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/core-api/workqueue.rst 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/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
50 #include <linux/sched/isolation.h>
51 #include <linux/nmi.h>
53 #include "workqueue_internal.h"
59 * A bound pool is either associated or disassociated with its CPU.
60 * While associated (!DISASSOCIATED), all workers are bound to the
61 * CPU and none has %WORKER_UNBOUND set and concurrency management
64 * While DISASSOCIATED, the cpu may be offline and all workers have
65 * %WORKER_UNBOUND set and concurrency management disabled, and may
66 * be executing on any CPU. The pool behaves as an unbound one.
68 * Note that DISASSOCIATED should be flipped only while holding
69 * attach_mutex to avoid changing binding state while
70 * worker_attach_to_pool() is in progress.
72 POOL_MANAGER_ACTIVE
= 1 << 0, /* being managed */
73 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
76 WORKER_DIE
= 1 << 1, /* die die die */
77 WORKER_IDLE
= 1 << 2, /* is idle */
78 WORKER_PREP
= 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
81 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
84 WORKER_UNBOUND
| WORKER_REBOUND
,
86 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
98 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give MIN_NICE.
104 RESCUER_NICE_LEVEL
= MIN_NICE
,
105 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * A: pool->attach_mutex protected.
128 * PL: wq_pool_mutex protected.
130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
135 * sched-RCU for reads.
137 * WQ: wq->mutex protected.
139 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
141 * MD: wq_mayday_lock protected.
144 /* struct worker is defined in workqueue_internal.h */
147 spinlock_t lock
; /* the pool lock */
148 int cpu
; /* I: the associated cpu */
149 int node
; /* I: the associated node ID */
150 int id
; /* I: pool ID */
151 unsigned int flags
; /* X: flags */
153 unsigned long watchdog_ts
; /* L: watchdog timestamp */
155 struct list_head worklist
; /* L: list of pending works */
156 int nr_workers
; /* L: total number of workers */
158 /* nr_idle includes the ones off idle_list for rebinding */
159 int nr_idle
; /* L: currently idle ones */
161 struct list_head idle_list
; /* X: list of idle workers */
162 struct timer_list idle_timer
; /* L: worker idle timeout */
163 struct timer_list mayday_timer
; /* L: SOS timer for workers */
165 /* a workers is either on busy_hash or idle_list, or the manager */
166 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
167 /* L: hash of busy workers */
169 /* see manage_workers() for details on the two manager mutexes */
170 struct worker
*manager
; /* L: purely informational */
171 struct mutex attach_mutex
; /* attach/detach exclusion */
172 struct list_head workers
; /* A: attached workers */
173 struct completion
*detach_completion
; /* all workers detached */
175 struct ida worker_ida
; /* worker IDs for task name */
177 struct workqueue_attrs
*attrs
; /* I: worker attributes */
178 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
179 int refcnt
; /* PL: refcnt for unbound pools */
182 * The current concurrency level. As it's likely to be accessed
183 * from other CPUs during try_to_wake_up(), put it in a separate
186 atomic_t nr_running ____cacheline_aligned_in_smp
;
189 * Destruction of pool is sched-RCU protected to allow dereferences
190 * from get_work_pool().
193 } ____cacheline_aligned_in_smp
;
196 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
197 * of work_struct->data are used for flags and the remaining high bits
198 * point to the pwq; thus, pwqs need to be aligned at two's power of the
199 * number of flag bits.
201 struct pool_workqueue
{
202 struct worker_pool
*pool
; /* I: the associated pool */
203 struct workqueue_struct
*wq
; /* I: the owning workqueue */
204 int work_color
; /* L: current color */
205 int flush_color
; /* L: flushing color */
206 int refcnt
; /* L: reference count */
207 int nr_in_flight
[WORK_NR_COLORS
];
208 /* L: nr of in_flight works */
209 int nr_active
; /* L: nr of active works */
210 int max_active
; /* L: max active works */
211 struct list_head delayed_works
; /* L: delayed works */
212 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
213 struct list_head mayday_node
; /* MD: node on wq->maydays */
216 * Release of unbound pwq is punted to system_wq. See put_pwq()
217 * and pwq_unbound_release_workfn() for details. pool_workqueue
218 * itself is also sched-RCU protected so that the first pwq can be
219 * determined without grabbing wq->mutex.
221 struct work_struct unbound_release_work
;
223 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
226 * Structure used to wait for workqueue flush.
229 struct list_head list
; /* WQ: list of flushers */
230 int flush_color
; /* WQ: flush color waiting for */
231 struct completion done
; /* flush completion */
237 * The externally visible workqueue. It relays the issued work items to
238 * the appropriate worker_pool through its pool_workqueues.
240 struct workqueue_struct
{
241 struct list_head pwqs
; /* WR: all pwqs of this wq */
242 struct list_head list
; /* PR: list of all workqueues */
244 struct mutex mutex
; /* protects this wq */
245 int work_color
; /* WQ: current work color */
246 int flush_color
; /* WQ: current flush color */
247 atomic_t nr_pwqs_to_flush
; /* flush in progress */
248 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
249 struct list_head flusher_queue
; /* WQ: flush waiters */
250 struct list_head flusher_overflow
; /* WQ: flush overflow list */
252 struct list_head maydays
; /* MD: pwqs requesting rescue */
253 struct worker
*rescuer
; /* I: rescue worker */
255 int nr_drainers
; /* WQ: drain in progress */
256 int saved_max_active
; /* WQ: saved pwq max_active */
258 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
259 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
262 struct wq_device
*wq_dev
; /* I: for sysfs interface */
264 #ifdef CONFIG_LOCKDEP
265 struct lockdep_map lockdep_map
;
267 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
270 * Destruction of workqueue_struct is sched-RCU protected to allow
271 * walking the workqueues list without grabbing wq_pool_mutex.
272 * This is used to dump all workqueues from sysrq.
276 /* hot fields used during command issue, aligned to cacheline */
277 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
278 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
279 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
282 static struct kmem_cache
*pwq_cache
;
284 static cpumask_var_t
*wq_numa_possible_cpumask
;
285 /* possible CPUs of each node */
287 static bool wq_disable_numa
;
288 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
290 /* see the comment above the definition of WQ_POWER_EFFICIENT */
291 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
292 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
294 static bool wq_online
; /* can kworkers be created yet? */
296 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
298 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
301 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
302 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
303 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait
); /* wait for manager to go away */
305 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
306 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
308 /* PL: allowable cpus for unbound wqs and work items */
309 static cpumask_var_t wq_unbound_cpumask
;
311 /* CPU where unbound work was last round robin scheduled from this CPU */
312 static DEFINE_PER_CPU(int, wq_rr_cpu_last
);
315 * Local execution of unbound work items is no longer guaranteed. The
316 * following always forces round-robin CPU selection on unbound work items
317 * to uncover usages which depend on it.
319 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
320 static bool wq_debug_force_rr_cpu
= true;
322 static bool wq_debug_force_rr_cpu
= false;
324 module_param_named(debug_force_rr_cpu
, wq_debug_force_rr_cpu
, bool, 0644);
326 /* the per-cpu worker pools */
327 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
], cpu_worker_pools
);
329 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
331 /* PL: hash of all unbound pools keyed by pool->attrs */
332 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
334 /* I: attributes used when instantiating standard unbound pools on demand */
335 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
337 /* I: attributes used when instantiating ordered pools on demand */
338 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
340 struct workqueue_struct
*system_wq __read_mostly
;
341 EXPORT_SYMBOL(system_wq
);
342 struct workqueue_struct
*system_highpri_wq __read_mostly
;
343 EXPORT_SYMBOL_GPL(system_highpri_wq
);
344 struct workqueue_struct
*system_long_wq __read_mostly
;
345 EXPORT_SYMBOL_GPL(system_long_wq
);
346 struct workqueue_struct
*system_unbound_wq __read_mostly
;
347 EXPORT_SYMBOL_GPL(system_unbound_wq
);
348 struct workqueue_struct
*system_freezable_wq __read_mostly
;
349 EXPORT_SYMBOL_GPL(system_freezable_wq
);
350 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
351 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
352 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
353 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
355 static int worker_thread(void *__worker
);
356 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
358 #define CREATE_TRACE_POINTS
359 #include <trace/events/workqueue.h>
361 #define assert_rcu_or_pool_mutex() \
362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
363 !lockdep_is_held(&wq_pool_mutex), \
364 "sched RCU or wq_pool_mutex should be held")
366 #define assert_rcu_or_wq_mutex(wq) \
367 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
368 !lockdep_is_held(&wq->mutex), \
369 "sched RCU or wq->mutex should be held")
371 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
373 !lockdep_is_held(&wq->mutex) && \
374 !lockdep_is_held(&wq_pool_mutex), \
375 "sched RCU, wq->mutex or wq_pool_mutex should be held")
377 #define for_each_cpu_worker_pool(pool, cpu) \
378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
383 * for_each_pool - iterate through all worker_pools in the system
384 * @pool: iteration cursor
385 * @pi: integer used for iteration
387 * This must be called either with wq_pool_mutex held or sched RCU read
388 * locked. If the pool needs to be used beyond the locking in effect, the
389 * caller is responsible for guaranteeing that the pool stays online.
391 * The if/else clause exists only for the lockdep assertion and can be
394 #define for_each_pool(pool, pi) \
395 idr_for_each_entry(&worker_pool_idr, pool, pi) \
396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
400 * for_each_pool_worker - iterate through all workers of a worker_pool
401 * @worker: iteration cursor
402 * @pool: worker_pool to iterate workers of
404 * This must be called with @pool->attach_mutex.
406 * The if/else clause exists only for the lockdep assertion and can be
409 #define for_each_pool_worker(worker, pool) \
410 list_for_each_entry((worker), &(pool)->workers, node) \
411 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
416 * @pwq: iteration cursor
417 * @wq: the target workqueue
419 * This must be called either with wq->mutex held or sched RCU read locked.
420 * If the pwq needs to be used beyond the locking in effect, the caller is
421 * responsible for guaranteeing that the pwq stays online.
423 * The if/else clause exists only for the lockdep assertion and can be
426 #define for_each_pwq(pwq, wq) \
427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
431 #ifdef CONFIG_DEBUG_OBJECTS_WORK
433 static struct debug_obj_descr work_debug_descr
;
435 static void *work_debug_hint(void *addr
)
437 return ((struct work_struct
*) addr
)->func
;
440 static bool work_is_static_object(void *addr
)
442 struct work_struct
*work
= addr
;
444 return test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
));
448 * fixup_init is called when:
449 * - an active object is initialized
451 static bool work_fixup_init(void *addr
, enum debug_obj_state state
)
453 struct work_struct
*work
= addr
;
456 case ODEBUG_STATE_ACTIVE
:
457 cancel_work_sync(work
);
458 debug_object_init(work
, &work_debug_descr
);
466 * fixup_free is called when:
467 * - an active object is freed
469 static bool work_fixup_free(void *addr
, enum debug_obj_state state
)
471 struct work_struct
*work
= addr
;
474 case ODEBUG_STATE_ACTIVE
:
475 cancel_work_sync(work
);
476 debug_object_free(work
, &work_debug_descr
);
483 static struct debug_obj_descr work_debug_descr
= {
484 .name
= "work_struct",
485 .debug_hint
= work_debug_hint
,
486 .is_static_object
= work_is_static_object
,
487 .fixup_init
= work_fixup_init
,
488 .fixup_free
= work_fixup_free
,
491 static inline void debug_work_activate(struct work_struct
*work
)
493 debug_object_activate(work
, &work_debug_descr
);
496 static inline void debug_work_deactivate(struct work_struct
*work
)
498 debug_object_deactivate(work
, &work_debug_descr
);
501 void __init_work(struct work_struct
*work
, int onstack
)
504 debug_object_init_on_stack(work
, &work_debug_descr
);
506 debug_object_init(work
, &work_debug_descr
);
508 EXPORT_SYMBOL_GPL(__init_work
);
510 void destroy_work_on_stack(struct work_struct
*work
)
512 debug_object_free(work
, &work_debug_descr
);
514 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
516 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
518 destroy_timer_on_stack(&work
->timer
);
519 debug_object_free(&work
->work
, &work_debug_descr
);
521 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
524 static inline void debug_work_activate(struct work_struct
*work
) { }
525 static inline void debug_work_deactivate(struct work_struct
*work
) { }
529 * worker_pool_assign_id - allocate ID and assing it to @pool
530 * @pool: the pool pointer of interest
532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
533 * successfully, -errno on failure.
535 static int worker_pool_assign_id(struct worker_pool
*pool
)
539 lockdep_assert_held(&wq_pool_mutex
);
541 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
552 * @wq: the target workqueue
555 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
557 * If the pwq needs to be used beyond the locking in effect, the caller is
558 * responsible for guaranteeing that the pwq stays online.
560 * Return: The unbound pool_workqueue for @node.
562 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
565 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
569 * delayed item is pending. The plan is to keep CPU -> NODE
570 * mapping valid and stable across CPU on/offlines. Once that
571 * happens, this workaround can be removed.
573 if (unlikely(node
== NUMA_NO_NODE
))
576 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
579 static unsigned int work_color_to_flags(int color
)
581 return color
<< WORK_STRUCT_COLOR_SHIFT
;
584 static int get_work_color(struct work_struct
*work
)
586 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
587 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
590 static int work_next_color(int color
)
592 return (color
+ 1) % WORK_NR_COLORS
;
596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597 * contain the pointer to the queued pwq. Once execution starts, the flag
598 * is cleared and the high bits contain OFFQ flags and pool ID.
600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601 * and clear_work_data() can be used to set the pwq, pool or clear
602 * work->data. These functions should only be called while the work is
603 * owned - ie. while the PENDING bit is set.
605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
606 * corresponding to a work. Pool is available once the work has been
607 * queued anywhere after initialization until it is sync canceled. pwq is
608 * available only while the work item is queued.
610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
611 * canceled. While being canceled, a work item may have its PENDING set
612 * but stay off timer and worklist for arbitrarily long and nobody should
613 * try to steal the PENDING bit.
615 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
618 WARN_ON_ONCE(!work_pending(work
));
619 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
622 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
623 unsigned long extra_flags
)
625 set_work_data(work
, (unsigned long)pwq
,
626 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
629 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
632 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
633 WORK_STRUCT_PENDING
);
636 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
640 * The following wmb is paired with the implied mb in
641 * test_and_set_bit(PENDING) and ensures all updates to @work made
642 * here are visible to and precede any updates by the next PENDING
646 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
648 * The following mb guarantees that previous clear of a PENDING bit
649 * will not be reordered with any speculative LOADS or STORES from
650 * work->current_func, which is executed afterwards. This possible
651 * reordering can lead to a missed execution on attempt to qeueue
652 * the same @work. E.g. consider this case:
655 * ---------------------------- --------------------------------
657 * 1 STORE event_indicated
658 * 2 queue_work_on() {
659 * 3 test_and_set_bit(PENDING)
660 * 4 } set_..._and_clear_pending() {
661 * 5 set_work_data() # clear bit
663 * 7 work->current_func() {
664 * 8 LOAD event_indicated
667 * Without an explicit full barrier speculative LOAD on line 8 can
668 * be executed before CPU#0 does STORE on line 1. If that happens,
669 * CPU#0 observes the PENDING bit is still set and new execution of
670 * a @work is not queued in a hope, that CPU#1 will eventually
671 * finish the queued @work. Meanwhile CPU#1 does not see
672 * event_indicated is set, because speculative LOAD was executed
673 * before actual STORE.
678 static void clear_work_data(struct work_struct
*work
)
680 smp_wmb(); /* see set_work_pool_and_clear_pending() */
681 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
684 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
686 unsigned long data
= atomic_long_read(&work
->data
);
688 if (data
& WORK_STRUCT_PWQ
)
689 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
695 * get_work_pool - return the worker_pool a given work was associated with
696 * @work: the work item of interest
698 * Pools are created and destroyed under wq_pool_mutex, and allows read
699 * access under sched-RCU read lock. As such, this function should be
700 * called under wq_pool_mutex or with preemption disabled.
702 * All fields of the returned pool are accessible as long as the above
703 * mentioned locking is in effect. If the returned pool needs to be used
704 * beyond the critical section, the caller is responsible for ensuring the
705 * returned pool is and stays online.
707 * Return: The worker_pool @work was last associated with. %NULL if none.
709 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
711 unsigned long data
= atomic_long_read(&work
->data
);
714 assert_rcu_or_pool_mutex();
716 if (data
& WORK_STRUCT_PWQ
)
717 return ((struct pool_workqueue
*)
718 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
720 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
721 if (pool_id
== WORK_OFFQ_POOL_NONE
)
724 return idr_find(&worker_pool_idr
, pool_id
);
728 * get_work_pool_id - return the worker pool ID a given work is associated with
729 * @work: the work item of interest
731 * Return: The worker_pool ID @work was last associated with.
732 * %WORK_OFFQ_POOL_NONE if none.
734 static int get_work_pool_id(struct work_struct
*work
)
736 unsigned long data
= atomic_long_read(&work
->data
);
738 if (data
& WORK_STRUCT_PWQ
)
739 return ((struct pool_workqueue
*)
740 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
742 return data
>> WORK_OFFQ_POOL_SHIFT
;
745 static void mark_work_canceling(struct work_struct
*work
)
747 unsigned long pool_id
= get_work_pool_id(work
);
749 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
750 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
753 static bool work_is_canceling(struct work_struct
*work
)
755 unsigned long data
= atomic_long_read(&work
->data
);
757 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
761 * Policy functions. These define the policies on how the global worker
762 * pools are managed. Unless noted otherwise, these functions assume that
763 * they're being called with pool->lock held.
766 static bool __need_more_worker(struct worker_pool
*pool
)
768 return !atomic_read(&pool
->nr_running
);
772 * Need to wake up a worker? Called from anything but currently
775 * Note that, because unbound workers never contribute to nr_running, this
776 * function will always return %true for unbound pools as long as the
777 * worklist isn't empty.
779 static bool need_more_worker(struct worker_pool
*pool
)
781 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
784 /* Can I start working? Called from busy but !running workers. */
785 static bool may_start_working(struct worker_pool
*pool
)
787 return pool
->nr_idle
;
790 /* Do I need to keep working? Called from currently running workers. */
791 static bool keep_working(struct worker_pool
*pool
)
793 return !list_empty(&pool
->worklist
) &&
794 atomic_read(&pool
->nr_running
) <= 1;
797 /* Do we need a new worker? Called from manager. */
798 static bool need_to_create_worker(struct worker_pool
*pool
)
800 return need_more_worker(pool
) && !may_start_working(pool
);
803 /* Do we have too many workers and should some go away? */
804 static bool too_many_workers(struct worker_pool
*pool
)
806 bool managing
= pool
->flags
& POOL_MANAGER_ACTIVE
;
807 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
808 int nr_busy
= pool
->nr_workers
- nr_idle
;
810 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
817 /* Return the first idle worker. Safe with preemption disabled */
818 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
820 if (unlikely(list_empty(&pool
->idle_list
)))
823 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
827 * wake_up_worker - wake up an idle worker
828 * @pool: worker pool to wake worker from
830 * Wake up the first idle worker of @pool.
833 * spin_lock_irq(pool->lock).
835 static void wake_up_worker(struct worker_pool
*pool
)
837 struct worker
*worker
= first_idle_worker(pool
);
840 wake_up_process(worker
->task
);
844 * wq_worker_waking_up - a worker is waking up
845 * @task: task waking up
846 * @cpu: CPU @task is waking up to
848 * This function is called during try_to_wake_up() when a worker is
852 * spin_lock_irq(rq->lock)
854 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
856 struct worker
*worker
= kthread_data(task
);
858 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
859 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
860 atomic_inc(&worker
->pool
->nr_running
);
865 * wq_worker_sleeping - a worker is going to sleep
866 * @task: task going to sleep
868 * This function is called during schedule() when a busy worker is
869 * going to sleep. Worker on the same cpu can be woken up by
870 * returning pointer to its task.
873 * spin_lock_irq(rq->lock)
876 * Worker task on @cpu to wake up, %NULL if none.
878 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
)
880 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
881 struct worker_pool
*pool
;
884 * Rescuers, which may not have all the fields set up like normal
885 * workers, also reach here, let's not access anything before
886 * checking NOT_RUNNING.
888 if (worker
->flags
& WORKER_NOT_RUNNING
)
893 /* this can only happen on the local cpu */
894 if (WARN_ON_ONCE(pool
->cpu
!= raw_smp_processor_id()))
898 * The counterpart of the following dec_and_test, implied mb,
899 * worklist not empty test sequence is in insert_work().
900 * Please read comment there.
902 * NOT_RUNNING is clear. This means that we're bound to and
903 * running on the local cpu w/ rq lock held and preemption
904 * disabled, which in turn means that none else could be
905 * manipulating idle_list, so dereferencing idle_list without pool
908 if (atomic_dec_and_test(&pool
->nr_running
) &&
909 !list_empty(&pool
->worklist
))
910 to_wakeup
= first_idle_worker(pool
);
911 return to_wakeup
? to_wakeup
->task
: NULL
;
915 * worker_set_flags - set worker flags and adjust nr_running accordingly
917 * @flags: flags to set
919 * Set @flags in @worker->flags and adjust nr_running accordingly.
922 * spin_lock_irq(pool->lock)
924 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
926 struct worker_pool
*pool
= worker
->pool
;
928 WARN_ON_ONCE(worker
->task
!= current
);
930 /* If transitioning into NOT_RUNNING, adjust nr_running. */
931 if ((flags
& WORKER_NOT_RUNNING
) &&
932 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
933 atomic_dec(&pool
->nr_running
);
936 worker
->flags
|= flags
;
940 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
942 * @flags: flags to clear
944 * Clear @flags in @worker->flags and adjust nr_running accordingly.
947 * spin_lock_irq(pool->lock)
949 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
951 struct worker_pool
*pool
= worker
->pool
;
952 unsigned int oflags
= worker
->flags
;
954 WARN_ON_ONCE(worker
->task
!= current
);
956 worker
->flags
&= ~flags
;
959 * If transitioning out of NOT_RUNNING, increment nr_running. Note
960 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
961 * of multiple flags, not a single flag.
963 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
964 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
965 atomic_inc(&pool
->nr_running
);
969 * find_worker_executing_work - find worker which is executing a work
970 * @pool: pool of interest
971 * @work: work to find worker for
973 * Find a worker which is executing @work on @pool by searching
974 * @pool->busy_hash which is keyed by the address of @work. For a worker
975 * to match, its current execution should match the address of @work and
976 * its work function. This is to avoid unwanted dependency between
977 * unrelated work executions through a work item being recycled while still
980 * This is a bit tricky. A work item may be freed once its execution
981 * starts and nothing prevents the freed area from being recycled for
982 * another work item. If the same work item address ends up being reused
983 * before the original execution finishes, workqueue will identify the
984 * recycled work item as currently executing and make it wait until the
985 * current execution finishes, introducing an unwanted dependency.
987 * This function checks the work item address and work function to avoid
988 * false positives. Note that this isn't complete as one may construct a
989 * work function which can introduce dependency onto itself through a
990 * recycled work item. Well, if somebody wants to shoot oneself in the
991 * foot that badly, there's only so much we can do, and if such deadlock
992 * actually occurs, it should be easy to locate the culprit work function.
995 * spin_lock_irq(pool->lock).
998 * Pointer to worker which is executing @work if found, %NULL
1001 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
1002 struct work_struct
*work
)
1004 struct worker
*worker
;
1006 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1007 (unsigned long)work
)
1008 if (worker
->current_work
== work
&&
1009 worker
->current_func
== work
->func
)
1016 * move_linked_works - move linked works to a list
1017 * @work: start of series of works to be scheduled
1018 * @head: target list to append @work to
1019 * @nextp: out parameter for nested worklist walking
1021 * Schedule linked works starting from @work to @head. Work series to
1022 * be scheduled starts at @work and includes any consecutive work with
1023 * WORK_STRUCT_LINKED set in its predecessor.
1025 * If @nextp is not NULL, it's updated to point to the next work of
1026 * the last scheduled work. This allows move_linked_works() to be
1027 * nested inside outer list_for_each_entry_safe().
1030 * spin_lock_irq(pool->lock).
1032 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1033 struct work_struct
**nextp
)
1035 struct work_struct
*n
;
1038 * Linked worklist will always end before the end of the list,
1039 * use NULL for list head.
1041 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1042 list_move_tail(&work
->entry
, head
);
1043 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1048 * If we're already inside safe list traversal and have moved
1049 * multiple works to the scheduled queue, the next position
1050 * needs to be updated.
1057 * get_pwq - get an extra reference on the specified pool_workqueue
1058 * @pwq: pool_workqueue to get
1060 * Obtain an extra reference on @pwq. The caller should guarantee that
1061 * @pwq has positive refcnt and be holding the matching pool->lock.
1063 static void get_pwq(struct pool_workqueue
*pwq
)
1065 lockdep_assert_held(&pwq
->pool
->lock
);
1066 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1071 * put_pwq - put a pool_workqueue reference
1072 * @pwq: pool_workqueue to put
1074 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1075 * destruction. The caller should be holding the matching pool->lock.
1077 static void put_pwq(struct pool_workqueue
*pwq
)
1079 lockdep_assert_held(&pwq
->pool
->lock
);
1080 if (likely(--pwq
->refcnt
))
1082 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1085 * @pwq can't be released under pool->lock, bounce to
1086 * pwq_unbound_release_workfn(). This never recurses on the same
1087 * pool->lock as this path is taken only for unbound workqueues and
1088 * the release work item is scheduled on a per-cpu workqueue. To
1089 * avoid lockdep warning, unbound pool->locks are given lockdep
1090 * subclass of 1 in get_unbound_pool().
1092 schedule_work(&pwq
->unbound_release_work
);
1096 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1097 * @pwq: pool_workqueue to put (can be %NULL)
1099 * put_pwq() with locking. This function also allows %NULL @pwq.
1101 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1105 * As both pwqs and pools are sched-RCU protected, the
1106 * following lock operations are safe.
1108 spin_lock_irq(&pwq
->pool
->lock
);
1110 spin_unlock_irq(&pwq
->pool
->lock
);
1114 static void pwq_activate_delayed_work(struct work_struct
*work
)
1116 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1118 trace_workqueue_activate_work(work
);
1119 if (list_empty(&pwq
->pool
->worklist
))
1120 pwq
->pool
->watchdog_ts
= jiffies
;
1121 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1122 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1126 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1128 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1129 struct work_struct
, entry
);
1131 pwq_activate_delayed_work(work
);
1135 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1136 * @pwq: pwq of interest
1137 * @color: color of work which left the queue
1139 * A work either has completed or is removed from pending queue,
1140 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1143 * spin_lock_irq(pool->lock).
1145 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1147 /* uncolored work items don't participate in flushing or nr_active */
1148 if (color
== WORK_NO_COLOR
)
1151 pwq
->nr_in_flight
[color
]--;
1154 if (!list_empty(&pwq
->delayed_works
)) {
1155 /* one down, submit a delayed one */
1156 if (pwq
->nr_active
< pwq
->max_active
)
1157 pwq_activate_first_delayed(pwq
);
1160 /* is flush in progress and are we at the flushing tip? */
1161 if (likely(pwq
->flush_color
!= color
))
1164 /* are there still in-flight works? */
1165 if (pwq
->nr_in_flight
[color
])
1168 /* this pwq is done, clear flush_color */
1169 pwq
->flush_color
= -1;
1172 * If this was the last pwq, wake up the first flusher. It
1173 * will handle the rest.
1175 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1176 complete(&pwq
->wq
->first_flusher
->done
);
1182 * try_to_grab_pending - steal work item from worklist and disable irq
1183 * @work: work item to steal
1184 * @is_dwork: @work is a delayed_work
1185 * @flags: place to store irq state
1187 * Try to grab PENDING bit of @work. This function can handle @work in any
1188 * stable state - idle, on timer or on worklist.
1191 * 1 if @work was pending and we successfully stole PENDING
1192 * 0 if @work was idle and we claimed PENDING
1193 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1194 * -ENOENT if someone else is canceling @work, this state may persist
1195 * for arbitrarily long
1198 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1199 * interrupted while holding PENDING and @work off queue, irq must be
1200 * disabled on entry. This, combined with delayed_work->timer being
1201 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1203 * On successful return, >= 0, irq is disabled and the caller is
1204 * responsible for releasing it using local_irq_restore(*@flags).
1206 * This function is safe to call from any context including IRQ handler.
1208 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1209 unsigned long *flags
)
1211 struct worker_pool
*pool
;
1212 struct pool_workqueue
*pwq
;
1214 local_irq_save(*flags
);
1216 /* try to steal the timer if it exists */
1218 struct delayed_work
*dwork
= to_delayed_work(work
);
1221 * dwork->timer is irqsafe. If del_timer() fails, it's
1222 * guaranteed that the timer is not queued anywhere and not
1223 * running on the local CPU.
1225 if (likely(del_timer(&dwork
->timer
)))
1229 /* try to claim PENDING the normal way */
1230 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1234 * The queueing is in progress, or it is already queued. Try to
1235 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1237 pool
= get_work_pool(work
);
1241 spin_lock(&pool
->lock
);
1243 * work->data is guaranteed to point to pwq only while the work
1244 * item is queued on pwq->wq, and both updating work->data to point
1245 * to pwq on queueing and to pool on dequeueing are done under
1246 * pwq->pool->lock. This in turn guarantees that, if work->data
1247 * points to pwq which is associated with a locked pool, the work
1248 * item is currently queued on that pool.
1250 pwq
= get_work_pwq(work
);
1251 if (pwq
&& pwq
->pool
== pool
) {
1252 debug_work_deactivate(work
);
1255 * A delayed work item cannot be grabbed directly because
1256 * it might have linked NO_COLOR work items which, if left
1257 * on the delayed_list, will confuse pwq->nr_active
1258 * management later on and cause stall. Make sure the work
1259 * item is activated before grabbing.
1261 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1262 pwq_activate_delayed_work(work
);
1264 list_del_init(&work
->entry
);
1265 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1267 /* work->data points to pwq iff queued, point to pool */
1268 set_work_pool_and_keep_pending(work
, pool
->id
);
1270 spin_unlock(&pool
->lock
);
1273 spin_unlock(&pool
->lock
);
1275 local_irq_restore(*flags
);
1276 if (work_is_canceling(work
))
1283 * insert_work - insert a work into a pool
1284 * @pwq: pwq @work belongs to
1285 * @work: work to insert
1286 * @head: insertion point
1287 * @extra_flags: extra WORK_STRUCT_* flags to set
1289 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1290 * work_struct flags.
1293 * spin_lock_irq(pool->lock).
1295 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1296 struct list_head
*head
, unsigned int extra_flags
)
1298 struct worker_pool
*pool
= pwq
->pool
;
1300 /* we own @work, set data and link */
1301 set_work_pwq(work
, pwq
, extra_flags
);
1302 list_add_tail(&work
->entry
, head
);
1306 * Ensure either wq_worker_sleeping() sees the above
1307 * list_add_tail() or we see zero nr_running to avoid workers lying
1308 * around lazily while there are works to be processed.
1312 if (__need_more_worker(pool
))
1313 wake_up_worker(pool
);
1317 * Test whether @work is being queued from another work executing on the
1320 static bool is_chained_work(struct workqueue_struct
*wq
)
1322 struct worker
*worker
;
1324 worker
= current_wq_worker();
1326 * Return %true iff I'm a worker execuing a work item on @wq. If
1327 * I'm @worker, it's safe to dereference it without locking.
1329 return worker
&& worker
->current_pwq
->wq
== wq
;
1333 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1334 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1335 * avoid perturbing sensitive tasks.
1337 static int wq_select_unbound_cpu(int cpu
)
1339 static bool printed_dbg_warning
;
1342 if (likely(!wq_debug_force_rr_cpu
)) {
1343 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1345 } else if (!printed_dbg_warning
) {
1346 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1347 printed_dbg_warning
= true;
1350 if (cpumask_empty(wq_unbound_cpumask
))
1353 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1354 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1355 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1356 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1357 if (unlikely(new_cpu
>= nr_cpu_ids
))
1360 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1365 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1366 struct work_struct
*work
)
1368 struct pool_workqueue
*pwq
;
1369 struct worker_pool
*last_pool
;
1370 struct list_head
*worklist
;
1371 unsigned int work_flags
;
1372 unsigned int req_cpu
= cpu
;
1375 * While a work item is PENDING && off queue, a task trying to
1376 * steal the PENDING will busy-loop waiting for it to either get
1377 * queued or lose PENDING. Grabbing PENDING and queueing should
1378 * happen with IRQ disabled.
1380 lockdep_assert_irqs_disabled();
1382 debug_work_activate(work
);
1384 /* if draining, only works from the same workqueue are allowed */
1385 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1386 WARN_ON_ONCE(!is_chained_work(wq
)))
1389 if (req_cpu
== WORK_CPU_UNBOUND
)
1390 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1392 /* pwq which will be used unless @work is executing elsewhere */
1393 if (!(wq
->flags
& WQ_UNBOUND
))
1394 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1396 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1399 * If @work was previously on a different pool, it might still be
1400 * running there, in which case the work needs to be queued on that
1401 * pool to guarantee non-reentrancy.
1403 last_pool
= get_work_pool(work
);
1404 if (last_pool
&& last_pool
!= pwq
->pool
) {
1405 struct worker
*worker
;
1407 spin_lock(&last_pool
->lock
);
1409 worker
= find_worker_executing_work(last_pool
, work
);
1411 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1412 pwq
= worker
->current_pwq
;
1414 /* meh... not running there, queue here */
1415 spin_unlock(&last_pool
->lock
);
1416 spin_lock(&pwq
->pool
->lock
);
1419 spin_lock(&pwq
->pool
->lock
);
1423 * pwq is determined and locked. For unbound pools, we could have
1424 * raced with pwq release and it could already be dead. If its
1425 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1426 * without another pwq replacing it in the numa_pwq_tbl or while
1427 * work items are executing on it, so the retrying is guaranteed to
1428 * make forward-progress.
1430 if (unlikely(!pwq
->refcnt
)) {
1431 if (wq
->flags
& WQ_UNBOUND
) {
1432 spin_unlock(&pwq
->pool
->lock
);
1437 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1441 /* pwq determined, queue */
1442 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1444 if (WARN_ON(!list_empty(&work
->entry
))) {
1445 spin_unlock(&pwq
->pool
->lock
);
1449 pwq
->nr_in_flight
[pwq
->work_color
]++;
1450 work_flags
= work_color_to_flags(pwq
->work_color
);
1452 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1453 trace_workqueue_activate_work(work
);
1455 worklist
= &pwq
->pool
->worklist
;
1456 if (list_empty(worklist
))
1457 pwq
->pool
->watchdog_ts
= jiffies
;
1459 work_flags
|= WORK_STRUCT_DELAYED
;
1460 worklist
= &pwq
->delayed_works
;
1463 insert_work(pwq
, work
, worklist
, work_flags
);
1465 spin_unlock(&pwq
->pool
->lock
);
1469 * queue_work_on - queue work on specific cpu
1470 * @cpu: CPU number to execute work on
1471 * @wq: workqueue to use
1472 * @work: work to queue
1474 * We queue the work to a specific CPU, the caller must ensure it
1477 * Return: %false if @work was already on a queue, %true otherwise.
1479 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1480 struct work_struct
*work
)
1483 unsigned long flags
;
1485 local_irq_save(flags
);
1487 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1488 __queue_work(cpu
, wq
, work
);
1492 local_irq_restore(flags
);
1495 EXPORT_SYMBOL(queue_work_on
);
1497 void delayed_work_timer_fn(struct timer_list
*t
)
1499 struct delayed_work
*dwork
= from_timer(dwork
, t
, timer
);
1501 /* should have been called from irqsafe timer with irq already off */
1502 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1504 EXPORT_SYMBOL(delayed_work_timer_fn
);
1506 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1507 struct delayed_work
*dwork
, unsigned long delay
)
1509 struct timer_list
*timer
= &dwork
->timer
;
1510 struct work_struct
*work
= &dwork
->work
;
1513 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
);
1514 WARN_ON_ONCE(timer_pending(timer
));
1515 WARN_ON_ONCE(!list_empty(&work
->entry
));
1518 * If @delay is 0, queue @dwork->work immediately. This is for
1519 * both optimization and correctness. The earliest @timer can
1520 * expire is on the closest next tick and delayed_work users depend
1521 * on that there's no such delay when @delay is 0.
1524 __queue_work(cpu
, wq
, &dwork
->work
);
1530 timer
->expires
= jiffies
+ delay
;
1532 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1533 add_timer_on(timer
, cpu
);
1539 * queue_delayed_work_on - queue work on specific CPU after delay
1540 * @cpu: CPU number to execute work on
1541 * @wq: workqueue to use
1542 * @dwork: work to queue
1543 * @delay: number of jiffies to wait before queueing
1545 * Return: %false if @work was already on a queue, %true otherwise. If
1546 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1549 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1550 struct delayed_work
*dwork
, unsigned long delay
)
1552 struct work_struct
*work
= &dwork
->work
;
1554 unsigned long flags
;
1556 /* read the comment in __queue_work() */
1557 local_irq_save(flags
);
1559 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1560 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1564 local_irq_restore(flags
);
1567 EXPORT_SYMBOL(queue_delayed_work_on
);
1570 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1571 * @cpu: CPU number to execute work on
1572 * @wq: workqueue to use
1573 * @dwork: work to queue
1574 * @delay: number of jiffies to wait before queueing
1576 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1577 * modify @dwork's timer so that it expires after @delay. If @delay is
1578 * zero, @work is guaranteed to be scheduled immediately regardless of its
1581 * Return: %false if @dwork was idle and queued, %true if @dwork was
1582 * pending and its timer was modified.
1584 * This function is safe to call from any context including IRQ handler.
1585 * See try_to_grab_pending() for details.
1587 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1588 struct delayed_work
*dwork
, unsigned long delay
)
1590 unsigned long flags
;
1594 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1595 } while (unlikely(ret
== -EAGAIN
));
1597 if (likely(ret
>= 0)) {
1598 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1599 local_irq_restore(flags
);
1602 /* -ENOENT from try_to_grab_pending() becomes %true */
1605 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1607 static void rcu_work_rcufn(struct rcu_head
*rcu
)
1609 struct rcu_work
*rwork
= container_of(rcu
, struct rcu_work
, rcu
);
1611 /* read the comment in __queue_work() */
1612 local_irq_disable();
1613 __queue_work(WORK_CPU_UNBOUND
, rwork
->wq
, &rwork
->work
);
1618 * queue_rcu_work - queue work after a RCU grace period
1619 * @wq: workqueue to use
1620 * @rwork: work to queue
1622 * Return: %false if @rwork was already pending, %true otherwise. Note
1623 * that a full RCU grace period is guaranteed only after a %true return.
1624 * While @rwork is guarnateed to be executed after a %false return, the
1625 * execution may happen before a full RCU grace period has passed.
1627 bool queue_rcu_work(struct workqueue_struct
*wq
, struct rcu_work
*rwork
)
1629 struct work_struct
*work
= &rwork
->work
;
1631 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1633 call_rcu(&rwork
->rcu
, rcu_work_rcufn
);
1639 EXPORT_SYMBOL(queue_rcu_work
);
1642 * worker_enter_idle - enter idle state
1643 * @worker: worker which is entering idle state
1645 * @worker is entering idle state. Update stats and idle timer if
1649 * spin_lock_irq(pool->lock).
1651 static void worker_enter_idle(struct worker
*worker
)
1653 struct worker_pool
*pool
= worker
->pool
;
1655 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1656 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1657 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1660 /* can't use worker_set_flags(), also called from create_worker() */
1661 worker
->flags
|= WORKER_IDLE
;
1663 worker
->last_active
= jiffies
;
1665 /* idle_list is LIFO */
1666 list_add(&worker
->entry
, &pool
->idle_list
);
1668 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1669 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1672 * Sanity check nr_running. Because unbind_workers() releases
1673 * pool->lock between setting %WORKER_UNBOUND and zapping
1674 * nr_running, the warning may trigger spuriously. Check iff
1675 * unbind is not in progress.
1677 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1678 pool
->nr_workers
== pool
->nr_idle
&&
1679 atomic_read(&pool
->nr_running
));
1683 * worker_leave_idle - leave idle state
1684 * @worker: worker which is leaving idle state
1686 * @worker is leaving idle state. Update stats.
1689 * spin_lock_irq(pool->lock).
1691 static void worker_leave_idle(struct worker
*worker
)
1693 struct worker_pool
*pool
= worker
->pool
;
1695 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1697 worker_clr_flags(worker
, WORKER_IDLE
);
1699 list_del_init(&worker
->entry
);
1702 static struct worker
*alloc_worker(int node
)
1704 struct worker
*worker
;
1706 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1708 INIT_LIST_HEAD(&worker
->entry
);
1709 INIT_LIST_HEAD(&worker
->scheduled
);
1710 INIT_LIST_HEAD(&worker
->node
);
1711 /* on creation a worker is in !idle && prep state */
1712 worker
->flags
= WORKER_PREP
;
1718 * worker_attach_to_pool() - attach a worker to a pool
1719 * @worker: worker to be attached
1720 * @pool: the target pool
1722 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1723 * cpu-binding of @worker are kept coordinated with the pool across
1726 static void worker_attach_to_pool(struct worker
*worker
,
1727 struct worker_pool
*pool
)
1729 mutex_lock(&pool
->attach_mutex
);
1732 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1733 * online CPUs. It'll be re-applied when any of the CPUs come up.
1735 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1738 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1739 * stable across this function. See the comments above the
1740 * flag definition for details.
1742 if (pool
->flags
& POOL_DISASSOCIATED
)
1743 worker
->flags
|= WORKER_UNBOUND
;
1745 list_add_tail(&worker
->node
, &pool
->workers
);
1747 mutex_unlock(&pool
->attach_mutex
);
1751 * worker_detach_from_pool() - detach a worker from its pool
1752 * @worker: worker which is attached to its pool
1753 * @pool: the pool @worker is attached to
1755 * Undo the attaching which had been done in worker_attach_to_pool(). The
1756 * caller worker shouldn't access to the pool after detached except it has
1757 * other reference to the pool.
1759 static void worker_detach_from_pool(struct worker
*worker
,
1760 struct worker_pool
*pool
)
1762 struct completion
*detach_completion
= NULL
;
1764 mutex_lock(&pool
->attach_mutex
);
1765 list_del(&worker
->node
);
1766 if (list_empty(&pool
->workers
))
1767 detach_completion
= pool
->detach_completion
;
1768 mutex_unlock(&pool
->attach_mutex
);
1770 /* clear leftover flags without pool->lock after it is detached */
1771 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1773 if (detach_completion
)
1774 complete(detach_completion
);
1778 * create_worker - create a new workqueue worker
1779 * @pool: pool the new worker will belong to
1781 * Create and start a new worker which is attached to @pool.
1784 * Might sleep. Does GFP_KERNEL allocations.
1787 * Pointer to the newly created worker.
1789 static struct worker
*create_worker(struct worker_pool
*pool
)
1791 struct worker
*worker
= NULL
;
1795 /* ID is needed to determine kthread name */
1796 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1800 worker
= alloc_worker(pool
->node
);
1804 worker
->pool
= pool
;
1808 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1809 pool
->attrs
->nice
< 0 ? "H" : "");
1811 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1813 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1814 "kworker/%s", id_buf
);
1815 if (IS_ERR(worker
->task
))
1818 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1819 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1821 /* successful, attach the worker to the pool */
1822 worker_attach_to_pool(worker
, pool
);
1824 /* start the newly created worker */
1825 spin_lock_irq(&pool
->lock
);
1826 worker
->pool
->nr_workers
++;
1827 worker_enter_idle(worker
);
1828 wake_up_process(worker
->task
);
1829 spin_unlock_irq(&pool
->lock
);
1835 ida_simple_remove(&pool
->worker_ida
, id
);
1841 * destroy_worker - destroy a workqueue worker
1842 * @worker: worker to be destroyed
1844 * Destroy @worker and adjust @pool stats accordingly. The worker should
1848 * spin_lock_irq(pool->lock).
1850 static void destroy_worker(struct worker
*worker
)
1852 struct worker_pool
*pool
= worker
->pool
;
1854 lockdep_assert_held(&pool
->lock
);
1856 /* sanity check frenzy */
1857 if (WARN_ON(worker
->current_work
) ||
1858 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1859 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1865 list_del_init(&worker
->entry
);
1866 worker
->flags
|= WORKER_DIE
;
1867 wake_up_process(worker
->task
);
1870 static void idle_worker_timeout(struct timer_list
*t
)
1872 struct worker_pool
*pool
= from_timer(pool
, t
, idle_timer
);
1874 spin_lock_irq(&pool
->lock
);
1876 while (too_many_workers(pool
)) {
1877 struct worker
*worker
;
1878 unsigned long expires
;
1880 /* idle_list is kept in LIFO order, check the last one */
1881 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1882 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1884 if (time_before(jiffies
, expires
)) {
1885 mod_timer(&pool
->idle_timer
, expires
);
1889 destroy_worker(worker
);
1892 spin_unlock_irq(&pool
->lock
);
1895 static void send_mayday(struct work_struct
*work
)
1897 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1898 struct workqueue_struct
*wq
= pwq
->wq
;
1900 lockdep_assert_held(&wq_mayday_lock
);
1905 /* mayday mayday mayday */
1906 if (list_empty(&pwq
->mayday_node
)) {
1908 * If @pwq is for an unbound wq, its base ref may be put at
1909 * any time due to an attribute change. Pin @pwq until the
1910 * rescuer is done with it.
1913 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1914 wake_up_process(wq
->rescuer
->task
);
1918 static void pool_mayday_timeout(struct timer_list
*t
)
1920 struct worker_pool
*pool
= from_timer(pool
, t
, mayday_timer
);
1921 struct work_struct
*work
;
1923 spin_lock_irq(&pool
->lock
);
1924 spin_lock(&wq_mayday_lock
); /* for wq->maydays */
1926 if (need_to_create_worker(pool
)) {
1928 * We've been trying to create a new worker but
1929 * haven't been successful. We might be hitting an
1930 * allocation deadlock. Send distress signals to
1933 list_for_each_entry(work
, &pool
->worklist
, entry
)
1937 spin_unlock(&wq_mayday_lock
);
1938 spin_unlock_irq(&pool
->lock
);
1940 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1944 * maybe_create_worker - create a new worker if necessary
1945 * @pool: pool to create a new worker for
1947 * Create a new worker for @pool if necessary. @pool is guaranteed to
1948 * have at least one idle worker on return from this function. If
1949 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1950 * sent to all rescuers with works scheduled on @pool to resolve
1951 * possible allocation deadlock.
1953 * On return, need_to_create_worker() is guaranteed to be %false and
1954 * may_start_working() %true.
1957 * spin_lock_irq(pool->lock) which may be released and regrabbed
1958 * multiple times. Does GFP_KERNEL allocations. Called only from
1961 static void maybe_create_worker(struct worker_pool
*pool
)
1962 __releases(&pool
->lock
)
1963 __acquires(&pool
->lock
)
1966 spin_unlock_irq(&pool
->lock
);
1968 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1969 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1972 if (create_worker(pool
) || !need_to_create_worker(pool
))
1975 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1977 if (!need_to_create_worker(pool
))
1981 del_timer_sync(&pool
->mayday_timer
);
1982 spin_lock_irq(&pool
->lock
);
1984 * This is necessary even after a new worker was just successfully
1985 * created as @pool->lock was dropped and the new worker might have
1986 * already become busy.
1988 if (need_to_create_worker(pool
))
1993 * manage_workers - manage worker pool
1996 * Assume the manager role and manage the worker pool @worker belongs
1997 * to. At any given time, there can be only zero or one manager per
1998 * pool. The exclusion is handled automatically by this function.
2000 * The caller can safely start processing works on false return. On
2001 * true return, it's guaranteed that need_to_create_worker() is false
2002 * and may_start_working() is true.
2005 * spin_lock_irq(pool->lock) which may be released and regrabbed
2006 * multiple times. Does GFP_KERNEL allocations.
2009 * %false if the pool doesn't need management and the caller can safely
2010 * start processing works, %true if management function was performed and
2011 * the conditions that the caller verified before calling the function may
2012 * no longer be true.
2014 static bool manage_workers(struct worker
*worker
)
2016 struct worker_pool
*pool
= worker
->pool
;
2018 if (pool
->flags
& POOL_MANAGER_ACTIVE
)
2021 pool
->flags
|= POOL_MANAGER_ACTIVE
;
2022 pool
->manager
= worker
;
2024 maybe_create_worker(pool
);
2026 pool
->manager
= NULL
;
2027 pool
->flags
&= ~POOL_MANAGER_ACTIVE
;
2028 wake_up(&wq_manager_wait
);
2033 * process_one_work - process single work
2035 * @work: work to process
2037 * Process @work. This function contains all the logics necessary to
2038 * process a single work including synchronization against and
2039 * interaction with other workers on the same cpu, queueing and
2040 * flushing. As long as context requirement is met, any worker can
2041 * call this function to process a work.
2044 * spin_lock_irq(pool->lock) which is released and regrabbed.
2046 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2047 __releases(&pool
->lock
)
2048 __acquires(&pool
->lock
)
2050 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2051 struct worker_pool
*pool
= worker
->pool
;
2052 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2054 struct worker
*collision
;
2055 #ifdef CONFIG_LOCKDEP
2057 * It is permissible to free the struct work_struct from
2058 * inside the function that is called from it, this we need to
2059 * take into account for lockdep too. To avoid bogus "held
2060 * lock freed" warnings as well as problems when looking into
2061 * work->lockdep_map, make a copy and use that here.
2063 struct lockdep_map lockdep_map
;
2065 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2067 /* ensure we're on the correct CPU */
2068 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2069 raw_smp_processor_id() != pool
->cpu
);
2072 * A single work shouldn't be executed concurrently by
2073 * multiple workers on a single cpu. Check whether anyone is
2074 * already processing the work. If so, defer the work to the
2075 * currently executing one.
2077 collision
= find_worker_executing_work(pool
, work
);
2078 if (unlikely(collision
)) {
2079 move_linked_works(work
, &collision
->scheduled
, NULL
);
2083 /* claim and dequeue */
2084 debug_work_deactivate(work
);
2085 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2086 worker
->current_work
= work
;
2087 worker
->current_func
= work
->func
;
2088 worker
->current_pwq
= pwq
;
2089 work_color
= get_work_color(work
);
2091 list_del_init(&work
->entry
);
2094 * CPU intensive works don't participate in concurrency management.
2095 * They're the scheduler's responsibility. This takes @worker out
2096 * of concurrency management and the next code block will chain
2097 * execution of the pending work items.
2099 if (unlikely(cpu_intensive
))
2100 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2103 * Wake up another worker if necessary. The condition is always
2104 * false for normal per-cpu workers since nr_running would always
2105 * be >= 1 at this point. This is used to chain execution of the
2106 * pending work items for WORKER_NOT_RUNNING workers such as the
2107 * UNBOUND and CPU_INTENSIVE ones.
2109 if (need_more_worker(pool
))
2110 wake_up_worker(pool
);
2113 * Record the last pool and clear PENDING which should be the last
2114 * update to @work. Also, do this inside @pool->lock so that
2115 * PENDING and queued state changes happen together while IRQ is
2118 set_work_pool_and_clear_pending(work
, pool
->id
);
2120 spin_unlock_irq(&pool
->lock
);
2122 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2123 lock_map_acquire(&lockdep_map
);
2125 * Strictly speaking we should mark the invariant state without holding
2126 * any locks, that is, before these two lock_map_acquire()'s.
2128 * However, that would result in:
2135 * Which would create W1->C->W1 dependencies, even though there is no
2136 * actual deadlock possible. There are two solutions, using a
2137 * read-recursive acquire on the work(queue) 'locks', but this will then
2138 * hit the lockdep limitation on recursive locks, or simply discard
2141 * AFAICT there is no possible deadlock scenario between the
2142 * flush_work() and complete() primitives (except for single-threaded
2143 * workqueues), so hiding them isn't a problem.
2145 lockdep_invariant_state(true);
2146 trace_workqueue_execute_start(work
);
2147 worker
->current_func(work
);
2149 * While we must be careful to not use "work" after this, the trace
2150 * point will only record its address.
2152 trace_workqueue_execute_end(work
);
2153 lock_map_release(&lockdep_map
);
2154 lock_map_release(&pwq
->wq
->lockdep_map
);
2156 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2157 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2158 " last function: %pf\n",
2159 current
->comm
, preempt_count(), task_pid_nr(current
),
2160 worker
->current_func
);
2161 debug_show_held_locks(current
);
2166 * The following prevents a kworker from hogging CPU on !PREEMPT
2167 * kernels, where a requeueing work item waiting for something to
2168 * happen could deadlock with stop_machine as such work item could
2169 * indefinitely requeue itself while all other CPUs are trapped in
2170 * stop_machine. At the same time, report a quiescent RCU state so
2171 * the same condition doesn't freeze RCU.
2173 cond_resched_rcu_qs();
2175 spin_lock_irq(&pool
->lock
);
2177 /* clear cpu intensive status */
2178 if (unlikely(cpu_intensive
))
2179 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2181 /* we're done with it, release */
2182 hash_del(&worker
->hentry
);
2183 worker
->current_work
= NULL
;
2184 worker
->current_func
= NULL
;
2185 worker
->current_pwq
= NULL
;
2186 worker
->desc_valid
= false;
2187 pwq_dec_nr_in_flight(pwq
, work_color
);
2191 * process_scheduled_works - process scheduled works
2194 * Process all scheduled works. Please note that the scheduled list
2195 * may change while processing a work, so this function repeatedly
2196 * fetches a work from the top and executes it.
2199 * spin_lock_irq(pool->lock) which may be released and regrabbed
2202 static void process_scheduled_works(struct worker
*worker
)
2204 while (!list_empty(&worker
->scheduled
)) {
2205 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2206 struct work_struct
, entry
);
2207 process_one_work(worker
, work
);
2212 * worker_thread - the worker thread function
2215 * The worker thread function. All workers belong to a worker_pool -
2216 * either a per-cpu one or dynamic unbound one. These workers process all
2217 * work items regardless of their specific target workqueue. The only
2218 * exception is work items which belong to workqueues with a rescuer which
2219 * will be explained in rescuer_thread().
2223 static int worker_thread(void *__worker
)
2225 struct worker
*worker
= __worker
;
2226 struct worker_pool
*pool
= worker
->pool
;
2228 /* tell the scheduler that this is a workqueue worker */
2229 worker
->task
->flags
|= PF_WQ_WORKER
;
2231 spin_lock_irq(&pool
->lock
);
2233 /* am I supposed to die? */
2234 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2235 spin_unlock_irq(&pool
->lock
);
2236 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2237 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2239 set_task_comm(worker
->task
, "kworker/dying");
2240 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2241 worker_detach_from_pool(worker
, pool
);
2246 worker_leave_idle(worker
);
2248 /* no more worker necessary? */
2249 if (!need_more_worker(pool
))
2252 /* do we need to manage? */
2253 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2257 * ->scheduled list can only be filled while a worker is
2258 * preparing to process a work or actually processing it.
2259 * Make sure nobody diddled with it while I was sleeping.
2261 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2264 * Finish PREP stage. We're guaranteed to have at least one idle
2265 * worker or that someone else has already assumed the manager
2266 * role. This is where @worker starts participating in concurrency
2267 * management if applicable and concurrency management is restored
2268 * after being rebound. See rebind_workers() for details.
2270 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2273 struct work_struct
*work
=
2274 list_first_entry(&pool
->worklist
,
2275 struct work_struct
, entry
);
2277 pool
->watchdog_ts
= jiffies
;
2279 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2280 /* optimization path, not strictly necessary */
2281 process_one_work(worker
, work
);
2282 if (unlikely(!list_empty(&worker
->scheduled
)))
2283 process_scheduled_works(worker
);
2285 move_linked_works(work
, &worker
->scheduled
, NULL
);
2286 process_scheduled_works(worker
);
2288 } while (keep_working(pool
));
2290 worker_set_flags(worker
, WORKER_PREP
);
2293 * pool->lock is held and there's no work to process and no need to
2294 * manage, sleep. Workers are woken up only while holding
2295 * pool->lock or from local cpu, so setting the current state
2296 * before releasing pool->lock is enough to prevent losing any
2299 worker_enter_idle(worker
);
2300 __set_current_state(TASK_IDLE
);
2301 spin_unlock_irq(&pool
->lock
);
2307 * rescuer_thread - the rescuer thread function
2310 * Workqueue rescuer thread function. There's one rescuer for each
2311 * workqueue which has WQ_MEM_RECLAIM set.
2313 * Regular work processing on a pool may block trying to create a new
2314 * worker which uses GFP_KERNEL allocation which has slight chance of
2315 * developing into deadlock if some works currently on the same queue
2316 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2317 * the problem rescuer solves.
2319 * When such condition is possible, the pool summons rescuers of all
2320 * workqueues which have works queued on the pool and let them process
2321 * those works so that forward progress can be guaranteed.
2323 * This should happen rarely.
2327 static int rescuer_thread(void *__rescuer
)
2329 struct worker
*rescuer
= __rescuer
;
2330 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2331 struct list_head
*scheduled
= &rescuer
->scheduled
;
2334 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2337 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2338 * doesn't participate in concurrency management.
2340 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2342 set_current_state(TASK_IDLE
);
2345 * By the time the rescuer is requested to stop, the workqueue
2346 * shouldn't have any work pending, but @wq->maydays may still have
2347 * pwq(s) queued. This can happen by non-rescuer workers consuming
2348 * all the work items before the rescuer got to them. Go through
2349 * @wq->maydays processing before acting on should_stop so that the
2350 * list is always empty on exit.
2352 should_stop
= kthread_should_stop();
2354 /* see whether any pwq is asking for help */
2355 spin_lock_irq(&wq_mayday_lock
);
2357 while (!list_empty(&wq
->maydays
)) {
2358 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2359 struct pool_workqueue
, mayday_node
);
2360 struct worker_pool
*pool
= pwq
->pool
;
2361 struct work_struct
*work
, *n
;
2364 __set_current_state(TASK_RUNNING
);
2365 list_del_init(&pwq
->mayday_node
);
2367 spin_unlock_irq(&wq_mayday_lock
);
2369 worker_attach_to_pool(rescuer
, pool
);
2371 spin_lock_irq(&pool
->lock
);
2372 rescuer
->pool
= pool
;
2375 * Slurp in all works issued via this workqueue and
2378 WARN_ON_ONCE(!list_empty(scheduled
));
2379 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2380 if (get_work_pwq(work
) == pwq
) {
2382 pool
->watchdog_ts
= jiffies
;
2383 move_linked_works(work
, scheduled
, &n
);
2388 if (!list_empty(scheduled
)) {
2389 process_scheduled_works(rescuer
);
2392 * The above execution of rescued work items could
2393 * have created more to rescue through
2394 * pwq_activate_first_delayed() or chained
2395 * queueing. Let's put @pwq back on mayday list so
2396 * that such back-to-back work items, which may be
2397 * being used to relieve memory pressure, don't
2398 * incur MAYDAY_INTERVAL delay inbetween.
2400 if (need_to_create_worker(pool
)) {
2401 spin_lock(&wq_mayday_lock
);
2403 * Queue iff we aren't racing destruction
2404 * and somebody else hasn't queued it already.
2406 if (wq
->rescuer
&& list_empty(&pwq
->mayday_node
)) {
2408 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
2410 spin_unlock(&wq_mayday_lock
);
2415 * Put the reference grabbed by send_mayday(). @pool won't
2416 * go away while we're still attached to it.
2421 * Leave this pool. If need_more_worker() is %true, notify a
2422 * regular worker; otherwise, we end up with 0 concurrency
2423 * and stalling the execution.
2425 if (need_more_worker(pool
))
2426 wake_up_worker(pool
);
2428 rescuer
->pool
= NULL
;
2429 spin_unlock_irq(&pool
->lock
);
2431 worker_detach_from_pool(rescuer
, pool
);
2433 spin_lock_irq(&wq_mayday_lock
);
2436 spin_unlock_irq(&wq_mayday_lock
);
2439 __set_current_state(TASK_RUNNING
);
2440 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2444 /* rescuers should never participate in concurrency management */
2445 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2451 * check_flush_dependency - check for flush dependency sanity
2452 * @target_wq: workqueue being flushed
2453 * @target_work: work item being flushed (NULL for workqueue flushes)
2455 * %current is trying to flush the whole @target_wq or @target_work on it.
2456 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2457 * reclaiming memory or running on a workqueue which doesn't have
2458 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2461 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2462 struct work_struct
*target_work
)
2464 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2465 struct worker
*worker
;
2467 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2470 worker
= current_wq_worker();
2472 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2473 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2474 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2475 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2476 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2477 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2478 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2479 target_wq
->name
, target_func
);
2483 struct work_struct work
;
2484 struct completion done
;
2485 struct task_struct
*task
; /* purely informational */
2488 static void wq_barrier_func(struct work_struct
*work
)
2490 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2491 complete(&barr
->done
);
2495 * insert_wq_barrier - insert a barrier work
2496 * @pwq: pwq to insert barrier into
2497 * @barr: wq_barrier to insert
2498 * @target: target work to attach @barr to
2499 * @worker: worker currently executing @target, NULL if @target is not executing
2501 * @barr is linked to @target such that @barr is completed only after
2502 * @target finishes execution. Please note that the ordering
2503 * guarantee is observed only with respect to @target and on the local
2506 * Currently, a queued barrier can't be canceled. This is because
2507 * try_to_grab_pending() can't determine whether the work to be
2508 * grabbed is at the head of the queue and thus can't clear LINKED
2509 * flag of the previous work while there must be a valid next work
2510 * after a work with LINKED flag set.
2512 * Note that when @worker is non-NULL, @target may be modified
2513 * underneath us, so we can't reliably determine pwq from @target.
2516 * spin_lock_irq(pool->lock).
2518 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2519 struct wq_barrier
*barr
,
2520 struct work_struct
*target
, struct worker
*worker
)
2522 struct list_head
*head
;
2523 unsigned int linked
= 0;
2526 * debugobject calls are safe here even with pool->lock locked
2527 * as we know for sure that this will not trigger any of the
2528 * checks and call back into the fixup functions where we
2531 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2532 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2534 init_completion_map(&barr
->done
, &target
->lockdep_map
);
2536 barr
->task
= current
;
2539 * If @target is currently being executed, schedule the
2540 * barrier to the worker; otherwise, put it after @target.
2543 head
= worker
->scheduled
.next
;
2545 unsigned long *bits
= work_data_bits(target
);
2547 head
= target
->entry
.next
;
2548 /* there can already be other linked works, inherit and set */
2549 linked
= *bits
& WORK_STRUCT_LINKED
;
2550 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2553 debug_work_activate(&barr
->work
);
2554 insert_work(pwq
, &barr
->work
, head
,
2555 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2559 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2560 * @wq: workqueue being flushed
2561 * @flush_color: new flush color, < 0 for no-op
2562 * @work_color: new work color, < 0 for no-op
2564 * Prepare pwqs for workqueue flushing.
2566 * If @flush_color is non-negative, flush_color on all pwqs should be
2567 * -1. If no pwq has in-flight commands at the specified color, all
2568 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2569 * has in flight commands, its pwq->flush_color is set to
2570 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2571 * wakeup logic is armed and %true is returned.
2573 * The caller should have initialized @wq->first_flusher prior to
2574 * calling this function with non-negative @flush_color. If
2575 * @flush_color is negative, no flush color update is done and %false
2578 * If @work_color is non-negative, all pwqs should have the same
2579 * work_color which is previous to @work_color and all will be
2580 * advanced to @work_color.
2583 * mutex_lock(wq->mutex).
2586 * %true if @flush_color >= 0 and there's something to flush. %false
2589 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2590 int flush_color
, int work_color
)
2593 struct pool_workqueue
*pwq
;
2595 if (flush_color
>= 0) {
2596 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2597 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2600 for_each_pwq(pwq
, wq
) {
2601 struct worker_pool
*pool
= pwq
->pool
;
2603 spin_lock_irq(&pool
->lock
);
2605 if (flush_color
>= 0) {
2606 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2608 if (pwq
->nr_in_flight
[flush_color
]) {
2609 pwq
->flush_color
= flush_color
;
2610 atomic_inc(&wq
->nr_pwqs_to_flush
);
2615 if (work_color
>= 0) {
2616 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2617 pwq
->work_color
= work_color
;
2620 spin_unlock_irq(&pool
->lock
);
2623 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2624 complete(&wq
->first_flusher
->done
);
2630 * flush_workqueue - ensure that any scheduled work has run to completion.
2631 * @wq: workqueue to flush
2633 * This function sleeps until all work items which were queued on entry
2634 * have finished execution, but it is not livelocked by new incoming ones.
2636 void flush_workqueue(struct workqueue_struct
*wq
)
2638 struct wq_flusher this_flusher
= {
2639 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2641 .done
= COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher
.done
, wq
->lockdep_map
),
2645 if (WARN_ON(!wq_online
))
2648 lock_map_acquire(&wq
->lockdep_map
);
2649 lock_map_release(&wq
->lockdep_map
);
2651 mutex_lock(&wq
->mutex
);
2654 * Start-to-wait phase
2656 next_color
= work_next_color(wq
->work_color
);
2658 if (next_color
!= wq
->flush_color
) {
2660 * Color space is not full. The current work_color
2661 * becomes our flush_color and work_color is advanced
2664 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2665 this_flusher
.flush_color
= wq
->work_color
;
2666 wq
->work_color
= next_color
;
2668 if (!wq
->first_flusher
) {
2669 /* no flush in progress, become the first flusher */
2670 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2672 wq
->first_flusher
= &this_flusher
;
2674 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2676 /* nothing to flush, done */
2677 wq
->flush_color
= next_color
;
2678 wq
->first_flusher
= NULL
;
2683 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2684 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2685 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2689 * Oops, color space is full, wait on overflow queue.
2690 * The next flush completion will assign us
2691 * flush_color and transfer to flusher_queue.
2693 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2696 check_flush_dependency(wq
, NULL
);
2698 mutex_unlock(&wq
->mutex
);
2700 wait_for_completion(&this_flusher
.done
);
2703 * Wake-up-and-cascade phase
2705 * First flushers are responsible for cascading flushes and
2706 * handling overflow. Non-first flushers can simply return.
2708 if (wq
->first_flusher
!= &this_flusher
)
2711 mutex_lock(&wq
->mutex
);
2713 /* we might have raced, check again with mutex held */
2714 if (wq
->first_flusher
!= &this_flusher
)
2717 wq
->first_flusher
= NULL
;
2719 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2720 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2723 struct wq_flusher
*next
, *tmp
;
2725 /* complete all the flushers sharing the current flush color */
2726 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2727 if (next
->flush_color
!= wq
->flush_color
)
2729 list_del_init(&next
->list
);
2730 complete(&next
->done
);
2733 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2734 wq
->flush_color
!= work_next_color(wq
->work_color
));
2736 /* this flush_color is finished, advance by one */
2737 wq
->flush_color
= work_next_color(wq
->flush_color
);
2739 /* one color has been freed, handle overflow queue */
2740 if (!list_empty(&wq
->flusher_overflow
)) {
2742 * Assign the same color to all overflowed
2743 * flushers, advance work_color and append to
2744 * flusher_queue. This is the start-to-wait
2745 * phase for these overflowed flushers.
2747 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2748 tmp
->flush_color
= wq
->work_color
;
2750 wq
->work_color
= work_next_color(wq
->work_color
);
2752 list_splice_tail_init(&wq
->flusher_overflow
,
2753 &wq
->flusher_queue
);
2754 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2757 if (list_empty(&wq
->flusher_queue
)) {
2758 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2763 * Need to flush more colors. Make the next flusher
2764 * the new first flusher and arm pwqs.
2766 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2767 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2769 list_del_init(&next
->list
);
2770 wq
->first_flusher
= next
;
2772 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2776 * Meh... this color is already done, clear first
2777 * flusher and repeat cascading.
2779 wq
->first_flusher
= NULL
;
2783 mutex_unlock(&wq
->mutex
);
2785 EXPORT_SYMBOL(flush_workqueue
);
2788 * drain_workqueue - drain a workqueue
2789 * @wq: workqueue to drain
2791 * Wait until the workqueue becomes empty. While draining is in progress,
2792 * only chain queueing is allowed. IOW, only currently pending or running
2793 * work items on @wq can queue further work items on it. @wq is flushed
2794 * repeatedly until it becomes empty. The number of flushing is determined
2795 * by the depth of chaining and should be relatively short. Whine if it
2798 void drain_workqueue(struct workqueue_struct
*wq
)
2800 unsigned int flush_cnt
= 0;
2801 struct pool_workqueue
*pwq
;
2804 * __queue_work() needs to test whether there are drainers, is much
2805 * hotter than drain_workqueue() and already looks at @wq->flags.
2806 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2808 mutex_lock(&wq
->mutex
);
2809 if (!wq
->nr_drainers
++)
2810 wq
->flags
|= __WQ_DRAINING
;
2811 mutex_unlock(&wq
->mutex
);
2813 flush_workqueue(wq
);
2815 mutex_lock(&wq
->mutex
);
2817 for_each_pwq(pwq
, wq
) {
2820 spin_lock_irq(&pwq
->pool
->lock
);
2821 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2822 spin_unlock_irq(&pwq
->pool
->lock
);
2827 if (++flush_cnt
== 10 ||
2828 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2829 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2830 wq
->name
, flush_cnt
);
2832 mutex_unlock(&wq
->mutex
);
2836 if (!--wq
->nr_drainers
)
2837 wq
->flags
&= ~__WQ_DRAINING
;
2838 mutex_unlock(&wq
->mutex
);
2840 EXPORT_SYMBOL_GPL(drain_workqueue
);
2842 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2845 struct worker
*worker
= NULL
;
2846 struct worker_pool
*pool
;
2847 struct pool_workqueue
*pwq
;
2851 local_irq_disable();
2852 pool
= get_work_pool(work
);
2858 spin_lock(&pool
->lock
);
2859 /* see the comment in try_to_grab_pending() with the same code */
2860 pwq
= get_work_pwq(work
);
2862 if (unlikely(pwq
->pool
!= pool
))
2865 worker
= find_worker_executing_work(pool
, work
);
2868 pwq
= worker
->current_pwq
;
2871 check_flush_dependency(pwq
->wq
, work
);
2873 insert_wq_barrier(pwq
, barr
, work
, worker
);
2874 spin_unlock_irq(&pool
->lock
);
2877 * Force a lock recursion deadlock when using flush_work() inside a
2878 * single-threaded or rescuer equipped workqueue.
2880 * For single threaded workqueues the deadlock happens when the work
2881 * is after the work issuing the flush_work(). For rescuer equipped
2882 * workqueues the deadlock happens when the rescuer stalls, blocking
2886 (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)) {
2887 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2888 lock_map_release(&pwq
->wq
->lockdep_map
);
2893 spin_unlock_irq(&pool
->lock
);
2897 static bool __flush_work(struct work_struct
*work
, bool from_cancel
)
2899 struct wq_barrier barr
;
2901 if (WARN_ON(!wq_online
))
2904 if (WARN_ON(!work
->func
))
2908 lock_map_acquire(&work
->lockdep_map
);
2909 lock_map_release(&work
->lockdep_map
);
2912 if (start_flush_work(work
, &barr
, from_cancel
)) {
2913 wait_for_completion(&barr
.done
);
2914 destroy_work_on_stack(&barr
.work
);
2922 * flush_work - wait for a work to finish executing the last queueing instance
2923 * @work: the work to flush
2925 * Wait until @work has finished execution. @work is guaranteed to be idle
2926 * on return if it hasn't been requeued since flush started.
2929 * %true if flush_work() waited for the work to finish execution,
2930 * %false if it was already idle.
2932 bool flush_work(struct work_struct
*work
)
2934 return __flush_work(work
, false);
2936 EXPORT_SYMBOL_GPL(flush_work
);
2939 wait_queue_entry_t wait
;
2940 struct work_struct
*work
;
2943 static int cwt_wakefn(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
2945 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
2947 if (cwait
->work
!= key
)
2949 return autoremove_wake_function(wait
, mode
, sync
, key
);
2952 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2954 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
2955 unsigned long flags
;
2959 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2961 * If someone else is already canceling, wait for it to
2962 * finish. flush_work() doesn't work for PREEMPT_NONE
2963 * because we may get scheduled between @work's completion
2964 * and the other canceling task resuming and clearing
2965 * CANCELING - flush_work() will return false immediately
2966 * as @work is no longer busy, try_to_grab_pending() will
2967 * return -ENOENT as @work is still being canceled and the
2968 * other canceling task won't be able to clear CANCELING as
2969 * we're hogging the CPU.
2971 * Let's wait for completion using a waitqueue. As this
2972 * may lead to the thundering herd problem, use a custom
2973 * wake function which matches @work along with exclusive
2976 if (unlikely(ret
== -ENOENT
)) {
2977 struct cwt_wait cwait
;
2979 init_wait(&cwait
.wait
);
2980 cwait
.wait
.func
= cwt_wakefn
;
2983 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
2984 TASK_UNINTERRUPTIBLE
);
2985 if (work_is_canceling(work
))
2987 finish_wait(&cancel_waitq
, &cwait
.wait
);
2989 } while (unlikely(ret
< 0));
2991 /* tell other tasks trying to grab @work to back off */
2992 mark_work_canceling(work
);
2993 local_irq_restore(flags
);
2996 * This allows canceling during early boot. We know that @work
3000 __flush_work(work
, true);
3002 clear_work_data(work
);
3005 * Paired with prepare_to_wait() above so that either
3006 * waitqueue_active() is visible here or !work_is_canceling() is
3010 if (waitqueue_active(&cancel_waitq
))
3011 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
3017 * cancel_work_sync - cancel a work and wait for it to finish
3018 * @work: the work to cancel
3020 * Cancel @work and wait for its execution to finish. This function
3021 * can be used even if the work re-queues itself or migrates to
3022 * another workqueue. On return from this function, @work is
3023 * guaranteed to be not pending or executing on any CPU.
3025 * cancel_work_sync(&delayed_work->work) must not be used for
3026 * delayed_work's. Use cancel_delayed_work_sync() instead.
3028 * The caller must ensure that the workqueue on which @work was last
3029 * queued can't be destroyed before this function returns.
3032 * %true if @work was pending, %false otherwise.
3034 bool cancel_work_sync(struct work_struct
*work
)
3036 return __cancel_work_timer(work
, false);
3038 EXPORT_SYMBOL_GPL(cancel_work_sync
);
3041 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3042 * @dwork: the delayed work to flush
3044 * Delayed timer is cancelled and the pending work is queued for
3045 * immediate execution. Like flush_work(), this function only
3046 * considers the last queueing instance of @dwork.
3049 * %true if flush_work() waited for the work to finish execution,
3050 * %false if it was already idle.
3052 bool flush_delayed_work(struct delayed_work
*dwork
)
3054 local_irq_disable();
3055 if (del_timer_sync(&dwork
->timer
))
3056 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3058 return flush_work(&dwork
->work
);
3060 EXPORT_SYMBOL(flush_delayed_work
);
3063 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3064 * @rwork: the rcu work to flush
3067 * %true if flush_rcu_work() waited for the work to finish execution,
3068 * %false if it was already idle.
3070 bool flush_rcu_work(struct rcu_work
*rwork
)
3072 if (test_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&rwork
->work
))) {
3074 flush_work(&rwork
->work
);
3077 return flush_work(&rwork
->work
);
3080 EXPORT_SYMBOL(flush_rcu_work
);
3082 static bool __cancel_work(struct work_struct
*work
, bool is_dwork
)
3084 unsigned long flags
;
3088 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
3089 } while (unlikely(ret
== -EAGAIN
));
3091 if (unlikely(ret
< 0))
3094 set_work_pool_and_clear_pending(work
, get_work_pool_id(work
));
3095 local_irq_restore(flags
);
3100 * See cancel_delayed_work()
3102 bool cancel_work(struct work_struct
*work
)
3104 return __cancel_work(work
, false);
3108 * cancel_delayed_work - cancel a delayed work
3109 * @dwork: delayed_work to cancel
3111 * Kill off a pending delayed_work.
3113 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3117 * The work callback function may still be running on return, unless
3118 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3119 * use cancel_delayed_work_sync() to wait on it.
3121 * This function is safe to call from any context including IRQ handler.
3123 bool cancel_delayed_work(struct delayed_work
*dwork
)
3125 return __cancel_work(&dwork
->work
, true);
3127 EXPORT_SYMBOL(cancel_delayed_work
);
3130 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3131 * @dwork: the delayed work cancel
3133 * This is cancel_work_sync() for delayed works.
3136 * %true if @dwork was pending, %false otherwise.
3138 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3140 return __cancel_work_timer(&dwork
->work
, true);
3142 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3145 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3146 * @func: the function to call
3148 * schedule_on_each_cpu() executes @func on each online CPU using the
3149 * system workqueue and blocks until all CPUs have completed.
3150 * schedule_on_each_cpu() is very slow.
3153 * 0 on success, -errno on failure.
3155 int schedule_on_each_cpu(work_func_t func
)
3158 struct work_struct __percpu
*works
;
3160 works
= alloc_percpu(struct work_struct
);
3166 for_each_online_cpu(cpu
) {
3167 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3169 INIT_WORK(work
, func
);
3170 schedule_work_on(cpu
, work
);
3173 for_each_online_cpu(cpu
)
3174 flush_work(per_cpu_ptr(works
, cpu
));
3182 * execute_in_process_context - reliably execute the routine with user context
3183 * @fn: the function to execute
3184 * @ew: guaranteed storage for the execute work structure (must
3185 * be available when the work executes)
3187 * Executes the function immediately if process context is available,
3188 * otherwise schedules the function for delayed execution.
3190 * Return: 0 - function was executed
3191 * 1 - function was scheduled for execution
3193 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3195 if (!in_interrupt()) {
3200 INIT_WORK(&ew
->work
, fn
);
3201 schedule_work(&ew
->work
);
3205 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3208 * free_workqueue_attrs - free a workqueue_attrs
3209 * @attrs: workqueue_attrs to free
3211 * Undo alloc_workqueue_attrs().
3213 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3216 free_cpumask_var(attrs
->cpumask
);
3222 * alloc_workqueue_attrs - allocate a workqueue_attrs
3223 * @gfp_mask: allocation mask to use
3225 * Allocate a new workqueue_attrs, initialize with default settings and
3228 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3230 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3232 struct workqueue_attrs
*attrs
;
3234 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3237 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3240 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3243 free_workqueue_attrs(attrs
);
3247 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3248 const struct workqueue_attrs
*from
)
3250 to
->nice
= from
->nice
;
3251 cpumask_copy(to
->cpumask
, from
->cpumask
);
3253 * Unlike hash and equality test, this function doesn't ignore
3254 * ->no_numa as it is used for both pool and wq attrs. Instead,
3255 * get_unbound_pool() explicitly clears ->no_numa after copying.
3257 to
->no_numa
= from
->no_numa
;
3260 /* hash value of the content of @attr */
3261 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3265 hash
= jhash_1word(attrs
->nice
, hash
);
3266 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3267 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3271 /* content equality test */
3272 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3273 const struct workqueue_attrs
*b
)
3275 if (a
->nice
!= b
->nice
)
3277 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3283 * init_worker_pool - initialize a newly zalloc'd worker_pool
3284 * @pool: worker_pool to initialize
3286 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3288 * Return: 0 on success, -errno on failure. Even on failure, all fields
3289 * inside @pool proper are initialized and put_unbound_pool() can be called
3290 * on @pool safely to release it.
3292 static int init_worker_pool(struct worker_pool
*pool
)
3294 spin_lock_init(&pool
->lock
);
3297 pool
->node
= NUMA_NO_NODE
;
3298 pool
->flags
|= POOL_DISASSOCIATED
;
3299 pool
->watchdog_ts
= jiffies
;
3300 INIT_LIST_HEAD(&pool
->worklist
);
3301 INIT_LIST_HEAD(&pool
->idle_list
);
3302 hash_init(pool
->busy_hash
);
3304 timer_setup(&pool
->idle_timer
, idle_worker_timeout
, TIMER_DEFERRABLE
);
3306 timer_setup(&pool
->mayday_timer
, pool_mayday_timeout
, 0);
3308 mutex_init(&pool
->attach_mutex
);
3309 INIT_LIST_HEAD(&pool
->workers
);
3311 ida_init(&pool
->worker_ida
);
3312 INIT_HLIST_NODE(&pool
->hash_node
);
3315 /* shouldn't fail above this point */
3316 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3322 static void rcu_free_wq(struct rcu_head
*rcu
)
3324 struct workqueue_struct
*wq
=
3325 container_of(rcu
, struct workqueue_struct
, rcu
);
3327 if (!(wq
->flags
& WQ_UNBOUND
))
3328 free_percpu(wq
->cpu_pwqs
);
3330 free_workqueue_attrs(wq
->unbound_attrs
);
3336 static void rcu_free_pool(struct rcu_head
*rcu
)
3338 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3340 ida_destroy(&pool
->worker_ida
);
3341 free_workqueue_attrs(pool
->attrs
);
3346 * put_unbound_pool - put a worker_pool
3347 * @pool: worker_pool to put
3349 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3350 * safe manner. get_unbound_pool() calls this function on its failure path
3351 * and this function should be able to release pools which went through,
3352 * successfully or not, init_worker_pool().
3354 * Should be called with wq_pool_mutex held.
3356 static void put_unbound_pool(struct worker_pool
*pool
)
3358 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3359 struct worker
*worker
;
3361 lockdep_assert_held(&wq_pool_mutex
);
3367 if (WARN_ON(!(pool
->cpu
< 0)) ||
3368 WARN_ON(!list_empty(&pool
->worklist
)))
3371 /* release id and unhash */
3373 idr_remove(&worker_pool_idr
, pool
->id
);
3374 hash_del(&pool
->hash_node
);
3377 * Become the manager and destroy all workers. This prevents
3378 * @pool's workers from blocking on attach_mutex. We're the last
3379 * manager and @pool gets freed with the flag set.
3381 spin_lock_irq(&pool
->lock
);
3382 wait_event_lock_irq(wq_manager_wait
,
3383 !(pool
->flags
& POOL_MANAGER_ACTIVE
), pool
->lock
);
3384 pool
->flags
|= POOL_MANAGER_ACTIVE
;
3386 while ((worker
= first_idle_worker(pool
)))
3387 destroy_worker(worker
);
3388 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3389 spin_unlock_irq(&pool
->lock
);
3391 mutex_lock(&pool
->attach_mutex
);
3392 if (!list_empty(&pool
->workers
))
3393 pool
->detach_completion
= &detach_completion
;
3394 mutex_unlock(&pool
->attach_mutex
);
3396 if (pool
->detach_completion
)
3397 wait_for_completion(pool
->detach_completion
);
3399 /* shut down the timers */
3400 del_timer_sync(&pool
->idle_timer
);
3401 del_timer_sync(&pool
->mayday_timer
);
3403 /* sched-RCU protected to allow dereferences from get_work_pool() */
3404 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3408 * get_unbound_pool - get a worker_pool with the specified attributes
3409 * @attrs: the attributes of the worker_pool to get
3411 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3412 * reference count and return it. If there already is a matching
3413 * worker_pool, it will be used; otherwise, this function attempts to
3416 * Should be called with wq_pool_mutex held.
3418 * Return: On success, a worker_pool with the same attributes as @attrs.
3419 * On failure, %NULL.
3421 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3423 u32 hash
= wqattrs_hash(attrs
);
3424 struct worker_pool
*pool
;
3426 int target_node
= NUMA_NO_NODE
;
3428 lockdep_assert_held(&wq_pool_mutex
);
3430 /* do we already have a matching pool? */
3431 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3432 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3438 /* if cpumask is contained inside a NUMA node, we belong to that node */
3439 if (wq_numa_enabled
) {
3440 for_each_node(node
) {
3441 if (cpumask_subset(attrs
->cpumask
,
3442 wq_numa_possible_cpumask
[node
])) {
3449 /* nope, create a new one */
3450 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3451 if (!pool
|| init_worker_pool(pool
) < 0)
3454 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3455 copy_workqueue_attrs(pool
->attrs
, attrs
);
3456 pool
->node
= target_node
;
3459 * no_numa isn't a worker_pool attribute, always clear it. See
3460 * 'struct workqueue_attrs' comments for detail.
3462 pool
->attrs
->no_numa
= false;
3464 if (worker_pool_assign_id(pool
) < 0)
3467 /* create and start the initial worker */
3468 if (wq_online
&& !create_worker(pool
))
3472 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3477 put_unbound_pool(pool
);
3481 static void rcu_free_pwq(struct rcu_head
*rcu
)
3483 kmem_cache_free(pwq_cache
,
3484 container_of(rcu
, struct pool_workqueue
, rcu
));
3488 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3489 * and needs to be destroyed.
3491 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3493 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3494 unbound_release_work
);
3495 struct workqueue_struct
*wq
= pwq
->wq
;
3496 struct worker_pool
*pool
= pwq
->pool
;
3499 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3502 mutex_lock(&wq
->mutex
);
3503 list_del_rcu(&pwq
->pwqs_node
);
3504 is_last
= list_empty(&wq
->pwqs
);
3505 mutex_unlock(&wq
->mutex
);
3507 mutex_lock(&wq_pool_mutex
);
3508 put_unbound_pool(pool
);
3509 mutex_unlock(&wq_pool_mutex
);
3511 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3514 * If we're the last pwq going away, @wq is already dead and no one
3515 * is gonna access it anymore. Schedule RCU free.
3518 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
3522 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3523 * @pwq: target pool_workqueue
3525 * If @pwq isn't freezing, set @pwq->max_active to the associated
3526 * workqueue's saved_max_active and activate delayed work items
3527 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3529 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3531 struct workqueue_struct
*wq
= pwq
->wq
;
3532 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3533 unsigned long flags
;
3535 /* for @wq->saved_max_active */
3536 lockdep_assert_held(&wq
->mutex
);
3538 /* fast exit for non-freezable wqs */
3539 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3542 /* this function can be called during early boot w/ irq disabled */
3543 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
3546 * During [un]freezing, the caller is responsible for ensuring that
3547 * this function is called at least once after @workqueue_freezing
3548 * is updated and visible.
3550 if (!freezable
|| !workqueue_freezing
) {
3551 pwq
->max_active
= wq
->saved_max_active
;
3553 while (!list_empty(&pwq
->delayed_works
) &&
3554 pwq
->nr_active
< pwq
->max_active
)
3555 pwq_activate_first_delayed(pwq
);
3558 * Need to kick a worker after thawed or an unbound wq's
3559 * max_active is bumped. It's a slow path. Do it always.
3561 wake_up_worker(pwq
->pool
);
3563 pwq
->max_active
= 0;
3566 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
3569 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3570 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3571 struct worker_pool
*pool
)
3573 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3575 memset(pwq
, 0, sizeof(*pwq
));
3579 pwq
->flush_color
= -1;
3581 INIT_LIST_HEAD(&pwq
->delayed_works
);
3582 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3583 INIT_LIST_HEAD(&pwq
->mayday_node
);
3584 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3587 /* sync @pwq with the current state of its associated wq and link it */
3588 static void link_pwq(struct pool_workqueue
*pwq
)
3590 struct workqueue_struct
*wq
= pwq
->wq
;
3592 lockdep_assert_held(&wq
->mutex
);
3594 /* may be called multiple times, ignore if already linked */
3595 if (!list_empty(&pwq
->pwqs_node
))
3598 /* set the matching work_color */
3599 pwq
->work_color
= wq
->work_color
;
3601 /* sync max_active to the current setting */
3602 pwq_adjust_max_active(pwq
);
3605 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3608 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3609 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3610 const struct workqueue_attrs
*attrs
)
3612 struct worker_pool
*pool
;
3613 struct pool_workqueue
*pwq
;
3615 lockdep_assert_held(&wq_pool_mutex
);
3617 pool
= get_unbound_pool(attrs
);
3621 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3623 put_unbound_pool(pool
);
3627 init_pwq(pwq
, wq
, pool
);
3632 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3633 * @attrs: the wq_attrs of the default pwq of the target workqueue
3634 * @node: the target NUMA node
3635 * @cpu_going_down: if >= 0, the CPU to consider as offline
3636 * @cpumask: outarg, the resulting cpumask
3638 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3639 * @cpu_going_down is >= 0, that cpu is considered offline during
3640 * calculation. The result is stored in @cpumask.
3642 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3643 * enabled and @node has online CPUs requested by @attrs, the returned
3644 * cpumask is the intersection of the possible CPUs of @node and
3647 * The caller is responsible for ensuring that the cpumask of @node stays
3650 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3653 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3654 int cpu_going_down
, cpumask_t
*cpumask
)
3656 if (!wq_numa_enabled
|| attrs
->no_numa
)
3659 /* does @node have any online CPUs @attrs wants? */
3660 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3661 if (cpu_going_down
>= 0)
3662 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3664 if (cpumask_empty(cpumask
))
3667 /* yeap, return possible CPUs in @node that @attrs wants */
3668 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3670 if (cpumask_empty(cpumask
)) {
3671 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3672 "possible intersect\n");
3676 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3679 cpumask_copy(cpumask
, attrs
->cpumask
);
3683 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3684 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3686 struct pool_workqueue
*pwq
)
3688 struct pool_workqueue
*old_pwq
;
3690 lockdep_assert_held(&wq_pool_mutex
);
3691 lockdep_assert_held(&wq
->mutex
);
3693 /* link_pwq() can handle duplicate calls */
3696 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3697 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3701 /* context to store the prepared attrs & pwqs before applying */
3702 struct apply_wqattrs_ctx
{
3703 struct workqueue_struct
*wq
; /* target workqueue */
3704 struct workqueue_attrs
*attrs
; /* attrs to apply */
3705 struct list_head list
; /* queued for batching commit */
3706 struct pool_workqueue
*dfl_pwq
;
3707 struct pool_workqueue
*pwq_tbl
[];
3710 /* free the resources after success or abort */
3711 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3717 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3718 put_pwq_unlocked(ctx
->dfl_pwq
);
3720 free_workqueue_attrs(ctx
->attrs
);
3726 /* allocate the attrs and pwqs for later installation */
3727 static struct apply_wqattrs_ctx
*
3728 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3729 const struct workqueue_attrs
*attrs
)
3731 struct apply_wqattrs_ctx
*ctx
;
3732 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3735 lockdep_assert_held(&wq_pool_mutex
);
3737 ctx
= kzalloc(sizeof(*ctx
) + nr_node_ids
* sizeof(ctx
->pwq_tbl
[0]),
3740 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3741 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3742 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3746 * Calculate the attrs of the default pwq.
3747 * If the user configured cpumask doesn't overlap with the
3748 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3750 copy_workqueue_attrs(new_attrs
, attrs
);
3751 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3752 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3753 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3756 * We may create multiple pwqs with differing cpumasks. Make a
3757 * copy of @new_attrs which will be modified and used to obtain
3760 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3763 * If something goes wrong during CPU up/down, we'll fall back to
3764 * the default pwq covering whole @attrs->cpumask. Always create
3765 * it even if we don't use it immediately.
3767 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3771 for_each_node(node
) {
3772 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3773 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3774 if (!ctx
->pwq_tbl
[node
])
3777 ctx
->dfl_pwq
->refcnt
++;
3778 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3782 /* save the user configured attrs and sanitize it. */
3783 copy_workqueue_attrs(new_attrs
, attrs
);
3784 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3785 ctx
->attrs
= new_attrs
;
3788 free_workqueue_attrs(tmp_attrs
);
3792 free_workqueue_attrs(tmp_attrs
);
3793 free_workqueue_attrs(new_attrs
);
3794 apply_wqattrs_cleanup(ctx
);
3798 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3799 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3803 /* all pwqs have been created successfully, let's install'em */
3804 mutex_lock(&ctx
->wq
->mutex
);
3806 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3808 /* save the previous pwq and install the new one */
3810 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
3811 ctx
->pwq_tbl
[node
]);
3813 /* @dfl_pwq might not have been used, ensure it's linked */
3814 link_pwq(ctx
->dfl_pwq
);
3815 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
3817 mutex_unlock(&ctx
->wq
->mutex
);
3820 static void apply_wqattrs_lock(void)
3822 /* CPUs should stay stable across pwq creations and installations */
3824 mutex_lock(&wq_pool_mutex
);
3827 static void apply_wqattrs_unlock(void)
3829 mutex_unlock(&wq_pool_mutex
);
3833 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
3834 const struct workqueue_attrs
*attrs
)
3836 struct apply_wqattrs_ctx
*ctx
;
3838 /* only unbound workqueues can change attributes */
3839 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3842 /* creating multiple pwqs breaks ordering guarantee */
3843 if (!list_empty(&wq
->pwqs
)) {
3844 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
3847 wq
->flags
&= ~__WQ_ORDERED
;
3850 ctx
= apply_wqattrs_prepare(wq
, attrs
);
3854 /* the ctx has been prepared successfully, let's commit it */
3855 apply_wqattrs_commit(ctx
);
3856 apply_wqattrs_cleanup(ctx
);
3862 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3863 * @wq: the target workqueue
3864 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3866 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3867 * machines, this function maps a separate pwq to each NUMA node with
3868 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3869 * NUMA node it was issued on. Older pwqs are released as in-flight work
3870 * items finish. Note that a work item which repeatedly requeues itself
3871 * back-to-back will stay on its current pwq.
3873 * Performs GFP_KERNEL allocations.
3875 * Return: 0 on success and -errno on failure.
3877 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3878 const struct workqueue_attrs
*attrs
)
3882 apply_wqattrs_lock();
3883 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
3884 apply_wqattrs_unlock();
3890 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3891 * @wq: the target workqueue
3892 * @cpu: the CPU coming up or going down
3893 * @online: whether @cpu is coming up or going down
3895 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3896 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3899 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3900 * falls back to @wq->dfl_pwq which may not be optimal but is always
3903 * Note that when the last allowed CPU of a NUMA node goes offline for a
3904 * workqueue with a cpumask spanning multiple nodes, the workers which were
3905 * already executing the work items for the workqueue will lose their CPU
3906 * affinity and may execute on any CPU. This is similar to how per-cpu
3907 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3908 * affinity, it's the user's responsibility to flush the work item from
3911 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3914 int node
= cpu_to_node(cpu
);
3915 int cpu_off
= online
? -1 : cpu
;
3916 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3917 struct workqueue_attrs
*target_attrs
;
3920 lockdep_assert_held(&wq_pool_mutex
);
3922 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
3923 wq
->unbound_attrs
->no_numa
)
3927 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3928 * Let's use a preallocated one. The following buf is protected by
3929 * CPU hotplug exclusion.
3931 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3932 cpumask
= target_attrs
->cpumask
;
3934 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3935 pwq
= unbound_pwq_by_node(wq
, node
);
3938 * Let's determine what needs to be done. If the target cpumask is
3939 * different from the default pwq's, we need to compare it to @pwq's
3940 * and create a new one if they don't match. If the target cpumask
3941 * equals the default pwq's, the default pwq should be used.
3943 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
3944 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3950 /* create a new pwq */
3951 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3953 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3958 /* Install the new pwq. */
3959 mutex_lock(&wq
->mutex
);
3960 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3964 mutex_lock(&wq
->mutex
);
3965 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3966 get_pwq(wq
->dfl_pwq
);
3967 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3968 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
3970 mutex_unlock(&wq
->mutex
);
3971 put_pwq_unlocked(old_pwq
);
3974 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3976 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3979 if (!(wq
->flags
& WQ_UNBOUND
)) {
3980 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3984 for_each_possible_cpu(cpu
) {
3985 struct pool_workqueue
*pwq
=
3986 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3987 struct worker_pool
*cpu_pools
=
3988 per_cpu(cpu_worker_pools
, cpu
);
3990 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
3992 mutex_lock(&wq
->mutex
);
3994 mutex_unlock(&wq
->mutex
);
3997 } else if (wq
->flags
& __WQ_ORDERED
) {
3998 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
3999 /* there should only be single pwq for ordering guarantee */
4000 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4001 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4002 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4005 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4009 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4012 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4014 if (max_active
< 1 || max_active
> lim
)
4015 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4016 max_active
, name
, 1, lim
);
4018 return clamp_val(max_active
, 1, lim
);
4021 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4024 struct lock_class_key
*key
,
4025 const char *lock_name
, ...)
4027 size_t tbl_size
= 0;
4029 struct workqueue_struct
*wq
;
4030 struct pool_workqueue
*pwq
;
4033 * Unbound && max_active == 1 used to imply ordered, which is no
4034 * longer the case on NUMA machines due to per-node pools. While
4035 * alloc_ordered_workqueue() is the right way to create an ordered
4036 * workqueue, keep the previous behavior to avoid subtle breakages
4039 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
4040 flags
|= __WQ_ORDERED
;
4042 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4043 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
4044 flags
|= WQ_UNBOUND
;
4046 /* allocate wq and format name */
4047 if (flags
& WQ_UNBOUND
)
4048 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
4050 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4054 if (flags
& WQ_UNBOUND
) {
4055 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4056 if (!wq
->unbound_attrs
)
4060 va_start(args
, lock_name
);
4061 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4064 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4065 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4069 wq
->saved_max_active
= max_active
;
4070 mutex_init(&wq
->mutex
);
4071 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4072 INIT_LIST_HEAD(&wq
->pwqs
);
4073 INIT_LIST_HEAD(&wq
->flusher_queue
);
4074 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4075 INIT_LIST_HEAD(&wq
->maydays
);
4077 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4078 INIT_LIST_HEAD(&wq
->list
);
4080 if (alloc_and_link_pwqs(wq
) < 0)
4084 * Workqueues which may be used during memory reclaim should
4085 * have a rescuer to guarantee forward progress.
4087 if (flags
& WQ_MEM_RECLAIM
) {
4088 struct worker
*rescuer
;
4090 rescuer
= alloc_worker(NUMA_NO_NODE
);
4094 rescuer
->rescue_wq
= wq
;
4095 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4097 if (IS_ERR(rescuer
->task
)) {
4102 wq
->rescuer
= rescuer
;
4103 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
4104 wake_up_process(rescuer
->task
);
4107 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4111 * wq_pool_mutex protects global freeze state and workqueues list.
4112 * Grab it, adjust max_active and add the new @wq to workqueues
4115 mutex_lock(&wq_pool_mutex
);
4117 mutex_lock(&wq
->mutex
);
4118 for_each_pwq(pwq
, wq
)
4119 pwq_adjust_max_active(pwq
);
4120 mutex_unlock(&wq
->mutex
);
4122 list_add_tail_rcu(&wq
->list
, &workqueues
);
4124 mutex_unlock(&wq_pool_mutex
);
4129 free_workqueue_attrs(wq
->unbound_attrs
);
4133 destroy_workqueue(wq
);
4136 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4139 * destroy_workqueue - safely terminate a workqueue
4140 * @wq: target workqueue
4142 * Safely destroy a workqueue. All work currently pending will be done first.
4144 void destroy_workqueue(struct workqueue_struct
*wq
)
4146 struct pool_workqueue
*pwq
;
4150 * Remove it from sysfs first so that sanity check failure doesn't
4151 * lead to sysfs name conflicts.
4153 workqueue_sysfs_unregister(wq
);
4155 /* drain it before proceeding with destruction */
4156 drain_workqueue(wq
);
4158 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4160 struct worker
*rescuer
= wq
->rescuer
;
4162 /* this prevents new queueing */
4163 spin_lock_irq(&wq_mayday_lock
);
4165 spin_unlock_irq(&wq_mayday_lock
);
4167 /* rescuer will empty maydays list before exiting */
4168 kthread_stop(rescuer
->task
);
4173 mutex_lock(&wq
->mutex
);
4174 for_each_pwq(pwq
, wq
) {
4177 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4178 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4179 mutex_unlock(&wq
->mutex
);
4180 show_workqueue_state();
4185 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4186 WARN_ON(pwq
->nr_active
) ||
4187 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4188 mutex_unlock(&wq
->mutex
);
4189 show_workqueue_state();
4193 mutex_unlock(&wq
->mutex
);
4196 * wq list is used to freeze wq, remove from list after
4197 * flushing is complete in case freeze races us.
4199 mutex_lock(&wq_pool_mutex
);
4200 list_del_rcu(&wq
->list
);
4201 mutex_unlock(&wq_pool_mutex
);
4203 if (!(wq
->flags
& WQ_UNBOUND
)) {
4205 * The base ref is never dropped on per-cpu pwqs. Directly
4206 * schedule RCU free.
4208 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
4211 * We're the sole accessor of @wq at this point. Directly
4212 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4213 * @wq will be freed when the last pwq is released.
4215 for_each_node(node
) {
4216 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4217 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4218 put_pwq_unlocked(pwq
);
4222 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4223 * put. Don't access it afterwards.
4227 put_pwq_unlocked(pwq
);
4230 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4233 * workqueue_set_max_active - adjust max_active of a workqueue
4234 * @wq: target workqueue
4235 * @max_active: new max_active value.
4237 * Set max_active of @wq to @max_active.
4240 * Don't call from IRQ context.
4242 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4244 struct pool_workqueue
*pwq
;
4246 /* disallow meddling with max_active for ordered workqueues */
4247 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4250 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4252 mutex_lock(&wq
->mutex
);
4254 wq
->flags
&= ~__WQ_ORDERED
;
4255 wq
->saved_max_active
= max_active
;
4257 for_each_pwq(pwq
, wq
)
4258 pwq_adjust_max_active(pwq
);
4260 mutex_unlock(&wq
->mutex
);
4262 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4265 * current_work - retrieve %current task's work struct
4267 * Determine if %current task is a workqueue worker and what it's working on.
4268 * Useful to find out the context that the %current task is running in.
4270 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4272 struct work_struct
*current_work(void)
4274 struct worker
*worker
= current_wq_worker();
4276 return worker
? worker
->current_work
: NULL
;
4278 EXPORT_SYMBOL(current_work
);
4281 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4283 * Determine whether %current is a workqueue rescuer. Can be used from
4284 * work functions to determine whether it's being run off the rescuer task.
4286 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4288 bool current_is_workqueue_rescuer(void)
4290 struct worker
*worker
= current_wq_worker();
4292 return worker
&& worker
->rescue_wq
;
4296 * workqueue_congested - test whether a workqueue is congested
4297 * @cpu: CPU in question
4298 * @wq: target workqueue
4300 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4301 * no synchronization around this function and the test result is
4302 * unreliable and only useful as advisory hints or for debugging.
4304 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4305 * Note that both per-cpu and unbound workqueues may be associated with
4306 * multiple pool_workqueues which have separate congested states. A
4307 * workqueue being congested on one CPU doesn't mean the workqueue is also
4308 * contested on other CPUs / NUMA nodes.
4311 * %true if congested, %false otherwise.
4313 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4315 struct pool_workqueue
*pwq
;
4318 rcu_read_lock_sched();
4320 if (cpu
== WORK_CPU_UNBOUND
)
4321 cpu
= smp_processor_id();
4323 if (!(wq
->flags
& WQ_UNBOUND
))
4324 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4326 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4328 ret
= !list_empty(&pwq
->delayed_works
);
4329 rcu_read_unlock_sched();
4333 EXPORT_SYMBOL_GPL(workqueue_congested
);
4336 * work_busy - test whether a work is currently pending or running
4337 * @work: the work to be tested
4339 * Test whether @work is currently pending or running. There is no
4340 * synchronization around this function and the test result is
4341 * unreliable and only useful as advisory hints or for debugging.
4344 * OR'd bitmask of WORK_BUSY_* bits.
4346 unsigned int work_busy(struct work_struct
*work
)
4348 struct worker_pool
*pool
;
4349 unsigned long flags
;
4350 unsigned int ret
= 0;
4352 if (work_pending(work
))
4353 ret
|= WORK_BUSY_PENDING
;
4355 local_irq_save(flags
);
4356 pool
= get_work_pool(work
);
4358 spin_lock(&pool
->lock
);
4359 if (find_worker_executing_work(pool
, work
))
4360 ret
|= WORK_BUSY_RUNNING
;
4361 spin_unlock(&pool
->lock
);
4363 local_irq_restore(flags
);
4367 EXPORT_SYMBOL_GPL(work_busy
);
4370 * set_worker_desc - set description for the current work item
4371 * @fmt: printf-style format string
4372 * @...: arguments for the format string
4374 * This function can be called by a running work function to describe what
4375 * the work item is about. If the worker task gets dumped, this
4376 * information will be printed out together to help debugging. The
4377 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4379 void set_worker_desc(const char *fmt
, ...)
4381 struct worker
*worker
= current_wq_worker();
4385 va_start(args
, fmt
);
4386 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4388 worker
->desc_valid
= true;
4393 * print_worker_info - print out worker information and description
4394 * @log_lvl: the log level to use when printing
4395 * @task: target task
4397 * If @task is a worker and currently executing a work item, print out the
4398 * name of the workqueue being serviced and worker description set with
4399 * set_worker_desc() by the currently executing work item.
4401 * This function can be safely called on any task as long as the
4402 * task_struct itself is accessible. While safe, this function isn't
4403 * synchronized and may print out mixups or garbages of limited length.
4405 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4407 work_func_t
*fn
= NULL
;
4408 char name
[WQ_NAME_LEN
] = { };
4409 char desc
[WORKER_DESC_LEN
] = { };
4410 struct pool_workqueue
*pwq
= NULL
;
4411 struct workqueue_struct
*wq
= NULL
;
4412 bool desc_valid
= false;
4413 struct worker
*worker
;
4415 if (!(task
->flags
& PF_WQ_WORKER
))
4419 * This function is called without any synchronization and @task
4420 * could be in any state. Be careful with dereferences.
4422 worker
= kthread_probe_data(task
);
4425 * Carefully copy the associated workqueue's workfn and name. Keep
4426 * the original last '\0' in case the original contains garbage.
4428 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4429 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4430 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4431 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4433 /* copy worker description */
4434 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4436 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4438 if (fn
|| name
[0] || desc
[0]) {
4439 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4441 pr_cont(" (%s)", desc
);
4446 static void pr_cont_pool_info(struct worker_pool
*pool
)
4448 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4449 if (pool
->node
!= NUMA_NO_NODE
)
4450 pr_cont(" node=%d", pool
->node
);
4451 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4454 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4456 if (work
->func
== wq_barrier_func
) {
4457 struct wq_barrier
*barr
;
4459 barr
= container_of(work
, struct wq_barrier
, work
);
4461 pr_cont("%s BAR(%d)", comma
? "," : "",
4462 task_pid_nr(barr
->task
));
4464 pr_cont("%s %pf", comma
? "," : "", work
->func
);
4468 static void show_pwq(struct pool_workqueue
*pwq
)
4470 struct worker_pool
*pool
= pwq
->pool
;
4471 struct work_struct
*work
;
4472 struct worker
*worker
;
4473 bool has_in_flight
= false, has_pending
= false;
4476 pr_info(" pwq %d:", pool
->id
);
4477 pr_cont_pool_info(pool
);
4479 pr_cont(" active=%d/%d refcnt=%d%s\n",
4480 pwq
->nr_active
, pwq
->max_active
, pwq
->refcnt
,
4481 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4483 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4484 if (worker
->current_pwq
== pwq
) {
4485 has_in_flight
= true;
4489 if (has_in_flight
) {
4492 pr_info(" in-flight:");
4493 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4494 if (worker
->current_pwq
!= pwq
)
4497 pr_cont("%s %d%s:%pf", comma
? "," : "",
4498 task_pid_nr(worker
->task
),
4499 worker
== pwq
->wq
->rescuer
? "(RESCUER)" : "",
4500 worker
->current_func
);
4501 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4502 pr_cont_work(false, work
);
4508 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4509 if (get_work_pwq(work
) == pwq
) {
4517 pr_info(" pending:");
4518 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4519 if (get_work_pwq(work
) != pwq
)
4522 pr_cont_work(comma
, work
);
4523 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4528 if (!list_empty(&pwq
->delayed_works
)) {
4531 pr_info(" delayed:");
4532 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4533 pr_cont_work(comma
, work
);
4534 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4541 * show_workqueue_state - dump workqueue state
4543 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4544 * all busy workqueues and pools.
4546 void show_workqueue_state(void)
4548 struct workqueue_struct
*wq
;
4549 struct worker_pool
*pool
;
4550 unsigned long flags
;
4553 rcu_read_lock_sched();
4555 pr_info("Showing busy workqueues and worker pools:\n");
4557 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4558 struct pool_workqueue
*pwq
;
4561 for_each_pwq(pwq
, wq
) {
4562 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4570 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4572 for_each_pwq(pwq
, wq
) {
4573 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4574 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4576 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4578 * We could be printing a lot from atomic context, e.g.
4579 * sysrq-t -> show_workqueue_state(). Avoid triggering
4582 touch_nmi_watchdog();
4586 for_each_pool(pool
, pi
) {
4587 struct worker
*worker
;
4590 spin_lock_irqsave(&pool
->lock
, flags
);
4591 if (pool
->nr_workers
== pool
->nr_idle
)
4594 pr_info("pool %d:", pool
->id
);
4595 pr_cont_pool_info(pool
);
4596 pr_cont(" hung=%us workers=%d",
4597 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4600 pr_cont(" manager: %d",
4601 task_pid_nr(pool
->manager
->task
));
4602 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4603 pr_cont(" %s%d", first
? "idle: " : "",
4604 task_pid_nr(worker
->task
));
4609 spin_unlock_irqrestore(&pool
->lock
, flags
);
4611 * We could be printing a lot from atomic context, e.g.
4612 * sysrq-t -> show_workqueue_state(). Avoid triggering
4615 touch_nmi_watchdog();
4618 rcu_read_unlock_sched();
4624 * There are two challenges in supporting CPU hotplug. Firstly, there
4625 * are a lot of assumptions on strong associations among work, pwq and
4626 * pool which make migrating pending and scheduled works very
4627 * difficult to implement without impacting hot paths. Secondly,
4628 * worker pools serve mix of short, long and very long running works making
4629 * blocked draining impractical.
4631 * This is solved by allowing the pools to be disassociated from the CPU
4632 * running as an unbound one and allowing it to be reattached later if the
4633 * cpu comes back online.
4636 static void unbind_workers(int cpu
)
4638 struct worker_pool
*pool
;
4639 struct worker
*worker
;
4641 for_each_cpu_worker_pool(pool
, cpu
) {
4642 mutex_lock(&pool
->attach_mutex
);
4643 spin_lock_irq(&pool
->lock
);
4646 * We've blocked all attach/detach operations. Make all workers
4647 * unbound and set DISASSOCIATED. Before this, all workers
4648 * except for the ones which are still executing works from
4649 * before the last CPU down must be on the cpu. After
4650 * this, they may become diasporas.
4652 for_each_pool_worker(worker
, pool
)
4653 worker
->flags
|= WORKER_UNBOUND
;
4655 pool
->flags
|= POOL_DISASSOCIATED
;
4657 spin_unlock_irq(&pool
->lock
);
4658 mutex_unlock(&pool
->attach_mutex
);
4661 * Call schedule() so that we cross rq->lock and thus can
4662 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4663 * This is necessary as scheduler callbacks may be invoked
4669 * Sched callbacks are disabled now. Zap nr_running.
4670 * After this, nr_running stays zero and need_more_worker()
4671 * and keep_working() are always true as long as the
4672 * worklist is not empty. This pool now behaves as an
4673 * unbound (in terms of concurrency management) pool which
4674 * are served by workers tied to the pool.
4676 atomic_set(&pool
->nr_running
, 0);
4679 * With concurrency management just turned off, a busy
4680 * worker blocking could lead to lengthy stalls. Kick off
4681 * unbound chain execution of currently pending work items.
4683 spin_lock_irq(&pool
->lock
);
4684 wake_up_worker(pool
);
4685 spin_unlock_irq(&pool
->lock
);
4690 * rebind_workers - rebind all workers of a pool to the associated CPU
4691 * @pool: pool of interest
4693 * @pool->cpu is coming online. Rebind all workers to the CPU.
4695 static void rebind_workers(struct worker_pool
*pool
)
4697 struct worker
*worker
;
4699 lockdep_assert_held(&pool
->attach_mutex
);
4702 * Restore CPU affinity of all workers. As all idle workers should
4703 * be on the run-queue of the associated CPU before any local
4704 * wake-ups for concurrency management happen, restore CPU affinity
4705 * of all workers first and then clear UNBOUND. As we're called
4706 * from CPU_ONLINE, the following shouldn't fail.
4708 for_each_pool_worker(worker
, pool
)
4709 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4710 pool
->attrs
->cpumask
) < 0);
4712 spin_lock_irq(&pool
->lock
);
4714 pool
->flags
&= ~POOL_DISASSOCIATED
;
4716 for_each_pool_worker(worker
, pool
) {
4717 unsigned int worker_flags
= worker
->flags
;
4720 * A bound idle worker should actually be on the runqueue
4721 * of the associated CPU for local wake-ups targeting it to
4722 * work. Kick all idle workers so that they migrate to the
4723 * associated CPU. Doing this in the same loop as
4724 * replacing UNBOUND with REBOUND is safe as no worker will
4725 * be bound before @pool->lock is released.
4727 if (worker_flags
& WORKER_IDLE
)
4728 wake_up_process(worker
->task
);
4731 * We want to clear UNBOUND but can't directly call
4732 * worker_clr_flags() or adjust nr_running. Atomically
4733 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4734 * @worker will clear REBOUND using worker_clr_flags() when
4735 * it initiates the next execution cycle thus restoring
4736 * concurrency management. Note that when or whether
4737 * @worker clears REBOUND doesn't affect correctness.
4739 * WRITE_ONCE() is necessary because @worker->flags may be
4740 * tested without holding any lock in
4741 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4742 * fail incorrectly leading to premature concurrency
4743 * management operations.
4745 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4746 worker_flags
|= WORKER_REBOUND
;
4747 worker_flags
&= ~WORKER_UNBOUND
;
4748 WRITE_ONCE(worker
->flags
, worker_flags
);
4751 spin_unlock_irq(&pool
->lock
);
4755 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4756 * @pool: unbound pool of interest
4757 * @cpu: the CPU which is coming up
4759 * An unbound pool may end up with a cpumask which doesn't have any online
4760 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4761 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4762 * online CPU before, cpus_allowed of all its workers should be restored.
4764 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4766 static cpumask_t cpumask
;
4767 struct worker
*worker
;
4769 lockdep_assert_held(&pool
->attach_mutex
);
4771 /* is @cpu allowed for @pool? */
4772 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4775 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4777 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4778 for_each_pool_worker(worker
, pool
)
4779 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
, &cpumask
) < 0);
4782 int workqueue_prepare_cpu(unsigned int cpu
)
4784 struct worker_pool
*pool
;
4786 for_each_cpu_worker_pool(pool
, cpu
) {
4787 if (pool
->nr_workers
)
4789 if (!create_worker(pool
))
4795 int workqueue_online_cpu(unsigned int cpu
)
4797 struct worker_pool
*pool
;
4798 struct workqueue_struct
*wq
;
4801 mutex_lock(&wq_pool_mutex
);
4803 for_each_pool(pool
, pi
) {
4804 mutex_lock(&pool
->attach_mutex
);
4806 if (pool
->cpu
== cpu
)
4807 rebind_workers(pool
);
4808 else if (pool
->cpu
< 0)
4809 restore_unbound_workers_cpumask(pool
, cpu
);
4811 mutex_unlock(&pool
->attach_mutex
);
4814 /* update NUMA affinity of unbound workqueues */
4815 list_for_each_entry(wq
, &workqueues
, list
)
4816 wq_update_unbound_numa(wq
, cpu
, true);
4818 mutex_unlock(&wq_pool_mutex
);
4822 int workqueue_offline_cpu(unsigned int cpu
)
4824 struct workqueue_struct
*wq
;
4826 /* unbinding per-cpu workers should happen on the local CPU */
4827 if (WARN_ON(cpu
!= smp_processor_id()))
4830 unbind_workers(cpu
);
4832 /* update NUMA affinity of unbound workqueues */
4833 mutex_lock(&wq_pool_mutex
);
4834 list_for_each_entry(wq
, &workqueues
, list
)
4835 wq_update_unbound_numa(wq
, cpu
, false);
4836 mutex_unlock(&wq_pool_mutex
);
4843 struct work_for_cpu
{
4844 struct work_struct work
;
4850 static void work_for_cpu_fn(struct work_struct
*work
)
4852 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4854 wfc
->ret
= wfc
->fn(wfc
->arg
);
4858 * work_on_cpu - run a function in thread context on a particular cpu
4859 * @cpu: the cpu to run on
4860 * @fn: the function to run
4861 * @arg: the function arg
4863 * It is up to the caller to ensure that the cpu doesn't go offline.
4864 * The caller must not hold any locks which would prevent @fn from completing.
4866 * Return: The value @fn returns.
4868 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4870 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4872 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4873 schedule_work_on(cpu
, &wfc
.work
);
4874 flush_work(&wfc
.work
);
4875 destroy_work_on_stack(&wfc
.work
);
4878 EXPORT_SYMBOL_GPL(work_on_cpu
);
4881 * work_on_cpu_safe - run a function in thread context on a particular cpu
4882 * @cpu: the cpu to run on
4883 * @fn: the function to run
4884 * @arg: the function argument
4886 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4887 * any locks which would prevent @fn from completing.
4889 * Return: The value @fn returns.
4891 long work_on_cpu_safe(int cpu
, long (*fn
)(void *), void *arg
)
4896 if (cpu_online(cpu
))
4897 ret
= work_on_cpu(cpu
, fn
, arg
);
4901 EXPORT_SYMBOL_GPL(work_on_cpu_safe
);
4902 #endif /* CONFIG_SMP */
4904 #ifdef CONFIG_FREEZER
4907 * freeze_workqueues_begin - begin freezing workqueues
4909 * Start freezing workqueues. After this function returns, all freezable
4910 * workqueues will queue new works to their delayed_works list instead of
4914 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4916 void freeze_workqueues_begin(void)
4918 struct workqueue_struct
*wq
;
4919 struct pool_workqueue
*pwq
;
4921 mutex_lock(&wq_pool_mutex
);
4923 WARN_ON_ONCE(workqueue_freezing
);
4924 workqueue_freezing
= true;
4926 list_for_each_entry(wq
, &workqueues
, list
) {
4927 mutex_lock(&wq
->mutex
);
4928 for_each_pwq(pwq
, wq
)
4929 pwq_adjust_max_active(pwq
);
4930 mutex_unlock(&wq
->mutex
);
4933 mutex_unlock(&wq_pool_mutex
);
4937 * freeze_workqueues_busy - are freezable workqueues still busy?
4939 * Check whether freezing is complete. This function must be called
4940 * between freeze_workqueues_begin() and thaw_workqueues().
4943 * Grabs and releases wq_pool_mutex.
4946 * %true if some freezable workqueues are still busy. %false if freezing
4949 bool freeze_workqueues_busy(void)
4952 struct workqueue_struct
*wq
;
4953 struct pool_workqueue
*pwq
;
4955 mutex_lock(&wq_pool_mutex
);
4957 WARN_ON_ONCE(!workqueue_freezing
);
4959 list_for_each_entry(wq
, &workqueues
, list
) {
4960 if (!(wq
->flags
& WQ_FREEZABLE
))
4963 * nr_active is monotonically decreasing. It's safe
4964 * to peek without lock.
4966 rcu_read_lock_sched();
4967 for_each_pwq(pwq
, wq
) {
4968 WARN_ON_ONCE(pwq
->nr_active
< 0);
4969 if (pwq
->nr_active
) {
4971 rcu_read_unlock_sched();
4975 rcu_read_unlock_sched();
4978 mutex_unlock(&wq_pool_mutex
);
4983 * thaw_workqueues - thaw workqueues
4985 * Thaw workqueues. Normal queueing is restored and all collected
4986 * frozen works are transferred to their respective pool worklists.
4989 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4991 void thaw_workqueues(void)
4993 struct workqueue_struct
*wq
;
4994 struct pool_workqueue
*pwq
;
4996 mutex_lock(&wq_pool_mutex
);
4998 if (!workqueue_freezing
)
5001 workqueue_freezing
= false;
5003 /* restore max_active and repopulate worklist */
5004 list_for_each_entry(wq
, &workqueues
, list
) {
5005 mutex_lock(&wq
->mutex
);
5006 for_each_pwq(pwq
, wq
)
5007 pwq_adjust_max_active(pwq
);
5008 mutex_unlock(&wq
->mutex
);
5012 mutex_unlock(&wq_pool_mutex
);
5014 #endif /* CONFIG_FREEZER */
5016 static int workqueue_apply_unbound_cpumask(void)
5020 struct workqueue_struct
*wq
;
5021 struct apply_wqattrs_ctx
*ctx
, *n
;
5023 lockdep_assert_held(&wq_pool_mutex
);
5025 list_for_each_entry(wq
, &workqueues
, list
) {
5026 if (!(wq
->flags
& WQ_UNBOUND
))
5028 /* creating multiple pwqs breaks ordering guarantee */
5029 if (wq
->flags
& __WQ_ORDERED
)
5032 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
5038 list_add_tail(&ctx
->list
, &ctxs
);
5041 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
5043 apply_wqattrs_commit(ctx
);
5044 apply_wqattrs_cleanup(ctx
);
5051 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5052 * @cpumask: the cpumask to set
5054 * The low-level workqueues cpumask is a global cpumask that limits
5055 * the affinity of all unbound workqueues. This function check the @cpumask
5056 * and apply it to all unbound workqueues and updates all pwqs of them.
5058 * Retun: 0 - Success
5059 * -EINVAL - Invalid @cpumask
5060 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5062 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
5065 cpumask_var_t saved_cpumask
;
5067 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
5071 * Not excluding isolated cpus on purpose.
5072 * If the user wishes to include them, we allow that.
5074 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
5075 if (!cpumask_empty(cpumask
)) {
5076 apply_wqattrs_lock();
5078 /* save the old wq_unbound_cpumask. */
5079 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
5081 /* update wq_unbound_cpumask at first and apply it to wqs. */
5082 cpumask_copy(wq_unbound_cpumask
, cpumask
);
5083 ret
= workqueue_apply_unbound_cpumask();
5085 /* restore the wq_unbound_cpumask when failed. */
5087 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
5089 apply_wqattrs_unlock();
5092 free_cpumask_var(saved_cpumask
);
5098 * Workqueues with WQ_SYSFS flag set is visible to userland via
5099 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5100 * following attributes.
5102 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5103 * max_active RW int : maximum number of in-flight work items
5105 * Unbound workqueues have the following extra attributes.
5107 * pool_ids RO int : the associated pool IDs for each node
5108 * nice RW int : nice value of the workers
5109 * cpumask RW mask : bitmask of allowed CPUs for the workers
5110 * numa RW bool : whether enable NUMA affinity
5113 struct workqueue_struct
*wq
;
5117 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
5119 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5124 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
5127 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5129 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
5131 static DEVICE_ATTR_RO(per_cpu
);
5133 static ssize_t
max_active_show(struct device
*dev
,
5134 struct device_attribute
*attr
, char *buf
)
5136 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5138 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
5141 static ssize_t
max_active_store(struct device
*dev
,
5142 struct device_attribute
*attr
, const char *buf
,
5145 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5148 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
5151 workqueue_set_max_active(wq
, val
);
5154 static DEVICE_ATTR_RW(max_active
);
5156 static struct attribute
*wq_sysfs_attrs
[] = {
5157 &dev_attr_per_cpu
.attr
,
5158 &dev_attr_max_active
.attr
,
5161 ATTRIBUTE_GROUPS(wq_sysfs
);
5163 static ssize_t
wq_pool_ids_show(struct device
*dev
,
5164 struct device_attribute
*attr
, char *buf
)
5166 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5167 const char *delim
= "";
5168 int node
, written
= 0;
5170 rcu_read_lock_sched();
5171 for_each_node(node
) {
5172 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
5173 "%s%d:%d", delim
, node
,
5174 unbound_pwq_by_node(wq
, node
)->pool
->id
);
5177 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
5178 rcu_read_unlock_sched();
5183 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
5186 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5189 mutex_lock(&wq
->mutex
);
5190 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
5191 mutex_unlock(&wq
->mutex
);
5196 /* prepare workqueue_attrs for sysfs store operations */
5197 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
5199 struct workqueue_attrs
*attrs
;
5201 lockdep_assert_held(&wq_pool_mutex
);
5203 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
5207 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
5211 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
5212 const char *buf
, size_t count
)
5214 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5215 struct workqueue_attrs
*attrs
;
5218 apply_wqattrs_lock();
5220 attrs
= wq_sysfs_prep_attrs(wq
);
5224 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5225 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5226 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5231 apply_wqattrs_unlock();
5232 free_workqueue_attrs(attrs
);
5233 return ret
?: count
;
5236 static ssize_t
wq_cpumask_show(struct device
*dev
,
5237 struct device_attribute
*attr
, char *buf
)
5239 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5242 mutex_lock(&wq
->mutex
);
5243 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5244 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5245 mutex_unlock(&wq
->mutex
);
5249 static ssize_t
wq_cpumask_store(struct device
*dev
,
5250 struct device_attribute
*attr
,
5251 const char *buf
, size_t count
)
5253 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5254 struct workqueue_attrs
*attrs
;
5257 apply_wqattrs_lock();
5259 attrs
= wq_sysfs_prep_attrs(wq
);
5263 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5265 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5268 apply_wqattrs_unlock();
5269 free_workqueue_attrs(attrs
);
5270 return ret
?: count
;
5273 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5276 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5279 mutex_lock(&wq
->mutex
);
5280 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5281 !wq
->unbound_attrs
->no_numa
);
5282 mutex_unlock(&wq
->mutex
);
5287 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5288 const char *buf
, size_t count
)
5290 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5291 struct workqueue_attrs
*attrs
;
5292 int v
, ret
= -ENOMEM
;
5294 apply_wqattrs_lock();
5296 attrs
= wq_sysfs_prep_attrs(wq
);
5301 if (sscanf(buf
, "%d", &v
) == 1) {
5302 attrs
->no_numa
= !v
;
5303 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5307 apply_wqattrs_unlock();
5308 free_workqueue_attrs(attrs
);
5309 return ret
?: count
;
5312 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5313 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5314 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5315 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5316 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5320 static struct bus_type wq_subsys
= {
5321 .name
= "workqueue",
5322 .dev_groups
= wq_sysfs_groups
,
5325 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5326 struct device_attribute
*attr
, char *buf
)
5330 mutex_lock(&wq_pool_mutex
);
5331 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5332 cpumask_pr_args(wq_unbound_cpumask
));
5333 mutex_unlock(&wq_pool_mutex
);
5338 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5339 struct device_attribute
*attr
, const char *buf
, size_t count
)
5341 cpumask_var_t cpumask
;
5344 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5347 ret
= cpumask_parse(buf
, cpumask
);
5349 ret
= workqueue_set_unbound_cpumask(cpumask
);
5351 free_cpumask_var(cpumask
);
5352 return ret
? ret
: count
;
5355 static struct device_attribute wq_sysfs_cpumask_attr
=
5356 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5357 wq_unbound_cpumask_store
);
5359 static int __init
wq_sysfs_init(void)
5363 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5367 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5369 core_initcall(wq_sysfs_init
);
5371 static void wq_device_release(struct device
*dev
)
5373 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5379 * workqueue_sysfs_register - make a workqueue visible in sysfs
5380 * @wq: the workqueue to register
5382 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5383 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5384 * which is the preferred method.
5386 * Workqueue user should use this function directly iff it wants to apply
5387 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5388 * apply_workqueue_attrs() may race against userland updating the
5391 * Return: 0 on success, -errno on failure.
5393 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5395 struct wq_device
*wq_dev
;
5399 * Adjusting max_active or creating new pwqs by applying
5400 * attributes breaks ordering guarantee. Disallow exposing ordered
5403 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
5406 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5411 wq_dev
->dev
.bus
= &wq_subsys
;
5412 wq_dev
->dev
.release
= wq_device_release
;
5413 dev_set_name(&wq_dev
->dev
, "%s", wq
->name
);
5416 * unbound_attrs are created separately. Suppress uevent until
5417 * everything is ready.
5419 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5421 ret
= device_register(&wq_dev
->dev
);
5423 put_device(&wq_dev
->dev
);
5428 if (wq
->flags
& WQ_UNBOUND
) {
5429 struct device_attribute
*attr
;
5431 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5432 ret
= device_create_file(&wq_dev
->dev
, attr
);
5434 device_unregister(&wq_dev
->dev
);
5441 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5442 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5447 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5448 * @wq: the workqueue to unregister
5450 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5452 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5454 struct wq_device
*wq_dev
= wq
->wq_dev
;
5460 device_unregister(&wq_dev
->dev
);
5462 #else /* CONFIG_SYSFS */
5463 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5464 #endif /* CONFIG_SYSFS */
5467 * Workqueue watchdog.
5469 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5470 * flush dependency, a concurrency managed work item which stays RUNNING
5471 * indefinitely. Workqueue stalls can be very difficult to debug as the
5472 * usual warning mechanisms don't trigger and internal workqueue state is
5475 * Workqueue watchdog monitors all worker pools periodically and dumps
5476 * state if some pools failed to make forward progress for a while where
5477 * forward progress is defined as the first item on ->worklist changing.
5479 * This mechanism is controlled through the kernel parameter
5480 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5481 * corresponding sysfs parameter file.
5483 #ifdef CONFIG_WQ_WATCHDOG
5485 static unsigned long wq_watchdog_thresh
= 30;
5486 static struct timer_list wq_watchdog_timer
;
5488 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5489 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5491 static void wq_watchdog_reset_touched(void)
5495 wq_watchdog_touched
= jiffies
;
5496 for_each_possible_cpu(cpu
)
5497 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5500 static void wq_watchdog_timer_fn(struct timer_list
*unused
)
5502 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5503 bool lockup_detected
= false;
5504 struct worker_pool
*pool
;
5512 for_each_pool(pool
, pi
) {
5513 unsigned long pool_ts
, touched
, ts
;
5515 if (list_empty(&pool
->worklist
))
5518 /* get the latest of pool and touched timestamps */
5519 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5520 touched
= READ_ONCE(wq_watchdog_touched
);
5522 if (time_after(pool_ts
, touched
))
5527 if (pool
->cpu
>= 0) {
5528 unsigned long cpu_touched
=
5529 READ_ONCE(per_cpu(wq_watchdog_touched_cpu
,
5531 if (time_after(cpu_touched
, ts
))
5536 if (time_after(jiffies
, ts
+ thresh
)) {
5537 lockup_detected
= true;
5538 pr_emerg("BUG: workqueue lockup - pool");
5539 pr_cont_pool_info(pool
);
5540 pr_cont(" stuck for %us!\n",
5541 jiffies_to_msecs(jiffies
- pool_ts
) / 1000);
5547 if (lockup_detected
)
5548 show_workqueue_state();
5550 wq_watchdog_reset_touched();
5551 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5554 notrace
void wq_watchdog_touch(int cpu
)
5557 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5559 wq_watchdog_touched
= jiffies
;
5562 static void wq_watchdog_set_thresh(unsigned long thresh
)
5564 wq_watchdog_thresh
= 0;
5565 del_timer_sync(&wq_watchdog_timer
);
5568 wq_watchdog_thresh
= thresh
;
5569 wq_watchdog_reset_touched();
5570 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5574 static int wq_watchdog_param_set_thresh(const char *val
,
5575 const struct kernel_param
*kp
)
5577 unsigned long thresh
;
5580 ret
= kstrtoul(val
, 0, &thresh
);
5585 wq_watchdog_set_thresh(thresh
);
5587 wq_watchdog_thresh
= thresh
;
5592 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5593 .set
= wq_watchdog_param_set_thresh
,
5594 .get
= param_get_ulong
,
5597 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5600 static void wq_watchdog_init(void)
5602 timer_setup(&wq_watchdog_timer
, wq_watchdog_timer_fn
, TIMER_DEFERRABLE
);
5603 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5606 #else /* CONFIG_WQ_WATCHDOG */
5608 static inline void wq_watchdog_init(void) { }
5610 #endif /* CONFIG_WQ_WATCHDOG */
5612 static void __init
wq_numa_init(void)
5617 if (num_possible_nodes() <= 1)
5620 if (wq_disable_numa
) {
5621 pr_info("workqueue: NUMA affinity support disabled\n");
5625 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
5626 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5629 * We want masks of possible CPUs of each node which isn't readily
5630 * available. Build one from cpu_to_node() which should have been
5631 * fully initialized by now.
5633 tbl
= kzalloc(nr_node_ids
* sizeof(tbl
[0]), GFP_KERNEL
);
5637 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5638 node_online(node
) ? node
: NUMA_NO_NODE
));
5640 for_each_possible_cpu(cpu
) {
5641 node
= cpu_to_node(cpu
);
5642 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5643 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5644 /* happens iff arch is bonkers, let's just proceed */
5647 cpumask_set_cpu(cpu
, tbl
[node
]);
5650 wq_numa_possible_cpumask
= tbl
;
5651 wq_numa_enabled
= true;
5655 * workqueue_init_early - early init for workqueue subsystem
5657 * This is the first half of two-staged workqueue subsystem initialization
5658 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5659 * idr are up. It sets up all the data structures and system workqueues
5660 * and allows early boot code to create workqueues and queue/cancel work
5661 * items. Actual work item execution starts only after kthreads can be
5662 * created and scheduled right before early initcalls.
5664 int __init
workqueue_init_early(void)
5666 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5669 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5671 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5672 cpumask_copy(wq_unbound_cpumask
, housekeeping_cpumask(HK_FLAG_DOMAIN
));
5674 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5676 /* initialize CPU pools */
5677 for_each_possible_cpu(cpu
) {
5678 struct worker_pool
*pool
;
5681 for_each_cpu_worker_pool(pool
, cpu
) {
5682 BUG_ON(init_worker_pool(pool
));
5684 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5685 pool
->attrs
->nice
= std_nice
[i
++];
5686 pool
->node
= cpu_to_node(cpu
);
5689 mutex_lock(&wq_pool_mutex
);
5690 BUG_ON(worker_pool_assign_id(pool
));
5691 mutex_unlock(&wq_pool_mutex
);
5695 /* create default unbound and ordered wq attrs */
5696 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5697 struct workqueue_attrs
*attrs
;
5699 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5700 attrs
->nice
= std_nice
[i
];
5701 unbound_std_wq_attrs
[i
] = attrs
;
5704 * An ordered wq should have only one pwq as ordering is
5705 * guaranteed by max_active which is enforced by pwqs.
5706 * Turn off NUMA so that dfl_pwq is used for all nodes.
5708 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5709 attrs
->nice
= std_nice
[i
];
5710 attrs
->no_numa
= true;
5711 ordered_wq_attrs
[i
] = attrs
;
5714 system_wq
= alloc_workqueue("events", 0, 0);
5715 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5716 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5717 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5718 WQ_UNBOUND_MAX_ACTIVE
);
5719 system_freezable_wq
= alloc_workqueue("events_freezable",
5721 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5722 WQ_POWER_EFFICIENT
, 0);
5723 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5724 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5726 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5727 !system_unbound_wq
|| !system_freezable_wq
||
5728 !system_power_efficient_wq
||
5729 !system_freezable_power_efficient_wq
);
5735 * workqueue_init - bring workqueue subsystem fully online
5737 * This is the latter half of two-staged workqueue subsystem initialization
5738 * and invoked as soon as kthreads can be created and scheduled.
5739 * Workqueues have been created and work items queued on them, but there
5740 * are no kworkers executing the work items yet. Populate the worker pools
5741 * with the initial workers and enable future kworker creations.
5743 int __init
workqueue_init(void)
5745 struct workqueue_struct
*wq
;
5746 struct worker_pool
*pool
;
5750 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5751 * CPU to node mapping may not be available that early on some
5752 * archs such as power and arm64. As per-cpu pools created
5753 * previously could be missing node hint and unbound pools NUMA
5754 * affinity, fix them up.
5758 mutex_lock(&wq_pool_mutex
);
5760 for_each_possible_cpu(cpu
) {
5761 for_each_cpu_worker_pool(pool
, cpu
) {
5762 pool
->node
= cpu_to_node(cpu
);
5766 list_for_each_entry(wq
, &workqueues
, list
)
5767 wq_update_unbound_numa(wq
, smp_processor_id(), true);
5769 mutex_unlock(&wq_pool_mutex
);
5771 /* create the initial workers */
5772 for_each_online_cpu(cpu
) {
5773 for_each_cpu_worker_pool(pool
, cpu
) {
5774 pool
->flags
&= ~POOL_DISASSOCIATED
;
5775 BUG_ON(!create_worker(pool
));
5779 hash_for_each(unbound_pool_hash
, bkt
, pool
, hash_node
)
5780 BUG_ON(!create_worker(pool
));