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 */
104 enum wq_internal_consts {
105 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
108 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
111 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
114 /* call for help after 10ms
116 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
117 CREATE_COOLDOWN = HZ, /* time to breath after fail */
120 * Rescue workers are used only on emergencies and shared by
121 * all cpus. Give MIN_NICE.
123 RESCUER_NICE_LEVEL = MIN_NICE,
124 HIGHPRI_NICE_LEVEL = MIN_NICE,
130 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
131 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
132 * msecs_to_jiffies() can't be an initializer.
134 #define BH_WORKER_JIFFIES msecs_to_jiffies(2)
135 #define BH_WORKER_RESTARTS 10
138 * Structure fields follow one of the following exclusion rules.
140 * I: Modifiable by initialization/destruction paths and read-only for
143 * P: Preemption protected. Disabling preemption is enough and should
144 * only be modified and accessed from the local cpu.
146 * L: pool->lock protected. Access with pool->lock held.
148 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
151 * K: Only modified by worker while holding pool->lock. Can be safely read by
152 * self, while holding pool->lock or from IRQ context if %current is the
155 * S: Only modified by worker self.
157 * A: wq_pool_attach_mutex protected.
159 * PL: wq_pool_mutex protected.
161 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
163 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
165 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168 * WQ: wq->mutex protected.
170 * WR: wq->mutex protected for writes. RCU protected for reads.
172 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
173 * with READ_ONCE() without locking.
175 * MD: wq_mayday_lock protected.
177 * WD: Used internally by the watchdog.
180 /* struct worker is defined in workqueue_internal.h */
183 raw_spinlock_t lock; /* the pool lock */
184 int cpu; /* I: the associated cpu */
185 int node; /* I: the associated node ID */
186 int id; /* I: pool ID */
187 unsigned int flags; /* L: flags */
189 unsigned long watchdog_ts; /* L: watchdog timestamp */
190 bool cpu_stall; /* WD: stalled cpu bound pool */
193 * The counter is incremented in a process context on the associated CPU
194 * w/ preemption disabled, and decremented or reset in the same context
195 * but w/ pool->lock held. The readers grab pool->lock and are
196 * guaranteed to see if the counter reached zero.
200 struct list_head worklist; /* L: list of pending works */
202 int nr_workers; /* L: total number of workers */
203 int nr_idle; /* L: currently idle workers */
205 struct list_head idle_list; /* L: list of idle workers */
206 struct timer_list idle_timer; /* L: worker idle timeout */
207 struct work_struct idle_cull_work; /* L: worker idle cleanup */
209 struct timer_list mayday_timer; /* L: SOS timer for workers */
211 /* a workers is either on busy_hash or idle_list, or the manager */
212 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
213 /* L: hash of busy workers */
215 struct worker *manager; /* L: purely informational */
216 struct list_head workers; /* A: attached workers */
217 struct list_head dying_workers; /* A: workers about to die */
218 struct completion *detach_completion; /* all workers detached */
220 struct ida worker_ida; /* worker IDs for task name */
222 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; /* PL: refcnt for unbound pools */
227 * Destruction of pool is RCU protected to allow dereferences
228 * from get_work_pool().
234 * Per-pool_workqueue statistics. These can be monitored using
235 * tools/workqueue/wq_monitor.py.
237 enum pool_workqueue_stats {
238 PWQ_STAT_STARTED, /* work items started execution */
239 PWQ_STAT_COMPLETED, /* work items completed execution */
240 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
242 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
243 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
244 PWQ_STAT_MAYDAY, /* maydays to rescuer */
245 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
251 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
252 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits.
256 struct pool_workqueue {
257 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; /* L: current color */
260 int flush_color; /* L: flushing color */
261 int refcnt; /* L: reference count */
262 int nr_in_flight[WORK_NR_COLORS];
263 /* L: nr of in_flight works */
264 bool plugged; /* L: execution suspended */
267 * nr_active management and WORK_STRUCT_INACTIVE:
269 * When pwq->nr_active >= max_active, new work item is queued to
270 * pwq->inactive_works instead of pool->worklist and marked with
271 * WORK_STRUCT_INACTIVE.
273 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
274 * nr_active and all work items in pwq->inactive_works are marked with
275 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
276 * in pwq->inactive_works. Some of them are ready to run in
277 * pool->worklist or worker->scheduled. Those work itmes are only struct
278 * wq_barrier which is used for flush_work() and should not participate
279 * in nr_active. For non-barrier work item, it is marked with
280 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
282 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works; /* L: inactive works */
284 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
285 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; /* MD: node on wq->maydays */
288 u64 stats[PWQ_NR_STATS];
291 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292 * and pwq_release_workfn() for details. pool_workqueue itself is also
293 * RCU protected so that the first pwq can be determined without
294 * grabbing wq->mutex.
296 struct kthread_work release_work;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
301 * Structure used to wait for workqueue flush.
304 struct list_head list; /* WQ: list of flushers */
305 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; /* flush completion */
312 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
313 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
314 * As sharing a single nr_active across multiple sockets can be very expensive,
315 * the counting and enforcement is per NUMA node.
317 * The following struct is used to enforce per-node max_active. When a pwq wants
318 * to start executing a work item, it should increment ->nr using
319 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
320 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
321 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
324 struct wq_node_nr_active {
325 int max; /* per-node max_active */
326 atomic_t nr; /* per-node nr_active */
327 raw_spinlock_t lock; /* nests inside pool locks */
328 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
332 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its pool_workqueues.
335 struct workqueue_struct {
336 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; /* PR: list of all workqueues */
339 struct mutex mutex; /* protects this wq */
340 int work_color; /* WQ: current work color */
341 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overflow; /* WQ: flush overflow list */
347 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; /* MD: rescue worker */
350 int nr_drainers; /* WQ: drain in progress */
352 /* See alloc_workqueue() function comment for info on min/max_active */
353 int max_active; /* WO: max active works */
354 int min_active; /* WO: min active works */
355 int saved_max_active; /* WQ: saved max_active */
356 int saved_min_active; /* WQ: saved min_active */
358 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
362 struct wq_device *wq_dev; /* I: for sysfs interface */
364 #ifdef CONFIG_LOCKDEP
366 struct lock_class_key key;
367 struct lockdep_map lockdep_map;
369 char name[WQ_NAME_LEN]; /* I: workqueue name */
372 * Destruction of workqueue_struct is RCU protected to allow walking
373 * the workqueues list without grabbing wq_pool_mutex.
374 * This is used to dump all workqueues from sysrq.
378 /* hot fields used during command issue, aligned to cacheline */
379 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
381 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
385 * Each pod type describes how CPUs should be grouped for unbound workqueues.
386 * See the comment above workqueue_attrs->affn_scope.
389 int nr_pods; /* number of pods */
390 cpumask_var_t *pod_cpus; /* pod -> cpus */
391 int *pod_node; /* pod -> node */
392 int *cpu_pod; /* cpu -> pod */
395 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
396 [WQ_AFFN_DFL] = "default",
397 [WQ_AFFN_CPU] = "cpu",
398 [WQ_AFFN_SMT] = "smt",
399 [WQ_AFFN_CACHE] = "cache",
400 [WQ_AFFN_NUMA] = "numa",
401 [WQ_AFFN_SYSTEM] = "system",
405 * Per-cpu work items which run for longer than the following threshold are
406 * automatically considered CPU intensive and excluded from concurrency
407 * management to prevent them from noticeably delaying other per-cpu work items.
408 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
409 * The actual value is initialized in wq_cpu_intensive_thresh_init().
411 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
412 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
413 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
414 static unsigned int wq_cpu_intensive_warning_thresh = 4;
415 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
418 /* see the comment above the definition of WQ_POWER_EFFICIENT */
419 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
420 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
422 static bool wq_online; /* can kworkers be created yet? */
423 static bool wq_topo_initialized __read_mostly = false;
425 static struct kmem_cache *pwq_cache;
427 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
428 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
430 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
431 static struct workqueue_attrs *wq_update_pod_attrs_buf;
433 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
434 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
435 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
436 /* wait for manager to go away */
437 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
439 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
440 static bool workqueue_freezing; /* PL: have wqs started freezing? */
442 /* PL&A: allowable cpus for unbound wqs and work items */
443 static cpumask_var_t wq_unbound_cpumask;
445 /* PL: user requested unbound cpumask via sysfs */
446 static cpumask_var_t wq_requested_unbound_cpumask;
448 /* PL: isolated cpumask to be excluded from unbound cpumask */
449 static cpumask_var_t wq_isolated_cpumask;
451 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
452 static struct cpumask wq_cmdline_cpumask __initdata;
454 /* CPU where unbound work was last round robin scheduled from this CPU */
455 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
458 * Local execution of unbound work items is no longer guaranteed. The
459 * following always forces round-robin CPU selection on unbound work items
460 * to uncover usages which depend on it.
462 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
463 static bool wq_debug_force_rr_cpu = true;
465 static bool wq_debug_force_rr_cpu = false;
467 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
469 /* to raise softirq for the BH worker pools on other CPUs */
470 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS],
473 /* the BH worker pools */
474 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
477 /* the per-cpu worker pools */
478 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
481 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
483 /* PL: hash of all unbound pools keyed by pool->attrs */
484 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
486 /* I: attributes used when instantiating standard unbound pools on demand */
487 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
489 /* I: attributes used when instantiating ordered pools on demand */
490 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
493 * Used to synchronize multiple cancel_sync attempts on the same work item. See
494 * work_grab_pending() and __cancel_work_sync().
496 static DECLARE_WAIT_QUEUE_HEAD(wq_cancel_waitq);
499 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
500 * process context while holding a pool lock. Bounce to a dedicated kthread
501 * worker to avoid A-A deadlocks.
503 static struct kthread_worker *pwq_release_worker __ro_after_init;
505 struct workqueue_struct *system_wq __ro_after_init;
506 EXPORT_SYMBOL(system_wq);
507 struct workqueue_struct *system_highpri_wq __ro_after_init;
508 EXPORT_SYMBOL_GPL(system_highpri_wq);
509 struct workqueue_struct *system_long_wq __ro_after_init;
510 EXPORT_SYMBOL_GPL(system_long_wq);
511 struct workqueue_struct *system_unbound_wq __ro_after_init;
512 EXPORT_SYMBOL_GPL(system_unbound_wq);
513 struct workqueue_struct *system_freezable_wq __ro_after_init;
514 EXPORT_SYMBOL_GPL(system_freezable_wq);
515 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
516 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
517 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
518 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
519 struct workqueue_struct *system_bh_wq;
520 EXPORT_SYMBOL_GPL(system_bh_wq);
521 struct workqueue_struct *system_bh_highpri_wq;
522 EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
524 static int worker_thread(void *__worker);
525 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
526 static void show_pwq(struct pool_workqueue *pwq);
527 static void show_one_worker_pool(struct worker_pool *pool);
529 #define CREATE_TRACE_POINTS
530 #include <trace/events/workqueue.h>
532 #define assert_rcu_or_pool_mutex() \
533 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
534 !lockdep_is_held(&wq_pool_mutex), \
535 "RCU or wq_pool_mutex should be held")
537 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
538 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
539 !lockdep_is_held(&wq->mutex) && \
540 !lockdep_is_held(&wq_pool_mutex), \
541 "RCU, wq->mutex or wq_pool_mutex should be held")
543 #define for_each_bh_worker_pool(pool, cpu) \
544 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
545 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
548 #define for_each_cpu_worker_pool(pool, cpu) \
549 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
550 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554 * for_each_pool - iterate through all worker_pools in the system
555 * @pool: iteration cursor
556 * @pi: integer used for iteration
558 * This must be called either with wq_pool_mutex held or RCU read
559 * locked. If the pool needs to be used beyond the locking in effect, the
560 * caller is responsible for guaranteeing that the pool stays online.
562 * The if/else clause exists only for the lockdep assertion and can be
565 #define for_each_pool(pool, pi) \
566 idr_for_each_entry(&worker_pool_idr, pool, pi) \
567 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571 * for_each_pool_worker - iterate through all workers of a worker_pool
572 * @worker: iteration cursor
573 * @pool: worker_pool to iterate workers of
575 * This must be called with wq_pool_attach_mutex.
577 * The if/else clause exists only for the lockdep assertion and can be
580 #define for_each_pool_worker(worker, pool) \
581 list_for_each_entry((worker), &(pool)->workers, node) \
582 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
586 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
587 * @pwq: iteration cursor
588 * @wq: the target workqueue
590 * This must be called either with wq->mutex held or RCU read locked.
591 * If the pwq needs to be used beyond the locking in effect, the caller is
592 * responsible for guaranteeing that the pwq stays online.
594 * The if/else clause exists only for the lockdep assertion and can be
597 #define for_each_pwq(pwq, wq) \
598 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
599 lockdep_is_held(&(wq->mutex)))
601 #ifdef CONFIG_DEBUG_OBJECTS_WORK
603 static const struct debug_obj_descr work_debug_descr;
605 static void *work_debug_hint(void *addr)
607 return ((struct work_struct *) addr)->func;
610 static bool work_is_static_object(void *addr)
612 struct work_struct *work = addr;
614 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
618 * fixup_init is called when:
619 * - an active object is initialized
621 static bool work_fixup_init(void *addr, enum debug_obj_state state)
623 struct work_struct *work = addr;
626 case ODEBUG_STATE_ACTIVE:
627 cancel_work_sync(work);
628 debug_object_init(work, &work_debug_descr);
636 * fixup_free is called when:
637 * - an active object is freed
639 static bool work_fixup_free(void *addr, enum debug_obj_state state)
641 struct work_struct *work = addr;
644 case ODEBUG_STATE_ACTIVE:
645 cancel_work_sync(work);
646 debug_object_free(work, &work_debug_descr);
653 static const struct debug_obj_descr work_debug_descr = {
654 .name = "work_struct",
655 .debug_hint = work_debug_hint,
656 .is_static_object = work_is_static_object,
657 .fixup_init = work_fixup_init,
658 .fixup_free = work_fixup_free,
661 static inline void debug_work_activate(struct work_struct *work)
663 debug_object_activate(work, &work_debug_descr);
666 static inline void debug_work_deactivate(struct work_struct *work)
668 debug_object_deactivate(work, &work_debug_descr);
671 void __init_work(struct work_struct *work, int onstack)
674 debug_object_init_on_stack(work, &work_debug_descr);
676 debug_object_init(work, &work_debug_descr);
678 EXPORT_SYMBOL_GPL(__init_work);
680 void destroy_work_on_stack(struct work_struct *work)
682 debug_object_free(work, &work_debug_descr);
684 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
686 void destroy_delayed_work_on_stack(struct delayed_work *work)
688 destroy_timer_on_stack(&work->timer);
689 debug_object_free(&work->work, &work_debug_descr);
691 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
694 static inline void debug_work_activate(struct work_struct *work) { }
695 static inline void debug_work_deactivate(struct work_struct *work) { }
699 * worker_pool_assign_id - allocate ID and assign it to @pool
700 * @pool: the pool pointer of interest
702 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
703 * successfully, -errno on failure.
705 static int worker_pool_assign_id(struct worker_pool *pool)
709 lockdep_assert_held(&wq_pool_mutex);
711 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
720 static struct pool_workqueue __rcu **
721 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
724 return per_cpu_ptr(wq->cpu_pwq, cpu);
729 /* @cpu < 0 for dfl_pwq */
730 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
732 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
733 lockdep_is_held(&wq_pool_mutex) ||
734 lockdep_is_held(&wq->mutex));
738 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
739 * @wq: workqueue of interest
741 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
742 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
743 * default pwq is always mapped to the pool with the current effective cpumask.
745 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
747 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
750 static unsigned int work_color_to_flags(int color)
752 return color << WORK_STRUCT_COLOR_SHIFT;
755 static int get_work_color(unsigned long work_data)
757 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
758 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
761 static int work_next_color(int color)
763 return (color + 1) % WORK_NR_COLORS;
767 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
768 * contain the pointer to the queued pwq. Once execution starts, the flag
769 * is cleared and the high bits contain OFFQ flags and pool ID.
771 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
772 * can be used to set the pwq, pool or clear work->data. These functions should
773 * only be called while the work is owned - ie. while the PENDING bit is set.
775 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
776 * corresponding to a work. Pool is available once the work has been
777 * queued anywhere after initialization until it is sync canceled. pwq is
778 * available only while the work item is queued.
780 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
781 * canceled. While being canceled, a work item may have its PENDING set
782 * but stay off timer and worklist for arbitrarily long and nobody should
783 * try to steal the PENDING bit.
785 static inline void set_work_data(struct work_struct *work, unsigned long data)
787 WARN_ON_ONCE(!work_pending(work));
788 atomic_long_set(&work->data, data | work_static(work));
791 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
794 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
795 WORK_STRUCT_PWQ | flags);
798 static void set_work_pool_and_keep_pending(struct work_struct *work,
799 int pool_id, unsigned long flags)
801 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
802 WORK_STRUCT_PENDING | flags);
805 static void set_work_pool_and_clear_pending(struct work_struct *work,
806 int pool_id, unsigned long flags)
809 * The following wmb is paired with the implied mb in
810 * test_and_set_bit(PENDING) and ensures all updates to @work made
811 * here are visible to and precede any updates by the next PENDING
815 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
818 * The following mb guarantees that previous clear of a PENDING bit
819 * will not be reordered with any speculative LOADS or STORES from
820 * work->current_func, which is executed afterwards. This possible
821 * reordering can lead to a missed execution on attempt to queue
822 * the same @work. E.g. consider this case:
825 * ---------------------------- --------------------------------
827 * 1 STORE event_indicated
828 * 2 queue_work_on() {
829 * 3 test_and_set_bit(PENDING)
830 * 4 } set_..._and_clear_pending() {
831 * 5 set_work_data() # clear bit
833 * 7 work->current_func() {
834 * 8 LOAD event_indicated
837 * Without an explicit full barrier speculative LOAD on line 8 can
838 * be executed before CPU#0 does STORE on line 1. If that happens,
839 * CPU#0 observes the PENDING bit is still set and new execution of
840 * a @work is not queued in a hope, that CPU#1 will eventually
841 * finish the queued @work. Meanwhile CPU#1 does not see
842 * event_indicated is set, because speculative LOAD was executed
843 * before actual STORE.
848 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
850 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
853 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
855 unsigned long data = atomic_long_read(&work->data);
857 if (data & WORK_STRUCT_PWQ)
858 return work_struct_pwq(data);
864 * get_work_pool - return the worker_pool a given work was associated with
865 * @work: the work item of interest
867 * Pools are created and destroyed under wq_pool_mutex, and allows read
868 * access under RCU read lock. As such, this function should be
869 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
871 * All fields of the returned pool are accessible as long as the above
872 * mentioned locking is in effect. If the returned pool needs to be used
873 * beyond the critical section, the caller is responsible for ensuring the
874 * returned pool is and stays online.
876 * Return: The worker_pool @work was last associated with. %NULL if none.
878 static struct worker_pool *get_work_pool(struct work_struct *work)
880 unsigned long data = atomic_long_read(&work->data);
883 assert_rcu_or_pool_mutex();
885 if (data & WORK_STRUCT_PWQ)
886 return work_struct_pwq(data)->pool;
888 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
889 if (pool_id == WORK_OFFQ_POOL_NONE)
892 return idr_find(&worker_pool_idr, pool_id);
896 * get_work_pool_id - return the worker pool ID a given work is associated with
897 * @work: the work item of interest
899 * Return: The worker_pool ID @work was last associated with.
900 * %WORK_OFFQ_POOL_NONE if none.
902 static int get_work_pool_id(struct work_struct *work)
904 unsigned long data = atomic_long_read(&work->data);
906 if (data & WORK_STRUCT_PWQ)
907 return work_struct_pwq(data)->pool->id;
909 return data >> WORK_OFFQ_POOL_SHIFT;
912 static void mark_work_canceling(struct work_struct *work)
914 unsigned long pool_id = get_work_pool_id(work);
916 pool_id <<= WORK_OFFQ_POOL_SHIFT;
917 set_work_data(work, pool_id | WORK_STRUCT_PENDING | WORK_OFFQ_CANCELING);
920 static bool work_is_canceling(struct work_struct *work)
922 unsigned long data = atomic_long_read(&work->data);
924 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
928 * Policy functions. These define the policies on how the global worker
929 * pools are managed. Unless noted otherwise, these functions assume that
930 * they're being called with pool->lock held.
934 * Need to wake up a worker? Called from anything but currently
937 * Note that, because unbound workers never contribute to nr_running, this
938 * function will always return %true for unbound pools as long as the
939 * worklist isn't empty.
941 static bool need_more_worker(struct worker_pool *pool)
943 return !list_empty(&pool->worklist) && !pool->nr_running;
946 /* Can I start working? Called from busy but !running workers. */
947 static bool may_start_working(struct worker_pool *pool)
949 return pool->nr_idle;
952 /* Do I need to keep working? Called from currently running workers. */
953 static bool keep_working(struct worker_pool *pool)
955 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
958 /* Do we need a new worker? Called from manager. */
959 static bool need_to_create_worker(struct worker_pool *pool)
961 return need_more_worker(pool) && !may_start_working(pool);
964 /* Do we have too many workers and should some go away? */
965 static bool too_many_workers(struct worker_pool *pool)
967 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
968 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
969 int nr_busy = pool->nr_workers - nr_idle;
971 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
975 * worker_set_flags - set worker flags and adjust nr_running accordingly
977 * @flags: flags to set
979 * Set @flags in @worker->flags and adjust nr_running accordingly.
981 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
983 struct worker_pool *pool = worker->pool;
985 lockdep_assert_held(&pool->lock);
987 /* If transitioning into NOT_RUNNING, adjust nr_running. */
988 if ((flags & WORKER_NOT_RUNNING) &&
989 !(worker->flags & WORKER_NOT_RUNNING)) {
993 worker->flags |= flags;
997 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
999 * @flags: flags to clear
1001 * Clear @flags in @worker->flags and adjust nr_running accordingly.
1003 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
1005 struct worker_pool *pool = worker->pool;
1006 unsigned int oflags = worker->flags;
1008 lockdep_assert_held(&pool->lock);
1010 worker->flags &= ~flags;
1013 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1014 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1015 * of multiple flags, not a single flag.
1017 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1018 if (!(worker->flags & WORKER_NOT_RUNNING))
1022 /* Return the first idle worker. Called with pool->lock held. */
1023 static struct worker *first_idle_worker(struct worker_pool *pool)
1025 if (unlikely(list_empty(&pool->idle_list)))
1028 return list_first_entry(&pool->idle_list, struct worker, entry);
1032 * worker_enter_idle - enter idle state
1033 * @worker: worker which is entering idle state
1035 * @worker is entering idle state. Update stats and idle timer if
1039 * raw_spin_lock_irq(pool->lock).
1041 static void worker_enter_idle(struct worker *worker)
1043 struct worker_pool *pool = worker->pool;
1045 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1046 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1047 (worker->hentry.next || worker->hentry.pprev)))
1050 /* can't use worker_set_flags(), also called from create_worker() */
1051 worker->flags |= WORKER_IDLE;
1053 worker->last_active = jiffies;
1055 /* idle_list is LIFO */
1056 list_add(&worker->entry, &pool->idle_list);
1058 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1059 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1061 /* Sanity check nr_running. */
1062 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1066 * worker_leave_idle - leave idle state
1067 * @worker: worker which is leaving idle state
1069 * @worker is leaving idle state. Update stats.
1072 * raw_spin_lock_irq(pool->lock).
1074 static void worker_leave_idle(struct worker *worker)
1076 struct worker_pool *pool = worker->pool;
1078 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1080 worker_clr_flags(worker, WORKER_IDLE);
1082 list_del_init(&worker->entry);
1086 * find_worker_executing_work - find worker which is executing a work
1087 * @pool: pool of interest
1088 * @work: work to find worker for
1090 * Find a worker which is executing @work on @pool by searching
1091 * @pool->busy_hash which is keyed by the address of @work. For a worker
1092 * to match, its current execution should match the address of @work and
1093 * its work function. This is to avoid unwanted dependency between
1094 * unrelated work executions through a work item being recycled while still
1097 * This is a bit tricky. A work item may be freed once its execution
1098 * starts and nothing prevents the freed area from being recycled for
1099 * another work item. If the same work item address ends up being reused
1100 * before the original execution finishes, workqueue will identify the
1101 * recycled work item as currently executing and make it wait until the
1102 * current execution finishes, introducing an unwanted dependency.
1104 * This function checks the work item address and work function to avoid
1105 * false positives. Note that this isn't complete as one may construct a
1106 * work function which can introduce dependency onto itself through a
1107 * recycled work item. Well, if somebody wants to shoot oneself in the
1108 * foot that badly, there's only so much we can do, and if such deadlock
1109 * actually occurs, it should be easy to locate the culprit work function.
1112 * raw_spin_lock_irq(pool->lock).
1115 * Pointer to worker which is executing @work if found, %NULL
1118 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1119 struct work_struct *work)
1121 struct worker *worker;
1123 hash_for_each_possible(pool->busy_hash, worker, hentry,
1124 (unsigned long)work)
1125 if (worker->current_work == work &&
1126 worker->current_func == work->func)
1133 * move_linked_works - move linked works to a list
1134 * @work: start of series of works to be scheduled
1135 * @head: target list to append @work to
1136 * @nextp: out parameter for nested worklist walking
1138 * Schedule linked works starting from @work to @head. Work series to be
1139 * scheduled starts at @work and includes any consecutive work with
1140 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1144 * raw_spin_lock_irq(pool->lock).
1146 static void move_linked_works(struct work_struct *work, struct list_head *head,
1147 struct work_struct **nextp)
1149 struct work_struct *n;
1152 * Linked worklist will always end before the end of the list,
1153 * use NULL for list head.
1155 list_for_each_entry_safe_from(work, n, NULL, entry) {
1156 list_move_tail(&work->entry, head);
1157 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1162 * If we're already inside safe list traversal and have moved
1163 * multiple works to the scheduled queue, the next position
1164 * needs to be updated.
1171 * assign_work - assign a work item and its linked work items to a worker
1172 * @work: work to assign
1173 * @worker: worker to assign to
1174 * @nextp: out parameter for nested worklist walking
1176 * Assign @work and its linked work items to @worker. If @work is already being
1177 * executed by another worker in the same pool, it'll be punted there.
1179 * If @nextp is not NULL, it's updated to point to the next work of the last
1180 * scheduled work. This allows assign_work() to be nested inside
1181 * list_for_each_entry_safe().
1183 * Returns %true if @work was successfully assigned to @worker. %false if @work
1184 * was punted to another worker already executing it.
1186 static bool assign_work(struct work_struct *work, struct worker *worker,
1187 struct work_struct **nextp)
1189 struct worker_pool *pool = worker->pool;
1190 struct worker *collision;
1192 lockdep_assert_held(&pool->lock);
1195 * A single work shouldn't be executed concurrently by multiple workers.
1196 * __queue_work() ensures that @work doesn't jump to a different pool
1197 * while still running in the previous pool. Here, we should ensure that
1198 * @work is not executed concurrently by multiple workers from the same
1199 * pool. Check whether anyone is already processing the work. If so,
1200 * defer the work to the currently executing one.
1202 collision = find_worker_executing_work(pool, work);
1203 if (unlikely(collision)) {
1204 move_linked_works(work, &collision->scheduled, nextp);
1208 move_linked_works(work, &worker->scheduled, nextp);
1212 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1214 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1216 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1219 static void kick_bh_pool(struct worker_pool *pool)
1222 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1223 if (unlikely(pool->cpu != smp_processor_id() &&
1224 !(pool->flags & POOL_BH_DRAINING))) {
1225 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1229 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1230 raise_softirq_irqoff(HI_SOFTIRQ);
1232 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1236 * kick_pool - wake up an idle worker if necessary
1237 * @pool: pool to kick
1239 * @pool may have pending work items. Wake up worker if necessary. Returns
1240 * whether a worker was woken up.
1242 static bool kick_pool(struct worker_pool *pool)
1244 struct worker *worker = first_idle_worker(pool);
1245 struct task_struct *p;
1247 lockdep_assert_held(&pool->lock);
1249 if (!need_more_worker(pool) || !worker)
1252 if (pool->flags & POOL_BH) {
1261 * Idle @worker is about to execute @work and waking up provides an
1262 * opportunity to migrate @worker at a lower cost by setting the task's
1263 * wake_cpu field. Let's see if we want to move @worker to improve
1264 * execution locality.
1266 * We're waking the worker that went idle the latest and there's some
1267 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1268 * so, setting the wake_cpu won't do anything. As this is a best-effort
1269 * optimization and the race window is narrow, let's leave as-is for
1270 * now. If this becomes pronounced, we can skip over workers which are
1271 * still on cpu when picking an idle worker.
1273 * If @pool has non-strict affinity, @worker might have ended up outside
1274 * its affinity scope. Repatriate.
1276 if (!pool->attrs->affn_strict &&
1277 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1278 struct work_struct *work = list_first_entry(&pool->worklist,
1279 struct work_struct, entry);
1280 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1282 if (wake_cpu < nr_cpu_ids) {
1283 p->wake_cpu = wake_cpu;
1284 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1292 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1295 * Concurrency-managed per-cpu work items that hog CPU for longer than
1296 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1297 * which prevents them from stalling other concurrency-managed work items. If a
1298 * work function keeps triggering this mechanism, it's likely that the work item
1299 * should be using an unbound workqueue instead.
1301 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1302 * and report them so that they can be examined and converted to use unbound
1303 * workqueues as appropriate. To avoid flooding the console, each violating work
1304 * function is tracked and reported with exponential backoff.
1306 #define WCI_MAX_ENTS 128
1311 struct hlist_node hash_node;
1314 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1315 static int wci_nr_ents;
1316 static DEFINE_RAW_SPINLOCK(wci_lock);
1317 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1319 static struct wci_ent *wci_find_ent(work_func_t func)
1321 struct wci_ent *ent;
1323 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1324 (unsigned long)func) {
1325 if (ent->func == func)
1331 static void wq_cpu_intensive_report(work_func_t func)
1333 struct wci_ent *ent;
1336 ent = wci_find_ent(func);
1341 * Start reporting from the warning_thresh and back off
1344 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1345 if (wq_cpu_intensive_warning_thresh &&
1346 cnt >= wq_cpu_intensive_warning_thresh &&
1347 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1348 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1349 ent->func, wq_cpu_intensive_thresh_us,
1350 atomic64_read(&ent->cnt));
1355 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1356 * is exhausted, something went really wrong and we probably made enough
1359 if (wci_nr_ents >= WCI_MAX_ENTS)
1362 raw_spin_lock(&wci_lock);
1364 if (wci_nr_ents >= WCI_MAX_ENTS) {
1365 raw_spin_unlock(&wci_lock);
1369 if (wci_find_ent(func)) {
1370 raw_spin_unlock(&wci_lock);
1374 ent = &wci_ents[wci_nr_ents++];
1376 atomic64_set(&ent->cnt, 0);
1377 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1379 raw_spin_unlock(&wci_lock);
1384 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1385 static void wq_cpu_intensive_report(work_func_t func) {}
1386 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1389 * wq_worker_running - a worker is running again
1390 * @task: task waking up
1392 * This function is called when a worker returns from schedule()
1394 void wq_worker_running(struct task_struct *task)
1396 struct worker *worker = kthread_data(task);
1398 if (!READ_ONCE(worker->sleeping))
1402 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1403 * and the nr_running increment below, we may ruin the nr_running reset
1404 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1405 * pool. Protect against such race.
1408 if (!(worker->flags & WORKER_NOT_RUNNING))
1409 worker->pool->nr_running++;
1413 * CPU intensive auto-detection cares about how long a work item hogged
1414 * CPU without sleeping. Reset the starting timestamp on wakeup.
1416 worker->current_at = worker->task->se.sum_exec_runtime;
1418 WRITE_ONCE(worker->sleeping, 0);
1422 * wq_worker_sleeping - a worker is going to sleep
1423 * @task: task going to sleep
1425 * This function is called from schedule() when a busy worker is
1428 void wq_worker_sleeping(struct task_struct *task)
1430 struct worker *worker = kthread_data(task);
1431 struct worker_pool *pool;
1434 * Rescuers, which may not have all the fields set up like normal
1435 * workers, also reach here, let's not access anything before
1436 * checking NOT_RUNNING.
1438 if (worker->flags & WORKER_NOT_RUNNING)
1441 pool = worker->pool;
1443 /* Return if preempted before wq_worker_running() was reached */
1444 if (READ_ONCE(worker->sleeping))
1447 WRITE_ONCE(worker->sleeping, 1);
1448 raw_spin_lock_irq(&pool->lock);
1451 * Recheck in case unbind_workers() preempted us. We don't
1452 * want to decrement nr_running after the worker is unbound
1453 * and nr_running has been reset.
1455 if (worker->flags & WORKER_NOT_RUNNING) {
1456 raw_spin_unlock_irq(&pool->lock);
1461 if (kick_pool(pool))
1462 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1464 raw_spin_unlock_irq(&pool->lock);
1468 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1469 * @task: task currently running
1471 * Called from sched_tick(). We're in the IRQ context and the current
1472 * worker's fields which follow the 'K' locking rule can be accessed safely.
1474 void wq_worker_tick(struct task_struct *task)
1476 struct worker *worker = kthread_data(task);
1477 struct pool_workqueue *pwq = worker->current_pwq;
1478 struct worker_pool *pool = worker->pool;
1483 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1485 if (!wq_cpu_intensive_thresh_us)
1489 * If the current worker is concurrency managed and hogged the CPU for
1490 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1491 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1493 * Set @worker->sleeping means that @worker is in the process of
1494 * switching out voluntarily and won't be contributing to
1495 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1496 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1497 * double decrements. The task is releasing the CPU anyway. Let's skip.
1498 * We probably want to make this prettier in the future.
1500 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1501 worker->task->se.sum_exec_runtime - worker->current_at <
1502 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1505 raw_spin_lock(&pool->lock);
1507 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1508 wq_cpu_intensive_report(worker->current_func);
1509 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1511 if (kick_pool(pool))
1512 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1514 raw_spin_unlock(&pool->lock);
1518 * wq_worker_last_func - retrieve worker's last work function
1519 * @task: Task to retrieve last work function of.
1521 * Determine the last function a worker executed. This is called from
1522 * the scheduler to get a worker's last known identity.
1525 * raw_spin_lock_irq(rq->lock)
1527 * This function is called during schedule() when a kworker is going
1528 * to sleep. It's used by psi to identify aggregation workers during
1529 * dequeuing, to allow periodic aggregation to shut-off when that
1530 * worker is the last task in the system or cgroup to go to sleep.
1532 * As this function doesn't involve any workqueue-related locking, it
1533 * only returns stable values when called from inside the scheduler's
1534 * queuing and dequeuing paths, when @task, which must be a kworker,
1535 * is guaranteed to not be processing any works.
1538 * The last work function %current executed as a worker, NULL if it
1539 * hasn't executed any work yet.
1541 work_func_t wq_worker_last_func(struct task_struct *task)
1543 struct worker *worker = kthread_data(task);
1545 return worker->last_func;
1549 * wq_node_nr_active - Determine wq_node_nr_active to use
1550 * @wq: workqueue of interest
1551 * @node: NUMA node, can be %NUMA_NO_NODE
1553 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1555 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1557 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1559 * - Otherwise, node_nr_active[@node].
1561 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1564 if (!(wq->flags & WQ_UNBOUND))
1567 if (node == NUMA_NO_NODE)
1570 return wq->node_nr_active[node];
1574 * wq_update_node_max_active - Update per-node max_actives to use
1575 * @wq: workqueue to update
1576 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1578 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1579 * distributed among nodes according to the proportions of numbers of online
1580 * cpus. The result is always between @wq->min_active and max_active.
1582 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1584 struct cpumask *effective = unbound_effective_cpumask(wq);
1585 int min_active = READ_ONCE(wq->min_active);
1586 int max_active = READ_ONCE(wq->max_active);
1587 int total_cpus, node;
1589 lockdep_assert_held(&wq->mutex);
1591 if (!wq_topo_initialized)
1594 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1597 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1601 /* If all CPUs of the wq get offline, use the default values */
1602 if (unlikely(!total_cpus)) {
1604 wq_node_nr_active(wq, node)->max = min_active;
1606 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1610 for_each_node(node) {
1613 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1614 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1617 wq_node_nr_active(wq, node)->max =
1618 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1619 min_active, max_active);
1622 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1626 * get_pwq - get an extra reference on the specified pool_workqueue
1627 * @pwq: pool_workqueue to get
1629 * Obtain an extra reference on @pwq. The caller should guarantee that
1630 * @pwq has positive refcnt and be holding the matching pool->lock.
1632 static void get_pwq(struct pool_workqueue *pwq)
1634 lockdep_assert_held(&pwq->pool->lock);
1635 WARN_ON_ONCE(pwq->refcnt <= 0);
1640 * put_pwq - put a pool_workqueue reference
1641 * @pwq: pool_workqueue to put
1643 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1644 * destruction. The caller should be holding the matching pool->lock.
1646 static void put_pwq(struct pool_workqueue *pwq)
1648 lockdep_assert_held(&pwq->pool->lock);
1649 if (likely(--pwq->refcnt))
1652 * @pwq can't be released under pool->lock, bounce to a dedicated
1653 * kthread_worker to avoid A-A deadlocks.
1655 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1659 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1660 * @pwq: pool_workqueue to put (can be %NULL)
1662 * put_pwq() with locking. This function also allows %NULL @pwq.
1664 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1668 * As both pwqs and pools are RCU protected, the
1669 * following lock operations are safe.
1671 raw_spin_lock_irq(&pwq->pool->lock);
1673 raw_spin_unlock_irq(&pwq->pool->lock);
1677 static bool pwq_is_empty(struct pool_workqueue *pwq)
1679 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1682 static void __pwq_activate_work(struct pool_workqueue *pwq,
1683 struct work_struct *work)
1685 unsigned long *wdb = work_data_bits(work);
1687 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1688 trace_workqueue_activate_work(work);
1689 if (list_empty(&pwq->pool->worklist))
1690 pwq->pool->watchdog_ts = jiffies;
1691 move_linked_works(work, &pwq->pool->worklist, NULL);
1692 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1696 * pwq_activate_work - Activate a work item if inactive
1697 * @pwq: pool_workqueue @work belongs to
1698 * @work: work item to activate
1700 * Returns %true if activated. %false if already active.
1702 static bool pwq_activate_work(struct pool_workqueue *pwq,
1703 struct work_struct *work)
1705 struct worker_pool *pool = pwq->pool;
1706 struct wq_node_nr_active *nna;
1708 lockdep_assert_held(&pool->lock);
1710 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE))
1713 nna = wq_node_nr_active(pwq->wq, pool->node);
1715 atomic_inc(&nna->nr);
1718 __pwq_activate_work(pwq, work);
1722 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1724 int max = READ_ONCE(nna->max);
1729 old = atomic_read(&nna->nr);
1732 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1739 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1740 * @pwq: pool_workqueue of interest
1741 * @fill: max_active may have increased, try to increase concurrency level
1743 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1744 * successfully obtained. %false otherwise.
1746 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1748 struct workqueue_struct *wq = pwq->wq;
1749 struct worker_pool *pool = pwq->pool;
1750 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1751 bool obtained = false;
1753 lockdep_assert_held(&pool->lock);
1756 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1757 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1761 if (unlikely(pwq->plugged))
1765 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1766 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1767 * concurrency level. Don't jump the line.
1769 * We need to ignore the pending test after max_active has increased as
1770 * pwq_dec_nr_active() can only maintain the concurrency level but not
1771 * increase it. This is indicated by @fill.
1773 if (!list_empty(&pwq->pending_node) && likely(!fill))
1776 obtained = tryinc_node_nr_active(nna);
1781 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1782 * and try again. The smp_mb() is paired with the implied memory barrier
1783 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1784 * we see the decremented $nna->nr or they see non-empty
1785 * $nna->pending_pwqs.
1787 raw_spin_lock(&nna->lock);
1789 if (list_empty(&pwq->pending_node))
1790 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1791 else if (likely(!fill))
1796 obtained = tryinc_node_nr_active(nna);
1799 * If @fill, @pwq might have already been pending. Being spuriously
1800 * pending in cold paths doesn't affect anything. Let's leave it be.
1802 if (obtained && likely(!fill))
1803 list_del_init(&pwq->pending_node);
1806 raw_spin_unlock(&nna->lock);
1814 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1815 * @pwq: pool_workqueue of interest
1816 * @fill: max_active may have increased, try to increase concurrency level
1818 * Activate the first inactive work item of @pwq if available and allowed by
1821 * Returns %true if an inactive work item has been activated. %false if no
1822 * inactive work item is found or max_active limit is reached.
1824 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1826 struct work_struct *work =
1827 list_first_entry_or_null(&pwq->inactive_works,
1828 struct work_struct, entry);
1830 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1831 __pwq_activate_work(pwq, work);
1839 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1840 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1842 * This function should only be called for ordered workqueues where only the
1843 * oldest pwq is unplugged, the others are plugged to suspend execution to
1844 * ensure proper work item ordering::
1846 * dfl_pwq --------------+ [P] - plugged
1849 * pwqs -> A -> B [P] -> C [P] (newest)
1855 * When the oldest pwq is drained and removed, this function should be called
1856 * to unplug the next oldest one to start its work item execution. Note that
1857 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1858 * the list is the oldest.
1860 static void unplug_oldest_pwq(struct workqueue_struct *wq)
1862 struct pool_workqueue *pwq;
1864 lockdep_assert_held(&wq->mutex);
1866 /* Caller should make sure that pwqs isn't empty before calling */
1867 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1869 raw_spin_lock_irq(&pwq->pool->lock);
1871 pwq->plugged = false;
1872 if (pwq_activate_first_inactive(pwq, true))
1873 kick_pool(pwq->pool);
1875 raw_spin_unlock_irq(&pwq->pool->lock);
1879 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1880 * @nna: wq_node_nr_active to activate a pending pwq for
1881 * @caller_pool: worker_pool the caller is locking
1883 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1884 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1886 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1887 struct worker_pool *caller_pool)
1889 struct worker_pool *locked_pool = caller_pool;
1890 struct pool_workqueue *pwq;
1891 struct work_struct *work;
1893 lockdep_assert_held(&caller_pool->lock);
1895 raw_spin_lock(&nna->lock);
1897 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1898 struct pool_workqueue, pending_node);
1903 * If @pwq is for a different pool than @locked_pool, we need to lock
1904 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1905 * / lock dance. For that, we also need to release @nna->lock as it's
1906 * nested inside pool locks.
1908 if (pwq->pool != locked_pool) {
1909 raw_spin_unlock(&locked_pool->lock);
1910 locked_pool = pwq->pool;
1911 if (!raw_spin_trylock(&locked_pool->lock)) {
1912 raw_spin_unlock(&nna->lock);
1913 raw_spin_lock(&locked_pool->lock);
1914 raw_spin_lock(&nna->lock);
1920 * $pwq may not have any inactive work items due to e.g. cancellations.
1921 * Drop it from pending_pwqs and see if there's another one.
1923 work = list_first_entry_or_null(&pwq->inactive_works,
1924 struct work_struct, entry);
1926 list_del_init(&pwq->pending_node);
1931 * Acquire an nr_active count and activate the inactive work item. If
1932 * $pwq still has inactive work items, rotate it to the end of the
1933 * pending_pwqs so that we round-robin through them. This means that
1934 * inactive work items are not activated in queueing order which is fine
1935 * given that there has never been any ordering across different pwqs.
1937 if (likely(tryinc_node_nr_active(nna))) {
1939 __pwq_activate_work(pwq, work);
1941 if (list_empty(&pwq->inactive_works))
1942 list_del_init(&pwq->pending_node);
1944 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1946 /* if activating a foreign pool, make sure it's running */
1947 if (pwq->pool != caller_pool)
1948 kick_pool(pwq->pool);
1952 raw_spin_unlock(&nna->lock);
1953 if (locked_pool != caller_pool) {
1954 raw_spin_unlock(&locked_pool->lock);
1955 raw_spin_lock(&caller_pool->lock);
1960 * pwq_dec_nr_active - Retire an active count
1961 * @pwq: pool_workqueue of interest
1963 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1964 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1966 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1968 struct worker_pool *pool = pwq->pool;
1969 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1971 lockdep_assert_held(&pool->lock);
1974 * @pwq->nr_active should be decremented for both percpu and unbound
1980 * For a percpu workqueue, it's simple. Just need to kick the first
1981 * inactive work item on @pwq itself.
1984 pwq_activate_first_inactive(pwq, false);
1989 * If @pwq is for an unbound workqueue, it's more complicated because
1990 * multiple pwqs and pools may be sharing the nr_active count. When a
1991 * pwq needs to wait for an nr_active count, it puts itself on
1992 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1993 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1994 * guarantee that either we see non-empty pending_pwqs or they see
1995 * decremented $nna->nr.
1997 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1998 * max_active gets updated. However, it is guaranteed to be equal to or
1999 * larger than @pwq->wq->min_active which is above zero unless freezing.
2000 * This maintains the forward progress guarantee.
2002 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
2005 if (!list_empty(&nna->pending_pwqs))
2006 node_activate_pending_pwq(nna, pool);
2010 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
2011 * @pwq: pwq of interest
2012 * @work_data: work_data of work which left the queue
2014 * A work either has completed or is removed from pending queue,
2015 * decrement nr_in_flight of its pwq and handle workqueue flushing.
2018 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
2019 * and thus should be called after all other state updates for the in-flight
2020 * work item is complete.
2023 * raw_spin_lock_irq(pool->lock).
2025 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
2027 int color = get_work_color(work_data);
2029 if (!(work_data & WORK_STRUCT_INACTIVE))
2030 pwq_dec_nr_active(pwq);
2032 pwq->nr_in_flight[color]--;
2034 /* is flush in progress and are we at the flushing tip? */
2035 if (likely(pwq->flush_color != color))
2038 /* are there still in-flight works? */
2039 if (pwq->nr_in_flight[color])
2042 /* this pwq is done, clear flush_color */
2043 pwq->flush_color = -1;
2046 * If this was the last pwq, wake up the first flusher. It
2047 * will handle the rest.
2049 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2050 complete(&pwq->wq->first_flusher->done);
2056 * try_to_grab_pending - steal work item from worklist and disable irq
2057 * @work: work item to steal
2058 * @cflags: %WORK_CANCEL_ flags
2059 * @irq_flags: place to store irq state
2061 * Try to grab PENDING bit of @work. This function can handle @work in any
2062 * stable state - idle, on timer or on worklist.
2066 * ======== ================================================================
2067 * 1 if @work was pending and we successfully stole PENDING
2068 * 0 if @work was idle and we claimed PENDING
2069 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2070 * -ENOENT if someone else is canceling @work, this state may persist
2071 * for arbitrarily long
2072 * ======== ================================================================
2075 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2076 * interrupted while holding PENDING and @work off queue, irq must be
2077 * disabled on entry. This, combined with delayed_work->timer being
2078 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2080 * On successful return, >= 0, irq is disabled and the caller is
2081 * responsible for releasing it using local_irq_restore(*@irq_flags).
2083 * This function is safe to call from any context including IRQ handler.
2085 static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2086 unsigned long *irq_flags)
2088 struct worker_pool *pool;
2089 struct pool_workqueue *pwq;
2091 local_irq_save(*irq_flags);
2093 /* try to steal the timer if it exists */
2094 if (cflags & WORK_CANCEL_DELAYED) {
2095 struct delayed_work *dwork = to_delayed_work(work);
2098 * dwork->timer is irqsafe. If del_timer() fails, it's
2099 * guaranteed that the timer is not queued anywhere and not
2100 * running on the local CPU.
2102 if (likely(del_timer(&dwork->timer)))
2106 /* try to claim PENDING the normal way */
2107 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2112 * The queueing is in progress, or it is already queued. Try to
2113 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2115 pool = get_work_pool(work);
2119 raw_spin_lock(&pool->lock);
2121 * work->data is guaranteed to point to pwq only while the work
2122 * item is queued on pwq->wq, and both updating work->data to point
2123 * to pwq on queueing and to pool on dequeueing are done under
2124 * pwq->pool->lock. This in turn guarantees that, if work->data
2125 * points to pwq which is associated with a locked pool, the work
2126 * item is currently queued on that pool.
2128 pwq = get_work_pwq(work);
2129 if (pwq && pwq->pool == pool) {
2130 unsigned long work_data;
2132 debug_work_deactivate(work);
2135 * A cancelable inactive work item must be in the
2136 * pwq->inactive_works since a queued barrier can't be
2137 * canceled (see the comments in insert_wq_barrier()).
2139 * An inactive work item cannot be grabbed directly because
2140 * it might have linked barrier work items which, if left
2141 * on the inactive_works list, will confuse pwq->nr_active
2142 * management later on and cause stall. Make sure the work
2143 * item is activated before grabbing.
2145 pwq_activate_work(pwq, work);
2147 list_del_init(&work->entry);
2150 * work->data points to pwq iff queued. Let's point to pool. As
2151 * this destroys work->data needed by the next step, stash it.
2153 work_data = *work_data_bits(work);
2154 set_work_pool_and_keep_pending(work, pool->id, 0);
2156 /* must be the last step, see the function comment */
2157 pwq_dec_nr_in_flight(pwq, work_data);
2159 raw_spin_unlock(&pool->lock);
2163 raw_spin_unlock(&pool->lock);
2166 local_irq_restore(*irq_flags);
2167 if (work_is_canceling(work))
2174 wait_queue_entry_t wait;
2175 struct work_struct *work;
2178 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2180 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2182 if (cwait->work != key)
2184 return autoremove_wake_function(wait, mode, sync, key);
2188 * work_grab_pending - steal work item from worklist and disable irq
2189 * @work: work item to steal
2190 * @cflags: %WORK_CANCEL_ flags
2191 * @irq_flags: place to store IRQ state
2193 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2196 * Must be called in process context. IRQ is disabled on return with IRQ state
2197 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2198 * local_irq_restore().
2200 * Returns %true if @work was pending. %false if idle.
2202 static bool work_grab_pending(struct work_struct *work, u32 cflags,
2203 unsigned long *irq_flags)
2205 struct cwt_wait cwait;
2210 ret = try_to_grab_pending(work, cflags, irq_flags);
2211 if (likely(ret >= 0))
2217 * Someone is already canceling. Wait for it to finish. flush_work()
2218 * doesn't work for PREEMPT_NONE because we may get woken up between
2219 * @work's completion and the other canceling task resuming and clearing
2220 * CANCELING - flush_work() will return false immediately as @work is no
2221 * longer busy, try_to_grab_pending() will return -ENOENT as @work is
2222 * still being canceled and the other canceling task won't be able to
2223 * clear CANCELING as we're hogging the CPU.
2225 * Let's wait for completion using a waitqueue. As this may lead to the
2226 * thundering herd problem, use a custom wake function which matches
2227 * @work along with exclusive wait and wakeup.
2229 init_wait(&cwait.wait);
2230 cwait.wait.func = cwt_wakefn;
2233 prepare_to_wait_exclusive(&wq_cancel_waitq, &cwait.wait,
2234 TASK_UNINTERRUPTIBLE);
2235 if (work_is_canceling(work))
2237 finish_wait(&wq_cancel_waitq, &cwait.wait);
2243 * insert_work - insert a work into a pool
2244 * @pwq: pwq @work belongs to
2245 * @work: work to insert
2246 * @head: insertion point
2247 * @extra_flags: extra WORK_STRUCT_* flags to set
2249 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2250 * work_struct flags.
2253 * raw_spin_lock_irq(pool->lock).
2255 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2256 struct list_head *head, unsigned int extra_flags)
2258 debug_work_activate(work);
2260 /* record the work call stack in order to print it in KASAN reports */
2261 kasan_record_aux_stack_noalloc(work);
2263 /* we own @work, set data and link */
2264 set_work_pwq(work, pwq, extra_flags);
2265 list_add_tail(&work->entry, head);
2270 * Test whether @work is being queued from another work executing on the
2273 static bool is_chained_work(struct workqueue_struct *wq)
2275 struct worker *worker;
2277 worker = current_wq_worker();
2279 * Return %true iff I'm a worker executing a work item on @wq. If
2280 * I'm @worker, it's safe to dereference it without locking.
2282 return worker && worker->current_pwq->wq == wq;
2286 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2287 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2288 * avoid perturbing sensitive tasks.
2290 static int wq_select_unbound_cpu(int cpu)
2294 if (likely(!wq_debug_force_rr_cpu)) {
2295 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2298 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2301 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2302 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2303 if (unlikely(new_cpu >= nr_cpu_ids)) {
2304 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2305 if (unlikely(new_cpu >= nr_cpu_ids))
2308 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2313 static void __queue_work(int cpu, struct workqueue_struct *wq,
2314 struct work_struct *work)
2316 struct pool_workqueue *pwq;
2317 struct worker_pool *last_pool, *pool;
2318 unsigned int work_flags;
2319 unsigned int req_cpu = cpu;
2322 * While a work item is PENDING && off queue, a task trying to
2323 * steal the PENDING will busy-loop waiting for it to either get
2324 * queued or lose PENDING. Grabbing PENDING and queueing should
2325 * happen with IRQ disabled.
2327 lockdep_assert_irqs_disabled();
2330 * For a draining wq, only works from the same workqueue are
2331 * allowed. The __WQ_DESTROYING helps to spot the issue that
2332 * queues a new work item to a wq after destroy_workqueue(wq).
2334 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2335 WARN_ON_ONCE(!is_chained_work(wq))))
2339 /* pwq which will be used unless @work is executing elsewhere */
2340 if (req_cpu == WORK_CPU_UNBOUND) {
2341 if (wq->flags & WQ_UNBOUND)
2342 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2344 cpu = raw_smp_processor_id();
2347 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2351 * If @work was previously on a different pool, it might still be
2352 * running there, in which case the work needs to be queued on that
2353 * pool to guarantee non-reentrancy.
2355 last_pool = get_work_pool(work);
2356 if (last_pool && last_pool != pool) {
2357 struct worker *worker;
2359 raw_spin_lock(&last_pool->lock);
2361 worker = find_worker_executing_work(last_pool, work);
2363 if (worker && worker->current_pwq->wq == wq) {
2364 pwq = worker->current_pwq;
2366 WARN_ON_ONCE(pool != last_pool);
2368 /* meh... not running there, queue here */
2369 raw_spin_unlock(&last_pool->lock);
2370 raw_spin_lock(&pool->lock);
2373 raw_spin_lock(&pool->lock);
2377 * pwq is determined and locked. For unbound pools, we could have raced
2378 * with pwq release and it could already be dead. If its refcnt is zero,
2379 * repeat pwq selection. Note that unbound pwqs never die without
2380 * another pwq replacing it in cpu_pwq or while work items are executing
2381 * on it, so the retrying is guaranteed to make forward-progress.
2383 if (unlikely(!pwq->refcnt)) {
2384 if (wq->flags & WQ_UNBOUND) {
2385 raw_spin_unlock(&pool->lock);
2390 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2394 /* pwq determined, queue */
2395 trace_workqueue_queue_work(req_cpu, pwq, work);
2397 if (WARN_ON(!list_empty(&work->entry)))
2400 pwq->nr_in_flight[pwq->work_color]++;
2401 work_flags = work_color_to_flags(pwq->work_color);
2404 * Limit the number of concurrently active work items to max_active.
2405 * @work must also queue behind existing inactive work items to maintain
2406 * ordering when max_active changes. See wq_adjust_max_active().
2408 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2409 if (list_empty(&pool->worklist))
2410 pool->watchdog_ts = jiffies;
2412 trace_workqueue_activate_work(work);
2413 insert_work(pwq, work, &pool->worklist, work_flags);
2416 work_flags |= WORK_STRUCT_INACTIVE;
2417 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2421 raw_spin_unlock(&pool->lock);
2426 * queue_work_on - queue work on specific cpu
2427 * @cpu: CPU number to execute work on
2428 * @wq: workqueue to use
2429 * @work: work to queue
2431 * We queue the work to a specific CPU, the caller must ensure it
2432 * can't go away. Callers that fail to ensure that the specified
2433 * CPU cannot go away will execute on a randomly chosen CPU.
2434 * But note well that callers specifying a CPU that never has been
2435 * online will get a splat.
2437 * Return: %false if @work was already on a queue, %true otherwise.
2439 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2440 struct work_struct *work)
2443 unsigned long irq_flags;
2445 local_irq_save(irq_flags);
2447 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2448 __queue_work(cpu, wq, work);
2452 local_irq_restore(irq_flags);
2455 EXPORT_SYMBOL(queue_work_on);
2458 * select_numa_node_cpu - Select a CPU based on NUMA node
2459 * @node: NUMA node ID that we want to select a CPU from
2461 * This function will attempt to find a "random" cpu available on a given
2462 * node. If there are no CPUs available on the given node it will return
2463 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2464 * available CPU if we need to schedule this work.
2466 static int select_numa_node_cpu(int node)
2470 /* Delay binding to CPU if node is not valid or online */
2471 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2472 return WORK_CPU_UNBOUND;
2474 /* Use local node/cpu if we are already there */
2475 cpu = raw_smp_processor_id();
2476 if (node == cpu_to_node(cpu))
2479 /* Use "random" otherwise know as "first" online CPU of node */
2480 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2482 /* If CPU is valid return that, otherwise just defer */
2483 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2487 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2488 * @node: NUMA node that we are targeting the work for
2489 * @wq: workqueue to use
2490 * @work: work to queue
2492 * We queue the work to a "random" CPU within a given NUMA node. The basic
2493 * idea here is to provide a way to somehow associate work with a given
2496 * This function will only make a best effort attempt at getting this onto
2497 * the right NUMA node. If no node is requested or the requested node is
2498 * offline then we just fall back to standard queue_work behavior.
2500 * Currently the "random" CPU ends up being the first available CPU in the
2501 * intersection of cpu_online_mask and the cpumask of the node, unless we
2502 * are running on the node. In that case we just use the current CPU.
2504 * Return: %false if @work was already on a queue, %true otherwise.
2506 bool queue_work_node(int node, struct workqueue_struct *wq,
2507 struct work_struct *work)
2509 unsigned long irq_flags;
2513 * This current implementation is specific to unbound workqueues.
2514 * Specifically we only return the first available CPU for a given
2515 * node instead of cycling through individual CPUs within the node.
2517 * If this is used with a per-cpu workqueue then the logic in
2518 * workqueue_select_cpu_near would need to be updated to allow for
2519 * some round robin type logic.
2521 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2523 local_irq_save(irq_flags);
2525 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2526 int cpu = select_numa_node_cpu(node);
2528 __queue_work(cpu, wq, work);
2532 local_irq_restore(irq_flags);
2535 EXPORT_SYMBOL_GPL(queue_work_node);
2537 void delayed_work_timer_fn(struct timer_list *t)
2539 struct delayed_work *dwork = from_timer(dwork, t, timer);
2541 /* should have been called from irqsafe timer with irq already off */
2542 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2544 EXPORT_SYMBOL(delayed_work_timer_fn);
2546 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2547 struct delayed_work *dwork, unsigned long delay)
2549 struct timer_list *timer = &dwork->timer;
2550 struct work_struct *work = &dwork->work;
2553 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2554 WARN_ON_ONCE(timer_pending(timer));
2555 WARN_ON_ONCE(!list_empty(&work->entry));
2558 * If @delay is 0, queue @dwork->work immediately. This is for
2559 * both optimization and correctness. The earliest @timer can
2560 * expire is on the closest next tick and delayed_work users depend
2561 * on that there's no such delay when @delay is 0.
2564 __queue_work(cpu, wq, &dwork->work);
2570 timer->expires = jiffies + delay;
2572 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2573 /* If the current cpu is a housekeeping cpu, use it. */
2574 cpu = smp_processor_id();
2575 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2576 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2577 add_timer_on(timer, cpu);
2579 if (likely(cpu == WORK_CPU_UNBOUND))
2580 add_timer_global(timer);
2582 add_timer_on(timer, cpu);
2587 * queue_delayed_work_on - queue work on specific CPU after delay
2588 * @cpu: CPU number to execute work on
2589 * @wq: workqueue to use
2590 * @dwork: work to queue
2591 * @delay: number of jiffies to wait before queueing
2593 * Return: %false if @work was already on a queue, %true otherwise. If
2594 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2597 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2598 struct delayed_work *dwork, unsigned long delay)
2600 struct work_struct *work = &dwork->work;
2602 unsigned long irq_flags;
2604 /* read the comment in __queue_work() */
2605 local_irq_save(irq_flags);
2607 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2608 __queue_delayed_work(cpu, wq, dwork, delay);
2612 local_irq_restore(irq_flags);
2615 EXPORT_SYMBOL(queue_delayed_work_on);
2618 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2619 * @cpu: CPU number to execute work on
2620 * @wq: workqueue to use
2621 * @dwork: work to queue
2622 * @delay: number of jiffies to wait before queueing
2624 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2625 * modify @dwork's timer so that it expires after @delay. If @delay is
2626 * zero, @work is guaranteed to be scheduled immediately regardless of its
2629 * Return: %false if @dwork was idle and queued, %true if @dwork was
2630 * pending and its timer was modified.
2632 * This function is safe to call from any context including IRQ handler.
2633 * See try_to_grab_pending() for details.
2635 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2636 struct delayed_work *dwork, unsigned long delay)
2638 unsigned long irq_flags;
2642 ret = try_to_grab_pending(&dwork->work, WORK_CANCEL_DELAYED,
2644 } while (unlikely(ret == -EAGAIN));
2646 if (likely(ret >= 0)) {
2647 __queue_delayed_work(cpu, wq, dwork, delay);
2648 local_irq_restore(irq_flags);
2651 /* -ENOENT from try_to_grab_pending() becomes %true */
2654 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2656 static void rcu_work_rcufn(struct rcu_head *rcu)
2658 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2660 /* read the comment in __queue_work() */
2661 local_irq_disable();
2662 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2667 * queue_rcu_work - queue work after a RCU grace period
2668 * @wq: workqueue to use
2669 * @rwork: work to queue
2671 * Return: %false if @rwork was already pending, %true otherwise. Note
2672 * that a full RCU grace period is guaranteed only after a %true return.
2673 * While @rwork is guaranteed to be executed after a %false return, the
2674 * execution may happen before a full RCU grace period has passed.
2676 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2678 struct work_struct *work = &rwork->work;
2680 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2682 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2688 EXPORT_SYMBOL(queue_rcu_work);
2690 static struct worker *alloc_worker(int node)
2692 struct worker *worker;
2694 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2696 INIT_LIST_HEAD(&worker->entry);
2697 INIT_LIST_HEAD(&worker->scheduled);
2698 INIT_LIST_HEAD(&worker->node);
2699 /* on creation a worker is in !idle && prep state */
2700 worker->flags = WORKER_PREP;
2705 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2707 if (pool->cpu < 0 && pool->attrs->affn_strict)
2708 return pool->attrs->__pod_cpumask;
2710 return pool->attrs->cpumask;
2714 * worker_attach_to_pool() - attach a worker to a pool
2715 * @worker: worker to be attached
2716 * @pool: the target pool
2718 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2719 * cpu-binding of @worker are kept coordinated with the pool across
2722 static void worker_attach_to_pool(struct worker *worker,
2723 struct worker_pool *pool)
2725 mutex_lock(&wq_pool_attach_mutex);
2728 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2729 * across this function. See the comments above the flag definition for
2730 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2732 if (pool->flags & POOL_DISASSOCIATED) {
2733 worker->flags |= WORKER_UNBOUND;
2735 WARN_ON_ONCE(pool->flags & POOL_BH);
2736 kthread_set_per_cpu(worker->task, pool->cpu);
2739 if (worker->rescue_wq)
2740 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2742 list_add_tail(&worker->node, &pool->workers);
2743 worker->pool = pool;
2745 mutex_unlock(&wq_pool_attach_mutex);
2749 * worker_detach_from_pool() - detach a worker from its pool
2750 * @worker: worker which is attached to its pool
2752 * Undo the attaching which had been done in worker_attach_to_pool(). The
2753 * caller worker shouldn't access to the pool after detached except it has
2754 * other reference to the pool.
2756 static void worker_detach_from_pool(struct worker *worker)
2758 struct worker_pool *pool = worker->pool;
2759 struct completion *detach_completion = NULL;
2761 /* there is one permanent BH worker per CPU which should never detach */
2762 WARN_ON_ONCE(pool->flags & POOL_BH);
2764 mutex_lock(&wq_pool_attach_mutex);
2766 kthread_set_per_cpu(worker->task, -1);
2767 list_del(&worker->node);
2768 worker->pool = NULL;
2770 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2771 detach_completion = pool->detach_completion;
2772 mutex_unlock(&wq_pool_attach_mutex);
2774 /* clear leftover flags without pool->lock after it is detached */
2775 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2777 if (detach_completion)
2778 complete(detach_completion);
2782 * create_worker - create a new workqueue worker
2783 * @pool: pool the new worker will belong to
2785 * Create and start a new worker which is attached to @pool.
2788 * Might sleep. Does GFP_KERNEL allocations.
2791 * Pointer to the newly created worker.
2793 static struct worker *create_worker(struct worker_pool *pool)
2795 struct worker *worker;
2799 /* ID is needed to determine kthread name */
2800 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2802 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2807 worker = alloc_worker(pool->node);
2809 pr_err_once("workqueue: Failed to allocate a worker\n");
2815 if (!(pool->flags & POOL_BH)) {
2817 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2818 pool->attrs->nice < 0 ? "H" : "");
2820 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2822 worker->task = kthread_create_on_node(worker_thread, worker,
2823 pool->node, "kworker/%s", id_buf);
2824 if (IS_ERR(worker->task)) {
2825 if (PTR_ERR(worker->task) == -EINTR) {
2826 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2829 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2835 set_user_nice(worker->task, pool->attrs->nice);
2836 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2839 /* successful, attach the worker to the pool */
2840 worker_attach_to_pool(worker, pool);
2842 /* start the newly created worker */
2843 raw_spin_lock_irq(&pool->lock);
2845 worker->pool->nr_workers++;
2846 worker_enter_idle(worker);
2849 * @worker is waiting on a completion in kthread() and will trigger hung
2850 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2851 * wake it up explicitly.
2854 wake_up_process(worker->task);
2856 raw_spin_unlock_irq(&pool->lock);
2861 ida_free(&pool->worker_ida, id);
2866 static void unbind_worker(struct worker *worker)
2868 lockdep_assert_held(&wq_pool_attach_mutex);
2870 kthread_set_per_cpu(worker->task, -1);
2871 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2872 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2874 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2877 static void wake_dying_workers(struct list_head *cull_list)
2879 struct worker *worker, *tmp;
2881 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2882 list_del_init(&worker->entry);
2883 unbind_worker(worker);
2885 * If the worker was somehow already running, then it had to be
2886 * in pool->idle_list when set_worker_dying() happened or we
2887 * wouldn't have gotten here.
2889 * Thus, the worker must either have observed the WORKER_DIE
2890 * flag, or have set its state to TASK_IDLE. Either way, the
2891 * below will be observed by the worker and is safe to do
2892 * outside of pool->lock.
2894 wake_up_process(worker->task);
2899 * set_worker_dying - Tag a worker for destruction
2900 * @worker: worker to be destroyed
2901 * @list: transfer worker away from its pool->idle_list and into list
2903 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2907 * raw_spin_lock_irq(pool->lock).
2909 static void set_worker_dying(struct worker *worker, struct list_head *list)
2911 struct worker_pool *pool = worker->pool;
2913 lockdep_assert_held(&pool->lock);
2914 lockdep_assert_held(&wq_pool_attach_mutex);
2916 /* sanity check frenzy */
2917 if (WARN_ON(worker->current_work) ||
2918 WARN_ON(!list_empty(&worker->scheduled)) ||
2919 WARN_ON(!(worker->flags & WORKER_IDLE)))
2925 worker->flags |= WORKER_DIE;
2927 list_move(&worker->entry, list);
2928 list_move(&worker->node, &pool->dying_workers);
2932 * idle_worker_timeout - check if some idle workers can now be deleted.
2933 * @t: The pool's idle_timer that just expired
2935 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2936 * worker_leave_idle(), as a worker flicking between idle and active while its
2937 * pool is at the too_many_workers() tipping point would cause too much timer
2938 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2939 * it expire and re-evaluate things from there.
2941 static void idle_worker_timeout(struct timer_list *t)
2943 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2944 bool do_cull = false;
2946 if (work_pending(&pool->idle_cull_work))
2949 raw_spin_lock_irq(&pool->lock);
2951 if (too_many_workers(pool)) {
2952 struct worker *worker;
2953 unsigned long expires;
2955 /* idle_list is kept in LIFO order, check the last one */
2956 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2957 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2958 do_cull = !time_before(jiffies, expires);
2961 mod_timer(&pool->idle_timer, expires);
2963 raw_spin_unlock_irq(&pool->lock);
2966 queue_work(system_unbound_wq, &pool->idle_cull_work);
2970 * idle_cull_fn - cull workers that have been idle for too long.
2971 * @work: the pool's work for handling these idle workers
2973 * This goes through a pool's idle workers and gets rid of those that have been
2974 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2976 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2977 * culled, so this also resets worker affinity. This requires a sleepable
2978 * context, hence the split between timer callback and work item.
2980 static void idle_cull_fn(struct work_struct *work)
2982 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2983 LIST_HEAD(cull_list);
2986 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2987 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2988 * path. This is required as a previously-preempted worker could run after
2989 * set_worker_dying() has happened but before wake_dying_workers() did.
2991 mutex_lock(&wq_pool_attach_mutex);
2992 raw_spin_lock_irq(&pool->lock);
2994 while (too_many_workers(pool)) {
2995 struct worker *worker;
2996 unsigned long expires;
2998 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2999 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
3001 if (time_before(jiffies, expires)) {
3002 mod_timer(&pool->idle_timer, expires);
3006 set_worker_dying(worker, &cull_list);
3009 raw_spin_unlock_irq(&pool->lock);
3010 wake_dying_workers(&cull_list);
3011 mutex_unlock(&wq_pool_attach_mutex);
3014 static void send_mayday(struct work_struct *work)
3016 struct pool_workqueue *pwq = get_work_pwq(work);
3017 struct workqueue_struct *wq = pwq->wq;
3019 lockdep_assert_held(&wq_mayday_lock);
3024 /* mayday mayday mayday */
3025 if (list_empty(&pwq->mayday_node)) {
3027 * If @pwq is for an unbound wq, its base ref may be put at
3028 * any time due to an attribute change. Pin @pwq until the
3029 * rescuer is done with it.
3032 list_add_tail(&pwq->mayday_node, &wq->maydays);
3033 wake_up_process(wq->rescuer->task);
3034 pwq->stats[PWQ_STAT_MAYDAY]++;
3038 static void pool_mayday_timeout(struct timer_list *t)
3040 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3041 struct work_struct *work;
3043 raw_spin_lock_irq(&pool->lock);
3044 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3046 if (need_to_create_worker(pool)) {
3048 * We've been trying to create a new worker but
3049 * haven't been successful. We might be hitting an
3050 * allocation deadlock. Send distress signals to
3053 list_for_each_entry(work, &pool->worklist, entry)
3057 raw_spin_unlock(&wq_mayday_lock);
3058 raw_spin_unlock_irq(&pool->lock);
3060 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3064 * maybe_create_worker - create a new worker if necessary
3065 * @pool: pool to create a new worker for
3067 * Create a new worker for @pool if necessary. @pool is guaranteed to
3068 * have at least one idle worker on return from this function. If
3069 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3070 * sent to all rescuers with works scheduled on @pool to resolve
3071 * possible allocation deadlock.
3073 * On return, need_to_create_worker() is guaranteed to be %false and
3074 * may_start_working() %true.
3077 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3078 * multiple times. Does GFP_KERNEL allocations. Called only from
3081 static void maybe_create_worker(struct worker_pool *pool)
3082 __releases(&pool->lock)
3083 __acquires(&pool->lock)
3086 raw_spin_unlock_irq(&pool->lock);
3088 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3089 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3092 if (create_worker(pool) || !need_to_create_worker(pool))
3095 schedule_timeout_interruptible(CREATE_COOLDOWN);
3097 if (!need_to_create_worker(pool))
3101 del_timer_sync(&pool->mayday_timer);
3102 raw_spin_lock_irq(&pool->lock);
3104 * This is necessary even after a new worker was just successfully
3105 * created as @pool->lock was dropped and the new worker might have
3106 * already become busy.
3108 if (need_to_create_worker(pool))
3113 * manage_workers - manage worker pool
3116 * Assume the manager role and manage the worker pool @worker belongs
3117 * to. At any given time, there can be only zero or one manager per
3118 * pool. The exclusion is handled automatically by this function.
3120 * The caller can safely start processing works on false return. On
3121 * true return, it's guaranteed that need_to_create_worker() is false
3122 * and may_start_working() is true.
3125 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3126 * multiple times. Does GFP_KERNEL allocations.
3129 * %false if the pool doesn't need management and the caller can safely
3130 * start processing works, %true if management function was performed and
3131 * the conditions that the caller verified before calling the function may
3132 * no longer be true.
3134 static bool manage_workers(struct worker *worker)
3136 struct worker_pool *pool = worker->pool;
3138 if (pool->flags & POOL_MANAGER_ACTIVE)
3141 pool->flags |= POOL_MANAGER_ACTIVE;
3142 pool->manager = worker;
3144 maybe_create_worker(pool);
3146 pool->manager = NULL;
3147 pool->flags &= ~POOL_MANAGER_ACTIVE;
3148 rcuwait_wake_up(&manager_wait);
3153 * process_one_work - process single work
3155 * @work: work to process
3157 * Process @work. This function contains all the logics necessary to
3158 * process a single work including synchronization against and
3159 * interaction with other workers on the same cpu, queueing and
3160 * flushing. As long as context requirement is met, any worker can
3161 * call this function to process a work.
3164 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3166 static void process_one_work(struct worker *worker, struct work_struct *work)
3167 __releases(&pool->lock)
3168 __acquires(&pool->lock)
3170 struct pool_workqueue *pwq = get_work_pwq(work);
3171 struct worker_pool *pool = worker->pool;
3172 unsigned long work_data;
3173 int lockdep_start_depth, rcu_start_depth;
3174 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3175 #ifdef CONFIG_LOCKDEP
3177 * It is permissible to free the struct work_struct from
3178 * inside the function that is called from it, this we need to
3179 * take into account for lockdep too. To avoid bogus "held
3180 * lock freed" warnings as well as problems when looking into
3181 * work->lockdep_map, make a copy and use that here.
3183 struct lockdep_map lockdep_map;
3185 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3187 /* ensure we're on the correct CPU */
3188 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3189 raw_smp_processor_id() != pool->cpu);
3191 /* claim and dequeue */
3192 debug_work_deactivate(work);
3193 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3194 worker->current_work = work;
3195 worker->current_func = work->func;
3196 worker->current_pwq = pwq;
3198 worker->current_at = worker->task->se.sum_exec_runtime;
3199 work_data = *work_data_bits(work);
3200 worker->current_color = get_work_color(work_data);
3203 * Record wq name for cmdline and debug reporting, may get
3204 * overridden through set_worker_desc().
3206 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3208 list_del_init(&work->entry);
3211 * CPU intensive works don't participate in concurrency management.
3212 * They're the scheduler's responsibility. This takes @worker out
3213 * of concurrency management and the next code block will chain
3214 * execution of the pending work items.
3216 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3217 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3220 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3221 * since nr_running would always be >= 1 at this point. This is used to
3222 * chain execution of the pending work items for WORKER_NOT_RUNNING
3223 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3228 * Record the last pool and clear PENDING which should be the last
3229 * update to @work. Also, do this inside @pool->lock so that
3230 * PENDING and queued state changes happen together while IRQ is
3233 set_work_pool_and_clear_pending(work, pool->id, 0);
3235 pwq->stats[PWQ_STAT_STARTED]++;
3236 raw_spin_unlock_irq(&pool->lock);
3238 rcu_start_depth = rcu_preempt_depth();
3239 lockdep_start_depth = lockdep_depth(current);
3240 /* see drain_dead_softirq_workfn() */
3242 lock_map_acquire(&pwq->wq->lockdep_map);
3243 lock_map_acquire(&lockdep_map);
3245 * Strictly speaking we should mark the invariant state without holding
3246 * any locks, that is, before these two lock_map_acquire()'s.
3248 * However, that would result in:
3255 * Which would create W1->C->W1 dependencies, even though there is no
3256 * actual deadlock possible. There are two solutions, using a
3257 * read-recursive acquire on the work(queue) 'locks', but this will then
3258 * hit the lockdep limitation on recursive locks, or simply discard
3261 * AFAICT there is no possible deadlock scenario between the
3262 * flush_work() and complete() primitives (except for single-threaded
3263 * workqueues), so hiding them isn't a problem.
3265 lockdep_invariant_state(true);
3266 trace_workqueue_execute_start(work);
3267 worker->current_func(work);
3269 * While we must be careful to not use "work" after this, the trace
3270 * point will only record its address.
3272 trace_workqueue_execute_end(work, worker->current_func);
3273 pwq->stats[PWQ_STAT_COMPLETED]++;
3274 lock_map_release(&lockdep_map);
3276 lock_map_release(&pwq->wq->lockdep_map);
3278 if (unlikely((worker->task && in_atomic()) ||
3279 lockdep_depth(current) != lockdep_start_depth ||
3280 rcu_preempt_depth() != rcu_start_depth)) {
3281 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3282 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3283 current->comm, task_pid_nr(current), preempt_count(),
3284 lockdep_start_depth, lockdep_depth(current),
3285 rcu_start_depth, rcu_preempt_depth(),
3286 worker->current_func);
3287 debug_show_held_locks(current);
3292 * The following prevents a kworker from hogging CPU on !PREEMPTION
3293 * kernels, where a requeueing work item waiting for something to
3294 * happen could deadlock with stop_machine as such work item could
3295 * indefinitely requeue itself while all other CPUs are trapped in
3296 * stop_machine. At the same time, report a quiescent RCU state so
3297 * the same condition doesn't freeze RCU.
3302 raw_spin_lock_irq(&pool->lock);
3305 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3306 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3307 * wq_cpu_intensive_thresh_us. Clear it.
3309 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3311 /* tag the worker for identification in schedule() */
3312 worker->last_func = worker->current_func;
3314 /* we're done with it, release */
3315 hash_del(&worker->hentry);
3316 worker->current_work = NULL;
3317 worker->current_func = NULL;
3318 worker->current_pwq = NULL;
3319 worker->current_color = INT_MAX;
3321 /* must be the last step, see the function comment */
3322 pwq_dec_nr_in_flight(pwq, work_data);
3326 * process_scheduled_works - process scheduled works
3329 * Process all scheduled works. Please note that the scheduled list
3330 * may change while processing a work, so this function repeatedly
3331 * fetches a work from the top and executes it.
3334 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3337 static void process_scheduled_works(struct worker *worker)
3339 struct work_struct *work;
3342 while ((work = list_first_entry_or_null(&worker->scheduled,
3343 struct work_struct, entry))) {
3345 worker->pool->watchdog_ts = jiffies;
3348 process_one_work(worker, work);
3352 static void set_pf_worker(bool val)
3354 mutex_lock(&wq_pool_attach_mutex);
3356 current->flags |= PF_WQ_WORKER;
3358 current->flags &= ~PF_WQ_WORKER;
3359 mutex_unlock(&wq_pool_attach_mutex);
3363 * worker_thread - the worker thread function
3366 * The worker thread function. All workers belong to a worker_pool -
3367 * either a per-cpu one or dynamic unbound one. These workers process all
3368 * work items regardless of their specific target workqueue. The only
3369 * exception is work items which belong to workqueues with a rescuer which
3370 * will be explained in rescuer_thread().
3374 static int worker_thread(void *__worker)
3376 struct worker *worker = __worker;
3377 struct worker_pool *pool = worker->pool;
3379 /* tell the scheduler that this is a workqueue worker */
3380 set_pf_worker(true);
3382 raw_spin_lock_irq(&pool->lock);
3384 /* am I supposed to die? */
3385 if (unlikely(worker->flags & WORKER_DIE)) {
3386 raw_spin_unlock_irq(&pool->lock);
3387 set_pf_worker(false);
3389 set_task_comm(worker->task, "kworker/dying");
3390 ida_free(&pool->worker_ida, worker->id);
3391 worker_detach_from_pool(worker);
3392 WARN_ON_ONCE(!list_empty(&worker->entry));
3397 worker_leave_idle(worker);
3399 /* no more worker necessary? */
3400 if (!need_more_worker(pool))
3403 /* do we need to manage? */
3404 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3408 * ->scheduled list can only be filled while a worker is
3409 * preparing to process a work or actually processing it.
3410 * Make sure nobody diddled with it while I was sleeping.
3412 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3415 * Finish PREP stage. We're guaranteed to have at least one idle
3416 * worker or that someone else has already assumed the manager
3417 * role. This is where @worker starts participating in concurrency
3418 * management if applicable and concurrency management is restored
3419 * after being rebound. See rebind_workers() for details.
3421 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3424 struct work_struct *work =
3425 list_first_entry(&pool->worklist,
3426 struct work_struct, entry);
3428 if (assign_work(work, worker, NULL))
3429 process_scheduled_works(worker);
3430 } while (keep_working(pool));
3432 worker_set_flags(worker, WORKER_PREP);
3435 * pool->lock is held and there's no work to process and no need to
3436 * manage, sleep. Workers are woken up only while holding
3437 * pool->lock or from local cpu, so setting the current state
3438 * before releasing pool->lock is enough to prevent losing any
3441 worker_enter_idle(worker);
3442 __set_current_state(TASK_IDLE);
3443 raw_spin_unlock_irq(&pool->lock);
3449 * rescuer_thread - the rescuer thread function
3452 * Workqueue rescuer thread function. There's one rescuer for each
3453 * workqueue which has WQ_MEM_RECLAIM set.
3455 * Regular work processing on a pool may block trying to create a new
3456 * worker which uses GFP_KERNEL allocation which has slight chance of
3457 * developing into deadlock if some works currently on the same queue
3458 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3459 * the problem rescuer solves.
3461 * When such condition is possible, the pool summons rescuers of all
3462 * workqueues which have works queued on the pool and let them process
3463 * those works so that forward progress can be guaranteed.
3465 * This should happen rarely.
3469 static int rescuer_thread(void *__rescuer)
3471 struct worker *rescuer = __rescuer;
3472 struct workqueue_struct *wq = rescuer->rescue_wq;
3475 set_user_nice(current, RESCUER_NICE_LEVEL);
3478 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3479 * doesn't participate in concurrency management.
3481 set_pf_worker(true);
3483 set_current_state(TASK_IDLE);
3486 * By the time the rescuer is requested to stop, the workqueue
3487 * shouldn't have any work pending, but @wq->maydays may still have
3488 * pwq(s) queued. This can happen by non-rescuer workers consuming
3489 * all the work items before the rescuer got to them. Go through
3490 * @wq->maydays processing before acting on should_stop so that the
3491 * list is always empty on exit.
3493 should_stop = kthread_should_stop();
3495 /* see whether any pwq is asking for help */
3496 raw_spin_lock_irq(&wq_mayday_lock);
3498 while (!list_empty(&wq->maydays)) {
3499 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3500 struct pool_workqueue, mayday_node);
3501 struct worker_pool *pool = pwq->pool;
3502 struct work_struct *work, *n;
3504 __set_current_state(TASK_RUNNING);
3505 list_del_init(&pwq->mayday_node);
3507 raw_spin_unlock_irq(&wq_mayday_lock);
3509 worker_attach_to_pool(rescuer, pool);
3511 raw_spin_lock_irq(&pool->lock);
3514 * Slurp in all works issued via this workqueue and
3517 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3518 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3519 if (get_work_pwq(work) == pwq &&
3520 assign_work(work, rescuer, &n))
3521 pwq->stats[PWQ_STAT_RESCUED]++;
3524 if (!list_empty(&rescuer->scheduled)) {
3525 process_scheduled_works(rescuer);
3528 * The above execution of rescued work items could
3529 * have created more to rescue through
3530 * pwq_activate_first_inactive() or chained
3531 * queueing. Let's put @pwq back on mayday list so
3532 * that such back-to-back work items, which may be
3533 * being used to relieve memory pressure, don't
3534 * incur MAYDAY_INTERVAL delay inbetween.
3536 if (pwq->nr_active && need_to_create_worker(pool)) {
3537 raw_spin_lock(&wq_mayday_lock);
3539 * Queue iff we aren't racing destruction
3540 * and somebody else hasn't queued it already.
3542 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3544 list_add_tail(&pwq->mayday_node, &wq->maydays);
3546 raw_spin_unlock(&wq_mayday_lock);
3551 * Put the reference grabbed by send_mayday(). @pool won't
3552 * go away while we're still attached to it.
3557 * Leave this pool. Notify regular workers; otherwise, we end up
3558 * with 0 concurrency and stalling the execution.
3562 raw_spin_unlock_irq(&pool->lock);
3564 worker_detach_from_pool(rescuer);
3566 raw_spin_lock_irq(&wq_mayday_lock);
3569 raw_spin_unlock_irq(&wq_mayday_lock);
3572 __set_current_state(TASK_RUNNING);
3573 set_pf_worker(false);
3577 /* rescuers should never participate in concurrency management */
3578 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3583 static void bh_worker(struct worker *worker)
3585 struct worker_pool *pool = worker->pool;
3586 int nr_restarts = BH_WORKER_RESTARTS;
3587 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3589 raw_spin_lock_irq(&pool->lock);
3590 worker_leave_idle(worker);
3593 * This function follows the structure of worker_thread(). See there for
3594 * explanations on each step.
3596 if (!need_more_worker(pool))
3599 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3600 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3603 struct work_struct *work =
3604 list_first_entry(&pool->worklist,
3605 struct work_struct, entry);
3607 if (assign_work(work, worker, NULL))
3608 process_scheduled_works(worker);
3609 } while (keep_working(pool) &&
3610 --nr_restarts && time_before(jiffies, end));
3612 worker_set_flags(worker, WORKER_PREP);
3614 worker_enter_idle(worker);
3616 raw_spin_unlock_irq(&pool->lock);
3620 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3622 * This is currently called from tasklet[_hi]action() and thus is also called
3623 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3624 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3627 * After full conversion, we'll add worker->softirq_action, directly use the
3628 * softirq action and obtain the worker pointer from the softirq_action pointer.
3630 void workqueue_softirq_action(bool highpri)
3632 struct worker_pool *pool =
3633 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3634 if (need_more_worker(pool))
3635 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3638 struct wq_drain_dead_softirq_work {
3639 struct work_struct work;
3640 struct worker_pool *pool;
3641 struct completion done;
3644 static void drain_dead_softirq_workfn(struct work_struct *work)
3646 struct wq_drain_dead_softirq_work *dead_work =
3647 container_of(work, struct wq_drain_dead_softirq_work, work);
3648 struct worker_pool *pool = dead_work->pool;
3652 * @pool's CPU is dead and we want to execute its still pending work
3653 * items from this BH work item which is running on a different CPU. As
3654 * its CPU is dead, @pool can't be kicked and, as work execution path
3655 * will be nested, a lockdep annotation needs to be suppressed. Mark
3656 * @pool with %POOL_BH_DRAINING for the special treatments.
3658 raw_spin_lock_irq(&pool->lock);
3659 pool->flags |= POOL_BH_DRAINING;
3660 raw_spin_unlock_irq(&pool->lock);
3662 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3664 raw_spin_lock_irq(&pool->lock);
3665 pool->flags &= ~POOL_BH_DRAINING;
3666 repeat = need_more_worker(pool);
3667 raw_spin_unlock_irq(&pool->lock);
3670 * bh_worker() might hit consecutive execution limit and bail. If there
3671 * still are pending work items, reschedule self and return so that we
3672 * don't hog this CPU's BH.
3675 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3676 queue_work(system_bh_highpri_wq, work);
3678 queue_work(system_bh_wq, work);
3680 complete(&dead_work->done);
3685 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3686 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3687 * have to worry about draining overlapping with CPU coming back online or
3688 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3689 * on). Let's keep it simple and drain them synchronously. These are BH work
3690 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3692 void workqueue_softirq_dead(unsigned int cpu)
3696 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3697 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3698 struct wq_drain_dead_softirq_work dead_work;
3700 if (!need_more_worker(pool))
3703 INIT_WORK(&dead_work.work, drain_dead_softirq_workfn);
3704 dead_work.pool = pool;
3705 init_completion(&dead_work.done);
3707 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3708 queue_work(system_bh_highpri_wq, &dead_work.work);
3710 queue_work(system_bh_wq, &dead_work.work);
3712 wait_for_completion(&dead_work.done);
3717 * check_flush_dependency - check for flush dependency sanity
3718 * @target_wq: workqueue being flushed
3719 * @target_work: work item being flushed (NULL for workqueue flushes)
3721 * %current is trying to flush the whole @target_wq or @target_work on it.
3722 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3723 * reclaiming memory or running on a workqueue which doesn't have
3724 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3727 static void check_flush_dependency(struct workqueue_struct *target_wq,
3728 struct work_struct *target_work)
3730 work_func_t target_func = target_work ? target_work->func : NULL;
3731 struct worker *worker;
3733 if (target_wq->flags & WQ_MEM_RECLAIM)
3736 worker = current_wq_worker();
3738 WARN_ONCE(current->flags & PF_MEMALLOC,
3739 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3740 current->pid, current->comm, target_wq->name, target_func);
3741 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3742 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3743 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3744 worker->current_pwq->wq->name, worker->current_func,
3745 target_wq->name, target_func);
3749 struct work_struct work;
3750 struct completion done;
3751 struct task_struct *task; /* purely informational */
3754 static void wq_barrier_func(struct work_struct *work)
3756 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3757 complete(&barr->done);
3761 * insert_wq_barrier - insert a barrier work
3762 * @pwq: pwq to insert barrier into
3763 * @barr: wq_barrier to insert
3764 * @target: target work to attach @barr to
3765 * @worker: worker currently executing @target, NULL if @target is not executing
3767 * @barr is linked to @target such that @barr is completed only after
3768 * @target finishes execution. Please note that the ordering
3769 * guarantee is observed only with respect to @target and on the local
3772 * Currently, a queued barrier can't be canceled. This is because
3773 * try_to_grab_pending() can't determine whether the work to be
3774 * grabbed is at the head of the queue and thus can't clear LINKED
3775 * flag of the previous work while there must be a valid next work
3776 * after a work with LINKED flag set.
3778 * Note that when @worker is non-NULL, @target may be modified
3779 * underneath us, so we can't reliably determine pwq from @target.
3782 * raw_spin_lock_irq(pool->lock).
3784 static void insert_wq_barrier(struct pool_workqueue *pwq,
3785 struct wq_barrier *barr,
3786 struct work_struct *target, struct worker *worker)
3788 static __maybe_unused struct lock_class_key bh_key, thr_key;
3789 unsigned int work_flags = 0;
3790 unsigned int work_color;
3791 struct list_head *head;
3794 * debugobject calls are safe here even with pool->lock locked
3795 * as we know for sure that this will not trigger any of the
3796 * checks and call back into the fixup functions where we
3799 * BH and threaded workqueues need separate lockdep keys to avoid
3800 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3803 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3804 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3805 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3807 init_completion_map(&barr->done, &target->lockdep_map);
3809 barr->task = current;
3811 /* The barrier work item does not participate in nr_active. */
3812 work_flags |= WORK_STRUCT_INACTIVE;
3815 * If @target is currently being executed, schedule the
3816 * barrier to the worker; otherwise, put it after @target.
3819 head = worker->scheduled.next;
3820 work_color = worker->current_color;
3822 unsigned long *bits = work_data_bits(target);
3824 head = target->entry.next;
3825 /* there can already be other linked works, inherit and set */
3826 work_flags |= *bits & WORK_STRUCT_LINKED;
3827 work_color = get_work_color(*bits);
3828 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3831 pwq->nr_in_flight[work_color]++;
3832 work_flags |= work_color_to_flags(work_color);
3834 insert_work(pwq, &barr->work, head, work_flags);
3838 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3839 * @wq: workqueue being flushed
3840 * @flush_color: new flush color, < 0 for no-op
3841 * @work_color: new work color, < 0 for no-op
3843 * Prepare pwqs for workqueue flushing.
3845 * If @flush_color is non-negative, flush_color on all pwqs should be
3846 * -1. If no pwq has in-flight commands at the specified color, all
3847 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3848 * has in flight commands, its pwq->flush_color is set to
3849 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3850 * wakeup logic is armed and %true is returned.
3852 * The caller should have initialized @wq->first_flusher prior to
3853 * calling this function with non-negative @flush_color. If
3854 * @flush_color is negative, no flush color update is done and %false
3857 * If @work_color is non-negative, all pwqs should have the same
3858 * work_color which is previous to @work_color and all will be
3859 * advanced to @work_color.
3862 * mutex_lock(wq->mutex).
3865 * %true if @flush_color >= 0 and there's something to flush. %false
3868 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3869 int flush_color, int work_color)
3872 struct pool_workqueue *pwq;
3874 if (flush_color >= 0) {
3875 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3876 atomic_set(&wq->nr_pwqs_to_flush, 1);
3879 for_each_pwq(pwq, wq) {
3880 struct worker_pool *pool = pwq->pool;
3882 raw_spin_lock_irq(&pool->lock);
3884 if (flush_color >= 0) {
3885 WARN_ON_ONCE(pwq->flush_color != -1);
3887 if (pwq->nr_in_flight[flush_color]) {
3888 pwq->flush_color = flush_color;
3889 atomic_inc(&wq->nr_pwqs_to_flush);
3894 if (work_color >= 0) {
3895 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3896 pwq->work_color = work_color;
3899 raw_spin_unlock_irq(&pool->lock);
3902 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3903 complete(&wq->first_flusher->done);
3908 static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3910 #ifdef CONFIG_LOCKDEP
3911 if (wq->flags & WQ_BH)
3914 lock_map_acquire(&wq->lockdep_map);
3915 lock_map_release(&wq->lockdep_map);
3917 if (wq->flags & WQ_BH)
3922 static void touch_work_lockdep_map(struct work_struct *work,
3923 struct workqueue_struct *wq)
3925 #ifdef CONFIG_LOCKDEP
3926 if (wq->flags & WQ_BH)
3929 lock_map_acquire(&work->lockdep_map);
3930 lock_map_release(&work->lockdep_map);
3932 if (wq->flags & WQ_BH)
3938 * __flush_workqueue - ensure that any scheduled work has run to completion.
3939 * @wq: workqueue to flush
3941 * This function sleeps until all work items which were queued on entry
3942 * have finished execution, but it is not livelocked by new incoming ones.
3944 void __flush_workqueue(struct workqueue_struct *wq)
3946 struct wq_flusher this_flusher = {
3947 .list = LIST_HEAD_INIT(this_flusher.list),
3949 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3953 if (WARN_ON(!wq_online))
3956 touch_wq_lockdep_map(wq);
3958 mutex_lock(&wq->mutex);
3961 * Start-to-wait phase
3963 next_color = work_next_color(wq->work_color);
3965 if (next_color != wq->flush_color) {
3967 * Color space is not full. The current work_color
3968 * becomes our flush_color and work_color is advanced
3971 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3972 this_flusher.flush_color = wq->work_color;
3973 wq->work_color = next_color;
3975 if (!wq->first_flusher) {
3976 /* no flush in progress, become the first flusher */
3977 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3979 wq->first_flusher = &this_flusher;
3981 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3983 /* nothing to flush, done */
3984 wq->flush_color = next_color;
3985 wq->first_flusher = NULL;
3990 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3991 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3992 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3996 * Oops, color space is full, wait on overflow queue.
3997 * The next flush completion will assign us
3998 * flush_color and transfer to flusher_queue.
4000 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
4003 check_flush_dependency(wq, NULL);
4005 mutex_unlock(&wq->mutex);
4007 wait_for_completion(&this_flusher.done);
4010 * Wake-up-and-cascade phase
4012 * First flushers are responsible for cascading flushes and
4013 * handling overflow. Non-first flushers can simply return.
4015 if (READ_ONCE(wq->first_flusher) != &this_flusher)
4018 mutex_lock(&wq->mutex);
4020 /* we might have raced, check again with mutex held */
4021 if (wq->first_flusher != &this_flusher)
4024 WRITE_ONCE(wq->first_flusher, NULL);
4026 WARN_ON_ONCE(!list_empty(&this_flusher.list));
4027 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4030 struct wq_flusher *next, *tmp;
4032 /* complete all the flushers sharing the current flush color */
4033 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
4034 if (next->flush_color != wq->flush_color)
4036 list_del_init(&next->list);
4037 complete(&next->done);
4040 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4041 wq->flush_color != work_next_color(wq->work_color));
4043 /* this flush_color is finished, advance by one */
4044 wq->flush_color = work_next_color(wq->flush_color);
4046 /* one color has been freed, handle overflow queue */
4047 if (!list_empty(&wq->flusher_overflow)) {
4049 * Assign the same color to all overflowed
4050 * flushers, advance work_color and append to
4051 * flusher_queue. This is the start-to-wait
4052 * phase for these overflowed flushers.
4054 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4055 tmp->flush_color = wq->work_color;
4057 wq->work_color = work_next_color(wq->work_color);
4059 list_splice_tail_init(&wq->flusher_overflow,
4060 &wq->flusher_queue);
4061 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4064 if (list_empty(&wq->flusher_queue)) {
4065 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4070 * Need to flush more colors. Make the next flusher
4071 * the new first flusher and arm pwqs.
4073 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4074 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4076 list_del_init(&next->list);
4077 wq->first_flusher = next;
4079 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4083 * Meh... this color is already done, clear first
4084 * flusher and repeat cascading.
4086 wq->first_flusher = NULL;
4090 mutex_unlock(&wq->mutex);
4092 EXPORT_SYMBOL(__flush_workqueue);
4095 * drain_workqueue - drain a workqueue
4096 * @wq: workqueue to drain
4098 * Wait until the workqueue becomes empty. While draining is in progress,
4099 * only chain queueing is allowed. IOW, only currently pending or running
4100 * work items on @wq can queue further work items on it. @wq is flushed
4101 * repeatedly until it becomes empty. The number of flushing is determined
4102 * by the depth of chaining and should be relatively short. Whine if it
4105 void drain_workqueue(struct workqueue_struct *wq)
4107 unsigned int flush_cnt = 0;
4108 struct pool_workqueue *pwq;
4111 * __queue_work() needs to test whether there are drainers, is much
4112 * hotter than drain_workqueue() and already looks at @wq->flags.
4113 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4115 mutex_lock(&wq->mutex);
4116 if (!wq->nr_drainers++)
4117 wq->flags |= __WQ_DRAINING;
4118 mutex_unlock(&wq->mutex);
4120 __flush_workqueue(wq);
4122 mutex_lock(&wq->mutex);
4124 for_each_pwq(pwq, wq) {
4127 raw_spin_lock_irq(&pwq->pool->lock);
4128 drained = pwq_is_empty(pwq);
4129 raw_spin_unlock_irq(&pwq->pool->lock);
4134 if (++flush_cnt == 10 ||
4135 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4136 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4137 wq->name, __func__, flush_cnt);
4139 mutex_unlock(&wq->mutex);
4143 if (!--wq->nr_drainers)
4144 wq->flags &= ~__WQ_DRAINING;
4145 mutex_unlock(&wq->mutex);
4147 EXPORT_SYMBOL_GPL(drain_workqueue);
4149 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4152 struct worker *worker = NULL;
4153 struct worker_pool *pool;
4154 struct pool_workqueue *pwq;
4155 struct workqueue_struct *wq;
4160 pool = get_work_pool(work);
4166 raw_spin_lock_irq(&pool->lock);
4167 /* see the comment in try_to_grab_pending() with the same code */
4168 pwq = get_work_pwq(work);
4170 if (unlikely(pwq->pool != pool))
4173 worker = find_worker_executing_work(pool, work);
4176 pwq = worker->current_pwq;
4180 check_flush_dependency(wq, work);
4182 insert_wq_barrier(pwq, barr, work, worker);
4183 raw_spin_unlock_irq(&pool->lock);
4185 touch_work_lockdep_map(work, wq);
4188 * Force a lock recursion deadlock when using flush_work() inside a
4189 * single-threaded or rescuer equipped workqueue.
4191 * For single threaded workqueues the deadlock happens when the work
4192 * is after the work issuing the flush_work(). For rescuer equipped
4193 * workqueues the deadlock happens when the rescuer stalls, blocking
4196 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4197 touch_wq_lockdep_map(wq);
4202 raw_spin_unlock_irq(&pool->lock);
4207 static bool __flush_work(struct work_struct *work, bool from_cancel)
4209 struct wq_barrier barr;
4211 if (WARN_ON(!wq_online))
4214 if (WARN_ON(!work->func))
4217 if (start_flush_work(work, &barr, from_cancel)) {
4218 wait_for_completion(&barr.done);
4219 destroy_work_on_stack(&barr.work);
4227 * flush_work - wait for a work to finish executing the last queueing instance
4228 * @work: the work to flush
4230 * Wait until @work has finished execution. @work is guaranteed to be idle
4231 * on return if it hasn't been requeued since flush started.
4234 * %true if flush_work() waited for the work to finish execution,
4235 * %false if it was already idle.
4237 bool flush_work(struct work_struct *work)
4239 return __flush_work(work, false);
4241 EXPORT_SYMBOL_GPL(flush_work);
4244 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4245 * @dwork: the delayed work to flush
4247 * Delayed timer is cancelled and the pending work is queued for
4248 * immediate execution. Like flush_work(), this function only
4249 * considers the last queueing instance of @dwork.
4252 * %true if flush_work() waited for the work to finish execution,
4253 * %false if it was already idle.
4255 bool flush_delayed_work(struct delayed_work *dwork)
4257 local_irq_disable();
4258 if (del_timer_sync(&dwork->timer))
4259 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4261 return flush_work(&dwork->work);
4263 EXPORT_SYMBOL(flush_delayed_work);
4266 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4267 * @rwork: the rcu work to flush
4270 * %true if flush_rcu_work() waited for the work to finish execution,
4271 * %false if it was already idle.
4273 bool flush_rcu_work(struct rcu_work *rwork)
4275 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4277 flush_work(&rwork->work);
4280 return flush_work(&rwork->work);
4283 EXPORT_SYMBOL(flush_rcu_work);
4285 static bool __cancel_work(struct work_struct *work, u32 cflags)
4287 unsigned long irq_flags;
4291 ret = try_to_grab_pending(work, cflags, &irq_flags);
4292 } while (unlikely(ret == -EAGAIN));
4294 if (unlikely(ret < 0))
4297 set_work_pool_and_clear_pending(work, get_work_pool_id(work), 0);
4298 local_irq_restore(irq_flags);
4302 static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4304 unsigned long irq_flags;
4307 /* claim @work and tell other tasks trying to grab @work to back off */
4308 ret = work_grab_pending(work, cflags, &irq_flags);
4309 mark_work_canceling(work);
4310 local_irq_restore(irq_flags);
4313 * Skip __flush_work() during early boot when we know that @work isn't
4314 * executing. This allows canceling during early boot.
4317 __flush_work(work, true);
4320 * smp_mb() at the end of set_work_pool_and_clear_pending() is paired
4321 * with prepare_to_wait() above so that either waitqueue_active() is
4322 * visible here or !work_is_canceling() is visible there.
4324 set_work_pool_and_clear_pending(work, WORK_OFFQ_POOL_NONE, 0);
4326 if (waitqueue_active(&wq_cancel_waitq))
4327 __wake_up(&wq_cancel_waitq, TASK_NORMAL, 1, work);
4333 * See cancel_delayed_work()
4335 bool cancel_work(struct work_struct *work)
4337 return __cancel_work(work, 0);
4339 EXPORT_SYMBOL(cancel_work);
4342 * cancel_work_sync - cancel a work and wait for it to finish
4343 * @work: the work to cancel
4345 * Cancel @work and wait for its execution to finish. This function
4346 * can be used even if the work re-queues itself or migrates to
4347 * another workqueue. On return from this function, @work is
4348 * guaranteed to be not pending or executing on any CPU.
4350 * cancel_work_sync(&delayed_work->work) must not be used for
4351 * delayed_work's. Use cancel_delayed_work_sync() instead.
4353 * The caller must ensure that the workqueue on which @work was last
4354 * queued can't be destroyed before this function returns.
4357 * %true if @work was pending, %false otherwise.
4359 bool cancel_work_sync(struct work_struct *work)
4361 return __cancel_work_sync(work, 0);
4363 EXPORT_SYMBOL_GPL(cancel_work_sync);
4366 * cancel_delayed_work - cancel a delayed work
4367 * @dwork: delayed_work to cancel
4369 * Kill off a pending delayed_work.
4371 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4375 * The work callback function may still be running on return, unless
4376 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4377 * use cancel_delayed_work_sync() to wait on it.
4379 * This function is safe to call from any context including IRQ handler.
4381 bool cancel_delayed_work(struct delayed_work *dwork)
4383 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4385 EXPORT_SYMBOL(cancel_delayed_work);
4388 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4389 * @dwork: the delayed work cancel
4391 * This is cancel_work_sync() for delayed works.
4394 * %true if @dwork was pending, %false otherwise.
4396 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4398 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4400 EXPORT_SYMBOL(cancel_delayed_work_sync);
4403 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4404 * @func: the function to call
4406 * schedule_on_each_cpu() executes @func on each online CPU using the
4407 * system workqueue and blocks until all CPUs have completed.
4408 * schedule_on_each_cpu() is very slow.
4411 * 0 on success, -errno on failure.
4413 int schedule_on_each_cpu(work_func_t func)
4416 struct work_struct __percpu *works;
4418 works = alloc_percpu(struct work_struct);
4424 for_each_online_cpu(cpu) {
4425 struct work_struct *work = per_cpu_ptr(works, cpu);
4427 INIT_WORK(work, func);
4428 schedule_work_on(cpu, work);
4431 for_each_online_cpu(cpu)
4432 flush_work(per_cpu_ptr(works, cpu));
4440 * execute_in_process_context - reliably execute the routine with user context
4441 * @fn: the function to execute
4442 * @ew: guaranteed storage for the execute work structure (must
4443 * be available when the work executes)
4445 * Executes the function immediately if process context is available,
4446 * otherwise schedules the function for delayed execution.
4448 * Return: 0 - function was executed
4449 * 1 - function was scheduled for execution
4451 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4453 if (!in_interrupt()) {
4458 INIT_WORK(&ew->work, fn);
4459 schedule_work(&ew->work);
4463 EXPORT_SYMBOL_GPL(execute_in_process_context);
4466 * free_workqueue_attrs - free a workqueue_attrs
4467 * @attrs: workqueue_attrs to free
4469 * Undo alloc_workqueue_attrs().
4471 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4474 free_cpumask_var(attrs->cpumask);
4475 free_cpumask_var(attrs->__pod_cpumask);
4481 * alloc_workqueue_attrs - allocate a workqueue_attrs
4483 * Allocate a new workqueue_attrs, initialize with default settings and
4486 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4488 struct workqueue_attrs *alloc_workqueue_attrs(void)
4490 struct workqueue_attrs *attrs;
4492 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4495 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4497 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4500 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4501 attrs->affn_scope = WQ_AFFN_DFL;
4504 free_workqueue_attrs(attrs);
4508 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4509 const struct workqueue_attrs *from)
4511 to->nice = from->nice;
4512 cpumask_copy(to->cpumask, from->cpumask);
4513 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4514 to->affn_strict = from->affn_strict;
4517 * Unlike hash and equality test, copying shouldn't ignore wq-only
4518 * fields as copying is used for both pool and wq attrs. Instead,
4519 * get_unbound_pool() explicitly clears the fields.
4521 to->affn_scope = from->affn_scope;
4522 to->ordered = from->ordered;
4526 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4527 * comments in 'struct workqueue_attrs' definition.
4529 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4531 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4532 attrs->ordered = false;
4535 /* hash value of the content of @attr */
4536 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4540 hash = jhash_1word(attrs->nice, hash);
4541 hash = jhash(cpumask_bits(attrs->cpumask),
4542 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4543 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4544 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4545 hash = jhash_1word(attrs->affn_strict, hash);
4549 /* content equality test */
4550 static bool wqattrs_equal(const struct workqueue_attrs *a,
4551 const struct workqueue_attrs *b)
4553 if (a->nice != b->nice)
4555 if (!cpumask_equal(a->cpumask, b->cpumask))
4557 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4559 if (a->affn_strict != b->affn_strict)
4564 /* Update @attrs with actually available CPUs */
4565 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4566 const cpumask_t *unbound_cpumask)
4569 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4570 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4573 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4574 if (unlikely(cpumask_empty(attrs->cpumask)))
4575 cpumask_copy(attrs->cpumask, unbound_cpumask);
4578 /* find wq_pod_type to use for @attrs */
4579 static const struct wq_pod_type *
4580 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4582 enum wq_affn_scope scope;
4583 struct wq_pod_type *pt;
4585 /* to synchronize access to wq_affn_dfl */
4586 lockdep_assert_held(&wq_pool_mutex);
4588 if (attrs->affn_scope == WQ_AFFN_DFL)
4589 scope = wq_affn_dfl;
4591 scope = attrs->affn_scope;
4593 pt = &wq_pod_types[scope];
4595 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4596 likely(pt->nr_pods))
4600 * Before workqueue_init_topology(), only SYSTEM is available which is
4601 * initialized in workqueue_init_early().
4603 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4604 BUG_ON(!pt->nr_pods);
4609 * init_worker_pool - initialize a newly zalloc'd worker_pool
4610 * @pool: worker_pool to initialize
4612 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4614 * Return: 0 on success, -errno on failure. Even on failure, all fields
4615 * inside @pool proper are initialized and put_unbound_pool() can be called
4616 * on @pool safely to release it.
4618 static int init_worker_pool(struct worker_pool *pool)
4620 raw_spin_lock_init(&pool->lock);
4623 pool->node = NUMA_NO_NODE;
4624 pool->flags |= POOL_DISASSOCIATED;
4625 pool->watchdog_ts = jiffies;
4626 INIT_LIST_HEAD(&pool->worklist);
4627 INIT_LIST_HEAD(&pool->idle_list);
4628 hash_init(pool->busy_hash);
4630 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4631 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4633 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4635 INIT_LIST_HEAD(&pool->workers);
4636 INIT_LIST_HEAD(&pool->dying_workers);
4638 ida_init(&pool->worker_ida);
4639 INIT_HLIST_NODE(&pool->hash_node);
4642 /* shouldn't fail above this point */
4643 pool->attrs = alloc_workqueue_attrs();
4647 wqattrs_clear_for_pool(pool->attrs);
4652 #ifdef CONFIG_LOCKDEP
4653 static void wq_init_lockdep(struct workqueue_struct *wq)
4657 lockdep_register_key(&wq->key);
4658 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4660 lock_name = wq->name;
4662 wq->lock_name = lock_name;
4663 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4666 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4668 lockdep_unregister_key(&wq->key);
4671 static void wq_free_lockdep(struct workqueue_struct *wq)
4673 if (wq->lock_name != wq->name)
4674 kfree(wq->lock_name);
4677 static void wq_init_lockdep(struct workqueue_struct *wq)
4681 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4685 static void wq_free_lockdep(struct workqueue_struct *wq)
4690 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4694 for_each_node(node) {
4695 kfree(nna_ar[node]);
4696 nna_ar[node] = NULL;
4699 kfree(nna_ar[nr_node_ids]);
4700 nna_ar[nr_node_ids] = NULL;
4703 static void init_node_nr_active(struct wq_node_nr_active *nna)
4705 nna->max = WQ_DFL_MIN_ACTIVE;
4706 atomic_set(&nna->nr, 0);
4707 raw_spin_lock_init(&nna->lock);
4708 INIT_LIST_HEAD(&nna->pending_pwqs);
4712 * Each node's nr_active counter will be accessed mostly from its own node and
4713 * should be allocated in the node.
4715 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4717 struct wq_node_nr_active *nna;
4720 for_each_node(node) {
4721 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4724 init_node_nr_active(nna);
4728 /* [nr_node_ids] is used as the fallback */
4729 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4732 init_node_nr_active(nna);
4733 nna_ar[nr_node_ids] = nna;
4738 free_node_nr_active(nna_ar);
4742 static void rcu_free_wq(struct rcu_head *rcu)
4744 struct workqueue_struct *wq =
4745 container_of(rcu, struct workqueue_struct, rcu);
4747 if (wq->flags & WQ_UNBOUND)
4748 free_node_nr_active(wq->node_nr_active);
4750 wq_free_lockdep(wq);
4751 free_percpu(wq->cpu_pwq);
4752 free_workqueue_attrs(wq->unbound_attrs);
4756 static void rcu_free_pool(struct rcu_head *rcu)
4758 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4760 ida_destroy(&pool->worker_ida);
4761 free_workqueue_attrs(pool->attrs);
4766 * put_unbound_pool - put a worker_pool
4767 * @pool: worker_pool to put
4769 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4770 * safe manner. get_unbound_pool() calls this function on its failure path
4771 * and this function should be able to release pools which went through,
4772 * successfully or not, init_worker_pool().
4774 * Should be called with wq_pool_mutex held.
4776 static void put_unbound_pool(struct worker_pool *pool)
4778 DECLARE_COMPLETION_ONSTACK(detach_completion);
4779 struct worker *worker;
4780 LIST_HEAD(cull_list);
4782 lockdep_assert_held(&wq_pool_mutex);
4788 if (WARN_ON(!(pool->cpu < 0)) ||
4789 WARN_ON(!list_empty(&pool->worklist)))
4792 /* release id and unhash */
4794 idr_remove(&worker_pool_idr, pool->id);
4795 hash_del(&pool->hash_node);
4798 * Become the manager and destroy all workers. This prevents
4799 * @pool's workers from blocking on attach_mutex. We're the last
4800 * manager and @pool gets freed with the flag set.
4802 * Having a concurrent manager is quite unlikely to happen as we can
4803 * only get here with
4804 * pwq->refcnt == pool->refcnt == 0
4805 * which implies no work queued to the pool, which implies no worker can
4806 * become the manager. However a worker could have taken the role of
4807 * manager before the refcnts dropped to 0, since maybe_create_worker()
4811 rcuwait_wait_event(&manager_wait,
4812 !(pool->flags & POOL_MANAGER_ACTIVE),
4813 TASK_UNINTERRUPTIBLE);
4815 mutex_lock(&wq_pool_attach_mutex);
4816 raw_spin_lock_irq(&pool->lock);
4817 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4818 pool->flags |= POOL_MANAGER_ACTIVE;
4821 raw_spin_unlock_irq(&pool->lock);
4822 mutex_unlock(&wq_pool_attach_mutex);
4825 while ((worker = first_idle_worker(pool)))
4826 set_worker_dying(worker, &cull_list);
4827 WARN_ON(pool->nr_workers || pool->nr_idle);
4828 raw_spin_unlock_irq(&pool->lock);
4830 wake_dying_workers(&cull_list);
4832 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4833 pool->detach_completion = &detach_completion;
4834 mutex_unlock(&wq_pool_attach_mutex);
4836 if (pool->detach_completion)
4837 wait_for_completion(pool->detach_completion);
4839 /* shut down the timers */
4840 del_timer_sync(&pool->idle_timer);
4841 cancel_work_sync(&pool->idle_cull_work);
4842 del_timer_sync(&pool->mayday_timer);
4844 /* RCU protected to allow dereferences from get_work_pool() */
4845 call_rcu(&pool->rcu, rcu_free_pool);
4849 * get_unbound_pool - get a worker_pool with the specified attributes
4850 * @attrs: the attributes of the worker_pool to get
4852 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4853 * reference count and return it. If there already is a matching
4854 * worker_pool, it will be used; otherwise, this function attempts to
4857 * Should be called with wq_pool_mutex held.
4859 * Return: On success, a worker_pool with the same attributes as @attrs.
4860 * On failure, %NULL.
4862 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4864 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4865 u32 hash = wqattrs_hash(attrs);
4866 struct worker_pool *pool;
4867 int pod, node = NUMA_NO_NODE;
4869 lockdep_assert_held(&wq_pool_mutex);
4871 /* do we already have a matching pool? */
4872 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4873 if (wqattrs_equal(pool->attrs, attrs)) {
4879 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4880 for (pod = 0; pod < pt->nr_pods; pod++) {
4881 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4882 node = pt->pod_node[pod];
4887 /* nope, create a new one */
4888 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4889 if (!pool || init_worker_pool(pool) < 0)
4893 copy_workqueue_attrs(pool->attrs, attrs);
4894 wqattrs_clear_for_pool(pool->attrs);
4896 if (worker_pool_assign_id(pool) < 0)
4899 /* create and start the initial worker */
4900 if (wq_online && !create_worker(pool))
4904 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4909 put_unbound_pool(pool);
4913 static void rcu_free_pwq(struct rcu_head *rcu)
4915 kmem_cache_free(pwq_cache,
4916 container_of(rcu, struct pool_workqueue, rcu));
4920 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4921 * refcnt and needs to be destroyed.
4923 static void pwq_release_workfn(struct kthread_work *work)
4925 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4927 struct workqueue_struct *wq = pwq->wq;
4928 struct worker_pool *pool = pwq->pool;
4929 bool is_last = false;
4932 * When @pwq is not linked, it doesn't hold any reference to the
4933 * @wq, and @wq is invalid to access.
4935 if (!list_empty(&pwq->pwqs_node)) {
4936 mutex_lock(&wq->mutex);
4937 list_del_rcu(&pwq->pwqs_node);
4938 is_last = list_empty(&wq->pwqs);
4941 * For ordered workqueue with a plugged dfl_pwq, restart it now.
4943 if (!is_last && (wq->flags & __WQ_ORDERED))
4944 unplug_oldest_pwq(wq);
4946 mutex_unlock(&wq->mutex);
4949 if (wq->flags & WQ_UNBOUND) {
4950 mutex_lock(&wq_pool_mutex);
4951 put_unbound_pool(pool);
4952 mutex_unlock(&wq_pool_mutex);
4955 if (!list_empty(&pwq->pending_node)) {
4956 struct wq_node_nr_active *nna =
4957 wq_node_nr_active(pwq->wq, pwq->pool->node);
4959 raw_spin_lock_irq(&nna->lock);
4960 list_del_init(&pwq->pending_node);
4961 raw_spin_unlock_irq(&nna->lock);
4964 call_rcu(&pwq->rcu, rcu_free_pwq);
4967 * If we're the last pwq going away, @wq is already dead and no one
4968 * is gonna access it anymore. Schedule RCU free.
4971 wq_unregister_lockdep(wq);
4972 call_rcu(&wq->rcu, rcu_free_wq);
4976 /* initialize newly allocated @pwq which is associated with @wq and @pool */
4977 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4978 struct worker_pool *pool)
4980 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
4982 memset(pwq, 0, sizeof(*pwq));
4986 pwq->flush_color = -1;
4988 INIT_LIST_HEAD(&pwq->inactive_works);
4989 INIT_LIST_HEAD(&pwq->pending_node);
4990 INIT_LIST_HEAD(&pwq->pwqs_node);
4991 INIT_LIST_HEAD(&pwq->mayday_node);
4992 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4995 /* sync @pwq with the current state of its associated wq and link it */
4996 static void link_pwq(struct pool_workqueue *pwq)
4998 struct workqueue_struct *wq = pwq->wq;
5000 lockdep_assert_held(&wq->mutex);
5002 /* may be called multiple times, ignore if already linked */
5003 if (!list_empty(&pwq->pwqs_node))
5006 /* set the matching work_color */
5007 pwq->work_color = wq->work_color;
5010 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5013 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5014 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5015 const struct workqueue_attrs *attrs)
5017 struct worker_pool *pool;
5018 struct pool_workqueue *pwq;
5020 lockdep_assert_held(&wq_pool_mutex);
5022 pool = get_unbound_pool(attrs);
5026 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5028 put_unbound_pool(pool);
5032 init_pwq(pwq, wq, pool);
5037 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5038 * @attrs: the wq_attrs of the default pwq of the target workqueue
5039 * @cpu: the target CPU
5040 * @cpu_going_down: if >= 0, the CPU to consider as offline
5042 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
5043 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
5044 * The result is stored in @attrs->__pod_cpumask.
5046 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5047 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5048 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5050 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5052 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
5055 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5056 int pod = pt->cpu_pod[cpu];
5058 /* does @pod have any online CPUs @attrs wants? */
5059 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5060 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
5061 if (cpu_going_down >= 0)
5062 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
5064 if (cpumask_empty(attrs->__pod_cpumask)) {
5065 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5069 /* yeap, return possible CPUs in @pod that @attrs wants */
5070 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
5072 if (cpumask_empty(attrs->__pod_cpumask))
5073 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
5074 "possible intersect\n");
5077 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5078 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5079 int cpu, struct pool_workqueue *pwq)
5081 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5082 struct pool_workqueue *old_pwq;
5084 lockdep_assert_held(&wq_pool_mutex);
5085 lockdep_assert_held(&wq->mutex);
5087 /* link_pwq() can handle duplicate calls */
5090 old_pwq = rcu_access_pointer(*slot);
5091 rcu_assign_pointer(*slot, pwq);
5095 /* context to store the prepared attrs & pwqs before applying */
5096 struct apply_wqattrs_ctx {
5097 struct workqueue_struct *wq; /* target workqueue */
5098 struct workqueue_attrs *attrs; /* attrs to apply */
5099 struct list_head list; /* queued for batching commit */
5100 struct pool_workqueue *dfl_pwq;
5101 struct pool_workqueue *pwq_tbl[];
5104 /* free the resources after success or abort */
5105 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5110 for_each_possible_cpu(cpu)
5111 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5112 put_pwq_unlocked(ctx->dfl_pwq);
5114 free_workqueue_attrs(ctx->attrs);
5120 /* allocate the attrs and pwqs for later installation */
5121 static struct apply_wqattrs_ctx *
5122 apply_wqattrs_prepare(struct workqueue_struct *wq,
5123 const struct workqueue_attrs *attrs,
5124 const cpumask_var_t unbound_cpumask)
5126 struct apply_wqattrs_ctx *ctx;
5127 struct workqueue_attrs *new_attrs;
5130 lockdep_assert_held(&wq_pool_mutex);
5132 if (WARN_ON(attrs->affn_scope < 0 ||
5133 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5134 return ERR_PTR(-EINVAL);
5136 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5138 new_attrs = alloc_workqueue_attrs();
5139 if (!ctx || !new_attrs)
5143 * If something goes wrong during CPU up/down, we'll fall back to
5144 * the default pwq covering whole @attrs->cpumask. Always create
5145 * it even if we don't use it immediately.
5147 copy_workqueue_attrs(new_attrs, attrs);
5148 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5149 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5150 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5154 for_each_possible_cpu(cpu) {
5155 if (new_attrs->ordered) {
5156 ctx->dfl_pwq->refcnt++;
5157 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5159 wq_calc_pod_cpumask(new_attrs, cpu, -1);
5160 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5161 if (!ctx->pwq_tbl[cpu])
5166 /* save the user configured attrs and sanitize it. */
5167 copy_workqueue_attrs(new_attrs, attrs);
5168 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5169 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5170 ctx->attrs = new_attrs;
5173 * For initialized ordered workqueues, there should only be one pwq
5174 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5175 * of newly queued work items until execution of older work items in
5176 * the old pwq's have completed.
5178 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5179 ctx->dfl_pwq->plugged = true;
5185 free_workqueue_attrs(new_attrs);
5186 apply_wqattrs_cleanup(ctx);
5187 return ERR_PTR(-ENOMEM);
5190 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5191 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5195 /* all pwqs have been created successfully, let's install'em */
5196 mutex_lock(&ctx->wq->mutex);
5198 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5200 /* save the previous pwqs and install the new ones */
5201 for_each_possible_cpu(cpu)
5202 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5204 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5206 /* update node_nr_active->max */
5207 wq_update_node_max_active(ctx->wq, -1);
5209 /* rescuer needs to respect wq cpumask changes */
5210 if (ctx->wq->rescuer)
5211 set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5212 unbound_effective_cpumask(ctx->wq));
5214 mutex_unlock(&ctx->wq->mutex);
5217 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5218 const struct workqueue_attrs *attrs)
5220 struct apply_wqattrs_ctx *ctx;
5222 /* only unbound workqueues can change attributes */
5223 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5226 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5228 return PTR_ERR(ctx);
5230 /* the ctx has been prepared successfully, let's commit it */
5231 apply_wqattrs_commit(ctx);
5232 apply_wqattrs_cleanup(ctx);
5238 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5239 * @wq: the target workqueue
5240 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5242 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5243 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5244 * work items are affine to the pod it was issued on. Older pwqs are released as
5245 * in-flight work items finish. Note that a work item which repeatedly requeues
5246 * itself back-to-back will stay on its current pwq.
5248 * Performs GFP_KERNEL allocations.
5250 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
5252 * Return: 0 on success and -errno on failure.
5254 int apply_workqueue_attrs(struct workqueue_struct *wq,
5255 const struct workqueue_attrs *attrs)
5259 lockdep_assert_cpus_held();
5261 mutex_lock(&wq_pool_mutex);
5262 ret = apply_workqueue_attrs_locked(wq, attrs);
5263 mutex_unlock(&wq_pool_mutex);
5269 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
5270 * @wq: the target workqueue
5271 * @cpu: the CPU to update pool association for
5272 * @hotplug_cpu: the CPU coming up or going down
5273 * @online: whether @cpu is coming up or going down
5275 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5276 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
5280 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5281 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5283 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5284 * with a cpumask spanning multiple pods, the workers which were already
5285 * executing the work items for the workqueue will lose their CPU affinity and
5286 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5287 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5288 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5290 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
5291 int hotplug_cpu, bool online)
5293 int off_cpu = online ? -1 : hotplug_cpu;
5294 struct pool_workqueue *old_pwq = NULL, *pwq;
5295 struct workqueue_attrs *target_attrs;
5297 lockdep_assert_held(&wq_pool_mutex);
5299 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5303 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5304 * Let's use a preallocated one. The following buf is protected by
5305 * CPU hotplug exclusion.
5307 target_attrs = wq_update_pod_attrs_buf;
5309 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5310 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5312 /* nothing to do if the target cpumask matches the current pwq */
5313 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
5314 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5317 /* create a new pwq */
5318 pwq = alloc_unbound_pwq(wq, target_attrs);
5320 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5325 /* Install the new pwq. */
5326 mutex_lock(&wq->mutex);
5327 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5331 mutex_lock(&wq->mutex);
5332 pwq = unbound_pwq(wq, -1);
5333 raw_spin_lock_irq(&pwq->pool->lock);
5335 raw_spin_unlock_irq(&pwq->pool->lock);
5336 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5338 mutex_unlock(&wq->mutex);
5339 put_pwq_unlocked(old_pwq);
5342 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5344 bool highpri = wq->flags & WQ_HIGHPRI;
5347 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5351 if (!(wq->flags & WQ_UNBOUND)) {
5352 for_each_possible_cpu(cpu) {
5353 struct pool_workqueue **pwq_p;
5354 struct worker_pool __percpu *pools;
5355 struct worker_pool *pool;
5357 if (wq->flags & WQ_BH)
5358 pools = bh_worker_pools;
5360 pools = cpu_worker_pools;
5362 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5363 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5365 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5370 init_pwq(*pwq_p, wq, pool);
5372 mutex_lock(&wq->mutex);
5374 mutex_unlock(&wq->mutex);
5380 if (wq->flags & __WQ_ORDERED) {
5381 struct pool_workqueue *dfl_pwq;
5383 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5384 /* there should only be single pwq for ordering guarantee */
5385 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5386 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5387 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5388 "ordering guarantee broken for workqueue %s\n", wq->name);
5390 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5394 /* for unbound pwq, flush the pwq_release_worker ensures that the
5395 * pwq_release_workfn() completes before calling kfree(wq).
5398 kthread_flush_worker(pwq_release_worker);
5404 for_each_possible_cpu(cpu) {
5405 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5408 kmem_cache_free(pwq_cache, pwq);
5410 free_percpu(wq->cpu_pwq);
5416 static int wq_clamp_max_active(int max_active, unsigned int flags,
5419 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5420 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5421 max_active, name, 1, WQ_MAX_ACTIVE);
5423 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5427 * Workqueues which may be used during memory reclaim should have a rescuer
5428 * to guarantee forward progress.
5430 static int init_rescuer(struct workqueue_struct *wq)
5432 struct worker *rescuer;
5435 if (!(wq->flags & WQ_MEM_RECLAIM))
5438 rescuer = alloc_worker(NUMA_NO_NODE);
5440 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5445 rescuer->rescue_wq = wq;
5446 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5447 if (IS_ERR(rescuer->task)) {
5448 ret = PTR_ERR(rescuer->task);
5449 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5450 wq->name, ERR_PTR(ret));
5455 wq->rescuer = rescuer;
5456 if (wq->flags & WQ_UNBOUND)
5457 kthread_bind_mask(rescuer->task, wq_unbound_cpumask);
5459 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5460 wake_up_process(rescuer->task);
5466 * wq_adjust_max_active - update a wq's max_active to the current setting
5467 * @wq: target workqueue
5469 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5470 * activate inactive work items accordingly. If @wq is freezing, clear
5471 * @wq->max_active to zero.
5473 static void wq_adjust_max_active(struct workqueue_struct *wq)
5476 int new_max, new_min;
5478 lockdep_assert_held(&wq->mutex);
5480 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5484 new_max = wq->saved_max_active;
5485 new_min = wq->saved_min_active;
5488 if (wq->max_active == new_max && wq->min_active == new_min)
5492 * Update @wq->max/min_active and then kick inactive work items if more
5493 * active work items are allowed. This doesn't break work item ordering
5494 * because new work items are always queued behind existing inactive
5495 * work items if there are any.
5497 WRITE_ONCE(wq->max_active, new_max);
5498 WRITE_ONCE(wq->min_active, new_min);
5500 if (wq->flags & WQ_UNBOUND)
5501 wq_update_node_max_active(wq, -1);
5507 * Round-robin through pwq's activating the first inactive work item
5508 * until max_active is filled.
5511 struct pool_workqueue *pwq;
5514 for_each_pwq(pwq, wq) {
5515 unsigned long irq_flags;
5517 /* can be called during early boot w/ irq disabled */
5518 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5519 if (pwq_activate_first_inactive(pwq, true)) {
5521 kick_pool(pwq->pool);
5523 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5525 } while (activated);
5529 struct workqueue_struct *alloc_workqueue(const char *fmt,
5531 int max_active, ...)
5534 struct workqueue_struct *wq;
5538 if (flags & WQ_BH) {
5539 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5541 if (WARN_ON_ONCE(max_active))
5545 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5546 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5547 flags |= WQ_UNBOUND;
5549 /* allocate wq and format name */
5550 if (flags & WQ_UNBOUND)
5551 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5553 wq_size = sizeof(*wq);
5555 wq = kzalloc(wq_size, GFP_KERNEL);
5559 if (flags & WQ_UNBOUND) {
5560 wq->unbound_attrs = alloc_workqueue_attrs();
5561 if (!wq->unbound_attrs)
5565 va_start(args, max_active);
5566 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5569 if (name_len >= WQ_NAME_LEN)
5570 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5573 if (flags & WQ_BH) {
5575 * BH workqueues always share a single execution context per CPU
5576 * and don't impose any max_active limit.
5578 max_active = INT_MAX;
5580 max_active = max_active ?: WQ_DFL_ACTIVE;
5581 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5586 wq->max_active = max_active;
5587 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5588 wq->saved_max_active = wq->max_active;
5589 wq->saved_min_active = wq->min_active;
5590 mutex_init(&wq->mutex);
5591 atomic_set(&wq->nr_pwqs_to_flush, 0);
5592 INIT_LIST_HEAD(&wq->pwqs);
5593 INIT_LIST_HEAD(&wq->flusher_queue);
5594 INIT_LIST_HEAD(&wq->flusher_overflow);
5595 INIT_LIST_HEAD(&wq->maydays);
5597 wq_init_lockdep(wq);
5598 INIT_LIST_HEAD(&wq->list);
5600 if (flags & WQ_UNBOUND) {
5601 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5602 goto err_unreg_lockdep;
5605 if (alloc_and_link_pwqs(wq) < 0)
5606 goto err_free_node_nr_active;
5608 if (wq_online && init_rescuer(wq) < 0)
5611 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5615 * wq_pool_mutex protects global freeze state and workqueues list.
5616 * Grab it, adjust max_active and add the new @wq to workqueues
5619 mutex_lock(&wq_pool_mutex);
5621 mutex_lock(&wq->mutex);
5622 wq_adjust_max_active(wq);
5623 mutex_unlock(&wq->mutex);
5625 list_add_tail_rcu(&wq->list, &workqueues);
5627 mutex_unlock(&wq_pool_mutex);
5631 err_free_node_nr_active:
5632 if (wq->flags & WQ_UNBOUND)
5633 free_node_nr_active(wq->node_nr_active);
5635 wq_unregister_lockdep(wq);
5636 wq_free_lockdep(wq);
5638 free_workqueue_attrs(wq->unbound_attrs);
5642 destroy_workqueue(wq);
5645 EXPORT_SYMBOL_GPL(alloc_workqueue);
5647 static bool pwq_busy(struct pool_workqueue *pwq)
5651 for (i = 0; i < WORK_NR_COLORS; i++)
5652 if (pwq->nr_in_flight[i])
5655 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5657 if (!pwq_is_empty(pwq))
5664 * destroy_workqueue - safely terminate a workqueue
5665 * @wq: target workqueue
5667 * Safely destroy a workqueue. All work currently pending will be done first.
5669 void destroy_workqueue(struct workqueue_struct *wq)
5671 struct pool_workqueue *pwq;
5675 * Remove it from sysfs first so that sanity check failure doesn't
5676 * lead to sysfs name conflicts.
5678 workqueue_sysfs_unregister(wq);
5680 /* mark the workqueue destruction is in progress */
5681 mutex_lock(&wq->mutex);
5682 wq->flags |= __WQ_DESTROYING;
5683 mutex_unlock(&wq->mutex);
5685 /* drain it before proceeding with destruction */
5686 drain_workqueue(wq);
5688 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5690 struct worker *rescuer = wq->rescuer;
5692 /* this prevents new queueing */
5693 raw_spin_lock_irq(&wq_mayday_lock);
5695 raw_spin_unlock_irq(&wq_mayday_lock);
5697 /* rescuer will empty maydays list before exiting */
5698 kthread_stop(rescuer->task);
5703 * Sanity checks - grab all the locks so that we wait for all
5704 * in-flight operations which may do put_pwq().
5706 mutex_lock(&wq_pool_mutex);
5707 mutex_lock(&wq->mutex);
5708 for_each_pwq(pwq, wq) {
5709 raw_spin_lock_irq(&pwq->pool->lock);
5710 if (WARN_ON(pwq_busy(pwq))) {
5711 pr_warn("%s: %s has the following busy pwq\n",
5712 __func__, wq->name);
5714 raw_spin_unlock_irq(&pwq->pool->lock);
5715 mutex_unlock(&wq->mutex);
5716 mutex_unlock(&wq_pool_mutex);
5717 show_one_workqueue(wq);
5720 raw_spin_unlock_irq(&pwq->pool->lock);
5722 mutex_unlock(&wq->mutex);
5725 * wq list is used to freeze wq, remove from list after
5726 * flushing is complete in case freeze races us.
5728 list_del_rcu(&wq->list);
5729 mutex_unlock(&wq_pool_mutex);
5732 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5733 * to put the base refs. @wq will be auto-destroyed from the last
5734 * pwq_put. RCU read lock prevents @wq from going away from under us.
5738 for_each_possible_cpu(cpu) {
5739 put_pwq_unlocked(unbound_pwq(wq, cpu));
5740 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5743 put_pwq_unlocked(unbound_pwq(wq, -1));
5744 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5748 EXPORT_SYMBOL_GPL(destroy_workqueue);
5751 * workqueue_set_max_active - adjust max_active of a workqueue
5752 * @wq: target workqueue
5753 * @max_active: new max_active value.
5755 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5759 * Don't call from IRQ context.
5761 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5763 /* max_active doesn't mean anything for BH workqueues */
5764 if (WARN_ON(wq->flags & WQ_BH))
5766 /* disallow meddling with max_active for ordered workqueues */
5767 if (WARN_ON(wq->flags & __WQ_ORDERED))
5770 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5772 mutex_lock(&wq->mutex);
5774 wq->saved_max_active = max_active;
5775 if (wq->flags & WQ_UNBOUND)
5776 wq->saved_min_active = min(wq->saved_min_active, max_active);
5778 wq_adjust_max_active(wq);
5780 mutex_unlock(&wq->mutex);
5782 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5785 * workqueue_set_min_active - adjust min_active of an unbound workqueue
5786 * @wq: target unbound workqueue
5787 * @min_active: new min_active value
5789 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5790 * unbound workqueue is not guaranteed to be able to process max_active
5791 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
5792 * able to process min_active number of interdependent work items which is
5793 * %WQ_DFL_MIN_ACTIVE by default.
5795 * Use this function to adjust the min_active value between 0 and the current
5798 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5800 /* min_active is only meaningful for non-ordered unbound workqueues */
5801 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
5805 mutex_lock(&wq->mutex);
5806 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5807 wq_adjust_max_active(wq);
5808 mutex_unlock(&wq->mutex);
5812 * current_work - retrieve %current task's work struct
5814 * Determine if %current task is a workqueue worker and what it's working on.
5815 * Useful to find out the context that the %current task is running in.
5817 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5819 struct work_struct *current_work(void)
5821 struct worker *worker = current_wq_worker();
5823 return worker ? worker->current_work : NULL;
5825 EXPORT_SYMBOL(current_work);
5828 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5830 * Determine whether %current is a workqueue rescuer. Can be used from
5831 * work functions to determine whether it's being run off the rescuer task.
5833 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5835 bool current_is_workqueue_rescuer(void)
5837 struct worker *worker = current_wq_worker();
5839 return worker && worker->rescue_wq;
5843 * workqueue_congested - test whether a workqueue is congested
5844 * @cpu: CPU in question
5845 * @wq: target workqueue
5847 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5848 * no synchronization around this function and the test result is
5849 * unreliable and only useful as advisory hints or for debugging.
5851 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5853 * With the exception of ordered workqueues, all workqueues have per-cpu
5854 * pool_workqueues, each with its own congested state. A workqueue being
5855 * congested on one CPU doesn't mean that the workqueue is contested on any
5859 * %true if congested, %false otherwise.
5861 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5863 struct pool_workqueue *pwq;
5869 if (cpu == WORK_CPU_UNBOUND)
5870 cpu = smp_processor_id();
5872 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5873 ret = !list_empty(&pwq->inactive_works);
5880 EXPORT_SYMBOL_GPL(workqueue_congested);
5883 * work_busy - test whether a work is currently pending or running
5884 * @work: the work to be tested
5886 * Test whether @work is currently pending or running. There is no
5887 * synchronization around this function and the test result is
5888 * unreliable and only useful as advisory hints or for debugging.
5891 * OR'd bitmask of WORK_BUSY_* bits.
5893 unsigned int work_busy(struct work_struct *work)
5895 struct worker_pool *pool;
5896 unsigned long irq_flags;
5897 unsigned int ret = 0;
5899 if (work_pending(work))
5900 ret |= WORK_BUSY_PENDING;
5903 pool = get_work_pool(work);
5905 raw_spin_lock_irqsave(&pool->lock, irq_flags);
5906 if (find_worker_executing_work(pool, work))
5907 ret |= WORK_BUSY_RUNNING;
5908 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
5914 EXPORT_SYMBOL_GPL(work_busy);
5917 * set_worker_desc - set description for the current work item
5918 * @fmt: printf-style format string
5919 * @...: arguments for the format string
5921 * This function can be called by a running work function to describe what
5922 * the work item is about. If the worker task gets dumped, this
5923 * information will be printed out together to help debugging. The
5924 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5926 void set_worker_desc(const char *fmt, ...)
5928 struct worker *worker = current_wq_worker();
5932 va_start(args, fmt);
5933 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5937 EXPORT_SYMBOL_GPL(set_worker_desc);
5940 * print_worker_info - print out worker information and description
5941 * @log_lvl: the log level to use when printing
5942 * @task: target task
5944 * If @task is a worker and currently executing a work item, print out the
5945 * name of the workqueue being serviced and worker description set with
5946 * set_worker_desc() by the currently executing work item.
5948 * This function can be safely called on any task as long as the
5949 * task_struct itself is accessible. While safe, this function isn't
5950 * synchronized and may print out mixups or garbages of limited length.
5952 void print_worker_info(const char *log_lvl, struct task_struct *task)
5954 work_func_t *fn = NULL;
5955 char name[WQ_NAME_LEN] = { };
5956 char desc[WORKER_DESC_LEN] = { };
5957 struct pool_workqueue *pwq = NULL;
5958 struct workqueue_struct *wq = NULL;
5959 struct worker *worker;
5961 if (!(task->flags & PF_WQ_WORKER))
5965 * This function is called without any synchronization and @task
5966 * could be in any state. Be careful with dereferences.
5968 worker = kthread_probe_data(task);
5971 * Carefully copy the associated workqueue's workfn, name and desc.
5972 * Keep the original last '\0' in case the original is garbage.
5974 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5975 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5976 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5977 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5978 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5980 if (fn || name[0] || desc[0]) {
5981 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5982 if (strcmp(name, desc))
5983 pr_cont(" (%s)", desc);
5988 static void pr_cont_pool_info(struct worker_pool *pool)
5990 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5991 if (pool->node != NUMA_NO_NODE)
5992 pr_cont(" node=%d", pool->node);
5993 pr_cont(" flags=0x%x", pool->flags);
5994 if (pool->flags & POOL_BH)
5996 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
5998 pr_cont(" nice=%d", pool->attrs->nice);
6001 static void pr_cont_worker_id(struct worker *worker)
6003 struct worker_pool *pool = worker->pool;
6005 if (pool->flags & WQ_BH)
6007 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6009 pr_cont("%d%s", task_pid_nr(worker->task),
6010 worker->rescue_wq ? "(RESCUER)" : "");
6013 struct pr_cont_work_struct {
6019 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6023 if (func == pcwsp->func) {
6027 if (pcwsp->ctr == 1)
6028 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6030 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6033 if ((long)func == -1L)
6035 pcwsp->comma = comma;
6040 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6042 if (work->func == wq_barrier_func) {
6043 struct wq_barrier *barr;
6045 barr = container_of(work, struct wq_barrier, work);
6047 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6048 pr_cont("%s BAR(%d)", comma ? "," : "",
6049 task_pid_nr(barr->task));
6052 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6053 pr_cont_work_flush(comma, work->func, pcwsp);
6057 static void show_pwq(struct pool_workqueue *pwq)
6059 struct pr_cont_work_struct pcws = { .ctr = 0, };
6060 struct worker_pool *pool = pwq->pool;
6061 struct work_struct *work;
6062 struct worker *worker;
6063 bool has_in_flight = false, has_pending = false;
6066 pr_info(" pwq %d:", pool->id);
6067 pr_cont_pool_info(pool);
6069 pr_cont(" active=%d refcnt=%d%s\n",
6070 pwq->nr_active, pwq->refcnt,
6071 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6073 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6074 if (worker->current_pwq == pwq) {
6075 has_in_flight = true;
6079 if (has_in_flight) {
6082 pr_info(" in-flight:");
6083 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6084 if (worker->current_pwq != pwq)
6087 pr_cont(" %s", comma ? "," : "");
6088 pr_cont_worker_id(worker);
6089 pr_cont(":%ps", worker->current_func);
6090 list_for_each_entry(work, &worker->scheduled, entry)
6091 pr_cont_work(false, work, &pcws);
6092 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6098 list_for_each_entry(work, &pool->worklist, entry) {
6099 if (get_work_pwq(work) == pwq) {
6107 pr_info(" pending:");
6108 list_for_each_entry(work, &pool->worklist, entry) {
6109 if (get_work_pwq(work) != pwq)
6112 pr_cont_work(comma, work, &pcws);
6113 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6115 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6119 if (!list_empty(&pwq->inactive_works)) {
6122 pr_info(" inactive:");
6123 list_for_each_entry(work, &pwq->inactive_works, entry) {
6124 pr_cont_work(comma, work, &pcws);
6125 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6127 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6133 * show_one_workqueue - dump state of specified workqueue
6134 * @wq: workqueue whose state will be printed
6136 void show_one_workqueue(struct workqueue_struct *wq)
6138 struct pool_workqueue *pwq;
6140 unsigned long irq_flags;
6142 for_each_pwq(pwq, wq) {
6143 if (!pwq_is_empty(pwq)) {
6148 if (idle) /* Nothing to print for idle workqueue */
6151 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6153 for_each_pwq(pwq, wq) {
6154 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6155 if (!pwq_is_empty(pwq)) {
6157 * Defer printing to avoid deadlocks in console
6158 * drivers that queue work while holding locks
6159 * also taken in their write paths.
6161 printk_deferred_enter();
6163 printk_deferred_exit();
6165 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6167 * We could be printing a lot from atomic context, e.g.
6168 * sysrq-t -> show_all_workqueues(). Avoid triggering
6171 touch_nmi_watchdog();
6177 * show_one_worker_pool - dump state of specified worker pool
6178 * @pool: worker pool whose state will be printed
6180 static void show_one_worker_pool(struct worker_pool *pool)
6182 struct worker *worker;
6184 unsigned long irq_flags;
6185 unsigned long hung = 0;
6187 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6188 if (pool->nr_workers == pool->nr_idle)
6191 /* How long the first pending work is waiting for a worker. */
6192 if (!list_empty(&pool->worklist))
6193 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6196 * Defer printing to avoid deadlocks in console drivers that
6197 * queue work while holding locks also taken in their write
6200 printk_deferred_enter();
6201 pr_info("pool %d:", pool->id);
6202 pr_cont_pool_info(pool);
6203 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6205 pr_cont(" manager: %d",
6206 task_pid_nr(pool->manager->task));
6207 list_for_each_entry(worker, &pool->idle_list, entry) {
6208 pr_cont(" %s", first ? "idle: " : "");
6209 pr_cont_worker_id(worker);
6213 printk_deferred_exit();
6215 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6217 * We could be printing a lot from atomic context, e.g.
6218 * sysrq-t -> show_all_workqueues(). Avoid triggering
6221 touch_nmi_watchdog();
6226 * show_all_workqueues - dump workqueue state
6228 * Called from a sysrq handler and prints out all busy workqueues and pools.
6230 void show_all_workqueues(void)
6232 struct workqueue_struct *wq;
6233 struct worker_pool *pool;
6238 pr_info("Showing busy workqueues and worker pools:\n");
6240 list_for_each_entry_rcu(wq, &workqueues, list)
6241 show_one_workqueue(wq);
6243 for_each_pool(pool, pi)
6244 show_one_worker_pool(pool);
6250 * show_freezable_workqueues - dump freezable workqueue state
6252 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6255 void show_freezable_workqueues(void)
6257 struct workqueue_struct *wq;
6261 pr_info("Showing freezable workqueues that are still busy:\n");
6263 list_for_each_entry_rcu(wq, &workqueues, list) {
6264 if (!(wq->flags & WQ_FREEZABLE))
6266 show_one_workqueue(wq);
6272 /* used to show worker information through /proc/PID/{comm,stat,status} */
6273 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6277 /* always show the actual comm */
6278 off = strscpy(buf, task->comm, size);
6282 /* stabilize PF_WQ_WORKER and worker pool association */
6283 mutex_lock(&wq_pool_attach_mutex);
6285 if (task->flags & PF_WQ_WORKER) {
6286 struct worker *worker = kthread_data(task);
6287 struct worker_pool *pool = worker->pool;
6290 raw_spin_lock_irq(&pool->lock);
6292 * ->desc tracks information (wq name or
6293 * set_worker_desc()) for the latest execution. If
6294 * current, prepend '+', otherwise '-'.
6296 if (worker->desc[0] != '\0') {
6297 if (worker->current_work)
6298 scnprintf(buf + off, size - off, "+%s",
6301 scnprintf(buf + off, size - off, "-%s",
6304 raw_spin_unlock_irq(&pool->lock);
6308 mutex_unlock(&wq_pool_attach_mutex);
6316 * There are two challenges in supporting CPU hotplug. Firstly, there
6317 * are a lot of assumptions on strong associations among work, pwq and
6318 * pool which make migrating pending and scheduled works very
6319 * difficult to implement without impacting hot paths. Secondly,
6320 * worker pools serve mix of short, long and very long running works making
6321 * blocked draining impractical.
6323 * This is solved by allowing the pools to be disassociated from the CPU
6324 * running as an unbound one and allowing it to be reattached later if the
6325 * cpu comes back online.
6328 static void unbind_workers(int cpu)
6330 struct worker_pool *pool;
6331 struct worker *worker;
6333 for_each_cpu_worker_pool(pool, cpu) {
6334 mutex_lock(&wq_pool_attach_mutex);
6335 raw_spin_lock_irq(&pool->lock);
6338 * We've blocked all attach/detach operations. Make all workers
6339 * unbound and set DISASSOCIATED. Before this, all workers
6340 * must be on the cpu. After this, they may become diasporas.
6341 * And the preemption disabled section in their sched callbacks
6342 * are guaranteed to see WORKER_UNBOUND since the code here
6343 * is on the same cpu.
6345 for_each_pool_worker(worker, pool)
6346 worker->flags |= WORKER_UNBOUND;
6348 pool->flags |= POOL_DISASSOCIATED;
6351 * The handling of nr_running in sched callbacks are disabled
6352 * now. Zap nr_running. After this, nr_running stays zero and
6353 * need_more_worker() and keep_working() are always true as
6354 * long as the worklist is not empty. This pool now behaves as
6355 * an unbound (in terms of concurrency management) pool which
6356 * are served by workers tied to the pool.
6358 pool->nr_running = 0;
6361 * With concurrency management just turned off, a busy
6362 * worker blocking could lead to lengthy stalls. Kick off
6363 * unbound chain execution of currently pending work items.
6367 raw_spin_unlock_irq(&pool->lock);
6369 for_each_pool_worker(worker, pool)
6370 unbind_worker(worker);
6372 mutex_unlock(&wq_pool_attach_mutex);
6377 * rebind_workers - rebind all workers of a pool to the associated CPU
6378 * @pool: pool of interest
6380 * @pool->cpu is coming online. Rebind all workers to the CPU.
6382 static void rebind_workers(struct worker_pool *pool)
6384 struct worker *worker;
6386 lockdep_assert_held(&wq_pool_attach_mutex);
6389 * Restore CPU affinity of all workers. As all idle workers should
6390 * be on the run-queue of the associated CPU before any local
6391 * wake-ups for concurrency management happen, restore CPU affinity
6392 * of all workers first and then clear UNBOUND. As we're called
6393 * from CPU_ONLINE, the following shouldn't fail.
6395 for_each_pool_worker(worker, pool) {
6396 kthread_set_per_cpu(worker->task, pool->cpu);
6397 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6398 pool_allowed_cpus(pool)) < 0);
6401 raw_spin_lock_irq(&pool->lock);
6403 pool->flags &= ~POOL_DISASSOCIATED;
6405 for_each_pool_worker(worker, pool) {
6406 unsigned int worker_flags = worker->flags;
6409 * We want to clear UNBOUND but can't directly call
6410 * worker_clr_flags() or adjust nr_running. Atomically
6411 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6412 * @worker will clear REBOUND using worker_clr_flags() when
6413 * it initiates the next execution cycle thus restoring
6414 * concurrency management. Note that when or whether
6415 * @worker clears REBOUND doesn't affect correctness.
6417 * WRITE_ONCE() is necessary because @worker->flags may be
6418 * tested without holding any lock in
6419 * wq_worker_running(). Without it, NOT_RUNNING test may
6420 * fail incorrectly leading to premature concurrency
6421 * management operations.
6423 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6424 worker_flags |= WORKER_REBOUND;
6425 worker_flags &= ~WORKER_UNBOUND;
6426 WRITE_ONCE(worker->flags, worker_flags);
6429 raw_spin_unlock_irq(&pool->lock);
6433 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6434 * @pool: unbound pool of interest
6435 * @cpu: the CPU which is coming up
6437 * An unbound pool may end up with a cpumask which doesn't have any online
6438 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6439 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6440 * online CPU before, cpus_allowed of all its workers should be restored.
6442 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6444 static cpumask_t cpumask;
6445 struct worker *worker;
6447 lockdep_assert_held(&wq_pool_attach_mutex);
6449 /* is @cpu allowed for @pool? */
6450 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6453 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6455 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6456 for_each_pool_worker(worker, pool)
6457 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6460 int workqueue_prepare_cpu(unsigned int cpu)
6462 struct worker_pool *pool;
6464 for_each_cpu_worker_pool(pool, cpu) {
6465 if (pool->nr_workers)
6467 if (!create_worker(pool))
6473 int workqueue_online_cpu(unsigned int cpu)
6475 struct worker_pool *pool;
6476 struct workqueue_struct *wq;
6479 mutex_lock(&wq_pool_mutex);
6481 for_each_pool(pool, pi) {
6482 /* BH pools aren't affected by hotplug */
6483 if (pool->flags & POOL_BH)
6486 mutex_lock(&wq_pool_attach_mutex);
6487 if (pool->cpu == cpu)
6488 rebind_workers(pool);
6489 else if (pool->cpu < 0)
6490 restore_unbound_workers_cpumask(pool, cpu);
6491 mutex_unlock(&wq_pool_attach_mutex);
6494 /* update pod affinity of unbound workqueues */
6495 list_for_each_entry(wq, &workqueues, list) {
6496 struct workqueue_attrs *attrs = wq->unbound_attrs;
6499 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6502 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6503 wq_update_pod(wq, tcpu, cpu, true);
6505 mutex_lock(&wq->mutex);
6506 wq_update_node_max_active(wq, -1);
6507 mutex_unlock(&wq->mutex);
6511 mutex_unlock(&wq_pool_mutex);
6515 int workqueue_offline_cpu(unsigned int cpu)
6517 struct workqueue_struct *wq;
6519 /* unbinding per-cpu workers should happen on the local CPU */
6520 if (WARN_ON(cpu != smp_processor_id()))
6523 unbind_workers(cpu);
6525 /* update pod affinity of unbound workqueues */
6526 mutex_lock(&wq_pool_mutex);
6527 list_for_each_entry(wq, &workqueues, list) {
6528 struct workqueue_attrs *attrs = wq->unbound_attrs;
6531 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6534 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6535 wq_update_pod(wq, tcpu, cpu, false);
6537 mutex_lock(&wq->mutex);
6538 wq_update_node_max_active(wq, cpu);
6539 mutex_unlock(&wq->mutex);
6542 mutex_unlock(&wq_pool_mutex);
6547 struct work_for_cpu {
6548 struct work_struct work;
6554 static void work_for_cpu_fn(struct work_struct *work)
6556 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6558 wfc->ret = wfc->fn(wfc->arg);
6562 * work_on_cpu_key - run a function in thread context on a particular cpu
6563 * @cpu: the cpu to run on
6564 * @fn: the function to run
6565 * @arg: the function arg
6566 * @key: The lock class key for lock debugging purposes
6568 * It is up to the caller to ensure that the cpu doesn't go offline.
6569 * The caller must not hold any locks which would prevent @fn from completing.
6571 * Return: The value @fn returns.
6573 long work_on_cpu_key(int cpu, long (*fn)(void *),
6574 void *arg, struct lock_class_key *key)
6576 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6578 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6579 schedule_work_on(cpu, &wfc.work);
6580 flush_work(&wfc.work);
6581 destroy_work_on_stack(&wfc.work);
6584 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6587 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6588 * @cpu: the cpu to run on
6589 * @fn: the function to run
6590 * @arg: the function argument
6591 * @key: The lock class key for lock debugging purposes
6593 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6594 * any locks which would prevent @fn from completing.
6596 * Return: The value @fn returns.
6598 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6599 void *arg, struct lock_class_key *key)
6604 if (cpu_online(cpu))
6605 ret = work_on_cpu_key(cpu, fn, arg, key);
6609 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6610 #endif /* CONFIG_SMP */
6612 #ifdef CONFIG_FREEZER
6615 * freeze_workqueues_begin - begin freezing workqueues
6617 * Start freezing workqueues. After this function returns, all freezable
6618 * workqueues will queue new works to their inactive_works list instead of
6622 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6624 void freeze_workqueues_begin(void)
6626 struct workqueue_struct *wq;
6628 mutex_lock(&wq_pool_mutex);
6630 WARN_ON_ONCE(workqueue_freezing);
6631 workqueue_freezing = true;
6633 list_for_each_entry(wq, &workqueues, list) {
6634 mutex_lock(&wq->mutex);
6635 wq_adjust_max_active(wq);
6636 mutex_unlock(&wq->mutex);
6639 mutex_unlock(&wq_pool_mutex);
6643 * freeze_workqueues_busy - are freezable workqueues still busy?
6645 * Check whether freezing is complete. This function must be called
6646 * between freeze_workqueues_begin() and thaw_workqueues().
6649 * Grabs and releases wq_pool_mutex.
6652 * %true if some freezable workqueues are still busy. %false if freezing
6655 bool freeze_workqueues_busy(void)
6658 struct workqueue_struct *wq;
6659 struct pool_workqueue *pwq;
6661 mutex_lock(&wq_pool_mutex);
6663 WARN_ON_ONCE(!workqueue_freezing);
6665 list_for_each_entry(wq, &workqueues, list) {
6666 if (!(wq->flags & WQ_FREEZABLE))
6669 * nr_active is monotonically decreasing. It's safe
6670 * to peek without lock.
6673 for_each_pwq(pwq, wq) {
6674 WARN_ON_ONCE(pwq->nr_active < 0);
6675 if (pwq->nr_active) {
6684 mutex_unlock(&wq_pool_mutex);
6689 * thaw_workqueues - thaw workqueues
6691 * Thaw workqueues. Normal queueing is restored and all collected
6692 * frozen works are transferred to their respective pool worklists.
6695 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6697 void thaw_workqueues(void)
6699 struct workqueue_struct *wq;
6701 mutex_lock(&wq_pool_mutex);
6703 if (!workqueue_freezing)
6706 workqueue_freezing = false;
6708 /* restore max_active and repopulate worklist */
6709 list_for_each_entry(wq, &workqueues, list) {
6710 mutex_lock(&wq->mutex);
6711 wq_adjust_max_active(wq);
6712 mutex_unlock(&wq->mutex);
6716 mutex_unlock(&wq_pool_mutex);
6718 #endif /* CONFIG_FREEZER */
6720 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6724 struct workqueue_struct *wq;
6725 struct apply_wqattrs_ctx *ctx, *n;
6727 lockdep_assert_held(&wq_pool_mutex);
6729 list_for_each_entry(wq, &workqueues, list) {
6730 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6733 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6739 list_add_tail(&ctx->list, &ctxs);
6742 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6744 apply_wqattrs_commit(ctx);
6745 apply_wqattrs_cleanup(ctx);
6749 mutex_lock(&wq_pool_attach_mutex);
6750 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6751 mutex_unlock(&wq_pool_attach_mutex);
6757 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6758 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6760 * This function can be called from cpuset code to provide a set of isolated
6761 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
6762 * either cpus_read_lock or cpus_write_lock.
6764 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6766 cpumask_var_t cpumask;
6769 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6772 lockdep_assert_cpus_held();
6773 mutex_lock(&wq_pool_mutex);
6775 /* Save the current isolated cpumask & export it via sysfs */
6776 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6779 * If the operation fails, it will fall back to
6780 * wq_requested_unbound_cpumask which is initially set to
6781 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6782 * by any subsequent write to workqueue/cpumask sysfs file.
6784 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6785 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6786 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6787 ret = workqueue_apply_unbound_cpumask(cpumask);
6789 mutex_unlock(&wq_pool_mutex);
6790 free_cpumask_var(cpumask);
6794 static int parse_affn_scope(const char *val)
6798 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6799 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6805 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6807 struct workqueue_struct *wq;
6810 affn = parse_affn_scope(val);
6813 if (affn == WQ_AFFN_DFL)
6817 mutex_lock(&wq_pool_mutex);
6821 list_for_each_entry(wq, &workqueues, list) {
6822 for_each_online_cpu(cpu) {
6823 wq_update_pod(wq, cpu, cpu, true);
6827 mutex_unlock(&wq_pool_mutex);
6833 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6835 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6838 static const struct kernel_param_ops wq_affn_dfl_ops = {
6839 .set = wq_affn_dfl_set,
6840 .get = wq_affn_dfl_get,
6843 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6847 * Workqueues with WQ_SYSFS flag set is visible to userland via
6848 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6849 * following attributes.
6851 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6852 * max_active RW int : maximum number of in-flight work items
6854 * Unbound workqueues have the following extra attributes.
6856 * nice RW int : nice value of the workers
6857 * cpumask RW mask : bitmask of allowed CPUs for the workers
6858 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6859 * affinity_strict RW bool : worker CPU affinity is strict
6862 struct workqueue_struct *wq;
6866 static struct workqueue_struct *dev_to_wq(struct device *dev)
6868 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6873 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6876 struct workqueue_struct *wq = dev_to_wq(dev);
6878 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6880 static DEVICE_ATTR_RO(per_cpu);
6882 static ssize_t max_active_show(struct device *dev,
6883 struct device_attribute *attr, char *buf)
6885 struct workqueue_struct *wq = dev_to_wq(dev);
6887 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6890 static ssize_t max_active_store(struct device *dev,
6891 struct device_attribute *attr, const char *buf,
6894 struct workqueue_struct *wq = dev_to_wq(dev);
6897 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
6900 workqueue_set_max_active(wq, val);
6903 static DEVICE_ATTR_RW(max_active);
6905 static struct attribute *wq_sysfs_attrs[] = {
6906 &dev_attr_per_cpu.attr,
6907 &dev_attr_max_active.attr,
6910 ATTRIBUTE_GROUPS(wq_sysfs);
6912 static void apply_wqattrs_lock(void)
6914 /* CPUs should stay stable across pwq creations and installations */
6916 mutex_lock(&wq_pool_mutex);
6919 static void apply_wqattrs_unlock(void)
6921 mutex_unlock(&wq_pool_mutex);
6925 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
6928 struct workqueue_struct *wq = dev_to_wq(dev);
6931 mutex_lock(&wq->mutex);
6932 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
6933 mutex_unlock(&wq->mutex);
6938 /* prepare workqueue_attrs for sysfs store operations */
6939 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
6941 struct workqueue_attrs *attrs;
6943 lockdep_assert_held(&wq_pool_mutex);
6945 attrs = alloc_workqueue_attrs();
6949 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6953 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6954 const char *buf, size_t count)
6956 struct workqueue_struct *wq = dev_to_wq(dev);
6957 struct workqueue_attrs *attrs;
6960 apply_wqattrs_lock();
6962 attrs = wq_sysfs_prep_attrs(wq);
6966 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6967 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6968 ret = apply_workqueue_attrs_locked(wq, attrs);
6973 apply_wqattrs_unlock();
6974 free_workqueue_attrs(attrs);
6975 return ret ?: count;
6978 static ssize_t wq_cpumask_show(struct device *dev,
6979 struct device_attribute *attr, char *buf)
6981 struct workqueue_struct *wq = dev_to_wq(dev);
6984 mutex_lock(&wq->mutex);
6985 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6986 cpumask_pr_args(wq->unbound_attrs->cpumask));
6987 mutex_unlock(&wq->mutex);
6991 static ssize_t wq_cpumask_store(struct device *dev,
6992 struct device_attribute *attr,
6993 const char *buf, size_t count)
6995 struct workqueue_struct *wq = dev_to_wq(dev);
6996 struct workqueue_attrs *attrs;
6999 apply_wqattrs_lock();
7001 attrs = wq_sysfs_prep_attrs(wq);
7005 ret = cpumask_parse(buf, attrs->cpumask);
7007 ret = apply_workqueue_attrs_locked(wq, attrs);
7010 apply_wqattrs_unlock();
7011 free_workqueue_attrs(attrs);
7012 return ret ?: count;
7015 static ssize_t wq_affn_scope_show(struct device *dev,
7016 struct device_attribute *attr, char *buf)
7018 struct workqueue_struct *wq = dev_to_wq(dev);
7021 mutex_lock(&wq->mutex);
7022 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7023 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7024 wq_affn_names[WQ_AFFN_DFL],
7025 wq_affn_names[wq_affn_dfl]);
7027 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7028 wq_affn_names[wq->unbound_attrs->affn_scope]);
7029 mutex_unlock(&wq->mutex);
7034 static ssize_t wq_affn_scope_store(struct device *dev,
7035 struct device_attribute *attr,
7036 const char *buf, size_t count)
7038 struct workqueue_struct *wq = dev_to_wq(dev);
7039 struct workqueue_attrs *attrs;
7040 int affn, ret = -ENOMEM;
7042 affn = parse_affn_scope(buf);
7046 apply_wqattrs_lock();
7047 attrs = wq_sysfs_prep_attrs(wq);
7049 attrs->affn_scope = affn;
7050 ret = apply_workqueue_attrs_locked(wq, attrs);
7052 apply_wqattrs_unlock();
7053 free_workqueue_attrs(attrs);
7054 return ret ?: count;
7057 static ssize_t wq_affinity_strict_show(struct device *dev,
7058 struct device_attribute *attr, char *buf)
7060 struct workqueue_struct *wq = dev_to_wq(dev);
7062 return scnprintf(buf, PAGE_SIZE, "%d\n",
7063 wq->unbound_attrs->affn_strict);
7066 static ssize_t wq_affinity_strict_store(struct device *dev,
7067 struct device_attribute *attr,
7068 const char *buf, size_t count)
7070 struct workqueue_struct *wq = dev_to_wq(dev);
7071 struct workqueue_attrs *attrs;
7072 int v, ret = -ENOMEM;
7074 if (sscanf(buf, "%d", &v) != 1)
7077 apply_wqattrs_lock();
7078 attrs = wq_sysfs_prep_attrs(wq);
7080 attrs->affn_strict = (bool)v;
7081 ret = apply_workqueue_attrs_locked(wq, attrs);
7083 apply_wqattrs_unlock();
7084 free_workqueue_attrs(attrs);
7085 return ret ?: count;
7088 static struct device_attribute wq_sysfs_unbound_attrs[] = {
7089 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7090 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7091 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7092 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7096 static const struct bus_type wq_subsys = {
7097 .name = "workqueue",
7098 .dev_groups = wq_sysfs_groups,
7102 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7103 * @cpumask: the cpumask to set
7105 * The low-level workqueues cpumask is a global cpumask that limits
7106 * the affinity of all unbound workqueues. This function check the @cpumask
7107 * and apply it to all unbound workqueues and updates all pwqs of them.
7109 * Return: 0 - Success
7110 * -EINVAL - Invalid @cpumask
7111 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7113 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7118 * Not excluding isolated cpus on purpose.
7119 * If the user wishes to include them, we allow that.
7121 cpumask_and(cpumask, cpumask, cpu_possible_mask);
7122 if (!cpumask_empty(cpumask)) {
7123 apply_wqattrs_lock();
7124 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7125 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
7130 ret = workqueue_apply_unbound_cpumask(cpumask);
7133 apply_wqattrs_unlock();
7139 static ssize_t __wq_cpumask_show(struct device *dev,
7140 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7144 mutex_lock(&wq_pool_mutex);
7145 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7146 mutex_unlock(&wq_pool_mutex);
7151 static ssize_t wq_unbound_cpumask_show(struct device *dev,
7152 struct device_attribute *attr, char *buf)
7154 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7157 static ssize_t wq_requested_cpumask_show(struct device *dev,
7158 struct device_attribute *attr, char *buf)
7160 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7163 static ssize_t wq_isolated_cpumask_show(struct device *dev,
7164 struct device_attribute *attr, char *buf)
7166 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7169 static ssize_t wq_unbound_cpumask_store(struct device *dev,
7170 struct device_attribute *attr, const char *buf, size_t count)
7172 cpumask_var_t cpumask;
7175 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7178 ret = cpumask_parse(buf, cpumask);
7180 ret = workqueue_set_unbound_cpumask(cpumask);
7182 free_cpumask_var(cpumask);
7183 return ret ? ret : count;
7186 static struct device_attribute wq_sysfs_cpumask_attrs[] = {
7187 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
7188 wq_unbound_cpumask_store),
7189 __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
7190 __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
7194 static int __init wq_sysfs_init(void)
7196 struct device *dev_root;
7199 err = subsys_virtual_register(&wq_subsys, NULL);
7203 dev_root = bus_get_dev_root(&wq_subsys);
7205 struct device_attribute *attr;
7207 for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
7208 err = device_create_file(dev_root, attr);
7212 put_device(dev_root);
7216 core_initcall(wq_sysfs_init);
7218 static void wq_device_release(struct device *dev)
7220 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7226 * workqueue_sysfs_register - make a workqueue visible in sysfs
7227 * @wq: the workqueue to register
7229 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7230 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7231 * which is the preferred method.
7233 * Workqueue user should use this function directly iff it wants to apply
7234 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7235 * apply_workqueue_attrs() may race against userland updating the
7238 * Return: 0 on success, -errno on failure.
7240 int workqueue_sysfs_register(struct workqueue_struct *wq)
7242 struct wq_device *wq_dev;
7246 * Adjusting max_active breaks ordering guarantee. Disallow exposing
7247 * ordered workqueues.
7249 if (WARN_ON(wq->flags & __WQ_ORDERED))
7252 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7257 wq_dev->dev.bus = &wq_subsys;
7258 wq_dev->dev.release = wq_device_release;
7259 dev_set_name(&wq_dev->dev, "%s", wq->name);
7262 * unbound_attrs are created separately. Suppress uevent until
7263 * everything is ready.
7265 dev_set_uevent_suppress(&wq_dev->dev, true);
7267 ret = device_register(&wq_dev->dev);
7269 put_device(&wq_dev->dev);
7274 if (wq->flags & WQ_UNBOUND) {
7275 struct device_attribute *attr;
7277 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7278 ret = device_create_file(&wq_dev->dev, attr);
7280 device_unregister(&wq_dev->dev);
7287 dev_set_uevent_suppress(&wq_dev->dev, false);
7288 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7293 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7294 * @wq: the workqueue to unregister
7296 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7298 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7300 struct wq_device *wq_dev = wq->wq_dev;
7306 device_unregister(&wq_dev->dev);
7308 #else /* CONFIG_SYSFS */
7309 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7310 #endif /* CONFIG_SYSFS */
7313 * Workqueue watchdog.
7315 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7316 * flush dependency, a concurrency managed work item which stays RUNNING
7317 * indefinitely. Workqueue stalls can be very difficult to debug as the
7318 * usual warning mechanisms don't trigger and internal workqueue state is
7321 * Workqueue watchdog monitors all worker pools periodically and dumps
7322 * state if some pools failed to make forward progress for a while where
7323 * forward progress is defined as the first item on ->worklist changing.
7325 * This mechanism is controlled through the kernel parameter
7326 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7327 * corresponding sysfs parameter file.
7329 #ifdef CONFIG_WQ_WATCHDOG
7331 static unsigned long wq_watchdog_thresh = 30;
7332 static struct timer_list wq_watchdog_timer;
7334 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7335 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7338 * Show workers that might prevent the processing of pending work items.
7339 * The only candidates are CPU-bound workers in the running state.
7340 * Pending work items should be handled by another idle worker
7341 * in all other situations.
7343 static void show_cpu_pool_hog(struct worker_pool *pool)
7345 struct worker *worker;
7346 unsigned long irq_flags;
7349 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7351 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7352 if (task_is_running(worker->task)) {
7354 * Defer printing to avoid deadlocks in console
7355 * drivers that queue work while holding locks
7356 * also taken in their write paths.
7358 printk_deferred_enter();
7360 pr_info("pool %d:\n", pool->id);
7361 sched_show_task(worker->task);
7363 printk_deferred_exit();
7367 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7370 static void show_cpu_pools_hogs(void)
7372 struct worker_pool *pool;
7375 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7379 for_each_pool(pool, pi) {
7380 if (pool->cpu_stall)
7381 show_cpu_pool_hog(pool);
7388 static void wq_watchdog_reset_touched(void)
7392 wq_watchdog_touched = jiffies;
7393 for_each_possible_cpu(cpu)
7394 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7397 static void wq_watchdog_timer_fn(struct timer_list *unused)
7399 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7400 bool lockup_detected = false;
7401 bool cpu_pool_stall = false;
7402 unsigned long now = jiffies;
7403 struct worker_pool *pool;
7411 for_each_pool(pool, pi) {
7412 unsigned long pool_ts, touched, ts;
7414 pool->cpu_stall = false;
7415 if (list_empty(&pool->worklist))
7419 * If a virtual machine is stopped by the host it can look to
7420 * the watchdog like a stall.
7422 kvm_check_and_clear_guest_paused();
7424 /* get the latest of pool and touched timestamps */
7426 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7428 touched = READ_ONCE(wq_watchdog_touched);
7429 pool_ts = READ_ONCE(pool->watchdog_ts);
7431 if (time_after(pool_ts, touched))
7437 if (time_after(now, ts + thresh)) {
7438 lockup_detected = true;
7439 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7440 pool->cpu_stall = true;
7441 cpu_pool_stall = true;
7443 pr_emerg("BUG: workqueue lockup - pool");
7444 pr_cont_pool_info(pool);
7445 pr_cont(" stuck for %us!\n",
7446 jiffies_to_msecs(now - pool_ts) / 1000);
7454 if (lockup_detected)
7455 show_all_workqueues();
7458 show_cpu_pools_hogs();
7460 wq_watchdog_reset_touched();
7461 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7464 notrace void wq_watchdog_touch(int cpu)
7467 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7469 wq_watchdog_touched = jiffies;
7472 static void wq_watchdog_set_thresh(unsigned long thresh)
7474 wq_watchdog_thresh = 0;
7475 del_timer_sync(&wq_watchdog_timer);
7478 wq_watchdog_thresh = thresh;
7479 wq_watchdog_reset_touched();
7480 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7484 static int wq_watchdog_param_set_thresh(const char *val,
7485 const struct kernel_param *kp)
7487 unsigned long thresh;
7490 ret = kstrtoul(val, 0, &thresh);
7495 wq_watchdog_set_thresh(thresh);
7497 wq_watchdog_thresh = thresh;
7502 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7503 .set = wq_watchdog_param_set_thresh,
7504 .get = param_get_ulong,
7507 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7510 static void wq_watchdog_init(void)
7512 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7513 wq_watchdog_set_thresh(wq_watchdog_thresh);
7516 #else /* CONFIG_WQ_WATCHDOG */
7518 static inline void wq_watchdog_init(void) { }
7520 #endif /* CONFIG_WQ_WATCHDOG */
7522 static void bh_pool_kick_normal(struct irq_work *irq_work)
7524 raise_softirq_irqoff(TASKLET_SOFTIRQ);
7527 static void bh_pool_kick_highpri(struct irq_work *irq_work)
7529 raise_softirq_irqoff(HI_SOFTIRQ);
7532 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7534 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7535 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7536 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7540 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7543 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7545 BUG_ON(init_worker_pool(pool));
7547 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7548 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7549 pool->attrs->nice = nice;
7550 pool->attrs->affn_strict = true;
7551 pool->node = cpu_to_node(cpu);
7554 mutex_lock(&wq_pool_mutex);
7555 BUG_ON(worker_pool_assign_id(pool));
7556 mutex_unlock(&wq_pool_mutex);
7560 * workqueue_init_early - early init for workqueue subsystem
7562 * This is the first step of three-staged workqueue subsystem initialization and
7563 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7564 * up. It sets up all the data structures and system workqueues and allows early
7565 * boot code to create workqueues and queue/cancel work items. Actual work item
7566 * execution starts only after kthreads can be created and scheduled right
7567 * before early initcalls.
7569 void __init workqueue_init_early(void)
7571 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7572 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7573 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7574 bh_pool_kick_highpri };
7577 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7579 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7580 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7581 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7583 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7584 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7585 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7586 if (!cpumask_empty(&wq_cmdline_cpumask))
7587 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7589 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7591 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7593 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7594 BUG_ON(!wq_update_pod_attrs_buf);
7597 * If nohz_full is enabled, set power efficient workqueue as unbound.
7598 * This allows workqueue items to be moved to HK CPUs.
7600 if (housekeeping_enabled(HK_TYPE_TICK))
7601 wq_power_efficient = true;
7603 /* initialize WQ_AFFN_SYSTEM pods */
7604 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7605 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7606 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7607 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7609 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7612 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7613 pt->pod_node[0] = NUMA_NO_NODE;
7616 /* initialize BH and CPU pools */
7617 for_each_possible_cpu(cpu) {
7618 struct worker_pool *pool;
7621 for_each_bh_worker_pool(pool, cpu) {
7622 init_cpu_worker_pool(pool, cpu, std_nice[i]);
7623 pool->flags |= POOL_BH;
7624 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7629 for_each_cpu_worker_pool(pool, cpu)
7630 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7633 /* create default unbound and ordered wq attrs */
7634 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7635 struct workqueue_attrs *attrs;
7637 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7638 attrs->nice = std_nice[i];
7639 unbound_std_wq_attrs[i] = attrs;
7642 * An ordered wq should have only one pwq as ordering is
7643 * guaranteed by max_active which is enforced by pwqs.
7645 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7646 attrs->nice = std_nice[i];
7647 attrs->ordered = true;
7648 ordered_wq_attrs[i] = attrs;
7651 system_wq = alloc_workqueue("events", 0, 0);
7652 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7653 system_long_wq = alloc_workqueue("events_long", 0, 0);
7654 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7656 system_freezable_wq = alloc_workqueue("events_freezable",
7658 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7659 WQ_POWER_EFFICIENT, 0);
7660 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7661 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7663 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7664 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7665 WQ_BH | WQ_HIGHPRI, 0);
7666 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7667 !system_unbound_wq || !system_freezable_wq ||
7668 !system_power_efficient_wq ||
7669 !system_freezable_power_efficient_wq ||
7670 !system_bh_wq || !system_bh_highpri_wq);
7673 static void __init wq_cpu_intensive_thresh_init(void)
7675 unsigned long thresh;
7678 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7679 BUG_ON(IS_ERR(pwq_release_worker));
7681 /* if the user set it to a specific value, keep it */
7682 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7686 * The default of 10ms is derived from the fact that most modern (as of
7687 * 2023) processors can do a lot in 10ms and that it's just below what
7688 * most consider human-perceivable. However, the kernel also runs on a
7689 * lot slower CPUs including microcontrollers where the threshold is way
7692 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7693 * This is by no means accurate but it doesn't have to be. The mechanism
7694 * is still useful even when the threshold is fully scaled up. Also, as
7695 * the reports would usually be applicable to everyone, some machines
7696 * operating on longer thresholds won't significantly diminish their
7699 thresh = 10 * USEC_PER_MSEC;
7701 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7702 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7704 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7706 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7707 loops_per_jiffy, bogo, thresh);
7709 wq_cpu_intensive_thresh_us = thresh;
7713 * workqueue_init - bring workqueue subsystem fully online
7715 * This is the second step of three-staged workqueue subsystem initialization
7716 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7717 * been created and work items queued on them, but there are no kworkers
7718 * executing the work items yet. Populate the worker pools with the initial
7719 * workers and enable future kworker creations.
7721 void __init workqueue_init(void)
7723 struct workqueue_struct *wq;
7724 struct worker_pool *pool;
7727 wq_cpu_intensive_thresh_init();
7729 mutex_lock(&wq_pool_mutex);
7732 * Per-cpu pools created earlier could be missing node hint. Fix them
7733 * up. Also, create a rescuer for workqueues that requested it.
7735 for_each_possible_cpu(cpu) {
7736 for_each_bh_worker_pool(pool, cpu)
7737 pool->node = cpu_to_node(cpu);
7738 for_each_cpu_worker_pool(pool, cpu)
7739 pool->node = cpu_to_node(cpu);
7742 list_for_each_entry(wq, &workqueues, list) {
7743 WARN(init_rescuer(wq),
7744 "workqueue: failed to create early rescuer for %s",
7748 mutex_unlock(&wq_pool_mutex);
7751 * Create the initial workers. A BH pool has one pseudo worker that
7752 * represents the shared BH execution context and thus doesn't get
7753 * affected by hotplug events. Create the BH pseudo workers for all
7754 * possible CPUs here.
7756 for_each_possible_cpu(cpu)
7757 for_each_bh_worker_pool(pool, cpu)
7758 BUG_ON(!create_worker(pool));
7760 for_each_online_cpu(cpu) {
7761 for_each_cpu_worker_pool(pool, cpu) {
7762 pool->flags &= ~POOL_DISASSOCIATED;
7763 BUG_ON(!create_worker(pool));
7767 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7768 BUG_ON(!create_worker(pool));
7775 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7776 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7777 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7779 static void __init init_pod_type(struct wq_pod_type *pt,
7780 bool (*cpus_share_pod)(int, int))
7782 int cur, pre, cpu, pod;
7786 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7787 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7788 BUG_ON(!pt->cpu_pod);
7790 for_each_possible_cpu(cur) {
7791 for_each_possible_cpu(pre) {
7793 pt->cpu_pod[cur] = pt->nr_pods++;
7796 if (cpus_share_pod(cur, pre)) {
7797 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7803 /* init the rest to match @pt->cpu_pod[] */
7804 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7805 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7806 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7808 for (pod = 0; pod < pt->nr_pods; pod++)
7809 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7811 for_each_possible_cpu(cpu) {
7812 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7813 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7817 static bool __init cpus_dont_share(int cpu0, int cpu1)
7822 static bool __init cpus_share_smt(int cpu0, int cpu1)
7824 #ifdef CONFIG_SCHED_SMT
7825 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7831 static bool __init cpus_share_numa(int cpu0, int cpu1)
7833 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7837 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7839 * This is the third step of three-staged workqueue subsystem initialization and
7840 * invoked after SMP and topology information are fully initialized. It
7841 * initializes the unbound CPU pods accordingly.
7843 void __init workqueue_init_topology(void)
7845 struct workqueue_struct *wq;
7848 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7849 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7850 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7851 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7853 wq_topo_initialized = true;
7855 mutex_lock(&wq_pool_mutex);
7858 * Workqueues allocated earlier would have all CPUs sharing the default
7859 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7860 * combinations to apply per-pod sharing.
7862 list_for_each_entry(wq, &workqueues, list) {
7863 for_each_online_cpu(cpu)
7864 wq_update_pod(wq, cpu, cpu, true);
7865 if (wq->flags & WQ_UNBOUND) {
7866 mutex_lock(&wq->mutex);
7867 wq_update_node_max_active(wq, -1);
7868 mutex_unlock(&wq->mutex);
7872 mutex_unlock(&wq_pool_mutex);
7875 void __warn_flushing_systemwide_wq(void)
7877 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7880 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7882 static int __init workqueue_unbound_cpus_setup(char *str)
7884 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7885 cpumask_clear(&wq_cmdline_cpumask);
7886 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7891 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);