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/interrupt.h>
33 #include <linux/signal.h>
34 #include <linux/completion.h>
35 #include <linux/workqueue.h>
36 #include <linux/slab.h>
37 #include <linux/cpu.h>
38 #include <linux/notifier.h>
39 #include <linux/kthread.h>
40 #include <linux/hardirq.h>
41 #include <linux/mempolicy.h>
42 #include <linux/freezer.h>
43 #include <linux/debug_locks.h>
44 #include <linux/lockdep.h>
45 #include <linux/idr.h>
46 #include <linux/jhash.h>
47 #include <linux/hashtable.h>
48 #include <linux/rculist.h>
49 #include <linux/nodemask.h>
50 #include <linux/moduleparam.h>
51 #include <linux/uaccess.h>
52 #include <linux/sched/isolation.h>
53 #include <linux/sched/debug.h>
54 #include <linux/nmi.h>
55 #include <linux/kvm_para.h>
56 #include <linux/delay.h>
58 #include "workqueue_internal.h"
60 enum worker_pool_flags {
64 * A bound pool is either associated or disassociated with its CPU.
65 * While associated (!DISASSOCIATED), all workers are bound to the
66 * CPU and none has %WORKER_UNBOUND set and concurrency management
69 * While DISASSOCIATED, the cpu may be offline and all workers have
70 * %WORKER_UNBOUND set and concurrency management disabled, and may
71 * be executing on any CPU. The pool behaves as an unbound one.
73 * Note that DISASSOCIATED should be flipped only while holding
74 * wq_pool_attach_mutex to avoid changing binding state while
75 * worker_attach_to_pool() is in progress.
77 * As there can only be one concurrent BH execution context per CPU, a
78 * BH pool is per-CPU and always DISASSOCIATED.
80 POOL_BH = 1 << 0, /* is a BH pool */
81 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
82 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
87 WORKER_DIE = 1 << 1, /* die die die */
88 WORKER_IDLE = 1 << 2, /* is idle */
89 WORKER_PREP = 1 << 3, /* preparing to run works */
90 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
91 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
92 WORKER_REBOUND = 1 << 8, /* worker was rebound */
94 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
95 WORKER_UNBOUND | WORKER_REBOUND,
98 enum wq_internal_consts {
99 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
101 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
102 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
104 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
105 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
107 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
108 /* call for help after 10ms
110 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
111 CREATE_COOLDOWN = HZ, /* time to breath after fail */
114 * Rescue workers are used only on emergencies and shared by
115 * all cpus. Give MIN_NICE.
117 RESCUER_NICE_LEVEL = MIN_NICE,
118 HIGHPRI_NICE_LEVEL = MIN_NICE,
124 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
125 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
126 * msecs_to_jiffies() can't be an initializer.
128 #define BH_WORKER_JIFFIES msecs_to_jiffies(2)
129 #define BH_WORKER_RESTARTS 10
132 * Structure fields follow one of the following exclusion rules.
134 * I: Modifiable by initialization/destruction paths and read-only for
137 * P: Preemption protected. Disabling preemption is enough and should
138 * only be modified and accessed from the local cpu.
140 * L: pool->lock protected. Access with pool->lock held.
142 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
145 * K: Only modified by worker while holding pool->lock. Can be safely read by
146 * self, while holding pool->lock or from IRQ context if %current is the
149 * S: Only modified by worker self.
151 * A: wq_pool_attach_mutex protected.
153 * PL: wq_pool_mutex protected.
155 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
157 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
159 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
162 * WQ: wq->mutex protected.
164 * WR: wq->mutex protected for writes. RCU protected for reads.
166 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
167 * with READ_ONCE() without locking.
169 * MD: wq_mayday_lock protected.
171 * WD: Used internally by the watchdog.
174 /* struct worker is defined in workqueue_internal.h */
177 raw_spinlock_t lock; /* the pool lock */
178 int cpu; /* I: the associated cpu */
179 int node; /* I: the associated node ID */
180 int id; /* I: pool ID */
181 unsigned int flags; /* L: flags */
183 unsigned long watchdog_ts; /* L: watchdog timestamp */
184 bool cpu_stall; /* WD: stalled cpu bound pool */
187 * The counter is incremented in a process context on the associated CPU
188 * w/ preemption disabled, and decremented or reset in the same context
189 * but w/ pool->lock held. The readers grab pool->lock and are
190 * guaranteed to see if the counter reached zero.
194 struct list_head worklist; /* L: list of pending works */
196 int nr_workers; /* L: total number of workers */
197 int nr_idle; /* L: currently idle workers */
199 struct list_head idle_list; /* L: list of idle workers */
200 struct timer_list idle_timer; /* L: worker idle timeout */
201 struct work_struct idle_cull_work; /* L: worker idle cleanup */
203 struct timer_list mayday_timer; /* L: SOS timer for workers */
205 /* a workers is either on busy_hash or idle_list, or the manager */
206 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
207 /* L: hash of busy workers */
209 struct worker *manager; /* L: purely informational */
210 struct list_head workers; /* A: attached workers */
211 struct list_head dying_workers; /* A: workers about to die */
212 struct completion *detach_completion; /* all workers detached */
214 struct ida worker_ida; /* worker IDs for task name */
216 struct workqueue_attrs *attrs; /* I: worker attributes */
217 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
218 int refcnt; /* PL: refcnt for unbound pools */
221 * Destruction of pool is RCU protected to allow dereferences
222 * from get_work_pool().
228 * Per-pool_workqueue statistics. These can be monitored using
229 * tools/workqueue/wq_monitor.py.
231 enum pool_workqueue_stats {
232 PWQ_STAT_STARTED, /* work items started execution */
233 PWQ_STAT_COMPLETED, /* work items completed execution */
234 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
235 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
236 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
237 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
238 PWQ_STAT_MAYDAY, /* maydays to rescuer */
239 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
245 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
246 * of work_struct->data are used for flags and the remaining high bits
247 * point to the pwq; thus, pwqs need to be aligned at two's power of the
248 * number of flag bits.
250 struct pool_workqueue {
251 struct worker_pool *pool; /* I: the associated pool */
252 struct workqueue_struct *wq; /* I: the owning workqueue */
253 int work_color; /* L: current color */
254 int flush_color; /* L: flushing color */
255 int refcnt; /* L: reference count */
256 int nr_in_flight[WORK_NR_COLORS];
257 /* L: nr of in_flight works */
260 * nr_active management and WORK_STRUCT_INACTIVE:
262 * When pwq->nr_active >= max_active, new work item is queued to
263 * pwq->inactive_works instead of pool->worklist and marked with
264 * WORK_STRUCT_INACTIVE.
266 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
267 * nr_active and all work items in pwq->inactive_works are marked with
268 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
269 * in pwq->inactive_works. Some of them are ready to run in
270 * pool->worklist or worker->scheduled. Those work itmes are only struct
271 * wq_barrier which is used for flush_work() and should not participate
272 * in nr_active. For non-barrier work item, it is marked with
273 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
275 int nr_active; /* L: nr of active works */
276 struct list_head inactive_works; /* L: inactive works */
277 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
278 struct list_head pwqs_node; /* WR: node on wq->pwqs */
279 struct list_head mayday_node; /* MD: node on wq->maydays */
281 u64 stats[PWQ_NR_STATS];
284 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
285 * and pwq_release_workfn() for details. pool_workqueue itself is also
286 * RCU protected so that the first pwq can be determined without
287 * grabbing wq->mutex.
289 struct kthread_work release_work;
291 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
294 * Structure used to wait for workqueue flush.
297 struct list_head list; /* WQ: list of flushers */
298 int flush_color; /* WQ: flush color waiting for */
299 struct completion done; /* flush completion */
305 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
306 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
307 * As sharing a single nr_active across multiple sockets can be very expensive,
308 * the counting and enforcement is per NUMA node.
310 * The following struct is used to enforce per-node max_active. When a pwq wants
311 * to start executing a work item, it should increment ->nr using
312 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
313 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
314 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
317 struct wq_node_nr_active {
318 int max; /* per-node max_active */
319 atomic_t nr; /* per-node nr_active */
320 raw_spinlock_t lock; /* nests inside pool locks */
321 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
325 * The externally visible workqueue. It relays the issued work items to
326 * the appropriate worker_pool through its pool_workqueues.
328 struct workqueue_struct {
329 struct list_head pwqs; /* WR: all pwqs of this wq */
330 struct list_head list; /* PR: list of all workqueues */
332 struct mutex mutex; /* protects this wq */
333 int work_color; /* WQ: current work color */
334 int flush_color; /* WQ: current flush color */
335 atomic_t nr_pwqs_to_flush; /* flush in progress */
336 struct wq_flusher *first_flusher; /* WQ: first flusher */
337 struct list_head flusher_queue; /* WQ: flush waiters */
338 struct list_head flusher_overflow; /* WQ: flush overflow list */
340 struct list_head maydays; /* MD: pwqs requesting rescue */
341 struct worker *rescuer; /* MD: rescue worker */
343 int nr_drainers; /* WQ: drain in progress */
345 /* See alloc_workqueue() function comment for info on min/max_active */
346 int max_active; /* WO: max active works */
347 int min_active; /* WO: min active works */
348 int saved_max_active; /* WQ: saved max_active */
349 int saved_min_active; /* WQ: saved min_active */
351 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
352 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
355 struct wq_device *wq_dev; /* I: for sysfs interface */
357 #ifdef CONFIG_LOCKDEP
359 struct lock_class_key key;
360 struct lockdep_map lockdep_map;
362 char name[WQ_NAME_LEN]; /* I: workqueue name */
365 * Destruction of workqueue_struct is RCU protected to allow walking
366 * the workqueues list without grabbing wq_pool_mutex.
367 * This is used to dump all workqueues from sysrq.
371 /* hot fields used during command issue, aligned to cacheline */
372 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
373 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
374 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
377 static struct kmem_cache *pwq_cache;
380 * Each pod type describes how CPUs should be grouped for unbound workqueues.
381 * See the comment above workqueue_attrs->affn_scope.
384 int nr_pods; /* number of pods */
385 cpumask_var_t *pod_cpus; /* pod -> cpus */
386 int *pod_node; /* pod -> node */
387 int *cpu_pod; /* cpu -> pod */
390 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
391 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
393 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
394 [WQ_AFFN_DFL] = "default",
395 [WQ_AFFN_CPU] = "cpu",
396 [WQ_AFFN_SMT] = "smt",
397 [WQ_AFFN_CACHE] = "cache",
398 [WQ_AFFN_NUMA] = "numa",
399 [WQ_AFFN_SYSTEM] = "system",
402 static bool wq_topo_initialized __read_mostly = false;
405 * Per-cpu work items which run for longer than the following threshold are
406 * automatically considered CPU intensive and excluded from concurrency
407 * management to prevent them from noticeably delaying other per-cpu work items.
408 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
409 * The actual value is initialized in wq_cpu_intensive_thresh_init().
411 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
412 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
414 /* see the comment above the definition of WQ_POWER_EFFICIENT */
415 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
416 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
418 static bool wq_online; /* can kworkers be created yet? */
420 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
421 static struct workqueue_attrs *wq_update_pod_attrs_buf;
423 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
424 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
425 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
426 /* wait for manager to go away */
427 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
429 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
430 static bool workqueue_freezing; /* PL: have wqs started freezing? */
432 /* PL&A: allowable cpus for unbound wqs and work items */
433 static cpumask_var_t wq_unbound_cpumask;
435 /* PL: user requested unbound cpumask via sysfs */
436 static cpumask_var_t wq_requested_unbound_cpumask;
438 /* PL: isolated cpumask to be excluded from unbound cpumask */
439 static cpumask_var_t wq_isolated_cpumask;
441 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
442 static struct cpumask wq_cmdline_cpumask __initdata;
444 /* CPU where unbound work was last round robin scheduled from this CPU */
445 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
448 * Local execution of unbound work items is no longer guaranteed. The
449 * following always forces round-robin CPU selection on unbound work items
450 * to uncover usages which depend on it.
452 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
453 static bool wq_debug_force_rr_cpu = true;
455 static bool wq_debug_force_rr_cpu = false;
457 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
459 /* the BH worker pools */
460 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
463 /* the per-cpu worker pools */
464 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
467 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
469 /* PL: hash of all unbound pools keyed by pool->attrs */
470 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
472 /* I: attributes used when instantiating standard unbound pools on demand */
473 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
475 /* I: attributes used when instantiating ordered pools on demand */
476 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
479 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
480 * process context while holding a pool lock. Bounce to a dedicated kthread
481 * worker to avoid A-A deadlocks.
483 static struct kthread_worker *pwq_release_worker __ro_after_init;
485 struct workqueue_struct *system_wq __ro_after_init;
486 EXPORT_SYMBOL(system_wq);
487 struct workqueue_struct *system_highpri_wq __ro_after_init;
488 EXPORT_SYMBOL_GPL(system_highpri_wq);
489 struct workqueue_struct *system_long_wq __ro_after_init;
490 EXPORT_SYMBOL_GPL(system_long_wq);
491 struct workqueue_struct *system_unbound_wq __ro_after_init;
492 EXPORT_SYMBOL_GPL(system_unbound_wq);
493 struct workqueue_struct *system_freezable_wq __ro_after_init;
494 EXPORT_SYMBOL_GPL(system_freezable_wq);
495 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
496 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
497 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
498 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
499 struct workqueue_struct *system_bh_wq;
500 EXPORT_SYMBOL_GPL(system_bh_wq);
501 struct workqueue_struct *system_bh_highpri_wq;
502 EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
504 static int worker_thread(void *__worker);
505 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
506 static void show_pwq(struct pool_workqueue *pwq);
507 static void show_one_worker_pool(struct worker_pool *pool);
509 #define CREATE_TRACE_POINTS
510 #include <trace/events/workqueue.h>
512 #define assert_rcu_or_pool_mutex() \
513 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
514 !lockdep_is_held(&wq_pool_mutex), \
515 "RCU or wq_pool_mutex should be held")
517 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
518 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
519 !lockdep_is_held(&wq->mutex) && \
520 !lockdep_is_held(&wq_pool_mutex), \
521 "RCU, wq->mutex or wq_pool_mutex should be held")
523 #define for_each_bh_worker_pool(pool, cpu) \
524 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
525 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
528 #define for_each_cpu_worker_pool(pool, cpu) \
529 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
530 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
534 * for_each_pool - iterate through all worker_pools in the system
535 * @pool: iteration cursor
536 * @pi: integer used for iteration
538 * This must be called either with wq_pool_mutex held or RCU read
539 * locked. If the pool needs to be used beyond the locking in effect, the
540 * caller is responsible for guaranteeing that the pool stays online.
542 * The if/else clause exists only for the lockdep assertion and can be
545 #define for_each_pool(pool, pi) \
546 idr_for_each_entry(&worker_pool_idr, pool, pi) \
547 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
551 * for_each_pool_worker - iterate through all workers of a worker_pool
552 * @worker: iteration cursor
553 * @pool: worker_pool to iterate workers of
555 * This must be called with wq_pool_attach_mutex.
557 * The if/else clause exists only for the lockdep assertion and can be
560 #define for_each_pool_worker(worker, pool) \
561 list_for_each_entry((worker), &(pool)->workers, node) \
562 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
566 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
567 * @pwq: iteration cursor
568 * @wq: the target workqueue
570 * This must be called either with wq->mutex held or RCU read locked.
571 * If the pwq needs to be used beyond the locking in effect, the caller is
572 * responsible for guaranteeing that the pwq stays online.
574 * The if/else clause exists only for the lockdep assertion and can be
577 #define for_each_pwq(pwq, wq) \
578 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
579 lockdep_is_held(&(wq->mutex)))
581 #ifdef CONFIG_DEBUG_OBJECTS_WORK
583 static const struct debug_obj_descr work_debug_descr;
585 static void *work_debug_hint(void *addr)
587 return ((struct work_struct *) addr)->func;
590 static bool work_is_static_object(void *addr)
592 struct work_struct *work = addr;
594 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
598 * fixup_init is called when:
599 * - an active object is initialized
601 static bool work_fixup_init(void *addr, enum debug_obj_state state)
603 struct work_struct *work = addr;
606 case ODEBUG_STATE_ACTIVE:
607 cancel_work_sync(work);
608 debug_object_init(work, &work_debug_descr);
616 * fixup_free is called when:
617 * - an active object is freed
619 static bool work_fixup_free(void *addr, enum debug_obj_state state)
621 struct work_struct *work = addr;
624 case ODEBUG_STATE_ACTIVE:
625 cancel_work_sync(work);
626 debug_object_free(work, &work_debug_descr);
633 static const struct debug_obj_descr work_debug_descr = {
634 .name = "work_struct",
635 .debug_hint = work_debug_hint,
636 .is_static_object = work_is_static_object,
637 .fixup_init = work_fixup_init,
638 .fixup_free = work_fixup_free,
641 static inline void debug_work_activate(struct work_struct *work)
643 debug_object_activate(work, &work_debug_descr);
646 static inline void debug_work_deactivate(struct work_struct *work)
648 debug_object_deactivate(work, &work_debug_descr);
651 void __init_work(struct work_struct *work, int onstack)
654 debug_object_init_on_stack(work, &work_debug_descr);
656 debug_object_init(work, &work_debug_descr);
658 EXPORT_SYMBOL_GPL(__init_work);
660 void destroy_work_on_stack(struct work_struct *work)
662 debug_object_free(work, &work_debug_descr);
664 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
666 void destroy_delayed_work_on_stack(struct delayed_work *work)
668 destroy_timer_on_stack(&work->timer);
669 debug_object_free(&work->work, &work_debug_descr);
671 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
674 static inline void debug_work_activate(struct work_struct *work) { }
675 static inline void debug_work_deactivate(struct work_struct *work) { }
679 * worker_pool_assign_id - allocate ID and assign it to @pool
680 * @pool: the pool pointer of interest
682 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
683 * successfully, -errno on failure.
685 static int worker_pool_assign_id(struct worker_pool *pool)
689 lockdep_assert_held(&wq_pool_mutex);
691 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
700 static struct pool_workqueue __rcu **
701 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
704 return per_cpu_ptr(wq->cpu_pwq, cpu);
709 /* @cpu < 0 for dfl_pwq */
710 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
712 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
713 lockdep_is_held(&wq_pool_mutex) ||
714 lockdep_is_held(&wq->mutex));
718 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
719 * @wq: workqueue of interest
721 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
722 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
723 * default pwq is always mapped to the pool with the current effective cpumask.
725 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
727 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
730 static unsigned int work_color_to_flags(int color)
732 return color << WORK_STRUCT_COLOR_SHIFT;
735 static int get_work_color(unsigned long work_data)
737 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
738 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
741 static int work_next_color(int color)
743 return (color + 1) % WORK_NR_COLORS;
747 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
748 * contain the pointer to the queued pwq. Once execution starts, the flag
749 * is cleared and the high bits contain OFFQ flags and pool ID.
751 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
752 * and clear_work_data() can be used to set the pwq, pool or clear
753 * work->data. These functions should only be called while the work is
754 * owned - ie. while the PENDING bit is set.
756 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
757 * corresponding to a work. Pool is available once the work has been
758 * queued anywhere after initialization until it is sync canceled. pwq is
759 * available only while the work item is queued.
761 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
762 * canceled. While being canceled, a work item may have its PENDING set
763 * but stay off timer and worklist for arbitrarily long and nobody should
764 * try to steal the PENDING bit.
766 static inline void set_work_data(struct work_struct *work, unsigned long data,
769 WARN_ON_ONCE(!work_pending(work));
770 atomic_long_set(&work->data, data | flags | work_static(work));
773 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
774 unsigned long extra_flags)
776 set_work_data(work, (unsigned long)pwq,
777 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
780 static void set_work_pool_and_keep_pending(struct work_struct *work,
783 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
784 WORK_STRUCT_PENDING);
787 static void set_work_pool_and_clear_pending(struct work_struct *work,
791 * The following wmb is paired with the implied mb in
792 * test_and_set_bit(PENDING) and ensures all updates to @work made
793 * here are visible to and precede any updates by the next PENDING
797 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
799 * The following mb guarantees that previous clear of a PENDING bit
800 * will not be reordered with any speculative LOADS or STORES from
801 * work->current_func, which is executed afterwards. This possible
802 * reordering can lead to a missed execution on attempt to queue
803 * the same @work. E.g. consider this case:
806 * ---------------------------- --------------------------------
808 * 1 STORE event_indicated
809 * 2 queue_work_on() {
810 * 3 test_and_set_bit(PENDING)
811 * 4 } set_..._and_clear_pending() {
812 * 5 set_work_data() # clear bit
814 * 7 work->current_func() {
815 * 8 LOAD event_indicated
818 * Without an explicit full barrier speculative LOAD on line 8 can
819 * be executed before CPU#0 does STORE on line 1. If that happens,
820 * CPU#0 observes the PENDING bit is still set and new execution of
821 * a @work is not queued in a hope, that CPU#1 will eventually
822 * finish the queued @work. Meanwhile CPU#1 does not see
823 * event_indicated is set, because speculative LOAD was executed
824 * before actual STORE.
829 static void clear_work_data(struct work_struct *work)
831 smp_wmb(); /* see set_work_pool_and_clear_pending() */
832 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
835 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
837 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
840 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
842 unsigned long data = atomic_long_read(&work->data);
844 if (data & WORK_STRUCT_PWQ)
845 return work_struct_pwq(data);
851 * get_work_pool - return the worker_pool a given work was associated with
852 * @work: the work item of interest
854 * Pools are created and destroyed under wq_pool_mutex, and allows read
855 * access under RCU read lock. As such, this function should be
856 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
858 * All fields of the returned pool are accessible as long as the above
859 * mentioned locking is in effect. If the returned pool needs to be used
860 * beyond the critical section, the caller is responsible for ensuring the
861 * returned pool is and stays online.
863 * Return: The worker_pool @work was last associated with. %NULL if none.
865 static struct worker_pool *get_work_pool(struct work_struct *work)
867 unsigned long data = atomic_long_read(&work->data);
870 assert_rcu_or_pool_mutex();
872 if (data & WORK_STRUCT_PWQ)
873 return work_struct_pwq(data)->pool;
875 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
876 if (pool_id == WORK_OFFQ_POOL_NONE)
879 return idr_find(&worker_pool_idr, pool_id);
883 * get_work_pool_id - return the worker pool ID a given work is associated with
884 * @work: the work item of interest
886 * Return: The worker_pool ID @work was last associated with.
887 * %WORK_OFFQ_POOL_NONE if none.
889 static int get_work_pool_id(struct work_struct *work)
891 unsigned long data = atomic_long_read(&work->data);
893 if (data & WORK_STRUCT_PWQ)
894 return work_struct_pwq(data)->pool->id;
896 return data >> WORK_OFFQ_POOL_SHIFT;
899 static void mark_work_canceling(struct work_struct *work)
901 unsigned long pool_id = get_work_pool_id(work);
903 pool_id <<= WORK_OFFQ_POOL_SHIFT;
904 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
907 static bool work_is_canceling(struct work_struct *work)
909 unsigned long data = atomic_long_read(&work->data);
911 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
915 * Policy functions. These define the policies on how the global worker
916 * pools are managed. Unless noted otherwise, these functions assume that
917 * they're being called with pool->lock held.
921 * Need to wake up a worker? Called from anything but currently
924 * Note that, because unbound workers never contribute to nr_running, this
925 * function will always return %true for unbound pools as long as the
926 * worklist isn't empty.
928 static bool need_more_worker(struct worker_pool *pool)
930 return !list_empty(&pool->worklist) && !pool->nr_running;
933 /* Can I start working? Called from busy but !running workers. */
934 static bool may_start_working(struct worker_pool *pool)
936 return pool->nr_idle;
939 /* Do I need to keep working? Called from currently running workers. */
940 static bool keep_working(struct worker_pool *pool)
942 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
945 /* Do we need a new worker? Called from manager. */
946 static bool need_to_create_worker(struct worker_pool *pool)
948 return need_more_worker(pool) && !may_start_working(pool);
951 /* Do we have too many workers and should some go away? */
952 static bool too_many_workers(struct worker_pool *pool)
954 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
955 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
956 int nr_busy = pool->nr_workers - nr_idle;
958 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
962 * worker_set_flags - set worker flags and adjust nr_running accordingly
964 * @flags: flags to set
966 * Set @flags in @worker->flags and adjust nr_running accordingly.
968 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
970 struct worker_pool *pool = worker->pool;
972 lockdep_assert_held(&pool->lock);
974 /* If transitioning into NOT_RUNNING, adjust nr_running. */
975 if ((flags & WORKER_NOT_RUNNING) &&
976 !(worker->flags & WORKER_NOT_RUNNING)) {
980 worker->flags |= flags;
984 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
986 * @flags: flags to clear
988 * Clear @flags in @worker->flags and adjust nr_running accordingly.
990 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
992 struct worker_pool *pool = worker->pool;
993 unsigned int oflags = worker->flags;
995 lockdep_assert_held(&pool->lock);
997 worker->flags &= ~flags;
1000 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1001 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1002 * of multiple flags, not a single flag.
1004 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1005 if (!(worker->flags & WORKER_NOT_RUNNING))
1009 /* Return the first idle worker. Called with pool->lock held. */
1010 static struct worker *first_idle_worker(struct worker_pool *pool)
1012 if (unlikely(list_empty(&pool->idle_list)))
1015 return list_first_entry(&pool->idle_list, struct worker, entry);
1019 * worker_enter_idle - enter idle state
1020 * @worker: worker which is entering idle state
1022 * @worker is entering idle state. Update stats and idle timer if
1026 * raw_spin_lock_irq(pool->lock).
1028 static void worker_enter_idle(struct worker *worker)
1030 struct worker_pool *pool = worker->pool;
1032 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1033 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1034 (worker->hentry.next || worker->hentry.pprev)))
1037 /* can't use worker_set_flags(), also called from create_worker() */
1038 worker->flags |= WORKER_IDLE;
1040 worker->last_active = jiffies;
1042 /* idle_list is LIFO */
1043 list_add(&worker->entry, &pool->idle_list);
1045 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1046 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1048 /* Sanity check nr_running. */
1049 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1053 * worker_leave_idle - leave idle state
1054 * @worker: worker which is leaving idle state
1056 * @worker is leaving idle state. Update stats.
1059 * raw_spin_lock_irq(pool->lock).
1061 static void worker_leave_idle(struct worker *worker)
1063 struct worker_pool *pool = worker->pool;
1065 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1067 worker_clr_flags(worker, WORKER_IDLE);
1069 list_del_init(&worker->entry);
1073 * find_worker_executing_work - find worker which is executing a work
1074 * @pool: pool of interest
1075 * @work: work to find worker for
1077 * Find a worker which is executing @work on @pool by searching
1078 * @pool->busy_hash which is keyed by the address of @work. For a worker
1079 * to match, its current execution should match the address of @work and
1080 * its work function. This is to avoid unwanted dependency between
1081 * unrelated work executions through a work item being recycled while still
1084 * This is a bit tricky. A work item may be freed once its execution
1085 * starts and nothing prevents the freed area from being recycled for
1086 * another work item. If the same work item address ends up being reused
1087 * before the original execution finishes, workqueue will identify the
1088 * recycled work item as currently executing and make it wait until the
1089 * current execution finishes, introducing an unwanted dependency.
1091 * This function checks the work item address and work function to avoid
1092 * false positives. Note that this isn't complete as one may construct a
1093 * work function which can introduce dependency onto itself through a
1094 * recycled work item. Well, if somebody wants to shoot oneself in the
1095 * foot that badly, there's only so much we can do, and if such deadlock
1096 * actually occurs, it should be easy to locate the culprit work function.
1099 * raw_spin_lock_irq(pool->lock).
1102 * Pointer to worker which is executing @work if found, %NULL
1105 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1106 struct work_struct *work)
1108 struct worker *worker;
1110 hash_for_each_possible(pool->busy_hash, worker, hentry,
1111 (unsigned long)work)
1112 if (worker->current_work == work &&
1113 worker->current_func == work->func)
1120 * move_linked_works - move linked works to a list
1121 * @work: start of series of works to be scheduled
1122 * @head: target list to append @work to
1123 * @nextp: out parameter for nested worklist walking
1125 * Schedule linked works starting from @work to @head. Work series to be
1126 * scheduled starts at @work and includes any consecutive work with
1127 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1131 * raw_spin_lock_irq(pool->lock).
1133 static void move_linked_works(struct work_struct *work, struct list_head *head,
1134 struct work_struct **nextp)
1136 struct work_struct *n;
1139 * Linked worklist will always end before the end of the list,
1140 * use NULL for list head.
1142 list_for_each_entry_safe_from(work, n, NULL, entry) {
1143 list_move_tail(&work->entry, head);
1144 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1149 * If we're already inside safe list traversal and have moved
1150 * multiple works to the scheduled queue, the next position
1151 * needs to be updated.
1158 * assign_work - assign a work item and its linked work items to a worker
1159 * @work: work to assign
1160 * @worker: worker to assign to
1161 * @nextp: out parameter for nested worklist walking
1163 * Assign @work and its linked work items to @worker. If @work is already being
1164 * executed by another worker in the same pool, it'll be punted there.
1166 * If @nextp is not NULL, it's updated to point to the next work of the last
1167 * scheduled work. This allows assign_work() to be nested inside
1168 * list_for_each_entry_safe().
1170 * Returns %true if @work was successfully assigned to @worker. %false if @work
1171 * was punted to another worker already executing it.
1173 static bool assign_work(struct work_struct *work, struct worker *worker,
1174 struct work_struct **nextp)
1176 struct worker_pool *pool = worker->pool;
1177 struct worker *collision;
1179 lockdep_assert_held(&pool->lock);
1182 * A single work shouldn't be executed concurrently by multiple workers.
1183 * __queue_work() ensures that @work doesn't jump to a different pool
1184 * while still running in the previous pool. Here, we should ensure that
1185 * @work is not executed concurrently by multiple workers from the same
1186 * pool. Check whether anyone is already processing the work. If so,
1187 * defer the work to the currently executing one.
1189 collision = find_worker_executing_work(pool, work);
1190 if (unlikely(collision)) {
1191 move_linked_works(work, &collision->scheduled, nextp);
1195 move_linked_works(work, &worker->scheduled, nextp);
1200 * kick_pool - wake up an idle worker if necessary
1201 * @pool: pool to kick
1203 * @pool may have pending work items. Wake up worker if necessary. Returns
1204 * whether a worker was woken up.
1206 static bool kick_pool(struct worker_pool *pool)
1208 struct worker *worker = first_idle_worker(pool);
1209 struct task_struct *p;
1211 lockdep_assert_held(&pool->lock);
1213 if (!need_more_worker(pool) || !worker)
1216 if (pool->flags & POOL_BH) {
1217 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1218 raise_softirq_irqoff(HI_SOFTIRQ);
1220 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1228 * Idle @worker is about to execute @work and waking up provides an
1229 * opportunity to migrate @worker at a lower cost by setting the task's
1230 * wake_cpu field. Let's see if we want to move @worker to improve
1231 * execution locality.
1233 * We're waking the worker that went idle the latest and there's some
1234 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1235 * so, setting the wake_cpu won't do anything. As this is a best-effort
1236 * optimization and the race window is narrow, let's leave as-is for
1237 * now. If this becomes pronounced, we can skip over workers which are
1238 * still on cpu when picking an idle worker.
1240 * If @pool has non-strict affinity, @worker might have ended up outside
1241 * its affinity scope. Repatriate.
1243 if (!pool->attrs->affn_strict &&
1244 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1245 struct work_struct *work = list_first_entry(&pool->worklist,
1246 struct work_struct, entry);
1247 p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1248 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1255 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1258 * Concurrency-managed per-cpu work items that hog CPU for longer than
1259 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1260 * which prevents them from stalling other concurrency-managed work items. If a
1261 * work function keeps triggering this mechanism, it's likely that the work item
1262 * should be using an unbound workqueue instead.
1264 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1265 * and report them so that they can be examined and converted to use unbound
1266 * workqueues as appropriate. To avoid flooding the console, each violating work
1267 * function is tracked and reported with exponential backoff.
1269 #define WCI_MAX_ENTS 128
1274 struct hlist_node hash_node;
1277 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1278 static int wci_nr_ents;
1279 static DEFINE_RAW_SPINLOCK(wci_lock);
1280 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1282 static struct wci_ent *wci_find_ent(work_func_t func)
1284 struct wci_ent *ent;
1286 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1287 (unsigned long)func) {
1288 if (ent->func == func)
1294 static void wq_cpu_intensive_report(work_func_t func)
1296 struct wci_ent *ent;
1299 ent = wci_find_ent(func);
1304 * Start reporting from the fourth time and back off
1307 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1308 if (cnt >= 4 && is_power_of_2(cnt))
1309 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1310 ent->func, wq_cpu_intensive_thresh_us,
1311 atomic64_read(&ent->cnt));
1316 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1317 * is exhausted, something went really wrong and we probably made enough
1320 if (wci_nr_ents >= WCI_MAX_ENTS)
1323 raw_spin_lock(&wci_lock);
1325 if (wci_nr_ents >= WCI_MAX_ENTS) {
1326 raw_spin_unlock(&wci_lock);
1330 if (wci_find_ent(func)) {
1331 raw_spin_unlock(&wci_lock);
1335 ent = &wci_ents[wci_nr_ents++];
1337 atomic64_set(&ent->cnt, 1);
1338 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1340 raw_spin_unlock(&wci_lock);
1343 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1344 static void wq_cpu_intensive_report(work_func_t func) {}
1345 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1348 * wq_worker_running - a worker is running again
1349 * @task: task waking up
1351 * This function is called when a worker returns from schedule()
1353 void wq_worker_running(struct task_struct *task)
1355 struct worker *worker = kthread_data(task);
1357 if (!READ_ONCE(worker->sleeping))
1361 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1362 * and the nr_running increment below, we may ruin the nr_running reset
1363 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1364 * pool. Protect against such race.
1367 if (!(worker->flags & WORKER_NOT_RUNNING))
1368 worker->pool->nr_running++;
1372 * CPU intensive auto-detection cares about how long a work item hogged
1373 * CPU without sleeping. Reset the starting timestamp on wakeup.
1375 worker->current_at = worker->task->se.sum_exec_runtime;
1377 WRITE_ONCE(worker->sleeping, 0);
1381 * wq_worker_sleeping - a worker is going to sleep
1382 * @task: task going to sleep
1384 * This function is called from schedule() when a busy worker is
1387 void wq_worker_sleeping(struct task_struct *task)
1389 struct worker *worker = kthread_data(task);
1390 struct worker_pool *pool;
1393 * Rescuers, which may not have all the fields set up like normal
1394 * workers, also reach here, let's not access anything before
1395 * checking NOT_RUNNING.
1397 if (worker->flags & WORKER_NOT_RUNNING)
1400 pool = worker->pool;
1402 /* Return if preempted before wq_worker_running() was reached */
1403 if (READ_ONCE(worker->sleeping))
1406 WRITE_ONCE(worker->sleeping, 1);
1407 raw_spin_lock_irq(&pool->lock);
1410 * Recheck in case unbind_workers() preempted us. We don't
1411 * want to decrement nr_running after the worker is unbound
1412 * and nr_running has been reset.
1414 if (worker->flags & WORKER_NOT_RUNNING) {
1415 raw_spin_unlock_irq(&pool->lock);
1420 if (kick_pool(pool))
1421 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1423 raw_spin_unlock_irq(&pool->lock);
1427 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1428 * @task: task currently running
1430 * Called from scheduler_tick(). We're in the IRQ context and the current
1431 * worker's fields which follow the 'K' locking rule can be accessed safely.
1433 void wq_worker_tick(struct task_struct *task)
1435 struct worker *worker = kthread_data(task);
1436 struct pool_workqueue *pwq = worker->current_pwq;
1437 struct worker_pool *pool = worker->pool;
1442 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1444 if (!wq_cpu_intensive_thresh_us)
1448 * If the current worker is concurrency managed and hogged the CPU for
1449 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1450 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1452 * Set @worker->sleeping means that @worker is in the process of
1453 * switching out voluntarily and won't be contributing to
1454 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1455 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1456 * double decrements. The task is releasing the CPU anyway. Let's skip.
1457 * We probably want to make this prettier in the future.
1459 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1460 worker->task->se.sum_exec_runtime - worker->current_at <
1461 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1464 raw_spin_lock(&pool->lock);
1466 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1467 wq_cpu_intensive_report(worker->current_func);
1468 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1470 if (kick_pool(pool))
1471 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1473 raw_spin_unlock(&pool->lock);
1477 * wq_worker_last_func - retrieve worker's last work function
1478 * @task: Task to retrieve last work function of.
1480 * Determine the last function a worker executed. This is called from
1481 * the scheduler to get a worker's last known identity.
1484 * raw_spin_lock_irq(rq->lock)
1486 * This function is called during schedule() when a kworker is going
1487 * to sleep. It's used by psi to identify aggregation workers during
1488 * dequeuing, to allow periodic aggregation to shut-off when that
1489 * worker is the last task in the system or cgroup to go to sleep.
1491 * As this function doesn't involve any workqueue-related locking, it
1492 * only returns stable values when called from inside the scheduler's
1493 * queuing and dequeuing paths, when @task, which must be a kworker,
1494 * is guaranteed to not be processing any works.
1497 * The last work function %current executed as a worker, NULL if it
1498 * hasn't executed any work yet.
1500 work_func_t wq_worker_last_func(struct task_struct *task)
1502 struct worker *worker = kthread_data(task);
1504 return worker->last_func;
1508 * wq_node_nr_active - Determine wq_node_nr_active to use
1509 * @wq: workqueue of interest
1510 * @node: NUMA node, can be %NUMA_NO_NODE
1512 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1514 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1516 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1518 * - Otherwise, node_nr_active[@node].
1520 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1523 if (!(wq->flags & WQ_UNBOUND))
1526 if (node == NUMA_NO_NODE)
1529 return wq->node_nr_active[node];
1533 * wq_update_node_max_active - Update per-node max_actives to use
1534 * @wq: workqueue to update
1535 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1537 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1538 * distributed among nodes according to the proportions of numbers of online
1539 * cpus. The result is always between @wq->min_active and max_active.
1541 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1543 struct cpumask *effective = unbound_effective_cpumask(wq);
1544 int min_active = READ_ONCE(wq->min_active);
1545 int max_active = READ_ONCE(wq->max_active);
1546 int total_cpus, node;
1548 lockdep_assert_held(&wq->mutex);
1550 if (!wq_topo_initialized)
1553 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1556 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1560 for_each_node(node) {
1563 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1564 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1567 wq_node_nr_active(wq, node)->max =
1568 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1569 min_active, max_active);
1572 wq_node_nr_active(wq, NUMA_NO_NODE)->max = min_active;
1576 * get_pwq - get an extra reference on the specified pool_workqueue
1577 * @pwq: pool_workqueue to get
1579 * Obtain an extra reference on @pwq. The caller should guarantee that
1580 * @pwq has positive refcnt and be holding the matching pool->lock.
1582 static void get_pwq(struct pool_workqueue *pwq)
1584 lockdep_assert_held(&pwq->pool->lock);
1585 WARN_ON_ONCE(pwq->refcnt <= 0);
1590 * put_pwq - put a pool_workqueue reference
1591 * @pwq: pool_workqueue to put
1593 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1594 * destruction. The caller should be holding the matching pool->lock.
1596 static void put_pwq(struct pool_workqueue *pwq)
1598 lockdep_assert_held(&pwq->pool->lock);
1599 if (likely(--pwq->refcnt))
1602 * @pwq can't be released under pool->lock, bounce to a dedicated
1603 * kthread_worker to avoid A-A deadlocks.
1605 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1609 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1610 * @pwq: pool_workqueue to put (can be %NULL)
1612 * put_pwq() with locking. This function also allows %NULL @pwq.
1614 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1618 * As both pwqs and pools are RCU protected, the
1619 * following lock operations are safe.
1621 raw_spin_lock_irq(&pwq->pool->lock);
1623 raw_spin_unlock_irq(&pwq->pool->lock);
1627 static bool pwq_is_empty(struct pool_workqueue *pwq)
1629 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1632 static void __pwq_activate_work(struct pool_workqueue *pwq,
1633 struct work_struct *work)
1635 unsigned long *wdb = work_data_bits(work);
1637 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1638 trace_workqueue_activate_work(work);
1639 if (list_empty(&pwq->pool->worklist))
1640 pwq->pool->watchdog_ts = jiffies;
1641 move_linked_works(work, &pwq->pool->worklist, NULL);
1642 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1646 * pwq_activate_work - Activate a work item if inactive
1647 * @pwq: pool_workqueue @work belongs to
1648 * @work: work item to activate
1650 * Returns %true if activated. %false if already active.
1652 static bool pwq_activate_work(struct pool_workqueue *pwq,
1653 struct work_struct *work)
1655 struct worker_pool *pool = pwq->pool;
1656 struct wq_node_nr_active *nna;
1658 lockdep_assert_held(&pool->lock);
1660 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE))
1663 nna = wq_node_nr_active(pwq->wq, pool->node);
1665 atomic_inc(&nna->nr);
1668 __pwq_activate_work(pwq, work);
1672 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1674 int max = READ_ONCE(nna->max);
1679 old = atomic_read(&nna->nr);
1682 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1689 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1690 * @pwq: pool_workqueue of interest
1691 * @fill: max_active may have increased, try to increase concurrency level
1693 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1694 * successfully obtained. %false otherwise.
1696 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1698 struct workqueue_struct *wq = pwq->wq;
1699 struct worker_pool *pool = pwq->pool;
1700 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1701 bool obtained = false;
1703 lockdep_assert_held(&pool->lock);
1706 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1707 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1712 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1713 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1714 * concurrency level. Don't jump the line.
1716 * We need to ignore the pending test after max_active has increased as
1717 * pwq_dec_nr_active() can only maintain the concurrency level but not
1718 * increase it. This is indicated by @fill.
1720 if (!list_empty(&pwq->pending_node) && likely(!fill))
1723 obtained = tryinc_node_nr_active(nna);
1728 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1729 * and try again. The smp_mb() is paired with the implied memory barrier
1730 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1731 * we see the decremented $nna->nr or they see non-empty
1732 * $nna->pending_pwqs.
1734 raw_spin_lock(&nna->lock);
1736 if (list_empty(&pwq->pending_node))
1737 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1738 else if (likely(!fill))
1743 obtained = tryinc_node_nr_active(nna);
1746 * If @fill, @pwq might have already been pending. Being spuriously
1747 * pending in cold paths doesn't affect anything. Let's leave it be.
1749 if (obtained && likely(!fill))
1750 list_del_init(&pwq->pending_node);
1753 raw_spin_unlock(&nna->lock);
1761 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1762 * @pwq: pool_workqueue of interest
1763 * @fill: max_active may have increased, try to increase concurrency level
1765 * Activate the first inactive work item of @pwq if available and allowed by
1768 * Returns %true if an inactive work item has been activated. %false if no
1769 * inactive work item is found or max_active limit is reached.
1771 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1773 struct work_struct *work =
1774 list_first_entry_or_null(&pwq->inactive_works,
1775 struct work_struct, entry);
1777 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1778 __pwq_activate_work(pwq, work);
1786 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1787 * @nna: wq_node_nr_active to activate a pending pwq for
1788 * @caller_pool: worker_pool the caller is locking
1790 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1791 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1793 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1794 struct worker_pool *caller_pool)
1796 struct worker_pool *locked_pool = caller_pool;
1797 struct pool_workqueue *pwq;
1798 struct work_struct *work;
1800 lockdep_assert_held(&caller_pool->lock);
1802 raw_spin_lock(&nna->lock);
1804 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1805 struct pool_workqueue, pending_node);
1810 * If @pwq is for a different pool than @locked_pool, we need to lock
1811 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1812 * / lock dance. For that, we also need to release @nna->lock as it's
1813 * nested inside pool locks.
1815 if (pwq->pool != locked_pool) {
1816 raw_spin_unlock(&locked_pool->lock);
1817 locked_pool = pwq->pool;
1818 if (!raw_spin_trylock(&locked_pool->lock)) {
1819 raw_spin_unlock(&nna->lock);
1820 raw_spin_lock(&locked_pool->lock);
1821 raw_spin_lock(&nna->lock);
1827 * $pwq may not have any inactive work items due to e.g. cancellations.
1828 * Drop it from pending_pwqs and see if there's another one.
1830 work = list_first_entry_or_null(&pwq->inactive_works,
1831 struct work_struct, entry);
1833 list_del_init(&pwq->pending_node);
1838 * Acquire an nr_active count and activate the inactive work item. If
1839 * $pwq still has inactive work items, rotate it to the end of the
1840 * pending_pwqs so that we round-robin through them. This means that
1841 * inactive work items are not activated in queueing order which is fine
1842 * given that there has never been any ordering across different pwqs.
1844 if (likely(tryinc_node_nr_active(nna))) {
1846 __pwq_activate_work(pwq, work);
1848 if (list_empty(&pwq->inactive_works))
1849 list_del_init(&pwq->pending_node);
1851 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1853 /* if activating a foreign pool, make sure it's running */
1854 if (pwq->pool != caller_pool)
1855 kick_pool(pwq->pool);
1859 raw_spin_unlock(&nna->lock);
1860 if (locked_pool != caller_pool) {
1861 raw_spin_unlock(&locked_pool->lock);
1862 raw_spin_lock(&caller_pool->lock);
1867 * pwq_dec_nr_active - Retire an active count
1868 * @pwq: pool_workqueue of interest
1870 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1871 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1873 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1875 struct worker_pool *pool = pwq->pool;
1876 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1878 lockdep_assert_held(&pool->lock);
1881 * @pwq->nr_active should be decremented for both percpu and unbound
1887 * For a percpu workqueue, it's simple. Just need to kick the first
1888 * inactive work item on @pwq itself.
1891 pwq_activate_first_inactive(pwq, false);
1896 * If @pwq is for an unbound workqueue, it's more complicated because
1897 * multiple pwqs and pools may be sharing the nr_active count. When a
1898 * pwq needs to wait for an nr_active count, it puts itself on
1899 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1900 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1901 * guarantee that either we see non-empty pending_pwqs or they see
1902 * decremented $nna->nr.
1904 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1905 * max_active gets updated. However, it is guaranteed to be equal to or
1906 * larger than @pwq->wq->min_active which is above zero unless freezing.
1907 * This maintains the forward progress guarantee.
1909 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1912 if (!list_empty(&nna->pending_pwqs))
1913 node_activate_pending_pwq(nna, pool);
1917 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1918 * @pwq: pwq of interest
1919 * @work_data: work_data of work which left the queue
1921 * A work either has completed or is removed from pending queue,
1922 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1925 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1926 * and thus should be called after all other state updates for the in-flight
1927 * work item is complete.
1930 * raw_spin_lock_irq(pool->lock).
1932 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1934 int color = get_work_color(work_data);
1936 if (!(work_data & WORK_STRUCT_INACTIVE))
1937 pwq_dec_nr_active(pwq);
1939 pwq->nr_in_flight[color]--;
1941 /* is flush in progress and are we at the flushing tip? */
1942 if (likely(pwq->flush_color != color))
1945 /* are there still in-flight works? */
1946 if (pwq->nr_in_flight[color])
1949 /* this pwq is done, clear flush_color */
1950 pwq->flush_color = -1;
1953 * If this was the last pwq, wake up the first flusher. It
1954 * will handle the rest.
1956 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1957 complete(&pwq->wq->first_flusher->done);
1963 * try_to_grab_pending - steal work item from worklist and disable irq
1964 * @work: work item to steal
1965 * @is_dwork: @work is a delayed_work
1966 * @flags: place to store irq state
1968 * Try to grab PENDING bit of @work. This function can handle @work in any
1969 * stable state - idle, on timer or on worklist.
1973 * ======== ================================================================
1974 * 1 if @work was pending and we successfully stole PENDING
1975 * 0 if @work was idle and we claimed PENDING
1976 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1977 * -ENOENT if someone else is canceling @work, this state may persist
1978 * for arbitrarily long
1979 * ======== ================================================================
1982 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1983 * interrupted while holding PENDING and @work off queue, irq must be
1984 * disabled on entry. This, combined with delayed_work->timer being
1985 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1987 * On successful return, >= 0, irq is disabled and the caller is
1988 * responsible for releasing it using local_irq_restore(*@flags).
1990 * This function is safe to call from any context including IRQ handler.
1992 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1993 unsigned long *flags)
1995 struct worker_pool *pool;
1996 struct pool_workqueue *pwq;
1998 local_irq_save(*flags);
2000 /* try to steal the timer if it exists */
2002 struct delayed_work *dwork = to_delayed_work(work);
2005 * dwork->timer is irqsafe. If del_timer() fails, it's
2006 * guaranteed that the timer is not queued anywhere and not
2007 * running on the local CPU.
2009 if (likely(del_timer(&dwork->timer)))
2013 /* try to claim PENDING the normal way */
2014 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2019 * The queueing is in progress, or it is already queued. Try to
2020 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2022 pool = get_work_pool(work);
2026 raw_spin_lock(&pool->lock);
2028 * work->data is guaranteed to point to pwq only while the work
2029 * item is queued on pwq->wq, and both updating work->data to point
2030 * to pwq on queueing and to pool on dequeueing are done under
2031 * pwq->pool->lock. This in turn guarantees that, if work->data
2032 * points to pwq which is associated with a locked pool, the work
2033 * item is currently queued on that pool.
2035 pwq = get_work_pwq(work);
2036 if (pwq && pwq->pool == pool) {
2037 unsigned long work_data;
2039 debug_work_deactivate(work);
2042 * A cancelable inactive work item must be in the
2043 * pwq->inactive_works since a queued barrier can't be
2044 * canceled (see the comments in insert_wq_barrier()).
2046 * An inactive work item cannot be grabbed directly because
2047 * it might have linked barrier work items which, if left
2048 * on the inactive_works list, will confuse pwq->nr_active
2049 * management later on and cause stall. Make sure the work
2050 * item is activated before grabbing.
2052 pwq_activate_work(pwq, work);
2054 list_del_init(&work->entry);
2057 * work->data points to pwq iff queued. Let's point to pool. As
2058 * this destroys work->data needed by the next step, stash it.
2060 work_data = *work_data_bits(work);
2061 set_work_pool_and_keep_pending(work, pool->id);
2063 /* must be the last step, see the function comment */
2064 pwq_dec_nr_in_flight(pwq, work_data);
2066 raw_spin_unlock(&pool->lock);
2070 raw_spin_unlock(&pool->lock);
2073 local_irq_restore(*flags);
2074 if (work_is_canceling(work))
2081 * insert_work - insert a work into a pool
2082 * @pwq: pwq @work belongs to
2083 * @work: work to insert
2084 * @head: insertion point
2085 * @extra_flags: extra WORK_STRUCT_* flags to set
2087 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2088 * work_struct flags.
2091 * raw_spin_lock_irq(pool->lock).
2093 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2094 struct list_head *head, unsigned int extra_flags)
2096 debug_work_activate(work);
2098 /* record the work call stack in order to print it in KASAN reports */
2099 kasan_record_aux_stack_noalloc(work);
2101 /* we own @work, set data and link */
2102 set_work_pwq(work, pwq, extra_flags);
2103 list_add_tail(&work->entry, head);
2108 * Test whether @work is being queued from another work executing on the
2111 static bool is_chained_work(struct workqueue_struct *wq)
2113 struct worker *worker;
2115 worker = current_wq_worker();
2117 * Return %true iff I'm a worker executing a work item on @wq. If
2118 * I'm @worker, it's safe to dereference it without locking.
2120 return worker && worker->current_pwq->wq == wq;
2124 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2125 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2126 * avoid perturbing sensitive tasks.
2128 static int wq_select_unbound_cpu(int cpu)
2132 if (likely(!wq_debug_force_rr_cpu)) {
2133 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2136 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2139 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2140 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2141 if (unlikely(new_cpu >= nr_cpu_ids)) {
2142 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2143 if (unlikely(new_cpu >= nr_cpu_ids))
2146 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2151 static void __queue_work(int cpu, struct workqueue_struct *wq,
2152 struct work_struct *work)
2154 struct pool_workqueue *pwq;
2155 struct worker_pool *last_pool, *pool;
2156 unsigned int work_flags;
2157 unsigned int req_cpu = cpu;
2160 * While a work item is PENDING && off queue, a task trying to
2161 * steal the PENDING will busy-loop waiting for it to either get
2162 * queued or lose PENDING. Grabbing PENDING and queueing should
2163 * happen with IRQ disabled.
2165 lockdep_assert_irqs_disabled();
2169 * For a draining wq, only works from the same workqueue are
2170 * allowed. The __WQ_DESTROYING helps to spot the issue that
2171 * queues a new work item to a wq after destroy_workqueue(wq).
2173 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2174 WARN_ON_ONCE(!is_chained_work(wq))))
2178 /* pwq which will be used unless @work is executing elsewhere */
2179 if (req_cpu == WORK_CPU_UNBOUND) {
2180 if (wq->flags & WQ_UNBOUND)
2181 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2183 cpu = raw_smp_processor_id();
2186 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2190 * If @work was previously on a different pool, it might still be
2191 * running there, in which case the work needs to be queued on that
2192 * pool to guarantee non-reentrancy.
2194 last_pool = get_work_pool(work);
2195 if (last_pool && last_pool != pool) {
2196 struct worker *worker;
2198 raw_spin_lock(&last_pool->lock);
2200 worker = find_worker_executing_work(last_pool, work);
2202 if (worker && worker->current_pwq->wq == wq) {
2203 pwq = worker->current_pwq;
2205 WARN_ON_ONCE(pool != last_pool);
2207 /* meh... not running there, queue here */
2208 raw_spin_unlock(&last_pool->lock);
2209 raw_spin_lock(&pool->lock);
2212 raw_spin_lock(&pool->lock);
2216 * pwq is determined and locked. For unbound pools, we could have raced
2217 * with pwq release and it could already be dead. If its refcnt is zero,
2218 * repeat pwq selection. Note that unbound pwqs never die without
2219 * another pwq replacing it in cpu_pwq or while work items are executing
2220 * on it, so the retrying is guaranteed to make forward-progress.
2222 if (unlikely(!pwq->refcnt)) {
2223 if (wq->flags & WQ_UNBOUND) {
2224 raw_spin_unlock(&pool->lock);
2229 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2233 /* pwq determined, queue */
2234 trace_workqueue_queue_work(req_cpu, pwq, work);
2236 if (WARN_ON(!list_empty(&work->entry)))
2239 pwq->nr_in_flight[pwq->work_color]++;
2240 work_flags = work_color_to_flags(pwq->work_color);
2243 * Limit the number of concurrently active work items to max_active.
2244 * @work must also queue behind existing inactive work items to maintain
2245 * ordering when max_active changes. See wq_adjust_max_active().
2247 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2248 if (list_empty(&pool->worklist))
2249 pool->watchdog_ts = jiffies;
2251 trace_workqueue_activate_work(work);
2252 insert_work(pwq, work, &pool->worklist, work_flags);
2255 work_flags |= WORK_STRUCT_INACTIVE;
2256 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2260 raw_spin_unlock(&pool->lock);
2265 * queue_work_on - queue work on specific cpu
2266 * @cpu: CPU number to execute work on
2267 * @wq: workqueue to use
2268 * @work: work to queue
2270 * We queue the work to a specific CPU, the caller must ensure it
2271 * can't go away. Callers that fail to ensure that the specified
2272 * CPU cannot go away will execute on a randomly chosen CPU.
2273 * But note well that callers specifying a CPU that never has been
2274 * online will get a splat.
2276 * Return: %false if @work was already on a queue, %true otherwise.
2278 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2279 struct work_struct *work)
2282 unsigned long flags;
2284 local_irq_save(flags);
2286 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2287 __queue_work(cpu, wq, work);
2291 local_irq_restore(flags);
2294 EXPORT_SYMBOL(queue_work_on);
2297 * select_numa_node_cpu - Select a CPU based on NUMA node
2298 * @node: NUMA node ID that we want to select a CPU from
2300 * This function will attempt to find a "random" cpu available on a given
2301 * node. If there are no CPUs available on the given node it will return
2302 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2303 * available CPU if we need to schedule this work.
2305 static int select_numa_node_cpu(int node)
2309 /* Delay binding to CPU if node is not valid or online */
2310 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2311 return WORK_CPU_UNBOUND;
2313 /* Use local node/cpu if we are already there */
2314 cpu = raw_smp_processor_id();
2315 if (node == cpu_to_node(cpu))
2318 /* Use "random" otherwise know as "first" online CPU of node */
2319 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2321 /* If CPU is valid return that, otherwise just defer */
2322 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2326 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2327 * @node: NUMA node that we are targeting the work for
2328 * @wq: workqueue to use
2329 * @work: work to queue
2331 * We queue the work to a "random" CPU within a given NUMA node. The basic
2332 * idea here is to provide a way to somehow associate work with a given
2335 * This function will only make a best effort attempt at getting this onto
2336 * the right NUMA node. If no node is requested or the requested node is
2337 * offline then we just fall back to standard queue_work behavior.
2339 * Currently the "random" CPU ends up being the first available CPU in the
2340 * intersection of cpu_online_mask and the cpumask of the node, unless we
2341 * are running on the node. In that case we just use the current CPU.
2343 * Return: %false if @work was already on a queue, %true otherwise.
2345 bool queue_work_node(int node, struct workqueue_struct *wq,
2346 struct work_struct *work)
2348 unsigned long flags;
2352 * This current implementation is specific to unbound workqueues.
2353 * Specifically we only return the first available CPU for a given
2354 * node instead of cycling through individual CPUs within the node.
2356 * If this is used with a per-cpu workqueue then the logic in
2357 * workqueue_select_cpu_near would need to be updated to allow for
2358 * some round robin type logic.
2360 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2362 local_irq_save(flags);
2364 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2365 int cpu = select_numa_node_cpu(node);
2367 __queue_work(cpu, wq, work);
2371 local_irq_restore(flags);
2374 EXPORT_SYMBOL_GPL(queue_work_node);
2376 void delayed_work_timer_fn(struct timer_list *t)
2378 struct delayed_work *dwork = from_timer(dwork, t, timer);
2380 /* should have been called from irqsafe timer with irq already off */
2381 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2383 EXPORT_SYMBOL(delayed_work_timer_fn);
2385 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2386 struct delayed_work *dwork, unsigned long delay)
2388 struct timer_list *timer = &dwork->timer;
2389 struct work_struct *work = &dwork->work;
2392 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2393 WARN_ON_ONCE(timer_pending(timer));
2394 WARN_ON_ONCE(!list_empty(&work->entry));
2397 * If @delay is 0, queue @dwork->work immediately. This is for
2398 * both optimization and correctness. The earliest @timer can
2399 * expire is on the closest next tick and delayed_work users depend
2400 * on that there's no such delay when @delay is 0.
2403 __queue_work(cpu, wq, &dwork->work);
2409 timer->expires = jiffies + delay;
2411 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2412 /* If the current cpu is a housekeeping cpu, use it. */
2413 cpu = smp_processor_id();
2414 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2415 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2416 add_timer_on(timer, cpu);
2418 if (likely(cpu == WORK_CPU_UNBOUND))
2421 add_timer_on(timer, cpu);
2426 * queue_delayed_work_on - queue work on specific CPU after delay
2427 * @cpu: CPU number to execute work on
2428 * @wq: workqueue to use
2429 * @dwork: work to queue
2430 * @delay: number of jiffies to wait before queueing
2432 * Return: %false if @work was already on a queue, %true otherwise. If
2433 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2436 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2437 struct delayed_work *dwork, unsigned long delay)
2439 struct work_struct *work = &dwork->work;
2441 unsigned long flags;
2443 /* read the comment in __queue_work() */
2444 local_irq_save(flags);
2446 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2447 __queue_delayed_work(cpu, wq, dwork, delay);
2451 local_irq_restore(flags);
2454 EXPORT_SYMBOL(queue_delayed_work_on);
2457 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2458 * @cpu: CPU number to execute work on
2459 * @wq: workqueue to use
2460 * @dwork: work to queue
2461 * @delay: number of jiffies to wait before queueing
2463 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2464 * modify @dwork's timer so that it expires after @delay. If @delay is
2465 * zero, @work is guaranteed to be scheduled immediately regardless of its
2468 * Return: %false if @dwork was idle and queued, %true if @dwork was
2469 * pending and its timer was modified.
2471 * This function is safe to call from any context including IRQ handler.
2472 * See try_to_grab_pending() for details.
2474 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2475 struct delayed_work *dwork, unsigned long delay)
2477 unsigned long flags;
2481 ret = try_to_grab_pending(&dwork->work, true, &flags);
2482 } while (unlikely(ret == -EAGAIN));
2484 if (likely(ret >= 0)) {
2485 __queue_delayed_work(cpu, wq, dwork, delay);
2486 local_irq_restore(flags);
2489 /* -ENOENT from try_to_grab_pending() becomes %true */
2492 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2494 static void rcu_work_rcufn(struct rcu_head *rcu)
2496 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2498 /* read the comment in __queue_work() */
2499 local_irq_disable();
2500 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2505 * queue_rcu_work - queue work after a RCU grace period
2506 * @wq: workqueue to use
2507 * @rwork: work to queue
2509 * Return: %false if @rwork was already pending, %true otherwise. Note
2510 * that a full RCU grace period is guaranteed only after a %true return.
2511 * While @rwork is guaranteed to be executed after a %false return, the
2512 * execution may happen before a full RCU grace period has passed.
2514 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2516 struct work_struct *work = &rwork->work;
2518 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2520 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2526 EXPORT_SYMBOL(queue_rcu_work);
2528 static struct worker *alloc_worker(int node)
2530 struct worker *worker;
2532 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2534 INIT_LIST_HEAD(&worker->entry);
2535 INIT_LIST_HEAD(&worker->scheduled);
2536 INIT_LIST_HEAD(&worker->node);
2537 /* on creation a worker is in !idle && prep state */
2538 worker->flags = WORKER_PREP;
2543 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2545 if (pool->cpu < 0 && pool->attrs->affn_strict)
2546 return pool->attrs->__pod_cpumask;
2548 return pool->attrs->cpumask;
2552 * worker_attach_to_pool() - attach a worker to a pool
2553 * @worker: worker to be attached
2554 * @pool: the target pool
2556 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2557 * cpu-binding of @worker are kept coordinated with the pool across
2560 static void worker_attach_to_pool(struct worker *worker,
2561 struct worker_pool *pool)
2563 mutex_lock(&wq_pool_attach_mutex);
2566 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2567 * across this function. See the comments above the flag definition for
2568 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2570 if (pool->flags & POOL_DISASSOCIATED) {
2571 worker->flags |= WORKER_UNBOUND;
2573 WARN_ON_ONCE(pool->flags & POOL_BH);
2574 kthread_set_per_cpu(worker->task, pool->cpu);
2577 if (worker->rescue_wq)
2578 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2580 list_add_tail(&worker->node, &pool->workers);
2581 worker->pool = pool;
2583 mutex_unlock(&wq_pool_attach_mutex);
2587 * worker_detach_from_pool() - detach a worker from its pool
2588 * @worker: worker which is attached to its pool
2590 * Undo the attaching which had been done in worker_attach_to_pool(). The
2591 * caller worker shouldn't access to the pool after detached except it has
2592 * other reference to the pool.
2594 static void worker_detach_from_pool(struct worker *worker)
2596 struct worker_pool *pool = worker->pool;
2597 struct completion *detach_completion = NULL;
2599 /* there is one permanent BH worker per CPU which should never detach */
2600 WARN_ON_ONCE(pool->flags & POOL_BH);
2602 mutex_lock(&wq_pool_attach_mutex);
2604 kthread_set_per_cpu(worker->task, -1);
2605 list_del(&worker->node);
2606 worker->pool = NULL;
2608 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2609 detach_completion = pool->detach_completion;
2610 mutex_unlock(&wq_pool_attach_mutex);
2612 /* clear leftover flags without pool->lock after it is detached */
2613 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2615 if (detach_completion)
2616 complete(detach_completion);
2620 * create_worker - create a new workqueue worker
2621 * @pool: pool the new worker will belong to
2623 * Create and start a new worker which is attached to @pool.
2626 * Might sleep. Does GFP_KERNEL allocations.
2629 * Pointer to the newly created worker.
2631 static struct worker *create_worker(struct worker_pool *pool)
2633 struct worker *worker;
2637 /* ID is needed to determine kthread name */
2638 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2640 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2645 worker = alloc_worker(pool->node);
2647 pr_err_once("workqueue: Failed to allocate a worker\n");
2653 if (!(pool->flags & POOL_BH)) {
2655 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2656 pool->attrs->nice < 0 ? "H" : "");
2658 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2660 worker->task = kthread_create_on_node(worker_thread, worker,
2661 pool->node, "kworker/%s", id_buf);
2662 if (IS_ERR(worker->task)) {
2663 if (PTR_ERR(worker->task) == -EINTR) {
2664 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2667 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2673 set_user_nice(worker->task, pool->attrs->nice);
2674 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2677 /* successful, attach the worker to the pool */
2678 worker_attach_to_pool(worker, pool);
2680 /* start the newly created worker */
2681 raw_spin_lock_irq(&pool->lock);
2683 worker->pool->nr_workers++;
2684 worker_enter_idle(worker);
2687 * @worker is waiting on a completion in kthread() and will trigger hung
2688 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2689 * wake it up explicitly.
2692 wake_up_process(worker->task);
2694 raw_spin_unlock_irq(&pool->lock);
2699 ida_free(&pool->worker_ida, id);
2704 static void unbind_worker(struct worker *worker)
2706 lockdep_assert_held(&wq_pool_attach_mutex);
2708 kthread_set_per_cpu(worker->task, -1);
2709 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2710 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2712 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2715 static void wake_dying_workers(struct list_head *cull_list)
2717 struct worker *worker, *tmp;
2719 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2720 list_del_init(&worker->entry);
2721 unbind_worker(worker);
2723 * If the worker was somehow already running, then it had to be
2724 * in pool->idle_list when set_worker_dying() happened or we
2725 * wouldn't have gotten here.
2727 * Thus, the worker must either have observed the WORKER_DIE
2728 * flag, or have set its state to TASK_IDLE. Either way, the
2729 * below will be observed by the worker and is safe to do
2730 * outside of pool->lock.
2732 wake_up_process(worker->task);
2737 * set_worker_dying - Tag a worker for destruction
2738 * @worker: worker to be destroyed
2739 * @list: transfer worker away from its pool->idle_list and into list
2741 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2745 * raw_spin_lock_irq(pool->lock).
2747 static void set_worker_dying(struct worker *worker, struct list_head *list)
2749 struct worker_pool *pool = worker->pool;
2751 lockdep_assert_held(&pool->lock);
2752 lockdep_assert_held(&wq_pool_attach_mutex);
2754 /* sanity check frenzy */
2755 if (WARN_ON(worker->current_work) ||
2756 WARN_ON(!list_empty(&worker->scheduled)) ||
2757 WARN_ON(!(worker->flags & WORKER_IDLE)))
2763 worker->flags |= WORKER_DIE;
2765 list_move(&worker->entry, list);
2766 list_move(&worker->node, &pool->dying_workers);
2770 * idle_worker_timeout - check if some idle workers can now be deleted.
2771 * @t: The pool's idle_timer that just expired
2773 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2774 * worker_leave_idle(), as a worker flicking between idle and active while its
2775 * pool is at the too_many_workers() tipping point would cause too much timer
2776 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2777 * it expire and re-evaluate things from there.
2779 static void idle_worker_timeout(struct timer_list *t)
2781 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2782 bool do_cull = false;
2784 if (work_pending(&pool->idle_cull_work))
2787 raw_spin_lock_irq(&pool->lock);
2789 if (too_many_workers(pool)) {
2790 struct worker *worker;
2791 unsigned long expires;
2793 /* idle_list is kept in LIFO order, check the last one */
2794 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2795 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2796 do_cull = !time_before(jiffies, expires);
2799 mod_timer(&pool->idle_timer, expires);
2801 raw_spin_unlock_irq(&pool->lock);
2804 queue_work(system_unbound_wq, &pool->idle_cull_work);
2808 * idle_cull_fn - cull workers that have been idle for too long.
2809 * @work: the pool's work for handling these idle workers
2811 * This goes through a pool's idle workers and gets rid of those that have been
2812 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2814 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2815 * culled, so this also resets worker affinity. This requires a sleepable
2816 * context, hence the split between timer callback and work item.
2818 static void idle_cull_fn(struct work_struct *work)
2820 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2821 LIST_HEAD(cull_list);
2824 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2825 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2826 * path. This is required as a previously-preempted worker could run after
2827 * set_worker_dying() has happened but before wake_dying_workers() did.
2829 mutex_lock(&wq_pool_attach_mutex);
2830 raw_spin_lock_irq(&pool->lock);
2832 while (too_many_workers(pool)) {
2833 struct worker *worker;
2834 unsigned long expires;
2836 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2837 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2839 if (time_before(jiffies, expires)) {
2840 mod_timer(&pool->idle_timer, expires);
2844 set_worker_dying(worker, &cull_list);
2847 raw_spin_unlock_irq(&pool->lock);
2848 wake_dying_workers(&cull_list);
2849 mutex_unlock(&wq_pool_attach_mutex);
2852 static void send_mayday(struct work_struct *work)
2854 struct pool_workqueue *pwq = get_work_pwq(work);
2855 struct workqueue_struct *wq = pwq->wq;
2857 lockdep_assert_held(&wq_mayday_lock);
2862 /* mayday mayday mayday */
2863 if (list_empty(&pwq->mayday_node)) {
2865 * If @pwq is for an unbound wq, its base ref may be put at
2866 * any time due to an attribute change. Pin @pwq until the
2867 * rescuer is done with it.
2870 list_add_tail(&pwq->mayday_node, &wq->maydays);
2871 wake_up_process(wq->rescuer->task);
2872 pwq->stats[PWQ_STAT_MAYDAY]++;
2876 static void pool_mayday_timeout(struct timer_list *t)
2878 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2879 struct work_struct *work;
2881 raw_spin_lock_irq(&pool->lock);
2882 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2884 if (need_to_create_worker(pool)) {
2886 * We've been trying to create a new worker but
2887 * haven't been successful. We might be hitting an
2888 * allocation deadlock. Send distress signals to
2891 list_for_each_entry(work, &pool->worklist, entry)
2895 raw_spin_unlock(&wq_mayday_lock);
2896 raw_spin_unlock_irq(&pool->lock);
2898 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2902 * maybe_create_worker - create a new worker if necessary
2903 * @pool: pool to create a new worker for
2905 * Create a new worker for @pool if necessary. @pool is guaranteed to
2906 * have at least one idle worker on return from this function. If
2907 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2908 * sent to all rescuers with works scheduled on @pool to resolve
2909 * possible allocation deadlock.
2911 * On return, need_to_create_worker() is guaranteed to be %false and
2912 * may_start_working() %true.
2915 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2916 * multiple times. Does GFP_KERNEL allocations. Called only from
2919 static void maybe_create_worker(struct worker_pool *pool)
2920 __releases(&pool->lock)
2921 __acquires(&pool->lock)
2924 raw_spin_unlock_irq(&pool->lock);
2926 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2927 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2930 if (create_worker(pool) || !need_to_create_worker(pool))
2933 schedule_timeout_interruptible(CREATE_COOLDOWN);
2935 if (!need_to_create_worker(pool))
2939 del_timer_sync(&pool->mayday_timer);
2940 raw_spin_lock_irq(&pool->lock);
2942 * This is necessary even after a new worker was just successfully
2943 * created as @pool->lock was dropped and the new worker might have
2944 * already become busy.
2946 if (need_to_create_worker(pool))
2951 * manage_workers - manage worker pool
2954 * Assume the manager role and manage the worker pool @worker belongs
2955 * to. At any given time, there can be only zero or one manager per
2956 * pool. The exclusion is handled automatically by this function.
2958 * The caller can safely start processing works on false return. On
2959 * true return, it's guaranteed that need_to_create_worker() is false
2960 * and may_start_working() is true.
2963 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2964 * multiple times. Does GFP_KERNEL allocations.
2967 * %false if the pool doesn't need management and the caller can safely
2968 * start processing works, %true if management function was performed and
2969 * the conditions that the caller verified before calling the function may
2970 * no longer be true.
2972 static bool manage_workers(struct worker *worker)
2974 struct worker_pool *pool = worker->pool;
2976 if (pool->flags & POOL_MANAGER_ACTIVE)
2979 pool->flags |= POOL_MANAGER_ACTIVE;
2980 pool->manager = worker;
2982 maybe_create_worker(pool);
2984 pool->manager = NULL;
2985 pool->flags &= ~POOL_MANAGER_ACTIVE;
2986 rcuwait_wake_up(&manager_wait);
2991 * process_one_work - process single work
2993 * @work: work to process
2995 * Process @work. This function contains all the logics necessary to
2996 * process a single work including synchronization against and
2997 * interaction with other workers on the same cpu, queueing and
2998 * flushing. As long as context requirement is met, any worker can
2999 * call this function to process a work.
3002 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3004 static void process_one_work(struct worker *worker, struct work_struct *work)
3005 __releases(&pool->lock)
3006 __acquires(&pool->lock)
3008 struct pool_workqueue *pwq = get_work_pwq(work);
3009 struct worker_pool *pool = worker->pool;
3010 unsigned long work_data;
3011 int lockdep_start_depth, rcu_start_depth;
3012 #ifdef CONFIG_LOCKDEP
3014 * It is permissible to free the struct work_struct from
3015 * inside the function that is called from it, this we need to
3016 * take into account for lockdep too. To avoid bogus "held
3017 * lock freed" warnings as well as problems when looking into
3018 * work->lockdep_map, make a copy and use that here.
3020 struct lockdep_map lockdep_map;
3022 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3024 /* ensure we're on the correct CPU */
3025 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3026 raw_smp_processor_id() != pool->cpu);
3028 /* claim and dequeue */
3029 debug_work_deactivate(work);
3030 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3031 worker->current_work = work;
3032 worker->current_func = work->func;
3033 worker->current_pwq = pwq;
3035 worker->current_at = worker->task->se.sum_exec_runtime;
3036 work_data = *work_data_bits(work);
3037 worker->current_color = get_work_color(work_data);
3040 * Record wq name for cmdline and debug reporting, may get
3041 * overridden through set_worker_desc().
3043 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3045 list_del_init(&work->entry);
3048 * CPU intensive works don't participate in concurrency management.
3049 * They're the scheduler's responsibility. This takes @worker out
3050 * of concurrency management and the next code block will chain
3051 * execution of the pending work items.
3053 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3054 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3057 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3058 * since nr_running would always be >= 1 at this point. This is used to
3059 * chain execution of the pending work items for WORKER_NOT_RUNNING
3060 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3065 * Record the last pool and clear PENDING which should be the last
3066 * update to @work. Also, do this inside @pool->lock so that
3067 * PENDING and queued state changes happen together while IRQ is
3070 set_work_pool_and_clear_pending(work, pool->id);
3072 pwq->stats[PWQ_STAT_STARTED]++;
3073 raw_spin_unlock_irq(&pool->lock);
3075 rcu_start_depth = rcu_preempt_depth();
3076 lockdep_start_depth = lockdep_depth(current);
3077 lock_map_acquire(&pwq->wq->lockdep_map);
3078 lock_map_acquire(&lockdep_map);
3080 * Strictly speaking we should mark the invariant state without holding
3081 * any locks, that is, before these two lock_map_acquire()'s.
3083 * However, that would result in:
3090 * Which would create W1->C->W1 dependencies, even though there is no
3091 * actual deadlock possible. There are two solutions, using a
3092 * read-recursive acquire on the work(queue) 'locks', but this will then
3093 * hit the lockdep limitation on recursive locks, or simply discard
3096 * AFAICT there is no possible deadlock scenario between the
3097 * flush_work() and complete() primitives (except for single-threaded
3098 * workqueues), so hiding them isn't a problem.
3100 lockdep_invariant_state(true);
3101 trace_workqueue_execute_start(work);
3102 worker->current_func(work);
3104 * While we must be careful to not use "work" after this, the trace
3105 * point will only record its address.
3107 trace_workqueue_execute_end(work, worker->current_func);
3108 pwq->stats[PWQ_STAT_COMPLETED]++;
3109 lock_map_release(&lockdep_map);
3110 lock_map_release(&pwq->wq->lockdep_map);
3112 if (unlikely((worker->task && in_atomic()) ||
3113 lockdep_depth(current) != lockdep_start_depth ||
3114 rcu_preempt_depth() != rcu_start_depth)) {
3115 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3116 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3117 current->comm, task_pid_nr(current), preempt_count(),
3118 lockdep_start_depth, lockdep_depth(current),
3119 rcu_start_depth, rcu_preempt_depth(),
3120 worker->current_func);
3121 debug_show_held_locks(current);
3126 * The following prevents a kworker from hogging CPU on !PREEMPTION
3127 * kernels, where a requeueing work item waiting for something to
3128 * happen could deadlock with stop_machine as such work item could
3129 * indefinitely requeue itself while all other CPUs are trapped in
3130 * stop_machine. At the same time, report a quiescent RCU state so
3131 * the same condition doesn't freeze RCU.
3136 raw_spin_lock_irq(&pool->lock);
3139 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3140 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3141 * wq_cpu_intensive_thresh_us. Clear it.
3143 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3145 /* tag the worker for identification in schedule() */
3146 worker->last_func = worker->current_func;
3148 /* we're done with it, release */
3149 hash_del(&worker->hentry);
3150 worker->current_work = NULL;
3151 worker->current_func = NULL;
3152 worker->current_pwq = NULL;
3153 worker->current_color = INT_MAX;
3155 /* must be the last step, see the function comment */
3156 pwq_dec_nr_in_flight(pwq, work_data);
3160 * process_scheduled_works - process scheduled works
3163 * Process all scheduled works. Please note that the scheduled list
3164 * may change while processing a work, so this function repeatedly
3165 * fetches a work from the top and executes it.
3168 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3171 static void process_scheduled_works(struct worker *worker)
3173 struct work_struct *work;
3176 while ((work = list_first_entry_or_null(&worker->scheduled,
3177 struct work_struct, entry))) {
3179 worker->pool->watchdog_ts = jiffies;
3182 process_one_work(worker, work);
3186 static void set_pf_worker(bool val)
3188 mutex_lock(&wq_pool_attach_mutex);
3190 current->flags |= PF_WQ_WORKER;
3192 current->flags &= ~PF_WQ_WORKER;
3193 mutex_unlock(&wq_pool_attach_mutex);
3197 * worker_thread - the worker thread function
3200 * The worker thread function. All workers belong to a worker_pool -
3201 * either a per-cpu one or dynamic unbound one. These workers process all
3202 * work items regardless of their specific target workqueue. The only
3203 * exception is work items which belong to workqueues with a rescuer which
3204 * will be explained in rescuer_thread().
3208 static int worker_thread(void *__worker)
3210 struct worker *worker = __worker;
3211 struct worker_pool *pool = worker->pool;
3213 /* tell the scheduler that this is a workqueue worker */
3214 set_pf_worker(true);
3216 raw_spin_lock_irq(&pool->lock);
3218 /* am I supposed to die? */
3219 if (unlikely(worker->flags & WORKER_DIE)) {
3220 raw_spin_unlock_irq(&pool->lock);
3221 set_pf_worker(false);
3223 set_task_comm(worker->task, "kworker/dying");
3224 ida_free(&pool->worker_ida, worker->id);
3225 worker_detach_from_pool(worker);
3226 WARN_ON_ONCE(!list_empty(&worker->entry));
3231 worker_leave_idle(worker);
3233 /* no more worker necessary? */
3234 if (!need_more_worker(pool))
3237 /* do we need to manage? */
3238 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3242 * ->scheduled list can only be filled while a worker is
3243 * preparing to process a work or actually processing it.
3244 * Make sure nobody diddled with it while I was sleeping.
3246 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3249 * Finish PREP stage. We're guaranteed to have at least one idle
3250 * worker or that someone else has already assumed the manager
3251 * role. This is where @worker starts participating in concurrency
3252 * management if applicable and concurrency management is restored
3253 * after being rebound. See rebind_workers() for details.
3255 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3258 struct work_struct *work =
3259 list_first_entry(&pool->worklist,
3260 struct work_struct, entry);
3262 if (assign_work(work, worker, NULL))
3263 process_scheduled_works(worker);
3264 } while (keep_working(pool));
3266 worker_set_flags(worker, WORKER_PREP);
3269 * pool->lock is held and there's no work to process and no need to
3270 * manage, sleep. Workers are woken up only while holding
3271 * pool->lock or from local cpu, so setting the current state
3272 * before releasing pool->lock is enough to prevent losing any
3275 worker_enter_idle(worker);
3276 __set_current_state(TASK_IDLE);
3277 raw_spin_unlock_irq(&pool->lock);
3283 * rescuer_thread - the rescuer thread function
3286 * Workqueue rescuer thread function. There's one rescuer for each
3287 * workqueue which has WQ_MEM_RECLAIM set.
3289 * Regular work processing on a pool may block trying to create a new
3290 * worker which uses GFP_KERNEL allocation which has slight chance of
3291 * developing into deadlock if some works currently on the same queue
3292 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3293 * the problem rescuer solves.
3295 * When such condition is possible, the pool summons rescuers of all
3296 * workqueues which have works queued on the pool and let them process
3297 * those works so that forward progress can be guaranteed.
3299 * This should happen rarely.
3303 static int rescuer_thread(void *__rescuer)
3305 struct worker *rescuer = __rescuer;
3306 struct workqueue_struct *wq = rescuer->rescue_wq;
3309 set_user_nice(current, RESCUER_NICE_LEVEL);
3312 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3313 * doesn't participate in concurrency management.
3315 set_pf_worker(true);
3317 set_current_state(TASK_IDLE);
3320 * By the time the rescuer is requested to stop, the workqueue
3321 * shouldn't have any work pending, but @wq->maydays may still have
3322 * pwq(s) queued. This can happen by non-rescuer workers consuming
3323 * all the work items before the rescuer got to them. Go through
3324 * @wq->maydays processing before acting on should_stop so that the
3325 * list is always empty on exit.
3327 should_stop = kthread_should_stop();
3329 /* see whether any pwq is asking for help */
3330 raw_spin_lock_irq(&wq_mayday_lock);
3332 while (!list_empty(&wq->maydays)) {
3333 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3334 struct pool_workqueue, mayday_node);
3335 struct worker_pool *pool = pwq->pool;
3336 struct work_struct *work, *n;
3338 __set_current_state(TASK_RUNNING);
3339 list_del_init(&pwq->mayday_node);
3341 raw_spin_unlock_irq(&wq_mayday_lock);
3343 worker_attach_to_pool(rescuer, pool);
3345 raw_spin_lock_irq(&pool->lock);
3348 * Slurp in all works issued via this workqueue and
3351 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3352 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3353 if (get_work_pwq(work) == pwq &&
3354 assign_work(work, rescuer, &n))
3355 pwq->stats[PWQ_STAT_RESCUED]++;
3358 if (!list_empty(&rescuer->scheduled)) {
3359 process_scheduled_works(rescuer);
3362 * The above execution of rescued work items could
3363 * have created more to rescue through
3364 * pwq_activate_first_inactive() or chained
3365 * queueing. Let's put @pwq back on mayday list so
3366 * that such back-to-back work items, which may be
3367 * being used to relieve memory pressure, don't
3368 * incur MAYDAY_INTERVAL delay inbetween.
3370 if (pwq->nr_active && need_to_create_worker(pool)) {
3371 raw_spin_lock(&wq_mayday_lock);
3373 * Queue iff we aren't racing destruction
3374 * and somebody else hasn't queued it already.
3376 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3378 list_add_tail(&pwq->mayday_node, &wq->maydays);
3380 raw_spin_unlock(&wq_mayday_lock);
3385 * Put the reference grabbed by send_mayday(). @pool won't
3386 * go away while we're still attached to it.
3391 * Leave this pool. Notify regular workers; otherwise, we end up
3392 * with 0 concurrency and stalling the execution.
3396 raw_spin_unlock_irq(&pool->lock);
3398 worker_detach_from_pool(rescuer);
3400 raw_spin_lock_irq(&wq_mayday_lock);
3403 raw_spin_unlock_irq(&wq_mayday_lock);
3406 __set_current_state(TASK_RUNNING);
3407 set_pf_worker(false);
3411 /* rescuers should never participate in concurrency management */
3412 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3417 static void bh_worker(struct worker *worker)
3419 struct worker_pool *pool = worker->pool;
3420 int nr_restarts = BH_WORKER_RESTARTS;
3421 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3423 raw_spin_lock_irq(&pool->lock);
3424 worker_leave_idle(worker);
3427 * This function follows the structure of worker_thread(). See there for
3428 * explanations on each step.
3430 if (!need_more_worker(pool))
3433 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3434 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3437 struct work_struct *work =
3438 list_first_entry(&pool->worklist,
3439 struct work_struct, entry);
3441 if (assign_work(work, worker, NULL))
3442 process_scheduled_works(worker);
3443 } while (keep_working(pool) &&
3444 --nr_restarts && time_before(jiffies, end));
3446 worker_set_flags(worker, WORKER_PREP);
3448 worker_enter_idle(worker);
3450 raw_spin_unlock_irq(&pool->lock);
3454 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3456 * This is currently called from tasklet[_hi]action() and thus is also called
3457 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3458 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3461 * After full conversion, we'll add worker->softirq_action, directly use the
3462 * softirq action and obtain the worker pointer from the softirq_action pointer.
3464 void workqueue_softirq_action(bool highpri)
3466 struct worker_pool *pool =
3467 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3468 if (need_more_worker(pool))
3469 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3473 * check_flush_dependency - check for flush dependency sanity
3474 * @target_wq: workqueue being flushed
3475 * @target_work: work item being flushed (NULL for workqueue flushes)
3477 * %current is trying to flush the whole @target_wq or @target_work on it.
3478 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3479 * reclaiming memory or running on a workqueue which doesn't have
3480 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3483 static void check_flush_dependency(struct workqueue_struct *target_wq,
3484 struct work_struct *target_work)
3486 work_func_t target_func = target_work ? target_work->func : NULL;
3487 struct worker *worker;
3489 if (target_wq->flags & WQ_MEM_RECLAIM)
3492 worker = current_wq_worker();
3494 WARN_ONCE(current->flags & PF_MEMALLOC,
3495 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3496 current->pid, current->comm, target_wq->name, target_func);
3497 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3498 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3499 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3500 worker->current_pwq->wq->name, worker->current_func,
3501 target_wq->name, target_func);
3505 struct work_struct work;
3506 struct completion done;
3507 struct task_struct *task; /* purely informational */
3510 static void wq_barrier_func(struct work_struct *work)
3512 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3513 complete(&barr->done);
3517 * insert_wq_barrier - insert a barrier work
3518 * @pwq: pwq to insert barrier into
3519 * @barr: wq_barrier to insert
3520 * @target: target work to attach @barr to
3521 * @worker: worker currently executing @target, NULL if @target is not executing
3523 * @barr is linked to @target such that @barr is completed only after
3524 * @target finishes execution. Please note that the ordering
3525 * guarantee is observed only with respect to @target and on the local
3528 * Currently, a queued barrier can't be canceled. This is because
3529 * try_to_grab_pending() can't determine whether the work to be
3530 * grabbed is at the head of the queue and thus can't clear LINKED
3531 * flag of the previous work while there must be a valid next work
3532 * after a work with LINKED flag set.
3534 * Note that when @worker is non-NULL, @target may be modified
3535 * underneath us, so we can't reliably determine pwq from @target.
3538 * raw_spin_lock_irq(pool->lock).
3540 static void insert_wq_barrier(struct pool_workqueue *pwq,
3541 struct wq_barrier *barr,
3542 struct work_struct *target, struct worker *worker)
3544 static __maybe_unused struct lock_class_key bh_key, thr_key;
3545 unsigned int work_flags = 0;
3546 unsigned int work_color;
3547 struct list_head *head;
3550 * debugobject calls are safe here even with pool->lock locked
3551 * as we know for sure that this will not trigger any of the
3552 * checks and call back into the fixup functions where we
3555 * BH and threaded workqueues need separate lockdep keys to avoid
3556 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3559 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3560 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3561 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3563 init_completion_map(&barr->done, &target->lockdep_map);
3565 barr->task = current;
3567 /* The barrier work item does not participate in nr_active. */
3568 work_flags |= WORK_STRUCT_INACTIVE;
3571 * If @target is currently being executed, schedule the
3572 * barrier to the worker; otherwise, put it after @target.
3575 head = worker->scheduled.next;
3576 work_color = worker->current_color;
3578 unsigned long *bits = work_data_bits(target);
3580 head = target->entry.next;
3581 /* there can already be other linked works, inherit and set */
3582 work_flags |= *bits & WORK_STRUCT_LINKED;
3583 work_color = get_work_color(*bits);
3584 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3587 pwq->nr_in_flight[work_color]++;
3588 work_flags |= work_color_to_flags(work_color);
3590 insert_work(pwq, &barr->work, head, work_flags);
3594 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3595 * @wq: workqueue being flushed
3596 * @flush_color: new flush color, < 0 for no-op
3597 * @work_color: new work color, < 0 for no-op
3599 * Prepare pwqs for workqueue flushing.
3601 * If @flush_color is non-negative, flush_color on all pwqs should be
3602 * -1. If no pwq has in-flight commands at the specified color, all
3603 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3604 * has in flight commands, its pwq->flush_color is set to
3605 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3606 * wakeup logic is armed and %true is returned.
3608 * The caller should have initialized @wq->first_flusher prior to
3609 * calling this function with non-negative @flush_color. If
3610 * @flush_color is negative, no flush color update is done and %false
3613 * If @work_color is non-negative, all pwqs should have the same
3614 * work_color which is previous to @work_color and all will be
3615 * advanced to @work_color.
3618 * mutex_lock(wq->mutex).
3621 * %true if @flush_color >= 0 and there's something to flush. %false
3624 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3625 int flush_color, int work_color)
3628 struct pool_workqueue *pwq;
3630 if (flush_color >= 0) {
3631 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3632 atomic_set(&wq->nr_pwqs_to_flush, 1);
3635 for_each_pwq(pwq, wq) {
3636 struct worker_pool *pool = pwq->pool;
3638 raw_spin_lock_irq(&pool->lock);
3640 if (flush_color >= 0) {
3641 WARN_ON_ONCE(pwq->flush_color != -1);
3643 if (pwq->nr_in_flight[flush_color]) {
3644 pwq->flush_color = flush_color;
3645 atomic_inc(&wq->nr_pwqs_to_flush);
3650 if (work_color >= 0) {
3651 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3652 pwq->work_color = work_color;
3655 raw_spin_unlock_irq(&pool->lock);
3658 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3659 complete(&wq->first_flusher->done);
3664 static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3666 #ifdef CONFIG_LOCKDEP
3667 if (wq->flags & WQ_BH)
3670 lock_map_acquire(&wq->lockdep_map);
3671 lock_map_release(&wq->lockdep_map);
3673 if (wq->flags & WQ_BH)
3678 static void touch_work_lockdep_map(struct work_struct *work,
3679 struct workqueue_struct *wq)
3681 #ifdef CONFIG_LOCKDEP
3682 if (wq->flags & WQ_BH)
3685 lock_map_acquire(&work->lockdep_map);
3686 lock_map_release(&work->lockdep_map);
3688 if (wq->flags & WQ_BH)
3694 * __flush_workqueue - ensure that any scheduled work has run to completion.
3695 * @wq: workqueue to flush
3697 * This function sleeps until all work items which were queued on entry
3698 * have finished execution, but it is not livelocked by new incoming ones.
3700 void __flush_workqueue(struct workqueue_struct *wq)
3702 struct wq_flusher this_flusher = {
3703 .list = LIST_HEAD_INIT(this_flusher.list),
3705 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3709 if (WARN_ON(!wq_online))
3712 touch_wq_lockdep_map(wq);
3714 mutex_lock(&wq->mutex);
3717 * Start-to-wait phase
3719 next_color = work_next_color(wq->work_color);
3721 if (next_color != wq->flush_color) {
3723 * Color space is not full. The current work_color
3724 * becomes our flush_color and work_color is advanced
3727 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3728 this_flusher.flush_color = wq->work_color;
3729 wq->work_color = next_color;
3731 if (!wq->first_flusher) {
3732 /* no flush in progress, become the first flusher */
3733 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3735 wq->first_flusher = &this_flusher;
3737 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3739 /* nothing to flush, done */
3740 wq->flush_color = next_color;
3741 wq->first_flusher = NULL;
3746 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3747 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3748 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3752 * Oops, color space is full, wait on overflow queue.
3753 * The next flush completion will assign us
3754 * flush_color and transfer to flusher_queue.
3756 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3759 check_flush_dependency(wq, NULL);
3761 mutex_unlock(&wq->mutex);
3763 wait_for_completion(&this_flusher.done);
3766 * Wake-up-and-cascade phase
3768 * First flushers are responsible for cascading flushes and
3769 * handling overflow. Non-first flushers can simply return.
3771 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3774 mutex_lock(&wq->mutex);
3776 /* we might have raced, check again with mutex held */
3777 if (wq->first_flusher != &this_flusher)
3780 WRITE_ONCE(wq->first_flusher, NULL);
3782 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3783 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3786 struct wq_flusher *next, *tmp;
3788 /* complete all the flushers sharing the current flush color */
3789 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3790 if (next->flush_color != wq->flush_color)
3792 list_del_init(&next->list);
3793 complete(&next->done);
3796 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3797 wq->flush_color != work_next_color(wq->work_color));
3799 /* this flush_color is finished, advance by one */
3800 wq->flush_color = work_next_color(wq->flush_color);
3802 /* one color has been freed, handle overflow queue */
3803 if (!list_empty(&wq->flusher_overflow)) {
3805 * Assign the same color to all overflowed
3806 * flushers, advance work_color and append to
3807 * flusher_queue. This is the start-to-wait
3808 * phase for these overflowed flushers.
3810 list_for_each_entry(tmp, &wq->flusher_overflow, list)
3811 tmp->flush_color = wq->work_color;
3813 wq->work_color = work_next_color(wq->work_color);
3815 list_splice_tail_init(&wq->flusher_overflow,
3816 &wq->flusher_queue);
3817 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3820 if (list_empty(&wq->flusher_queue)) {
3821 WARN_ON_ONCE(wq->flush_color != wq->work_color);
3826 * Need to flush more colors. Make the next flusher
3827 * the new first flusher and arm pwqs.
3829 WARN_ON_ONCE(wq->flush_color == wq->work_color);
3830 WARN_ON_ONCE(wq->flush_color != next->flush_color);
3832 list_del_init(&next->list);
3833 wq->first_flusher = next;
3835 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3839 * Meh... this color is already done, clear first
3840 * flusher and repeat cascading.
3842 wq->first_flusher = NULL;
3846 mutex_unlock(&wq->mutex);
3848 EXPORT_SYMBOL(__flush_workqueue);
3851 * drain_workqueue - drain a workqueue
3852 * @wq: workqueue to drain
3854 * Wait until the workqueue becomes empty. While draining is in progress,
3855 * only chain queueing is allowed. IOW, only currently pending or running
3856 * work items on @wq can queue further work items on it. @wq is flushed
3857 * repeatedly until it becomes empty. The number of flushing is determined
3858 * by the depth of chaining and should be relatively short. Whine if it
3861 void drain_workqueue(struct workqueue_struct *wq)
3863 unsigned int flush_cnt = 0;
3864 struct pool_workqueue *pwq;
3867 * __queue_work() needs to test whether there are drainers, is much
3868 * hotter than drain_workqueue() and already looks at @wq->flags.
3869 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3871 mutex_lock(&wq->mutex);
3872 if (!wq->nr_drainers++)
3873 wq->flags |= __WQ_DRAINING;
3874 mutex_unlock(&wq->mutex);
3876 __flush_workqueue(wq);
3878 mutex_lock(&wq->mutex);
3880 for_each_pwq(pwq, wq) {
3883 raw_spin_lock_irq(&pwq->pool->lock);
3884 drained = pwq_is_empty(pwq);
3885 raw_spin_unlock_irq(&pwq->pool->lock);
3890 if (++flush_cnt == 10 ||
3891 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3892 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3893 wq->name, __func__, flush_cnt);
3895 mutex_unlock(&wq->mutex);
3899 if (!--wq->nr_drainers)
3900 wq->flags &= ~__WQ_DRAINING;
3901 mutex_unlock(&wq->mutex);
3903 EXPORT_SYMBOL_GPL(drain_workqueue);
3905 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3908 struct worker *worker = NULL;
3909 struct worker_pool *pool;
3910 struct pool_workqueue *pwq;
3911 struct workqueue_struct *wq;
3916 pool = get_work_pool(work);
3922 raw_spin_lock_irq(&pool->lock);
3923 /* see the comment in try_to_grab_pending() with the same code */
3924 pwq = get_work_pwq(work);
3926 if (unlikely(pwq->pool != pool))
3929 worker = find_worker_executing_work(pool, work);
3932 pwq = worker->current_pwq;
3936 check_flush_dependency(wq, work);
3938 insert_wq_barrier(pwq, barr, work, worker);
3939 raw_spin_unlock_irq(&pool->lock);
3941 touch_work_lockdep_map(work, wq);
3944 * Force a lock recursion deadlock when using flush_work() inside a
3945 * single-threaded or rescuer equipped workqueue.
3947 * For single threaded workqueues the deadlock happens when the work
3948 * is after the work issuing the flush_work(). For rescuer equipped
3949 * workqueues the deadlock happens when the rescuer stalls, blocking
3952 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
3953 touch_wq_lockdep_map(wq);
3958 raw_spin_unlock_irq(&pool->lock);
3963 static bool __flush_work(struct work_struct *work, bool from_cancel)
3965 struct wq_barrier barr;
3967 if (WARN_ON(!wq_online))
3970 if (WARN_ON(!work->func))
3973 if (start_flush_work(work, &barr, from_cancel)) {
3974 wait_for_completion(&barr.done);
3975 destroy_work_on_stack(&barr.work);
3983 * flush_work - wait for a work to finish executing the last queueing instance
3984 * @work: the work to flush
3986 * Wait until @work has finished execution. @work is guaranteed to be idle
3987 * on return if it hasn't been requeued since flush started.
3990 * %true if flush_work() waited for the work to finish execution,
3991 * %false if it was already idle.
3993 bool flush_work(struct work_struct *work)
3995 return __flush_work(work, false);
3997 EXPORT_SYMBOL_GPL(flush_work);
4000 wait_queue_entry_t wait;
4001 struct work_struct *work;
4004 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
4006 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
4008 if (cwait->work != key)
4010 return autoremove_wake_function(wait, mode, sync, key);
4013 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
4015 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
4016 unsigned long flags;
4020 ret = try_to_grab_pending(work, is_dwork, &flags);
4022 * If someone else is already canceling, wait for it to
4023 * finish. flush_work() doesn't work for PREEMPT_NONE
4024 * because we may get scheduled between @work's completion
4025 * and the other canceling task resuming and clearing
4026 * CANCELING - flush_work() will return false immediately
4027 * as @work is no longer busy, try_to_grab_pending() will
4028 * return -ENOENT as @work is still being canceled and the
4029 * other canceling task won't be able to clear CANCELING as
4030 * we're hogging the CPU.
4032 * Let's wait for completion using a waitqueue. As this
4033 * may lead to the thundering herd problem, use a custom
4034 * wake function which matches @work along with exclusive
4037 if (unlikely(ret == -ENOENT)) {
4038 struct cwt_wait cwait;
4040 init_wait(&cwait.wait);
4041 cwait.wait.func = cwt_wakefn;
4044 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
4045 TASK_UNINTERRUPTIBLE);
4046 if (work_is_canceling(work))
4048 finish_wait(&cancel_waitq, &cwait.wait);
4050 } while (unlikely(ret < 0));
4052 /* tell other tasks trying to grab @work to back off */
4053 mark_work_canceling(work);
4054 local_irq_restore(flags);
4057 * This allows canceling during early boot. We know that @work
4061 __flush_work(work, true);
4063 clear_work_data(work);
4066 * Paired with prepare_to_wait() above so that either
4067 * waitqueue_active() is visible here or !work_is_canceling() is
4071 if (waitqueue_active(&cancel_waitq))
4072 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
4078 * cancel_work_sync - cancel a work and wait for it to finish
4079 * @work: the work to cancel
4081 * Cancel @work and wait for its execution to finish. This function
4082 * can be used even if the work re-queues itself or migrates to
4083 * another workqueue. On return from this function, @work is
4084 * guaranteed to be not pending or executing on any CPU.
4086 * cancel_work_sync(&delayed_work->work) must not be used for
4087 * delayed_work's. Use cancel_delayed_work_sync() instead.
4089 * The caller must ensure that the workqueue on which @work was last
4090 * queued can't be destroyed before this function returns.
4093 * %true if @work was pending, %false otherwise.
4095 bool cancel_work_sync(struct work_struct *work)
4097 return __cancel_work_timer(work, false);
4099 EXPORT_SYMBOL_GPL(cancel_work_sync);
4102 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4103 * @dwork: the delayed work to flush
4105 * Delayed timer is cancelled and the pending work is queued for
4106 * immediate execution. Like flush_work(), this function only
4107 * considers the last queueing instance of @dwork.
4110 * %true if flush_work() waited for the work to finish execution,
4111 * %false if it was already idle.
4113 bool flush_delayed_work(struct delayed_work *dwork)
4115 local_irq_disable();
4116 if (del_timer_sync(&dwork->timer))
4117 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4119 return flush_work(&dwork->work);
4121 EXPORT_SYMBOL(flush_delayed_work);
4124 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4125 * @rwork: the rcu work to flush
4128 * %true if flush_rcu_work() waited for the work to finish execution,
4129 * %false if it was already idle.
4131 bool flush_rcu_work(struct rcu_work *rwork)
4133 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4135 flush_work(&rwork->work);
4138 return flush_work(&rwork->work);
4141 EXPORT_SYMBOL(flush_rcu_work);
4143 static bool __cancel_work(struct work_struct *work, bool is_dwork)
4145 unsigned long flags;
4149 ret = try_to_grab_pending(work, is_dwork, &flags);
4150 } while (unlikely(ret == -EAGAIN));
4152 if (unlikely(ret < 0))
4155 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
4156 local_irq_restore(flags);
4161 * See cancel_delayed_work()
4163 bool cancel_work(struct work_struct *work)
4165 return __cancel_work(work, false);
4167 EXPORT_SYMBOL(cancel_work);
4170 * cancel_delayed_work - cancel a delayed work
4171 * @dwork: delayed_work to cancel
4173 * Kill off a pending delayed_work.
4175 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4179 * The work callback function may still be running on return, unless
4180 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4181 * use cancel_delayed_work_sync() to wait on it.
4183 * This function is safe to call from any context including IRQ handler.
4185 bool cancel_delayed_work(struct delayed_work *dwork)
4187 return __cancel_work(&dwork->work, true);
4189 EXPORT_SYMBOL(cancel_delayed_work);
4192 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4193 * @dwork: the delayed work cancel
4195 * This is cancel_work_sync() for delayed works.
4198 * %true if @dwork was pending, %false otherwise.
4200 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4202 return __cancel_work_timer(&dwork->work, true);
4204 EXPORT_SYMBOL(cancel_delayed_work_sync);
4207 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4208 * @func: the function to call
4210 * schedule_on_each_cpu() executes @func on each online CPU using the
4211 * system workqueue and blocks until all CPUs have completed.
4212 * schedule_on_each_cpu() is very slow.
4215 * 0 on success, -errno on failure.
4217 int schedule_on_each_cpu(work_func_t func)
4220 struct work_struct __percpu *works;
4222 works = alloc_percpu(struct work_struct);
4228 for_each_online_cpu(cpu) {
4229 struct work_struct *work = per_cpu_ptr(works, cpu);
4231 INIT_WORK(work, func);
4232 schedule_work_on(cpu, work);
4235 for_each_online_cpu(cpu)
4236 flush_work(per_cpu_ptr(works, cpu));
4244 * execute_in_process_context - reliably execute the routine with user context
4245 * @fn: the function to execute
4246 * @ew: guaranteed storage for the execute work structure (must
4247 * be available when the work executes)
4249 * Executes the function immediately if process context is available,
4250 * otherwise schedules the function for delayed execution.
4252 * Return: 0 - function was executed
4253 * 1 - function was scheduled for execution
4255 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4257 if (!in_interrupt()) {
4262 INIT_WORK(&ew->work, fn);
4263 schedule_work(&ew->work);
4267 EXPORT_SYMBOL_GPL(execute_in_process_context);
4270 * free_workqueue_attrs - free a workqueue_attrs
4271 * @attrs: workqueue_attrs to free
4273 * Undo alloc_workqueue_attrs().
4275 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4278 free_cpumask_var(attrs->cpumask);
4279 free_cpumask_var(attrs->__pod_cpumask);
4285 * alloc_workqueue_attrs - allocate a workqueue_attrs
4287 * Allocate a new workqueue_attrs, initialize with default settings and
4290 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4292 struct workqueue_attrs *alloc_workqueue_attrs(void)
4294 struct workqueue_attrs *attrs;
4296 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4299 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4301 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4304 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4305 attrs->affn_scope = WQ_AFFN_DFL;
4308 free_workqueue_attrs(attrs);
4312 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4313 const struct workqueue_attrs *from)
4315 to->nice = from->nice;
4316 cpumask_copy(to->cpumask, from->cpumask);
4317 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4318 to->affn_strict = from->affn_strict;
4321 * Unlike hash and equality test, copying shouldn't ignore wq-only
4322 * fields as copying is used for both pool and wq attrs. Instead,
4323 * get_unbound_pool() explicitly clears the fields.
4325 to->affn_scope = from->affn_scope;
4326 to->ordered = from->ordered;
4330 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4331 * comments in 'struct workqueue_attrs' definition.
4333 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4335 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4336 attrs->ordered = false;
4339 /* hash value of the content of @attr */
4340 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4344 hash = jhash_1word(attrs->nice, hash);
4345 hash = jhash(cpumask_bits(attrs->cpumask),
4346 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4347 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4348 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4349 hash = jhash_1word(attrs->affn_strict, hash);
4353 /* content equality test */
4354 static bool wqattrs_equal(const struct workqueue_attrs *a,
4355 const struct workqueue_attrs *b)
4357 if (a->nice != b->nice)
4359 if (!cpumask_equal(a->cpumask, b->cpumask))
4361 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4363 if (a->affn_strict != b->affn_strict)
4368 /* Update @attrs with actually available CPUs */
4369 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4370 const cpumask_t *unbound_cpumask)
4373 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4374 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4377 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4378 if (unlikely(cpumask_empty(attrs->cpumask)))
4379 cpumask_copy(attrs->cpumask, unbound_cpumask);
4382 /* find wq_pod_type to use for @attrs */
4383 static const struct wq_pod_type *
4384 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4386 enum wq_affn_scope scope;
4387 struct wq_pod_type *pt;
4389 /* to synchronize access to wq_affn_dfl */
4390 lockdep_assert_held(&wq_pool_mutex);
4392 if (attrs->affn_scope == WQ_AFFN_DFL)
4393 scope = wq_affn_dfl;
4395 scope = attrs->affn_scope;
4397 pt = &wq_pod_types[scope];
4399 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4400 likely(pt->nr_pods))
4404 * Before workqueue_init_topology(), only SYSTEM is available which is
4405 * initialized in workqueue_init_early().
4407 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4408 BUG_ON(!pt->nr_pods);
4413 * init_worker_pool - initialize a newly zalloc'd worker_pool
4414 * @pool: worker_pool to initialize
4416 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4418 * Return: 0 on success, -errno on failure. Even on failure, all fields
4419 * inside @pool proper are initialized and put_unbound_pool() can be called
4420 * on @pool safely to release it.
4422 static int init_worker_pool(struct worker_pool *pool)
4424 raw_spin_lock_init(&pool->lock);
4427 pool->node = NUMA_NO_NODE;
4428 pool->flags |= POOL_DISASSOCIATED;
4429 pool->watchdog_ts = jiffies;
4430 INIT_LIST_HEAD(&pool->worklist);
4431 INIT_LIST_HEAD(&pool->idle_list);
4432 hash_init(pool->busy_hash);
4434 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4435 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4437 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4439 INIT_LIST_HEAD(&pool->workers);
4440 INIT_LIST_HEAD(&pool->dying_workers);
4442 ida_init(&pool->worker_ida);
4443 INIT_HLIST_NODE(&pool->hash_node);
4446 /* shouldn't fail above this point */
4447 pool->attrs = alloc_workqueue_attrs();
4451 wqattrs_clear_for_pool(pool->attrs);
4456 #ifdef CONFIG_LOCKDEP
4457 static void wq_init_lockdep(struct workqueue_struct *wq)
4461 lockdep_register_key(&wq->key);
4462 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4464 lock_name = wq->name;
4466 wq->lock_name = lock_name;
4467 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4470 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4472 lockdep_unregister_key(&wq->key);
4475 static void wq_free_lockdep(struct workqueue_struct *wq)
4477 if (wq->lock_name != wq->name)
4478 kfree(wq->lock_name);
4481 static void wq_init_lockdep(struct workqueue_struct *wq)
4485 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4489 static void wq_free_lockdep(struct workqueue_struct *wq)
4494 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4498 for_each_node(node) {
4499 kfree(nna_ar[node]);
4500 nna_ar[node] = NULL;
4503 kfree(nna_ar[nr_node_ids]);
4504 nna_ar[nr_node_ids] = NULL;
4507 static void init_node_nr_active(struct wq_node_nr_active *nna)
4509 nna->max = WQ_DFL_MIN_ACTIVE;
4510 atomic_set(&nna->nr, 0);
4511 raw_spin_lock_init(&nna->lock);
4512 INIT_LIST_HEAD(&nna->pending_pwqs);
4516 * Each node's nr_active counter will be accessed mostly from its own node and
4517 * should be allocated in the node.
4519 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4521 struct wq_node_nr_active *nna;
4524 for_each_node(node) {
4525 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4528 init_node_nr_active(nna);
4532 /* [nr_node_ids] is used as the fallback */
4533 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4536 init_node_nr_active(nna);
4537 nna_ar[nr_node_ids] = nna;
4542 free_node_nr_active(nna_ar);
4546 static void rcu_free_wq(struct rcu_head *rcu)
4548 struct workqueue_struct *wq =
4549 container_of(rcu, struct workqueue_struct, rcu);
4551 if (wq->flags & WQ_UNBOUND)
4552 free_node_nr_active(wq->node_nr_active);
4554 wq_free_lockdep(wq);
4555 free_percpu(wq->cpu_pwq);
4556 free_workqueue_attrs(wq->unbound_attrs);
4560 static void rcu_free_pool(struct rcu_head *rcu)
4562 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4564 ida_destroy(&pool->worker_ida);
4565 free_workqueue_attrs(pool->attrs);
4570 * put_unbound_pool - put a worker_pool
4571 * @pool: worker_pool to put
4573 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4574 * safe manner. get_unbound_pool() calls this function on its failure path
4575 * and this function should be able to release pools which went through,
4576 * successfully or not, init_worker_pool().
4578 * Should be called with wq_pool_mutex held.
4580 static void put_unbound_pool(struct worker_pool *pool)
4582 DECLARE_COMPLETION_ONSTACK(detach_completion);
4583 struct worker *worker;
4584 LIST_HEAD(cull_list);
4586 lockdep_assert_held(&wq_pool_mutex);
4592 if (WARN_ON(!(pool->cpu < 0)) ||
4593 WARN_ON(!list_empty(&pool->worklist)))
4596 /* release id and unhash */
4598 idr_remove(&worker_pool_idr, pool->id);
4599 hash_del(&pool->hash_node);
4602 * Become the manager and destroy all workers. This prevents
4603 * @pool's workers from blocking on attach_mutex. We're the last
4604 * manager and @pool gets freed with the flag set.
4606 * Having a concurrent manager is quite unlikely to happen as we can
4607 * only get here with
4608 * pwq->refcnt == pool->refcnt == 0
4609 * which implies no work queued to the pool, which implies no worker can
4610 * become the manager. However a worker could have taken the role of
4611 * manager before the refcnts dropped to 0, since maybe_create_worker()
4615 rcuwait_wait_event(&manager_wait,
4616 !(pool->flags & POOL_MANAGER_ACTIVE),
4617 TASK_UNINTERRUPTIBLE);
4619 mutex_lock(&wq_pool_attach_mutex);
4620 raw_spin_lock_irq(&pool->lock);
4621 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4622 pool->flags |= POOL_MANAGER_ACTIVE;
4625 raw_spin_unlock_irq(&pool->lock);
4626 mutex_unlock(&wq_pool_attach_mutex);
4629 while ((worker = first_idle_worker(pool)))
4630 set_worker_dying(worker, &cull_list);
4631 WARN_ON(pool->nr_workers || pool->nr_idle);
4632 raw_spin_unlock_irq(&pool->lock);
4634 wake_dying_workers(&cull_list);
4636 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4637 pool->detach_completion = &detach_completion;
4638 mutex_unlock(&wq_pool_attach_mutex);
4640 if (pool->detach_completion)
4641 wait_for_completion(pool->detach_completion);
4643 /* shut down the timers */
4644 del_timer_sync(&pool->idle_timer);
4645 cancel_work_sync(&pool->idle_cull_work);
4646 del_timer_sync(&pool->mayday_timer);
4648 /* RCU protected to allow dereferences from get_work_pool() */
4649 call_rcu(&pool->rcu, rcu_free_pool);
4653 * get_unbound_pool - get a worker_pool with the specified attributes
4654 * @attrs: the attributes of the worker_pool to get
4656 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4657 * reference count and return it. If there already is a matching
4658 * worker_pool, it will be used; otherwise, this function attempts to
4661 * Should be called with wq_pool_mutex held.
4663 * Return: On success, a worker_pool with the same attributes as @attrs.
4664 * On failure, %NULL.
4666 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4668 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4669 u32 hash = wqattrs_hash(attrs);
4670 struct worker_pool *pool;
4671 int pod, node = NUMA_NO_NODE;
4673 lockdep_assert_held(&wq_pool_mutex);
4675 /* do we already have a matching pool? */
4676 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4677 if (wqattrs_equal(pool->attrs, attrs)) {
4683 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4684 for (pod = 0; pod < pt->nr_pods; pod++) {
4685 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4686 node = pt->pod_node[pod];
4691 /* nope, create a new one */
4692 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4693 if (!pool || init_worker_pool(pool) < 0)
4697 copy_workqueue_attrs(pool->attrs, attrs);
4698 wqattrs_clear_for_pool(pool->attrs);
4700 if (worker_pool_assign_id(pool) < 0)
4703 /* create and start the initial worker */
4704 if (wq_online && !create_worker(pool))
4708 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4713 put_unbound_pool(pool);
4717 static void rcu_free_pwq(struct rcu_head *rcu)
4719 kmem_cache_free(pwq_cache,
4720 container_of(rcu, struct pool_workqueue, rcu));
4724 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4725 * refcnt and needs to be destroyed.
4727 static void pwq_release_workfn(struct kthread_work *work)
4729 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4731 struct workqueue_struct *wq = pwq->wq;
4732 struct worker_pool *pool = pwq->pool;
4733 bool is_last = false;
4736 * When @pwq is not linked, it doesn't hold any reference to the
4737 * @wq, and @wq is invalid to access.
4739 if (!list_empty(&pwq->pwqs_node)) {
4740 mutex_lock(&wq->mutex);
4741 list_del_rcu(&pwq->pwqs_node);
4742 is_last = list_empty(&wq->pwqs);
4743 mutex_unlock(&wq->mutex);
4746 if (wq->flags & WQ_UNBOUND) {
4747 mutex_lock(&wq_pool_mutex);
4748 put_unbound_pool(pool);
4749 mutex_unlock(&wq_pool_mutex);
4752 if (!list_empty(&pwq->pending_node)) {
4753 struct wq_node_nr_active *nna =
4754 wq_node_nr_active(pwq->wq, pwq->pool->node);
4756 raw_spin_lock_irq(&nna->lock);
4757 list_del_init(&pwq->pending_node);
4758 raw_spin_unlock_irq(&nna->lock);
4761 call_rcu(&pwq->rcu, rcu_free_pwq);
4764 * If we're the last pwq going away, @wq is already dead and no one
4765 * is gonna access it anymore. Schedule RCU free.
4768 wq_unregister_lockdep(wq);
4769 call_rcu(&wq->rcu, rcu_free_wq);
4773 /* initialize newly allocated @pwq which is associated with @wq and @pool */
4774 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4775 struct worker_pool *pool)
4777 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4779 memset(pwq, 0, sizeof(*pwq));
4783 pwq->flush_color = -1;
4785 INIT_LIST_HEAD(&pwq->inactive_works);
4786 INIT_LIST_HEAD(&pwq->pending_node);
4787 INIT_LIST_HEAD(&pwq->pwqs_node);
4788 INIT_LIST_HEAD(&pwq->mayday_node);
4789 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4792 /* sync @pwq with the current state of its associated wq and link it */
4793 static void link_pwq(struct pool_workqueue *pwq)
4795 struct workqueue_struct *wq = pwq->wq;
4797 lockdep_assert_held(&wq->mutex);
4799 /* may be called multiple times, ignore if already linked */
4800 if (!list_empty(&pwq->pwqs_node))
4803 /* set the matching work_color */
4804 pwq->work_color = wq->work_color;
4807 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4810 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4811 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4812 const struct workqueue_attrs *attrs)
4814 struct worker_pool *pool;
4815 struct pool_workqueue *pwq;
4817 lockdep_assert_held(&wq_pool_mutex);
4819 pool = get_unbound_pool(attrs);
4823 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4825 put_unbound_pool(pool);
4829 init_pwq(pwq, wq, pool);
4834 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4835 * @attrs: the wq_attrs of the default pwq of the target workqueue
4836 * @cpu: the target CPU
4837 * @cpu_going_down: if >= 0, the CPU to consider as offline
4839 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4840 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4841 * The result is stored in @attrs->__pod_cpumask.
4843 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4844 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4845 * intersection of the possible CPUs of @pod and @attrs->cpumask.
4847 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4849 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4852 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4853 int pod = pt->cpu_pod[cpu];
4855 /* does @pod have any online CPUs @attrs wants? */
4856 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4857 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4858 if (cpu_going_down >= 0)
4859 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4861 if (cpumask_empty(attrs->__pod_cpumask)) {
4862 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4866 /* yeap, return possible CPUs in @pod that @attrs wants */
4867 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4869 if (cpumask_empty(attrs->__pod_cpumask))
4870 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4871 "possible intersect\n");
4874 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
4875 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4876 int cpu, struct pool_workqueue *pwq)
4878 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
4879 struct pool_workqueue *old_pwq;
4881 lockdep_assert_held(&wq_pool_mutex);
4882 lockdep_assert_held(&wq->mutex);
4884 /* link_pwq() can handle duplicate calls */
4887 old_pwq = rcu_access_pointer(*slot);
4888 rcu_assign_pointer(*slot, pwq);
4892 /* context to store the prepared attrs & pwqs before applying */
4893 struct apply_wqattrs_ctx {
4894 struct workqueue_struct *wq; /* target workqueue */
4895 struct workqueue_attrs *attrs; /* attrs to apply */
4896 struct list_head list; /* queued for batching commit */
4897 struct pool_workqueue *dfl_pwq;
4898 struct pool_workqueue *pwq_tbl[];
4901 /* free the resources after success or abort */
4902 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4907 for_each_possible_cpu(cpu)
4908 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4909 put_pwq_unlocked(ctx->dfl_pwq);
4911 free_workqueue_attrs(ctx->attrs);
4917 /* allocate the attrs and pwqs for later installation */
4918 static struct apply_wqattrs_ctx *
4919 apply_wqattrs_prepare(struct workqueue_struct *wq,
4920 const struct workqueue_attrs *attrs,
4921 const cpumask_var_t unbound_cpumask)
4923 struct apply_wqattrs_ctx *ctx;
4924 struct workqueue_attrs *new_attrs;
4927 lockdep_assert_held(&wq_pool_mutex);
4929 if (WARN_ON(attrs->affn_scope < 0 ||
4930 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4931 return ERR_PTR(-EINVAL);
4933 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4935 new_attrs = alloc_workqueue_attrs();
4936 if (!ctx || !new_attrs)
4940 * If something goes wrong during CPU up/down, we'll fall back to
4941 * the default pwq covering whole @attrs->cpumask. Always create
4942 * it even if we don't use it immediately.
4944 copy_workqueue_attrs(new_attrs, attrs);
4945 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4946 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4947 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4951 for_each_possible_cpu(cpu) {
4952 if (new_attrs->ordered) {
4953 ctx->dfl_pwq->refcnt++;
4954 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4956 wq_calc_pod_cpumask(new_attrs, cpu, -1);
4957 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4958 if (!ctx->pwq_tbl[cpu])
4963 /* save the user configured attrs and sanitize it. */
4964 copy_workqueue_attrs(new_attrs, attrs);
4965 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4966 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4967 ctx->attrs = new_attrs;
4973 free_workqueue_attrs(new_attrs);
4974 apply_wqattrs_cleanup(ctx);
4975 return ERR_PTR(-ENOMEM);
4978 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4979 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4983 /* all pwqs have been created successfully, let's install'em */
4984 mutex_lock(&ctx->wq->mutex);
4986 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4988 /* save the previous pwqs and install the new ones */
4989 for_each_possible_cpu(cpu)
4990 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4992 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
4994 /* update node_nr_active->max */
4995 wq_update_node_max_active(ctx->wq, -1);
4997 mutex_unlock(&ctx->wq->mutex);
5000 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5001 const struct workqueue_attrs *attrs)
5003 struct apply_wqattrs_ctx *ctx;
5005 /* only unbound workqueues can change attributes */
5006 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5009 /* creating multiple pwqs breaks ordering guarantee */
5010 if (!list_empty(&wq->pwqs) && WARN_ON(wq->flags & __WQ_ORDERED))
5013 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5015 return PTR_ERR(ctx);
5017 /* the ctx has been prepared successfully, let's commit it */
5018 apply_wqattrs_commit(ctx);
5019 apply_wqattrs_cleanup(ctx);
5025 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5026 * @wq: the target workqueue
5027 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5029 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5030 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5031 * work items are affine to the pod it was issued on. Older pwqs are released as
5032 * in-flight work items finish. Note that a work item which repeatedly requeues
5033 * itself back-to-back will stay on its current pwq.
5035 * Performs GFP_KERNEL allocations.
5037 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
5039 * Return: 0 on success and -errno on failure.
5041 int apply_workqueue_attrs(struct workqueue_struct *wq,
5042 const struct workqueue_attrs *attrs)
5046 lockdep_assert_cpus_held();
5048 mutex_lock(&wq_pool_mutex);
5049 ret = apply_workqueue_attrs_locked(wq, attrs);
5050 mutex_unlock(&wq_pool_mutex);
5056 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
5057 * @wq: the target workqueue
5058 * @cpu: the CPU to update pool association for
5059 * @hotplug_cpu: the CPU coming up or going down
5060 * @online: whether @cpu is coming up or going down
5062 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5063 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
5067 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5068 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5070 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5071 * with a cpumask spanning multiple pods, the workers which were already
5072 * executing the work items for the workqueue will lose their CPU affinity and
5073 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5074 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5075 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5077 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
5078 int hotplug_cpu, bool online)
5080 int off_cpu = online ? -1 : hotplug_cpu;
5081 struct pool_workqueue *old_pwq = NULL, *pwq;
5082 struct workqueue_attrs *target_attrs;
5084 lockdep_assert_held(&wq_pool_mutex);
5086 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5090 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5091 * Let's use a preallocated one. The following buf is protected by
5092 * CPU hotplug exclusion.
5094 target_attrs = wq_update_pod_attrs_buf;
5096 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5097 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5099 /* nothing to do if the target cpumask matches the current pwq */
5100 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
5101 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5104 /* create a new pwq */
5105 pwq = alloc_unbound_pwq(wq, target_attrs);
5107 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5112 /* Install the new pwq. */
5113 mutex_lock(&wq->mutex);
5114 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5118 mutex_lock(&wq->mutex);
5119 pwq = unbound_pwq(wq, -1);
5120 raw_spin_lock_irq(&pwq->pool->lock);
5122 raw_spin_unlock_irq(&pwq->pool->lock);
5123 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5125 mutex_unlock(&wq->mutex);
5126 put_pwq_unlocked(old_pwq);
5129 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5131 bool highpri = wq->flags & WQ_HIGHPRI;
5134 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5138 if (!(wq->flags & WQ_UNBOUND)) {
5139 for_each_possible_cpu(cpu) {
5140 struct pool_workqueue **pwq_p;
5141 struct worker_pool __percpu *pools;
5142 struct worker_pool *pool;
5144 if (wq->flags & WQ_BH)
5145 pools = bh_worker_pools;
5147 pools = cpu_worker_pools;
5149 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5150 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5152 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5157 init_pwq(*pwq_p, wq, pool);
5159 mutex_lock(&wq->mutex);
5161 mutex_unlock(&wq->mutex);
5167 if (wq->flags & __WQ_ORDERED) {
5168 struct pool_workqueue *dfl_pwq;
5170 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5171 /* there should only be single pwq for ordering guarantee */
5172 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5173 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5174 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5175 "ordering guarantee broken for workqueue %s\n", wq->name);
5177 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5181 /* for unbound pwq, flush the pwq_release_worker ensures that the
5182 * pwq_release_workfn() completes before calling kfree(wq).
5185 kthread_flush_worker(pwq_release_worker);
5191 for_each_possible_cpu(cpu) {
5192 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5195 kmem_cache_free(pwq_cache, pwq);
5197 free_percpu(wq->cpu_pwq);
5203 static int wq_clamp_max_active(int max_active, unsigned int flags,
5206 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5207 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5208 max_active, name, 1, WQ_MAX_ACTIVE);
5210 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5214 * Workqueues which may be used during memory reclaim should have a rescuer
5215 * to guarantee forward progress.
5217 static int init_rescuer(struct workqueue_struct *wq)
5219 struct worker *rescuer;
5222 if (!(wq->flags & WQ_MEM_RECLAIM))
5225 rescuer = alloc_worker(NUMA_NO_NODE);
5227 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5232 rescuer->rescue_wq = wq;
5233 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5234 if (IS_ERR(rescuer->task)) {
5235 ret = PTR_ERR(rescuer->task);
5236 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5237 wq->name, ERR_PTR(ret));
5242 wq->rescuer = rescuer;
5243 if (wq->flags & WQ_UNBOUND)
5244 kthread_bind_mask(rescuer->task, wq->unbound_attrs->cpumask);
5246 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5247 wake_up_process(rescuer->task);
5253 * wq_adjust_max_active - update a wq's max_active to the current setting
5254 * @wq: target workqueue
5256 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5257 * activate inactive work items accordingly. If @wq is freezing, clear
5258 * @wq->max_active to zero.
5260 static void wq_adjust_max_active(struct workqueue_struct *wq)
5263 int new_max, new_min;
5265 lockdep_assert_held(&wq->mutex);
5267 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5271 new_max = wq->saved_max_active;
5272 new_min = wq->saved_min_active;
5275 if (wq->max_active == new_max && wq->min_active == new_min)
5279 * Update @wq->max/min_active and then kick inactive work items if more
5280 * active work items are allowed. This doesn't break work item ordering
5281 * because new work items are always queued behind existing inactive
5282 * work items if there are any.
5284 WRITE_ONCE(wq->max_active, new_max);
5285 WRITE_ONCE(wq->min_active, new_min);
5287 if (wq->flags & WQ_UNBOUND)
5288 wq_update_node_max_active(wq, -1);
5294 * Round-robin through pwq's activating the first inactive work item
5295 * until max_active is filled.
5298 struct pool_workqueue *pwq;
5301 for_each_pwq(pwq, wq) {
5302 unsigned long flags;
5304 /* can be called during early boot w/ irq disabled */
5305 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5306 if (pwq_activate_first_inactive(pwq, true)) {
5308 kick_pool(pwq->pool);
5310 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5312 } while (activated);
5316 struct workqueue_struct *alloc_workqueue(const char *fmt,
5318 int max_active, ...)
5321 struct workqueue_struct *wq;
5325 if (flags & WQ_BH) {
5326 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5328 if (WARN_ON_ONCE(max_active))
5332 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5333 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5334 flags |= WQ_UNBOUND;
5336 /* allocate wq and format name */
5337 if (flags & WQ_UNBOUND)
5338 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5340 wq_size = sizeof(*wq);
5342 wq = kzalloc(wq_size, GFP_KERNEL);
5346 if (flags & WQ_UNBOUND) {
5347 wq->unbound_attrs = alloc_workqueue_attrs();
5348 if (!wq->unbound_attrs)
5352 va_start(args, max_active);
5353 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5356 if (name_len >= WQ_NAME_LEN)
5357 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5360 if (flags & WQ_BH) {
5362 * BH workqueues always share a single execution context per CPU
5363 * and don't impose any max_active limit.
5365 max_active = INT_MAX;
5367 max_active = max_active ?: WQ_DFL_ACTIVE;
5368 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5373 wq->max_active = max_active;
5374 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5375 wq->saved_max_active = wq->max_active;
5376 wq->saved_min_active = wq->min_active;
5377 mutex_init(&wq->mutex);
5378 atomic_set(&wq->nr_pwqs_to_flush, 0);
5379 INIT_LIST_HEAD(&wq->pwqs);
5380 INIT_LIST_HEAD(&wq->flusher_queue);
5381 INIT_LIST_HEAD(&wq->flusher_overflow);
5382 INIT_LIST_HEAD(&wq->maydays);
5384 wq_init_lockdep(wq);
5385 INIT_LIST_HEAD(&wq->list);
5387 if (flags & WQ_UNBOUND) {
5388 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5389 goto err_unreg_lockdep;
5392 if (alloc_and_link_pwqs(wq) < 0)
5393 goto err_free_node_nr_active;
5395 if (wq_online && init_rescuer(wq) < 0)
5398 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5402 * wq_pool_mutex protects global freeze state and workqueues list.
5403 * Grab it, adjust max_active and add the new @wq to workqueues
5406 mutex_lock(&wq_pool_mutex);
5408 mutex_lock(&wq->mutex);
5409 wq_adjust_max_active(wq);
5410 mutex_unlock(&wq->mutex);
5412 list_add_tail_rcu(&wq->list, &workqueues);
5414 mutex_unlock(&wq_pool_mutex);
5418 err_free_node_nr_active:
5419 if (wq->flags & WQ_UNBOUND)
5420 free_node_nr_active(wq->node_nr_active);
5422 wq_unregister_lockdep(wq);
5423 wq_free_lockdep(wq);
5425 free_workqueue_attrs(wq->unbound_attrs);
5429 destroy_workqueue(wq);
5432 EXPORT_SYMBOL_GPL(alloc_workqueue);
5434 static bool pwq_busy(struct pool_workqueue *pwq)
5438 for (i = 0; i < WORK_NR_COLORS; i++)
5439 if (pwq->nr_in_flight[i])
5442 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5444 if (!pwq_is_empty(pwq))
5451 * destroy_workqueue - safely terminate a workqueue
5452 * @wq: target workqueue
5454 * Safely destroy a workqueue. All work currently pending will be done first.
5456 void destroy_workqueue(struct workqueue_struct *wq)
5458 struct pool_workqueue *pwq;
5462 * Remove it from sysfs first so that sanity check failure doesn't
5463 * lead to sysfs name conflicts.
5465 workqueue_sysfs_unregister(wq);
5467 /* mark the workqueue destruction is in progress */
5468 mutex_lock(&wq->mutex);
5469 wq->flags |= __WQ_DESTROYING;
5470 mutex_unlock(&wq->mutex);
5472 /* drain it before proceeding with destruction */
5473 drain_workqueue(wq);
5475 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5477 struct worker *rescuer = wq->rescuer;
5479 /* this prevents new queueing */
5480 raw_spin_lock_irq(&wq_mayday_lock);
5482 raw_spin_unlock_irq(&wq_mayday_lock);
5484 /* rescuer will empty maydays list before exiting */
5485 kthread_stop(rescuer->task);
5490 * Sanity checks - grab all the locks so that we wait for all
5491 * in-flight operations which may do put_pwq().
5493 mutex_lock(&wq_pool_mutex);
5494 mutex_lock(&wq->mutex);
5495 for_each_pwq(pwq, wq) {
5496 raw_spin_lock_irq(&pwq->pool->lock);
5497 if (WARN_ON(pwq_busy(pwq))) {
5498 pr_warn("%s: %s has the following busy pwq\n",
5499 __func__, wq->name);
5501 raw_spin_unlock_irq(&pwq->pool->lock);
5502 mutex_unlock(&wq->mutex);
5503 mutex_unlock(&wq_pool_mutex);
5504 show_one_workqueue(wq);
5507 raw_spin_unlock_irq(&pwq->pool->lock);
5509 mutex_unlock(&wq->mutex);
5512 * wq list is used to freeze wq, remove from list after
5513 * flushing is complete in case freeze races us.
5515 list_del_rcu(&wq->list);
5516 mutex_unlock(&wq_pool_mutex);
5519 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5520 * to put the base refs. @wq will be auto-destroyed from the last
5521 * pwq_put. RCU read lock prevents @wq from going away from under us.
5525 for_each_possible_cpu(cpu) {
5526 put_pwq_unlocked(unbound_pwq(wq, cpu));
5527 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5530 put_pwq_unlocked(unbound_pwq(wq, -1));
5531 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5535 EXPORT_SYMBOL_GPL(destroy_workqueue);
5538 * workqueue_set_max_active - adjust max_active of a workqueue
5539 * @wq: target workqueue
5540 * @max_active: new max_active value.
5542 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5546 * Don't call from IRQ context.
5548 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5550 /* max_active doesn't mean anything for BH workqueues */
5551 if (WARN_ON(wq->flags & WQ_BH))
5553 /* disallow meddling with max_active for ordered workqueues */
5554 if (WARN_ON(wq->flags & __WQ_ORDERED))
5557 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5559 mutex_lock(&wq->mutex);
5561 wq->saved_max_active = max_active;
5562 if (wq->flags & WQ_UNBOUND)
5563 wq->saved_min_active = min(wq->saved_min_active, max_active);
5565 wq_adjust_max_active(wq);
5567 mutex_unlock(&wq->mutex);
5569 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5572 * current_work - retrieve %current task's work struct
5574 * Determine if %current task is a workqueue worker and what it's working on.
5575 * Useful to find out the context that the %current task is running in.
5577 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5579 struct work_struct *current_work(void)
5581 struct worker *worker = current_wq_worker();
5583 return worker ? worker->current_work : NULL;
5585 EXPORT_SYMBOL(current_work);
5588 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5590 * Determine whether %current is a workqueue rescuer. Can be used from
5591 * work functions to determine whether it's being run off the rescuer task.
5593 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5595 bool current_is_workqueue_rescuer(void)
5597 struct worker *worker = current_wq_worker();
5599 return worker && worker->rescue_wq;
5603 * workqueue_congested - test whether a workqueue is congested
5604 * @cpu: CPU in question
5605 * @wq: target workqueue
5607 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5608 * no synchronization around this function and the test result is
5609 * unreliable and only useful as advisory hints or for debugging.
5611 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5613 * With the exception of ordered workqueues, all workqueues have per-cpu
5614 * pool_workqueues, each with its own congested state. A workqueue being
5615 * congested on one CPU doesn't mean that the workqueue is contested on any
5619 * %true if congested, %false otherwise.
5621 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5623 struct pool_workqueue *pwq;
5629 if (cpu == WORK_CPU_UNBOUND)
5630 cpu = smp_processor_id();
5632 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5633 ret = !list_empty(&pwq->inactive_works);
5640 EXPORT_SYMBOL_GPL(workqueue_congested);
5643 * work_busy - test whether a work is currently pending or running
5644 * @work: the work to be tested
5646 * Test whether @work is currently pending or running. There is no
5647 * synchronization around this function and the test result is
5648 * unreliable and only useful as advisory hints or for debugging.
5651 * OR'd bitmask of WORK_BUSY_* bits.
5653 unsigned int work_busy(struct work_struct *work)
5655 struct worker_pool *pool;
5656 unsigned long flags;
5657 unsigned int ret = 0;
5659 if (work_pending(work))
5660 ret |= WORK_BUSY_PENDING;
5663 pool = get_work_pool(work);
5665 raw_spin_lock_irqsave(&pool->lock, flags);
5666 if (find_worker_executing_work(pool, work))
5667 ret |= WORK_BUSY_RUNNING;
5668 raw_spin_unlock_irqrestore(&pool->lock, flags);
5674 EXPORT_SYMBOL_GPL(work_busy);
5677 * set_worker_desc - set description for the current work item
5678 * @fmt: printf-style format string
5679 * @...: arguments for the format string
5681 * This function can be called by a running work function to describe what
5682 * the work item is about. If the worker task gets dumped, this
5683 * information will be printed out together to help debugging. The
5684 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5686 void set_worker_desc(const char *fmt, ...)
5688 struct worker *worker = current_wq_worker();
5692 va_start(args, fmt);
5693 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5697 EXPORT_SYMBOL_GPL(set_worker_desc);
5700 * print_worker_info - print out worker information and description
5701 * @log_lvl: the log level to use when printing
5702 * @task: target task
5704 * If @task is a worker and currently executing a work item, print out the
5705 * name of the workqueue being serviced and worker description set with
5706 * set_worker_desc() by the currently executing work item.
5708 * This function can be safely called on any task as long as the
5709 * task_struct itself is accessible. While safe, this function isn't
5710 * synchronized and may print out mixups or garbages of limited length.
5712 void print_worker_info(const char *log_lvl, struct task_struct *task)
5714 work_func_t *fn = NULL;
5715 char name[WQ_NAME_LEN] = { };
5716 char desc[WORKER_DESC_LEN] = { };
5717 struct pool_workqueue *pwq = NULL;
5718 struct workqueue_struct *wq = NULL;
5719 struct worker *worker;
5721 if (!(task->flags & PF_WQ_WORKER))
5725 * This function is called without any synchronization and @task
5726 * could be in any state. Be careful with dereferences.
5728 worker = kthread_probe_data(task);
5731 * Carefully copy the associated workqueue's workfn, name and desc.
5732 * Keep the original last '\0' in case the original is garbage.
5734 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5735 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5736 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5737 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5738 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5740 if (fn || name[0] || desc[0]) {
5741 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5742 if (strcmp(name, desc))
5743 pr_cont(" (%s)", desc);
5748 static void pr_cont_pool_info(struct worker_pool *pool)
5750 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5751 if (pool->node != NUMA_NO_NODE)
5752 pr_cont(" node=%d", pool->node);
5753 pr_cont(" flags=0x%x", pool->flags);
5754 if (pool->flags & POOL_BH)
5756 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
5758 pr_cont(" nice=%d", pool->attrs->nice);
5761 static void pr_cont_worker_id(struct worker *worker)
5763 struct worker_pool *pool = worker->pool;
5765 if (pool->flags & WQ_BH)
5767 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
5769 pr_cont("%d%s", task_pid_nr(worker->task),
5770 worker->rescue_wq ? "(RESCUER)" : "");
5773 struct pr_cont_work_struct {
5779 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5783 if (func == pcwsp->func) {
5787 if (pcwsp->ctr == 1)
5788 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5790 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5793 if ((long)func == -1L)
5795 pcwsp->comma = comma;
5800 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5802 if (work->func == wq_barrier_func) {
5803 struct wq_barrier *barr;
5805 barr = container_of(work, struct wq_barrier, work);
5807 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5808 pr_cont("%s BAR(%d)", comma ? "," : "",
5809 task_pid_nr(barr->task));
5812 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5813 pr_cont_work_flush(comma, work->func, pcwsp);
5817 static void show_pwq(struct pool_workqueue *pwq)
5819 struct pr_cont_work_struct pcws = { .ctr = 0, };
5820 struct worker_pool *pool = pwq->pool;
5821 struct work_struct *work;
5822 struct worker *worker;
5823 bool has_in_flight = false, has_pending = false;
5826 pr_info(" pwq %d:", pool->id);
5827 pr_cont_pool_info(pool);
5829 pr_cont(" active=%d refcnt=%d%s\n",
5830 pwq->nr_active, pwq->refcnt,
5831 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5833 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5834 if (worker->current_pwq == pwq) {
5835 has_in_flight = true;
5839 if (has_in_flight) {
5842 pr_info(" in-flight:");
5843 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5844 if (worker->current_pwq != pwq)
5847 pr_cont(" %s", comma ? "," : "");
5848 pr_cont_worker_id(worker);
5849 pr_cont(":%ps", worker->current_func);
5850 list_for_each_entry(work, &worker->scheduled, entry)
5851 pr_cont_work(false, work, &pcws);
5852 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5858 list_for_each_entry(work, &pool->worklist, entry) {
5859 if (get_work_pwq(work) == pwq) {
5867 pr_info(" pending:");
5868 list_for_each_entry(work, &pool->worklist, entry) {
5869 if (get_work_pwq(work) != pwq)
5872 pr_cont_work(comma, work, &pcws);
5873 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5875 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5879 if (!list_empty(&pwq->inactive_works)) {
5882 pr_info(" inactive:");
5883 list_for_each_entry(work, &pwq->inactive_works, entry) {
5884 pr_cont_work(comma, work, &pcws);
5885 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5887 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5893 * show_one_workqueue - dump state of specified workqueue
5894 * @wq: workqueue whose state will be printed
5896 void show_one_workqueue(struct workqueue_struct *wq)
5898 struct pool_workqueue *pwq;
5900 unsigned long flags;
5902 for_each_pwq(pwq, wq) {
5903 if (!pwq_is_empty(pwq)) {
5908 if (idle) /* Nothing to print for idle workqueue */
5911 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5913 for_each_pwq(pwq, wq) {
5914 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5915 if (!pwq_is_empty(pwq)) {
5917 * Defer printing to avoid deadlocks in console
5918 * drivers that queue work while holding locks
5919 * also taken in their write paths.
5921 printk_deferred_enter();
5923 printk_deferred_exit();
5925 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5927 * We could be printing a lot from atomic context, e.g.
5928 * sysrq-t -> show_all_workqueues(). Avoid triggering
5931 touch_nmi_watchdog();
5937 * show_one_worker_pool - dump state of specified worker pool
5938 * @pool: worker pool whose state will be printed
5940 static void show_one_worker_pool(struct worker_pool *pool)
5942 struct worker *worker;
5944 unsigned long flags;
5945 unsigned long hung = 0;
5947 raw_spin_lock_irqsave(&pool->lock, flags);
5948 if (pool->nr_workers == pool->nr_idle)
5951 /* How long the first pending work is waiting for a worker. */
5952 if (!list_empty(&pool->worklist))
5953 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5956 * Defer printing to avoid deadlocks in console drivers that
5957 * queue work while holding locks also taken in their write
5960 printk_deferred_enter();
5961 pr_info("pool %d:", pool->id);
5962 pr_cont_pool_info(pool);
5963 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5965 pr_cont(" manager: %d",
5966 task_pid_nr(pool->manager->task));
5967 list_for_each_entry(worker, &pool->idle_list, entry) {
5968 pr_cont(" %s", first ? "idle: " : "");
5969 pr_cont_worker_id(worker);
5973 printk_deferred_exit();
5975 raw_spin_unlock_irqrestore(&pool->lock, flags);
5977 * We could be printing a lot from atomic context, e.g.
5978 * sysrq-t -> show_all_workqueues(). Avoid triggering
5981 touch_nmi_watchdog();
5986 * show_all_workqueues - dump workqueue state
5988 * Called from a sysrq handler and prints out all busy workqueues and pools.
5990 void show_all_workqueues(void)
5992 struct workqueue_struct *wq;
5993 struct worker_pool *pool;
5998 pr_info("Showing busy workqueues and worker pools:\n");
6000 list_for_each_entry_rcu(wq, &workqueues, list)
6001 show_one_workqueue(wq);
6003 for_each_pool(pool, pi)
6004 show_one_worker_pool(pool);
6010 * show_freezable_workqueues - dump freezable workqueue state
6012 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6015 void show_freezable_workqueues(void)
6017 struct workqueue_struct *wq;
6021 pr_info("Showing freezable workqueues that are still busy:\n");
6023 list_for_each_entry_rcu(wq, &workqueues, list) {
6024 if (!(wq->flags & WQ_FREEZABLE))
6026 show_one_workqueue(wq);
6032 /* used to show worker information through /proc/PID/{comm,stat,status} */
6033 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6037 /* always show the actual comm */
6038 off = strscpy(buf, task->comm, size);
6042 /* stabilize PF_WQ_WORKER and worker pool association */
6043 mutex_lock(&wq_pool_attach_mutex);
6045 if (task->flags & PF_WQ_WORKER) {
6046 struct worker *worker = kthread_data(task);
6047 struct worker_pool *pool = worker->pool;
6050 raw_spin_lock_irq(&pool->lock);
6052 * ->desc tracks information (wq name or
6053 * set_worker_desc()) for the latest execution. If
6054 * current, prepend '+', otherwise '-'.
6056 if (worker->desc[0] != '\0') {
6057 if (worker->current_work)
6058 scnprintf(buf + off, size - off, "+%s",
6061 scnprintf(buf + off, size - off, "-%s",
6064 raw_spin_unlock_irq(&pool->lock);
6068 mutex_unlock(&wq_pool_attach_mutex);
6076 * There are two challenges in supporting CPU hotplug. Firstly, there
6077 * are a lot of assumptions on strong associations among work, pwq and
6078 * pool which make migrating pending and scheduled works very
6079 * difficult to implement without impacting hot paths. Secondly,
6080 * worker pools serve mix of short, long and very long running works making
6081 * blocked draining impractical.
6083 * This is solved by allowing the pools to be disassociated from the CPU
6084 * running as an unbound one and allowing it to be reattached later if the
6085 * cpu comes back online.
6088 static void unbind_workers(int cpu)
6090 struct worker_pool *pool;
6091 struct worker *worker;
6093 for_each_cpu_worker_pool(pool, cpu) {
6094 mutex_lock(&wq_pool_attach_mutex);
6095 raw_spin_lock_irq(&pool->lock);
6098 * We've blocked all attach/detach operations. Make all workers
6099 * unbound and set DISASSOCIATED. Before this, all workers
6100 * must be on the cpu. After this, they may become diasporas.
6101 * And the preemption disabled section in their sched callbacks
6102 * are guaranteed to see WORKER_UNBOUND since the code here
6103 * is on the same cpu.
6105 for_each_pool_worker(worker, pool)
6106 worker->flags |= WORKER_UNBOUND;
6108 pool->flags |= POOL_DISASSOCIATED;
6111 * The handling of nr_running in sched callbacks are disabled
6112 * now. Zap nr_running. After this, nr_running stays zero and
6113 * need_more_worker() and keep_working() are always true as
6114 * long as the worklist is not empty. This pool now behaves as
6115 * an unbound (in terms of concurrency management) pool which
6116 * are served by workers tied to the pool.
6118 pool->nr_running = 0;
6121 * With concurrency management just turned off, a busy
6122 * worker blocking could lead to lengthy stalls. Kick off
6123 * unbound chain execution of currently pending work items.
6127 raw_spin_unlock_irq(&pool->lock);
6129 for_each_pool_worker(worker, pool)
6130 unbind_worker(worker);
6132 mutex_unlock(&wq_pool_attach_mutex);
6137 * rebind_workers - rebind all workers of a pool to the associated CPU
6138 * @pool: pool of interest
6140 * @pool->cpu is coming online. Rebind all workers to the CPU.
6142 static void rebind_workers(struct worker_pool *pool)
6144 struct worker *worker;
6146 lockdep_assert_held(&wq_pool_attach_mutex);
6149 * Restore CPU affinity of all workers. As all idle workers should
6150 * be on the run-queue of the associated CPU before any local
6151 * wake-ups for concurrency management happen, restore CPU affinity
6152 * of all workers first and then clear UNBOUND. As we're called
6153 * from CPU_ONLINE, the following shouldn't fail.
6155 for_each_pool_worker(worker, pool) {
6156 kthread_set_per_cpu(worker->task, pool->cpu);
6157 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6158 pool_allowed_cpus(pool)) < 0);
6161 raw_spin_lock_irq(&pool->lock);
6163 pool->flags &= ~POOL_DISASSOCIATED;
6165 for_each_pool_worker(worker, pool) {
6166 unsigned int worker_flags = worker->flags;
6169 * We want to clear UNBOUND but can't directly call
6170 * worker_clr_flags() or adjust nr_running. Atomically
6171 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6172 * @worker will clear REBOUND using worker_clr_flags() when
6173 * it initiates the next execution cycle thus restoring
6174 * concurrency management. Note that when or whether
6175 * @worker clears REBOUND doesn't affect correctness.
6177 * WRITE_ONCE() is necessary because @worker->flags may be
6178 * tested without holding any lock in
6179 * wq_worker_running(). Without it, NOT_RUNNING test may
6180 * fail incorrectly leading to premature concurrency
6181 * management operations.
6183 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6184 worker_flags |= WORKER_REBOUND;
6185 worker_flags &= ~WORKER_UNBOUND;
6186 WRITE_ONCE(worker->flags, worker_flags);
6189 raw_spin_unlock_irq(&pool->lock);
6193 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6194 * @pool: unbound pool of interest
6195 * @cpu: the CPU which is coming up
6197 * An unbound pool may end up with a cpumask which doesn't have any online
6198 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6199 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6200 * online CPU before, cpus_allowed of all its workers should be restored.
6202 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6204 static cpumask_t cpumask;
6205 struct worker *worker;
6207 lockdep_assert_held(&wq_pool_attach_mutex);
6209 /* is @cpu allowed for @pool? */
6210 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6213 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6215 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6216 for_each_pool_worker(worker, pool)
6217 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6220 int workqueue_prepare_cpu(unsigned int cpu)
6222 struct worker_pool *pool;
6224 for_each_cpu_worker_pool(pool, cpu) {
6225 if (pool->nr_workers)
6227 if (!create_worker(pool))
6233 int workqueue_online_cpu(unsigned int cpu)
6235 struct worker_pool *pool;
6236 struct workqueue_struct *wq;
6239 mutex_lock(&wq_pool_mutex);
6241 for_each_pool(pool, pi) {
6242 /* BH pools aren't affected by hotplug */
6243 if (pool->flags & POOL_BH)
6246 mutex_lock(&wq_pool_attach_mutex);
6247 if (pool->cpu == cpu)
6248 rebind_workers(pool);
6249 else if (pool->cpu < 0)
6250 restore_unbound_workers_cpumask(pool, cpu);
6251 mutex_unlock(&wq_pool_attach_mutex);
6254 /* update pod affinity of unbound workqueues */
6255 list_for_each_entry(wq, &workqueues, list) {
6256 struct workqueue_attrs *attrs = wq->unbound_attrs;
6259 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6262 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6263 wq_update_pod(wq, tcpu, cpu, true);
6265 mutex_lock(&wq->mutex);
6266 wq_update_node_max_active(wq, -1);
6267 mutex_unlock(&wq->mutex);
6271 mutex_unlock(&wq_pool_mutex);
6275 int workqueue_offline_cpu(unsigned int cpu)
6277 struct workqueue_struct *wq;
6279 /* unbinding per-cpu workers should happen on the local CPU */
6280 if (WARN_ON(cpu != smp_processor_id()))
6283 unbind_workers(cpu);
6285 /* update pod affinity of unbound workqueues */
6286 mutex_lock(&wq_pool_mutex);
6287 list_for_each_entry(wq, &workqueues, list) {
6288 struct workqueue_attrs *attrs = wq->unbound_attrs;
6291 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6294 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6295 wq_update_pod(wq, tcpu, cpu, false);
6297 mutex_lock(&wq->mutex);
6298 wq_update_node_max_active(wq, cpu);
6299 mutex_unlock(&wq->mutex);
6302 mutex_unlock(&wq_pool_mutex);
6307 struct work_for_cpu {
6308 struct work_struct work;
6314 static void work_for_cpu_fn(struct work_struct *work)
6316 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6318 wfc->ret = wfc->fn(wfc->arg);
6322 * work_on_cpu_key - run a function in thread context on a particular cpu
6323 * @cpu: the cpu to run on
6324 * @fn: the function to run
6325 * @arg: the function arg
6326 * @key: The lock class key for lock debugging purposes
6328 * It is up to the caller to ensure that the cpu doesn't go offline.
6329 * The caller must not hold any locks which would prevent @fn from completing.
6331 * Return: The value @fn returns.
6333 long work_on_cpu_key(int cpu, long (*fn)(void *),
6334 void *arg, struct lock_class_key *key)
6336 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6338 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6339 schedule_work_on(cpu, &wfc.work);
6340 flush_work(&wfc.work);
6341 destroy_work_on_stack(&wfc.work);
6344 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6347 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6348 * @cpu: the cpu to run on
6349 * @fn: the function to run
6350 * @arg: the function argument
6351 * @key: The lock class key for lock debugging purposes
6353 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6354 * any locks which would prevent @fn from completing.
6356 * Return: The value @fn returns.
6358 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6359 void *arg, struct lock_class_key *key)
6364 if (cpu_online(cpu))
6365 ret = work_on_cpu_key(cpu, fn, arg, key);
6369 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6370 #endif /* CONFIG_SMP */
6372 #ifdef CONFIG_FREEZER
6375 * freeze_workqueues_begin - begin freezing workqueues
6377 * Start freezing workqueues. After this function returns, all freezable
6378 * workqueues will queue new works to their inactive_works list instead of
6382 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6384 void freeze_workqueues_begin(void)
6386 struct workqueue_struct *wq;
6388 mutex_lock(&wq_pool_mutex);
6390 WARN_ON_ONCE(workqueue_freezing);
6391 workqueue_freezing = true;
6393 list_for_each_entry(wq, &workqueues, list) {
6394 mutex_lock(&wq->mutex);
6395 wq_adjust_max_active(wq);
6396 mutex_unlock(&wq->mutex);
6399 mutex_unlock(&wq_pool_mutex);
6403 * freeze_workqueues_busy - are freezable workqueues still busy?
6405 * Check whether freezing is complete. This function must be called
6406 * between freeze_workqueues_begin() and thaw_workqueues().
6409 * Grabs and releases wq_pool_mutex.
6412 * %true if some freezable workqueues are still busy. %false if freezing
6415 bool freeze_workqueues_busy(void)
6418 struct workqueue_struct *wq;
6419 struct pool_workqueue *pwq;
6421 mutex_lock(&wq_pool_mutex);
6423 WARN_ON_ONCE(!workqueue_freezing);
6425 list_for_each_entry(wq, &workqueues, list) {
6426 if (!(wq->flags & WQ_FREEZABLE))
6429 * nr_active is monotonically decreasing. It's safe
6430 * to peek without lock.
6433 for_each_pwq(pwq, wq) {
6434 WARN_ON_ONCE(pwq->nr_active < 0);
6435 if (pwq->nr_active) {
6444 mutex_unlock(&wq_pool_mutex);
6449 * thaw_workqueues - thaw workqueues
6451 * Thaw workqueues. Normal queueing is restored and all collected
6452 * frozen works are transferred to their respective pool worklists.
6455 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6457 void thaw_workqueues(void)
6459 struct workqueue_struct *wq;
6461 mutex_lock(&wq_pool_mutex);
6463 if (!workqueue_freezing)
6466 workqueue_freezing = false;
6468 /* restore max_active and repopulate worklist */
6469 list_for_each_entry(wq, &workqueues, list) {
6470 mutex_lock(&wq->mutex);
6471 wq_adjust_max_active(wq);
6472 mutex_unlock(&wq->mutex);
6476 mutex_unlock(&wq_pool_mutex);
6478 #endif /* CONFIG_FREEZER */
6480 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6484 struct workqueue_struct *wq;
6485 struct apply_wqattrs_ctx *ctx, *n;
6487 lockdep_assert_held(&wq_pool_mutex);
6489 list_for_each_entry(wq, &workqueues, list) {
6490 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6493 /* creating multiple pwqs breaks ordering guarantee */
6494 if (!list_empty(&wq->pwqs)) {
6495 if (wq->flags & __WQ_ORDERED_EXPLICIT)
6497 wq->flags &= ~__WQ_ORDERED;
6500 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6506 list_add_tail(&ctx->list, &ctxs);
6509 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6511 apply_wqattrs_commit(ctx);
6512 apply_wqattrs_cleanup(ctx);
6516 mutex_lock(&wq_pool_attach_mutex);
6517 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6518 mutex_unlock(&wq_pool_attach_mutex);
6524 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6525 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6527 * This function can be called from cpuset code to provide a set of isolated
6528 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
6529 * either cpus_read_lock or cpus_write_lock.
6531 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6533 cpumask_var_t cpumask;
6536 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6539 lockdep_assert_cpus_held();
6540 mutex_lock(&wq_pool_mutex);
6542 /* Save the current isolated cpumask & export it via sysfs */
6543 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6546 * If the operation fails, it will fall back to
6547 * wq_requested_unbound_cpumask which is initially set to
6548 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6549 * by any subsequent write to workqueue/cpumask sysfs file.
6551 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6552 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6553 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6554 ret = workqueue_apply_unbound_cpumask(cpumask);
6556 mutex_unlock(&wq_pool_mutex);
6557 free_cpumask_var(cpumask);
6561 static int parse_affn_scope(const char *val)
6565 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6566 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6572 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6574 struct workqueue_struct *wq;
6577 affn = parse_affn_scope(val);
6580 if (affn == WQ_AFFN_DFL)
6584 mutex_lock(&wq_pool_mutex);
6588 list_for_each_entry(wq, &workqueues, list) {
6589 for_each_online_cpu(cpu) {
6590 wq_update_pod(wq, cpu, cpu, true);
6594 mutex_unlock(&wq_pool_mutex);
6600 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6602 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6605 static const struct kernel_param_ops wq_affn_dfl_ops = {
6606 .set = wq_affn_dfl_set,
6607 .get = wq_affn_dfl_get,
6610 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6614 * Workqueues with WQ_SYSFS flag set is visible to userland via
6615 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6616 * following attributes.
6618 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6619 * max_active RW int : maximum number of in-flight work items
6621 * Unbound workqueues have the following extra attributes.
6623 * nice RW int : nice value of the workers
6624 * cpumask RW mask : bitmask of allowed CPUs for the workers
6625 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6626 * affinity_strict RW bool : worker CPU affinity is strict
6629 struct workqueue_struct *wq;
6633 static struct workqueue_struct *dev_to_wq(struct device *dev)
6635 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6640 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6643 struct workqueue_struct *wq = dev_to_wq(dev);
6645 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6647 static DEVICE_ATTR_RO(per_cpu);
6649 static ssize_t max_active_show(struct device *dev,
6650 struct device_attribute *attr, char *buf)
6652 struct workqueue_struct *wq = dev_to_wq(dev);
6654 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6657 static ssize_t max_active_store(struct device *dev,
6658 struct device_attribute *attr, const char *buf,
6661 struct workqueue_struct *wq = dev_to_wq(dev);
6664 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
6667 workqueue_set_max_active(wq, val);
6670 static DEVICE_ATTR_RW(max_active);
6672 static struct attribute *wq_sysfs_attrs[] = {
6673 &dev_attr_per_cpu.attr,
6674 &dev_attr_max_active.attr,
6677 ATTRIBUTE_GROUPS(wq_sysfs);
6679 static void apply_wqattrs_lock(void)
6681 /* CPUs should stay stable across pwq creations and installations */
6683 mutex_lock(&wq_pool_mutex);
6686 static void apply_wqattrs_unlock(void)
6688 mutex_unlock(&wq_pool_mutex);
6692 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
6695 struct workqueue_struct *wq = dev_to_wq(dev);
6698 mutex_lock(&wq->mutex);
6699 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
6700 mutex_unlock(&wq->mutex);
6705 /* prepare workqueue_attrs for sysfs store operations */
6706 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
6708 struct workqueue_attrs *attrs;
6710 lockdep_assert_held(&wq_pool_mutex);
6712 attrs = alloc_workqueue_attrs();
6716 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6720 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6721 const char *buf, size_t count)
6723 struct workqueue_struct *wq = dev_to_wq(dev);
6724 struct workqueue_attrs *attrs;
6727 apply_wqattrs_lock();
6729 attrs = wq_sysfs_prep_attrs(wq);
6733 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6734 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6735 ret = apply_workqueue_attrs_locked(wq, attrs);
6740 apply_wqattrs_unlock();
6741 free_workqueue_attrs(attrs);
6742 return ret ?: count;
6745 static ssize_t wq_cpumask_show(struct device *dev,
6746 struct device_attribute *attr, char *buf)
6748 struct workqueue_struct *wq = dev_to_wq(dev);
6751 mutex_lock(&wq->mutex);
6752 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6753 cpumask_pr_args(wq->unbound_attrs->cpumask));
6754 mutex_unlock(&wq->mutex);
6758 static ssize_t wq_cpumask_store(struct device *dev,
6759 struct device_attribute *attr,
6760 const char *buf, size_t count)
6762 struct workqueue_struct *wq = dev_to_wq(dev);
6763 struct workqueue_attrs *attrs;
6766 apply_wqattrs_lock();
6768 attrs = wq_sysfs_prep_attrs(wq);
6772 ret = cpumask_parse(buf, attrs->cpumask);
6774 ret = apply_workqueue_attrs_locked(wq, attrs);
6777 apply_wqattrs_unlock();
6778 free_workqueue_attrs(attrs);
6779 return ret ?: count;
6782 static ssize_t wq_affn_scope_show(struct device *dev,
6783 struct device_attribute *attr, char *buf)
6785 struct workqueue_struct *wq = dev_to_wq(dev);
6788 mutex_lock(&wq->mutex);
6789 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6790 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6791 wq_affn_names[WQ_AFFN_DFL],
6792 wq_affn_names[wq_affn_dfl]);
6794 written = scnprintf(buf, PAGE_SIZE, "%s\n",
6795 wq_affn_names[wq->unbound_attrs->affn_scope]);
6796 mutex_unlock(&wq->mutex);
6801 static ssize_t wq_affn_scope_store(struct device *dev,
6802 struct device_attribute *attr,
6803 const char *buf, size_t count)
6805 struct workqueue_struct *wq = dev_to_wq(dev);
6806 struct workqueue_attrs *attrs;
6807 int affn, ret = -ENOMEM;
6809 affn = parse_affn_scope(buf);
6813 apply_wqattrs_lock();
6814 attrs = wq_sysfs_prep_attrs(wq);
6816 attrs->affn_scope = affn;
6817 ret = apply_workqueue_attrs_locked(wq, attrs);
6819 apply_wqattrs_unlock();
6820 free_workqueue_attrs(attrs);
6821 return ret ?: count;
6824 static ssize_t wq_affinity_strict_show(struct device *dev,
6825 struct device_attribute *attr, char *buf)
6827 struct workqueue_struct *wq = dev_to_wq(dev);
6829 return scnprintf(buf, PAGE_SIZE, "%d\n",
6830 wq->unbound_attrs->affn_strict);
6833 static ssize_t wq_affinity_strict_store(struct device *dev,
6834 struct device_attribute *attr,
6835 const char *buf, size_t count)
6837 struct workqueue_struct *wq = dev_to_wq(dev);
6838 struct workqueue_attrs *attrs;
6839 int v, ret = -ENOMEM;
6841 if (sscanf(buf, "%d", &v) != 1)
6844 apply_wqattrs_lock();
6845 attrs = wq_sysfs_prep_attrs(wq);
6847 attrs->affn_strict = (bool)v;
6848 ret = apply_workqueue_attrs_locked(wq, attrs);
6850 apply_wqattrs_unlock();
6851 free_workqueue_attrs(attrs);
6852 return ret ?: count;
6855 static struct device_attribute wq_sysfs_unbound_attrs[] = {
6856 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6857 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6858 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6859 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6863 static struct bus_type wq_subsys = {
6864 .name = "workqueue",
6865 .dev_groups = wq_sysfs_groups,
6869 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6870 * @cpumask: the cpumask to set
6872 * The low-level workqueues cpumask is a global cpumask that limits
6873 * the affinity of all unbound workqueues. This function check the @cpumask
6874 * and apply it to all unbound workqueues and updates all pwqs of them.
6876 * Return: 0 - Success
6877 * -EINVAL - Invalid @cpumask
6878 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6880 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6885 * Not excluding isolated cpus on purpose.
6886 * If the user wishes to include them, we allow that.
6888 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6889 if (!cpumask_empty(cpumask)) {
6890 apply_wqattrs_lock();
6891 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
6892 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
6897 ret = workqueue_apply_unbound_cpumask(cpumask);
6900 apply_wqattrs_unlock();
6906 static ssize_t __wq_cpumask_show(struct device *dev,
6907 struct device_attribute *attr, char *buf, cpumask_var_t mask)
6911 mutex_lock(&wq_pool_mutex);
6912 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
6913 mutex_unlock(&wq_pool_mutex);
6918 static ssize_t wq_unbound_cpumask_show(struct device *dev,
6919 struct device_attribute *attr, char *buf)
6921 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
6924 static ssize_t wq_requested_cpumask_show(struct device *dev,
6925 struct device_attribute *attr, char *buf)
6927 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
6930 static ssize_t wq_isolated_cpumask_show(struct device *dev,
6931 struct device_attribute *attr, char *buf)
6933 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
6936 static ssize_t wq_unbound_cpumask_store(struct device *dev,
6937 struct device_attribute *attr, const char *buf, size_t count)
6939 cpumask_var_t cpumask;
6942 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6945 ret = cpumask_parse(buf, cpumask);
6947 ret = workqueue_set_unbound_cpumask(cpumask);
6949 free_cpumask_var(cpumask);
6950 return ret ? ret : count;
6953 static struct device_attribute wq_sysfs_cpumask_attrs[] = {
6954 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6955 wq_unbound_cpumask_store),
6956 __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
6957 __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
6961 static int __init wq_sysfs_init(void)
6963 struct device *dev_root;
6966 err = subsys_virtual_register(&wq_subsys, NULL);
6970 dev_root = bus_get_dev_root(&wq_subsys);
6972 struct device_attribute *attr;
6974 for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
6975 err = device_create_file(dev_root, attr);
6979 put_device(dev_root);
6983 core_initcall(wq_sysfs_init);
6985 static void wq_device_release(struct device *dev)
6987 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6993 * workqueue_sysfs_register - make a workqueue visible in sysfs
6994 * @wq: the workqueue to register
6996 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6997 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6998 * which is the preferred method.
7000 * Workqueue user should use this function directly iff it wants to apply
7001 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7002 * apply_workqueue_attrs() may race against userland updating the
7005 * Return: 0 on success, -errno on failure.
7007 int workqueue_sysfs_register(struct workqueue_struct *wq)
7009 struct wq_device *wq_dev;
7013 * Adjusting max_active or creating new pwqs by applying
7014 * attributes breaks ordering guarantee. Disallow exposing ordered
7017 if (WARN_ON(wq->flags & __WQ_ORDERED))
7020 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7025 wq_dev->dev.bus = &wq_subsys;
7026 wq_dev->dev.release = wq_device_release;
7027 dev_set_name(&wq_dev->dev, "%s", wq->name);
7030 * unbound_attrs are created separately. Suppress uevent until
7031 * everything is ready.
7033 dev_set_uevent_suppress(&wq_dev->dev, true);
7035 ret = device_register(&wq_dev->dev);
7037 put_device(&wq_dev->dev);
7042 if (wq->flags & WQ_UNBOUND) {
7043 struct device_attribute *attr;
7045 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7046 ret = device_create_file(&wq_dev->dev, attr);
7048 device_unregister(&wq_dev->dev);
7055 dev_set_uevent_suppress(&wq_dev->dev, false);
7056 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7061 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7062 * @wq: the workqueue to unregister
7064 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7066 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7068 struct wq_device *wq_dev = wq->wq_dev;
7074 device_unregister(&wq_dev->dev);
7076 #else /* CONFIG_SYSFS */
7077 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7078 #endif /* CONFIG_SYSFS */
7081 * Workqueue watchdog.
7083 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7084 * flush dependency, a concurrency managed work item which stays RUNNING
7085 * indefinitely. Workqueue stalls can be very difficult to debug as the
7086 * usual warning mechanisms don't trigger and internal workqueue state is
7089 * Workqueue watchdog monitors all worker pools periodically and dumps
7090 * state if some pools failed to make forward progress for a while where
7091 * forward progress is defined as the first item on ->worklist changing.
7093 * This mechanism is controlled through the kernel parameter
7094 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7095 * corresponding sysfs parameter file.
7097 #ifdef CONFIG_WQ_WATCHDOG
7099 static unsigned long wq_watchdog_thresh = 30;
7100 static struct timer_list wq_watchdog_timer;
7102 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7103 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7106 * Show workers that might prevent the processing of pending work items.
7107 * The only candidates are CPU-bound workers in the running state.
7108 * Pending work items should be handled by another idle worker
7109 * in all other situations.
7111 static void show_cpu_pool_hog(struct worker_pool *pool)
7113 struct worker *worker;
7114 unsigned long flags;
7117 raw_spin_lock_irqsave(&pool->lock, flags);
7119 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7120 if (task_is_running(worker->task)) {
7122 * Defer printing to avoid deadlocks in console
7123 * drivers that queue work while holding locks
7124 * also taken in their write paths.
7126 printk_deferred_enter();
7128 pr_info("pool %d:\n", pool->id);
7129 sched_show_task(worker->task);
7131 printk_deferred_exit();
7135 raw_spin_unlock_irqrestore(&pool->lock, flags);
7138 static void show_cpu_pools_hogs(void)
7140 struct worker_pool *pool;
7143 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7147 for_each_pool(pool, pi) {
7148 if (pool->cpu_stall)
7149 show_cpu_pool_hog(pool);
7156 static void wq_watchdog_reset_touched(void)
7160 wq_watchdog_touched = jiffies;
7161 for_each_possible_cpu(cpu)
7162 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7165 static void wq_watchdog_timer_fn(struct timer_list *unused)
7167 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7168 bool lockup_detected = false;
7169 bool cpu_pool_stall = false;
7170 unsigned long now = jiffies;
7171 struct worker_pool *pool;
7179 for_each_pool(pool, pi) {
7180 unsigned long pool_ts, touched, ts;
7182 pool->cpu_stall = false;
7183 if (list_empty(&pool->worklist))
7187 * If a virtual machine is stopped by the host it can look to
7188 * the watchdog like a stall.
7190 kvm_check_and_clear_guest_paused();
7192 /* get the latest of pool and touched timestamps */
7194 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7196 touched = READ_ONCE(wq_watchdog_touched);
7197 pool_ts = READ_ONCE(pool->watchdog_ts);
7199 if (time_after(pool_ts, touched))
7205 if (time_after(now, ts + thresh)) {
7206 lockup_detected = true;
7207 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7208 pool->cpu_stall = true;
7209 cpu_pool_stall = true;
7211 pr_emerg("BUG: workqueue lockup - pool");
7212 pr_cont_pool_info(pool);
7213 pr_cont(" stuck for %us!\n",
7214 jiffies_to_msecs(now - pool_ts) / 1000);
7222 if (lockup_detected)
7223 show_all_workqueues();
7226 show_cpu_pools_hogs();
7228 wq_watchdog_reset_touched();
7229 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7232 notrace void wq_watchdog_touch(int cpu)
7235 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7237 wq_watchdog_touched = jiffies;
7240 static void wq_watchdog_set_thresh(unsigned long thresh)
7242 wq_watchdog_thresh = 0;
7243 del_timer_sync(&wq_watchdog_timer);
7246 wq_watchdog_thresh = thresh;
7247 wq_watchdog_reset_touched();
7248 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7252 static int wq_watchdog_param_set_thresh(const char *val,
7253 const struct kernel_param *kp)
7255 unsigned long thresh;
7258 ret = kstrtoul(val, 0, &thresh);
7263 wq_watchdog_set_thresh(thresh);
7265 wq_watchdog_thresh = thresh;
7270 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7271 .set = wq_watchdog_param_set_thresh,
7272 .get = param_get_ulong,
7275 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7278 static void wq_watchdog_init(void)
7280 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7281 wq_watchdog_set_thresh(wq_watchdog_thresh);
7284 #else /* CONFIG_WQ_WATCHDOG */
7286 static inline void wq_watchdog_init(void) { }
7288 #endif /* CONFIG_WQ_WATCHDOG */
7290 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7292 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7293 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7294 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7298 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7301 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7303 BUG_ON(init_worker_pool(pool));
7305 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7306 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7307 pool->attrs->nice = nice;
7308 pool->attrs->affn_strict = true;
7309 pool->node = cpu_to_node(cpu);
7312 mutex_lock(&wq_pool_mutex);
7313 BUG_ON(worker_pool_assign_id(pool));
7314 mutex_unlock(&wq_pool_mutex);
7318 * workqueue_init_early - early init for workqueue subsystem
7320 * This is the first step of three-staged workqueue subsystem initialization and
7321 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7322 * up. It sets up all the data structures and system workqueues and allows early
7323 * boot code to create workqueues and queue/cancel work items. Actual work item
7324 * execution starts only after kthreads can be created and scheduled right
7325 * before early initcalls.
7327 void __init workqueue_init_early(void)
7329 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7330 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7333 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7335 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7336 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7337 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7339 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7340 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7341 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7342 if (!cpumask_empty(&wq_cmdline_cpumask))
7343 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7345 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7347 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7349 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7350 BUG_ON(!wq_update_pod_attrs_buf);
7353 * If nohz_full is enabled, set power efficient workqueue as unbound.
7354 * This allows workqueue items to be moved to HK CPUs.
7356 if (housekeeping_enabled(HK_TYPE_TICK))
7357 wq_power_efficient = true;
7359 /* initialize WQ_AFFN_SYSTEM pods */
7360 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7361 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7362 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7363 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7365 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7368 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7369 pt->pod_node[0] = NUMA_NO_NODE;
7372 /* initialize BH and CPU pools */
7373 for_each_possible_cpu(cpu) {
7374 struct worker_pool *pool;
7377 for_each_bh_worker_pool(pool, cpu) {
7378 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7379 pool->flags |= POOL_BH;
7383 for_each_cpu_worker_pool(pool, cpu)
7384 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7387 /* create default unbound and ordered wq attrs */
7388 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7389 struct workqueue_attrs *attrs;
7391 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7392 attrs->nice = std_nice[i];
7393 unbound_std_wq_attrs[i] = attrs;
7396 * An ordered wq should have only one pwq as ordering is
7397 * guaranteed by max_active which is enforced by pwqs.
7399 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7400 attrs->nice = std_nice[i];
7401 attrs->ordered = true;
7402 ordered_wq_attrs[i] = attrs;
7405 system_wq = alloc_workqueue("events", 0, 0);
7406 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7407 system_long_wq = alloc_workqueue("events_long", 0, 0);
7408 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7410 system_freezable_wq = alloc_workqueue("events_freezable",
7412 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7413 WQ_POWER_EFFICIENT, 0);
7414 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7415 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7417 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7418 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7419 WQ_BH | WQ_HIGHPRI, 0);
7420 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7421 !system_unbound_wq || !system_freezable_wq ||
7422 !system_power_efficient_wq ||
7423 !system_freezable_power_efficient_wq ||
7424 !system_bh_wq || !system_bh_highpri_wq);
7427 static void __init wq_cpu_intensive_thresh_init(void)
7429 unsigned long thresh;
7432 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7433 BUG_ON(IS_ERR(pwq_release_worker));
7435 /* if the user set it to a specific value, keep it */
7436 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7440 * The default of 10ms is derived from the fact that most modern (as of
7441 * 2023) processors can do a lot in 10ms and that it's just below what
7442 * most consider human-perceivable. However, the kernel also runs on a
7443 * lot slower CPUs including microcontrollers where the threshold is way
7446 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7447 * This is by no means accurate but it doesn't have to be. The mechanism
7448 * is still useful even when the threshold is fully scaled up. Also, as
7449 * the reports would usually be applicable to everyone, some machines
7450 * operating on longer thresholds won't significantly diminish their
7453 thresh = 10 * USEC_PER_MSEC;
7455 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7456 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7458 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7460 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7461 loops_per_jiffy, bogo, thresh);
7463 wq_cpu_intensive_thresh_us = thresh;
7467 * workqueue_init - bring workqueue subsystem fully online
7469 * This is the second step of three-staged workqueue subsystem initialization
7470 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7471 * been created and work items queued on them, but there are no kworkers
7472 * executing the work items yet. Populate the worker pools with the initial
7473 * workers and enable future kworker creations.
7475 void __init workqueue_init(void)
7477 struct workqueue_struct *wq;
7478 struct worker_pool *pool;
7481 wq_cpu_intensive_thresh_init();
7483 mutex_lock(&wq_pool_mutex);
7486 * Per-cpu pools created earlier could be missing node hint. Fix them
7487 * up. Also, create a rescuer for workqueues that requested it.
7489 for_each_possible_cpu(cpu) {
7490 for_each_bh_worker_pool(pool, cpu)
7491 pool->node = cpu_to_node(cpu);
7492 for_each_cpu_worker_pool(pool, cpu)
7493 pool->node = cpu_to_node(cpu);
7496 list_for_each_entry(wq, &workqueues, list) {
7497 WARN(init_rescuer(wq),
7498 "workqueue: failed to create early rescuer for %s",
7502 mutex_unlock(&wq_pool_mutex);
7505 * Create the initial workers. A BH pool has one pseudo worker that
7506 * represents the shared BH execution context and thus doesn't get
7507 * affected by hotplug events. Create the BH pseudo workers for all
7508 * possible CPUs here.
7510 for_each_possible_cpu(cpu)
7511 for_each_bh_worker_pool(pool, cpu)
7512 BUG_ON(!create_worker(pool));
7514 for_each_online_cpu(cpu) {
7515 for_each_cpu_worker_pool(pool, cpu) {
7516 pool->flags &= ~POOL_DISASSOCIATED;
7517 BUG_ON(!create_worker(pool));
7521 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7522 BUG_ON(!create_worker(pool));
7529 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7530 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7531 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7533 static void __init init_pod_type(struct wq_pod_type *pt,
7534 bool (*cpus_share_pod)(int, int))
7536 int cur, pre, cpu, pod;
7540 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7541 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7542 BUG_ON(!pt->cpu_pod);
7544 for_each_possible_cpu(cur) {
7545 for_each_possible_cpu(pre) {
7547 pt->cpu_pod[cur] = pt->nr_pods++;
7550 if (cpus_share_pod(cur, pre)) {
7551 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7557 /* init the rest to match @pt->cpu_pod[] */
7558 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7559 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7560 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7562 for (pod = 0; pod < pt->nr_pods; pod++)
7563 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7565 for_each_possible_cpu(cpu) {
7566 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7567 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7571 static bool __init cpus_dont_share(int cpu0, int cpu1)
7576 static bool __init cpus_share_smt(int cpu0, int cpu1)
7578 #ifdef CONFIG_SCHED_SMT
7579 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7585 static bool __init cpus_share_numa(int cpu0, int cpu1)
7587 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7591 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7593 * This is the third step of three-staged workqueue subsystem initialization and
7594 * invoked after SMP and topology information are fully initialized. It
7595 * initializes the unbound CPU pods accordingly.
7597 void __init workqueue_init_topology(void)
7599 struct workqueue_struct *wq;
7602 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7603 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7604 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7605 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7607 wq_topo_initialized = true;
7609 mutex_lock(&wq_pool_mutex);
7612 * Workqueues allocated earlier would have all CPUs sharing the default
7613 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7614 * combinations to apply per-pod sharing.
7616 list_for_each_entry(wq, &workqueues, list) {
7617 for_each_online_cpu(cpu)
7618 wq_update_pod(wq, cpu, cpu, true);
7619 if (wq->flags & WQ_UNBOUND) {
7620 mutex_lock(&wq->mutex);
7621 wq_update_node_max_active(wq, -1);
7622 mutex_unlock(&wq->mutex);
7626 mutex_unlock(&wq_pool_mutex);
7629 void __warn_flushing_systemwide_wq(void)
7631 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7634 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7636 static int __init workqueue_unbound_cpus_setup(char *str)
7638 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7639 cpumask_clear(&wq_cmdline_cpumask);
7640 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7645 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);