2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE
= 1 << 1, /* die die die */
75 WORKER_IDLE
= 1 << 2, /* is idle */
76 WORKER_PREP
= 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
79 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
82 WORKER_UNBOUND
| WORKER_REBOUND
,
84 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
96 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL
= MIN_NICE
,
103 HIGHPRI_NICE_LEVEL
= MIN_NICE
,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
135 * WQ: wq->mutex protected.
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
139 * MD: wq_mayday_lock protected.
142 /* struct worker is defined in workqueue_internal.h */
145 spinlock_t lock
; /* the pool lock */
146 int cpu
; /* I: the associated cpu */
147 int node
; /* I: the associated node ID */
148 int id
; /* I: pool ID */
149 unsigned int flags
; /* X: flags */
151 unsigned long watchdog_ts
; /* L: watchdog timestamp */
153 struct list_head worklist
; /* L: list of pending works */
154 int nr_workers
; /* L: total number of workers */
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle
; /* L: currently idle ones */
159 struct list_head idle_list
; /* X: list of idle workers */
160 struct timer_list idle_timer
; /* L: worker idle timeout */
161 struct timer_list mayday_timer
; /* L: SOS timer for workers */
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
165 /* L: hash of busy workers */
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb
; /* manager arbitration */
169 struct worker
*manager
; /* L: purely informational */
170 struct mutex attach_mutex
; /* attach/detach exclusion */
171 struct list_head workers
; /* A: attached workers */
172 struct completion
*detach_completion
; /* all workers detached */
174 struct ida worker_ida
; /* worker IDs for task name */
176 struct workqueue_attrs
*attrs
; /* I: worker attributes */
177 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
178 int refcnt
; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp
;
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp
;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue
{
201 struct worker_pool
*pool
; /* I: the associated pool */
202 struct workqueue_struct
*wq
; /* I: the owning workqueue */
203 int work_color
; /* L: current color */
204 int flush_color
; /* L: flushing color */
205 int refcnt
; /* L: reference count */
206 int nr_in_flight
[WORK_NR_COLORS
];
207 /* L: nr of in_flight works */
208 int nr_active
; /* L: nr of active works */
209 int max_active
; /* L: max active works */
210 struct list_head delayed_works
; /* L: delayed works */
211 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
212 struct list_head mayday_node
; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work
;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
225 * Structure used to wait for workqueue flush.
228 struct list_head list
; /* WQ: list of flushers */
229 int flush_color
; /* WQ: flush color waiting for */
230 struct completion done
; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct
{
240 struct list_head pwqs
; /* WR: all pwqs of this wq */
241 struct list_head list
; /* PR: list of all workqueues */
243 struct mutex mutex
; /* protects this wq */
244 int work_color
; /* WQ: current work color */
245 int flush_color
; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush
; /* flush in progress */
247 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
248 struct list_head flusher_queue
; /* WQ: flush waiters */
249 struct list_head flusher_overflow
; /* WQ: flush overflow list */
251 struct list_head maydays
; /* MD: pwqs requesting rescue */
252 struct worker
*rescuer
; /* I: rescue worker */
254 int nr_drainers
; /* WQ: drain in progress */
255 int saved_max_active
; /* WQ: saved pwq max_active */
257 struct workqueue_attrs
*unbound_attrs
; /* PW: only for unbound wqs */
258 struct pool_workqueue
*dfl_pwq
; /* PW: only for unbound wqs */
261 struct wq_device
*wq_dev
; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map
;
266 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache
*pwq_cache
;
283 static cpumask_var_t
*wq_numa_possible_cpumask
;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa
;
287 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient
= IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT
);
291 module_param_named(power_efficient
, wq_power_efficient
, bool, 0444);
293 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
295 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
298 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
299 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
301 static LIST_HEAD(workqueues
); /* PR: list of all workqueues */
302 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
304 /* PL: allowable cpus for unbound wqs and work items */
305 static cpumask_var_t wq_unbound_cpumask
;
307 /* CPU where unbound work was last round robin scheduled from this CPU */
308 static DEFINE_PER_CPU(int, wq_rr_cpu_last
);
310 /* the per-cpu worker pools */
311 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
314 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
316 /* PL: hash of all unbound pools keyed by pool->attrs */
317 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
319 /* I: attributes used when instantiating standard unbound pools on demand */
320 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
322 /* I: attributes used when instantiating ordered pools on demand */
323 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
325 struct workqueue_struct
*system_wq __read_mostly
;
326 EXPORT_SYMBOL(system_wq
);
327 struct workqueue_struct
*system_highpri_wq __read_mostly
;
328 EXPORT_SYMBOL_GPL(system_highpri_wq
);
329 struct workqueue_struct
*system_long_wq __read_mostly
;
330 EXPORT_SYMBOL_GPL(system_long_wq
);
331 struct workqueue_struct
*system_unbound_wq __read_mostly
;
332 EXPORT_SYMBOL_GPL(system_unbound_wq
);
333 struct workqueue_struct
*system_freezable_wq __read_mostly
;
334 EXPORT_SYMBOL_GPL(system_freezable_wq
);
335 struct workqueue_struct
*system_power_efficient_wq __read_mostly
;
336 EXPORT_SYMBOL_GPL(system_power_efficient_wq
);
337 struct workqueue_struct
*system_freezable_power_efficient_wq __read_mostly
;
338 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq
);
340 static int worker_thread(void *__worker
);
341 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
);
343 #define CREATE_TRACE_POINTS
344 #include <trace/events/workqueue.h>
346 #define assert_rcu_or_pool_mutex() \
347 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
348 !lockdep_is_held(&wq_pool_mutex), \
349 "sched RCU or wq_pool_mutex should be held")
351 #define assert_rcu_or_wq_mutex(wq) \
352 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
353 !lockdep_is_held(&wq->mutex), \
354 "sched RCU or wq->mutex should be held")
356 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
357 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
358 !lockdep_is_held(&wq->mutex) && \
359 !lockdep_is_held(&wq_pool_mutex), \
360 "sched RCU, wq->mutex or wq_pool_mutex should be held")
362 #define for_each_cpu_worker_pool(pool, cpu) \
363 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
364 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
368 * for_each_pool - iterate through all worker_pools in the system
369 * @pool: iteration cursor
370 * @pi: integer used for iteration
372 * This must be called either with wq_pool_mutex held or sched RCU read
373 * locked. If the pool needs to be used beyond the locking in effect, the
374 * caller is responsible for guaranteeing that the pool stays online.
376 * The if/else clause exists only for the lockdep assertion and can be
379 #define for_each_pool(pool, pi) \
380 idr_for_each_entry(&worker_pool_idr, pool, pi) \
381 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
385 * for_each_pool_worker - iterate through all workers of a worker_pool
386 * @worker: iteration cursor
387 * @pool: worker_pool to iterate workers of
389 * This must be called with @pool->attach_mutex.
391 * The if/else clause exists only for the lockdep assertion and can be
394 #define for_each_pool_worker(worker, pool) \
395 list_for_each_entry((worker), &(pool)->workers, node) \
396 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
400 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
401 * @pwq: iteration cursor
402 * @wq: the target workqueue
404 * This must be called either with wq->mutex held or sched RCU read locked.
405 * If the pwq needs to be used beyond the locking in effect, the caller is
406 * responsible for guaranteeing that the pwq stays online.
408 * The if/else clause exists only for the lockdep assertion and can be
411 #define for_each_pwq(pwq, wq) \
412 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
413 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
416 #ifdef CONFIG_DEBUG_OBJECTS_WORK
418 static struct debug_obj_descr work_debug_descr
;
420 static void *work_debug_hint(void *addr
)
422 return ((struct work_struct
*) addr
)->func
;
426 * fixup_init is called when:
427 * - an active object is initialized
429 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
431 struct work_struct
*work
= addr
;
434 case ODEBUG_STATE_ACTIVE
:
435 cancel_work_sync(work
);
436 debug_object_init(work
, &work_debug_descr
);
444 * fixup_activate is called when:
445 * - an active object is activated
446 * - an unknown object is activated (might be a statically initialized object)
448 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
450 struct work_struct
*work
= addr
;
454 case ODEBUG_STATE_NOTAVAILABLE
:
456 * This is not really a fixup. The work struct was
457 * statically initialized. We just make sure that it
458 * is tracked in the object tracker.
460 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
461 debug_object_init(work
, &work_debug_descr
);
462 debug_object_activate(work
, &work_debug_descr
);
468 case ODEBUG_STATE_ACTIVE
:
477 * fixup_free is called when:
478 * - an active object is freed
480 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
482 struct work_struct
*work
= addr
;
485 case ODEBUG_STATE_ACTIVE
:
486 cancel_work_sync(work
);
487 debug_object_free(work
, &work_debug_descr
);
494 static struct debug_obj_descr work_debug_descr
= {
495 .name
= "work_struct",
496 .debug_hint
= work_debug_hint
,
497 .fixup_init
= work_fixup_init
,
498 .fixup_activate
= work_fixup_activate
,
499 .fixup_free
= work_fixup_free
,
502 static inline void debug_work_activate(struct work_struct
*work
)
504 debug_object_activate(work
, &work_debug_descr
);
507 static inline void debug_work_deactivate(struct work_struct
*work
)
509 debug_object_deactivate(work
, &work_debug_descr
);
512 void __init_work(struct work_struct
*work
, int onstack
)
515 debug_object_init_on_stack(work
, &work_debug_descr
);
517 debug_object_init(work
, &work_debug_descr
);
519 EXPORT_SYMBOL_GPL(__init_work
);
521 void destroy_work_on_stack(struct work_struct
*work
)
523 debug_object_free(work
, &work_debug_descr
);
525 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
527 void destroy_delayed_work_on_stack(struct delayed_work
*work
)
529 destroy_timer_on_stack(&work
->timer
);
530 debug_object_free(&work
->work
, &work_debug_descr
);
532 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack
);
535 static inline void debug_work_activate(struct work_struct
*work
) { }
536 static inline void debug_work_deactivate(struct work_struct
*work
) { }
540 * worker_pool_assign_id - allocate ID and assing it to @pool
541 * @pool: the pool pointer of interest
543 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
544 * successfully, -errno on failure.
546 static int worker_pool_assign_id(struct worker_pool
*pool
)
550 lockdep_assert_held(&wq_pool_mutex
);
552 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, WORK_OFFQ_POOL_NONE
,
562 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
563 * @wq: the target workqueue
566 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
568 * If the pwq needs to be used beyond the locking in effect, the caller is
569 * responsible for guaranteeing that the pwq stays online.
571 * Return: The unbound pool_workqueue for @node.
573 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
576 assert_rcu_or_wq_mutex_or_pool_mutex(wq
);
577 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
580 static unsigned int work_color_to_flags(int color
)
582 return color
<< WORK_STRUCT_COLOR_SHIFT
;
585 static int get_work_color(struct work_struct
*work
)
587 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
588 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
591 static int work_next_color(int color
)
593 return (color
+ 1) % WORK_NR_COLORS
;
597 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
598 * contain the pointer to the queued pwq. Once execution starts, the flag
599 * is cleared and the high bits contain OFFQ flags and pool ID.
601 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
602 * and clear_work_data() can be used to set the pwq, pool or clear
603 * work->data. These functions should only be called while the work is
604 * owned - ie. while the PENDING bit is set.
606 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
607 * corresponding to a work. Pool is available once the work has been
608 * queued anywhere after initialization until it is sync canceled. pwq is
609 * available only while the work item is queued.
611 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
612 * canceled. While being canceled, a work item may have its PENDING set
613 * but stay off timer and worklist for arbitrarily long and nobody should
614 * try to steal the PENDING bit.
616 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
619 WARN_ON_ONCE(!work_pending(work
));
620 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
623 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
624 unsigned long extra_flags
)
626 set_work_data(work
, (unsigned long)pwq
,
627 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
630 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
633 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
634 WORK_STRUCT_PENDING
);
637 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
641 * The following wmb is paired with the implied mb in
642 * test_and_set_bit(PENDING) and ensures all updates to @work made
643 * here are visible to and precede any updates by the next PENDING
647 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
650 static void clear_work_data(struct work_struct
*work
)
652 smp_wmb(); /* see set_work_pool_and_clear_pending() */
653 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
656 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
658 unsigned long data
= atomic_long_read(&work
->data
);
660 if (data
& WORK_STRUCT_PWQ
)
661 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
667 * get_work_pool - return the worker_pool a given work was associated with
668 * @work: the work item of interest
670 * Pools are created and destroyed under wq_pool_mutex, and allows read
671 * access under sched-RCU read lock. As such, this function should be
672 * called under wq_pool_mutex or with preemption disabled.
674 * All fields of the returned pool are accessible as long as the above
675 * mentioned locking is in effect. If the returned pool needs to be used
676 * beyond the critical section, the caller is responsible for ensuring the
677 * returned pool is and stays online.
679 * Return: The worker_pool @work was last associated with. %NULL if none.
681 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
683 unsigned long data
= atomic_long_read(&work
->data
);
686 assert_rcu_or_pool_mutex();
688 if (data
& WORK_STRUCT_PWQ
)
689 return ((struct pool_workqueue
*)
690 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
692 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
693 if (pool_id
== WORK_OFFQ_POOL_NONE
)
696 return idr_find(&worker_pool_idr
, pool_id
);
700 * get_work_pool_id - return the worker pool ID a given work is associated with
701 * @work: the work item of interest
703 * Return: The worker_pool ID @work was last associated with.
704 * %WORK_OFFQ_POOL_NONE if none.
706 static int get_work_pool_id(struct work_struct
*work
)
708 unsigned long data
= atomic_long_read(&work
->data
);
710 if (data
& WORK_STRUCT_PWQ
)
711 return ((struct pool_workqueue
*)
712 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
714 return data
>> WORK_OFFQ_POOL_SHIFT
;
717 static void mark_work_canceling(struct work_struct
*work
)
719 unsigned long pool_id
= get_work_pool_id(work
);
721 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
722 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
725 static bool work_is_canceling(struct work_struct
*work
)
727 unsigned long data
= atomic_long_read(&work
->data
);
729 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
733 * Policy functions. These define the policies on how the global worker
734 * pools are managed. Unless noted otherwise, these functions assume that
735 * they're being called with pool->lock held.
738 static bool __need_more_worker(struct worker_pool
*pool
)
740 return !atomic_read(&pool
->nr_running
);
744 * Need to wake up a worker? Called from anything but currently
747 * Note that, because unbound workers never contribute to nr_running, this
748 * function will always return %true for unbound pools as long as the
749 * worklist isn't empty.
751 static bool need_more_worker(struct worker_pool
*pool
)
753 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
756 /* Can I start working? Called from busy but !running workers. */
757 static bool may_start_working(struct worker_pool
*pool
)
759 return pool
->nr_idle
;
762 /* Do I need to keep working? Called from currently running workers. */
763 static bool keep_working(struct worker_pool
*pool
)
765 return !list_empty(&pool
->worklist
) &&
766 atomic_read(&pool
->nr_running
) <= 1;
769 /* Do we need a new worker? Called from manager. */
770 static bool need_to_create_worker(struct worker_pool
*pool
)
772 return need_more_worker(pool
) && !may_start_working(pool
);
775 /* Do we have too many workers and should some go away? */
776 static bool too_many_workers(struct worker_pool
*pool
)
778 bool managing
= mutex_is_locked(&pool
->manager_arb
);
779 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
780 int nr_busy
= pool
->nr_workers
- nr_idle
;
782 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
789 /* Return the first idle worker. Safe with preemption disabled */
790 static struct worker
*first_idle_worker(struct worker_pool
*pool
)
792 if (unlikely(list_empty(&pool
->idle_list
)))
795 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
799 * wake_up_worker - wake up an idle worker
800 * @pool: worker pool to wake worker from
802 * Wake up the first idle worker of @pool.
805 * spin_lock_irq(pool->lock).
807 static void wake_up_worker(struct worker_pool
*pool
)
809 struct worker
*worker
= first_idle_worker(pool
);
812 wake_up_process(worker
->task
);
816 * wq_worker_waking_up - a worker is waking up
817 * @task: task waking up
818 * @cpu: CPU @task is waking up to
820 * This function is called during try_to_wake_up() when a worker is
824 * spin_lock_irq(rq->lock)
826 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
828 struct worker
*worker
= kthread_data(task
);
830 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
831 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
832 atomic_inc(&worker
->pool
->nr_running
);
837 * wq_worker_sleeping - a worker is going to sleep
838 * @task: task going to sleep
839 * @cpu: CPU in question, must be the current CPU number
841 * This function is called during schedule() when a busy worker is
842 * going to sleep. Worker on the same cpu can be woken up by
843 * returning pointer to its task.
846 * spin_lock_irq(rq->lock)
849 * Worker task on @cpu to wake up, %NULL if none.
851 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
853 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
854 struct worker_pool
*pool
;
857 * Rescuers, which may not have all the fields set up like normal
858 * workers, also reach here, let's not access anything before
859 * checking NOT_RUNNING.
861 if (worker
->flags
& WORKER_NOT_RUNNING
)
866 /* this can only happen on the local cpu */
867 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id() || pool
->cpu
!= cpu
))
871 * The counterpart of the following dec_and_test, implied mb,
872 * worklist not empty test sequence is in insert_work().
873 * Please read comment there.
875 * NOT_RUNNING is clear. This means that we're bound to and
876 * running on the local cpu w/ rq lock held and preemption
877 * disabled, which in turn means that none else could be
878 * manipulating idle_list, so dereferencing idle_list without pool
881 if (atomic_dec_and_test(&pool
->nr_running
) &&
882 !list_empty(&pool
->worklist
))
883 to_wakeup
= first_idle_worker(pool
);
884 return to_wakeup
? to_wakeup
->task
: NULL
;
888 * worker_set_flags - set worker flags and adjust nr_running accordingly
890 * @flags: flags to set
892 * Set @flags in @worker->flags and adjust nr_running accordingly.
895 * spin_lock_irq(pool->lock)
897 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
)
899 struct worker_pool
*pool
= worker
->pool
;
901 WARN_ON_ONCE(worker
->task
!= current
);
903 /* If transitioning into NOT_RUNNING, adjust nr_running. */
904 if ((flags
& WORKER_NOT_RUNNING
) &&
905 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
906 atomic_dec(&pool
->nr_running
);
909 worker
->flags
|= flags
;
913 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
915 * @flags: flags to clear
917 * Clear @flags in @worker->flags and adjust nr_running accordingly.
920 * spin_lock_irq(pool->lock)
922 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
924 struct worker_pool
*pool
= worker
->pool
;
925 unsigned int oflags
= worker
->flags
;
927 WARN_ON_ONCE(worker
->task
!= current
);
929 worker
->flags
&= ~flags
;
932 * If transitioning out of NOT_RUNNING, increment nr_running. Note
933 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
934 * of multiple flags, not a single flag.
936 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
937 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
938 atomic_inc(&pool
->nr_running
);
942 * find_worker_executing_work - find worker which is executing a work
943 * @pool: pool of interest
944 * @work: work to find worker for
946 * Find a worker which is executing @work on @pool by searching
947 * @pool->busy_hash which is keyed by the address of @work. For a worker
948 * to match, its current execution should match the address of @work and
949 * its work function. This is to avoid unwanted dependency between
950 * unrelated work executions through a work item being recycled while still
953 * This is a bit tricky. A work item may be freed once its execution
954 * starts and nothing prevents the freed area from being recycled for
955 * another work item. If the same work item address ends up being reused
956 * before the original execution finishes, workqueue will identify the
957 * recycled work item as currently executing and make it wait until the
958 * current execution finishes, introducing an unwanted dependency.
960 * This function checks the work item address and work function to avoid
961 * false positives. Note that this isn't complete as one may construct a
962 * work function which can introduce dependency onto itself through a
963 * recycled work item. Well, if somebody wants to shoot oneself in the
964 * foot that badly, there's only so much we can do, and if such deadlock
965 * actually occurs, it should be easy to locate the culprit work function.
968 * spin_lock_irq(pool->lock).
971 * Pointer to worker which is executing @work if found, %NULL
974 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
975 struct work_struct
*work
)
977 struct worker
*worker
;
979 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
981 if (worker
->current_work
== work
&&
982 worker
->current_func
== work
->func
)
989 * move_linked_works - move linked works to a list
990 * @work: start of series of works to be scheduled
991 * @head: target list to append @work to
992 * @nextp: out parameter for nested worklist walking
994 * Schedule linked works starting from @work to @head. Work series to
995 * be scheduled starts at @work and includes any consecutive work with
996 * WORK_STRUCT_LINKED set in its predecessor.
998 * If @nextp is not NULL, it's updated to point to the next work of
999 * the last scheduled work. This allows move_linked_works() to be
1000 * nested inside outer list_for_each_entry_safe().
1003 * spin_lock_irq(pool->lock).
1005 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1006 struct work_struct
**nextp
)
1008 struct work_struct
*n
;
1011 * Linked worklist will always end before the end of the list,
1012 * use NULL for list head.
1014 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1015 list_move_tail(&work
->entry
, head
);
1016 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1021 * If we're already inside safe list traversal and have moved
1022 * multiple works to the scheduled queue, the next position
1023 * needs to be updated.
1030 * get_pwq - get an extra reference on the specified pool_workqueue
1031 * @pwq: pool_workqueue to get
1033 * Obtain an extra reference on @pwq. The caller should guarantee that
1034 * @pwq has positive refcnt and be holding the matching pool->lock.
1036 static void get_pwq(struct pool_workqueue
*pwq
)
1038 lockdep_assert_held(&pwq
->pool
->lock
);
1039 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1044 * put_pwq - put a pool_workqueue reference
1045 * @pwq: pool_workqueue to put
1047 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1048 * destruction. The caller should be holding the matching pool->lock.
1050 static void put_pwq(struct pool_workqueue
*pwq
)
1052 lockdep_assert_held(&pwq
->pool
->lock
);
1053 if (likely(--pwq
->refcnt
))
1055 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1058 * @pwq can't be released under pool->lock, bounce to
1059 * pwq_unbound_release_workfn(). This never recurses on the same
1060 * pool->lock as this path is taken only for unbound workqueues and
1061 * the release work item is scheduled on a per-cpu workqueue. To
1062 * avoid lockdep warning, unbound pool->locks are given lockdep
1063 * subclass of 1 in get_unbound_pool().
1065 schedule_work(&pwq
->unbound_release_work
);
1069 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1070 * @pwq: pool_workqueue to put (can be %NULL)
1072 * put_pwq() with locking. This function also allows %NULL @pwq.
1074 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1078 * As both pwqs and pools are sched-RCU protected, the
1079 * following lock operations are safe.
1081 spin_lock_irq(&pwq
->pool
->lock
);
1083 spin_unlock_irq(&pwq
->pool
->lock
);
1087 static void pwq_activate_delayed_work(struct work_struct
*work
)
1089 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1091 trace_workqueue_activate_work(work
);
1092 if (list_empty(&pwq
->pool
->worklist
))
1093 pwq
->pool
->watchdog_ts
= jiffies
;
1094 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1095 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1099 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1101 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1102 struct work_struct
, entry
);
1104 pwq_activate_delayed_work(work
);
1108 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1109 * @pwq: pwq of interest
1110 * @color: color of work which left the queue
1112 * A work either has completed or is removed from pending queue,
1113 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1116 * spin_lock_irq(pool->lock).
1118 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1120 /* uncolored work items don't participate in flushing or nr_active */
1121 if (color
== WORK_NO_COLOR
)
1124 pwq
->nr_in_flight
[color
]--;
1127 if (!list_empty(&pwq
->delayed_works
)) {
1128 /* one down, submit a delayed one */
1129 if (pwq
->nr_active
< pwq
->max_active
)
1130 pwq_activate_first_delayed(pwq
);
1133 /* is flush in progress and are we at the flushing tip? */
1134 if (likely(pwq
->flush_color
!= color
))
1137 /* are there still in-flight works? */
1138 if (pwq
->nr_in_flight
[color
])
1141 /* this pwq is done, clear flush_color */
1142 pwq
->flush_color
= -1;
1145 * If this was the last pwq, wake up the first flusher. It
1146 * will handle the rest.
1148 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1149 complete(&pwq
->wq
->first_flusher
->done
);
1155 * try_to_grab_pending - steal work item from worklist and disable irq
1156 * @work: work item to steal
1157 * @is_dwork: @work is a delayed_work
1158 * @flags: place to store irq state
1160 * Try to grab PENDING bit of @work. This function can handle @work in any
1161 * stable state - idle, on timer or on worklist.
1164 * 1 if @work was pending and we successfully stole PENDING
1165 * 0 if @work was idle and we claimed PENDING
1166 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1167 * -ENOENT if someone else is canceling @work, this state may persist
1168 * for arbitrarily long
1171 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1172 * interrupted while holding PENDING and @work off queue, irq must be
1173 * disabled on entry. This, combined with delayed_work->timer being
1174 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1176 * On successful return, >= 0, irq is disabled and the caller is
1177 * responsible for releasing it using local_irq_restore(*@flags).
1179 * This function is safe to call from any context including IRQ handler.
1181 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1182 unsigned long *flags
)
1184 struct worker_pool
*pool
;
1185 struct pool_workqueue
*pwq
;
1187 local_irq_save(*flags
);
1189 /* try to steal the timer if it exists */
1191 struct delayed_work
*dwork
= to_delayed_work(work
);
1194 * dwork->timer is irqsafe. If del_timer() fails, it's
1195 * guaranteed that the timer is not queued anywhere and not
1196 * running on the local CPU.
1198 if (likely(del_timer(&dwork
->timer
)))
1202 /* try to claim PENDING the normal way */
1203 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1207 * The queueing is in progress, or it is already queued. Try to
1208 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1210 pool
= get_work_pool(work
);
1214 spin_lock(&pool
->lock
);
1216 * work->data is guaranteed to point to pwq only while the work
1217 * item is queued on pwq->wq, and both updating work->data to point
1218 * to pwq on queueing and to pool on dequeueing are done under
1219 * pwq->pool->lock. This in turn guarantees that, if work->data
1220 * points to pwq which is associated with a locked pool, the work
1221 * item is currently queued on that pool.
1223 pwq
= get_work_pwq(work
);
1224 if (pwq
&& pwq
->pool
== pool
) {
1225 debug_work_deactivate(work
);
1228 * A delayed work item cannot be grabbed directly because
1229 * it might have linked NO_COLOR work items which, if left
1230 * on the delayed_list, will confuse pwq->nr_active
1231 * management later on and cause stall. Make sure the work
1232 * item is activated before grabbing.
1234 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1235 pwq_activate_delayed_work(work
);
1237 list_del_init(&work
->entry
);
1238 pwq_dec_nr_in_flight(pwq
, get_work_color(work
));
1240 /* work->data points to pwq iff queued, point to pool */
1241 set_work_pool_and_keep_pending(work
, pool
->id
);
1243 spin_unlock(&pool
->lock
);
1246 spin_unlock(&pool
->lock
);
1248 local_irq_restore(*flags
);
1249 if (work_is_canceling(work
))
1256 * insert_work - insert a work into a pool
1257 * @pwq: pwq @work belongs to
1258 * @work: work to insert
1259 * @head: insertion point
1260 * @extra_flags: extra WORK_STRUCT_* flags to set
1262 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1263 * work_struct flags.
1266 * spin_lock_irq(pool->lock).
1268 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1269 struct list_head
*head
, unsigned int extra_flags
)
1271 struct worker_pool
*pool
= pwq
->pool
;
1273 /* we own @work, set data and link */
1274 set_work_pwq(work
, pwq
, extra_flags
);
1275 list_add_tail(&work
->entry
, head
);
1279 * Ensure either wq_worker_sleeping() sees the above
1280 * list_add_tail() or we see zero nr_running to avoid workers lying
1281 * around lazily while there are works to be processed.
1285 if (__need_more_worker(pool
))
1286 wake_up_worker(pool
);
1290 * Test whether @work is being queued from another work executing on the
1293 static bool is_chained_work(struct workqueue_struct
*wq
)
1295 struct worker
*worker
;
1297 worker
= current_wq_worker();
1299 * Return %true iff I'm a worker execuing a work item on @wq. If
1300 * I'm @worker, it's safe to dereference it without locking.
1302 return worker
&& worker
->current_pwq
->wq
== wq
;
1306 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1307 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1308 * avoid perturbing sensitive tasks.
1310 static int wq_select_unbound_cpu(int cpu
)
1314 if (cpumask_test_cpu(cpu
, wq_unbound_cpumask
))
1316 if (cpumask_empty(wq_unbound_cpumask
))
1319 new_cpu
= __this_cpu_read(wq_rr_cpu_last
);
1320 new_cpu
= cpumask_next_and(new_cpu
, wq_unbound_cpumask
, cpu_online_mask
);
1321 if (unlikely(new_cpu
>= nr_cpu_ids
)) {
1322 new_cpu
= cpumask_first_and(wq_unbound_cpumask
, cpu_online_mask
);
1323 if (unlikely(new_cpu
>= nr_cpu_ids
))
1326 __this_cpu_write(wq_rr_cpu_last
, new_cpu
);
1331 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1332 struct work_struct
*work
)
1334 struct pool_workqueue
*pwq
;
1335 struct worker_pool
*last_pool
;
1336 struct list_head
*worklist
;
1337 unsigned int work_flags
;
1338 unsigned int req_cpu
= cpu
;
1341 * While a work item is PENDING && off queue, a task trying to
1342 * steal the PENDING will busy-loop waiting for it to either get
1343 * queued or lose PENDING. Grabbing PENDING and queueing should
1344 * happen with IRQ disabled.
1346 WARN_ON_ONCE(!irqs_disabled());
1348 debug_work_activate(work
);
1350 /* if draining, only works from the same workqueue are allowed */
1351 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1352 WARN_ON_ONCE(!is_chained_work(wq
)))
1355 if (req_cpu
== WORK_CPU_UNBOUND
)
1356 cpu
= wq_select_unbound_cpu(raw_smp_processor_id());
1358 /* pwq which will be used unless @work is executing elsewhere */
1359 if (!(wq
->flags
& WQ_UNBOUND
))
1360 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1362 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1365 * If @work was previously on a different pool, it might still be
1366 * running there, in which case the work needs to be queued on that
1367 * pool to guarantee non-reentrancy.
1369 last_pool
= get_work_pool(work
);
1370 if (last_pool
&& last_pool
!= pwq
->pool
) {
1371 struct worker
*worker
;
1373 spin_lock(&last_pool
->lock
);
1375 worker
= find_worker_executing_work(last_pool
, work
);
1377 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1378 pwq
= worker
->current_pwq
;
1380 /* meh... not running there, queue here */
1381 spin_unlock(&last_pool
->lock
);
1382 spin_lock(&pwq
->pool
->lock
);
1385 spin_lock(&pwq
->pool
->lock
);
1389 * pwq is determined and locked. For unbound pools, we could have
1390 * raced with pwq release and it could already be dead. If its
1391 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1392 * without another pwq replacing it in the numa_pwq_tbl or while
1393 * work items are executing on it, so the retrying is guaranteed to
1394 * make forward-progress.
1396 if (unlikely(!pwq
->refcnt
)) {
1397 if (wq
->flags
& WQ_UNBOUND
) {
1398 spin_unlock(&pwq
->pool
->lock
);
1403 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1407 /* pwq determined, queue */
1408 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1410 if (WARN_ON(!list_empty(&work
->entry
))) {
1411 spin_unlock(&pwq
->pool
->lock
);
1415 pwq
->nr_in_flight
[pwq
->work_color
]++;
1416 work_flags
= work_color_to_flags(pwq
->work_color
);
1418 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1419 trace_workqueue_activate_work(work
);
1421 worklist
= &pwq
->pool
->worklist
;
1422 if (list_empty(worklist
))
1423 pwq
->pool
->watchdog_ts
= jiffies
;
1425 work_flags
|= WORK_STRUCT_DELAYED
;
1426 worklist
= &pwq
->delayed_works
;
1429 insert_work(pwq
, work
, worklist
, work_flags
);
1431 spin_unlock(&pwq
->pool
->lock
);
1435 * queue_work_on - queue work on specific cpu
1436 * @cpu: CPU number to execute work on
1437 * @wq: workqueue to use
1438 * @work: work to queue
1440 * We queue the work to a specific CPU, the caller must ensure it
1443 * Return: %false if @work was already on a queue, %true otherwise.
1445 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1446 struct work_struct
*work
)
1449 unsigned long flags
;
1451 local_irq_save(flags
);
1453 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1454 __queue_work(cpu
, wq
, work
);
1458 local_irq_restore(flags
);
1461 EXPORT_SYMBOL(queue_work_on
);
1463 void delayed_work_timer_fn(unsigned long __data
)
1465 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1467 /* should have been called from irqsafe timer with irq already off */
1468 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1470 EXPORT_SYMBOL(delayed_work_timer_fn
);
1472 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1473 struct delayed_work
*dwork
, unsigned long delay
)
1475 struct timer_list
*timer
= &dwork
->timer
;
1476 struct work_struct
*work
= &dwork
->work
;
1478 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1479 timer
->data
!= (unsigned long)dwork
);
1480 WARN_ON_ONCE(timer_pending(timer
));
1481 WARN_ON_ONCE(!list_empty(&work
->entry
));
1484 * If @delay is 0, queue @dwork->work immediately. This is for
1485 * both optimization and correctness. The earliest @timer can
1486 * expire is on the closest next tick and delayed_work users depend
1487 * on that there's no such delay when @delay is 0.
1490 __queue_work(cpu
, wq
, &dwork
->work
);
1494 timer_stats_timer_set_start_info(&dwork
->timer
);
1498 timer
->expires
= jiffies
+ delay
;
1500 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1501 add_timer_on(timer
, cpu
);
1507 * queue_delayed_work_on - queue work on specific CPU after delay
1508 * @cpu: CPU number to execute work on
1509 * @wq: workqueue to use
1510 * @dwork: work to queue
1511 * @delay: number of jiffies to wait before queueing
1513 * Return: %false if @work was already on a queue, %true otherwise. If
1514 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1517 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1518 struct delayed_work
*dwork
, unsigned long delay
)
1520 struct work_struct
*work
= &dwork
->work
;
1522 unsigned long flags
;
1524 /* read the comment in __queue_work() */
1525 local_irq_save(flags
);
1527 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1528 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1532 local_irq_restore(flags
);
1535 EXPORT_SYMBOL(queue_delayed_work_on
);
1538 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1539 * @cpu: CPU number to execute work on
1540 * @wq: workqueue to use
1541 * @dwork: work to queue
1542 * @delay: number of jiffies to wait before queueing
1544 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1545 * modify @dwork's timer so that it expires after @delay. If @delay is
1546 * zero, @work is guaranteed to be scheduled immediately regardless of its
1549 * Return: %false if @dwork was idle and queued, %true if @dwork was
1550 * pending and its timer was modified.
1552 * This function is safe to call from any context including IRQ handler.
1553 * See try_to_grab_pending() for details.
1555 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1556 struct delayed_work
*dwork
, unsigned long delay
)
1558 unsigned long flags
;
1562 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1563 } while (unlikely(ret
== -EAGAIN
));
1565 if (likely(ret
>= 0)) {
1566 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1567 local_irq_restore(flags
);
1570 /* -ENOENT from try_to_grab_pending() becomes %true */
1573 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1576 * worker_enter_idle - enter idle state
1577 * @worker: worker which is entering idle state
1579 * @worker is entering idle state. Update stats and idle timer if
1583 * spin_lock_irq(pool->lock).
1585 static void worker_enter_idle(struct worker
*worker
)
1587 struct worker_pool
*pool
= worker
->pool
;
1589 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1590 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1591 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1594 /* can't use worker_set_flags(), also called from create_worker() */
1595 worker
->flags
|= WORKER_IDLE
;
1597 worker
->last_active
= jiffies
;
1599 /* idle_list is LIFO */
1600 list_add(&worker
->entry
, &pool
->idle_list
);
1602 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1603 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1606 * Sanity check nr_running. Because wq_unbind_fn() releases
1607 * pool->lock between setting %WORKER_UNBOUND and zapping
1608 * nr_running, the warning may trigger spuriously. Check iff
1609 * unbind is not in progress.
1611 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1612 pool
->nr_workers
== pool
->nr_idle
&&
1613 atomic_read(&pool
->nr_running
));
1617 * worker_leave_idle - leave idle state
1618 * @worker: worker which is leaving idle state
1620 * @worker is leaving idle state. Update stats.
1623 * spin_lock_irq(pool->lock).
1625 static void worker_leave_idle(struct worker
*worker
)
1627 struct worker_pool
*pool
= worker
->pool
;
1629 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1631 worker_clr_flags(worker
, WORKER_IDLE
);
1633 list_del_init(&worker
->entry
);
1636 static struct worker
*alloc_worker(int node
)
1638 struct worker
*worker
;
1640 worker
= kzalloc_node(sizeof(*worker
), GFP_KERNEL
, node
);
1642 INIT_LIST_HEAD(&worker
->entry
);
1643 INIT_LIST_HEAD(&worker
->scheduled
);
1644 INIT_LIST_HEAD(&worker
->node
);
1645 /* on creation a worker is in !idle && prep state */
1646 worker
->flags
= WORKER_PREP
;
1652 * worker_attach_to_pool() - attach a worker to a pool
1653 * @worker: worker to be attached
1654 * @pool: the target pool
1656 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1657 * cpu-binding of @worker are kept coordinated with the pool across
1660 static void worker_attach_to_pool(struct worker
*worker
,
1661 struct worker_pool
*pool
)
1663 mutex_lock(&pool
->attach_mutex
);
1666 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1667 * online CPUs. It'll be re-applied when any of the CPUs come up.
1669 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1672 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1673 * stable across this function. See the comments above the
1674 * flag definition for details.
1676 if (pool
->flags
& POOL_DISASSOCIATED
)
1677 worker
->flags
|= WORKER_UNBOUND
;
1679 list_add_tail(&worker
->node
, &pool
->workers
);
1681 mutex_unlock(&pool
->attach_mutex
);
1685 * worker_detach_from_pool() - detach a worker from its pool
1686 * @worker: worker which is attached to its pool
1687 * @pool: the pool @worker is attached to
1689 * Undo the attaching which had been done in worker_attach_to_pool(). The
1690 * caller worker shouldn't access to the pool after detached except it has
1691 * other reference to the pool.
1693 static void worker_detach_from_pool(struct worker
*worker
,
1694 struct worker_pool
*pool
)
1696 struct completion
*detach_completion
= NULL
;
1698 mutex_lock(&pool
->attach_mutex
);
1699 list_del(&worker
->node
);
1700 if (list_empty(&pool
->workers
))
1701 detach_completion
= pool
->detach_completion
;
1702 mutex_unlock(&pool
->attach_mutex
);
1704 /* clear leftover flags without pool->lock after it is detached */
1705 worker
->flags
&= ~(WORKER_UNBOUND
| WORKER_REBOUND
);
1707 if (detach_completion
)
1708 complete(detach_completion
);
1712 * create_worker - create a new workqueue worker
1713 * @pool: pool the new worker will belong to
1715 * Create and start a new worker which is attached to @pool.
1718 * Might sleep. Does GFP_KERNEL allocations.
1721 * Pointer to the newly created worker.
1723 static struct worker
*create_worker(struct worker_pool
*pool
)
1725 struct worker
*worker
= NULL
;
1729 /* ID is needed to determine kthread name */
1730 id
= ida_simple_get(&pool
->worker_ida
, 0, 0, GFP_KERNEL
);
1734 worker
= alloc_worker(pool
->node
);
1738 worker
->pool
= pool
;
1742 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1743 pool
->attrs
->nice
< 0 ? "H" : "");
1745 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1747 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1748 "kworker/%s", id_buf
);
1749 if (IS_ERR(worker
->task
))
1752 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1753 kthread_bind_mask(worker
->task
, pool
->attrs
->cpumask
);
1755 /* successful, attach the worker to the pool */
1756 worker_attach_to_pool(worker
, pool
);
1758 /* start the newly created worker */
1759 spin_lock_irq(&pool
->lock
);
1760 worker
->pool
->nr_workers
++;
1761 worker_enter_idle(worker
);
1762 wake_up_process(worker
->task
);
1763 spin_unlock_irq(&pool
->lock
);
1769 ida_simple_remove(&pool
->worker_ida
, id
);
1775 * destroy_worker - destroy a workqueue worker
1776 * @worker: worker to be destroyed
1778 * Destroy @worker and adjust @pool stats accordingly. The worker should
1782 * spin_lock_irq(pool->lock).
1784 static void destroy_worker(struct worker
*worker
)
1786 struct worker_pool
*pool
= worker
->pool
;
1788 lockdep_assert_held(&pool
->lock
);
1790 /* sanity check frenzy */
1791 if (WARN_ON(worker
->current_work
) ||
1792 WARN_ON(!list_empty(&worker
->scheduled
)) ||
1793 WARN_ON(!(worker
->flags
& WORKER_IDLE
)))
1799 list_del_init(&worker
->entry
);
1800 worker
->flags
|= WORKER_DIE
;
1801 wake_up_process(worker
->task
);
1804 static void idle_worker_timeout(unsigned long __pool
)
1806 struct worker_pool
*pool
= (void *)__pool
;
1808 spin_lock_irq(&pool
->lock
);
1810 while (too_many_workers(pool
)) {
1811 struct worker
*worker
;
1812 unsigned long expires
;
1814 /* idle_list is kept in LIFO order, check the last one */
1815 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1816 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1818 if (time_before(jiffies
, expires
)) {
1819 mod_timer(&pool
->idle_timer
, expires
);
1823 destroy_worker(worker
);
1826 spin_unlock_irq(&pool
->lock
);
1829 static void send_mayday(struct work_struct
*work
)
1831 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1832 struct workqueue_struct
*wq
= pwq
->wq
;
1834 lockdep_assert_held(&wq_mayday_lock
);
1839 /* mayday mayday mayday */
1840 if (list_empty(&pwq
->mayday_node
)) {
1842 * If @pwq is for an unbound wq, its base ref may be put at
1843 * any time due to an attribute change. Pin @pwq until the
1844 * rescuer is done with it.
1847 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1848 wake_up_process(wq
->rescuer
->task
);
1852 static void pool_mayday_timeout(unsigned long __pool
)
1854 struct worker_pool
*pool
= (void *)__pool
;
1855 struct work_struct
*work
;
1857 spin_lock_irq(&pool
->lock
);
1858 spin_lock(&wq_mayday_lock
); /* for wq->maydays */
1860 if (need_to_create_worker(pool
)) {
1862 * We've been trying to create a new worker but
1863 * haven't been successful. We might be hitting an
1864 * allocation deadlock. Send distress signals to
1867 list_for_each_entry(work
, &pool
->worklist
, entry
)
1871 spin_unlock(&wq_mayday_lock
);
1872 spin_unlock_irq(&pool
->lock
);
1874 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1878 * maybe_create_worker - create a new worker if necessary
1879 * @pool: pool to create a new worker for
1881 * Create a new worker for @pool if necessary. @pool is guaranteed to
1882 * have at least one idle worker on return from this function. If
1883 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1884 * sent to all rescuers with works scheduled on @pool to resolve
1885 * possible allocation deadlock.
1887 * On return, need_to_create_worker() is guaranteed to be %false and
1888 * may_start_working() %true.
1891 * spin_lock_irq(pool->lock) which may be released and regrabbed
1892 * multiple times. Does GFP_KERNEL allocations. Called only from
1895 static void maybe_create_worker(struct worker_pool
*pool
)
1896 __releases(&pool
->lock
)
1897 __acquires(&pool
->lock
)
1900 spin_unlock_irq(&pool
->lock
);
1902 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1903 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1906 if (create_worker(pool
) || !need_to_create_worker(pool
))
1909 schedule_timeout_interruptible(CREATE_COOLDOWN
);
1911 if (!need_to_create_worker(pool
))
1915 del_timer_sync(&pool
->mayday_timer
);
1916 spin_lock_irq(&pool
->lock
);
1918 * This is necessary even after a new worker was just successfully
1919 * created as @pool->lock was dropped and the new worker might have
1920 * already become busy.
1922 if (need_to_create_worker(pool
))
1927 * manage_workers - manage worker pool
1930 * Assume the manager role and manage the worker pool @worker belongs
1931 * to. At any given time, there can be only zero or one manager per
1932 * pool. The exclusion is handled automatically by this function.
1934 * The caller can safely start processing works on false return. On
1935 * true return, it's guaranteed that need_to_create_worker() is false
1936 * and may_start_working() is true.
1939 * spin_lock_irq(pool->lock) which may be released and regrabbed
1940 * multiple times. Does GFP_KERNEL allocations.
1943 * %false if the pool doesn't need management and the caller can safely
1944 * start processing works, %true if management function was performed and
1945 * the conditions that the caller verified before calling the function may
1946 * no longer be true.
1948 static bool manage_workers(struct worker
*worker
)
1950 struct worker_pool
*pool
= worker
->pool
;
1953 * Anyone who successfully grabs manager_arb wins the arbitration
1954 * and becomes the manager. mutex_trylock() on pool->manager_arb
1955 * failure while holding pool->lock reliably indicates that someone
1956 * else is managing the pool and the worker which failed trylock
1957 * can proceed to executing work items. This means that anyone
1958 * grabbing manager_arb is responsible for actually performing
1959 * manager duties. If manager_arb is grabbed and released without
1960 * actual management, the pool may stall indefinitely.
1962 if (!mutex_trylock(&pool
->manager_arb
))
1964 pool
->manager
= worker
;
1966 maybe_create_worker(pool
);
1968 pool
->manager
= NULL
;
1969 mutex_unlock(&pool
->manager_arb
);
1974 * process_one_work - process single work
1976 * @work: work to process
1978 * Process @work. This function contains all the logics necessary to
1979 * process a single work including synchronization against and
1980 * interaction with other workers on the same cpu, queueing and
1981 * flushing. As long as context requirement is met, any worker can
1982 * call this function to process a work.
1985 * spin_lock_irq(pool->lock) which is released and regrabbed.
1987 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1988 __releases(&pool
->lock
)
1989 __acquires(&pool
->lock
)
1991 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1992 struct worker_pool
*pool
= worker
->pool
;
1993 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1995 struct worker
*collision
;
1996 #ifdef CONFIG_LOCKDEP
1998 * It is permissible to free the struct work_struct from
1999 * inside the function that is called from it, this we need to
2000 * take into account for lockdep too. To avoid bogus "held
2001 * lock freed" warnings as well as problems when looking into
2002 * work->lockdep_map, make a copy and use that here.
2004 struct lockdep_map lockdep_map
;
2006 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2008 /* ensure we're on the correct CPU */
2009 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
2010 raw_smp_processor_id() != pool
->cpu
);
2013 * A single work shouldn't be executed concurrently by
2014 * multiple workers on a single cpu. Check whether anyone is
2015 * already processing the work. If so, defer the work to the
2016 * currently executing one.
2018 collision
= find_worker_executing_work(pool
, work
);
2019 if (unlikely(collision
)) {
2020 move_linked_works(work
, &collision
->scheduled
, NULL
);
2024 /* claim and dequeue */
2025 debug_work_deactivate(work
);
2026 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2027 worker
->current_work
= work
;
2028 worker
->current_func
= work
->func
;
2029 worker
->current_pwq
= pwq
;
2030 work_color
= get_work_color(work
);
2032 list_del_init(&work
->entry
);
2035 * CPU intensive works don't participate in concurrency management.
2036 * They're the scheduler's responsibility. This takes @worker out
2037 * of concurrency management and the next code block will chain
2038 * execution of the pending work items.
2040 if (unlikely(cpu_intensive
))
2041 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
);
2044 * Wake up another worker if necessary. The condition is always
2045 * false for normal per-cpu workers since nr_running would always
2046 * be >= 1 at this point. This is used to chain execution of the
2047 * pending work items for WORKER_NOT_RUNNING workers such as the
2048 * UNBOUND and CPU_INTENSIVE ones.
2050 if (need_more_worker(pool
))
2051 wake_up_worker(pool
);
2054 * Record the last pool and clear PENDING which should be the last
2055 * update to @work. Also, do this inside @pool->lock so that
2056 * PENDING and queued state changes happen together while IRQ is
2059 set_work_pool_and_clear_pending(work
, pool
->id
);
2061 spin_unlock_irq(&pool
->lock
);
2063 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2064 lock_map_acquire(&lockdep_map
);
2065 trace_workqueue_execute_start(work
);
2066 worker
->current_func(work
);
2068 * While we must be careful to not use "work" after this, the trace
2069 * point will only record its address.
2071 trace_workqueue_execute_end(work
);
2072 lock_map_release(&lockdep_map
);
2073 lock_map_release(&pwq
->wq
->lockdep_map
);
2075 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2076 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2077 " last function: %pf\n",
2078 current
->comm
, preempt_count(), task_pid_nr(current
),
2079 worker
->current_func
);
2080 debug_show_held_locks(current
);
2085 * The following prevents a kworker from hogging CPU on !PREEMPT
2086 * kernels, where a requeueing work item waiting for something to
2087 * happen could deadlock with stop_machine as such work item could
2088 * indefinitely requeue itself while all other CPUs are trapped in
2089 * stop_machine. At the same time, report a quiescent RCU state so
2090 * the same condition doesn't freeze RCU.
2092 cond_resched_rcu_qs();
2094 spin_lock_irq(&pool
->lock
);
2096 /* clear cpu intensive status */
2097 if (unlikely(cpu_intensive
))
2098 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2100 /* we're done with it, release */
2101 hash_del(&worker
->hentry
);
2102 worker
->current_work
= NULL
;
2103 worker
->current_func
= NULL
;
2104 worker
->current_pwq
= NULL
;
2105 worker
->desc_valid
= false;
2106 pwq_dec_nr_in_flight(pwq
, work_color
);
2110 * process_scheduled_works - process scheduled works
2113 * Process all scheduled works. Please note that the scheduled list
2114 * may change while processing a work, so this function repeatedly
2115 * fetches a work from the top and executes it.
2118 * spin_lock_irq(pool->lock) which may be released and regrabbed
2121 static void process_scheduled_works(struct worker
*worker
)
2123 while (!list_empty(&worker
->scheduled
)) {
2124 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2125 struct work_struct
, entry
);
2126 process_one_work(worker
, work
);
2131 * worker_thread - the worker thread function
2134 * The worker thread function. All workers belong to a worker_pool -
2135 * either a per-cpu one or dynamic unbound one. These workers process all
2136 * work items regardless of their specific target workqueue. The only
2137 * exception is work items which belong to workqueues with a rescuer which
2138 * will be explained in rescuer_thread().
2142 static int worker_thread(void *__worker
)
2144 struct worker
*worker
= __worker
;
2145 struct worker_pool
*pool
= worker
->pool
;
2147 /* tell the scheduler that this is a workqueue worker */
2148 worker
->task
->flags
|= PF_WQ_WORKER
;
2150 spin_lock_irq(&pool
->lock
);
2152 /* am I supposed to die? */
2153 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2154 spin_unlock_irq(&pool
->lock
);
2155 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2156 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2158 set_task_comm(worker
->task
, "kworker/dying");
2159 ida_simple_remove(&pool
->worker_ida
, worker
->id
);
2160 worker_detach_from_pool(worker
, pool
);
2165 worker_leave_idle(worker
);
2167 /* no more worker necessary? */
2168 if (!need_more_worker(pool
))
2171 /* do we need to manage? */
2172 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2176 * ->scheduled list can only be filled while a worker is
2177 * preparing to process a work or actually processing it.
2178 * Make sure nobody diddled with it while I was sleeping.
2180 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2183 * Finish PREP stage. We're guaranteed to have at least one idle
2184 * worker or that someone else has already assumed the manager
2185 * role. This is where @worker starts participating in concurrency
2186 * management if applicable and concurrency management is restored
2187 * after being rebound. See rebind_workers() for details.
2189 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2192 struct work_struct
*work
=
2193 list_first_entry(&pool
->worklist
,
2194 struct work_struct
, entry
);
2196 pool
->watchdog_ts
= jiffies
;
2198 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2199 /* optimization path, not strictly necessary */
2200 process_one_work(worker
, work
);
2201 if (unlikely(!list_empty(&worker
->scheduled
)))
2202 process_scheduled_works(worker
);
2204 move_linked_works(work
, &worker
->scheduled
, NULL
);
2205 process_scheduled_works(worker
);
2207 } while (keep_working(pool
));
2209 worker_set_flags(worker
, WORKER_PREP
);
2212 * pool->lock is held and there's no work to process and no need to
2213 * manage, sleep. Workers are woken up only while holding
2214 * pool->lock or from local cpu, so setting the current state
2215 * before releasing pool->lock is enough to prevent losing any
2218 worker_enter_idle(worker
);
2219 __set_current_state(TASK_INTERRUPTIBLE
);
2220 spin_unlock_irq(&pool
->lock
);
2226 * rescuer_thread - the rescuer thread function
2229 * Workqueue rescuer thread function. There's one rescuer for each
2230 * workqueue which has WQ_MEM_RECLAIM set.
2232 * Regular work processing on a pool may block trying to create a new
2233 * worker which uses GFP_KERNEL allocation which has slight chance of
2234 * developing into deadlock if some works currently on the same queue
2235 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2236 * the problem rescuer solves.
2238 * When such condition is possible, the pool summons rescuers of all
2239 * workqueues which have works queued on the pool and let them process
2240 * those works so that forward progress can be guaranteed.
2242 * This should happen rarely.
2246 static int rescuer_thread(void *__rescuer
)
2248 struct worker
*rescuer
= __rescuer
;
2249 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2250 struct list_head
*scheduled
= &rescuer
->scheduled
;
2253 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2256 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2257 * doesn't participate in concurrency management.
2259 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2261 set_current_state(TASK_INTERRUPTIBLE
);
2264 * By the time the rescuer is requested to stop, the workqueue
2265 * shouldn't have any work pending, but @wq->maydays may still have
2266 * pwq(s) queued. This can happen by non-rescuer workers consuming
2267 * all the work items before the rescuer got to them. Go through
2268 * @wq->maydays processing before acting on should_stop so that the
2269 * list is always empty on exit.
2271 should_stop
= kthread_should_stop();
2273 /* see whether any pwq is asking for help */
2274 spin_lock_irq(&wq_mayday_lock
);
2276 while (!list_empty(&wq
->maydays
)) {
2277 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2278 struct pool_workqueue
, mayday_node
);
2279 struct worker_pool
*pool
= pwq
->pool
;
2280 struct work_struct
*work
, *n
;
2283 __set_current_state(TASK_RUNNING
);
2284 list_del_init(&pwq
->mayday_node
);
2286 spin_unlock_irq(&wq_mayday_lock
);
2288 worker_attach_to_pool(rescuer
, pool
);
2290 spin_lock_irq(&pool
->lock
);
2291 rescuer
->pool
= pool
;
2294 * Slurp in all works issued via this workqueue and
2297 WARN_ON_ONCE(!list_empty(scheduled
));
2298 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
) {
2299 if (get_work_pwq(work
) == pwq
) {
2301 pool
->watchdog_ts
= jiffies
;
2302 move_linked_works(work
, scheduled
, &n
);
2307 if (!list_empty(scheduled
)) {
2308 process_scheduled_works(rescuer
);
2311 * The above execution of rescued work items could
2312 * have created more to rescue through
2313 * pwq_activate_first_delayed() or chained
2314 * queueing. Let's put @pwq back on mayday list so
2315 * that such back-to-back work items, which may be
2316 * being used to relieve memory pressure, don't
2317 * incur MAYDAY_INTERVAL delay inbetween.
2319 if (need_to_create_worker(pool
)) {
2320 spin_lock(&wq_mayday_lock
);
2322 list_move_tail(&pwq
->mayday_node
, &wq
->maydays
);
2323 spin_unlock(&wq_mayday_lock
);
2328 * Put the reference grabbed by send_mayday(). @pool won't
2329 * go away while we're still attached to it.
2334 * Leave this pool. If need_more_worker() is %true, notify a
2335 * regular worker; otherwise, we end up with 0 concurrency
2336 * and stalling the execution.
2338 if (need_more_worker(pool
))
2339 wake_up_worker(pool
);
2341 rescuer
->pool
= NULL
;
2342 spin_unlock_irq(&pool
->lock
);
2344 worker_detach_from_pool(rescuer
, pool
);
2346 spin_lock_irq(&wq_mayday_lock
);
2349 spin_unlock_irq(&wq_mayday_lock
);
2352 __set_current_state(TASK_RUNNING
);
2353 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2357 /* rescuers should never participate in concurrency management */
2358 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2364 * check_flush_dependency - check for flush dependency sanity
2365 * @target_wq: workqueue being flushed
2366 * @target_work: work item being flushed (NULL for workqueue flushes)
2368 * %current is trying to flush the whole @target_wq or @target_work on it.
2369 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2370 * reclaiming memory or running on a workqueue which doesn't have
2371 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2374 static void check_flush_dependency(struct workqueue_struct
*target_wq
,
2375 struct work_struct
*target_work
)
2377 work_func_t target_func
= target_work
? target_work
->func
: NULL
;
2378 struct worker
*worker
;
2380 if (target_wq
->flags
& WQ_MEM_RECLAIM
)
2383 worker
= current_wq_worker();
2385 WARN_ONCE(current
->flags
& PF_MEMALLOC
,
2386 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2387 current
->pid
, current
->comm
, target_wq
->name
, target_func
);
2388 WARN_ONCE(worker
&& ((worker
->current_pwq
->wq
->flags
&
2389 (WQ_MEM_RECLAIM
| __WQ_LEGACY
)) == WQ_MEM_RECLAIM
),
2390 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2391 worker
->current_pwq
->wq
->name
, worker
->current_func
,
2392 target_wq
->name
, target_func
);
2396 struct work_struct work
;
2397 struct completion done
;
2398 struct task_struct
*task
; /* purely informational */
2401 static void wq_barrier_func(struct work_struct
*work
)
2403 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2404 complete(&barr
->done
);
2408 * insert_wq_barrier - insert a barrier work
2409 * @pwq: pwq to insert barrier into
2410 * @barr: wq_barrier to insert
2411 * @target: target work to attach @barr to
2412 * @worker: worker currently executing @target, NULL if @target is not executing
2414 * @barr is linked to @target such that @barr is completed only after
2415 * @target finishes execution. Please note that the ordering
2416 * guarantee is observed only with respect to @target and on the local
2419 * Currently, a queued barrier can't be canceled. This is because
2420 * try_to_grab_pending() can't determine whether the work to be
2421 * grabbed is at the head of the queue and thus can't clear LINKED
2422 * flag of the previous work while there must be a valid next work
2423 * after a work with LINKED flag set.
2425 * Note that when @worker is non-NULL, @target may be modified
2426 * underneath us, so we can't reliably determine pwq from @target.
2429 * spin_lock_irq(pool->lock).
2431 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2432 struct wq_barrier
*barr
,
2433 struct work_struct
*target
, struct worker
*worker
)
2435 struct list_head
*head
;
2436 unsigned int linked
= 0;
2439 * debugobject calls are safe here even with pool->lock locked
2440 * as we know for sure that this will not trigger any of the
2441 * checks and call back into the fixup functions where we
2444 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2445 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2446 init_completion(&barr
->done
);
2447 barr
->task
= current
;
2450 * If @target is currently being executed, schedule the
2451 * barrier to the worker; otherwise, put it after @target.
2454 head
= worker
->scheduled
.next
;
2456 unsigned long *bits
= work_data_bits(target
);
2458 head
= target
->entry
.next
;
2459 /* there can already be other linked works, inherit and set */
2460 linked
= *bits
& WORK_STRUCT_LINKED
;
2461 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2464 debug_work_activate(&barr
->work
);
2465 insert_work(pwq
, &barr
->work
, head
,
2466 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2470 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2471 * @wq: workqueue being flushed
2472 * @flush_color: new flush color, < 0 for no-op
2473 * @work_color: new work color, < 0 for no-op
2475 * Prepare pwqs for workqueue flushing.
2477 * If @flush_color is non-negative, flush_color on all pwqs should be
2478 * -1. If no pwq has in-flight commands at the specified color, all
2479 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2480 * has in flight commands, its pwq->flush_color is set to
2481 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2482 * wakeup logic is armed and %true is returned.
2484 * The caller should have initialized @wq->first_flusher prior to
2485 * calling this function with non-negative @flush_color. If
2486 * @flush_color is negative, no flush color update is done and %false
2489 * If @work_color is non-negative, all pwqs should have the same
2490 * work_color which is previous to @work_color and all will be
2491 * advanced to @work_color.
2494 * mutex_lock(wq->mutex).
2497 * %true if @flush_color >= 0 and there's something to flush. %false
2500 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2501 int flush_color
, int work_color
)
2504 struct pool_workqueue
*pwq
;
2506 if (flush_color
>= 0) {
2507 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2508 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2511 for_each_pwq(pwq
, wq
) {
2512 struct worker_pool
*pool
= pwq
->pool
;
2514 spin_lock_irq(&pool
->lock
);
2516 if (flush_color
>= 0) {
2517 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2519 if (pwq
->nr_in_flight
[flush_color
]) {
2520 pwq
->flush_color
= flush_color
;
2521 atomic_inc(&wq
->nr_pwqs_to_flush
);
2526 if (work_color
>= 0) {
2527 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2528 pwq
->work_color
= work_color
;
2531 spin_unlock_irq(&pool
->lock
);
2534 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2535 complete(&wq
->first_flusher
->done
);
2541 * flush_workqueue - ensure that any scheduled work has run to completion.
2542 * @wq: workqueue to flush
2544 * This function sleeps until all work items which were queued on entry
2545 * have finished execution, but it is not livelocked by new incoming ones.
2547 void flush_workqueue(struct workqueue_struct
*wq
)
2549 struct wq_flusher this_flusher
= {
2550 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2552 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2556 lock_map_acquire(&wq
->lockdep_map
);
2557 lock_map_release(&wq
->lockdep_map
);
2559 mutex_lock(&wq
->mutex
);
2562 * Start-to-wait phase
2564 next_color
= work_next_color(wq
->work_color
);
2566 if (next_color
!= wq
->flush_color
) {
2568 * Color space is not full. The current work_color
2569 * becomes our flush_color and work_color is advanced
2572 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2573 this_flusher
.flush_color
= wq
->work_color
;
2574 wq
->work_color
= next_color
;
2576 if (!wq
->first_flusher
) {
2577 /* no flush in progress, become the first flusher */
2578 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2580 wq
->first_flusher
= &this_flusher
;
2582 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2584 /* nothing to flush, done */
2585 wq
->flush_color
= next_color
;
2586 wq
->first_flusher
= NULL
;
2591 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2592 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2593 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2597 * Oops, color space is full, wait on overflow queue.
2598 * The next flush completion will assign us
2599 * flush_color and transfer to flusher_queue.
2601 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2604 check_flush_dependency(wq
, NULL
);
2606 mutex_unlock(&wq
->mutex
);
2608 wait_for_completion(&this_flusher
.done
);
2611 * Wake-up-and-cascade phase
2613 * First flushers are responsible for cascading flushes and
2614 * handling overflow. Non-first flushers can simply return.
2616 if (wq
->first_flusher
!= &this_flusher
)
2619 mutex_lock(&wq
->mutex
);
2621 /* we might have raced, check again with mutex held */
2622 if (wq
->first_flusher
!= &this_flusher
)
2625 wq
->first_flusher
= NULL
;
2627 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2628 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2631 struct wq_flusher
*next
, *tmp
;
2633 /* complete all the flushers sharing the current flush color */
2634 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2635 if (next
->flush_color
!= wq
->flush_color
)
2637 list_del_init(&next
->list
);
2638 complete(&next
->done
);
2641 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2642 wq
->flush_color
!= work_next_color(wq
->work_color
));
2644 /* this flush_color is finished, advance by one */
2645 wq
->flush_color
= work_next_color(wq
->flush_color
);
2647 /* one color has been freed, handle overflow queue */
2648 if (!list_empty(&wq
->flusher_overflow
)) {
2650 * Assign the same color to all overflowed
2651 * flushers, advance work_color and append to
2652 * flusher_queue. This is the start-to-wait
2653 * phase for these overflowed flushers.
2655 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2656 tmp
->flush_color
= wq
->work_color
;
2658 wq
->work_color
= work_next_color(wq
->work_color
);
2660 list_splice_tail_init(&wq
->flusher_overflow
,
2661 &wq
->flusher_queue
);
2662 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2665 if (list_empty(&wq
->flusher_queue
)) {
2666 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2671 * Need to flush more colors. Make the next flusher
2672 * the new first flusher and arm pwqs.
2674 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2675 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2677 list_del_init(&next
->list
);
2678 wq
->first_flusher
= next
;
2680 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2684 * Meh... this color is already done, clear first
2685 * flusher and repeat cascading.
2687 wq
->first_flusher
= NULL
;
2691 mutex_unlock(&wq
->mutex
);
2693 EXPORT_SYMBOL(flush_workqueue
);
2696 * drain_workqueue - drain a workqueue
2697 * @wq: workqueue to drain
2699 * Wait until the workqueue becomes empty. While draining is in progress,
2700 * only chain queueing is allowed. IOW, only currently pending or running
2701 * work items on @wq can queue further work items on it. @wq is flushed
2702 * repeatedly until it becomes empty. The number of flushing is determined
2703 * by the depth of chaining and should be relatively short. Whine if it
2706 void drain_workqueue(struct workqueue_struct
*wq
)
2708 unsigned int flush_cnt
= 0;
2709 struct pool_workqueue
*pwq
;
2712 * __queue_work() needs to test whether there are drainers, is much
2713 * hotter than drain_workqueue() and already looks at @wq->flags.
2714 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2716 mutex_lock(&wq
->mutex
);
2717 if (!wq
->nr_drainers
++)
2718 wq
->flags
|= __WQ_DRAINING
;
2719 mutex_unlock(&wq
->mutex
);
2721 flush_workqueue(wq
);
2723 mutex_lock(&wq
->mutex
);
2725 for_each_pwq(pwq
, wq
) {
2728 spin_lock_irq(&pwq
->pool
->lock
);
2729 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2730 spin_unlock_irq(&pwq
->pool
->lock
);
2735 if (++flush_cnt
== 10 ||
2736 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2737 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2738 wq
->name
, flush_cnt
);
2740 mutex_unlock(&wq
->mutex
);
2744 if (!--wq
->nr_drainers
)
2745 wq
->flags
&= ~__WQ_DRAINING
;
2746 mutex_unlock(&wq
->mutex
);
2748 EXPORT_SYMBOL_GPL(drain_workqueue
);
2750 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2752 struct worker
*worker
= NULL
;
2753 struct worker_pool
*pool
;
2754 struct pool_workqueue
*pwq
;
2758 local_irq_disable();
2759 pool
= get_work_pool(work
);
2765 spin_lock(&pool
->lock
);
2766 /* see the comment in try_to_grab_pending() with the same code */
2767 pwq
= get_work_pwq(work
);
2769 if (unlikely(pwq
->pool
!= pool
))
2772 worker
= find_worker_executing_work(pool
, work
);
2775 pwq
= worker
->current_pwq
;
2778 check_flush_dependency(pwq
->wq
, work
);
2780 insert_wq_barrier(pwq
, barr
, work
, worker
);
2781 spin_unlock_irq(&pool
->lock
);
2784 * If @max_active is 1 or rescuer is in use, flushing another work
2785 * item on the same workqueue may lead to deadlock. Make sure the
2786 * flusher is not running on the same workqueue by verifying write
2789 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2790 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2792 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2793 lock_map_release(&pwq
->wq
->lockdep_map
);
2797 spin_unlock_irq(&pool
->lock
);
2802 * flush_work - wait for a work to finish executing the last queueing instance
2803 * @work: the work to flush
2805 * Wait until @work has finished execution. @work is guaranteed to be idle
2806 * on return if it hasn't been requeued since flush started.
2809 * %true if flush_work() waited for the work to finish execution,
2810 * %false if it was already idle.
2812 bool flush_work(struct work_struct
*work
)
2814 struct wq_barrier barr
;
2816 lock_map_acquire(&work
->lockdep_map
);
2817 lock_map_release(&work
->lockdep_map
);
2819 if (start_flush_work(work
, &barr
)) {
2820 wait_for_completion(&barr
.done
);
2821 destroy_work_on_stack(&barr
.work
);
2827 EXPORT_SYMBOL_GPL(flush_work
);
2831 struct work_struct
*work
;
2834 static int cwt_wakefn(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
2836 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
2838 if (cwait
->work
!= key
)
2840 return autoremove_wake_function(wait
, mode
, sync
, key
);
2843 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2845 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
2846 unsigned long flags
;
2850 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2852 * If someone else is already canceling, wait for it to
2853 * finish. flush_work() doesn't work for PREEMPT_NONE
2854 * because we may get scheduled between @work's completion
2855 * and the other canceling task resuming and clearing
2856 * CANCELING - flush_work() will return false immediately
2857 * as @work is no longer busy, try_to_grab_pending() will
2858 * return -ENOENT as @work is still being canceled and the
2859 * other canceling task won't be able to clear CANCELING as
2860 * we're hogging the CPU.
2862 * Let's wait for completion using a waitqueue. As this
2863 * may lead to the thundering herd problem, use a custom
2864 * wake function which matches @work along with exclusive
2867 if (unlikely(ret
== -ENOENT
)) {
2868 struct cwt_wait cwait
;
2870 init_wait(&cwait
.wait
);
2871 cwait
.wait
.func
= cwt_wakefn
;
2874 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
2875 TASK_UNINTERRUPTIBLE
);
2876 if (work_is_canceling(work
))
2878 finish_wait(&cancel_waitq
, &cwait
.wait
);
2880 } while (unlikely(ret
< 0));
2882 /* tell other tasks trying to grab @work to back off */
2883 mark_work_canceling(work
);
2884 local_irq_restore(flags
);
2887 clear_work_data(work
);
2890 * Paired with prepare_to_wait() above so that either
2891 * waitqueue_active() is visible here or !work_is_canceling() is
2895 if (waitqueue_active(&cancel_waitq
))
2896 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
2902 * cancel_work_sync - cancel a work and wait for it to finish
2903 * @work: the work to cancel
2905 * Cancel @work and wait for its execution to finish. This function
2906 * can be used even if the work re-queues itself or migrates to
2907 * another workqueue. On return from this function, @work is
2908 * guaranteed to be not pending or executing on any CPU.
2910 * cancel_work_sync(&delayed_work->work) must not be used for
2911 * delayed_work's. Use cancel_delayed_work_sync() instead.
2913 * The caller must ensure that the workqueue on which @work was last
2914 * queued can't be destroyed before this function returns.
2917 * %true if @work was pending, %false otherwise.
2919 bool cancel_work_sync(struct work_struct
*work
)
2921 return __cancel_work_timer(work
, false);
2923 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2926 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2927 * @dwork: the delayed work to flush
2929 * Delayed timer is cancelled and the pending work is queued for
2930 * immediate execution. Like flush_work(), this function only
2931 * considers the last queueing instance of @dwork.
2934 * %true if flush_work() waited for the work to finish execution,
2935 * %false if it was already idle.
2937 bool flush_delayed_work(struct delayed_work
*dwork
)
2939 local_irq_disable();
2940 if (del_timer_sync(&dwork
->timer
))
2941 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2943 return flush_work(&dwork
->work
);
2945 EXPORT_SYMBOL(flush_delayed_work
);
2948 * cancel_delayed_work - cancel a delayed work
2949 * @dwork: delayed_work to cancel
2951 * Kill off a pending delayed_work.
2953 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2957 * The work callback function may still be running on return, unless
2958 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2959 * use cancel_delayed_work_sync() to wait on it.
2961 * This function is safe to call from any context including IRQ handler.
2963 bool cancel_delayed_work(struct delayed_work
*dwork
)
2965 unsigned long flags
;
2969 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2970 } while (unlikely(ret
== -EAGAIN
));
2972 if (unlikely(ret
< 0))
2975 set_work_pool_and_clear_pending(&dwork
->work
,
2976 get_work_pool_id(&dwork
->work
));
2977 local_irq_restore(flags
);
2980 EXPORT_SYMBOL(cancel_delayed_work
);
2983 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2984 * @dwork: the delayed work cancel
2986 * This is cancel_work_sync() for delayed works.
2989 * %true if @dwork was pending, %false otherwise.
2991 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2993 return __cancel_work_timer(&dwork
->work
, true);
2995 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2998 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2999 * @func: the function to call
3001 * schedule_on_each_cpu() executes @func on each online CPU using the
3002 * system workqueue and blocks until all CPUs have completed.
3003 * schedule_on_each_cpu() is very slow.
3006 * 0 on success, -errno on failure.
3008 int schedule_on_each_cpu(work_func_t func
)
3011 struct work_struct __percpu
*works
;
3013 works
= alloc_percpu(struct work_struct
);
3019 for_each_online_cpu(cpu
) {
3020 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3022 INIT_WORK(work
, func
);
3023 schedule_work_on(cpu
, work
);
3026 for_each_online_cpu(cpu
)
3027 flush_work(per_cpu_ptr(works
, cpu
));
3035 * execute_in_process_context - reliably execute the routine with user context
3036 * @fn: the function to execute
3037 * @ew: guaranteed storage for the execute work structure (must
3038 * be available when the work executes)
3040 * Executes the function immediately if process context is available,
3041 * otherwise schedules the function for delayed execution.
3043 * Return: 0 - function was executed
3044 * 1 - function was scheduled for execution
3046 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3048 if (!in_interrupt()) {
3053 INIT_WORK(&ew
->work
, fn
);
3054 schedule_work(&ew
->work
);
3058 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3061 * free_workqueue_attrs - free a workqueue_attrs
3062 * @attrs: workqueue_attrs to free
3064 * Undo alloc_workqueue_attrs().
3066 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3069 free_cpumask_var(attrs
->cpumask
);
3075 * alloc_workqueue_attrs - allocate a workqueue_attrs
3076 * @gfp_mask: allocation mask to use
3078 * Allocate a new workqueue_attrs, initialize with default settings and
3081 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3083 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3085 struct workqueue_attrs
*attrs
;
3087 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3090 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3093 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3096 free_workqueue_attrs(attrs
);
3100 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3101 const struct workqueue_attrs
*from
)
3103 to
->nice
= from
->nice
;
3104 cpumask_copy(to
->cpumask
, from
->cpumask
);
3106 * Unlike hash and equality test, this function doesn't ignore
3107 * ->no_numa as it is used for both pool and wq attrs. Instead,
3108 * get_unbound_pool() explicitly clears ->no_numa after copying.
3110 to
->no_numa
= from
->no_numa
;
3113 /* hash value of the content of @attr */
3114 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3118 hash
= jhash_1word(attrs
->nice
, hash
);
3119 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3120 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3124 /* content equality test */
3125 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3126 const struct workqueue_attrs
*b
)
3128 if (a
->nice
!= b
->nice
)
3130 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3136 * init_worker_pool - initialize a newly zalloc'd worker_pool
3137 * @pool: worker_pool to initialize
3139 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3141 * Return: 0 on success, -errno on failure. Even on failure, all fields
3142 * inside @pool proper are initialized and put_unbound_pool() can be called
3143 * on @pool safely to release it.
3145 static int init_worker_pool(struct worker_pool
*pool
)
3147 spin_lock_init(&pool
->lock
);
3150 pool
->node
= NUMA_NO_NODE
;
3151 pool
->flags
|= POOL_DISASSOCIATED
;
3152 pool
->watchdog_ts
= jiffies
;
3153 INIT_LIST_HEAD(&pool
->worklist
);
3154 INIT_LIST_HEAD(&pool
->idle_list
);
3155 hash_init(pool
->busy_hash
);
3157 init_timer_deferrable(&pool
->idle_timer
);
3158 pool
->idle_timer
.function
= idle_worker_timeout
;
3159 pool
->idle_timer
.data
= (unsigned long)pool
;
3161 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3162 (unsigned long)pool
);
3164 mutex_init(&pool
->manager_arb
);
3165 mutex_init(&pool
->attach_mutex
);
3166 INIT_LIST_HEAD(&pool
->workers
);
3168 ida_init(&pool
->worker_ida
);
3169 INIT_HLIST_NODE(&pool
->hash_node
);
3172 /* shouldn't fail above this point */
3173 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3179 static void rcu_free_wq(struct rcu_head
*rcu
)
3181 struct workqueue_struct
*wq
=
3182 container_of(rcu
, struct workqueue_struct
, rcu
);
3184 if (!(wq
->flags
& WQ_UNBOUND
))
3185 free_percpu(wq
->cpu_pwqs
);
3187 free_workqueue_attrs(wq
->unbound_attrs
);
3193 static void rcu_free_pool(struct rcu_head
*rcu
)
3195 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3197 ida_destroy(&pool
->worker_ida
);
3198 free_workqueue_attrs(pool
->attrs
);
3203 * put_unbound_pool - put a worker_pool
3204 * @pool: worker_pool to put
3206 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3207 * safe manner. get_unbound_pool() calls this function on its failure path
3208 * and this function should be able to release pools which went through,
3209 * successfully or not, init_worker_pool().
3211 * Should be called with wq_pool_mutex held.
3213 static void put_unbound_pool(struct worker_pool
*pool
)
3215 DECLARE_COMPLETION_ONSTACK(detach_completion
);
3216 struct worker
*worker
;
3218 lockdep_assert_held(&wq_pool_mutex
);
3224 if (WARN_ON(!(pool
->cpu
< 0)) ||
3225 WARN_ON(!list_empty(&pool
->worklist
)))
3228 /* release id and unhash */
3230 idr_remove(&worker_pool_idr
, pool
->id
);
3231 hash_del(&pool
->hash_node
);
3234 * Become the manager and destroy all workers. Grabbing
3235 * manager_arb prevents @pool's workers from blocking on
3238 mutex_lock(&pool
->manager_arb
);
3240 spin_lock_irq(&pool
->lock
);
3241 while ((worker
= first_idle_worker(pool
)))
3242 destroy_worker(worker
);
3243 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3244 spin_unlock_irq(&pool
->lock
);
3246 mutex_lock(&pool
->attach_mutex
);
3247 if (!list_empty(&pool
->workers
))
3248 pool
->detach_completion
= &detach_completion
;
3249 mutex_unlock(&pool
->attach_mutex
);
3251 if (pool
->detach_completion
)
3252 wait_for_completion(pool
->detach_completion
);
3254 mutex_unlock(&pool
->manager_arb
);
3256 /* shut down the timers */
3257 del_timer_sync(&pool
->idle_timer
);
3258 del_timer_sync(&pool
->mayday_timer
);
3260 /* sched-RCU protected to allow dereferences from get_work_pool() */
3261 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3265 * get_unbound_pool - get a worker_pool with the specified attributes
3266 * @attrs: the attributes of the worker_pool to get
3268 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3269 * reference count and return it. If there already is a matching
3270 * worker_pool, it will be used; otherwise, this function attempts to
3273 * Should be called with wq_pool_mutex held.
3275 * Return: On success, a worker_pool with the same attributes as @attrs.
3276 * On failure, %NULL.
3278 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3280 u32 hash
= wqattrs_hash(attrs
);
3281 struct worker_pool
*pool
;
3283 int target_node
= NUMA_NO_NODE
;
3285 lockdep_assert_held(&wq_pool_mutex
);
3287 /* do we already have a matching pool? */
3288 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3289 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3295 /* if cpumask is contained inside a NUMA node, we belong to that node */
3296 if (wq_numa_enabled
) {
3297 for_each_node(node
) {
3298 if (cpumask_subset(attrs
->cpumask
,
3299 wq_numa_possible_cpumask
[node
])) {
3306 /* nope, create a new one */
3307 pool
= kzalloc_node(sizeof(*pool
), GFP_KERNEL
, target_node
);
3308 if (!pool
|| init_worker_pool(pool
) < 0)
3311 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3312 copy_workqueue_attrs(pool
->attrs
, attrs
);
3313 pool
->node
= target_node
;
3316 * no_numa isn't a worker_pool attribute, always clear it. See
3317 * 'struct workqueue_attrs' comments for detail.
3319 pool
->attrs
->no_numa
= false;
3321 if (worker_pool_assign_id(pool
) < 0)
3324 /* create and start the initial worker */
3325 if (!create_worker(pool
))
3329 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3334 put_unbound_pool(pool
);
3338 static void rcu_free_pwq(struct rcu_head
*rcu
)
3340 kmem_cache_free(pwq_cache
,
3341 container_of(rcu
, struct pool_workqueue
, rcu
));
3345 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3346 * and needs to be destroyed.
3348 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3350 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3351 unbound_release_work
);
3352 struct workqueue_struct
*wq
= pwq
->wq
;
3353 struct worker_pool
*pool
= pwq
->pool
;
3356 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3359 mutex_lock(&wq
->mutex
);
3360 list_del_rcu(&pwq
->pwqs_node
);
3361 is_last
= list_empty(&wq
->pwqs
);
3362 mutex_unlock(&wq
->mutex
);
3364 mutex_lock(&wq_pool_mutex
);
3365 put_unbound_pool(pool
);
3366 mutex_unlock(&wq_pool_mutex
);
3368 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3371 * If we're the last pwq going away, @wq is already dead and no one
3372 * is gonna access it anymore. Schedule RCU free.
3375 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
3379 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3380 * @pwq: target pool_workqueue
3382 * If @pwq isn't freezing, set @pwq->max_active to the associated
3383 * workqueue's saved_max_active and activate delayed work items
3384 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3386 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3388 struct workqueue_struct
*wq
= pwq
->wq
;
3389 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3391 /* for @wq->saved_max_active */
3392 lockdep_assert_held(&wq
->mutex
);
3394 /* fast exit for non-freezable wqs */
3395 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3398 spin_lock_irq(&pwq
->pool
->lock
);
3401 * During [un]freezing, the caller is responsible for ensuring that
3402 * this function is called at least once after @workqueue_freezing
3403 * is updated and visible.
3405 if (!freezable
|| !workqueue_freezing
) {
3406 pwq
->max_active
= wq
->saved_max_active
;
3408 while (!list_empty(&pwq
->delayed_works
) &&
3409 pwq
->nr_active
< pwq
->max_active
)
3410 pwq_activate_first_delayed(pwq
);
3413 * Need to kick a worker after thawed or an unbound wq's
3414 * max_active is bumped. It's a slow path. Do it always.
3416 wake_up_worker(pwq
->pool
);
3418 pwq
->max_active
= 0;
3421 spin_unlock_irq(&pwq
->pool
->lock
);
3424 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3425 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3426 struct worker_pool
*pool
)
3428 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3430 memset(pwq
, 0, sizeof(*pwq
));
3434 pwq
->flush_color
= -1;
3436 INIT_LIST_HEAD(&pwq
->delayed_works
);
3437 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3438 INIT_LIST_HEAD(&pwq
->mayday_node
);
3439 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3442 /* sync @pwq with the current state of its associated wq and link it */
3443 static void link_pwq(struct pool_workqueue
*pwq
)
3445 struct workqueue_struct
*wq
= pwq
->wq
;
3447 lockdep_assert_held(&wq
->mutex
);
3449 /* may be called multiple times, ignore if already linked */
3450 if (!list_empty(&pwq
->pwqs_node
))
3453 /* set the matching work_color */
3454 pwq
->work_color
= wq
->work_color
;
3456 /* sync max_active to the current setting */
3457 pwq_adjust_max_active(pwq
);
3460 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3463 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3464 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3465 const struct workqueue_attrs
*attrs
)
3467 struct worker_pool
*pool
;
3468 struct pool_workqueue
*pwq
;
3470 lockdep_assert_held(&wq_pool_mutex
);
3472 pool
= get_unbound_pool(attrs
);
3476 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3478 put_unbound_pool(pool
);
3482 init_pwq(pwq
, wq
, pool
);
3487 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3488 * @attrs: the wq_attrs of the default pwq of the target workqueue
3489 * @node: the target NUMA node
3490 * @cpu_going_down: if >= 0, the CPU to consider as offline
3491 * @cpumask: outarg, the resulting cpumask
3493 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3494 * @cpu_going_down is >= 0, that cpu is considered offline during
3495 * calculation. The result is stored in @cpumask.
3497 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3498 * enabled and @node has online CPUs requested by @attrs, the returned
3499 * cpumask is the intersection of the possible CPUs of @node and
3502 * The caller is responsible for ensuring that the cpumask of @node stays
3505 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3508 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3509 int cpu_going_down
, cpumask_t
*cpumask
)
3511 if (!wq_numa_enabled
|| attrs
->no_numa
)
3514 /* does @node have any online CPUs @attrs wants? */
3515 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3516 if (cpu_going_down
>= 0)
3517 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3519 if (cpumask_empty(cpumask
))
3522 /* yeap, return possible CPUs in @node that @attrs wants */
3523 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3524 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3527 cpumask_copy(cpumask
, attrs
->cpumask
);
3531 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3532 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3534 struct pool_workqueue
*pwq
)
3536 struct pool_workqueue
*old_pwq
;
3538 lockdep_assert_held(&wq_pool_mutex
);
3539 lockdep_assert_held(&wq
->mutex
);
3541 /* link_pwq() can handle duplicate calls */
3544 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3545 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3549 /* context to store the prepared attrs & pwqs before applying */
3550 struct apply_wqattrs_ctx
{
3551 struct workqueue_struct
*wq
; /* target workqueue */
3552 struct workqueue_attrs
*attrs
; /* attrs to apply */
3553 struct list_head list
; /* queued for batching commit */
3554 struct pool_workqueue
*dfl_pwq
;
3555 struct pool_workqueue
*pwq_tbl
[];
3558 /* free the resources after success or abort */
3559 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx
*ctx
)
3565 put_pwq_unlocked(ctx
->pwq_tbl
[node
]);
3566 put_pwq_unlocked(ctx
->dfl_pwq
);
3568 free_workqueue_attrs(ctx
->attrs
);
3574 /* allocate the attrs and pwqs for later installation */
3575 static struct apply_wqattrs_ctx
*
3576 apply_wqattrs_prepare(struct workqueue_struct
*wq
,
3577 const struct workqueue_attrs
*attrs
)
3579 struct apply_wqattrs_ctx
*ctx
;
3580 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3583 lockdep_assert_held(&wq_pool_mutex
);
3585 ctx
= kzalloc(sizeof(*ctx
) + nr_node_ids
* sizeof(ctx
->pwq_tbl
[0]),
3588 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3589 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3590 if (!ctx
|| !new_attrs
|| !tmp_attrs
)
3594 * Calculate the attrs of the default pwq.
3595 * If the user configured cpumask doesn't overlap with the
3596 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3598 copy_workqueue_attrs(new_attrs
, attrs
);
3599 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, wq_unbound_cpumask
);
3600 if (unlikely(cpumask_empty(new_attrs
->cpumask
)))
3601 cpumask_copy(new_attrs
->cpumask
, wq_unbound_cpumask
);
3604 * We may create multiple pwqs with differing cpumasks. Make a
3605 * copy of @new_attrs which will be modified and used to obtain
3608 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3611 * If something goes wrong during CPU up/down, we'll fall back to
3612 * the default pwq covering whole @attrs->cpumask. Always create
3613 * it even if we don't use it immediately.
3615 ctx
->dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3619 for_each_node(node
) {
3620 if (wq_calc_node_cpumask(new_attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3621 ctx
->pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3622 if (!ctx
->pwq_tbl
[node
])
3625 ctx
->dfl_pwq
->refcnt
++;
3626 ctx
->pwq_tbl
[node
] = ctx
->dfl_pwq
;
3630 /* save the user configured attrs and sanitize it. */
3631 copy_workqueue_attrs(new_attrs
, attrs
);
3632 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3633 ctx
->attrs
= new_attrs
;
3636 free_workqueue_attrs(tmp_attrs
);
3640 free_workqueue_attrs(tmp_attrs
);
3641 free_workqueue_attrs(new_attrs
);
3642 apply_wqattrs_cleanup(ctx
);
3646 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3647 static void apply_wqattrs_commit(struct apply_wqattrs_ctx
*ctx
)
3651 /* all pwqs have been created successfully, let's install'em */
3652 mutex_lock(&ctx
->wq
->mutex
);
3654 copy_workqueue_attrs(ctx
->wq
->unbound_attrs
, ctx
->attrs
);
3656 /* save the previous pwq and install the new one */
3658 ctx
->pwq_tbl
[node
] = numa_pwq_tbl_install(ctx
->wq
, node
,
3659 ctx
->pwq_tbl
[node
]);
3661 /* @dfl_pwq might not have been used, ensure it's linked */
3662 link_pwq(ctx
->dfl_pwq
);
3663 swap(ctx
->wq
->dfl_pwq
, ctx
->dfl_pwq
);
3665 mutex_unlock(&ctx
->wq
->mutex
);
3668 static void apply_wqattrs_lock(void)
3670 /* CPUs should stay stable across pwq creations and installations */
3672 mutex_lock(&wq_pool_mutex
);
3675 static void apply_wqattrs_unlock(void)
3677 mutex_unlock(&wq_pool_mutex
);
3681 static int apply_workqueue_attrs_locked(struct workqueue_struct
*wq
,
3682 const struct workqueue_attrs
*attrs
)
3684 struct apply_wqattrs_ctx
*ctx
;
3686 /* only unbound workqueues can change attributes */
3687 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3690 /* creating multiple pwqs breaks ordering guarantee */
3691 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3694 ctx
= apply_wqattrs_prepare(wq
, attrs
);
3698 /* the ctx has been prepared successfully, let's commit it */
3699 apply_wqattrs_commit(ctx
);
3700 apply_wqattrs_cleanup(ctx
);
3706 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3707 * @wq: the target workqueue
3708 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3710 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3711 * machines, this function maps a separate pwq to each NUMA node with
3712 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3713 * NUMA node it was issued on. Older pwqs are released as in-flight work
3714 * items finish. Note that a work item which repeatedly requeues itself
3715 * back-to-back will stay on its current pwq.
3717 * Performs GFP_KERNEL allocations.
3719 * Return: 0 on success and -errno on failure.
3721 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3722 const struct workqueue_attrs
*attrs
)
3726 apply_wqattrs_lock();
3727 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
3728 apply_wqattrs_unlock();
3734 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3735 * @wq: the target workqueue
3736 * @cpu: the CPU coming up or going down
3737 * @online: whether @cpu is coming up or going down
3739 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3740 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3743 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3744 * falls back to @wq->dfl_pwq which may not be optimal but is always
3747 * Note that when the last allowed CPU of a NUMA node goes offline for a
3748 * workqueue with a cpumask spanning multiple nodes, the workers which were
3749 * already executing the work items for the workqueue will lose their CPU
3750 * affinity and may execute on any CPU. This is similar to how per-cpu
3751 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3752 * affinity, it's the user's responsibility to flush the work item from
3755 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3758 int node
= cpu_to_node(cpu
);
3759 int cpu_off
= online
? -1 : cpu
;
3760 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3761 struct workqueue_attrs
*target_attrs
;
3764 lockdep_assert_held(&wq_pool_mutex
);
3766 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
) ||
3767 wq
->unbound_attrs
->no_numa
)
3771 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3772 * Let's use a preallocated one. The following buf is protected by
3773 * CPU hotplug exclusion.
3775 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3776 cpumask
= target_attrs
->cpumask
;
3778 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3779 pwq
= unbound_pwq_by_node(wq
, node
);
3782 * Let's determine what needs to be done. If the target cpumask is
3783 * different from the default pwq's, we need to compare it to @pwq's
3784 * and create a new one if they don't match. If the target cpumask
3785 * equals the default pwq's, the default pwq should be used.
3787 if (wq_calc_node_cpumask(wq
->dfl_pwq
->pool
->attrs
, node
, cpu_off
, cpumask
)) {
3788 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3794 /* create a new pwq */
3795 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3797 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3802 /* Install the new pwq. */
3803 mutex_lock(&wq
->mutex
);
3804 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3808 mutex_lock(&wq
->mutex
);
3809 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3810 get_pwq(wq
->dfl_pwq
);
3811 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3812 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
3814 mutex_unlock(&wq
->mutex
);
3815 put_pwq_unlocked(old_pwq
);
3818 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3820 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3823 if (!(wq
->flags
& WQ_UNBOUND
)) {
3824 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3828 for_each_possible_cpu(cpu
) {
3829 struct pool_workqueue
*pwq
=
3830 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3831 struct worker_pool
*cpu_pools
=
3832 per_cpu(cpu_worker_pools
, cpu
);
3834 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
3836 mutex_lock(&wq
->mutex
);
3838 mutex_unlock(&wq
->mutex
);
3841 } else if (wq
->flags
& __WQ_ORDERED
) {
3842 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
3843 /* there should only be single pwq for ordering guarantee */
3844 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
3845 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
3846 "ordering guarantee broken for workqueue %s\n", wq
->name
);
3849 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3853 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3856 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3858 if (max_active
< 1 || max_active
> lim
)
3859 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3860 max_active
, name
, 1, lim
);
3862 return clamp_val(max_active
, 1, lim
);
3865 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3868 struct lock_class_key
*key
,
3869 const char *lock_name
, ...)
3871 size_t tbl_size
= 0;
3873 struct workqueue_struct
*wq
;
3874 struct pool_workqueue
*pwq
;
3876 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3877 if ((flags
& WQ_POWER_EFFICIENT
) && wq_power_efficient
)
3878 flags
|= WQ_UNBOUND
;
3880 /* allocate wq and format name */
3881 if (flags
& WQ_UNBOUND
)
3882 tbl_size
= nr_node_ids
* sizeof(wq
->numa_pwq_tbl
[0]);
3884 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
3888 if (flags
& WQ_UNBOUND
) {
3889 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3890 if (!wq
->unbound_attrs
)
3894 va_start(args
, lock_name
);
3895 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
3898 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3899 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3903 wq
->saved_max_active
= max_active
;
3904 mutex_init(&wq
->mutex
);
3905 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3906 INIT_LIST_HEAD(&wq
->pwqs
);
3907 INIT_LIST_HEAD(&wq
->flusher_queue
);
3908 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3909 INIT_LIST_HEAD(&wq
->maydays
);
3911 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3912 INIT_LIST_HEAD(&wq
->list
);
3914 if (alloc_and_link_pwqs(wq
) < 0)
3918 * Workqueues which may be used during memory reclaim should
3919 * have a rescuer to guarantee forward progress.
3921 if (flags
& WQ_MEM_RECLAIM
) {
3922 struct worker
*rescuer
;
3924 rescuer
= alloc_worker(NUMA_NO_NODE
);
3928 rescuer
->rescue_wq
= wq
;
3929 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3931 if (IS_ERR(rescuer
->task
)) {
3936 wq
->rescuer
= rescuer
;
3937 kthread_bind_mask(rescuer
->task
, cpu_possible_mask
);
3938 wake_up_process(rescuer
->task
);
3941 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
3945 * wq_pool_mutex protects global freeze state and workqueues list.
3946 * Grab it, adjust max_active and add the new @wq to workqueues
3949 mutex_lock(&wq_pool_mutex
);
3951 mutex_lock(&wq
->mutex
);
3952 for_each_pwq(pwq
, wq
)
3953 pwq_adjust_max_active(pwq
);
3954 mutex_unlock(&wq
->mutex
);
3956 list_add_tail_rcu(&wq
->list
, &workqueues
);
3958 mutex_unlock(&wq_pool_mutex
);
3963 free_workqueue_attrs(wq
->unbound_attrs
);
3967 destroy_workqueue(wq
);
3970 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3973 * destroy_workqueue - safely terminate a workqueue
3974 * @wq: target workqueue
3976 * Safely destroy a workqueue. All work currently pending will be done first.
3978 void destroy_workqueue(struct workqueue_struct
*wq
)
3980 struct pool_workqueue
*pwq
;
3983 /* drain it before proceeding with destruction */
3984 drain_workqueue(wq
);
3987 mutex_lock(&wq
->mutex
);
3988 for_each_pwq(pwq
, wq
) {
3991 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
3992 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
3993 mutex_unlock(&wq
->mutex
);
3998 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
3999 WARN_ON(pwq
->nr_active
) ||
4000 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4001 mutex_unlock(&wq
->mutex
);
4005 mutex_unlock(&wq
->mutex
);
4008 * wq list is used to freeze wq, remove from list after
4009 * flushing is complete in case freeze races us.
4011 mutex_lock(&wq_pool_mutex
);
4012 list_del_rcu(&wq
->list
);
4013 mutex_unlock(&wq_pool_mutex
);
4015 workqueue_sysfs_unregister(wq
);
4018 kthread_stop(wq
->rescuer
->task
);
4020 if (!(wq
->flags
& WQ_UNBOUND
)) {
4022 * The base ref is never dropped on per-cpu pwqs. Directly
4023 * schedule RCU free.
4025 call_rcu_sched(&wq
->rcu
, rcu_free_wq
);
4028 * We're the sole accessor of @wq at this point. Directly
4029 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4030 * @wq will be freed when the last pwq is released.
4032 for_each_node(node
) {
4033 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4034 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4035 put_pwq_unlocked(pwq
);
4039 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4040 * put. Don't access it afterwards.
4044 put_pwq_unlocked(pwq
);
4047 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4050 * workqueue_set_max_active - adjust max_active of a workqueue
4051 * @wq: target workqueue
4052 * @max_active: new max_active value.
4054 * Set max_active of @wq to @max_active.
4057 * Don't call from IRQ context.
4059 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4061 struct pool_workqueue
*pwq
;
4063 /* disallow meddling with max_active for ordered workqueues */
4064 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4067 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4069 mutex_lock(&wq
->mutex
);
4071 wq
->saved_max_active
= max_active
;
4073 for_each_pwq(pwq
, wq
)
4074 pwq_adjust_max_active(pwq
);
4076 mutex_unlock(&wq
->mutex
);
4078 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4081 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4083 * Determine whether %current is a workqueue rescuer. Can be used from
4084 * work functions to determine whether it's being run off the rescuer task.
4086 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4088 bool current_is_workqueue_rescuer(void)
4090 struct worker
*worker
= current_wq_worker();
4092 return worker
&& worker
->rescue_wq
;
4096 * workqueue_congested - test whether a workqueue is congested
4097 * @cpu: CPU in question
4098 * @wq: target workqueue
4100 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4101 * no synchronization around this function and the test result is
4102 * unreliable and only useful as advisory hints or for debugging.
4104 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4105 * Note that both per-cpu and unbound workqueues may be associated with
4106 * multiple pool_workqueues which have separate congested states. A
4107 * workqueue being congested on one CPU doesn't mean the workqueue is also
4108 * contested on other CPUs / NUMA nodes.
4111 * %true if congested, %false otherwise.
4113 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4115 struct pool_workqueue
*pwq
;
4118 rcu_read_lock_sched();
4120 if (cpu
== WORK_CPU_UNBOUND
)
4121 cpu
= smp_processor_id();
4123 if (!(wq
->flags
& WQ_UNBOUND
))
4124 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4126 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4128 ret
= !list_empty(&pwq
->delayed_works
);
4129 rcu_read_unlock_sched();
4133 EXPORT_SYMBOL_GPL(workqueue_congested
);
4136 * work_busy - test whether a work is currently pending or running
4137 * @work: the work to be tested
4139 * Test whether @work is currently pending or running. There is no
4140 * synchronization around this function and the test result is
4141 * unreliable and only useful as advisory hints or for debugging.
4144 * OR'd bitmask of WORK_BUSY_* bits.
4146 unsigned int work_busy(struct work_struct
*work
)
4148 struct worker_pool
*pool
;
4149 unsigned long flags
;
4150 unsigned int ret
= 0;
4152 if (work_pending(work
))
4153 ret
|= WORK_BUSY_PENDING
;
4155 local_irq_save(flags
);
4156 pool
= get_work_pool(work
);
4158 spin_lock(&pool
->lock
);
4159 if (find_worker_executing_work(pool
, work
))
4160 ret
|= WORK_BUSY_RUNNING
;
4161 spin_unlock(&pool
->lock
);
4163 local_irq_restore(flags
);
4167 EXPORT_SYMBOL_GPL(work_busy
);
4170 * set_worker_desc - set description for the current work item
4171 * @fmt: printf-style format string
4172 * @...: arguments for the format string
4174 * This function can be called by a running work function to describe what
4175 * the work item is about. If the worker task gets dumped, this
4176 * information will be printed out together to help debugging. The
4177 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4179 void set_worker_desc(const char *fmt
, ...)
4181 struct worker
*worker
= current_wq_worker();
4185 va_start(args
, fmt
);
4186 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4188 worker
->desc_valid
= true;
4193 * print_worker_info - print out worker information and description
4194 * @log_lvl: the log level to use when printing
4195 * @task: target task
4197 * If @task is a worker and currently executing a work item, print out the
4198 * name of the workqueue being serviced and worker description set with
4199 * set_worker_desc() by the currently executing work item.
4201 * This function can be safely called on any task as long as the
4202 * task_struct itself is accessible. While safe, this function isn't
4203 * synchronized and may print out mixups or garbages of limited length.
4205 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4207 work_func_t
*fn
= NULL
;
4208 char name
[WQ_NAME_LEN
] = { };
4209 char desc
[WORKER_DESC_LEN
] = { };
4210 struct pool_workqueue
*pwq
= NULL
;
4211 struct workqueue_struct
*wq
= NULL
;
4212 bool desc_valid
= false;
4213 struct worker
*worker
;
4215 if (!(task
->flags
& PF_WQ_WORKER
))
4219 * This function is called without any synchronization and @task
4220 * could be in any state. Be careful with dereferences.
4222 worker
= probe_kthread_data(task
);
4225 * Carefully copy the associated workqueue's workfn and name. Keep
4226 * the original last '\0' in case the original contains garbage.
4228 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4229 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4230 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4231 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4233 /* copy worker description */
4234 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4236 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4238 if (fn
|| name
[0] || desc
[0]) {
4239 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4241 pr_cont(" (%s)", desc
);
4246 static void pr_cont_pool_info(struct worker_pool
*pool
)
4248 pr_cont(" cpus=%*pbl", nr_cpumask_bits
, pool
->attrs
->cpumask
);
4249 if (pool
->node
!= NUMA_NO_NODE
)
4250 pr_cont(" node=%d", pool
->node
);
4251 pr_cont(" flags=0x%x nice=%d", pool
->flags
, pool
->attrs
->nice
);
4254 static void pr_cont_work(bool comma
, struct work_struct
*work
)
4256 if (work
->func
== wq_barrier_func
) {
4257 struct wq_barrier
*barr
;
4259 barr
= container_of(work
, struct wq_barrier
, work
);
4261 pr_cont("%s BAR(%d)", comma
? "," : "",
4262 task_pid_nr(barr
->task
));
4264 pr_cont("%s %pf", comma
? "," : "", work
->func
);
4268 static void show_pwq(struct pool_workqueue
*pwq
)
4270 struct worker_pool
*pool
= pwq
->pool
;
4271 struct work_struct
*work
;
4272 struct worker
*worker
;
4273 bool has_in_flight
= false, has_pending
= false;
4276 pr_info(" pwq %d:", pool
->id
);
4277 pr_cont_pool_info(pool
);
4279 pr_cont(" active=%d/%d%s\n", pwq
->nr_active
, pwq
->max_active
,
4280 !list_empty(&pwq
->mayday_node
) ? " MAYDAY" : "");
4282 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4283 if (worker
->current_pwq
== pwq
) {
4284 has_in_flight
= true;
4288 if (has_in_flight
) {
4291 pr_info(" in-flight:");
4292 hash_for_each(pool
->busy_hash
, bkt
, worker
, hentry
) {
4293 if (worker
->current_pwq
!= pwq
)
4296 pr_cont("%s %d%s:%pf", comma
? "," : "",
4297 task_pid_nr(worker
->task
),
4298 worker
== pwq
->wq
->rescuer
? "(RESCUER)" : "",
4299 worker
->current_func
);
4300 list_for_each_entry(work
, &worker
->scheduled
, entry
)
4301 pr_cont_work(false, work
);
4307 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4308 if (get_work_pwq(work
) == pwq
) {
4316 pr_info(" pending:");
4317 list_for_each_entry(work
, &pool
->worklist
, entry
) {
4318 if (get_work_pwq(work
) != pwq
)
4321 pr_cont_work(comma
, work
);
4322 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4327 if (!list_empty(&pwq
->delayed_works
)) {
4330 pr_info(" delayed:");
4331 list_for_each_entry(work
, &pwq
->delayed_works
, entry
) {
4332 pr_cont_work(comma
, work
);
4333 comma
= !(*work_data_bits(work
) & WORK_STRUCT_LINKED
);
4340 * show_workqueue_state - dump workqueue state
4342 * Called from a sysrq handler and prints out all busy workqueues and
4345 void show_workqueue_state(void)
4347 struct workqueue_struct
*wq
;
4348 struct worker_pool
*pool
;
4349 unsigned long flags
;
4352 rcu_read_lock_sched();
4354 pr_info("Showing busy workqueues and worker pools:\n");
4356 list_for_each_entry_rcu(wq
, &workqueues
, list
) {
4357 struct pool_workqueue
*pwq
;
4360 for_each_pwq(pwq
, wq
) {
4361 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
)) {
4369 pr_info("workqueue %s: flags=0x%x\n", wq
->name
, wq
->flags
);
4371 for_each_pwq(pwq
, wq
) {
4372 spin_lock_irqsave(&pwq
->pool
->lock
, flags
);
4373 if (pwq
->nr_active
|| !list_empty(&pwq
->delayed_works
))
4375 spin_unlock_irqrestore(&pwq
->pool
->lock
, flags
);
4379 for_each_pool(pool
, pi
) {
4380 struct worker
*worker
;
4383 spin_lock_irqsave(&pool
->lock
, flags
);
4384 if (pool
->nr_workers
== pool
->nr_idle
)
4387 pr_info("pool %d:", pool
->id
);
4388 pr_cont_pool_info(pool
);
4389 pr_cont(" hung=%us workers=%d",
4390 jiffies_to_msecs(jiffies
- pool
->watchdog_ts
) / 1000,
4393 pr_cont(" manager: %d",
4394 task_pid_nr(pool
->manager
->task
));
4395 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
4396 pr_cont(" %s%d", first
? "idle: " : "",
4397 task_pid_nr(worker
->task
));
4402 spin_unlock_irqrestore(&pool
->lock
, flags
);
4405 rcu_read_unlock_sched();
4411 * There are two challenges in supporting CPU hotplug. Firstly, there
4412 * are a lot of assumptions on strong associations among work, pwq and
4413 * pool which make migrating pending and scheduled works very
4414 * difficult to implement without impacting hot paths. Secondly,
4415 * worker pools serve mix of short, long and very long running works making
4416 * blocked draining impractical.
4418 * This is solved by allowing the pools to be disassociated from the CPU
4419 * running as an unbound one and allowing it to be reattached later if the
4420 * cpu comes back online.
4423 static void wq_unbind_fn(struct work_struct
*work
)
4425 int cpu
= smp_processor_id();
4426 struct worker_pool
*pool
;
4427 struct worker
*worker
;
4429 for_each_cpu_worker_pool(pool
, cpu
) {
4430 mutex_lock(&pool
->attach_mutex
);
4431 spin_lock_irq(&pool
->lock
);
4434 * We've blocked all attach/detach operations. Make all workers
4435 * unbound and set DISASSOCIATED. Before this, all workers
4436 * except for the ones which are still executing works from
4437 * before the last CPU down must be on the cpu. After
4438 * this, they may become diasporas.
4440 for_each_pool_worker(worker
, pool
)
4441 worker
->flags
|= WORKER_UNBOUND
;
4443 pool
->flags
|= POOL_DISASSOCIATED
;
4445 spin_unlock_irq(&pool
->lock
);
4446 mutex_unlock(&pool
->attach_mutex
);
4449 * Call schedule() so that we cross rq->lock and thus can
4450 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4451 * This is necessary as scheduler callbacks may be invoked
4457 * Sched callbacks are disabled now. Zap nr_running.
4458 * After this, nr_running stays zero and need_more_worker()
4459 * and keep_working() are always true as long as the
4460 * worklist is not empty. This pool now behaves as an
4461 * unbound (in terms of concurrency management) pool which
4462 * are served by workers tied to the pool.
4464 atomic_set(&pool
->nr_running
, 0);
4467 * With concurrency management just turned off, a busy
4468 * worker blocking could lead to lengthy stalls. Kick off
4469 * unbound chain execution of currently pending work items.
4471 spin_lock_irq(&pool
->lock
);
4472 wake_up_worker(pool
);
4473 spin_unlock_irq(&pool
->lock
);
4478 * rebind_workers - rebind all workers of a pool to the associated CPU
4479 * @pool: pool of interest
4481 * @pool->cpu is coming online. Rebind all workers to the CPU.
4483 static void rebind_workers(struct worker_pool
*pool
)
4485 struct worker
*worker
;
4487 lockdep_assert_held(&pool
->attach_mutex
);
4490 * Restore CPU affinity of all workers. As all idle workers should
4491 * be on the run-queue of the associated CPU before any local
4492 * wake-ups for concurrency management happen, restore CPU affinity
4493 * of all workers first and then clear UNBOUND. As we're called
4494 * from CPU_ONLINE, the following shouldn't fail.
4496 for_each_pool_worker(worker
, pool
)
4497 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4498 pool
->attrs
->cpumask
) < 0);
4500 spin_lock_irq(&pool
->lock
);
4501 pool
->flags
&= ~POOL_DISASSOCIATED
;
4503 for_each_pool_worker(worker
, pool
) {
4504 unsigned int worker_flags
= worker
->flags
;
4507 * A bound idle worker should actually be on the runqueue
4508 * of the associated CPU for local wake-ups targeting it to
4509 * work. Kick all idle workers so that they migrate to the
4510 * associated CPU. Doing this in the same loop as
4511 * replacing UNBOUND with REBOUND is safe as no worker will
4512 * be bound before @pool->lock is released.
4514 if (worker_flags
& WORKER_IDLE
)
4515 wake_up_process(worker
->task
);
4518 * We want to clear UNBOUND but can't directly call
4519 * worker_clr_flags() or adjust nr_running. Atomically
4520 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4521 * @worker will clear REBOUND using worker_clr_flags() when
4522 * it initiates the next execution cycle thus restoring
4523 * concurrency management. Note that when or whether
4524 * @worker clears REBOUND doesn't affect correctness.
4526 * ACCESS_ONCE() is necessary because @worker->flags may be
4527 * tested without holding any lock in
4528 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4529 * fail incorrectly leading to premature concurrency
4530 * management operations.
4532 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4533 worker_flags
|= WORKER_REBOUND
;
4534 worker_flags
&= ~WORKER_UNBOUND
;
4535 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4538 spin_unlock_irq(&pool
->lock
);
4542 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4543 * @pool: unbound pool of interest
4544 * @cpu: the CPU which is coming up
4546 * An unbound pool may end up with a cpumask which doesn't have any online
4547 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4548 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4549 * online CPU before, cpus_allowed of all its workers should be restored.
4551 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4553 static cpumask_t cpumask
;
4554 struct worker
*worker
;
4556 lockdep_assert_held(&pool
->attach_mutex
);
4558 /* is @cpu allowed for @pool? */
4559 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4562 /* is @cpu the only online CPU? */
4563 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4564 if (cpumask_weight(&cpumask
) != 1)
4567 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4568 for_each_pool_worker(worker
, pool
)
4569 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4570 pool
->attrs
->cpumask
) < 0);
4574 * Workqueues should be brought up before normal priority CPU notifiers.
4575 * This will be registered high priority CPU notifier.
4577 static int workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4578 unsigned long action
,
4581 int cpu
= (unsigned long)hcpu
;
4582 struct worker_pool
*pool
;
4583 struct workqueue_struct
*wq
;
4586 switch (action
& ~CPU_TASKS_FROZEN
) {
4587 case CPU_UP_PREPARE
:
4588 for_each_cpu_worker_pool(pool
, cpu
) {
4589 if (pool
->nr_workers
)
4591 if (!create_worker(pool
))
4596 case CPU_DOWN_FAILED
:
4598 mutex_lock(&wq_pool_mutex
);
4600 for_each_pool(pool
, pi
) {
4601 mutex_lock(&pool
->attach_mutex
);
4603 if (pool
->cpu
== cpu
)
4604 rebind_workers(pool
);
4605 else if (pool
->cpu
< 0)
4606 restore_unbound_workers_cpumask(pool
, cpu
);
4608 mutex_unlock(&pool
->attach_mutex
);
4611 /* update NUMA affinity of unbound workqueues */
4612 list_for_each_entry(wq
, &workqueues
, list
)
4613 wq_update_unbound_numa(wq
, cpu
, true);
4615 mutex_unlock(&wq_pool_mutex
);
4622 * Workqueues should be brought down after normal priority CPU notifiers.
4623 * This will be registered as low priority CPU notifier.
4625 static int workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4626 unsigned long action
,
4629 int cpu
= (unsigned long)hcpu
;
4630 struct work_struct unbind_work
;
4631 struct workqueue_struct
*wq
;
4633 switch (action
& ~CPU_TASKS_FROZEN
) {
4634 case CPU_DOWN_PREPARE
:
4635 /* unbinding per-cpu workers should happen on the local CPU */
4636 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4637 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4639 /* update NUMA affinity of unbound workqueues */
4640 mutex_lock(&wq_pool_mutex
);
4641 list_for_each_entry(wq
, &workqueues
, list
)
4642 wq_update_unbound_numa(wq
, cpu
, false);
4643 mutex_unlock(&wq_pool_mutex
);
4645 /* wait for per-cpu unbinding to finish */
4646 flush_work(&unbind_work
);
4647 destroy_work_on_stack(&unbind_work
);
4655 struct work_for_cpu
{
4656 struct work_struct work
;
4662 static void work_for_cpu_fn(struct work_struct
*work
)
4664 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4666 wfc
->ret
= wfc
->fn(wfc
->arg
);
4670 * work_on_cpu - run a function in user context on a particular cpu
4671 * @cpu: the cpu to run on
4672 * @fn: the function to run
4673 * @arg: the function arg
4675 * It is up to the caller to ensure that the cpu doesn't go offline.
4676 * The caller must not hold any locks which would prevent @fn from completing.
4678 * Return: The value @fn returns.
4680 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4682 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4684 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4685 schedule_work_on(cpu
, &wfc
.work
);
4686 flush_work(&wfc
.work
);
4687 destroy_work_on_stack(&wfc
.work
);
4690 EXPORT_SYMBOL_GPL(work_on_cpu
);
4691 #endif /* CONFIG_SMP */
4693 #ifdef CONFIG_FREEZER
4696 * freeze_workqueues_begin - begin freezing workqueues
4698 * Start freezing workqueues. After this function returns, all freezable
4699 * workqueues will queue new works to their delayed_works list instead of
4703 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4705 void freeze_workqueues_begin(void)
4707 struct workqueue_struct
*wq
;
4708 struct pool_workqueue
*pwq
;
4710 mutex_lock(&wq_pool_mutex
);
4712 WARN_ON_ONCE(workqueue_freezing
);
4713 workqueue_freezing
= true;
4715 list_for_each_entry(wq
, &workqueues
, list
) {
4716 mutex_lock(&wq
->mutex
);
4717 for_each_pwq(pwq
, wq
)
4718 pwq_adjust_max_active(pwq
);
4719 mutex_unlock(&wq
->mutex
);
4722 mutex_unlock(&wq_pool_mutex
);
4726 * freeze_workqueues_busy - are freezable workqueues still busy?
4728 * Check whether freezing is complete. This function must be called
4729 * between freeze_workqueues_begin() and thaw_workqueues().
4732 * Grabs and releases wq_pool_mutex.
4735 * %true if some freezable workqueues are still busy. %false if freezing
4738 bool freeze_workqueues_busy(void)
4741 struct workqueue_struct
*wq
;
4742 struct pool_workqueue
*pwq
;
4744 mutex_lock(&wq_pool_mutex
);
4746 WARN_ON_ONCE(!workqueue_freezing
);
4748 list_for_each_entry(wq
, &workqueues
, list
) {
4749 if (!(wq
->flags
& WQ_FREEZABLE
))
4752 * nr_active is monotonically decreasing. It's safe
4753 * to peek without lock.
4755 rcu_read_lock_sched();
4756 for_each_pwq(pwq
, wq
) {
4757 WARN_ON_ONCE(pwq
->nr_active
< 0);
4758 if (pwq
->nr_active
) {
4760 rcu_read_unlock_sched();
4764 rcu_read_unlock_sched();
4767 mutex_unlock(&wq_pool_mutex
);
4772 * thaw_workqueues - thaw workqueues
4774 * Thaw workqueues. Normal queueing is restored and all collected
4775 * frozen works are transferred to their respective pool worklists.
4778 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4780 void thaw_workqueues(void)
4782 struct workqueue_struct
*wq
;
4783 struct pool_workqueue
*pwq
;
4785 mutex_lock(&wq_pool_mutex
);
4787 if (!workqueue_freezing
)
4790 workqueue_freezing
= false;
4792 /* restore max_active and repopulate worklist */
4793 list_for_each_entry(wq
, &workqueues
, list
) {
4794 mutex_lock(&wq
->mutex
);
4795 for_each_pwq(pwq
, wq
)
4796 pwq_adjust_max_active(pwq
);
4797 mutex_unlock(&wq
->mutex
);
4801 mutex_unlock(&wq_pool_mutex
);
4803 #endif /* CONFIG_FREEZER */
4805 static int workqueue_apply_unbound_cpumask(void)
4809 struct workqueue_struct
*wq
;
4810 struct apply_wqattrs_ctx
*ctx
, *n
;
4812 lockdep_assert_held(&wq_pool_mutex
);
4814 list_for_each_entry(wq
, &workqueues
, list
) {
4815 if (!(wq
->flags
& WQ_UNBOUND
))
4817 /* creating multiple pwqs breaks ordering guarantee */
4818 if (wq
->flags
& __WQ_ORDERED
)
4821 ctx
= apply_wqattrs_prepare(wq
, wq
->unbound_attrs
);
4827 list_add_tail(&ctx
->list
, &ctxs
);
4830 list_for_each_entry_safe(ctx
, n
, &ctxs
, list
) {
4832 apply_wqattrs_commit(ctx
);
4833 apply_wqattrs_cleanup(ctx
);
4840 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4841 * @cpumask: the cpumask to set
4843 * The low-level workqueues cpumask is a global cpumask that limits
4844 * the affinity of all unbound workqueues. This function check the @cpumask
4845 * and apply it to all unbound workqueues and updates all pwqs of them.
4847 * Retun: 0 - Success
4848 * -EINVAL - Invalid @cpumask
4849 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4851 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask
)
4854 cpumask_var_t saved_cpumask
;
4856 if (!zalloc_cpumask_var(&saved_cpumask
, GFP_KERNEL
))
4859 cpumask_and(cpumask
, cpumask
, cpu_possible_mask
);
4860 if (!cpumask_empty(cpumask
)) {
4861 apply_wqattrs_lock();
4863 /* save the old wq_unbound_cpumask. */
4864 cpumask_copy(saved_cpumask
, wq_unbound_cpumask
);
4866 /* update wq_unbound_cpumask at first and apply it to wqs. */
4867 cpumask_copy(wq_unbound_cpumask
, cpumask
);
4868 ret
= workqueue_apply_unbound_cpumask();
4870 /* restore the wq_unbound_cpumask when failed. */
4872 cpumask_copy(wq_unbound_cpumask
, saved_cpumask
);
4874 apply_wqattrs_unlock();
4877 free_cpumask_var(saved_cpumask
);
4883 * Workqueues with WQ_SYSFS flag set is visible to userland via
4884 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4885 * following attributes.
4887 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4888 * max_active RW int : maximum number of in-flight work items
4890 * Unbound workqueues have the following extra attributes.
4892 * id RO int : the associated pool ID
4893 * nice RW int : nice value of the workers
4894 * cpumask RW mask : bitmask of allowed CPUs for the workers
4897 struct workqueue_struct
*wq
;
4901 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
4903 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
4908 static ssize_t
per_cpu_show(struct device
*dev
, struct device_attribute
*attr
,
4911 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4913 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
4915 static DEVICE_ATTR_RO(per_cpu
);
4917 static ssize_t
max_active_show(struct device
*dev
,
4918 struct device_attribute
*attr
, char *buf
)
4920 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4922 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
4925 static ssize_t
max_active_store(struct device
*dev
,
4926 struct device_attribute
*attr
, const char *buf
,
4929 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4932 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
4935 workqueue_set_max_active(wq
, val
);
4938 static DEVICE_ATTR_RW(max_active
);
4940 static struct attribute
*wq_sysfs_attrs
[] = {
4941 &dev_attr_per_cpu
.attr
,
4942 &dev_attr_max_active
.attr
,
4945 ATTRIBUTE_GROUPS(wq_sysfs
);
4947 static ssize_t
wq_pool_ids_show(struct device
*dev
,
4948 struct device_attribute
*attr
, char *buf
)
4950 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4951 const char *delim
= "";
4952 int node
, written
= 0;
4954 rcu_read_lock_sched();
4955 for_each_node(node
) {
4956 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
4957 "%s%d:%d", delim
, node
,
4958 unbound_pwq_by_node(wq
, node
)->pool
->id
);
4961 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
4962 rcu_read_unlock_sched();
4967 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
4970 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4973 mutex_lock(&wq
->mutex
);
4974 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
4975 mutex_unlock(&wq
->mutex
);
4980 /* prepare workqueue_attrs for sysfs store operations */
4981 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
4983 struct workqueue_attrs
*attrs
;
4985 lockdep_assert_held(&wq_pool_mutex
);
4987 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4991 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
4995 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
4996 const char *buf
, size_t count
)
4998 struct workqueue_struct
*wq
= dev_to_wq(dev
);
4999 struct workqueue_attrs
*attrs
;
5002 apply_wqattrs_lock();
5004 attrs
= wq_sysfs_prep_attrs(wq
);
5008 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
5009 attrs
->nice
>= MIN_NICE
&& attrs
->nice
<= MAX_NICE
)
5010 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5015 apply_wqattrs_unlock();
5016 free_workqueue_attrs(attrs
);
5017 return ret
?: count
;
5020 static ssize_t
wq_cpumask_show(struct device
*dev
,
5021 struct device_attribute
*attr
, char *buf
)
5023 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5026 mutex_lock(&wq
->mutex
);
5027 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5028 cpumask_pr_args(wq
->unbound_attrs
->cpumask
));
5029 mutex_unlock(&wq
->mutex
);
5033 static ssize_t
wq_cpumask_store(struct device
*dev
,
5034 struct device_attribute
*attr
,
5035 const char *buf
, size_t count
)
5037 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5038 struct workqueue_attrs
*attrs
;
5041 apply_wqattrs_lock();
5043 attrs
= wq_sysfs_prep_attrs(wq
);
5047 ret
= cpumask_parse(buf
, attrs
->cpumask
);
5049 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5052 apply_wqattrs_unlock();
5053 free_workqueue_attrs(attrs
);
5054 return ret
?: count
;
5057 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
5060 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5063 mutex_lock(&wq
->mutex
);
5064 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
5065 !wq
->unbound_attrs
->no_numa
);
5066 mutex_unlock(&wq
->mutex
);
5071 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
5072 const char *buf
, size_t count
)
5074 struct workqueue_struct
*wq
= dev_to_wq(dev
);
5075 struct workqueue_attrs
*attrs
;
5076 int v
, ret
= -ENOMEM
;
5078 apply_wqattrs_lock();
5080 attrs
= wq_sysfs_prep_attrs(wq
);
5085 if (sscanf(buf
, "%d", &v
) == 1) {
5086 attrs
->no_numa
= !v
;
5087 ret
= apply_workqueue_attrs_locked(wq
, attrs
);
5091 apply_wqattrs_unlock();
5092 free_workqueue_attrs(attrs
);
5093 return ret
?: count
;
5096 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
5097 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
5098 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
5099 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
5100 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
5104 static struct bus_type wq_subsys
= {
5105 .name
= "workqueue",
5106 .dev_groups
= wq_sysfs_groups
,
5109 static ssize_t
wq_unbound_cpumask_show(struct device
*dev
,
5110 struct device_attribute
*attr
, char *buf
)
5114 mutex_lock(&wq_pool_mutex
);
5115 written
= scnprintf(buf
, PAGE_SIZE
, "%*pb\n",
5116 cpumask_pr_args(wq_unbound_cpumask
));
5117 mutex_unlock(&wq_pool_mutex
);
5122 static ssize_t
wq_unbound_cpumask_store(struct device
*dev
,
5123 struct device_attribute
*attr
, const char *buf
, size_t count
)
5125 cpumask_var_t cpumask
;
5128 if (!zalloc_cpumask_var(&cpumask
, GFP_KERNEL
))
5131 ret
= cpumask_parse(buf
, cpumask
);
5133 ret
= workqueue_set_unbound_cpumask(cpumask
);
5135 free_cpumask_var(cpumask
);
5136 return ret
? ret
: count
;
5139 static struct device_attribute wq_sysfs_cpumask_attr
=
5140 __ATTR(cpumask
, 0644, wq_unbound_cpumask_show
,
5141 wq_unbound_cpumask_store
);
5143 static int __init
wq_sysfs_init(void)
5147 err
= subsys_virtual_register(&wq_subsys
, NULL
);
5151 return device_create_file(wq_subsys
.dev_root
, &wq_sysfs_cpumask_attr
);
5153 core_initcall(wq_sysfs_init
);
5155 static void wq_device_release(struct device
*dev
)
5157 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
5163 * workqueue_sysfs_register - make a workqueue visible in sysfs
5164 * @wq: the workqueue to register
5166 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5167 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5168 * which is the preferred method.
5170 * Workqueue user should use this function directly iff it wants to apply
5171 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5172 * apply_workqueue_attrs() may race against userland updating the
5175 * Return: 0 on success, -errno on failure.
5177 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
5179 struct wq_device
*wq_dev
;
5183 * Adjusting max_active or creating new pwqs by applying
5184 * attributes breaks ordering guarantee. Disallow exposing ordered
5187 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
5190 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
5195 wq_dev
->dev
.bus
= &wq_subsys
;
5196 wq_dev
->dev
.init_name
= wq
->name
;
5197 wq_dev
->dev
.release
= wq_device_release
;
5200 * unbound_attrs are created separately. Suppress uevent until
5201 * everything is ready.
5203 dev_set_uevent_suppress(&wq_dev
->dev
, true);
5205 ret
= device_register(&wq_dev
->dev
);
5212 if (wq
->flags
& WQ_UNBOUND
) {
5213 struct device_attribute
*attr
;
5215 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
5216 ret
= device_create_file(&wq_dev
->dev
, attr
);
5218 device_unregister(&wq_dev
->dev
);
5225 dev_set_uevent_suppress(&wq_dev
->dev
, false);
5226 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
5231 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5232 * @wq: the workqueue to unregister
5234 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5236 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
5238 struct wq_device
*wq_dev
= wq
->wq_dev
;
5244 device_unregister(&wq_dev
->dev
);
5246 #else /* CONFIG_SYSFS */
5247 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
5248 #endif /* CONFIG_SYSFS */
5251 * Workqueue watchdog.
5253 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5254 * flush dependency, a concurrency managed work item which stays RUNNING
5255 * indefinitely. Workqueue stalls can be very difficult to debug as the
5256 * usual warning mechanisms don't trigger and internal workqueue state is
5259 * Workqueue watchdog monitors all worker pools periodically and dumps
5260 * state if some pools failed to make forward progress for a while where
5261 * forward progress is defined as the first item on ->worklist changing.
5263 * This mechanism is controlled through the kernel parameter
5264 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5265 * corresponding sysfs parameter file.
5267 #ifdef CONFIG_WQ_WATCHDOG
5269 static void wq_watchdog_timer_fn(unsigned long data
);
5271 static unsigned long wq_watchdog_thresh
= 30;
5272 static struct timer_list wq_watchdog_timer
=
5273 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn
, 0, 0);
5275 static unsigned long wq_watchdog_touched
= INITIAL_JIFFIES
;
5276 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu
) = INITIAL_JIFFIES
;
5278 static void wq_watchdog_reset_touched(void)
5282 wq_watchdog_touched
= jiffies
;
5283 for_each_possible_cpu(cpu
)
5284 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5287 static void wq_watchdog_timer_fn(unsigned long data
)
5289 unsigned long thresh
= READ_ONCE(wq_watchdog_thresh
) * HZ
;
5290 bool lockup_detected
= false;
5291 struct worker_pool
*pool
;
5299 for_each_pool(pool
, pi
) {
5300 unsigned long pool_ts
, touched
, ts
;
5302 if (list_empty(&pool
->worklist
))
5305 /* get the latest of pool and touched timestamps */
5306 pool_ts
= READ_ONCE(pool
->watchdog_ts
);
5307 touched
= READ_ONCE(wq_watchdog_touched
);
5309 if (time_after(pool_ts
, touched
))
5314 if (pool
->cpu
>= 0) {
5315 unsigned long cpu_touched
=
5316 READ_ONCE(per_cpu(wq_watchdog_touched_cpu
,
5318 if (time_after(cpu_touched
, ts
))
5323 if (time_after(jiffies
, ts
+ thresh
)) {
5324 lockup_detected
= true;
5325 pr_emerg("BUG: workqueue lockup - pool");
5326 pr_cont_pool_info(pool
);
5327 pr_cont(" stuck for %us!\n",
5328 jiffies_to_msecs(jiffies
- pool_ts
) / 1000);
5334 if (lockup_detected
)
5335 show_workqueue_state();
5337 wq_watchdog_reset_touched();
5338 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
);
5341 void wq_watchdog_touch(int cpu
)
5344 per_cpu(wq_watchdog_touched_cpu
, cpu
) = jiffies
;
5346 wq_watchdog_touched
= jiffies
;
5349 static void wq_watchdog_set_thresh(unsigned long thresh
)
5351 wq_watchdog_thresh
= 0;
5352 del_timer_sync(&wq_watchdog_timer
);
5355 wq_watchdog_thresh
= thresh
;
5356 wq_watchdog_reset_touched();
5357 mod_timer(&wq_watchdog_timer
, jiffies
+ thresh
* HZ
);
5361 static int wq_watchdog_param_set_thresh(const char *val
,
5362 const struct kernel_param
*kp
)
5364 unsigned long thresh
;
5367 ret
= kstrtoul(val
, 0, &thresh
);
5372 wq_watchdog_set_thresh(thresh
);
5374 wq_watchdog_thresh
= thresh
;
5379 static const struct kernel_param_ops wq_watchdog_thresh_ops
= {
5380 .set
= wq_watchdog_param_set_thresh
,
5381 .get
= param_get_ulong
,
5384 module_param_cb(watchdog_thresh
, &wq_watchdog_thresh_ops
, &wq_watchdog_thresh
,
5387 static void wq_watchdog_init(void)
5389 wq_watchdog_set_thresh(wq_watchdog_thresh
);
5392 #else /* CONFIG_WQ_WATCHDOG */
5394 static inline void wq_watchdog_init(void) { }
5396 #endif /* CONFIG_WQ_WATCHDOG */
5398 static void __init
wq_numa_init(void)
5403 if (num_possible_nodes() <= 1)
5406 if (wq_disable_numa
) {
5407 pr_info("workqueue: NUMA affinity support disabled\n");
5411 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
5412 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5415 * We want masks of possible CPUs of each node which isn't readily
5416 * available. Build one from cpu_to_node() which should have been
5417 * fully initialized by now.
5419 tbl
= kzalloc(nr_node_ids
* sizeof(tbl
[0]), GFP_KERNEL
);
5423 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5424 node_online(node
) ? node
: NUMA_NO_NODE
));
5426 for_each_possible_cpu(cpu
) {
5427 node
= cpu_to_node(cpu
);
5428 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5429 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5430 /* happens iff arch is bonkers, let's just proceed */
5433 cpumask_set_cpu(cpu
, tbl
[node
]);
5436 wq_numa_possible_cpumask
= tbl
;
5437 wq_numa_enabled
= true;
5440 static int __init
init_workqueues(void)
5442 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5445 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5447 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask
, GFP_KERNEL
));
5448 cpumask_copy(wq_unbound_cpumask
, cpu_possible_mask
);
5450 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5452 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
5453 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
5457 /* initialize CPU pools */
5458 for_each_possible_cpu(cpu
) {
5459 struct worker_pool
*pool
;
5462 for_each_cpu_worker_pool(pool
, cpu
) {
5463 BUG_ON(init_worker_pool(pool
));
5465 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5466 pool
->attrs
->nice
= std_nice
[i
++];
5467 pool
->node
= cpu_to_node(cpu
);
5470 mutex_lock(&wq_pool_mutex
);
5471 BUG_ON(worker_pool_assign_id(pool
));
5472 mutex_unlock(&wq_pool_mutex
);
5476 /* create the initial worker */
5477 for_each_online_cpu(cpu
) {
5478 struct worker_pool
*pool
;
5480 for_each_cpu_worker_pool(pool
, cpu
) {
5481 pool
->flags
&= ~POOL_DISASSOCIATED
;
5482 BUG_ON(!create_worker(pool
));
5486 /* create default unbound and ordered wq attrs */
5487 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5488 struct workqueue_attrs
*attrs
;
5490 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5491 attrs
->nice
= std_nice
[i
];
5492 unbound_std_wq_attrs
[i
] = attrs
;
5495 * An ordered wq should have only one pwq as ordering is
5496 * guaranteed by max_active which is enforced by pwqs.
5497 * Turn off NUMA so that dfl_pwq is used for all nodes.
5499 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5500 attrs
->nice
= std_nice
[i
];
5501 attrs
->no_numa
= true;
5502 ordered_wq_attrs
[i
] = attrs
;
5505 system_wq
= alloc_workqueue("events", 0, 0);
5506 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5507 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5508 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5509 WQ_UNBOUND_MAX_ACTIVE
);
5510 system_freezable_wq
= alloc_workqueue("events_freezable",
5512 system_power_efficient_wq
= alloc_workqueue("events_power_efficient",
5513 WQ_POWER_EFFICIENT
, 0);
5514 system_freezable_power_efficient_wq
= alloc_workqueue("events_freezable_power_efficient",
5515 WQ_FREEZABLE
| WQ_POWER_EFFICIENT
,
5517 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5518 !system_unbound_wq
|| !system_freezable_wq
||
5519 !system_power_efficient_wq
||
5520 !system_freezable_power_efficient_wq
);
5526 early_initcall(init_workqueues
);