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",
389 static bool wq_topo_initialized __read_mostly = false;
392 * Per-cpu work items which run for longer than the following threshold are
393 * automatically considered CPU intensive and excluded from concurrency
394 * management to prevent them from noticeably delaying other per-cpu work items.
395 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
396 * The actual value is initialized in wq_cpu_intensive_thresh_init().
398 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
399 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
401 /* see the comment above the definition of WQ_POWER_EFFICIENT */
402 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
403 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
405 static bool wq_online; /* can kworkers be created yet? */
407 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
408 static struct workqueue_attrs *wq_update_pod_attrs_buf;
410 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
411 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
412 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
413 /* wait for manager to go away */
414 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
416 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
417 static bool workqueue_freezing; /* PL: have wqs started freezing? */
419 /* PL&A: allowable cpus for unbound wqs and work items */
420 static cpumask_var_t wq_unbound_cpumask;
422 /* PL: user requested unbound cpumask via sysfs */
423 static cpumask_var_t wq_requested_unbound_cpumask;
425 /* PL: isolated cpumask to be excluded from unbound cpumask */
426 static cpumask_var_t wq_isolated_cpumask;
428 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
429 static struct cpumask wq_cmdline_cpumask __initdata;
431 /* CPU where unbound work was last round robin scheduled from this CPU */
432 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
435 * Local execution of unbound work items is no longer guaranteed. The
436 * following always forces round-robin CPU selection on unbound work items
437 * to uncover usages which depend on it.
439 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
440 static bool wq_debug_force_rr_cpu = true;
442 static bool wq_debug_force_rr_cpu = false;
444 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
446 /* the per-cpu worker pools */
447 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
449 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
451 /* PL: hash of all unbound pools keyed by pool->attrs */
452 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
454 /* I: attributes used when instantiating standard unbound pools on demand */
455 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
457 /* I: attributes used when instantiating ordered pools on demand */
458 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
461 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
462 * process context while holding a pool lock. Bounce to a dedicated kthread
463 * worker to avoid A-A deadlocks.
465 static struct kthread_worker *pwq_release_worker __ro_after_init;
467 struct workqueue_struct *system_wq __ro_after_init;
468 EXPORT_SYMBOL(system_wq);
469 struct workqueue_struct *system_highpri_wq __ro_after_init;
470 EXPORT_SYMBOL_GPL(system_highpri_wq);
471 struct workqueue_struct *system_long_wq __ro_after_init;
472 EXPORT_SYMBOL_GPL(system_long_wq);
473 struct workqueue_struct *system_unbound_wq __ro_after_init;
474 EXPORT_SYMBOL_GPL(system_unbound_wq);
475 struct workqueue_struct *system_freezable_wq __ro_after_init;
476 EXPORT_SYMBOL_GPL(system_freezable_wq);
477 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
478 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
479 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
480 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
482 static int worker_thread(void *__worker);
483 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
484 static void show_pwq(struct pool_workqueue *pwq);
485 static void show_one_worker_pool(struct worker_pool *pool);
487 #define CREATE_TRACE_POINTS
488 #include <trace/events/workqueue.h>
490 #define assert_rcu_or_pool_mutex() \
491 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
492 !lockdep_is_held(&wq_pool_mutex), \
493 "RCU or wq_pool_mutex should be held")
495 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
496 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
497 !lockdep_is_held(&wq->mutex) && \
498 !lockdep_is_held(&wq_pool_mutex), \
499 "RCU, wq->mutex or wq_pool_mutex should be held")
501 #define for_each_cpu_worker_pool(pool, cpu) \
502 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
503 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
507 * for_each_pool - iterate through all worker_pools in the system
508 * @pool: iteration cursor
509 * @pi: integer used for iteration
511 * This must be called either with wq_pool_mutex held or RCU read
512 * locked. If the pool needs to be used beyond the locking in effect, the
513 * caller is responsible for guaranteeing that the pool stays online.
515 * The if/else clause exists only for the lockdep assertion and can be
518 #define for_each_pool(pool, pi) \
519 idr_for_each_entry(&worker_pool_idr, pool, pi) \
520 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
524 * for_each_pool_worker - iterate through all workers of a worker_pool
525 * @worker: iteration cursor
526 * @pool: worker_pool to iterate workers of
528 * This must be called with wq_pool_attach_mutex.
530 * The if/else clause exists only for the lockdep assertion and can be
533 #define for_each_pool_worker(worker, pool) \
534 list_for_each_entry((worker), &(pool)->workers, node) \
535 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
539 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
540 * @pwq: iteration cursor
541 * @wq: the target workqueue
543 * This must be called either with wq->mutex held or RCU read locked.
544 * If the pwq needs to be used beyond the locking in effect, the caller is
545 * responsible for guaranteeing that the pwq stays online.
547 * The if/else clause exists only for the lockdep assertion and can be
550 #define for_each_pwq(pwq, wq) \
551 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
552 lockdep_is_held(&(wq->mutex)))
554 #ifdef CONFIG_DEBUG_OBJECTS_WORK
556 static const struct debug_obj_descr work_debug_descr;
558 static void *work_debug_hint(void *addr)
560 return ((struct work_struct *) addr)->func;
563 static bool work_is_static_object(void *addr)
565 struct work_struct *work = addr;
567 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
571 * fixup_init is called when:
572 * - an active object is initialized
574 static bool work_fixup_init(void *addr, enum debug_obj_state state)
576 struct work_struct *work = addr;
579 case ODEBUG_STATE_ACTIVE:
580 cancel_work_sync(work);
581 debug_object_init(work, &work_debug_descr);
589 * fixup_free is called when:
590 * - an active object is freed
592 static bool work_fixup_free(void *addr, enum debug_obj_state state)
594 struct work_struct *work = addr;
597 case ODEBUG_STATE_ACTIVE:
598 cancel_work_sync(work);
599 debug_object_free(work, &work_debug_descr);
606 static const struct debug_obj_descr work_debug_descr = {
607 .name = "work_struct",
608 .debug_hint = work_debug_hint,
609 .is_static_object = work_is_static_object,
610 .fixup_init = work_fixup_init,
611 .fixup_free = work_fixup_free,
614 static inline void debug_work_activate(struct work_struct *work)
616 debug_object_activate(work, &work_debug_descr);
619 static inline void debug_work_deactivate(struct work_struct *work)
621 debug_object_deactivate(work, &work_debug_descr);
624 void __init_work(struct work_struct *work, int onstack)
627 debug_object_init_on_stack(work, &work_debug_descr);
629 debug_object_init(work, &work_debug_descr);
631 EXPORT_SYMBOL_GPL(__init_work);
633 void destroy_work_on_stack(struct work_struct *work)
635 debug_object_free(work, &work_debug_descr);
637 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
639 void destroy_delayed_work_on_stack(struct delayed_work *work)
641 destroy_timer_on_stack(&work->timer);
642 debug_object_free(&work->work, &work_debug_descr);
644 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
647 static inline void debug_work_activate(struct work_struct *work) { }
648 static inline void debug_work_deactivate(struct work_struct *work) { }
652 * worker_pool_assign_id - allocate ID and assign it to @pool
653 * @pool: the pool pointer of interest
655 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
656 * successfully, -errno on failure.
658 static int worker_pool_assign_id(struct worker_pool *pool)
662 lockdep_assert_held(&wq_pool_mutex);
664 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
673 static struct pool_workqueue __rcu **
674 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
677 return per_cpu_ptr(wq->cpu_pwq, cpu);
682 /* @cpu < 0 for dfl_pwq */
683 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
685 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
686 lockdep_is_held(&wq_pool_mutex) ||
687 lockdep_is_held(&wq->mutex));
691 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
692 * @wq: workqueue of interest
694 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
695 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
696 * default pwq is always mapped to the pool with the current effective cpumask.
698 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
700 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
703 static unsigned int work_color_to_flags(int color)
705 return color << WORK_STRUCT_COLOR_SHIFT;
708 static int get_work_color(unsigned long work_data)
710 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
711 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
714 static int work_next_color(int color)
716 return (color + 1) % WORK_NR_COLORS;
720 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
721 * contain the pointer to the queued pwq. Once execution starts, the flag
722 * is cleared and the high bits contain OFFQ flags and pool ID.
724 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
725 * and clear_work_data() can be used to set the pwq, pool or clear
726 * work->data. These functions should only be called while the work is
727 * owned - ie. while the PENDING bit is set.
729 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
730 * corresponding to a work. Pool is available once the work has been
731 * queued anywhere after initialization until it is sync canceled. pwq is
732 * available only while the work item is queued.
734 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
735 * canceled. While being canceled, a work item may have its PENDING set
736 * but stay off timer and worklist for arbitrarily long and nobody should
737 * try to steal the PENDING bit.
739 static inline void set_work_data(struct work_struct *work, unsigned long data,
742 WARN_ON_ONCE(!work_pending(work));
743 atomic_long_set(&work->data, data | flags | work_static(work));
746 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
747 unsigned long extra_flags)
749 set_work_data(work, (unsigned long)pwq,
750 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
753 static void set_work_pool_and_keep_pending(struct work_struct *work,
756 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
757 WORK_STRUCT_PENDING);
760 static void set_work_pool_and_clear_pending(struct work_struct *work,
764 * The following wmb is paired with the implied mb in
765 * test_and_set_bit(PENDING) and ensures all updates to @work made
766 * here are visible to and precede any updates by the next PENDING
770 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
772 * The following mb guarantees that previous clear of a PENDING bit
773 * will not be reordered with any speculative LOADS or STORES from
774 * work->current_func, which is executed afterwards. This possible
775 * reordering can lead to a missed execution on attempt to queue
776 * the same @work. E.g. consider this case:
779 * ---------------------------- --------------------------------
781 * 1 STORE event_indicated
782 * 2 queue_work_on() {
783 * 3 test_and_set_bit(PENDING)
784 * 4 } set_..._and_clear_pending() {
785 * 5 set_work_data() # clear bit
787 * 7 work->current_func() {
788 * 8 LOAD event_indicated
791 * Without an explicit full barrier speculative LOAD on line 8 can
792 * be executed before CPU#0 does STORE on line 1. If that happens,
793 * CPU#0 observes the PENDING bit is still set and new execution of
794 * a @work is not queued in a hope, that CPU#1 will eventually
795 * finish the queued @work. Meanwhile CPU#1 does not see
796 * event_indicated is set, because speculative LOAD was executed
797 * before actual STORE.
802 static void clear_work_data(struct work_struct *work)
804 smp_wmb(); /* see set_work_pool_and_clear_pending() */
805 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
808 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
810 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
813 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
815 unsigned long data = atomic_long_read(&work->data);
817 if (data & WORK_STRUCT_PWQ)
818 return work_struct_pwq(data);
824 * get_work_pool - return the worker_pool a given work was associated with
825 * @work: the work item of interest
827 * Pools are created and destroyed under wq_pool_mutex, and allows read
828 * access under RCU read lock. As such, this function should be
829 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
831 * All fields of the returned pool are accessible as long as the above
832 * mentioned locking is in effect. If the returned pool needs to be used
833 * beyond the critical section, the caller is responsible for ensuring the
834 * returned pool is and stays online.
836 * Return: The worker_pool @work was last associated with. %NULL if none.
838 static struct worker_pool *get_work_pool(struct work_struct *work)
840 unsigned long data = atomic_long_read(&work->data);
843 assert_rcu_or_pool_mutex();
845 if (data & WORK_STRUCT_PWQ)
846 return work_struct_pwq(data)->pool;
848 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
849 if (pool_id == WORK_OFFQ_POOL_NONE)
852 return idr_find(&worker_pool_idr, pool_id);
856 * get_work_pool_id - return the worker pool ID a given work is associated with
857 * @work: the work item of interest
859 * Return: The worker_pool ID @work was last associated with.
860 * %WORK_OFFQ_POOL_NONE if none.
862 static int get_work_pool_id(struct work_struct *work)
864 unsigned long data = atomic_long_read(&work->data);
866 if (data & WORK_STRUCT_PWQ)
867 return work_struct_pwq(data)->pool->id;
869 return data >> WORK_OFFQ_POOL_SHIFT;
872 static void mark_work_canceling(struct work_struct *work)
874 unsigned long pool_id = get_work_pool_id(work);
876 pool_id <<= WORK_OFFQ_POOL_SHIFT;
877 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
880 static bool work_is_canceling(struct work_struct *work)
882 unsigned long data = atomic_long_read(&work->data);
884 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
888 * Policy functions. These define the policies on how the global worker
889 * pools are managed. Unless noted otherwise, these functions assume that
890 * they're being called with pool->lock held.
894 * Need to wake up a worker? Called from anything but currently
897 * Note that, because unbound workers never contribute to nr_running, this
898 * function will always return %true for unbound pools as long as the
899 * worklist isn't empty.
901 static bool need_more_worker(struct worker_pool *pool)
903 return !list_empty(&pool->worklist) && !pool->nr_running;
906 /* Can I start working? Called from busy but !running workers. */
907 static bool may_start_working(struct worker_pool *pool)
909 return pool->nr_idle;
912 /* Do I need to keep working? Called from currently running workers. */
913 static bool keep_working(struct worker_pool *pool)
915 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
918 /* Do we need a new worker? Called from manager. */
919 static bool need_to_create_worker(struct worker_pool *pool)
921 return need_more_worker(pool) && !may_start_working(pool);
924 /* Do we have too many workers and should some go away? */
925 static bool too_many_workers(struct worker_pool *pool)
927 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
928 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
929 int nr_busy = pool->nr_workers - nr_idle;
931 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
935 * worker_set_flags - set worker flags and adjust nr_running accordingly
937 * @flags: flags to set
939 * Set @flags in @worker->flags and adjust nr_running accordingly.
941 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
943 struct worker_pool *pool = worker->pool;
945 lockdep_assert_held(&pool->lock);
947 /* If transitioning into NOT_RUNNING, adjust nr_running. */
948 if ((flags & WORKER_NOT_RUNNING) &&
949 !(worker->flags & WORKER_NOT_RUNNING)) {
953 worker->flags |= flags;
957 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
959 * @flags: flags to clear
961 * Clear @flags in @worker->flags and adjust nr_running accordingly.
963 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
965 struct worker_pool *pool = worker->pool;
966 unsigned int oflags = worker->flags;
968 lockdep_assert_held(&pool->lock);
970 worker->flags &= ~flags;
973 * If transitioning out of NOT_RUNNING, increment nr_running. Note
974 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
975 * of multiple flags, not a single flag.
977 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
978 if (!(worker->flags & WORKER_NOT_RUNNING))
982 /* Return the first idle worker. Called with pool->lock held. */
983 static struct worker *first_idle_worker(struct worker_pool *pool)
985 if (unlikely(list_empty(&pool->idle_list)))
988 return list_first_entry(&pool->idle_list, struct worker, entry);
992 * worker_enter_idle - enter idle state
993 * @worker: worker which is entering idle state
995 * @worker is entering idle state. Update stats and idle timer if
999 * raw_spin_lock_irq(pool->lock).
1001 static void worker_enter_idle(struct worker *worker)
1003 struct worker_pool *pool = worker->pool;
1005 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1006 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1007 (worker->hentry.next || worker->hentry.pprev)))
1010 /* can't use worker_set_flags(), also called from create_worker() */
1011 worker->flags |= WORKER_IDLE;
1013 worker->last_active = jiffies;
1015 /* idle_list is LIFO */
1016 list_add(&worker->entry, &pool->idle_list);
1018 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1019 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1021 /* Sanity check nr_running. */
1022 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1026 * worker_leave_idle - leave idle state
1027 * @worker: worker which is leaving idle state
1029 * @worker is leaving idle state. Update stats.
1032 * raw_spin_lock_irq(pool->lock).
1034 static void worker_leave_idle(struct worker *worker)
1036 struct worker_pool *pool = worker->pool;
1038 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1040 worker_clr_flags(worker, WORKER_IDLE);
1042 list_del_init(&worker->entry);
1046 * find_worker_executing_work - find worker which is executing a work
1047 * @pool: pool of interest
1048 * @work: work to find worker for
1050 * Find a worker which is executing @work on @pool by searching
1051 * @pool->busy_hash which is keyed by the address of @work. For a worker
1052 * to match, its current execution should match the address of @work and
1053 * its work function. This is to avoid unwanted dependency between
1054 * unrelated work executions through a work item being recycled while still
1057 * This is a bit tricky. A work item may be freed once its execution
1058 * starts and nothing prevents the freed area from being recycled for
1059 * another work item. If the same work item address ends up being reused
1060 * before the original execution finishes, workqueue will identify the
1061 * recycled work item as currently executing and make it wait until the
1062 * current execution finishes, introducing an unwanted dependency.
1064 * This function checks the work item address and work function to avoid
1065 * false positives. Note that this isn't complete as one may construct a
1066 * work function which can introduce dependency onto itself through a
1067 * recycled work item. Well, if somebody wants to shoot oneself in the
1068 * foot that badly, there's only so much we can do, and if such deadlock
1069 * actually occurs, it should be easy to locate the culprit work function.
1072 * raw_spin_lock_irq(pool->lock).
1075 * Pointer to worker which is executing @work if found, %NULL
1078 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1079 struct work_struct *work)
1081 struct worker *worker;
1083 hash_for_each_possible(pool->busy_hash, worker, hentry,
1084 (unsigned long)work)
1085 if (worker->current_work == work &&
1086 worker->current_func == work->func)
1093 * move_linked_works - move linked works to a list
1094 * @work: start of series of works to be scheduled
1095 * @head: target list to append @work to
1096 * @nextp: out parameter for nested worklist walking
1098 * Schedule linked works starting from @work to @head. Work series to be
1099 * scheduled starts at @work and includes any consecutive work with
1100 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1104 * raw_spin_lock_irq(pool->lock).
1106 static void move_linked_works(struct work_struct *work, struct list_head *head,
1107 struct work_struct **nextp)
1109 struct work_struct *n;
1112 * Linked worklist will always end before the end of the list,
1113 * use NULL for list head.
1115 list_for_each_entry_safe_from(work, n, NULL, entry) {
1116 list_move_tail(&work->entry, head);
1117 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1122 * If we're already inside safe list traversal and have moved
1123 * multiple works to the scheduled queue, the next position
1124 * needs to be updated.
1131 * assign_work - assign a work item and its linked work items to a worker
1132 * @work: work to assign
1133 * @worker: worker to assign to
1134 * @nextp: out parameter for nested worklist walking
1136 * Assign @work and its linked work items to @worker. If @work is already being
1137 * executed by another worker in the same pool, it'll be punted there.
1139 * If @nextp is not NULL, it's updated to point to the next work of the last
1140 * scheduled work. This allows assign_work() to be nested inside
1141 * list_for_each_entry_safe().
1143 * Returns %true if @work was successfully assigned to @worker. %false if @work
1144 * was punted to another worker already executing it.
1146 static bool assign_work(struct work_struct *work, struct worker *worker,
1147 struct work_struct **nextp)
1149 struct worker_pool *pool = worker->pool;
1150 struct worker *collision;
1152 lockdep_assert_held(&pool->lock);
1155 * A single work shouldn't be executed concurrently by multiple workers.
1156 * __queue_work() ensures that @work doesn't jump to a different pool
1157 * while still running in the previous pool. Here, we should ensure that
1158 * @work is not executed concurrently by multiple workers from the same
1159 * pool. Check whether anyone is already processing the work. If so,
1160 * defer the work to the currently executing one.
1162 collision = find_worker_executing_work(pool, work);
1163 if (unlikely(collision)) {
1164 move_linked_works(work, &collision->scheduled, nextp);
1168 move_linked_works(work, &worker->scheduled, nextp);
1173 * kick_pool - wake up an idle worker if necessary
1174 * @pool: pool to kick
1176 * @pool may have pending work items. Wake up worker if necessary. Returns
1177 * whether a worker was woken up.
1179 static bool kick_pool(struct worker_pool *pool)
1181 struct worker *worker = first_idle_worker(pool);
1182 struct task_struct *p;
1184 lockdep_assert_held(&pool->lock);
1186 if (!need_more_worker(pool) || !worker)
1193 * Idle @worker is about to execute @work and waking up provides an
1194 * opportunity to migrate @worker at a lower cost by setting the task's
1195 * wake_cpu field. Let's see if we want to move @worker to improve
1196 * execution locality.
1198 * We're waking the worker that went idle the latest and there's some
1199 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1200 * so, setting the wake_cpu won't do anything. As this is a best-effort
1201 * optimization and the race window is narrow, let's leave as-is for
1202 * now. If this becomes pronounced, we can skip over workers which are
1203 * still on cpu when picking an idle worker.
1205 * If @pool has non-strict affinity, @worker might have ended up outside
1206 * its affinity scope. Repatriate.
1208 if (!pool->attrs->affn_strict &&
1209 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1210 struct work_struct *work = list_first_entry(&pool->worklist,
1211 struct work_struct, entry);
1212 p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1213 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1220 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1223 * Concurrency-managed per-cpu work items that hog CPU for longer than
1224 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1225 * which prevents them from stalling other concurrency-managed work items. If a
1226 * work function keeps triggering this mechanism, it's likely that the work item
1227 * should be using an unbound workqueue instead.
1229 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1230 * and report them so that they can be examined and converted to use unbound
1231 * workqueues as appropriate. To avoid flooding the console, each violating work
1232 * function is tracked and reported with exponential backoff.
1234 #define WCI_MAX_ENTS 128
1239 struct hlist_node hash_node;
1242 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1243 static int wci_nr_ents;
1244 static DEFINE_RAW_SPINLOCK(wci_lock);
1245 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1247 static struct wci_ent *wci_find_ent(work_func_t func)
1249 struct wci_ent *ent;
1251 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1252 (unsigned long)func) {
1253 if (ent->func == func)
1259 static void wq_cpu_intensive_report(work_func_t func)
1261 struct wci_ent *ent;
1264 ent = wci_find_ent(func);
1269 * Start reporting from the fourth time and back off
1272 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1273 if (cnt >= 4 && is_power_of_2(cnt))
1274 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1275 ent->func, wq_cpu_intensive_thresh_us,
1276 atomic64_read(&ent->cnt));
1281 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1282 * is exhausted, something went really wrong and we probably made enough
1285 if (wci_nr_ents >= WCI_MAX_ENTS)
1288 raw_spin_lock(&wci_lock);
1290 if (wci_nr_ents >= WCI_MAX_ENTS) {
1291 raw_spin_unlock(&wci_lock);
1295 if (wci_find_ent(func)) {
1296 raw_spin_unlock(&wci_lock);
1300 ent = &wci_ents[wci_nr_ents++];
1302 atomic64_set(&ent->cnt, 1);
1303 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1305 raw_spin_unlock(&wci_lock);
1308 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1309 static void wq_cpu_intensive_report(work_func_t func) {}
1310 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1313 * wq_worker_running - a worker is running again
1314 * @task: task waking up
1316 * This function is called when a worker returns from schedule()
1318 void wq_worker_running(struct task_struct *task)
1320 struct worker *worker = kthread_data(task);
1322 if (!READ_ONCE(worker->sleeping))
1326 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1327 * and the nr_running increment below, we may ruin the nr_running reset
1328 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1329 * pool. Protect against such race.
1332 if (!(worker->flags & WORKER_NOT_RUNNING))
1333 worker->pool->nr_running++;
1337 * CPU intensive auto-detection cares about how long a work item hogged
1338 * CPU without sleeping. Reset the starting timestamp on wakeup.
1340 worker->current_at = worker->task->se.sum_exec_runtime;
1342 WRITE_ONCE(worker->sleeping, 0);
1346 * wq_worker_sleeping - a worker is going to sleep
1347 * @task: task going to sleep
1349 * This function is called from schedule() when a busy worker is
1352 void wq_worker_sleeping(struct task_struct *task)
1354 struct worker *worker = kthread_data(task);
1355 struct worker_pool *pool;
1358 * Rescuers, which may not have all the fields set up like normal
1359 * workers, also reach here, let's not access anything before
1360 * checking NOT_RUNNING.
1362 if (worker->flags & WORKER_NOT_RUNNING)
1365 pool = worker->pool;
1367 /* Return if preempted before wq_worker_running() was reached */
1368 if (READ_ONCE(worker->sleeping))
1371 WRITE_ONCE(worker->sleeping, 1);
1372 raw_spin_lock_irq(&pool->lock);
1375 * Recheck in case unbind_workers() preempted us. We don't
1376 * want to decrement nr_running after the worker is unbound
1377 * and nr_running has been reset.
1379 if (worker->flags & WORKER_NOT_RUNNING) {
1380 raw_spin_unlock_irq(&pool->lock);
1385 if (kick_pool(pool))
1386 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1388 raw_spin_unlock_irq(&pool->lock);
1392 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1393 * @task: task currently running
1395 * Called from scheduler_tick(). We're in the IRQ context and the current
1396 * worker's fields which follow the 'K' locking rule can be accessed safely.
1398 void wq_worker_tick(struct task_struct *task)
1400 struct worker *worker = kthread_data(task);
1401 struct pool_workqueue *pwq = worker->current_pwq;
1402 struct worker_pool *pool = worker->pool;
1407 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1409 if (!wq_cpu_intensive_thresh_us)
1413 * If the current worker is concurrency managed and hogged the CPU for
1414 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1415 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1417 * Set @worker->sleeping means that @worker is in the process of
1418 * switching out voluntarily and won't be contributing to
1419 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1420 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1421 * double decrements. The task is releasing the CPU anyway. Let's skip.
1422 * We probably want to make this prettier in the future.
1424 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1425 worker->task->se.sum_exec_runtime - worker->current_at <
1426 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1429 raw_spin_lock(&pool->lock);
1431 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1432 wq_cpu_intensive_report(worker->current_func);
1433 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1435 if (kick_pool(pool))
1436 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1438 raw_spin_unlock(&pool->lock);
1442 * wq_worker_last_func - retrieve worker's last work function
1443 * @task: Task to retrieve last work function of.
1445 * Determine the last function a worker executed. This is called from
1446 * the scheduler to get a worker's last known identity.
1449 * raw_spin_lock_irq(rq->lock)
1451 * This function is called during schedule() when a kworker is going
1452 * to sleep. It's used by psi to identify aggregation workers during
1453 * dequeuing, to allow periodic aggregation to shut-off when that
1454 * worker is the last task in the system or cgroup to go to sleep.
1456 * As this function doesn't involve any workqueue-related locking, it
1457 * only returns stable values when called from inside the scheduler's
1458 * queuing and dequeuing paths, when @task, which must be a kworker,
1459 * is guaranteed to not be processing any works.
1462 * The last work function %current executed as a worker, NULL if it
1463 * hasn't executed any work yet.
1465 work_func_t wq_worker_last_func(struct task_struct *task)
1467 struct worker *worker = kthread_data(task);
1469 return worker->last_func;
1473 * wq_node_nr_active - Determine wq_node_nr_active to use
1474 * @wq: workqueue of interest
1475 * @node: NUMA node, can be %NUMA_NO_NODE
1477 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1479 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1481 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1483 * - Otherwise, node_nr_active[@node].
1485 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1488 if (!(wq->flags & WQ_UNBOUND))
1491 if (node == NUMA_NO_NODE)
1494 return wq->node_nr_active[node];
1498 * wq_update_node_max_active - Update per-node max_actives to use
1499 * @wq: workqueue to update
1500 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1502 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1503 * distributed among nodes according to the proportions of numbers of online
1504 * cpus. The result is always between @wq->min_active and max_active.
1506 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1508 struct cpumask *effective = unbound_effective_cpumask(wq);
1509 int min_active = READ_ONCE(wq->min_active);
1510 int max_active = READ_ONCE(wq->max_active);
1511 int total_cpus, node;
1513 lockdep_assert_held(&wq->mutex);
1515 if (!wq_topo_initialized)
1518 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1521 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1525 for_each_node(node) {
1528 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1529 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1532 wq_node_nr_active(wq, node)->max =
1533 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1534 min_active, max_active);
1537 wq_node_nr_active(wq, NUMA_NO_NODE)->max = min_active;
1541 * get_pwq - get an extra reference on the specified pool_workqueue
1542 * @pwq: pool_workqueue to get
1544 * Obtain an extra reference on @pwq. The caller should guarantee that
1545 * @pwq has positive refcnt and be holding the matching pool->lock.
1547 static void get_pwq(struct pool_workqueue *pwq)
1549 lockdep_assert_held(&pwq->pool->lock);
1550 WARN_ON_ONCE(pwq->refcnt <= 0);
1555 * put_pwq - put a pool_workqueue reference
1556 * @pwq: pool_workqueue to put
1558 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1559 * destruction. The caller should be holding the matching pool->lock.
1561 static void put_pwq(struct pool_workqueue *pwq)
1563 lockdep_assert_held(&pwq->pool->lock);
1564 if (likely(--pwq->refcnt))
1567 * @pwq can't be released under pool->lock, bounce to a dedicated
1568 * kthread_worker to avoid A-A deadlocks.
1570 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1574 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1575 * @pwq: pool_workqueue to put (can be %NULL)
1577 * put_pwq() with locking. This function also allows %NULL @pwq.
1579 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1583 * As both pwqs and pools are RCU protected, the
1584 * following lock operations are safe.
1586 raw_spin_lock_irq(&pwq->pool->lock);
1588 raw_spin_unlock_irq(&pwq->pool->lock);
1592 static bool pwq_is_empty(struct pool_workqueue *pwq)
1594 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1597 static void __pwq_activate_work(struct pool_workqueue *pwq,
1598 struct work_struct *work)
1600 unsigned long *wdb = work_data_bits(work);
1602 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1603 trace_workqueue_activate_work(work);
1604 if (list_empty(&pwq->pool->worklist))
1605 pwq->pool->watchdog_ts = jiffies;
1606 move_linked_works(work, &pwq->pool->worklist, NULL);
1607 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1611 * pwq_activate_work - Activate a work item if inactive
1612 * @pwq: pool_workqueue @work belongs to
1613 * @work: work item to activate
1615 * Returns %true if activated. %false if already active.
1617 static bool pwq_activate_work(struct pool_workqueue *pwq,
1618 struct work_struct *work)
1620 struct worker_pool *pool = pwq->pool;
1621 struct wq_node_nr_active *nna;
1623 lockdep_assert_held(&pool->lock);
1625 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE))
1628 nna = wq_node_nr_active(pwq->wq, pool->node);
1630 atomic_inc(&nna->nr);
1633 __pwq_activate_work(pwq, work);
1637 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1639 int max = READ_ONCE(nna->max);
1644 old = atomic_read(&nna->nr);
1647 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1654 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1655 * @pwq: pool_workqueue of interest
1656 * @fill: max_active may have increased, try to increase concurrency level
1658 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1659 * successfully obtained. %false otherwise.
1661 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1663 struct workqueue_struct *wq = pwq->wq;
1664 struct worker_pool *pool = pwq->pool;
1665 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1666 bool obtained = false;
1668 lockdep_assert_held(&pool->lock);
1671 /* per-cpu workqueue, pwq->nr_active is sufficient */
1672 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1677 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1678 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1679 * concurrency level. Don't jump the line.
1681 * We need to ignore the pending test after max_active has increased as
1682 * pwq_dec_nr_active() can only maintain the concurrency level but not
1683 * increase it. This is indicated by @fill.
1685 if (!list_empty(&pwq->pending_node) && likely(!fill))
1688 obtained = tryinc_node_nr_active(nna);
1693 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1694 * and try again. The smp_mb() is paired with the implied memory barrier
1695 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1696 * we see the decremented $nna->nr or they see non-empty
1697 * $nna->pending_pwqs.
1699 raw_spin_lock(&nna->lock);
1701 if (list_empty(&pwq->pending_node))
1702 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1703 else if (likely(!fill))
1708 obtained = tryinc_node_nr_active(nna);
1711 * If @fill, @pwq might have already been pending. Being spuriously
1712 * pending in cold paths doesn't affect anything. Let's leave it be.
1714 if (obtained && likely(!fill))
1715 list_del_init(&pwq->pending_node);
1718 raw_spin_unlock(&nna->lock);
1726 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1727 * @pwq: pool_workqueue of interest
1728 * @fill: max_active may have increased, try to increase concurrency level
1730 * Activate the first inactive work item of @pwq if available and allowed by
1733 * Returns %true if an inactive work item has been activated. %false if no
1734 * inactive work item is found or max_active limit is reached.
1736 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1738 struct work_struct *work =
1739 list_first_entry_or_null(&pwq->inactive_works,
1740 struct work_struct, entry);
1742 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1743 __pwq_activate_work(pwq, work);
1751 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1752 * @nna: wq_node_nr_active to activate a pending pwq for
1753 * @caller_pool: worker_pool the caller is locking
1755 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1756 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1758 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1759 struct worker_pool *caller_pool)
1761 struct worker_pool *locked_pool = caller_pool;
1762 struct pool_workqueue *pwq;
1763 struct work_struct *work;
1765 lockdep_assert_held(&caller_pool->lock);
1767 raw_spin_lock(&nna->lock);
1769 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1770 struct pool_workqueue, pending_node);
1775 * If @pwq is for a different pool than @locked_pool, we need to lock
1776 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1777 * / lock dance. For that, we also need to release @nna->lock as it's
1778 * nested inside pool locks.
1780 if (pwq->pool != locked_pool) {
1781 raw_spin_unlock(&locked_pool->lock);
1782 locked_pool = pwq->pool;
1783 if (!raw_spin_trylock(&locked_pool->lock)) {
1784 raw_spin_unlock(&nna->lock);
1785 raw_spin_lock(&locked_pool->lock);
1786 raw_spin_lock(&nna->lock);
1792 * $pwq may not have any inactive work items due to e.g. cancellations.
1793 * Drop it from pending_pwqs and see if there's another one.
1795 work = list_first_entry_or_null(&pwq->inactive_works,
1796 struct work_struct, entry);
1798 list_del_init(&pwq->pending_node);
1803 * Acquire an nr_active count and activate the inactive work item. If
1804 * $pwq still has inactive work items, rotate it to the end of the
1805 * pending_pwqs so that we round-robin through them. This means that
1806 * inactive work items are not activated in queueing order which is fine
1807 * given that there has never been any ordering across different pwqs.
1809 if (likely(tryinc_node_nr_active(nna))) {
1811 __pwq_activate_work(pwq, work);
1813 if (list_empty(&pwq->inactive_works))
1814 list_del_init(&pwq->pending_node);
1816 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1818 /* if activating a foreign pool, make sure it's running */
1819 if (pwq->pool != caller_pool)
1820 kick_pool(pwq->pool);
1824 raw_spin_unlock(&nna->lock);
1825 if (locked_pool != caller_pool) {
1826 raw_spin_unlock(&locked_pool->lock);
1827 raw_spin_lock(&caller_pool->lock);
1832 * pwq_dec_nr_active - Retire an active count
1833 * @pwq: pool_workqueue of interest
1835 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1836 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1838 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1840 struct worker_pool *pool = pwq->pool;
1841 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1843 lockdep_assert_held(&pool->lock);
1846 * @pwq->nr_active should be decremented for both percpu and unbound
1852 * For a percpu workqueue, it's simple. Just need to kick the first
1853 * inactive work item on @pwq itself.
1856 pwq_activate_first_inactive(pwq, false);
1861 * If @pwq is for an unbound workqueue, it's more complicated because
1862 * multiple pwqs and pools may be sharing the nr_active count. When a
1863 * pwq needs to wait for an nr_active count, it puts itself on
1864 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1865 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1866 * guarantee that either we see non-empty pending_pwqs or they see
1867 * decremented $nna->nr.
1869 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1870 * max_active gets updated. However, it is guaranteed to be equal to or
1871 * larger than @pwq->wq->min_active which is above zero unless freezing.
1872 * This maintains the forward progress guarantee.
1874 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1877 if (!list_empty(&nna->pending_pwqs))
1878 node_activate_pending_pwq(nna, pool);
1882 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1883 * @pwq: pwq of interest
1884 * @work_data: work_data of work which left the queue
1886 * A work either has completed or is removed from pending queue,
1887 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1890 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1891 * and thus should be called after all other state updates for the in-flight
1892 * work item is complete.
1895 * raw_spin_lock_irq(pool->lock).
1897 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1899 int color = get_work_color(work_data);
1901 if (!(work_data & WORK_STRUCT_INACTIVE))
1902 pwq_dec_nr_active(pwq);
1904 pwq->nr_in_flight[color]--;
1906 /* is flush in progress and are we at the flushing tip? */
1907 if (likely(pwq->flush_color != color))
1910 /* are there still in-flight works? */
1911 if (pwq->nr_in_flight[color])
1914 /* this pwq is done, clear flush_color */
1915 pwq->flush_color = -1;
1918 * If this was the last pwq, wake up the first flusher. It
1919 * will handle the rest.
1921 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1922 complete(&pwq->wq->first_flusher->done);
1928 * try_to_grab_pending - steal work item from worklist and disable irq
1929 * @work: work item to steal
1930 * @is_dwork: @work is a delayed_work
1931 * @flags: place to store irq state
1933 * Try to grab PENDING bit of @work. This function can handle @work in any
1934 * stable state - idle, on timer or on worklist.
1938 * ======== ================================================================
1939 * 1 if @work was pending and we successfully stole PENDING
1940 * 0 if @work was idle and we claimed PENDING
1941 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1942 * -ENOENT if someone else is canceling @work, this state may persist
1943 * for arbitrarily long
1944 * ======== ================================================================
1947 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1948 * interrupted while holding PENDING and @work off queue, irq must be
1949 * disabled on entry. This, combined with delayed_work->timer being
1950 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1952 * On successful return, >= 0, irq is disabled and the caller is
1953 * responsible for releasing it using local_irq_restore(*@flags).
1955 * This function is safe to call from any context including IRQ handler.
1957 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1958 unsigned long *flags)
1960 struct worker_pool *pool;
1961 struct pool_workqueue *pwq;
1963 local_irq_save(*flags);
1965 /* try to steal the timer if it exists */
1967 struct delayed_work *dwork = to_delayed_work(work);
1970 * dwork->timer is irqsafe. If del_timer() fails, it's
1971 * guaranteed that the timer is not queued anywhere and not
1972 * running on the local CPU.
1974 if (likely(del_timer(&dwork->timer)))
1978 /* try to claim PENDING the normal way */
1979 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1984 * The queueing is in progress, or it is already queued. Try to
1985 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1987 pool = get_work_pool(work);
1991 raw_spin_lock(&pool->lock);
1993 * work->data is guaranteed to point to pwq only while the work
1994 * item is queued on pwq->wq, and both updating work->data to point
1995 * to pwq on queueing and to pool on dequeueing are done under
1996 * pwq->pool->lock. This in turn guarantees that, if work->data
1997 * points to pwq which is associated with a locked pool, the work
1998 * item is currently queued on that pool.
2000 pwq = get_work_pwq(work);
2001 if (pwq && pwq->pool == pool) {
2002 unsigned long work_data;
2004 debug_work_deactivate(work);
2007 * A cancelable inactive work item must be in the
2008 * pwq->inactive_works since a queued barrier can't be
2009 * canceled (see the comments in insert_wq_barrier()).
2011 * An inactive work item cannot be grabbed directly because
2012 * it might have linked barrier work items which, if left
2013 * on the inactive_works list, will confuse pwq->nr_active
2014 * management later on and cause stall. Make sure the work
2015 * item is activated before grabbing.
2017 pwq_activate_work(pwq, work);
2019 list_del_init(&work->entry);
2022 * work->data points to pwq iff queued. Let's point to pool. As
2023 * this destroys work->data needed by the next step, stash it.
2025 work_data = *work_data_bits(work);
2026 set_work_pool_and_keep_pending(work, pool->id);
2028 /* must be the last step, see the function comment */
2029 pwq_dec_nr_in_flight(pwq, work_data);
2031 raw_spin_unlock(&pool->lock);
2035 raw_spin_unlock(&pool->lock);
2038 local_irq_restore(*flags);
2039 if (work_is_canceling(work))
2046 * insert_work - insert a work into a pool
2047 * @pwq: pwq @work belongs to
2048 * @work: work to insert
2049 * @head: insertion point
2050 * @extra_flags: extra WORK_STRUCT_* flags to set
2052 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2053 * work_struct flags.
2056 * raw_spin_lock_irq(pool->lock).
2058 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2059 struct list_head *head, unsigned int extra_flags)
2061 debug_work_activate(work);
2063 /* record the work call stack in order to print it in KASAN reports */
2064 kasan_record_aux_stack_noalloc(work);
2066 /* we own @work, set data and link */
2067 set_work_pwq(work, pwq, extra_flags);
2068 list_add_tail(&work->entry, head);
2073 * Test whether @work is being queued from another work executing on the
2076 static bool is_chained_work(struct workqueue_struct *wq)
2078 struct worker *worker;
2080 worker = current_wq_worker();
2082 * Return %true iff I'm a worker executing a work item on @wq. If
2083 * I'm @worker, it's safe to dereference it without locking.
2085 return worker && worker->current_pwq->wq == wq;
2089 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2090 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2091 * avoid perturbing sensitive tasks.
2093 static int wq_select_unbound_cpu(int cpu)
2097 if (likely(!wq_debug_force_rr_cpu)) {
2098 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2101 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2104 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2105 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2106 if (unlikely(new_cpu >= nr_cpu_ids)) {
2107 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2108 if (unlikely(new_cpu >= nr_cpu_ids))
2111 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2116 static void __queue_work(int cpu, struct workqueue_struct *wq,
2117 struct work_struct *work)
2119 struct pool_workqueue *pwq;
2120 struct worker_pool *last_pool, *pool;
2121 unsigned int work_flags;
2122 unsigned int req_cpu = cpu;
2125 * While a work item is PENDING && off queue, a task trying to
2126 * steal the PENDING will busy-loop waiting for it to either get
2127 * queued or lose PENDING. Grabbing PENDING and queueing should
2128 * happen with IRQ disabled.
2130 lockdep_assert_irqs_disabled();
2134 * For a draining wq, only works from the same workqueue are
2135 * allowed. The __WQ_DESTROYING helps to spot the issue that
2136 * queues a new work item to a wq after destroy_workqueue(wq).
2138 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2139 WARN_ON_ONCE(!is_chained_work(wq))))
2143 /* pwq which will be used unless @work is executing elsewhere */
2144 if (req_cpu == WORK_CPU_UNBOUND) {
2145 if (wq->flags & WQ_UNBOUND)
2146 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2148 cpu = raw_smp_processor_id();
2151 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2155 * If @work was previously on a different pool, it might still be
2156 * running there, in which case the work needs to be queued on that
2157 * pool to guarantee non-reentrancy.
2159 last_pool = get_work_pool(work);
2160 if (last_pool && last_pool != pool) {
2161 struct worker *worker;
2163 raw_spin_lock(&last_pool->lock);
2165 worker = find_worker_executing_work(last_pool, work);
2167 if (worker && worker->current_pwq->wq == wq) {
2168 pwq = worker->current_pwq;
2170 WARN_ON_ONCE(pool != last_pool);
2172 /* meh... not running there, queue here */
2173 raw_spin_unlock(&last_pool->lock);
2174 raw_spin_lock(&pool->lock);
2177 raw_spin_lock(&pool->lock);
2181 * pwq is determined and locked. For unbound pools, we could have raced
2182 * with pwq release and it could already be dead. If its refcnt is zero,
2183 * repeat pwq selection. Note that unbound pwqs never die without
2184 * another pwq replacing it in cpu_pwq or while work items are executing
2185 * on it, so the retrying is guaranteed to make forward-progress.
2187 if (unlikely(!pwq->refcnt)) {
2188 if (wq->flags & WQ_UNBOUND) {
2189 raw_spin_unlock(&pool->lock);
2194 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2198 /* pwq determined, queue */
2199 trace_workqueue_queue_work(req_cpu, pwq, work);
2201 if (WARN_ON(!list_empty(&work->entry)))
2204 pwq->nr_in_flight[pwq->work_color]++;
2205 work_flags = work_color_to_flags(pwq->work_color);
2208 * Limit the number of concurrently active work items to max_active.
2209 * @work must also queue behind existing inactive work items to maintain
2210 * ordering when max_active changes. See wq_adjust_max_active().
2212 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2213 if (list_empty(&pool->worklist))
2214 pool->watchdog_ts = jiffies;
2216 trace_workqueue_activate_work(work);
2217 insert_work(pwq, work, &pool->worklist, work_flags);
2220 work_flags |= WORK_STRUCT_INACTIVE;
2221 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2225 raw_spin_unlock(&pool->lock);
2230 * queue_work_on - queue work on specific cpu
2231 * @cpu: CPU number to execute work on
2232 * @wq: workqueue to use
2233 * @work: work to queue
2235 * We queue the work to a specific CPU, the caller must ensure it
2236 * can't go away. Callers that fail to ensure that the specified
2237 * CPU cannot go away will execute on a randomly chosen CPU.
2238 * But note well that callers specifying a CPU that never has been
2239 * online will get a splat.
2241 * Return: %false if @work was already on a queue, %true otherwise.
2243 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2244 struct work_struct *work)
2247 unsigned long flags;
2249 local_irq_save(flags);
2251 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2252 __queue_work(cpu, wq, work);
2256 local_irq_restore(flags);
2259 EXPORT_SYMBOL(queue_work_on);
2262 * select_numa_node_cpu - Select a CPU based on NUMA node
2263 * @node: NUMA node ID that we want to select a CPU from
2265 * This function will attempt to find a "random" cpu available on a given
2266 * node. If there are no CPUs available on the given node it will return
2267 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2268 * available CPU if we need to schedule this work.
2270 static int select_numa_node_cpu(int node)
2274 /* Delay binding to CPU if node is not valid or online */
2275 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2276 return WORK_CPU_UNBOUND;
2278 /* Use local node/cpu if we are already there */
2279 cpu = raw_smp_processor_id();
2280 if (node == cpu_to_node(cpu))
2283 /* Use "random" otherwise know as "first" online CPU of node */
2284 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2286 /* If CPU is valid return that, otherwise just defer */
2287 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2291 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2292 * @node: NUMA node that we are targeting the work for
2293 * @wq: workqueue to use
2294 * @work: work to queue
2296 * We queue the work to a "random" CPU within a given NUMA node. The basic
2297 * idea here is to provide a way to somehow associate work with a given
2300 * This function will only make a best effort attempt at getting this onto
2301 * the right NUMA node. If no node is requested or the requested node is
2302 * offline then we just fall back to standard queue_work behavior.
2304 * Currently the "random" CPU ends up being the first available CPU in the
2305 * intersection of cpu_online_mask and the cpumask of the node, unless we
2306 * are running on the node. In that case we just use the current CPU.
2308 * Return: %false if @work was already on a queue, %true otherwise.
2310 bool queue_work_node(int node, struct workqueue_struct *wq,
2311 struct work_struct *work)
2313 unsigned long flags;
2317 * This current implementation is specific to unbound workqueues.
2318 * Specifically we only return the first available CPU for a given
2319 * node instead of cycling through individual CPUs within the node.
2321 * If this is used with a per-cpu workqueue then the logic in
2322 * workqueue_select_cpu_near would need to be updated to allow for
2323 * some round robin type logic.
2325 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2327 local_irq_save(flags);
2329 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2330 int cpu = select_numa_node_cpu(node);
2332 __queue_work(cpu, wq, work);
2336 local_irq_restore(flags);
2339 EXPORT_SYMBOL_GPL(queue_work_node);
2341 void delayed_work_timer_fn(struct timer_list *t)
2343 struct delayed_work *dwork = from_timer(dwork, t, timer);
2345 /* should have been called from irqsafe timer with irq already off */
2346 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2348 EXPORT_SYMBOL(delayed_work_timer_fn);
2350 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2351 struct delayed_work *dwork, unsigned long delay)
2353 struct timer_list *timer = &dwork->timer;
2354 struct work_struct *work = &dwork->work;
2357 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2358 WARN_ON_ONCE(timer_pending(timer));
2359 WARN_ON_ONCE(!list_empty(&work->entry));
2362 * If @delay is 0, queue @dwork->work immediately. This is for
2363 * both optimization and correctness. The earliest @timer can
2364 * expire is on the closest next tick and delayed_work users depend
2365 * on that there's no such delay when @delay is 0.
2368 __queue_work(cpu, wq, &dwork->work);
2374 timer->expires = jiffies + delay;
2376 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2377 /* If the current cpu is a housekeeping cpu, use it. */
2378 cpu = smp_processor_id();
2379 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2380 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2381 add_timer_on(timer, cpu);
2383 if (likely(cpu == WORK_CPU_UNBOUND))
2386 add_timer_on(timer, cpu);
2391 * queue_delayed_work_on - queue work on specific CPU after delay
2392 * @cpu: CPU number to execute work on
2393 * @wq: workqueue to use
2394 * @dwork: work to queue
2395 * @delay: number of jiffies to wait before queueing
2397 * Return: %false if @work was already on a queue, %true otherwise. If
2398 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2401 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2402 struct delayed_work *dwork, unsigned long delay)
2404 struct work_struct *work = &dwork->work;
2406 unsigned long flags;
2408 /* read the comment in __queue_work() */
2409 local_irq_save(flags);
2411 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2412 __queue_delayed_work(cpu, wq, dwork, delay);
2416 local_irq_restore(flags);
2419 EXPORT_SYMBOL(queue_delayed_work_on);
2422 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2423 * @cpu: CPU number to execute work on
2424 * @wq: workqueue to use
2425 * @dwork: work to queue
2426 * @delay: number of jiffies to wait before queueing
2428 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2429 * modify @dwork's timer so that it expires after @delay. If @delay is
2430 * zero, @work is guaranteed to be scheduled immediately regardless of its
2433 * Return: %false if @dwork was idle and queued, %true if @dwork was
2434 * pending and its timer was modified.
2436 * This function is safe to call from any context including IRQ handler.
2437 * See try_to_grab_pending() for details.
2439 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2440 struct delayed_work *dwork, unsigned long delay)
2442 unsigned long flags;
2446 ret = try_to_grab_pending(&dwork->work, true, &flags);
2447 } while (unlikely(ret == -EAGAIN));
2449 if (likely(ret >= 0)) {
2450 __queue_delayed_work(cpu, wq, dwork, delay);
2451 local_irq_restore(flags);
2454 /* -ENOENT from try_to_grab_pending() becomes %true */
2457 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2459 static void rcu_work_rcufn(struct rcu_head *rcu)
2461 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2463 /* read the comment in __queue_work() */
2464 local_irq_disable();
2465 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2470 * queue_rcu_work - queue work after a RCU grace period
2471 * @wq: workqueue to use
2472 * @rwork: work to queue
2474 * Return: %false if @rwork was already pending, %true otherwise. Note
2475 * that a full RCU grace period is guaranteed only after a %true return.
2476 * While @rwork is guaranteed to be executed after a %false return, the
2477 * execution may happen before a full RCU grace period has passed.
2479 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2481 struct work_struct *work = &rwork->work;
2483 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2485 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2491 EXPORT_SYMBOL(queue_rcu_work);
2493 static struct worker *alloc_worker(int node)
2495 struct worker *worker;
2497 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2499 INIT_LIST_HEAD(&worker->entry);
2500 INIT_LIST_HEAD(&worker->scheduled);
2501 INIT_LIST_HEAD(&worker->node);
2502 /* on creation a worker is in !idle && prep state */
2503 worker->flags = WORKER_PREP;
2508 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2510 if (pool->cpu < 0 && pool->attrs->affn_strict)
2511 return pool->attrs->__pod_cpumask;
2513 return pool->attrs->cpumask;
2517 * worker_attach_to_pool() - attach a worker to a pool
2518 * @worker: worker to be attached
2519 * @pool: the target pool
2521 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2522 * cpu-binding of @worker are kept coordinated with the pool across
2525 static void worker_attach_to_pool(struct worker *worker,
2526 struct worker_pool *pool)
2528 mutex_lock(&wq_pool_attach_mutex);
2531 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2532 * stable across this function. See the comments above the flag
2533 * definition for details.
2535 if (pool->flags & POOL_DISASSOCIATED)
2536 worker->flags |= WORKER_UNBOUND;
2538 kthread_set_per_cpu(worker->task, pool->cpu);
2540 if (worker->rescue_wq)
2541 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2543 list_add_tail(&worker->node, &pool->workers);
2544 worker->pool = pool;
2546 mutex_unlock(&wq_pool_attach_mutex);
2550 * worker_detach_from_pool() - detach a worker from its pool
2551 * @worker: worker which is attached to its pool
2553 * Undo the attaching which had been done in worker_attach_to_pool(). The
2554 * caller worker shouldn't access to the pool after detached except it has
2555 * other reference to the pool.
2557 static void worker_detach_from_pool(struct worker *worker)
2559 struct worker_pool *pool = worker->pool;
2560 struct completion *detach_completion = NULL;
2562 mutex_lock(&wq_pool_attach_mutex);
2564 kthread_set_per_cpu(worker->task, -1);
2565 list_del(&worker->node);
2566 worker->pool = NULL;
2568 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2569 detach_completion = pool->detach_completion;
2570 mutex_unlock(&wq_pool_attach_mutex);
2572 /* clear leftover flags without pool->lock after it is detached */
2573 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2575 if (detach_completion)
2576 complete(detach_completion);
2580 * create_worker - create a new workqueue worker
2581 * @pool: pool the new worker will belong to
2583 * Create and start a new worker which is attached to @pool.
2586 * Might sleep. Does GFP_KERNEL allocations.
2589 * Pointer to the newly created worker.
2591 static struct worker *create_worker(struct worker_pool *pool)
2593 struct worker *worker;
2597 /* ID is needed to determine kthread name */
2598 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2600 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2605 worker = alloc_worker(pool->node);
2607 pr_err_once("workqueue: Failed to allocate a worker\n");
2614 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2615 pool->attrs->nice < 0 ? "H" : "");
2617 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2619 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2620 "kworker/%s", id_buf);
2621 if (IS_ERR(worker->task)) {
2622 if (PTR_ERR(worker->task) == -EINTR) {
2623 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2626 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2632 set_user_nice(worker->task, pool->attrs->nice);
2633 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2635 /* successful, attach the worker to the pool */
2636 worker_attach_to_pool(worker, pool);
2638 /* start the newly created worker */
2639 raw_spin_lock_irq(&pool->lock);
2641 worker->pool->nr_workers++;
2642 worker_enter_idle(worker);
2645 * @worker is waiting on a completion in kthread() and will trigger hung
2646 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2647 * wake it up explicitly.
2649 wake_up_process(worker->task);
2651 raw_spin_unlock_irq(&pool->lock);
2656 ida_free(&pool->worker_ida, id);
2661 static void unbind_worker(struct worker *worker)
2663 lockdep_assert_held(&wq_pool_attach_mutex);
2665 kthread_set_per_cpu(worker->task, -1);
2666 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2667 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2669 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2672 static void wake_dying_workers(struct list_head *cull_list)
2674 struct worker *worker, *tmp;
2676 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2677 list_del_init(&worker->entry);
2678 unbind_worker(worker);
2680 * If the worker was somehow already running, then it had to be
2681 * in pool->idle_list when set_worker_dying() happened or we
2682 * wouldn't have gotten here.
2684 * Thus, the worker must either have observed the WORKER_DIE
2685 * flag, or have set its state to TASK_IDLE. Either way, the
2686 * below will be observed by the worker and is safe to do
2687 * outside of pool->lock.
2689 wake_up_process(worker->task);
2694 * set_worker_dying - Tag a worker for destruction
2695 * @worker: worker to be destroyed
2696 * @list: transfer worker away from its pool->idle_list and into list
2698 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2702 * raw_spin_lock_irq(pool->lock).
2704 static void set_worker_dying(struct worker *worker, struct list_head *list)
2706 struct worker_pool *pool = worker->pool;
2708 lockdep_assert_held(&pool->lock);
2709 lockdep_assert_held(&wq_pool_attach_mutex);
2711 /* sanity check frenzy */
2712 if (WARN_ON(worker->current_work) ||
2713 WARN_ON(!list_empty(&worker->scheduled)) ||
2714 WARN_ON(!(worker->flags & WORKER_IDLE)))
2720 worker->flags |= WORKER_DIE;
2722 list_move(&worker->entry, list);
2723 list_move(&worker->node, &pool->dying_workers);
2727 * idle_worker_timeout - check if some idle workers can now be deleted.
2728 * @t: The pool's idle_timer that just expired
2730 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2731 * worker_leave_idle(), as a worker flicking between idle and active while its
2732 * pool is at the too_many_workers() tipping point would cause too much timer
2733 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2734 * it expire and re-evaluate things from there.
2736 static void idle_worker_timeout(struct timer_list *t)
2738 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2739 bool do_cull = false;
2741 if (work_pending(&pool->idle_cull_work))
2744 raw_spin_lock_irq(&pool->lock);
2746 if (too_many_workers(pool)) {
2747 struct worker *worker;
2748 unsigned long expires;
2750 /* idle_list is kept in LIFO order, check the last one */
2751 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2752 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2753 do_cull = !time_before(jiffies, expires);
2756 mod_timer(&pool->idle_timer, expires);
2758 raw_spin_unlock_irq(&pool->lock);
2761 queue_work(system_unbound_wq, &pool->idle_cull_work);
2765 * idle_cull_fn - cull workers that have been idle for too long.
2766 * @work: the pool's work for handling these idle workers
2768 * This goes through a pool's idle workers and gets rid of those that have been
2769 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2771 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2772 * culled, so this also resets worker affinity. This requires a sleepable
2773 * context, hence the split between timer callback and work item.
2775 static void idle_cull_fn(struct work_struct *work)
2777 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2778 LIST_HEAD(cull_list);
2781 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2782 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2783 * path. This is required as a previously-preempted worker could run after
2784 * set_worker_dying() has happened but before wake_dying_workers() did.
2786 mutex_lock(&wq_pool_attach_mutex);
2787 raw_spin_lock_irq(&pool->lock);
2789 while (too_many_workers(pool)) {
2790 struct worker *worker;
2791 unsigned long expires;
2793 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2794 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2796 if (time_before(jiffies, expires)) {
2797 mod_timer(&pool->idle_timer, expires);
2801 set_worker_dying(worker, &cull_list);
2804 raw_spin_unlock_irq(&pool->lock);
2805 wake_dying_workers(&cull_list);
2806 mutex_unlock(&wq_pool_attach_mutex);
2809 static void send_mayday(struct work_struct *work)
2811 struct pool_workqueue *pwq = get_work_pwq(work);
2812 struct workqueue_struct *wq = pwq->wq;
2814 lockdep_assert_held(&wq_mayday_lock);
2819 /* mayday mayday mayday */
2820 if (list_empty(&pwq->mayday_node)) {
2822 * If @pwq is for an unbound wq, its base ref may be put at
2823 * any time due to an attribute change. Pin @pwq until the
2824 * rescuer is done with it.
2827 list_add_tail(&pwq->mayday_node, &wq->maydays);
2828 wake_up_process(wq->rescuer->task);
2829 pwq->stats[PWQ_STAT_MAYDAY]++;
2833 static void pool_mayday_timeout(struct timer_list *t)
2835 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2836 struct work_struct *work;
2838 raw_spin_lock_irq(&pool->lock);
2839 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2841 if (need_to_create_worker(pool)) {
2843 * We've been trying to create a new worker but
2844 * haven't been successful. We might be hitting an
2845 * allocation deadlock. Send distress signals to
2848 list_for_each_entry(work, &pool->worklist, entry)
2852 raw_spin_unlock(&wq_mayday_lock);
2853 raw_spin_unlock_irq(&pool->lock);
2855 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2859 * maybe_create_worker - create a new worker if necessary
2860 * @pool: pool to create a new worker for
2862 * Create a new worker for @pool if necessary. @pool is guaranteed to
2863 * have at least one idle worker on return from this function. If
2864 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2865 * sent to all rescuers with works scheduled on @pool to resolve
2866 * possible allocation deadlock.
2868 * On return, need_to_create_worker() is guaranteed to be %false and
2869 * may_start_working() %true.
2872 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2873 * multiple times. Does GFP_KERNEL allocations. Called only from
2876 static void maybe_create_worker(struct worker_pool *pool)
2877 __releases(&pool->lock)
2878 __acquires(&pool->lock)
2881 raw_spin_unlock_irq(&pool->lock);
2883 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2884 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2887 if (create_worker(pool) || !need_to_create_worker(pool))
2890 schedule_timeout_interruptible(CREATE_COOLDOWN);
2892 if (!need_to_create_worker(pool))
2896 del_timer_sync(&pool->mayday_timer);
2897 raw_spin_lock_irq(&pool->lock);
2899 * This is necessary even after a new worker was just successfully
2900 * created as @pool->lock was dropped and the new worker might have
2901 * already become busy.
2903 if (need_to_create_worker(pool))
2908 * manage_workers - manage worker pool
2911 * Assume the manager role and manage the worker pool @worker belongs
2912 * to. At any given time, there can be only zero or one manager per
2913 * pool. The exclusion is handled automatically by this function.
2915 * The caller can safely start processing works on false return. On
2916 * true return, it's guaranteed that need_to_create_worker() is false
2917 * and may_start_working() is true.
2920 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2921 * multiple times. Does GFP_KERNEL allocations.
2924 * %false if the pool doesn't need management and the caller can safely
2925 * start processing works, %true if management function was performed and
2926 * the conditions that the caller verified before calling the function may
2927 * no longer be true.
2929 static bool manage_workers(struct worker *worker)
2931 struct worker_pool *pool = worker->pool;
2933 if (pool->flags & POOL_MANAGER_ACTIVE)
2936 pool->flags |= POOL_MANAGER_ACTIVE;
2937 pool->manager = worker;
2939 maybe_create_worker(pool);
2941 pool->manager = NULL;
2942 pool->flags &= ~POOL_MANAGER_ACTIVE;
2943 rcuwait_wake_up(&manager_wait);
2948 * process_one_work - process single work
2950 * @work: work to process
2952 * Process @work. This function contains all the logics necessary to
2953 * process a single work including synchronization against and
2954 * interaction with other workers on the same cpu, queueing and
2955 * flushing. As long as context requirement is met, any worker can
2956 * call this function to process a work.
2959 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2961 static void process_one_work(struct worker *worker, struct work_struct *work)
2962 __releases(&pool->lock)
2963 __acquires(&pool->lock)
2965 struct pool_workqueue *pwq = get_work_pwq(work);
2966 struct worker_pool *pool = worker->pool;
2967 unsigned long work_data;
2968 int lockdep_start_depth, rcu_start_depth;
2969 #ifdef CONFIG_LOCKDEP
2971 * It is permissible to free the struct work_struct from
2972 * inside the function that is called from it, this we need to
2973 * take into account for lockdep too. To avoid bogus "held
2974 * lock freed" warnings as well as problems when looking into
2975 * work->lockdep_map, make a copy and use that here.
2977 struct lockdep_map lockdep_map;
2979 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2981 /* ensure we're on the correct CPU */
2982 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2983 raw_smp_processor_id() != pool->cpu);
2985 /* claim and dequeue */
2986 debug_work_deactivate(work);
2987 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2988 worker->current_work = work;
2989 worker->current_func = work->func;
2990 worker->current_pwq = pwq;
2991 worker->current_at = worker->task->se.sum_exec_runtime;
2992 work_data = *work_data_bits(work);
2993 worker->current_color = get_work_color(work_data);
2996 * Record wq name for cmdline and debug reporting, may get
2997 * overridden through set_worker_desc().
2999 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3001 list_del_init(&work->entry);
3004 * CPU intensive works don't participate in concurrency management.
3005 * They're the scheduler's responsibility. This takes @worker out
3006 * of concurrency management and the next code block will chain
3007 * execution of the pending work items.
3009 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3010 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3013 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3014 * since nr_running would always be >= 1 at this point. This is used to
3015 * chain execution of the pending work items for WORKER_NOT_RUNNING
3016 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3021 * Record the last pool and clear PENDING which should be the last
3022 * update to @work. Also, do this inside @pool->lock so that
3023 * PENDING and queued state changes happen together while IRQ is
3026 set_work_pool_and_clear_pending(work, pool->id);
3028 pwq->stats[PWQ_STAT_STARTED]++;
3029 raw_spin_unlock_irq(&pool->lock);
3031 rcu_start_depth = rcu_preempt_depth();
3032 lockdep_start_depth = lockdep_depth(current);
3033 lock_map_acquire(&pwq->wq->lockdep_map);
3034 lock_map_acquire(&lockdep_map);
3036 * Strictly speaking we should mark the invariant state without holding
3037 * any locks, that is, before these two lock_map_acquire()'s.
3039 * However, that would result in:
3046 * Which would create W1->C->W1 dependencies, even though there is no
3047 * actual deadlock possible. There are two solutions, using a
3048 * read-recursive acquire on the work(queue) 'locks', but this will then
3049 * hit the lockdep limitation on recursive locks, or simply discard
3052 * AFAICT there is no possible deadlock scenario between the
3053 * flush_work() and complete() primitives (except for single-threaded
3054 * workqueues), so hiding them isn't a problem.
3056 lockdep_invariant_state(true);
3057 trace_workqueue_execute_start(work);
3058 worker->current_func(work);
3060 * While we must be careful to not use "work" after this, the trace
3061 * point will only record its address.
3063 trace_workqueue_execute_end(work, worker->current_func);
3064 pwq->stats[PWQ_STAT_COMPLETED]++;
3065 lock_map_release(&lockdep_map);
3066 lock_map_release(&pwq->wq->lockdep_map);
3068 if (unlikely((worker->task && in_atomic()) ||
3069 lockdep_depth(current) != lockdep_start_depth ||
3070 rcu_preempt_depth() != rcu_start_depth)) {
3071 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3072 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3073 current->comm, task_pid_nr(current), preempt_count(),
3074 lockdep_start_depth, lockdep_depth(current),
3075 rcu_start_depth, rcu_preempt_depth(),
3076 worker->current_func);
3077 debug_show_held_locks(current);
3082 * The following prevents a kworker from hogging CPU on !PREEMPTION
3083 * kernels, where a requeueing work item waiting for something to
3084 * happen could deadlock with stop_machine as such work item could
3085 * indefinitely requeue itself while all other CPUs are trapped in
3086 * stop_machine. At the same time, report a quiescent RCU state so
3087 * the same condition doesn't freeze RCU.
3091 raw_spin_lock_irq(&pool->lock);
3094 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3095 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3096 * wq_cpu_intensive_thresh_us. Clear it.
3098 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3100 /* tag the worker for identification in schedule() */
3101 worker->last_func = worker->current_func;
3103 /* we're done with it, release */
3104 hash_del(&worker->hentry);
3105 worker->current_work = NULL;
3106 worker->current_func = NULL;
3107 worker->current_pwq = NULL;
3108 worker->current_color = INT_MAX;
3110 /* must be the last step, see the function comment */
3111 pwq_dec_nr_in_flight(pwq, work_data);
3115 * process_scheduled_works - process scheduled works
3118 * Process all scheduled works. Please note that the scheduled list
3119 * may change while processing a work, so this function repeatedly
3120 * fetches a work from the top and executes it.
3123 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3126 static void process_scheduled_works(struct worker *worker)
3128 struct work_struct *work;
3131 while ((work = list_first_entry_or_null(&worker->scheduled,
3132 struct work_struct, entry))) {
3134 worker->pool->watchdog_ts = jiffies;
3137 process_one_work(worker, work);
3141 static void set_pf_worker(bool val)
3143 mutex_lock(&wq_pool_attach_mutex);
3145 current->flags |= PF_WQ_WORKER;
3147 current->flags &= ~PF_WQ_WORKER;
3148 mutex_unlock(&wq_pool_attach_mutex);
3152 * worker_thread - the worker thread function
3155 * The worker thread function. All workers belong to a worker_pool -
3156 * either a per-cpu one or dynamic unbound one. These workers process all
3157 * work items regardless of their specific target workqueue. The only
3158 * exception is work items which belong to workqueues with a rescuer which
3159 * will be explained in rescuer_thread().
3163 static int worker_thread(void *__worker)
3165 struct worker *worker = __worker;
3166 struct worker_pool *pool = worker->pool;
3168 /* tell the scheduler that this is a workqueue worker */
3169 set_pf_worker(true);
3171 raw_spin_lock_irq(&pool->lock);
3173 /* am I supposed to die? */
3174 if (unlikely(worker->flags & WORKER_DIE)) {
3175 raw_spin_unlock_irq(&pool->lock);
3176 set_pf_worker(false);
3178 set_task_comm(worker->task, "kworker/dying");
3179 ida_free(&pool->worker_ida, worker->id);
3180 worker_detach_from_pool(worker);
3181 WARN_ON_ONCE(!list_empty(&worker->entry));
3186 worker_leave_idle(worker);
3188 /* no more worker necessary? */
3189 if (!need_more_worker(pool))
3192 /* do we need to manage? */
3193 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3197 * ->scheduled list can only be filled while a worker is
3198 * preparing to process a work or actually processing it.
3199 * Make sure nobody diddled with it while I was sleeping.
3201 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3204 * Finish PREP stage. We're guaranteed to have at least one idle
3205 * worker or that someone else has already assumed the manager
3206 * role. This is where @worker starts participating in concurrency
3207 * management if applicable and concurrency management is restored
3208 * after being rebound. See rebind_workers() for details.
3210 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3213 struct work_struct *work =
3214 list_first_entry(&pool->worklist,
3215 struct work_struct, entry);
3217 if (assign_work(work, worker, NULL))
3218 process_scheduled_works(worker);
3219 } while (keep_working(pool));
3221 worker_set_flags(worker, WORKER_PREP);
3224 * pool->lock is held and there's no work to process and no need to
3225 * manage, sleep. Workers are woken up only while holding
3226 * pool->lock or from local cpu, so setting the current state
3227 * before releasing pool->lock is enough to prevent losing any
3230 worker_enter_idle(worker);
3231 __set_current_state(TASK_IDLE);
3232 raw_spin_unlock_irq(&pool->lock);
3238 * rescuer_thread - the rescuer thread function
3241 * Workqueue rescuer thread function. There's one rescuer for each
3242 * workqueue which has WQ_MEM_RECLAIM set.
3244 * Regular work processing on a pool may block trying to create a new
3245 * worker which uses GFP_KERNEL allocation which has slight chance of
3246 * developing into deadlock if some works currently on the same queue
3247 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3248 * the problem rescuer solves.
3250 * When such condition is possible, the pool summons rescuers of all
3251 * workqueues which have works queued on the pool and let them process
3252 * those works so that forward progress can be guaranteed.
3254 * This should happen rarely.
3258 static int rescuer_thread(void *__rescuer)
3260 struct worker *rescuer = __rescuer;
3261 struct workqueue_struct *wq = rescuer->rescue_wq;
3264 set_user_nice(current, RESCUER_NICE_LEVEL);
3267 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3268 * doesn't participate in concurrency management.
3270 set_pf_worker(true);
3272 set_current_state(TASK_IDLE);
3275 * By the time the rescuer is requested to stop, the workqueue
3276 * shouldn't have any work pending, but @wq->maydays may still have
3277 * pwq(s) queued. This can happen by non-rescuer workers consuming
3278 * all the work items before the rescuer got to them. Go through
3279 * @wq->maydays processing before acting on should_stop so that the
3280 * list is always empty on exit.
3282 should_stop = kthread_should_stop();
3284 /* see whether any pwq is asking for help */
3285 raw_spin_lock_irq(&wq_mayday_lock);
3287 while (!list_empty(&wq->maydays)) {
3288 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3289 struct pool_workqueue, mayday_node);
3290 struct worker_pool *pool = pwq->pool;
3291 struct work_struct *work, *n;
3293 __set_current_state(TASK_RUNNING);
3294 list_del_init(&pwq->mayday_node);
3296 raw_spin_unlock_irq(&wq_mayday_lock);
3298 worker_attach_to_pool(rescuer, pool);
3300 raw_spin_lock_irq(&pool->lock);
3303 * Slurp in all works issued via this workqueue and
3306 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3307 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3308 if (get_work_pwq(work) == pwq &&
3309 assign_work(work, rescuer, &n))
3310 pwq->stats[PWQ_STAT_RESCUED]++;
3313 if (!list_empty(&rescuer->scheduled)) {
3314 process_scheduled_works(rescuer);
3317 * The above execution of rescued work items could
3318 * have created more to rescue through
3319 * pwq_activate_first_inactive() or chained
3320 * queueing. Let's put @pwq back on mayday list so
3321 * that such back-to-back work items, which may be
3322 * being used to relieve memory pressure, don't
3323 * incur MAYDAY_INTERVAL delay inbetween.
3325 if (pwq->nr_active && need_to_create_worker(pool)) {
3326 raw_spin_lock(&wq_mayday_lock);
3328 * Queue iff we aren't racing destruction
3329 * and somebody else hasn't queued it already.
3331 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3333 list_add_tail(&pwq->mayday_node, &wq->maydays);
3335 raw_spin_unlock(&wq_mayday_lock);
3340 * Put the reference grabbed by send_mayday(). @pool won't
3341 * go away while we're still attached to it.
3346 * Leave this pool. Notify regular workers; otherwise, we end up
3347 * with 0 concurrency and stalling the execution.
3351 raw_spin_unlock_irq(&pool->lock);
3353 worker_detach_from_pool(rescuer);
3355 raw_spin_lock_irq(&wq_mayday_lock);
3358 raw_spin_unlock_irq(&wq_mayday_lock);
3361 __set_current_state(TASK_RUNNING);
3362 set_pf_worker(false);
3366 /* rescuers should never participate in concurrency management */
3367 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3373 * check_flush_dependency - check for flush dependency sanity
3374 * @target_wq: workqueue being flushed
3375 * @target_work: work item being flushed (NULL for workqueue flushes)
3377 * %current is trying to flush the whole @target_wq or @target_work on it.
3378 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3379 * reclaiming memory or running on a workqueue which doesn't have
3380 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3383 static void check_flush_dependency(struct workqueue_struct *target_wq,
3384 struct work_struct *target_work)
3386 work_func_t target_func = target_work ? target_work->func : NULL;
3387 struct worker *worker;
3389 if (target_wq->flags & WQ_MEM_RECLAIM)
3392 worker = current_wq_worker();
3394 WARN_ONCE(current->flags & PF_MEMALLOC,
3395 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3396 current->pid, current->comm, target_wq->name, target_func);
3397 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3398 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3399 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3400 worker->current_pwq->wq->name, worker->current_func,
3401 target_wq->name, target_func);
3405 struct work_struct work;
3406 struct completion done;
3407 struct task_struct *task; /* purely informational */
3410 static void wq_barrier_func(struct work_struct *work)
3412 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3413 complete(&barr->done);
3417 * insert_wq_barrier - insert a barrier work
3418 * @pwq: pwq to insert barrier into
3419 * @barr: wq_barrier to insert
3420 * @target: target work to attach @barr to
3421 * @worker: worker currently executing @target, NULL if @target is not executing
3423 * @barr is linked to @target such that @barr is completed only after
3424 * @target finishes execution. Please note that the ordering
3425 * guarantee is observed only with respect to @target and on the local
3428 * Currently, a queued barrier can't be canceled. This is because
3429 * try_to_grab_pending() can't determine whether the work to be
3430 * grabbed is at the head of the queue and thus can't clear LINKED
3431 * flag of the previous work while there must be a valid next work
3432 * after a work with LINKED flag set.
3434 * Note that when @worker is non-NULL, @target may be modified
3435 * underneath us, so we can't reliably determine pwq from @target.
3438 * raw_spin_lock_irq(pool->lock).
3440 static void insert_wq_barrier(struct pool_workqueue *pwq,
3441 struct wq_barrier *barr,
3442 struct work_struct *target, struct worker *worker)
3444 unsigned int work_flags = 0;
3445 unsigned int work_color;
3446 struct list_head *head;
3449 * debugobject calls are safe here even with pool->lock locked
3450 * as we know for sure that this will not trigger any of the
3451 * checks and call back into the fixup functions where we
3454 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3455 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3457 init_completion_map(&barr->done, &target->lockdep_map);
3459 barr->task = current;
3461 /* The barrier work item does not participate in nr_active. */
3462 work_flags |= WORK_STRUCT_INACTIVE;
3465 * If @target is currently being executed, schedule the
3466 * barrier to the worker; otherwise, put it after @target.
3469 head = worker->scheduled.next;
3470 work_color = worker->current_color;
3472 unsigned long *bits = work_data_bits(target);
3474 head = target->entry.next;
3475 /* there can already be other linked works, inherit and set */
3476 work_flags |= *bits & WORK_STRUCT_LINKED;
3477 work_color = get_work_color(*bits);
3478 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3481 pwq->nr_in_flight[work_color]++;
3482 work_flags |= work_color_to_flags(work_color);
3484 insert_work(pwq, &barr->work, head, work_flags);
3488 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3489 * @wq: workqueue being flushed
3490 * @flush_color: new flush color, < 0 for no-op
3491 * @work_color: new work color, < 0 for no-op
3493 * Prepare pwqs for workqueue flushing.
3495 * If @flush_color is non-negative, flush_color on all pwqs should be
3496 * -1. If no pwq has in-flight commands at the specified color, all
3497 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3498 * has in flight commands, its pwq->flush_color is set to
3499 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3500 * wakeup logic is armed and %true is returned.
3502 * The caller should have initialized @wq->first_flusher prior to
3503 * calling this function with non-negative @flush_color. If
3504 * @flush_color is negative, no flush color update is done and %false
3507 * If @work_color is non-negative, all pwqs should have the same
3508 * work_color which is previous to @work_color and all will be
3509 * advanced to @work_color.
3512 * mutex_lock(wq->mutex).
3515 * %true if @flush_color >= 0 and there's something to flush. %false
3518 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3519 int flush_color, int work_color)
3522 struct pool_workqueue *pwq;
3524 if (flush_color >= 0) {
3525 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3526 atomic_set(&wq->nr_pwqs_to_flush, 1);
3529 for_each_pwq(pwq, wq) {
3530 struct worker_pool *pool = pwq->pool;
3532 raw_spin_lock_irq(&pool->lock);
3534 if (flush_color >= 0) {
3535 WARN_ON_ONCE(pwq->flush_color != -1);
3537 if (pwq->nr_in_flight[flush_color]) {
3538 pwq->flush_color = flush_color;
3539 atomic_inc(&wq->nr_pwqs_to_flush);
3544 if (work_color >= 0) {
3545 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3546 pwq->work_color = work_color;
3549 raw_spin_unlock_irq(&pool->lock);
3552 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3553 complete(&wq->first_flusher->done);
3558 static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3560 lock_map_acquire(&wq->lockdep_map);
3561 lock_map_release(&wq->lockdep_map);
3564 static void touch_work_lockdep_map(struct work_struct *work,
3565 struct workqueue_struct *wq)
3567 lock_map_acquire(&work->lockdep_map);
3568 lock_map_release(&work->lockdep_map);
3572 * __flush_workqueue - ensure that any scheduled work has run to completion.
3573 * @wq: workqueue to flush
3575 * This function sleeps until all work items which were queued on entry
3576 * have finished execution, but it is not livelocked by new incoming ones.
3578 void __flush_workqueue(struct workqueue_struct *wq)
3580 struct wq_flusher this_flusher = {
3581 .list = LIST_HEAD_INIT(this_flusher.list),
3583 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3587 if (WARN_ON(!wq_online))
3590 touch_wq_lockdep_map(wq);
3592 mutex_lock(&wq->mutex);
3595 * Start-to-wait phase
3597 next_color = work_next_color(wq->work_color);
3599 if (next_color != wq->flush_color) {
3601 * Color space is not full. The current work_color
3602 * becomes our flush_color and work_color is advanced
3605 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3606 this_flusher.flush_color = wq->work_color;
3607 wq->work_color = next_color;
3609 if (!wq->first_flusher) {
3610 /* no flush in progress, become the first flusher */
3611 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3613 wq->first_flusher = &this_flusher;
3615 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3617 /* nothing to flush, done */
3618 wq->flush_color = next_color;
3619 wq->first_flusher = NULL;
3624 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3625 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3626 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3630 * Oops, color space is full, wait on overflow queue.
3631 * The next flush completion will assign us
3632 * flush_color and transfer to flusher_queue.
3634 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3637 check_flush_dependency(wq, NULL);
3639 mutex_unlock(&wq->mutex);
3641 wait_for_completion(&this_flusher.done);
3644 * Wake-up-and-cascade phase
3646 * First flushers are responsible for cascading flushes and
3647 * handling overflow. Non-first flushers can simply return.
3649 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3652 mutex_lock(&wq->mutex);
3654 /* we might have raced, check again with mutex held */
3655 if (wq->first_flusher != &this_flusher)
3658 WRITE_ONCE(wq->first_flusher, NULL);
3660 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3661 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3664 struct wq_flusher *next, *tmp;
3666 /* complete all the flushers sharing the current flush color */
3667 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3668 if (next->flush_color != wq->flush_color)
3670 list_del_init(&next->list);
3671 complete(&next->done);
3674 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3675 wq->flush_color != work_next_color(wq->work_color));
3677 /* this flush_color is finished, advance by one */
3678 wq->flush_color = work_next_color(wq->flush_color);
3680 /* one color has been freed, handle overflow queue */
3681 if (!list_empty(&wq->flusher_overflow)) {
3683 * Assign the same color to all overflowed
3684 * flushers, advance work_color and append to
3685 * flusher_queue. This is the start-to-wait
3686 * phase for these overflowed flushers.
3688 list_for_each_entry(tmp, &wq->flusher_overflow, list)
3689 tmp->flush_color = wq->work_color;
3691 wq->work_color = work_next_color(wq->work_color);
3693 list_splice_tail_init(&wq->flusher_overflow,
3694 &wq->flusher_queue);
3695 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3698 if (list_empty(&wq->flusher_queue)) {
3699 WARN_ON_ONCE(wq->flush_color != wq->work_color);
3704 * Need to flush more colors. Make the next flusher
3705 * the new first flusher and arm pwqs.
3707 WARN_ON_ONCE(wq->flush_color == wq->work_color);
3708 WARN_ON_ONCE(wq->flush_color != next->flush_color);
3710 list_del_init(&next->list);
3711 wq->first_flusher = next;
3713 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3717 * Meh... this color is already done, clear first
3718 * flusher and repeat cascading.
3720 wq->first_flusher = NULL;
3724 mutex_unlock(&wq->mutex);
3726 EXPORT_SYMBOL(__flush_workqueue);
3729 * drain_workqueue - drain a workqueue
3730 * @wq: workqueue to drain
3732 * Wait until the workqueue becomes empty. While draining is in progress,
3733 * only chain queueing is allowed. IOW, only currently pending or running
3734 * work items on @wq can queue further work items on it. @wq is flushed
3735 * repeatedly until it becomes empty. The number of flushing is determined
3736 * by the depth of chaining and should be relatively short. Whine if it
3739 void drain_workqueue(struct workqueue_struct *wq)
3741 unsigned int flush_cnt = 0;
3742 struct pool_workqueue *pwq;
3745 * __queue_work() needs to test whether there are drainers, is much
3746 * hotter than drain_workqueue() and already looks at @wq->flags.
3747 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3749 mutex_lock(&wq->mutex);
3750 if (!wq->nr_drainers++)
3751 wq->flags |= __WQ_DRAINING;
3752 mutex_unlock(&wq->mutex);
3754 __flush_workqueue(wq);
3756 mutex_lock(&wq->mutex);
3758 for_each_pwq(pwq, wq) {
3761 raw_spin_lock_irq(&pwq->pool->lock);
3762 drained = pwq_is_empty(pwq);
3763 raw_spin_unlock_irq(&pwq->pool->lock);
3768 if (++flush_cnt == 10 ||
3769 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3770 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3771 wq->name, __func__, flush_cnt);
3773 mutex_unlock(&wq->mutex);
3777 if (!--wq->nr_drainers)
3778 wq->flags &= ~__WQ_DRAINING;
3779 mutex_unlock(&wq->mutex);
3781 EXPORT_SYMBOL_GPL(drain_workqueue);
3783 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3786 struct worker *worker = NULL;
3787 struct worker_pool *pool;
3788 struct pool_workqueue *pwq;
3789 struct workqueue_struct *wq;
3794 pool = get_work_pool(work);
3800 raw_spin_lock_irq(&pool->lock);
3801 /* see the comment in try_to_grab_pending() with the same code */
3802 pwq = get_work_pwq(work);
3804 if (unlikely(pwq->pool != pool))
3807 worker = find_worker_executing_work(pool, work);
3810 pwq = worker->current_pwq;
3814 check_flush_dependency(wq, work);
3816 insert_wq_barrier(pwq, barr, work, worker);
3817 raw_spin_unlock_irq(&pool->lock);
3819 touch_work_lockdep_map(work, wq);
3822 * Force a lock recursion deadlock when using flush_work() inside a
3823 * single-threaded or rescuer equipped workqueue.
3825 * For single threaded workqueues the deadlock happens when the work
3826 * is after the work issuing the flush_work(). For rescuer equipped
3827 * workqueues the deadlock happens when the rescuer stalls, blocking
3830 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
3831 touch_wq_lockdep_map(wq);
3836 raw_spin_unlock_irq(&pool->lock);
3841 static bool __flush_work(struct work_struct *work, bool from_cancel)
3843 struct wq_barrier barr;
3845 if (WARN_ON(!wq_online))
3848 if (WARN_ON(!work->func))
3851 if (start_flush_work(work, &barr, from_cancel)) {
3852 wait_for_completion(&barr.done);
3853 destroy_work_on_stack(&barr.work);
3861 * flush_work - wait for a work to finish executing the last queueing instance
3862 * @work: the work to flush
3864 * Wait until @work has finished execution. @work is guaranteed to be idle
3865 * on return if it hasn't been requeued since flush started.
3868 * %true if flush_work() waited for the work to finish execution,
3869 * %false if it was already idle.
3871 bool flush_work(struct work_struct *work)
3873 return __flush_work(work, false);
3875 EXPORT_SYMBOL_GPL(flush_work);
3878 wait_queue_entry_t wait;
3879 struct work_struct *work;
3882 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3884 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3886 if (cwait->work != key)
3888 return autoremove_wake_function(wait, mode, sync, key);
3891 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3893 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3894 unsigned long flags;
3898 ret = try_to_grab_pending(work, is_dwork, &flags);
3900 * If someone else is already canceling, wait for it to
3901 * finish. flush_work() doesn't work for PREEMPT_NONE
3902 * because we may get scheduled between @work's completion
3903 * and the other canceling task resuming and clearing
3904 * CANCELING - flush_work() will return false immediately
3905 * as @work is no longer busy, try_to_grab_pending() will
3906 * return -ENOENT as @work is still being canceled and the
3907 * other canceling task won't be able to clear CANCELING as
3908 * we're hogging the CPU.
3910 * Let's wait for completion using a waitqueue. As this
3911 * may lead to the thundering herd problem, use a custom
3912 * wake function which matches @work along with exclusive
3915 if (unlikely(ret == -ENOENT)) {
3916 struct cwt_wait cwait;
3918 init_wait(&cwait.wait);
3919 cwait.wait.func = cwt_wakefn;
3922 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3923 TASK_UNINTERRUPTIBLE);
3924 if (work_is_canceling(work))
3926 finish_wait(&cancel_waitq, &cwait.wait);
3928 } while (unlikely(ret < 0));
3930 /* tell other tasks trying to grab @work to back off */
3931 mark_work_canceling(work);
3932 local_irq_restore(flags);
3935 * This allows canceling during early boot. We know that @work
3939 __flush_work(work, true);
3941 clear_work_data(work);
3944 * Paired with prepare_to_wait() above so that either
3945 * waitqueue_active() is visible here or !work_is_canceling() is
3949 if (waitqueue_active(&cancel_waitq))
3950 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3956 * cancel_work_sync - cancel a work and wait for it to finish
3957 * @work: the work to cancel
3959 * Cancel @work and wait for its execution to finish. This function
3960 * can be used even if the work re-queues itself or migrates to
3961 * another workqueue. On return from this function, @work is
3962 * guaranteed to be not pending or executing on any CPU.
3964 * cancel_work_sync(&delayed_work->work) must not be used for
3965 * delayed_work's. Use cancel_delayed_work_sync() instead.
3967 * The caller must ensure that the workqueue on which @work was last
3968 * queued can't be destroyed before this function returns.
3971 * %true if @work was pending, %false otherwise.
3973 bool cancel_work_sync(struct work_struct *work)
3975 return __cancel_work_timer(work, false);
3977 EXPORT_SYMBOL_GPL(cancel_work_sync);
3980 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3981 * @dwork: the delayed work to flush
3983 * Delayed timer is cancelled and the pending work is queued for
3984 * immediate execution. Like flush_work(), this function only
3985 * considers the last queueing instance of @dwork.
3988 * %true if flush_work() waited for the work to finish execution,
3989 * %false if it was already idle.
3991 bool flush_delayed_work(struct delayed_work *dwork)
3993 local_irq_disable();
3994 if (del_timer_sync(&dwork->timer))
3995 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3997 return flush_work(&dwork->work);
3999 EXPORT_SYMBOL(flush_delayed_work);
4002 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4003 * @rwork: the rcu work to flush
4006 * %true if flush_rcu_work() waited for the work to finish execution,
4007 * %false if it was already idle.
4009 bool flush_rcu_work(struct rcu_work *rwork)
4011 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4013 flush_work(&rwork->work);
4016 return flush_work(&rwork->work);
4019 EXPORT_SYMBOL(flush_rcu_work);
4021 static bool __cancel_work(struct work_struct *work, bool is_dwork)
4023 unsigned long flags;
4027 ret = try_to_grab_pending(work, is_dwork, &flags);
4028 } while (unlikely(ret == -EAGAIN));
4030 if (unlikely(ret < 0))
4033 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
4034 local_irq_restore(flags);
4039 * See cancel_delayed_work()
4041 bool cancel_work(struct work_struct *work)
4043 return __cancel_work(work, false);
4045 EXPORT_SYMBOL(cancel_work);
4048 * cancel_delayed_work - cancel a delayed work
4049 * @dwork: delayed_work to cancel
4051 * Kill off a pending delayed_work.
4053 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4057 * The work callback function may still be running on return, unless
4058 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4059 * use cancel_delayed_work_sync() to wait on it.
4061 * This function is safe to call from any context including IRQ handler.
4063 bool cancel_delayed_work(struct delayed_work *dwork)
4065 return __cancel_work(&dwork->work, true);
4067 EXPORT_SYMBOL(cancel_delayed_work);
4070 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4071 * @dwork: the delayed work cancel
4073 * This is cancel_work_sync() for delayed works.
4076 * %true if @dwork was pending, %false otherwise.
4078 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4080 return __cancel_work_timer(&dwork->work, true);
4082 EXPORT_SYMBOL(cancel_delayed_work_sync);
4085 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4086 * @func: the function to call
4088 * schedule_on_each_cpu() executes @func on each online CPU using the
4089 * system workqueue and blocks until all CPUs have completed.
4090 * schedule_on_each_cpu() is very slow.
4093 * 0 on success, -errno on failure.
4095 int schedule_on_each_cpu(work_func_t func)
4098 struct work_struct __percpu *works;
4100 works = alloc_percpu(struct work_struct);
4106 for_each_online_cpu(cpu) {
4107 struct work_struct *work = per_cpu_ptr(works, cpu);
4109 INIT_WORK(work, func);
4110 schedule_work_on(cpu, work);
4113 for_each_online_cpu(cpu)
4114 flush_work(per_cpu_ptr(works, cpu));
4122 * execute_in_process_context - reliably execute the routine with user context
4123 * @fn: the function to execute
4124 * @ew: guaranteed storage for the execute work structure (must
4125 * be available when the work executes)
4127 * Executes the function immediately if process context is available,
4128 * otherwise schedules the function for delayed execution.
4130 * Return: 0 - function was executed
4131 * 1 - function was scheduled for execution
4133 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4135 if (!in_interrupt()) {
4140 INIT_WORK(&ew->work, fn);
4141 schedule_work(&ew->work);
4145 EXPORT_SYMBOL_GPL(execute_in_process_context);
4148 * free_workqueue_attrs - free a workqueue_attrs
4149 * @attrs: workqueue_attrs to free
4151 * Undo alloc_workqueue_attrs().
4153 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4156 free_cpumask_var(attrs->cpumask);
4157 free_cpumask_var(attrs->__pod_cpumask);
4163 * alloc_workqueue_attrs - allocate a workqueue_attrs
4165 * Allocate a new workqueue_attrs, initialize with default settings and
4168 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4170 struct workqueue_attrs *alloc_workqueue_attrs(void)
4172 struct workqueue_attrs *attrs;
4174 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4177 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4179 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4182 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4183 attrs->affn_scope = WQ_AFFN_DFL;
4186 free_workqueue_attrs(attrs);
4190 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4191 const struct workqueue_attrs *from)
4193 to->nice = from->nice;
4194 cpumask_copy(to->cpumask, from->cpumask);
4195 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4196 to->affn_strict = from->affn_strict;
4199 * Unlike hash and equality test, copying shouldn't ignore wq-only
4200 * fields as copying is used for both pool and wq attrs. Instead,
4201 * get_unbound_pool() explicitly clears the fields.
4203 to->affn_scope = from->affn_scope;
4204 to->ordered = from->ordered;
4208 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4209 * comments in 'struct workqueue_attrs' definition.
4211 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4213 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4214 attrs->ordered = false;
4217 /* hash value of the content of @attr */
4218 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4222 hash = jhash_1word(attrs->nice, hash);
4223 hash = jhash(cpumask_bits(attrs->cpumask),
4224 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4225 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4226 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4227 hash = jhash_1word(attrs->affn_strict, hash);
4231 /* content equality test */
4232 static bool wqattrs_equal(const struct workqueue_attrs *a,
4233 const struct workqueue_attrs *b)
4235 if (a->nice != b->nice)
4237 if (!cpumask_equal(a->cpumask, b->cpumask))
4239 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4241 if (a->affn_strict != b->affn_strict)
4246 /* Update @attrs with actually available CPUs */
4247 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4248 const cpumask_t *unbound_cpumask)
4251 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4252 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4255 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4256 if (unlikely(cpumask_empty(attrs->cpumask)))
4257 cpumask_copy(attrs->cpumask, unbound_cpumask);
4260 /* find wq_pod_type to use for @attrs */
4261 static const struct wq_pod_type *
4262 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4264 enum wq_affn_scope scope;
4265 struct wq_pod_type *pt;
4267 /* to synchronize access to wq_affn_dfl */
4268 lockdep_assert_held(&wq_pool_mutex);
4270 if (attrs->affn_scope == WQ_AFFN_DFL)
4271 scope = wq_affn_dfl;
4273 scope = attrs->affn_scope;
4275 pt = &wq_pod_types[scope];
4277 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4278 likely(pt->nr_pods))
4282 * Before workqueue_init_topology(), only SYSTEM is available which is
4283 * initialized in workqueue_init_early().
4285 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4286 BUG_ON(!pt->nr_pods);
4291 * init_worker_pool - initialize a newly zalloc'd worker_pool
4292 * @pool: worker_pool to initialize
4294 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4296 * Return: 0 on success, -errno on failure. Even on failure, all fields
4297 * inside @pool proper are initialized and put_unbound_pool() can be called
4298 * on @pool safely to release it.
4300 static int init_worker_pool(struct worker_pool *pool)
4302 raw_spin_lock_init(&pool->lock);
4305 pool->node = NUMA_NO_NODE;
4306 pool->flags |= POOL_DISASSOCIATED;
4307 pool->watchdog_ts = jiffies;
4308 INIT_LIST_HEAD(&pool->worklist);
4309 INIT_LIST_HEAD(&pool->idle_list);
4310 hash_init(pool->busy_hash);
4312 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4313 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4315 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4317 INIT_LIST_HEAD(&pool->workers);
4318 INIT_LIST_HEAD(&pool->dying_workers);
4320 ida_init(&pool->worker_ida);
4321 INIT_HLIST_NODE(&pool->hash_node);
4324 /* shouldn't fail above this point */
4325 pool->attrs = alloc_workqueue_attrs();
4329 wqattrs_clear_for_pool(pool->attrs);
4334 #ifdef CONFIG_LOCKDEP
4335 static void wq_init_lockdep(struct workqueue_struct *wq)
4339 lockdep_register_key(&wq->key);
4340 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4342 lock_name = wq->name;
4344 wq->lock_name = lock_name;
4345 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4348 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4350 lockdep_unregister_key(&wq->key);
4353 static void wq_free_lockdep(struct workqueue_struct *wq)
4355 if (wq->lock_name != wq->name)
4356 kfree(wq->lock_name);
4359 static void wq_init_lockdep(struct workqueue_struct *wq)
4363 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4367 static void wq_free_lockdep(struct workqueue_struct *wq)
4372 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4376 for_each_node(node) {
4377 kfree(nna_ar[node]);
4378 nna_ar[node] = NULL;
4381 kfree(nna_ar[nr_node_ids]);
4382 nna_ar[nr_node_ids] = NULL;
4385 static void init_node_nr_active(struct wq_node_nr_active *nna)
4387 nna->max = WQ_DFL_MIN_ACTIVE;
4388 atomic_set(&nna->nr, 0);
4389 raw_spin_lock_init(&nna->lock);
4390 INIT_LIST_HEAD(&nna->pending_pwqs);
4394 * Each node's nr_active counter will be accessed mostly from its own node and
4395 * should be allocated in the node.
4397 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4399 struct wq_node_nr_active *nna;
4402 for_each_node(node) {
4403 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4406 init_node_nr_active(nna);
4410 /* [nr_node_ids] is used as the fallback */
4411 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4414 init_node_nr_active(nna);
4415 nna_ar[nr_node_ids] = nna;
4420 free_node_nr_active(nna_ar);
4424 static void rcu_free_wq(struct rcu_head *rcu)
4426 struct workqueue_struct *wq =
4427 container_of(rcu, struct workqueue_struct, rcu);
4429 if (wq->flags & WQ_UNBOUND)
4430 free_node_nr_active(wq->node_nr_active);
4432 wq_free_lockdep(wq);
4433 free_percpu(wq->cpu_pwq);
4434 free_workqueue_attrs(wq->unbound_attrs);
4438 static void rcu_free_pool(struct rcu_head *rcu)
4440 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4442 ida_destroy(&pool->worker_ida);
4443 free_workqueue_attrs(pool->attrs);
4448 * put_unbound_pool - put a worker_pool
4449 * @pool: worker_pool to put
4451 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4452 * safe manner. get_unbound_pool() calls this function on its failure path
4453 * and this function should be able to release pools which went through,
4454 * successfully or not, init_worker_pool().
4456 * Should be called with wq_pool_mutex held.
4458 static void put_unbound_pool(struct worker_pool *pool)
4460 DECLARE_COMPLETION_ONSTACK(detach_completion);
4461 struct worker *worker;
4462 LIST_HEAD(cull_list);
4464 lockdep_assert_held(&wq_pool_mutex);
4470 if (WARN_ON(!(pool->cpu < 0)) ||
4471 WARN_ON(!list_empty(&pool->worklist)))
4474 /* release id and unhash */
4476 idr_remove(&worker_pool_idr, pool->id);
4477 hash_del(&pool->hash_node);
4480 * Become the manager and destroy all workers. This prevents
4481 * @pool's workers from blocking on attach_mutex. We're the last
4482 * manager and @pool gets freed with the flag set.
4484 * Having a concurrent manager is quite unlikely to happen as we can
4485 * only get here with
4486 * pwq->refcnt == pool->refcnt == 0
4487 * which implies no work queued to the pool, which implies no worker can
4488 * become the manager. However a worker could have taken the role of
4489 * manager before the refcnts dropped to 0, since maybe_create_worker()
4493 rcuwait_wait_event(&manager_wait,
4494 !(pool->flags & POOL_MANAGER_ACTIVE),
4495 TASK_UNINTERRUPTIBLE);
4497 mutex_lock(&wq_pool_attach_mutex);
4498 raw_spin_lock_irq(&pool->lock);
4499 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4500 pool->flags |= POOL_MANAGER_ACTIVE;
4503 raw_spin_unlock_irq(&pool->lock);
4504 mutex_unlock(&wq_pool_attach_mutex);
4507 while ((worker = first_idle_worker(pool)))
4508 set_worker_dying(worker, &cull_list);
4509 WARN_ON(pool->nr_workers || pool->nr_idle);
4510 raw_spin_unlock_irq(&pool->lock);
4512 wake_dying_workers(&cull_list);
4514 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4515 pool->detach_completion = &detach_completion;
4516 mutex_unlock(&wq_pool_attach_mutex);
4518 if (pool->detach_completion)
4519 wait_for_completion(pool->detach_completion);
4521 /* shut down the timers */
4522 del_timer_sync(&pool->idle_timer);
4523 cancel_work_sync(&pool->idle_cull_work);
4524 del_timer_sync(&pool->mayday_timer);
4526 /* RCU protected to allow dereferences from get_work_pool() */
4527 call_rcu(&pool->rcu, rcu_free_pool);
4531 * get_unbound_pool - get a worker_pool with the specified attributes
4532 * @attrs: the attributes of the worker_pool to get
4534 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4535 * reference count and return it. If there already is a matching
4536 * worker_pool, it will be used; otherwise, this function attempts to
4539 * Should be called with wq_pool_mutex held.
4541 * Return: On success, a worker_pool with the same attributes as @attrs.
4542 * On failure, %NULL.
4544 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4546 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4547 u32 hash = wqattrs_hash(attrs);
4548 struct worker_pool *pool;
4549 int pod, node = NUMA_NO_NODE;
4551 lockdep_assert_held(&wq_pool_mutex);
4553 /* do we already have a matching pool? */
4554 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4555 if (wqattrs_equal(pool->attrs, attrs)) {
4561 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4562 for (pod = 0; pod < pt->nr_pods; pod++) {
4563 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4564 node = pt->pod_node[pod];
4569 /* nope, create a new one */
4570 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4571 if (!pool || init_worker_pool(pool) < 0)
4575 copy_workqueue_attrs(pool->attrs, attrs);
4576 wqattrs_clear_for_pool(pool->attrs);
4578 if (worker_pool_assign_id(pool) < 0)
4581 /* create and start the initial worker */
4582 if (wq_online && !create_worker(pool))
4586 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4591 put_unbound_pool(pool);
4595 static void rcu_free_pwq(struct rcu_head *rcu)
4597 kmem_cache_free(pwq_cache,
4598 container_of(rcu, struct pool_workqueue, rcu));
4602 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4603 * refcnt and needs to be destroyed.
4605 static void pwq_release_workfn(struct kthread_work *work)
4607 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4609 struct workqueue_struct *wq = pwq->wq;
4610 struct worker_pool *pool = pwq->pool;
4611 bool is_last = false;
4614 * When @pwq is not linked, it doesn't hold any reference to the
4615 * @wq, and @wq is invalid to access.
4617 if (!list_empty(&pwq->pwqs_node)) {
4618 mutex_lock(&wq->mutex);
4619 list_del_rcu(&pwq->pwqs_node);
4620 is_last = list_empty(&wq->pwqs);
4621 mutex_unlock(&wq->mutex);
4624 if (wq->flags & WQ_UNBOUND) {
4625 mutex_lock(&wq_pool_mutex);
4626 put_unbound_pool(pool);
4627 mutex_unlock(&wq_pool_mutex);
4630 if (!list_empty(&pwq->pending_node)) {
4631 struct wq_node_nr_active *nna =
4632 wq_node_nr_active(pwq->wq, pwq->pool->node);
4634 raw_spin_lock_irq(&nna->lock);
4635 list_del_init(&pwq->pending_node);
4636 raw_spin_unlock_irq(&nna->lock);
4639 call_rcu(&pwq->rcu, rcu_free_pwq);
4642 * If we're the last pwq going away, @wq is already dead and no one
4643 * is gonna access it anymore. Schedule RCU free.
4646 wq_unregister_lockdep(wq);
4647 call_rcu(&wq->rcu, rcu_free_wq);
4651 /* initialize newly allocated @pwq which is associated with @wq and @pool */
4652 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4653 struct worker_pool *pool)
4655 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4657 memset(pwq, 0, sizeof(*pwq));
4661 pwq->flush_color = -1;
4663 INIT_LIST_HEAD(&pwq->inactive_works);
4664 INIT_LIST_HEAD(&pwq->pending_node);
4665 INIT_LIST_HEAD(&pwq->pwqs_node);
4666 INIT_LIST_HEAD(&pwq->mayday_node);
4667 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4670 /* sync @pwq with the current state of its associated wq and link it */
4671 static void link_pwq(struct pool_workqueue *pwq)
4673 struct workqueue_struct *wq = pwq->wq;
4675 lockdep_assert_held(&wq->mutex);
4677 /* may be called multiple times, ignore if already linked */
4678 if (!list_empty(&pwq->pwqs_node))
4681 /* set the matching work_color */
4682 pwq->work_color = wq->work_color;
4685 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4688 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4689 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4690 const struct workqueue_attrs *attrs)
4692 struct worker_pool *pool;
4693 struct pool_workqueue *pwq;
4695 lockdep_assert_held(&wq_pool_mutex);
4697 pool = get_unbound_pool(attrs);
4701 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4703 put_unbound_pool(pool);
4707 init_pwq(pwq, wq, pool);
4712 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4713 * @attrs: the wq_attrs of the default pwq of the target workqueue
4714 * @cpu: the target CPU
4715 * @cpu_going_down: if >= 0, the CPU to consider as offline
4717 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4718 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4719 * The result is stored in @attrs->__pod_cpumask.
4721 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4722 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4723 * intersection of the possible CPUs of @pod and @attrs->cpumask.
4725 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4727 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4730 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4731 int pod = pt->cpu_pod[cpu];
4733 /* does @pod have any online CPUs @attrs wants? */
4734 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4735 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4736 if (cpu_going_down >= 0)
4737 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4739 if (cpumask_empty(attrs->__pod_cpumask)) {
4740 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4744 /* yeap, return possible CPUs in @pod that @attrs wants */
4745 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4747 if (cpumask_empty(attrs->__pod_cpumask))
4748 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4749 "possible intersect\n");
4752 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
4753 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4754 int cpu, struct pool_workqueue *pwq)
4756 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
4757 struct pool_workqueue *old_pwq;
4759 lockdep_assert_held(&wq_pool_mutex);
4760 lockdep_assert_held(&wq->mutex);
4762 /* link_pwq() can handle duplicate calls */
4765 old_pwq = rcu_access_pointer(*slot);
4766 rcu_assign_pointer(*slot, pwq);
4770 /* context to store the prepared attrs & pwqs before applying */
4771 struct apply_wqattrs_ctx {
4772 struct workqueue_struct *wq; /* target workqueue */
4773 struct workqueue_attrs *attrs; /* attrs to apply */
4774 struct list_head list; /* queued for batching commit */
4775 struct pool_workqueue *dfl_pwq;
4776 struct pool_workqueue *pwq_tbl[];
4779 /* free the resources after success or abort */
4780 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4785 for_each_possible_cpu(cpu)
4786 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4787 put_pwq_unlocked(ctx->dfl_pwq);
4789 free_workqueue_attrs(ctx->attrs);
4795 /* allocate the attrs and pwqs for later installation */
4796 static struct apply_wqattrs_ctx *
4797 apply_wqattrs_prepare(struct workqueue_struct *wq,
4798 const struct workqueue_attrs *attrs,
4799 const cpumask_var_t unbound_cpumask)
4801 struct apply_wqattrs_ctx *ctx;
4802 struct workqueue_attrs *new_attrs;
4805 lockdep_assert_held(&wq_pool_mutex);
4807 if (WARN_ON(attrs->affn_scope < 0 ||
4808 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4809 return ERR_PTR(-EINVAL);
4811 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4813 new_attrs = alloc_workqueue_attrs();
4814 if (!ctx || !new_attrs)
4818 * If something goes wrong during CPU up/down, we'll fall back to
4819 * the default pwq covering whole @attrs->cpumask. Always create
4820 * it even if we don't use it immediately.
4822 copy_workqueue_attrs(new_attrs, attrs);
4823 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4824 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4825 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4829 for_each_possible_cpu(cpu) {
4830 if (new_attrs->ordered) {
4831 ctx->dfl_pwq->refcnt++;
4832 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4834 wq_calc_pod_cpumask(new_attrs, cpu, -1);
4835 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4836 if (!ctx->pwq_tbl[cpu])
4841 /* save the user configured attrs and sanitize it. */
4842 copy_workqueue_attrs(new_attrs, attrs);
4843 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4844 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4845 ctx->attrs = new_attrs;
4851 free_workqueue_attrs(new_attrs);
4852 apply_wqattrs_cleanup(ctx);
4853 return ERR_PTR(-ENOMEM);
4856 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4857 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4861 /* all pwqs have been created successfully, let's install'em */
4862 mutex_lock(&ctx->wq->mutex);
4864 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4866 /* save the previous pwqs and install the new ones */
4867 for_each_possible_cpu(cpu)
4868 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4870 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
4872 /* update node_nr_active->max */
4873 wq_update_node_max_active(ctx->wq, -1);
4875 mutex_unlock(&ctx->wq->mutex);
4878 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4879 const struct workqueue_attrs *attrs)
4881 struct apply_wqattrs_ctx *ctx;
4883 /* only unbound workqueues can change attributes */
4884 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4887 /* creating multiple pwqs breaks ordering guarantee */
4888 if (!list_empty(&wq->pwqs)) {
4889 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4892 wq->flags &= ~__WQ_ORDERED;
4895 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4897 return PTR_ERR(ctx);
4899 /* the ctx has been prepared successfully, let's commit it */
4900 apply_wqattrs_commit(ctx);
4901 apply_wqattrs_cleanup(ctx);
4907 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4908 * @wq: the target workqueue
4909 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4911 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4912 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4913 * work items are affine to the pod it was issued on. Older pwqs are released as
4914 * in-flight work items finish. Note that a work item which repeatedly requeues
4915 * itself back-to-back will stay on its current pwq.
4917 * Performs GFP_KERNEL allocations.
4919 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4921 * Return: 0 on success and -errno on failure.
4923 int apply_workqueue_attrs(struct workqueue_struct *wq,
4924 const struct workqueue_attrs *attrs)
4928 lockdep_assert_cpus_held();
4930 mutex_lock(&wq_pool_mutex);
4931 ret = apply_workqueue_attrs_locked(wq, attrs);
4932 mutex_unlock(&wq_pool_mutex);
4938 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4939 * @wq: the target workqueue
4940 * @cpu: the CPU to update pool association for
4941 * @hotplug_cpu: the CPU coming up or going down
4942 * @online: whether @cpu is coming up or going down
4944 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4945 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
4949 * If pod affinity can't be adjusted due to memory allocation failure, it falls
4950 * back to @wq->dfl_pwq which may not be optimal but is always correct.
4952 * Note that when the last allowed CPU of a pod goes offline for a workqueue
4953 * with a cpumask spanning multiple pods, the workers which were already
4954 * executing the work items for the workqueue will lose their CPU affinity and
4955 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
4956 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4957 * responsibility to flush the work item from CPU_DOWN_PREPARE.
4959 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4960 int hotplug_cpu, bool online)
4962 int off_cpu = online ? -1 : hotplug_cpu;
4963 struct pool_workqueue *old_pwq = NULL, *pwq;
4964 struct workqueue_attrs *target_attrs;
4966 lockdep_assert_held(&wq_pool_mutex);
4968 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
4972 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4973 * Let's use a preallocated one. The following buf is protected by
4974 * CPU hotplug exclusion.
4976 target_attrs = wq_update_pod_attrs_buf;
4978 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4979 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4981 /* nothing to do if the target cpumask matches the current pwq */
4982 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4983 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
4986 /* create a new pwq */
4987 pwq = alloc_unbound_pwq(wq, target_attrs);
4989 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4994 /* Install the new pwq. */
4995 mutex_lock(&wq->mutex);
4996 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5000 mutex_lock(&wq->mutex);
5001 pwq = unbound_pwq(wq, -1);
5002 raw_spin_lock_irq(&pwq->pool->lock);
5004 raw_spin_unlock_irq(&pwq->pool->lock);
5005 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5007 mutex_unlock(&wq->mutex);
5008 put_pwq_unlocked(old_pwq);
5011 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5013 bool highpri = wq->flags & WQ_HIGHPRI;
5016 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5020 if (!(wq->flags & WQ_UNBOUND)) {
5021 for_each_possible_cpu(cpu) {
5022 struct pool_workqueue **pwq_p =
5023 per_cpu_ptr(wq->cpu_pwq, cpu);
5024 struct worker_pool *pool =
5025 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
5027 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5032 init_pwq(*pwq_p, wq, pool);
5034 mutex_lock(&wq->mutex);
5036 mutex_unlock(&wq->mutex);
5042 if (wq->flags & __WQ_ORDERED) {
5043 struct pool_workqueue *dfl_pwq;
5045 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5046 /* there should only be single pwq for ordering guarantee */
5047 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5048 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5049 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5050 "ordering guarantee broken for workqueue %s\n", wq->name);
5052 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5056 /* for unbound pwq, flush the pwq_release_worker ensures that the
5057 * pwq_release_workfn() completes before calling kfree(wq).
5060 kthread_flush_worker(pwq_release_worker);
5066 for_each_possible_cpu(cpu) {
5067 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5070 kmem_cache_free(pwq_cache, pwq);
5072 free_percpu(wq->cpu_pwq);
5078 static int wq_clamp_max_active(int max_active, unsigned int flags,
5081 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5082 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5083 max_active, name, 1, WQ_MAX_ACTIVE);
5085 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5089 * Workqueues which may be used during memory reclaim should have a rescuer
5090 * to guarantee forward progress.
5092 static int init_rescuer(struct workqueue_struct *wq)
5094 struct worker *rescuer;
5097 if (!(wq->flags & WQ_MEM_RECLAIM))
5100 rescuer = alloc_worker(NUMA_NO_NODE);
5102 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5107 rescuer->rescue_wq = wq;
5108 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5109 if (IS_ERR(rescuer->task)) {
5110 ret = PTR_ERR(rescuer->task);
5111 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5112 wq->name, ERR_PTR(ret));
5117 wq->rescuer = rescuer;
5118 if (wq->flags & WQ_UNBOUND)
5119 kthread_bind_mask(rescuer->task, wq->unbound_attrs->cpumask);
5121 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5122 wake_up_process(rescuer->task);
5128 * wq_adjust_max_active - update a wq's max_active to the current setting
5129 * @wq: target workqueue
5131 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5132 * activate inactive work items accordingly. If @wq is freezing, clear
5133 * @wq->max_active to zero.
5135 static void wq_adjust_max_active(struct workqueue_struct *wq)
5138 int new_max, new_min;
5140 lockdep_assert_held(&wq->mutex);
5142 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5146 new_max = wq->saved_max_active;
5147 new_min = wq->saved_min_active;
5150 if (wq->max_active == new_max && wq->min_active == new_min)
5154 * Update @wq->max/min_active and then kick inactive work items if more
5155 * active work items are allowed. This doesn't break work item ordering
5156 * because new work items are always queued behind existing inactive
5157 * work items if there are any.
5159 WRITE_ONCE(wq->max_active, new_max);
5160 WRITE_ONCE(wq->min_active, new_min);
5162 if (wq->flags & WQ_UNBOUND)
5163 wq_update_node_max_active(wq, -1);
5169 * Round-robin through pwq's activating the first inactive work item
5170 * until max_active is filled.
5173 struct pool_workqueue *pwq;
5176 for_each_pwq(pwq, wq) {
5177 unsigned long flags;
5179 /* can be called during early boot w/ irq disabled */
5180 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5181 if (pwq_activate_first_inactive(pwq, true)) {
5183 kick_pool(pwq->pool);
5185 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5187 } while (activated);
5191 struct workqueue_struct *alloc_workqueue(const char *fmt,
5193 int max_active, ...)
5196 struct workqueue_struct *wq;
5201 * Unbound && max_active == 1 used to imply ordered, which is no longer
5202 * the case on many machines due to per-pod pools. While
5203 * alloc_ordered_workqueue() is the right way to create an ordered
5204 * workqueue, keep the previous behavior to avoid subtle breakages.
5206 if ((flags & WQ_UNBOUND) && max_active == 1)
5207 flags |= __WQ_ORDERED;
5209 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5210 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5211 flags |= WQ_UNBOUND;
5213 /* allocate wq and format name */
5214 if (flags & WQ_UNBOUND)
5215 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5217 wq_size = sizeof(*wq);
5219 wq = kzalloc(wq_size, GFP_KERNEL);
5223 if (flags & WQ_UNBOUND) {
5224 wq->unbound_attrs = alloc_workqueue_attrs();
5225 if (!wq->unbound_attrs)
5229 va_start(args, max_active);
5230 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5233 if (name_len >= WQ_NAME_LEN)
5234 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5237 max_active = max_active ?: WQ_DFL_ACTIVE;
5238 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5242 wq->max_active = max_active;
5243 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5244 wq->saved_max_active = wq->max_active;
5245 wq->saved_min_active = wq->min_active;
5246 mutex_init(&wq->mutex);
5247 atomic_set(&wq->nr_pwqs_to_flush, 0);
5248 INIT_LIST_HEAD(&wq->pwqs);
5249 INIT_LIST_HEAD(&wq->flusher_queue);
5250 INIT_LIST_HEAD(&wq->flusher_overflow);
5251 INIT_LIST_HEAD(&wq->maydays);
5253 wq_init_lockdep(wq);
5254 INIT_LIST_HEAD(&wq->list);
5256 if (flags & WQ_UNBOUND) {
5257 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5258 goto err_unreg_lockdep;
5261 if (alloc_and_link_pwqs(wq) < 0)
5262 goto err_free_node_nr_active;
5264 if (wq_online && init_rescuer(wq) < 0)
5267 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5271 * wq_pool_mutex protects global freeze state and workqueues list.
5272 * Grab it, adjust max_active and add the new @wq to workqueues
5275 mutex_lock(&wq_pool_mutex);
5277 mutex_lock(&wq->mutex);
5278 wq_adjust_max_active(wq);
5279 mutex_unlock(&wq->mutex);
5281 list_add_tail_rcu(&wq->list, &workqueues);
5283 mutex_unlock(&wq_pool_mutex);
5287 err_free_node_nr_active:
5288 if (wq->flags & WQ_UNBOUND)
5289 free_node_nr_active(wq->node_nr_active);
5291 wq_unregister_lockdep(wq);
5292 wq_free_lockdep(wq);
5294 free_workqueue_attrs(wq->unbound_attrs);
5298 destroy_workqueue(wq);
5301 EXPORT_SYMBOL_GPL(alloc_workqueue);
5303 static bool pwq_busy(struct pool_workqueue *pwq)
5307 for (i = 0; i < WORK_NR_COLORS; i++)
5308 if (pwq->nr_in_flight[i])
5311 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5313 if (!pwq_is_empty(pwq))
5320 * destroy_workqueue - safely terminate a workqueue
5321 * @wq: target workqueue
5323 * Safely destroy a workqueue. All work currently pending will be done first.
5325 void destroy_workqueue(struct workqueue_struct *wq)
5327 struct pool_workqueue *pwq;
5331 * Remove it from sysfs first so that sanity check failure doesn't
5332 * lead to sysfs name conflicts.
5334 workqueue_sysfs_unregister(wq);
5336 /* mark the workqueue destruction is in progress */
5337 mutex_lock(&wq->mutex);
5338 wq->flags |= __WQ_DESTROYING;
5339 mutex_unlock(&wq->mutex);
5341 /* drain it before proceeding with destruction */
5342 drain_workqueue(wq);
5344 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5346 struct worker *rescuer = wq->rescuer;
5348 /* this prevents new queueing */
5349 raw_spin_lock_irq(&wq_mayday_lock);
5351 raw_spin_unlock_irq(&wq_mayday_lock);
5353 /* rescuer will empty maydays list before exiting */
5354 kthread_stop(rescuer->task);
5359 * Sanity checks - grab all the locks so that we wait for all
5360 * in-flight operations which may do put_pwq().
5362 mutex_lock(&wq_pool_mutex);
5363 mutex_lock(&wq->mutex);
5364 for_each_pwq(pwq, wq) {
5365 raw_spin_lock_irq(&pwq->pool->lock);
5366 if (WARN_ON(pwq_busy(pwq))) {
5367 pr_warn("%s: %s has the following busy pwq\n",
5368 __func__, wq->name);
5370 raw_spin_unlock_irq(&pwq->pool->lock);
5371 mutex_unlock(&wq->mutex);
5372 mutex_unlock(&wq_pool_mutex);
5373 show_one_workqueue(wq);
5376 raw_spin_unlock_irq(&pwq->pool->lock);
5378 mutex_unlock(&wq->mutex);
5381 * wq list is used to freeze wq, remove from list after
5382 * flushing is complete in case freeze races us.
5384 list_del_rcu(&wq->list);
5385 mutex_unlock(&wq_pool_mutex);
5388 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5389 * to put the base refs. @wq will be auto-destroyed from the last
5390 * pwq_put. RCU read lock prevents @wq from going away from under us.
5394 for_each_possible_cpu(cpu) {
5395 put_pwq_unlocked(unbound_pwq(wq, cpu));
5396 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5399 put_pwq_unlocked(unbound_pwq(wq, -1));
5400 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5404 EXPORT_SYMBOL_GPL(destroy_workqueue);
5407 * workqueue_set_max_active - adjust max_active of a workqueue
5408 * @wq: target workqueue
5409 * @max_active: new max_active value.
5411 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5415 * Don't call from IRQ context.
5417 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5419 /* disallow meddling with max_active for ordered workqueues */
5420 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5423 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5425 mutex_lock(&wq->mutex);
5427 wq->flags &= ~__WQ_ORDERED;
5428 wq->saved_max_active = max_active;
5429 if (wq->flags & WQ_UNBOUND)
5430 wq->saved_min_active = min(wq->saved_min_active, max_active);
5432 wq_adjust_max_active(wq);
5434 mutex_unlock(&wq->mutex);
5436 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5439 * current_work - retrieve %current task's work struct
5441 * Determine if %current task is a workqueue worker and what it's working on.
5442 * Useful to find out the context that the %current task is running in.
5444 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5446 struct work_struct *current_work(void)
5448 struct worker *worker = current_wq_worker();
5450 return worker ? worker->current_work : NULL;
5452 EXPORT_SYMBOL(current_work);
5455 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5457 * Determine whether %current is a workqueue rescuer. Can be used from
5458 * work functions to determine whether it's being run off the rescuer task.
5460 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5462 bool current_is_workqueue_rescuer(void)
5464 struct worker *worker = current_wq_worker();
5466 return worker && worker->rescue_wq;
5470 * workqueue_congested - test whether a workqueue is congested
5471 * @cpu: CPU in question
5472 * @wq: target workqueue
5474 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5475 * no synchronization around this function and the test result is
5476 * unreliable and only useful as advisory hints or for debugging.
5478 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5480 * With the exception of ordered workqueues, all workqueues have per-cpu
5481 * pool_workqueues, each with its own congested state. A workqueue being
5482 * congested on one CPU doesn't mean that the workqueue is contested on any
5486 * %true if congested, %false otherwise.
5488 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5490 struct pool_workqueue *pwq;
5496 if (cpu == WORK_CPU_UNBOUND)
5497 cpu = smp_processor_id();
5499 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5500 ret = !list_empty(&pwq->inactive_works);
5507 EXPORT_SYMBOL_GPL(workqueue_congested);
5510 * work_busy - test whether a work is currently pending or running
5511 * @work: the work to be tested
5513 * Test whether @work is currently pending or running. There is no
5514 * synchronization around this function and the test result is
5515 * unreliable and only useful as advisory hints or for debugging.
5518 * OR'd bitmask of WORK_BUSY_* bits.
5520 unsigned int work_busy(struct work_struct *work)
5522 struct worker_pool *pool;
5523 unsigned long flags;
5524 unsigned int ret = 0;
5526 if (work_pending(work))
5527 ret |= WORK_BUSY_PENDING;
5530 pool = get_work_pool(work);
5532 raw_spin_lock_irqsave(&pool->lock, flags);
5533 if (find_worker_executing_work(pool, work))
5534 ret |= WORK_BUSY_RUNNING;
5535 raw_spin_unlock_irqrestore(&pool->lock, flags);
5541 EXPORT_SYMBOL_GPL(work_busy);
5544 * set_worker_desc - set description for the current work item
5545 * @fmt: printf-style format string
5546 * @...: arguments for the format string
5548 * This function can be called by a running work function to describe what
5549 * the work item is about. If the worker task gets dumped, this
5550 * information will be printed out together to help debugging. The
5551 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5553 void set_worker_desc(const char *fmt, ...)
5555 struct worker *worker = current_wq_worker();
5559 va_start(args, fmt);
5560 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5564 EXPORT_SYMBOL_GPL(set_worker_desc);
5567 * print_worker_info - print out worker information and description
5568 * @log_lvl: the log level to use when printing
5569 * @task: target task
5571 * If @task is a worker and currently executing a work item, print out the
5572 * name of the workqueue being serviced and worker description set with
5573 * set_worker_desc() by the currently executing work item.
5575 * This function can be safely called on any task as long as the
5576 * task_struct itself is accessible. While safe, this function isn't
5577 * synchronized and may print out mixups or garbages of limited length.
5579 void print_worker_info(const char *log_lvl, struct task_struct *task)
5581 work_func_t *fn = NULL;
5582 char name[WQ_NAME_LEN] = { };
5583 char desc[WORKER_DESC_LEN] = { };
5584 struct pool_workqueue *pwq = NULL;
5585 struct workqueue_struct *wq = NULL;
5586 struct worker *worker;
5588 if (!(task->flags & PF_WQ_WORKER))
5592 * This function is called without any synchronization and @task
5593 * could be in any state. Be careful with dereferences.
5595 worker = kthread_probe_data(task);
5598 * Carefully copy the associated workqueue's workfn, name and desc.
5599 * Keep the original last '\0' in case the original is garbage.
5601 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5602 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5603 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5604 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5605 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5607 if (fn || name[0] || desc[0]) {
5608 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5609 if (strcmp(name, desc))
5610 pr_cont(" (%s)", desc);
5615 static void pr_cont_pool_info(struct worker_pool *pool)
5617 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5618 if (pool->node != NUMA_NO_NODE)
5619 pr_cont(" node=%d", pool->node);
5620 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5623 struct pr_cont_work_struct {
5629 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5633 if (func == pcwsp->func) {
5637 if (pcwsp->ctr == 1)
5638 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5640 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5643 if ((long)func == -1L)
5645 pcwsp->comma = comma;
5650 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5652 if (work->func == wq_barrier_func) {
5653 struct wq_barrier *barr;
5655 barr = container_of(work, struct wq_barrier, work);
5657 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5658 pr_cont("%s BAR(%d)", comma ? "," : "",
5659 task_pid_nr(barr->task));
5662 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5663 pr_cont_work_flush(comma, work->func, pcwsp);
5667 static void show_pwq(struct pool_workqueue *pwq)
5669 struct pr_cont_work_struct pcws = { .ctr = 0, };
5670 struct worker_pool *pool = pwq->pool;
5671 struct work_struct *work;
5672 struct worker *worker;
5673 bool has_in_flight = false, has_pending = false;
5676 pr_info(" pwq %d:", pool->id);
5677 pr_cont_pool_info(pool);
5679 pr_cont(" active=%d refcnt=%d%s\n",
5680 pwq->nr_active, pwq->refcnt,
5681 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5683 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5684 if (worker->current_pwq == pwq) {
5685 has_in_flight = true;
5689 if (has_in_flight) {
5692 pr_info(" in-flight:");
5693 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5694 if (worker->current_pwq != pwq)
5697 pr_cont("%s %d%s:%ps", comma ? "," : "",
5698 task_pid_nr(worker->task),
5699 worker->rescue_wq ? "(RESCUER)" : "",
5700 worker->current_func);
5701 list_for_each_entry(work, &worker->scheduled, entry)
5702 pr_cont_work(false, work, &pcws);
5703 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5709 list_for_each_entry(work, &pool->worklist, entry) {
5710 if (get_work_pwq(work) == pwq) {
5718 pr_info(" pending:");
5719 list_for_each_entry(work, &pool->worklist, entry) {
5720 if (get_work_pwq(work) != pwq)
5723 pr_cont_work(comma, work, &pcws);
5724 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5726 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5730 if (!list_empty(&pwq->inactive_works)) {
5733 pr_info(" inactive:");
5734 list_for_each_entry(work, &pwq->inactive_works, entry) {
5735 pr_cont_work(comma, work, &pcws);
5736 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5738 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5744 * show_one_workqueue - dump state of specified workqueue
5745 * @wq: workqueue whose state will be printed
5747 void show_one_workqueue(struct workqueue_struct *wq)
5749 struct pool_workqueue *pwq;
5751 unsigned long flags;
5753 for_each_pwq(pwq, wq) {
5754 if (!pwq_is_empty(pwq)) {
5759 if (idle) /* Nothing to print for idle workqueue */
5762 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5764 for_each_pwq(pwq, wq) {
5765 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5766 if (!pwq_is_empty(pwq)) {
5768 * Defer printing to avoid deadlocks in console
5769 * drivers that queue work while holding locks
5770 * also taken in their write paths.
5772 printk_deferred_enter();
5774 printk_deferred_exit();
5776 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5778 * We could be printing a lot from atomic context, e.g.
5779 * sysrq-t -> show_all_workqueues(). Avoid triggering
5782 touch_nmi_watchdog();
5788 * show_one_worker_pool - dump state of specified worker pool
5789 * @pool: worker pool whose state will be printed
5791 static void show_one_worker_pool(struct worker_pool *pool)
5793 struct worker *worker;
5795 unsigned long flags;
5796 unsigned long hung = 0;
5798 raw_spin_lock_irqsave(&pool->lock, flags);
5799 if (pool->nr_workers == pool->nr_idle)
5802 /* How long the first pending work is waiting for a worker. */
5803 if (!list_empty(&pool->worklist))
5804 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5807 * Defer printing to avoid deadlocks in console drivers that
5808 * queue work while holding locks also taken in their write
5811 printk_deferred_enter();
5812 pr_info("pool %d:", pool->id);
5813 pr_cont_pool_info(pool);
5814 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5816 pr_cont(" manager: %d",
5817 task_pid_nr(pool->manager->task));
5818 list_for_each_entry(worker, &pool->idle_list, entry) {
5819 pr_cont(" %s%d", first ? "idle: " : "",
5820 task_pid_nr(worker->task));
5824 printk_deferred_exit();
5826 raw_spin_unlock_irqrestore(&pool->lock, flags);
5828 * We could be printing a lot from atomic context, e.g.
5829 * sysrq-t -> show_all_workqueues(). Avoid triggering
5832 touch_nmi_watchdog();
5837 * show_all_workqueues - dump workqueue state
5839 * Called from a sysrq handler and prints out all busy workqueues and pools.
5841 void show_all_workqueues(void)
5843 struct workqueue_struct *wq;
5844 struct worker_pool *pool;
5849 pr_info("Showing busy workqueues and worker pools:\n");
5851 list_for_each_entry_rcu(wq, &workqueues, list)
5852 show_one_workqueue(wq);
5854 for_each_pool(pool, pi)
5855 show_one_worker_pool(pool);
5861 * show_freezable_workqueues - dump freezable workqueue state
5863 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
5866 void show_freezable_workqueues(void)
5868 struct workqueue_struct *wq;
5872 pr_info("Showing freezable workqueues that are still busy:\n");
5874 list_for_each_entry_rcu(wq, &workqueues, list) {
5875 if (!(wq->flags & WQ_FREEZABLE))
5877 show_one_workqueue(wq);
5883 /* used to show worker information through /proc/PID/{comm,stat,status} */
5884 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
5888 /* always show the actual comm */
5889 off = strscpy(buf, task->comm, size);
5893 /* stabilize PF_WQ_WORKER and worker pool association */
5894 mutex_lock(&wq_pool_attach_mutex);
5896 if (task->flags & PF_WQ_WORKER) {
5897 struct worker *worker = kthread_data(task);
5898 struct worker_pool *pool = worker->pool;
5901 raw_spin_lock_irq(&pool->lock);
5903 * ->desc tracks information (wq name or
5904 * set_worker_desc()) for the latest execution. If
5905 * current, prepend '+', otherwise '-'.
5907 if (worker->desc[0] != '\0') {
5908 if (worker->current_work)
5909 scnprintf(buf + off, size - off, "+%s",
5912 scnprintf(buf + off, size - off, "-%s",
5915 raw_spin_unlock_irq(&pool->lock);
5919 mutex_unlock(&wq_pool_attach_mutex);
5927 * There are two challenges in supporting CPU hotplug. Firstly, there
5928 * are a lot of assumptions on strong associations among work, pwq and
5929 * pool which make migrating pending and scheduled works very
5930 * difficult to implement without impacting hot paths. Secondly,
5931 * worker pools serve mix of short, long and very long running works making
5932 * blocked draining impractical.
5934 * This is solved by allowing the pools to be disassociated from the CPU
5935 * running as an unbound one and allowing it to be reattached later if the
5936 * cpu comes back online.
5939 static void unbind_workers(int cpu)
5941 struct worker_pool *pool;
5942 struct worker *worker;
5944 for_each_cpu_worker_pool(pool, cpu) {
5945 mutex_lock(&wq_pool_attach_mutex);
5946 raw_spin_lock_irq(&pool->lock);
5949 * We've blocked all attach/detach operations. Make all workers
5950 * unbound and set DISASSOCIATED. Before this, all workers
5951 * must be on the cpu. After this, they may become diasporas.
5952 * And the preemption disabled section in their sched callbacks
5953 * are guaranteed to see WORKER_UNBOUND since the code here
5954 * is on the same cpu.
5956 for_each_pool_worker(worker, pool)
5957 worker->flags |= WORKER_UNBOUND;
5959 pool->flags |= POOL_DISASSOCIATED;
5962 * The handling of nr_running in sched callbacks are disabled
5963 * now. Zap nr_running. After this, nr_running stays zero and
5964 * need_more_worker() and keep_working() are always true as
5965 * long as the worklist is not empty. This pool now behaves as
5966 * an unbound (in terms of concurrency management) pool which
5967 * are served by workers tied to the pool.
5969 pool->nr_running = 0;
5972 * With concurrency management just turned off, a busy
5973 * worker blocking could lead to lengthy stalls. Kick off
5974 * unbound chain execution of currently pending work items.
5978 raw_spin_unlock_irq(&pool->lock);
5980 for_each_pool_worker(worker, pool)
5981 unbind_worker(worker);
5983 mutex_unlock(&wq_pool_attach_mutex);
5988 * rebind_workers - rebind all workers of a pool to the associated CPU
5989 * @pool: pool of interest
5991 * @pool->cpu is coming online. Rebind all workers to the CPU.
5993 static void rebind_workers(struct worker_pool *pool)
5995 struct worker *worker;
5997 lockdep_assert_held(&wq_pool_attach_mutex);
6000 * Restore CPU affinity of all workers. As all idle workers should
6001 * be on the run-queue of the associated CPU before any local
6002 * wake-ups for concurrency management happen, restore CPU affinity
6003 * of all workers first and then clear UNBOUND. As we're called
6004 * from CPU_ONLINE, the following shouldn't fail.
6006 for_each_pool_worker(worker, pool) {
6007 kthread_set_per_cpu(worker->task, pool->cpu);
6008 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6009 pool_allowed_cpus(pool)) < 0);
6012 raw_spin_lock_irq(&pool->lock);
6014 pool->flags &= ~POOL_DISASSOCIATED;
6016 for_each_pool_worker(worker, pool) {
6017 unsigned int worker_flags = worker->flags;
6020 * We want to clear UNBOUND but can't directly call
6021 * worker_clr_flags() or adjust nr_running. Atomically
6022 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6023 * @worker will clear REBOUND using worker_clr_flags() when
6024 * it initiates the next execution cycle thus restoring
6025 * concurrency management. Note that when or whether
6026 * @worker clears REBOUND doesn't affect correctness.
6028 * WRITE_ONCE() is necessary because @worker->flags may be
6029 * tested without holding any lock in
6030 * wq_worker_running(). Without it, NOT_RUNNING test may
6031 * fail incorrectly leading to premature concurrency
6032 * management operations.
6034 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6035 worker_flags |= WORKER_REBOUND;
6036 worker_flags &= ~WORKER_UNBOUND;
6037 WRITE_ONCE(worker->flags, worker_flags);
6040 raw_spin_unlock_irq(&pool->lock);
6044 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6045 * @pool: unbound pool of interest
6046 * @cpu: the CPU which is coming up
6048 * An unbound pool may end up with a cpumask which doesn't have any online
6049 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6050 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6051 * online CPU before, cpus_allowed of all its workers should be restored.
6053 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6055 static cpumask_t cpumask;
6056 struct worker *worker;
6058 lockdep_assert_held(&wq_pool_attach_mutex);
6060 /* is @cpu allowed for @pool? */
6061 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6064 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6066 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6067 for_each_pool_worker(worker, pool)
6068 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6071 int workqueue_prepare_cpu(unsigned int cpu)
6073 struct worker_pool *pool;
6075 for_each_cpu_worker_pool(pool, cpu) {
6076 if (pool->nr_workers)
6078 if (!create_worker(pool))
6084 int workqueue_online_cpu(unsigned int cpu)
6086 struct worker_pool *pool;
6087 struct workqueue_struct *wq;
6090 mutex_lock(&wq_pool_mutex);
6092 for_each_pool(pool, pi) {
6093 mutex_lock(&wq_pool_attach_mutex);
6095 if (pool->cpu == cpu)
6096 rebind_workers(pool);
6097 else if (pool->cpu < 0)
6098 restore_unbound_workers_cpumask(pool, cpu);
6100 mutex_unlock(&wq_pool_attach_mutex);
6103 /* update pod affinity of unbound workqueues */
6104 list_for_each_entry(wq, &workqueues, list) {
6105 struct workqueue_attrs *attrs = wq->unbound_attrs;
6108 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6111 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6112 wq_update_pod(wq, tcpu, cpu, true);
6114 mutex_lock(&wq->mutex);
6115 wq_update_node_max_active(wq, -1);
6116 mutex_unlock(&wq->mutex);
6120 mutex_unlock(&wq_pool_mutex);
6124 int workqueue_offline_cpu(unsigned int cpu)
6126 struct workqueue_struct *wq;
6128 /* unbinding per-cpu workers should happen on the local CPU */
6129 if (WARN_ON(cpu != smp_processor_id()))
6132 unbind_workers(cpu);
6134 /* update pod affinity of unbound workqueues */
6135 mutex_lock(&wq_pool_mutex);
6136 list_for_each_entry(wq, &workqueues, list) {
6137 struct workqueue_attrs *attrs = wq->unbound_attrs;
6140 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6143 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6144 wq_update_pod(wq, tcpu, cpu, false);
6146 mutex_lock(&wq->mutex);
6147 wq_update_node_max_active(wq, cpu);
6148 mutex_unlock(&wq->mutex);
6151 mutex_unlock(&wq_pool_mutex);
6156 struct work_for_cpu {
6157 struct work_struct work;
6163 static void work_for_cpu_fn(struct work_struct *work)
6165 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6167 wfc->ret = wfc->fn(wfc->arg);
6171 * work_on_cpu_key - run a function in thread context on a particular cpu
6172 * @cpu: the cpu to run on
6173 * @fn: the function to run
6174 * @arg: the function arg
6175 * @key: The lock class key for lock debugging purposes
6177 * It is up to the caller to ensure that the cpu doesn't go offline.
6178 * The caller must not hold any locks which would prevent @fn from completing.
6180 * Return: The value @fn returns.
6182 long work_on_cpu_key(int cpu, long (*fn)(void *),
6183 void *arg, struct lock_class_key *key)
6185 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6187 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6188 schedule_work_on(cpu, &wfc.work);
6189 flush_work(&wfc.work);
6190 destroy_work_on_stack(&wfc.work);
6193 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6196 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6197 * @cpu: the cpu to run on
6198 * @fn: the function to run
6199 * @arg: the function argument
6200 * @key: The lock class key for lock debugging purposes
6202 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6203 * any locks which would prevent @fn from completing.
6205 * Return: The value @fn returns.
6207 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6208 void *arg, struct lock_class_key *key)
6213 if (cpu_online(cpu))
6214 ret = work_on_cpu_key(cpu, fn, arg, key);
6218 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6219 #endif /* CONFIG_SMP */
6221 #ifdef CONFIG_FREEZER
6224 * freeze_workqueues_begin - begin freezing workqueues
6226 * Start freezing workqueues. After this function returns, all freezable
6227 * workqueues will queue new works to their inactive_works list instead of
6231 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6233 void freeze_workqueues_begin(void)
6235 struct workqueue_struct *wq;
6237 mutex_lock(&wq_pool_mutex);
6239 WARN_ON_ONCE(workqueue_freezing);
6240 workqueue_freezing = true;
6242 list_for_each_entry(wq, &workqueues, list) {
6243 mutex_lock(&wq->mutex);
6244 wq_adjust_max_active(wq);
6245 mutex_unlock(&wq->mutex);
6248 mutex_unlock(&wq_pool_mutex);
6252 * freeze_workqueues_busy - are freezable workqueues still busy?
6254 * Check whether freezing is complete. This function must be called
6255 * between freeze_workqueues_begin() and thaw_workqueues().
6258 * Grabs and releases wq_pool_mutex.
6261 * %true if some freezable workqueues are still busy. %false if freezing
6264 bool freeze_workqueues_busy(void)
6267 struct workqueue_struct *wq;
6268 struct pool_workqueue *pwq;
6270 mutex_lock(&wq_pool_mutex);
6272 WARN_ON_ONCE(!workqueue_freezing);
6274 list_for_each_entry(wq, &workqueues, list) {
6275 if (!(wq->flags & WQ_FREEZABLE))
6278 * nr_active is monotonically decreasing. It's safe
6279 * to peek without lock.
6282 for_each_pwq(pwq, wq) {
6283 WARN_ON_ONCE(pwq->nr_active < 0);
6284 if (pwq->nr_active) {
6293 mutex_unlock(&wq_pool_mutex);
6298 * thaw_workqueues - thaw workqueues
6300 * Thaw workqueues. Normal queueing is restored and all collected
6301 * frozen works are transferred to their respective pool worklists.
6304 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6306 void thaw_workqueues(void)
6308 struct workqueue_struct *wq;
6310 mutex_lock(&wq_pool_mutex);
6312 if (!workqueue_freezing)
6315 workqueue_freezing = false;
6317 /* restore max_active and repopulate worklist */
6318 list_for_each_entry(wq, &workqueues, list) {
6319 mutex_lock(&wq->mutex);
6320 wq_adjust_max_active(wq);
6321 mutex_unlock(&wq->mutex);
6325 mutex_unlock(&wq_pool_mutex);
6327 #endif /* CONFIG_FREEZER */
6329 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6333 struct workqueue_struct *wq;
6334 struct apply_wqattrs_ctx *ctx, *n;
6336 lockdep_assert_held(&wq_pool_mutex);
6338 list_for_each_entry(wq, &workqueues, list) {
6339 if (!(wq->flags & WQ_UNBOUND))
6342 /* creating multiple pwqs breaks ordering guarantee */
6343 if (!list_empty(&wq->pwqs)) {
6344 if (wq->flags & __WQ_ORDERED_EXPLICIT)
6346 wq->flags &= ~__WQ_ORDERED;
6349 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6355 list_add_tail(&ctx->list, &ctxs);
6358 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6360 apply_wqattrs_commit(ctx);
6361 apply_wqattrs_cleanup(ctx);
6365 mutex_lock(&wq_pool_attach_mutex);
6366 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6367 mutex_unlock(&wq_pool_attach_mutex);
6373 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6374 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6376 * This function can be called from cpuset code to provide a set of isolated
6377 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
6378 * either cpus_read_lock or cpus_write_lock.
6380 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6382 cpumask_var_t cpumask;
6385 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6388 lockdep_assert_cpus_held();
6389 mutex_lock(&wq_pool_mutex);
6391 /* Save the current isolated cpumask & export it via sysfs */
6392 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6395 * If the operation fails, it will fall back to
6396 * wq_requested_unbound_cpumask which is initially set to
6397 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6398 * by any subsequent write to workqueue/cpumask sysfs file.
6400 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6401 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6402 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6403 ret = workqueue_apply_unbound_cpumask(cpumask);
6405 mutex_unlock(&wq_pool_mutex);
6406 free_cpumask_var(cpumask);
6410 static int parse_affn_scope(const char *val)
6414 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6415 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6421 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6423 struct workqueue_struct *wq;
6426 affn = parse_affn_scope(val);
6429 if (affn == WQ_AFFN_DFL)
6433 mutex_lock(&wq_pool_mutex);
6437 list_for_each_entry(wq, &workqueues, list) {
6438 for_each_online_cpu(cpu) {
6439 wq_update_pod(wq, cpu, cpu, true);
6443 mutex_unlock(&wq_pool_mutex);
6449 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6451 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6454 static const struct kernel_param_ops wq_affn_dfl_ops = {
6455 .set = wq_affn_dfl_set,
6456 .get = wq_affn_dfl_get,
6459 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6463 * Workqueues with WQ_SYSFS flag set is visible to userland via
6464 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6465 * following attributes.
6467 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6468 * max_active RW int : maximum number of in-flight work items
6470 * Unbound workqueues have the following extra attributes.
6472 * nice RW int : nice value of the workers
6473 * cpumask RW mask : bitmask of allowed CPUs for the workers
6474 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6475 * affinity_strict RW bool : worker CPU affinity is strict
6478 struct workqueue_struct *wq;
6482 static struct workqueue_struct *dev_to_wq(struct device *dev)
6484 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6489 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6492 struct workqueue_struct *wq = dev_to_wq(dev);
6494 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6496 static DEVICE_ATTR_RO(per_cpu);
6498 static ssize_t max_active_show(struct device *dev,
6499 struct device_attribute *attr, char *buf)
6501 struct workqueue_struct *wq = dev_to_wq(dev);
6503 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6506 static ssize_t max_active_store(struct device *dev,
6507 struct device_attribute *attr, const char *buf,
6510 struct workqueue_struct *wq = dev_to_wq(dev);
6513 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
6516 workqueue_set_max_active(wq, val);
6519 static DEVICE_ATTR_RW(max_active);
6521 static struct attribute *wq_sysfs_attrs[] = {
6522 &dev_attr_per_cpu.attr,
6523 &dev_attr_max_active.attr,
6526 ATTRIBUTE_GROUPS(wq_sysfs);
6528 static void apply_wqattrs_lock(void)
6530 /* CPUs should stay stable across pwq creations and installations */
6532 mutex_lock(&wq_pool_mutex);
6535 static void apply_wqattrs_unlock(void)
6537 mutex_unlock(&wq_pool_mutex);
6541 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
6544 struct workqueue_struct *wq = dev_to_wq(dev);
6547 mutex_lock(&wq->mutex);
6548 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
6549 mutex_unlock(&wq->mutex);
6554 /* prepare workqueue_attrs for sysfs store operations */
6555 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
6557 struct workqueue_attrs *attrs;
6559 lockdep_assert_held(&wq_pool_mutex);
6561 attrs = alloc_workqueue_attrs();
6565 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6569 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6570 const char *buf, size_t count)
6572 struct workqueue_struct *wq = dev_to_wq(dev);
6573 struct workqueue_attrs *attrs;
6576 apply_wqattrs_lock();
6578 attrs = wq_sysfs_prep_attrs(wq);
6582 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6583 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6584 ret = apply_workqueue_attrs_locked(wq, attrs);
6589 apply_wqattrs_unlock();
6590 free_workqueue_attrs(attrs);
6591 return ret ?: count;
6594 static ssize_t wq_cpumask_show(struct device *dev,
6595 struct device_attribute *attr, char *buf)
6597 struct workqueue_struct *wq = dev_to_wq(dev);
6600 mutex_lock(&wq->mutex);
6601 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6602 cpumask_pr_args(wq->unbound_attrs->cpumask));
6603 mutex_unlock(&wq->mutex);
6607 static ssize_t wq_cpumask_store(struct device *dev,
6608 struct device_attribute *attr,
6609 const char *buf, size_t count)
6611 struct workqueue_struct *wq = dev_to_wq(dev);
6612 struct workqueue_attrs *attrs;
6615 apply_wqattrs_lock();
6617 attrs = wq_sysfs_prep_attrs(wq);
6621 ret = cpumask_parse(buf, attrs->cpumask);
6623 ret = apply_workqueue_attrs_locked(wq, attrs);
6626 apply_wqattrs_unlock();
6627 free_workqueue_attrs(attrs);
6628 return ret ?: count;
6631 static ssize_t wq_affn_scope_show(struct device *dev,
6632 struct device_attribute *attr, char *buf)
6634 struct workqueue_struct *wq = dev_to_wq(dev);
6637 mutex_lock(&wq->mutex);
6638 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6639 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6640 wq_affn_names[WQ_AFFN_DFL],
6641 wq_affn_names[wq_affn_dfl]);
6643 written = scnprintf(buf, PAGE_SIZE, "%s\n",
6644 wq_affn_names[wq->unbound_attrs->affn_scope]);
6645 mutex_unlock(&wq->mutex);
6650 static ssize_t wq_affn_scope_store(struct device *dev,
6651 struct device_attribute *attr,
6652 const char *buf, size_t count)
6654 struct workqueue_struct *wq = dev_to_wq(dev);
6655 struct workqueue_attrs *attrs;
6656 int affn, ret = -ENOMEM;
6658 affn = parse_affn_scope(buf);
6662 apply_wqattrs_lock();
6663 attrs = wq_sysfs_prep_attrs(wq);
6665 attrs->affn_scope = affn;
6666 ret = apply_workqueue_attrs_locked(wq, attrs);
6668 apply_wqattrs_unlock();
6669 free_workqueue_attrs(attrs);
6670 return ret ?: count;
6673 static ssize_t wq_affinity_strict_show(struct device *dev,
6674 struct device_attribute *attr, char *buf)
6676 struct workqueue_struct *wq = dev_to_wq(dev);
6678 return scnprintf(buf, PAGE_SIZE, "%d\n",
6679 wq->unbound_attrs->affn_strict);
6682 static ssize_t wq_affinity_strict_store(struct device *dev,
6683 struct device_attribute *attr,
6684 const char *buf, size_t count)
6686 struct workqueue_struct *wq = dev_to_wq(dev);
6687 struct workqueue_attrs *attrs;
6688 int v, ret = -ENOMEM;
6690 if (sscanf(buf, "%d", &v) != 1)
6693 apply_wqattrs_lock();
6694 attrs = wq_sysfs_prep_attrs(wq);
6696 attrs->affn_strict = (bool)v;
6697 ret = apply_workqueue_attrs_locked(wq, attrs);
6699 apply_wqattrs_unlock();
6700 free_workqueue_attrs(attrs);
6701 return ret ?: count;
6704 static struct device_attribute wq_sysfs_unbound_attrs[] = {
6705 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6706 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6707 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6708 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6712 static const struct bus_type wq_subsys = {
6713 .name = "workqueue",
6714 .dev_groups = wq_sysfs_groups,
6718 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6719 * @cpumask: the cpumask to set
6721 * The low-level workqueues cpumask is a global cpumask that limits
6722 * the affinity of all unbound workqueues. This function check the @cpumask
6723 * and apply it to all unbound workqueues and updates all pwqs of them.
6725 * Return: 0 - Success
6726 * -EINVAL - Invalid @cpumask
6727 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6729 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6734 * Not excluding isolated cpus on purpose.
6735 * If the user wishes to include them, we allow that.
6737 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6738 if (!cpumask_empty(cpumask)) {
6739 apply_wqattrs_lock();
6740 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
6741 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
6746 ret = workqueue_apply_unbound_cpumask(cpumask);
6749 apply_wqattrs_unlock();
6755 static ssize_t __wq_cpumask_show(struct device *dev,
6756 struct device_attribute *attr, char *buf, cpumask_var_t mask)
6760 mutex_lock(&wq_pool_mutex);
6761 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
6762 mutex_unlock(&wq_pool_mutex);
6767 static ssize_t wq_unbound_cpumask_show(struct device *dev,
6768 struct device_attribute *attr, char *buf)
6770 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
6773 static ssize_t wq_requested_cpumask_show(struct device *dev,
6774 struct device_attribute *attr, char *buf)
6776 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
6779 static ssize_t wq_isolated_cpumask_show(struct device *dev,
6780 struct device_attribute *attr, char *buf)
6782 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
6785 static ssize_t wq_unbound_cpumask_store(struct device *dev,
6786 struct device_attribute *attr, const char *buf, size_t count)
6788 cpumask_var_t cpumask;
6791 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6794 ret = cpumask_parse(buf, cpumask);
6796 ret = workqueue_set_unbound_cpumask(cpumask);
6798 free_cpumask_var(cpumask);
6799 return ret ? ret : count;
6802 static struct device_attribute wq_sysfs_cpumask_attrs[] = {
6803 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6804 wq_unbound_cpumask_store),
6805 __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
6806 __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
6810 static int __init wq_sysfs_init(void)
6812 struct device *dev_root;
6815 err = subsys_virtual_register(&wq_subsys, NULL);
6819 dev_root = bus_get_dev_root(&wq_subsys);
6821 struct device_attribute *attr;
6823 for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
6824 err = device_create_file(dev_root, attr);
6828 put_device(dev_root);
6832 core_initcall(wq_sysfs_init);
6834 static void wq_device_release(struct device *dev)
6836 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6842 * workqueue_sysfs_register - make a workqueue visible in sysfs
6843 * @wq: the workqueue to register
6845 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6846 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6847 * which is the preferred method.
6849 * Workqueue user should use this function directly iff it wants to apply
6850 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6851 * apply_workqueue_attrs() may race against userland updating the
6854 * Return: 0 on success, -errno on failure.
6856 int workqueue_sysfs_register(struct workqueue_struct *wq)
6858 struct wq_device *wq_dev;
6862 * Adjusting max_active or creating new pwqs by applying
6863 * attributes breaks ordering guarantee. Disallow exposing ordered
6866 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6869 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6874 wq_dev->dev.bus = &wq_subsys;
6875 wq_dev->dev.release = wq_device_release;
6876 dev_set_name(&wq_dev->dev, "%s", wq->name);
6879 * unbound_attrs are created separately. Suppress uevent until
6880 * everything is ready.
6882 dev_set_uevent_suppress(&wq_dev->dev, true);
6884 ret = device_register(&wq_dev->dev);
6886 put_device(&wq_dev->dev);
6891 if (wq->flags & WQ_UNBOUND) {
6892 struct device_attribute *attr;
6894 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6895 ret = device_create_file(&wq_dev->dev, attr);
6897 device_unregister(&wq_dev->dev);
6904 dev_set_uevent_suppress(&wq_dev->dev, false);
6905 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6910 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6911 * @wq: the workqueue to unregister
6913 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6915 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6917 struct wq_device *wq_dev = wq->wq_dev;
6923 device_unregister(&wq_dev->dev);
6925 #else /* CONFIG_SYSFS */
6926 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
6927 #endif /* CONFIG_SYSFS */
6930 * Workqueue watchdog.
6932 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6933 * flush dependency, a concurrency managed work item which stays RUNNING
6934 * indefinitely. Workqueue stalls can be very difficult to debug as the
6935 * usual warning mechanisms don't trigger and internal workqueue state is
6938 * Workqueue watchdog monitors all worker pools periodically and dumps
6939 * state if some pools failed to make forward progress for a while where
6940 * forward progress is defined as the first item on ->worklist changing.
6942 * This mechanism is controlled through the kernel parameter
6943 * "workqueue.watchdog_thresh" which can be updated at runtime through the
6944 * corresponding sysfs parameter file.
6946 #ifdef CONFIG_WQ_WATCHDOG
6948 static unsigned long wq_watchdog_thresh = 30;
6949 static struct timer_list wq_watchdog_timer;
6951 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6952 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6955 * Show workers that might prevent the processing of pending work items.
6956 * The only candidates are CPU-bound workers in the running state.
6957 * Pending work items should be handled by another idle worker
6958 * in all other situations.
6960 static void show_cpu_pool_hog(struct worker_pool *pool)
6962 struct worker *worker;
6963 unsigned long flags;
6966 raw_spin_lock_irqsave(&pool->lock, flags);
6968 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6969 if (task_is_running(worker->task)) {
6971 * Defer printing to avoid deadlocks in console
6972 * drivers that queue work while holding locks
6973 * also taken in their write paths.
6975 printk_deferred_enter();
6977 pr_info("pool %d:\n", pool->id);
6978 sched_show_task(worker->task);
6980 printk_deferred_exit();
6984 raw_spin_unlock_irqrestore(&pool->lock, flags);
6987 static void show_cpu_pools_hogs(void)
6989 struct worker_pool *pool;
6992 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6996 for_each_pool(pool, pi) {
6997 if (pool->cpu_stall)
6998 show_cpu_pool_hog(pool);
7005 static void wq_watchdog_reset_touched(void)
7009 wq_watchdog_touched = jiffies;
7010 for_each_possible_cpu(cpu)
7011 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7014 static void wq_watchdog_timer_fn(struct timer_list *unused)
7016 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7017 bool lockup_detected = false;
7018 bool cpu_pool_stall = false;
7019 unsigned long now = jiffies;
7020 struct worker_pool *pool;
7028 for_each_pool(pool, pi) {
7029 unsigned long pool_ts, touched, ts;
7031 pool->cpu_stall = false;
7032 if (list_empty(&pool->worklist))
7036 * If a virtual machine is stopped by the host it can look to
7037 * the watchdog like a stall.
7039 kvm_check_and_clear_guest_paused();
7041 /* get the latest of pool and touched timestamps */
7043 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7045 touched = READ_ONCE(wq_watchdog_touched);
7046 pool_ts = READ_ONCE(pool->watchdog_ts);
7048 if (time_after(pool_ts, touched))
7054 if (time_after(now, ts + thresh)) {
7055 lockup_detected = true;
7056 if (pool->cpu >= 0) {
7057 pool->cpu_stall = true;
7058 cpu_pool_stall = true;
7060 pr_emerg("BUG: workqueue lockup - pool");
7061 pr_cont_pool_info(pool);
7062 pr_cont(" stuck for %us!\n",
7063 jiffies_to_msecs(now - pool_ts) / 1000);
7071 if (lockup_detected)
7072 show_all_workqueues();
7075 show_cpu_pools_hogs();
7077 wq_watchdog_reset_touched();
7078 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7081 notrace void wq_watchdog_touch(int cpu)
7084 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7086 wq_watchdog_touched = jiffies;
7089 static void wq_watchdog_set_thresh(unsigned long thresh)
7091 wq_watchdog_thresh = 0;
7092 del_timer_sync(&wq_watchdog_timer);
7095 wq_watchdog_thresh = thresh;
7096 wq_watchdog_reset_touched();
7097 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7101 static int wq_watchdog_param_set_thresh(const char *val,
7102 const struct kernel_param *kp)
7104 unsigned long thresh;
7107 ret = kstrtoul(val, 0, &thresh);
7112 wq_watchdog_set_thresh(thresh);
7114 wq_watchdog_thresh = thresh;
7119 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7120 .set = wq_watchdog_param_set_thresh,
7121 .get = param_get_ulong,
7124 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7127 static void wq_watchdog_init(void)
7129 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7130 wq_watchdog_set_thresh(wq_watchdog_thresh);
7133 #else /* CONFIG_WQ_WATCHDOG */
7135 static inline void wq_watchdog_init(void) { }
7137 #endif /* CONFIG_WQ_WATCHDOG */
7139 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7141 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7142 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7143 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7147 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7150 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7152 BUG_ON(init_worker_pool(pool));
7154 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7155 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7156 pool->attrs->nice = nice;
7157 pool->attrs->affn_strict = true;
7158 pool->node = cpu_to_node(cpu);
7161 mutex_lock(&wq_pool_mutex);
7162 BUG_ON(worker_pool_assign_id(pool));
7163 mutex_unlock(&wq_pool_mutex);
7167 * workqueue_init_early - early init for workqueue subsystem
7169 * This is the first step of three-staged workqueue subsystem initialization and
7170 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7171 * up. It sets up all the data structures and system workqueues and allows early
7172 * boot code to create workqueues and queue/cancel work items. Actual work item
7173 * execution starts only after kthreads can be created and scheduled right
7174 * before early initcalls.
7176 void __init workqueue_init_early(void)
7178 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7179 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7182 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7184 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7185 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7186 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7188 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7189 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7190 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7191 if (!cpumask_empty(&wq_cmdline_cpumask))
7192 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7194 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7196 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7198 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7199 BUG_ON(!wq_update_pod_attrs_buf);
7202 * If nohz_full is enabled, set power efficient workqueue as unbound.
7203 * This allows workqueue items to be moved to HK CPUs.
7205 if (housekeeping_enabled(HK_TYPE_TICK))
7206 wq_power_efficient = true;
7208 /* initialize WQ_AFFN_SYSTEM pods */
7209 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7210 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7211 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7212 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7214 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7217 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7218 pt->pod_node[0] = NUMA_NO_NODE;
7221 /* initialize CPU pools */
7222 for_each_possible_cpu(cpu) {
7223 struct worker_pool *pool;
7226 for_each_cpu_worker_pool(pool, cpu)
7227 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7230 /* create default unbound and ordered wq attrs */
7231 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7232 struct workqueue_attrs *attrs;
7234 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7235 attrs->nice = std_nice[i];
7236 unbound_std_wq_attrs[i] = attrs;
7239 * An ordered wq should have only one pwq as ordering is
7240 * guaranteed by max_active which is enforced by pwqs.
7242 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7243 attrs->nice = std_nice[i];
7244 attrs->ordered = true;
7245 ordered_wq_attrs[i] = attrs;
7248 system_wq = alloc_workqueue("events", 0, 0);
7249 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7250 system_long_wq = alloc_workqueue("events_long", 0, 0);
7251 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7253 system_freezable_wq = alloc_workqueue("events_freezable",
7255 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7256 WQ_POWER_EFFICIENT, 0);
7257 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7258 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7260 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7261 !system_unbound_wq || !system_freezable_wq ||
7262 !system_power_efficient_wq ||
7263 !system_freezable_power_efficient_wq);
7266 static void __init wq_cpu_intensive_thresh_init(void)
7268 unsigned long thresh;
7271 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7272 BUG_ON(IS_ERR(pwq_release_worker));
7274 /* if the user set it to a specific value, keep it */
7275 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7279 * The default of 10ms is derived from the fact that most modern (as of
7280 * 2023) processors can do a lot in 10ms and that it's just below what
7281 * most consider human-perceivable. However, the kernel also runs on a
7282 * lot slower CPUs including microcontrollers where the threshold is way
7285 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7286 * This is by no means accurate but it doesn't have to be. The mechanism
7287 * is still useful even when the threshold is fully scaled up. Also, as
7288 * the reports would usually be applicable to everyone, some machines
7289 * operating on longer thresholds won't significantly diminish their
7292 thresh = 10 * USEC_PER_MSEC;
7294 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7295 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7297 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7299 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7300 loops_per_jiffy, bogo, thresh);
7302 wq_cpu_intensive_thresh_us = thresh;
7306 * workqueue_init - bring workqueue subsystem fully online
7308 * This is the second step of three-staged workqueue subsystem initialization
7309 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7310 * been created and work items queued on them, but there are no kworkers
7311 * executing the work items yet. Populate the worker pools with the initial
7312 * workers and enable future kworker creations.
7314 void __init workqueue_init(void)
7316 struct workqueue_struct *wq;
7317 struct worker_pool *pool;
7320 wq_cpu_intensive_thresh_init();
7322 mutex_lock(&wq_pool_mutex);
7325 * Per-cpu pools created earlier could be missing node hint. Fix them
7326 * up. Also, create a rescuer for workqueues that requested it.
7328 for_each_possible_cpu(cpu) {
7329 for_each_cpu_worker_pool(pool, cpu) {
7330 pool->node = cpu_to_node(cpu);
7334 list_for_each_entry(wq, &workqueues, list) {
7335 WARN(init_rescuer(wq),
7336 "workqueue: failed to create early rescuer for %s",
7340 mutex_unlock(&wq_pool_mutex);
7342 /* create the initial workers */
7343 for_each_online_cpu(cpu) {
7344 for_each_cpu_worker_pool(pool, cpu) {
7345 pool->flags &= ~POOL_DISASSOCIATED;
7346 BUG_ON(!create_worker(pool));
7350 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7351 BUG_ON(!create_worker(pool));
7358 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7359 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7360 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7362 static void __init init_pod_type(struct wq_pod_type *pt,
7363 bool (*cpus_share_pod)(int, int))
7365 int cur, pre, cpu, pod;
7369 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7370 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7371 BUG_ON(!pt->cpu_pod);
7373 for_each_possible_cpu(cur) {
7374 for_each_possible_cpu(pre) {
7376 pt->cpu_pod[cur] = pt->nr_pods++;
7379 if (cpus_share_pod(cur, pre)) {
7380 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7386 /* init the rest to match @pt->cpu_pod[] */
7387 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7388 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7389 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7391 for (pod = 0; pod < pt->nr_pods; pod++)
7392 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7394 for_each_possible_cpu(cpu) {
7395 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7396 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7400 static bool __init cpus_dont_share(int cpu0, int cpu1)
7405 static bool __init cpus_share_smt(int cpu0, int cpu1)
7407 #ifdef CONFIG_SCHED_SMT
7408 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7414 static bool __init cpus_share_numa(int cpu0, int cpu1)
7416 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7420 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7422 * This is the third step of there-staged workqueue subsystem initialization and
7423 * invoked after SMP and topology information are fully initialized. It
7424 * initializes the unbound CPU pods accordingly.
7426 void __init workqueue_init_topology(void)
7428 struct workqueue_struct *wq;
7431 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7432 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7433 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7434 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7436 wq_topo_initialized = true;
7438 mutex_lock(&wq_pool_mutex);
7441 * Workqueues allocated earlier would have all CPUs sharing the default
7442 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7443 * combinations to apply per-pod sharing.
7445 list_for_each_entry(wq, &workqueues, list) {
7446 for_each_online_cpu(cpu)
7447 wq_update_pod(wq, cpu, cpu, true);
7448 if (wq->flags & WQ_UNBOUND) {
7449 mutex_lock(&wq->mutex);
7450 wq_update_node_max_active(wq, -1);
7451 mutex_unlock(&wq->mutex);
7455 mutex_unlock(&wq_pool_mutex);
7458 void __warn_flushing_systemwide_wq(void)
7460 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7463 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7465 static int __init workqueue_unbound_cpus_setup(char *str)
7467 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7468 cpumask_clear(&wq_cmdline_cpumask);
7469 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7474 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);