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 is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
50 #include "workqueue_internal.h"
56 * A bound pool is either associated or disassociated with its CPU.
57 * While associated (!DISASSOCIATED), all workers are bound to the
58 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 * While DISASSOCIATED, the cpu may be offline and all workers have
62 * %WORKER_UNBOUND set and concurrency management disabled, and may
63 * be executing on any CPU. The pool behaves as an unbound one.
65 * Note that DISASSOCIATED should be flipped only while holding
66 * manager_mutex to avoid changing binding state while
67 * create_worker() is in progress.
69 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
70 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
71 POOL_FREEZING
= 1 << 3, /* freeze in progress */
74 WORKER_STARTED
= 1 << 0, /* started */
75 WORKER_DIE
= 1 << 1, /* die die die */
76 WORKER_IDLE
= 1 << 2, /* is idle */
77 WORKER_PREP
= 1 << 3, /* preparing to run works */
78 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
79 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
80 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
82 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
83 WORKER_UNBOUND
| WORKER_REBOUND
,
85 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
87 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
88 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
90 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
91 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
93 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
94 /* call for help after 10ms
96 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
97 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
100 * Rescue workers are used only on emergencies and shared by
101 * all cpus. Give -20.
103 RESCUER_NICE_LEVEL
= -20,
104 HIGHPRI_NICE_LEVEL
= -20,
110 * Structure fields follow one of the following exclusion rules.
112 * I: Modifiable by initialization/destruction paths and read-only for
115 * P: Preemption protected. Disabling preemption is enough and should
116 * only be modified and accessed from the local cpu.
118 * L: pool->lock protected. Access with pool->lock held.
120 * X: During normal operation, modification requires pool->lock and should
121 * be done only from local cpu. Either disabling preemption on local
122 * cpu or grabbing pool->lock is enough for read access. If
123 * POOL_DISASSOCIATED is set, it's identical to L.
125 * MG: pool->manager_mutex and pool->lock protected. Writes require both
126 * locks. Reads can happen under either lock.
128 * PL: wq_pool_mutex protected.
130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
132 * WQ: wq->mutex protected.
134 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
136 * MD: wq_mayday_lock protected.
139 /* struct worker is defined in workqueue_internal.h */
142 spinlock_t lock
; /* the pool lock */
143 int cpu
; /* I: the associated cpu */
144 int node
; /* I: the associated node ID */
145 int id
; /* I: pool ID */
146 unsigned int flags
; /* X: flags */
148 struct list_head worklist
; /* L: list of pending works */
149 int nr_workers
; /* L: total number of workers */
151 /* nr_idle includes the ones off idle_list for rebinding */
152 int nr_idle
; /* L: currently idle ones */
154 struct list_head idle_list
; /* X: list of idle workers */
155 struct timer_list idle_timer
; /* L: worker idle timeout */
156 struct timer_list mayday_timer
; /* L: SOS timer for workers */
158 /* a workers is either on busy_hash or idle_list, or the manager */
159 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
160 /* L: hash of busy workers */
162 /* see manage_workers() for details on the two manager mutexes */
163 struct mutex manager_arb
; /* manager arbitration */
164 struct mutex manager_mutex
; /* manager exclusion */
165 struct idr worker_idr
; /* MG: worker IDs and iteration */
167 struct workqueue_attrs
*attrs
; /* I: worker attributes */
168 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
169 int refcnt
; /* PL: refcnt for unbound pools */
172 * The current concurrency level. As it's likely to be accessed
173 * from other CPUs during try_to_wake_up(), put it in a separate
176 atomic_t nr_running ____cacheline_aligned_in_smp
;
179 * Destruction of pool is sched-RCU protected to allow dereferences
180 * from get_work_pool().
183 } ____cacheline_aligned_in_smp
;
186 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
187 * of work_struct->data are used for flags and the remaining high bits
188 * point to the pwq; thus, pwqs need to be aligned at two's power of the
189 * number of flag bits.
191 struct pool_workqueue
{
192 struct worker_pool
*pool
; /* I: the associated pool */
193 struct workqueue_struct
*wq
; /* I: the owning workqueue */
194 int work_color
; /* L: current color */
195 int flush_color
; /* L: flushing color */
196 int refcnt
; /* L: reference count */
197 int nr_in_flight
[WORK_NR_COLORS
];
198 /* L: nr of in_flight works */
199 int nr_active
; /* L: nr of active works */
200 int max_active
; /* L: max active works */
201 struct list_head delayed_works
; /* L: delayed works */
202 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
203 struct list_head mayday_node
; /* MD: node on wq->maydays */
206 * Release of unbound pwq is punted to system_wq. See put_pwq()
207 * and pwq_unbound_release_workfn() for details. pool_workqueue
208 * itself is also sched-RCU protected so that the first pwq can be
209 * determined without grabbing wq->mutex.
211 struct work_struct unbound_release_work
;
213 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
216 * Structure used to wait for workqueue flush.
219 struct list_head list
; /* WQ: list of flushers */
220 int flush_color
; /* WQ: flush color waiting for */
221 struct completion done
; /* flush completion */
227 * The externally visible workqueue. It relays the issued work items to
228 * the appropriate worker_pool through its pool_workqueues.
230 struct workqueue_struct
{
231 struct list_head pwqs
; /* WR: all pwqs of this wq */
232 struct list_head list
; /* PL: list of all workqueues */
234 struct mutex mutex
; /* protects this wq */
235 int work_color
; /* WQ: current work color */
236 int flush_color
; /* WQ: current flush color */
237 atomic_t nr_pwqs_to_flush
; /* flush in progress */
238 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
239 struct list_head flusher_queue
; /* WQ: flush waiters */
240 struct list_head flusher_overflow
; /* WQ: flush overflow list */
242 struct list_head maydays
; /* MD: pwqs requesting rescue */
243 struct worker
*rescuer
; /* I: rescue worker */
245 int nr_drainers
; /* WQ: drain in progress */
246 int saved_max_active
; /* WQ: saved pwq max_active */
248 struct workqueue_attrs
*unbound_attrs
; /* WQ: only for unbound wqs */
249 struct pool_workqueue
*dfl_pwq
; /* WQ: only for unbound wqs */
252 struct wq_device
*wq_dev
; /* I: for sysfs interface */
254 #ifdef CONFIG_LOCKDEP
255 struct lockdep_map lockdep_map
;
257 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
259 /* hot fields used during command issue, aligned to cacheline */
260 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
261 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
262 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* FR: unbound pwqs indexed by node */
265 static struct kmem_cache
*pwq_cache
;
267 static int wq_numa_tbl_len
; /* highest possible NUMA node id + 1 */
268 static cpumask_var_t
*wq_numa_possible_cpumask
;
269 /* possible CPUs of each node */
271 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
273 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
274 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
276 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
277 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
279 static LIST_HEAD(workqueues
); /* PL: list of all workqueues */
280 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
282 /* the per-cpu worker pools */
283 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
286 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
288 /* PL: hash of all unbound pools keyed by pool->attrs */
289 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
291 /* I: attributes used when instantiating standard unbound pools on demand */
292 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
294 struct workqueue_struct
*system_wq __read_mostly
;
295 EXPORT_SYMBOL_GPL(system_wq
);
296 struct workqueue_struct
*system_highpri_wq __read_mostly
;
297 EXPORT_SYMBOL_GPL(system_highpri_wq
);
298 struct workqueue_struct
*system_long_wq __read_mostly
;
299 EXPORT_SYMBOL_GPL(system_long_wq
);
300 struct workqueue_struct
*system_unbound_wq __read_mostly
;
301 EXPORT_SYMBOL_GPL(system_unbound_wq
);
302 struct workqueue_struct
*system_freezable_wq __read_mostly
;
303 EXPORT_SYMBOL_GPL(system_freezable_wq
);
305 static int worker_thread(void *__worker
);
306 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
307 const struct workqueue_attrs
*from
);
309 #define CREATE_TRACE_POINTS
310 #include <trace/events/workqueue.h>
312 #define assert_rcu_or_pool_mutex() \
313 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
314 lockdep_is_held(&wq_pool_mutex), \
315 "sched RCU or wq_pool_mutex should be held")
317 #define assert_rcu_or_wq_mutex(wq) \
318 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
319 lockdep_is_held(&wq->mutex), \
320 "sched RCU or wq->mutex should be held")
322 #ifdef CONFIG_LOCKDEP
323 #define assert_manager_or_pool_lock(pool) \
324 WARN_ONCE(debug_locks && \
325 !lockdep_is_held(&(pool)->manager_mutex) && \
326 !lockdep_is_held(&(pool)->lock), \
327 "pool->manager_mutex or ->lock should be held")
329 #define assert_manager_or_pool_lock(pool) do { } while (0)
332 #define for_each_cpu_worker_pool(pool, cpu) \
333 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
334 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
338 * for_each_pool - iterate through all worker_pools in the system
339 * @pool: iteration cursor
340 * @pi: integer used for iteration
342 * This must be called either with wq_pool_mutex held or sched RCU read
343 * locked. If the pool needs to be used beyond the locking in effect, the
344 * caller is responsible for guaranteeing that the pool stays online.
346 * The if/else clause exists only for the lockdep assertion and can be
349 #define for_each_pool(pool, pi) \
350 idr_for_each_entry(&worker_pool_idr, pool, pi) \
351 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
355 * for_each_pool_worker - iterate through all workers of a worker_pool
356 * @worker: iteration cursor
357 * @wi: integer used for iteration
358 * @pool: worker_pool to iterate workers of
360 * This must be called with either @pool->manager_mutex or ->lock held.
362 * The if/else clause exists only for the lockdep assertion and can be
365 #define for_each_pool_worker(worker, wi, pool) \
366 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
367 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
371 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
372 * @pwq: iteration cursor
373 * @wq: the target workqueue
375 * This must be called either with wq->mutex held or sched RCU read locked.
376 * If the pwq needs to be used beyond the locking in effect, the caller is
377 * responsible for guaranteeing that the pwq stays online.
379 * The if/else clause exists only for the lockdep assertion and can be
382 #define for_each_pwq(pwq, wq) \
383 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
384 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
387 #ifdef CONFIG_DEBUG_OBJECTS_WORK
389 static struct debug_obj_descr work_debug_descr
;
391 static void *work_debug_hint(void *addr
)
393 return ((struct work_struct
*) addr
)->func
;
397 * fixup_init is called when:
398 * - an active object is initialized
400 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
402 struct work_struct
*work
= addr
;
405 case ODEBUG_STATE_ACTIVE
:
406 cancel_work_sync(work
);
407 debug_object_init(work
, &work_debug_descr
);
415 * fixup_activate is called when:
416 * - an active object is activated
417 * - an unknown object is activated (might be a statically initialized object)
419 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
421 struct work_struct
*work
= addr
;
425 case ODEBUG_STATE_NOTAVAILABLE
:
427 * This is not really a fixup. The work struct was
428 * statically initialized. We just make sure that it
429 * is tracked in the object tracker.
431 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
432 debug_object_init(work
, &work_debug_descr
);
433 debug_object_activate(work
, &work_debug_descr
);
439 case ODEBUG_STATE_ACTIVE
:
448 * fixup_free is called when:
449 * - an active object is freed
451 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
453 struct work_struct
*work
= addr
;
456 case ODEBUG_STATE_ACTIVE
:
457 cancel_work_sync(work
);
458 debug_object_free(work
, &work_debug_descr
);
465 static struct debug_obj_descr work_debug_descr
= {
466 .name
= "work_struct",
467 .debug_hint
= work_debug_hint
,
468 .fixup_init
= work_fixup_init
,
469 .fixup_activate
= work_fixup_activate
,
470 .fixup_free
= work_fixup_free
,
473 static inline void debug_work_activate(struct work_struct
*work
)
475 debug_object_activate(work
, &work_debug_descr
);
478 static inline void debug_work_deactivate(struct work_struct
*work
)
480 debug_object_deactivate(work
, &work_debug_descr
);
483 void __init_work(struct work_struct
*work
, int onstack
)
486 debug_object_init_on_stack(work
, &work_debug_descr
);
488 debug_object_init(work
, &work_debug_descr
);
490 EXPORT_SYMBOL_GPL(__init_work
);
492 void destroy_work_on_stack(struct work_struct
*work
)
494 debug_object_free(work
, &work_debug_descr
);
496 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
499 static inline void debug_work_activate(struct work_struct
*work
) { }
500 static inline void debug_work_deactivate(struct work_struct
*work
) { }
503 /* allocate ID and assign it to @pool */
504 static int worker_pool_assign_id(struct worker_pool
*pool
)
508 lockdep_assert_held(&wq_pool_mutex
);
511 if (!idr_pre_get(&worker_pool_idr
, GFP_KERNEL
))
513 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
514 } while (ret
== -EAGAIN
);
520 * first_pwq - return the first pool_workqueue of the specified workqueue
521 * @wq: the target workqueue
523 * This must be called either with wq->mutex held or sched RCU read locked.
524 * If the pwq needs to be used beyond the locking in effect, the caller is
525 * responsible for guaranteeing that the pwq stays online.
527 static struct pool_workqueue
*first_pwq(struct workqueue_struct
*wq
)
529 assert_rcu_or_wq_mutex(wq
);
530 return list_first_or_null_rcu(&wq
->pwqs
, struct pool_workqueue
,
535 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
536 * @wq: the target workqueue
539 * This must be called either with pwq_lock held or sched RCU read locked.
540 * If the pwq needs to be used beyond the locking in effect, the caller is
541 * responsible for guaranteeing that the pwq stays online.
543 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
546 assert_rcu_or_wq_mutex(wq
);
547 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
550 static unsigned int work_color_to_flags(int color
)
552 return color
<< WORK_STRUCT_COLOR_SHIFT
;
555 static int get_work_color(struct work_struct
*work
)
557 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
558 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
561 static int work_next_color(int color
)
563 return (color
+ 1) % WORK_NR_COLORS
;
567 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
568 * contain the pointer to the queued pwq. Once execution starts, the flag
569 * is cleared and the high bits contain OFFQ flags and pool ID.
571 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
572 * and clear_work_data() can be used to set the pwq, pool or clear
573 * work->data. These functions should only be called while the work is
574 * owned - ie. while the PENDING bit is set.
576 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
577 * corresponding to a work. Pool is available once the work has been
578 * queued anywhere after initialization until it is sync canceled. pwq is
579 * available only while the work item is queued.
581 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
582 * canceled. While being canceled, a work item may have its PENDING set
583 * but stay off timer and worklist for arbitrarily long and nobody should
584 * try to steal the PENDING bit.
586 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
589 WARN_ON_ONCE(!work_pending(work
));
590 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
593 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
594 unsigned long extra_flags
)
596 set_work_data(work
, (unsigned long)pwq
,
597 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
600 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
603 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
604 WORK_STRUCT_PENDING
);
607 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
611 * The following wmb is paired with the implied mb in
612 * test_and_set_bit(PENDING) and ensures all updates to @work made
613 * here are visible to and precede any updates by the next PENDING
617 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
620 static void clear_work_data(struct work_struct
*work
)
622 smp_wmb(); /* see set_work_pool_and_clear_pending() */
623 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
626 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
628 unsigned long data
= atomic_long_read(&work
->data
);
630 if (data
& WORK_STRUCT_PWQ
)
631 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
637 * get_work_pool - return the worker_pool a given work was associated with
638 * @work: the work item of interest
640 * Return the worker_pool @work was last associated with. %NULL if none.
642 * Pools are created and destroyed under wq_pool_mutex, and allows read
643 * access under sched-RCU read lock. As such, this function should be
644 * called under wq_pool_mutex or with preemption disabled.
646 * All fields of the returned pool are accessible as long as the above
647 * mentioned locking is in effect. If the returned pool needs to be used
648 * beyond the critical section, the caller is responsible for ensuring the
649 * returned pool is and stays online.
651 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
653 unsigned long data
= atomic_long_read(&work
->data
);
656 assert_rcu_or_pool_mutex();
658 if (data
& WORK_STRUCT_PWQ
)
659 return ((struct pool_workqueue
*)
660 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
662 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
663 if (pool_id
== WORK_OFFQ_POOL_NONE
)
666 return idr_find(&worker_pool_idr
, pool_id
);
670 * get_work_pool_id - return the worker pool ID a given work is associated with
671 * @work: the work item of interest
673 * Return the worker_pool ID @work was last associated with.
674 * %WORK_OFFQ_POOL_NONE if none.
676 static int get_work_pool_id(struct work_struct
*work
)
678 unsigned long data
= atomic_long_read(&work
->data
);
680 if (data
& WORK_STRUCT_PWQ
)
681 return ((struct pool_workqueue
*)
682 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
684 return data
>> WORK_OFFQ_POOL_SHIFT
;
687 static void mark_work_canceling(struct work_struct
*work
)
689 unsigned long pool_id
= get_work_pool_id(work
);
691 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
692 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
695 static bool work_is_canceling(struct work_struct
*work
)
697 unsigned long data
= atomic_long_read(&work
->data
);
699 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
703 * Policy functions. These define the policies on how the global worker
704 * pools are managed. Unless noted otherwise, these functions assume that
705 * they're being called with pool->lock held.
708 static bool __need_more_worker(struct worker_pool
*pool
)
710 return !atomic_read(&pool
->nr_running
);
714 * Need to wake up a worker? Called from anything but currently
717 * Note that, because unbound workers never contribute to nr_running, this
718 * function will always return %true for unbound pools as long as the
719 * worklist isn't empty.
721 static bool need_more_worker(struct worker_pool
*pool
)
723 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
726 /* Can I start working? Called from busy but !running workers. */
727 static bool may_start_working(struct worker_pool
*pool
)
729 return pool
->nr_idle
;
732 /* Do I need to keep working? Called from currently running workers. */
733 static bool keep_working(struct worker_pool
*pool
)
735 return !list_empty(&pool
->worklist
) &&
736 atomic_read(&pool
->nr_running
) <= 1;
739 /* Do we need a new worker? Called from manager. */
740 static bool need_to_create_worker(struct worker_pool
*pool
)
742 return need_more_worker(pool
) && !may_start_working(pool
);
745 /* Do I need to be the manager? */
746 static bool need_to_manage_workers(struct worker_pool
*pool
)
748 return need_to_create_worker(pool
) ||
749 (pool
->flags
& POOL_MANAGE_WORKERS
);
752 /* Do we have too many workers and should some go away? */
753 static bool too_many_workers(struct worker_pool
*pool
)
755 bool managing
= mutex_is_locked(&pool
->manager_arb
);
756 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
757 int nr_busy
= pool
->nr_workers
- nr_idle
;
760 * nr_idle and idle_list may disagree if idle rebinding is in
761 * progress. Never return %true if idle_list is empty.
763 if (list_empty(&pool
->idle_list
))
766 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
773 /* Return the first worker. Safe with preemption disabled */
774 static struct worker
*first_worker(struct worker_pool
*pool
)
776 if (unlikely(list_empty(&pool
->idle_list
)))
779 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
783 * wake_up_worker - wake up an idle worker
784 * @pool: worker pool to wake worker from
786 * Wake up the first idle worker of @pool.
789 * spin_lock_irq(pool->lock).
791 static void wake_up_worker(struct worker_pool
*pool
)
793 struct worker
*worker
= first_worker(pool
);
796 wake_up_process(worker
->task
);
800 * wq_worker_waking_up - a worker is waking up
801 * @task: task waking up
802 * @cpu: CPU @task is waking up to
804 * This function is called during try_to_wake_up() when a worker is
808 * spin_lock_irq(rq->lock)
810 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
812 struct worker
*worker
= kthread_data(task
);
814 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
815 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
816 atomic_inc(&worker
->pool
->nr_running
);
821 * wq_worker_sleeping - a worker is going to sleep
822 * @task: task going to sleep
823 * @cpu: CPU in question, must be the current CPU number
825 * This function is called during schedule() when a busy worker is
826 * going to sleep. Worker on the same cpu can be woken up by
827 * returning pointer to its task.
830 * spin_lock_irq(rq->lock)
833 * Worker task on @cpu to wake up, %NULL if none.
835 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
837 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
838 struct worker_pool
*pool
;
841 * Rescuers, which may not have all the fields set up like normal
842 * workers, also reach here, let's not access anything before
843 * checking NOT_RUNNING.
845 if (worker
->flags
& WORKER_NOT_RUNNING
)
850 /* this can only happen on the local cpu */
851 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
855 * The counterpart of the following dec_and_test, implied mb,
856 * worklist not empty test sequence is in insert_work().
857 * Please read comment there.
859 * NOT_RUNNING is clear. This means that we're bound to and
860 * running on the local cpu w/ rq lock held and preemption
861 * disabled, which in turn means that none else could be
862 * manipulating idle_list, so dereferencing idle_list without pool
865 if (atomic_dec_and_test(&pool
->nr_running
) &&
866 !list_empty(&pool
->worklist
))
867 to_wakeup
= first_worker(pool
);
868 return to_wakeup
? to_wakeup
->task
: NULL
;
872 * worker_set_flags - set worker flags and adjust nr_running accordingly
874 * @flags: flags to set
875 * @wakeup: wakeup an idle worker if necessary
877 * Set @flags in @worker->flags and adjust nr_running accordingly. If
878 * nr_running becomes zero and @wakeup is %true, an idle worker is
882 * spin_lock_irq(pool->lock)
884 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
887 struct worker_pool
*pool
= worker
->pool
;
889 WARN_ON_ONCE(worker
->task
!= current
);
892 * If transitioning into NOT_RUNNING, adjust nr_running and
893 * wake up an idle worker as necessary if requested by
896 if ((flags
& WORKER_NOT_RUNNING
) &&
897 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
899 if (atomic_dec_and_test(&pool
->nr_running
) &&
900 !list_empty(&pool
->worklist
))
901 wake_up_worker(pool
);
903 atomic_dec(&pool
->nr_running
);
906 worker
->flags
|= flags
;
910 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
912 * @flags: flags to clear
914 * Clear @flags in @worker->flags and adjust nr_running accordingly.
917 * spin_lock_irq(pool->lock)
919 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
921 struct worker_pool
*pool
= worker
->pool
;
922 unsigned int oflags
= worker
->flags
;
924 WARN_ON_ONCE(worker
->task
!= current
);
926 worker
->flags
&= ~flags
;
929 * If transitioning out of NOT_RUNNING, increment nr_running. Note
930 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
931 * of multiple flags, not a single flag.
933 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
934 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
935 atomic_inc(&pool
->nr_running
);
939 * find_worker_executing_work - find worker which is executing a work
940 * @pool: pool of interest
941 * @work: work to find worker for
943 * Find a worker which is executing @work on @pool by searching
944 * @pool->busy_hash which is keyed by the address of @work. For a worker
945 * to match, its current execution should match the address of @work and
946 * its work function. This is to avoid unwanted dependency between
947 * unrelated work executions through a work item being recycled while still
950 * This is a bit tricky. A work item may be freed once its execution
951 * starts and nothing prevents the freed area from being recycled for
952 * another work item. If the same work item address ends up being reused
953 * before the original execution finishes, workqueue will identify the
954 * recycled work item as currently executing and make it wait until the
955 * current execution finishes, introducing an unwanted dependency.
957 * This function checks the work item address and work function to avoid
958 * false positives. Note that this isn't complete as one may construct a
959 * work function which can introduce dependency onto itself through a
960 * recycled work item. Well, if somebody wants to shoot oneself in the
961 * foot that badly, there's only so much we can do, and if such deadlock
962 * actually occurs, it should be easy to locate the culprit work function.
965 * spin_lock_irq(pool->lock).
968 * Pointer to worker which is executing @work if found, NULL
971 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
972 struct work_struct
*work
)
974 struct worker
*worker
;
976 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
978 if (worker
->current_work
== work
&&
979 worker
->current_func
== work
->func
)
986 * move_linked_works - move linked works to a list
987 * @work: start of series of works to be scheduled
988 * @head: target list to append @work to
989 * @nextp: out paramter for nested worklist walking
991 * Schedule linked works starting from @work to @head. Work series to
992 * be scheduled starts at @work and includes any consecutive work with
993 * WORK_STRUCT_LINKED set in its predecessor.
995 * If @nextp is not NULL, it's updated to point to the next work of
996 * the last scheduled work. This allows move_linked_works() to be
997 * nested inside outer list_for_each_entry_safe().
1000 * spin_lock_irq(pool->lock).
1002 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1003 struct work_struct
**nextp
)
1005 struct work_struct
*n
;
1008 * Linked worklist will always end before the end of the list,
1009 * use NULL for list head.
1011 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1012 list_move_tail(&work
->entry
, head
);
1013 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1018 * If we're already inside safe list traversal and have moved
1019 * multiple works to the scheduled queue, the next position
1020 * needs to be updated.
1027 * get_pwq - get an extra reference on the specified pool_workqueue
1028 * @pwq: pool_workqueue to get
1030 * Obtain an extra reference on @pwq. The caller should guarantee that
1031 * @pwq has positive refcnt and be holding the matching pool->lock.
1033 static void get_pwq(struct pool_workqueue
*pwq
)
1035 lockdep_assert_held(&pwq
->pool
->lock
);
1036 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1041 * put_pwq - put a pool_workqueue reference
1042 * @pwq: pool_workqueue to put
1044 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1045 * destruction. The caller should be holding the matching pool->lock.
1047 static void put_pwq(struct pool_workqueue
*pwq
)
1049 lockdep_assert_held(&pwq
->pool
->lock
);
1050 if (likely(--pwq
->refcnt
))
1052 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1055 * @pwq can't be released under pool->lock, bounce to
1056 * pwq_unbound_release_workfn(). This never recurses on the same
1057 * pool->lock as this path is taken only for unbound workqueues and
1058 * the release work item is scheduled on a per-cpu workqueue. To
1059 * avoid lockdep warning, unbound pool->locks are given lockdep
1060 * subclass of 1 in get_unbound_pool().
1062 schedule_work(&pwq
->unbound_release_work
);
1066 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1067 * @pwq: pool_workqueue to put (can be %NULL)
1069 * put_pwq() with locking. This function also allows %NULL @pwq.
1071 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1075 * As both pwqs and pools are sched-RCU protected, the
1076 * following lock operations are safe.
1078 spin_lock_irq(&pwq
->pool
->lock
);
1080 spin_unlock_irq(&pwq
->pool
->lock
);
1084 static void pwq_activate_delayed_work(struct work_struct
*work
)
1086 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1088 trace_workqueue_activate_work(work
);
1089 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1090 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1094 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1096 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1097 struct work_struct
, entry
);
1099 pwq_activate_delayed_work(work
);
1103 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1104 * @pwq: pwq of interest
1105 * @color: color of work which left the queue
1107 * A work either has completed or is removed from pending queue,
1108 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1111 * spin_lock_irq(pool->lock).
1113 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1115 /* uncolored work items don't participate in flushing or nr_active */
1116 if (color
== WORK_NO_COLOR
)
1119 pwq
->nr_in_flight
[color
]--;
1122 if (!list_empty(&pwq
->delayed_works
)) {
1123 /* one down, submit a delayed one */
1124 if (pwq
->nr_active
< pwq
->max_active
)
1125 pwq_activate_first_delayed(pwq
);
1128 /* is flush in progress and are we at the flushing tip? */
1129 if (likely(pwq
->flush_color
!= color
))
1132 /* are there still in-flight works? */
1133 if (pwq
->nr_in_flight
[color
])
1136 /* this pwq is done, clear flush_color */
1137 pwq
->flush_color
= -1;
1140 * If this was the last pwq, wake up the first flusher. It
1141 * will handle the rest.
1143 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1144 complete(&pwq
->wq
->first_flusher
->done
);
1150 * try_to_grab_pending - steal work item from worklist and disable irq
1151 * @work: work item to steal
1152 * @is_dwork: @work is a delayed_work
1153 * @flags: place to store irq state
1155 * Try to grab PENDING bit of @work. This function can handle @work in any
1156 * stable state - idle, on timer or on worklist. Return values are
1158 * 1 if @work was pending and we successfully stole PENDING
1159 * 0 if @work was idle and we claimed PENDING
1160 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1161 * -ENOENT if someone else is canceling @work, this state may persist
1162 * for arbitrarily long
1164 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1165 * interrupted while holding PENDING and @work off queue, irq must be
1166 * disabled on entry. This, combined with delayed_work->timer being
1167 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1169 * On successful return, >= 0, irq is disabled and the caller is
1170 * responsible for releasing it using local_irq_restore(*@flags).
1172 * This function is safe to call from any context including IRQ handler.
1174 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1175 unsigned long *flags
)
1177 struct worker_pool
*pool
;
1178 struct pool_workqueue
*pwq
;
1180 local_irq_save(*flags
);
1182 /* try to steal the timer if it exists */
1184 struct delayed_work
*dwork
= to_delayed_work(work
);
1187 * dwork->timer is irqsafe. If del_timer() fails, it's
1188 * guaranteed that the timer is not queued anywhere and not
1189 * running on the local CPU.
1191 if (likely(del_timer(&dwork
->timer
)))
1195 /* try to claim PENDING the normal way */
1196 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1200 * The queueing is in progress, or it is already queued. Try to
1201 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1203 pool
= get_work_pool(work
);
1207 spin_lock(&pool
->lock
);
1209 * work->data is guaranteed to point to pwq only while the work
1210 * item is queued on pwq->wq, and both updating work->data to point
1211 * to pwq on queueing and to pool on dequeueing are done under
1212 * pwq->pool->lock. This in turn guarantees that, if work->data
1213 * points to pwq which is associated with a locked pool, the work
1214 * item is currently queued on that pool.
1216 pwq
= get_work_pwq(work
);
1217 if (pwq
&& pwq
->pool
== pool
) {
1218 debug_work_deactivate(work
);
1221 * A delayed work item cannot be grabbed directly because
1222 * it might have linked NO_COLOR work items which, if left
1223 * on the delayed_list, will confuse pwq->nr_active
1224 * management later on and cause stall. Make sure the work
1225 * item is activated before grabbing.
1227 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1228 pwq_activate_delayed_work(work
);
1230 list_del_init(&work
->entry
);
1231 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1233 /* work->data points to pwq iff queued, point to pool */
1234 set_work_pool_and_keep_pending(work
, pool
->id
);
1236 spin_unlock(&pool
->lock
);
1239 spin_unlock(&pool
->lock
);
1241 local_irq_restore(*flags
);
1242 if (work_is_canceling(work
))
1249 * insert_work - insert a work into a pool
1250 * @pwq: pwq @work belongs to
1251 * @work: work to insert
1252 * @head: insertion point
1253 * @extra_flags: extra WORK_STRUCT_* flags to set
1255 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1256 * work_struct flags.
1259 * spin_lock_irq(pool->lock).
1261 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1262 struct list_head
*head
, unsigned int extra_flags
)
1264 struct worker_pool
*pool
= pwq
->pool
;
1266 /* we own @work, set data and link */
1267 set_work_pwq(work
, pwq
, extra_flags
);
1268 list_add_tail(&work
->entry
, head
);
1272 * Ensure either wq_worker_sleeping() sees the above
1273 * list_add_tail() or we see zero nr_running to avoid workers lying
1274 * around lazily while there are works to be processed.
1278 if (__need_more_worker(pool
))
1279 wake_up_worker(pool
);
1283 * Test whether @work is being queued from another work executing on the
1286 static bool is_chained_work(struct workqueue_struct
*wq
)
1288 struct worker
*worker
;
1290 worker
= current_wq_worker();
1292 * Return %true iff I'm a worker execuing a work item on @wq. If
1293 * I'm @worker, it's safe to dereference it without locking.
1295 return worker
&& worker
->current_pwq
->wq
== wq
;
1298 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1299 struct work_struct
*work
)
1301 struct pool_workqueue
*pwq
;
1302 struct worker_pool
*last_pool
;
1303 struct list_head
*worklist
;
1304 unsigned int work_flags
;
1305 unsigned int req_cpu
= cpu
;
1308 * While a work item is PENDING && off queue, a task trying to
1309 * steal the PENDING will busy-loop waiting for it to either get
1310 * queued or lose PENDING. Grabbing PENDING and queueing should
1311 * happen with IRQ disabled.
1313 WARN_ON_ONCE(!irqs_disabled());
1315 debug_work_activate(work
);
1317 /* if dying, only works from the same workqueue are allowed */
1318 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1319 WARN_ON_ONCE(!is_chained_work(wq
)))
1322 if (req_cpu
== WORK_CPU_UNBOUND
)
1323 cpu
= raw_smp_processor_id();
1325 /* pwq which will be used unless @work is executing elsewhere */
1326 if (!(wq
->flags
& WQ_UNBOUND
))
1327 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1329 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1332 * If @work was previously on a different pool, it might still be
1333 * running there, in which case the work needs to be queued on that
1334 * pool to guarantee non-reentrancy.
1336 last_pool
= get_work_pool(work
);
1337 if (last_pool
&& last_pool
!= pwq
->pool
) {
1338 struct worker
*worker
;
1340 spin_lock(&last_pool
->lock
);
1342 worker
= find_worker_executing_work(last_pool
, work
);
1344 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1345 pwq
= worker
->current_pwq
;
1347 /* meh... not running there, queue here */
1348 spin_unlock(&last_pool
->lock
);
1349 spin_lock(&pwq
->pool
->lock
);
1352 spin_lock(&pwq
->pool
->lock
);
1356 * pwq is determined and locked. For unbound pools, we could have
1357 * raced with pwq release and it could already be dead. If its
1358 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1359 * without another pwq replacing it in the numa_pwq_tbl or while
1360 * work items are executing on it, so the retrying is guaranteed to
1361 * make forward-progress.
1363 if (unlikely(!pwq
->refcnt
)) {
1364 if (wq
->flags
& WQ_UNBOUND
) {
1365 spin_unlock(&pwq
->pool
->lock
);
1370 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1374 /* pwq determined, queue */
1375 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1377 if (WARN_ON(!list_empty(&work
->entry
))) {
1378 spin_unlock(&pwq
->pool
->lock
);
1382 pwq
->nr_in_flight
[pwq
->work_color
]++;
1383 work_flags
= work_color_to_flags(pwq
->work_color
);
1385 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1386 trace_workqueue_activate_work(work
);
1388 worklist
= &pwq
->pool
->worklist
;
1390 work_flags
|= WORK_STRUCT_DELAYED
;
1391 worklist
= &pwq
->delayed_works
;
1394 insert_work(pwq
, work
, worklist
, work_flags
);
1396 spin_unlock(&pwq
->pool
->lock
);
1400 * queue_work_on - queue work on specific cpu
1401 * @cpu: CPU number to execute work on
1402 * @wq: workqueue to use
1403 * @work: work to queue
1405 * Returns %false if @work was already on a queue, %true otherwise.
1407 * We queue the work to a specific CPU, the caller must ensure it
1410 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1411 struct work_struct
*work
)
1414 unsigned long flags
;
1416 local_irq_save(flags
);
1418 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1419 __queue_work(cpu
, wq
, work
);
1423 local_irq_restore(flags
);
1426 EXPORT_SYMBOL_GPL(queue_work_on
);
1428 void delayed_work_timer_fn(unsigned long __data
)
1430 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1432 /* should have been called from irqsafe timer with irq already off */
1433 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1435 EXPORT_SYMBOL(delayed_work_timer_fn
);
1437 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1438 struct delayed_work
*dwork
, unsigned long delay
)
1440 struct timer_list
*timer
= &dwork
->timer
;
1441 struct work_struct
*work
= &dwork
->work
;
1443 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1444 timer
->data
!= (unsigned long)dwork
);
1445 WARN_ON_ONCE(timer_pending(timer
));
1446 WARN_ON_ONCE(!list_empty(&work
->entry
));
1449 * If @delay is 0, queue @dwork->work immediately. This is for
1450 * both optimization and correctness. The earliest @timer can
1451 * expire is on the closest next tick and delayed_work users depend
1452 * on that there's no such delay when @delay is 0.
1455 __queue_work(cpu
, wq
, &dwork
->work
);
1459 timer_stats_timer_set_start_info(&dwork
->timer
);
1463 timer
->expires
= jiffies
+ delay
;
1465 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1466 add_timer_on(timer
, cpu
);
1472 * queue_delayed_work_on - queue work on specific CPU after delay
1473 * @cpu: CPU number to execute work on
1474 * @wq: workqueue to use
1475 * @dwork: work to queue
1476 * @delay: number of jiffies to wait before queueing
1478 * Returns %false if @work was already on a queue, %true otherwise. If
1479 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1482 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1483 struct delayed_work
*dwork
, unsigned long delay
)
1485 struct work_struct
*work
= &dwork
->work
;
1487 unsigned long flags
;
1489 /* read the comment in __queue_work() */
1490 local_irq_save(flags
);
1492 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1493 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1497 local_irq_restore(flags
);
1500 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1503 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1504 * @cpu: CPU number to execute work on
1505 * @wq: workqueue to use
1506 * @dwork: work to queue
1507 * @delay: number of jiffies to wait before queueing
1509 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1510 * modify @dwork's timer so that it expires after @delay. If @delay is
1511 * zero, @work is guaranteed to be scheduled immediately regardless of its
1514 * Returns %false if @dwork was idle and queued, %true if @dwork was
1515 * pending and its timer was modified.
1517 * This function is safe to call from any context including IRQ handler.
1518 * See try_to_grab_pending() for details.
1520 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1521 struct delayed_work
*dwork
, unsigned long delay
)
1523 unsigned long flags
;
1527 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1528 } while (unlikely(ret
== -EAGAIN
));
1530 if (likely(ret
>= 0)) {
1531 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1532 local_irq_restore(flags
);
1535 /* -ENOENT from try_to_grab_pending() becomes %true */
1538 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1541 * worker_enter_idle - enter idle state
1542 * @worker: worker which is entering idle state
1544 * @worker is entering idle state. Update stats and idle timer if
1548 * spin_lock_irq(pool->lock).
1550 static void worker_enter_idle(struct worker
*worker
)
1552 struct worker_pool
*pool
= worker
->pool
;
1554 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1555 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1556 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1559 /* can't use worker_set_flags(), also called from start_worker() */
1560 worker
->flags
|= WORKER_IDLE
;
1562 worker
->last_active
= jiffies
;
1564 /* idle_list is LIFO */
1565 list_add(&worker
->entry
, &pool
->idle_list
);
1567 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1568 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1571 * Sanity check nr_running. Because wq_unbind_fn() releases
1572 * pool->lock between setting %WORKER_UNBOUND and zapping
1573 * nr_running, the warning may trigger spuriously. Check iff
1574 * unbind is not in progress.
1576 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1577 pool
->nr_workers
== pool
->nr_idle
&&
1578 atomic_read(&pool
->nr_running
));
1582 * worker_leave_idle - leave idle state
1583 * @worker: worker which is leaving idle state
1585 * @worker is leaving idle state. Update stats.
1588 * spin_lock_irq(pool->lock).
1590 static void worker_leave_idle(struct worker
*worker
)
1592 struct worker_pool
*pool
= worker
->pool
;
1594 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1596 worker_clr_flags(worker
, WORKER_IDLE
);
1598 list_del_init(&worker
->entry
);
1602 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1603 * @pool: target worker_pool
1605 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1607 * Works which are scheduled while the cpu is online must at least be
1608 * scheduled to a worker which is bound to the cpu so that if they are
1609 * flushed from cpu callbacks while cpu is going down, they are
1610 * guaranteed to execute on the cpu.
1612 * This function is to be used by unbound workers and rescuers to bind
1613 * themselves to the target cpu and may race with cpu going down or
1614 * coming online. kthread_bind() can't be used because it may put the
1615 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1616 * verbatim as it's best effort and blocking and pool may be
1617 * [dis]associated in the meantime.
1619 * This function tries set_cpus_allowed() and locks pool and verifies the
1620 * binding against %POOL_DISASSOCIATED which is set during
1621 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1622 * enters idle state or fetches works without dropping lock, it can
1623 * guarantee the scheduling requirement described in the first paragraph.
1626 * Might sleep. Called without any lock but returns with pool->lock
1630 * %true if the associated pool is online (@worker is successfully
1631 * bound), %false if offline.
1633 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1634 __acquires(&pool
->lock
)
1638 * The following call may fail, succeed or succeed
1639 * without actually migrating the task to the cpu if
1640 * it races with cpu hotunplug operation. Verify
1641 * against POOL_DISASSOCIATED.
1643 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1644 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1646 spin_lock_irq(&pool
->lock
);
1647 if (pool
->flags
& POOL_DISASSOCIATED
)
1649 if (task_cpu(current
) == pool
->cpu
&&
1650 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1652 spin_unlock_irq(&pool
->lock
);
1655 * We've raced with CPU hot[un]plug. Give it a breather
1656 * and retry migration. cond_resched() is required here;
1657 * otherwise, we might deadlock against cpu_stop trying to
1658 * bring down the CPU on non-preemptive kernel.
1665 static struct worker
*alloc_worker(void)
1667 struct worker
*worker
;
1669 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1671 INIT_LIST_HEAD(&worker
->entry
);
1672 INIT_LIST_HEAD(&worker
->scheduled
);
1673 /* on creation a worker is in !idle && prep state */
1674 worker
->flags
= WORKER_PREP
;
1680 * create_worker - create a new workqueue worker
1681 * @pool: pool the new worker will belong to
1683 * Create a new worker which is bound to @pool. The returned worker
1684 * can be started by calling start_worker() or destroyed using
1688 * Might sleep. Does GFP_KERNEL allocations.
1691 * Pointer to the newly created worker.
1693 static struct worker
*create_worker(struct worker_pool
*pool
)
1695 struct worker
*worker
= NULL
;
1699 lockdep_assert_held(&pool
->manager_mutex
);
1702 * ID is needed to determine kthread name. Allocate ID first
1703 * without installing the pointer.
1705 idr_preload(GFP_KERNEL
);
1706 spin_lock_irq(&pool
->lock
);
1708 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1710 spin_unlock_irq(&pool
->lock
);
1715 worker
= alloc_worker();
1719 worker
->pool
= pool
;
1723 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1724 pool
->attrs
->nice
< 0 ? "H" : "");
1726 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1728 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1729 "kworker/%s", id_buf
);
1730 if (IS_ERR(worker
->task
))
1734 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1735 * online CPUs. It'll be re-applied when any of the CPUs come up.
1737 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1738 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1740 /* prevent userland from meddling with cpumask of workqueue workers */
1741 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1744 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1745 * remains stable across this function. See the comments above the
1746 * flag definition for details.
1748 if (pool
->flags
& POOL_DISASSOCIATED
)
1749 worker
->flags
|= WORKER_UNBOUND
;
1751 /* successful, commit the pointer to idr */
1752 spin_lock_irq(&pool
->lock
);
1753 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1754 spin_unlock_irq(&pool
->lock
);
1760 spin_lock_irq(&pool
->lock
);
1761 idr_remove(&pool
->worker_idr
, id
);
1762 spin_unlock_irq(&pool
->lock
);
1769 * start_worker - start a newly created worker
1770 * @worker: worker to start
1772 * Make the pool aware of @worker and start it.
1775 * spin_lock_irq(pool->lock).
1777 static void start_worker(struct worker
*worker
)
1779 worker
->flags
|= WORKER_STARTED
;
1780 worker
->pool
->nr_workers
++;
1781 worker_enter_idle(worker
);
1782 wake_up_process(worker
->task
);
1786 * create_and_start_worker - create and start a worker for a pool
1787 * @pool: the target pool
1789 * Grab the managership of @pool and create and start a new worker for it.
1791 static int create_and_start_worker(struct worker_pool
*pool
)
1793 struct worker
*worker
;
1795 mutex_lock(&pool
->manager_mutex
);
1797 worker
= create_worker(pool
);
1799 spin_lock_irq(&pool
->lock
);
1800 start_worker(worker
);
1801 spin_unlock_irq(&pool
->lock
);
1804 mutex_unlock(&pool
->manager_mutex
);
1806 return worker
? 0 : -ENOMEM
;
1810 * destroy_worker - destroy a workqueue worker
1811 * @worker: worker to be destroyed
1813 * Destroy @worker and adjust @pool stats accordingly.
1816 * spin_lock_irq(pool->lock) which is released and regrabbed.
1818 static void destroy_worker(struct worker
*worker
)
1820 struct worker_pool
*pool
= worker
->pool
;
1822 lockdep_assert_held(&pool
->manager_mutex
);
1823 lockdep_assert_held(&pool
->lock
);
1825 /* sanity check frenzy */
1826 if (WARN_ON(worker
->current_work
) ||
1827 WARN_ON(!list_empty(&worker
->scheduled
)))
1830 if (worker
->flags
& WORKER_STARTED
)
1832 if (worker
->flags
& WORKER_IDLE
)
1835 list_del_init(&worker
->entry
);
1836 worker
->flags
|= WORKER_DIE
;
1838 idr_remove(&pool
->worker_idr
, worker
->id
);
1840 spin_unlock_irq(&pool
->lock
);
1842 kthread_stop(worker
->task
);
1845 spin_lock_irq(&pool
->lock
);
1848 static void idle_worker_timeout(unsigned long __pool
)
1850 struct worker_pool
*pool
= (void *)__pool
;
1852 spin_lock_irq(&pool
->lock
);
1854 if (too_many_workers(pool
)) {
1855 struct worker
*worker
;
1856 unsigned long expires
;
1858 /* idle_list is kept in LIFO order, check the last one */
1859 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1860 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1862 if (time_before(jiffies
, expires
))
1863 mod_timer(&pool
->idle_timer
, expires
);
1865 /* it's been idle for too long, wake up manager */
1866 pool
->flags
|= POOL_MANAGE_WORKERS
;
1867 wake_up_worker(pool
);
1871 spin_unlock_irq(&pool
->lock
);
1874 static void send_mayday(struct work_struct
*work
)
1876 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1877 struct workqueue_struct
*wq
= pwq
->wq
;
1879 lockdep_assert_held(&wq_mayday_lock
);
1884 /* mayday mayday mayday */
1885 if (list_empty(&pwq
->mayday_node
)) {
1886 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1887 wake_up_process(wq
->rescuer
->task
);
1891 static void pool_mayday_timeout(unsigned long __pool
)
1893 struct worker_pool
*pool
= (void *)__pool
;
1894 struct work_struct
*work
;
1896 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1897 spin_lock(&pool
->lock
);
1899 if (need_to_create_worker(pool
)) {
1901 * We've been trying to create a new worker but
1902 * haven't been successful. We might be hitting an
1903 * allocation deadlock. Send distress signals to
1906 list_for_each_entry(work
, &pool
->worklist
, entry
)
1910 spin_unlock(&pool
->lock
);
1911 spin_unlock_irq(&wq_mayday_lock
);
1913 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1917 * maybe_create_worker - create a new worker if necessary
1918 * @pool: pool to create a new worker for
1920 * Create a new worker for @pool if necessary. @pool is guaranteed to
1921 * have at least one idle worker on return from this function. If
1922 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1923 * sent to all rescuers with works scheduled on @pool to resolve
1924 * possible allocation deadlock.
1926 * On return, need_to_create_worker() is guaranteed to be %false and
1927 * may_start_working() %true.
1930 * spin_lock_irq(pool->lock) which may be released and regrabbed
1931 * multiple times. Does GFP_KERNEL allocations. Called only from
1935 * %false if no action was taken and pool->lock stayed locked, %true
1938 static bool maybe_create_worker(struct worker_pool
*pool
)
1939 __releases(&pool
->lock
)
1940 __acquires(&pool
->lock
)
1942 if (!need_to_create_worker(pool
))
1945 spin_unlock_irq(&pool
->lock
);
1947 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1948 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1951 struct worker
*worker
;
1953 worker
= create_worker(pool
);
1955 del_timer_sync(&pool
->mayday_timer
);
1956 spin_lock_irq(&pool
->lock
);
1957 start_worker(worker
);
1958 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1963 if (!need_to_create_worker(pool
))
1966 __set_current_state(TASK_INTERRUPTIBLE
);
1967 schedule_timeout(CREATE_COOLDOWN
);
1969 if (!need_to_create_worker(pool
))
1973 del_timer_sync(&pool
->mayday_timer
);
1974 spin_lock_irq(&pool
->lock
);
1975 if (need_to_create_worker(pool
))
1981 * maybe_destroy_worker - destroy workers which have been idle for a while
1982 * @pool: pool to destroy workers for
1984 * Destroy @pool workers which have been idle for longer than
1985 * IDLE_WORKER_TIMEOUT.
1988 * spin_lock_irq(pool->lock) which may be released and regrabbed
1989 * multiple times. Called only from manager.
1992 * %false if no action was taken and pool->lock stayed locked, %true
1995 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1999 while (too_many_workers(pool
)) {
2000 struct worker
*worker
;
2001 unsigned long expires
;
2003 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2004 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2006 if (time_before(jiffies
, expires
)) {
2007 mod_timer(&pool
->idle_timer
, expires
);
2011 destroy_worker(worker
);
2019 * manage_workers - manage worker pool
2022 * Assume the manager role and manage the worker pool @worker belongs
2023 * to. At any given time, there can be only zero or one manager per
2024 * pool. The exclusion is handled automatically by this function.
2026 * The caller can safely start processing works on false return. On
2027 * true return, it's guaranteed that need_to_create_worker() is false
2028 * and may_start_working() is true.
2031 * spin_lock_irq(pool->lock) which may be released and regrabbed
2032 * multiple times. Does GFP_KERNEL allocations.
2035 * spin_lock_irq(pool->lock) which may be released and regrabbed
2036 * multiple times. Does GFP_KERNEL allocations.
2038 static bool manage_workers(struct worker
*worker
)
2040 struct worker_pool
*pool
= worker
->pool
;
2044 * Managership is governed by two mutexes - manager_arb and
2045 * manager_mutex. manager_arb handles arbitration of manager role.
2046 * Anyone who successfully grabs manager_arb wins the arbitration
2047 * and becomes the manager. mutex_trylock() on pool->manager_arb
2048 * failure while holding pool->lock reliably indicates that someone
2049 * else is managing the pool and the worker which failed trylock
2050 * can proceed to executing work items. This means that anyone
2051 * grabbing manager_arb is responsible for actually performing
2052 * manager duties. If manager_arb is grabbed and released without
2053 * actual management, the pool may stall indefinitely.
2055 * manager_mutex is used for exclusion of actual management
2056 * operations. The holder of manager_mutex can be sure that none
2057 * of management operations, including creation and destruction of
2058 * workers, won't take place until the mutex is released. Because
2059 * manager_mutex doesn't interfere with manager role arbitration,
2060 * it is guaranteed that the pool's management, while may be
2061 * delayed, won't be disturbed by someone else grabbing
2064 if (!mutex_trylock(&pool
->manager_arb
))
2068 * With manager arbitration won, manager_mutex would be free in
2069 * most cases. trylock first without dropping @pool->lock.
2071 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2072 spin_unlock_irq(&pool
->lock
);
2073 mutex_lock(&pool
->manager_mutex
);
2077 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2080 * Destroy and then create so that may_start_working() is true
2083 ret
|= maybe_destroy_workers(pool
);
2084 ret
|= maybe_create_worker(pool
);
2086 mutex_unlock(&pool
->manager_mutex
);
2087 mutex_unlock(&pool
->manager_arb
);
2092 * process_one_work - process single work
2094 * @work: work to process
2096 * Process @work. This function contains all the logics necessary to
2097 * process a single work including synchronization against and
2098 * interaction with other workers on the same cpu, queueing and
2099 * flushing. As long as context requirement is met, any worker can
2100 * call this function to process a work.
2103 * spin_lock_irq(pool->lock) which is released and regrabbed.
2105 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2106 __releases(&pool
->lock
)
2107 __acquires(&pool
->lock
)
2109 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2110 struct worker_pool
*pool
= worker
->pool
;
2111 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2113 struct worker
*collision
;
2114 #ifdef CONFIG_LOCKDEP
2116 * It is permissible to free the struct work_struct from
2117 * inside the function that is called from it, this we need to
2118 * take into account for lockdep too. To avoid bogus "held
2119 * lock freed" warnings as well as problems when looking into
2120 * work->lockdep_map, make a copy and use that here.
2122 struct lockdep_map lockdep_map
;
2124 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2127 * Ensure we're on the correct CPU. DISASSOCIATED test is
2128 * necessary to avoid spurious warnings from rescuers servicing the
2129 * unbound or a disassociated pool.
2131 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2132 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2133 raw_smp_processor_id() != pool
->cpu
);
2136 * A single work shouldn't be executed concurrently by
2137 * multiple workers on a single cpu. Check whether anyone is
2138 * already processing the work. If so, defer the work to the
2139 * currently executing one.
2141 collision
= find_worker_executing_work(pool
, work
);
2142 if (unlikely(collision
)) {
2143 move_linked_works(work
, &collision
->scheduled
, NULL
);
2147 /* claim and dequeue */
2148 debug_work_deactivate(work
);
2149 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2150 worker
->current_work
= work
;
2151 worker
->current_func
= work
->func
;
2152 worker
->current_pwq
= pwq
;
2153 work_color
= get_work_color(work
);
2155 list_del_init(&work
->entry
);
2158 * CPU intensive works don't participate in concurrency
2159 * management. They're the scheduler's responsibility.
2161 if (unlikely(cpu_intensive
))
2162 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2165 * Unbound pool isn't concurrency managed and work items should be
2166 * executed ASAP. Wake up another worker if necessary.
2168 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2169 wake_up_worker(pool
);
2172 * Record the last pool and clear PENDING which should be the last
2173 * update to @work. Also, do this inside @pool->lock so that
2174 * PENDING and queued state changes happen together while IRQ is
2177 set_work_pool_and_clear_pending(work
, pool
->id
);
2179 spin_unlock_irq(&pool
->lock
);
2181 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2182 lock_map_acquire(&lockdep_map
);
2183 trace_workqueue_execute_start(work
);
2184 worker
->current_func(work
);
2186 * While we must be careful to not use "work" after this, the trace
2187 * point will only record its address.
2189 trace_workqueue_execute_end(work
);
2190 lock_map_release(&lockdep_map
);
2191 lock_map_release(&pwq
->wq
->lockdep_map
);
2193 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2194 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2195 " last function: %pf\n",
2196 current
->comm
, preempt_count(), task_pid_nr(current
),
2197 worker
->current_func
);
2198 debug_show_held_locks(current
);
2202 spin_lock_irq(&pool
->lock
);
2204 /* clear cpu intensive status */
2205 if (unlikely(cpu_intensive
))
2206 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2208 /* we're done with it, release */
2209 hash_del(&worker
->hentry
);
2210 worker
->current_work
= NULL
;
2211 worker
->current_func
= NULL
;
2212 worker
->current_pwq
= NULL
;
2213 pwq_dec_nr_in_flight(pwq
, work_color
);
2217 * process_scheduled_works - process scheduled works
2220 * Process all scheduled works. Please note that the scheduled list
2221 * may change while processing a work, so this function repeatedly
2222 * fetches a work from the top and executes it.
2225 * spin_lock_irq(pool->lock) which may be released and regrabbed
2228 static void process_scheduled_works(struct worker
*worker
)
2230 while (!list_empty(&worker
->scheduled
)) {
2231 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2232 struct work_struct
, entry
);
2233 process_one_work(worker
, work
);
2238 * worker_thread - the worker thread function
2241 * The worker thread function. All workers belong to a worker_pool -
2242 * either a per-cpu one or dynamic unbound one. These workers process all
2243 * work items regardless of their specific target workqueue. The only
2244 * exception is work items which belong to workqueues with a rescuer which
2245 * will be explained in rescuer_thread().
2247 static int worker_thread(void *__worker
)
2249 struct worker
*worker
= __worker
;
2250 struct worker_pool
*pool
= worker
->pool
;
2252 /* tell the scheduler that this is a workqueue worker */
2253 worker
->task
->flags
|= PF_WQ_WORKER
;
2255 spin_lock_irq(&pool
->lock
);
2257 /* am I supposed to die? */
2258 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2259 spin_unlock_irq(&pool
->lock
);
2260 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2261 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2265 worker_leave_idle(worker
);
2267 /* no more worker necessary? */
2268 if (!need_more_worker(pool
))
2271 /* do we need to manage? */
2272 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2276 * ->scheduled list can only be filled while a worker is
2277 * preparing to process a work or actually processing it.
2278 * Make sure nobody diddled with it while I was sleeping.
2280 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2283 * Finish PREP stage. We're guaranteed to have at least one idle
2284 * worker or that someone else has already assumed the manager
2285 * role. This is where @worker starts participating in concurrency
2286 * management if applicable and concurrency management is restored
2287 * after being rebound. See rebind_workers() for details.
2289 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2292 struct work_struct
*work
=
2293 list_first_entry(&pool
->worklist
,
2294 struct work_struct
, entry
);
2296 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2297 /* optimization path, not strictly necessary */
2298 process_one_work(worker
, work
);
2299 if (unlikely(!list_empty(&worker
->scheduled
)))
2300 process_scheduled_works(worker
);
2302 move_linked_works(work
, &worker
->scheduled
, NULL
);
2303 process_scheduled_works(worker
);
2305 } while (keep_working(pool
));
2307 worker_set_flags(worker
, WORKER_PREP
, false);
2309 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2313 * pool->lock is held and there's no work to process and no need to
2314 * manage, sleep. Workers are woken up only while holding
2315 * pool->lock or from local cpu, so setting the current state
2316 * before releasing pool->lock is enough to prevent losing any
2319 worker_enter_idle(worker
);
2320 __set_current_state(TASK_INTERRUPTIBLE
);
2321 spin_unlock_irq(&pool
->lock
);
2327 * rescuer_thread - the rescuer thread function
2330 * Workqueue rescuer thread function. There's one rescuer for each
2331 * workqueue which has WQ_MEM_RECLAIM set.
2333 * Regular work processing on a pool may block trying to create a new
2334 * worker which uses GFP_KERNEL allocation which has slight chance of
2335 * developing into deadlock if some works currently on the same queue
2336 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2337 * the problem rescuer solves.
2339 * When such condition is possible, the pool summons rescuers of all
2340 * workqueues which have works queued on the pool and let them process
2341 * those works so that forward progress can be guaranteed.
2343 * This should happen rarely.
2345 static int rescuer_thread(void *__rescuer
)
2347 struct worker
*rescuer
= __rescuer
;
2348 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2349 struct list_head
*scheduled
= &rescuer
->scheduled
;
2351 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2354 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2355 * doesn't participate in concurrency management.
2357 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2359 set_current_state(TASK_INTERRUPTIBLE
);
2361 if (kthread_should_stop()) {
2362 __set_current_state(TASK_RUNNING
);
2363 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2367 /* see whether any pwq is asking for help */
2368 spin_lock_irq(&wq_mayday_lock
);
2370 while (!list_empty(&wq
->maydays
)) {
2371 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2372 struct pool_workqueue
, mayday_node
);
2373 struct worker_pool
*pool
= pwq
->pool
;
2374 struct work_struct
*work
, *n
;
2376 __set_current_state(TASK_RUNNING
);
2377 list_del_init(&pwq
->mayday_node
);
2379 spin_unlock_irq(&wq_mayday_lock
);
2381 /* migrate to the target cpu if possible */
2382 worker_maybe_bind_and_lock(pool
);
2383 rescuer
->pool
= pool
;
2386 * Slurp in all works issued via this workqueue and
2389 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2390 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2391 if (get_work_pwq(work
) == pwq
)
2392 move_linked_works(work
, scheduled
, &n
);
2394 process_scheduled_works(rescuer
);
2397 * Leave this pool. If keep_working() is %true, notify a
2398 * regular worker; otherwise, we end up with 0 concurrency
2399 * and stalling the execution.
2401 if (keep_working(pool
))
2402 wake_up_worker(pool
);
2404 rescuer
->pool
= NULL
;
2405 spin_unlock(&pool
->lock
);
2406 spin_lock(&wq_mayday_lock
);
2409 spin_unlock_irq(&wq_mayday_lock
);
2411 /* rescuers should never participate in concurrency management */
2412 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2418 struct work_struct work
;
2419 struct completion done
;
2422 static void wq_barrier_func(struct work_struct
*work
)
2424 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2425 complete(&barr
->done
);
2429 * insert_wq_barrier - insert a barrier work
2430 * @pwq: pwq to insert barrier into
2431 * @barr: wq_barrier to insert
2432 * @target: target work to attach @barr to
2433 * @worker: worker currently executing @target, NULL if @target is not executing
2435 * @barr is linked to @target such that @barr is completed only after
2436 * @target finishes execution. Please note that the ordering
2437 * guarantee is observed only with respect to @target and on the local
2440 * Currently, a queued barrier can't be canceled. This is because
2441 * try_to_grab_pending() can't determine whether the work to be
2442 * grabbed is at the head of the queue and thus can't clear LINKED
2443 * flag of the previous work while there must be a valid next work
2444 * after a work with LINKED flag set.
2446 * Note that when @worker is non-NULL, @target may be modified
2447 * underneath us, so we can't reliably determine pwq from @target.
2450 * spin_lock_irq(pool->lock).
2452 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2453 struct wq_barrier
*barr
,
2454 struct work_struct
*target
, struct worker
*worker
)
2456 struct list_head
*head
;
2457 unsigned int linked
= 0;
2460 * debugobject calls are safe here even with pool->lock locked
2461 * as we know for sure that this will not trigger any of the
2462 * checks and call back into the fixup functions where we
2465 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2466 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2467 init_completion(&barr
->done
);
2470 * If @target is currently being executed, schedule the
2471 * barrier to the worker; otherwise, put it after @target.
2474 head
= worker
->scheduled
.next
;
2476 unsigned long *bits
= work_data_bits(target
);
2478 head
= target
->entry
.next
;
2479 /* there can already be other linked works, inherit and set */
2480 linked
= *bits
& WORK_STRUCT_LINKED
;
2481 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2484 debug_work_activate(&barr
->work
);
2485 insert_work(pwq
, &barr
->work
, head
,
2486 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2490 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2491 * @wq: workqueue being flushed
2492 * @flush_color: new flush color, < 0 for no-op
2493 * @work_color: new work color, < 0 for no-op
2495 * Prepare pwqs for workqueue flushing.
2497 * If @flush_color is non-negative, flush_color on all pwqs should be
2498 * -1. If no pwq has in-flight commands at the specified color, all
2499 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2500 * has in flight commands, its pwq->flush_color is set to
2501 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2502 * wakeup logic is armed and %true is returned.
2504 * The caller should have initialized @wq->first_flusher prior to
2505 * calling this function with non-negative @flush_color. If
2506 * @flush_color is negative, no flush color update is done and %false
2509 * If @work_color is non-negative, all pwqs should have the same
2510 * work_color which is previous to @work_color and all will be
2511 * advanced to @work_color.
2514 * mutex_lock(wq->mutex).
2517 * %true if @flush_color >= 0 and there's something to flush. %false
2520 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2521 int flush_color
, int work_color
)
2524 struct pool_workqueue
*pwq
;
2526 if (flush_color
>= 0) {
2527 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2528 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2531 for_each_pwq(pwq
, wq
) {
2532 struct worker_pool
*pool
= pwq
->pool
;
2534 spin_lock_irq(&pool
->lock
);
2536 if (flush_color
>= 0) {
2537 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2539 if (pwq
->nr_in_flight
[flush_color
]) {
2540 pwq
->flush_color
= flush_color
;
2541 atomic_inc(&wq
->nr_pwqs_to_flush
);
2546 if (work_color
>= 0) {
2547 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2548 pwq
->work_color
= work_color
;
2551 spin_unlock_irq(&pool
->lock
);
2554 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2555 complete(&wq
->first_flusher
->done
);
2561 * flush_workqueue - ensure that any scheduled work has run to completion.
2562 * @wq: workqueue to flush
2564 * This function sleeps until all work items which were queued on entry
2565 * have finished execution, but it is not livelocked by new incoming ones.
2567 void flush_workqueue(struct workqueue_struct
*wq
)
2569 struct wq_flusher this_flusher
= {
2570 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2572 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2576 lock_map_acquire(&wq
->lockdep_map
);
2577 lock_map_release(&wq
->lockdep_map
);
2579 mutex_lock(&wq
->mutex
);
2582 * Start-to-wait phase
2584 next_color
= work_next_color(wq
->work_color
);
2586 if (next_color
!= wq
->flush_color
) {
2588 * Color space is not full. The current work_color
2589 * becomes our flush_color and work_color is advanced
2592 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2593 this_flusher
.flush_color
= wq
->work_color
;
2594 wq
->work_color
= next_color
;
2596 if (!wq
->first_flusher
) {
2597 /* no flush in progress, become the first flusher */
2598 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2600 wq
->first_flusher
= &this_flusher
;
2602 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2604 /* nothing to flush, done */
2605 wq
->flush_color
= next_color
;
2606 wq
->first_flusher
= NULL
;
2611 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2612 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2613 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2617 * Oops, color space is full, wait on overflow queue.
2618 * The next flush completion will assign us
2619 * flush_color and transfer to flusher_queue.
2621 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2624 mutex_unlock(&wq
->mutex
);
2626 wait_for_completion(&this_flusher
.done
);
2629 * Wake-up-and-cascade phase
2631 * First flushers are responsible for cascading flushes and
2632 * handling overflow. Non-first flushers can simply return.
2634 if (wq
->first_flusher
!= &this_flusher
)
2637 mutex_lock(&wq
->mutex
);
2639 /* we might have raced, check again with mutex held */
2640 if (wq
->first_flusher
!= &this_flusher
)
2643 wq
->first_flusher
= NULL
;
2645 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2646 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2649 struct wq_flusher
*next
, *tmp
;
2651 /* complete all the flushers sharing the current flush color */
2652 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2653 if (next
->flush_color
!= wq
->flush_color
)
2655 list_del_init(&next
->list
);
2656 complete(&next
->done
);
2659 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2660 wq
->flush_color
!= work_next_color(wq
->work_color
));
2662 /* this flush_color is finished, advance by one */
2663 wq
->flush_color
= work_next_color(wq
->flush_color
);
2665 /* one color has been freed, handle overflow queue */
2666 if (!list_empty(&wq
->flusher_overflow
)) {
2668 * Assign the same color to all overflowed
2669 * flushers, advance work_color and append to
2670 * flusher_queue. This is the start-to-wait
2671 * phase for these overflowed flushers.
2673 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2674 tmp
->flush_color
= wq
->work_color
;
2676 wq
->work_color
= work_next_color(wq
->work_color
);
2678 list_splice_tail_init(&wq
->flusher_overflow
,
2679 &wq
->flusher_queue
);
2680 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2683 if (list_empty(&wq
->flusher_queue
)) {
2684 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2689 * Need to flush more colors. Make the next flusher
2690 * the new first flusher and arm pwqs.
2692 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2693 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2695 list_del_init(&next
->list
);
2696 wq
->first_flusher
= next
;
2698 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2702 * Meh... this color is already done, clear first
2703 * flusher and repeat cascading.
2705 wq
->first_flusher
= NULL
;
2709 mutex_unlock(&wq
->mutex
);
2711 EXPORT_SYMBOL_GPL(flush_workqueue
);
2714 * drain_workqueue - drain a workqueue
2715 * @wq: workqueue to drain
2717 * Wait until the workqueue becomes empty. While draining is in progress,
2718 * only chain queueing is allowed. IOW, only currently pending or running
2719 * work items on @wq can queue further work items on it. @wq is flushed
2720 * repeatedly until it becomes empty. The number of flushing is detemined
2721 * by the depth of chaining and should be relatively short. Whine if it
2724 void drain_workqueue(struct workqueue_struct
*wq
)
2726 unsigned int flush_cnt
= 0;
2727 struct pool_workqueue
*pwq
;
2730 * __queue_work() needs to test whether there are drainers, is much
2731 * hotter than drain_workqueue() and already looks at @wq->flags.
2732 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2734 mutex_lock(&wq
->mutex
);
2735 if (!wq
->nr_drainers
++)
2736 wq
->flags
|= __WQ_DRAINING
;
2737 mutex_unlock(&wq
->mutex
);
2739 flush_workqueue(wq
);
2741 mutex_lock(&wq
->mutex
);
2743 for_each_pwq(pwq
, wq
) {
2746 spin_lock_irq(&pwq
->pool
->lock
);
2747 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2748 spin_unlock_irq(&pwq
->pool
->lock
);
2753 if (++flush_cnt
== 10 ||
2754 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2755 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2756 wq
->name
, flush_cnt
);
2758 mutex_unlock(&wq
->mutex
);
2762 if (!--wq
->nr_drainers
)
2763 wq
->flags
&= ~__WQ_DRAINING
;
2764 mutex_unlock(&wq
->mutex
);
2766 EXPORT_SYMBOL_GPL(drain_workqueue
);
2768 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2770 struct worker
*worker
= NULL
;
2771 struct worker_pool
*pool
;
2772 struct pool_workqueue
*pwq
;
2776 local_irq_disable();
2777 pool
= get_work_pool(work
);
2783 spin_lock(&pool
->lock
);
2784 /* see the comment in try_to_grab_pending() with the same code */
2785 pwq
= get_work_pwq(work
);
2787 if (unlikely(pwq
->pool
!= pool
))
2790 worker
= find_worker_executing_work(pool
, work
);
2793 pwq
= worker
->current_pwq
;
2796 insert_wq_barrier(pwq
, barr
, work
, worker
);
2797 spin_unlock_irq(&pool
->lock
);
2800 * If @max_active is 1 or rescuer is in use, flushing another work
2801 * item on the same workqueue may lead to deadlock. Make sure the
2802 * flusher is not running on the same workqueue by verifying write
2805 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2806 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2808 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2809 lock_map_release(&pwq
->wq
->lockdep_map
);
2813 spin_unlock_irq(&pool
->lock
);
2818 * flush_work - wait for a work to finish executing the last queueing instance
2819 * @work: the work to flush
2821 * Wait until @work has finished execution. @work is guaranteed to be idle
2822 * on return if it hasn't been requeued since flush started.
2825 * %true if flush_work() waited for the work to finish execution,
2826 * %false if it was already idle.
2828 bool flush_work(struct work_struct
*work
)
2830 struct wq_barrier barr
;
2832 lock_map_acquire(&work
->lockdep_map
);
2833 lock_map_release(&work
->lockdep_map
);
2835 if (start_flush_work(work
, &barr
)) {
2836 wait_for_completion(&barr
.done
);
2837 destroy_work_on_stack(&barr
.work
);
2843 EXPORT_SYMBOL_GPL(flush_work
);
2845 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2847 unsigned long flags
;
2851 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2853 * If someone else is canceling, wait for the same event it
2854 * would be waiting for before retrying.
2856 if (unlikely(ret
== -ENOENT
))
2858 } while (unlikely(ret
< 0));
2860 /* tell other tasks trying to grab @work to back off */
2861 mark_work_canceling(work
);
2862 local_irq_restore(flags
);
2865 clear_work_data(work
);
2870 * cancel_work_sync - cancel a work and wait for it to finish
2871 * @work: the work to cancel
2873 * Cancel @work and wait for its execution to finish. This function
2874 * can be used even if the work re-queues itself or migrates to
2875 * another workqueue. On return from this function, @work is
2876 * guaranteed to be not pending or executing on any CPU.
2878 * cancel_work_sync(&delayed_work->work) must not be used for
2879 * delayed_work's. Use cancel_delayed_work_sync() instead.
2881 * The caller must ensure that the workqueue on which @work was last
2882 * queued can't be destroyed before this function returns.
2885 * %true if @work was pending, %false otherwise.
2887 bool cancel_work_sync(struct work_struct
*work
)
2889 return __cancel_work_timer(work
, false);
2891 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2894 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2895 * @dwork: the delayed work to flush
2897 * Delayed timer is cancelled and the pending work is queued for
2898 * immediate execution. Like flush_work(), this function only
2899 * considers the last queueing instance of @dwork.
2902 * %true if flush_work() waited for the work to finish execution,
2903 * %false if it was already idle.
2905 bool flush_delayed_work(struct delayed_work
*dwork
)
2907 local_irq_disable();
2908 if (del_timer_sync(&dwork
->timer
))
2909 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2911 return flush_work(&dwork
->work
);
2913 EXPORT_SYMBOL(flush_delayed_work
);
2916 * cancel_delayed_work - cancel a delayed work
2917 * @dwork: delayed_work to cancel
2919 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2920 * and canceled; %false if wasn't pending. Note that the work callback
2921 * function may still be running on return, unless it returns %true and the
2922 * work doesn't re-arm itself. Explicitly flush or use
2923 * cancel_delayed_work_sync() to wait on it.
2925 * This function is safe to call from any context including IRQ handler.
2927 bool cancel_delayed_work(struct delayed_work
*dwork
)
2929 unsigned long flags
;
2933 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2934 } while (unlikely(ret
== -EAGAIN
));
2936 if (unlikely(ret
< 0))
2939 set_work_pool_and_clear_pending(&dwork
->work
,
2940 get_work_pool_id(&dwork
->work
));
2941 local_irq_restore(flags
);
2944 EXPORT_SYMBOL(cancel_delayed_work
);
2947 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2948 * @dwork: the delayed work cancel
2950 * This is cancel_work_sync() for delayed works.
2953 * %true if @dwork was pending, %false otherwise.
2955 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2957 return __cancel_work_timer(&dwork
->work
, true);
2959 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2962 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2963 * @func: the function to call
2965 * schedule_on_each_cpu() executes @func on each online CPU using the
2966 * system workqueue and blocks until all CPUs have completed.
2967 * schedule_on_each_cpu() is very slow.
2970 * 0 on success, -errno on failure.
2972 int schedule_on_each_cpu(work_func_t func
)
2975 struct work_struct __percpu
*works
;
2977 works
= alloc_percpu(struct work_struct
);
2983 for_each_online_cpu(cpu
) {
2984 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2986 INIT_WORK(work
, func
);
2987 schedule_work_on(cpu
, work
);
2990 for_each_online_cpu(cpu
)
2991 flush_work(per_cpu_ptr(works
, cpu
));
2999 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3001 * Forces execution of the kernel-global workqueue and blocks until its
3004 * Think twice before calling this function! It's very easy to get into
3005 * trouble if you don't take great care. Either of the following situations
3006 * will lead to deadlock:
3008 * One of the work items currently on the workqueue needs to acquire
3009 * a lock held by your code or its caller.
3011 * Your code is running in the context of a work routine.
3013 * They will be detected by lockdep when they occur, but the first might not
3014 * occur very often. It depends on what work items are on the workqueue and
3015 * what locks they need, which you have no control over.
3017 * In most situations flushing the entire workqueue is overkill; you merely
3018 * need to know that a particular work item isn't queued and isn't running.
3019 * In such cases you should use cancel_delayed_work_sync() or
3020 * cancel_work_sync() instead.
3022 void flush_scheduled_work(void)
3024 flush_workqueue(system_wq
);
3026 EXPORT_SYMBOL(flush_scheduled_work
);
3029 * execute_in_process_context - reliably execute the routine with user context
3030 * @fn: the function to execute
3031 * @ew: guaranteed storage for the execute work structure (must
3032 * be available when the work executes)
3034 * Executes the function immediately if process context is available,
3035 * otherwise schedules the function for delayed execution.
3037 * Returns: 0 - function was executed
3038 * 1 - function was scheduled for execution
3040 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3042 if (!in_interrupt()) {
3047 INIT_WORK(&ew
->work
, fn
);
3048 schedule_work(&ew
->work
);
3052 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3056 * Workqueues with WQ_SYSFS flag set is visible to userland via
3057 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3058 * following attributes.
3060 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3061 * max_active RW int : maximum number of in-flight work items
3063 * Unbound workqueues have the following extra attributes.
3065 * id RO int : the associated pool ID
3066 * nice RW int : nice value of the workers
3067 * cpumask RW mask : bitmask of allowed CPUs for the workers
3070 struct workqueue_struct
*wq
;
3074 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3076 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3081 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3082 struct device_attribute
*attr
, char *buf
)
3084 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3086 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3089 static ssize_t
wq_max_active_show(struct device
*dev
,
3090 struct device_attribute
*attr
, char *buf
)
3092 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3094 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3097 static ssize_t
wq_max_active_store(struct device
*dev
,
3098 struct device_attribute
*attr
,
3099 const char *buf
, size_t count
)
3101 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3104 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3107 workqueue_set_max_active(wq
, val
);
3111 static struct device_attribute wq_sysfs_attrs
[] = {
3112 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3113 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3117 static ssize_t
wq_pool_id_show(struct device
*dev
,
3118 struct device_attribute
*attr
, char *buf
)
3120 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3121 struct worker_pool
*pool
;
3124 rcu_read_lock_sched();
3125 pool
= first_pwq(wq
)->pool
;
3126 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", pool
->id
);
3127 rcu_read_unlock_sched();
3132 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3135 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3138 mutex_lock(&wq
->mutex
);
3139 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3140 mutex_unlock(&wq
->mutex
);
3145 /* prepare workqueue_attrs for sysfs store operations */
3146 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3148 struct workqueue_attrs
*attrs
;
3150 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3154 mutex_lock(&wq
->mutex
);
3155 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3156 mutex_unlock(&wq
->mutex
);
3160 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3161 const char *buf
, size_t count
)
3163 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3164 struct workqueue_attrs
*attrs
;
3167 attrs
= wq_sysfs_prep_attrs(wq
);
3171 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3172 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3173 ret
= apply_workqueue_attrs(wq
, attrs
);
3177 free_workqueue_attrs(attrs
);
3178 return ret
?: count
;
3181 static ssize_t
wq_cpumask_show(struct device
*dev
,
3182 struct device_attribute
*attr
, char *buf
)
3184 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3187 mutex_lock(&wq
->mutex
);
3188 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3189 mutex_unlock(&wq
->mutex
);
3191 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3195 static ssize_t
wq_cpumask_store(struct device
*dev
,
3196 struct device_attribute
*attr
,
3197 const char *buf
, size_t count
)
3199 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3200 struct workqueue_attrs
*attrs
;
3203 attrs
= wq_sysfs_prep_attrs(wq
);
3207 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3209 ret
= apply_workqueue_attrs(wq
, attrs
);
3211 free_workqueue_attrs(attrs
);
3212 return ret
?: count
;
3215 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3216 __ATTR(pool_id
, 0444, wq_pool_id_show
, NULL
),
3217 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3218 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3222 static struct bus_type wq_subsys
= {
3223 .name
= "workqueue",
3224 .dev_attrs
= wq_sysfs_attrs
,
3227 static int __init
wq_sysfs_init(void)
3229 return subsys_virtual_register(&wq_subsys
, NULL
);
3231 core_initcall(wq_sysfs_init
);
3233 static void wq_device_release(struct device
*dev
)
3235 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3241 * workqueue_sysfs_register - make a workqueue visible in sysfs
3242 * @wq: the workqueue to register
3244 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3245 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3246 * which is the preferred method.
3248 * Workqueue user should use this function directly iff it wants to apply
3249 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3250 * apply_workqueue_attrs() may race against userland updating the
3253 * Returns 0 on success, -errno on failure.
3255 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3257 struct wq_device
*wq_dev
;
3261 * Adjusting max_active or creating new pwqs by applyting
3262 * attributes breaks ordering guarantee. Disallow exposing ordered
3265 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3268 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3273 wq_dev
->dev
.bus
= &wq_subsys
;
3274 wq_dev
->dev
.init_name
= wq
->name
;
3275 wq_dev
->dev
.release
= wq_device_release
;
3278 * unbound_attrs are created separately. Suppress uevent until
3279 * everything is ready.
3281 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3283 ret
= device_register(&wq_dev
->dev
);
3290 if (wq
->flags
& WQ_UNBOUND
) {
3291 struct device_attribute
*attr
;
3293 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3294 ret
= device_create_file(&wq_dev
->dev
, attr
);
3296 device_unregister(&wq_dev
->dev
);
3303 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3308 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3309 * @wq: the workqueue to unregister
3311 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3313 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3315 struct wq_device
*wq_dev
= wq
->wq_dev
;
3321 device_unregister(&wq_dev
->dev
);
3323 #else /* CONFIG_SYSFS */
3324 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3325 #endif /* CONFIG_SYSFS */
3328 * free_workqueue_attrs - free a workqueue_attrs
3329 * @attrs: workqueue_attrs to free
3331 * Undo alloc_workqueue_attrs().
3333 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3336 free_cpumask_var(attrs
->cpumask
);
3342 * alloc_workqueue_attrs - allocate a workqueue_attrs
3343 * @gfp_mask: allocation mask to use
3345 * Allocate a new workqueue_attrs, initialize with default settings and
3346 * return it. Returns NULL on failure.
3348 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3350 struct workqueue_attrs
*attrs
;
3352 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3355 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3358 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3361 free_workqueue_attrs(attrs
);
3365 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3366 const struct workqueue_attrs
*from
)
3368 to
->nice
= from
->nice
;
3369 cpumask_copy(to
->cpumask
, from
->cpumask
);
3372 /* hash value of the content of @attr */
3373 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3377 hash
= jhash_1word(attrs
->nice
, hash
);
3378 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3379 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3383 /* content equality test */
3384 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3385 const struct workqueue_attrs
*b
)
3387 if (a
->nice
!= b
->nice
)
3389 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3395 * init_worker_pool - initialize a newly zalloc'd worker_pool
3396 * @pool: worker_pool to initialize
3398 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3399 * Returns 0 on success, -errno on failure. Even on failure, all fields
3400 * inside @pool proper are initialized and put_unbound_pool() can be called
3401 * on @pool safely to release it.
3403 static int init_worker_pool(struct worker_pool
*pool
)
3405 spin_lock_init(&pool
->lock
);
3408 pool
->node
= NUMA_NO_NODE
;
3409 pool
->flags
|= POOL_DISASSOCIATED
;
3410 INIT_LIST_HEAD(&pool
->worklist
);
3411 INIT_LIST_HEAD(&pool
->idle_list
);
3412 hash_init(pool
->busy_hash
);
3414 init_timer_deferrable(&pool
->idle_timer
);
3415 pool
->idle_timer
.function
= idle_worker_timeout
;
3416 pool
->idle_timer
.data
= (unsigned long)pool
;
3418 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3419 (unsigned long)pool
);
3421 mutex_init(&pool
->manager_arb
);
3422 mutex_init(&pool
->manager_mutex
);
3423 idr_init(&pool
->worker_idr
);
3425 INIT_HLIST_NODE(&pool
->hash_node
);
3428 /* shouldn't fail above this point */
3429 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3435 static void rcu_free_pool(struct rcu_head
*rcu
)
3437 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3439 idr_destroy(&pool
->worker_idr
);
3440 free_workqueue_attrs(pool
->attrs
);
3445 * put_unbound_pool - put a worker_pool
3446 * @pool: worker_pool to put
3448 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3449 * safe manner. get_unbound_pool() calls this function on its failure path
3450 * and this function should be able to release pools which went through,
3451 * successfully or not, init_worker_pool().
3453 * Should be called with wq_pool_mutex held.
3455 static void put_unbound_pool(struct worker_pool
*pool
)
3457 struct worker
*worker
;
3459 lockdep_assert_held(&wq_pool_mutex
);
3465 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3466 WARN_ON(!list_empty(&pool
->worklist
)))
3469 /* release id and unhash */
3471 idr_remove(&worker_pool_idr
, pool
->id
);
3472 hash_del(&pool
->hash_node
);
3475 * Become the manager and destroy all workers. Grabbing
3476 * manager_arb prevents @pool's workers from blocking on
3479 mutex_lock(&pool
->manager_arb
);
3480 mutex_lock(&pool
->manager_mutex
);
3481 spin_lock_irq(&pool
->lock
);
3483 while ((worker
= first_worker(pool
)))
3484 destroy_worker(worker
);
3485 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3487 spin_unlock_irq(&pool
->lock
);
3488 mutex_unlock(&pool
->manager_mutex
);
3489 mutex_unlock(&pool
->manager_arb
);
3491 /* shut down the timers */
3492 del_timer_sync(&pool
->idle_timer
);
3493 del_timer_sync(&pool
->mayday_timer
);
3495 /* sched-RCU protected to allow dereferences from get_work_pool() */
3496 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3500 * get_unbound_pool - get a worker_pool with the specified attributes
3501 * @attrs: the attributes of the worker_pool to get
3503 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3504 * reference count and return it. If there already is a matching
3505 * worker_pool, it will be used; otherwise, this function attempts to
3506 * create a new one. On failure, returns NULL.
3508 * Should be called with wq_pool_mutex held.
3510 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3512 u32 hash
= wqattrs_hash(attrs
);
3513 struct worker_pool
*pool
;
3516 lockdep_assert_held(&wq_pool_mutex
);
3518 /* do we already have a matching pool? */
3519 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3520 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3526 /* nope, create a new one */
3527 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3528 if (!pool
|| init_worker_pool(pool
) < 0)
3531 if (workqueue_freezing
)
3532 pool
->flags
|= POOL_FREEZING
;
3534 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3535 copy_workqueue_attrs(pool
->attrs
, attrs
);
3537 /* if cpumask is contained inside a NUMA node, we belong to that node */
3538 if (wq_numa_enabled
) {
3539 for_each_node(node
) {
3540 if (cpumask_subset(pool
->attrs
->cpumask
,
3541 wq_numa_possible_cpumask
[node
])) {
3548 if (worker_pool_assign_id(pool
) < 0)
3551 /* create and start the initial worker */
3552 if (create_and_start_worker(pool
) < 0)
3556 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3561 put_unbound_pool(pool
);
3565 static void rcu_free_pwq(struct rcu_head
*rcu
)
3567 kmem_cache_free(pwq_cache
,
3568 container_of(rcu
, struct pool_workqueue
, rcu
));
3572 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3573 * and needs to be destroyed.
3575 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3577 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3578 unbound_release_work
);
3579 struct workqueue_struct
*wq
= pwq
->wq
;
3580 struct worker_pool
*pool
= pwq
->pool
;
3583 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3587 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3588 * necessary on release but do it anyway. It's easier to verify
3589 * and consistent with the linking path.
3591 mutex_lock(&wq
->mutex
);
3592 list_del_rcu(&pwq
->pwqs_node
);
3593 is_last
= list_empty(&wq
->pwqs
);
3594 mutex_unlock(&wq
->mutex
);
3596 mutex_lock(&wq_pool_mutex
);
3597 put_unbound_pool(pool
);
3598 mutex_unlock(&wq_pool_mutex
);
3600 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3603 * If we're the last pwq going away, @wq is already dead and no one
3604 * is gonna access it anymore. Free it.
3607 free_workqueue_attrs(wq
->unbound_attrs
);
3613 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3614 * @pwq: target pool_workqueue
3616 * If @pwq isn't freezing, set @pwq->max_active to the associated
3617 * workqueue's saved_max_active and activate delayed work items
3618 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3620 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3622 struct workqueue_struct
*wq
= pwq
->wq
;
3623 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3625 /* for @wq->saved_max_active */
3626 lockdep_assert_held(&wq
->mutex
);
3628 /* fast exit for non-freezable wqs */
3629 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3632 spin_lock_irq(&pwq
->pool
->lock
);
3634 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3635 pwq
->max_active
= wq
->saved_max_active
;
3637 while (!list_empty(&pwq
->delayed_works
) &&
3638 pwq
->nr_active
< pwq
->max_active
)
3639 pwq_activate_first_delayed(pwq
);
3642 * Need to kick a worker after thawed or an unbound wq's
3643 * max_active is bumped. It's a slow path. Do it always.
3645 wake_up_worker(pwq
->pool
);
3647 pwq
->max_active
= 0;
3650 spin_unlock_irq(&pwq
->pool
->lock
);
3653 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3654 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3655 struct worker_pool
*pool
)
3657 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3659 memset(pwq
, 0, sizeof(*pwq
));
3663 pwq
->flush_color
= -1;
3665 INIT_LIST_HEAD(&pwq
->delayed_works
);
3666 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3667 INIT_LIST_HEAD(&pwq
->mayday_node
);
3668 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3671 /* sync @pwq with the current state of its associated wq and link it */
3672 static void link_pwq(struct pool_workqueue
*pwq
)
3674 struct workqueue_struct
*wq
= pwq
->wq
;
3676 lockdep_assert_held(&wq
->mutex
);
3678 /* may be called multiple times, ignore if already linked */
3679 if (!list_empty(&pwq
->pwqs_node
))
3683 * Set the matching work_color. This is synchronized with
3684 * wq->mutex to avoid confusing flush_workqueue().
3686 pwq
->work_color
= wq
->work_color
;
3688 /* sync max_active to the current setting */
3689 pwq_adjust_max_active(pwq
);
3692 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3695 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3696 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3697 const struct workqueue_attrs
*attrs
)
3699 struct worker_pool
*pool
;
3700 struct pool_workqueue
*pwq
;
3702 lockdep_assert_held(&wq_pool_mutex
);
3704 pool
= get_unbound_pool(attrs
);
3708 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3710 put_unbound_pool(pool
);
3714 init_pwq(pwq
, wq
, pool
);
3718 /* undo alloc_unbound_pwq(), used only in the error path */
3719 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3721 lockdep_assert_held(&wq_pool_mutex
);
3724 put_unbound_pool(pwq
->pool
);
3730 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3731 * @attrs: the wq_attrs of interest
3732 * @node: the target NUMA node
3733 * @cpu_going_down: if >= 0, the CPU to consider as offline
3734 * @cpumask: outarg, the resulting cpumask
3736 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3737 * @cpu_going_down is >= 0, that cpu is considered offline during
3738 * calculation. The result is stored in @cpumask. This function returns
3739 * %true if the resulting @cpumask is different from @attrs->cpumask,
3742 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3743 * enabled and @node has online CPUs requested by @attrs, the returned
3744 * cpumask is the intersection of the possible CPUs of @node and
3747 * The caller is responsible for ensuring that the cpumask of @node stays
3750 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3751 int cpu_going_down
, cpumask_t
*cpumask
)
3753 if (!wq_numa_enabled
)
3756 /* does @node have any online CPUs @attrs wants? */
3757 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3758 if (cpu_going_down
>= 0)
3759 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3761 if (cpumask_empty(cpumask
))
3764 /* yeap, return possible CPUs in @node that @attrs wants */
3765 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3766 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3769 cpumask_copy(cpumask
, attrs
->cpumask
);
3773 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3774 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3776 struct pool_workqueue
*pwq
)
3778 struct pool_workqueue
*old_pwq
;
3780 lockdep_assert_held(&wq
->mutex
);
3782 /* link_pwq() can handle duplicate calls */
3785 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3786 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3791 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3792 * @wq: the target workqueue
3793 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3795 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3796 * machines, this function maps a separate pwq to each NUMA node with
3797 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3798 * NUMA node it was issued on. Older pwqs are released as in-flight work
3799 * items finish. Note that a work item which repeatedly requeues itself
3800 * back-to-back will stay on its current pwq.
3802 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3805 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3806 const struct workqueue_attrs
*attrs
)
3808 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3809 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3812 /* only unbound workqueues can change attributes */
3813 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3816 /* creating multiple pwqs breaks ordering guarantee */
3817 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3820 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3821 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3822 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3823 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3826 /* make a copy of @attrs and sanitize it */
3827 copy_workqueue_attrs(new_attrs
, attrs
);
3828 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3831 * We may create multiple pwqs with differing cpumasks. Make a
3832 * copy of @new_attrs which will be modified and used to obtain
3835 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3838 * CPUs should stay stable across pwq creations and installations.
3839 * Pin CPUs, determine the target cpumask for each node and create
3844 mutex_lock(&wq_pool_mutex
);
3847 * If something goes wrong during CPU up/down, we'll fall back to
3848 * the default pwq covering whole @attrs->cpumask. Always create
3849 * it even if we don't use it immediately.
3851 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3855 for_each_node(node
) {
3856 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3857 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3862 pwq_tbl
[node
] = dfl_pwq
;
3866 mutex_unlock(&wq_pool_mutex
);
3868 /* all pwqs have been created successfully, let's install'em */
3869 mutex_lock(&wq
->mutex
);
3871 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3873 /* save the previous pwq and install the new one */
3875 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3877 /* @dfl_pwq might not have been used, ensure it's linked */
3879 swap(wq
->dfl_pwq
, dfl_pwq
);
3881 mutex_unlock(&wq
->mutex
);
3883 /* put the old pwqs */
3885 put_pwq_unlocked(pwq_tbl
[node
]);
3886 put_pwq_unlocked(dfl_pwq
);
3892 free_workqueue_attrs(tmp_attrs
);
3893 free_workqueue_attrs(new_attrs
);
3898 free_unbound_pwq(dfl_pwq
);
3900 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
3901 free_unbound_pwq(pwq_tbl
[node
]);
3902 mutex_unlock(&wq_pool_mutex
);
3910 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3911 * @wq: the target workqueue
3912 * @cpu: the CPU coming up or going down
3913 * @online: whether @cpu is coming up or going down
3915 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3916 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3919 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3920 * falls back to @wq->dfl_pwq which may not be optimal but is always
3923 * Note that when the last allowed CPU of a NUMA node goes offline for a
3924 * workqueue with a cpumask spanning multiple nodes, the workers which were
3925 * already executing the work items for the workqueue will lose their CPU
3926 * affinity and may execute on any CPU. This is similar to how per-cpu
3927 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3928 * affinity, it's the user's responsibility to flush the work item from
3931 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3934 int node
= cpu_to_node(cpu
);
3935 int cpu_off
= online
? -1 : cpu
;
3936 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3937 struct workqueue_attrs
*target_attrs
;
3940 lockdep_assert_held(&wq_pool_mutex
);
3942 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
3946 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3947 * Let's use a preallocated one. The following buf is protected by
3948 * CPU hotplug exclusion.
3950 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3951 cpumask
= target_attrs
->cpumask
;
3953 mutex_lock(&wq
->mutex
);
3955 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3956 pwq
= unbound_pwq_by_node(wq
, node
);
3959 * Let's determine what needs to be done. If the target cpumask is
3960 * different from wq's, we need to compare it to @pwq's and create
3961 * a new one if they don't match. If the target cpumask equals
3962 * wq's, the default pwq should be used. If @pwq is already the
3963 * default one, nothing to do; otherwise, install the default one.
3965 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
3966 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3969 if (pwq
== wq
->dfl_pwq
)
3975 mutex_unlock(&wq
->mutex
);
3977 /* create a new pwq */
3978 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
3980 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3986 * Install the new pwq. As this function is called only from CPU
3987 * hotplug callbacks and applying a new attrs is wrapped with
3988 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3991 mutex_lock(&wq
->mutex
);
3992 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
3996 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
3997 get_pwq(wq
->dfl_pwq
);
3998 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
3999 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4001 mutex_unlock(&wq
->mutex
);
4002 put_pwq_unlocked(old_pwq
);
4005 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4007 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4010 if (!(wq
->flags
& WQ_UNBOUND
)) {
4011 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4015 for_each_possible_cpu(cpu
) {
4016 struct pool_workqueue
*pwq
=
4017 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4018 struct worker_pool
*cpu_pools
=
4019 per_cpu(cpu_worker_pools
, cpu
);
4021 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4023 mutex_lock(&wq
->mutex
);
4025 mutex_unlock(&wq
->mutex
);
4029 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4033 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4036 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4038 if (max_active
< 1 || max_active
> lim
)
4039 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4040 max_active
, name
, 1, lim
);
4042 return clamp_val(max_active
, 1, lim
);
4045 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4048 struct lock_class_key
*key
,
4049 const char *lock_name
, ...)
4051 size_t tbl_size
= 0;
4053 struct workqueue_struct
*wq
;
4054 struct pool_workqueue
*pwq
;
4056 /* allocate wq and format name */
4057 if (flags
& WQ_UNBOUND
)
4058 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4060 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4064 if (flags
& WQ_UNBOUND
) {
4065 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4066 if (!wq
->unbound_attrs
)
4070 va_start(args
, lock_name
);
4071 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4074 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4075 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4079 wq
->saved_max_active
= max_active
;
4080 mutex_init(&wq
->mutex
);
4081 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4082 INIT_LIST_HEAD(&wq
->pwqs
);
4083 INIT_LIST_HEAD(&wq
->flusher_queue
);
4084 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4085 INIT_LIST_HEAD(&wq
->maydays
);
4087 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4088 INIT_LIST_HEAD(&wq
->list
);
4090 if (alloc_and_link_pwqs(wq
) < 0)
4094 * Workqueues which may be used during memory reclaim should
4095 * have a rescuer to guarantee forward progress.
4097 if (flags
& WQ_MEM_RECLAIM
) {
4098 struct worker
*rescuer
;
4100 rescuer
= alloc_worker();
4104 rescuer
->rescue_wq
= wq
;
4105 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4107 if (IS_ERR(rescuer
->task
)) {
4112 wq
->rescuer
= rescuer
;
4113 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4114 wake_up_process(rescuer
->task
);
4117 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4121 * wq_pool_mutex protects global freeze state and workqueues list.
4122 * Grab it, adjust max_active and add the new @wq to workqueues
4125 mutex_lock(&wq_pool_mutex
);
4127 mutex_lock(&wq
->mutex
);
4128 for_each_pwq(pwq
, wq
)
4129 pwq_adjust_max_active(pwq
);
4130 mutex_unlock(&wq
->mutex
);
4132 list_add(&wq
->list
, &workqueues
);
4134 mutex_unlock(&wq_pool_mutex
);
4139 free_workqueue_attrs(wq
->unbound_attrs
);
4143 destroy_workqueue(wq
);
4146 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4149 * destroy_workqueue - safely terminate a workqueue
4150 * @wq: target workqueue
4152 * Safely destroy a workqueue. All work currently pending will be done first.
4154 void destroy_workqueue(struct workqueue_struct
*wq
)
4156 struct pool_workqueue
*pwq
;
4159 /* drain it before proceeding with destruction */
4160 drain_workqueue(wq
);
4163 mutex_lock(&wq
->mutex
);
4164 for_each_pwq(pwq
, wq
) {
4167 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4168 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4169 mutex_unlock(&wq
->mutex
);
4174 if (WARN_ON(pwq
->refcnt
> 1) ||
4175 WARN_ON(pwq
->nr_active
) ||
4176 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4177 mutex_unlock(&wq
->mutex
);
4181 mutex_unlock(&wq
->mutex
);
4184 * wq list is used to freeze wq, remove from list after
4185 * flushing is complete in case freeze races us.
4187 mutex_lock(&wq_pool_mutex
);
4188 list_del_init(&wq
->list
);
4189 mutex_unlock(&wq_pool_mutex
);
4191 workqueue_sysfs_unregister(wq
);
4194 kthread_stop(wq
->rescuer
->task
);
4199 if (!(wq
->flags
& WQ_UNBOUND
)) {
4201 * The base ref is never dropped on per-cpu pwqs. Directly
4202 * free the pwqs and wq.
4204 free_percpu(wq
->cpu_pwqs
);
4208 * We're the sole accessor of @wq at this point. Directly
4209 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4210 * @wq will be freed when the last pwq is released.
4212 for_each_node(node
) {
4213 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4214 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4215 put_pwq_unlocked(pwq
);
4219 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4220 * put. Don't access it afterwards.
4224 put_pwq_unlocked(pwq
);
4227 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4230 * workqueue_set_max_active - adjust max_active of a workqueue
4231 * @wq: target workqueue
4232 * @max_active: new max_active value.
4234 * Set max_active of @wq to @max_active.
4237 * Don't call from IRQ context.
4239 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4241 struct pool_workqueue
*pwq
;
4243 /* disallow meddling with max_active for ordered workqueues */
4244 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4247 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4249 mutex_lock(&wq
->mutex
);
4251 wq
->saved_max_active
= max_active
;
4253 for_each_pwq(pwq
, wq
)
4254 pwq_adjust_max_active(pwq
);
4256 mutex_unlock(&wq
->mutex
);
4258 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4261 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4263 * Determine whether %current is a workqueue rescuer. Can be used from
4264 * work functions to determine whether it's being run off the rescuer task.
4266 bool current_is_workqueue_rescuer(void)
4268 struct worker
*worker
= current_wq_worker();
4270 return worker
&& worker
->rescue_wq
;
4274 * workqueue_congested - test whether a workqueue is congested
4275 * @cpu: CPU in question
4276 * @wq: target workqueue
4278 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4279 * no synchronization around this function and the test result is
4280 * unreliable and only useful as advisory hints or for debugging.
4283 * %true if congested, %false otherwise.
4285 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4287 struct pool_workqueue
*pwq
;
4290 rcu_read_lock_sched();
4292 if (!(wq
->flags
& WQ_UNBOUND
))
4293 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4295 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4297 ret
= !list_empty(&pwq
->delayed_works
);
4298 rcu_read_unlock_sched();
4302 EXPORT_SYMBOL_GPL(workqueue_congested
);
4305 * work_busy - test whether a work is currently pending or running
4306 * @work: the work to be tested
4308 * Test whether @work is currently pending or running. There is no
4309 * synchronization around this function and the test result is
4310 * unreliable and only useful as advisory hints or for debugging.
4313 * OR'd bitmask of WORK_BUSY_* bits.
4315 unsigned int work_busy(struct work_struct
*work
)
4317 struct worker_pool
*pool
;
4318 unsigned long flags
;
4319 unsigned int ret
= 0;
4321 if (work_pending(work
))
4322 ret
|= WORK_BUSY_PENDING
;
4324 local_irq_save(flags
);
4325 pool
= get_work_pool(work
);
4327 spin_lock(&pool
->lock
);
4328 if (find_worker_executing_work(pool
, work
))
4329 ret
|= WORK_BUSY_RUNNING
;
4330 spin_unlock(&pool
->lock
);
4332 local_irq_restore(flags
);
4336 EXPORT_SYMBOL_GPL(work_busy
);
4341 * There are two challenges in supporting CPU hotplug. Firstly, there
4342 * are a lot of assumptions on strong associations among work, pwq and
4343 * pool which make migrating pending and scheduled works very
4344 * difficult to implement without impacting hot paths. Secondly,
4345 * worker pools serve mix of short, long and very long running works making
4346 * blocked draining impractical.
4348 * This is solved by allowing the pools to be disassociated from the CPU
4349 * running as an unbound one and allowing it to be reattached later if the
4350 * cpu comes back online.
4353 static void wq_unbind_fn(struct work_struct
*work
)
4355 int cpu
= smp_processor_id();
4356 struct worker_pool
*pool
;
4357 struct worker
*worker
;
4360 for_each_cpu_worker_pool(pool
, cpu
) {
4361 WARN_ON_ONCE(cpu
!= smp_processor_id());
4363 mutex_lock(&pool
->manager_mutex
);
4364 spin_lock_irq(&pool
->lock
);
4367 * We've blocked all manager operations. Make all workers
4368 * unbound and set DISASSOCIATED. Before this, all workers
4369 * except for the ones which are still executing works from
4370 * before the last CPU down must be on the cpu. After
4371 * this, they may become diasporas.
4373 for_each_pool_worker(worker
, wi
, pool
)
4374 worker
->flags
|= WORKER_UNBOUND
;
4376 pool
->flags
|= POOL_DISASSOCIATED
;
4378 spin_unlock_irq(&pool
->lock
);
4379 mutex_unlock(&pool
->manager_mutex
);
4383 * Call schedule() so that we cross rq->lock and thus can guarantee
4384 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4385 * as scheduler callbacks may be invoked from other cpus.
4390 * Sched callbacks are disabled now. Zap nr_running. After this,
4391 * nr_running stays zero and need_more_worker() and keep_working()
4392 * are always true as long as the worklist is not empty. Pools on
4393 * @cpu now behave as unbound (in terms of concurrency management)
4394 * pools which are served by workers tied to the CPU.
4396 * On return from this function, the current worker would trigger
4397 * unbound chain execution of pending work items if other workers
4400 for_each_cpu_worker_pool(pool
, cpu
)
4401 atomic_set(&pool
->nr_running
, 0);
4405 * rebind_workers - rebind all workers of a pool to the associated CPU
4406 * @pool: pool of interest
4408 * @pool->cpu is coming online. Rebind all workers to the CPU.
4410 static void rebind_workers(struct worker_pool
*pool
)
4412 struct worker
*worker
;
4415 lockdep_assert_held(&pool
->manager_mutex
);
4418 * Restore CPU affinity of all workers. As all idle workers should
4419 * be on the run-queue of the associated CPU before any local
4420 * wake-ups for concurrency management happen, restore CPU affinty
4421 * of all workers first and then clear UNBOUND. As we're called
4422 * from CPU_ONLINE, the following shouldn't fail.
4424 for_each_pool_worker(worker
, wi
, pool
)
4425 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4426 pool
->attrs
->cpumask
) < 0);
4428 spin_lock_irq(&pool
->lock
);
4430 for_each_pool_worker(worker
, wi
, pool
) {
4431 unsigned int worker_flags
= worker
->flags
;
4434 * A bound idle worker should actually be on the runqueue
4435 * of the associated CPU for local wake-ups targeting it to
4436 * work. Kick all idle workers so that they migrate to the
4437 * associated CPU. Doing this in the same loop as
4438 * replacing UNBOUND with REBOUND is safe as no worker will
4439 * be bound before @pool->lock is released.
4441 if (worker_flags
& WORKER_IDLE
)
4442 wake_up_process(worker
->task
);
4445 * We want to clear UNBOUND but can't directly call
4446 * worker_clr_flags() or adjust nr_running. Atomically
4447 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4448 * @worker will clear REBOUND using worker_clr_flags() when
4449 * it initiates the next execution cycle thus restoring
4450 * concurrency management. Note that when or whether
4451 * @worker clears REBOUND doesn't affect correctness.
4453 * ACCESS_ONCE() is necessary because @worker->flags may be
4454 * tested without holding any lock in
4455 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4456 * fail incorrectly leading to premature concurrency
4457 * management operations.
4459 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4460 worker_flags
|= WORKER_REBOUND
;
4461 worker_flags
&= ~WORKER_UNBOUND
;
4462 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4465 spin_unlock_irq(&pool
->lock
);
4469 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4470 * @pool: unbound pool of interest
4471 * @cpu: the CPU which is coming up
4473 * An unbound pool may end up with a cpumask which doesn't have any online
4474 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4475 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4476 * online CPU before, cpus_allowed of all its workers should be restored.
4478 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4480 static cpumask_t cpumask
;
4481 struct worker
*worker
;
4484 lockdep_assert_held(&pool
->manager_mutex
);
4486 /* is @cpu allowed for @pool? */
4487 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4490 /* is @cpu the only online CPU? */
4491 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4492 if (cpumask_weight(&cpumask
) != 1)
4495 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4496 for_each_pool_worker(worker
, wi
, pool
)
4497 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4498 pool
->attrs
->cpumask
) < 0);
4502 * Workqueues should be brought up before normal priority CPU notifiers.
4503 * This will be registered high priority CPU notifier.
4505 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4506 unsigned long action
,
4509 int cpu
= (unsigned long)hcpu
;
4510 struct worker_pool
*pool
;
4511 struct workqueue_struct
*wq
;
4514 switch (action
& ~CPU_TASKS_FROZEN
) {
4515 case CPU_UP_PREPARE
:
4516 for_each_cpu_worker_pool(pool
, cpu
) {
4517 if (pool
->nr_workers
)
4519 if (create_and_start_worker(pool
) < 0)
4524 case CPU_DOWN_FAILED
:
4526 mutex_lock(&wq_pool_mutex
);
4528 for_each_pool(pool
, pi
) {
4529 mutex_lock(&pool
->manager_mutex
);
4531 if (pool
->cpu
== cpu
) {
4532 spin_lock_irq(&pool
->lock
);
4533 pool
->flags
&= ~POOL_DISASSOCIATED
;
4534 spin_unlock_irq(&pool
->lock
);
4536 rebind_workers(pool
);
4537 } else if (pool
->cpu
< 0) {
4538 restore_unbound_workers_cpumask(pool
, cpu
);
4541 mutex_unlock(&pool
->manager_mutex
);
4544 /* update NUMA affinity of unbound workqueues */
4545 list_for_each_entry(wq
, &workqueues
, list
)
4546 wq_update_unbound_numa(wq
, cpu
, true);
4548 mutex_unlock(&wq_pool_mutex
);
4555 * Workqueues should be brought down after normal priority CPU notifiers.
4556 * This will be registered as low priority CPU notifier.
4558 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4559 unsigned long action
,
4562 int cpu
= (unsigned long)hcpu
;
4563 struct work_struct unbind_work
;
4564 struct workqueue_struct
*wq
;
4566 switch (action
& ~CPU_TASKS_FROZEN
) {
4567 case CPU_DOWN_PREPARE
:
4568 /* unbinding per-cpu workers should happen on the local CPU */
4569 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4570 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4572 /* update NUMA affinity of unbound workqueues */
4573 mutex_lock(&wq_pool_mutex
);
4574 list_for_each_entry(wq
, &workqueues
, list
)
4575 wq_update_unbound_numa(wq
, cpu
, false);
4576 mutex_unlock(&wq_pool_mutex
);
4578 /* wait for per-cpu unbinding to finish */
4579 flush_work(&unbind_work
);
4587 struct work_for_cpu
{
4588 struct work_struct work
;
4594 static void work_for_cpu_fn(struct work_struct
*work
)
4596 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4598 wfc
->ret
= wfc
->fn(wfc
->arg
);
4602 * work_on_cpu - run a function in user context on a particular cpu
4603 * @cpu: the cpu to run on
4604 * @fn: the function to run
4605 * @arg: the function arg
4607 * This will return the value @fn returns.
4608 * It is up to the caller to ensure that the cpu doesn't go offline.
4609 * The caller must not hold any locks which would prevent @fn from completing.
4611 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4613 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4615 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4616 schedule_work_on(cpu
, &wfc
.work
);
4617 flush_work(&wfc
.work
);
4620 EXPORT_SYMBOL_GPL(work_on_cpu
);
4621 #endif /* CONFIG_SMP */
4623 #ifdef CONFIG_FREEZER
4626 * freeze_workqueues_begin - begin freezing workqueues
4628 * Start freezing workqueues. After this function returns, all freezable
4629 * workqueues will queue new works to their delayed_works list instead of
4633 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4635 void freeze_workqueues_begin(void)
4637 struct worker_pool
*pool
;
4638 struct workqueue_struct
*wq
;
4639 struct pool_workqueue
*pwq
;
4642 mutex_lock(&wq_pool_mutex
);
4644 WARN_ON_ONCE(workqueue_freezing
);
4645 workqueue_freezing
= true;
4648 for_each_pool(pool
, pi
) {
4649 spin_lock_irq(&pool
->lock
);
4650 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4651 pool
->flags
|= POOL_FREEZING
;
4652 spin_unlock_irq(&pool
->lock
);
4655 list_for_each_entry(wq
, &workqueues
, list
) {
4656 mutex_lock(&wq
->mutex
);
4657 for_each_pwq(pwq
, wq
)
4658 pwq_adjust_max_active(pwq
);
4659 mutex_unlock(&wq
->mutex
);
4662 mutex_unlock(&wq_pool_mutex
);
4666 * freeze_workqueues_busy - are freezable workqueues still busy?
4668 * Check whether freezing is complete. This function must be called
4669 * between freeze_workqueues_begin() and thaw_workqueues().
4672 * Grabs and releases wq_pool_mutex.
4675 * %true if some freezable workqueues are still busy. %false if freezing
4678 bool freeze_workqueues_busy(void)
4681 struct workqueue_struct
*wq
;
4682 struct pool_workqueue
*pwq
;
4684 mutex_lock(&wq_pool_mutex
);
4686 WARN_ON_ONCE(!workqueue_freezing
);
4688 list_for_each_entry(wq
, &workqueues
, list
) {
4689 if (!(wq
->flags
& WQ_FREEZABLE
))
4692 * nr_active is monotonically decreasing. It's safe
4693 * to peek without lock.
4695 rcu_read_lock_sched();
4696 for_each_pwq(pwq
, wq
) {
4697 WARN_ON_ONCE(pwq
->nr_active
< 0);
4698 if (pwq
->nr_active
) {
4700 rcu_read_unlock_sched();
4704 rcu_read_unlock_sched();
4707 mutex_unlock(&wq_pool_mutex
);
4712 * thaw_workqueues - thaw workqueues
4714 * Thaw workqueues. Normal queueing is restored and all collected
4715 * frozen works are transferred to their respective pool worklists.
4718 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4720 void thaw_workqueues(void)
4722 struct workqueue_struct
*wq
;
4723 struct pool_workqueue
*pwq
;
4724 struct worker_pool
*pool
;
4727 mutex_lock(&wq_pool_mutex
);
4729 if (!workqueue_freezing
)
4732 /* clear FREEZING */
4733 for_each_pool(pool
, pi
) {
4734 spin_lock_irq(&pool
->lock
);
4735 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4736 pool
->flags
&= ~POOL_FREEZING
;
4737 spin_unlock_irq(&pool
->lock
);
4740 /* restore max_active and repopulate worklist */
4741 list_for_each_entry(wq
, &workqueues
, list
) {
4742 mutex_lock(&wq
->mutex
);
4743 for_each_pwq(pwq
, wq
)
4744 pwq_adjust_max_active(pwq
);
4745 mutex_unlock(&wq
->mutex
);
4748 workqueue_freezing
= false;
4750 mutex_unlock(&wq_pool_mutex
);
4752 #endif /* CONFIG_FREEZER */
4754 static void __init
wq_numa_init(void)
4759 /* determine NUMA pwq table len - highest node id + 1 */
4761 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4763 if (num_possible_nodes() <= 1)
4766 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4767 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4770 * We want masks of possible CPUs of each node which isn't readily
4771 * available. Build one from cpu_to_node() which should have been
4772 * fully initialized by now.
4774 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
4778 BUG_ON(!alloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
, node
));
4780 for_each_possible_cpu(cpu
) {
4781 node
= cpu_to_node(cpu
);
4782 if (WARN_ON(node
== NUMA_NO_NODE
)) {
4783 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
4784 /* happens iff arch is bonkers, let's just proceed */
4787 cpumask_set_cpu(cpu
, tbl
[node
]);
4790 wq_numa_possible_cpumask
= tbl
;
4791 wq_numa_enabled
= true;
4794 static int __init
init_workqueues(void)
4796 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4799 /* make sure we have enough bits for OFFQ pool ID */
4800 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
4801 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
4803 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4805 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4807 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4808 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4812 /* initialize CPU pools */
4813 for_each_possible_cpu(cpu
) {
4814 struct worker_pool
*pool
;
4817 for_each_cpu_worker_pool(pool
, cpu
) {
4818 BUG_ON(init_worker_pool(pool
));
4820 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4821 pool
->attrs
->nice
= std_nice
[i
++];
4822 pool
->node
= cpu_to_node(cpu
);
4825 mutex_lock(&wq_pool_mutex
);
4826 BUG_ON(worker_pool_assign_id(pool
));
4827 mutex_unlock(&wq_pool_mutex
);
4831 /* create the initial worker */
4832 for_each_online_cpu(cpu
) {
4833 struct worker_pool
*pool
;
4835 for_each_cpu_worker_pool(pool
, cpu
) {
4836 pool
->flags
&= ~POOL_DISASSOCIATED
;
4837 BUG_ON(create_and_start_worker(pool
) < 0);
4841 /* create default unbound wq attrs */
4842 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4843 struct workqueue_attrs
*attrs
;
4845 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4846 attrs
->nice
= std_nice
[i
];
4847 unbound_std_wq_attrs
[i
] = attrs
;
4850 system_wq
= alloc_workqueue("events", 0, 0);
4851 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4852 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4853 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4854 WQ_UNBOUND_MAX_ACTIVE
);
4855 system_freezable_wq
= alloc_workqueue("events_freezable",
4857 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
4858 !system_unbound_wq
|| !system_freezable_wq
);
4861 early_initcall(init_workqueues
);