3548 void __flush_workqueue(struct workqueue_struct *wq)
3550 struct wq_flusher this_flusher = {
3551 .list = LIST_HEAD_INIT(this_flusher.list),
3553 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3557 if (WARN_ON(!wq_online))
3560 lock_map_acquire(&wq->lockdep_map);
3561 lock_map_release(&wq->lockdep_map);
3563 mutex_lock(&wq->mutex);
3566 * Start-to-wait phase
3568 next_color = work_next_color(wq->work_color);
3570 if (next_color != wq->flush_color) {
3572 * Color space is not full. The current work_color
3573 * becomes our flush_color and work_color is advanced
3576 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3577 this_flusher.flush_color = wq->work_color;
3578 wq->work_color = next_color;
3580 if (!wq->first_flusher) {
3581 /* no flush in progress, become the first flusher */
3582 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3584 wq->first_flusher = &this_flusher;
3586 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3588 /* nothing to flush, done */
3589 wq->flush_color = next_color;
3590 wq->first_flusher = NULL;
3595 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3596 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3597 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3601 * Oops, color space is full, wait on overflow queue.
3602 * The next flush completion will assign us
3603 * flush_color and transfer to flusher_queue.
3605 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3608 check_flush_dependency(wq, NULL);
3610 mutex_unlock(&wq->mutex);
3612 wait_for_completion(&this_flusher.done);
3615 * Wake-up-and-cascade phase
3617 * First flushers are responsible for cascading flushes and
3618 * handling overflow. Non-first flushers can simply return.
3620 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3623 mutex_lock(&wq->mutex);
3625 /* we might have raced, check again with mutex held */
3626 if (wq->first_flusher != &this_flusher)
3629 WRITE_ONCE(wq->first_flusher, NULL);
3631 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3632 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3635 struct wq_flusher *next, *tmp;
3637 /* complete all the flushers sharing the current flush color */
3638 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3639 if (next->flush_color != wq->flush_color)
3641 list_del_init(&next->list);
3642 complete(&next->done);
3645 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3646 wq->flush_color != work_next_color(wq->work_color));
3648 /* this flush_color is finished, advance by one */
3649 wq->flush_color = work_next_color(wq->flush_color);
3651 /* one color has been freed, handle overflow queue */
3652 if (!list_empty(&wq->flusher_overflow)) {
3654 * Assign the same color to all overflowed
3655 * flushers, advance work_color and append to
3656 * flusher_queue. This is the start-to-wait
3657 * phase for these overflowed flushers.
3659 list_for_each_entry(tmp, &wq->flusher_overflow, list)
3660 tmp->flush_color = wq->work_color;
3662 wq->work_color = work_next_color(wq->work_color);
3664 list_splice_tail_init(&wq->flusher_overflow,
3665 &wq->flusher_queue);
3666 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3669 if (list_empty(&wq->flusher_queue)) {
3670 WARN_ON_ONCE(wq->flush_color != wq->work_color);
3675 * Need to flush more colors. Make the next flusher
3676 * the new first flusher and arm pwqs.
3678 WARN_ON_ONCE(wq->flush_color == wq->work_color);
3679 WARN_ON_ONCE(wq->flush_color != next->flush_color);
3681 list_del_init(&next->list);
3682 wq->first_flusher = next;
3684 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3688 * Meh... this color is already done, clear first
3689 * flusher and repeat cascading.
3691 wq->first_flusher = NULL;
3695 mutex_unlock(&wq->mutex);
3697 EXPORT_SYMBOL(__flush_workqueue);
3700 * drain_workqueue - drain a workqueue
3701 * @wq: workqueue to drain
3703 * Wait until the workqueue becomes empty. While draining is in progress,
3704 * only chain queueing is allowed. IOW, only currently pending or running
3705 * work items on @wq can queue further work items on it. @wq is flushed
3706 * repeatedly until it becomes empty. The number of flushing is determined
3707 * by the depth of chaining and should be relatively short. Whine if it
3710 void drain_workqueue(struct workqueue_struct *wq)
3712 unsigned int flush_cnt = 0;
3713 struct pool_workqueue *pwq;
3716 * __queue_work() needs to test whether there are drainers, is much
3717 * hotter than drain_workqueue() and already looks at @wq->flags.
3718 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3720 mutex_lock(&wq->mutex);
3721 if (!wq->nr_drainers++)
3722 wq->flags |= __WQ_DRAINING;
3723 mutex_unlock(&wq->mutex);
3725 __flush_workqueue(wq);
3727 mutex_lock(&wq->mutex);
3729 for_each_pwq(pwq, wq) {
3732 raw_spin_lock_irq(&pwq->pool->lock);
3733 drained = pwq_is_empty(pwq);
3734 raw_spin_unlock_irq(&pwq->pool->lock);
3739 if (++flush_cnt == 10 ||
3740 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3741 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3742 wq->name, __func__, flush_cnt);
3744 mutex_unlock(&wq->mutex);
3748 if (!--wq->nr_drainers)
3749 wq->flags &= ~__WQ_DRAINING;
3750 mutex_unlock(&wq->mutex);
3752 EXPORT_SYMBOL_GPL(drain_workqueue);
3754 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3757 struct worker *worker = NULL;
3758 struct worker_pool *pool;
3759 struct pool_workqueue *pwq;
3764 pool = get_work_pool(work);
3770 raw_spin_lock_irq(&pool->lock);
3771 /* see the comment in try_to_grab_pending() with the same code */
3772 pwq = get_work_pwq(work);
3774 if (unlikely(pwq->pool != pool))
3777 worker = find_worker_executing_work(pool, work);
3780 pwq = worker->current_pwq;
3783 check_flush_dependency(pwq->wq, work);
3785 insert_wq_barrier(pwq, barr, work, worker);
3786 raw_spin_unlock_irq(&pool->lock);
3789 * Force a lock recursion deadlock when using flush_work() inside a
3790 * single-threaded or rescuer equipped workqueue.
3792 * For single threaded workqueues the deadlock happens when the work
3793 * is after the work issuing the flush_work(). For rescuer equipped
3794 * workqueues the deadlock happens when the rescuer stalls, blocking
3798 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3799 lock_map_acquire(&pwq->wq->lockdep_map);
3800 lock_map_release(&pwq->wq->lockdep_map);
3805 raw_spin_unlock_irq(&pool->lock);
3810 static bool __flush_work(struct work_struct *work, bool from_cancel)
3812 struct wq_barrier barr;
3814 if (WARN_ON(!wq_online))
3817 if (WARN_ON(!work->func))
3820 lock_map_acquire(&work->lockdep_map);
3821 lock_map_release(&work->lockdep_map);
3823 if (start_flush_work(work, &barr, from_cancel)) {
3824 wait_for_completion(&barr.done);
3825 destroy_work_on_stack(&barr.work);
3833 * flush_work - wait for a work to finish executing the last queueing instance
3834 * @work: the work to flush
3836 * Wait until @work has finished execution. @work is guaranteed to be idle
3837 * on return if it hasn't been requeued since flush started.
3840 * %true if flush_work() waited for the work to finish execution,
3841 * %false if it was already idle.
3843 bool flush_work(struct work_struct *work)
3845 return __flush_work(work, false);
3847 EXPORT_SYMBOL_GPL(flush_work);
3850 wait_queue_entry_t wait;
3851 struct work_struct *work;
3854 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3856 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3858 if (cwait->work != key)
3860 return autoremove_wake_function(wait, mode, sync, key);
3863 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3865 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3866 unsigned long flags;
3870 ret = try_to_grab_pending(work, is_dwork, &flags);
3872 * If someone else is already canceling, wait for it to
3873 * finish. flush_work() doesn't work for PREEMPT_NONE
3874 * because we may get scheduled between @work's completion
3875 * and the other canceling task resuming and clearing
3876 * CANCELING - flush_work() will return false immediately
3877 * as @work is no longer busy, try_to_grab_pending() will
3878 * return -ENOENT as @work is still being canceled and the
3879 * other canceling task won't be able to clear CANCELING as
3880 * we're hogging the CPU.
3882 * Let's wait for completion using a waitqueue. As this
3883 * may lead to the thundering herd problem, use a custom
3884 * wake function which matches @work along with exclusive
3887 if (unlikely(ret == -ENOENT)) {
3888 struct cwt_wait cwait;
3890 init_wait(&cwait.wait);
3891 cwait.wait.func = cwt_wakefn;
3894 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3895 TASK_UNINTERRUPTIBLE);
3896 if (work_is_canceling(work))
3898 finish_wait(&cancel_waitq, &cwait.wait);
3900 } while (unlikely(ret < 0));
3902 /* tell other tasks trying to grab @work to back off */
3903 mark_work_canceling(work);
3904 local_irq_restore(flags);
3907 * This allows canceling during early boot. We know that @work
3911 __flush_work(work, true);
3913 clear_work_data(work);
3916 * Paired with prepare_to_wait() above so that either
3917 * waitqueue_active() is visible here or !work_is_canceling() is
3921 if (waitqueue_active(&cancel_waitq))
3922 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3928 * cancel_work_sync - cancel a work and wait for it to finish
3929 * @work: the work to cancel
3931 * Cancel @work and wait for its execution to finish. This function
3932 * can be used even if the work re-queues itself or migrates to
3933 * another workqueue. On return from this function, @work is
3934 * guaranteed to be not pending or executing on any CPU.
3936 * cancel_work_sync(&delayed_work->work) must not be used for
3937 * delayed_work's. Use cancel_delayed_work_sync() instead.
3939 * The caller must ensure that the workqueue on which @work was last
3940 * queued can't be destroyed before this function returns.
3943 * %true if @work was pending, %false otherwise.
3945 bool cancel_work_sync(struct work_struct *work)
3947 return __cancel_work_timer(work, false);
3949 EXPORT_SYMBOL_GPL(cancel_work_sync);
3952 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3953 * @dwork: the delayed work to flush
3955 * Delayed timer is cancelled and the pending work is queued for
3956 * immediate execution. Like flush_work(), this function only
3957 * considers the last queueing instance of @dwork.
3960 * %true if flush_work() waited for the work to finish execution,
3961 * %false if it was already idle.
3963 bool flush_delayed_work(struct delayed_work *dwork)
3965 local_irq_disable();
3966 if (del_timer_sync(&dwork->timer))
3967 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3969 return flush_work(&dwork->work);
3971 EXPORT_SYMBOL(flush_delayed_work);
3974 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3975 * @rwork: the rcu work to flush
3978 * %true if flush_rcu_work() waited for the work to finish execution,
3979 * %false if it was already idle.
3981 bool flush_rcu_work(struct rcu_work *rwork)
3983 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3985 flush_work(&rwork->work);
3988 return flush_work(&rwork->work);
3991 EXPORT_SYMBOL(flush_rcu_work);
3993 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3995 unsigned long flags;
3999 ret = try_to_grab_pending(work, is_dwork, &flags);
4000 } while (unlikely(ret == -EAGAIN));
4002 if (unlikely(ret < 0))
4005 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
4006 local_irq_restore(flags);
4011 * See cancel_delayed_work()
4013 bool cancel_work(struct work_struct *work)
4015 return __cancel_work(work, false);
4017 EXPORT_SYMBOL(cancel_work);
4020 * cancel_delayed_work - cancel a delayed work
4021 * @dwork: delayed_work to cancel
4023 * Kill off a pending delayed_work.
4025 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4029 * The work callback function may still be running on return, unless
4030 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4031 * use cancel_delayed_work_sync() to wait on it.
4033 * This function is safe to call from any context including IRQ handler.
4035 bool cancel_delayed_work(struct delayed_work *dwork)
4037 return __cancel_work(&dwork->work, true);
4039 EXPORT_SYMBOL(cancel_delayed_work);
4042 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4043 * @dwork: the delayed work cancel
4045 * This is cancel_work_sync() for delayed works.
4048 * %true if @dwork was pending, %false otherwise.
4050 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4052 return __cancel_work_timer(&dwork->work, true);
4054 EXPORT_SYMBOL(cancel_delayed_work_sync);
4057 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4058 * @func: the function to call
4060 * schedule_on_each_cpu() executes @func on each online CPU using the
4061 * system workqueue and blocks until all CPUs have completed.
4062 * schedule_on_each_cpu() is very slow.
4065 * 0 on success, -errno on failure.
4067 int schedule_on_each_cpu(work_func_t func)
4070 struct work_struct __percpu *works;
4072 works = alloc_percpu(struct work_struct);
4078 for_each_online_cpu(cpu) {
4079 struct work_struct *work = per_cpu_ptr(works, cpu);
4081 INIT_WORK(work, func);
4082 schedule_work_on(cpu, work);
4085 for_each_online_cpu(cpu)
4086 flush_work(per_cpu_ptr(works, cpu));
4094 * execute_in_process_context - reliably execute the routine with user context
4095 * @fn: the function to execute
4096 * @ew: guaranteed storage for the execute work structure (must
4097 * be available when the work executes)
4099 * Executes the function immediately if process context is available,
4100 * otherwise schedules the function for delayed execution.
4102 * Return: 0 - function was executed
4103 * 1 - function was scheduled for execution
4105 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4107 if (!in_interrupt()) {
4112 INIT_WORK(&ew->work, fn);
4113 schedule_work(&ew->work);
4117 EXPORT_SYMBOL_GPL(execute_in_process_context);
4120 * free_workqueue_attrs - free a workqueue_attrs
4121 * @attrs: workqueue_attrs to free
4123 * Undo alloc_workqueue_attrs().
4125 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4128 free_cpumask_var(attrs->cpumask);
4129 free_cpumask_var(attrs->__pod_cpumask);
4135 * alloc_workqueue_attrs - allocate a workqueue_attrs
4137 * Allocate a new workqueue_attrs, initialize with default settings and
4140 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4142 struct workqueue_attrs *alloc_workqueue_attrs(void)
4144 struct workqueue_attrs *attrs;
4146 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4149 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4151 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4154 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4155 attrs->affn_scope = WQ_AFFN_DFL;
4158 free_workqueue_attrs(attrs);
4162 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4163 const struct workqueue_attrs *from)
4165 to->nice = from->nice;
4166 cpumask_copy(to->cpumask, from->cpumask);
4167 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4168 to->affn_strict = from->affn_strict;
4171 * Unlike hash and equality test, copying shouldn't ignore wq-only
4172 * fields as copying is used for both pool and wq attrs. Instead,
4173 * get_unbound_pool() explicitly clears the fields.
4175 to->affn_scope = from->affn_scope;
4176 to->ordered = from->ordered;
4180 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4181 * comments in 'struct workqueue_attrs' definition.
4183 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4185 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4186 attrs->ordered = false;
4189 /* hash value of the content of @attr */
4190 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4194 hash = jhash_1word(attrs->nice, hash);
4195 hash = jhash(cpumask_bits(attrs->cpumask),
4196 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4197 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4198 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4199 hash = jhash_1word(attrs->affn_strict, hash);
4203 /* content equality test */
4204 static bool wqattrs_equal(const struct workqueue_attrs *a,
4205 const struct workqueue_attrs *b)
4207 if (a->nice != b->nice)
4209 if (!cpumask_equal(a->cpumask, b->cpumask))
4211 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4213 if (a->affn_strict != b->affn_strict)
4218 /* Update @attrs with actually available CPUs */
4219 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4220 const cpumask_t *unbound_cpumask)
4223 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4224 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4227 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4228 if (unlikely(cpumask_empty(attrs->cpumask)))
4229 cpumask_copy(attrs->cpumask, unbound_cpumask);
4232 /* find wq_pod_type to use for @attrs */
4233 static const struct wq_pod_type *
4234 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4236 enum wq_affn_scope scope;
4237 struct wq_pod_type *pt;
4239 /* to synchronize access to wq_affn_dfl */
4240 lockdep_assert_held(&wq_pool_mutex);
4242 if (attrs->affn_scope == WQ_AFFN_DFL)
4243 scope = wq_affn_dfl;
4245 scope = attrs->affn_scope;
4247 pt = &wq_pod_types[scope];
4249 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4250 likely(pt->nr_pods))
4254 * Before workqueue_init_topology(), only SYSTEM is available which is
4255 * initialized in workqueue_init_early().
4257 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4258 BUG_ON(!pt->nr_pods);
4263 * init_worker_pool - initialize a newly zalloc'd worker_pool
4264 * @pool: worker_pool to initialize
4266 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4268 * Return: 0 on success, -errno on failure. Even on failure, all fields
4269 * inside @pool proper are initialized and put_unbound_pool() can be called
4270 * on @pool safely to release it.
4272 static int init_worker_pool(struct worker_pool *pool)
4274 raw_spin_lock_init(&pool->lock);
4277 pool->node = NUMA_NO_NODE;
4278 pool->flags |= POOL_DISASSOCIATED;
4279 pool->watchdog_ts = jiffies;
4280 INIT_LIST_HEAD(&pool->worklist);
4281 INIT_LIST_HEAD(&pool->idle_list);
4282 hash_init(pool->busy_hash);
4284 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4285 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4287 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4289 INIT_LIST_HEAD(&pool->workers);
4290 INIT_LIST_HEAD(&pool->dying_workers);
4292 ida_init(&pool->worker_ida);
4293 INIT_HLIST_NODE(&pool->hash_node);
4296 /* shouldn't fail above this point */
4297 pool->attrs = alloc_workqueue_attrs();
4301 wqattrs_clear_for_pool(pool->attrs);
4306 #ifdef CONFIG_LOCKDEP
4307 static void wq_init_lockdep(struct workqueue_struct *wq)
4311 lockdep_register_key(&wq->key);
4312 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4314 lock_name = wq->name;
4316 wq->lock_name = lock_name;
4317 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4320 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4322 lockdep_unregister_key(&wq->key);
4325 static void wq_free_lockdep(struct workqueue_struct *wq)
4327 if (wq->lock_name != wq->name)
4328 kfree(wq->lock_name);
4331 static void wq_init_lockdep(struct workqueue_struct *wq)
4335 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4339 static void wq_free_lockdep(struct workqueue_struct *wq)
4344 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4348 for_each_node(node) {
4349 kfree(nna_ar[node]);
4350 nna_ar[node] = NULL;
4353 kfree(nna_ar[nr_node_ids]);
4354 nna_ar[nr_node_ids] = NULL;
4357 static void init_node_nr_active(struct wq_node_nr_active *nna)
4359 atomic_set(&nna->nr, 0);
4360 raw_spin_lock_init(&nna->lock);
4361 INIT_LIST_HEAD(&nna->pending_pwqs);
4365 * Each node's nr_active counter will be accessed mostly from its own node and
4366 * should be allocated in the node.
4368 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4370 struct wq_node_nr_active *nna;
4373 for_each_node(node) {
4374 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4377 init_node_nr_active(nna);
4381 /* [nr_node_ids] is used as the fallback */
4382 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4385 init_node_nr_active(nna);
4386 nna_ar[nr_node_ids] = nna;
4391 free_node_nr_active(nna_ar);
4395 static void rcu_free_wq(struct rcu_head *rcu)
4397 struct workqueue_struct *wq =
4398 container_of(rcu, struct workqueue_struct, rcu);
4400 if (wq->flags & WQ_UNBOUND)
4401 free_node_nr_active(wq->node_nr_active);
4403 wq_free_lockdep(wq);
4404 free_percpu(wq->cpu_pwq);
4405 free_workqueue_attrs(wq->unbound_attrs);
4409 static void rcu_free_pool(struct rcu_head *rcu)
4411 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4413 ida_destroy(&pool->worker_ida);
4414 free_workqueue_attrs(pool->attrs);
4419 * put_unbound_pool - put a worker_pool
4420 * @pool: worker_pool to put
4422 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4423 * safe manner. get_unbound_pool() calls this function on its failure path
4424 * and this function should be able to release pools which went through,
4425 * successfully or not, init_worker_pool().
4427 * Should be called with wq_pool_mutex held.
4429 static void put_unbound_pool(struct worker_pool *pool)
4431 DECLARE_COMPLETION_ONSTACK(detach_completion);
4432 struct worker *worker;
4433 LIST_HEAD(cull_list);
4435 lockdep_assert_held(&wq_pool_mutex);
4441 if (WARN_ON(!(pool->cpu < 0)) ||
4442 WARN_ON(!list_empty(&pool->worklist)))
4445 /* release id and unhash */
4447 idr_remove(&worker_pool_idr, pool->id);
4448 hash_del(&pool->hash_node);
4451 * Become the manager and destroy all workers. This prevents
4452 * @pool's workers from blocking on attach_mutex. We're the last
4453 * manager and @pool gets freed with the flag set.
4455 * Having a concurrent manager is quite unlikely to happen as we can
4456 * only get here with
4457 * pwq->refcnt == pool->refcnt == 0
4458 * which implies no work queued to the pool, which implies no worker can
4459 * become the manager. However a worker could have taken the role of
4460 * manager before the refcnts dropped to 0, since maybe_create_worker()
4464 rcuwait_wait_event(&manager_wait,
4465 !(pool->flags & POOL_MANAGER_ACTIVE),
4466 TASK_UNINTERRUPTIBLE);
4468 mutex_lock(&wq_pool_attach_mutex);
4469 raw_spin_lock_irq(&pool->lock);
4470 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4471 pool->flags |= POOL_MANAGER_ACTIVE;
4474 raw_spin_unlock_irq(&pool->lock);
4475 mutex_unlock(&wq_pool_attach_mutex);
4478 while ((worker = first_idle_worker(pool)))
4479 set_worker_dying(worker, &cull_list);
4480 WARN_ON(pool->nr_workers || pool->nr_idle);
4481 raw_spin_unlock_irq(&pool->lock);
4483 wake_dying_workers(&cull_list);
4485 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4486 pool->detach_completion = &detach_completion;
4487 mutex_unlock(&wq_pool_attach_mutex);
4489 if (pool->detach_completion)
4490 wait_for_completion(pool->detach_completion);
4492 /* shut down the timers */
4493 del_timer_sync(&pool->idle_timer);
4494 cancel_work_sync(&pool->idle_cull_work);
4495 del_timer_sync(&pool->mayday_timer);
4497 /* RCU protected to allow dereferences from get_work_pool() */
4498 call_rcu(&pool->rcu, rcu_free_pool);
4502 * get_unbound_pool - get a worker_pool with the specified attributes
4503 * @attrs: the attributes of the worker_pool to get
4505 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4506 * reference count and return it. If there already is a matching
4507 * worker_pool, it will be used; otherwise, this function attempts to
4510 * Should be called with wq_pool_mutex held.
4512 * Return: On success, a worker_pool with the same attributes as @attrs.
4513 * On failure, %NULL.
4515 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4517 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4518 u32 hash = wqattrs_hash(attrs);
4519 struct worker_pool *pool;
4520 int pod, node = NUMA_NO_NODE;
4522 lockdep_assert_held(&wq_pool_mutex);
4524 /* do we already have a matching pool? */
4525 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4526 if (wqattrs_equal(pool->attrs, attrs)) {
4532 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4533 for (pod = 0; pod < pt->nr_pods; pod++) {
4534 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4535 node = pt->pod_node[pod];
4540 /* nope, create a new one */
4541 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4542 if (!pool || init_worker_pool(pool) < 0)
4546 copy_workqueue_attrs(pool->attrs, attrs);
4547 wqattrs_clear_for_pool(pool->attrs);
4549 if (worker_pool_assign_id(pool) < 0)
4552 /* create and start the initial worker */
4553 if (wq_online && !create_worker(pool))
4557 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4562 put_unbound_pool(pool);
4566 static void rcu_free_pwq(struct rcu_head *rcu)
4568 kmem_cache_free(pwq_cache,
4569 container_of(rcu, struct pool_workqueue, rcu));
4573 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4574 * refcnt and needs to be destroyed.
4576 static void pwq_release_workfn(struct kthread_work *work)
4578 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4580 struct workqueue_struct *wq = pwq->wq;
4581 struct worker_pool *pool = pwq->pool;
4582 bool is_last = false;
4585 * When @pwq is not linked, it doesn't hold any reference to the
4586 * @wq, and @wq is invalid to access.
4588 if (!list_empty(&pwq->pwqs_node)) {
4589 mutex_lock(&wq->mutex);
4590 list_del_rcu(&pwq->pwqs_node);
4591 is_last = list_empty(&wq->pwqs);
4592 mutex_unlock(&wq->mutex);
4595 if (wq->flags & WQ_UNBOUND) {
4596 mutex_lock(&wq_pool_mutex);
4597 put_unbound_pool(pool);
4598 mutex_unlock(&wq_pool_mutex);
4601 if (!list_empty(&pwq->pending_node)) {
4602 struct wq_node_nr_active *nna =
4603 wq_node_nr_active(pwq->wq, pwq->pool->node);
4605 raw_spin_lock_irq(&nna->lock);
4606 list_del_init(&pwq->pending_node);
4607 raw_spin_unlock_irq(&nna->lock);
4610 call_rcu(&pwq->rcu, rcu_free_pwq);
4613 * If we're the last pwq going away, @wq is already dead and no one
4614 * is gonna access it anymore. Schedule RCU free.
4617 wq_unregister_lockdep(wq);
4618 call_rcu(&wq->rcu, rcu_free_wq);
4622 /* initialize newly allocated @pwq which is associated with @wq and @pool */
4623 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4624 struct worker_pool *pool)
4626 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4628 memset(pwq, 0, sizeof(*pwq));
4632 pwq->flush_color = -1;
4634 INIT_LIST_HEAD(&pwq->inactive_works);
4635 INIT_LIST_HEAD(&pwq->pending_node);
4636 INIT_LIST_HEAD(&pwq->pwqs_node);
4637 INIT_LIST_HEAD(&pwq->mayday_node);
4638 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4641 /* sync @pwq with the current state of its associated wq and link it */
4642 static void link_pwq(struct pool_workqueue *pwq)
4644 struct workqueue_struct *wq = pwq->wq;
4646 lockdep_assert_held(&wq->mutex);
4648 /* may be called multiple times, ignore if already linked */
4649 if (!list_empty(&pwq->pwqs_node))
4652 /* set the matching work_color */
4653 pwq->work_color = wq->work_color;
4656 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4659 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4660 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4661 const struct workqueue_attrs *attrs)
4663 struct worker_pool *pool;
4664 struct pool_workqueue *pwq;
4666 lockdep_assert_held(&wq_pool_mutex);
4668 pool = get_unbound_pool(attrs);
4672 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4674 put_unbound_pool(pool);
4678 init_pwq(pwq, wq, pool);
4683 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4684 * @attrs: the wq_attrs of the default pwq of the target workqueue
4685 * @cpu: the target CPU
4686 * @cpu_going_down: if >= 0, the CPU to consider as offline
4688 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4689 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4690 * The result is stored in @attrs->__pod_cpumask.
4692 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4693 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4694 * intersection of the possible CPUs of @pod and @attrs->cpumask.
4696 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4698 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4701 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4702 int pod = pt->cpu_pod[cpu];
4704 /* does @pod have any online CPUs @attrs wants? */
4705 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4706 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4707 if (cpu_going_down >= 0)
4708 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4710 if (cpumask_empty(attrs->__pod_cpumask)) {
4711 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4715 /* yeap, return possible CPUs in @pod that @attrs wants */
4716 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4718 if (cpumask_empty(attrs->__pod_cpumask))
4719 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4720 "possible intersect\n");
4723 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
4724 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4725 int cpu, struct pool_workqueue *pwq)
4727 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
4728 struct pool_workqueue *old_pwq;
4730 lockdep_assert_held(&wq_pool_mutex);
4731 lockdep_assert_held(&wq->mutex);
4733 /* link_pwq() can handle duplicate calls */
4736 old_pwq = rcu_access_pointer(*slot);
4737 rcu_assign_pointer(*slot, pwq);
4741 /* context to store the prepared attrs & pwqs before applying */
4742 struct apply_wqattrs_ctx {
4743 struct workqueue_struct *wq; /* target workqueue */
4744 struct workqueue_attrs *attrs; /* attrs to apply */
4745 struct list_head list; /* queued for batching commit */
4746 struct pool_workqueue *dfl_pwq;
4747 struct pool_workqueue *pwq_tbl[];
4750 /* free the resources after success or abort */
4751 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4756 for_each_possible_cpu(cpu)
4757 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4758 put_pwq_unlocked(ctx->dfl_pwq);
4760 free_workqueue_attrs(ctx->attrs);
4766 /* allocate the attrs and pwqs for later installation */
4767 static struct apply_wqattrs_ctx *
4768 apply_wqattrs_prepare(struct workqueue_struct *wq,
4769 const struct workqueue_attrs *attrs,
4770 const cpumask_var_t unbound_cpumask)
4772 struct apply_wqattrs_ctx *ctx;
4773 struct workqueue_attrs *new_attrs;
4776 lockdep_assert_held(&wq_pool_mutex);
4778 if (WARN_ON(attrs->affn_scope < 0 ||
4779 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4780 return ERR_PTR(-EINVAL);
4782 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4784 new_attrs = alloc_workqueue_attrs();
4785 if (!ctx || !new_attrs)
4789 * If something goes wrong during CPU up/down, we'll fall back to
4790 * the default pwq covering whole @attrs->cpumask. Always create
4791 * it even if we don't use it immediately.
4793 copy_workqueue_attrs(new_attrs, attrs);
4794 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4795 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4796 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4800 for_each_possible_cpu(cpu) {
4801 if (new_attrs->ordered) {
4802 ctx->dfl_pwq->refcnt++;
4803 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4805 wq_calc_pod_cpumask(new_attrs, cpu, -1);
4806 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4807 if (!ctx->pwq_tbl[cpu])
4812 /* save the user configured attrs and sanitize it. */
4813 copy_workqueue_attrs(new_attrs, attrs);
4814 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4815 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4816 ctx->attrs = new_attrs;
4822 free_workqueue_attrs(new_attrs);
4823 apply_wqattrs_cleanup(ctx);
4824 return ERR_PTR(-ENOMEM);
4827 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4828 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4832 /* all pwqs have been created successfully, let's install'em */
4833 mutex_lock(&ctx->wq->mutex);
4835 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4837 /* save the previous pwqs and install the new ones */
4838 for_each_possible_cpu(cpu)
4839 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4841 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
4843 /* update node_nr_active->max */
4844 wq_update_node_max_active(ctx->wq, -1);
4846 mutex_unlock(&ctx->wq->mutex);
4849 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4850 const struct workqueue_attrs *attrs)
4852 struct apply_wqattrs_ctx *ctx;
4854 /* only unbound workqueues can change attributes */
4855 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4858 /* creating multiple pwqs breaks ordering guarantee */
4859 if (!list_empty(&wq->pwqs)) {
4860 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4863 wq->flags &= ~__WQ_ORDERED;
4866 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4868 return PTR_ERR(ctx);
4870 /* the ctx has been prepared successfully, let's commit it */
4871 apply_wqattrs_commit(ctx);
4872 apply_wqattrs_cleanup(ctx);
4878 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4879 * @wq: the target workqueue
4880 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4882 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4883 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4884 * work items are affine to the pod it was issued on. Older pwqs are released as
4885 * in-flight work items finish. Note that a work item which repeatedly requeues
4886 * itself back-to-back will stay on its current pwq.
4888 * Performs GFP_KERNEL allocations.
4890 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4892 * Return: 0 on success and -errno on failure.
4894 int apply_workqueue_attrs(struct workqueue_struct *wq,
4895 const struct workqueue_attrs *attrs)
4899 lockdep_assert_cpus_held();
4901 mutex_lock(&wq_pool_mutex);
4902 ret = apply_workqueue_attrs_locked(wq, attrs);
4903 mutex_unlock(&wq_pool_mutex);
4909 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4910 * @wq: the target workqueue
4911 * @cpu: the CPU to update pool association for
4912 * @hotplug_cpu: the CPU coming up or going down
4913 * @online: whether @cpu is coming up or going down
4915 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4916 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
4920 * If pod affinity can't be adjusted due to memory allocation failure, it falls
4921 * back to @wq->dfl_pwq which may not be optimal but is always correct.
4923 * Note that when the last allowed CPU of a pod goes offline for a workqueue
4924 * with a cpumask spanning multiple pods, the workers which were already
4925 * executing the work items for the workqueue will lose their CPU affinity and
4926 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
4927 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4928 * responsibility to flush the work item from CPU_DOWN_PREPARE.
4930 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4931 int hotplug_cpu, bool online)
4933 int off_cpu = online ? -1 : hotplug_cpu;
4934 struct pool_workqueue *old_pwq = NULL, *pwq;
4935 struct workqueue_attrs *target_attrs;
4937 lockdep_assert_held(&wq_pool_mutex);
4939 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
4943 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4944 * Let's use a preallocated one. The following buf is protected by
4945 * CPU hotplug exclusion.
4947 target_attrs = wq_update_pod_attrs_buf;
4949 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4950 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4952 /* nothing to do if the target cpumask matches the current pwq */
4953 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4954 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
4957 /* create a new pwq */
4958 pwq = alloc_unbound_pwq(wq, target_attrs);
4960 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4965 /* Install the new pwq. */
4966 mutex_lock(&wq->mutex);
4967 old_pwq = install_unbound_pwq(wq, cpu, pwq);
4971 mutex_lock(&wq->mutex);
4972 pwq = unbound_pwq(wq, -1);
4973 raw_spin_lock_irq(&pwq->pool->lock);
4975 raw_spin_unlock_irq(&pwq->pool->lock);
4976 old_pwq = install_unbound_pwq(wq, cpu, pwq);
4978 mutex_unlock(&wq->mutex);
4979 put_pwq_unlocked(old_pwq);
4982 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4984 bool highpri = wq->flags & WQ_HIGHPRI;
4987 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
4991 if (!(wq->flags & WQ_UNBOUND)) {
4992 for_each_possible_cpu(cpu) {
4993 struct pool_workqueue **pwq_p =
4994 per_cpu_ptr(wq->cpu_pwq, cpu);
4995 struct worker_pool *pool =
4996 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
4998 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5003 init_pwq(*pwq_p, wq, pool);
5005 mutex_lock(&wq->mutex);
5007 mutex_unlock(&wq->mutex);
5013 if (wq->flags & __WQ_ORDERED) {
5014 struct pool_workqueue *dfl_pwq;
5016 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5017 /* there should only be single pwq for ordering guarantee */
5018 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5019 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5020 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5021 "ordering guarantee broken for workqueue %s\n", wq->name);
5023 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5027 /* for unbound pwq, flush the pwq_release_worker ensures that the
5028 * pwq_release_workfn() completes before calling kfree(wq).
5031 kthread_flush_worker(pwq_release_worker);
5037 for_each_possible_cpu(cpu) {
5038 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5041 kmem_cache_free(pwq_cache, pwq);
5043 free_percpu(wq->cpu_pwq);
5049 static int wq_clamp_max_active(int max_active, unsigned int flags,
5052 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5053 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5054 max_active, name, 1, WQ_MAX_ACTIVE);
5056 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5060 * Workqueues which may be used during memory reclaim should have a rescuer
5061 * to guarantee forward progress.
5063 static int init_rescuer(struct workqueue_struct *wq)
5065 struct worker *rescuer;
5068 if (!(wq->flags & WQ_MEM_RECLAIM))
5071 rescuer = alloc_worker(NUMA_NO_NODE);
5073 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5078 rescuer->rescue_wq = wq;
5079 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5080 if (IS_ERR(rescuer->task)) {
5081 ret = PTR_ERR(rescuer->task);
5082 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5083 wq->name, ERR_PTR(ret));
5088 wq->rescuer = rescuer;
5089 if (wq->flags & WQ_UNBOUND)
5090 kthread_bind_mask(rescuer->task, wq->unbound_attrs->cpumask);
5092 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5093 wake_up_process(rescuer->task);
5099 * wq_adjust_max_active - update a wq's max_active to the current setting
5100 * @wq: target workqueue
5102 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5103 * activate inactive work items accordingly. If @wq is freezing, clear
5104 * @wq->max_active to zero.
5106 static void wq_adjust_max_active(struct workqueue_struct *wq)
5109 int new_max, new_min;
5111 lockdep_assert_held(&wq->mutex);
5113 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5117 new_max = wq->saved_max_active;
5118 new_min = wq->saved_min_active;
5121 if (wq->max_active == new_max && wq->min_active == new_min)
5125 * Update @wq->max/min_active and then kick inactive work items if more
5126 * active work items are allowed. This doesn't break work item ordering
5127 * because new work items are always queued behind existing inactive
5128 * work items if there are any.
5130 WRITE_ONCE(wq->max_active, new_max);
5131 WRITE_ONCE(wq->min_active, new_min);
5133 if (wq->flags & WQ_UNBOUND)
5134 wq_update_node_max_active(wq, -1);
5140 * Round-robin through pwq's activating the first inactive work item
5141 * until max_active is filled.
5144 struct pool_workqueue *pwq;
5147 for_each_pwq(pwq, wq) {
5148 unsigned long flags;
5150 /* can be called during early boot w/ irq disabled */
5151 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5152 if (pwq_activate_first_inactive(pwq, true)) {
5154 kick_pool(pwq->pool);
5156 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5158 } while (activated);
5162 struct workqueue_struct *alloc_workqueue(const char *fmt,
5164 int max_active, ...)
5167 struct workqueue_struct *wq;
5172 * Unbound && max_active == 1 used to imply ordered, which is no longer
5173 * the case on many machines due to per-pod pools. While
5174 * alloc_ordered_workqueue() is the right way to create an ordered
5175 * workqueue, keep the previous behavior to avoid subtle breakages.
5177 if ((flags & WQ_UNBOUND) && max_active == 1)
5178 flags |= __WQ_ORDERED;
5180 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5181 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5182 flags |= WQ_UNBOUND;
5184 /* allocate wq and format name */
5185 if (flags & WQ_UNBOUND)
5186 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5188 wq_size = sizeof(*wq);
5190 wq = kzalloc(wq_size, GFP_KERNEL);
5194 if (flags & WQ_UNBOUND) {
5195 wq->unbound_attrs = alloc_workqueue_attrs();
5196 if (!wq->unbound_attrs)
5200 va_start(args, max_active);
5201 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5204 if (name_len >= WQ_NAME_LEN)
5205 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5208 max_active = max_active ?: WQ_DFL_ACTIVE;
5209 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5213 wq->max_active = max_active;
5214 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5215 wq->saved_max_active = wq->max_active;
5216 wq->saved_min_active = wq->min_active;
5217 mutex_init(&wq->mutex);
5218 atomic_set(&wq->nr_pwqs_to_flush, 0);
5219 INIT_LIST_HEAD(&wq->pwqs);
5220 INIT_LIST_HEAD(&wq->flusher_queue);
5221 INIT_LIST_HEAD(&wq->flusher_overflow);
5222 INIT_LIST_HEAD(&wq->maydays);
5224 wq_init_lockdep(wq);
5225 INIT_LIST_HEAD(&wq->list);
5227 if (flags & WQ_UNBOUND) {
5228 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5229 goto err_unreg_lockdep;
5232 if (alloc_and_link_pwqs(wq) < 0)
5233 goto err_free_node_nr_active;
5235 if (wq_online && init_rescuer(wq) < 0)
5238 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5242 * wq_pool_mutex protects global freeze state and workqueues list.
5243 * Grab it, adjust max_active and add the new @wq to workqueues
5246 mutex_lock(&wq_pool_mutex);
5248 mutex_lock(&wq->mutex);
5249 wq_adjust_max_active(wq);
5250 mutex_unlock(&wq->mutex);
5252 list_add_tail_rcu(&wq->list, &workqueues);
5254 mutex_unlock(&wq_pool_mutex);
5258 err_free_node_nr_active:
5259 if (wq->flags & WQ_UNBOUND)
5260 free_node_nr_active(wq->node_nr_active);
5262 wq_unregister_lockdep(wq);
5263 wq_free_lockdep(wq);
5265 free_workqueue_attrs(wq->unbound_attrs);
5269 destroy_workqueue(wq);
5272 EXPORT_SYMBOL_GPL(alloc_workqueue);
5274 static bool pwq_busy(struct pool_workqueue *pwq)
5278 for (i = 0; i < WORK_NR_COLORS; i++)
5279 if (pwq->nr_in_flight[i])
5282 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5284 if (!pwq_is_empty(pwq))
5291 * destroy_workqueue - safely terminate a workqueue
5292 * @wq: target workqueue
5294 * Safely destroy a workqueue. All work currently pending will be done first.
5296 void destroy_workqueue(struct workqueue_struct *wq)
5298 struct pool_workqueue *pwq;
5302 * Remove it from sysfs first so that sanity check failure doesn't
5303 * lead to sysfs name conflicts.
5305 workqueue_sysfs_unregister(wq);
5307 /* mark the workqueue destruction is in progress */
5308 mutex_lock(&wq->mutex);
5309 wq->flags |= __WQ_DESTROYING;
5310 mutex_unlock(&wq->mutex);
5312 /* drain it before proceeding with destruction */
5313 drain_workqueue(wq);
5315 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5317 struct worker *rescuer = wq->rescuer;
5319 /* this prevents new queueing */
5320 raw_spin_lock_irq(&wq_mayday_lock);
5322 raw_spin_unlock_irq(&wq_mayday_lock);
5324 /* rescuer will empty maydays list before exiting */
5325 kthread_stop(rescuer->task);
5330 * Sanity checks - grab all the locks so that we wait for all
5331 * in-flight operations which may do put_pwq().
5333 mutex_lock(&wq_pool_mutex);
5334 mutex_lock(&wq->mutex);
5335 for_each_pwq(pwq, wq) {
5336 raw_spin_lock_irq(&pwq->pool->lock);
5337 if (WARN_ON(pwq_busy(pwq))) {
5338 pr_warn("%s: %s has the following busy pwq\n",
5339 __func__, wq->name);
5341 raw_spin_unlock_irq(&pwq->pool->lock);
5342 mutex_unlock(&wq->mutex);
5343 mutex_unlock(&wq_pool_mutex);
5344 show_one_workqueue(wq);
5347 raw_spin_unlock_irq(&pwq->pool->lock);
5349 mutex_unlock(&wq->mutex);
5352 * wq list is used to freeze wq, remove from list after
5353 * flushing is complete in case freeze races us.
5355 list_del_rcu(&wq->list);
5356 mutex_unlock(&wq_pool_mutex);
5359 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5360 * to put the base refs. @wq will be auto-destroyed from the last
5361 * pwq_put. RCU read lock prevents @wq from going away from under us.
5365 for_each_possible_cpu(cpu) {
5366 put_pwq_unlocked(unbound_pwq(wq, cpu));
5367 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5370 put_pwq_unlocked(unbound_pwq(wq, -1));
5371 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5375 EXPORT_SYMBOL_GPL(destroy_workqueue);
5378 * workqueue_set_max_active - adjust max_active of a workqueue
5379 * @wq: target workqueue
5380 * @max_active: new max_active value.
5382 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5386 * Don't call from IRQ context.
5388 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5390 /* disallow meddling with max_active for ordered workqueues */
5391 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5394 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5396 mutex_lock(&wq->mutex);
5398 wq->flags &= ~__WQ_ORDERED;
5399 wq->saved_max_active = max_active;
5400 if (wq->flags & WQ_UNBOUND)
5401 wq->saved_min_active = min(wq->saved_min_active, max_active);
5403 wq_adjust_max_active(wq);
5405 mutex_unlock(&wq->mutex);
5407 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5410 * current_work - retrieve %current task's work struct
5412 * Determine if %current task is a workqueue worker and what it's working on.
5413 * Useful to find out the context that the %current task is running in.
5415 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5417 struct work_struct *current_work(void)
5419 struct worker *worker = current_wq_worker();
5421 return worker ? worker->current_work : NULL;
5423 EXPORT_SYMBOL(current_work);
5426 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5428 * Determine whether %current is a workqueue rescuer. Can be used from
5429 * work functions to determine whether it's being run off the rescuer task.
5431 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5433 bool current_is_workqueue_rescuer(void)
5435 struct worker *worker = current_wq_worker();
5437 return worker && worker->rescue_wq;
5441 * workqueue_congested - test whether a workqueue is congested
5442 * @cpu: CPU in question
5443 * @wq: target workqueue
5445 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5446 * no synchronization around this function and the test result is
5447 * unreliable and only useful as advisory hints or for debugging.
5449 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5451 * With the exception of ordered workqueues, all workqueues have per-cpu
5452 * pool_workqueues, each with its own congested state. A workqueue being
5453 * congested on one CPU doesn't mean that the workqueue is contested on any
5457 * %true if congested, %false otherwise.
5459 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5461 struct pool_workqueue *pwq;
5467 if (cpu == WORK_CPU_UNBOUND)
5468 cpu = smp_processor_id();
5470 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5471 ret = !list_empty(&pwq->inactive_works);
5478 EXPORT_SYMBOL_GPL(workqueue_congested);
5481 * work_busy - test whether a work is currently pending or running
5482 * @work: the work to be tested
5484 * Test whether @work is currently pending or running. There is no
5485 * synchronization around this function and the test result is
5486 * unreliable and only useful as advisory hints or for debugging.
5489 * OR'd bitmask of WORK_BUSY_* bits.
5491 unsigned int work_busy(struct work_struct *work)
5493 struct worker_pool *pool;
5494 unsigned long flags;
5495 unsigned int ret = 0;
5497 if (work_pending(work))
5498 ret |= WORK_BUSY_PENDING;
5501 pool = get_work_pool(work);
5503 raw_spin_lock_irqsave(&pool->lock, flags);
5504 if (find_worker_executing_work(pool, work))
5505 ret |= WORK_BUSY_RUNNING;
5506 raw_spin_unlock_irqrestore(&pool->lock, flags);
5512 EXPORT_SYMBOL_GPL(work_busy);
5515 * set_worker_desc - set description for the current work item
5516 * @fmt: printf-style format string
5517 * @...: arguments for the format string
5519 * This function can be called by a running work function to describe what
5520 * the work item is about. If the worker task gets dumped, this
5521 * information will be printed out together to help debugging. The
5522 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5524 void set_worker_desc(const char *fmt, ...)
5526 struct worker *worker = current_wq_worker();
5530 va_start(args, fmt);
5531 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5535 EXPORT_SYMBOL_GPL(set_worker_desc);
5538 * print_worker_info - print out worker information and description
5539 * @log_lvl: the log level to use when printing
5540 * @task: target task
5542 * If @task is a worker and currently executing a work item, print out the
5543 * name of the workqueue being serviced and worker description set with
5544 * set_worker_desc() by the currently executing work item.
5546 * This function can be safely called on any task as long as the
5547 * task_struct itself is accessible. While safe, this function isn't
5548 * synchronized and may print out mixups or garbages of limited length.
5550 void print_worker_info(const char *log_lvl, struct task_struct *task)
5552 work_func_t *fn = NULL;
5553 char name[WQ_NAME_LEN] = { };
5554 char desc[WORKER_DESC_LEN] = { };
5555 struct pool_workqueue *pwq = NULL;
5556 struct workqueue_struct *wq = NULL;
5557 struct worker *worker;
5559 if (!(task->flags & PF_WQ_WORKER))
5563 * This function is called without any synchronization and @task
5564 * could be in any state. Be careful with dereferences.
5566 worker = kthread_probe_data(task);
5569 * Carefully copy the associated workqueue's workfn, name and desc.
5570 * Keep the original last '\0' in case the original is garbage.
5572 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5573 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5574 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5575 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5576 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5578 if (fn || name[0] || desc[0]) {
5579 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5580 if (strcmp(name, desc))
5581 pr_cont(" (%s)", desc);
5586 static void pr_cont_pool_info(struct worker_pool *pool)
5588 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5589 if (pool->node != NUMA_NO_NODE)
5590 pr_cont(" node=%d", pool->node);
5591 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5594 struct pr_cont_work_struct {
5600 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5604 if (func == pcwsp->func) {
5608 if (pcwsp->ctr == 1)
5609 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5611 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5614 if ((long)func == -1L)
5616 pcwsp->comma = comma;
5621 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5623 if (work->func == wq_barrier_func) {
5624 struct wq_barrier *barr;
5626 barr = container_of(work, struct wq_barrier, work);
5628 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5629 pr_cont("%s BAR(%d)", comma ? "," : "",
5630 task_pid_nr(barr->task));
5633 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5634 pr_cont_work_flush(comma, work->func, pcwsp);
5638 static void show_pwq(struct pool_workqueue *pwq)
5640 struct pr_cont_work_struct pcws = { .ctr = 0, };
5641 struct worker_pool *pool = pwq->pool;
5642 struct work_struct *work;
5643 struct worker *worker;
5644 bool has_in_flight = false, has_pending = false;
5647 pr_info(" pwq %d:", pool->id);
5648 pr_cont_pool_info(pool);
5650 pr_cont(" active=%d refcnt=%d%s\n",
5651 pwq->nr_active, pwq->refcnt,
5652 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5654 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5655 if (worker->current_pwq == pwq) {
5656 has_in_flight = true;
5660 if (has_in_flight) {
5663 pr_info(" in-flight:");
5664 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5665 if (worker->current_pwq != pwq)
5668 pr_cont("%s %d%s:%ps", comma ? "," : "",
5669 task_pid_nr(worker->task),
5670 worker->rescue_wq ? "(RESCUER)" : "",
5671 worker->current_func);
5672 list_for_each_entry(work, &worker->scheduled, entry)
5673 pr_cont_work(false, work, &pcws);
5674 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5680 list_for_each_entry(work, &pool->worklist, entry) {
5681 if (get_work_pwq(work) == pwq) {
5689 pr_info(" pending:");
5690 list_for_each_entry(work, &pool->worklist, entry) {
5691 if (get_work_pwq(work) != pwq)
5694 pr_cont_work(comma, work, &pcws);
5695 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5697 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5701 if (!list_empty(&pwq->inactive_works)) {
5704 pr_info(" inactive:");
5705 list_for_each_entry(work, &pwq->inactive_works, entry) {
5706 pr_cont_work(comma, work, &pcws);
5707 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5709 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5715 * show_one_workqueue - dump state of specified workqueue
5716 * @wq: workqueue whose state will be printed
5718 void show_one_workqueue(struct workqueue_struct *wq)
5720 struct pool_workqueue *pwq;
5722 unsigned long flags;
5724 for_each_pwq(pwq, wq) {
5725 if (!pwq_is_empty(pwq)) {
5730 if (idle) /* Nothing to print for idle workqueue */
5733 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5735 for_each_pwq(pwq, wq) {
5736 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5737 if (!pwq_is_empty(pwq)) {
5739 * Defer printing to avoid deadlocks in console
5740 * drivers that queue work while holding locks
5741 * also taken in their write paths.
5743 printk_deferred_enter();
5745 printk_deferred_exit();
5747 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5749 * We could be printing a lot from atomic context, e.g.
5750 * sysrq-t -> show_all_workqueues(). Avoid triggering
5753 touch_nmi_watchdog();
5759 * show_one_worker_pool - dump state of specified worker pool
5760 * @pool: worker pool whose state will be printed
5762 static void show_one_worker_pool(struct worker_pool *pool)
5764 struct worker *worker;
5766 unsigned long flags;
5767 unsigned long hung = 0;
5769 raw_spin_lock_irqsave(&pool->lock, flags);
5770 if (pool->nr_workers == pool->nr_idle)
5773 /* How long the first pending work is waiting for a worker. */
5774 if (!list_empty(&pool->worklist))
5775 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5778 * Defer printing to avoid deadlocks in console drivers that
5779 * queue work while holding locks also taken in their write
5782 printk_deferred_enter();
5783 pr_info("pool %d:", pool->id);
5784 pr_cont_pool_info(pool);
5785 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5787 pr_cont(" manager: %d",
5788 task_pid_nr(pool->manager->task));
5789 list_for_each_entry(worker, &pool->idle_list, entry) {
5790 pr_cont(" %s%d", first ? "idle: " : "",
5791 task_pid_nr(worker->task));
5795 printk_deferred_exit();
5797 raw_spin_unlock_irqrestore(&pool->lock, flags);
5799 * We could be printing a lot from atomic context, e.g.
5800 * sysrq-t -> show_all_workqueues(). Avoid triggering
5803 touch_nmi_watchdog();
5808 * show_all_workqueues - dump workqueue state
5810 * Called from a sysrq handler and prints out all busy workqueues and pools.
5812 void show_all_workqueues(void)
5814 struct workqueue_struct *wq;
5815 struct worker_pool *pool;
5820 pr_info("Showing busy workqueues and worker pools:\n");
5822 list_for_each_entry_rcu(wq, &workqueues, list)
5823 show_one_workqueue(wq);
5825 for_each_pool(pool, pi)
5826 show_one_worker_pool(pool);
5832 * show_freezable_workqueues - dump freezable workqueue state
5834 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
5837 void show_freezable_workqueues(void)
5839 struct workqueue_struct *wq;
5843 pr_info("Showing freezable workqueues that are still busy:\n");
5845 list_for_each_entry_rcu(wq, &workqueues, list) {
5846 if (!(wq->flags & WQ_FREEZABLE))
5848 show_one_workqueue(wq);
5854 /* used to show worker information through /proc/PID/{comm,stat,status} */
5855 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
5859 /* always show the actual comm */
5860 off = strscpy(buf, task->comm, size);
5864 /* stabilize PF_WQ_WORKER and worker pool association */
5865 mutex_lock(&wq_pool_attach_mutex);
5867 if (task->flags & PF_WQ_WORKER) {
5868 struct worker *worker = kthread_data(task);
5869 struct worker_pool *pool = worker->pool;
5872 raw_spin_lock_irq(&pool->lock);
5874 * ->desc tracks information (wq name or
5875 * set_worker_desc()) for the latest execution. If
5876 * current, prepend '+', otherwise '-'.
5878 if (worker->desc[0] != '\0') {
5879 if (worker->current_work)
5880 scnprintf(buf + off, size - off, "+%s",
5883 scnprintf(buf + off, size - off, "-%s",
5886 raw_spin_unlock_irq(&pool->lock);
5890 mutex_unlock(&wq_pool_attach_mutex);
5898 * There are two challenges in supporting CPU hotplug. Firstly, there
5899 * are a lot of assumptions on strong associations among work, pwq and
5900 * pool which make migrating pending and scheduled works very
5901 * difficult to implement without impacting hot paths. Secondly,
5902 * worker pools serve mix of short, long and very long running works making
5903 * blocked draining impractical.
5905 * This is solved by allowing the pools to be disassociated from the CPU
5906 * running as an unbound one and allowing it to be reattached later if the
5907 * cpu comes back online.
5910 static void unbind_workers(int cpu)
5912 struct worker_pool *pool;
5913 struct worker *worker;
5915 for_each_cpu_worker_pool(pool, cpu) {
5916 mutex_lock(&wq_pool_attach_mutex);
5917 raw_spin_lock_irq(&pool->lock);
5920 * We've blocked all attach/detach operations. Make all workers
5921 * unbound and set DISASSOCIATED. Before this, all workers
5922 * must be on the cpu. After this, they may become diasporas.
5923 * And the preemption disabled section in their sched callbacks
5924 * are guaranteed to see WORKER_UNBOUND since the code here
5925 * is on the same cpu.
5927 for_each_pool_worker(worker, pool)
5928 worker->flags |= WORKER_UNBOUND;
5930 pool->flags |= POOL_DISASSOCIATED;
5933 * The handling of nr_running in sched callbacks are disabled
5934 * now. Zap nr_running. After this, nr_running stays zero and
5935 * need_more_worker() and keep_working() are always true as
5936 * long as the worklist is not empty. This pool now behaves as
5937 * an unbound (in terms of concurrency management) pool which
5938 * are served by workers tied to the pool.
5940 pool->nr_running = 0;
5943 * With concurrency management just turned off, a busy
5944 * worker blocking could lead to lengthy stalls. Kick off
5945 * unbound chain execution of currently pending work items.
5949 raw_spin_unlock_irq(&pool->lock);
5951 for_each_pool_worker(worker, pool)
5952 unbind_worker(worker);
5954 mutex_unlock(&wq_pool_attach_mutex);
5959 * rebind_workers - rebind all workers of a pool to the associated CPU
5960 * @pool: pool of interest
5962 * @pool->cpu is coming online. Rebind all workers to the CPU.
5964 static void rebind_workers(struct worker_pool *pool)
5966 struct worker *worker;
5968 lockdep_assert_held(&wq_pool_attach_mutex);
5971 * Restore CPU affinity of all workers. As all idle workers should
5972 * be on the run-queue of the associated CPU before any local
5973 * wake-ups for concurrency management happen, restore CPU affinity
5974 * of all workers first and then clear UNBOUND. As we're called
5975 * from CPU_ONLINE, the following shouldn't fail.
5977 for_each_pool_worker(worker, pool) {
5978 kthread_set_per_cpu(worker->task, pool->cpu);
5979 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5980 pool_allowed_cpus(pool)) < 0);
5983 raw_spin_lock_irq(&pool->lock);
5985 pool->flags &= ~POOL_DISASSOCIATED;
5987 for_each_pool_worker(worker, pool) {
5988 unsigned int worker_flags = worker->flags;
5991 * We want to clear UNBOUND but can't directly call
5992 * worker_clr_flags() or adjust nr_running. Atomically
5993 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5994 * @worker will clear REBOUND using worker_clr_flags() when
5995 * it initiates the next execution cycle thus restoring
5996 * concurrency management. Note that when or whether
5997 * @worker clears REBOUND doesn't affect correctness.
5999 * WRITE_ONCE() is necessary because @worker->flags may be
6000 * tested without holding any lock in
6001 * wq_worker_running(). Without it, NOT_RUNNING test may
6002 * fail incorrectly leading to premature concurrency
6003 * management operations.
6005 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6006 worker_flags |= WORKER_REBOUND;
6007 worker_flags &= ~WORKER_UNBOUND;
6008 WRITE_ONCE(worker->flags, worker_flags);
6011 raw_spin_unlock_irq(&pool->lock);
6015 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6016 * @pool: unbound pool of interest
6017 * @cpu: the CPU which is coming up
6019 * An unbound pool may end up with a cpumask which doesn't have any online
6020 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6021 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6022 * online CPU before, cpus_allowed of all its workers should be restored.
6024 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6026 static cpumask_t cpumask;
6027 struct worker *worker;
6029 lockdep_assert_held(&wq_pool_attach_mutex);
6031 /* is @cpu allowed for @pool? */
6032 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6035 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6037 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6038 for_each_pool_worker(worker, pool)
6039 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6042 int workqueue_prepare_cpu(unsigned int cpu)
6044 struct worker_pool *pool;
6046 for_each_cpu_worker_pool(pool, cpu) {
6047 if (pool->nr_workers)
6049 if (!create_worker(pool))
6055 int workqueue_online_cpu(unsigned int cpu)
6057 struct worker_pool *pool;
6058 struct workqueue_struct *wq;
6061 mutex_lock(&wq_pool_mutex);
6063 for_each_pool(pool, pi) {
6064 mutex_lock(&wq_pool_attach_mutex);
6066 if (pool->cpu == cpu)
6067 rebind_workers(pool);
6068 else if (pool->cpu < 0)
6069 restore_unbound_workers_cpumask(pool, cpu);
6071 mutex_unlock(&wq_pool_attach_mutex);
6074 /* update pod affinity of unbound workqueues */
6075 list_for_each_entry(wq, &workqueues, list) {
6076 struct workqueue_attrs *attrs = wq->unbound_attrs;
6079 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6082 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6083 wq_update_pod(wq, tcpu, cpu, true);
6085 mutex_lock(&wq->mutex);
6086 wq_update_node_max_active(wq, -1);
6087 mutex_unlock(&wq->mutex);
6091 mutex_unlock(&wq_pool_mutex);
6095 int workqueue_offline_cpu(unsigned int cpu)
6097 struct workqueue_struct *wq;
6099 /* unbinding per-cpu workers should happen on the local CPU */
6100 if (WARN_ON(cpu != smp_processor_id()))
6103 unbind_workers(cpu);
6105 /* update pod affinity of unbound workqueues */
6106 mutex_lock(&wq_pool_mutex);
6107 list_for_each_entry(wq, &workqueues, list) {
6108 struct workqueue_attrs *attrs = wq->unbound_attrs;
6111 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6114 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6115 wq_update_pod(wq, tcpu, cpu, false);
6117 mutex_lock(&wq->mutex);
6118 wq_update_node_max_active(wq, cpu);
6119 mutex_unlock(&wq->mutex);
6122 mutex_unlock(&wq_pool_mutex);
6127 struct work_for_cpu {
6128 struct work_struct work;
6134 static void work_for_cpu_fn(struct work_struct *work)
6136 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6138 wfc->ret = wfc->fn(wfc->arg);
6142 * work_on_cpu_key - run a function in thread context on a particular cpu
6143 * @cpu: the cpu to run on
6144 * @fn: the function to run
6145 * @arg: the function arg
6146 * @key: The lock class key for lock debugging purposes
6148 * It is up to the caller to ensure that the cpu doesn't go offline.
6149 * The caller must not hold any locks which would prevent @fn from completing.
6151 * Return: The value @fn returns.
6153 long work_on_cpu_key(int cpu, long (*fn)(void *),
6154 void *arg, struct lock_class_key *key)
6156 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6158 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6159 schedule_work_on(cpu, &wfc.work);
6160 flush_work(&wfc.work);
6161 destroy_work_on_stack(&wfc.work);
6164 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6167 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6168 * @cpu: the cpu to run on
6169 * @fn: the function to run
6170 * @arg: the function argument
6171 * @key: The lock class key for lock debugging purposes
6173 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6174 * any locks which would prevent @fn from completing.
6176 * Return: The value @fn returns.
6178 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6179 void *arg, struct lock_class_key *key)
6184 if (cpu_online(cpu))
6185 ret = work_on_cpu_key(cpu, fn, arg, key);
6189 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6190 #endif /* CONFIG_SMP */
6192 #ifdef CONFIG_FREEZER
6195 * freeze_workqueues_begin - begin freezing workqueues
6197 * Start freezing workqueues. After this function returns, all freezable
6198 * workqueues will queue new works to their inactive_works list instead of
6202 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6204 void freeze_workqueues_begin(void)
6206 struct workqueue_struct *wq;
6208 mutex_lock(&wq_pool_mutex);
6210 WARN_ON_ONCE(workqueue_freezing);
6211 workqueue_freezing = true;
6213 list_for_each_entry(wq, &workqueues, list) {
6214 mutex_lock(&wq->mutex);
6215 wq_adjust_max_active(wq);
6216 mutex_unlock(&wq->mutex);
6219 mutex_unlock(&wq_pool_mutex);
6223 * freeze_workqueues_busy - are freezable workqueues still busy?
6225 * Check whether freezing is complete. This function must be called
6226 * between freeze_workqueues_begin() and thaw_workqueues().
6229 * Grabs and releases wq_pool_mutex.
6232 * %true if some freezable workqueues are still busy. %false if freezing
6235 bool freeze_workqueues_busy(void)
6238 struct workqueue_struct *wq;
6239 struct pool_workqueue *pwq;
6241 mutex_lock(&wq_pool_mutex);
6243 WARN_ON_ONCE(!workqueue_freezing);
6245 list_for_each_entry(wq, &workqueues, list) {
6246 if (!(wq->flags & WQ_FREEZABLE))
6249 * nr_active is monotonically decreasing. It's safe
6250 * to peek without lock.
6253 for_each_pwq(pwq, wq) {
6254 WARN_ON_ONCE(pwq->nr_active < 0);
6255 if (pwq->nr_active) {
6264 mutex_unlock(&wq_pool_mutex);
6269 * thaw_workqueues - thaw workqueues
6271 * Thaw workqueues. Normal queueing is restored and all collected
6272 * frozen works are transferred to their respective pool worklists.
6275 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6277 void thaw_workqueues(void)
6279 struct workqueue_struct *wq;
6281 mutex_lock(&wq_pool_mutex);
6283 if (!workqueue_freezing)
6286 workqueue_freezing = false;
6288 /* restore max_active and repopulate worklist */
6289 list_for_each_entry(wq, &workqueues, list) {
6290 mutex_lock(&wq->mutex);
6291 wq_adjust_max_active(wq);
6292 mutex_unlock(&wq->mutex);
6296 mutex_unlock(&wq_pool_mutex);
6298 #endif /* CONFIG_FREEZER */
6300 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6304 struct workqueue_struct *wq;
6305 struct apply_wqattrs_ctx *ctx, *n;
6307 lockdep_assert_held(&wq_pool_mutex);
6309 list_for_each_entry(wq, &workqueues, list) {
6310 if (!(wq->flags & WQ_UNBOUND))
6313 /* creating multiple pwqs breaks ordering guarantee */
6314 if (!list_empty(&wq->pwqs)) {
6315 if (wq->flags & __WQ_ORDERED_EXPLICIT)
6317 wq->flags &= ~__WQ_ORDERED;
6320 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6326 list_add_tail(&ctx->list, &ctxs);
6329 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6331 apply_wqattrs_commit(ctx);
6332 apply_wqattrs_cleanup(ctx);
6336 mutex_lock(&wq_pool_attach_mutex);
6337 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6338 mutex_unlock(&wq_pool_attach_mutex);
6344 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6345 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6347 * This function can be called from cpuset code to provide a set of isolated
6348 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
6349 * either cpus_read_lock or cpus_write_lock.
6351 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6353 cpumask_var_t cpumask;
6356 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6359 lockdep_assert_cpus_held();
6360 mutex_lock(&wq_pool_mutex);
6362 /* Save the current isolated cpumask & export it via sysfs */
6363 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6366 * If the operation fails, it will fall back to
6367 * wq_requested_unbound_cpumask which is initially set to
6368 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6369 * by any subsequent write to workqueue/cpumask sysfs file.
6371 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6372 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6373 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6374 ret = workqueue_apply_unbound_cpumask(cpumask);
6376 mutex_unlock(&wq_pool_mutex);
6377 free_cpumask_var(cpumask);
6381 static int parse_affn_scope(const char *val)
6385 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6386 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6392 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6394 struct workqueue_struct *wq;
6397 affn = parse_affn_scope(val);
6400 if (affn == WQ_AFFN_DFL)
6404 mutex_lock(&wq_pool_mutex);
6408 list_for_each_entry(wq, &workqueues, list) {
6409 for_each_online_cpu(cpu) {
6410 wq_update_pod(wq, cpu, cpu, true);
6414 mutex_unlock(&wq_pool_mutex);
6420 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6422 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6425 static const struct kernel_param_ops wq_affn_dfl_ops = {
6426 .set = wq_affn_dfl_set,
6427 .get = wq_affn_dfl_get,
6430 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6434 * Workqueues with WQ_SYSFS flag set is visible to userland via
6435 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6436 * following attributes.
6438 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6439 * max_active RW int : maximum number of in-flight work items
6441 * Unbound workqueues have the following extra attributes.
6443 * nice RW int : nice value of the workers
6444 * cpumask RW mask : bitmask of allowed CPUs for the workers
6445 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6446 * affinity_strict RW bool : worker CPU affinity is strict
6449 struct workqueue_struct *wq;
6453 static struct workqueue_struct *dev_to_wq(struct device *dev)
6455 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6460 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6463 struct workqueue_struct *wq = dev_to_wq(dev);
6465 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6467 static DEVICE_ATTR_RO(per_cpu);
6469 static ssize_t max_active_show(struct device *dev,
6470 struct device_attribute *attr, char *buf)
6472 struct workqueue_struct *wq = dev_to_wq(dev);
6474 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6477 static ssize_t max_active_store(struct device *dev,
6478 struct device_attribute *attr, const char *buf,
6481 struct workqueue_struct *wq = dev_to_wq(dev);
6484 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
6487 workqueue_set_max_active(wq, val);
6490 static DEVICE_ATTR_RW(max_active);
6492 static struct attribute *wq_sysfs_attrs[] = {
6493 &dev_attr_per_cpu.attr,
6494 &dev_attr_max_active.attr,
6497 ATTRIBUTE_GROUPS(wq_sysfs);
6499 static void apply_wqattrs_lock(void)
6501 /* CPUs should stay stable across pwq creations and installations */
6503 mutex_lock(&wq_pool_mutex);
6506 static void apply_wqattrs_unlock(void)
6508 mutex_unlock(&wq_pool_mutex);
6512 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
6515 struct workqueue_struct *wq = dev_to_wq(dev);
6518 mutex_lock(&wq->mutex);
6519 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
6520 mutex_unlock(&wq->mutex);
6525 /* prepare workqueue_attrs for sysfs store operations */
6526 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
6528 struct workqueue_attrs *attrs;
6530 lockdep_assert_held(&wq_pool_mutex);
6532 attrs = alloc_workqueue_attrs();
6536 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6540 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6541 const char *buf, size_t count)
6543 struct workqueue_struct *wq = dev_to_wq(dev);
6544 struct workqueue_attrs *attrs;
6547 apply_wqattrs_lock();
6549 attrs = wq_sysfs_prep_attrs(wq);
6553 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6554 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6555 ret = apply_workqueue_attrs_locked(wq, attrs);
6560 apply_wqattrs_unlock();
6561 free_workqueue_attrs(attrs);
6562 return ret ?: count;
6565 static ssize_t wq_cpumask_show(struct device *dev,
6566 struct device_attribute *attr, char *buf)
6568 struct workqueue_struct *wq = dev_to_wq(dev);
6571 mutex_lock(&wq->mutex);
6572 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6573 cpumask_pr_args(wq->unbound_attrs->cpumask));
6574 mutex_unlock(&wq->mutex);
6578 static ssize_t wq_cpumask_store(struct device *dev,
6579 struct device_attribute *attr,
6580 const char *buf, size_t count)
6582 struct workqueue_struct *wq = dev_to_wq(dev);
6583 struct workqueue_attrs *attrs;
6586 apply_wqattrs_lock();
6588 attrs = wq_sysfs_prep_attrs(wq);
6592 ret = cpumask_parse(buf, attrs->cpumask);
6594 ret = apply_workqueue_attrs_locked(wq, attrs);
6597 apply_wqattrs_unlock();
6598 free_workqueue_attrs(attrs);
6599 return ret ?: count;
6602 static ssize_t wq_affn_scope_show(struct device *dev,
6603 struct device_attribute *attr, char *buf)
6605 struct workqueue_struct *wq = dev_to_wq(dev);
6608 mutex_lock(&wq->mutex);
6609 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6610 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6611 wq_affn_names[WQ_AFFN_DFL],
6612 wq_affn_names[wq_affn_dfl]);
6614 written = scnprintf(buf, PAGE_SIZE, "%s\n",
6615 wq_affn_names[wq->unbound_attrs->affn_scope]);
6616 mutex_unlock(&wq->mutex);
6621 static ssize_t wq_affn_scope_store(struct device *dev,
6622 struct device_attribute *attr,
6623 const char *buf, size_t count)
6625 struct workqueue_struct *wq = dev_to_wq(dev);
6626 struct workqueue_attrs *attrs;
6627 int affn, ret = -ENOMEM;
6629 affn = parse_affn_scope(buf);
6633 apply_wqattrs_lock();
6634 attrs = wq_sysfs_prep_attrs(wq);
6636 attrs->affn_scope = affn;
6637 ret = apply_workqueue_attrs_locked(wq, attrs);
6639 apply_wqattrs_unlock();
6640 free_workqueue_attrs(attrs);
6641 return ret ?: count;
6644 static ssize_t wq_affinity_strict_show(struct device *dev,
6645 struct device_attribute *attr, char *buf)
6647 struct workqueue_struct *wq = dev_to_wq(dev);
6649 return scnprintf(buf, PAGE_SIZE, "%d\n",
6650 wq->unbound_attrs->affn_strict);
6653 static ssize_t wq_affinity_strict_store(struct device *dev,
6654 struct device_attribute *attr,
6655 const char *buf, size_t count)
6657 struct workqueue_struct *wq = dev_to_wq(dev);
6658 struct workqueue_attrs *attrs;
6659 int v, ret = -ENOMEM;
6661 if (sscanf(buf, "%d", &v) != 1)
6664 apply_wqattrs_lock();
6665 attrs = wq_sysfs_prep_attrs(wq);
6667 attrs->affn_strict = (bool)v;
6668 ret = apply_workqueue_attrs_locked(wq, attrs);
6670 apply_wqattrs_unlock();
6671 free_workqueue_attrs(attrs);
6672 return ret ?: count;
6675 static struct device_attribute wq_sysfs_unbound_attrs[] = {
6676 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6677 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6678 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6679 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6683 static struct bus_type wq_subsys = {
6684 .name = "workqueue",
6685 .dev_groups = wq_sysfs_groups,
6689 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6690 * @cpumask: the cpumask to set
6692 * The low-level workqueues cpumask is a global cpumask that limits
6693 * the affinity of all unbound workqueues. This function check the @cpumask
6694 * and apply it to all unbound workqueues and updates all pwqs of them.
6696 * Return: 0 - Success
6697 * -EINVAL - Invalid @cpumask
6698 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6700 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6705 * Not excluding isolated cpus on purpose.
6706 * If the user wishes to include them, we allow that.
6708 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6709 if (!cpumask_empty(cpumask)) {
6710 apply_wqattrs_lock();
6711 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
6712 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
6717 ret = workqueue_apply_unbound_cpumask(cpumask);
6720 apply_wqattrs_unlock();
6726 static ssize_t __wq_cpumask_show(struct device *dev,
6727 struct device_attribute *attr, char *buf, cpumask_var_t mask)
6731 mutex_lock(&wq_pool_mutex);
6732 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
6733 mutex_unlock(&wq_pool_mutex);
6738 static ssize_t wq_unbound_cpumask_show(struct device *dev,
6739 struct device_attribute *attr, char *buf)
6741 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
6744 static ssize_t wq_requested_cpumask_show(struct device *dev,
6745 struct device_attribute *attr, char *buf)
6747 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
6750 static ssize_t wq_isolated_cpumask_show(struct device *dev,
6751 struct device_attribute *attr, char *buf)
6753 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
6756 static ssize_t wq_unbound_cpumask_store(struct device *dev,
6757 struct device_attribute *attr, const char *buf, size_t count)
6759 cpumask_var_t cpumask;
6762 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6765 ret = cpumask_parse(buf, cpumask);
6767 ret = workqueue_set_unbound_cpumask(cpumask);
6769 free_cpumask_var(cpumask);
6770 return ret ? ret : count;
6773 static struct device_attribute wq_sysfs_cpumask_attrs[] = {
6774 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6775 wq_unbound_cpumask_store),
6776 __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
6777 __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
6781 static int __init wq_sysfs_init(void)
6783 struct device *dev_root;
6786 err = subsys_virtual_register(&wq_subsys, NULL);
6790 dev_root = bus_get_dev_root(&wq_subsys);
6792 struct device_attribute *attr;
6794 for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
6795 err = device_create_file(dev_root, attr);
6799 put_device(dev_root);
6803 core_initcall(wq_sysfs_init);
6805 static void wq_device_release(struct device *dev)
6807 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6813 * workqueue_sysfs_register - make a workqueue visible in sysfs
6814 * @wq: the workqueue to register
6816 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6817 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6818 * which is the preferred method.
6820 * Workqueue user should use this function directly iff it wants to apply
6821 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6822 * apply_workqueue_attrs() may race against userland updating the
6825 * Return: 0 on success, -errno on failure.
6827 int workqueue_sysfs_register(struct workqueue_struct *wq)
6829 struct wq_device *wq_dev;
6833 * Adjusting max_active or creating new pwqs by applying
6834 * attributes breaks ordering guarantee. Disallow exposing ordered
6837 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6840 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6845 wq_dev->dev.bus = &wq_subsys;
6846 wq_dev->dev.release = wq_device_release;
6847 dev_set_name(&wq_dev->dev, "%s", wq->name);
6850 * unbound_attrs are created separately. Suppress uevent until
6851 * everything is ready.
6853 dev_set_uevent_suppress(&wq_dev->dev, true);
6855 ret = device_register(&wq_dev->dev);
6857 put_device(&wq_dev->dev);
6862 if (wq->flags & WQ_UNBOUND) {
6863 struct device_attribute *attr;
6865 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6866 ret = device_create_file(&wq_dev->dev, attr);
6868 device_unregister(&wq_dev->dev);
6875 dev_set_uevent_suppress(&wq_dev->dev, false);
6876 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6881 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6882 * @wq: the workqueue to unregister
6884 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6886 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6888 struct wq_device *wq_dev = wq->wq_dev;
6894 device_unregister(&wq_dev->dev);
6896 #else /* CONFIG_SYSFS */
6897 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
6898 #endif /* CONFIG_SYSFS */
6901 * Workqueue watchdog.
6903 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6904 * flush dependency, a concurrency managed work item which stays RUNNING
6905 * indefinitely. Workqueue stalls can be very difficult to debug as the
6906 * usual warning mechanisms don't trigger and internal workqueue state is
6909 * Workqueue watchdog monitors all worker pools periodically and dumps
6910 * state if some pools failed to make forward progress for a while where
6911 * forward progress is defined as the first item on ->worklist changing.
6913 * This mechanism is controlled through the kernel parameter
6914 * "workqueue.watchdog_thresh" which can be updated at runtime through the
6915 * corresponding sysfs parameter file.
6917 #ifdef CONFIG_WQ_WATCHDOG
6919 static unsigned long wq_watchdog_thresh = 30;
6920 static struct timer_list wq_watchdog_timer;
6922 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6923 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6926 * Show workers that might prevent the processing of pending work items.
6927 * The only candidates are CPU-bound workers in the running state.
6928 * Pending work items should be handled by another idle worker
6929 * in all other situations.
6931 static void show_cpu_pool_hog(struct worker_pool *pool)
6933 struct worker *worker;
6934 unsigned long flags;
6937 raw_spin_lock_irqsave(&pool->lock, flags);
6939 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6940 if (task_is_running(worker->task)) {
6942 * Defer printing to avoid deadlocks in console
6943 * drivers that queue work while holding locks
6944 * also taken in their write paths.
6946 printk_deferred_enter();
6948 pr_info("pool %d:\n", pool->id);
6949 sched_show_task(worker->task);
6951 printk_deferred_exit();
6955 raw_spin_unlock_irqrestore(&pool->lock, flags);
6958 static void show_cpu_pools_hogs(void)
6960 struct worker_pool *pool;
6963 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6967 for_each_pool(pool, pi) {
6968 if (pool->cpu_stall)
6969 show_cpu_pool_hog(pool);
6976 static void wq_watchdog_reset_touched(void)
6980 wq_watchdog_touched = jiffies;
6981 for_each_possible_cpu(cpu)
6982 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6985 static void wq_watchdog_timer_fn(struct timer_list *unused)
6987 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
6988 bool lockup_detected = false;
6989 bool cpu_pool_stall = false;
6990 unsigned long now = jiffies;
6991 struct worker_pool *pool;
6999 for_each_pool(pool, pi) {
7000 unsigned long pool_ts, touched, ts;
7002 pool->cpu_stall = false;
7003 if (list_empty(&pool->worklist))
7007 * If a virtual machine is stopped by the host it can look to
7008 * the watchdog like a stall.
7010 kvm_check_and_clear_guest_paused();
7012 /* get the latest of pool and touched timestamps */
7014 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7016 touched = READ_ONCE(wq_watchdog_touched);
7017 pool_ts = READ_ONCE(pool->watchdog_ts);
7019 if (time_after(pool_ts, touched))
7025 if (time_after(now, ts + thresh)) {
7026 lockup_detected = true;
7027 if (pool->cpu >= 0) {
7028 pool->cpu_stall = true;
7029 cpu_pool_stall = true;
7031 pr_emerg("BUG: workqueue lockup - pool");
7032 pr_cont_pool_info(pool);
7033 pr_cont(" stuck for %us!\n",
7034 jiffies_to_msecs(now - pool_ts) / 1000);
7042 if (lockup_detected)
7043 show_all_workqueues();
7046 show_cpu_pools_hogs();
7048 wq_watchdog_reset_touched();
7049 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7052 notrace void wq_watchdog_touch(int cpu)
7055 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7057 wq_watchdog_touched = jiffies;
7060 static void wq_watchdog_set_thresh(unsigned long thresh)
7062 wq_watchdog_thresh = 0;
7063 del_timer_sync(&wq_watchdog_timer);
7066 wq_watchdog_thresh = thresh;
7067 wq_watchdog_reset_touched();
7068 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7072 static int wq_watchdog_param_set_thresh(const char *val,
7073 const struct kernel_param *kp)
7075 unsigned long thresh;
7078 ret = kstrtoul(val, 0, &thresh);
7083 wq_watchdog_set_thresh(thresh);
7085 wq_watchdog_thresh = thresh;
7090 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7091 .set = wq_watchdog_param_set_thresh,
7092 .get = param_get_ulong,
7095 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7098 static void wq_watchdog_init(void)
7100 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7101 wq_watchdog_set_thresh(wq_watchdog_thresh);
7104 #else /* CONFIG_WQ_WATCHDOG */
7106 static inline void wq_watchdog_init(void) { }
7108 #endif /* CONFIG_WQ_WATCHDOG */
7110 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7112 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7113 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7114 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7118 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7122 * workqueue_init_early - early init for workqueue subsystem
7124 * This is the first step of three-staged workqueue subsystem initialization and
7125 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7126 * up. It sets up all the data structures and system workqueues and allows early
7127 * boot code to create workqueues and queue/cancel work items. Actual work item
7128 * execution starts only after kthreads can be created and scheduled right
7129 * before early initcalls.
7131 void __init workqueue_init_early(void)
7133 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7134 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7137 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7139 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7140 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7141 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7143 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7144 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7145 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7146 if (!cpumask_empty(&wq_cmdline_cpumask))
7147 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7149 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7151 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7153 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7154 BUG_ON(!wq_update_pod_attrs_buf);
7157 * If nohz_full is enabled, set power efficient workqueue as unbound.
7158 * This allows workqueue items to be moved to HK CPUs.
7160 if (housekeeping_enabled(HK_TYPE_TICK))
7161 wq_power_efficient = true;
7163 /* initialize WQ_AFFN_SYSTEM pods */
7164 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7165 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7166 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7167 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7169 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7172 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7173 pt->pod_node[0] = NUMA_NO_NODE;
7176 /* initialize CPU pools */
7177 for_each_possible_cpu(cpu) {
7178 struct worker_pool *pool;
7181 for_each_cpu_worker_pool(pool, cpu) {
7182 BUG_ON(init_worker_pool(pool));
7184 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7185 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7186 pool->attrs->nice = std_nice[i++];
7187 pool->attrs->affn_strict = true;
7188 pool->node = cpu_to_node(cpu);
7191 mutex_lock(&wq_pool_mutex);
7192 BUG_ON(worker_pool_assign_id(pool));
7193 mutex_unlock(&wq_pool_mutex);
7197 /* create default unbound and ordered wq attrs */
7198 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7199 struct workqueue_attrs *attrs;
7201 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7202 attrs->nice = std_nice[i];
7203 unbound_std_wq_attrs[i] = attrs;
7206 * An ordered wq should have only one pwq as ordering is
7207 * guaranteed by max_active which is enforced by pwqs.
7209 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7210 attrs->nice = std_nice[i];
7211 attrs->ordered = true;
7212 ordered_wq_attrs[i] = attrs;
7215 system_wq = alloc_workqueue("events", 0, 0);
7216 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7217 system_long_wq = alloc_workqueue("events_long", 0, 0);
7218 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7220 system_freezable_wq = alloc_workqueue("events_freezable",
7222 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7223 WQ_POWER_EFFICIENT, 0);
7224 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7225 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7227 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7228 !system_unbound_wq || !system_freezable_wq ||
7229 !system_power_efficient_wq ||
7230 !system_freezable_power_efficient_wq);
7233 static void __init wq_cpu_intensive_thresh_init(void)
7235 unsigned long thresh;
7238 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7239 BUG_ON(IS_ERR(pwq_release_worker));
7241 /* if the user set it to a specific value, keep it */
7242 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7246 * The default of 10ms is derived from the fact that most modern (as of
7247 * 2023) processors can do a lot in 10ms and that it's just below what
7248 * most consider human-perceivable. However, the kernel also runs on a
7249 * lot slower CPUs including microcontrollers where the threshold is way
7252 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7253 * This is by no means accurate but it doesn't have to be. The mechanism
7254 * is still useful even when the threshold is fully scaled up. Also, as
7255 * the reports would usually be applicable to everyone, some machines
7256 * operating on longer thresholds won't significantly diminish their
7259 thresh = 10 * USEC_PER_MSEC;
7261 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7262 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7264 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7266 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7267 loops_per_jiffy, bogo, thresh);
7269 wq_cpu_intensive_thresh_us = thresh;
7273 * workqueue_init - bring workqueue subsystem fully online
7275 * This is the second step of three-staged workqueue subsystem initialization
7276 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7277 * been created and work items queued on them, but there are no kworkers
7278 * executing the work items yet. Populate the worker pools with the initial
7279 * workers and enable future kworker creations.
7281 void __init workqueue_init(void)
7283 struct workqueue_struct *wq;
7284 struct worker_pool *pool;
7287 wq_cpu_intensive_thresh_init();
7289 mutex_lock(&wq_pool_mutex);
7292 * Per-cpu pools created earlier could be missing node hint. Fix them
7293 * up. Also, create a rescuer for workqueues that requested it.
7295 for_each_possible_cpu(cpu) {
7296 for_each_cpu_worker_pool(pool, cpu) {
7297 pool->node = cpu_to_node(cpu);
7301 list_for_each_entry(wq, &workqueues, list) {
7302 WARN(init_rescuer(wq),
7303 "workqueue: failed to create early rescuer for %s",
7307 mutex_unlock(&wq_pool_mutex);
7309 /* create the initial workers */
7310 for_each_online_cpu(cpu) {
7311 for_each_cpu_worker_pool(pool, cpu) {
7312 pool->flags &= ~POOL_DISASSOCIATED;
7313 BUG_ON(!create_worker(pool));
7317 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7318 BUG_ON(!create_worker(pool));
7325 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7326 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7327 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7329 static void __init init_pod_type(struct wq_pod_type *pt,
7330 bool (*cpus_share_pod)(int, int))
7332 int cur, pre, cpu, pod;
7336 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7337 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7338 BUG_ON(!pt->cpu_pod);
7340 for_each_possible_cpu(cur) {
7341 for_each_possible_cpu(pre) {
7343 pt->cpu_pod[cur] = pt->nr_pods++;
7346 if (cpus_share_pod(cur, pre)) {
7347 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7353 /* init the rest to match @pt->cpu_pod[] */
7354 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7355 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7356 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7358 for (pod = 0; pod < pt->nr_pods; pod++)
7359 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7361 for_each_possible_cpu(cpu) {
7362 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7363 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7367 static bool __init cpus_dont_share(int cpu0, int cpu1)
7372 static bool __init cpus_share_smt(int cpu0, int cpu1)
7374 #ifdef CONFIG_SCHED_SMT
7375 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7381 static bool __init cpus_share_numa(int cpu0, int cpu1)
7383 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7387 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7389 * This is the third step of there-staged workqueue subsystem initialization and
7390 * invoked after SMP and topology information are fully initialized. It
7391 * initializes the unbound CPU pods accordingly.
7393 void __init workqueue_init_topology(void)
7395 struct workqueue_struct *wq;
7398 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7399 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7400 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7401 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7403 mutex_lock(&wq_pool_mutex);
7406 * Workqueues allocated earlier would have all CPUs sharing the default
7407 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7408 * combinations to apply per-pod sharing.
7410 list_for_each_entry(wq, &workqueues, list) {
7411 for_each_online_cpu(cpu)
7412 wq_update_pod(wq, cpu, cpu, true);
7413 if (wq->flags & WQ_UNBOUND) {
7414 mutex_lock(&wq->mutex);
7415 wq_update_node_max_active(wq, -1);
7416 mutex_unlock(&wq->mutex);
7420 mutex_unlock(&wq_pool_mutex);
7423 void __warn_flushing_systemwide_wq(void)
7425 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7428 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7430 static int __init workqueue_unbound_cpus_setup(char *str)
7432 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7433 cpumask_clear(&wq_cmdline_cpumask);
7434 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7439 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);