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_disable_numa
;
272 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
274 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
276 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
277 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
279 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
280 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
282 static LIST_HEAD(workqueues
); /* PL: list of all workqueues */
283 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
285 /* the per-cpu worker pools */
286 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
289 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
291 /* PL: hash of all unbound pools keyed by pool->attrs */
292 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
294 /* I: attributes used when instantiating standard unbound pools on demand */
295 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
297 struct workqueue_struct
*system_wq __read_mostly
;
298 EXPORT_SYMBOL_GPL(system_wq
);
299 struct workqueue_struct
*system_highpri_wq __read_mostly
;
300 EXPORT_SYMBOL_GPL(system_highpri_wq
);
301 struct workqueue_struct
*system_long_wq __read_mostly
;
302 EXPORT_SYMBOL_GPL(system_long_wq
);
303 struct workqueue_struct
*system_unbound_wq __read_mostly
;
304 EXPORT_SYMBOL_GPL(system_unbound_wq
);
305 struct workqueue_struct
*system_freezable_wq __read_mostly
;
306 EXPORT_SYMBOL_GPL(system_freezable_wq
);
308 static int worker_thread(void *__worker
);
309 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
310 const struct workqueue_attrs
*from
);
312 #define CREATE_TRACE_POINTS
313 #include <trace/events/workqueue.h>
315 #define assert_rcu_or_pool_mutex() \
316 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
317 lockdep_is_held(&wq_pool_mutex), \
318 "sched RCU or wq_pool_mutex should be held")
320 #define assert_rcu_or_wq_mutex(wq) \
321 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
322 lockdep_is_held(&wq->mutex), \
323 "sched RCU or wq->mutex should be held")
325 #ifdef CONFIG_LOCKDEP
326 #define assert_manager_or_pool_lock(pool) \
327 WARN_ONCE(debug_locks && \
328 !lockdep_is_held(&(pool)->manager_mutex) && \
329 !lockdep_is_held(&(pool)->lock), \
330 "pool->manager_mutex or ->lock should be held")
332 #define assert_manager_or_pool_lock(pool) do { } while (0)
335 #define for_each_cpu_worker_pool(pool, cpu) \
336 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
337 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
341 * for_each_pool - iterate through all worker_pools in the system
342 * @pool: iteration cursor
343 * @pi: integer used for iteration
345 * This must be called either with wq_pool_mutex held or sched RCU read
346 * locked. If the pool needs to be used beyond the locking in effect, the
347 * caller is responsible for guaranteeing that the pool stays online.
349 * The if/else clause exists only for the lockdep assertion and can be
352 #define for_each_pool(pool, pi) \
353 idr_for_each_entry(&worker_pool_idr, pool, pi) \
354 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
358 * for_each_pool_worker - iterate through all workers of a worker_pool
359 * @worker: iteration cursor
360 * @wi: integer used for iteration
361 * @pool: worker_pool to iterate workers of
363 * This must be called with either @pool->manager_mutex or ->lock held.
365 * The if/else clause exists only for the lockdep assertion and can be
368 #define for_each_pool_worker(worker, wi, pool) \
369 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
370 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
374 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
375 * @pwq: iteration cursor
376 * @wq: the target workqueue
378 * This must be called either with wq->mutex held or sched RCU read locked.
379 * If the pwq needs to be used beyond the locking in effect, the caller is
380 * responsible for guaranteeing that the pwq stays online.
382 * The if/else clause exists only for the lockdep assertion and can be
385 #define for_each_pwq(pwq, wq) \
386 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
387 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
390 #ifdef CONFIG_DEBUG_OBJECTS_WORK
392 static struct debug_obj_descr work_debug_descr
;
394 static void *work_debug_hint(void *addr
)
396 return ((struct work_struct
*) addr
)->func
;
400 * fixup_init is called when:
401 * - an active object is initialized
403 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
405 struct work_struct
*work
= addr
;
408 case ODEBUG_STATE_ACTIVE
:
409 cancel_work_sync(work
);
410 debug_object_init(work
, &work_debug_descr
);
418 * fixup_activate is called when:
419 * - an active object is activated
420 * - an unknown object is activated (might be a statically initialized object)
422 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
424 struct work_struct
*work
= addr
;
428 case ODEBUG_STATE_NOTAVAILABLE
:
430 * This is not really a fixup. The work struct was
431 * statically initialized. We just make sure that it
432 * is tracked in the object tracker.
434 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
435 debug_object_init(work
, &work_debug_descr
);
436 debug_object_activate(work
, &work_debug_descr
);
442 case ODEBUG_STATE_ACTIVE
:
451 * fixup_free is called when:
452 * - an active object is freed
454 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
456 struct work_struct
*work
= addr
;
459 case ODEBUG_STATE_ACTIVE
:
460 cancel_work_sync(work
);
461 debug_object_free(work
, &work_debug_descr
);
468 static struct debug_obj_descr work_debug_descr
= {
469 .name
= "work_struct",
470 .debug_hint
= work_debug_hint
,
471 .fixup_init
= work_fixup_init
,
472 .fixup_activate
= work_fixup_activate
,
473 .fixup_free
= work_fixup_free
,
476 static inline void debug_work_activate(struct work_struct
*work
)
478 debug_object_activate(work
, &work_debug_descr
);
481 static inline void debug_work_deactivate(struct work_struct
*work
)
483 debug_object_deactivate(work
, &work_debug_descr
);
486 void __init_work(struct work_struct
*work
, int onstack
)
489 debug_object_init_on_stack(work
, &work_debug_descr
);
491 debug_object_init(work
, &work_debug_descr
);
493 EXPORT_SYMBOL_GPL(__init_work
);
495 void destroy_work_on_stack(struct work_struct
*work
)
497 debug_object_free(work
, &work_debug_descr
);
499 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
502 static inline void debug_work_activate(struct work_struct
*work
) { }
503 static inline void debug_work_deactivate(struct work_struct
*work
) { }
506 /* allocate ID and assign it to @pool */
507 static int worker_pool_assign_id(struct worker_pool
*pool
)
511 lockdep_assert_held(&wq_pool_mutex
);
513 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, 0, GFP_KERNEL
);
522 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
523 * @wq: the target workqueue
526 * This must be called either with pwq_lock held or sched RCU read locked.
527 * If the pwq needs to be used beyond the locking in effect, the caller is
528 * responsible for guaranteeing that the pwq stays online.
530 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
533 assert_rcu_or_wq_mutex(wq
);
534 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
537 static unsigned int work_color_to_flags(int color
)
539 return color
<< WORK_STRUCT_COLOR_SHIFT
;
542 static int get_work_color(struct work_struct
*work
)
544 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
545 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
548 static int work_next_color(int color
)
550 return (color
+ 1) % WORK_NR_COLORS
;
554 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
555 * contain the pointer to the queued pwq. Once execution starts, the flag
556 * is cleared and the high bits contain OFFQ flags and pool ID.
558 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
559 * and clear_work_data() can be used to set the pwq, pool or clear
560 * work->data. These functions should only be called while the work is
561 * owned - ie. while the PENDING bit is set.
563 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
564 * corresponding to a work. Pool is available once the work has been
565 * queued anywhere after initialization until it is sync canceled. pwq is
566 * available only while the work item is queued.
568 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
569 * canceled. While being canceled, a work item may have its PENDING set
570 * but stay off timer and worklist for arbitrarily long and nobody should
571 * try to steal the PENDING bit.
573 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
576 WARN_ON_ONCE(!work_pending(work
));
577 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
580 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
581 unsigned long extra_flags
)
583 set_work_data(work
, (unsigned long)pwq
,
584 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
587 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
590 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
591 WORK_STRUCT_PENDING
);
594 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
598 * The following wmb is paired with the implied mb in
599 * test_and_set_bit(PENDING) and ensures all updates to @work made
600 * here are visible to and precede any updates by the next PENDING
604 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
607 static void clear_work_data(struct work_struct
*work
)
609 smp_wmb(); /* see set_work_pool_and_clear_pending() */
610 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
613 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
615 unsigned long data
= atomic_long_read(&work
->data
);
617 if (data
& WORK_STRUCT_PWQ
)
618 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
624 * get_work_pool - return the worker_pool a given work was associated with
625 * @work: the work item of interest
627 * Return the worker_pool @work was last associated with. %NULL if none.
629 * Pools are created and destroyed under wq_pool_mutex, and allows read
630 * access under sched-RCU read lock. As such, this function should be
631 * called under wq_pool_mutex or with preemption disabled.
633 * All fields of the returned pool are accessible as long as the above
634 * mentioned locking is in effect. If the returned pool needs to be used
635 * beyond the critical section, the caller is responsible for ensuring the
636 * returned pool is and stays online.
638 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
640 unsigned long data
= atomic_long_read(&work
->data
);
643 assert_rcu_or_pool_mutex();
645 if (data
& WORK_STRUCT_PWQ
)
646 return ((struct pool_workqueue
*)
647 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
649 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
650 if (pool_id
== WORK_OFFQ_POOL_NONE
)
653 return idr_find(&worker_pool_idr
, pool_id
);
657 * get_work_pool_id - return the worker pool ID a given work is associated with
658 * @work: the work item of interest
660 * Return the worker_pool ID @work was last associated with.
661 * %WORK_OFFQ_POOL_NONE if none.
663 static int get_work_pool_id(struct work_struct
*work
)
665 unsigned long data
= atomic_long_read(&work
->data
);
667 if (data
& WORK_STRUCT_PWQ
)
668 return ((struct pool_workqueue
*)
669 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
671 return data
>> WORK_OFFQ_POOL_SHIFT
;
674 static void mark_work_canceling(struct work_struct
*work
)
676 unsigned long pool_id
= get_work_pool_id(work
);
678 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
679 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
682 static bool work_is_canceling(struct work_struct
*work
)
684 unsigned long data
= atomic_long_read(&work
->data
);
686 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
690 * Policy functions. These define the policies on how the global worker
691 * pools are managed. Unless noted otherwise, these functions assume that
692 * they're being called with pool->lock held.
695 static bool __need_more_worker(struct worker_pool
*pool
)
697 return !atomic_read(&pool
->nr_running
);
701 * Need to wake up a worker? Called from anything but currently
704 * Note that, because unbound workers never contribute to nr_running, this
705 * function will always return %true for unbound pools as long as the
706 * worklist isn't empty.
708 static bool need_more_worker(struct worker_pool
*pool
)
710 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
713 /* Can I start working? Called from busy but !running workers. */
714 static bool may_start_working(struct worker_pool
*pool
)
716 return pool
->nr_idle
;
719 /* Do I need to keep working? Called from currently running workers. */
720 static bool keep_working(struct worker_pool
*pool
)
722 return !list_empty(&pool
->worklist
) &&
723 atomic_read(&pool
->nr_running
) <= 1;
726 /* Do we need a new worker? Called from manager. */
727 static bool need_to_create_worker(struct worker_pool
*pool
)
729 return need_more_worker(pool
) && !may_start_working(pool
);
732 /* Do I need to be the manager? */
733 static bool need_to_manage_workers(struct worker_pool
*pool
)
735 return need_to_create_worker(pool
) ||
736 (pool
->flags
& POOL_MANAGE_WORKERS
);
739 /* Do we have too many workers and should some go away? */
740 static bool too_many_workers(struct worker_pool
*pool
)
742 bool managing
= mutex_is_locked(&pool
->manager_arb
);
743 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
744 int nr_busy
= pool
->nr_workers
- nr_idle
;
747 * nr_idle and idle_list may disagree if idle rebinding is in
748 * progress. Never return %true if idle_list is empty.
750 if (list_empty(&pool
->idle_list
))
753 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
760 /* Return the first worker. Safe with preemption disabled */
761 static struct worker
*first_worker(struct worker_pool
*pool
)
763 if (unlikely(list_empty(&pool
->idle_list
)))
766 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
770 * wake_up_worker - wake up an idle worker
771 * @pool: worker pool to wake worker from
773 * Wake up the first idle worker of @pool.
776 * spin_lock_irq(pool->lock).
778 static void wake_up_worker(struct worker_pool
*pool
)
780 struct worker
*worker
= first_worker(pool
);
783 wake_up_process(worker
->task
);
787 * wq_worker_waking_up - a worker is waking up
788 * @task: task waking up
789 * @cpu: CPU @task is waking up to
791 * This function is called during try_to_wake_up() when a worker is
795 * spin_lock_irq(rq->lock)
797 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
799 struct worker
*worker
= kthread_data(task
);
801 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
802 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
803 atomic_inc(&worker
->pool
->nr_running
);
808 * wq_worker_sleeping - a worker is going to sleep
809 * @task: task going to sleep
810 * @cpu: CPU in question, must be the current CPU number
812 * This function is called during schedule() when a busy worker is
813 * going to sleep. Worker on the same cpu can be woken up by
814 * returning pointer to its task.
817 * spin_lock_irq(rq->lock)
820 * Worker task on @cpu to wake up, %NULL if none.
822 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
824 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
825 struct worker_pool
*pool
;
828 * Rescuers, which may not have all the fields set up like normal
829 * workers, also reach here, let's not access anything before
830 * checking NOT_RUNNING.
832 if (worker
->flags
& WORKER_NOT_RUNNING
)
837 /* this can only happen on the local cpu */
838 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
842 * The counterpart of the following dec_and_test, implied mb,
843 * worklist not empty test sequence is in insert_work().
844 * Please read comment there.
846 * NOT_RUNNING is clear. This means that we're bound to and
847 * running on the local cpu w/ rq lock held and preemption
848 * disabled, which in turn means that none else could be
849 * manipulating idle_list, so dereferencing idle_list without pool
852 if (atomic_dec_and_test(&pool
->nr_running
) &&
853 !list_empty(&pool
->worklist
))
854 to_wakeup
= first_worker(pool
);
855 return to_wakeup
? to_wakeup
->task
: NULL
;
859 * worker_set_flags - set worker flags and adjust nr_running accordingly
861 * @flags: flags to set
862 * @wakeup: wakeup an idle worker if necessary
864 * Set @flags in @worker->flags and adjust nr_running accordingly. If
865 * nr_running becomes zero and @wakeup is %true, an idle worker is
869 * spin_lock_irq(pool->lock)
871 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
874 struct worker_pool
*pool
= worker
->pool
;
876 WARN_ON_ONCE(worker
->task
!= current
);
879 * If transitioning into NOT_RUNNING, adjust nr_running and
880 * wake up an idle worker as necessary if requested by
883 if ((flags
& WORKER_NOT_RUNNING
) &&
884 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
886 if (atomic_dec_and_test(&pool
->nr_running
) &&
887 !list_empty(&pool
->worklist
))
888 wake_up_worker(pool
);
890 atomic_dec(&pool
->nr_running
);
893 worker
->flags
|= flags
;
897 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
899 * @flags: flags to clear
901 * Clear @flags in @worker->flags and adjust nr_running accordingly.
904 * spin_lock_irq(pool->lock)
906 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
908 struct worker_pool
*pool
= worker
->pool
;
909 unsigned int oflags
= worker
->flags
;
911 WARN_ON_ONCE(worker
->task
!= current
);
913 worker
->flags
&= ~flags
;
916 * If transitioning out of NOT_RUNNING, increment nr_running. Note
917 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
918 * of multiple flags, not a single flag.
920 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
921 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
922 atomic_inc(&pool
->nr_running
);
926 * find_worker_executing_work - find worker which is executing a work
927 * @pool: pool of interest
928 * @work: work to find worker for
930 * Find a worker which is executing @work on @pool by searching
931 * @pool->busy_hash which is keyed by the address of @work. For a worker
932 * to match, its current execution should match the address of @work and
933 * its work function. This is to avoid unwanted dependency between
934 * unrelated work executions through a work item being recycled while still
937 * This is a bit tricky. A work item may be freed once its execution
938 * starts and nothing prevents the freed area from being recycled for
939 * another work item. If the same work item address ends up being reused
940 * before the original execution finishes, workqueue will identify the
941 * recycled work item as currently executing and make it wait until the
942 * current execution finishes, introducing an unwanted dependency.
944 * This function checks the work item address and work function to avoid
945 * false positives. Note that this isn't complete as one may construct a
946 * work function which can introduce dependency onto itself through a
947 * recycled work item. Well, if somebody wants to shoot oneself in the
948 * foot that badly, there's only so much we can do, and if such deadlock
949 * actually occurs, it should be easy to locate the culprit work function.
952 * spin_lock_irq(pool->lock).
955 * Pointer to worker which is executing @work if found, NULL
958 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
959 struct work_struct
*work
)
961 struct worker
*worker
;
963 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
965 if (worker
->current_work
== work
&&
966 worker
->current_func
== work
->func
)
973 * move_linked_works - move linked works to a list
974 * @work: start of series of works to be scheduled
975 * @head: target list to append @work to
976 * @nextp: out paramter for nested worklist walking
978 * Schedule linked works starting from @work to @head. Work series to
979 * be scheduled starts at @work and includes any consecutive work with
980 * WORK_STRUCT_LINKED set in its predecessor.
982 * If @nextp is not NULL, it's updated to point to the next work of
983 * the last scheduled work. This allows move_linked_works() to be
984 * nested inside outer list_for_each_entry_safe().
987 * spin_lock_irq(pool->lock).
989 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
990 struct work_struct
**nextp
)
992 struct work_struct
*n
;
995 * Linked worklist will always end before the end of the list,
996 * use NULL for list head.
998 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
999 list_move_tail(&work
->entry
, head
);
1000 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1005 * If we're already inside safe list traversal and have moved
1006 * multiple works to the scheduled queue, the next position
1007 * needs to be updated.
1014 * get_pwq - get an extra reference on the specified pool_workqueue
1015 * @pwq: pool_workqueue to get
1017 * Obtain an extra reference on @pwq. The caller should guarantee that
1018 * @pwq has positive refcnt and be holding the matching pool->lock.
1020 static void get_pwq(struct pool_workqueue
*pwq
)
1022 lockdep_assert_held(&pwq
->pool
->lock
);
1023 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1028 * put_pwq - put a pool_workqueue reference
1029 * @pwq: pool_workqueue to put
1031 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1032 * destruction. The caller should be holding the matching pool->lock.
1034 static void put_pwq(struct pool_workqueue
*pwq
)
1036 lockdep_assert_held(&pwq
->pool
->lock
);
1037 if (likely(--pwq
->refcnt
))
1039 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1042 * @pwq can't be released under pool->lock, bounce to
1043 * pwq_unbound_release_workfn(). This never recurses on the same
1044 * pool->lock as this path is taken only for unbound workqueues and
1045 * the release work item is scheduled on a per-cpu workqueue. To
1046 * avoid lockdep warning, unbound pool->locks are given lockdep
1047 * subclass of 1 in get_unbound_pool().
1049 schedule_work(&pwq
->unbound_release_work
);
1053 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1054 * @pwq: pool_workqueue to put (can be %NULL)
1056 * put_pwq() with locking. This function also allows %NULL @pwq.
1058 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1062 * As both pwqs and pools are sched-RCU protected, the
1063 * following lock operations are safe.
1065 spin_lock_irq(&pwq
->pool
->lock
);
1067 spin_unlock_irq(&pwq
->pool
->lock
);
1071 static void pwq_activate_delayed_work(struct work_struct
*work
)
1073 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1075 trace_workqueue_activate_work(work
);
1076 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1077 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1081 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1083 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1084 struct work_struct
, entry
);
1086 pwq_activate_delayed_work(work
);
1090 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1091 * @pwq: pwq of interest
1092 * @color: color of work which left the queue
1094 * A work either has completed or is removed from pending queue,
1095 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1098 * spin_lock_irq(pool->lock).
1100 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1102 /* uncolored work items don't participate in flushing or nr_active */
1103 if (color
== WORK_NO_COLOR
)
1106 pwq
->nr_in_flight
[color
]--;
1109 if (!list_empty(&pwq
->delayed_works
)) {
1110 /* one down, submit a delayed one */
1111 if (pwq
->nr_active
< pwq
->max_active
)
1112 pwq_activate_first_delayed(pwq
);
1115 /* is flush in progress and are we at the flushing tip? */
1116 if (likely(pwq
->flush_color
!= color
))
1119 /* are there still in-flight works? */
1120 if (pwq
->nr_in_flight
[color
])
1123 /* this pwq is done, clear flush_color */
1124 pwq
->flush_color
= -1;
1127 * If this was the last pwq, wake up the first flusher. It
1128 * will handle the rest.
1130 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1131 complete(&pwq
->wq
->first_flusher
->done
);
1137 * try_to_grab_pending - steal work item from worklist and disable irq
1138 * @work: work item to steal
1139 * @is_dwork: @work is a delayed_work
1140 * @flags: place to store irq state
1142 * Try to grab PENDING bit of @work. This function can handle @work in any
1143 * stable state - idle, on timer or on worklist. Return values are
1145 * 1 if @work was pending and we successfully stole PENDING
1146 * 0 if @work was idle and we claimed PENDING
1147 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1148 * -ENOENT if someone else is canceling @work, this state may persist
1149 * for arbitrarily long
1151 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1152 * interrupted while holding PENDING and @work off queue, irq must be
1153 * disabled on entry. This, combined with delayed_work->timer being
1154 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1156 * On successful return, >= 0, irq is disabled and the caller is
1157 * responsible for releasing it using local_irq_restore(*@flags).
1159 * This function is safe to call from any context including IRQ handler.
1161 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1162 unsigned long *flags
)
1164 struct worker_pool
*pool
;
1165 struct pool_workqueue
*pwq
;
1167 local_irq_save(*flags
);
1169 /* try to steal the timer if it exists */
1171 struct delayed_work
*dwork
= to_delayed_work(work
);
1174 * dwork->timer is irqsafe. If del_timer() fails, it's
1175 * guaranteed that the timer is not queued anywhere and not
1176 * running on the local CPU.
1178 if (likely(del_timer(&dwork
->timer
)))
1182 /* try to claim PENDING the normal way */
1183 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1187 * The queueing is in progress, or it is already queued. Try to
1188 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1190 pool
= get_work_pool(work
);
1194 spin_lock(&pool
->lock
);
1196 * work->data is guaranteed to point to pwq only while the work
1197 * item is queued on pwq->wq, and both updating work->data to point
1198 * to pwq on queueing and to pool on dequeueing are done under
1199 * pwq->pool->lock. This in turn guarantees that, if work->data
1200 * points to pwq which is associated with a locked pool, the work
1201 * item is currently queued on that pool.
1203 pwq
= get_work_pwq(work
);
1204 if (pwq
&& pwq
->pool
== pool
) {
1205 debug_work_deactivate(work
);
1208 * A delayed work item cannot be grabbed directly because
1209 * it might have linked NO_COLOR work items which, if left
1210 * on the delayed_list, will confuse pwq->nr_active
1211 * management later on and cause stall. Make sure the work
1212 * item is activated before grabbing.
1214 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1215 pwq_activate_delayed_work(work
);
1217 list_del_init(&work
->entry
);
1218 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1220 /* work->data points to pwq iff queued, point to pool */
1221 set_work_pool_and_keep_pending(work
, pool
->id
);
1223 spin_unlock(&pool
->lock
);
1226 spin_unlock(&pool
->lock
);
1228 local_irq_restore(*flags
);
1229 if (work_is_canceling(work
))
1236 * insert_work - insert a work into a pool
1237 * @pwq: pwq @work belongs to
1238 * @work: work to insert
1239 * @head: insertion point
1240 * @extra_flags: extra WORK_STRUCT_* flags to set
1242 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1243 * work_struct flags.
1246 * spin_lock_irq(pool->lock).
1248 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1249 struct list_head
*head
, unsigned int extra_flags
)
1251 struct worker_pool
*pool
= pwq
->pool
;
1253 /* we own @work, set data and link */
1254 set_work_pwq(work
, pwq
, extra_flags
);
1255 list_add_tail(&work
->entry
, head
);
1259 * Ensure either wq_worker_sleeping() sees the above
1260 * list_add_tail() or we see zero nr_running to avoid workers lying
1261 * around lazily while there are works to be processed.
1265 if (__need_more_worker(pool
))
1266 wake_up_worker(pool
);
1270 * Test whether @work is being queued from another work executing on the
1273 static bool is_chained_work(struct workqueue_struct
*wq
)
1275 struct worker
*worker
;
1277 worker
= current_wq_worker();
1279 * Return %true iff I'm a worker execuing a work item on @wq. If
1280 * I'm @worker, it's safe to dereference it without locking.
1282 return worker
&& worker
->current_pwq
->wq
== wq
;
1285 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1286 struct work_struct
*work
)
1288 struct pool_workqueue
*pwq
;
1289 struct worker_pool
*last_pool
;
1290 struct list_head
*worklist
;
1291 unsigned int work_flags
;
1292 unsigned int req_cpu
= cpu
;
1295 * While a work item is PENDING && off queue, a task trying to
1296 * steal the PENDING will busy-loop waiting for it to either get
1297 * queued or lose PENDING. Grabbing PENDING and queueing should
1298 * happen with IRQ disabled.
1300 WARN_ON_ONCE(!irqs_disabled());
1302 debug_work_activate(work
);
1304 /* if dying, only works from the same workqueue are allowed */
1305 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1306 WARN_ON_ONCE(!is_chained_work(wq
)))
1309 if (req_cpu
== WORK_CPU_UNBOUND
)
1310 cpu
= raw_smp_processor_id();
1312 /* pwq which will be used unless @work is executing elsewhere */
1313 if (!(wq
->flags
& WQ_UNBOUND
))
1314 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1316 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1319 * If @work was previously on a different pool, it might still be
1320 * running there, in which case the work needs to be queued on that
1321 * pool to guarantee non-reentrancy.
1323 last_pool
= get_work_pool(work
);
1324 if (last_pool
&& last_pool
!= pwq
->pool
) {
1325 struct worker
*worker
;
1327 spin_lock(&last_pool
->lock
);
1329 worker
= find_worker_executing_work(last_pool
, work
);
1331 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1332 pwq
= worker
->current_pwq
;
1334 /* meh... not running there, queue here */
1335 spin_unlock(&last_pool
->lock
);
1336 spin_lock(&pwq
->pool
->lock
);
1339 spin_lock(&pwq
->pool
->lock
);
1343 * pwq is determined and locked. For unbound pools, we could have
1344 * raced with pwq release and it could already be dead. If its
1345 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1346 * without another pwq replacing it in the numa_pwq_tbl or while
1347 * work items are executing on it, so the retrying is guaranteed to
1348 * make forward-progress.
1350 if (unlikely(!pwq
->refcnt
)) {
1351 if (wq
->flags
& WQ_UNBOUND
) {
1352 spin_unlock(&pwq
->pool
->lock
);
1357 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1361 /* pwq determined, queue */
1362 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1364 if (WARN_ON(!list_empty(&work
->entry
))) {
1365 spin_unlock(&pwq
->pool
->lock
);
1369 pwq
->nr_in_flight
[pwq
->work_color
]++;
1370 work_flags
= work_color_to_flags(pwq
->work_color
);
1372 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1373 trace_workqueue_activate_work(work
);
1375 worklist
= &pwq
->pool
->worklist
;
1377 work_flags
|= WORK_STRUCT_DELAYED
;
1378 worklist
= &pwq
->delayed_works
;
1381 insert_work(pwq
, work
, worklist
, work_flags
);
1383 spin_unlock(&pwq
->pool
->lock
);
1387 * queue_work_on - queue work on specific cpu
1388 * @cpu: CPU number to execute work on
1389 * @wq: workqueue to use
1390 * @work: work to queue
1392 * Returns %false if @work was already on a queue, %true otherwise.
1394 * We queue the work to a specific CPU, the caller must ensure it
1397 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1398 struct work_struct
*work
)
1401 unsigned long flags
;
1403 local_irq_save(flags
);
1405 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1406 __queue_work(cpu
, wq
, work
);
1410 local_irq_restore(flags
);
1413 EXPORT_SYMBOL_GPL(queue_work_on
);
1415 void delayed_work_timer_fn(unsigned long __data
)
1417 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1419 /* should have been called from irqsafe timer with irq already off */
1420 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1422 EXPORT_SYMBOL(delayed_work_timer_fn
);
1424 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1425 struct delayed_work
*dwork
, unsigned long delay
)
1427 struct timer_list
*timer
= &dwork
->timer
;
1428 struct work_struct
*work
= &dwork
->work
;
1430 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1431 timer
->data
!= (unsigned long)dwork
);
1432 WARN_ON_ONCE(timer_pending(timer
));
1433 WARN_ON_ONCE(!list_empty(&work
->entry
));
1436 * If @delay is 0, queue @dwork->work immediately. This is for
1437 * both optimization and correctness. The earliest @timer can
1438 * expire is on the closest next tick and delayed_work users depend
1439 * on that there's no such delay when @delay is 0.
1442 __queue_work(cpu
, wq
, &dwork
->work
);
1446 timer_stats_timer_set_start_info(&dwork
->timer
);
1450 timer
->expires
= jiffies
+ delay
;
1452 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1453 add_timer_on(timer
, cpu
);
1459 * queue_delayed_work_on - queue work on specific CPU after delay
1460 * @cpu: CPU number to execute work on
1461 * @wq: workqueue to use
1462 * @dwork: work to queue
1463 * @delay: number of jiffies to wait before queueing
1465 * Returns %false if @work was already on a queue, %true otherwise. If
1466 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1469 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1470 struct delayed_work
*dwork
, unsigned long delay
)
1472 struct work_struct
*work
= &dwork
->work
;
1474 unsigned long flags
;
1476 /* read the comment in __queue_work() */
1477 local_irq_save(flags
);
1479 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1480 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1484 local_irq_restore(flags
);
1487 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1490 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1491 * @cpu: CPU number to execute work on
1492 * @wq: workqueue to use
1493 * @dwork: work to queue
1494 * @delay: number of jiffies to wait before queueing
1496 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1497 * modify @dwork's timer so that it expires after @delay. If @delay is
1498 * zero, @work is guaranteed to be scheduled immediately regardless of its
1501 * Returns %false if @dwork was idle and queued, %true if @dwork was
1502 * pending and its timer was modified.
1504 * This function is safe to call from any context including IRQ handler.
1505 * See try_to_grab_pending() for details.
1507 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1508 struct delayed_work
*dwork
, unsigned long delay
)
1510 unsigned long flags
;
1514 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1515 } while (unlikely(ret
== -EAGAIN
));
1517 if (likely(ret
>= 0)) {
1518 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1519 local_irq_restore(flags
);
1522 /* -ENOENT from try_to_grab_pending() becomes %true */
1525 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1528 * worker_enter_idle - enter idle state
1529 * @worker: worker which is entering idle state
1531 * @worker is entering idle state. Update stats and idle timer if
1535 * spin_lock_irq(pool->lock).
1537 static void worker_enter_idle(struct worker
*worker
)
1539 struct worker_pool
*pool
= worker
->pool
;
1541 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1542 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1543 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1546 /* can't use worker_set_flags(), also called from start_worker() */
1547 worker
->flags
|= WORKER_IDLE
;
1549 worker
->last_active
= jiffies
;
1551 /* idle_list is LIFO */
1552 list_add(&worker
->entry
, &pool
->idle_list
);
1554 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1555 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1558 * Sanity check nr_running. Because wq_unbind_fn() releases
1559 * pool->lock between setting %WORKER_UNBOUND and zapping
1560 * nr_running, the warning may trigger spuriously. Check iff
1561 * unbind is not in progress.
1563 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1564 pool
->nr_workers
== pool
->nr_idle
&&
1565 atomic_read(&pool
->nr_running
));
1569 * worker_leave_idle - leave idle state
1570 * @worker: worker which is leaving idle state
1572 * @worker is leaving idle state. Update stats.
1575 * spin_lock_irq(pool->lock).
1577 static void worker_leave_idle(struct worker
*worker
)
1579 struct worker_pool
*pool
= worker
->pool
;
1581 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1583 worker_clr_flags(worker
, WORKER_IDLE
);
1585 list_del_init(&worker
->entry
);
1589 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1590 * @pool: target worker_pool
1592 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1594 * Works which are scheduled while the cpu is online must at least be
1595 * scheduled to a worker which is bound to the cpu so that if they are
1596 * flushed from cpu callbacks while cpu is going down, they are
1597 * guaranteed to execute on the cpu.
1599 * This function is to be used by unbound workers and rescuers to bind
1600 * themselves to the target cpu and may race with cpu going down or
1601 * coming online. kthread_bind() can't be used because it may put the
1602 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1603 * verbatim as it's best effort and blocking and pool may be
1604 * [dis]associated in the meantime.
1606 * This function tries set_cpus_allowed() and locks pool and verifies the
1607 * binding against %POOL_DISASSOCIATED which is set during
1608 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1609 * enters idle state or fetches works without dropping lock, it can
1610 * guarantee the scheduling requirement described in the first paragraph.
1613 * Might sleep. Called without any lock but returns with pool->lock
1617 * %true if the associated pool is online (@worker is successfully
1618 * bound), %false if offline.
1620 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1621 __acquires(&pool
->lock
)
1625 * The following call may fail, succeed or succeed
1626 * without actually migrating the task to the cpu if
1627 * it races with cpu hotunplug operation. Verify
1628 * against POOL_DISASSOCIATED.
1630 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1631 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1633 spin_lock_irq(&pool
->lock
);
1634 if (pool
->flags
& POOL_DISASSOCIATED
)
1636 if (task_cpu(current
) == pool
->cpu
&&
1637 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1639 spin_unlock_irq(&pool
->lock
);
1642 * We've raced with CPU hot[un]plug. Give it a breather
1643 * and retry migration. cond_resched() is required here;
1644 * otherwise, we might deadlock against cpu_stop trying to
1645 * bring down the CPU on non-preemptive kernel.
1652 static struct worker
*alloc_worker(void)
1654 struct worker
*worker
;
1656 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1658 INIT_LIST_HEAD(&worker
->entry
);
1659 INIT_LIST_HEAD(&worker
->scheduled
);
1660 /* on creation a worker is in !idle && prep state */
1661 worker
->flags
= WORKER_PREP
;
1667 * create_worker - create a new workqueue worker
1668 * @pool: pool the new worker will belong to
1670 * Create a new worker which is bound to @pool. The returned worker
1671 * can be started by calling start_worker() or destroyed using
1675 * Might sleep. Does GFP_KERNEL allocations.
1678 * Pointer to the newly created worker.
1680 static struct worker
*create_worker(struct worker_pool
*pool
)
1682 struct worker
*worker
= NULL
;
1686 lockdep_assert_held(&pool
->manager_mutex
);
1689 * ID is needed to determine kthread name. Allocate ID first
1690 * without installing the pointer.
1692 idr_preload(GFP_KERNEL
);
1693 spin_lock_irq(&pool
->lock
);
1695 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1697 spin_unlock_irq(&pool
->lock
);
1702 worker
= alloc_worker();
1706 worker
->pool
= pool
;
1710 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1711 pool
->attrs
->nice
< 0 ? "H" : "");
1713 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1715 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1716 "kworker/%s", id_buf
);
1717 if (IS_ERR(worker
->task
))
1721 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1722 * online CPUs. It'll be re-applied when any of the CPUs come up.
1724 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1725 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1727 /* prevent userland from meddling with cpumask of workqueue workers */
1728 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1731 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1732 * remains stable across this function. See the comments above the
1733 * flag definition for details.
1735 if (pool
->flags
& POOL_DISASSOCIATED
)
1736 worker
->flags
|= WORKER_UNBOUND
;
1738 /* successful, commit the pointer to idr */
1739 spin_lock_irq(&pool
->lock
);
1740 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1741 spin_unlock_irq(&pool
->lock
);
1747 spin_lock_irq(&pool
->lock
);
1748 idr_remove(&pool
->worker_idr
, id
);
1749 spin_unlock_irq(&pool
->lock
);
1756 * start_worker - start a newly created worker
1757 * @worker: worker to start
1759 * Make the pool aware of @worker and start it.
1762 * spin_lock_irq(pool->lock).
1764 static void start_worker(struct worker
*worker
)
1766 worker
->flags
|= WORKER_STARTED
;
1767 worker
->pool
->nr_workers
++;
1768 worker_enter_idle(worker
);
1769 wake_up_process(worker
->task
);
1773 * create_and_start_worker - create and start a worker for a pool
1774 * @pool: the target pool
1776 * Grab the managership of @pool and create and start a new worker for it.
1778 static int create_and_start_worker(struct worker_pool
*pool
)
1780 struct worker
*worker
;
1782 mutex_lock(&pool
->manager_mutex
);
1784 worker
= create_worker(pool
);
1786 spin_lock_irq(&pool
->lock
);
1787 start_worker(worker
);
1788 spin_unlock_irq(&pool
->lock
);
1791 mutex_unlock(&pool
->manager_mutex
);
1793 return worker
? 0 : -ENOMEM
;
1797 * destroy_worker - destroy a workqueue worker
1798 * @worker: worker to be destroyed
1800 * Destroy @worker and adjust @pool stats accordingly.
1803 * spin_lock_irq(pool->lock) which is released and regrabbed.
1805 static void destroy_worker(struct worker
*worker
)
1807 struct worker_pool
*pool
= worker
->pool
;
1809 lockdep_assert_held(&pool
->manager_mutex
);
1810 lockdep_assert_held(&pool
->lock
);
1812 /* sanity check frenzy */
1813 if (WARN_ON(worker
->current_work
) ||
1814 WARN_ON(!list_empty(&worker
->scheduled
)))
1817 if (worker
->flags
& WORKER_STARTED
)
1819 if (worker
->flags
& WORKER_IDLE
)
1822 list_del_init(&worker
->entry
);
1823 worker
->flags
|= WORKER_DIE
;
1825 idr_remove(&pool
->worker_idr
, worker
->id
);
1827 spin_unlock_irq(&pool
->lock
);
1829 kthread_stop(worker
->task
);
1832 spin_lock_irq(&pool
->lock
);
1835 static void idle_worker_timeout(unsigned long __pool
)
1837 struct worker_pool
*pool
= (void *)__pool
;
1839 spin_lock_irq(&pool
->lock
);
1841 if (too_many_workers(pool
)) {
1842 struct worker
*worker
;
1843 unsigned long expires
;
1845 /* idle_list is kept in LIFO order, check the last one */
1846 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1847 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1849 if (time_before(jiffies
, expires
))
1850 mod_timer(&pool
->idle_timer
, expires
);
1852 /* it's been idle for too long, wake up manager */
1853 pool
->flags
|= POOL_MANAGE_WORKERS
;
1854 wake_up_worker(pool
);
1858 spin_unlock_irq(&pool
->lock
);
1861 static void send_mayday(struct work_struct
*work
)
1863 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1864 struct workqueue_struct
*wq
= pwq
->wq
;
1866 lockdep_assert_held(&wq_mayday_lock
);
1871 /* mayday mayday mayday */
1872 if (list_empty(&pwq
->mayday_node
)) {
1873 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1874 wake_up_process(wq
->rescuer
->task
);
1878 static void pool_mayday_timeout(unsigned long __pool
)
1880 struct worker_pool
*pool
= (void *)__pool
;
1881 struct work_struct
*work
;
1883 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1884 spin_lock(&pool
->lock
);
1886 if (need_to_create_worker(pool
)) {
1888 * We've been trying to create a new worker but
1889 * haven't been successful. We might be hitting an
1890 * allocation deadlock. Send distress signals to
1893 list_for_each_entry(work
, &pool
->worklist
, entry
)
1897 spin_unlock(&pool
->lock
);
1898 spin_unlock_irq(&wq_mayday_lock
);
1900 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1904 * maybe_create_worker - create a new worker if necessary
1905 * @pool: pool to create a new worker for
1907 * Create a new worker for @pool if necessary. @pool is guaranteed to
1908 * have at least one idle worker on return from this function. If
1909 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1910 * sent to all rescuers with works scheduled on @pool to resolve
1911 * possible allocation deadlock.
1913 * On return, need_to_create_worker() is guaranteed to be %false and
1914 * may_start_working() %true.
1917 * spin_lock_irq(pool->lock) which may be released and regrabbed
1918 * multiple times. Does GFP_KERNEL allocations. Called only from
1922 * %false if no action was taken and pool->lock stayed locked, %true
1925 static bool maybe_create_worker(struct worker_pool
*pool
)
1926 __releases(&pool
->lock
)
1927 __acquires(&pool
->lock
)
1929 if (!need_to_create_worker(pool
))
1932 spin_unlock_irq(&pool
->lock
);
1934 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1935 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1938 struct worker
*worker
;
1940 worker
= create_worker(pool
);
1942 del_timer_sync(&pool
->mayday_timer
);
1943 spin_lock_irq(&pool
->lock
);
1944 start_worker(worker
);
1945 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1950 if (!need_to_create_worker(pool
))
1953 __set_current_state(TASK_INTERRUPTIBLE
);
1954 schedule_timeout(CREATE_COOLDOWN
);
1956 if (!need_to_create_worker(pool
))
1960 del_timer_sync(&pool
->mayday_timer
);
1961 spin_lock_irq(&pool
->lock
);
1962 if (need_to_create_worker(pool
))
1968 * maybe_destroy_worker - destroy workers which have been idle for a while
1969 * @pool: pool to destroy workers for
1971 * Destroy @pool workers which have been idle for longer than
1972 * IDLE_WORKER_TIMEOUT.
1975 * spin_lock_irq(pool->lock) which may be released and regrabbed
1976 * multiple times. Called only from manager.
1979 * %false if no action was taken and pool->lock stayed locked, %true
1982 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1986 while (too_many_workers(pool
)) {
1987 struct worker
*worker
;
1988 unsigned long expires
;
1990 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1991 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1993 if (time_before(jiffies
, expires
)) {
1994 mod_timer(&pool
->idle_timer
, expires
);
1998 destroy_worker(worker
);
2006 * manage_workers - manage worker pool
2009 * Assume the manager role and manage the worker pool @worker belongs
2010 * to. At any given time, there can be only zero or one manager per
2011 * pool. The exclusion is handled automatically by this function.
2013 * The caller can safely start processing works on false return. On
2014 * true return, it's guaranteed that need_to_create_worker() is false
2015 * and may_start_working() is true.
2018 * spin_lock_irq(pool->lock) which may be released and regrabbed
2019 * multiple times. Does GFP_KERNEL allocations.
2022 * spin_lock_irq(pool->lock) which may be released and regrabbed
2023 * multiple times. Does GFP_KERNEL allocations.
2025 static bool manage_workers(struct worker
*worker
)
2027 struct worker_pool
*pool
= worker
->pool
;
2031 * Managership is governed by two mutexes - manager_arb and
2032 * manager_mutex. manager_arb handles arbitration of manager role.
2033 * Anyone who successfully grabs manager_arb wins the arbitration
2034 * and becomes the manager. mutex_trylock() on pool->manager_arb
2035 * failure while holding pool->lock reliably indicates that someone
2036 * else is managing the pool and the worker which failed trylock
2037 * can proceed to executing work items. This means that anyone
2038 * grabbing manager_arb is responsible for actually performing
2039 * manager duties. If manager_arb is grabbed and released without
2040 * actual management, the pool may stall indefinitely.
2042 * manager_mutex is used for exclusion of actual management
2043 * operations. The holder of manager_mutex can be sure that none
2044 * of management operations, including creation and destruction of
2045 * workers, won't take place until the mutex is released. Because
2046 * manager_mutex doesn't interfere with manager role arbitration,
2047 * it is guaranteed that the pool's management, while may be
2048 * delayed, won't be disturbed by someone else grabbing
2051 if (!mutex_trylock(&pool
->manager_arb
))
2055 * With manager arbitration won, manager_mutex would be free in
2056 * most cases. trylock first without dropping @pool->lock.
2058 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2059 spin_unlock_irq(&pool
->lock
);
2060 mutex_lock(&pool
->manager_mutex
);
2064 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2067 * Destroy and then create so that may_start_working() is true
2070 ret
|= maybe_destroy_workers(pool
);
2071 ret
|= maybe_create_worker(pool
);
2073 mutex_unlock(&pool
->manager_mutex
);
2074 mutex_unlock(&pool
->manager_arb
);
2079 * process_one_work - process single work
2081 * @work: work to process
2083 * Process @work. This function contains all the logics necessary to
2084 * process a single work including synchronization against and
2085 * interaction with other workers on the same cpu, queueing and
2086 * flushing. As long as context requirement is met, any worker can
2087 * call this function to process a work.
2090 * spin_lock_irq(pool->lock) which is released and regrabbed.
2092 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2093 __releases(&pool
->lock
)
2094 __acquires(&pool
->lock
)
2096 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2097 struct worker_pool
*pool
= worker
->pool
;
2098 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2100 struct worker
*collision
;
2101 #ifdef CONFIG_LOCKDEP
2103 * It is permissible to free the struct work_struct from
2104 * inside the function that is called from it, this we need to
2105 * take into account for lockdep too. To avoid bogus "held
2106 * lock freed" warnings as well as problems when looking into
2107 * work->lockdep_map, make a copy and use that here.
2109 struct lockdep_map lockdep_map
;
2111 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2114 * Ensure we're on the correct CPU. DISASSOCIATED test is
2115 * necessary to avoid spurious warnings from rescuers servicing the
2116 * unbound or a disassociated pool.
2118 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2119 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2120 raw_smp_processor_id() != pool
->cpu
);
2123 * A single work shouldn't be executed concurrently by
2124 * multiple workers on a single cpu. Check whether anyone is
2125 * already processing the work. If so, defer the work to the
2126 * currently executing one.
2128 collision
= find_worker_executing_work(pool
, work
);
2129 if (unlikely(collision
)) {
2130 move_linked_works(work
, &collision
->scheduled
, NULL
);
2134 /* claim and dequeue */
2135 debug_work_deactivate(work
);
2136 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2137 worker
->current_work
= work
;
2138 worker
->current_func
= work
->func
;
2139 worker
->current_pwq
= pwq
;
2140 work_color
= get_work_color(work
);
2142 list_del_init(&work
->entry
);
2145 * CPU intensive works don't participate in concurrency
2146 * management. They're the scheduler's responsibility.
2148 if (unlikely(cpu_intensive
))
2149 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2152 * Unbound pool isn't concurrency managed and work items should be
2153 * executed ASAP. Wake up another worker if necessary.
2155 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2156 wake_up_worker(pool
);
2159 * Record the last pool and clear PENDING which should be the last
2160 * update to @work. Also, do this inside @pool->lock so that
2161 * PENDING and queued state changes happen together while IRQ is
2164 set_work_pool_and_clear_pending(work
, pool
->id
);
2166 spin_unlock_irq(&pool
->lock
);
2168 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2169 lock_map_acquire(&lockdep_map
);
2170 trace_workqueue_execute_start(work
);
2171 worker
->current_func(work
);
2173 * While we must be careful to not use "work" after this, the trace
2174 * point will only record its address.
2176 trace_workqueue_execute_end(work
);
2177 lock_map_release(&lockdep_map
);
2178 lock_map_release(&pwq
->wq
->lockdep_map
);
2180 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2181 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2182 " last function: %pf\n",
2183 current
->comm
, preempt_count(), task_pid_nr(current
),
2184 worker
->current_func
);
2185 debug_show_held_locks(current
);
2189 spin_lock_irq(&pool
->lock
);
2191 /* clear cpu intensive status */
2192 if (unlikely(cpu_intensive
))
2193 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2195 /* we're done with it, release */
2196 hash_del(&worker
->hentry
);
2197 worker
->current_work
= NULL
;
2198 worker
->current_func
= NULL
;
2199 worker
->current_pwq
= NULL
;
2200 pwq_dec_nr_in_flight(pwq
, work_color
);
2204 * process_scheduled_works - process scheduled works
2207 * Process all scheduled works. Please note that the scheduled list
2208 * may change while processing a work, so this function repeatedly
2209 * fetches a work from the top and executes it.
2212 * spin_lock_irq(pool->lock) which may be released and regrabbed
2215 static void process_scheduled_works(struct worker
*worker
)
2217 while (!list_empty(&worker
->scheduled
)) {
2218 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2219 struct work_struct
, entry
);
2220 process_one_work(worker
, work
);
2225 * worker_thread - the worker thread function
2228 * The worker thread function. All workers belong to a worker_pool -
2229 * either a per-cpu one or dynamic unbound one. These workers process all
2230 * work items regardless of their specific target workqueue. The only
2231 * exception is work items which belong to workqueues with a rescuer which
2232 * will be explained in rescuer_thread().
2234 static int worker_thread(void *__worker
)
2236 struct worker
*worker
= __worker
;
2237 struct worker_pool
*pool
= worker
->pool
;
2239 /* tell the scheduler that this is a workqueue worker */
2240 worker
->task
->flags
|= PF_WQ_WORKER
;
2242 spin_lock_irq(&pool
->lock
);
2244 /* am I supposed to die? */
2245 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2246 spin_unlock_irq(&pool
->lock
);
2247 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2248 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2252 worker_leave_idle(worker
);
2254 /* no more worker necessary? */
2255 if (!need_more_worker(pool
))
2258 /* do we need to manage? */
2259 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2263 * ->scheduled list can only be filled while a worker is
2264 * preparing to process a work or actually processing it.
2265 * Make sure nobody diddled with it while I was sleeping.
2267 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2270 * Finish PREP stage. We're guaranteed to have at least one idle
2271 * worker or that someone else has already assumed the manager
2272 * role. This is where @worker starts participating in concurrency
2273 * management if applicable and concurrency management is restored
2274 * after being rebound. See rebind_workers() for details.
2276 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2279 struct work_struct
*work
=
2280 list_first_entry(&pool
->worklist
,
2281 struct work_struct
, entry
);
2283 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2284 /* optimization path, not strictly necessary */
2285 process_one_work(worker
, work
);
2286 if (unlikely(!list_empty(&worker
->scheduled
)))
2287 process_scheduled_works(worker
);
2289 move_linked_works(work
, &worker
->scheduled
, NULL
);
2290 process_scheduled_works(worker
);
2292 } while (keep_working(pool
));
2294 worker_set_flags(worker
, WORKER_PREP
, false);
2296 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2300 * pool->lock is held and there's no work to process and no need to
2301 * manage, sleep. Workers are woken up only while holding
2302 * pool->lock or from local cpu, so setting the current state
2303 * before releasing pool->lock is enough to prevent losing any
2306 worker_enter_idle(worker
);
2307 __set_current_state(TASK_INTERRUPTIBLE
);
2308 spin_unlock_irq(&pool
->lock
);
2314 * rescuer_thread - the rescuer thread function
2317 * Workqueue rescuer thread function. There's one rescuer for each
2318 * workqueue which has WQ_MEM_RECLAIM set.
2320 * Regular work processing on a pool may block trying to create a new
2321 * worker which uses GFP_KERNEL allocation which has slight chance of
2322 * developing into deadlock if some works currently on the same queue
2323 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2324 * the problem rescuer solves.
2326 * When such condition is possible, the pool summons rescuers of all
2327 * workqueues which have works queued on the pool and let them process
2328 * those works so that forward progress can be guaranteed.
2330 * This should happen rarely.
2332 static int rescuer_thread(void *__rescuer
)
2334 struct worker
*rescuer
= __rescuer
;
2335 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2336 struct list_head
*scheduled
= &rescuer
->scheduled
;
2338 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2341 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2342 * doesn't participate in concurrency management.
2344 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2346 set_current_state(TASK_INTERRUPTIBLE
);
2348 if (kthread_should_stop()) {
2349 __set_current_state(TASK_RUNNING
);
2350 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2354 /* see whether any pwq is asking for help */
2355 spin_lock_irq(&wq_mayday_lock
);
2357 while (!list_empty(&wq
->maydays
)) {
2358 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2359 struct pool_workqueue
, mayday_node
);
2360 struct worker_pool
*pool
= pwq
->pool
;
2361 struct work_struct
*work
, *n
;
2363 __set_current_state(TASK_RUNNING
);
2364 list_del_init(&pwq
->mayday_node
);
2366 spin_unlock_irq(&wq_mayday_lock
);
2368 /* migrate to the target cpu if possible */
2369 worker_maybe_bind_and_lock(pool
);
2370 rescuer
->pool
= pool
;
2373 * Slurp in all works issued via this workqueue and
2376 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2377 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2378 if (get_work_pwq(work
) == pwq
)
2379 move_linked_works(work
, scheduled
, &n
);
2381 process_scheduled_works(rescuer
);
2384 * Leave this pool. If keep_working() is %true, notify a
2385 * regular worker; otherwise, we end up with 0 concurrency
2386 * and stalling the execution.
2388 if (keep_working(pool
))
2389 wake_up_worker(pool
);
2391 rescuer
->pool
= NULL
;
2392 spin_unlock(&pool
->lock
);
2393 spin_lock(&wq_mayday_lock
);
2396 spin_unlock_irq(&wq_mayday_lock
);
2398 /* rescuers should never participate in concurrency management */
2399 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2405 struct work_struct work
;
2406 struct completion done
;
2409 static void wq_barrier_func(struct work_struct
*work
)
2411 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2412 complete(&barr
->done
);
2416 * insert_wq_barrier - insert a barrier work
2417 * @pwq: pwq to insert barrier into
2418 * @barr: wq_barrier to insert
2419 * @target: target work to attach @barr to
2420 * @worker: worker currently executing @target, NULL if @target is not executing
2422 * @barr is linked to @target such that @barr is completed only after
2423 * @target finishes execution. Please note that the ordering
2424 * guarantee is observed only with respect to @target and on the local
2427 * Currently, a queued barrier can't be canceled. This is because
2428 * try_to_grab_pending() can't determine whether the work to be
2429 * grabbed is at the head of the queue and thus can't clear LINKED
2430 * flag of the previous work while there must be a valid next work
2431 * after a work with LINKED flag set.
2433 * Note that when @worker is non-NULL, @target may be modified
2434 * underneath us, so we can't reliably determine pwq from @target.
2437 * spin_lock_irq(pool->lock).
2439 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2440 struct wq_barrier
*barr
,
2441 struct work_struct
*target
, struct worker
*worker
)
2443 struct list_head
*head
;
2444 unsigned int linked
= 0;
2447 * debugobject calls are safe here even with pool->lock locked
2448 * as we know for sure that this will not trigger any of the
2449 * checks and call back into the fixup functions where we
2452 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2453 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2454 init_completion(&barr
->done
);
2457 * If @target is currently being executed, schedule the
2458 * barrier to the worker; otherwise, put it after @target.
2461 head
= worker
->scheduled
.next
;
2463 unsigned long *bits
= work_data_bits(target
);
2465 head
= target
->entry
.next
;
2466 /* there can already be other linked works, inherit and set */
2467 linked
= *bits
& WORK_STRUCT_LINKED
;
2468 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2471 debug_work_activate(&barr
->work
);
2472 insert_work(pwq
, &barr
->work
, head
,
2473 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2477 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2478 * @wq: workqueue being flushed
2479 * @flush_color: new flush color, < 0 for no-op
2480 * @work_color: new work color, < 0 for no-op
2482 * Prepare pwqs for workqueue flushing.
2484 * If @flush_color is non-negative, flush_color on all pwqs should be
2485 * -1. If no pwq has in-flight commands at the specified color, all
2486 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2487 * has in flight commands, its pwq->flush_color is set to
2488 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2489 * wakeup logic is armed and %true is returned.
2491 * The caller should have initialized @wq->first_flusher prior to
2492 * calling this function with non-negative @flush_color. If
2493 * @flush_color is negative, no flush color update is done and %false
2496 * If @work_color is non-negative, all pwqs should have the same
2497 * work_color which is previous to @work_color and all will be
2498 * advanced to @work_color.
2501 * mutex_lock(wq->mutex).
2504 * %true if @flush_color >= 0 and there's something to flush. %false
2507 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2508 int flush_color
, int work_color
)
2511 struct pool_workqueue
*pwq
;
2513 if (flush_color
>= 0) {
2514 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2515 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2518 for_each_pwq(pwq
, wq
) {
2519 struct worker_pool
*pool
= pwq
->pool
;
2521 spin_lock_irq(&pool
->lock
);
2523 if (flush_color
>= 0) {
2524 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2526 if (pwq
->nr_in_flight
[flush_color
]) {
2527 pwq
->flush_color
= flush_color
;
2528 atomic_inc(&wq
->nr_pwqs_to_flush
);
2533 if (work_color
>= 0) {
2534 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2535 pwq
->work_color
= work_color
;
2538 spin_unlock_irq(&pool
->lock
);
2541 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2542 complete(&wq
->first_flusher
->done
);
2548 * flush_workqueue - ensure that any scheduled work has run to completion.
2549 * @wq: workqueue to flush
2551 * This function sleeps until all work items which were queued on entry
2552 * have finished execution, but it is not livelocked by new incoming ones.
2554 void flush_workqueue(struct workqueue_struct
*wq
)
2556 struct wq_flusher this_flusher
= {
2557 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2559 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2563 lock_map_acquire(&wq
->lockdep_map
);
2564 lock_map_release(&wq
->lockdep_map
);
2566 mutex_lock(&wq
->mutex
);
2569 * Start-to-wait phase
2571 next_color
= work_next_color(wq
->work_color
);
2573 if (next_color
!= wq
->flush_color
) {
2575 * Color space is not full. The current work_color
2576 * becomes our flush_color and work_color is advanced
2579 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2580 this_flusher
.flush_color
= wq
->work_color
;
2581 wq
->work_color
= next_color
;
2583 if (!wq
->first_flusher
) {
2584 /* no flush in progress, become the first flusher */
2585 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2587 wq
->first_flusher
= &this_flusher
;
2589 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2591 /* nothing to flush, done */
2592 wq
->flush_color
= next_color
;
2593 wq
->first_flusher
= NULL
;
2598 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2599 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2600 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2604 * Oops, color space is full, wait on overflow queue.
2605 * The next flush completion will assign us
2606 * flush_color and transfer to flusher_queue.
2608 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2611 mutex_unlock(&wq
->mutex
);
2613 wait_for_completion(&this_flusher
.done
);
2616 * Wake-up-and-cascade phase
2618 * First flushers are responsible for cascading flushes and
2619 * handling overflow. Non-first flushers can simply return.
2621 if (wq
->first_flusher
!= &this_flusher
)
2624 mutex_lock(&wq
->mutex
);
2626 /* we might have raced, check again with mutex held */
2627 if (wq
->first_flusher
!= &this_flusher
)
2630 wq
->first_flusher
= NULL
;
2632 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2633 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2636 struct wq_flusher
*next
, *tmp
;
2638 /* complete all the flushers sharing the current flush color */
2639 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2640 if (next
->flush_color
!= wq
->flush_color
)
2642 list_del_init(&next
->list
);
2643 complete(&next
->done
);
2646 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2647 wq
->flush_color
!= work_next_color(wq
->work_color
));
2649 /* this flush_color is finished, advance by one */
2650 wq
->flush_color
= work_next_color(wq
->flush_color
);
2652 /* one color has been freed, handle overflow queue */
2653 if (!list_empty(&wq
->flusher_overflow
)) {
2655 * Assign the same color to all overflowed
2656 * flushers, advance work_color and append to
2657 * flusher_queue. This is the start-to-wait
2658 * phase for these overflowed flushers.
2660 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2661 tmp
->flush_color
= wq
->work_color
;
2663 wq
->work_color
= work_next_color(wq
->work_color
);
2665 list_splice_tail_init(&wq
->flusher_overflow
,
2666 &wq
->flusher_queue
);
2667 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2670 if (list_empty(&wq
->flusher_queue
)) {
2671 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2676 * Need to flush more colors. Make the next flusher
2677 * the new first flusher and arm pwqs.
2679 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2680 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2682 list_del_init(&next
->list
);
2683 wq
->first_flusher
= next
;
2685 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2689 * Meh... this color is already done, clear first
2690 * flusher and repeat cascading.
2692 wq
->first_flusher
= NULL
;
2696 mutex_unlock(&wq
->mutex
);
2698 EXPORT_SYMBOL_GPL(flush_workqueue
);
2701 * drain_workqueue - drain a workqueue
2702 * @wq: workqueue to drain
2704 * Wait until the workqueue becomes empty. While draining is in progress,
2705 * only chain queueing is allowed. IOW, only currently pending or running
2706 * work items on @wq can queue further work items on it. @wq is flushed
2707 * repeatedly until it becomes empty. The number of flushing is detemined
2708 * by the depth of chaining and should be relatively short. Whine if it
2711 void drain_workqueue(struct workqueue_struct
*wq
)
2713 unsigned int flush_cnt
= 0;
2714 struct pool_workqueue
*pwq
;
2717 * __queue_work() needs to test whether there are drainers, is much
2718 * hotter than drain_workqueue() and already looks at @wq->flags.
2719 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2721 mutex_lock(&wq
->mutex
);
2722 if (!wq
->nr_drainers
++)
2723 wq
->flags
|= __WQ_DRAINING
;
2724 mutex_unlock(&wq
->mutex
);
2726 flush_workqueue(wq
);
2728 mutex_lock(&wq
->mutex
);
2730 for_each_pwq(pwq
, wq
) {
2733 spin_lock_irq(&pwq
->pool
->lock
);
2734 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2735 spin_unlock_irq(&pwq
->pool
->lock
);
2740 if (++flush_cnt
== 10 ||
2741 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2742 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2743 wq
->name
, flush_cnt
);
2745 mutex_unlock(&wq
->mutex
);
2749 if (!--wq
->nr_drainers
)
2750 wq
->flags
&= ~__WQ_DRAINING
;
2751 mutex_unlock(&wq
->mutex
);
2753 EXPORT_SYMBOL_GPL(drain_workqueue
);
2755 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2757 struct worker
*worker
= NULL
;
2758 struct worker_pool
*pool
;
2759 struct pool_workqueue
*pwq
;
2763 local_irq_disable();
2764 pool
= get_work_pool(work
);
2770 spin_lock(&pool
->lock
);
2771 /* see the comment in try_to_grab_pending() with the same code */
2772 pwq
= get_work_pwq(work
);
2774 if (unlikely(pwq
->pool
!= pool
))
2777 worker
= find_worker_executing_work(pool
, work
);
2780 pwq
= worker
->current_pwq
;
2783 insert_wq_barrier(pwq
, barr
, work
, worker
);
2784 spin_unlock_irq(&pool
->lock
);
2787 * If @max_active is 1 or rescuer is in use, flushing another work
2788 * item on the same workqueue may lead to deadlock. Make sure the
2789 * flusher is not running on the same workqueue by verifying write
2792 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2793 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2795 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2796 lock_map_release(&pwq
->wq
->lockdep_map
);
2800 spin_unlock_irq(&pool
->lock
);
2805 * flush_work - wait for a work to finish executing the last queueing instance
2806 * @work: the work to flush
2808 * Wait until @work has finished execution. @work is guaranteed to be idle
2809 * on return if it hasn't been requeued since flush started.
2812 * %true if flush_work() waited for the work to finish execution,
2813 * %false if it was already idle.
2815 bool flush_work(struct work_struct
*work
)
2817 struct wq_barrier barr
;
2819 lock_map_acquire(&work
->lockdep_map
);
2820 lock_map_release(&work
->lockdep_map
);
2822 if (start_flush_work(work
, &barr
)) {
2823 wait_for_completion(&barr
.done
);
2824 destroy_work_on_stack(&barr
.work
);
2830 EXPORT_SYMBOL_GPL(flush_work
);
2832 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2834 unsigned long flags
;
2838 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2840 * If someone else is canceling, wait for the same event it
2841 * would be waiting for before retrying.
2843 if (unlikely(ret
== -ENOENT
))
2845 } while (unlikely(ret
< 0));
2847 /* tell other tasks trying to grab @work to back off */
2848 mark_work_canceling(work
);
2849 local_irq_restore(flags
);
2852 clear_work_data(work
);
2857 * cancel_work_sync - cancel a work and wait for it to finish
2858 * @work: the work to cancel
2860 * Cancel @work and wait for its execution to finish. This function
2861 * can be used even if the work re-queues itself or migrates to
2862 * another workqueue. On return from this function, @work is
2863 * guaranteed to be not pending or executing on any CPU.
2865 * cancel_work_sync(&delayed_work->work) must not be used for
2866 * delayed_work's. Use cancel_delayed_work_sync() instead.
2868 * The caller must ensure that the workqueue on which @work was last
2869 * queued can't be destroyed before this function returns.
2872 * %true if @work was pending, %false otherwise.
2874 bool cancel_work_sync(struct work_struct
*work
)
2876 return __cancel_work_timer(work
, false);
2878 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2881 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2882 * @dwork: the delayed work to flush
2884 * Delayed timer is cancelled and the pending work is queued for
2885 * immediate execution. Like flush_work(), this function only
2886 * considers the last queueing instance of @dwork.
2889 * %true if flush_work() waited for the work to finish execution,
2890 * %false if it was already idle.
2892 bool flush_delayed_work(struct delayed_work
*dwork
)
2894 local_irq_disable();
2895 if (del_timer_sync(&dwork
->timer
))
2896 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2898 return flush_work(&dwork
->work
);
2900 EXPORT_SYMBOL(flush_delayed_work
);
2903 * cancel_delayed_work - cancel a delayed work
2904 * @dwork: delayed_work to cancel
2906 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2907 * and canceled; %false if wasn't pending. Note that the work callback
2908 * function may still be running on return, unless it returns %true and the
2909 * work doesn't re-arm itself. Explicitly flush or use
2910 * cancel_delayed_work_sync() to wait on it.
2912 * This function is safe to call from any context including IRQ handler.
2914 bool cancel_delayed_work(struct delayed_work
*dwork
)
2916 unsigned long flags
;
2920 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2921 } while (unlikely(ret
== -EAGAIN
));
2923 if (unlikely(ret
< 0))
2926 set_work_pool_and_clear_pending(&dwork
->work
,
2927 get_work_pool_id(&dwork
->work
));
2928 local_irq_restore(flags
);
2931 EXPORT_SYMBOL(cancel_delayed_work
);
2934 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2935 * @dwork: the delayed work cancel
2937 * This is cancel_work_sync() for delayed works.
2940 * %true if @dwork was pending, %false otherwise.
2942 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2944 return __cancel_work_timer(&dwork
->work
, true);
2946 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2949 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2950 * @func: the function to call
2952 * schedule_on_each_cpu() executes @func on each online CPU using the
2953 * system workqueue and blocks until all CPUs have completed.
2954 * schedule_on_each_cpu() is very slow.
2957 * 0 on success, -errno on failure.
2959 int schedule_on_each_cpu(work_func_t func
)
2962 struct work_struct __percpu
*works
;
2964 works
= alloc_percpu(struct work_struct
);
2970 for_each_online_cpu(cpu
) {
2971 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2973 INIT_WORK(work
, func
);
2974 schedule_work_on(cpu
, work
);
2977 for_each_online_cpu(cpu
)
2978 flush_work(per_cpu_ptr(works
, cpu
));
2986 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2988 * Forces execution of the kernel-global workqueue and blocks until its
2991 * Think twice before calling this function! It's very easy to get into
2992 * trouble if you don't take great care. Either of the following situations
2993 * will lead to deadlock:
2995 * One of the work items currently on the workqueue needs to acquire
2996 * a lock held by your code or its caller.
2998 * Your code is running in the context of a work routine.
3000 * They will be detected by lockdep when they occur, but the first might not
3001 * occur very often. It depends on what work items are on the workqueue and
3002 * what locks they need, which you have no control over.
3004 * In most situations flushing the entire workqueue is overkill; you merely
3005 * need to know that a particular work item isn't queued and isn't running.
3006 * In such cases you should use cancel_delayed_work_sync() or
3007 * cancel_work_sync() instead.
3009 void flush_scheduled_work(void)
3011 flush_workqueue(system_wq
);
3013 EXPORT_SYMBOL(flush_scheduled_work
);
3016 * execute_in_process_context - reliably execute the routine with user context
3017 * @fn: the function to execute
3018 * @ew: guaranteed storage for the execute work structure (must
3019 * be available when the work executes)
3021 * Executes the function immediately if process context is available,
3022 * otherwise schedules the function for delayed execution.
3024 * Returns: 0 - function was executed
3025 * 1 - function was scheduled for execution
3027 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3029 if (!in_interrupt()) {
3034 INIT_WORK(&ew
->work
, fn
);
3035 schedule_work(&ew
->work
);
3039 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3043 * Workqueues with WQ_SYSFS flag set is visible to userland via
3044 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3045 * following attributes.
3047 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3048 * max_active RW int : maximum number of in-flight work items
3050 * Unbound workqueues have the following extra attributes.
3052 * id RO int : the associated pool ID
3053 * nice RW int : nice value of the workers
3054 * cpumask RW mask : bitmask of allowed CPUs for the workers
3057 struct workqueue_struct
*wq
;
3061 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3063 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3068 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3069 struct device_attribute
*attr
, char *buf
)
3071 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3073 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3076 static ssize_t
wq_max_active_show(struct device
*dev
,
3077 struct device_attribute
*attr
, char *buf
)
3079 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3081 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3084 static ssize_t
wq_max_active_store(struct device
*dev
,
3085 struct device_attribute
*attr
,
3086 const char *buf
, size_t count
)
3088 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3091 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3094 workqueue_set_max_active(wq
, val
);
3098 static struct device_attribute wq_sysfs_attrs
[] = {
3099 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3100 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3104 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3105 struct device_attribute
*attr
, char *buf
)
3107 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3108 const char *delim
= "";
3109 int node
, written
= 0;
3111 rcu_read_lock_sched();
3112 for_each_node(node
) {
3113 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3114 "%s%d:%d", delim
, node
,
3115 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3118 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3119 rcu_read_unlock_sched();
3124 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3127 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3130 mutex_lock(&wq
->mutex
);
3131 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3132 mutex_unlock(&wq
->mutex
);
3137 /* prepare workqueue_attrs for sysfs store operations */
3138 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3140 struct workqueue_attrs
*attrs
;
3142 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3146 mutex_lock(&wq
->mutex
);
3147 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3148 mutex_unlock(&wq
->mutex
);
3152 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3153 const char *buf
, size_t count
)
3155 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3156 struct workqueue_attrs
*attrs
;
3159 attrs
= wq_sysfs_prep_attrs(wq
);
3163 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3164 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3165 ret
= apply_workqueue_attrs(wq
, attrs
);
3169 free_workqueue_attrs(attrs
);
3170 return ret
?: count
;
3173 static ssize_t
wq_cpumask_show(struct device
*dev
,
3174 struct device_attribute
*attr
, char *buf
)
3176 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3179 mutex_lock(&wq
->mutex
);
3180 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3181 mutex_unlock(&wq
->mutex
);
3183 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3187 static ssize_t
wq_cpumask_store(struct device
*dev
,
3188 struct device_attribute
*attr
,
3189 const char *buf
, size_t count
)
3191 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3192 struct workqueue_attrs
*attrs
;
3195 attrs
= wq_sysfs_prep_attrs(wq
);
3199 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3201 ret
= apply_workqueue_attrs(wq
, attrs
);
3203 free_workqueue_attrs(attrs
);
3204 return ret
?: count
;
3207 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3210 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3213 mutex_lock(&wq
->mutex
);
3214 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3215 !wq
->unbound_attrs
->no_numa
);
3216 mutex_unlock(&wq
->mutex
);
3221 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3222 const char *buf
, size_t count
)
3224 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3225 struct workqueue_attrs
*attrs
;
3228 attrs
= wq_sysfs_prep_attrs(wq
);
3233 if (sscanf(buf
, "%d", &v
) == 1) {
3234 attrs
->no_numa
= !v
;
3235 ret
= apply_workqueue_attrs(wq
, attrs
);
3238 free_workqueue_attrs(attrs
);
3239 return ret
?: count
;
3242 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3243 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3244 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3245 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3246 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3250 static struct bus_type wq_subsys
= {
3251 .name
= "workqueue",
3252 .dev_attrs
= wq_sysfs_attrs
,
3255 static int __init
wq_sysfs_init(void)
3257 return subsys_virtual_register(&wq_subsys
, NULL
);
3259 core_initcall(wq_sysfs_init
);
3261 static void wq_device_release(struct device
*dev
)
3263 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3269 * workqueue_sysfs_register - make a workqueue visible in sysfs
3270 * @wq: the workqueue to register
3272 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3273 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3274 * which is the preferred method.
3276 * Workqueue user should use this function directly iff it wants to apply
3277 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3278 * apply_workqueue_attrs() may race against userland updating the
3281 * Returns 0 on success, -errno on failure.
3283 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3285 struct wq_device
*wq_dev
;
3289 * Adjusting max_active or creating new pwqs by applyting
3290 * attributes breaks ordering guarantee. Disallow exposing ordered
3293 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3296 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3301 wq_dev
->dev
.bus
= &wq_subsys
;
3302 wq_dev
->dev
.init_name
= wq
->name
;
3303 wq_dev
->dev
.release
= wq_device_release
;
3306 * unbound_attrs are created separately. Suppress uevent until
3307 * everything is ready.
3309 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3311 ret
= device_register(&wq_dev
->dev
);
3318 if (wq
->flags
& WQ_UNBOUND
) {
3319 struct device_attribute
*attr
;
3321 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3322 ret
= device_create_file(&wq_dev
->dev
, attr
);
3324 device_unregister(&wq_dev
->dev
);
3331 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3336 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3337 * @wq: the workqueue to unregister
3339 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3341 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3343 struct wq_device
*wq_dev
= wq
->wq_dev
;
3349 device_unregister(&wq_dev
->dev
);
3351 #else /* CONFIG_SYSFS */
3352 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3353 #endif /* CONFIG_SYSFS */
3356 * free_workqueue_attrs - free a workqueue_attrs
3357 * @attrs: workqueue_attrs to free
3359 * Undo alloc_workqueue_attrs().
3361 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3364 free_cpumask_var(attrs
->cpumask
);
3370 * alloc_workqueue_attrs - allocate a workqueue_attrs
3371 * @gfp_mask: allocation mask to use
3373 * Allocate a new workqueue_attrs, initialize with default settings and
3374 * return it. Returns NULL on failure.
3376 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3378 struct workqueue_attrs
*attrs
;
3380 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3383 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3386 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3389 free_workqueue_attrs(attrs
);
3393 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3394 const struct workqueue_attrs
*from
)
3396 to
->nice
= from
->nice
;
3397 cpumask_copy(to
->cpumask
, from
->cpumask
);
3400 /* hash value of the content of @attr */
3401 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3405 hash
= jhash_1word(attrs
->nice
, hash
);
3406 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3407 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3411 /* content equality test */
3412 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3413 const struct workqueue_attrs
*b
)
3415 if (a
->nice
!= b
->nice
)
3417 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3423 * init_worker_pool - initialize a newly zalloc'd worker_pool
3424 * @pool: worker_pool to initialize
3426 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3427 * Returns 0 on success, -errno on failure. Even on failure, all fields
3428 * inside @pool proper are initialized and put_unbound_pool() can be called
3429 * on @pool safely to release it.
3431 static int init_worker_pool(struct worker_pool
*pool
)
3433 spin_lock_init(&pool
->lock
);
3436 pool
->node
= NUMA_NO_NODE
;
3437 pool
->flags
|= POOL_DISASSOCIATED
;
3438 INIT_LIST_HEAD(&pool
->worklist
);
3439 INIT_LIST_HEAD(&pool
->idle_list
);
3440 hash_init(pool
->busy_hash
);
3442 init_timer_deferrable(&pool
->idle_timer
);
3443 pool
->idle_timer
.function
= idle_worker_timeout
;
3444 pool
->idle_timer
.data
= (unsigned long)pool
;
3446 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3447 (unsigned long)pool
);
3449 mutex_init(&pool
->manager_arb
);
3450 mutex_init(&pool
->manager_mutex
);
3451 idr_init(&pool
->worker_idr
);
3453 INIT_HLIST_NODE(&pool
->hash_node
);
3456 /* shouldn't fail above this point */
3457 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3463 static void rcu_free_pool(struct rcu_head
*rcu
)
3465 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3467 idr_destroy(&pool
->worker_idr
);
3468 free_workqueue_attrs(pool
->attrs
);
3473 * put_unbound_pool - put a worker_pool
3474 * @pool: worker_pool to put
3476 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3477 * safe manner. get_unbound_pool() calls this function on its failure path
3478 * and this function should be able to release pools which went through,
3479 * successfully or not, init_worker_pool().
3481 * Should be called with wq_pool_mutex held.
3483 static void put_unbound_pool(struct worker_pool
*pool
)
3485 struct worker
*worker
;
3487 lockdep_assert_held(&wq_pool_mutex
);
3493 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3494 WARN_ON(!list_empty(&pool
->worklist
)))
3497 /* release id and unhash */
3499 idr_remove(&worker_pool_idr
, pool
->id
);
3500 hash_del(&pool
->hash_node
);
3503 * Become the manager and destroy all workers. Grabbing
3504 * manager_arb prevents @pool's workers from blocking on
3507 mutex_lock(&pool
->manager_arb
);
3508 mutex_lock(&pool
->manager_mutex
);
3509 spin_lock_irq(&pool
->lock
);
3511 while ((worker
= first_worker(pool
)))
3512 destroy_worker(worker
);
3513 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3515 spin_unlock_irq(&pool
->lock
);
3516 mutex_unlock(&pool
->manager_mutex
);
3517 mutex_unlock(&pool
->manager_arb
);
3519 /* shut down the timers */
3520 del_timer_sync(&pool
->idle_timer
);
3521 del_timer_sync(&pool
->mayday_timer
);
3523 /* sched-RCU protected to allow dereferences from get_work_pool() */
3524 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3528 * get_unbound_pool - get a worker_pool with the specified attributes
3529 * @attrs: the attributes of the worker_pool to get
3531 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3532 * reference count and return it. If there already is a matching
3533 * worker_pool, it will be used; otherwise, this function attempts to
3534 * create a new one. On failure, returns NULL.
3536 * Should be called with wq_pool_mutex held.
3538 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3540 u32 hash
= wqattrs_hash(attrs
);
3541 struct worker_pool
*pool
;
3544 lockdep_assert_held(&wq_pool_mutex
);
3546 /* do we already have a matching pool? */
3547 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3548 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3554 /* nope, create a new one */
3555 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3556 if (!pool
|| init_worker_pool(pool
) < 0)
3559 if (workqueue_freezing
)
3560 pool
->flags
|= POOL_FREEZING
;
3562 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3563 copy_workqueue_attrs(pool
->attrs
, attrs
);
3565 /* if cpumask is contained inside a NUMA node, we belong to that node */
3566 if (wq_numa_enabled
) {
3567 for_each_node(node
) {
3568 if (cpumask_subset(pool
->attrs
->cpumask
,
3569 wq_numa_possible_cpumask
[node
])) {
3576 if (worker_pool_assign_id(pool
) < 0)
3579 /* create and start the initial worker */
3580 if (create_and_start_worker(pool
) < 0)
3584 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3589 put_unbound_pool(pool
);
3593 static void rcu_free_pwq(struct rcu_head
*rcu
)
3595 kmem_cache_free(pwq_cache
,
3596 container_of(rcu
, struct pool_workqueue
, rcu
));
3600 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3601 * and needs to be destroyed.
3603 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3605 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3606 unbound_release_work
);
3607 struct workqueue_struct
*wq
= pwq
->wq
;
3608 struct worker_pool
*pool
= pwq
->pool
;
3611 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3615 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3616 * necessary on release but do it anyway. It's easier to verify
3617 * and consistent with the linking path.
3619 mutex_lock(&wq
->mutex
);
3620 list_del_rcu(&pwq
->pwqs_node
);
3621 is_last
= list_empty(&wq
->pwqs
);
3622 mutex_unlock(&wq
->mutex
);
3624 mutex_lock(&wq_pool_mutex
);
3625 put_unbound_pool(pool
);
3626 mutex_unlock(&wq_pool_mutex
);
3628 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3631 * If we're the last pwq going away, @wq is already dead and no one
3632 * is gonna access it anymore. Free it.
3635 free_workqueue_attrs(wq
->unbound_attrs
);
3641 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3642 * @pwq: target pool_workqueue
3644 * If @pwq isn't freezing, set @pwq->max_active to the associated
3645 * workqueue's saved_max_active and activate delayed work items
3646 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3648 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3650 struct workqueue_struct
*wq
= pwq
->wq
;
3651 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3653 /* for @wq->saved_max_active */
3654 lockdep_assert_held(&wq
->mutex
);
3656 /* fast exit for non-freezable wqs */
3657 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3660 spin_lock_irq(&pwq
->pool
->lock
);
3662 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3663 pwq
->max_active
= wq
->saved_max_active
;
3665 while (!list_empty(&pwq
->delayed_works
) &&
3666 pwq
->nr_active
< pwq
->max_active
)
3667 pwq_activate_first_delayed(pwq
);
3670 * Need to kick a worker after thawed or an unbound wq's
3671 * max_active is bumped. It's a slow path. Do it always.
3673 wake_up_worker(pwq
->pool
);
3675 pwq
->max_active
= 0;
3678 spin_unlock_irq(&pwq
->pool
->lock
);
3681 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3682 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3683 struct worker_pool
*pool
)
3685 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3687 memset(pwq
, 0, sizeof(*pwq
));
3691 pwq
->flush_color
= -1;
3693 INIT_LIST_HEAD(&pwq
->delayed_works
);
3694 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3695 INIT_LIST_HEAD(&pwq
->mayday_node
);
3696 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3699 /* sync @pwq with the current state of its associated wq and link it */
3700 static void link_pwq(struct pool_workqueue
*pwq
)
3702 struct workqueue_struct
*wq
= pwq
->wq
;
3704 lockdep_assert_held(&wq
->mutex
);
3706 /* may be called multiple times, ignore if already linked */
3707 if (!list_empty(&pwq
->pwqs_node
))
3711 * Set the matching work_color. This is synchronized with
3712 * wq->mutex to avoid confusing flush_workqueue().
3714 pwq
->work_color
= wq
->work_color
;
3716 /* sync max_active to the current setting */
3717 pwq_adjust_max_active(pwq
);
3720 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3723 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3724 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3725 const struct workqueue_attrs
*attrs
)
3727 struct worker_pool
*pool
;
3728 struct pool_workqueue
*pwq
;
3730 lockdep_assert_held(&wq_pool_mutex
);
3732 pool
= get_unbound_pool(attrs
);
3736 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3738 put_unbound_pool(pool
);
3742 init_pwq(pwq
, wq
, pool
);
3746 /* undo alloc_unbound_pwq(), used only in the error path */
3747 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3749 lockdep_assert_held(&wq_pool_mutex
);
3752 put_unbound_pool(pwq
->pool
);
3753 kmem_cache_free(pwq_cache
, pwq
);
3758 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3759 * @attrs: the wq_attrs of interest
3760 * @node: the target NUMA node
3761 * @cpu_going_down: if >= 0, the CPU to consider as offline
3762 * @cpumask: outarg, the resulting cpumask
3764 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3765 * @cpu_going_down is >= 0, that cpu is considered offline during
3766 * calculation. The result is stored in @cpumask. This function returns
3767 * %true if the resulting @cpumask is different from @attrs->cpumask,
3770 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3771 * enabled and @node has online CPUs requested by @attrs, the returned
3772 * cpumask is the intersection of the possible CPUs of @node and
3775 * The caller is responsible for ensuring that the cpumask of @node stays
3778 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3779 int cpu_going_down
, cpumask_t
*cpumask
)
3781 if (!wq_numa_enabled
|| attrs
->no_numa
)
3784 /* does @node have any online CPUs @attrs wants? */
3785 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3786 if (cpu_going_down
>= 0)
3787 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3789 if (cpumask_empty(cpumask
))
3792 /* yeap, return possible CPUs in @node that @attrs wants */
3793 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3794 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3797 cpumask_copy(cpumask
, attrs
->cpumask
);
3801 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3802 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3804 struct pool_workqueue
*pwq
)
3806 struct pool_workqueue
*old_pwq
;
3808 lockdep_assert_held(&wq
->mutex
);
3810 /* link_pwq() can handle duplicate calls */
3813 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3814 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3819 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3820 * @wq: the target workqueue
3821 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3823 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3824 * machines, this function maps a separate pwq to each NUMA node with
3825 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3826 * NUMA node it was issued on. Older pwqs are released as in-flight work
3827 * items finish. Note that a work item which repeatedly requeues itself
3828 * back-to-back will stay on its current pwq.
3830 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3833 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3834 const struct workqueue_attrs
*attrs
)
3836 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3837 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3840 /* only unbound workqueues can change attributes */
3841 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3844 /* creating multiple pwqs breaks ordering guarantee */
3845 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3848 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3849 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3850 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3851 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3854 /* make a copy of @attrs and sanitize it */
3855 copy_workqueue_attrs(new_attrs
, attrs
);
3856 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3859 * We may create multiple pwqs with differing cpumasks. Make a
3860 * copy of @new_attrs which will be modified and used to obtain
3863 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3866 * CPUs should stay stable across pwq creations and installations.
3867 * Pin CPUs, determine the target cpumask for each node and create
3872 mutex_lock(&wq_pool_mutex
);
3875 * If something goes wrong during CPU up/down, we'll fall back to
3876 * the default pwq covering whole @attrs->cpumask. Always create
3877 * it even if we don't use it immediately.
3879 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3883 for_each_node(node
) {
3884 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3885 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3890 pwq_tbl
[node
] = dfl_pwq
;
3894 mutex_unlock(&wq_pool_mutex
);
3896 /* all pwqs have been created successfully, let's install'em */
3897 mutex_lock(&wq
->mutex
);
3899 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3901 /* save the previous pwq and install the new one */
3903 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3905 /* @dfl_pwq might not have been used, ensure it's linked */
3907 swap(wq
->dfl_pwq
, dfl_pwq
);
3909 mutex_unlock(&wq
->mutex
);
3911 /* put the old pwqs */
3913 put_pwq_unlocked(pwq_tbl
[node
]);
3914 put_pwq_unlocked(dfl_pwq
);
3920 free_workqueue_attrs(tmp_attrs
);
3921 free_workqueue_attrs(new_attrs
);
3926 free_unbound_pwq(dfl_pwq
);
3928 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
3929 free_unbound_pwq(pwq_tbl
[node
]);
3930 mutex_unlock(&wq_pool_mutex
);
3938 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3939 * @wq: the target workqueue
3940 * @cpu: the CPU coming up or going down
3941 * @online: whether @cpu is coming up or going down
3943 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3944 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3947 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3948 * falls back to @wq->dfl_pwq which may not be optimal but is always
3951 * Note that when the last allowed CPU of a NUMA node goes offline for a
3952 * workqueue with a cpumask spanning multiple nodes, the workers which were
3953 * already executing the work items for the workqueue will lose their CPU
3954 * affinity and may execute on any CPU. This is similar to how per-cpu
3955 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3956 * affinity, it's the user's responsibility to flush the work item from
3959 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3962 int node
= cpu_to_node(cpu
);
3963 int cpu_off
= online
? -1 : cpu
;
3964 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3965 struct workqueue_attrs
*target_attrs
;
3968 lockdep_assert_held(&wq_pool_mutex
);
3970 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
3974 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3975 * Let's use a preallocated one. The following buf is protected by
3976 * CPU hotplug exclusion.
3978 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3979 cpumask
= target_attrs
->cpumask
;
3981 mutex_lock(&wq
->mutex
);
3982 if (wq
->unbound_attrs
->no_numa
)
3985 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
3986 pwq
= unbound_pwq_by_node(wq
, node
);
3989 * Let's determine what needs to be done. If the target cpumask is
3990 * different from wq's, we need to compare it to @pwq's and create
3991 * a new one if they don't match. If the target cpumask equals
3992 * wq's, the default pwq should be used. If @pwq is already the
3993 * default one, nothing to do; otherwise, install the default one.
3995 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
3996 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
3999 if (pwq
== wq
->dfl_pwq
)
4005 mutex_unlock(&wq
->mutex
);
4007 /* create a new pwq */
4008 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4010 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4016 * Install the new pwq. As this function is called only from CPU
4017 * hotplug callbacks and applying a new attrs is wrapped with
4018 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4021 mutex_lock(&wq
->mutex
);
4022 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4026 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4027 get_pwq(wq
->dfl_pwq
);
4028 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4029 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4031 mutex_unlock(&wq
->mutex
);
4032 put_pwq_unlocked(old_pwq
);
4035 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4037 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4040 if (!(wq
->flags
& WQ_UNBOUND
)) {
4041 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4045 for_each_possible_cpu(cpu
) {
4046 struct pool_workqueue
*pwq
=
4047 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4048 struct worker_pool
*cpu_pools
=
4049 per_cpu(cpu_worker_pools
, cpu
);
4051 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4053 mutex_lock(&wq
->mutex
);
4055 mutex_unlock(&wq
->mutex
);
4059 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4063 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4066 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4068 if (max_active
< 1 || max_active
> lim
)
4069 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4070 max_active
, name
, 1, lim
);
4072 return clamp_val(max_active
, 1, lim
);
4075 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4078 struct lock_class_key
*key
,
4079 const char *lock_name
, ...)
4081 size_t tbl_size
= 0;
4083 struct workqueue_struct
*wq
;
4084 struct pool_workqueue
*pwq
;
4086 /* allocate wq and format name */
4087 if (flags
& WQ_UNBOUND
)
4088 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4090 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4094 if (flags
& WQ_UNBOUND
) {
4095 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4096 if (!wq
->unbound_attrs
)
4100 va_start(args
, lock_name
);
4101 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4104 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4105 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4109 wq
->saved_max_active
= max_active
;
4110 mutex_init(&wq
->mutex
);
4111 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4112 INIT_LIST_HEAD(&wq
->pwqs
);
4113 INIT_LIST_HEAD(&wq
->flusher_queue
);
4114 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4115 INIT_LIST_HEAD(&wq
->maydays
);
4117 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4118 INIT_LIST_HEAD(&wq
->list
);
4120 if (alloc_and_link_pwqs(wq
) < 0)
4124 * Workqueues which may be used during memory reclaim should
4125 * have a rescuer to guarantee forward progress.
4127 if (flags
& WQ_MEM_RECLAIM
) {
4128 struct worker
*rescuer
;
4130 rescuer
= alloc_worker();
4134 rescuer
->rescue_wq
= wq
;
4135 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4137 if (IS_ERR(rescuer
->task
)) {
4142 wq
->rescuer
= rescuer
;
4143 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4144 wake_up_process(rescuer
->task
);
4147 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4151 * wq_pool_mutex protects global freeze state and workqueues list.
4152 * Grab it, adjust max_active and add the new @wq to workqueues
4155 mutex_lock(&wq_pool_mutex
);
4157 mutex_lock(&wq
->mutex
);
4158 for_each_pwq(pwq
, wq
)
4159 pwq_adjust_max_active(pwq
);
4160 mutex_unlock(&wq
->mutex
);
4162 list_add(&wq
->list
, &workqueues
);
4164 mutex_unlock(&wq_pool_mutex
);
4169 free_workqueue_attrs(wq
->unbound_attrs
);
4173 destroy_workqueue(wq
);
4176 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4179 * destroy_workqueue - safely terminate a workqueue
4180 * @wq: target workqueue
4182 * Safely destroy a workqueue. All work currently pending will be done first.
4184 void destroy_workqueue(struct workqueue_struct
*wq
)
4186 struct pool_workqueue
*pwq
;
4189 /* drain it before proceeding with destruction */
4190 drain_workqueue(wq
);
4193 mutex_lock(&wq
->mutex
);
4194 for_each_pwq(pwq
, wq
) {
4197 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4198 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4199 mutex_unlock(&wq
->mutex
);
4204 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4205 WARN_ON(pwq
->nr_active
) ||
4206 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4207 mutex_unlock(&wq
->mutex
);
4211 mutex_unlock(&wq
->mutex
);
4214 * wq list is used to freeze wq, remove from list after
4215 * flushing is complete in case freeze races us.
4217 mutex_lock(&wq_pool_mutex
);
4218 list_del_init(&wq
->list
);
4219 mutex_unlock(&wq_pool_mutex
);
4221 workqueue_sysfs_unregister(wq
);
4224 kthread_stop(wq
->rescuer
->task
);
4229 if (!(wq
->flags
& WQ_UNBOUND
)) {
4231 * The base ref is never dropped on per-cpu pwqs. Directly
4232 * free the pwqs and wq.
4234 free_percpu(wq
->cpu_pwqs
);
4238 * We're the sole accessor of @wq at this point. Directly
4239 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4240 * @wq will be freed when the last pwq is released.
4242 for_each_node(node
) {
4243 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4244 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4245 put_pwq_unlocked(pwq
);
4249 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4250 * put. Don't access it afterwards.
4254 put_pwq_unlocked(pwq
);
4257 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4260 * workqueue_set_max_active - adjust max_active of a workqueue
4261 * @wq: target workqueue
4262 * @max_active: new max_active value.
4264 * Set max_active of @wq to @max_active.
4267 * Don't call from IRQ context.
4269 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4271 struct pool_workqueue
*pwq
;
4273 /* disallow meddling with max_active for ordered workqueues */
4274 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4277 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4279 mutex_lock(&wq
->mutex
);
4281 wq
->saved_max_active
= max_active
;
4283 for_each_pwq(pwq
, wq
)
4284 pwq_adjust_max_active(pwq
);
4286 mutex_unlock(&wq
->mutex
);
4288 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4291 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4293 * Determine whether %current is a workqueue rescuer. Can be used from
4294 * work functions to determine whether it's being run off the rescuer task.
4296 bool current_is_workqueue_rescuer(void)
4298 struct worker
*worker
= current_wq_worker();
4300 return worker
&& worker
->rescue_wq
;
4304 * workqueue_congested - test whether a workqueue is congested
4305 * @cpu: CPU in question
4306 * @wq: target workqueue
4308 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4309 * no synchronization around this function and the test result is
4310 * unreliable and only useful as advisory hints or for debugging.
4313 * %true if congested, %false otherwise.
4315 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4317 struct pool_workqueue
*pwq
;
4320 rcu_read_lock_sched();
4322 if (!(wq
->flags
& WQ_UNBOUND
))
4323 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4325 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4327 ret
= !list_empty(&pwq
->delayed_works
);
4328 rcu_read_unlock_sched();
4332 EXPORT_SYMBOL_GPL(workqueue_congested
);
4335 * work_busy - test whether a work is currently pending or running
4336 * @work: the work to be tested
4338 * Test whether @work is currently pending or running. There is no
4339 * synchronization around this function and the test result is
4340 * unreliable and only useful as advisory hints or for debugging.
4343 * OR'd bitmask of WORK_BUSY_* bits.
4345 unsigned int work_busy(struct work_struct
*work
)
4347 struct worker_pool
*pool
;
4348 unsigned long flags
;
4349 unsigned int ret
= 0;
4351 if (work_pending(work
))
4352 ret
|= WORK_BUSY_PENDING
;
4354 local_irq_save(flags
);
4355 pool
= get_work_pool(work
);
4357 spin_lock(&pool
->lock
);
4358 if (find_worker_executing_work(pool
, work
))
4359 ret
|= WORK_BUSY_RUNNING
;
4360 spin_unlock(&pool
->lock
);
4362 local_irq_restore(flags
);
4366 EXPORT_SYMBOL_GPL(work_busy
);
4371 * There are two challenges in supporting CPU hotplug. Firstly, there
4372 * are a lot of assumptions on strong associations among work, pwq and
4373 * pool which make migrating pending and scheduled works very
4374 * difficult to implement without impacting hot paths. Secondly,
4375 * worker pools serve mix of short, long and very long running works making
4376 * blocked draining impractical.
4378 * This is solved by allowing the pools to be disassociated from the CPU
4379 * running as an unbound one and allowing it to be reattached later if the
4380 * cpu comes back online.
4383 static void wq_unbind_fn(struct work_struct
*work
)
4385 int cpu
= smp_processor_id();
4386 struct worker_pool
*pool
;
4387 struct worker
*worker
;
4390 for_each_cpu_worker_pool(pool
, cpu
) {
4391 WARN_ON_ONCE(cpu
!= smp_processor_id());
4393 mutex_lock(&pool
->manager_mutex
);
4394 spin_lock_irq(&pool
->lock
);
4397 * We've blocked all manager operations. Make all workers
4398 * unbound and set DISASSOCIATED. Before this, all workers
4399 * except for the ones which are still executing works from
4400 * before the last CPU down must be on the cpu. After
4401 * this, they may become diasporas.
4403 for_each_pool_worker(worker
, wi
, pool
)
4404 worker
->flags
|= WORKER_UNBOUND
;
4406 pool
->flags
|= POOL_DISASSOCIATED
;
4408 spin_unlock_irq(&pool
->lock
);
4409 mutex_unlock(&pool
->manager_mutex
);
4412 * Call schedule() so that we cross rq->lock and thus can
4413 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4414 * This is necessary as scheduler callbacks may be invoked
4420 * Sched callbacks are disabled now. Zap nr_running.
4421 * After this, nr_running stays zero and need_more_worker()
4422 * and keep_working() are always true as long as the
4423 * worklist is not empty. This pool now behaves as an
4424 * unbound (in terms of concurrency management) pool which
4425 * are served by workers tied to the pool.
4427 atomic_set(&pool
->nr_running
, 0);
4430 * With concurrency management just turned off, a busy
4431 * worker blocking could lead to lengthy stalls. Kick off
4432 * unbound chain execution of currently pending work items.
4434 spin_lock_irq(&pool
->lock
);
4435 wake_up_worker(pool
);
4436 spin_unlock_irq(&pool
->lock
);
4441 * rebind_workers - rebind all workers of a pool to the associated CPU
4442 * @pool: pool of interest
4444 * @pool->cpu is coming online. Rebind all workers to the CPU.
4446 static void rebind_workers(struct worker_pool
*pool
)
4448 struct worker
*worker
;
4451 lockdep_assert_held(&pool
->manager_mutex
);
4454 * Restore CPU affinity of all workers. As all idle workers should
4455 * be on the run-queue of the associated CPU before any local
4456 * wake-ups for concurrency management happen, restore CPU affinty
4457 * of all workers first and then clear UNBOUND. As we're called
4458 * from CPU_ONLINE, the following shouldn't fail.
4460 for_each_pool_worker(worker
, wi
, pool
)
4461 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4462 pool
->attrs
->cpumask
) < 0);
4464 spin_lock_irq(&pool
->lock
);
4466 for_each_pool_worker(worker
, wi
, pool
) {
4467 unsigned int worker_flags
= worker
->flags
;
4470 * A bound idle worker should actually be on the runqueue
4471 * of the associated CPU for local wake-ups targeting it to
4472 * work. Kick all idle workers so that they migrate to the
4473 * associated CPU. Doing this in the same loop as
4474 * replacing UNBOUND with REBOUND is safe as no worker will
4475 * be bound before @pool->lock is released.
4477 if (worker_flags
& WORKER_IDLE
)
4478 wake_up_process(worker
->task
);
4481 * We want to clear UNBOUND but can't directly call
4482 * worker_clr_flags() or adjust nr_running. Atomically
4483 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4484 * @worker will clear REBOUND using worker_clr_flags() when
4485 * it initiates the next execution cycle thus restoring
4486 * concurrency management. Note that when or whether
4487 * @worker clears REBOUND doesn't affect correctness.
4489 * ACCESS_ONCE() is necessary because @worker->flags may be
4490 * tested without holding any lock in
4491 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4492 * fail incorrectly leading to premature concurrency
4493 * management operations.
4495 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4496 worker_flags
|= WORKER_REBOUND
;
4497 worker_flags
&= ~WORKER_UNBOUND
;
4498 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4501 spin_unlock_irq(&pool
->lock
);
4505 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4506 * @pool: unbound pool of interest
4507 * @cpu: the CPU which is coming up
4509 * An unbound pool may end up with a cpumask which doesn't have any online
4510 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4511 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4512 * online CPU before, cpus_allowed of all its workers should be restored.
4514 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4516 static cpumask_t cpumask
;
4517 struct worker
*worker
;
4520 lockdep_assert_held(&pool
->manager_mutex
);
4522 /* is @cpu allowed for @pool? */
4523 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4526 /* is @cpu the only online CPU? */
4527 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4528 if (cpumask_weight(&cpumask
) != 1)
4531 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4532 for_each_pool_worker(worker
, wi
, pool
)
4533 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4534 pool
->attrs
->cpumask
) < 0);
4538 * Workqueues should be brought up before normal priority CPU notifiers.
4539 * This will be registered high priority CPU notifier.
4541 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4542 unsigned long action
,
4545 int cpu
= (unsigned long)hcpu
;
4546 struct worker_pool
*pool
;
4547 struct workqueue_struct
*wq
;
4550 switch (action
& ~CPU_TASKS_FROZEN
) {
4551 case CPU_UP_PREPARE
:
4552 for_each_cpu_worker_pool(pool
, cpu
) {
4553 if (pool
->nr_workers
)
4555 if (create_and_start_worker(pool
) < 0)
4560 case CPU_DOWN_FAILED
:
4562 mutex_lock(&wq_pool_mutex
);
4564 for_each_pool(pool
, pi
) {
4565 mutex_lock(&pool
->manager_mutex
);
4567 if (pool
->cpu
== cpu
) {
4568 spin_lock_irq(&pool
->lock
);
4569 pool
->flags
&= ~POOL_DISASSOCIATED
;
4570 spin_unlock_irq(&pool
->lock
);
4572 rebind_workers(pool
);
4573 } else if (pool
->cpu
< 0) {
4574 restore_unbound_workers_cpumask(pool
, cpu
);
4577 mutex_unlock(&pool
->manager_mutex
);
4580 /* update NUMA affinity of unbound workqueues */
4581 list_for_each_entry(wq
, &workqueues
, list
)
4582 wq_update_unbound_numa(wq
, cpu
, true);
4584 mutex_unlock(&wq_pool_mutex
);
4591 * Workqueues should be brought down after normal priority CPU notifiers.
4592 * This will be registered as low priority CPU notifier.
4594 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4595 unsigned long action
,
4598 int cpu
= (unsigned long)hcpu
;
4599 struct work_struct unbind_work
;
4600 struct workqueue_struct
*wq
;
4602 switch (action
& ~CPU_TASKS_FROZEN
) {
4603 case CPU_DOWN_PREPARE
:
4604 /* unbinding per-cpu workers should happen on the local CPU */
4605 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4606 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4608 /* update NUMA affinity of unbound workqueues */
4609 mutex_lock(&wq_pool_mutex
);
4610 list_for_each_entry(wq
, &workqueues
, list
)
4611 wq_update_unbound_numa(wq
, cpu
, false);
4612 mutex_unlock(&wq_pool_mutex
);
4614 /* wait for per-cpu unbinding to finish */
4615 flush_work(&unbind_work
);
4623 struct work_for_cpu
{
4624 struct work_struct work
;
4630 static void work_for_cpu_fn(struct work_struct
*work
)
4632 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4634 wfc
->ret
= wfc
->fn(wfc
->arg
);
4638 * work_on_cpu - run a function in user context on a particular cpu
4639 * @cpu: the cpu to run on
4640 * @fn: the function to run
4641 * @arg: the function arg
4643 * This will return the value @fn returns.
4644 * It is up to the caller to ensure that the cpu doesn't go offline.
4645 * The caller must not hold any locks which would prevent @fn from completing.
4647 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4649 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4651 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4652 schedule_work_on(cpu
, &wfc
.work
);
4653 flush_work(&wfc
.work
);
4656 EXPORT_SYMBOL_GPL(work_on_cpu
);
4657 #endif /* CONFIG_SMP */
4659 #ifdef CONFIG_FREEZER
4662 * freeze_workqueues_begin - begin freezing workqueues
4664 * Start freezing workqueues. After this function returns, all freezable
4665 * workqueues will queue new works to their delayed_works list instead of
4669 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4671 void freeze_workqueues_begin(void)
4673 struct worker_pool
*pool
;
4674 struct workqueue_struct
*wq
;
4675 struct pool_workqueue
*pwq
;
4678 mutex_lock(&wq_pool_mutex
);
4680 WARN_ON_ONCE(workqueue_freezing
);
4681 workqueue_freezing
= true;
4684 for_each_pool(pool
, pi
) {
4685 spin_lock_irq(&pool
->lock
);
4686 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4687 pool
->flags
|= POOL_FREEZING
;
4688 spin_unlock_irq(&pool
->lock
);
4691 list_for_each_entry(wq
, &workqueues
, list
) {
4692 mutex_lock(&wq
->mutex
);
4693 for_each_pwq(pwq
, wq
)
4694 pwq_adjust_max_active(pwq
);
4695 mutex_unlock(&wq
->mutex
);
4698 mutex_unlock(&wq_pool_mutex
);
4702 * freeze_workqueues_busy - are freezable workqueues still busy?
4704 * Check whether freezing is complete. This function must be called
4705 * between freeze_workqueues_begin() and thaw_workqueues().
4708 * Grabs and releases wq_pool_mutex.
4711 * %true if some freezable workqueues are still busy. %false if freezing
4714 bool freeze_workqueues_busy(void)
4717 struct workqueue_struct
*wq
;
4718 struct pool_workqueue
*pwq
;
4720 mutex_lock(&wq_pool_mutex
);
4722 WARN_ON_ONCE(!workqueue_freezing
);
4724 list_for_each_entry(wq
, &workqueues
, list
) {
4725 if (!(wq
->flags
& WQ_FREEZABLE
))
4728 * nr_active is monotonically decreasing. It's safe
4729 * to peek without lock.
4731 rcu_read_lock_sched();
4732 for_each_pwq(pwq
, wq
) {
4733 WARN_ON_ONCE(pwq
->nr_active
< 0);
4734 if (pwq
->nr_active
) {
4736 rcu_read_unlock_sched();
4740 rcu_read_unlock_sched();
4743 mutex_unlock(&wq_pool_mutex
);
4748 * thaw_workqueues - thaw workqueues
4750 * Thaw workqueues. Normal queueing is restored and all collected
4751 * frozen works are transferred to their respective pool worklists.
4754 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4756 void thaw_workqueues(void)
4758 struct workqueue_struct
*wq
;
4759 struct pool_workqueue
*pwq
;
4760 struct worker_pool
*pool
;
4763 mutex_lock(&wq_pool_mutex
);
4765 if (!workqueue_freezing
)
4768 /* clear FREEZING */
4769 for_each_pool(pool
, pi
) {
4770 spin_lock_irq(&pool
->lock
);
4771 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4772 pool
->flags
&= ~POOL_FREEZING
;
4773 spin_unlock_irq(&pool
->lock
);
4776 /* restore max_active and repopulate worklist */
4777 list_for_each_entry(wq
, &workqueues
, list
) {
4778 mutex_lock(&wq
->mutex
);
4779 for_each_pwq(pwq
, wq
)
4780 pwq_adjust_max_active(pwq
);
4781 mutex_unlock(&wq
->mutex
);
4784 workqueue_freezing
= false;
4786 mutex_unlock(&wq_pool_mutex
);
4788 #endif /* CONFIG_FREEZER */
4790 static void __init
wq_numa_init(void)
4795 /* determine NUMA pwq table len - highest node id + 1 */
4797 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4799 if (num_possible_nodes() <= 1)
4802 if (wq_disable_numa
) {
4803 pr_info("workqueue: NUMA affinity support disabled\n");
4807 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4808 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4811 * We want masks of possible CPUs of each node which isn't readily
4812 * available. Build one from cpu_to_node() which should have been
4813 * fully initialized by now.
4815 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
4819 BUG_ON(!alloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
, node
));
4821 for_each_possible_cpu(cpu
) {
4822 node
= cpu_to_node(cpu
);
4823 if (WARN_ON(node
== NUMA_NO_NODE
)) {
4824 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
4825 /* happens iff arch is bonkers, let's just proceed */
4828 cpumask_set_cpu(cpu
, tbl
[node
]);
4831 wq_numa_possible_cpumask
= tbl
;
4832 wq_numa_enabled
= true;
4835 static int __init
init_workqueues(void)
4837 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4840 /* make sure we have enough bits for OFFQ pool ID */
4841 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
4842 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
4844 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4846 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4848 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4849 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4853 /* initialize CPU pools */
4854 for_each_possible_cpu(cpu
) {
4855 struct worker_pool
*pool
;
4858 for_each_cpu_worker_pool(pool
, cpu
) {
4859 BUG_ON(init_worker_pool(pool
));
4861 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4862 pool
->attrs
->nice
= std_nice
[i
++];
4863 pool
->node
= cpu_to_node(cpu
);
4866 mutex_lock(&wq_pool_mutex
);
4867 BUG_ON(worker_pool_assign_id(pool
));
4868 mutex_unlock(&wq_pool_mutex
);
4872 /* create the initial worker */
4873 for_each_online_cpu(cpu
) {
4874 struct worker_pool
*pool
;
4876 for_each_cpu_worker_pool(pool
, cpu
) {
4877 pool
->flags
&= ~POOL_DISASSOCIATED
;
4878 BUG_ON(create_and_start_worker(pool
) < 0);
4882 /* create default unbound wq attrs */
4883 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4884 struct workqueue_attrs
*attrs
;
4886 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4887 attrs
->nice
= std_nice
[i
];
4888 unbound_std_wq_attrs
[i
] = attrs
;
4891 system_wq
= alloc_workqueue("events", 0, 0);
4892 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4893 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4894 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4895 WQ_UNBOUND_MAX_ACTIVE
);
4896 system_freezable_wq
= alloc_workqueue("events_freezable",
4898 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
4899 !system_unbound_wq
|| !system_freezable_wq
);
4902 early_initcall(init_workqueues
);