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>
57 #include <linux/irq_work.h>
59 #include "workqueue_internal.h"
61 enum worker_pool_flags {
65 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND set and concurrency management
70 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool behaves as an unbound one.
74 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progress.
78 * As there can only be one concurrent BH execution context per CPU, a
79 * BH pool is per-CPU and always DISASSOCIATED.
81 POOL_BH = 1 << 0, /* is a BH pool */
82 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
83 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */
89 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, /* worker was rebound */
96 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOUND | WORKER_REBOUND,
100 enum work_cancel_flags {
101 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
102 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */
105 enum wq_internal_consts {
106 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
108 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
111 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 /* call for help after 10ms
117 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, /* time to breath after fail */
121 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE.
124 RESCUER_NICE_LEVEL = MIN_NICE,
125 HIGHPRI_NICE_LEVEL = MIN_NICE,
131 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
132 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
133 * msecs_to_jiffies() can't be an initializer.
135 #define BH_WORKER_JIFFIES msecs_to_jiffies(2)
136 #define BH_WORKER_RESTARTS 10
139 * Structure fields follow one of the following exclusion rules.
141 * I: Modifiable by initialization/destruction paths and read-only for
144 * P: Preemption protected. Disabling preemption is enough and should
145 * only be modified and accessed from the local cpu.
147 * L: pool->lock protected. Access with pool->lock held.
149 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
152 * K: Only modified by worker while holding pool->lock. Can be safely read by
153 * self, while holding pool->lock or from IRQ context if %current is the
156 * S: Only modified by worker self.
158 * A: wq_pool_attach_mutex protected.
160 * PL: wq_pool_mutex protected.
162 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
164 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
169 * WQ: wq->mutex protected.
171 * WR: wq->mutex protected for writes. RCU protected for reads.
173 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
174 * with READ_ONCE() without locking.
176 * MD: wq_mayday_lock protected.
178 * WD: Used internally by the watchdog.
181 /* struct worker is defined in workqueue_internal.h */
184 raw_spinlock_t lock; /* the pool lock */
185 int cpu; /* I: the associated cpu */
186 int node; /* I: the associated node ID */
187 int id; /* I: pool ID */
188 unsigned int flags; /* L: flags */
190 unsigned long watchdog_ts; /* L: watchdog timestamp */
191 bool cpu_stall; /* WD: stalled cpu bound pool */
194 * The counter is incremented in a process context on the associated CPU
195 * w/ preemption disabled, and decremented or reset in the same context
196 * but w/ pool->lock held. The readers grab pool->lock and are
197 * guaranteed to see if the counter reached zero.
201 struct list_head worklist; /* L: list of pending works */
203 int nr_workers; /* L: total number of workers */
204 int nr_idle; /* L: currently idle workers */
206 struct list_head idle_list; /* L: list of idle workers */
207 struct timer_list idle_timer; /* L: worker idle timeout */
208 struct work_struct idle_cull_work; /* L: worker idle cleanup */
210 struct timer_list mayday_timer; /* L: SOS timer for workers */
212 /* a workers is either on busy_hash or idle_list, or the manager */
213 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
214 /* L: hash of busy workers */
216 struct worker *manager; /* L: purely informational */
217 struct list_head workers; /* A: attached workers */
218 struct list_head dying_workers; /* A: workers about to die */
219 struct completion *detach_completion; /* all workers detached */
221 struct ida worker_ida; /* worker IDs for task name */
223 struct workqueue_attrs *attrs; /* I: worker attributes */
224 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
225 int refcnt; /* PL: refcnt for unbound pools */
228 * Destruction of pool is RCU protected to allow dereferences
229 * from get_work_pool().
235 * Per-pool_workqueue statistics. These can be monitored using
236 * tools/workqueue/wq_monitor.py.
238 enum pool_workqueue_stats {
239 PWQ_STAT_STARTED, /* work items started execution */
240 PWQ_STAT_COMPLETED, /* work items completed execution */
241 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
242 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
243 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
244 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
245 PWQ_STAT_MAYDAY, /* maydays to rescuer */
246 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
252 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
253 * of work_struct->data are used for flags and the remaining high bits
254 * point to the pwq; thus, pwqs need to be aligned at two's power of the
255 * number of flag bits.
257 struct pool_workqueue {
258 struct worker_pool *pool; /* I: the associated pool */
259 struct workqueue_struct *wq; /* I: the owning workqueue */
260 int work_color; /* L: current color */
261 int flush_color; /* L: flushing color */
262 int refcnt; /* L: reference count */
263 int nr_in_flight[WORK_NR_COLORS];
264 /* L: nr of in_flight works */
265 bool plugged; /* L: execution suspended */
268 * nr_active management and WORK_STRUCT_INACTIVE:
270 * When pwq->nr_active >= max_active, new work item is queued to
271 * pwq->inactive_works instead of pool->worklist and marked with
272 * WORK_STRUCT_INACTIVE.
274 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
275 * nr_active and all work items in pwq->inactive_works are marked with
276 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
277 * in pwq->inactive_works. Some of them are ready to run in
278 * pool->worklist or worker->scheduled. Those work itmes are only struct
279 * wq_barrier which is used for flush_work() and should not participate
280 * in nr_active. For non-barrier work item, it is marked with
281 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
283 int nr_active; /* L: nr of active works */
284 struct list_head inactive_works; /* L: inactive works */
285 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
286 struct list_head pwqs_node; /* WR: node on wq->pwqs */
287 struct list_head mayday_node; /* MD: node on wq->maydays */
289 u64 stats[PWQ_NR_STATS];
292 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
293 * and pwq_release_workfn() for details. pool_workqueue itself is also
294 * RCU protected so that the first pwq can be determined without
295 * grabbing wq->mutex.
297 struct kthread_work release_work;
299 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
302 * Structure used to wait for workqueue flush.
305 struct list_head list; /* WQ: list of flushers */
306 int flush_color; /* WQ: flush color waiting for */
307 struct completion done; /* flush completion */
313 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
314 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
315 * As sharing a single nr_active across multiple sockets can be very expensive,
316 * the counting and enforcement is per NUMA node.
318 * The following struct is used to enforce per-node max_active. When a pwq wants
319 * to start executing a work item, it should increment ->nr using
320 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
321 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
322 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
325 struct wq_node_nr_active {
326 int max; /* per-node max_active */
327 atomic_t nr; /* per-node nr_active */
328 raw_spinlock_t lock; /* nests inside pool locks */
329 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
333 * The externally visible workqueue. It relays the issued work items to
334 * the appropriate worker_pool through its pool_workqueues.
336 struct workqueue_struct {
337 struct list_head pwqs; /* WR: all pwqs of this wq */
338 struct list_head list; /* PR: list of all workqueues */
340 struct mutex mutex; /* protects this wq */
341 int work_color; /* WQ: current work color */
342 int flush_color; /* WQ: current flush color */
343 atomic_t nr_pwqs_to_flush; /* flush in progress */
344 struct wq_flusher *first_flusher; /* WQ: first flusher */
345 struct list_head flusher_queue; /* WQ: flush waiters */
346 struct list_head flusher_overflow; /* WQ: flush overflow list */
348 struct list_head maydays; /* MD: pwqs requesting rescue */
349 struct worker *rescuer; /* MD: rescue worker */
351 int nr_drainers; /* WQ: drain in progress */
353 /* See alloc_workqueue() function comment for info on min/max_active */
354 int max_active; /* WO: max active works */
355 int min_active; /* WO: min active works */
356 int saved_max_active; /* WQ: saved max_active */
357 int saved_min_active; /* WQ: saved min_active */
359 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
360 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
363 struct wq_device *wq_dev; /* I: for sysfs interface */
365 #ifdef CONFIG_LOCKDEP
367 struct lock_class_key key;
368 struct lockdep_map lockdep_map;
370 char name[WQ_NAME_LEN]; /* I: workqueue name */
373 * Destruction of workqueue_struct is RCU protected to allow walking
374 * the workqueues list without grabbing wq_pool_mutex.
375 * This is used to dump all workqueues from sysrq.
379 /* hot fields used during command issue, aligned to cacheline */
380 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
381 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
382 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
386 * Each pod type describes how CPUs should be grouped for unbound workqueues.
387 * See the comment above workqueue_attrs->affn_scope.
390 int nr_pods; /* number of pods */
391 cpumask_var_t *pod_cpus; /* pod -> cpus */
392 int *pod_node; /* pod -> node */
393 int *cpu_pod; /* cpu -> pod */
396 struct work_offq_data {
402 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
403 [WQ_AFFN_DFL] = "default",
404 [WQ_AFFN_CPU] = "cpu",
405 [WQ_AFFN_SMT] = "smt",
406 [WQ_AFFN_CACHE] = "cache",
407 [WQ_AFFN_NUMA] = "numa",
408 [WQ_AFFN_SYSTEM] = "system",
412 * Per-cpu work items which run for longer than the following threshold are
413 * automatically considered CPU intensive and excluded from concurrency
414 * management to prevent them from noticeably delaying other per-cpu work items.
415 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
416 * The actual value is initialized in wq_cpu_intensive_thresh_init().
418 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
419 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
420 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
421 static unsigned int wq_cpu_intensive_warning_thresh = 4;
422 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
425 /* see the comment above the definition of WQ_POWER_EFFICIENT */
426 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
427 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
429 static bool wq_online; /* can kworkers be created yet? */
430 static bool wq_topo_initialized __read_mostly = false;
432 static struct kmem_cache *pwq_cache;
434 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
435 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
437 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
438 static struct workqueue_attrs *wq_update_pod_attrs_buf;
440 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
441 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
442 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
443 /* wait for manager to go away */
444 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
446 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
447 static bool workqueue_freezing; /* PL: have wqs started freezing? */
449 /* PL&A: allowable cpus for unbound wqs and work items */
450 static cpumask_var_t wq_unbound_cpumask;
452 /* PL: user requested unbound cpumask via sysfs */
453 static cpumask_var_t wq_requested_unbound_cpumask;
455 /* PL: isolated cpumask to be excluded from unbound cpumask */
456 static cpumask_var_t wq_isolated_cpumask;
458 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
459 static struct cpumask wq_cmdline_cpumask __initdata;
461 /* CPU where unbound work was last round robin scheduled from this CPU */
462 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
465 * Local execution of unbound work items is no longer guaranteed. The
466 * following always forces round-robin CPU selection on unbound work items
467 * to uncover usages which depend on it.
469 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
470 static bool wq_debug_force_rr_cpu = true;
472 static bool wq_debug_force_rr_cpu = false;
474 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
476 /* to raise softirq for the BH worker pools on other CPUs */
477 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS],
480 /* the BH worker pools */
481 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
484 /* the per-cpu worker pools */
485 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
488 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
490 /* PL: hash of all unbound pools keyed by pool->attrs */
491 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
493 /* I: attributes used when instantiating standard unbound pools on demand */
494 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
496 /* I: attributes used when instantiating ordered pools on demand */
497 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
501 * process context while holding a pool lock. Bounce to a dedicated kthread
502 * worker to avoid A-A deadlocks.
504 static struct kthread_worker *pwq_release_worker __ro_after_init;
506 struct workqueue_struct *system_wq __ro_after_init;
507 EXPORT_SYMBOL(system_wq);
508 struct workqueue_struct *system_highpri_wq __ro_after_init;
509 EXPORT_SYMBOL_GPL(system_highpri_wq);
510 struct workqueue_struct *system_long_wq __ro_after_init;
511 EXPORT_SYMBOL_GPL(system_long_wq);
512 struct workqueue_struct *system_unbound_wq __ro_after_init;
513 EXPORT_SYMBOL_GPL(system_unbound_wq);
514 struct workqueue_struct *system_freezable_wq __ro_after_init;
515 EXPORT_SYMBOL_GPL(system_freezable_wq);
516 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
517 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
518 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
519 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
520 struct workqueue_struct *system_bh_wq;
521 EXPORT_SYMBOL_GPL(system_bh_wq);
522 struct workqueue_struct *system_bh_highpri_wq;
523 EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
525 static int worker_thread(void *__worker);
526 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
527 static void show_pwq(struct pool_workqueue *pwq);
528 static void show_one_worker_pool(struct worker_pool *pool);
530 #define CREATE_TRACE_POINTS
531 #include <trace/events/workqueue.h>
533 #define assert_rcu_or_pool_mutex() \
534 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
535 !lockdep_is_held(&wq_pool_mutex), \
536 "RCU or wq_pool_mutex should be held")
538 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
539 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
540 !lockdep_is_held(&wq->mutex) && \
541 !lockdep_is_held(&wq_pool_mutex), \
542 "RCU, wq->mutex or wq_pool_mutex should be held")
544 #define for_each_bh_worker_pool(pool, cpu) \
545 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
546 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
549 #define for_each_cpu_worker_pool(pool, cpu) \
550 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
551 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
555 * for_each_pool - iterate through all worker_pools in the system
556 * @pool: iteration cursor
557 * @pi: integer used for iteration
559 * This must be called either with wq_pool_mutex held or RCU read
560 * locked. If the pool needs to be used beyond the locking in effect, the
561 * caller is responsible for guaranteeing that the pool stays online.
563 * The if/else clause exists only for the lockdep assertion and can be
566 #define for_each_pool(pool, pi) \
567 idr_for_each_entry(&worker_pool_idr, pool, pi) \
568 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
572 * for_each_pool_worker - iterate through all workers of a worker_pool
573 * @worker: iteration cursor
574 * @pool: worker_pool to iterate workers of
576 * This must be called with wq_pool_attach_mutex.
578 * The if/else clause exists only for the lockdep assertion and can be
581 #define for_each_pool_worker(worker, pool) \
582 list_for_each_entry((worker), &(pool)->workers, node) \
583 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
587 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
588 * @pwq: iteration cursor
589 * @wq: the target workqueue
591 * This must be called either with wq->mutex held or RCU read locked.
592 * If the pwq needs to be used beyond the locking in effect, the caller is
593 * responsible for guaranteeing that the pwq stays online.
595 * The if/else clause exists only for the lockdep assertion and can be
598 #define for_each_pwq(pwq, wq) \
599 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
600 lockdep_is_held(&(wq->mutex)))
602 #ifdef CONFIG_DEBUG_OBJECTS_WORK
604 static const struct debug_obj_descr work_debug_descr;
606 static void *work_debug_hint(void *addr)
608 return ((struct work_struct *) addr)->func;
611 static bool work_is_static_object(void *addr)
613 struct work_struct *work = addr;
615 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
619 * fixup_init is called when:
620 * - an active object is initialized
622 static bool work_fixup_init(void *addr, enum debug_obj_state state)
624 struct work_struct *work = addr;
627 case ODEBUG_STATE_ACTIVE:
628 cancel_work_sync(work);
629 debug_object_init(work, &work_debug_descr);
637 * fixup_free is called when:
638 * - an active object is freed
640 static bool work_fixup_free(void *addr, enum debug_obj_state state)
642 struct work_struct *work = addr;
645 case ODEBUG_STATE_ACTIVE:
646 cancel_work_sync(work);
647 debug_object_free(work, &work_debug_descr);
654 static const struct debug_obj_descr work_debug_descr = {
655 .name = "work_struct",
656 .debug_hint = work_debug_hint,
657 .is_static_object = work_is_static_object,
658 .fixup_init = work_fixup_init,
659 .fixup_free = work_fixup_free,
662 static inline void debug_work_activate(struct work_struct *work)
664 debug_object_activate(work, &work_debug_descr);
667 static inline void debug_work_deactivate(struct work_struct *work)
669 debug_object_deactivate(work, &work_debug_descr);
672 void __init_work(struct work_struct *work, int onstack)
675 debug_object_init_on_stack(work, &work_debug_descr);
677 debug_object_init(work, &work_debug_descr);
679 EXPORT_SYMBOL_GPL(__init_work);
681 void destroy_work_on_stack(struct work_struct *work)
683 debug_object_free(work, &work_debug_descr);
685 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
687 void destroy_delayed_work_on_stack(struct delayed_work *work)
689 destroy_timer_on_stack(&work->timer);
690 debug_object_free(&work->work, &work_debug_descr);
692 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
695 static inline void debug_work_activate(struct work_struct *work) { }
696 static inline void debug_work_deactivate(struct work_struct *work) { }
700 * worker_pool_assign_id - allocate ID and assign it to @pool
701 * @pool: the pool pointer of interest
703 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
704 * successfully, -errno on failure.
706 static int worker_pool_assign_id(struct worker_pool *pool)
710 lockdep_assert_held(&wq_pool_mutex);
712 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
721 static struct pool_workqueue __rcu **
722 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
725 return per_cpu_ptr(wq->cpu_pwq, cpu);
730 /* @cpu < 0 for dfl_pwq */
731 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
733 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
734 lockdep_is_held(&wq_pool_mutex) ||
735 lockdep_is_held(&wq->mutex));
739 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
740 * @wq: workqueue of interest
742 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
743 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
744 * default pwq is always mapped to the pool with the current effective cpumask.
746 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
748 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
751 static unsigned int work_color_to_flags(int color)
753 return color << WORK_STRUCT_COLOR_SHIFT;
756 static int get_work_color(unsigned long work_data)
758 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
759 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
762 static int work_next_color(int color)
764 return (color + 1) % WORK_NR_COLORS;
767 static unsigned long pool_offq_flags(struct worker_pool *pool)
769 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
773 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
774 * contain the pointer to the queued pwq. Once execution starts, the flag
775 * is cleared and the high bits contain OFFQ flags and pool ID.
777 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
778 * can be used to set the pwq, pool or clear work->data. These functions should
779 * only be called while the work is owned - ie. while the PENDING bit is set.
781 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
782 * corresponding to a work. Pool is available once the work has been
783 * queued anywhere after initialization until it is sync canceled. pwq is
784 * available only while the work item is queued.
786 static inline void set_work_data(struct work_struct *work, unsigned long data)
788 WARN_ON_ONCE(!work_pending(work));
789 atomic_long_set(&work->data, data | work_static(work));
792 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
795 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
796 WORK_STRUCT_PWQ | flags);
799 static void set_work_pool_and_keep_pending(struct work_struct *work,
800 int pool_id, unsigned long flags)
802 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
803 WORK_STRUCT_PENDING | flags);
806 static void set_work_pool_and_clear_pending(struct work_struct *work,
807 int pool_id, unsigned long flags)
810 * The following wmb is paired with the implied mb in
811 * test_and_set_bit(PENDING) and ensures all updates to @work made
812 * here are visible to and precede any updates by the next PENDING
816 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
819 * The following mb guarantees that previous clear of a PENDING bit
820 * will not be reordered with any speculative LOADS or STORES from
821 * work->current_func, which is executed afterwards. This possible
822 * reordering can lead to a missed execution on attempt to queue
823 * the same @work. E.g. consider this case:
826 * ---------------------------- --------------------------------
828 * 1 STORE event_indicated
829 * 2 queue_work_on() {
830 * 3 test_and_set_bit(PENDING)
831 * 4 } set_..._and_clear_pending() {
832 * 5 set_work_data() # clear bit
834 * 7 work->current_func() {
835 * 8 LOAD event_indicated
838 * Without an explicit full barrier speculative LOAD on line 8 can
839 * be executed before CPU#0 does STORE on line 1. If that happens,
840 * CPU#0 observes the PENDING bit is still set and new execution of
841 * a @work is not queued in a hope, that CPU#1 will eventually
842 * finish the queued @work. Meanwhile CPU#1 does not see
843 * event_indicated is set, because speculative LOAD was executed
844 * before actual STORE.
849 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
851 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
854 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
856 unsigned long data = atomic_long_read(&work->data);
858 if (data & WORK_STRUCT_PWQ)
859 return work_struct_pwq(data);
865 * get_work_pool - return the worker_pool a given work was associated with
866 * @work: the work item of interest
868 * Pools are created and destroyed under wq_pool_mutex, and allows read
869 * access under RCU read lock. As such, this function should be
870 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
872 * All fields of the returned pool are accessible as long as the above
873 * mentioned locking is in effect. If the returned pool needs to be used
874 * beyond the critical section, the caller is responsible for ensuring the
875 * returned pool is and stays online.
877 * Return: The worker_pool @work was last associated with. %NULL if none.
879 static struct worker_pool *get_work_pool(struct work_struct *work)
881 unsigned long data = atomic_long_read(&work->data);
884 assert_rcu_or_pool_mutex();
886 if (data & WORK_STRUCT_PWQ)
887 return work_struct_pwq(data)->pool;
889 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
890 if (pool_id == WORK_OFFQ_POOL_NONE)
893 return idr_find(&worker_pool_idr, pool_id);
896 static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
898 return (v >> shift) & ((1 << bits) - 1);
901 static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
903 WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
905 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT,
906 WORK_OFFQ_POOL_BITS);
907 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT,
908 WORK_OFFQ_DISABLE_BITS);
909 offqd->flags = data & WORK_OFFQ_FLAG_MASK;
912 static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
914 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) |
915 ((unsigned long)offqd->flags);
919 * Policy functions. These define the policies on how the global worker
920 * pools are managed. Unless noted otherwise, these functions assume that
921 * they're being called with pool->lock held.
925 * Need to wake up a worker? Called from anything but currently
928 * Note that, because unbound workers never contribute to nr_running, this
929 * function will always return %true for unbound pools as long as the
930 * worklist isn't empty.
932 static bool need_more_worker(struct worker_pool *pool)
934 return !list_empty(&pool->worklist) && !pool->nr_running;
937 /* Can I start working? Called from busy but !running workers. */
938 static bool may_start_working(struct worker_pool *pool)
940 return pool->nr_idle;
943 /* Do I need to keep working? Called from currently running workers. */
944 static bool keep_working(struct worker_pool *pool)
946 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
949 /* Do we need a new worker? Called from manager. */
950 static bool need_to_create_worker(struct worker_pool *pool)
952 return need_more_worker(pool) && !may_start_working(pool);
955 /* Do we have too many workers and should some go away? */
956 static bool too_many_workers(struct worker_pool *pool)
958 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
959 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
960 int nr_busy = pool->nr_workers - nr_idle;
962 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
966 * worker_set_flags - set worker flags and adjust nr_running accordingly
968 * @flags: flags to set
970 * Set @flags in @worker->flags and adjust nr_running accordingly.
972 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
974 struct worker_pool *pool = worker->pool;
976 lockdep_assert_held(&pool->lock);
978 /* If transitioning into NOT_RUNNING, adjust nr_running. */
979 if ((flags & WORKER_NOT_RUNNING) &&
980 !(worker->flags & WORKER_NOT_RUNNING)) {
984 worker->flags |= flags;
988 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
990 * @flags: flags to clear
992 * Clear @flags in @worker->flags and adjust nr_running accordingly.
994 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
996 struct worker_pool *pool = worker->pool;
997 unsigned int oflags = worker->flags;
999 lockdep_assert_held(&pool->lock);
1001 worker->flags &= ~flags;
1004 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1005 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1006 * of multiple flags, not a single flag.
1008 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1009 if (!(worker->flags & WORKER_NOT_RUNNING))
1013 /* Return the first idle worker. Called with pool->lock held. */
1014 static struct worker *first_idle_worker(struct worker_pool *pool)
1016 if (unlikely(list_empty(&pool->idle_list)))
1019 return list_first_entry(&pool->idle_list, struct worker, entry);
1023 * worker_enter_idle - enter idle state
1024 * @worker: worker which is entering idle state
1026 * @worker is entering idle state. Update stats and idle timer if
1030 * raw_spin_lock_irq(pool->lock).
1032 static void worker_enter_idle(struct worker *worker)
1034 struct worker_pool *pool = worker->pool;
1036 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1037 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1038 (worker->hentry.next || worker->hentry.pprev)))
1041 /* can't use worker_set_flags(), also called from create_worker() */
1042 worker->flags |= WORKER_IDLE;
1044 worker->last_active = jiffies;
1046 /* idle_list is LIFO */
1047 list_add(&worker->entry, &pool->idle_list);
1049 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1050 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1052 /* Sanity check nr_running. */
1053 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1057 * worker_leave_idle - leave idle state
1058 * @worker: worker which is leaving idle state
1060 * @worker is leaving idle state. Update stats.
1063 * raw_spin_lock_irq(pool->lock).
1065 static void worker_leave_idle(struct worker *worker)
1067 struct worker_pool *pool = worker->pool;
1069 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1071 worker_clr_flags(worker, WORKER_IDLE);
1073 list_del_init(&worker->entry);
1077 * find_worker_executing_work - find worker which is executing a work
1078 * @pool: pool of interest
1079 * @work: work to find worker for
1081 * Find a worker which is executing @work on @pool by searching
1082 * @pool->busy_hash which is keyed by the address of @work. For a worker
1083 * to match, its current execution should match the address of @work and
1084 * its work function. This is to avoid unwanted dependency between
1085 * unrelated work executions through a work item being recycled while still
1088 * This is a bit tricky. A work item may be freed once its execution
1089 * starts and nothing prevents the freed area from being recycled for
1090 * another work item. If the same work item address ends up being reused
1091 * before the original execution finishes, workqueue will identify the
1092 * recycled work item as currently executing and make it wait until the
1093 * current execution finishes, introducing an unwanted dependency.
1095 * This function checks the work item address and work function to avoid
1096 * false positives. Note that this isn't complete as one may construct a
1097 * work function which can introduce dependency onto itself through a
1098 * recycled work item. Well, if somebody wants to shoot oneself in the
1099 * foot that badly, there's only so much we can do, and if such deadlock
1100 * actually occurs, it should be easy to locate the culprit work function.
1103 * raw_spin_lock_irq(pool->lock).
1106 * Pointer to worker which is executing @work if found, %NULL
1109 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1110 struct work_struct *work)
1112 struct worker *worker;
1114 hash_for_each_possible(pool->busy_hash, worker, hentry,
1115 (unsigned long)work)
1116 if (worker->current_work == work &&
1117 worker->current_func == work->func)
1124 * move_linked_works - move linked works to a list
1125 * @work: start of series of works to be scheduled
1126 * @head: target list to append @work to
1127 * @nextp: out parameter for nested worklist walking
1129 * Schedule linked works starting from @work to @head. Work series to be
1130 * scheduled starts at @work and includes any consecutive work with
1131 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1135 * raw_spin_lock_irq(pool->lock).
1137 static void move_linked_works(struct work_struct *work, struct list_head *head,
1138 struct work_struct **nextp)
1140 struct work_struct *n;
1143 * Linked worklist will always end before the end of the list,
1144 * use NULL for list head.
1146 list_for_each_entry_safe_from(work, n, NULL, entry) {
1147 list_move_tail(&work->entry, head);
1148 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1153 * If we're already inside safe list traversal and have moved
1154 * multiple works to the scheduled queue, the next position
1155 * needs to be updated.
1162 * assign_work - assign a work item and its linked work items to a worker
1163 * @work: work to assign
1164 * @worker: worker to assign to
1165 * @nextp: out parameter for nested worklist walking
1167 * Assign @work and its linked work items to @worker. If @work is already being
1168 * executed by another worker in the same pool, it'll be punted there.
1170 * If @nextp is not NULL, it's updated to point to the next work of the last
1171 * scheduled work. This allows assign_work() to be nested inside
1172 * list_for_each_entry_safe().
1174 * Returns %true if @work was successfully assigned to @worker. %false if @work
1175 * was punted to another worker already executing it.
1177 static bool assign_work(struct work_struct *work, struct worker *worker,
1178 struct work_struct **nextp)
1180 struct worker_pool *pool = worker->pool;
1181 struct worker *collision;
1183 lockdep_assert_held(&pool->lock);
1186 * A single work shouldn't be executed concurrently by multiple workers.
1187 * __queue_work() ensures that @work doesn't jump to a different pool
1188 * while still running in the previous pool. Here, we should ensure that
1189 * @work is not executed concurrently by multiple workers from the same
1190 * pool. Check whether anyone is already processing the work. If so,
1191 * defer the work to the currently executing one.
1193 collision = find_worker_executing_work(pool, work);
1194 if (unlikely(collision)) {
1195 move_linked_works(work, &collision->scheduled, nextp);
1199 move_linked_works(work, &worker->scheduled, nextp);
1203 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1205 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1207 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1210 static void kick_bh_pool(struct worker_pool *pool)
1213 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1214 if (unlikely(pool->cpu != smp_processor_id() &&
1215 !(pool->flags & POOL_BH_DRAINING))) {
1216 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1220 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1221 raise_softirq_irqoff(HI_SOFTIRQ);
1223 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1227 * kick_pool - wake up an idle worker if necessary
1228 * @pool: pool to kick
1230 * @pool may have pending work items. Wake up worker if necessary. Returns
1231 * whether a worker was woken up.
1233 static bool kick_pool(struct worker_pool *pool)
1235 struct worker *worker = first_idle_worker(pool);
1236 struct task_struct *p;
1238 lockdep_assert_held(&pool->lock);
1240 if (!need_more_worker(pool) || !worker)
1243 if (pool->flags & POOL_BH) {
1252 * Idle @worker is about to execute @work and waking up provides an
1253 * opportunity to migrate @worker at a lower cost by setting the task's
1254 * wake_cpu field. Let's see if we want to move @worker to improve
1255 * execution locality.
1257 * We're waking the worker that went idle the latest and there's some
1258 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1259 * so, setting the wake_cpu won't do anything. As this is a best-effort
1260 * optimization and the race window is narrow, let's leave as-is for
1261 * now. If this becomes pronounced, we can skip over workers which are
1262 * still on cpu when picking an idle worker.
1264 * If @pool has non-strict affinity, @worker might have ended up outside
1265 * its affinity scope. Repatriate.
1267 if (!pool->attrs->affn_strict &&
1268 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1269 struct work_struct *work = list_first_entry(&pool->worklist,
1270 struct work_struct, entry);
1271 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1273 if (wake_cpu < nr_cpu_ids) {
1274 p->wake_cpu = wake_cpu;
1275 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1283 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1286 * Concurrency-managed per-cpu work items that hog CPU for longer than
1287 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1288 * which prevents them from stalling other concurrency-managed work items. If a
1289 * work function keeps triggering this mechanism, it's likely that the work item
1290 * should be using an unbound workqueue instead.
1292 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1293 * and report them so that they can be examined and converted to use unbound
1294 * workqueues as appropriate. To avoid flooding the console, each violating work
1295 * function is tracked and reported with exponential backoff.
1297 #define WCI_MAX_ENTS 128
1302 struct hlist_node hash_node;
1305 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1306 static int wci_nr_ents;
1307 static DEFINE_RAW_SPINLOCK(wci_lock);
1308 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1310 static struct wci_ent *wci_find_ent(work_func_t func)
1312 struct wci_ent *ent;
1314 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1315 (unsigned long)func) {
1316 if (ent->func == func)
1322 static void wq_cpu_intensive_report(work_func_t func)
1324 struct wci_ent *ent;
1327 ent = wci_find_ent(func);
1332 * Start reporting from the warning_thresh and back off
1335 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1336 if (wq_cpu_intensive_warning_thresh &&
1337 cnt >= wq_cpu_intensive_warning_thresh &&
1338 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1339 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1340 ent->func, wq_cpu_intensive_thresh_us,
1341 atomic64_read(&ent->cnt));
1346 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1347 * is exhausted, something went really wrong and we probably made enough
1350 if (wci_nr_ents >= WCI_MAX_ENTS)
1353 raw_spin_lock(&wci_lock);
1355 if (wci_nr_ents >= WCI_MAX_ENTS) {
1356 raw_spin_unlock(&wci_lock);
1360 if (wci_find_ent(func)) {
1361 raw_spin_unlock(&wci_lock);
1365 ent = &wci_ents[wci_nr_ents++];
1367 atomic64_set(&ent->cnt, 0);
1368 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1370 raw_spin_unlock(&wci_lock);
1375 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1376 static void wq_cpu_intensive_report(work_func_t func) {}
1377 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1380 * wq_worker_running - a worker is running again
1381 * @task: task waking up
1383 * This function is called when a worker returns from schedule()
1385 void wq_worker_running(struct task_struct *task)
1387 struct worker *worker = kthread_data(task);
1389 if (!READ_ONCE(worker->sleeping))
1393 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1394 * and the nr_running increment below, we may ruin the nr_running reset
1395 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1396 * pool. Protect against such race.
1399 if (!(worker->flags & WORKER_NOT_RUNNING))
1400 worker->pool->nr_running++;
1404 * CPU intensive auto-detection cares about how long a work item hogged
1405 * CPU without sleeping. Reset the starting timestamp on wakeup.
1407 worker->current_at = worker->task->se.sum_exec_runtime;
1409 WRITE_ONCE(worker->sleeping, 0);
1413 * wq_worker_sleeping - a worker is going to sleep
1414 * @task: task going to sleep
1416 * This function is called from schedule() when a busy worker is
1419 void wq_worker_sleeping(struct task_struct *task)
1421 struct worker *worker = kthread_data(task);
1422 struct worker_pool *pool;
1425 * Rescuers, which may not have all the fields set up like normal
1426 * workers, also reach here, let's not access anything before
1427 * checking NOT_RUNNING.
1429 if (worker->flags & WORKER_NOT_RUNNING)
1432 pool = worker->pool;
1434 /* Return if preempted before wq_worker_running() was reached */
1435 if (READ_ONCE(worker->sleeping))
1438 WRITE_ONCE(worker->sleeping, 1);
1439 raw_spin_lock_irq(&pool->lock);
1442 * Recheck in case unbind_workers() preempted us. We don't
1443 * want to decrement nr_running after the worker is unbound
1444 * and nr_running has been reset.
1446 if (worker->flags & WORKER_NOT_RUNNING) {
1447 raw_spin_unlock_irq(&pool->lock);
1452 if (kick_pool(pool))
1453 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1455 raw_spin_unlock_irq(&pool->lock);
1459 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1460 * @task: task currently running
1462 * Called from sched_tick(). We're in the IRQ context and the current
1463 * worker's fields which follow the 'K' locking rule can be accessed safely.
1465 void wq_worker_tick(struct task_struct *task)
1467 struct worker *worker = kthread_data(task);
1468 struct pool_workqueue *pwq = worker->current_pwq;
1469 struct worker_pool *pool = worker->pool;
1474 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1476 if (!wq_cpu_intensive_thresh_us)
1480 * If the current worker is concurrency managed and hogged the CPU for
1481 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1482 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1484 * Set @worker->sleeping means that @worker is in the process of
1485 * switching out voluntarily and won't be contributing to
1486 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1487 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1488 * double decrements. The task is releasing the CPU anyway. Let's skip.
1489 * We probably want to make this prettier in the future.
1491 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1492 worker->task->se.sum_exec_runtime - worker->current_at <
1493 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1496 raw_spin_lock(&pool->lock);
1498 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1499 wq_cpu_intensive_report(worker->current_func);
1500 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1502 if (kick_pool(pool))
1503 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1505 raw_spin_unlock(&pool->lock);
1509 * wq_worker_last_func - retrieve worker's last work function
1510 * @task: Task to retrieve last work function of.
1512 * Determine the last function a worker executed. This is called from
1513 * the scheduler to get a worker's last known identity.
1516 * raw_spin_lock_irq(rq->lock)
1518 * This function is called during schedule() when a kworker is going
1519 * to sleep. It's used by psi to identify aggregation workers during
1520 * dequeuing, to allow periodic aggregation to shut-off when that
1521 * worker is the last task in the system or cgroup to go to sleep.
1523 * As this function doesn't involve any workqueue-related locking, it
1524 * only returns stable values when called from inside the scheduler's
1525 * queuing and dequeuing paths, when @task, which must be a kworker,
1526 * is guaranteed to not be processing any works.
1529 * The last work function %current executed as a worker, NULL if it
1530 * hasn't executed any work yet.
1532 work_func_t wq_worker_last_func(struct task_struct *task)
1534 struct worker *worker = kthread_data(task);
1536 return worker->last_func;
1540 * wq_node_nr_active - Determine wq_node_nr_active to use
1541 * @wq: workqueue of interest
1542 * @node: NUMA node, can be %NUMA_NO_NODE
1544 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1546 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1548 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1550 * - Otherwise, node_nr_active[@node].
1552 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1555 if (!(wq->flags & WQ_UNBOUND))
1558 if (node == NUMA_NO_NODE)
1561 return wq->node_nr_active[node];
1565 * wq_update_node_max_active - Update per-node max_actives to use
1566 * @wq: workqueue to update
1567 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1569 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1570 * distributed among nodes according to the proportions of numbers of online
1571 * cpus. The result is always between @wq->min_active and max_active.
1573 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1575 struct cpumask *effective = unbound_effective_cpumask(wq);
1576 int min_active = READ_ONCE(wq->min_active);
1577 int max_active = READ_ONCE(wq->max_active);
1578 int total_cpus, node;
1580 lockdep_assert_held(&wq->mutex);
1582 if (!wq_topo_initialized)
1585 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1588 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1592 /* If all CPUs of the wq get offline, use the default values */
1593 if (unlikely(!total_cpus)) {
1595 wq_node_nr_active(wq, node)->max = min_active;
1597 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1601 for_each_node(node) {
1604 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1605 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1608 wq_node_nr_active(wq, node)->max =
1609 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1610 min_active, max_active);
1613 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1617 * get_pwq - get an extra reference on the specified pool_workqueue
1618 * @pwq: pool_workqueue to get
1620 * Obtain an extra reference on @pwq. The caller should guarantee that
1621 * @pwq has positive refcnt and be holding the matching pool->lock.
1623 static void get_pwq(struct pool_workqueue *pwq)
1625 lockdep_assert_held(&pwq->pool->lock);
1626 WARN_ON_ONCE(pwq->refcnt <= 0);
1631 * put_pwq - put a pool_workqueue reference
1632 * @pwq: pool_workqueue to put
1634 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1635 * destruction. The caller should be holding the matching pool->lock.
1637 static void put_pwq(struct pool_workqueue *pwq)
1639 lockdep_assert_held(&pwq->pool->lock);
1640 if (likely(--pwq->refcnt))
1643 * @pwq can't be released under pool->lock, bounce to a dedicated
1644 * kthread_worker to avoid A-A deadlocks.
1646 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1650 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1651 * @pwq: pool_workqueue to put (can be %NULL)
1653 * put_pwq() with locking. This function also allows %NULL @pwq.
1655 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1659 * As both pwqs and pools are RCU protected, the
1660 * following lock operations are safe.
1662 raw_spin_lock_irq(&pwq->pool->lock);
1664 raw_spin_unlock_irq(&pwq->pool->lock);
1668 static bool pwq_is_empty(struct pool_workqueue *pwq)
1670 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1673 static void __pwq_activate_work(struct pool_workqueue *pwq,
1674 struct work_struct *work)
1676 unsigned long *wdb = work_data_bits(work);
1678 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1679 trace_workqueue_activate_work(work);
1680 if (list_empty(&pwq->pool->worklist))
1681 pwq->pool->watchdog_ts = jiffies;
1682 move_linked_works(work, &pwq->pool->worklist, NULL);
1683 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1687 * pwq_activate_work - Activate a work item if inactive
1688 * @pwq: pool_workqueue @work belongs to
1689 * @work: work item to activate
1691 * Returns %true if activated. %false if already active.
1693 static bool pwq_activate_work(struct pool_workqueue *pwq,
1694 struct work_struct *work)
1696 struct worker_pool *pool = pwq->pool;
1697 struct wq_node_nr_active *nna;
1699 lockdep_assert_held(&pool->lock);
1701 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE))
1704 nna = wq_node_nr_active(pwq->wq, pool->node);
1706 atomic_inc(&nna->nr);
1709 __pwq_activate_work(pwq, work);
1713 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1715 int max = READ_ONCE(nna->max);
1720 old = atomic_read(&nna->nr);
1723 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1730 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1731 * @pwq: pool_workqueue of interest
1732 * @fill: max_active may have increased, try to increase concurrency level
1734 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1735 * successfully obtained. %false otherwise.
1737 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1739 struct workqueue_struct *wq = pwq->wq;
1740 struct worker_pool *pool = pwq->pool;
1741 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1742 bool obtained = false;
1744 lockdep_assert_held(&pool->lock);
1747 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1748 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1752 if (unlikely(pwq->plugged))
1756 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1757 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1758 * concurrency level. Don't jump the line.
1760 * We need to ignore the pending test after max_active has increased as
1761 * pwq_dec_nr_active() can only maintain the concurrency level but not
1762 * increase it. This is indicated by @fill.
1764 if (!list_empty(&pwq->pending_node) && likely(!fill))
1767 obtained = tryinc_node_nr_active(nna);
1772 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1773 * and try again. The smp_mb() is paired with the implied memory barrier
1774 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1775 * we see the decremented $nna->nr or they see non-empty
1776 * $nna->pending_pwqs.
1778 raw_spin_lock(&nna->lock);
1780 if (list_empty(&pwq->pending_node))
1781 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1782 else if (likely(!fill))
1787 obtained = tryinc_node_nr_active(nna);
1790 * If @fill, @pwq might have already been pending. Being spuriously
1791 * pending in cold paths doesn't affect anything. Let's leave it be.
1793 if (obtained && likely(!fill))
1794 list_del_init(&pwq->pending_node);
1797 raw_spin_unlock(&nna->lock);
1805 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1806 * @pwq: pool_workqueue of interest
1807 * @fill: max_active may have increased, try to increase concurrency level
1809 * Activate the first inactive work item of @pwq if available and allowed by
1812 * Returns %true if an inactive work item has been activated. %false if no
1813 * inactive work item is found or max_active limit is reached.
1815 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1817 struct work_struct *work =
1818 list_first_entry_or_null(&pwq->inactive_works,
1819 struct work_struct, entry);
1821 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1822 __pwq_activate_work(pwq, work);
1830 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1831 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1833 * This function should only be called for ordered workqueues where only the
1834 * oldest pwq is unplugged, the others are plugged to suspend execution to
1835 * ensure proper work item ordering::
1837 * dfl_pwq --------------+ [P] - plugged
1840 * pwqs -> A -> B [P] -> C [P] (newest)
1846 * When the oldest pwq is drained and removed, this function should be called
1847 * to unplug the next oldest one to start its work item execution. Note that
1848 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1849 * the list is the oldest.
1851 static void unplug_oldest_pwq(struct workqueue_struct *wq)
1853 struct pool_workqueue *pwq;
1855 lockdep_assert_held(&wq->mutex);
1857 /* Caller should make sure that pwqs isn't empty before calling */
1858 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1860 raw_spin_lock_irq(&pwq->pool->lock);
1862 pwq->plugged = false;
1863 if (pwq_activate_first_inactive(pwq, true))
1864 kick_pool(pwq->pool);
1866 raw_spin_unlock_irq(&pwq->pool->lock);
1870 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1871 * @nna: wq_node_nr_active to activate a pending pwq for
1872 * @caller_pool: worker_pool the caller is locking
1874 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1875 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1877 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1878 struct worker_pool *caller_pool)
1880 struct worker_pool *locked_pool = caller_pool;
1881 struct pool_workqueue *pwq;
1882 struct work_struct *work;
1884 lockdep_assert_held(&caller_pool->lock);
1886 raw_spin_lock(&nna->lock);
1888 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1889 struct pool_workqueue, pending_node);
1894 * If @pwq is for a different pool than @locked_pool, we need to lock
1895 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1896 * / lock dance. For that, we also need to release @nna->lock as it's
1897 * nested inside pool locks.
1899 if (pwq->pool != locked_pool) {
1900 raw_spin_unlock(&locked_pool->lock);
1901 locked_pool = pwq->pool;
1902 if (!raw_spin_trylock(&locked_pool->lock)) {
1903 raw_spin_unlock(&nna->lock);
1904 raw_spin_lock(&locked_pool->lock);
1905 raw_spin_lock(&nna->lock);
1911 * $pwq may not have any inactive work items due to e.g. cancellations.
1912 * Drop it from pending_pwqs and see if there's another one.
1914 work = list_first_entry_or_null(&pwq->inactive_works,
1915 struct work_struct, entry);
1917 list_del_init(&pwq->pending_node);
1922 * Acquire an nr_active count and activate the inactive work item. If
1923 * $pwq still has inactive work items, rotate it to the end of the
1924 * pending_pwqs so that we round-robin through them. This means that
1925 * inactive work items are not activated in queueing order which is fine
1926 * given that there has never been any ordering across different pwqs.
1928 if (likely(tryinc_node_nr_active(nna))) {
1930 __pwq_activate_work(pwq, work);
1932 if (list_empty(&pwq->inactive_works))
1933 list_del_init(&pwq->pending_node);
1935 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1937 /* if activating a foreign pool, make sure it's running */
1938 if (pwq->pool != caller_pool)
1939 kick_pool(pwq->pool);
1943 raw_spin_unlock(&nna->lock);
1944 if (locked_pool != caller_pool) {
1945 raw_spin_unlock(&locked_pool->lock);
1946 raw_spin_lock(&caller_pool->lock);
1951 * pwq_dec_nr_active - Retire an active count
1952 * @pwq: pool_workqueue of interest
1954 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1955 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1957 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1959 struct worker_pool *pool = pwq->pool;
1960 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1962 lockdep_assert_held(&pool->lock);
1965 * @pwq->nr_active should be decremented for both percpu and unbound
1971 * For a percpu workqueue, it's simple. Just need to kick the first
1972 * inactive work item on @pwq itself.
1975 pwq_activate_first_inactive(pwq, false);
1980 * If @pwq is for an unbound workqueue, it's more complicated because
1981 * multiple pwqs and pools may be sharing the nr_active count. When a
1982 * pwq needs to wait for an nr_active count, it puts itself on
1983 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1984 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1985 * guarantee that either we see non-empty pending_pwqs or they see
1986 * decremented $nna->nr.
1988 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1989 * max_active gets updated. However, it is guaranteed to be equal to or
1990 * larger than @pwq->wq->min_active which is above zero unless freezing.
1991 * This maintains the forward progress guarantee.
1993 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1996 if (!list_empty(&nna->pending_pwqs))
1997 node_activate_pending_pwq(nna, pool);
2001 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
2002 * @pwq: pwq of interest
2003 * @work_data: work_data of work which left the queue
2005 * A work either has completed or is removed from pending queue,
2006 * decrement nr_in_flight of its pwq and handle workqueue flushing.
2009 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
2010 * and thus should be called after all other state updates for the in-flight
2011 * work item is complete.
2014 * raw_spin_lock_irq(pool->lock).
2016 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
2018 int color = get_work_color(work_data);
2020 if (!(work_data & WORK_STRUCT_INACTIVE))
2021 pwq_dec_nr_active(pwq);
2023 pwq->nr_in_flight[color]--;
2025 /* is flush in progress and are we at the flushing tip? */
2026 if (likely(pwq->flush_color != color))
2029 /* are there still in-flight works? */
2030 if (pwq->nr_in_flight[color])
2033 /* this pwq is done, clear flush_color */
2034 pwq->flush_color = -1;
2037 * If this was the last pwq, wake up the first flusher. It
2038 * will handle the rest.
2040 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2041 complete(&pwq->wq->first_flusher->done);
2047 * try_to_grab_pending - steal work item from worklist and disable irq
2048 * @work: work item to steal
2049 * @cflags: %WORK_CANCEL_ flags
2050 * @irq_flags: place to store irq state
2052 * Try to grab PENDING bit of @work. This function can handle @work in any
2053 * stable state - idle, on timer or on worklist.
2057 * ======== ================================================================
2058 * 1 if @work was pending and we successfully stole PENDING
2059 * 0 if @work was idle and we claimed PENDING
2060 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2061 * ======== ================================================================
2064 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2065 * interrupted while holding PENDING and @work off queue, irq must be
2066 * disabled on entry. This, combined with delayed_work->timer being
2067 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2069 * On successful return, >= 0, irq is disabled and the caller is
2070 * responsible for releasing it using local_irq_restore(*@irq_flags).
2072 * This function is safe to call from any context including IRQ handler.
2074 static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2075 unsigned long *irq_flags)
2077 struct worker_pool *pool;
2078 struct pool_workqueue *pwq;
2080 local_irq_save(*irq_flags);
2082 /* try to steal the timer if it exists */
2083 if (cflags & WORK_CANCEL_DELAYED) {
2084 struct delayed_work *dwork = to_delayed_work(work);
2087 * dwork->timer is irqsafe. If del_timer() fails, it's
2088 * guaranteed that the timer is not queued anywhere and not
2089 * running on the local CPU.
2091 if (likely(del_timer(&dwork->timer)))
2095 /* try to claim PENDING the normal way */
2096 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2101 * The queueing is in progress, or it is already queued. Try to
2102 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2104 pool = get_work_pool(work);
2108 raw_spin_lock(&pool->lock);
2110 * work->data is guaranteed to point to pwq only while the work
2111 * item is queued on pwq->wq, and both updating work->data to point
2112 * to pwq on queueing and to pool on dequeueing are done under
2113 * pwq->pool->lock. This in turn guarantees that, if work->data
2114 * points to pwq which is associated with a locked pool, the work
2115 * item is currently queued on that pool.
2117 pwq = get_work_pwq(work);
2118 if (pwq && pwq->pool == pool) {
2119 unsigned long work_data;
2121 debug_work_deactivate(work);
2124 * A cancelable inactive work item must be in the
2125 * pwq->inactive_works since a queued barrier can't be
2126 * canceled (see the comments in insert_wq_barrier()).
2128 * An inactive work item cannot be grabbed directly because
2129 * it might have linked barrier work items which, if left
2130 * on the inactive_works list, will confuse pwq->nr_active
2131 * management later on and cause stall. Make sure the work
2132 * item is activated before grabbing.
2134 pwq_activate_work(pwq, work);
2136 list_del_init(&work->entry);
2139 * work->data points to pwq iff queued. Let's point to pool. As
2140 * this destroys work->data needed by the next step, stash it.
2142 work_data = *work_data_bits(work);
2143 set_work_pool_and_keep_pending(work, pool->id,
2144 pool_offq_flags(pool));
2146 /* must be the last step, see the function comment */
2147 pwq_dec_nr_in_flight(pwq, work_data);
2149 raw_spin_unlock(&pool->lock);
2153 raw_spin_unlock(&pool->lock);
2156 local_irq_restore(*irq_flags);
2161 * work_grab_pending - steal work item from worklist and disable irq
2162 * @work: work item to steal
2163 * @cflags: %WORK_CANCEL_ flags
2164 * @irq_flags: place to store IRQ state
2166 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2169 * Can be called from any context. IRQ is disabled on return with IRQ state
2170 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2171 * local_irq_restore().
2173 * Returns %true if @work was pending. %false if idle.
2175 static bool work_grab_pending(struct work_struct *work, u32 cflags,
2176 unsigned long *irq_flags)
2181 ret = try_to_grab_pending(work, cflags, irq_flags);
2189 * insert_work - insert a work into a pool
2190 * @pwq: pwq @work belongs to
2191 * @work: work to insert
2192 * @head: insertion point
2193 * @extra_flags: extra WORK_STRUCT_* flags to set
2195 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2196 * work_struct flags.
2199 * raw_spin_lock_irq(pool->lock).
2201 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2202 struct list_head *head, unsigned int extra_flags)
2204 debug_work_activate(work);
2206 /* record the work call stack in order to print it in KASAN reports */
2207 kasan_record_aux_stack_noalloc(work);
2209 /* we own @work, set data and link */
2210 set_work_pwq(work, pwq, extra_flags);
2211 list_add_tail(&work->entry, head);
2216 * Test whether @work is being queued from another work executing on the
2219 static bool is_chained_work(struct workqueue_struct *wq)
2221 struct worker *worker;
2223 worker = current_wq_worker();
2225 * Return %true iff I'm a worker executing a work item on @wq. If
2226 * I'm @worker, it's safe to dereference it without locking.
2228 return worker && worker->current_pwq->wq == wq;
2232 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2233 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2234 * avoid perturbing sensitive tasks.
2236 static int wq_select_unbound_cpu(int cpu)
2240 if (likely(!wq_debug_force_rr_cpu)) {
2241 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2244 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2247 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2248 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2249 if (unlikely(new_cpu >= nr_cpu_ids)) {
2250 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2251 if (unlikely(new_cpu >= nr_cpu_ids))
2254 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2259 static void __queue_work(int cpu, struct workqueue_struct *wq,
2260 struct work_struct *work)
2262 struct pool_workqueue *pwq;
2263 struct worker_pool *last_pool, *pool;
2264 unsigned int work_flags;
2265 unsigned int req_cpu = cpu;
2268 * While a work item is PENDING && off queue, a task trying to
2269 * steal the PENDING will busy-loop waiting for it to either get
2270 * queued or lose PENDING. Grabbing PENDING and queueing should
2271 * happen with IRQ disabled.
2273 lockdep_assert_irqs_disabled();
2276 * For a draining wq, only works from the same workqueue are
2277 * allowed. The __WQ_DESTROYING helps to spot the issue that
2278 * queues a new work item to a wq after destroy_workqueue(wq).
2280 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2281 WARN_ON_ONCE(!is_chained_work(wq))))
2285 /* pwq which will be used unless @work is executing elsewhere */
2286 if (req_cpu == WORK_CPU_UNBOUND) {
2287 if (wq->flags & WQ_UNBOUND)
2288 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2290 cpu = raw_smp_processor_id();
2293 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2297 * If @work was previously on a different pool, it might still be
2298 * running there, in which case the work needs to be queued on that
2299 * pool to guarantee non-reentrancy.
2301 last_pool = get_work_pool(work);
2302 if (last_pool && last_pool != pool) {
2303 struct worker *worker;
2305 raw_spin_lock(&last_pool->lock);
2307 worker = find_worker_executing_work(last_pool, work);
2309 if (worker && worker->current_pwq->wq == wq) {
2310 pwq = worker->current_pwq;
2312 WARN_ON_ONCE(pool != last_pool);
2314 /* meh... not running there, queue here */
2315 raw_spin_unlock(&last_pool->lock);
2316 raw_spin_lock(&pool->lock);
2319 raw_spin_lock(&pool->lock);
2323 * pwq is determined and locked. For unbound pools, we could have raced
2324 * with pwq release and it could already be dead. If its refcnt is zero,
2325 * repeat pwq selection. Note that unbound pwqs never die without
2326 * another pwq replacing it in cpu_pwq or while work items are executing
2327 * on it, so the retrying is guaranteed to make forward-progress.
2329 if (unlikely(!pwq->refcnt)) {
2330 if (wq->flags & WQ_UNBOUND) {
2331 raw_spin_unlock(&pool->lock);
2336 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2340 /* pwq determined, queue */
2341 trace_workqueue_queue_work(req_cpu, pwq, work);
2343 if (WARN_ON(!list_empty(&work->entry)))
2346 pwq->nr_in_flight[pwq->work_color]++;
2347 work_flags = work_color_to_flags(pwq->work_color);
2350 * Limit the number of concurrently active work items to max_active.
2351 * @work must also queue behind existing inactive work items to maintain
2352 * ordering when max_active changes. See wq_adjust_max_active().
2354 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2355 if (list_empty(&pool->worklist))
2356 pool->watchdog_ts = jiffies;
2358 trace_workqueue_activate_work(work);
2359 insert_work(pwq, work, &pool->worklist, work_flags);
2362 work_flags |= WORK_STRUCT_INACTIVE;
2363 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2367 raw_spin_unlock(&pool->lock);
2371 static bool clear_pending_if_disabled(struct work_struct *work)
2373 unsigned long data = *work_data_bits(work);
2374 struct work_offq_data offqd;
2376 if (likely((data & WORK_STRUCT_PWQ) ||
2377 !(data & WORK_OFFQ_DISABLE_MASK)))
2380 work_offqd_unpack(&offqd, data);
2381 set_work_pool_and_clear_pending(work, offqd.pool_id,
2382 work_offqd_pack_flags(&offqd));
2387 * queue_work_on - queue work on specific cpu
2388 * @cpu: CPU number to execute work on
2389 * @wq: workqueue to use
2390 * @work: work to queue
2392 * We queue the work to a specific CPU, the caller must ensure it
2393 * can't go away. Callers that fail to ensure that the specified
2394 * CPU cannot go away will execute on a randomly chosen CPU.
2395 * But note well that callers specifying a CPU that never has been
2396 * online will get a splat.
2398 * Return: %false if @work was already on a queue, %true otherwise.
2400 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2401 struct work_struct *work)
2404 unsigned long irq_flags;
2406 local_irq_save(irq_flags);
2408 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2409 !clear_pending_if_disabled(work)) {
2410 __queue_work(cpu, wq, work);
2414 local_irq_restore(irq_flags);
2417 EXPORT_SYMBOL(queue_work_on);
2420 * select_numa_node_cpu - Select a CPU based on NUMA node
2421 * @node: NUMA node ID that we want to select a CPU from
2423 * This function will attempt to find a "random" cpu available on a given
2424 * node. If there are no CPUs available on the given node it will return
2425 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2426 * available CPU if we need to schedule this work.
2428 static int select_numa_node_cpu(int node)
2432 /* Delay binding to CPU if node is not valid or online */
2433 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2434 return WORK_CPU_UNBOUND;
2436 /* Use local node/cpu if we are already there */
2437 cpu = raw_smp_processor_id();
2438 if (node == cpu_to_node(cpu))
2441 /* Use "random" otherwise know as "first" online CPU of node */
2442 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2444 /* If CPU is valid return that, otherwise just defer */
2445 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2449 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2450 * @node: NUMA node that we are targeting the work for
2451 * @wq: workqueue to use
2452 * @work: work to queue
2454 * We queue the work to a "random" CPU within a given NUMA node. The basic
2455 * idea here is to provide a way to somehow associate work with a given
2458 * This function will only make a best effort attempt at getting this onto
2459 * the right NUMA node. If no node is requested or the requested node is
2460 * offline then we just fall back to standard queue_work behavior.
2462 * Currently the "random" CPU ends up being the first available CPU in the
2463 * intersection of cpu_online_mask and the cpumask of the node, unless we
2464 * are running on the node. In that case we just use the current CPU.
2466 * Return: %false if @work was already on a queue, %true otherwise.
2468 bool queue_work_node(int node, struct workqueue_struct *wq,
2469 struct work_struct *work)
2471 unsigned long irq_flags;
2475 * This current implementation is specific to unbound workqueues.
2476 * Specifically we only return the first available CPU for a given
2477 * node instead of cycling through individual CPUs within the node.
2479 * If this is used with a per-cpu workqueue then the logic in
2480 * workqueue_select_cpu_near would need to be updated to allow for
2481 * some round robin type logic.
2483 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2485 local_irq_save(irq_flags);
2487 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2488 !clear_pending_if_disabled(work)) {
2489 int cpu = select_numa_node_cpu(node);
2491 __queue_work(cpu, wq, work);
2495 local_irq_restore(irq_flags);
2498 EXPORT_SYMBOL_GPL(queue_work_node);
2500 void delayed_work_timer_fn(struct timer_list *t)
2502 struct delayed_work *dwork = from_timer(dwork, t, timer);
2504 /* should have been called from irqsafe timer with irq already off */
2505 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2507 EXPORT_SYMBOL(delayed_work_timer_fn);
2509 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2510 struct delayed_work *dwork, unsigned long delay)
2512 struct timer_list *timer = &dwork->timer;
2513 struct work_struct *work = &dwork->work;
2516 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2517 WARN_ON_ONCE(timer_pending(timer));
2518 WARN_ON_ONCE(!list_empty(&work->entry));
2521 * If @delay is 0, queue @dwork->work immediately. This is for
2522 * both optimization and correctness. The earliest @timer can
2523 * expire is on the closest next tick and delayed_work users depend
2524 * on that there's no such delay when @delay is 0.
2527 __queue_work(cpu, wq, &dwork->work);
2533 timer->expires = jiffies + delay;
2535 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2536 /* If the current cpu is a housekeeping cpu, use it. */
2537 cpu = smp_processor_id();
2538 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2539 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2540 add_timer_on(timer, cpu);
2542 if (likely(cpu == WORK_CPU_UNBOUND))
2543 add_timer_global(timer);
2545 add_timer_on(timer, cpu);
2550 * queue_delayed_work_on - queue work on specific CPU after delay
2551 * @cpu: CPU number to execute work on
2552 * @wq: workqueue to use
2553 * @dwork: work to queue
2554 * @delay: number of jiffies to wait before queueing
2556 * Return: %false if @work was already on a queue, %true otherwise. If
2557 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2560 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2561 struct delayed_work *dwork, unsigned long delay)
2563 struct work_struct *work = &dwork->work;
2565 unsigned long irq_flags;
2567 /* read the comment in __queue_work() */
2568 local_irq_save(irq_flags);
2570 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2571 !clear_pending_if_disabled(work)) {
2572 __queue_delayed_work(cpu, wq, dwork, delay);
2576 local_irq_restore(irq_flags);
2579 EXPORT_SYMBOL(queue_delayed_work_on);
2582 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2583 * @cpu: CPU number to execute work on
2584 * @wq: workqueue to use
2585 * @dwork: work to queue
2586 * @delay: number of jiffies to wait before queueing
2588 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2589 * modify @dwork's timer so that it expires after @delay. If @delay is
2590 * zero, @work is guaranteed to be scheduled immediately regardless of its
2593 * Return: %false if @dwork was idle and queued, %true if @dwork was
2594 * pending and its timer was modified.
2596 * This function is safe to call from any context including IRQ handler.
2597 * See try_to_grab_pending() for details.
2599 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2600 struct delayed_work *dwork, unsigned long delay)
2602 unsigned long irq_flags;
2605 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags);
2607 if (!clear_pending_if_disabled(&dwork->work))
2608 __queue_delayed_work(cpu, wq, dwork, delay);
2610 local_irq_restore(irq_flags);
2613 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2615 static void rcu_work_rcufn(struct rcu_head *rcu)
2617 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2619 /* read the comment in __queue_work() */
2620 local_irq_disable();
2621 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2626 * queue_rcu_work - queue work after a RCU grace period
2627 * @wq: workqueue to use
2628 * @rwork: work to queue
2630 * Return: %false if @rwork was already pending, %true otherwise. Note
2631 * that a full RCU grace period is guaranteed only after a %true return.
2632 * While @rwork is guaranteed to be executed after a %false return, the
2633 * execution may happen before a full RCU grace period has passed.
2635 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2637 struct work_struct *work = &rwork->work;
2640 * rcu_work can't be canceled or disabled. Warn if the user reached
2641 * inside @rwork and disabled the inner work.
2643 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2644 !WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2646 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2652 EXPORT_SYMBOL(queue_rcu_work);
2654 static struct worker *alloc_worker(int node)
2656 struct worker *worker;
2658 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2660 INIT_LIST_HEAD(&worker->entry);
2661 INIT_LIST_HEAD(&worker->scheduled);
2662 INIT_LIST_HEAD(&worker->node);
2663 /* on creation a worker is in !idle && prep state */
2664 worker->flags = WORKER_PREP;
2669 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2671 if (pool->cpu < 0 && pool->attrs->affn_strict)
2672 return pool->attrs->__pod_cpumask;
2674 return pool->attrs->cpumask;
2678 * worker_attach_to_pool() - attach a worker to a pool
2679 * @worker: worker to be attached
2680 * @pool: the target pool
2682 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2683 * cpu-binding of @worker are kept coordinated with the pool across
2686 static void worker_attach_to_pool(struct worker *worker,
2687 struct worker_pool *pool)
2689 mutex_lock(&wq_pool_attach_mutex);
2692 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2693 * across this function. See the comments above the flag definition for
2694 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2696 if (pool->flags & POOL_DISASSOCIATED) {
2697 worker->flags |= WORKER_UNBOUND;
2699 WARN_ON_ONCE(pool->flags & POOL_BH);
2700 kthread_set_per_cpu(worker->task, pool->cpu);
2703 if (worker->rescue_wq)
2704 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2706 list_add_tail(&worker->node, &pool->workers);
2707 worker->pool = pool;
2709 mutex_unlock(&wq_pool_attach_mutex);
2713 * worker_detach_from_pool() - detach a worker from its pool
2714 * @worker: worker which is attached to its pool
2716 * Undo the attaching which had been done in worker_attach_to_pool(). The
2717 * caller worker shouldn't access to the pool after detached except it has
2718 * other reference to the pool.
2720 static void worker_detach_from_pool(struct worker *worker)
2722 struct worker_pool *pool = worker->pool;
2723 struct completion *detach_completion = NULL;
2725 /* there is one permanent BH worker per CPU which should never detach */
2726 WARN_ON_ONCE(pool->flags & POOL_BH);
2728 mutex_lock(&wq_pool_attach_mutex);
2730 kthread_set_per_cpu(worker->task, -1);
2731 list_del(&worker->node);
2732 worker->pool = NULL;
2734 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2735 detach_completion = pool->detach_completion;
2736 mutex_unlock(&wq_pool_attach_mutex);
2738 /* clear leftover flags without pool->lock after it is detached */
2739 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2741 if (detach_completion)
2742 complete(detach_completion);
2746 * create_worker - create a new workqueue worker
2747 * @pool: pool the new worker will belong to
2749 * Create and start a new worker which is attached to @pool.
2752 * Might sleep. Does GFP_KERNEL allocations.
2755 * Pointer to the newly created worker.
2757 static struct worker *create_worker(struct worker_pool *pool)
2759 struct worker *worker;
2763 /* ID is needed to determine kthread name */
2764 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2766 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2771 worker = alloc_worker(pool->node);
2773 pr_err_once("workqueue: Failed to allocate a worker\n");
2779 if (!(pool->flags & POOL_BH)) {
2781 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2782 pool->attrs->nice < 0 ? "H" : "");
2784 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2786 worker->task = kthread_create_on_node(worker_thread, worker,
2787 pool->node, "kworker/%s", id_buf);
2788 if (IS_ERR(worker->task)) {
2789 if (PTR_ERR(worker->task) == -EINTR) {
2790 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2793 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2799 set_user_nice(worker->task, pool->attrs->nice);
2800 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2803 /* successful, attach the worker to the pool */
2804 worker_attach_to_pool(worker, pool);
2806 /* start the newly created worker */
2807 raw_spin_lock_irq(&pool->lock);
2809 worker->pool->nr_workers++;
2810 worker_enter_idle(worker);
2813 * @worker is waiting on a completion in kthread() and will trigger hung
2814 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2815 * wake it up explicitly.
2818 wake_up_process(worker->task);
2820 raw_spin_unlock_irq(&pool->lock);
2825 ida_free(&pool->worker_ida, id);
2830 static void unbind_worker(struct worker *worker)
2832 lockdep_assert_held(&wq_pool_attach_mutex);
2834 kthread_set_per_cpu(worker->task, -1);
2835 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2836 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2838 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2841 static void wake_dying_workers(struct list_head *cull_list)
2843 struct worker *worker, *tmp;
2845 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2846 list_del_init(&worker->entry);
2847 unbind_worker(worker);
2849 * If the worker was somehow already running, then it had to be
2850 * in pool->idle_list when set_worker_dying() happened or we
2851 * wouldn't have gotten here.
2853 * Thus, the worker must either have observed the WORKER_DIE
2854 * flag, or have set its state to TASK_IDLE. Either way, the
2855 * below will be observed by the worker and is safe to do
2856 * outside of pool->lock.
2858 wake_up_process(worker->task);
2863 * set_worker_dying - Tag a worker for destruction
2864 * @worker: worker to be destroyed
2865 * @list: transfer worker away from its pool->idle_list and into list
2867 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2871 * raw_spin_lock_irq(pool->lock).
2873 static void set_worker_dying(struct worker *worker, struct list_head *list)
2875 struct worker_pool *pool = worker->pool;
2877 lockdep_assert_held(&pool->lock);
2878 lockdep_assert_held(&wq_pool_attach_mutex);
2880 /* sanity check frenzy */
2881 if (WARN_ON(worker->current_work) ||
2882 WARN_ON(!list_empty(&worker->scheduled)) ||
2883 WARN_ON(!(worker->flags & WORKER_IDLE)))
2889 worker->flags |= WORKER_DIE;
2891 list_move(&worker->entry, list);
2892 list_move(&worker->node, &pool->dying_workers);
2896 * idle_worker_timeout - check if some idle workers can now be deleted.
2897 * @t: The pool's idle_timer that just expired
2899 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2900 * worker_leave_idle(), as a worker flicking between idle and active while its
2901 * pool is at the too_many_workers() tipping point would cause too much timer
2902 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2903 * it expire and re-evaluate things from there.
2905 static void idle_worker_timeout(struct timer_list *t)
2907 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2908 bool do_cull = false;
2910 if (work_pending(&pool->idle_cull_work))
2913 raw_spin_lock_irq(&pool->lock);
2915 if (too_many_workers(pool)) {
2916 struct worker *worker;
2917 unsigned long expires;
2919 /* idle_list is kept in LIFO order, check the last one */
2920 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2921 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2922 do_cull = !time_before(jiffies, expires);
2925 mod_timer(&pool->idle_timer, expires);
2927 raw_spin_unlock_irq(&pool->lock);
2930 queue_work(system_unbound_wq, &pool->idle_cull_work);
2934 * idle_cull_fn - cull workers that have been idle for too long.
2935 * @work: the pool's work for handling these idle workers
2937 * This goes through a pool's idle workers and gets rid of those that have been
2938 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2940 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2941 * culled, so this also resets worker affinity. This requires a sleepable
2942 * context, hence the split between timer callback and work item.
2944 static void idle_cull_fn(struct work_struct *work)
2946 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2947 LIST_HEAD(cull_list);
2950 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2951 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2952 * path. This is required as a previously-preempted worker could run after
2953 * set_worker_dying() has happened but before wake_dying_workers() did.
2955 mutex_lock(&wq_pool_attach_mutex);
2956 raw_spin_lock_irq(&pool->lock);
2958 while (too_many_workers(pool)) {
2959 struct worker *worker;
2960 unsigned long expires;
2962 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2963 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2965 if (time_before(jiffies, expires)) {
2966 mod_timer(&pool->idle_timer, expires);
2970 set_worker_dying(worker, &cull_list);
2973 raw_spin_unlock_irq(&pool->lock);
2974 wake_dying_workers(&cull_list);
2975 mutex_unlock(&wq_pool_attach_mutex);
2978 static void send_mayday(struct work_struct *work)
2980 struct pool_workqueue *pwq = get_work_pwq(work);
2981 struct workqueue_struct *wq = pwq->wq;
2983 lockdep_assert_held(&wq_mayday_lock);
2988 /* mayday mayday mayday */
2989 if (list_empty(&pwq->mayday_node)) {
2991 * If @pwq is for an unbound wq, its base ref may be put at
2992 * any time due to an attribute change. Pin @pwq until the
2993 * rescuer is done with it.
2996 list_add_tail(&pwq->mayday_node, &wq->maydays);
2997 wake_up_process(wq->rescuer->task);
2998 pwq->stats[PWQ_STAT_MAYDAY]++;
3002 static void pool_mayday_timeout(struct timer_list *t)
3004 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3005 struct work_struct *work;
3007 raw_spin_lock_irq(&pool->lock);
3008 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3010 if (need_to_create_worker(pool)) {
3012 * We've been trying to create a new worker but
3013 * haven't been successful. We might be hitting an
3014 * allocation deadlock. Send distress signals to
3017 list_for_each_entry(work, &pool->worklist, entry)
3021 raw_spin_unlock(&wq_mayday_lock);
3022 raw_spin_unlock_irq(&pool->lock);
3024 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3028 * maybe_create_worker - create a new worker if necessary
3029 * @pool: pool to create a new worker for
3031 * Create a new worker for @pool if necessary. @pool is guaranteed to
3032 * have at least one idle worker on return from this function. If
3033 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034 * sent to all rescuers with works scheduled on @pool to resolve
3035 * possible allocation deadlock.
3037 * On return, need_to_create_worker() is guaranteed to be %false and
3038 * may_start_working() %true.
3041 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3042 * multiple times. Does GFP_KERNEL allocations. Called only from
3045 static void maybe_create_worker(struct worker_pool *pool)
3046 __releases(&pool->lock)
3047 __acquires(&pool->lock)
3050 raw_spin_unlock_irq(&pool->lock);
3052 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3056 if (create_worker(pool) || !need_to_create_worker(pool))
3059 schedule_timeout_interruptible(CREATE_COOLDOWN);
3061 if (!need_to_create_worker(pool))
3065 del_timer_sync(&pool->mayday_timer);
3066 raw_spin_lock_irq(&pool->lock);
3068 * This is necessary even after a new worker was just successfully
3069 * created as @pool->lock was dropped and the new worker might have
3070 * already become busy.
3072 if (need_to_create_worker(pool))
3077 * manage_workers - manage worker pool
3080 * Assume the manager role and manage the worker pool @worker belongs
3081 * to. At any given time, there can be only zero or one manager per
3082 * pool. The exclusion is handled automatically by this function.
3084 * The caller can safely start processing works on false return. On
3085 * true return, it's guaranteed that need_to_create_worker() is false
3086 * and may_start_working() is true.
3089 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3090 * multiple times. Does GFP_KERNEL allocations.
3093 * %false if the pool doesn't need management and the caller can safely
3094 * start processing works, %true if management function was performed and
3095 * the conditions that the caller verified before calling the function may
3096 * no longer be true.
3098 static bool manage_workers(struct worker *worker)
3100 struct worker_pool *pool = worker->pool;
3102 if (pool->flags & POOL_MANAGER_ACTIVE)
3105 pool->flags |= POOL_MANAGER_ACTIVE;
3106 pool->manager = worker;
3108 maybe_create_worker(pool);
3110 pool->manager = NULL;
3111 pool->flags &= ~POOL_MANAGER_ACTIVE;
3112 rcuwait_wake_up(&manager_wait);
3117 * process_one_work - process single work
3119 * @work: work to process
3121 * Process @work. This function contains all the logics necessary to
3122 * process a single work including synchronization against and
3123 * interaction with other workers on the same cpu, queueing and
3124 * flushing. As long as context requirement is met, any worker can
3125 * call this function to process a work.
3128 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3130 static void process_one_work(struct worker *worker, struct work_struct *work)
3131 __releases(&pool->lock)
3132 __acquires(&pool->lock)
3134 struct pool_workqueue *pwq = get_work_pwq(work);
3135 struct worker_pool *pool = worker->pool;
3136 unsigned long work_data;
3137 int lockdep_start_depth, rcu_start_depth;
3138 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3139 #ifdef CONFIG_LOCKDEP
3141 * It is permissible to free the struct work_struct from
3142 * inside the function that is called from it, this we need to
3143 * take into account for lockdep too. To avoid bogus "held
3144 * lock freed" warnings as well as problems when looking into
3145 * work->lockdep_map, make a copy and use that here.
3147 struct lockdep_map lockdep_map;
3149 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3151 /* ensure we're on the correct CPU */
3152 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3153 raw_smp_processor_id() != pool->cpu);
3155 /* claim and dequeue */
3156 debug_work_deactivate(work);
3157 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158 worker->current_work = work;
3159 worker->current_func = work->func;
3160 worker->current_pwq = pwq;
3162 worker->current_at = worker->task->se.sum_exec_runtime;
3163 work_data = *work_data_bits(work);
3164 worker->current_color = get_work_color(work_data);
3167 * Record wq name for cmdline and debug reporting, may get
3168 * overridden through set_worker_desc().
3170 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3172 list_del_init(&work->entry);
3175 * CPU intensive works don't participate in concurrency management.
3176 * They're the scheduler's responsibility. This takes @worker out
3177 * of concurrency management and the next code block will chain
3178 * execution of the pending work items.
3180 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3181 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3184 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3185 * since nr_running would always be >= 1 at this point. This is used to
3186 * chain execution of the pending work items for WORKER_NOT_RUNNING
3187 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3192 * Record the last pool and clear PENDING which should be the last
3193 * update to @work. Also, do this inside @pool->lock so that
3194 * PENDING and queued state changes happen together while IRQ is
3197 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool));
3199 pwq->stats[PWQ_STAT_STARTED]++;
3200 raw_spin_unlock_irq(&pool->lock);
3202 rcu_start_depth = rcu_preempt_depth();
3203 lockdep_start_depth = lockdep_depth(current);
3204 /* see drain_dead_softirq_workfn() */
3206 lock_map_acquire(&pwq->wq->lockdep_map);
3207 lock_map_acquire(&lockdep_map);
3209 * Strictly speaking we should mark the invariant state without holding
3210 * any locks, that is, before these two lock_map_acquire()'s.
3212 * However, that would result in:
3219 * Which would create W1->C->W1 dependencies, even though there is no
3220 * actual deadlock possible. There are two solutions, using a
3221 * read-recursive acquire on the work(queue) 'locks', but this will then
3222 * hit the lockdep limitation on recursive locks, or simply discard
3225 * AFAICT there is no possible deadlock scenario between the
3226 * flush_work() and complete() primitives (except for single-threaded
3227 * workqueues), so hiding them isn't a problem.
3229 lockdep_invariant_state(true);
3230 trace_workqueue_execute_start(work);
3231 worker->current_func(work);
3233 * While we must be careful to not use "work" after this, the trace
3234 * point will only record its address.
3236 trace_workqueue_execute_end(work, worker->current_func);
3237 pwq->stats[PWQ_STAT_COMPLETED]++;
3238 lock_map_release(&lockdep_map);
3240 lock_map_release(&pwq->wq->lockdep_map);
3242 if (unlikely((worker->task && in_atomic()) ||
3243 lockdep_depth(current) != lockdep_start_depth ||
3244 rcu_preempt_depth() != rcu_start_depth)) {
3245 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3246 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3247 current->comm, task_pid_nr(current), preempt_count(),
3248 lockdep_start_depth, lockdep_depth(current),
3249 rcu_start_depth, rcu_preempt_depth(),
3250 worker->current_func);
3251 debug_show_held_locks(current);
3256 * The following prevents a kworker from hogging CPU on !PREEMPTION
3257 * kernels, where a requeueing work item waiting for something to
3258 * happen could deadlock with stop_machine as such work item could
3259 * indefinitely requeue itself while all other CPUs are trapped in
3260 * stop_machine. At the same time, report a quiescent RCU state so
3261 * the same condition doesn't freeze RCU.
3266 raw_spin_lock_irq(&pool->lock);
3269 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3270 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3271 * wq_cpu_intensive_thresh_us. Clear it.
3273 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3275 /* tag the worker for identification in schedule() */
3276 worker->last_func = worker->current_func;
3278 /* we're done with it, release */
3279 hash_del(&worker->hentry);
3280 worker->current_work = NULL;
3281 worker->current_func = NULL;
3282 worker->current_pwq = NULL;
3283 worker->current_color = INT_MAX;
3285 /* must be the last step, see the function comment */
3286 pwq_dec_nr_in_flight(pwq, work_data);
3290 * process_scheduled_works - process scheduled works
3293 * Process all scheduled works. Please note that the scheduled list
3294 * may change while processing a work, so this function repeatedly
3295 * fetches a work from the top and executes it.
3298 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3301 static void process_scheduled_works(struct worker *worker)
3303 struct work_struct *work;
3306 while ((work = list_first_entry_or_null(&worker->scheduled,
3307 struct work_struct, entry))) {
3309 worker->pool->watchdog_ts = jiffies;
3312 process_one_work(worker, work);
3316 static void set_pf_worker(bool val)
3318 mutex_lock(&wq_pool_attach_mutex);
3320 current->flags |= PF_WQ_WORKER;
3322 current->flags &= ~PF_WQ_WORKER;
3323 mutex_unlock(&wq_pool_attach_mutex);
3327 * worker_thread - the worker thread function
3330 * The worker thread function. All workers belong to a worker_pool -
3331 * either a per-cpu one or dynamic unbound one. These workers process all
3332 * work items regardless of their specific target workqueue. The only
3333 * exception is work items which belong to workqueues with a rescuer which
3334 * will be explained in rescuer_thread().
3338 static int worker_thread(void *__worker)
3340 struct worker *worker = __worker;
3341 struct worker_pool *pool = worker->pool;
3343 /* tell the scheduler that this is a workqueue worker */
3344 set_pf_worker(true);
3346 raw_spin_lock_irq(&pool->lock);
3348 /* am I supposed to die? */
3349 if (unlikely(worker->flags & WORKER_DIE)) {
3350 raw_spin_unlock_irq(&pool->lock);
3351 set_pf_worker(false);
3353 set_task_comm(worker->task, "kworker/dying");
3354 ida_free(&pool->worker_ida, worker->id);
3355 worker_detach_from_pool(worker);
3356 WARN_ON_ONCE(!list_empty(&worker->entry));
3361 worker_leave_idle(worker);
3363 /* no more worker necessary? */
3364 if (!need_more_worker(pool))
3367 /* do we need to manage? */
3368 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3372 * ->scheduled list can only be filled while a worker is
3373 * preparing to process a work or actually processing it.
3374 * Make sure nobody diddled with it while I was sleeping.
3376 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3379 * Finish PREP stage. We're guaranteed to have at least one idle
3380 * worker or that someone else has already assumed the manager
3381 * role. This is where @worker starts participating in concurrency
3382 * management if applicable and concurrency management is restored
3383 * after being rebound. See rebind_workers() for details.
3385 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3388 struct work_struct *work =
3389 list_first_entry(&pool->worklist,
3390 struct work_struct, entry);
3392 if (assign_work(work, worker, NULL))
3393 process_scheduled_works(worker);
3394 } while (keep_working(pool));
3396 worker_set_flags(worker, WORKER_PREP);
3399 * pool->lock is held and there's no work to process and no need to
3400 * manage, sleep. Workers are woken up only while holding
3401 * pool->lock or from local cpu, so setting the current state
3402 * before releasing pool->lock is enough to prevent losing any
3405 worker_enter_idle(worker);
3406 __set_current_state(TASK_IDLE);
3407 raw_spin_unlock_irq(&pool->lock);
3413 * rescuer_thread - the rescuer thread function
3416 * Workqueue rescuer thread function. There's one rescuer for each
3417 * workqueue which has WQ_MEM_RECLAIM set.
3419 * Regular work processing on a pool may block trying to create a new
3420 * worker which uses GFP_KERNEL allocation which has slight chance of
3421 * developing into deadlock if some works currently on the same queue
3422 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3423 * the problem rescuer solves.
3425 * When such condition is possible, the pool summons rescuers of all
3426 * workqueues which have works queued on the pool and let them process
3427 * those works so that forward progress can be guaranteed.
3429 * This should happen rarely.
3433 static int rescuer_thread(void *__rescuer)
3435 struct worker *rescuer = __rescuer;
3436 struct workqueue_struct *wq = rescuer->rescue_wq;
3439 set_user_nice(current, RESCUER_NICE_LEVEL);
3442 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3443 * doesn't participate in concurrency management.
3445 set_pf_worker(true);
3447 set_current_state(TASK_IDLE);
3450 * By the time the rescuer is requested to stop, the workqueue
3451 * shouldn't have any work pending, but @wq->maydays may still have
3452 * pwq(s) queued. This can happen by non-rescuer workers consuming
3453 * all the work items before the rescuer got to them. Go through
3454 * @wq->maydays processing before acting on should_stop so that the
3455 * list is always empty on exit.
3457 should_stop = kthread_should_stop();
3459 /* see whether any pwq is asking for help */
3460 raw_spin_lock_irq(&wq_mayday_lock);
3462 while (!list_empty(&wq->maydays)) {
3463 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3464 struct pool_workqueue, mayday_node);
3465 struct worker_pool *pool = pwq->pool;
3466 struct work_struct *work, *n;
3468 __set_current_state(TASK_RUNNING);
3469 list_del_init(&pwq->mayday_node);
3471 raw_spin_unlock_irq(&wq_mayday_lock);
3473 worker_attach_to_pool(rescuer, pool);
3475 raw_spin_lock_irq(&pool->lock);
3478 * Slurp in all works issued via this workqueue and
3481 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3482 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3483 if (get_work_pwq(work) == pwq &&
3484 assign_work(work, rescuer, &n))
3485 pwq->stats[PWQ_STAT_RESCUED]++;
3488 if (!list_empty(&rescuer->scheduled)) {
3489 process_scheduled_works(rescuer);
3492 * The above execution of rescued work items could
3493 * have created more to rescue through
3494 * pwq_activate_first_inactive() or chained
3495 * queueing. Let's put @pwq back on mayday list so
3496 * that such back-to-back work items, which may be
3497 * being used to relieve memory pressure, don't
3498 * incur MAYDAY_INTERVAL delay inbetween.
3500 if (pwq->nr_active && need_to_create_worker(pool)) {
3501 raw_spin_lock(&wq_mayday_lock);
3503 * Queue iff we aren't racing destruction
3504 * and somebody else hasn't queued it already.
3506 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3508 list_add_tail(&pwq->mayday_node, &wq->maydays);
3510 raw_spin_unlock(&wq_mayday_lock);
3515 * Put the reference grabbed by send_mayday(). @pool won't
3516 * go away while we're still attached to it.
3521 * Leave this pool. Notify regular workers; otherwise, we end up
3522 * with 0 concurrency and stalling the execution.
3526 raw_spin_unlock_irq(&pool->lock);
3528 worker_detach_from_pool(rescuer);
3530 raw_spin_lock_irq(&wq_mayday_lock);
3533 raw_spin_unlock_irq(&wq_mayday_lock);
3536 __set_current_state(TASK_RUNNING);
3537 set_pf_worker(false);
3541 /* rescuers should never participate in concurrency management */
3542 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3547 static void bh_worker(struct worker *worker)
3549 struct worker_pool *pool = worker->pool;
3550 int nr_restarts = BH_WORKER_RESTARTS;
3551 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3553 raw_spin_lock_irq(&pool->lock);
3554 worker_leave_idle(worker);
3557 * This function follows the structure of worker_thread(). See there for
3558 * explanations on each step.
3560 if (!need_more_worker(pool))
3563 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3564 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3567 struct work_struct *work =
3568 list_first_entry(&pool->worklist,
3569 struct work_struct, entry);
3571 if (assign_work(work, worker, NULL))
3572 process_scheduled_works(worker);
3573 } while (keep_working(pool) &&
3574 --nr_restarts && time_before(jiffies, end));
3576 worker_set_flags(worker, WORKER_PREP);
3578 worker_enter_idle(worker);
3580 raw_spin_unlock_irq(&pool->lock);
3584 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3586 * This is currently called from tasklet[_hi]action() and thus is also called
3587 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3588 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3591 * After full conversion, we'll add worker->softirq_action, directly use the
3592 * softirq action and obtain the worker pointer from the softirq_action pointer.
3594 void workqueue_softirq_action(bool highpri)
3596 struct worker_pool *pool =
3597 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3598 if (need_more_worker(pool))
3599 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3602 struct wq_drain_dead_softirq_work {
3603 struct work_struct work;
3604 struct worker_pool *pool;
3605 struct completion done;
3608 static void drain_dead_softirq_workfn(struct work_struct *work)
3610 struct wq_drain_dead_softirq_work *dead_work =
3611 container_of(work, struct wq_drain_dead_softirq_work, work);
3612 struct worker_pool *pool = dead_work->pool;
3616 * @pool's CPU is dead and we want to execute its still pending work
3617 * items from this BH work item which is running on a different CPU. As
3618 * its CPU is dead, @pool can't be kicked and, as work execution path
3619 * will be nested, a lockdep annotation needs to be suppressed. Mark
3620 * @pool with %POOL_BH_DRAINING for the special treatments.
3622 raw_spin_lock_irq(&pool->lock);
3623 pool->flags |= POOL_BH_DRAINING;
3624 raw_spin_unlock_irq(&pool->lock);
3626 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3628 raw_spin_lock_irq(&pool->lock);
3629 pool->flags &= ~POOL_BH_DRAINING;
3630 repeat = need_more_worker(pool);
3631 raw_spin_unlock_irq(&pool->lock);
3634 * bh_worker() might hit consecutive execution limit and bail. If there
3635 * still are pending work items, reschedule self and return so that we
3636 * don't hog this CPU's BH.
3639 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3640 queue_work(system_bh_highpri_wq, work);
3642 queue_work(system_bh_wq, work);
3644 complete(&dead_work->done);
3649 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3650 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3651 * have to worry about draining overlapping with CPU coming back online or
3652 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3653 * on). Let's keep it simple and drain them synchronously. These are BH work
3654 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3656 void workqueue_softirq_dead(unsigned int cpu)
3660 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3661 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3662 struct wq_drain_dead_softirq_work dead_work;
3664 if (!need_more_worker(pool))
3667 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3668 dead_work.pool = pool;
3669 init_completion(&dead_work.done);
3671 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3672 queue_work(system_bh_highpri_wq, &dead_work.work);
3674 queue_work(system_bh_wq, &dead_work.work);
3676 wait_for_completion(&dead_work.done);
3677 destroy_work_on_stack(&dead_work.work);
3682 * check_flush_dependency - check for flush dependency sanity
3683 * @target_wq: workqueue being flushed
3684 * @target_work: work item being flushed (NULL for workqueue flushes)
3686 * %current is trying to flush the whole @target_wq or @target_work on it.
3687 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3688 * reclaiming memory or running on a workqueue which doesn't have
3689 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3692 static void check_flush_dependency(struct workqueue_struct *target_wq,
3693 struct work_struct *target_work)
3695 work_func_t target_func = target_work ? target_work->func : NULL;
3696 struct worker *worker;
3698 if (target_wq->flags & WQ_MEM_RECLAIM)
3701 worker = current_wq_worker();
3703 WARN_ONCE(current->flags & PF_MEMALLOC,
3704 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3705 current->pid, current->comm, target_wq->name, target_func);
3706 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3707 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3708 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3709 worker->current_pwq->wq->name, worker->current_func,
3710 target_wq->name, target_func);
3714 struct work_struct work;
3715 struct completion done;
3716 struct task_struct *task; /* purely informational */
3719 static void wq_barrier_func(struct work_struct *work)
3721 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3722 complete(&barr->done);
3726 * insert_wq_barrier - insert a barrier work
3727 * @pwq: pwq to insert barrier into
3728 * @barr: wq_barrier to insert
3729 * @target: target work to attach @barr to
3730 * @worker: worker currently executing @target, NULL if @target is not executing
3732 * @barr is linked to @target such that @barr is completed only after
3733 * @target finishes execution. Please note that the ordering
3734 * guarantee is observed only with respect to @target and on the local
3737 * Currently, a queued barrier can't be canceled. This is because
3738 * try_to_grab_pending() can't determine whether the work to be
3739 * grabbed is at the head of the queue and thus can't clear LINKED
3740 * flag of the previous work while there must be a valid next work
3741 * after a work with LINKED flag set.
3743 * Note that when @worker is non-NULL, @target may be modified
3744 * underneath us, so we can't reliably determine pwq from @target.
3747 * raw_spin_lock_irq(pool->lock).
3749 static void insert_wq_barrier(struct pool_workqueue *pwq,
3750 struct wq_barrier *barr,
3751 struct work_struct *target, struct worker *worker)
3753 static __maybe_unused struct lock_class_key bh_key, thr_key;
3754 unsigned int work_flags = 0;
3755 unsigned int work_color;
3756 struct list_head *head;
3759 * debugobject calls are safe here even with pool->lock locked
3760 * as we know for sure that this will not trigger any of the
3761 * checks and call back into the fixup functions where we
3764 * BH and threaded workqueues need separate lockdep keys to avoid
3765 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3768 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3769 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3770 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3772 init_completion_map(&barr->done, &target->lockdep_map);
3774 barr->task = current;
3776 /* The barrier work item does not participate in nr_active. */
3777 work_flags |= WORK_STRUCT_INACTIVE;
3780 * If @target is currently being executed, schedule the
3781 * barrier to the worker; otherwise, put it after @target.
3784 head = worker->scheduled.next;
3785 work_color = worker->current_color;
3787 unsigned long *bits = work_data_bits(target);
3789 head = target->entry.next;
3790 /* there can already be other linked works, inherit and set */
3791 work_flags |= *bits & WORK_STRUCT_LINKED;
3792 work_color = get_work_color(*bits);
3793 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3796 pwq->nr_in_flight[work_color]++;
3797 work_flags |= work_color_to_flags(work_color);
3799 insert_work(pwq, &barr->work, head, work_flags);
3803 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3804 * @wq: workqueue being flushed
3805 * @flush_color: new flush color, < 0 for no-op
3806 * @work_color: new work color, < 0 for no-op
3808 * Prepare pwqs for workqueue flushing.
3810 * If @flush_color is non-negative, flush_color on all pwqs should be
3811 * -1. If no pwq has in-flight commands at the specified color, all
3812 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3813 * has in flight commands, its pwq->flush_color is set to
3814 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3815 * wakeup logic is armed and %true is returned.
3817 * The caller should have initialized @wq->first_flusher prior to
3818 * calling this function with non-negative @flush_color. If
3819 * @flush_color is negative, no flush color update is done and %false
3822 * If @work_color is non-negative, all pwqs should have the same
3823 * work_color which is previous to @work_color and all will be
3824 * advanced to @work_color.
3827 * mutex_lock(wq->mutex).
3830 * %true if @flush_color >= 0 and there's something to flush. %false
3833 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3834 int flush_color, int work_color)
3837 struct pool_workqueue *pwq;
3839 if (flush_color >= 0) {
3840 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3841 atomic_set(&wq->nr_pwqs_to_flush, 1);
3844 for_each_pwq(pwq, wq) {
3845 struct worker_pool *pool = pwq->pool;
3847 raw_spin_lock_irq(&pool->lock);
3849 if (flush_color >= 0) {
3850 WARN_ON_ONCE(pwq->flush_color != -1);
3852 if (pwq->nr_in_flight[flush_color]) {
3853 pwq->flush_color = flush_color;
3854 atomic_inc(&wq->nr_pwqs_to_flush);
3859 if (work_color >= 0) {
3860 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3861 pwq->work_color = work_color;
3864 raw_spin_unlock_irq(&pool->lock);
3867 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3868 complete(&wq->first_flusher->done);
3873 static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3875 #ifdef CONFIG_LOCKDEP
3876 if (wq->flags & WQ_BH)
3879 lock_map_acquire(&wq->lockdep_map);
3880 lock_map_release(&wq->lockdep_map);
3882 if (wq->flags & WQ_BH)
3887 static void touch_work_lockdep_map(struct work_struct *work,
3888 struct workqueue_struct *wq)
3890 #ifdef CONFIG_LOCKDEP
3891 if (wq->flags & WQ_BH)
3894 lock_map_acquire(&work->lockdep_map);
3895 lock_map_release(&work->lockdep_map);
3897 if (wq->flags & WQ_BH)
3903 * __flush_workqueue - ensure that any scheduled work has run to completion.
3904 * @wq: workqueue to flush
3906 * This function sleeps until all work items which were queued on entry
3907 * have finished execution, but it is not livelocked by new incoming ones.
3909 void __flush_workqueue(struct workqueue_struct *wq)
3911 struct wq_flusher this_flusher = {
3912 .list = LIST_HEAD_INIT(this_flusher.list),
3914 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3918 if (WARN_ON(!wq_online))
3921 touch_wq_lockdep_map(wq);
3923 mutex_lock(&wq->mutex);
3926 * Start-to-wait phase
3928 next_color = work_next_color(wq->work_color);
3930 if (next_color != wq->flush_color) {
3932 * Color space is not full. The current work_color
3933 * becomes our flush_color and work_color is advanced
3936 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3937 this_flusher.flush_color = wq->work_color;
3938 wq->work_color = next_color;
3940 if (!wq->first_flusher) {
3941 /* no flush in progress, become the first flusher */
3942 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3944 wq->first_flusher = &this_flusher;
3946 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3948 /* nothing to flush, done */
3949 wq->flush_color = next_color;
3950 wq->first_flusher = NULL;
3955 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3956 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3957 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3961 * Oops, color space is full, wait on overflow queue.
3962 * The next flush completion will assign us
3963 * flush_color and transfer to flusher_queue.
3965 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3968 check_flush_dependency(wq, NULL);
3970 mutex_unlock(&wq->mutex);
3972 wait_for_completion(&this_flusher.done);
3975 * Wake-up-and-cascade phase
3977 * First flushers are responsible for cascading flushes and
3978 * handling overflow. Non-first flushers can simply return.
3980 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3983 mutex_lock(&wq->mutex);
3985 /* we might have raced, check again with mutex held */
3986 if (wq->first_flusher != &this_flusher)
3989 WRITE_ONCE(wq->first_flusher, NULL);
3991 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3992 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3995 struct wq_flusher *next, *tmp;
3997 /* complete all the flushers sharing the current flush color */
3998 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3999 if (next->flush_color != wq->flush_color)
4001 list_del_init(&next->list);
4002 complete(&next->done);
4005 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4006 wq->flush_color != work_next_color(wq->work_color));
4008 /* this flush_color is finished, advance by one */
4009 wq->flush_color = work_next_color(wq->flush_color);
4011 /* one color has been freed, handle overflow queue */
4012 if (!list_empty(&wq->flusher_overflow)) {
4014 * Assign the same color to all overflowed
4015 * flushers, advance work_color and append to
4016 * flusher_queue. This is the start-to-wait
4017 * phase for these overflowed flushers.
4019 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4020 tmp->flush_color = wq->work_color;
4022 wq->work_color = work_next_color(wq->work_color);
4024 list_splice_tail_init(&wq->flusher_overflow,
4025 &wq->flusher_queue);
4026 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4029 if (list_empty(&wq->flusher_queue)) {
4030 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4035 * Need to flush more colors. Make the next flusher
4036 * the new first flusher and arm pwqs.
4038 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4039 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4041 list_del_init(&next->list);
4042 wq->first_flusher = next;
4044 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4048 * Meh... this color is already done, clear first
4049 * flusher and repeat cascading.
4051 wq->first_flusher = NULL;
4055 mutex_unlock(&wq->mutex);
4057 EXPORT_SYMBOL(__flush_workqueue);
4060 * drain_workqueue - drain a workqueue
4061 * @wq: workqueue to drain
4063 * Wait until the workqueue becomes empty. While draining is in progress,
4064 * only chain queueing is allowed. IOW, only currently pending or running
4065 * work items on @wq can queue further work items on it. @wq is flushed
4066 * repeatedly until it becomes empty. The number of flushing is determined
4067 * by the depth of chaining and should be relatively short. Whine if it
4070 void drain_workqueue(struct workqueue_struct *wq)
4072 unsigned int flush_cnt = 0;
4073 struct pool_workqueue *pwq;
4076 * __queue_work() needs to test whether there are drainers, is much
4077 * hotter than drain_workqueue() and already looks at @wq->flags.
4078 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4080 mutex_lock(&wq->mutex);
4081 if (!wq->nr_drainers++)
4082 wq->flags |= __WQ_DRAINING;
4083 mutex_unlock(&wq->mutex);
4085 __flush_workqueue(wq);
4087 mutex_lock(&wq->mutex);
4089 for_each_pwq(pwq, wq) {
4092 raw_spin_lock_irq(&pwq->pool->lock);
4093 drained = pwq_is_empty(pwq);
4094 raw_spin_unlock_irq(&pwq->pool->lock);
4099 if (++flush_cnt == 10 ||
4100 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4101 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4102 wq->name, __func__, flush_cnt);
4104 mutex_unlock(&wq->mutex);
4108 if (!--wq->nr_drainers)
4109 wq->flags &= ~__WQ_DRAINING;
4110 mutex_unlock(&wq->mutex);
4112 EXPORT_SYMBOL_GPL(drain_workqueue);
4114 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4117 struct worker *worker = NULL;
4118 struct worker_pool *pool;
4119 struct pool_workqueue *pwq;
4120 struct workqueue_struct *wq;
4123 pool = get_work_pool(work);
4129 raw_spin_lock_irq(&pool->lock);
4130 /* see the comment in try_to_grab_pending() with the same code */
4131 pwq = get_work_pwq(work);
4133 if (unlikely(pwq->pool != pool))
4136 worker = find_worker_executing_work(pool, work);
4139 pwq = worker->current_pwq;
4143 check_flush_dependency(wq, work);
4145 insert_wq_barrier(pwq, barr, work, worker);
4146 raw_spin_unlock_irq(&pool->lock);
4148 touch_work_lockdep_map(work, wq);
4151 * Force a lock recursion deadlock when using flush_work() inside a
4152 * single-threaded or rescuer equipped workqueue.
4154 * For single threaded workqueues the deadlock happens when the work
4155 * is after the work issuing the flush_work(). For rescuer equipped
4156 * workqueues the deadlock happens when the rescuer stalls, blocking
4159 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4160 touch_wq_lockdep_map(wq);
4165 raw_spin_unlock_irq(&pool->lock);
4170 static bool __flush_work(struct work_struct *work, bool from_cancel)
4172 struct wq_barrier barr;
4175 if (WARN_ON(!wq_online))
4178 if (WARN_ON(!work->func))
4181 if (!start_flush_work(work, &barr, from_cancel))
4185 * start_flush_work() returned %true. If @from_cancel is set, we know
4186 * that @work must have been executing during start_flush_work() and
4187 * can't currently be queued. Its data must contain OFFQ bits. If @work
4188 * was queued on a BH workqueue, we also know that it was running in the
4189 * BH context and thus can be busy-waited.
4191 data = *work_data_bits(work);
4193 !WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_BH)) {
4195 * On RT, prevent a live lock when %current preempted soft
4196 * interrupt processing or prevents ksoftirqd from running by
4197 * keeping flipping BH. If the BH work item runs on a different
4198 * CPU then this has no effect other than doing the BH
4199 * disable/enable dance for nothing. This is copied from
4200 * kernel/softirq.c::tasklet_unlock_spin_wait().
4202 while (!try_wait_for_completion(&barr.done)) {
4203 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4211 wait_for_completion(&barr.done);
4214 destroy_work_on_stack(&barr.work);
4219 * flush_work - wait for a work to finish executing the last queueing instance
4220 * @work: the work to flush
4222 * Wait until @work has finished execution. @work is guaranteed to be idle
4223 * on return if it hasn't been requeued since flush started.
4226 * %true if flush_work() waited for the work to finish execution,
4227 * %false if it was already idle.
4229 bool flush_work(struct work_struct *work)
4232 return __flush_work(work, false);
4234 EXPORT_SYMBOL_GPL(flush_work);
4237 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4238 * @dwork: the delayed work to flush
4240 * Delayed timer is cancelled and the pending work is queued for
4241 * immediate execution. Like flush_work(), this function only
4242 * considers the last queueing instance of @dwork.
4245 * %true if flush_work() waited for the work to finish execution,
4246 * %false if it was already idle.
4248 bool flush_delayed_work(struct delayed_work *dwork)
4250 local_irq_disable();
4251 if (del_timer_sync(&dwork->timer))
4252 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4254 return flush_work(&dwork->work);
4256 EXPORT_SYMBOL(flush_delayed_work);
4259 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4260 * @rwork: the rcu work to flush
4263 * %true if flush_rcu_work() waited for the work to finish execution,
4264 * %false if it was already idle.
4266 bool flush_rcu_work(struct rcu_work *rwork)
4268 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4270 flush_work(&rwork->work);
4273 return flush_work(&rwork->work);
4276 EXPORT_SYMBOL(flush_rcu_work);
4278 static void work_offqd_disable(struct work_offq_data *offqd)
4280 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4282 if (likely(offqd->disable < max))
4285 WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4288 static void work_offqd_enable(struct work_offq_data *offqd)
4290 if (likely(offqd->disable > 0))
4293 WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4296 static bool __cancel_work(struct work_struct *work, u32 cflags)
4298 struct work_offq_data offqd;
4299 unsigned long irq_flags;
4302 ret = work_grab_pending(work, cflags, &irq_flags);
4304 work_offqd_unpack(&offqd, *work_data_bits(work));
4306 if (cflags & WORK_CANCEL_DISABLE)
4307 work_offqd_disable(&offqd);
4309 set_work_pool_and_clear_pending(work, offqd.pool_id,
4310 work_offqd_pack_flags(&offqd));
4311 local_irq_restore(irq_flags);
4315 static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4319 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE);
4321 if (*work_data_bits(work) & WORK_OFFQ_BH)
4322 WARN_ON_ONCE(in_hardirq());
4327 * Skip __flush_work() during early boot when we know that @work isn't
4328 * executing. This allows canceling during early boot.
4331 __flush_work(work, true);
4333 if (!(cflags & WORK_CANCEL_DISABLE))
4340 * See cancel_delayed_work()
4342 bool cancel_work(struct work_struct *work)
4344 return __cancel_work(work, 0);
4346 EXPORT_SYMBOL(cancel_work);
4349 * cancel_work_sync - cancel a work and wait for it to finish
4350 * @work: the work to cancel
4352 * Cancel @work and wait for its execution to finish. This function can be used
4353 * even if the work re-queues itself or migrates to another workqueue. On return
4354 * from this function, @work is guaranteed to be not pending or executing on any
4355 * CPU as long as there aren't racing enqueues.
4357 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4358 * Use cancel_delayed_work_sync() instead.
4360 * Must be called from a sleepable context if @work was last queued on a non-BH
4361 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4362 * if @work was last queued on a BH workqueue.
4364 * Returns %true if @work was pending, %false otherwise.
4366 bool cancel_work_sync(struct work_struct *work)
4368 return __cancel_work_sync(work, 0);
4370 EXPORT_SYMBOL_GPL(cancel_work_sync);
4373 * cancel_delayed_work - cancel a delayed work
4374 * @dwork: delayed_work to cancel
4376 * Kill off a pending delayed_work.
4378 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4382 * The work callback function may still be running on return, unless
4383 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4384 * use cancel_delayed_work_sync() to wait on it.
4386 * This function is safe to call from any context including IRQ handler.
4388 bool cancel_delayed_work(struct delayed_work *dwork)
4390 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4392 EXPORT_SYMBOL(cancel_delayed_work);
4395 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4396 * @dwork: the delayed work cancel
4398 * This is cancel_work_sync() for delayed works.
4401 * %true if @dwork was pending, %false otherwise.
4403 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4405 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4407 EXPORT_SYMBOL(cancel_delayed_work_sync);
4410 * disable_work - Disable and cancel a work item
4411 * @work: work item to disable
4413 * Disable @work by incrementing its disable count and cancel it if currently
4414 * pending. As long as the disable count is non-zero, any attempt to queue @work
4415 * will fail and return %false. The maximum supported disable depth is 2 to the
4416 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4418 * Can be called from any context. Returns %true if @work was pending, %false
4421 bool disable_work(struct work_struct *work)
4423 return __cancel_work(work, WORK_CANCEL_DISABLE);
4425 EXPORT_SYMBOL_GPL(disable_work);
4428 * disable_work_sync - Disable, cancel and drain a work item
4429 * @work: work item to disable
4431 * Similar to disable_work() but also wait for @work to finish if currently
4434 * Must be called from a sleepable context if @work was last queued on a non-BH
4435 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4436 * if @work was last queued on a BH workqueue.
4438 * Returns %true if @work was pending, %false otherwise.
4440 bool disable_work_sync(struct work_struct *work)
4442 return __cancel_work_sync(work, WORK_CANCEL_DISABLE);
4444 EXPORT_SYMBOL_GPL(disable_work_sync);
4447 * enable_work - Enable a work item
4448 * @work: work item to enable
4450 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4451 * only be queued if its disable count is 0.
4453 * Can be called from any context. Returns %true if the disable count reached 0.
4454 * Otherwise, %false.
4456 bool enable_work(struct work_struct *work)
4458 struct work_offq_data offqd;
4459 unsigned long irq_flags;
4461 work_grab_pending(work, 0, &irq_flags);
4463 work_offqd_unpack(&offqd, *work_data_bits(work));
4464 work_offqd_enable(&offqd);
4465 set_work_pool_and_clear_pending(work, offqd.pool_id,
4466 work_offqd_pack_flags(&offqd));
4467 local_irq_restore(irq_flags);
4469 return !offqd.disable;
4471 EXPORT_SYMBOL_GPL(enable_work);
4474 * disable_delayed_work - Disable and cancel a delayed work item
4475 * @dwork: delayed work item to disable
4477 * disable_work() for delayed work items.
4479 bool disable_delayed_work(struct delayed_work *dwork)
4481 return __cancel_work(&dwork->work,
4482 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4484 EXPORT_SYMBOL_GPL(disable_delayed_work);
4487 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4488 * @dwork: delayed work item to disable
4490 * disable_work_sync() for delayed work items.
4492 bool disable_delayed_work_sync(struct delayed_work *dwork)
4494 return __cancel_work_sync(&dwork->work,
4495 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4497 EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4500 * enable_delayed_work - Enable a delayed work item
4501 * @dwork: delayed work item to enable
4503 * enable_work() for delayed work items.
4505 bool enable_delayed_work(struct delayed_work *dwork)
4507 return enable_work(&dwork->work);
4509 EXPORT_SYMBOL_GPL(enable_delayed_work);
4512 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4513 * @func: the function to call
4515 * schedule_on_each_cpu() executes @func on each online CPU using the
4516 * system workqueue and blocks until all CPUs have completed.
4517 * schedule_on_each_cpu() is very slow.
4520 * 0 on success, -errno on failure.
4522 int schedule_on_each_cpu(work_func_t func)
4525 struct work_struct __percpu *works;
4527 works = alloc_percpu(struct work_struct);
4533 for_each_online_cpu(cpu) {
4534 struct work_struct *work = per_cpu_ptr(works, cpu);
4536 INIT_WORK(work, func);
4537 schedule_work_on(cpu, work);
4540 for_each_online_cpu(cpu)
4541 flush_work(per_cpu_ptr(works, cpu));
4549 * execute_in_process_context - reliably execute the routine with user context
4550 * @fn: the function to execute
4551 * @ew: guaranteed storage for the execute work structure (must
4552 * be available when the work executes)
4554 * Executes the function immediately if process context is available,
4555 * otherwise schedules the function for delayed execution.
4557 * Return: 0 - function was executed
4558 * 1 - function was scheduled for execution
4560 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4562 if (!in_interrupt()) {
4567 INIT_WORK(&ew->work, fn);
4568 schedule_work(&ew->work);
4572 EXPORT_SYMBOL_GPL(execute_in_process_context);
4575 * free_workqueue_attrs - free a workqueue_attrs
4576 * @attrs: workqueue_attrs to free
4578 * Undo alloc_workqueue_attrs().
4580 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4583 free_cpumask_var(attrs->cpumask);
4584 free_cpumask_var(attrs->__pod_cpumask);
4590 * alloc_workqueue_attrs - allocate a workqueue_attrs
4592 * Allocate a new workqueue_attrs, initialize with default settings and
4595 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4597 struct workqueue_attrs *alloc_workqueue_attrs(void)
4599 struct workqueue_attrs *attrs;
4601 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4604 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4606 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4609 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4610 attrs->affn_scope = WQ_AFFN_DFL;
4613 free_workqueue_attrs(attrs);
4617 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4618 const struct workqueue_attrs *from)
4620 to->nice = from->nice;
4621 cpumask_copy(to->cpumask, from->cpumask);
4622 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4623 to->affn_strict = from->affn_strict;
4626 * Unlike hash and equality test, copying shouldn't ignore wq-only
4627 * fields as copying is used for both pool and wq attrs. Instead,
4628 * get_unbound_pool() explicitly clears the fields.
4630 to->affn_scope = from->affn_scope;
4631 to->ordered = from->ordered;
4635 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4636 * comments in 'struct workqueue_attrs' definition.
4638 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4640 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4641 attrs->ordered = false;
4642 if (attrs->affn_strict)
4643 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4646 /* hash value of the content of @attr */
4647 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4651 hash = jhash_1word(attrs->nice, hash);
4652 hash = jhash_1word(attrs->affn_strict, hash);
4653 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4654 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4655 if (!attrs->affn_strict)
4656 hash = jhash(cpumask_bits(attrs->cpumask),
4657 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4661 /* content equality test */
4662 static bool wqattrs_equal(const struct workqueue_attrs *a,
4663 const struct workqueue_attrs *b)
4665 if (a->nice != b->nice)
4667 if (a->affn_strict != b->affn_strict)
4669 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4671 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask))
4676 /* Update @attrs with actually available CPUs */
4677 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4678 const cpumask_t *unbound_cpumask)
4681 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4682 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4685 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4686 if (unlikely(cpumask_empty(attrs->cpumask)))
4687 cpumask_copy(attrs->cpumask, unbound_cpumask);
4690 /* find wq_pod_type to use for @attrs */
4691 static const struct wq_pod_type *
4692 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4694 enum wq_affn_scope scope;
4695 struct wq_pod_type *pt;
4697 /* to synchronize access to wq_affn_dfl */
4698 lockdep_assert_held(&wq_pool_mutex);
4700 if (attrs->affn_scope == WQ_AFFN_DFL)
4701 scope = wq_affn_dfl;
4703 scope = attrs->affn_scope;
4705 pt = &wq_pod_types[scope];
4707 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4708 likely(pt->nr_pods))
4712 * Before workqueue_init_topology(), only SYSTEM is available which is
4713 * initialized in workqueue_init_early().
4715 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4716 BUG_ON(!pt->nr_pods);
4721 * init_worker_pool - initialize a newly zalloc'd worker_pool
4722 * @pool: worker_pool to initialize
4724 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4726 * Return: 0 on success, -errno on failure. Even on failure, all fields
4727 * inside @pool proper are initialized and put_unbound_pool() can be called
4728 * on @pool safely to release it.
4730 static int init_worker_pool(struct worker_pool *pool)
4732 raw_spin_lock_init(&pool->lock);
4735 pool->node = NUMA_NO_NODE;
4736 pool->flags |= POOL_DISASSOCIATED;
4737 pool->watchdog_ts = jiffies;
4738 INIT_LIST_HEAD(&pool->worklist);
4739 INIT_LIST_HEAD(&pool->idle_list);
4740 hash_init(pool->busy_hash);
4742 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4743 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4745 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4747 INIT_LIST_HEAD(&pool->workers);
4748 INIT_LIST_HEAD(&pool->dying_workers);
4750 ida_init(&pool->worker_ida);
4751 INIT_HLIST_NODE(&pool->hash_node);
4754 /* shouldn't fail above this point */
4755 pool->attrs = alloc_workqueue_attrs();
4759 wqattrs_clear_for_pool(pool->attrs);
4764 #ifdef CONFIG_LOCKDEP
4765 static void wq_init_lockdep(struct workqueue_struct *wq)
4769 lockdep_register_key(&wq->key);
4770 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4772 lock_name = wq->name;
4774 wq->lock_name = lock_name;
4775 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4778 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4780 lockdep_unregister_key(&wq->key);
4783 static void wq_free_lockdep(struct workqueue_struct *wq)
4785 if (wq->lock_name != wq->name)
4786 kfree(wq->lock_name);
4789 static void wq_init_lockdep(struct workqueue_struct *wq)
4793 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4797 static void wq_free_lockdep(struct workqueue_struct *wq)
4802 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4806 for_each_node(node) {
4807 kfree(nna_ar[node]);
4808 nna_ar[node] = NULL;
4811 kfree(nna_ar[nr_node_ids]);
4812 nna_ar[nr_node_ids] = NULL;
4815 static void init_node_nr_active(struct wq_node_nr_active *nna)
4817 nna->max = WQ_DFL_MIN_ACTIVE;
4818 atomic_set(&nna->nr, 0);
4819 raw_spin_lock_init(&nna->lock);
4820 INIT_LIST_HEAD(&nna->pending_pwqs);
4824 * Each node's nr_active counter will be accessed mostly from its own node and
4825 * should be allocated in the node.
4827 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4829 struct wq_node_nr_active *nna;
4832 for_each_node(node) {
4833 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4836 init_node_nr_active(nna);
4840 /* [nr_node_ids] is used as the fallback */
4841 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4844 init_node_nr_active(nna);
4845 nna_ar[nr_node_ids] = nna;
4850 free_node_nr_active(nna_ar);
4854 static void rcu_free_wq(struct rcu_head *rcu)
4856 struct workqueue_struct *wq =
4857 container_of(rcu, struct workqueue_struct, rcu);
4859 if (wq->flags & WQ_UNBOUND)
4860 free_node_nr_active(wq->node_nr_active);
4862 wq_free_lockdep(wq);
4863 free_percpu(wq->cpu_pwq);
4864 free_workqueue_attrs(wq->unbound_attrs);
4868 static void rcu_free_pool(struct rcu_head *rcu)
4870 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4872 ida_destroy(&pool->worker_ida);
4873 free_workqueue_attrs(pool->attrs);
4878 * put_unbound_pool - put a worker_pool
4879 * @pool: worker_pool to put
4881 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4882 * safe manner. get_unbound_pool() calls this function on its failure path
4883 * and this function should be able to release pools which went through,
4884 * successfully or not, init_worker_pool().
4886 * Should be called with wq_pool_mutex held.
4888 static void put_unbound_pool(struct worker_pool *pool)
4890 DECLARE_COMPLETION_ONSTACK(detach_completion);
4891 struct worker *worker;
4892 LIST_HEAD(cull_list);
4894 lockdep_assert_held(&wq_pool_mutex);
4900 if (WARN_ON(!(pool->cpu < 0)) ||
4901 WARN_ON(!list_empty(&pool->worklist)))
4904 /* release id and unhash */
4906 idr_remove(&worker_pool_idr, pool->id);
4907 hash_del(&pool->hash_node);
4910 * Become the manager and destroy all workers. This prevents
4911 * @pool's workers from blocking on attach_mutex. We're the last
4912 * manager and @pool gets freed with the flag set.
4914 * Having a concurrent manager is quite unlikely to happen as we can
4915 * only get here with
4916 * pwq->refcnt == pool->refcnt == 0
4917 * which implies no work queued to the pool, which implies no worker can
4918 * become the manager. However a worker could have taken the role of
4919 * manager before the refcnts dropped to 0, since maybe_create_worker()
4923 rcuwait_wait_event(&manager_wait,
4924 !(pool->flags & POOL_MANAGER_ACTIVE),
4925 TASK_UNINTERRUPTIBLE);
4927 mutex_lock(&wq_pool_attach_mutex);
4928 raw_spin_lock_irq(&pool->lock);
4929 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4930 pool->flags |= POOL_MANAGER_ACTIVE;
4933 raw_spin_unlock_irq(&pool->lock);
4934 mutex_unlock(&wq_pool_attach_mutex);
4937 while ((worker = first_idle_worker(pool)))
4938 set_worker_dying(worker, &cull_list);
4939 WARN_ON(pool->nr_workers || pool->nr_idle);
4940 raw_spin_unlock_irq(&pool->lock);
4942 wake_dying_workers(&cull_list);
4944 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4945 pool->detach_completion = &detach_completion;
4946 mutex_unlock(&wq_pool_attach_mutex);
4948 if (pool->detach_completion)
4949 wait_for_completion(pool->detach_completion);
4951 /* shut down the timers */
4952 del_timer_sync(&pool->idle_timer);
4953 cancel_work_sync(&pool->idle_cull_work);
4954 del_timer_sync(&pool->mayday_timer);
4956 /* RCU protected to allow dereferences from get_work_pool() */
4957 call_rcu(&pool->rcu, rcu_free_pool);
4961 * get_unbound_pool - get a worker_pool with the specified attributes
4962 * @attrs: the attributes of the worker_pool to get
4964 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4965 * reference count and return it. If there already is a matching
4966 * worker_pool, it will be used; otherwise, this function attempts to
4969 * Should be called with wq_pool_mutex held.
4971 * Return: On success, a worker_pool with the same attributes as @attrs.
4972 * On failure, %NULL.
4974 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4976 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4977 u32 hash = wqattrs_hash(attrs);
4978 struct worker_pool *pool;
4979 int pod, node = NUMA_NO_NODE;
4981 lockdep_assert_held(&wq_pool_mutex);
4983 /* do we already have a matching pool? */
4984 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4985 if (wqattrs_equal(pool->attrs, attrs)) {
4991 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4992 for (pod = 0; pod < pt->nr_pods; pod++) {
4993 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4994 node = pt->pod_node[pod];
4999 /* nope, create a new one */
5000 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
5001 if (!pool || init_worker_pool(pool) < 0)
5005 copy_workqueue_attrs(pool->attrs, attrs);
5006 wqattrs_clear_for_pool(pool->attrs);
5008 if (worker_pool_assign_id(pool) < 0)
5011 /* create and start the initial worker */
5012 if (wq_online && !create_worker(pool))
5016 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5021 put_unbound_pool(pool);
5025 static void rcu_free_pwq(struct rcu_head *rcu)
5027 kmem_cache_free(pwq_cache,
5028 container_of(rcu, struct pool_workqueue, rcu));
5032 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5033 * refcnt and needs to be destroyed.
5035 static void pwq_release_workfn(struct kthread_work *work)
5037 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5039 struct workqueue_struct *wq = pwq->wq;
5040 struct worker_pool *pool = pwq->pool;
5041 bool is_last = false;
5044 * When @pwq is not linked, it doesn't hold any reference to the
5045 * @wq, and @wq is invalid to access.
5047 if (!list_empty(&pwq->pwqs_node)) {
5048 mutex_lock(&wq->mutex);
5049 list_del_rcu(&pwq->pwqs_node);
5050 is_last = list_empty(&wq->pwqs);
5053 * For ordered workqueue with a plugged dfl_pwq, restart it now.
5055 if (!is_last && (wq->flags & __WQ_ORDERED))
5056 unplug_oldest_pwq(wq);
5058 mutex_unlock(&wq->mutex);
5061 if (wq->flags & WQ_UNBOUND) {
5062 mutex_lock(&wq_pool_mutex);
5063 put_unbound_pool(pool);
5064 mutex_unlock(&wq_pool_mutex);
5067 if (!list_empty(&pwq->pending_node)) {
5068 struct wq_node_nr_active *nna =
5069 wq_node_nr_active(pwq->wq, pwq->pool->node);
5071 raw_spin_lock_irq(&nna->lock);
5072 list_del_init(&pwq->pending_node);
5073 raw_spin_unlock_irq(&nna->lock);
5076 call_rcu(&pwq->rcu, rcu_free_pwq);
5079 * If we're the last pwq going away, @wq is already dead and no one
5080 * is gonna access it anymore. Schedule RCU free.
5083 wq_unregister_lockdep(wq);
5084 call_rcu(&wq->rcu, rcu_free_wq);
5088 /* initialize newly allocated @pwq which is associated with @wq and @pool */
5089 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5090 struct worker_pool *pool)
5092 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5094 memset(pwq, 0, sizeof(*pwq));
5098 pwq->flush_color = -1;
5100 INIT_LIST_HEAD(&pwq->inactive_works);
5101 INIT_LIST_HEAD(&pwq->pending_node);
5102 INIT_LIST_HEAD(&pwq->pwqs_node);
5103 INIT_LIST_HEAD(&pwq->mayday_node);
5104 kthread_init_work(&pwq->release_work, pwq_release_workfn);
5107 /* sync @pwq with the current state of its associated wq and link it */
5108 static void link_pwq(struct pool_workqueue *pwq)
5110 struct workqueue_struct *wq = pwq->wq;
5112 lockdep_assert_held(&wq->mutex);
5114 /* may be called multiple times, ignore if already linked */
5115 if (!list_empty(&pwq->pwqs_node))
5118 /* set the matching work_color */
5119 pwq->work_color = wq->work_color;
5122 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5125 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5126 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5127 const struct workqueue_attrs *attrs)
5129 struct worker_pool *pool;
5130 struct pool_workqueue *pwq;
5132 lockdep_assert_held(&wq_pool_mutex);
5134 pool = get_unbound_pool(attrs);
5138 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5140 put_unbound_pool(pool);
5144 init_pwq(pwq, wq, pool);
5149 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5150 * @attrs: the wq_attrs of the default pwq of the target workqueue
5151 * @cpu: the target CPU
5152 * @cpu_going_down: if >= 0, the CPU to consider as offline
5154 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
5155 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
5156 * The result is stored in @attrs->__pod_cpumask.
5158 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5159 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5160 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5162 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5164 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
5167 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5168 int pod = pt->cpu_pod[cpu];
5170 /* does @pod have any online CPUs @attrs wants? */
5171 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5172 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
5173 if (cpu_going_down >= 0)
5174 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
5176 if (cpumask_empty(attrs->__pod_cpumask)) {
5177 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5181 /* yeap, return possible CPUs in @pod that @attrs wants */
5182 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
5184 if (cpumask_empty(attrs->__pod_cpumask))
5185 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
5186 "possible intersect\n");
5189 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5190 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5191 int cpu, struct pool_workqueue *pwq)
5193 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5194 struct pool_workqueue *old_pwq;
5196 lockdep_assert_held(&wq_pool_mutex);
5197 lockdep_assert_held(&wq->mutex);
5199 /* link_pwq() can handle duplicate calls */
5202 old_pwq = rcu_access_pointer(*slot);
5203 rcu_assign_pointer(*slot, pwq);
5207 /* context to store the prepared attrs & pwqs before applying */
5208 struct apply_wqattrs_ctx {
5209 struct workqueue_struct *wq; /* target workqueue */
5210 struct workqueue_attrs *attrs; /* attrs to apply */
5211 struct list_head list; /* queued for batching commit */
5212 struct pool_workqueue *dfl_pwq;
5213 struct pool_workqueue *pwq_tbl[];
5216 /* free the resources after success or abort */
5217 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5222 for_each_possible_cpu(cpu)
5223 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5224 put_pwq_unlocked(ctx->dfl_pwq);
5226 free_workqueue_attrs(ctx->attrs);
5232 /* allocate the attrs and pwqs for later installation */
5233 static struct apply_wqattrs_ctx *
5234 apply_wqattrs_prepare(struct workqueue_struct *wq,
5235 const struct workqueue_attrs *attrs,
5236 const cpumask_var_t unbound_cpumask)
5238 struct apply_wqattrs_ctx *ctx;
5239 struct workqueue_attrs *new_attrs;
5242 lockdep_assert_held(&wq_pool_mutex);
5244 if (WARN_ON(attrs->affn_scope < 0 ||
5245 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5246 return ERR_PTR(-EINVAL);
5248 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5250 new_attrs = alloc_workqueue_attrs();
5251 if (!ctx || !new_attrs)
5255 * If something goes wrong during CPU up/down, we'll fall back to
5256 * the default pwq covering whole @attrs->cpumask. Always create
5257 * it even if we don't use it immediately.
5259 copy_workqueue_attrs(new_attrs, attrs);
5260 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5261 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5262 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5266 for_each_possible_cpu(cpu) {
5267 if (new_attrs->ordered) {
5268 ctx->dfl_pwq->refcnt++;
5269 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5271 wq_calc_pod_cpumask(new_attrs, cpu, -1);
5272 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5273 if (!ctx->pwq_tbl[cpu])
5278 /* save the user configured attrs and sanitize it. */
5279 copy_workqueue_attrs(new_attrs, attrs);
5280 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5281 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5282 ctx->attrs = new_attrs;
5285 * For initialized ordered workqueues, there should only be one pwq
5286 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5287 * of newly queued work items until execution of older work items in
5288 * the old pwq's have completed.
5290 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5291 ctx->dfl_pwq->plugged = true;
5297 free_workqueue_attrs(new_attrs);
5298 apply_wqattrs_cleanup(ctx);
5299 return ERR_PTR(-ENOMEM);
5302 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5303 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5307 /* all pwqs have been created successfully, let's install'em */
5308 mutex_lock(&ctx->wq->mutex);
5310 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5312 /* save the previous pwqs and install the new ones */
5313 for_each_possible_cpu(cpu)
5314 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5316 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5318 /* update node_nr_active->max */
5319 wq_update_node_max_active(ctx->wq, -1);
5321 /* rescuer needs to respect wq cpumask changes */
5322 if (ctx->wq->rescuer)
5323 set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5324 unbound_effective_cpumask(ctx->wq));
5326 mutex_unlock(&ctx->wq->mutex);
5329 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5330 const struct workqueue_attrs *attrs)
5332 struct apply_wqattrs_ctx *ctx;
5334 /* only unbound workqueues can change attributes */
5335 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5338 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5340 return PTR_ERR(ctx);
5342 /* the ctx has been prepared successfully, let's commit it */
5343 apply_wqattrs_commit(ctx);
5344 apply_wqattrs_cleanup(ctx);
5350 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5351 * @wq: the target workqueue
5352 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5354 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5355 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5356 * work items are affine to the pod it was issued on. Older pwqs are released as
5357 * in-flight work items finish. Note that a work item which repeatedly requeues
5358 * itself back-to-back will stay on its current pwq.
5360 * Performs GFP_KERNEL allocations.
5362 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
5364 * Return: 0 on success and -errno on failure.
5366 int apply_workqueue_attrs(struct workqueue_struct *wq,
5367 const struct workqueue_attrs *attrs)
5371 lockdep_assert_cpus_held();
5373 mutex_lock(&wq_pool_mutex);
5374 ret = apply_workqueue_attrs_locked(wq, attrs);
5375 mutex_unlock(&wq_pool_mutex);
5381 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
5382 * @wq: the target workqueue
5383 * @cpu: the CPU to update pool association for
5384 * @hotplug_cpu: the CPU coming up or going down
5385 * @online: whether @cpu is coming up or going down
5387 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5388 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
5392 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5393 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5395 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5396 * with a cpumask spanning multiple pods, the workers which were already
5397 * executing the work items for the workqueue will lose their CPU affinity and
5398 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5399 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5400 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5402 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
5403 int hotplug_cpu, bool online)
5405 int off_cpu = online ? -1 : hotplug_cpu;
5406 struct pool_workqueue *old_pwq = NULL, *pwq;
5407 struct workqueue_attrs *target_attrs;
5409 lockdep_assert_held(&wq_pool_mutex);
5411 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5415 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5416 * Let's use a preallocated one. The following buf is protected by
5417 * CPU hotplug exclusion.
5419 target_attrs = wq_update_pod_attrs_buf;
5421 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5422 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5424 /* nothing to do if the target cpumask matches the current pwq */
5425 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
5426 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5429 /* create a new pwq */
5430 pwq = alloc_unbound_pwq(wq, target_attrs);
5432 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5437 /* Install the new pwq. */
5438 mutex_lock(&wq->mutex);
5439 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5443 mutex_lock(&wq->mutex);
5444 pwq = unbound_pwq(wq, -1);
5445 raw_spin_lock_irq(&pwq->pool->lock);
5447 raw_spin_unlock_irq(&pwq->pool->lock);
5448 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5450 mutex_unlock(&wq->mutex);
5451 put_pwq_unlocked(old_pwq);
5454 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5456 bool highpri = wq->flags & WQ_HIGHPRI;
5459 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5463 if (!(wq->flags & WQ_UNBOUND)) {
5464 for_each_possible_cpu(cpu) {
5465 struct pool_workqueue **pwq_p;
5466 struct worker_pool __percpu *pools;
5467 struct worker_pool *pool;
5469 if (wq->flags & WQ_BH)
5470 pools = bh_worker_pools;
5472 pools = cpu_worker_pools;
5474 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5475 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5477 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5482 init_pwq(*pwq_p, wq, pool);
5484 mutex_lock(&wq->mutex);
5486 mutex_unlock(&wq->mutex);
5492 if (wq->flags & __WQ_ORDERED) {
5493 struct pool_workqueue *dfl_pwq;
5495 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5496 /* there should only be single pwq for ordering guarantee */
5497 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5498 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5499 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5500 "ordering guarantee broken for workqueue %s\n", wq->name);
5502 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5506 /* for unbound pwq, flush the pwq_release_worker ensures that the
5507 * pwq_release_workfn() completes before calling kfree(wq).
5510 kthread_flush_worker(pwq_release_worker);
5516 for_each_possible_cpu(cpu) {
5517 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5520 kmem_cache_free(pwq_cache, pwq);
5522 free_percpu(wq->cpu_pwq);
5528 static int wq_clamp_max_active(int max_active, unsigned int flags,
5531 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5532 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5533 max_active, name, 1, WQ_MAX_ACTIVE);
5535 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5539 * Workqueues which may be used during memory reclaim should have a rescuer
5540 * to guarantee forward progress.
5542 static int init_rescuer(struct workqueue_struct *wq)
5544 struct worker *rescuer;
5547 if (!(wq->flags & WQ_MEM_RECLAIM))
5550 rescuer = alloc_worker(NUMA_NO_NODE);
5552 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5557 rescuer->rescue_wq = wq;
5558 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5559 if (IS_ERR(rescuer->task)) {
5560 ret = PTR_ERR(rescuer->task);
5561 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5562 wq->name, ERR_PTR(ret));
5567 wq->rescuer = rescuer;
5568 if (wq->flags & WQ_UNBOUND)
5569 kthread_bind_mask(rescuer->task, wq_unbound_cpumask);
5571 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5572 wake_up_process(rescuer->task);
5578 * wq_adjust_max_active - update a wq's max_active to the current setting
5579 * @wq: target workqueue
5581 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5582 * activate inactive work items accordingly. If @wq is freezing, clear
5583 * @wq->max_active to zero.
5585 static void wq_adjust_max_active(struct workqueue_struct *wq)
5588 int new_max, new_min;
5590 lockdep_assert_held(&wq->mutex);
5592 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5596 new_max = wq->saved_max_active;
5597 new_min = wq->saved_min_active;
5600 if (wq->max_active == new_max && wq->min_active == new_min)
5604 * Update @wq->max/min_active and then kick inactive work items if more
5605 * active work items are allowed. This doesn't break work item ordering
5606 * because new work items are always queued behind existing inactive
5607 * work items if there are any.
5609 WRITE_ONCE(wq->max_active, new_max);
5610 WRITE_ONCE(wq->min_active, new_min);
5612 if (wq->flags & WQ_UNBOUND)
5613 wq_update_node_max_active(wq, -1);
5619 * Round-robin through pwq's activating the first inactive work item
5620 * until max_active is filled.
5623 struct pool_workqueue *pwq;
5626 for_each_pwq(pwq, wq) {
5627 unsigned long irq_flags;
5629 /* can be called during early boot w/ irq disabled */
5630 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5631 if (pwq_activate_first_inactive(pwq, true)) {
5633 kick_pool(pwq->pool);
5635 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5637 } while (activated);
5641 struct workqueue_struct *alloc_workqueue(const char *fmt,
5643 int max_active, ...)
5646 struct workqueue_struct *wq;
5650 if (flags & WQ_BH) {
5651 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5653 if (WARN_ON_ONCE(max_active))
5657 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5658 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5659 flags |= WQ_UNBOUND;
5661 /* allocate wq and format name */
5662 if (flags & WQ_UNBOUND)
5663 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5665 wq_size = sizeof(*wq);
5667 wq = kzalloc(wq_size, GFP_KERNEL);
5671 if (flags & WQ_UNBOUND) {
5672 wq->unbound_attrs = alloc_workqueue_attrs();
5673 if (!wq->unbound_attrs)
5677 va_start(args, max_active);
5678 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5681 if (name_len >= WQ_NAME_LEN)
5682 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5685 if (flags & WQ_BH) {
5687 * BH workqueues always share a single execution context per CPU
5688 * and don't impose any max_active limit.
5690 max_active = INT_MAX;
5692 max_active = max_active ?: WQ_DFL_ACTIVE;
5693 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5698 wq->max_active = max_active;
5699 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5700 wq->saved_max_active = wq->max_active;
5701 wq->saved_min_active = wq->min_active;
5702 mutex_init(&wq->mutex);
5703 atomic_set(&wq->nr_pwqs_to_flush, 0);
5704 INIT_LIST_HEAD(&wq->pwqs);
5705 INIT_LIST_HEAD(&wq->flusher_queue);
5706 INIT_LIST_HEAD(&wq->flusher_overflow);
5707 INIT_LIST_HEAD(&wq->maydays);
5709 wq_init_lockdep(wq);
5710 INIT_LIST_HEAD(&wq->list);
5712 if (flags & WQ_UNBOUND) {
5713 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5714 goto err_unreg_lockdep;
5717 if (alloc_and_link_pwqs(wq) < 0)
5718 goto err_free_node_nr_active;
5720 if (wq_online && init_rescuer(wq) < 0)
5723 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5727 * wq_pool_mutex protects global freeze state and workqueues list.
5728 * Grab it, adjust max_active and add the new @wq to workqueues
5731 mutex_lock(&wq_pool_mutex);
5733 mutex_lock(&wq->mutex);
5734 wq_adjust_max_active(wq);
5735 mutex_unlock(&wq->mutex);
5737 list_add_tail_rcu(&wq->list, &workqueues);
5739 mutex_unlock(&wq_pool_mutex);
5743 err_free_node_nr_active:
5744 if (wq->flags & WQ_UNBOUND)
5745 free_node_nr_active(wq->node_nr_active);
5747 wq_unregister_lockdep(wq);
5748 wq_free_lockdep(wq);
5750 free_workqueue_attrs(wq->unbound_attrs);
5754 destroy_workqueue(wq);
5757 EXPORT_SYMBOL_GPL(alloc_workqueue);
5759 static bool pwq_busy(struct pool_workqueue *pwq)
5763 for (i = 0; i < WORK_NR_COLORS; i++)
5764 if (pwq->nr_in_flight[i])
5767 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5769 if (!pwq_is_empty(pwq))
5776 * destroy_workqueue - safely terminate a workqueue
5777 * @wq: target workqueue
5779 * Safely destroy a workqueue. All work currently pending will be done first.
5781 void destroy_workqueue(struct workqueue_struct *wq)
5783 struct pool_workqueue *pwq;
5787 * Remove it from sysfs first so that sanity check failure doesn't
5788 * lead to sysfs name conflicts.
5790 workqueue_sysfs_unregister(wq);
5792 /* mark the workqueue destruction is in progress */
5793 mutex_lock(&wq->mutex);
5794 wq->flags |= __WQ_DESTROYING;
5795 mutex_unlock(&wq->mutex);
5797 /* drain it before proceeding with destruction */
5798 drain_workqueue(wq);
5800 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5802 struct worker *rescuer = wq->rescuer;
5804 /* this prevents new queueing */
5805 raw_spin_lock_irq(&wq_mayday_lock);
5807 raw_spin_unlock_irq(&wq_mayday_lock);
5809 /* rescuer will empty maydays list before exiting */
5810 kthread_stop(rescuer->task);
5815 * Sanity checks - grab all the locks so that we wait for all
5816 * in-flight operations which may do put_pwq().
5818 mutex_lock(&wq_pool_mutex);
5819 mutex_lock(&wq->mutex);
5820 for_each_pwq(pwq, wq) {
5821 raw_spin_lock_irq(&pwq->pool->lock);
5822 if (WARN_ON(pwq_busy(pwq))) {
5823 pr_warn("%s: %s has the following busy pwq\n",
5824 __func__, wq->name);
5826 raw_spin_unlock_irq(&pwq->pool->lock);
5827 mutex_unlock(&wq->mutex);
5828 mutex_unlock(&wq_pool_mutex);
5829 show_one_workqueue(wq);
5832 raw_spin_unlock_irq(&pwq->pool->lock);
5834 mutex_unlock(&wq->mutex);
5837 * wq list is used to freeze wq, remove from list after
5838 * flushing is complete in case freeze races us.
5840 list_del_rcu(&wq->list);
5841 mutex_unlock(&wq_pool_mutex);
5844 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5845 * to put the base refs. @wq will be auto-destroyed from the last
5846 * pwq_put. RCU read lock prevents @wq from going away from under us.
5850 for_each_possible_cpu(cpu) {
5851 put_pwq_unlocked(unbound_pwq(wq, cpu));
5852 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5855 put_pwq_unlocked(unbound_pwq(wq, -1));
5856 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5860 EXPORT_SYMBOL_GPL(destroy_workqueue);
5863 * workqueue_set_max_active - adjust max_active of a workqueue
5864 * @wq: target workqueue
5865 * @max_active: new max_active value.
5867 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5871 * Don't call from IRQ context.
5873 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5875 /* max_active doesn't mean anything for BH workqueues */
5876 if (WARN_ON(wq->flags & WQ_BH))
5878 /* disallow meddling with max_active for ordered workqueues */
5879 if (WARN_ON(wq->flags & __WQ_ORDERED))
5882 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5884 mutex_lock(&wq->mutex);
5886 wq->saved_max_active = max_active;
5887 if (wq->flags & WQ_UNBOUND)
5888 wq->saved_min_active = min(wq->saved_min_active, max_active);
5890 wq_adjust_max_active(wq);
5892 mutex_unlock(&wq->mutex);
5894 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5897 * workqueue_set_min_active - adjust min_active of an unbound workqueue
5898 * @wq: target unbound workqueue
5899 * @min_active: new min_active value
5901 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5902 * unbound workqueue is not guaranteed to be able to process max_active
5903 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
5904 * able to process min_active number of interdependent work items which is
5905 * %WQ_DFL_MIN_ACTIVE by default.
5907 * Use this function to adjust the min_active value between 0 and the current
5910 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5912 /* min_active is only meaningful for non-ordered unbound workqueues */
5913 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
5917 mutex_lock(&wq->mutex);
5918 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5919 wq_adjust_max_active(wq);
5920 mutex_unlock(&wq->mutex);
5924 * current_work - retrieve %current task's work struct
5926 * Determine if %current task is a workqueue worker and what it's working on.
5927 * Useful to find out the context that the %current task is running in.
5929 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5931 struct work_struct *current_work(void)
5933 struct worker *worker = current_wq_worker();
5935 return worker ? worker->current_work : NULL;
5937 EXPORT_SYMBOL(current_work);
5940 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5942 * Determine whether %current is a workqueue rescuer. Can be used from
5943 * work functions to determine whether it's being run off the rescuer task.
5945 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5947 bool current_is_workqueue_rescuer(void)
5949 struct worker *worker = current_wq_worker();
5951 return worker && worker->rescue_wq;
5955 * workqueue_congested - test whether a workqueue is congested
5956 * @cpu: CPU in question
5957 * @wq: target workqueue
5959 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5960 * no synchronization around this function and the test result is
5961 * unreliable and only useful as advisory hints or for debugging.
5963 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5965 * With the exception of ordered workqueues, all workqueues have per-cpu
5966 * pool_workqueues, each with its own congested state. A workqueue being
5967 * congested on one CPU doesn't mean that the workqueue is contested on any
5971 * %true if congested, %false otherwise.
5973 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5975 struct pool_workqueue *pwq;
5981 if (cpu == WORK_CPU_UNBOUND)
5982 cpu = smp_processor_id();
5984 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5985 ret = !list_empty(&pwq->inactive_works);
5992 EXPORT_SYMBOL_GPL(workqueue_congested);
5995 * work_busy - test whether a work is currently pending or running
5996 * @work: the work to be tested
5998 * Test whether @work is currently pending or running. There is no
5999 * synchronization around this function and the test result is
6000 * unreliable and only useful as advisory hints or for debugging.
6003 * OR'd bitmask of WORK_BUSY_* bits.
6005 unsigned int work_busy(struct work_struct *work)
6007 struct worker_pool *pool;
6008 unsigned long irq_flags;
6009 unsigned int ret = 0;
6011 if (work_pending(work))
6012 ret |= WORK_BUSY_PENDING;
6015 pool = get_work_pool(work);
6017 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6018 if (find_worker_executing_work(pool, work))
6019 ret |= WORK_BUSY_RUNNING;
6020 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6026 EXPORT_SYMBOL_GPL(work_busy);
6029 * set_worker_desc - set description for the current work item
6030 * @fmt: printf-style format string
6031 * @...: arguments for the format string
6033 * This function can be called by a running work function to describe what
6034 * the work item is about. If the worker task gets dumped, this
6035 * information will be printed out together to help debugging. The
6036 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6038 void set_worker_desc(const char *fmt, ...)
6040 struct worker *worker = current_wq_worker();
6044 va_start(args, fmt);
6045 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6049 EXPORT_SYMBOL_GPL(set_worker_desc);
6052 * print_worker_info - print out worker information and description
6053 * @log_lvl: the log level to use when printing
6054 * @task: target task
6056 * If @task is a worker and currently executing a work item, print out the
6057 * name of the workqueue being serviced and worker description set with
6058 * set_worker_desc() by the currently executing work item.
6060 * This function can be safely called on any task as long as the
6061 * task_struct itself is accessible. While safe, this function isn't
6062 * synchronized and may print out mixups or garbages of limited length.
6064 void print_worker_info(const char *log_lvl, struct task_struct *task)
6066 work_func_t *fn = NULL;
6067 char name[WQ_NAME_LEN] = { };
6068 char desc[WORKER_DESC_LEN] = { };
6069 struct pool_workqueue *pwq = NULL;
6070 struct workqueue_struct *wq = NULL;
6071 struct worker *worker;
6073 if (!(task->flags & PF_WQ_WORKER))
6077 * This function is called without any synchronization and @task
6078 * could be in any state. Be careful with dereferences.
6080 worker = kthread_probe_data(task);
6083 * Carefully copy the associated workqueue's workfn, name and desc.
6084 * Keep the original last '\0' in case the original is garbage.
6086 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6087 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6088 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6089 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6090 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6092 if (fn || name[0] || desc[0]) {
6093 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6094 if (strcmp(name, desc))
6095 pr_cont(" (%s)", desc);
6100 static void pr_cont_pool_info(struct worker_pool *pool)
6102 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6103 if (pool->node != NUMA_NO_NODE)
6104 pr_cont(" node=%d", pool->node);
6105 pr_cont(" flags=0x%x", pool->flags);
6106 if (pool->flags & POOL_BH)
6108 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6110 pr_cont(" nice=%d", pool->attrs->nice);
6113 static void pr_cont_worker_id(struct worker *worker)
6115 struct worker_pool *pool = worker->pool;
6117 if (pool->flags & WQ_BH)
6119 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6121 pr_cont("%d%s", task_pid_nr(worker->task),
6122 worker->rescue_wq ? "(RESCUER)" : "");
6125 struct pr_cont_work_struct {
6131 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6135 if (func == pcwsp->func) {
6139 if (pcwsp->ctr == 1)
6140 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6142 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6145 if ((long)func == -1L)
6147 pcwsp->comma = comma;
6152 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6154 if (work->func == wq_barrier_func) {
6155 struct wq_barrier *barr;
6157 barr = container_of(work, struct wq_barrier, work);
6159 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6160 pr_cont("%s BAR(%d)", comma ? "," : "",
6161 task_pid_nr(barr->task));
6164 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6165 pr_cont_work_flush(comma, work->func, pcwsp);
6169 static void show_pwq(struct pool_workqueue *pwq)
6171 struct pr_cont_work_struct pcws = { .ctr = 0, };
6172 struct worker_pool *pool = pwq->pool;
6173 struct work_struct *work;
6174 struct worker *worker;
6175 bool has_in_flight = false, has_pending = false;
6178 pr_info(" pwq %d:", pool->id);
6179 pr_cont_pool_info(pool);
6181 pr_cont(" active=%d refcnt=%d%s\n",
6182 pwq->nr_active, pwq->refcnt,
6183 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6185 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6186 if (worker->current_pwq == pwq) {
6187 has_in_flight = true;
6191 if (has_in_flight) {
6194 pr_info(" in-flight:");
6195 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6196 if (worker->current_pwq != pwq)
6199 pr_cont(" %s", comma ? "," : "");
6200 pr_cont_worker_id(worker);
6201 pr_cont(":%ps", worker->current_func);
6202 list_for_each_entry(work, &worker->scheduled, entry)
6203 pr_cont_work(false, work, &pcws);
6204 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6210 list_for_each_entry(work, &pool->worklist, entry) {
6211 if (get_work_pwq(work) == pwq) {
6219 pr_info(" pending:");
6220 list_for_each_entry(work, &pool->worklist, entry) {
6221 if (get_work_pwq(work) != pwq)
6224 pr_cont_work(comma, work, &pcws);
6225 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6227 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6231 if (!list_empty(&pwq->inactive_works)) {
6234 pr_info(" inactive:");
6235 list_for_each_entry(work, &pwq->inactive_works, entry) {
6236 pr_cont_work(comma, work, &pcws);
6237 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6239 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6245 * show_one_workqueue - dump state of specified workqueue
6246 * @wq: workqueue whose state will be printed
6248 void show_one_workqueue(struct workqueue_struct *wq)
6250 struct pool_workqueue *pwq;
6252 unsigned long irq_flags;
6254 for_each_pwq(pwq, wq) {
6255 if (!pwq_is_empty(pwq)) {
6260 if (idle) /* Nothing to print for idle workqueue */
6263 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6265 for_each_pwq(pwq, wq) {
6266 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6267 if (!pwq_is_empty(pwq)) {
6269 * Defer printing to avoid deadlocks in console
6270 * drivers that queue work while holding locks
6271 * also taken in their write paths.
6273 printk_deferred_enter();
6275 printk_deferred_exit();
6277 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6279 * We could be printing a lot from atomic context, e.g.
6280 * sysrq-t -> show_all_workqueues(). Avoid triggering
6283 touch_nmi_watchdog();
6289 * show_one_worker_pool - dump state of specified worker pool
6290 * @pool: worker pool whose state will be printed
6292 static void show_one_worker_pool(struct worker_pool *pool)
6294 struct worker *worker;
6296 unsigned long irq_flags;
6297 unsigned long hung = 0;
6299 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6300 if (pool->nr_workers == pool->nr_idle)
6303 /* How long the first pending work is waiting for a worker. */
6304 if (!list_empty(&pool->worklist))
6305 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6308 * Defer printing to avoid deadlocks in console drivers that
6309 * queue work while holding locks also taken in their write
6312 printk_deferred_enter();
6313 pr_info("pool %d:", pool->id);
6314 pr_cont_pool_info(pool);
6315 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6317 pr_cont(" manager: %d",
6318 task_pid_nr(pool->manager->task));
6319 list_for_each_entry(worker, &pool->idle_list, entry) {
6320 pr_cont(" %s", first ? "idle: " : "");
6321 pr_cont_worker_id(worker);
6325 printk_deferred_exit();
6327 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6329 * We could be printing a lot from atomic context, e.g.
6330 * sysrq-t -> show_all_workqueues(). Avoid triggering
6333 touch_nmi_watchdog();
6338 * show_all_workqueues - dump workqueue state
6340 * Called from a sysrq handler and prints out all busy workqueues and pools.
6342 void show_all_workqueues(void)
6344 struct workqueue_struct *wq;
6345 struct worker_pool *pool;
6350 pr_info("Showing busy workqueues and worker pools:\n");
6352 list_for_each_entry_rcu(wq, &workqueues, list)
6353 show_one_workqueue(wq);
6355 for_each_pool(pool, pi)
6356 show_one_worker_pool(pool);
6362 * show_freezable_workqueues - dump freezable workqueue state
6364 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6367 void show_freezable_workqueues(void)
6369 struct workqueue_struct *wq;
6373 pr_info("Showing freezable workqueues that are still busy:\n");
6375 list_for_each_entry_rcu(wq, &workqueues, list) {
6376 if (!(wq->flags & WQ_FREEZABLE))
6378 show_one_workqueue(wq);
6384 /* used to show worker information through /proc/PID/{comm,stat,status} */
6385 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6389 /* always show the actual comm */
6390 off = strscpy(buf, task->comm, size);
6394 /* stabilize PF_WQ_WORKER and worker pool association */
6395 mutex_lock(&wq_pool_attach_mutex);
6397 if (task->flags & PF_WQ_WORKER) {
6398 struct worker *worker = kthread_data(task);
6399 struct worker_pool *pool = worker->pool;
6402 raw_spin_lock_irq(&pool->lock);
6404 * ->desc tracks information (wq name or
6405 * set_worker_desc()) for the latest execution. If
6406 * current, prepend '+', otherwise '-'.
6408 if (worker->desc[0] != '\0') {
6409 if (worker->current_work)
6410 scnprintf(buf + off, size - off, "+%s",
6413 scnprintf(buf + off, size - off, "-%s",
6416 raw_spin_unlock_irq(&pool->lock);
6420 mutex_unlock(&wq_pool_attach_mutex);
6428 * There are two challenges in supporting CPU hotplug. Firstly, there
6429 * are a lot of assumptions on strong associations among work, pwq and
6430 * pool which make migrating pending and scheduled works very
6431 * difficult to implement without impacting hot paths. Secondly,
6432 * worker pools serve mix of short, long and very long running works making
6433 * blocked draining impractical.
6435 * This is solved by allowing the pools to be disassociated from the CPU
6436 * running as an unbound one and allowing it to be reattached later if the
6437 * cpu comes back online.
6440 static void unbind_workers(int cpu)
6442 struct worker_pool *pool;
6443 struct worker *worker;
6445 for_each_cpu_worker_pool(pool, cpu) {
6446 mutex_lock(&wq_pool_attach_mutex);
6447 raw_spin_lock_irq(&pool->lock);
6450 * We've blocked all attach/detach operations. Make all workers
6451 * unbound and set DISASSOCIATED. Before this, all workers
6452 * must be on the cpu. After this, they may become diasporas.
6453 * And the preemption disabled section in their sched callbacks
6454 * are guaranteed to see WORKER_UNBOUND since the code here
6455 * is on the same cpu.
6457 for_each_pool_worker(worker, pool)
6458 worker->flags |= WORKER_UNBOUND;
6460 pool->flags |= POOL_DISASSOCIATED;
6463 * The handling of nr_running in sched callbacks are disabled
6464 * now. Zap nr_running. After this, nr_running stays zero and
6465 * need_more_worker() and keep_working() are always true as
6466 * long as the worklist is not empty. This pool now behaves as
6467 * an unbound (in terms of concurrency management) pool which
6468 * are served by workers tied to the pool.
6470 pool->nr_running = 0;
6473 * With concurrency management just turned off, a busy
6474 * worker blocking could lead to lengthy stalls. Kick off
6475 * unbound chain execution of currently pending work items.
6479 raw_spin_unlock_irq(&pool->lock);
6481 for_each_pool_worker(worker, pool)
6482 unbind_worker(worker);
6484 mutex_unlock(&wq_pool_attach_mutex);
6489 * rebind_workers - rebind all workers of a pool to the associated CPU
6490 * @pool: pool of interest
6492 * @pool->cpu is coming online. Rebind all workers to the CPU.
6494 static void rebind_workers(struct worker_pool *pool)
6496 struct worker *worker;
6498 lockdep_assert_held(&wq_pool_attach_mutex);
6501 * Restore CPU affinity of all workers. As all idle workers should
6502 * be on the run-queue of the associated CPU before any local
6503 * wake-ups for concurrency management happen, restore CPU affinity
6504 * of all workers first and then clear UNBOUND. As we're called
6505 * from CPU_ONLINE, the following shouldn't fail.
6507 for_each_pool_worker(worker, pool) {
6508 kthread_set_per_cpu(worker->task, pool->cpu);
6509 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6510 pool_allowed_cpus(pool)) < 0);
6513 raw_spin_lock_irq(&pool->lock);
6515 pool->flags &= ~POOL_DISASSOCIATED;
6517 for_each_pool_worker(worker, pool) {
6518 unsigned int worker_flags = worker->flags;
6521 * We want to clear UNBOUND but can't directly call
6522 * worker_clr_flags() or adjust nr_running. Atomically
6523 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6524 * @worker will clear REBOUND using worker_clr_flags() when
6525 * it initiates the next execution cycle thus restoring
6526 * concurrency management. Note that when or whether
6527 * @worker clears REBOUND doesn't affect correctness.
6529 * WRITE_ONCE() is necessary because @worker->flags may be
6530 * tested without holding any lock in
6531 * wq_worker_running(). Without it, NOT_RUNNING test may
6532 * fail incorrectly leading to premature concurrency
6533 * management operations.
6535 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6536 worker_flags |= WORKER_REBOUND;
6537 worker_flags &= ~WORKER_UNBOUND;
6538 WRITE_ONCE(worker->flags, worker_flags);
6541 raw_spin_unlock_irq(&pool->lock);
6545 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6546 * @pool: unbound pool of interest
6547 * @cpu: the CPU which is coming up
6549 * An unbound pool may end up with a cpumask which doesn't have any online
6550 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6551 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6552 * online CPU before, cpus_allowed of all its workers should be restored.
6554 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6556 static cpumask_t cpumask;
6557 struct worker *worker;
6559 lockdep_assert_held(&wq_pool_attach_mutex);
6561 /* is @cpu allowed for @pool? */
6562 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6565 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6567 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6568 for_each_pool_worker(worker, pool)
6569 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6572 int workqueue_prepare_cpu(unsigned int cpu)
6574 struct worker_pool *pool;
6576 for_each_cpu_worker_pool(pool, cpu) {
6577 if (pool->nr_workers)
6579 if (!create_worker(pool))
6585 int workqueue_online_cpu(unsigned int cpu)
6587 struct worker_pool *pool;
6588 struct workqueue_struct *wq;
6591 mutex_lock(&wq_pool_mutex);
6593 for_each_pool(pool, pi) {
6594 /* BH pools aren't affected by hotplug */
6595 if (pool->flags & POOL_BH)
6598 mutex_lock(&wq_pool_attach_mutex);
6599 if (pool->cpu == cpu)
6600 rebind_workers(pool);
6601 else if (pool->cpu < 0)
6602 restore_unbound_workers_cpumask(pool, cpu);
6603 mutex_unlock(&wq_pool_attach_mutex);
6606 /* update pod affinity of unbound workqueues */
6607 list_for_each_entry(wq, &workqueues, list) {
6608 struct workqueue_attrs *attrs = wq->unbound_attrs;
6611 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6614 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6615 wq_update_pod(wq, tcpu, cpu, true);
6617 mutex_lock(&wq->mutex);
6618 wq_update_node_max_active(wq, -1);
6619 mutex_unlock(&wq->mutex);
6623 mutex_unlock(&wq_pool_mutex);
6627 int workqueue_offline_cpu(unsigned int cpu)
6629 struct workqueue_struct *wq;
6631 /* unbinding per-cpu workers should happen on the local CPU */
6632 if (WARN_ON(cpu != smp_processor_id()))
6635 unbind_workers(cpu);
6637 /* update pod affinity of unbound workqueues */
6638 mutex_lock(&wq_pool_mutex);
6639 list_for_each_entry(wq, &workqueues, list) {
6640 struct workqueue_attrs *attrs = wq->unbound_attrs;
6643 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6646 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6647 wq_update_pod(wq, tcpu, cpu, false);
6649 mutex_lock(&wq->mutex);
6650 wq_update_node_max_active(wq, cpu);
6651 mutex_unlock(&wq->mutex);
6654 mutex_unlock(&wq_pool_mutex);
6659 struct work_for_cpu {
6660 struct work_struct work;
6666 static void work_for_cpu_fn(struct work_struct *work)
6668 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6670 wfc->ret = wfc->fn(wfc->arg);
6674 * work_on_cpu_key - run a function in thread context on a particular cpu
6675 * @cpu: the cpu to run on
6676 * @fn: the function to run
6677 * @arg: the function arg
6678 * @key: The lock class key for lock debugging purposes
6680 * It is up to the caller to ensure that the cpu doesn't go offline.
6681 * The caller must not hold any locks which would prevent @fn from completing.
6683 * Return: The value @fn returns.
6685 long work_on_cpu_key(int cpu, long (*fn)(void *),
6686 void *arg, struct lock_class_key *key)
6688 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6690 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6691 schedule_work_on(cpu, &wfc.work);
6692 flush_work(&wfc.work);
6693 destroy_work_on_stack(&wfc.work);
6696 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6699 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6700 * @cpu: the cpu to run on
6701 * @fn: the function to run
6702 * @arg: the function argument
6703 * @key: The lock class key for lock debugging purposes
6705 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6706 * any locks which would prevent @fn from completing.
6708 * Return: The value @fn returns.
6710 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6711 void *arg, struct lock_class_key *key)
6716 if (cpu_online(cpu))
6717 ret = work_on_cpu_key(cpu, fn, arg, key);
6721 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6722 #endif /* CONFIG_SMP */
6724 #ifdef CONFIG_FREEZER
6727 * freeze_workqueues_begin - begin freezing workqueues
6729 * Start freezing workqueues. After this function returns, all freezable
6730 * workqueues will queue new works to their inactive_works list instead of
6734 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6736 void freeze_workqueues_begin(void)
6738 struct workqueue_struct *wq;
6740 mutex_lock(&wq_pool_mutex);
6742 WARN_ON_ONCE(workqueue_freezing);
6743 workqueue_freezing = true;
6745 list_for_each_entry(wq, &workqueues, list) {
6746 mutex_lock(&wq->mutex);
6747 wq_adjust_max_active(wq);
6748 mutex_unlock(&wq->mutex);
6751 mutex_unlock(&wq_pool_mutex);
6755 * freeze_workqueues_busy - are freezable workqueues still busy?
6757 * Check whether freezing is complete. This function must be called
6758 * between freeze_workqueues_begin() and thaw_workqueues().
6761 * Grabs and releases wq_pool_mutex.
6764 * %true if some freezable workqueues are still busy. %false if freezing
6767 bool freeze_workqueues_busy(void)
6770 struct workqueue_struct *wq;
6771 struct pool_workqueue *pwq;
6773 mutex_lock(&wq_pool_mutex);
6775 WARN_ON_ONCE(!workqueue_freezing);
6777 list_for_each_entry(wq, &workqueues, list) {
6778 if (!(wq->flags & WQ_FREEZABLE))
6781 * nr_active is monotonically decreasing. It's safe
6782 * to peek without lock.
6785 for_each_pwq(pwq, wq) {
6786 WARN_ON_ONCE(pwq->nr_active < 0);
6787 if (pwq->nr_active) {
6796 mutex_unlock(&wq_pool_mutex);
6801 * thaw_workqueues - thaw workqueues
6803 * Thaw workqueues. Normal queueing is restored and all collected
6804 * frozen works are transferred to their respective pool worklists.
6807 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6809 void thaw_workqueues(void)
6811 struct workqueue_struct *wq;
6813 mutex_lock(&wq_pool_mutex);
6815 if (!workqueue_freezing)
6818 workqueue_freezing = false;
6820 /* restore max_active and repopulate worklist */
6821 list_for_each_entry(wq, &workqueues, list) {
6822 mutex_lock(&wq->mutex);
6823 wq_adjust_max_active(wq);
6824 mutex_unlock(&wq->mutex);
6828 mutex_unlock(&wq_pool_mutex);
6830 #endif /* CONFIG_FREEZER */
6832 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6836 struct workqueue_struct *wq;
6837 struct apply_wqattrs_ctx *ctx, *n;
6839 lockdep_assert_held(&wq_pool_mutex);
6841 list_for_each_entry(wq, &workqueues, list) {
6842 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6845 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6851 list_add_tail(&ctx->list, &ctxs);
6854 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6856 apply_wqattrs_commit(ctx);
6857 apply_wqattrs_cleanup(ctx);
6861 mutex_lock(&wq_pool_attach_mutex);
6862 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6863 mutex_unlock(&wq_pool_attach_mutex);
6869 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6870 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6872 * This function can be called from cpuset code to provide a set of isolated
6873 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
6874 * either cpus_read_lock or cpus_write_lock.
6876 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6878 cpumask_var_t cpumask;
6881 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6884 lockdep_assert_cpus_held();
6885 mutex_lock(&wq_pool_mutex);
6887 /* Save the current isolated cpumask & export it via sysfs */
6888 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6891 * If the operation fails, it will fall back to
6892 * wq_requested_unbound_cpumask which is initially set to
6893 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6894 * by any subsequent write to workqueue/cpumask sysfs file.
6896 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6897 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6898 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6899 ret = workqueue_apply_unbound_cpumask(cpumask);
6901 mutex_unlock(&wq_pool_mutex);
6902 free_cpumask_var(cpumask);
6906 static int parse_affn_scope(const char *val)
6910 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6911 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6917 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6919 struct workqueue_struct *wq;
6922 affn = parse_affn_scope(val);
6925 if (affn == WQ_AFFN_DFL)
6929 mutex_lock(&wq_pool_mutex);
6933 list_for_each_entry(wq, &workqueues, list) {
6934 for_each_online_cpu(cpu) {
6935 wq_update_pod(wq, cpu, cpu, true);
6939 mutex_unlock(&wq_pool_mutex);
6945 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6947 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6950 static const struct kernel_param_ops wq_affn_dfl_ops = {
6951 .set = wq_affn_dfl_set,
6952 .get = wq_affn_dfl_get,
6955 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6959 * Workqueues with WQ_SYSFS flag set is visible to userland via
6960 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6961 * following attributes.
6963 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6964 * max_active RW int : maximum number of in-flight work items
6966 * Unbound workqueues have the following extra attributes.
6968 * nice RW int : nice value of the workers
6969 * cpumask RW mask : bitmask of allowed CPUs for the workers
6970 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6971 * affinity_strict RW bool : worker CPU affinity is strict
6974 struct workqueue_struct *wq;
6978 static struct workqueue_struct *dev_to_wq(struct device *dev)
6980 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6985 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6988 struct workqueue_struct *wq = dev_to_wq(dev);
6990 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6992 static DEVICE_ATTR_RO(per_cpu);
6994 static ssize_t max_active_show(struct device *dev,
6995 struct device_attribute *attr, char *buf)
6997 struct workqueue_struct *wq = dev_to_wq(dev);
6999 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
7002 static ssize_t max_active_store(struct device *dev,
7003 struct device_attribute *attr, const char *buf,
7006 struct workqueue_struct *wq = dev_to_wq(dev);
7009 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7012 workqueue_set_max_active(wq, val);
7015 static DEVICE_ATTR_RW(max_active);
7017 static struct attribute *wq_sysfs_attrs[] = {
7018 &dev_attr_per_cpu.attr,
7019 &dev_attr_max_active.attr,
7022 ATTRIBUTE_GROUPS(wq_sysfs);
7024 static void apply_wqattrs_lock(void)
7026 /* CPUs should stay stable across pwq creations and installations */
7028 mutex_lock(&wq_pool_mutex);
7031 static void apply_wqattrs_unlock(void)
7033 mutex_unlock(&wq_pool_mutex);
7037 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7040 struct workqueue_struct *wq = dev_to_wq(dev);
7043 mutex_lock(&wq->mutex);
7044 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7045 mutex_unlock(&wq->mutex);
7050 /* prepare workqueue_attrs for sysfs store operations */
7051 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7053 struct workqueue_attrs *attrs;
7055 lockdep_assert_held(&wq_pool_mutex);
7057 attrs = alloc_workqueue_attrs();
7061 copy_workqueue_attrs(attrs, wq->unbound_attrs);
7065 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7066 const char *buf, size_t count)
7068 struct workqueue_struct *wq = dev_to_wq(dev);
7069 struct workqueue_attrs *attrs;
7072 apply_wqattrs_lock();
7074 attrs = wq_sysfs_prep_attrs(wq);
7078 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7079 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7080 ret = apply_workqueue_attrs_locked(wq, attrs);
7085 apply_wqattrs_unlock();
7086 free_workqueue_attrs(attrs);
7087 return ret ?: count;
7090 static ssize_t wq_cpumask_show(struct device *dev,
7091 struct device_attribute *attr, char *buf)
7093 struct workqueue_struct *wq = dev_to_wq(dev);
7096 mutex_lock(&wq->mutex);
7097 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7098 cpumask_pr_args(wq->unbound_attrs->cpumask));
7099 mutex_unlock(&wq->mutex);
7103 static ssize_t wq_cpumask_store(struct device *dev,
7104 struct device_attribute *attr,
7105 const char *buf, size_t count)
7107 struct workqueue_struct *wq = dev_to_wq(dev);
7108 struct workqueue_attrs *attrs;
7111 apply_wqattrs_lock();
7113 attrs = wq_sysfs_prep_attrs(wq);
7117 ret = cpumask_parse(buf, attrs->cpumask);
7119 ret = apply_workqueue_attrs_locked(wq, attrs);
7122 apply_wqattrs_unlock();
7123 free_workqueue_attrs(attrs);
7124 return ret ?: count;
7127 static ssize_t wq_affn_scope_show(struct device *dev,
7128 struct device_attribute *attr, char *buf)
7130 struct workqueue_struct *wq = dev_to_wq(dev);
7133 mutex_lock(&wq->mutex);
7134 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7135 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7136 wq_affn_names[WQ_AFFN_DFL],
7137 wq_affn_names[wq_affn_dfl]);
7139 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7140 wq_affn_names[wq->unbound_attrs->affn_scope]);
7141 mutex_unlock(&wq->mutex);
7146 static ssize_t wq_affn_scope_store(struct device *dev,
7147 struct device_attribute *attr,
7148 const char *buf, size_t count)
7150 struct workqueue_struct *wq = dev_to_wq(dev);
7151 struct workqueue_attrs *attrs;
7152 int affn, ret = -ENOMEM;
7154 affn = parse_affn_scope(buf);
7158 apply_wqattrs_lock();
7159 attrs = wq_sysfs_prep_attrs(wq);
7161 attrs->affn_scope = affn;
7162 ret = apply_workqueue_attrs_locked(wq, attrs);
7164 apply_wqattrs_unlock();
7165 free_workqueue_attrs(attrs);
7166 return ret ?: count;
7169 static ssize_t wq_affinity_strict_show(struct device *dev,
7170 struct device_attribute *attr, char *buf)
7172 struct workqueue_struct *wq = dev_to_wq(dev);
7174 return scnprintf(buf, PAGE_SIZE, "%d\n",
7175 wq->unbound_attrs->affn_strict);
7178 static ssize_t wq_affinity_strict_store(struct device *dev,
7179 struct device_attribute *attr,
7180 const char *buf, size_t count)
7182 struct workqueue_struct *wq = dev_to_wq(dev);
7183 struct workqueue_attrs *attrs;
7184 int v, ret = -ENOMEM;
7186 if (sscanf(buf, "%d", &v) != 1)
7189 apply_wqattrs_lock();
7190 attrs = wq_sysfs_prep_attrs(wq);
7192 attrs->affn_strict = (bool)v;
7193 ret = apply_workqueue_attrs_locked(wq, attrs);
7195 apply_wqattrs_unlock();
7196 free_workqueue_attrs(attrs);
7197 return ret ?: count;
7200 static struct device_attribute wq_sysfs_unbound_attrs[] = {
7201 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7202 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7203 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7204 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7208 static const struct bus_type wq_subsys = {
7209 .name = "workqueue",
7210 .dev_groups = wq_sysfs_groups,
7214 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7215 * @cpumask: the cpumask to set
7217 * The low-level workqueues cpumask is a global cpumask that limits
7218 * the affinity of all unbound workqueues. This function check the @cpumask
7219 * and apply it to all unbound workqueues and updates all pwqs of them.
7221 * Return: 0 - Success
7222 * -EINVAL - Invalid @cpumask
7223 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7225 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7230 * Not excluding isolated cpus on purpose.
7231 * If the user wishes to include them, we allow that.
7233 cpumask_and(cpumask, cpumask, cpu_possible_mask);
7234 if (!cpumask_empty(cpumask)) {
7235 apply_wqattrs_lock();
7236 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7237 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
7242 ret = workqueue_apply_unbound_cpumask(cpumask);
7245 apply_wqattrs_unlock();
7251 static ssize_t __wq_cpumask_show(struct device *dev,
7252 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7256 mutex_lock(&wq_pool_mutex);
7257 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7258 mutex_unlock(&wq_pool_mutex);
7263 static ssize_t cpumask_requested_show(struct device *dev,
7264 struct device_attribute *attr, char *buf)
7266 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7268 static DEVICE_ATTR_RO(cpumask_requested);
7270 static ssize_t cpumask_isolated_show(struct device *dev,
7271 struct device_attribute *attr, char *buf)
7273 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7275 static DEVICE_ATTR_RO(cpumask_isolated);
7277 static ssize_t cpumask_show(struct device *dev,
7278 struct device_attribute *attr, char *buf)
7280 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7283 static ssize_t cpumask_store(struct device *dev,
7284 struct device_attribute *attr, const char *buf, size_t count)
7286 cpumask_var_t cpumask;
7289 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7292 ret = cpumask_parse(buf, cpumask);
7294 ret = workqueue_set_unbound_cpumask(cpumask);
7296 free_cpumask_var(cpumask);
7297 return ret ? ret : count;
7299 static DEVICE_ATTR_RW(cpumask);
7301 static struct attribute *wq_sysfs_cpumask_attrs[] = {
7302 &dev_attr_cpumask.attr,
7303 &dev_attr_cpumask_requested.attr,
7304 &dev_attr_cpumask_isolated.attr,
7307 ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7309 static int __init wq_sysfs_init(void)
7311 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups);
7313 core_initcall(wq_sysfs_init);
7315 static void wq_device_release(struct device *dev)
7317 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7323 * workqueue_sysfs_register - make a workqueue visible in sysfs
7324 * @wq: the workqueue to register
7326 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7327 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7328 * which is the preferred method.
7330 * Workqueue user should use this function directly iff it wants to apply
7331 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7332 * apply_workqueue_attrs() may race against userland updating the
7335 * Return: 0 on success, -errno on failure.
7337 int workqueue_sysfs_register(struct workqueue_struct *wq)
7339 struct wq_device *wq_dev;
7343 * Adjusting max_active breaks ordering guarantee. Disallow exposing
7344 * ordered workqueues.
7346 if (WARN_ON(wq->flags & __WQ_ORDERED))
7349 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7354 wq_dev->dev.bus = &wq_subsys;
7355 wq_dev->dev.release = wq_device_release;
7356 dev_set_name(&wq_dev->dev, "%s", wq->name);
7359 * unbound_attrs are created separately. Suppress uevent until
7360 * everything is ready.
7362 dev_set_uevent_suppress(&wq_dev->dev, true);
7364 ret = device_register(&wq_dev->dev);
7366 put_device(&wq_dev->dev);
7371 if (wq->flags & WQ_UNBOUND) {
7372 struct device_attribute *attr;
7374 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7375 ret = device_create_file(&wq_dev->dev, attr);
7377 device_unregister(&wq_dev->dev);
7384 dev_set_uevent_suppress(&wq_dev->dev, false);
7385 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7390 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7391 * @wq: the workqueue to unregister
7393 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7395 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7397 struct wq_device *wq_dev = wq->wq_dev;
7403 device_unregister(&wq_dev->dev);
7405 #else /* CONFIG_SYSFS */
7406 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7407 #endif /* CONFIG_SYSFS */
7410 * Workqueue watchdog.
7412 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7413 * flush dependency, a concurrency managed work item which stays RUNNING
7414 * indefinitely. Workqueue stalls can be very difficult to debug as the
7415 * usual warning mechanisms don't trigger and internal workqueue state is
7418 * Workqueue watchdog monitors all worker pools periodically and dumps
7419 * state if some pools failed to make forward progress for a while where
7420 * forward progress is defined as the first item on ->worklist changing.
7422 * This mechanism is controlled through the kernel parameter
7423 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7424 * corresponding sysfs parameter file.
7426 #ifdef CONFIG_WQ_WATCHDOG
7428 static unsigned long wq_watchdog_thresh = 30;
7429 static struct timer_list wq_watchdog_timer;
7431 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7432 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7435 * Show workers that might prevent the processing of pending work items.
7436 * The only candidates are CPU-bound workers in the running state.
7437 * Pending work items should be handled by another idle worker
7438 * in all other situations.
7440 static void show_cpu_pool_hog(struct worker_pool *pool)
7442 struct worker *worker;
7443 unsigned long irq_flags;
7446 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7448 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7449 if (task_is_running(worker->task)) {
7451 * Defer printing to avoid deadlocks in console
7452 * drivers that queue work while holding locks
7453 * also taken in their write paths.
7455 printk_deferred_enter();
7457 pr_info("pool %d:\n", pool->id);
7458 sched_show_task(worker->task);
7460 printk_deferred_exit();
7464 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7467 static void show_cpu_pools_hogs(void)
7469 struct worker_pool *pool;
7472 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7476 for_each_pool(pool, pi) {
7477 if (pool->cpu_stall)
7478 show_cpu_pool_hog(pool);
7485 static void wq_watchdog_reset_touched(void)
7489 wq_watchdog_touched = jiffies;
7490 for_each_possible_cpu(cpu)
7491 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7494 static void wq_watchdog_timer_fn(struct timer_list *unused)
7496 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7497 bool lockup_detected = false;
7498 bool cpu_pool_stall = false;
7499 unsigned long now = jiffies;
7500 struct worker_pool *pool;
7508 for_each_pool(pool, pi) {
7509 unsigned long pool_ts, touched, ts;
7511 pool->cpu_stall = false;
7512 if (list_empty(&pool->worklist))
7516 * If a virtual machine is stopped by the host it can look to
7517 * the watchdog like a stall.
7519 kvm_check_and_clear_guest_paused();
7521 /* get the latest of pool and touched timestamps */
7523 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7525 touched = READ_ONCE(wq_watchdog_touched);
7526 pool_ts = READ_ONCE(pool->watchdog_ts);
7528 if (time_after(pool_ts, touched))
7534 if (time_after(now, ts + thresh)) {
7535 lockup_detected = true;
7536 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7537 pool->cpu_stall = true;
7538 cpu_pool_stall = true;
7540 pr_emerg("BUG: workqueue lockup - pool");
7541 pr_cont_pool_info(pool);
7542 pr_cont(" stuck for %us!\n",
7543 jiffies_to_msecs(now - pool_ts) / 1000);
7551 if (lockup_detected)
7552 show_all_workqueues();
7555 show_cpu_pools_hogs();
7557 wq_watchdog_reset_touched();
7558 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7561 notrace void wq_watchdog_touch(int cpu)
7564 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7566 wq_watchdog_touched = jiffies;
7569 static void wq_watchdog_set_thresh(unsigned long thresh)
7571 wq_watchdog_thresh = 0;
7572 del_timer_sync(&wq_watchdog_timer);
7575 wq_watchdog_thresh = thresh;
7576 wq_watchdog_reset_touched();
7577 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7581 static int wq_watchdog_param_set_thresh(const char *val,
7582 const struct kernel_param *kp)
7584 unsigned long thresh;
7587 ret = kstrtoul(val, 0, &thresh);
7592 wq_watchdog_set_thresh(thresh);
7594 wq_watchdog_thresh = thresh;
7599 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7600 .set = wq_watchdog_param_set_thresh,
7601 .get = param_get_ulong,
7604 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7607 static void wq_watchdog_init(void)
7609 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7610 wq_watchdog_set_thresh(wq_watchdog_thresh);
7613 #else /* CONFIG_WQ_WATCHDOG */
7615 static inline void wq_watchdog_init(void) { }
7617 #endif /* CONFIG_WQ_WATCHDOG */
7619 static void bh_pool_kick_normal(struct irq_work *irq_work)
7621 raise_softirq_irqoff(TASKLET_SOFTIRQ);
7624 static void bh_pool_kick_highpri(struct irq_work *irq_work)
7626 raise_softirq_irqoff(HI_SOFTIRQ);
7629 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7631 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7632 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7633 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7637 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7640 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7642 BUG_ON(init_worker_pool(pool));
7644 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7645 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7646 pool->attrs->nice = nice;
7647 pool->attrs->affn_strict = true;
7648 pool->node = cpu_to_node(cpu);
7651 mutex_lock(&wq_pool_mutex);
7652 BUG_ON(worker_pool_assign_id(pool));
7653 mutex_unlock(&wq_pool_mutex);
7657 * workqueue_init_early - early init for workqueue subsystem
7659 * This is the first step of three-staged workqueue subsystem initialization and
7660 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7661 * up. It sets up all the data structures and system workqueues and allows early
7662 * boot code to create workqueues and queue/cancel work items. Actual work item
7663 * execution starts only after kthreads can be created and scheduled right
7664 * before early initcalls.
7666 void __init workqueue_init_early(void)
7668 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7669 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7670 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7671 bh_pool_kick_highpri };
7674 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7676 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7677 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7678 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7680 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7681 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7682 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7683 if (!cpumask_empty(&wq_cmdline_cpumask))
7684 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7686 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7688 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7690 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7691 BUG_ON(!wq_update_pod_attrs_buf);
7694 * If nohz_full is enabled, set power efficient workqueue as unbound.
7695 * This allows workqueue items to be moved to HK CPUs.
7697 if (housekeeping_enabled(HK_TYPE_TICK))
7698 wq_power_efficient = true;
7700 /* initialize WQ_AFFN_SYSTEM pods */
7701 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7702 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7703 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7704 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7706 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7709 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7710 pt->pod_node[0] = NUMA_NO_NODE;
7713 /* initialize BH and CPU pools */
7714 for_each_possible_cpu(cpu) {
7715 struct worker_pool *pool;
7718 for_each_bh_worker_pool(pool, cpu) {
7719 init_cpu_worker_pool(pool, cpu, std_nice[i]);
7720 pool->flags |= POOL_BH;
7721 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7726 for_each_cpu_worker_pool(pool, cpu)
7727 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7730 /* create default unbound and ordered wq attrs */
7731 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7732 struct workqueue_attrs *attrs;
7734 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7735 attrs->nice = std_nice[i];
7736 unbound_std_wq_attrs[i] = attrs;
7739 * An ordered wq should have only one pwq as ordering is
7740 * guaranteed by max_active which is enforced by pwqs.
7742 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7743 attrs->nice = std_nice[i];
7744 attrs->ordered = true;
7745 ordered_wq_attrs[i] = attrs;
7748 system_wq = alloc_workqueue("events", 0, 0);
7749 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7750 system_long_wq = alloc_workqueue("events_long", 0, 0);
7751 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7753 system_freezable_wq = alloc_workqueue("events_freezable",
7755 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7756 WQ_POWER_EFFICIENT, 0);
7757 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7758 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7760 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7761 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7762 WQ_BH | WQ_HIGHPRI, 0);
7763 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7764 !system_unbound_wq || !system_freezable_wq ||
7765 !system_power_efficient_wq ||
7766 !system_freezable_power_efficient_wq ||
7767 !system_bh_wq || !system_bh_highpri_wq);
7770 static void __init wq_cpu_intensive_thresh_init(void)
7772 unsigned long thresh;
7775 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7776 BUG_ON(IS_ERR(pwq_release_worker));
7778 /* if the user set it to a specific value, keep it */
7779 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7783 * The default of 10ms is derived from the fact that most modern (as of
7784 * 2023) processors can do a lot in 10ms and that it's just below what
7785 * most consider human-perceivable. However, the kernel also runs on a
7786 * lot slower CPUs including microcontrollers where the threshold is way
7789 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7790 * This is by no means accurate but it doesn't have to be. The mechanism
7791 * is still useful even when the threshold is fully scaled up. Also, as
7792 * the reports would usually be applicable to everyone, some machines
7793 * operating on longer thresholds won't significantly diminish their
7796 thresh = 10 * USEC_PER_MSEC;
7798 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7799 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7801 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7803 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7804 loops_per_jiffy, bogo, thresh);
7806 wq_cpu_intensive_thresh_us = thresh;
7810 * workqueue_init - bring workqueue subsystem fully online
7812 * This is the second step of three-staged workqueue subsystem initialization
7813 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7814 * been created and work items queued on them, but there are no kworkers
7815 * executing the work items yet. Populate the worker pools with the initial
7816 * workers and enable future kworker creations.
7818 void __init workqueue_init(void)
7820 struct workqueue_struct *wq;
7821 struct worker_pool *pool;
7824 wq_cpu_intensive_thresh_init();
7826 mutex_lock(&wq_pool_mutex);
7829 * Per-cpu pools created earlier could be missing node hint. Fix them
7830 * up. Also, create a rescuer for workqueues that requested it.
7832 for_each_possible_cpu(cpu) {
7833 for_each_bh_worker_pool(pool, cpu)
7834 pool->node = cpu_to_node(cpu);
7835 for_each_cpu_worker_pool(pool, cpu)
7836 pool->node = cpu_to_node(cpu);
7839 list_for_each_entry(wq, &workqueues, list) {
7840 WARN(init_rescuer(wq),
7841 "workqueue: failed to create early rescuer for %s",
7845 mutex_unlock(&wq_pool_mutex);
7848 * Create the initial workers. A BH pool has one pseudo worker that
7849 * represents the shared BH execution context and thus doesn't get
7850 * affected by hotplug events. Create the BH pseudo workers for all
7851 * possible CPUs here.
7853 for_each_possible_cpu(cpu)
7854 for_each_bh_worker_pool(pool, cpu)
7855 BUG_ON(!create_worker(pool));
7857 for_each_online_cpu(cpu) {
7858 for_each_cpu_worker_pool(pool, cpu) {
7859 pool->flags &= ~POOL_DISASSOCIATED;
7860 BUG_ON(!create_worker(pool));
7864 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7865 BUG_ON(!create_worker(pool));
7872 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7873 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7874 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7876 static void __init init_pod_type(struct wq_pod_type *pt,
7877 bool (*cpus_share_pod)(int, int))
7879 int cur, pre, cpu, pod;
7883 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7884 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7885 BUG_ON(!pt->cpu_pod);
7887 for_each_possible_cpu(cur) {
7888 for_each_possible_cpu(pre) {
7890 pt->cpu_pod[cur] = pt->nr_pods++;
7893 if (cpus_share_pod(cur, pre)) {
7894 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7900 /* init the rest to match @pt->cpu_pod[] */
7901 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7902 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7903 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7905 for (pod = 0; pod < pt->nr_pods; pod++)
7906 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7908 for_each_possible_cpu(cpu) {
7909 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7910 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7914 static bool __init cpus_dont_share(int cpu0, int cpu1)
7919 static bool __init cpus_share_smt(int cpu0, int cpu1)
7921 #ifdef CONFIG_SCHED_SMT
7922 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7928 static bool __init cpus_share_numa(int cpu0, int cpu1)
7930 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7934 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7936 * This is the third step of three-staged workqueue subsystem initialization and
7937 * invoked after SMP and topology information are fully initialized. It
7938 * initializes the unbound CPU pods accordingly.
7940 void __init workqueue_init_topology(void)
7942 struct workqueue_struct *wq;
7945 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7946 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7947 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7948 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7950 wq_topo_initialized = true;
7952 mutex_lock(&wq_pool_mutex);
7955 * Workqueues allocated earlier would have all CPUs sharing the default
7956 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7957 * combinations to apply per-pod sharing.
7959 list_for_each_entry(wq, &workqueues, list) {
7960 for_each_online_cpu(cpu)
7961 wq_update_pod(wq, cpu, cpu, true);
7962 if (wq->flags & WQ_UNBOUND) {
7963 mutex_lock(&wq->mutex);
7964 wq_update_node_max_active(wq, -1);
7965 mutex_unlock(&wq->mutex);
7969 mutex_unlock(&wq_pool_mutex);
7972 void __warn_flushing_systemwide_wq(void)
7974 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7977 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7979 static int __init workqueue_unbound_cpus_setup(char *str)
7981 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7982 cpumask_clear(&wq_cmdline_cpumask);
7983 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7988 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);