2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE
= 1 << 1, /* die die die */
75 WORKER_IDLE
= 1 << 2, /* is idle */
76 WORKER_PREP
= 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
79 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
82 WORKER_UNBOUND
| WORKER_REBOUND
,
84 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
96 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL
= MIN_NICE
,
103 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
135 * WQ: wq->mutex protected.
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
139 * MD: wq_mayday_lock protected.
142 /* struct worker is defined in workqueue_internal.h */
145 spinlock_t lock
; /* the pool lock */
146 int cpu
; /* I: the associated cpu */
147 int node
; /* I: the associated node ID */
148 int id
; /* I: pool ID */
149 unsigned int flags
; /* X: flags */
151 unsigned long watchdog_ts
; /* L: watchdog timestamp */
153 struct list_head worklist
; /* L: list of pending works */
154 int nr_workers
; /* L: total number of workers */
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle
; /* L: currently idle ones */
159 struct list_head idle_list
; /* X: list of idle workers */
160 struct timer_list idle_timer
; /* L: worker idle timeout */
161 struct timer_list mayday_timer
; /* L: SOS timer for workers */
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
165 /* L: hash of busy workers */
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb
; /* manager arbitration */
169 struct worker
*manager
; /* L: purely informational */
170 struct mutex attach_mutex
; /* attach/detach exclusion */
171 struct list_head workers
; /* A: attached workers */
172 struct completion
*detach_completion
; /* all workers detached */
174 struct ida worker_ida
; /* worker IDs for task name */
176 struct workqueue_attrs
*attrs
; /* I: worker attributes */
177 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
178 int refcnt
; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp
;
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp
;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue
{
201 struct worker_pool
*pool
; /* I: the associated pool */
202 struct workqueue_struct
*wq
; /* I: the owning workqueue */
203 int work_color
; /* L: current color */
204 int flush_color
; /* L: flushing color */
205 int refcnt
; /* L: reference count */
206 int nr_in_flight
[WORK_NR_COLORS
];
207 /* L: nr of in_flight works */
208 int nr_active
; /* L: nr of active works */
209 int max_active
; /* L: max active works */
210 struct list_head delayed_works
; /* L: delayed works */
211 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
212 struct list_head mayday_node
; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work
;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
225 * Structure used to wait for workqueue flush.
228 struct list_head list
; /* WQ: list of flushers */
229 int flush_color
; /* WQ: flush color waiting for */
230 struct completion done
; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct
{
240 struct list_head pwqs
; /* WR: all pwqs of this wq */
241 struct list_head list
; /* PR: list of all workqueues */
243 struct mutex mutex
; /* protects this wq */
244 int work_color
; /* WQ: current work color */
245 int flush_color
; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush
; /* flush in progress */
247 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
248 struct list_head flusher_queue
; /* WQ: flush waiters */
249 struct list_head flusher_overflow
; /* WQ: flush overflow list */
251 struct list_head maydays
; /* MD: pwqs requesting rescue */
252 struct worker
*rescuer
; /* I: rescue worker */
254 int nr_drainers
; /* WQ: drain in progress */
255 int saved_max_active
; /* WQ: saved pwq max_active */
257 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
258 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
261 struct wq_device
*wq_dev
; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map
;
266 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache
*pwq_cache
;
283 static cpumask_var_t
*wq_numa_possible_cpumask
;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa
;
287 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
291 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
293 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
295 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
298 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
299 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
301 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
302 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
304 /* PL: allowable cpus for unbound wqs and work items */
305 static cpumask_var_t wq_unbound_cpumask
;
307 /* CPU where unbound work was last round robin scheduled from this CPU */
308 static DEFINE_PER_CPU(int, wq_rr_cpu_last
);
311 * Local execution of unbound work items is no longer guaranteed. The
312 * following always forces round-robin CPU selection on unbound work items
313 * to uncover usages which depend on it.
315 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316 static bool wq_debug_force_rr_cpu
= true;
318 static bool wq_debug_force_rr_cpu
= false;
320 module_param_named(debug_force_rr_cpu
, wq_debug_force_rr_cpu
, bool, 0644);
322 /* the per-cpu worker pools */
323 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
], cpu_worker_pools
);
325 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
327 /* PL: hash of all unbound pools keyed by pool->attrs */
328 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
330 /* I: attributes used when instantiating standard unbound pools on demand */
331 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
333 /* I: attributes used when instantiating ordered pools on demand */
334 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
336 struct workqueue_struct
*system_wq __read_mostly
;
337 EXPORT_SYMBOL(system_wq
);
338 struct workqueue_struct
*system_highpri_wq __read_mostly
;
339 EXPORT_SYMBOL_GPL(system_highpri_wq
);
340 struct workqueue_struct
*system_long_wq __read_mostly
;
341 EXPORT_SYMBOL_GPL(system_long_wq
);
342 struct workqueue_struct
*system_unbound_wq __read_mostly
;
343 EXPORT_SYMBOL_GPL(system_unbound_wq
);
344 struct workqueue_struct
*system_freezable_wq __read_mostly
;
345 EXPORT_SYMBOL_GPL(system_freezable_wq
);
346 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
347 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
348 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
349 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
351 static int worker_thread(void *__worker
);
352 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
354 #define CREATE_TRACE_POINTS
355 #include <trace/events/workqueue.h>
357 #define assert_rcu_or_pool_mutex() \
358 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
359 !lockdep_is_held(&wq_pool_mutex), \
360 "sched RCU or wq_pool_mutex should be held")
362 #define assert_rcu_or_wq_mutex(wq) \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq->mutex), \
365 "sched RCU or wq->mutex should be held")
367 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex) && \
370 !lockdep_is_held(&wq_pool_mutex), \
371 "sched RCU, wq->mutex or wq_pool_mutex should be held")
373 #define for_each_cpu_worker_pool(pool, cpu) \
374 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
375 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
379 * for_each_pool - iterate through all worker_pools in the system
380 * @pool: iteration cursor
381 * @pi: integer used for iteration
383 * This must be called either with wq_pool_mutex held or sched RCU read
384 * locked. If the pool needs to be used beyond the locking in effect, the
385 * caller is responsible for guaranteeing that the pool stays online.
387 * The if/else clause exists only for the lockdep assertion and can be
390 #define for_each_pool(pool, pi) \
391 idr_for_each_entry(&worker_pool_idr, pool, pi) \
392 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
396 * for_each_pool_worker - iterate through all workers of a worker_pool
397 * @worker: iteration cursor
398 * @pool: worker_pool to iterate workers of
400 * This must be called with @pool->attach_mutex.
402 * The if/else clause exists only for the lockdep assertion and can be
405 #define for_each_pool_worker(worker, pool) \
406 list_for_each_entry((worker), &(pool)->workers, node) \
407 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
411 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
412 * @pwq: iteration cursor
413 * @wq: the target workqueue
415 * This must be called either with wq->mutex held or sched RCU read locked.
416 * If the pwq needs to be used beyond the locking in effect, the caller is
417 * responsible for guaranteeing that the pwq stays online.
419 * The if/else clause exists only for the lockdep assertion and can be
422 #define for_each_pwq(pwq, wq) \
423 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
424 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
427 #ifdef CONFIG_DEBUG_OBJECTS_WORK
429 static struct debug_obj_descr work_debug_descr
;
431 static void *work_debug_hint(void *addr
)
433 return ((struct work_struct
*) addr
)->func
;
437 * fixup_init is called when:
438 * - an active object is initialized
440 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
442 struct work_struct
*work
= addr
;
445 case ODEBUG_STATE_ACTIVE
:
446 cancel_work_sync(work
);
447 debug_object_init(work
, &work_debug_descr
);
455 * fixup_activate is called when:
456 * - an active object is activated
457 * - an unknown object is activated (might be a statically initialized object)
459 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
461 struct work_struct
*work
= addr
;
465 case ODEBUG_STATE_NOTAVAILABLE
:
467 * This is not really a fixup. The work struct was
468 * statically initialized. We just make sure that it
469 * is tracked in the object tracker.
471 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
472 debug_object_init(work
, &work_debug_descr
);
473 debug_object_activate(work
, &work_debug_descr
);
479 case ODEBUG_STATE_ACTIVE
:
488 * fixup_free is called when:
489 * - an active object is freed
491 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
493 struct work_struct
*work
= addr
;
496 case ODEBUG_STATE_ACTIVE
:
497 cancel_work_sync(work
);
498 debug_object_free(work
, &work_debug_descr
);
505 static struct debug_obj_descr work_debug_descr
= {
506 .name
= "work_struct",
507 .debug_hint
= work_debug_hint
,
508 .fixup_init
= work_fixup_init
,
509 .fixup_activate
= work_fixup_activate
,
510 .fixup_free
= work_fixup_free
,
513 static inline void debug_work_activate(struct work_struct
*work
)
515 debug_object_activate(work
, &work_debug_descr
);
518 static inline void debug_work_deactivate(struct work_struct
*work
)
520 debug_object_deactivate(work
, &work_debug_descr
);
523 void __init_work(struct work_struct
*work
, int onstack
)
526 debug_object_init_on_stack(work
, &work_debug_descr
);
528 debug_object_init(work
, &work_debug_descr
);
530 EXPORT_SYMBOL_GPL(__init_work
);
532 void destroy_work_on_stack(struct work_struct
*work
)
534 debug_object_free(work
, &work_debug_descr
);
536 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
538 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
540 destroy_timer_on_stack(&work
->timer
);
541 debug_object_free(&work
->work
, &work_debug_descr
);
543 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
546 static inline void debug_work_activate(struct work_struct
*work
) { }
547 static inline void debug_work_deactivate(struct work_struct
*work
) { }
551 * worker_pool_assign_id - allocate ID and assing it to @pool
552 * @pool: the pool pointer of interest
554 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
555 * successfully, -errno on failure.
557 static int worker_pool_assign_id(struct worker_pool
*pool
)
561 lockdep_assert_held(&wq_pool_mutex
);
563 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
573 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
574 * @wq: the target workqueue
577 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
579 * If the pwq needs to be used beyond the locking in effect, the caller is
580 * responsible for guaranteeing that the pwq stays online.
582 * Return: The unbound pool_workqueue for @node.
584 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
587 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
590 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
591 * delayed item is pending. The plan is to keep CPU -> NODE
592 * mapping valid and stable across CPU on/offlines. Once that
593 * happens, this workaround can be removed.
595 if (unlikely(node
== NUMA_NO_NODE
))
598 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
601 static unsigned int work_color_to_flags(int color
)
603 return color
<< WORK_STRUCT_COLOR_SHIFT
;
606 static int get_work_color(struct work_struct
*work
)
608 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
609 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
612 static int work_next_color(int color
)
614 return (color
+ 1) % WORK_NR_COLORS
;
618 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
619 * contain the pointer to the queued pwq. Once execution starts, the flag
620 * is cleared and the high bits contain OFFQ flags and pool ID.
622 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
623 * and clear_work_data() can be used to set the pwq, pool or clear
624 * work->data. These functions should only be called while the work is
625 * owned - ie. while the PENDING bit is set.
627 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
628 * corresponding to a work. Pool is available once the work has been
629 * queued anywhere after initialization until it is sync canceled. pwq is
630 * available only while the work item is queued.
632 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
633 * canceled. While being canceled, a work item may have its PENDING set
634 * but stay off timer and worklist for arbitrarily long and nobody should
635 * try to steal the PENDING bit.
637 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
640 WARN_ON_ONCE(!work_pending(work
));
641 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
644 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
645 unsigned long extra_flags
)
647 set_work_data(work
, (unsigned long)pwq
,
648 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
651 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
654 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
655 WORK_STRUCT_PENDING
);
658 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
662 * The following wmb is paired with the implied mb in
663 * test_and_set_bit(PENDING) and ensures all updates to @work made
664 * here are visible to and precede any updates by the next PENDING
668 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
670 * The following mb guarantees that previous clear of a PENDING bit
671 * will not be reordered with any speculative LOADS or STORES from
672 * work->current_func, which is executed afterwards. This possible
673 * reordering can lead to a missed execution on attempt to qeueue
674 * the same @work. E.g. consider this case:
677 * ---------------------------- --------------------------------
679 * 1 STORE event_indicated
680 * 2 queue_work_on() {
681 * 3 test_and_set_bit(PENDING)
682 * 4 } set_..._and_clear_pending() {
683 * 5 set_work_data() # clear bit
685 * 7 work->current_func() {
686 * 8 LOAD event_indicated
689 * Without an explicit full barrier speculative LOAD on line 8 can
690 * be executed before CPU#0 does STORE on line 1. If that happens,
691 * CPU#0 observes the PENDING bit is still set and new execution of
692 * a @work is not queued in a hope, that CPU#1 will eventually
693 * finish the queued @work. Meanwhile CPU#1 does not see
694 * event_indicated is set, because speculative LOAD was executed
695 * before actual STORE.
700 static void clear_work_data(struct work_struct
*work
)
702 smp_wmb(); /* see set_work_pool_and_clear_pending() */
703 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
706 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
708 unsigned long data
= atomic_long_read(&work
->data
);
710 if (data
& WORK_STRUCT_PWQ
)
711 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
717 * get_work_pool - return the worker_pool a given work was associated with
718 * @work: the work item of interest
720 * Pools are created and destroyed under wq_pool_mutex, and allows read
721 * access under sched-RCU read lock. As such, this function should be
722 * called under wq_pool_mutex or with preemption disabled.
724 * All fields of the returned pool are accessible as long as the above
725 * mentioned locking is in effect. If the returned pool needs to be used
726 * beyond the critical section, the caller is responsible for ensuring the
727 * returned pool is and stays online.
729 * Return: The worker_pool @work was last associated with. %NULL if none.
731 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
733 unsigned long data
= atomic_long_read(&work
->data
);
736 assert_rcu_or_pool_mutex();
738 if (data
& WORK_STRUCT_PWQ
)
739 return ((struct pool_workqueue
*)
740 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
742 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
743 if (pool_id
== WORK_OFFQ_POOL_NONE
)
746 return idr_find(&worker_pool_idr
, pool_id
);
750 * get_work_pool_id - return the worker pool ID a given work is associated with
751 * @work: the work item of interest
753 * Return: The worker_pool ID @work was last associated with.
754 * %WORK_OFFQ_POOL_NONE if none.
756 static int get_work_pool_id(struct work_struct
*work
)
758 unsigned long data
= atomic_long_read(&work
->data
);
760 if (data
& WORK_STRUCT_PWQ
)
761 return ((struct pool_workqueue
*)
762 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
764 return data
>> WORK_OFFQ_POOL_SHIFT
;
767 static void mark_work_canceling(struct work_struct
*work
)
769 unsigned long pool_id
= get_work_pool_id(work
);
771 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
772 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
775 static bool work_is_canceling(struct work_struct
*work
)
777 unsigned long data
= atomic_long_read(&work
->data
);
779 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
783 * Policy functions. These define the policies on how the global worker
784 * pools are managed. Unless noted otherwise, these functions assume that
785 * they're being called with pool->lock held.
788 static bool __need_more_worker(struct worker_pool
*pool
)
790 return !atomic_read(&pool
->nr_running
);
794 * Need to wake up a worker? Called from anything but currently
797 * Note that, because unbound workers never contribute to nr_running, this
798 * function will always return %true for unbound pools as long as the
799 * worklist isn't empty.
801 static bool need_more_worker(struct worker_pool
*pool
)
803 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
806 /* Can I start working? Called from busy but !running workers. */
807 static bool may_start_working(struct worker_pool
*pool
)
809 return pool
->nr_idle
;
812 /* Do I need to keep working? Called from currently running workers. */
813 static bool keep_working(struct worker_pool
*pool
)
815 return !list_empty(&pool
->worklist
) &&
816 atomic_read(&pool
->nr_running
) <= 1;
819 /* Do we need a new worker? Called from manager. */
820 static bool need_to_create_worker(struct worker_pool
*pool
)
822 return need_more_worker(pool
) && !may_start_working(pool
);
825 /* Do we have too many workers and should some go away? */
826 static bool too_many_workers(struct worker_pool
*pool
)
828 bool managing
= mutex_is_locked(&pool
->manager_arb
);
829 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
830 int nr_busy
= pool
->nr_workers
- nr_idle
;
832 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
839 /* Return the first idle worker. Safe with preemption disabled */
840 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
842 if (unlikely(list_empty(&pool
->idle_list
)))
845 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
849 * wake_up_worker - wake up an idle worker
850 * @pool: worker pool to wake worker from
852 * Wake up the first idle worker of @pool.
855 * spin_lock_irq(pool->lock).
857 static void wake_up_worker(struct worker_pool
*pool
)
859 struct worker
*worker
= first_idle_worker(pool
);
862 wake_up_process(worker
->task
);
866 * wq_worker_waking_up - a worker is waking up
867 * @task: task waking up
868 * @cpu: CPU @task is waking up to
870 * This function is called during try_to_wake_up() when a worker is
874 * spin_lock_irq(rq->lock)
876 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
878 struct worker
*worker
= kthread_data(task
);
880 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
881 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
882 atomic_inc(&worker
->pool
->nr_running
);
887 * wq_worker_sleeping - a worker is going to sleep
888 * @task: task going to sleep
889 * @cpu: CPU in question, must be the current CPU number
891 * This function is called during schedule() when a busy worker is
892 * going to sleep. Worker on the same cpu can be woken up by
893 * returning pointer to its task.
896 * spin_lock_irq(rq->lock)
899 * Worker task on @cpu to wake up, %NULL if none.
901 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
903 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
904 struct worker_pool
*pool
;
907 * Rescuers, which may not have all the fields set up like normal
908 * workers, also reach here, let's not access anything before
909 * checking NOT_RUNNING.
911 if (worker
->flags
& WORKER_NOT_RUNNING
)
916 /* this can only happen on the local cpu */
917 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id() || pool
->cpu
!= cpu
))
921 * The counterpart of the following dec_and_test, implied mb,
922 * worklist not empty test sequence is in insert_work().
923 * Please read comment there.
925 * NOT_RUNNING is clear. This means that we're bound to and
926 * running on the local cpu w/ rq lock held and preemption
927 * disabled, which in turn means that none else could be
928 * manipulating idle_list, so dereferencing idle_list without pool
931 if (atomic_dec_and_test(&pool
->nr_running
) &&
932 !list_empty(&pool
->worklist
))
933 to_wakeup
= first_idle_worker(pool
);
934 return to_wakeup
? to_wakeup
->task
: NULL
;
938 * worker_set_flags - set worker flags and adjust nr_running accordingly
940 * @flags: flags to set
942 * Set @flags in @worker->flags and adjust nr_running accordingly.
945 * spin_lock_irq(pool->lock)
947 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
949 struct worker_pool
*pool
= worker
->pool
;
951 WARN_ON_ONCE(worker
->task
!= current
);
953 /* If transitioning into NOT_RUNNING, adjust nr_running. */
954 if ((flags
& WORKER_NOT_RUNNING
) &&
955 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
956 atomic_dec(&pool
->nr_running
);
959 worker
->flags
|= flags
;
963 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
965 * @flags: flags to clear
967 * Clear @flags in @worker->flags and adjust nr_running accordingly.
970 * spin_lock_irq(pool->lock)
972 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
974 struct worker_pool
*pool
= worker
->pool
;
975 unsigned int oflags
= worker
->flags
;
977 WARN_ON_ONCE(worker
->task
!= current
);
979 worker
->flags
&= ~flags
;
982 * If transitioning out of NOT_RUNNING, increment nr_running. Note
983 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
984 * of multiple flags, not a single flag.
986 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
987 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
988 atomic_inc(&pool
->nr_running
);
992 * find_worker_executing_work - find worker which is executing a work
993 * @pool: pool of interest
994 * @work: work to find worker for
996 * Find a worker which is executing @work on @pool by searching
997 * @pool->busy_hash which is keyed by the address of @work. For a worker
998 * to match, its current execution should match the address of @work and
999 * its work function. This is to avoid unwanted dependency between
1000 * unrelated work executions through a work item being recycled while still
1003 * This is a bit tricky. A work item may be freed once its execution
1004 * starts and nothing prevents the freed area from being recycled for
1005 * another work item. If the same work item address ends up being reused
1006 * before the original execution finishes, workqueue will identify the
1007 * recycled work item as currently executing and make it wait until the
1008 * current execution finishes, introducing an unwanted dependency.
1010 * This function checks the work item address and work function to avoid
1011 * false positives. Note that this isn't complete as one may construct a
1012 * work function which can introduce dependency onto itself through a
1013 * recycled work item. Well, if somebody wants to shoot oneself in the
1014 * foot that badly, there's only so much we can do, and if such deadlock
1015 * actually occurs, it should be easy to locate the culprit work function.
1018 * spin_lock_irq(pool->lock).
1021 * Pointer to worker which is executing @work if found, %NULL
1024 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
1025 struct work_struct
*work
)
1027 struct worker
*worker
;
1029 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1030 (unsigned long)work
)
1031 if (worker
->current_work
== work
&&
1032 worker
->current_func
== work
->func
)
1039 * move_linked_works - move linked works to a list
1040 * @work: start of series of works to be scheduled
1041 * @head: target list to append @work to
1042 * @nextp: out parameter for nested worklist walking
1044 * Schedule linked works starting from @work to @head. Work series to
1045 * be scheduled starts at @work and includes any consecutive work with
1046 * WORK_STRUCT_LINKED set in its predecessor.
1048 * If @nextp is not NULL, it's updated to point to the next work of
1049 * the last scheduled work. This allows move_linked_works() to be
1050 * nested inside outer list_for_each_entry_safe().
1053 * spin_lock_irq(pool->lock).
1055 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1056 struct work_struct
**nextp
)
1058 struct work_struct
*n
;
1061 * Linked worklist will always end before the end of the list,
1062 * use NULL for list head.
1064 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1065 list_move_tail(&work
->entry
, head
);
1066 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1071 * If we're already inside safe list traversal and have moved
1072 * multiple works to the scheduled queue, the next position
1073 * needs to be updated.
1080 * get_pwq - get an extra reference on the specified pool_workqueue
1081 * @pwq: pool_workqueue to get
1083 * Obtain an extra reference on @pwq. The caller should guarantee that
1084 * @pwq has positive refcnt and be holding the matching pool->lock.
1086 static void get_pwq(struct pool_workqueue
*pwq
)
1088 lockdep_assert_held(&pwq
->pool
->lock
);
1089 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1094 * put_pwq - put a pool_workqueue reference
1095 * @pwq: pool_workqueue to put
1097 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1098 * destruction. The caller should be holding the matching pool->lock.
1100 static void put_pwq(struct pool_workqueue
*pwq
)
1102 lockdep_assert_held(&pwq
->pool
->lock
);
1103 if (likely(--pwq
->refcnt
))
1105 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1108 * @pwq can't be released under pool->lock, bounce to
1109 * pwq_unbound_release_workfn(). This never recurses on the same
1110 * pool->lock as this path is taken only for unbound workqueues and
1111 * the release work item is scheduled on a per-cpu workqueue. To
1112 * avoid lockdep warning, unbound pool->locks are given lockdep
1113 * subclass of 1 in get_unbound_pool().
1115 schedule_work(&pwq
->unbound_release_work
);
1119 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1120 * @pwq: pool_workqueue to put (can be %NULL)
1122 * put_pwq() with locking. This function also allows %NULL @pwq.
1124 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1128 * As both pwqs and pools are sched-RCU protected, the
1129 * following lock operations are safe.
1131 spin_lock_irq(&pwq
->pool
->lock
);
1133 spin_unlock_irq(&pwq
->pool
->lock
);
1137 static void pwq_activate_delayed_work(struct work_struct
*work
)
1139 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1141 trace_workqueue_activate_work(work
);
1142 if (list_empty(&pwq
->pool
->worklist
))
1143 pwq
->pool
->watchdog_ts
= jiffies
;
1144 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1145 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1149 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1151 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1152 struct work_struct
, entry
);
1154 pwq_activate_delayed_work(work
);
1158 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1159 * @pwq: pwq of interest
1160 * @color: color of work which left the queue
1162 * A work either has completed or is removed from pending queue,
1163 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1166 * spin_lock_irq(pool->lock).
1168 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1170 /* uncolored work items don't participate in flushing or nr_active */
1171 if (color
== WORK_NO_COLOR
)
1174 pwq
->nr_in_flight
[color
]--;
1177 if (!list_empty(&pwq
->delayed_works
)) {
1178 /* one down, submit a delayed one */
1179 if (pwq
->nr_active
< pwq
->max_active
)
1180 pwq_activate_first_delayed(pwq
);
1183 /* is flush in progress and are we at the flushing tip? */
1184 if (likely(pwq
->flush_color
!= color
))
1187 /* are there still in-flight works? */
1188 if (pwq
->nr_in_flight
[color
])
1191 /* this pwq is done, clear flush_color */
1192 pwq
->flush_color
= -1;
1195 * If this was the last pwq, wake up the first flusher. It
1196 * will handle the rest.
1198 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1199 complete(&pwq
->wq
->first_flusher
->done
);
1205 * try_to_grab_pending - steal work item from worklist and disable irq
1206 * @work: work item to steal
1207 * @is_dwork: @work is a delayed_work
1208 * @flags: place to store irq state
1210 * Try to grab PENDING bit of @work. This function can handle @work in any
1211 * stable state - idle, on timer or on worklist.
1214 * 1 if @work was pending and we successfully stole PENDING
1215 * 0 if @work was idle and we claimed PENDING
1216 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1217 * -ENOENT if someone else is canceling @work, this state may persist
1218 * for arbitrarily long
1221 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1222 * interrupted while holding PENDING and @work off queue, irq must be
1223 * disabled on entry. This, combined with delayed_work->timer being
1224 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1226 * On successful return, >= 0, irq is disabled and the caller is
1227 * responsible for releasing it using local_irq_restore(*@flags).
1229 * This function is safe to call from any context including IRQ handler.
1231 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1232 unsigned long *flags
)
1234 struct worker_pool
*pool
;
1235 struct pool_workqueue
*pwq
;
1237 local_irq_save(*flags
);
1239 /* try to steal the timer if it exists */
1241 struct delayed_work
*dwork
= to_delayed_work(work
);
1244 * dwork->timer is irqsafe. If del_timer() fails, it's
1245 * guaranteed that the timer is not queued anywhere and not
1246 * running on the local CPU.
1248 if (likely(del_timer(&dwork
->timer
)))
1252 /* try to claim PENDING the normal way */
1253 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1257 * The queueing is in progress, or it is already queued. Try to
1258 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1260 pool
= get_work_pool(work
);
1264 spin_lock(&pool
->lock
);
1266 * work->data is guaranteed to point to pwq only while the work
1267 * item is queued on pwq->wq, and both updating work->data to point
1268 * to pwq on queueing and to pool on dequeueing are done under
1269 * pwq->pool->lock. This in turn guarantees that, if work->data
1270 * points to pwq which is associated with a locked pool, the work
1271 * item is currently queued on that pool.
1273 pwq
= get_work_pwq(work
);
1274 if (pwq
&& pwq
->pool
== pool
) {
1275 debug_work_deactivate(work
);
1278 * A delayed work item cannot be grabbed directly because
1279 * it might have linked NO_COLOR work items which, if left
1280 * on the delayed_list, will confuse pwq->nr_active
1281 * management later on and cause stall. Make sure the work
1282 * item is activated before grabbing.
1284 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1285 pwq_activate_delayed_work(work
);
1287 list_del_init(&work
->entry
);
1288 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1290 /* work->data points to pwq iff queued, point to pool */
1291 set_work_pool_and_keep_pending(work
, pool
->id
);
1293 spin_unlock(&pool
->lock
);
1296 spin_unlock(&pool
->lock
);
1298 local_irq_restore(*flags
);
1299 if (work_is_canceling(work
))
1306 * insert_work - insert a work into a pool
1307 * @pwq: pwq @work belongs to
1308 * @work: work to insert
1309 * @head: insertion point
1310 * @extra_flags: extra WORK_STRUCT_* flags to set
1312 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1313 * work_struct flags.
1316 * spin_lock_irq(pool->lock).
1318 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1319 struct list_head
*head
, unsigned int extra_flags
)
1321 struct worker_pool
*pool
= pwq
->pool
;
1323 /* we own @work, set data and link */
1324 set_work_pwq(work
, pwq
, extra_flags
);
1325 list_add_tail(&work
->entry
, head
);
1329 * Ensure either wq_worker_sleeping() sees the above
1330 * list_add_tail() or we see zero nr_running to avoid workers lying
1331 * around lazily while there are works to be processed.
1335 if (__need_more_worker(pool
))
1336 wake_up_worker(pool
);
1340 * Test whether @work is being queued from another work executing on the
1343 static bool is_chained_work(struct workqueue_struct
*wq
)
1345 struct worker
*worker
;
1347 worker
= current_wq_worker();
1349 * Return %true iff I'm a worker execuing a work item on @wq. If
1350 * I'm @worker, it's safe to dereference it without locking.
1352 return worker
&& worker
->current_pwq
->wq
== wq
;
1356 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1357 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1358 * avoid perturbing sensitive tasks.
1360 static int wq_select_unbound_cpu(int cpu
)
1362 static bool printed_dbg_warning
;
1365 if (likely(!wq_debug_force_rr_cpu
)) {
1366 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1368 } else if (!printed_dbg_warning
) {
1369 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1370 printed_dbg_warning
= true;
1373 if (cpumask_empty(wq_unbound_cpumask
))
1376 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1377 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1378 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1379 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1380 if (unlikely(new_cpu
>= nr_cpu_ids
))
1383 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1388 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1389 struct work_struct
*work
)
1391 struct pool_workqueue
*pwq
;
1392 struct worker_pool
*last_pool
;
1393 struct list_head
*worklist
;
1394 unsigned int work_flags
;
1395 unsigned int req_cpu
= cpu
;
1398 * While a work item is PENDING && off queue, a task trying to
1399 * steal the PENDING will busy-loop waiting for it to either get
1400 * queued or lose PENDING. Grabbing PENDING and queueing should
1401 * happen with IRQ disabled.
1403 WARN_ON_ONCE(!irqs_disabled());
1405 debug_work_activate(work
);
1407 /* if draining, only works from the same workqueue are allowed */
1408 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1409 WARN_ON_ONCE(!is_chained_work(wq
)))
1412 if (req_cpu
== WORK_CPU_UNBOUND
)
1413 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1415 /* pwq which will be used unless @work is executing elsewhere */
1416 if (!(wq
->flags
& WQ_UNBOUND
))
1417 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1419 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1422 * If @work was previously on a different pool, it might still be
1423 * running there, in which case the work needs to be queued on that
1424 * pool to guarantee non-reentrancy.
1426 last_pool
= get_work_pool(work
);
1427 if (last_pool
&& last_pool
!= pwq
->pool
) {
1428 struct worker
*worker
;
1430 spin_lock(&last_pool
->lock
);
1432 worker
= find_worker_executing_work(last_pool
, work
);
1434 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1435 pwq
= worker
->current_pwq
;
1437 /* meh... not running there, queue here */
1438 spin_unlock(&last_pool
->lock
);
1439 spin_lock(&pwq
->pool
->lock
);
1442 spin_lock(&pwq
->pool
->lock
);
1446 * pwq is determined and locked. For unbound pools, we could have
1447 * raced with pwq release and it could already be dead. If its
1448 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1449 * without another pwq replacing it in the numa_pwq_tbl or while
1450 * work items are executing on it, so the retrying is guaranteed to
1451 * make forward-progress.
1453 if (unlikely(!pwq
->refcnt
)) {
1454 if (wq
->flags
& WQ_UNBOUND
) {
1455 spin_unlock(&pwq
->pool
->lock
);
1460 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1464 /* pwq determined, queue */
1465 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1467 if (WARN_ON(!list_empty(&work
->entry
))) {
1468 spin_unlock(&pwq
->pool
->lock
);
1472 pwq
->nr_in_flight
[pwq
->work_color
]++;
1473 work_flags
= work_color_to_flags(pwq
->work_color
);
1475 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1476 trace_workqueue_activate_work(work
);
1478 worklist
= &pwq
->pool
->worklist
;
1479 if (list_empty(worklist
))
1480 pwq
->pool
->watchdog_ts
= jiffies
;
1482 work_flags
|= WORK_STRUCT_DELAYED
;
1483 worklist
= &pwq
->delayed_works
;
1486 insert_work(pwq
, work
, worklist
, work_flags
);
1488 spin_unlock(&pwq
->pool
->lock
);
1492 * queue_work_on - queue work on specific cpu
1493 * @cpu: CPU number to execute work on
1494 * @wq: workqueue to use
1495 * @work: work to queue
1497 * We queue the work to a specific CPU, the caller must ensure it
1500 * Return: %false if @work was already on a queue, %true otherwise.
1502 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1503 struct work_struct
*work
)
1506 unsigned long flags
;
1508 local_irq_save(flags
);
1510 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1511 __queue_work(cpu
, wq
, work
);
1515 local_irq_restore(flags
);
1518 EXPORT_SYMBOL(queue_work_on
);
1520 void delayed_work_timer_fn(unsigned long __data
)
1522 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1524 /* should have been called from irqsafe timer with irq already off */
1525 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1527 EXPORT_SYMBOL(delayed_work_timer_fn
);
1529 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1530 struct delayed_work
*dwork
, unsigned long delay
)
1532 struct timer_list
*timer
= &dwork
->timer
;
1533 struct work_struct
*work
= &dwork
->work
;
1535 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1536 timer
->data
!= (unsigned long)dwork
);
1537 WARN_ON_ONCE(timer_pending(timer
));
1538 WARN_ON_ONCE(!list_empty(&work
->entry
));
1541 * If @delay is 0, queue @dwork->work immediately. This is for
1542 * both optimization and correctness. The earliest @timer can
1543 * expire is on the closest next tick and delayed_work users depend
1544 * on that there's no such delay when @delay is 0.
1547 __queue_work(cpu
, wq
, &dwork
->work
);
1551 timer_stats_timer_set_start_info(&dwork
->timer
);
1555 timer
->expires
= jiffies
+ delay
;
1557 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1558 add_timer_on(timer
, cpu
);
1564 * queue_delayed_work_on - queue work on specific CPU after delay
1565 * @cpu: CPU number to execute work on
1566 * @wq: workqueue to use
1567 * @dwork: work to queue
1568 * @delay: number of jiffies to wait before queueing
1570 * Return: %false if @work was already on a queue, %true otherwise. If
1571 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1574 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1575 struct delayed_work
*dwork
, unsigned long delay
)
1577 struct work_struct
*work
= &dwork
->work
;
1579 unsigned long flags
;
1581 /* read the comment in __queue_work() */
1582 local_irq_save(flags
);
1584 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1585 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1589 local_irq_restore(flags
);
1592 EXPORT_SYMBOL(queue_delayed_work_on
);
1595 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1596 * @cpu: CPU number to execute work on
1597 * @wq: workqueue to use
1598 * @dwork: work to queue
1599 * @delay: number of jiffies to wait before queueing
1601 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1602 * modify @dwork's timer so that it expires after @delay. If @delay is
1603 * zero, @work is guaranteed to be scheduled immediately regardless of its
1606 * Return: %false if @dwork was idle and queued, %true if @dwork was
1607 * pending and its timer was modified.
1609 * This function is safe to call from any context including IRQ handler.
1610 * See try_to_grab_pending() for details.
1612 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1613 struct delayed_work
*dwork
, unsigned long delay
)
1615 unsigned long flags
;
1619 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1620 } while (unlikely(ret
== -EAGAIN
));
1622 if (likely(ret
>= 0)) {
1623 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1624 local_irq_restore(flags
);
1627 /* -ENOENT from try_to_grab_pending() becomes %true */
1630 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1633 * worker_enter_idle - enter idle state
1634 * @worker: worker which is entering idle state
1636 * @worker is entering idle state. Update stats and idle timer if
1640 * spin_lock_irq(pool->lock).
1642 static void worker_enter_idle(struct worker
*worker
)
1644 struct worker_pool
*pool
= worker
->pool
;
1646 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1647 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1648 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1651 /* can't use worker_set_flags(), also called from create_worker() */
1652 worker
->flags
|= WORKER_IDLE
;
1654 worker
->last_active
= jiffies
;
1656 /* idle_list is LIFO */
1657 list_add(&worker
->entry
, &pool
->idle_list
);
1659 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1660 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1663 * Sanity check nr_running. Because wq_unbind_fn() releases
1664 * pool->lock between setting %WORKER_UNBOUND and zapping
1665 * nr_running, the warning may trigger spuriously. Check iff
1666 * unbind is not in progress.
1668 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1669 pool
->nr_workers
== pool
->nr_idle
&&
1670 atomic_read(&pool
->nr_running
));
1674 * worker_leave_idle - leave idle state
1675 * @worker: worker which is leaving idle state
1677 * @worker is leaving idle state. Update stats.
1680 * spin_lock_irq(pool->lock).
1682 static void worker_leave_idle(struct worker
*worker
)
1684 struct worker_pool
*pool
= worker
->pool
;
1686 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1688 worker_clr_flags(worker
, WORKER_IDLE
);
1690 list_del_init(&worker
->entry
);
1693 static struct worker
*alloc_worker(int node
)
1695 struct worker
*worker
;
1697 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1699 INIT_LIST_HEAD(&worker
->entry
);
1700 INIT_LIST_HEAD(&worker
->scheduled
);
1701 INIT_LIST_HEAD(&worker
->node
);
1702 /* on creation a worker is in !idle && prep state */
1703 worker
->flags
= WORKER_PREP
;
1709 * worker_attach_to_pool() - attach a worker to a pool
1710 * @worker: worker to be attached
1711 * @pool: the target pool
1713 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1714 * cpu-binding of @worker are kept coordinated with the pool across
1717 static void worker_attach_to_pool(struct worker
*worker
,
1718 struct worker_pool
*pool
)
1720 mutex_lock(&pool
->attach_mutex
);
1723 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1724 * online CPUs. It'll be re-applied when any of the CPUs come up.
1726 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1729 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1730 * stable across this function. See the comments above the
1731 * flag definition for details.
1733 if (pool
->flags
& POOL_DISASSOCIATED
)
1734 worker
->flags
|= WORKER_UNBOUND
;
1736 list_add_tail(&worker
->node
, &pool
->workers
);
1738 mutex_unlock(&pool
->attach_mutex
);
1742 * worker_detach_from_pool() - detach a worker from its pool
1743 * @worker: worker which is attached to its pool
1744 * @pool: the pool @worker is attached to
1746 * Undo the attaching which had been done in worker_attach_to_pool(). The
1747 * caller worker shouldn't access to the pool after detached except it has
1748 * other reference to the pool.
1750 static void worker_detach_from_pool(struct worker
*worker
,
1751 struct worker_pool
*pool
)
1753 struct completion
*detach_completion
= NULL
;
1755 mutex_lock(&pool
->attach_mutex
);
1756 list_del(&worker
->node
);
1757 if (list_empty(&pool
->workers
))
1758 detach_completion
= pool
->detach_completion
;
1759 mutex_unlock(&pool
->attach_mutex
);
1761 /* clear leftover flags without pool->lock after it is detached */
1762 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1764 if (detach_completion
)
1765 complete(detach_completion
);
1769 * create_worker - create a new workqueue worker
1770 * @pool: pool the new worker will belong to
1772 * Create and start a new worker which is attached to @pool.
1775 * Might sleep. Does GFP_KERNEL allocations.
1778 * Pointer to the newly created worker.
1780 static struct worker
*create_worker(struct worker_pool
*pool
)
1782 struct worker
*worker
= NULL
;
1786 /* ID is needed to determine kthread name */
1787 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1791 worker
= alloc_worker(pool
->node
);
1795 worker
->pool
= pool
;
1799 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1800 pool
->attrs
->nice
< 0 ? "H" : "");
1802 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1804 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1805 "kworker/%s", id_buf
);
1806 if (IS_ERR(worker
->task
))
1809 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1810 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1812 /* successful, attach the worker to the pool */
1813 worker_attach_to_pool(worker
, pool
);
1815 /* start the newly created worker */
1816 spin_lock_irq(&pool
->lock
);
1817 worker
->pool
->nr_workers
++;
1818 worker_enter_idle(worker
);
1819 wake_up_process(worker
->task
);
1820 spin_unlock_irq(&pool
->lock
);
1826 ida_simple_remove(&pool
->worker_ida
, id
);
1832 * destroy_worker - destroy a workqueue worker
1833 * @worker: worker to be destroyed
1835 * Destroy @worker and adjust @pool stats accordingly. The worker should
1839 * spin_lock_irq(pool->lock).
1841 static void destroy_worker(struct worker
*worker
)
1843 struct worker_pool
*pool
= worker
->pool
;
1845 lockdep_assert_held(&pool
->lock
);
1847 /* sanity check frenzy */
1848 if (WARN_ON(worker
->current_work
) ||
1849 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1850 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1856 list_del_init(&worker
->entry
);
1857 worker
->flags
|= WORKER_DIE
;
1858 wake_up_process(worker
->task
);
1861 static void idle_worker_timeout(unsigned long __pool
)
1863 struct worker_pool
*pool
= (void *)__pool
;
1865 spin_lock_irq(&pool
->lock
);
1867 while (too_many_workers(pool
)) {
1868 struct worker
*worker
;
1869 unsigned long expires
;
1871 /* idle_list is kept in LIFO order, check the last one */
1872 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1873 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1875 if (time_before(jiffies
, expires
)) {
1876 mod_timer(&pool
->idle_timer
, expires
);
1880 destroy_worker(worker
);
1883 spin_unlock_irq(&pool
->lock
);
1886 static void send_mayday(struct work_struct
*work
)
1888 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1889 struct workqueue_struct
*wq
= pwq
->wq
;
1891 lockdep_assert_held(&wq_mayday_lock
);
1896 /* mayday mayday mayday */
1897 if (list_empty(&pwq
->mayday_node
)) {
1899 * If @pwq is for an unbound wq, its base ref may be put at
1900 * any time due to an attribute change. Pin @pwq until the
1901 * rescuer is done with it.
1904 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1905 wake_up_process(wq
->rescuer
->task
);
1909 static void pool_mayday_timeout(unsigned long __pool
)
1911 struct worker_pool
*pool
= (void *)__pool
;
1912 struct work_struct
*work
;
1914 spin_lock_irq(&pool
->lock
);
1915 spin_lock(&wq_mayday_lock
); /* for wq->maydays */
1917 if (need_to_create_worker(pool
)) {
1919 * We've been trying to create a new worker but
1920 * haven't been successful. We might be hitting an
1921 * allocation deadlock. Send distress signals to
1924 list_for_each_entry(work
, &pool
->worklist
, entry
)
1928 spin_unlock(&wq_mayday_lock
);
1929 spin_unlock_irq(&pool
->lock
);
1931 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1935 * maybe_create_worker - create a new worker if necessary
1936 * @pool: pool to create a new worker for
1938 * Create a new worker for @pool if necessary. @pool is guaranteed to
1939 * have at least one idle worker on return from this function. If
1940 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1941 * sent to all rescuers with works scheduled on @pool to resolve
1942 * possible allocation deadlock.
1944 * On return, need_to_create_worker() is guaranteed to be %false and
1945 * may_start_working() %true.
1948 * spin_lock_irq(pool->lock) which may be released and regrabbed
1949 * multiple times. Does GFP_KERNEL allocations. Called only from
1952 static void maybe_create_worker(struct worker_pool
*pool
)
1953 __releases(&pool
->lock
)
1954 __acquires(&pool
->lock
)
1957 spin_unlock_irq(&pool
->lock
);
1959 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1960 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1963 if (create_worker(pool
) || !need_to_create_worker(pool
))
1966 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1968 if (!need_to_create_worker(pool
))
1972 del_timer_sync(&pool
->mayday_timer
);
1973 spin_lock_irq(&pool
->lock
);
1975 * This is necessary even after a new worker was just successfully
1976 * created as @pool->lock was dropped and the new worker might have
1977 * already become busy.
1979 if (need_to_create_worker(pool
))
1984 * manage_workers - manage worker pool
1987 * Assume the manager role and manage the worker pool @worker belongs
1988 * to. At any given time, there can be only zero or one manager per
1989 * pool. The exclusion is handled automatically by this function.
1991 * The caller can safely start processing works on false return. On
1992 * true return, it's guaranteed that need_to_create_worker() is false
1993 * and may_start_working() is true.
1996 * spin_lock_irq(pool->lock) which may be released and regrabbed
1997 * multiple times. Does GFP_KERNEL allocations.
2000 * %false if the pool doesn't need management and the caller can safely
2001 * start processing works, %true if management function was performed and
2002 * the conditions that the caller verified before calling the function may
2003 * no longer be true.
2005 static bool manage_workers(struct worker
*worker
)
2007 struct worker_pool
*pool
= worker
->pool
;
2010 * Anyone who successfully grabs manager_arb wins the arbitration
2011 * and becomes the manager. mutex_trylock() on pool->manager_arb
2012 * failure while holding pool->lock reliably indicates that someone
2013 * else is managing the pool and the worker which failed trylock
2014 * can proceed to executing work items. This means that anyone
2015 * grabbing manager_arb is responsible for actually performing
2016 * manager duties. If manager_arb is grabbed and released without
2017 * actual management, the pool may stall indefinitely.
2019 if (!mutex_trylock(&pool
->manager_arb
))
2021 pool
->manager
= worker
;
2023 maybe_create_worker(pool
);
2025 pool
->manager
= NULL
;
2026 mutex_unlock(&pool
->manager_arb
);
2031 * process_one_work - process single work
2033 * @work: work to process
2035 * Process @work. This function contains all the logics necessary to
2036 * process a single work including synchronization against and
2037 * interaction with other workers on the same cpu, queueing and
2038 * flushing. As long as context requirement is met, any worker can
2039 * call this function to process a work.
2042 * spin_lock_irq(pool->lock) which is released and regrabbed.
2044 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2045 __releases(&pool
->lock
)
2046 __acquires(&pool
->lock
)
2048 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2049 struct worker_pool
*pool
= worker
->pool
;
2050 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2052 struct worker
*collision
;
2053 #ifdef CONFIG_LOCKDEP
2055 * It is permissible to free the struct work_struct from
2056 * inside the function that is called from it, this we need to
2057 * take into account for lockdep too. To avoid bogus "held
2058 * lock freed" warnings as well as problems when looking into
2059 * work->lockdep_map, make a copy and use that here.
2061 struct lockdep_map lockdep_map
;
2063 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2065 /* ensure we're on the correct CPU */
2066 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2067 raw_smp_processor_id() != pool
->cpu
);
2070 * A single work shouldn't be executed concurrently by
2071 * multiple workers on a single cpu. Check whether anyone is
2072 * already processing the work. If so, defer the work to the
2073 * currently executing one.
2075 collision
= find_worker_executing_work(pool
, work
);
2076 if (unlikely(collision
)) {
2077 move_linked_works(work
, &collision
->scheduled
, NULL
);
2081 /* claim and dequeue */
2082 debug_work_deactivate(work
);
2083 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2084 worker
->current_work
= work
;
2085 worker
->current_func
= work
->func
;
2086 worker
->current_pwq
= pwq
;
2087 work_color
= get_work_color(work
);
2089 list_del_init(&work
->entry
);
2092 * CPU intensive works don't participate in concurrency management.
2093 * They're the scheduler's responsibility. This takes @worker out
2094 * of concurrency management and the next code block will chain
2095 * execution of the pending work items.
2097 if (unlikely(cpu_intensive
))
2098 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2101 * Wake up another worker if necessary. The condition is always
2102 * false for normal per-cpu workers since nr_running would always
2103 * be >= 1 at this point. This is used to chain execution of the
2104 * pending work items for WORKER_NOT_RUNNING workers such as the
2105 * UNBOUND and CPU_INTENSIVE ones.
2107 if (need_more_worker(pool
))
2108 wake_up_worker(pool
);
2111 * Record the last pool and clear PENDING which should be the last
2112 * update to @work. Also, do this inside @pool->lock so that
2113 * PENDING and queued state changes happen together while IRQ is
2116 set_work_pool_and_clear_pending(work
, pool
->id
);
2118 spin_unlock_irq(&pool
->lock
);
2120 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2121 lock_map_acquire(&lockdep_map
);
2122 trace_workqueue_execute_start(work
);
2123 worker
->current_func(work
);
2125 * While we must be careful to not use "work" after this, the trace
2126 * point will only record its address.
2128 trace_workqueue_execute_end(work
);
2129 lock_map_release(&lockdep_map
);
2130 lock_map_release(&pwq
->wq
->lockdep_map
);
2132 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2133 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2134 " last function: %pf\n",
2135 current
->comm
, preempt_count(), task_pid_nr(current
),
2136 worker
->current_func
);
2137 debug_show_held_locks(current
);
2142 * The following prevents a kworker from hogging CPU on !PREEMPT
2143 * kernels, where a requeueing work item waiting for something to
2144 * happen could deadlock with stop_machine as such work item could
2145 * indefinitely requeue itself while all other CPUs are trapped in
2146 * stop_machine. At the same time, report a quiescent RCU state so
2147 * the same condition doesn't freeze RCU.
2149 cond_resched_rcu_qs();
2151 spin_lock_irq(&pool
->lock
);
2153 /* clear cpu intensive status */
2154 if (unlikely(cpu_intensive
))
2155 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2157 /* we're done with it, release */
2158 hash_del(&worker
->hentry
);
2159 worker
->current_work
= NULL
;
2160 worker
->current_func
= NULL
;
2161 worker
->current_pwq
= NULL
;
2162 worker
->desc_valid
= false;
2163 pwq_dec_nr_in_flight(pwq
, work_color
);
2167 * process_scheduled_works - process scheduled works
2170 * Process all scheduled works. Please note that the scheduled list
2171 * may change while processing a work, so this function repeatedly
2172 * fetches a work from the top and executes it.
2175 * spin_lock_irq(pool->lock) which may be released and regrabbed
2178 static void process_scheduled_works(struct worker
*worker
)
2180 while (!list_empty(&worker
->scheduled
)) {
2181 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2182 struct work_struct
, entry
);
2183 process_one_work(worker
, work
);
2188 * worker_thread - the worker thread function
2191 * The worker thread function. All workers belong to a worker_pool -
2192 * either a per-cpu one or dynamic unbound one. These workers process all
2193 * work items regardless of their specific target workqueue. The only
2194 * exception is work items which belong to workqueues with a rescuer which
2195 * will be explained in rescuer_thread().
2199 static int worker_thread(void *__worker
)
2201 struct worker
*worker
= __worker
;
2202 struct worker_pool
*pool
= worker
->pool
;
2204 /* tell the scheduler that this is a workqueue worker */
2205 worker
->task
->flags
|= PF_WQ_WORKER
;
2207 spin_lock_irq(&pool
->lock
);
2209 /* am I supposed to die? */
2210 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2211 spin_unlock_irq(&pool
->lock
);
2212 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2213 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2215 set_task_comm(worker
->task
, "kworker/dying");
2216 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2217 worker_detach_from_pool(worker
, pool
);
2222 worker_leave_idle(worker
);
2224 /* no more worker necessary? */
2225 if (!need_more_worker(pool
))
2228 /* do we need to manage? */
2229 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2233 * ->scheduled list can only be filled while a worker is
2234 * preparing to process a work or actually processing it.
2235 * Make sure nobody diddled with it while I was sleeping.
2237 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2240 * Finish PREP stage. We're guaranteed to have at least one idle
2241 * worker or that someone else has already assumed the manager
2242 * role. This is where @worker starts participating in concurrency
2243 * management if applicable and concurrency management is restored
2244 * after being rebound. See rebind_workers() for details.
2246 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2249 struct work_struct
*work
=
2250 list_first_entry(&pool
->worklist
,
2251 struct work_struct
, entry
);
2253 pool
->watchdog_ts
= jiffies
;
2255 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2256 /* optimization path, not strictly necessary */
2257 process_one_work(worker
, work
);
2258 if (unlikely(!list_empty(&worker
->scheduled
)))
2259 process_scheduled_works(worker
);
2261 move_linked_works(work
, &worker
->scheduled
, NULL
);
2262 process_scheduled_works(worker
);
2264 } while (keep_working(pool
));
2266 worker_set_flags(worker
, WORKER_PREP
);
2269 * pool->lock is held and there's no work to process and no need to
2270 * manage, sleep. Workers are woken up only while holding
2271 * pool->lock or from local cpu, so setting the current state
2272 * before releasing pool->lock is enough to prevent losing any
2275 worker_enter_idle(worker
);
2276 __set_current_state(TASK_INTERRUPTIBLE
);
2277 spin_unlock_irq(&pool
->lock
);
2283 * rescuer_thread - the rescuer thread function
2286 * Workqueue rescuer thread function. There's one rescuer for each
2287 * workqueue which has WQ_MEM_RECLAIM set.
2289 * Regular work processing on a pool may block trying to create a new
2290 * worker which uses GFP_KERNEL allocation which has slight chance of
2291 * developing into deadlock if some works currently on the same queue
2292 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2293 * the problem rescuer solves.
2295 * When such condition is possible, the pool summons rescuers of all
2296 * workqueues which have works queued on the pool and let them process
2297 * those works so that forward progress can be guaranteed.
2299 * This should happen rarely.
2303 static int rescuer_thread(void *__rescuer
)
2305 struct worker
*rescuer
= __rescuer
;
2306 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2307 struct list_head
*scheduled
= &rescuer
->scheduled
;
2310 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2313 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2314 * doesn't participate in concurrency management.
2316 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2318 set_current_state(TASK_INTERRUPTIBLE
);
2321 * By the time the rescuer is requested to stop, the workqueue
2322 * shouldn't have any work pending, but @wq->maydays may still have
2323 * pwq(s) queued. This can happen by non-rescuer workers consuming
2324 * all the work items before the rescuer got to them. Go through
2325 * @wq->maydays processing before acting on should_stop so that the
2326 * list is always empty on exit.
2328 should_stop
= kthread_should_stop();
2330 /* see whether any pwq is asking for help */
2331 spin_lock_irq(&wq_mayday_lock
);
2333 while (!list_empty(&wq
->maydays
)) {
2334 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2335 struct pool_workqueue
, mayday_node
);
2336 struct worker_pool
*pool
= pwq
->pool
;
2337 struct work_struct
*work
, *n
;
2340 __set_current_state(TASK_RUNNING
);
2341 list_del_init(&pwq
->mayday_node
);
2343 spin_unlock_irq(&wq_mayday_lock
);
2345 worker_attach_to_pool(rescuer
, pool
);
2347 spin_lock_irq(&pool
->lock
);
2348 rescuer
->pool
= pool
;
2351 * Slurp in all works issued via this workqueue and
2354 WARN_ON_ONCE(!list_empty(scheduled
));
2355 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2356 if (get_work_pwq(work
) == pwq
) {
2358 pool
->watchdog_ts
= jiffies
;
2359 move_linked_works(work
, scheduled
, &n
);
2364 if (!list_empty(scheduled
)) {
2365 process_scheduled_works(rescuer
);
2368 * The above execution of rescued work items could
2369 * have created more to rescue through
2370 * pwq_activate_first_delayed() or chained
2371 * queueing. Let's put @pwq back on mayday list so
2372 * that such back-to-back work items, which may be
2373 * being used to relieve memory pressure, don't
2374 * incur MAYDAY_INTERVAL delay inbetween.
2376 if (need_to_create_worker(pool
)) {
2377 spin_lock(&wq_mayday_lock
);
2379 list_move_tail(&pwq
->mayday_node
, &wq
->maydays
);
2380 spin_unlock(&wq_mayday_lock
);
2385 * Put the reference grabbed by send_mayday(). @pool won't
2386 * go away while we're still attached to it.
2391 * Leave this pool. If need_more_worker() is %true, notify a
2392 * regular worker; otherwise, we end up with 0 concurrency
2393 * and stalling the execution.
2395 if (need_more_worker(pool
))
2396 wake_up_worker(pool
);
2398 rescuer
->pool
= NULL
;
2399 spin_unlock_irq(&pool
->lock
);
2401 worker_detach_from_pool(rescuer
, pool
);
2403 spin_lock_irq(&wq_mayday_lock
);
2406 spin_unlock_irq(&wq_mayday_lock
);
2409 __set_current_state(TASK_RUNNING
);
2410 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2414 /* rescuers should never participate in concurrency management */
2415 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2421 * check_flush_dependency - check for flush dependency sanity
2422 * @target_wq: workqueue being flushed
2423 * @target_work: work item being flushed (NULL for workqueue flushes)
2425 * %current is trying to flush the whole @target_wq or @target_work on it.
2426 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2427 * reclaiming memory or running on a workqueue which doesn't have
2428 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2431 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2432 struct work_struct
*target_work
)
2434 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2435 struct worker
*worker
;
2437 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2440 worker
= current_wq_worker();
2442 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2443 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2444 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2445 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2446 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2447 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2448 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2449 target_wq
->name
, target_func
);
2453 struct work_struct work
;
2454 struct completion done
;
2455 struct task_struct
*task
; /* purely informational */
2458 static void wq_barrier_func(struct work_struct
*work
)
2460 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2461 complete(&barr
->done
);
2465 * insert_wq_barrier - insert a barrier work
2466 * @pwq: pwq to insert barrier into
2467 * @barr: wq_barrier to insert
2468 * @target: target work to attach @barr to
2469 * @worker: worker currently executing @target, NULL if @target is not executing
2471 * @barr is linked to @target such that @barr is completed only after
2472 * @target finishes execution. Please note that the ordering
2473 * guarantee is observed only with respect to @target and on the local
2476 * Currently, a queued barrier can't be canceled. This is because
2477 * try_to_grab_pending() can't determine whether the work to be
2478 * grabbed is at the head of the queue and thus can't clear LINKED
2479 * flag of the previous work while there must be a valid next work
2480 * after a work with LINKED flag set.
2482 * Note that when @worker is non-NULL, @target may be modified
2483 * underneath us, so we can't reliably determine pwq from @target.
2486 * spin_lock_irq(pool->lock).
2488 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2489 struct wq_barrier
*barr
,
2490 struct work_struct
*target
, struct worker
*worker
)
2492 struct list_head
*head
;
2493 unsigned int linked
= 0;
2496 * debugobject calls are safe here even with pool->lock locked
2497 * as we know for sure that this will not trigger any of the
2498 * checks and call back into the fixup functions where we
2501 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2502 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2503 init_completion(&barr
->done
);
2504 barr
->task
= current
;
2507 * If @target is currently being executed, schedule the
2508 * barrier to the worker; otherwise, put it after @target.
2511 head
= worker
->scheduled
.next
;
2513 unsigned long *bits
= work_data_bits(target
);
2515 head
= target
->entry
.next
;
2516 /* there can already be other linked works, inherit and set */
2517 linked
= *bits
& WORK_STRUCT_LINKED
;
2518 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2521 debug_work_activate(&barr
->work
);
2522 insert_work(pwq
, &barr
->work
, head
,
2523 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2527 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2528 * @wq: workqueue being flushed
2529 * @flush_color: new flush color, < 0 for no-op
2530 * @work_color: new work color, < 0 for no-op
2532 * Prepare pwqs for workqueue flushing.
2534 * If @flush_color is non-negative, flush_color on all pwqs should be
2535 * -1. If no pwq has in-flight commands at the specified color, all
2536 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2537 * has in flight commands, its pwq->flush_color is set to
2538 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2539 * wakeup logic is armed and %true is returned.
2541 * The caller should have initialized @wq->first_flusher prior to
2542 * calling this function with non-negative @flush_color. If
2543 * @flush_color is negative, no flush color update is done and %false
2546 * If @work_color is non-negative, all pwqs should have the same
2547 * work_color which is previous to @work_color and all will be
2548 * advanced to @work_color.
2551 * mutex_lock(wq->mutex).
2554 * %true if @flush_color >= 0 and there's something to flush. %false
2557 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2558 int flush_color
, int work_color
)
2561 struct pool_workqueue
*pwq
;
2563 if (flush_color
>= 0) {
2564 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2565 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2568 for_each_pwq(pwq
, wq
) {
2569 struct worker_pool
*pool
= pwq
->pool
;
2571 spin_lock_irq(&pool
->lock
);
2573 if (flush_color
>= 0) {
2574 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2576 if (pwq
->nr_in_flight
[flush_color
]) {
2577 pwq
->flush_color
= flush_color
;
2578 atomic_inc(&wq
->nr_pwqs_to_flush
);
2583 if (work_color
>= 0) {
2584 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2585 pwq
->work_color
= work_color
;
2588 spin_unlock_irq(&pool
->lock
);
2591 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2592 complete(&wq
->first_flusher
->done
);
2598 * flush_workqueue - ensure that any scheduled work has run to completion.
2599 * @wq: workqueue to flush
2601 * This function sleeps until all work items which were queued on entry
2602 * have finished execution, but it is not livelocked by new incoming ones.
2604 void flush_workqueue(struct workqueue_struct
*wq
)
2606 struct wq_flusher this_flusher
= {
2607 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2609 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2613 lock_map_acquire(&wq
->lockdep_map
);
2614 lock_map_release(&wq
->lockdep_map
);
2616 mutex_lock(&wq
->mutex
);
2619 * Start-to-wait phase
2621 next_color
= work_next_color(wq
->work_color
);
2623 if (next_color
!= wq
->flush_color
) {
2625 * Color space is not full. The current work_color
2626 * becomes our flush_color and work_color is advanced
2629 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2630 this_flusher
.flush_color
= wq
->work_color
;
2631 wq
->work_color
= next_color
;
2633 if (!wq
->first_flusher
) {
2634 /* no flush in progress, become the first flusher */
2635 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2637 wq
->first_flusher
= &this_flusher
;
2639 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2641 /* nothing to flush, done */
2642 wq
->flush_color
= next_color
;
2643 wq
->first_flusher
= NULL
;
2648 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2649 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2650 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2654 * Oops, color space is full, wait on overflow queue.
2655 * The next flush completion will assign us
2656 * flush_color and transfer to flusher_queue.
2658 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2661 check_flush_dependency(wq
, NULL
);
2663 mutex_unlock(&wq
->mutex
);
2665 wait_for_completion(&this_flusher
.done
);
2668 * Wake-up-and-cascade phase
2670 * First flushers are responsible for cascading flushes and
2671 * handling overflow. Non-first flushers can simply return.
2673 if (wq
->first_flusher
!= &this_flusher
)
2676 mutex_lock(&wq
->mutex
);
2678 /* we might have raced, check again with mutex held */
2679 if (wq
->first_flusher
!= &this_flusher
)
2682 wq
->first_flusher
= NULL
;
2684 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2685 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2688 struct wq_flusher
*next
, *tmp
;
2690 /* complete all the flushers sharing the current flush color */
2691 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2692 if (next
->flush_color
!= wq
->flush_color
)
2694 list_del_init(&next
->list
);
2695 complete(&next
->done
);
2698 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2699 wq
->flush_color
!= work_next_color(wq
->work_color
));
2701 /* this flush_color is finished, advance by one */
2702 wq
->flush_color
= work_next_color(wq
->flush_color
);
2704 /* one color has been freed, handle overflow queue */
2705 if (!list_empty(&wq
->flusher_overflow
)) {
2707 * Assign the same color to all overflowed
2708 * flushers, advance work_color and append to
2709 * flusher_queue. This is the start-to-wait
2710 * phase for these overflowed flushers.
2712 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2713 tmp
->flush_color
= wq
->work_color
;
2715 wq
->work_color
= work_next_color(wq
->work_color
);
2717 list_splice_tail_init(&wq
->flusher_overflow
,
2718 &wq
->flusher_queue
);
2719 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2722 if (list_empty(&wq
->flusher_queue
)) {
2723 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2728 * Need to flush more colors. Make the next flusher
2729 * the new first flusher and arm pwqs.
2731 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2732 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2734 list_del_init(&next
->list
);
2735 wq
->first_flusher
= next
;
2737 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2741 * Meh... this color is already done, clear first
2742 * flusher and repeat cascading.
2744 wq
->first_flusher
= NULL
;
2748 mutex_unlock(&wq
->mutex
);
2750 EXPORT_SYMBOL(flush_workqueue
);
2753 * drain_workqueue - drain a workqueue
2754 * @wq: workqueue to drain
2756 * Wait until the workqueue becomes empty. While draining is in progress,
2757 * only chain queueing is allowed. IOW, only currently pending or running
2758 * work items on @wq can queue further work items on it. @wq is flushed
2759 * repeatedly until it becomes empty. The number of flushing is determined
2760 * by the depth of chaining and should be relatively short. Whine if it
2763 void drain_workqueue(struct workqueue_struct
*wq
)
2765 unsigned int flush_cnt
= 0;
2766 struct pool_workqueue
*pwq
;
2769 * __queue_work() needs to test whether there are drainers, is much
2770 * hotter than drain_workqueue() and already looks at @wq->flags.
2771 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2773 mutex_lock(&wq
->mutex
);
2774 if (!wq
->nr_drainers
++)
2775 wq
->flags
|= __WQ_DRAINING
;
2776 mutex_unlock(&wq
->mutex
);
2778 flush_workqueue(wq
);
2780 mutex_lock(&wq
->mutex
);
2782 for_each_pwq(pwq
, wq
) {
2785 spin_lock_irq(&pwq
->pool
->lock
);
2786 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2787 spin_unlock_irq(&pwq
->pool
->lock
);
2792 if (++flush_cnt
== 10 ||
2793 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2794 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2795 wq
->name
, flush_cnt
);
2797 mutex_unlock(&wq
->mutex
);
2801 if (!--wq
->nr_drainers
)
2802 wq
->flags
&= ~__WQ_DRAINING
;
2803 mutex_unlock(&wq
->mutex
);
2805 EXPORT_SYMBOL_GPL(drain_workqueue
);
2807 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2809 struct worker
*worker
= NULL
;
2810 struct worker_pool
*pool
;
2811 struct pool_workqueue
*pwq
;
2815 local_irq_disable();
2816 pool
= get_work_pool(work
);
2822 spin_lock(&pool
->lock
);
2823 /* see the comment in try_to_grab_pending() with the same code */
2824 pwq
= get_work_pwq(work
);
2826 if (unlikely(pwq
->pool
!= pool
))
2829 worker
= find_worker_executing_work(pool
, work
);
2832 pwq
= worker
->current_pwq
;
2835 check_flush_dependency(pwq
->wq
, work
);
2837 insert_wq_barrier(pwq
, barr
, work
, worker
);
2838 spin_unlock_irq(&pool
->lock
);
2841 * If @max_active is 1 or rescuer is in use, flushing another work
2842 * item on the same workqueue may lead to deadlock. Make sure the
2843 * flusher is not running on the same workqueue by verifying write
2846 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2847 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2849 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2850 lock_map_release(&pwq
->wq
->lockdep_map
);
2854 spin_unlock_irq(&pool
->lock
);
2859 * flush_work - wait for a work to finish executing the last queueing instance
2860 * @work: the work to flush
2862 * Wait until @work has finished execution. @work is guaranteed to be idle
2863 * on return if it hasn't been requeued since flush started.
2866 * %true if flush_work() waited for the work to finish execution,
2867 * %false if it was already idle.
2869 bool flush_work(struct work_struct
*work
)
2871 struct wq_barrier barr
;
2873 lock_map_acquire(&work
->lockdep_map
);
2874 lock_map_release(&work
->lockdep_map
);
2876 if (start_flush_work(work
, &barr
)) {
2877 wait_for_completion(&barr
.done
);
2878 destroy_work_on_stack(&barr
.work
);
2884 EXPORT_SYMBOL_GPL(flush_work
);
2888 struct work_struct
*work
;
2891 static int cwt_wakefn(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
2893 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
2895 if (cwait
->work
!= key
)
2897 return autoremove_wake_function(wait
, mode
, sync
, key
);
2900 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2902 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
2903 unsigned long flags
;
2907 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2909 * If someone else is already canceling, wait for it to
2910 * finish. flush_work() doesn't work for PREEMPT_NONE
2911 * because we may get scheduled between @work's completion
2912 * and the other canceling task resuming and clearing
2913 * CANCELING - flush_work() will return false immediately
2914 * as @work is no longer busy, try_to_grab_pending() will
2915 * return -ENOENT as @work is still being canceled and the
2916 * other canceling task won't be able to clear CANCELING as
2917 * we're hogging the CPU.
2919 * Let's wait for completion using a waitqueue. As this
2920 * may lead to the thundering herd problem, use a custom
2921 * wake function which matches @work along with exclusive
2924 if (unlikely(ret
== -ENOENT
)) {
2925 struct cwt_wait cwait
;
2927 init_wait(&cwait
.wait
);
2928 cwait
.wait
.func
= cwt_wakefn
;
2931 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
2932 TASK_UNINTERRUPTIBLE
);
2933 if (work_is_canceling(work
))
2935 finish_wait(&cancel_waitq
, &cwait
.wait
);
2937 } while (unlikely(ret
< 0));
2939 /* tell other tasks trying to grab @work to back off */
2940 mark_work_canceling(work
);
2941 local_irq_restore(flags
);
2944 clear_work_data(work
);
2947 * Paired with prepare_to_wait() above so that either
2948 * waitqueue_active() is visible here or !work_is_canceling() is
2952 if (waitqueue_active(&cancel_waitq
))
2953 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
2959 * cancel_work_sync - cancel a work and wait for it to finish
2960 * @work: the work to cancel
2962 * Cancel @work and wait for its execution to finish. This function
2963 * can be used even if the work re-queues itself or migrates to
2964 * another workqueue. On return from this function, @work is
2965 * guaranteed to be not pending or executing on any CPU.
2967 * cancel_work_sync(&delayed_work->work) must not be used for
2968 * delayed_work's. Use cancel_delayed_work_sync() instead.
2970 * The caller must ensure that the workqueue on which @work was last
2971 * queued can't be destroyed before this function returns.
2974 * %true if @work was pending, %false otherwise.
2976 bool cancel_work_sync(struct work_struct
*work
)
2978 return __cancel_work_timer(work
, false);
2980 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2983 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2984 * @dwork: the delayed work to flush
2986 * Delayed timer is cancelled and the pending work is queued for
2987 * immediate execution. Like flush_work(), this function only
2988 * considers the last queueing instance of @dwork.
2991 * %true if flush_work() waited for the work to finish execution,
2992 * %false if it was already idle.
2994 bool flush_delayed_work(struct delayed_work
*dwork
)
2996 local_irq_disable();
2997 if (del_timer_sync(&dwork
->timer
))
2998 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3000 return flush_work(&dwork
->work
);
3002 EXPORT_SYMBOL(flush_delayed_work
);
3005 * cancel_delayed_work - cancel a delayed work
3006 * @dwork: delayed_work to cancel
3008 * Kill off a pending delayed_work.
3010 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3014 * The work callback function may still be running on return, unless
3015 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3016 * use cancel_delayed_work_sync() to wait on it.
3018 * This function is safe to call from any context including IRQ handler.
3020 bool cancel_delayed_work(struct delayed_work
*dwork
)
3022 unsigned long flags
;
3026 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
3027 } while (unlikely(ret
== -EAGAIN
));
3029 if (unlikely(ret
< 0))
3032 set_work_pool_and_clear_pending(&dwork
->work
,
3033 get_work_pool_id(&dwork
->work
));
3034 local_irq_restore(flags
);
3037 EXPORT_SYMBOL(cancel_delayed_work
);
3040 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3041 * @dwork: the delayed work cancel
3043 * This is cancel_work_sync() for delayed works.
3046 * %true if @dwork was pending, %false otherwise.
3048 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3050 return __cancel_work_timer(&dwork
->work
, true);
3052 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3055 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3056 * @func: the function to call
3058 * schedule_on_each_cpu() executes @func on each online CPU using the
3059 * system workqueue and blocks until all CPUs have completed.
3060 * schedule_on_each_cpu() is very slow.
3063 * 0 on success, -errno on failure.
3065 int schedule_on_each_cpu(work_func_t func
)
3068 struct work_struct __percpu
*works
;
3070 works
= alloc_percpu(struct work_struct
);
3076 for_each_online_cpu(cpu
) {
3077 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3079 INIT_WORK(work
, func
);
3080 schedule_work_on(cpu
, work
);
3083 for_each_online_cpu(cpu
)
3084 flush_work(per_cpu_ptr(works
, cpu
));
3092 * execute_in_process_context - reliably execute the routine with user context
3093 * @fn: the function to execute
3094 * @ew: guaranteed storage for the execute work structure (must
3095 * be available when the work executes)
3097 * Executes the function immediately if process context is available,
3098 * otherwise schedules the function for delayed execution.
3100 * Return: 0 - function was executed
3101 * 1 - function was scheduled for execution
3103 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3105 if (!in_interrupt()) {
3110 INIT_WORK(&ew
->work
, fn
);
3111 schedule_work(&ew
->work
);
3115 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3118 * free_workqueue_attrs - free a workqueue_attrs
3119 * @attrs: workqueue_attrs to free
3121 * Undo alloc_workqueue_attrs().
3123 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3126 free_cpumask_var(attrs
->cpumask
);
3132 * alloc_workqueue_attrs - allocate a workqueue_attrs
3133 * @gfp_mask: allocation mask to use
3135 * Allocate a new workqueue_attrs, initialize with default settings and
3138 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3140 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3142 struct workqueue_attrs
*attrs
;
3144 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3147 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3150 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3153 free_workqueue_attrs(attrs
);
3157 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3158 const struct workqueue_attrs
*from
)
3160 to
->nice
= from
->nice
;
3161 cpumask_copy(to
->cpumask
, from
->cpumask
);
3163 * Unlike hash and equality test, this function doesn't ignore
3164 * ->no_numa as it is used for both pool and wq attrs. Instead,
3165 * get_unbound_pool() explicitly clears ->no_numa after copying.
3167 to
->no_numa
= from
->no_numa
;
3170 /* hash value of the content of @attr */
3171 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3175 hash
= jhash_1word(attrs
->nice
, hash
);
3176 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3177 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3181 /* content equality test */
3182 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3183 const struct workqueue_attrs
*b
)
3185 if (a
->nice
!= b
->nice
)
3187 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3193 * init_worker_pool - initialize a newly zalloc'd worker_pool
3194 * @pool: worker_pool to initialize
3196 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3198 * Return: 0 on success, -errno on failure. Even on failure, all fields
3199 * inside @pool proper are initialized and put_unbound_pool() can be called
3200 * on @pool safely to release it.
3202 static int init_worker_pool(struct worker_pool
*pool
)
3204 spin_lock_init(&pool
->lock
);
3207 pool
->node
= NUMA_NO_NODE
;
3208 pool
->flags
|= POOL_DISASSOCIATED
;
3209 pool
->watchdog_ts
= jiffies
;
3210 INIT_LIST_HEAD(&pool
->worklist
);
3211 INIT_LIST_HEAD(&pool
->idle_list
);
3212 hash_init(pool
->busy_hash
);
3214 init_timer_deferrable(&pool
->idle_timer
);
3215 pool
->idle_timer
.function
= idle_worker_timeout
;
3216 pool
->idle_timer
.data
= (unsigned long)pool
;
3218 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3219 (unsigned long)pool
);
3221 mutex_init(&pool
->manager_arb
);
3222 mutex_init(&pool
->attach_mutex
);
3223 INIT_LIST_HEAD(&pool
->workers
);
3225 ida_init(&pool
->worker_ida
);
3226 INIT_HLIST_NODE(&pool
->hash_node
);
3229 /* shouldn't fail above this point */
3230 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3236 static void rcu_free_wq(struct rcu_head
*rcu
)
3238 struct workqueue_struct
*wq
=
3239 container_of(rcu
, struct workqueue_struct
, rcu
);
3241 if (!(wq
->flags
& WQ_UNBOUND
))
3242 free_percpu(wq
->cpu_pwqs
);
3244 free_workqueue_attrs(wq
->unbound_attrs
);
3250 static void rcu_free_pool(struct rcu_head
*rcu
)
3252 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3254 ida_destroy(&pool
->worker_ida
);
3255 free_workqueue_attrs(pool
->attrs
);
3260 * put_unbound_pool - put a worker_pool
3261 * @pool: worker_pool to put
3263 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3264 * safe manner. get_unbound_pool() calls this function on its failure path
3265 * and this function should be able to release pools which went through,
3266 * successfully or not, init_worker_pool().
3268 * Should be called with wq_pool_mutex held.
3270 static void put_unbound_pool(struct worker_pool
*pool
)
3272 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3273 struct worker
*worker
;
3275 lockdep_assert_held(&wq_pool_mutex
);
3281 if (WARN_ON(!(pool
->cpu
< 0)) ||
3282 WARN_ON(!list_empty(&pool
->worklist
)))
3285 /* release id and unhash */
3287 idr_remove(&worker_pool_idr
, pool
->id
);
3288 hash_del(&pool
->hash_node
);
3291 * Become the manager and destroy all workers. Grabbing
3292 * manager_arb prevents @pool's workers from blocking on
3295 mutex_lock(&pool
->manager_arb
);
3297 spin_lock_irq(&pool
->lock
);
3298 while ((worker
= first_idle_worker(pool
)))
3299 destroy_worker(worker
);
3300 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3301 spin_unlock_irq(&pool
->lock
);
3303 mutex_lock(&pool
->attach_mutex
);
3304 if (!list_empty(&pool
->workers
))
3305 pool
->detach_completion
= &detach_completion
;
3306 mutex_unlock(&pool
->attach_mutex
);
3308 if (pool
->detach_completion
)
3309 wait_for_completion(pool
->detach_completion
);
3311 mutex_unlock(&pool
->manager_arb
);
3313 /* shut down the timers */
3314 del_timer_sync(&pool
->idle_timer
);
3315 del_timer_sync(&pool
->mayday_timer
);
3317 /* sched-RCU protected to allow dereferences from get_work_pool() */
3318 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3322 * get_unbound_pool - get a worker_pool with the specified attributes
3323 * @attrs: the attributes of the worker_pool to get
3325 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3326 * reference count and return it. If there already is a matching
3327 * worker_pool, it will be used; otherwise, this function attempts to
3330 * Should be called with wq_pool_mutex held.
3332 * Return: On success, a worker_pool with the same attributes as @attrs.
3333 * On failure, %NULL.
3335 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3337 u32 hash
= wqattrs_hash(attrs
);
3338 struct worker_pool
*pool
;
3340 int target_node
= NUMA_NO_NODE
;
3342 lockdep_assert_held(&wq_pool_mutex
);
3344 /* do we already have a matching pool? */
3345 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3346 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3352 /* if cpumask is contained inside a NUMA node, we belong to that node */
3353 if (wq_numa_enabled
) {
3354 for_each_node(node
) {
3355 if (cpumask_subset(attrs
->cpumask
,
3356 wq_numa_possible_cpumask
[node
])) {
3363 /* nope, create a new one */
3364 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3365 if (!pool
|| init_worker_pool(pool
) < 0)
3368 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3369 copy_workqueue_attrs(pool
->attrs
, attrs
);
3370 pool
->node
= target_node
;
3373 * no_numa isn't a worker_pool attribute, always clear it. See
3374 * 'struct workqueue_attrs' comments for detail.
3376 pool
->attrs
->no_numa
= false;
3378 if (worker_pool_assign_id(pool
) < 0)
3381 /* create and start the initial worker */
3382 if (!create_worker(pool
))
3386 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3391 put_unbound_pool(pool
);
3395 static void rcu_free_pwq(struct rcu_head
*rcu
)
3397 kmem_cache_free(pwq_cache
,
3398 container_of(rcu
, struct pool_workqueue
, rcu
));
3402 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3403 * and needs to be destroyed.
3405 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3407 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3408 unbound_release_work
);
3409 struct workqueue_struct
*wq
= pwq
->wq
;
3410 struct worker_pool
*pool
= pwq
->pool
;
3413 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3416 mutex_lock(&wq
->mutex
);
3417 list_del_rcu(&pwq
->pwqs_node
);
3418 is_last
= list_empty(&wq
->pwqs
);
3419 mutex_unlock(&wq
->mutex
);
3421 mutex_lock(&wq_pool_mutex
);
3422 put_unbound_pool(pool
);
3423 mutex_unlock(&wq_pool_mutex
);
3425 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3428 * If we're the last pwq going away, @wq is already dead and no one
3429 * is gonna access it anymore. Schedule RCU free.
3432 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
3436 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3437 * @pwq: target pool_workqueue
3439 * If @pwq isn't freezing, set @pwq->max_active to the associated
3440 * workqueue's saved_max_active and activate delayed work items
3441 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3443 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3445 struct workqueue_struct
*wq
= pwq
->wq
;
3446 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3448 /* for @wq->saved_max_active */
3449 lockdep_assert_held(&wq
->mutex
);
3451 /* fast exit for non-freezable wqs */
3452 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3455 spin_lock_irq(&pwq
->pool
->lock
);
3458 * During [un]freezing, the caller is responsible for ensuring that
3459 * this function is called at least once after @workqueue_freezing
3460 * is updated and visible.
3462 if (!freezable
|| !workqueue_freezing
) {
3463 pwq
->max_active
= wq
->saved_max_active
;
3465 while (!list_empty(&pwq
->delayed_works
) &&
3466 pwq
->nr_active
< pwq
->max_active
)
3467 pwq_activate_first_delayed(pwq
);
3470 * Need to kick a worker after thawed or an unbound wq's
3471 * max_active is bumped. It's a slow path. Do it always.
3473 wake_up_worker(pwq
->pool
);
3475 pwq
->max_active
= 0;
3478 spin_unlock_irq(&pwq
->pool
->lock
);
3481 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3482 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3483 struct worker_pool
*pool
)
3485 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3487 memset(pwq
, 0, sizeof(*pwq
));
3491 pwq
->flush_color
= -1;
3493 INIT_LIST_HEAD(&pwq
->delayed_works
);
3494 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3495 INIT_LIST_HEAD(&pwq
->mayday_node
);
3496 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3499 /* sync @pwq with the current state of its associated wq and link it */
3500 static void link_pwq(struct pool_workqueue
*pwq
)
3502 struct workqueue_struct
*wq
= pwq
->wq
;
3504 lockdep_assert_held(&wq
->mutex
);
3506 /* may be called multiple times, ignore if already linked */
3507 if (!list_empty(&pwq
->pwqs_node
))
3510 /* set the matching work_color */
3511 pwq
->work_color
= wq
->work_color
;
3513 /* sync max_active to the current setting */
3514 pwq_adjust_max_active(pwq
);
3517 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3520 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3521 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3522 const struct workqueue_attrs
*attrs
)
3524 struct worker_pool
*pool
;
3525 struct pool_workqueue
*pwq
;
3527 lockdep_assert_held(&wq_pool_mutex
);
3529 pool
= get_unbound_pool(attrs
);
3533 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3535 put_unbound_pool(pool
);
3539 init_pwq(pwq
, wq
, pool
);
3544 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3545 * @attrs: the wq_attrs of the default pwq of the target workqueue
3546 * @node: the target NUMA node
3547 * @cpu_going_down: if >= 0, the CPU to consider as offline
3548 * @cpumask: outarg, the resulting cpumask
3550 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3551 * @cpu_going_down is >= 0, that cpu is considered offline during
3552 * calculation. The result is stored in @cpumask.
3554 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3555 * enabled and @node has online CPUs requested by @attrs, the returned
3556 * cpumask is the intersection of the possible CPUs of @node and
3559 * The caller is responsible for ensuring that the cpumask of @node stays
3562 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3565 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3566 int cpu_going_down
, cpumask_t
*cpumask
)
3568 if (!wq_numa_enabled
|| attrs
->no_numa
)
3571 /* does @node have any online CPUs @attrs wants? */
3572 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3573 if (cpu_going_down
>= 0)
3574 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3576 if (cpumask_empty(cpumask
))
3579 /* yeap, return possible CPUs in @node that @attrs wants */
3580 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3581 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3584 cpumask_copy(cpumask
, attrs
->cpumask
);
3588 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3589 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3591 struct pool_workqueue
*pwq
)
3593 struct pool_workqueue
*old_pwq
;
3595 lockdep_assert_held(&wq_pool_mutex
);
3596 lockdep_assert_held(&wq
->mutex
);
3598 /* link_pwq() can handle duplicate calls */
3601 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3602 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3606 /* context to store the prepared attrs & pwqs before applying */
3607 struct apply_wqattrs_ctx
{
3608 struct workqueue_struct
*wq
; /* target workqueue */
3609 struct workqueue_attrs
*attrs
; /* attrs to apply */
3610 struct list_head list
; /* queued for batching commit */
3611 struct pool_workqueue
*dfl_pwq
;
3612 struct pool_workqueue
*pwq_tbl
[];
3615 /* free the resources after success or abort */
3616 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3622 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3623 put_pwq_unlocked(ctx
->dfl_pwq
);
3625 free_workqueue_attrs(ctx
->attrs
);
3631 /* allocate the attrs and pwqs for later installation */
3632 static struct apply_wqattrs_ctx
*
3633 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3634 const struct workqueue_attrs
*attrs
)
3636 struct apply_wqattrs_ctx
*ctx
;
3637 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3640 lockdep_assert_held(&wq_pool_mutex
);
3642 ctx
= kzalloc(sizeof(*ctx
) + nr_node_ids
* sizeof(ctx
->pwq_tbl
[0]),
3645 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3646 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3647 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3651 * Calculate the attrs of the default pwq.
3652 * If the user configured cpumask doesn't overlap with the
3653 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3655 copy_workqueue_attrs(new_attrs
, attrs
);
3656 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3657 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3658 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3661 * We may create multiple pwqs with differing cpumasks. Make a
3662 * copy of @new_attrs which will be modified and used to obtain
3665 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3668 * If something goes wrong during CPU up/down, we'll fall back to
3669 * the default pwq covering whole @attrs->cpumask. Always create
3670 * it even if we don't use it immediately.
3672 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3676 for_each_node(node
) {
3677 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3678 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3679 if (!ctx
->pwq_tbl
[node
])
3682 ctx
->dfl_pwq
->refcnt
++;
3683 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3687 /* save the user configured attrs and sanitize it. */
3688 copy_workqueue_attrs(new_attrs
, attrs
);
3689 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3690 ctx
->attrs
= new_attrs
;
3693 free_workqueue_attrs(tmp_attrs
);
3697 free_workqueue_attrs(tmp_attrs
);
3698 free_workqueue_attrs(new_attrs
);
3699 apply_wqattrs_cleanup(ctx
);
3703 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3704 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3708 /* all pwqs have been created successfully, let's install'em */
3709 mutex_lock(&ctx
->wq
->mutex
);
3711 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3713 /* save the previous pwq and install the new one */
3715 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
3716 ctx
->pwq_tbl
[node
]);
3718 /* @dfl_pwq might not have been used, ensure it's linked */
3719 link_pwq(ctx
->dfl_pwq
);
3720 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
3722 mutex_unlock(&ctx
->wq
->mutex
);
3725 static void apply_wqattrs_lock(void)
3727 /* CPUs should stay stable across pwq creations and installations */
3729 mutex_lock(&wq_pool_mutex
);
3732 static void apply_wqattrs_unlock(void)
3734 mutex_unlock(&wq_pool_mutex
);
3738 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
3739 const struct workqueue_attrs
*attrs
)
3741 struct apply_wqattrs_ctx
*ctx
;
3743 /* only unbound workqueues can change attributes */
3744 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3747 /* creating multiple pwqs breaks ordering guarantee */
3748 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3751 ctx
= apply_wqattrs_prepare(wq
, attrs
);
3755 /* the ctx has been prepared successfully, let's commit it */
3756 apply_wqattrs_commit(ctx
);
3757 apply_wqattrs_cleanup(ctx
);
3763 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3764 * @wq: the target workqueue
3765 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3767 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3768 * machines, this function maps a separate pwq to each NUMA node with
3769 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3770 * NUMA node it was issued on. Older pwqs are released as in-flight work
3771 * items finish. Note that a work item which repeatedly requeues itself
3772 * back-to-back will stay on its current pwq.
3774 * Performs GFP_KERNEL allocations.
3776 * Return: 0 on success and -errno on failure.
3778 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3779 const struct workqueue_attrs
*attrs
)
3783 apply_wqattrs_lock();
3784 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
3785 apply_wqattrs_unlock();
3791 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3792 * @wq: the target workqueue
3793 * @cpu: the CPU coming up or going down
3794 * @online: whether @cpu is coming up or going down
3796 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3797 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3800 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3801 * falls back to @wq->dfl_pwq which may not be optimal but is always
3804 * Note that when the last allowed CPU of a NUMA node goes offline for a
3805 * workqueue with a cpumask spanning multiple nodes, the workers which were
3806 * already executing the work items for the workqueue will lose their CPU
3807 * affinity and may execute on any CPU. This is similar to how per-cpu
3808 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3809 * affinity, it's the user's responsibility to flush the work item from
3812 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3815 int node
= cpu_to_node(cpu
);
3816 int cpu_off
= online
? -1 : cpu
;
3817 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3818 struct workqueue_attrs
*target_attrs
;
3821 lockdep_assert_held(&wq_pool_mutex
);
3823 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
3824 wq
->unbound_attrs
->no_numa
)
3828 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3829 * Let's use a preallocated one. The following buf is protected by
3830 * CPU hotplug exclusion.
3832 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3833 cpumask
= target_attrs
->cpumask
;
3835 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3836 pwq
= unbound_pwq_by_node(wq
, node
);
3839 * Let's determine what needs to be done. If the target cpumask is
3840 * different from the default pwq's, we need to compare it to @pwq's
3841 * and create a new one if they don't match. If the target cpumask
3842 * equals the default pwq's, the default pwq should be used.
3844 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
3845 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3851 /* create a new pwq */
3852 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3854 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3859 /* Install the new pwq. */
3860 mutex_lock(&wq
->mutex
);
3861 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3865 mutex_lock(&wq
->mutex
);
3866 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3867 get_pwq(wq
->dfl_pwq
);
3868 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3869 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
3871 mutex_unlock(&wq
->mutex
);
3872 put_pwq_unlocked(old_pwq
);
3875 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3877 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3880 if (!(wq
->flags
& WQ_UNBOUND
)) {
3881 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3885 for_each_possible_cpu(cpu
) {
3886 struct pool_workqueue
*pwq
=
3887 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3888 struct worker_pool
*cpu_pools
=
3889 per_cpu(cpu_worker_pools
, cpu
);
3891 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
3893 mutex_lock(&wq
->mutex
);
3895 mutex_unlock(&wq
->mutex
);
3898 } else if (wq
->flags
& __WQ_ORDERED
) {
3899 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
3900 /* there should only be single pwq for ordering guarantee */
3901 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
3902 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
3903 "ordering guarantee broken for workqueue %s\n", wq
->name
);
3906 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3910 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3913 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3915 if (max_active
< 1 || max_active
> lim
)
3916 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3917 max_active
, name
, 1, lim
);
3919 return clamp_val(max_active
, 1, lim
);
3922 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3925 struct lock_class_key
*key
,
3926 const char *lock_name
, ...)
3928 size_t tbl_size
= 0;
3930 struct workqueue_struct
*wq
;
3931 struct pool_workqueue
*pwq
;
3933 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3934 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
3935 flags
|= WQ_UNBOUND
;
3937 /* allocate wq and format name */
3938 if (flags
& WQ_UNBOUND
)
3939 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
3941 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
3945 if (flags
& WQ_UNBOUND
) {
3946 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3947 if (!wq
->unbound_attrs
)
3951 va_start(args
, lock_name
);
3952 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
3955 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3956 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3960 wq
->saved_max_active
= max_active
;
3961 mutex_init(&wq
->mutex
);
3962 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3963 INIT_LIST_HEAD(&wq
->pwqs
);
3964 INIT_LIST_HEAD(&wq
->flusher_queue
);
3965 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3966 INIT_LIST_HEAD(&wq
->maydays
);
3968 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3969 INIT_LIST_HEAD(&wq
->list
);
3971 if (alloc_and_link_pwqs(wq
) < 0)
3975 * Workqueues which may be used during memory reclaim should
3976 * have a rescuer to guarantee forward progress.
3978 if (flags
& WQ_MEM_RECLAIM
) {
3979 struct worker
*rescuer
;
3981 rescuer
= alloc_worker(NUMA_NO_NODE
);
3985 rescuer
->rescue_wq
= wq
;
3986 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3988 if (IS_ERR(rescuer
->task
)) {
3993 wq
->rescuer
= rescuer
;
3994 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
3995 wake_up_process(rescuer
->task
);
3998 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4002 * wq_pool_mutex protects global freeze state and workqueues list.
4003 * Grab it, adjust max_active and add the new @wq to workqueues
4006 mutex_lock(&wq_pool_mutex
);
4008 mutex_lock(&wq
->mutex
);
4009 for_each_pwq(pwq
, wq
)
4010 pwq_adjust_max_active(pwq
);
4011 mutex_unlock(&wq
->mutex
);
4013 list_add_tail_rcu(&wq
->list
, &workqueues
);
4015 mutex_unlock(&wq_pool_mutex
);
4020 free_workqueue_attrs(wq
->unbound_attrs
);
4024 destroy_workqueue(wq
);
4027 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4030 * destroy_workqueue - safely terminate a workqueue
4031 * @wq: target workqueue
4033 * Safely destroy a workqueue. All work currently pending will be done first.
4035 void destroy_workqueue(struct workqueue_struct
*wq
)
4037 struct pool_workqueue
*pwq
;
4040 /* drain it before proceeding with destruction */
4041 drain_workqueue(wq
);
4044 mutex_lock(&wq
->mutex
);
4045 for_each_pwq(pwq
, wq
) {
4048 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4049 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4050 mutex_unlock(&wq
->mutex
);
4055 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4056 WARN_ON(pwq
->nr_active
) ||
4057 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4058 mutex_unlock(&wq
->mutex
);
4062 mutex_unlock(&wq
->mutex
);
4065 * wq list is used to freeze wq, remove from list after
4066 * flushing is complete in case freeze races us.
4068 mutex_lock(&wq_pool_mutex
);
4069 list_del_rcu(&wq
->list
);
4070 mutex_unlock(&wq_pool_mutex
);
4072 workqueue_sysfs_unregister(wq
);
4075 kthread_stop(wq
->rescuer
->task
);
4077 if (!(wq
->flags
& WQ_UNBOUND
)) {
4079 * The base ref is never dropped on per-cpu pwqs. Directly
4080 * schedule RCU free.
4082 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
4085 * We're the sole accessor of @wq at this point. Directly
4086 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4087 * @wq will be freed when the last pwq is released.
4089 for_each_node(node
) {
4090 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4091 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4092 put_pwq_unlocked(pwq
);
4096 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4097 * put. Don't access it afterwards.
4101 put_pwq_unlocked(pwq
);
4104 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4107 * workqueue_set_max_active - adjust max_active of a workqueue
4108 * @wq: target workqueue
4109 * @max_active: new max_active value.
4111 * Set max_active of @wq to @max_active.
4114 * Don't call from IRQ context.
4116 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4118 struct pool_workqueue
*pwq
;
4120 /* disallow meddling with max_active for ordered workqueues */
4121 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4124 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4126 mutex_lock(&wq
->mutex
);
4128 wq
->saved_max_active
= max_active
;
4130 for_each_pwq(pwq
, wq
)
4131 pwq_adjust_max_active(pwq
);
4133 mutex_unlock(&wq
->mutex
);
4135 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4138 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4140 * Determine whether %current is a workqueue rescuer. Can be used from
4141 * work functions to determine whether it's being run off the rescuer task.
4143 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4145 bool current_is_workqueue_rescuer(void)
4147 struct worker
*worker
= current_wq_worker();
4149 return worker
&& worker
->rescue_wq
;
4153 * workqueue_congested - test whether a workqueue is congested
4154 * @cpu: CPU in question
4155 * @wq: target workqueue
4157 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4158 * no synchronization around this function and the test result is
4159 * unreliable and only useful as advisory hints or for debugging.
4161 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4162 * Note that both per-cpu and unbound workqueues may be associated with
4163 * multiple pool_workqueues which have separate congested states. A
4164 * workqueue being congested on one CPU doesn't mean the workqueue is also
4165 * contested on other CPUs / NUMA nodes.
4168 * %true if congested, %false otherwise.
4170 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4172 struct pool_workqueue
*pwq
;
4175 rcu_read_lock_sched();
4177 if (cpu
== WORK_CPU_UNBOUND
)
4178 cpu
= smp_processor_id();
4180 if (!(wq
->flags
& WQ_UNBOUND
))
4181 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4183 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4185 ret
= !list_empty(&pwq
->delayed_works
);
4186 rcu_read_unlock_sched();
4190 EXPORT_SYMBOL_GPL(workqueue_congested
);
4193 * work_busy - test whether a work is currently pending or running
4194 * @work: the work to be tested
4196 * Test whether @work is currently pending or running. There is no
4197 * synchronization around this function and the test result is
4198 * unreliable and only useful as advisory hints or for debugging.
4201 * OR'd bitmask of WORK_BUSY_* bits.
4203 unsigned int work_busy(struct work_struct
*work
)
4205 struct worker_pool
*pool
;
4206 unsigned long flags
;
4207 unsigned int ret
= 0;
4209 if (work_pending(work
))
4210 ret
|= WORK_BUSY_PENDING
;
4212 local_irq_save(flags
);
4213 pool
= get_work_pool(work
);
4215 spin_lock(&pool
->lock
);
4216 if (find_worker_executing_work(pool
, work
))
4217 ret
|= WORK_BUSY_RUNNING
;
4218 spin_unlock(&pool
->lock
);
4220 local_irq_restore(flags
);
4224 EXPORT_SYMBOL_GPL(work_busy
);
4227 * set_worker_desc - set description for the current work item
4228 * @fmt: printf-style format string
4229 * @...: arguments for the format string
4231 * This function can be called by a running work function to describe what
4232 * the work item is about. If the worker task gets dumped, this
4233 * information will be printed out together to help debugging. The
4234 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4236 void set_worker_desc(const char *fmt
, ...)
4238 struct worker
*worker
= current_wq_worker();
4242 va_start(args
, fmt
);
4243 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4245 worker
->desc_valid
= true;
4250 * print_worker_info - print out worker information and description
4251 * @log_lvl: the log level to use when printing
4252 * @task: target task
4254 * If @task is a worker and currently executing a work item, print out the
4255 * name of the workqueue being serviced and worker description set with
4256 * set_worker_desc() by the currently executing work item.
4258 * This function can be safely called on any task as long as the
4259 * task_struct itself is accessible. While safe, this function isn't
4260 * synchronized and may print out mixups or garbages of limited length.
4262 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4264 work_func_t
*fn
= NULL
;
4265 char name
[WQ_NAME_LEN
] = { };
4266 char desc
[WORKER_DESC_LEN
] = { };
4267 struct pool_workqueue
*pwq
= NULL
;
4268 struct workqueue_struct
*wq
= NULL
;
4269 bool desc_valid
= false;
4270 struct worker
*worker
;
4272 if (!(task
->flags
& PF_WQ_WORKER
))
4276 * This function is called without any synchronization and @task
4277 * could be in any state. Be careful with dereferences.
4279 worker
= probe_kthread_data(task
);
4282 * Carefully copy the associated workqueue's workfn and name. Keep
4283 * the original last '\0' in case the original contains garbage.
4285 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4286 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4287 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4288 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4290 /* copy worker description */
4291 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4293 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4295 if (fn
|| name
[0] || desc
[0]) {
4296 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4298 pr_cont(" (%s)", desc
);
4303 static void pr_cont_pool_info(struct worker_pool
*pool
)
4305 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4306 if (pool
->node
!= NUMA_NO_NODE
)
4307 pr_cont(" node=%d", pool
->node
);
4308 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4311 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4313 if (work
->func
== wq_barrier_func
) {
4314 struct wq_barrier
*barr
;
4316 barr
= container_of(work
, struct wq_barrier
, work
);
4318 pr_cont("%s BAR(%d)", comma
? "," : "",
4319 task_pid_nr(barr
->task
));
4321 pr_cont("%s %pf", comma
? "," : "", work
->func
);
4325 static void show_pwq(struct pool_workqueue
*pwq
)
4327 struct worker_pool
*pool
= pwq
->pool
;
4328 struct work_struct
*work
;
4329 struct worker
*worker
;
4330 bool has_in_flight
= false, has_pending
= false;
4333 pr_info(" pwq %d:", pool
->id
);
4334 pr_cont_pool_info(pool
);
4336 pr_cont(" active=%d/%d%s\n", pwq
->nr_active
, pwq
->max_active
,
4337 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4339 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4340 if (worker
->current_pwq
== pwq
) {
4341 has_in_flight
= true;
4345 if (has_in_flight
) {
4348 pr_info(" in-flight:");
4349 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4350 if (worker
->current_pwq
!= pwq
)
4353 pr_cont("%s %d%s:%pf", comma
? "," : "",
4354 task_pid_nr(worker
->task
),
4355 worker
== pwq
->wq
->rescuer
? "(RESCUER)" : "",
4356 worker
->current_func
);
4357 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4358 pr_cont_work(false, work
);
4364 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4365 if (get_work_pwq(work
) == pwq
) {
4373 pr_info(" pending:");
4374 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4375 if (get_work_pwq(work
) != pwq
)
4378 pr_cont_work(comma
, work
);
4379 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4384 if (!list_empty(&pwq
->delayed_works
)) {
4387 pr_info(" delayed:");
4388 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4389 pr_cont_work(comma
, work
);
4390 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4397 * show_workqueue_state - dump workqueue state
4399 * Called from a sysrq handler and prints out all busy workqueues and
4402 void show_workqueue_state(void)
4404 struct workqueue_struct
*wq
;
4405 struct worker_pool
*pool
;
4406 unsigned long flags
;
4409 rcu_read_lock_sched();
4411 pr_info("Showing busy workqueues and worker pools:\n");
4413 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4414 struct pool_workqueue
*pwq
;
4417 for_each_pwq(pwq
, wq
) {
4418 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4426 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4428 for_each_pwq(pwq
, wq
) {
4429 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4430 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4432 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4436 for_each_pool(pool
, pi
) {
4437 struct worker
*worker
;
4440 spin_lock_irqsave(&pool
->lock
, flags
);
4441 if (pool
->nr_workers
== pool
->nr_idle
)
4444 pr_info("pool %d:", pool
->id
);
4445 pr_cont_pool_info(pool
);
4446 pr_cont(" hung=%us workers=%d",
4447 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4450 pr_cont(" manager: %d",
4451 task_pid_nr(pool
->manager
->task
));
4452 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4453 pr_cont(" %s%d", first
? "idle: " : "",
4454 task_pid_nr(worker
->task
));
4459 spin_unlock_irqrestore(&pool
->lock
, flags
);
4462 rcu_read_unlock_sched();
4468 * There are two challenges in supporting CPU hotplug. Firstly, there
4469 * are a lot of assumptions on strong associations among work, pwq and
4470 * pool which make migrating pending and scheduled works very
4471 * difficult to implement without impacting hot paths. Secondly,
4472 * worker pools serve mix of short, long and very long running works making
4473 * blocked draining impractical.
4475 * This is solved by allowing the pools to be disassociated from the CPU
4476 * running as an unbound one and allowing it to be reattached later if the
4477 * cpu comes back online.
4480 static void wq_unbind_fn(struct work_struct
*work
)
4482 int cpu
= smp_processor_id();
4483 struct worker_pool
*pool
;
4484 struct worker
*worker
;
4486 for_each_cpu_worker_pool(pool
, cpu
) {
4487 mutex_lock(&pool
->attach_mutex
);
4488 spin_lock_irq(&pool
->lock
);
4491 * We've blocked all attach/detach operations. Make all workers
4492 * unbound and set DISASSOCIATED. Before this, all workers
4493 * except for the ones which are still executing works from
4494 * before the last CPU down must be on the cpu. After
4495 * this, they may become diasporas.
4497 for_each_pool_worker(worker
, pool
)
4498 worker
->flags
|= WORKER_UNBOUND
;
4500 pool
->flags
|= POOL_DISASSOCIATED
;
4502 spin_unlock_irq(&pool
->lock
);
4503 mutex_unlock(&pool
->attach_mutex
);
4506 * Call schedule() so that we cross rq->lock and thus can
4507 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4508 * This is necessary as scheduler callbacks may be invoked
4514 * Sched callbacks are disabled now. Zap nr_running.
4515 * After this, nr_running stays zero and need_more_worker()
4516 * and keep_working() are always true as long as the
4517 * worklist is not empty. This pool now behaves as an
4518 * unbound (in terms of concurrency management) pool which
4519 * are served by workers tied to the pool.
4521 atomic_set(&pool
->nr_running
, 0);
4524 * With concurrency management just turned off, a busy
4525 * worker blocking could lead to lengthy stalls. Kick off
4526 * unbound chain execution of currently pending work items.
4528 spin_lock_irq(&pool
->lock
);
4529 wake_up_worker(pool
);
4530 spin_unlock_irq(&pool
->lock
);
4535 * rebind_workers - rebind all workers of a pool to the associated CPU
4536 * @pool: pool of interest
4538 * @pool->cpu is coming online. Rebind all workers to the CPU.
4540 static void rebind_workers(struct worker_pool
*pool
)
4542 struct worker
*worker
;
4544 lockdep_assert_held(&pool
->attach_mutex
);
4547 * Restore CPU affinity of all workers. As all idle workers should
4548 * be on the run-queue of the associated CPU before any local
4549 * wake-ups for concurrency management happen, restore CPU affinity
4550 * of all workers first and then clear UNBOUND. As we're called
4551 * from CPU_ONLINE, the following shouldn't fail.
4553 for_each_pool_worker(worker
, pool
)
4554 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4555 pool
->attrs
->cpumask
) < 0);
4557 spin_lock_irq(&pool
->lock
);
4558 pool
->flags
&= ~POOL_DISASSOCIATED
;
4560 for_each_pool_worker(worker
, pool
) {
4561 unsigned int worker_flags
= worker
->flags
;
4564 * A bound idle worker should actually be on the runqueue
4565 * of the associated CPU for local wake-ups targeting it to
4566 * work. Kick all idle workers so that they migrate to the
4567 * associated CPU. Doing this in the same loop as
4568 * replacing UNBOUND with REBOUND is safe as no worker will
4569 * be bound before @pool->lock is released.
4571 if (worker_flags
& WORKER_IDLE
)
4572 wake_up_process(worker
->task
);
4575 * We want to clear UNBOUND but can't directly call
4576 * worker_clr_flags() or adjust nr_running. Atomically
4577 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4578 * @worker will clear REBOUND using worker_clr_flags() when
4579 * it initiates the next execution cycle thus restoring
4580 * concurrency management. Note that when or whether
4581 * @worker clears REBOUND doesn't affect correctness.
4583 * ACCESS_ONCE() is necessary because @worker->flags may be
4584 * tested without holding any lock in
4585 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4586 * fail incorrectly leading to premature concurrency
4587 * management operations.
4589 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4590 worker_flags
|= WORKER_REBOUND
;
4591 worker_flags
&= ~WORKER_UNBOUND
;
4592 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4595 spin_unlock_irq(&pool
->lock
);
4599 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4600 * @pool: unbound pool of interest
4601 * @cpu: the CPU which is coming up
4603 * An unbound pool may end up with a cpumask which doesn't have any online
4604 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4605 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4606 * online CPU before, cpus_allowed of all its workers should be restored.
4608 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4610 static cpumask_t cpumask
;
4611 struct worker
*worker
;
4613 lockdep_assert_held(&pool
->attach_mutex
);
4615 /* is @cpu allowed for @pool? */
4616 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4619 /* is @cpu the only online CPU? */
4620 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4621 if (cpumask_weight(&cpumask
) != 1)
4624 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4625 for_each_pool_worker(worker
, pool
)
4626 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4627 pool
->attrs
->cpumask
) < 0);
4631 * Workqueues should be brought up before normal priority CPU notifiers.
4632 * This will be registered high priority CPU notifier.
4634 static int workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4635 unsigned long action
,
4638 int cpu
= (unsigned long)hcpu
;
4639 struct worker_pool
*pool
;
4640 struct workqueue_struct
*wq
;
4643 switch (action
& ~CPU_TASKS_FROZEN
) {
4644 case CPU_UP_PREPARE
:
4645 for_each_cpu_worker_pool(pool
, cpu
) {
4646 if (pool
->nr_workers
)
4648 if (!create_worker(pool
))
4653 case CPU_DOWN_FAILED
:
4655 mutex_lock(&wq_pool_mutex
);
4657 for_each_pool(pool
, pi
) {
4658 mutex_lock(&pool
->attach_mutex
);
4660 if (pool
->cpu
== cpu
)
4661 rebind_workers(pool
);
4662 else if (pool
->cpu
< 0)
4663 restore_unbound_workers_cpumask(pool
, cpu
);
4665 mutex_unlock(&pool
->attach_mutex
);
4668 /* update NUMA affinity of unbound workqueues */
4669 list_for_each_entry(wq
, &workqueues
, list
)
4670 wq_update_unbound_numa(wq
, cpu
, true);
4672 mutex_unlock(&wq_pool_mutex
);
4679 * Workqueues should be brought down after normal priority CPU notifiers.
4680 * This will be registered as low priority CPU notifier.
4682 static int workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4683 unsigned long action
,
4686 int cpu
= (unsigned long)hcpu
;
4687 struct work_struct unbind_work
;
4688 struct workqueue_struct
*wq
;
4690 switch (action
& ~CPU_TASKS_FROZEN
) {
4691 case CPU_DOWN_PREPARE
:
4692 /* unbinding per-cpu workers should happen on the local CPU */
4693 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4694 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4696 /* update NUMA affinity of unbound workqueues */
4697 mutex_lock(&wq_pool_mutex
);
4698 list_for_each_entry(wq
, &workqueues
, list
)
4699 wq_update_unbound_numa(wq
, cpu
, false);
4700 mutex_unlock(&wq_pool_mutex
);
4702 /* wait for per-cpu unbinding to finish */
4703 flush_work(&unbind_work
);
4704 destroy_work_on_stack(&unbind_work
);
4712 struct work_for_cpu
{
4713 struct work_struct work
;
4719 static void work_for_cpu_fn(struct work_struct
*work
)
4721 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4723 wfc
->ret
= wfc
->fn(wfc
->arg
);
4727 * work_on_cpu - run a function in user context on a particular cpu
4728 * @cpu: the cpu to run on
4729 * @fn: the function to run
4730 * @arg: the function arg
4732 * It is up to the caller to ensure that the cpu doesn't go offline.
4733 * The caller must not hold any locks which would prevent @fn from completing.
4735 * Return: The value @fn returns.
4737 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4739 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4741 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4742 schedule_work_on(cpu
, &wfc
.work
);
4743 flush_work(&wfc
.work
);
4744 destroy_work_on_stack(&wfc
.work
);
4747 EXPORT_SYMBOL_GPL(work_on_cpu
);
4748 #endif /* CONFIG_SMP */
4750 #ifdef CONFIG_FREEZER
4753 * freeze_workqueues_begin - begin freezing workqueues
4755 * Start freezing workqueues. After this function returns, all freezable
4756 * workqueues will queue new works to their delayed_works list instead of
4760 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4762 void freeze_workqueues_begin(void)
4764 struct workqueue_struct
*wq
;
4765 struct pool_workqueue
*pwq
;
4767 mutex_lock(&wq_pool_mutex
);
4769 WARN_ON_ONCE(workqueue_freezing
);
4770 workqueue_freezing
= true;
4772 list_for_each_entry(wq
, &workqueues
, list
) {
4773 mutex_lock(&wq
->mutex
);
4774 for_each_pwq(pwq
, wq
)
4775 pwq_adjust_max_active(pwq
);
4776 mutex_unlock(&wq
->mutex
);
4779 mutex_unlock(&wq_pool_mutex
);
4783 * freeze_workqueues_busy - are freezable workqueues still busy?
4785 * Check whether freezing is complete. This function must be called
4786 * between freeze_workqueues_begin() and thaw_workqueues().
4789 * Grabs and releases wq_pool_mutex.
4792 * %true if some freezable workqueues are still busy. %false if freezing
4795 bool freeze_workqueues_busy(void)
4798 struct workqueue_struct
*wq
;
4799 struct pool_workqueue
*pwq
;
4801 mutex_lock(&wq_pool_mutex
);
4803 WARN_ON_ONCE(!workqueue_freezing
);
4805 list_for_each_entry(wq
, &workqueues
, list
) {
4806 if (!(wq
->flags
& WQ_FREEZABLE
))
4809 * nr_active is monotonically decreasing. It's safe
4810 * to peek without lock.
4812 rcu_read_lock_sched();
4813 for_each_pwq(pwq
, wq
) {
4814 WARN_ON_ONCE(pwq
->nr_active
< 0);
4815 if (pwq
->nr_active
) {
4817 rcu_read_unlock_sched();
4821 rcu_read_unlock_sched();
4824 mutex_unlock(&wq_pool_mutex
);
4829 * thaw_workqueues - thaw workqueues
4831 * Thaw workqueues. Normal queueing is restored and all collected
4832 * frozen works are transferred to their respective pool worklists.
4835 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4837 void thaw_workqueues(void)
4839 struct workqueue_struct
*wq
;
4840 struct pool_workqueue
*pwq
;
4842 mutex_lock(&wq_pool_mutex
);
4844 if (!workqueue_freezing
)
4847 workqueue_freezing
= false;
4849 /* restore max_active and repopulate worklist */
4850 list_for_each_entry(wq
, &workqueues
, list
) {
4851 mutex_lock(&wq
->mutex
);
4852 for_each_pwq(pwq
, wq
)
4853 pwq_adjust_max_active(pwq
);
4854 mutex_unlock(&wq
->mutex
);
4858 mutex_unlock(&wq_pool_mutex
);
4860 #endif /* CONFIG_FREEZER */
4862 static int workqueue_apply_unbound_cpumask(void)
4866 struct workqueue_struct
*wq
;
4867 struct apply_wqattrs_ctx
*ctx
, *n
;
4869 lockdep_assert_held(&wq_pool_mutex
);
4871 list_for_each_entry(wq
, &workqueues
, list
) {
4872 if (!(wq
->flags
& WQ_UNBOUND
))
4874 /* creating multiple pwqs breaks ordering guarantee */
4875 if (wq
->flags
& __WQ_ORDERED
)
4878 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
4884 list_add_tail(&ctx
->list
, &ctxs
);
4887 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
4889 apply_wqattrs_commit(ctx
);
4890 apply_wqattrs_cleanup(ctx
);
4897 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4898 * @cpumask: the cpumask to set
4900 * The low-level workqueues cpumask is a global cpumask that limits
4901 * the affinity of all unbound workqueues. This function check the @cpumask
4902 * and apply it to all unbound workqueues and updates all pwqs of them.
4904 * Retun: 0 - Success
4905 * -EINVAL - Invalid @cpumask
4906 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4908 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
4911 cpumask_var_t saved_cpumask
;
4913 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
4916 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
4917 if (!cpumask_empty(cpumask
)) {
4918 apply_wqattrs_lock();
4920 /* save the old wq_unbound_cpumask. */
4921 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
4923 /* update wq_unbound_cpumask at first and apply it to wqs. */
4924 cpumask_copy(wq_unbound_cpumask
, cpumask
);
4925 ret
= workqueue_apply_unbound_cpumask();
4927 /* restore the wq_unbound_cpumask when failed. */
4929 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
4931 apply_wqattrs_unlock();
4934 free_cpumask_var(saved_cpumask
);
4940 * Workqueues with WQ_SYSFS flag set is visible to userland via
4941 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4942 * following attributes.
4944 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4945 * max_active RW int : maximum number of in-flight work items
4947 * Unbound workqueues have the following extra attributes.
4949 * id RO int : the associated pool ID
4950 * nice RW int : nice value of the workers
4951 * cpumask RW mask : bitmask of allowed CPUs for the workers
4954 struct workqueue_struct
*wq
;
4958 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
4960 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
4965 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
4968 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4970 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
4972 static DEVICE_ATTR_RO(per_cpu
);
4974 static ssize_t
max_active_show(struct device
*dev
,
4975 struct device_attribute
*attr
, char *buf
)
4977 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4979 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
4982 static ssize_t
max_active_store(struct device
*dev
,
4983 struct device_attribute
*attr
, const char *buf
,
4986 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4989 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
4992 workqueue_set_max_active(wq
, val
);
4995 static DEVICE_ATTR_RW(max_active
);
4997 static struct attribute
*wq_sysfs_attrs
[] = {
4998 &dev_attr_per_cpu
.attr
,
4999 &dev_attr_max_active
.attr
,
5002 ATTRIBUTE_GROUPS(wq_sysfs
);
5004 static ssize_t
wq_pool_ids_show(struct device
*dev
,
5005 struct device_attribute
*attr
, char *buf
)
5007 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5008 const char *delim
= "";
5009 int node
, written
= 0;
5011 rcu_read_lock_sched();
5012 for_each_node(node
) {
5013 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
5014 "%s%d:%d", delim
, node
,
5015 unbound_pwq_by_node(wq
, node
)->pool
->id
);
5018 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
5019 rcu_read_unlock_sched();
5024 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
5027 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5030 mutex_lock(&wq
->mutex
);
5031 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
5032 mutex_unlock(&wq
->mutex
);
5037 /* prepare workqueue_attrs for sysfs store operations */
5038 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
5040 struct workqueue_attrs
*attrs
;
5042 lockdep_assert_held(&wq_pool_mutex
);
5044 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
5048 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
5052 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
5053 const char *buf
, size_t count
)
5055 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5056 struct workqueue_attrs
*attrs
;
5059 apply_wqattrs_lock();
5061 attrs
= wq_sysfs_prep_attrs(wq
);
5065 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5066 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5067 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5072 apply_wqattrs_unlock();
5073 free_workqueue_attrs(attrs
);
5074 return ret
?: count
;
5077 static ssize_t
wq_cpumask_show(struct device
*dev
,
5078 struct device_attribute
*attr
, char *buf
)
5080 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5083 mutex_lock(&wq
->mutex
);
5084 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5085 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5086 mutex_unlock(&wq
->mutex
);
5090 static ssize_t
wq_cpumask_store(struct device
*dev
,
5091 struct device_attribute
*attr
,
5092 const char *buf
, size_t count
)
5094 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5095 struct workqueue_attrs
*attrs
;
5098 apply_wqattrs_lock();
5100 attrs
= wq_sysfs_prep_attrs(wq
);
5104 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5106 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5109 apply_wqattrs_unlock();
5110 free_workqueue_attrs(attrs
);
5111 return ret
?: count
;
5114 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5117 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5120 mutex_lock(&wq
->mutex
);
5121 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5122 !wq
->unbound_attrs
->no_numa
);
5123 mutex_unlock(&wq
->mutex
);
5128 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5129 const char *buf
, size_t count
)
5131 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5132 struct workqueue_attrs
*attrs
;
5133 int v
, ret
= -ENOMEM
;
5135 apply_wqattrs_lock();
5137 attrs
= wq_sysfs_prep_attrs(wq
);
5142 if (sscanf(buf
, "%d", &v
) == 1) {
5143 attrs
->no_numa
= !v
;
5144 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5148 apply_wqattrs_unlock();
5149 free_workqueue_attrs(attrs
);
5150 return ret
?: count
;
5153 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5154 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5155 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5156 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5157 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5161 static struct bus_type wq_subsys
= {
5162 .name
= "workqueue",
5163 .dev_groups
= wq_sysfs_groups
,
5166 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5167 struct device_attribute
*attr
, char *buf
)
5171 mutex_lock(&wq_pool_mutex
);
5172 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5173 cpumask_pr_args(wq_unbound_cpumask
));
5174 mutex_unlock(&wq_pool_mutex
);
5179 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5180 struct device_attribute
*attr
, const char *buf
, size_t count
)
5182 cpumask_var_t cpumask
;
5185 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5188 ret
= cpumask_parse(buf
, cpumask
);
5190 ret
= workqueue_set_unbound_cpumask(cpumask
);
5192 free_cpumask_var(cpumask
);
5193 return ret
? ret
: count
;
5196 static struct device_attribute wq_sysfs_cpumask_attr
=
5197 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5198 wq_unbound_cpumask_store
);
5200 static int __init
wq_sysfs_init(void)
5204 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5208 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5210 core_initcall(wq_sysfs_init
);
5212 static void wq_device_release(struct device
*dev
)
5214 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5220 * workqueue_sysfs_register - make a workqueue visible in sysfs
5221 * @wq: the workqueue to register
5223 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5224 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5225 * which is the preferred method.
5227 * Workqueue user should use this function directly iff it wants to apply
5228 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5229 * apply_workqueue_attrs() may race against userland updating the
5232 * Return: 0 on success, -errno on failure.
5234 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5236 struct wq_device
*wq_dev
;
5240 * Adjusting max_active or creating new pwqs by applying
5241 * attributes breaks ordering guarantee. Disallow exposing ordered
5244 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
5247 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5252 wq_dev
->dev
.bus
= &wq_subsys
;
5253 wq_dev
->dev
.init_name
= wq
->name
;
5254 wq_dev
->dev
.release
= wq_device_release
;
5257 * unbound_attrs are created separately. Suppress uevent until
5258 * everything is ready.
5260 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5262 ret
= device_register(&wq_dev
->dev
);
5269 if (wq
->flags
& WQ_UNBOUND
) {
5270 struct device_attribute
*attr
;
5272 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5273 ret
= device_create_file(&wq_dev
->dev
, attr
);
5275 device_unregister(&wq_dev
->dev
);
5282 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5283 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5288 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5289 * @wq: the workqueue to unregister
5291 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5293 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5295 struct wq_device
*wq_dev
= wq
->wq_dev
;
5301 device_unregister(&wq_dev
->dev
);
5303 #else /* CONFIG_SYSFS */
5304 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5305 #endif /* CONFIG_SYSFS */
5308 * Workqueue watchdog.
5310 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5311 * flush dependency, a concurrency managed work item which stays RUNNING
5312 * indefinitely. Workqueue stalls can be very difficult to debug as the
5313 * usual warning mechanisms don't trigger and internal workqueue state is
5316 * Workqueue watchdog monitors all worker pools periodically and dumps
5317 * state if some pools failed to make forward progress for a while where
5318 * forward progress is defined as the first item on ->worklist changing.
5320 * This mechanism is controlled through the kernel parameter
5321 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5322 * corresponding sysfs parameter file.
5324 #ifdef CONFIG_WQ_WATCHDOG
5326 static void wq_watchdog_timer_fn(unsigned long data
);
5328 static unsigned long wq_watchdog_thresh
= 30;
5329 static struct timer_list wq_watchdog_timer
=
5330 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn
, 0, 0);
5332 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5333 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5335 static void wq_watchdog_reset_touched(void)
5339 wq_watchdog_touched
= jiffies
;
5340 for_each_possible_cpu(cpu
)
5341 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5344 static void wq_watchdog_timer_fn(unsigned long data
)
5346 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5347 bool lockup_detected
= false;
5348 struct worker_pool
*pool
;
5356 for_each_pool(pool
, pi
) {
5357 unsigned long pool_ts
, touched
, ts
;
5359 if (list_empty(&pool
->worklist
))
5362 /* get the latest of pool and touched timestamps */
5363 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5364 touched
= READ_ONCE(wq_watchdog_touched
);
5366 if (time_after(pool_ts
, touched
))
5371 if (pool
->cpu
>= 0) {
5372 unsigned long cpu_touched
=
5373 READ_ONCE(per_cpu(wq_watchdog_touched_cpu
,
5375 if (time_after(cpu_touched
, ts
))
5380 if (time_after(jiffies
, ts
+ thresh
)) {
5381 lockup_detected
= true;
5382 pr_emerg("BUG: workqueue lockup - pool");
5383 pr_cont_pool_info(pool
);
5384 pr_cont(" stuck for %us!\n",
5385 jiffies_to_msecs(jiffies
- pool_ts
) / 1000);
5391 if (lockup_detected
)
5392 show_workqueue_state();
5394 wq_watchdog_reset_touched();
5395 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5398 void wq_watchdog_touch(int cpu
)
5401 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5403 wq_watchdog_touched
= jiffies
;
5406 static void wq_watchdog_set_thresh(unsigned long thresh
)
5408 wq_watchdog_thresh
= 0;
5409 del_timer_sync(&wq_watchdog_timer
);
5412 wq_watchdog_thresh
= thresh
;
5413 wq_watchdog_reset_touched();
5414 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5418 static int wq_watchdog_param_set_thresh(const char *val
,
5419 const struct kernel_param
*kp
)
5421 unsigned long thresh
;
5424 ret
= kstrtoul(val
, 0, &thresh
);
5429 wq_watchdog_set_thresh(thresh
);
5431 wq_watchdog_thresh
= thresh
;
5436 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5437 .set
= wq_watchdog_param_set_thresh
,
5438 .get
= param_get_ulong
,
5441 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5444 static void wq_watchdog_init(void)
5446 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5449 #else /* CONFIG_WQ_WATCHDOG */
5451 static inline void wq_watchdog_init(void) { }
5453 #endif /* CONFIG_WQ_WATCHDOG */
5455 static void __init
wq_numa_init(void)
5460 if (num_possible_nodes() <= 1)
5463 if (wq_disable_numa
) {
5464 pr_info("workqueue: NUMA affinity support disabled\n");
5468 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
5469 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5472 * We want masks of possible CPUs of each node which isn't readily
5473 * available. Build one from cpu_to_node() which should have been
5474 * fully initialized by now.
5476 tbl
= kzalloc(nr_node_ids
* sizeof(tbl
[0]), GFP_KERNEL
);
5480 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5481 node_online(node
) ? node
: NUMA_NO_NODE
));
5483 for_each_possible_cpu(cpu
) {
5484 node
= cpu_to_node(cpu
);
5485 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5486 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5487 /* happens iff arch is bonkers, let's just proceed */
5490 cpumask_set_cpu(cpu
, tbl
[node
]);
5493 wq_numa_possible_cpumask
= tbl
;
5494 wq_numa_enabled
= true;
5497 static int __init
init_workqueues(void)
5499 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5502 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5504 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5505 cpumask_copy(wq_unbound_cpumask
, cpu_possible_mask
);
5507 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5509 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
5510 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
5514 /* initialize CPU pools */
5515 for_each_possible_cpu(cpu
) {
5516 struct worker_pool
*pool
;
5519 for_each_cpu_worker_pool(pool
, cpu
) {
5520 BUG_ON(init_worker_pool(pool
));
5522 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5523 pool
->attrs
->nice
= std_nice
[i
++];
5524 pool
->node
= cpu_to_node(cpu
);
5527 mutex_lock(&wq_pool_mutex
);
5528 BUG_ON(worker_pool_assign_id(pool
));
5529 mutex_unlock(&wq_pool_mutex
);
5533 /* create the initial worker */
5534 for_each_online_cpu(cpu
) {
5535 struct worker_pool
*pool
;
5537 for_each_cpu_worker_pool(pool
, cpu
) {
5538 pool
->flags
&= ~POOL_DISASSOCIATED
;
5539 BUG_ON(!create_worker(pool
));
5543 /* create default unbound and ordered wq attrs */
5544 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5545 struct workqueue_attrs
*attrs
;
5547 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5548 attrs
->nice
= std_nice
[i
];
5549 unbound_std_wq_attrs
[i
] = attrs
;
5552 * An ordered wq should have only one pwq as ordering is
5553 * guaranteed by max_active which is enforced by pwqs.
5554 * Turn off NUMA so that dfl_pwq is used for all nodes.
5556 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5557 attrs
->nice
= std_nice
[i
];
5558 attrs
->no_numa
= true;
5559 ordered_wq_attrs
[i
] = attrs
;
5562 system_wq
= alloc_workqueue("events", 0, 0);
5563 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5564 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5565 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5566 WQ_UNBOUND_MAX_ACTIVE
);
5567 system_freezable_wq
= alloc_workqueue("events_freezable",
5569 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5570 WQ_POWER_EFFICIENT
, 0);
5571 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5572 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5574 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5575 !system_unbound_wq
|| !system_freezable_wq
||
5576 !system_power_efficient_wq
||
5577 !system_freezable_power_efficient_wq
);
5583 early_initcall(init_workqueues
);