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>
49 #include <linux/uaccess.h>
51 #include "workqueue_internal.h"
57 * A bound pool is either associated or disassociated with its CPU.
58 * While associated (!DISASSOCIATED), all workers are bound to the
59 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * While DISASSOCIATED, the cpu may be offline and all workers have
63 * %WORKER_UNBOUND set and concurrency management disabled, and may
64 * be executing on any CPU. The pool behaves as an unbound one.
66 * Note that DISASSOCIATED should be flipped only while holding
67 * manager_mutex to avoid changing binding state while
68 * create_worker() is in progress.
70 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
71 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
72 POOL_FREEZING
= 1 << 3, /* freeze in progress */
75 WORKER_STARTED
= 1 << 0, /* started */
76 WORKER_DIE
= 1 << 1, /* die die die */
77 WORKER_IDLE
= 1 << 2, /* is idle */
78 WORKER_PREP
= 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
81 WORKER_REBOUND
= 1 << 8, /* worker was rebound */
83 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_CPU_INTENSIVE
|
84 WORKER_UNBOUND
| WORKER_REBOUND
,
86 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
88 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
91 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
94 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
95 /* call for help after 10ms
97 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
98 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give -20.
104 RESCUER_NICE_LEVEL
= -20,
105 HIGHPRI_NICE_LEVEL
= -20,
111 * Structure fields follow one of the following exclusion rules.
113 * I: Modifiable by initialization/destruction paths and read-only for
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
119 * L: pool->lock protected. Access with pool->lock held.
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
126 * MG: pool->manager_mutex and pool->lock protected. Writes require both
127 * locks. Reads can happen under either lock.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * WQ: wq->mutex protected.
135 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
137 * MD: wq_mayday_lock protected.
140 /* struct worker is defined in workqueue_internal.h */
143 spinlock_t lock
; /* the pool lock */
144 int cpu
; /* I: the associated cpu */
145 int node
; /* I: the associated node ID */
146 int id
; /* I: pool ID */
147 unsigned int flags
; /* X: flags */
149 struct list_head worklist
; /* L: list of pending works */
150 int nr_workers
; /* L: total number of workers */
152 /* nr_idle includes the ones off idle_list for rebinding */
153 int nr_idle
; /* L: currently idle ones */
155 struct list_head idle_list
; /* X: list of idle workers */
156 struct timer_list idle_timer
; /* L: worker idle timeout */
157 struct timer_list mayday_timer
; /* L: SOS timer for workers */
159 /* a workers is either on busy_hash or idle_list, or the manager */
160 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
161 /* L: hash of busy workers */
163 /* see manage_workers() for details on the two manager mutexes */
164 struct mutex manager_arb
; /* manager arbitration */
165 struct mutex manager_mutex
; /* manager exclusion */
166 struct idr worker_idr
; /* MG: worker IDs and iteration */
168 struct workqueue_attrs
*attrs
; /* I: worker attributes */
169 struct hlist_node hash_node
; /* PL: unbound_pool_hash node */
170 int refcnt
; /* PL: refcnt for unbound pools */
173 * The current concurrency level. As it's likely to be accessed
174 * from other CPUs during try_to_wake_up(), put it in a separate
177 atomic_t nr_running ____cacheline_aligned_in_smp
;
180 * Destruction of pool is sched-RCU protected to allow dereferences
181 * from get_work_pool().
184 } ____cacheline_aligned_in_smp
;
187 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 * of work_struct->data are used for flags and the remaining high bits
189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 * number of flag bits.
192 struct pool_workqueue
{
193 struct worker_pool
*pool
; /* I: the associated pool */
194 struct workqueue_struct
*wq
; /* I: the owning workqueue */
195 int work_color
; /* L: current color */
196 int flush_color
; /* L: flushing color */
197 int refcnt
; /* L: reference count */
198 int nr_in_flight
[WORK_NR_COLORS
];
199 /* L: nr of in_flight works */
200 int nr_active
; /* L: nr of active works */
201 int max_active
; /* L: max active works */
202 struct list_head delayed_works
; /* L: delayed works */
203 struct list_head pwqs_node
; /* WR: node on wq->pwqs */
204 struct list_head mayday_node
; /* MD: node on wq->maydays */
207 * Release of unbound pwq is punted to system_wq. See put_pwq()
208 * and pwq_unbound_release_workfn() for details. pool_workqueue
209 * itself is also sched-RCU protected so that the first pwq can be
210 * determined without grabbing wq->mutex.
212 struct work_struct unbound_release_work
;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
217 * Structure used to wait for workqueue flush.
220 struct list_head list
; /* WQ: list of flushers */
221 int flush_color
; /* WQ: flush color waiting for */
222 struct completion done
; /* flush completion */
228 * The externally visible workqueue. It relays the issued work items to
229 * the appropriate worker_pool through its pool_workqueues.
231 struct workqueue_struct
{
232 struct list_head pwqs
; /* WR: all pwqs of this wq */
233 struct list_head list
; /* PL: list of all workqueues */
235 struct mutex mutex
; /* protects this wq */
236 int work_color
; /* WQ: current work color */
237 int flush_color
; /* WQ: current flush color */
238 atomic_t nr_pwqs_to_flush
; /* flush in progress */
239 struct wq_flusher
*first_flusher
; /* WQ: first flusher */
240 struct list_head flusher_queue
; /* WQ: flush waiters */
241 struct list_head flusher_overflow
; /* WQ: flush overflow list */
243 struct list_head maydays
; /* MD: pwqs requesting rescue */
244 struct worker
*rescuer
; /* I: rescue worker */
246 int nr_drainers
; /* WQ: drain in progress */
247 int saved_max_active
; /* WQ: saved pwq max_active */
249 struct workqueue_attrs
*unbound_attrs
; /* WQ: only for unbound wqs */
250 struct pool_workqueue
*dfl_pwq
; /* WQ: only for unbound wqs */
253 struct wq_device
*wq_dev
; /* I: for sysfs interface */
255 #ifdef CONFIG_LOCKDEP
256 struct lockdep_map lockdep_map
;
258 char name
[WQ_NAME_LEN
]; /* I: workqueue name */
260 /* hot fields used during command issue, aligned to cacheline */
261 unsigned int flags ____cacheline_aligned
; /* WQ: WQ_* flags */
262 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwqs */
263 struct pool_workqueue __rcu
*numa_pwq_tbl
[]; /* FR: unbound pwqs indexed by node */
266 static struct kmem_cache
*pwq_cache
;
268 static int wq_numa_tbl_len
; /* highest possible NUMA node id + 1 */
269 static cpumask_var_t
*wq_numa_possible_cpumask
;
270 /* possible CPUs of each node */
272 static bool wq_disable_numa
;
273 module_param_named(disable_numa
, wq_disable_numa
, bool, 0444);
275 static bool wq_numa_enabled
; /* unbound NUMA affinity enabled */
277 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
278 static struct workqueue_attrs
*wq_update_unbound_numa_attrs_buf
;
280 static DEFINE_MUTEX(wq_pool_mutex
); /* protects pools and workqueues list */
281 static DEFINE_SPINLOCK(wq_mayday_lock
); /* protects wq->maydays list */
283 static LIST_HEAD(workqueues
); /* PL: list of all workqueues */
284 static bool workqueue_freezing
; /* PL: have wqs started freezing? */
286 /* the per-cpu worker pools */
287 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
290 static DEFINE_IDR(worker_pool_idr
); /* PR: idr of all pools */
292 /* PL: hash of all unbound pools keyed by pool->attrs */
293 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
295 /* I: attributes used when instantiating standard unbound pools on demand */
296 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
298 /* I: attributes used when instantiating ordered pools on demand */
299 static struct workqueue_attrs
*ordered_wq_attrs
[NR_STD_WORKER_POOLS
];
301 struct workqueue_struct
*system_wq __read_mostly
;
302 EXPORT_SYMBOL(system_wq
);
303 struct workqueue_struct
*system_highpri_wq __read_mostly
;
304 EXPORT_SYMBOL_GPL(system_highpri_wq
);
305 struct workqueue_struct
*system_long_wq __read_mostly
;
306 EXPORT_SYMBOL_GPL(system_long_wq
);
307 struct workqueue_struct
*system_unbound_wq __read_mostly
;
308 EXPORT_SYMBOL_GPL(system_unbound_wq
);
309 struct workqueue_struct
*system_freezable_wq __read_mostly
;
310 EXPORT_SYMBOL_GPL(system_freezable_wq
);
312 static int worker_thread(void *__worker
);
313 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
314 const struct workqueue_attrs
*from
);
316 #define CREATE_TRACE_POINTS
317 #include <trace/events/workqueue.h>
319 #define assert_rcu_or_pool_mutex() \
320 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
321 lockdep_is_held(&wq_pool_mutex), \
322 "sched RCU or wq_pool_mutex should be held")
324 #define assert_rcu_or_wq_mutex(wq) \
325 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
326 lockdep_is_held(&wq->mutex), \
327 "sched RCU or wq->mutex should be held")
329 #ifdef CONFIG_LOCKDEP
330 #define assert_manager_or_pool_lock(pool) \
331 WARN_ONCE(debug_locks && \
332 !lockdep_is_held(&(pool)->manager_mutex) && \
333 !lockdep_is_held(&(pool)->lock), \
334 "pool->manager_mutex or ->lock should be held")
336 #define assert_manager_or_pool_lock(pool) do { } while (0)
339 #define for_each_cpu_worker_pool(pool, cpu) \
340 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
341 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
345 * for_each_pool - iterate through all worker_pools in the system
346 * @pool: iteration cursor
347 * @pi: integer used for iteration
349 * This must be called either with wq_pool_mutex held or sched RCU read
350 * locked. If the pool needs to be used beyond the locking in effect, the
351 * caller is responsible for guaranteeing that the pool stays online.
353 * The if/else clause exists only for the lockdep assertion and can be
356 #define for_each_pool(pool, pi) \
357 idr_for_each_entry(&worker_pool_idr, pool, pi) \
358 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
362 * for_each_pool_worker - iterate through all workers of a worker_pool
363 * @worker: iteration cursor
364 * @wi: integer used for iteration
365 * @pool: worker_pool to iterate workers of
367 * This must be called with either @pool->manager_mutex or ->lock held.
369 * The if/else clause exists only for the lockdep assertion and can be
372 #define for_each_pool_worker(worker, wi, pool) \
373 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
374 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
378 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
379 * @pwq: iteration cursor
380 * @wq: the target workqueue
382 * This must be called either with wq->mutex held or sched RCU read locked.
383 * If the pwq needs to be used beyond the locking in effect, the caller is
384 * responsible for guaranteeing that the pwq stays online.
386 * The if/else clause exists only for the lockdep assertion and can be
389 #define for_each_pwq(pwq, wq) \
390 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
391 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
394 #ifdef CONFIG_DEBUG_OBJECTS_WORK
396 static struct debug_obj_descr work_debug_descr
;
398 static void *work_debug_hint(void *addr
)
400 return ((struct work_struct
*) addr
)->func
;
404 * fixup_init is called when:
405 * - an active object is initialized
407 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
409 struct work_struct
*work
= addr
;
412 case ODEBUG_STATE_ACTIVE
:
413 cancel_work_sync(work
);
414 debug_object_init(work
, &work_debug_descr
);
422 * fixup_activate is called when:
423 * - an active object is activated
424 * - an unknown object is activated (might be a statically initialized object)
426 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
428 struct work_struct
*work
= addr
;
432 case ODEBUG_STATE_NOTAVAILABLE
:
434 * This is not really a fixup. The work struct was
435 * statically initialized. We just make sure that it
436 * is tracked in the object tracker.
438 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
439 debug_object_init(work
, &work_debug_descr
);
440 debug_object_activate(work
, &work_debug_descr
);
446 case ODEBUG_STATE_ACTIVE
:
455 * fixup_free is called when:
456 * - an active object is freed
458 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
460 struct work_struct
*work
= addr
;
463 case ODEBUG_STATE_ACTIVE
:
464 cancel_work_sync(work
);
465 debug_object_free(work
, &work_debug_descr
);
472 static struct debug_obj_descr work_debug_descr
= {
473 .name
= "work_struct",
474 .debug_hint
= work_debug_hint
,
475 .fixup_init
= work_fixup_init
,
476 .fixup_activate
= work_fixup_activate
,
477 .fixup_free
= work_fixup_free
,
480 static inline void debug_work_activate(struct work_struct
*work
)
482 debug_object_activate(work
, &work_debug_descr
);
485 static inline void debug_work_deactivate(struct work_struct
*work
)
487 debug_object_deactivate(work
, &work_debug_descr
);
490 void __init_work(struct work_struct
*work
, int onstack
)
493 debug_object_init_on_stack(work
, &work_debug_descr
);
495 debug_object_init(work
, &work_debug_descr
);
497 EXPORT_SYMBOL_GPL(__init_work
);
499 void destroy_work_on_stack(struct work_struct
*work
)
501 debug_object_free(work
, &work_debug_descr
);
503 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
506 static inline void debug_work_activate(struct work_struct
*work
) { }
507 static inline void debug_work_deactivate(struct work_struct
*work
) { }
510 #ifdef CONFIG_MTK_WQ_DEBUG
511 extern void mttrace_workqueue_execute_work(struct work_struct
*work
);
512 extern void mttrace_workqueue_activate_work(struct work_struct
*work
);
513 extern void mttrace_workqueue_queue_work(unsigned int req_cpu
, struct work_struct
*work
);
514 extern void mttrace_workqueue_execute_end(struct work_struct
*work
);
515 #endif //CONFIG_MTK_WQ_DEBUG
517 /* allocate ID and assign it to @pool */
518 static int worker_pool_assign_id(struct worker_pool
*pool
)
522 lockdep_assert_held(&wq_pool_mutex
);
524 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, 0, GFP_KERNEL
);
533 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
534 * @wq: the target workqueue
537 * This must be called either with pwq_lock held or sched RCU read locked.
538 * If the pwq needs to be used beyond the locking in effect, the caller is
539 * responsible for guaranteeing that the pwq stays online.
541 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
544 assert_rcu_or_wq_mutex(wq
);
545 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
548 static unsigned int work_color_to_flags(int color
)
550 return color
<< WORK_STRUCT_COLOR_SHIFT
;
553 static int get_work_color(struct work_struct
*work
)
555 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
556 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
559 static int work_next_color(int color
)
561 return (color
+ 1) % WORK_NR_COLORS
;
565 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
566 * contain the pointer to the queued pwq. Once execution starts, the flag
567 * is cleared and the high bits contain OFFQ flags and pool ID.
569 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
570 * and clear_work_data() can be used to set the pwq, pool or clear
571 * work->data. These functions should only be called while the work is
572 * owned - ie. while the PENDING bit is set.
574 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
575 * corresponding to a work. Pool is available once the work has been
576 * queued anywhere after initialization until it is sync canceled. pwq is
577 * available only while the work item is queued.
579 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
580 * canceled. While being canceled, a work item may have its PENDING set
581 * but stay off timer and worklist for arbitrarily long and nobody should
582 * try to steal the PENDING bit.
584 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
587 WARN_ON_ONCE(!work_pending(work
));
588 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
591 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
592 unsigned long extra_flags
)
594 set_work_data(work
, (unsigned long)pwq
,
595 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
598 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
601 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
602 WORK_STRUCT_PENDING
);
605 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
609 * The following wmb is paired with the implied mb in
610 * test_and_set_bit(PENDING) and ensures all updates to @work made
611 * here are visible to and precede any updates by the next PENDING
615 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
618 static void clear_work_data(struct work_struct
*work
)
620 smp_wmb(); /* see set_work_pool_and_clear_pending() */
621 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
624 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
626 unsigned long data
= atomic_long_read(&work
->data
);
628 if (data
& WORK_STRUCT_PWQ
)
629 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
635 * get_work_pool - return the worker_pool a given work was associated with
636 * @work: the work item of interest
638 * Return the worker_pool @work was last associated with. %NULL if none.
640 * Pools are created and destroyed under wq_pool_mutex, and allows read
641 * access under sched-RCU read lock. As such, this function should be
642 * called under wq_pool_mutex or with preemption disabled.
644 * All fields of the returned pool are accessible as long as the above
645 * mentioned locking is in effect. If the returned pool needs to be used
646 * beyond the critical section, the caller is responsible for ensuring the
647 * returned pool is and stays online.
649 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
651 unsigned long data
= atomic_long_read(&work
->data
);
654 assert_rcu_or_pool_mutex();
656 if (data
& WORK_STRUCT_PWQ
)
657 return ((struct pool_workqueue
*)
658 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
660 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
661 if (pool_id
== WORK_OFFQ_POOL_NONE
)
664 return idr_find(&worker_pool_idr
, pool_id
);
668 * get_work_pool_id - return the worker pool ID a given work is associated with
669 * @work: the work item of interest
671 * Return the worker_pool ID @work was last associated with.
672 * %WORK_OFFQ_POOL_NONE if none.
674 static int get_work_pool_id(struct work_struct
*work
)
676 unsigned long data
= atomic_long_read(&work
->data
);
678 if (data
& WORK_STRUCT_PWQ
)
679 return ((struct pool_workqueue
*)
680 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
682 return data
>> WORK_OFFQ_POOL_SHIFT
;
685 static void mark_work_canceling(struct work_struct
*work
)
687 unsigned long pool_id
= get_work_pool_id(work
);
689 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
690 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
693 static bool work_is_canceling(struct work_struct
*work
)
695 unsigned long data
= atomic_long_read(&work
->data
);
697 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
701 * Policy functions. These define the policies on how the global worker
702 * pools are managed. Unless noted otherwise, these functions assume that
703 * they're being called with pool->lock held.
706 static bool __need_more_worker(struct worker_pool
*pool
)
708 return !atomic_read(&pool
->nr_running
);
712 * Need to wake up a worker? Called from anything but currently
715 * Note that, because unbound workers never contribute to nr_running, this
716 * function will always return %true for unbound pools as long as the
717 * worklist isn't empty.
719 static bool need_more_worker(struct worker_pool
*pool
)
721 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
724 /* Can I start working? Called from busy but !running workers. */
725 static bool may_start_working(struct worker_pool
*pool
)
727 return pool
->nr_idle
;
730 /* Do I need to keep working? Called from currently running workers. */
731 static bool keep_working(struct worker_pool
*pool
)
733 return !list_empty(&pool
->worklist
) &&
734 atomic_read(&pool
->nr_running
) <= 1;
737 /* Do we need a new worker? Called from manager. */
738 static bool need_to_create_worker(struct worker_pool
*pool
)
740 return need_more_worker(pool
) && !may_start_working(pool
);
743 /* Do I need to be the manager? */
744 static bool need_to_manage_workers(struct worker_pool
*pool
)
746 return need_to_create_worker(pool
) ||
747 (pool
->flags
& POOL_MANAGE_WORKERS
);
750 /* Do we have too many workers and should some go away? */
751 static bool too_many_workers(struct worker_pool
*pool
)
753 bool managing
= mutex_is_locked(&pool
->manager_arb
);
754 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
755 int nr_busy
= pool
->nr_workers
- nr_idle
;
758 * nr_idle and idle_list may disagree if idle rebinding is in
759 * progress. Never return %true if idle_list is empty.
761 if (list_empty(&pool
->idle_list
))
764 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
771 /* Return the first worker. Safe with preemption disabled */
772 static struct worker
*first_worker(struct worker_pool
*pool
)
774 if (unlikely(list_empty(&pool
->idle_list
)))
777 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
781 * wake_up_worker - wake up an idle worker
782 * @pool: worker pool to wake worker from
784 * Wake up the first idle worker of @pool.
787 * spin_lock_irq(pool->lock).
789 static void wake_up_worker(struct worker_pool
*pool
)
791 struct worker
*worker
= first_worker(pool
);
794 wake_up_process(worker
->task
);
798 * wq_worker_waking_up - a worker is waking up
799 * @task: task waking up
800 * @cpu: CPU @task is waking up to
802 * This function is called during try_to_wake_up() when a worker is
806 * spin_lock_irq(rq->lock)
808 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
810 struct worker
*worker
= kthread_data(task
);
812 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
813 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
814 atomic_inc(&worker
->pool
->nr_running
);
819 * wq_worker_sleeping - a worker is going to sleep
820 * @task: task going to sleep
821 * @cpu: CPU in question, must be the current CPU number
823 * This function is called during schedule() when a busy worker is
824 * going to sleep. Worker on the same cpu can be woken up by
825 * returning pointer to its task.
828 * spin_lock_irq(rq->lock)
831 * Worker task on @cpu to wake up, %NULL if none.
833 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
835 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
836 struct worker_pool
*pool
;
839 * Rescuers, which may not have all the fields set up like normal
840 * workers, also reach here, let's not access anything before
841 * checking NOT_RUNNING.
843 if (worker
->flags
& WORKER_NOT_RUNNING
)
848 /* this can only happen on the local cpu */
849 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
853 * The counterpart of the following dec_and_test, implied mb,
854 * worklist not empty test sequence is in insert_work().
855 * Please read comment there.
857 * NOT_RUNNING is clear. This means that we're bound to and
858 * running on the local cpu w/ rq lock held and preemption
859 * disabled, which in turn means that none else could be
860 * manipulating idle_list, so dereferencing idle_list without pool
863 if (atomic_dec_and_test(&pool
->nr_running
) &&
864 !list_empty(&pool
->worklist
))
865 to_wakeup
= first_worker(pool
);
866 return to_wakeup
? to_wakeup
->task
: NULL
;
870 * worker_set_flags - set worker flags and adjust nr_running accordingly
872 * @flags: flags to set
873 * @wakeup: wakeup an idle worker if necessary
875 * Set @flags in @worker->flags and adjust nr_running accordingly. If
876 * nr_running becomes zero and @wakeup is %true, an idle worker is
880 * spin_lock_irq(pool->lock)
882 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
885 struct worker_pool
*pool
= worker
->pool
;
887 WARN_ON_ONCE(worker
->task
!= current
);
890 * If transitioning into NOT_RUNNING, adjust nr_running and
891 * wake up an idle worker as necessary if requested by
894 if ((flags
& WORKER_NOT_RUNNING
) &&
895 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
897 if (atomic_dec_and_test(&pool
->nr_running
) &&
898 !list_empty(&pool
->worklist
))
899 wake_up_worker(pool
);
901 atomic_dec(&pool
->nr_running
);
904 worker
->flags
|= flags
;
908 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
910 * @flags: flags to clear
912 * Clear @flags in @worker->flags and adjust nr_running accordingly.
915 * spin_lock_irq(pool->lock)
917 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
919 struct worker_pool
*pool
= worker
->pool
;
920 unsigned int oflags
= worker
->flags
;
922 WARN_ON_ONCE(worker
->task
!= current
);
924 worker
->flags
&= ~flags
;
927 * If transitioning out of NOT_RUNNING, increment nr_running. Note
928 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
929 * of multiple flags, not a single flag.
931 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
932 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
933 atomic_inc(&pool
->nr_running
);
937 * find_worker_executing_work - find worker which is executing a work
938 * @pool: pool of interest
939 * @work: work to find worker for
941 * Find a worker which is executing @work on @pool by searching
942 * @pool->busy_hash which is keyed by the address of @work. For a worker
943 * to match, its current execution should match the address of @work and
944 * its work function. This is to avoid unwanted dependency between
945 * unrelated work executions through a work item being recycled while still
948 * This is a bit tricky. A work item may be freed once its execution
949 * starts and nothing prevents the freed area from being recycled for
950 * another work item. If the same work item address ends up being reused
951 * before the original execution finishes, workqueue will identify the
952 * recycled work item as currently executing and make it wait until the
953 * current execution finishes, introducing an unwanted dependency.
955 * This function checks the work item address and work function to avoid
956 * false positives. Note that this isn't complete as one may construct a
957 * work function which can introduce dependency onto itself through a
958 * recycled work item. Well, if somebody wants to shoot oneself in the
959 * foot that badly, there's only so much we can do, and if such deadlock
960 * actually occurs, it should be easy to locate the culprit work function.
963 * spin_lock_irq(pool->lock).
966 * Pointer to worker which is executing @work if found, NULL
969 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
970 struct work_struct
*work
)
972 struct worker
*worker
;
974 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
976 if (worker
->current_work
== work
&&
977 worker
->current_func
== work
->func
)
984 * move_linked_works - move linked works to a list
985 * @work: start of series of works to be scheduled
986 * @head: target list to append @work to
987 * @nextp: out paramter for nested worklist walking
989 * Schedule linked works starting from @work to @head. Work series to
990 * be scheduled starts at @work and includes any consecutive work with
991 * WORK_STRUCT_LINKED set in its predecessor.
993 * If @nextp is not NULL, it's updated to point to the next work of
994 * the last scheduled work. This allows move_linked_works() to be
995 * nested inside outer list_for_each_entry_safe().
998 * spin_lock_irq(pool->lock).
1000 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1001 struct work_struct
**nextp
)
1003 struct work_struct
*n
;
1006 * Linked worklist will always end before the end of the list,
1007 * use NULL for list head.
1009 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1010 list_move_tail(&work
->entry
, head
);
1011 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1016 * If we're already inside safe list traversal and have moved
1017 * multiple works to the scheduled queue, the next position
1018 * needs to be updated.
1025 * get_pwq - get an extra reference on the specified pool_workqueue
1026 * @pwq: pool_workqueue to get
1028 * Obtain an extra reference on @pwq. The caller should guarantee that
1029 * @pwq has positive refcnt and be holding the matching pool->lock.
1031 static void get_pwq(struct pool_workqueue
*pwq
)
1033 lockdep_assert_held(&pwq
->pool
->lock
);
1034 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1039 * put_pwq - put a pool_workqueue reference
1040 * @pwq: pool_workqueue to put
1042 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1043 * destruction. The caller should be holding the matching pool->lock.
1045 static void put_pwq(struct pool_workqueue
*pwq
)
1047 lockdep_assert_held(&pwq
->pool
->lock
);
1048 if (likely(--pwq
->refcnt
))
1050 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1053 * @pwq can't be released under pool->lock, bounce to
1054 * pwq_unbound_release_workfn(). This never recurses on the same
1055 * pool->lock as this path is taken only for unbound workqueues and
1056 * the release work item is scheduled on a per-cpu workqueue. To
1057 * avoid lockdep warning, unbound pool->locks are given lockdep
1058 * subclass of 1 in get_unbound_pool().
1060 schedule_work(&pwq
->unbound_release_work
);
1064 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1065 * @pwq: pool_workqueue to put (can be %NULL)
1067 * put_pwq() with locking. This function also allows %NULL @pwq.
1069 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1073 * As both pwqs and pools are sched-RCU protected, the
1074 * following lock operations are safe.
1076 spin_lock_irq(&pwq
->pool
->lock
);
1078 spin_unlock_irq(&pwq
->pool
->lock
);
1082 static void pwq_activate_delayed_work(struct work_struct
*work
)
1084 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1086 trace_workqueue_activate_work(work
);
1087 #ifdef CONFIG_MTK_WQ_DEBUG
1088 mttrace_workqueue_activate_work(work
);
1089 #endif //CONFIG_MTK_WQ_DEBUG
1090 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1091 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1095 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1097 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1098 struct work_struct
, entry
);
1100 pwq_activate_delayed_work(work
);
1104 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1105 * @pwq: pwq of interest
1106 * @color: color of work which left the queue
1108 * A work either has completed or is removed from pending queue,
1109 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1112 * spin_lock_irq(pool->lock).
1114 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1116 /* uncolored work items don't participate in flushing or nr_active */
1117 if (color
== WORK_NO_COLOR
)
1120 pwq
->nr_in_flight
[color
]--;
1123 if (!list_empty(&pwq
->delayed_works
)) {
1124 /* one down, submit a delayed one */
1125 if (pwq
->nr_active
< pwq
->max_active
)
1126 pwq_activate_first_delayed(pwq
);
1129 /* is flush in progress and are we at the flushing tip? */
1130 if (likely(pwq
->flush_color
!= color
))
1133 /* are there still in-flight works? */
1134 if (pwq
->nr_in_flight
[color
])
1137 /* this pwq is done, clear flush_color */
1138 pwq
->flush_color
= -1;
1141 * If this was the last pwq, wake up the first flusher. It
1142 * will handle the rest.
1144 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1145 complete(&pwq
->wq
->first_flusher
->done
);
1151 * try_to_grab_pending - steal work item from worklist and disable irq
1152 * @work: work item to steal
1153 * @is_dwork: @work is a delayed_work
1154 * @flags: place to store irq state
1156 * Try to grab PENDING bit of @work. This function can handle @work in any
1157 * stable state - idle, on timer or on worklist. Return values are
1159 * 1 if @work was pending and we successfully stole PENDING
1160 * 0 if @work was idle and we claimed PENDING
1161 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1162 * -ENOENT if someone else is canceling @work, this state may persist
1163 * for arbitrarily long
1165 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1166 * interrupted while holding PENDING and @work off queue, irq must be
1167 * disabled on entry. This, combined with delayed_work->timer being
1168 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1170 * On successful return, >= 0, irq is disabled and the caller is
1171 * responsible for releasing it using local_irq_restore(*@flags).
1173 * This function is safe to call from any context including IRQ handler.
1175 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1176 unsigned long *flags
)
1178 struct worker_pool
*pool
;
1179 struct pool_workqueue
*pwq
;
1181 local_irq_save(*flags
);
1183 /* try to steal the timer if it exists */
1185 struct delayed_work
*dwork
= to_delayed_work(work
);
1188 * dwork->timer is irqsafe. If del_timer() fails, it's
1189 * guaranteed that the timer is not queued anywhere and not
1190 * running on the local CPU.
1192 if (likely(del_timer(&dwork
->timer
)))
1196 /* try to claim PENDING the normal way */
1197 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1201 * The queueing is in progress, or it is already queued. Try to
1202 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1204 pool
= get_work_pool(work
);
1208 spin_lock(&pool
->lock
);
1210 * work->data is guaranteed to point to pwq only while the work
1211 * item is queued on pwq->wq, and both updating work->data to point
1212 * to pwq on queueing and to pool on dequeueing are done under
1213 * pwq->pool->lock. This in turn guarantees that, if work->data
1214 * points to pwq which is associated with a locked pool, the work
1215 * item is currently queued on that pool.
1217 pwq
= get_work_pwq(work
);
1218 if (pwq
&& pwq
->pool
== pool
) {
1219 debug_work_deactivate(work
);
1222 * A delayed work item cannot be grabbed directly because
1223 * it might have linked NO_COLOR work items which, if left
1224 * on the delayed_list, will confuse pwq->nr_active
1225 * management later on and cause stall. Make sure the work
1226 * item is activated before grabbing.
1228 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1229 pwq_activate_delayed_work(work
);
1231 list_del_init(&work
->entry
);
1232 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1234 /* work->data points to pwq iff queued, point to pool */
1235 set_work_pool_and_keep_pending(work
, pool
->id
);
1237 spin_unlock(&pool
->lock
);
1240 spin_unlock(&pool
->lock
);
1242 local_irq_restore(*flags
);
1243 if (work_is_canceling(work
))
1250 * insert_work - insert a work into a pool
1251 * @pwq: pwq @work belongs to
1252 * @work: work to insert
1253 * @head: insertion point
1254 * @extra_flags: extra WORK_STRUCT_* flags to set
1256 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1257 * work_struct flags.
1260 * spin_lock_irq(pool->lock).
1262 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1263 struct list_head
*head
, unsigned int extra_flags
)
1265 struct worker_pool
*pool
= pwq
->pool
;
1267 /* we own @work, set data and link */
1268 set_work_pwq(work
, pwq
, extra_flags
);
1269 list_add_tail(&work
->entry
, head
);
1273 * Ensure either wq_worker_sleeping() sees the above
1274 * list_add_tail() or we see zero nr_running to avoid workers lying
1275 * around lazily while there are works to be processed.
1279 if (__need_more_worker(pool
))
1280 wake_up_worker(pool
);
1284 * Test whether @work is being queued from another work executing on the
1287 static bool is_chained_work(struct workqueue_struct
*wq
)
1289 struct worker
*worker
;
1291 worker
= current_wq_worker();
1293 * Return %true iff I'm a worker execuing a work item on @wq. If
1294 * I'm @worker, it's safe to dereference it without locking.
1296 return worker
&& worker
->current_pwq
->wq
== wq
;
1299 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1300 struct work_struct
*work
)
1302 struct pool_workqueue
*pwq
;
1303 struct worker_pool
*last_pool
;
1304 struct list_head
*worklist
;
1305 unsigned int work_flags
;
1306 unsigned int req_cpu
= cpu
;
1309 * While a work item is PENDING && off queue, a task trying to
1310 * steal the PENDING will busy-loop waiting for it to either get
1311 * queued or lose PENDING. Grabbing PENDING and queueing should
1312 * happen with IRQ disabled.
1314 WARN_ON_ONCE(!irqs_disabled());
1316 debug_work_activate(work
);
1318 /* if dying, only works from the same workqueue are allowed */
1319 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1320 WARN_ON_ONCE(!is_chained_work(wq
)))
1323 if (req_cpu
== WORK_CPU_UNBOUND
)
1324 cpu
= raw_smp_processor_id();
1326 /* pwq which will be used unless @work is executing elsewhere */
1327 if (!(wq
->flags
& WQ_UNBOUND
))
1328 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1330 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1333 * If @work was previously on a different pool, it might still be
1334 * running there, in which case the work needs to be queued on that
1335 * pool to guarantee non-reentrancy.
1337 last_pool
= get_work_pool(work
);
1338 if (last_pool
&& last_pool
!= pwq
->pool
) {
1339 struct worker
*worker
;
1341 spin_lock(&last_pool
->lock
);
1343 worker
= find_worker_executing_work(last_pool
, work
);
1345 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1346 pwq
= worker
->current_pwq
;
1348 /* meh... not running there, queue here */
1349 spin_unlock(&last_pool
->lock
);
1350 spin_lock(&pwq
->pool
->lock
);
1353 spin_lock(&pwq
->pool
->lock
);
1357 * pwq is determined and locked. For unbound pools, we could have
1358 * raced with pwq release and it could already be dead. If its
1359 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1360 * without another pwq replacing it in the numa_pwq_tbl or while
1361 * work items are executing on it, so the retrying is guaranteed to
1362 * make forward-progress.
1364 if (unlikely(!pwq
->refcnt
)) {
1365 if (wq
->flags
& WQ_UNBOUND
) {
1366 spin_unlock(&pwq
->pool
->lock
);
1371 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1375 /* pwq determined, queue */
1376 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1377 #ifdef CONFIG_MTK_WQ_DEBUG
1378 mttrace_workqueue_queue_work(cpu
, work
);
1379 #endif //CONFIG_MTK_WQ_DEBUG
1381 if (WARN_ON(!list_empty(&work
->entry
))) {
1382 spin_unlock(&pwq
->pool
->lock
);
1386 pwq
->nr_in_flight
[pwq
->work_color
]++;
1387 work_flags
= work_color_to_flags(pwq
->work_color
);
1389 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1390 trace_workqueue_activate_work(work
);
1391 #ifdef CONFIG_MTK_WQ_DEBUG
1392 mttrace_workqueue_activate_work(work
);
1393 #endif //CONFIG_MTK_WQ_DEBUG
1395 worklist
= &pwq
->pool
->worklist
;
1397 work_flags
|= WORK_STRUCT_DELAYED
;
1398 worklist
= &pwq
->delayed_works
;
1401 insert_work(pwq
, work
, worklist
, work_flags
);
1403 spin_unlock(&pwq
->pool
->lock
);
1407 * queue_work_on - queue work on specific cpu
1408 * @cpu: CPU number to execute work on
1409 * @wq: workqueue to use
1410 * @work: work to queue
1412 * Returns %false if @work was already on a queue, %true otherwise.
1414 * We queue the work to a specific CPU, the caller must ensure it
1417 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1418 struct work_struct
*work
)
1421 unsigned long flags
;
1423 local_irq_save(flags
);
1425 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1426 __queue_work(cpu
, wq
, work
);
1430 local_irq_restore(flags
);
1433 EXPORT_SYMBOL(queue_work_on
);
1435 void delayed_work_timer_fn(unsigned long __data
)
1437 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1439 /* should have been called from irqsafe timer with irq already off */
1440 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1442 EXPORT_SYMBOL(delayed_work_timer_fn
);
1444 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1445 struct delayed_work
*dwork
, unsigned long delay
)
1447 struct timer_list
*timer
= &dwork
->timer
;
1448 struct work_struct
*work
= &dwork
->work
;
1450 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1451 timer
->data
!= (unsigned long)dwork
);
1452 WARN_ON_ONCE(timer_pending(timer
));
1453 WARN_ON_ONCE(!list_empty(&work
->entry
));
1456 * If @delay is 0, queue @dwork->work immediately. This is for
1457 * both optimization and correctness. The earliest @timer can
1458 * expire is on the closest next tick and delayed_work users depend
1459 * on that there's no such delay when @delay is 0.
1462 __queue_work(cpu
, wq
, &dwork
->work
);
1466 timer_stats_timer_set_start_info(&dwork
->timer
);
1470 timer
->expires
= jiffies
+ delay
;
1472 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1473 add_timer_on(timer
, cpu
);
1479 * queue_delayed_work_on - queue work on specific CPU after delay
1480 * @cpu: CPU number to execute work on
1481 * @wq: workqueue to use
1482 * @dwork: work to queue
1483 * @delay: number of jiffies to wait before queueing
1485 * Returns %false if @work was already on a queue, %true otherwise. If
1486 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1489 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1490 struct delayed_work
*dwork
, unsigned long delay
)
1492 struct work_struct
*work
= &dwork
->work
;
1494 unsigned long flags
;
1496 /* read the comment in __queue_work() */
1497 local_irq_save(flags
);
1499 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1500 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1504 local_irq_restore(flags
);
1507 EXPORT_SYMBOL(queue_delayed_work_on
);
1510 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1511 * @cpu: CPU number to execute work on
1512 * @wq: workqueue to use
1513 * @dwork: work to queue
1514 * @delay: number of jiffies to wait before queueing
1516 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1517 * modify @dwork's timer so that it expires after @delay. If @delay is
1518 * zero, @work is guaranteed to be scheduled immediately regardless of its
1521 * Returns %false if @dwork was idle and queued, %true if @dwork was
1522 * pending and its timer was modified.
1524 * This function is safe to call from any context including IRQ handler.
1525 * See try_to_grab_pending() for details.
1527 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1528 struct delayed_work
*dwork
, unsigned long delay
)
1530 unsigned long flags
;
1534 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1535 } while (unlikely(ret
== -EAGAIN
));
1537 if (likely(ret
>= 0)) {
1538 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1539 local_irq_restore(flags
);
1542 /* -ENOENT from try_to_grab_pending() becomes %true */
1545 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1548 * worker_enter_idle - enter idle state
1549 * @worker: worker which is entering idle state
1551 * @worker is entering idle state. Update stats and idle timer if
1555 * spin_lock_irq(pool->lock).
1557 static void worker_enter_idle(struct worker
*worker
)
1559 struct worker_pool
*pool
= worker
->pool
;
1561 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1562 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1563 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1566 /* can't use worker_set_flags(), also called from start_worker() */
1567 worker
->flags
|= WORKER_IDLE
;
1569 worker
->last_active
= jiffies
;
1571 /* idle_list is LIFO */
1572 list_add(&worker
->entry
, &pool
->idle_list
);
1574 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1575 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1578 * Sanity check nr_running. Because wq_unbind_fn() releases
1579 * pool->lock between setting %WORKER_UNBOUND and zapping
1580 * nr_running, the warning may trigger spuriously. Check iff
1581 * unbind is not in progress.
1583 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1584 pool
->nr_workers
== pool
->nr_idle
&&
1585 atomic_read(&pool
->nr_running
));
1589 * worker_leave_idle - leave idle state
1590 * @worker: worker which is leaving idle state
1592 * @worker is leaving idle state. Update stats.
1595 * spin_lock_irq(pool->lock).
1597 static void worker_leave_idle(struct worker
*worker
)
1599 struct worker_pool
*pool
= worker
->pool
;
1601 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1603 worker_clr_flags(worker
, WORKER_IDLE
);
1605 list_del_init(&worker
->entry
);
1609 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1610 * @pool: target worker_pool
1612 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1614 * Works which are scheduled while the cpu is online must at least be
1615 * scheduled to a worker which is bound to the cpu so that if they are
1616 * flushed from cpu callbacks while cpu is going down, they are
1617 * guaranteed to execute on the cpu.
1619 * This function is to be used by unbound workers and rescuers to bind
1620 * themselves to the target cpu and may race with cpu going down or
1621 * coming online. kthread_bind() can't be used because it may put the
1622 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1623 * verbatim as it's best effort and blocking and pool may be
1624 * [dis]associated in the meantime.
1626 * This function tries set_cpus_allowed() and locks pool and verifies the
1627 * binding against %POOL_DISASSOCIATED which is set during
1628 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1629 * enters idle state or fetches works without dropping lock, it can
1630 * guarantee the scheduling requirement described in the first paragraph.
1633 * Might sleep. Called without any lock but returns with pool->lock
1637 * %true if the associated pool is online (@worker is successfully
1638 * bound), %false if offline.
1640 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1641 __acquires(&pool
->lock
)
1645 * The following call may fail, succeed or succeed
1646 * without actually migrating the task to the cpu if
1647 * it races with cpu hotunplug operation. Verify
1648 * against POOL_DISASSOCIATED.
1650 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1651 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1653 spin_lock_irq(&pool
->lock
);
1654 if (pool
->flags
& POOL_DISASSOCIATED
)
1656 if (task_cpu(current
) == pool
->cpu
&&
1657 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1659 spin_unlock_irq(&pool
->lock
);
1662 * We've raced with CPU hot[un]plug. Give it a breather
1663 * and retry migration. cond_resched() is required here;
1664 * otherwise, we might deadlock against cpu_stop trying to
1665 * bring down the CPU on non-preemptive kernel.
1672 static struct worker
*alloc_worker(void)
1674 struct worker
*worker
;
1676 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1678 INIT_LIST_HEAD(&worker
->entry
);
1679 INIT_LIST_HEAD(&worker
->scheduled
);
1680 /* on creation a worker is in !idle && prep state */
1681 worker
->flags
= WORKER_PREP
;
1687 * create_worker - create a new workqueue worker
1688 * @pool: pool the new worker will belong to
1690 * Create a new worker which is bound to @pool. The returned worker
1691 * can be started by calling start_worker() or destroyed using
1695 * Might sleep. Does GFP_KERNEL allocations.
1698 * Pointer to the newly created worker.
1700 static struct worker
*create_worker(struct worker_pool
*pool
)
1702 struct worker
*worker
= NULL
;
1706 lockdep_assert_held(&pool
->manager_mutex
);
1709 * ID is needed to determine kthread name. Allocate ID first
1710 * without installing the pointer.
1712 idr_preload(GFP_KERNEL
);
1713 spin_lock_irq(&pool
->lock
);
1715 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1717 spin_unlock_irq(&pool
->lock
);
1722 worker
= alloc_worker();
1726 worker
->pool
= pool
;
1730 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1731 pool
->attrs
->nice
< 0 ? "H" : "");
1733 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1735 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1736 "kworker/%s", id_buf
);
1737 if (IS_ERR(worker
->task
))
1741 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1742 * online CPUs. It'll be re-applied when any of the CPUs come up.
1744 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1745 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1747 /* prevent userland from meddling with cpumask of workqueue workers */
1748 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1751 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1752 * remains stable across this function. See the comments above the
1753 * flag definition for details.
1755 if (pool
->flags
& POOL_DISASSOCIATED
)
1756 worker
->flags
|= WORKER_UNBOUND
;
1758 /* successful, commit the pointer to idr */
1759 spin_lock_irq(&pool
->lock
);
1760 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1761 spin_unlock_irq(&pool
->lock
);
1767 spin_lock_irq(&pool
->lock
);
1768 idr_remove(&pool
->worker_idr
, id
);
1769 spin_unlock_irq(&pool
->lock
);
1776 * start_worker - start a newly created worker
1777 * @worker: worker to start
1779 * Make the pool aware of @worker and start it.
1782 * spin_lock_irq(pool->lock).
1784 static void start_worker(struct worker
*worker
)
1786 worker
->flags
|= WORKER_STARTED
;
1787 worker
->pool
->nr_workers
++;
1788 worker_enter_idle(worker
);
1789 wake_up_process(worker
->task
);
1793 * create_and_start_worker - create and start a worker for a pool
1794 * @pool: the target pool
1796 * Grab the managership of @pool and create and start a new worker for it.
1798 static int create_and_start_worker(struct worker_pool
*pool
)
1800 struct worker
*worker
;
1802 mutex_lock(&pool
->manager_mutex
);
1804 worker
= create_worker(pool
);
1806 spin_lock_irq(&pool
->lock
);
1807 start_worker(worker
);
1808 spin_unlock_irq(&pool
->lock
);
1811 mutex_unlock(&pool
->manager_mutex
);
1813 return worker
? 0 : -ENOMEM
;
1817 * destroy_worker - destroy a workqueue worker
1818 * @worker: worker to be destroyed
1820 * Destroy @worker and adjust @pool stats accordingly.
1823 * spin_lock_irq(pool->lock) which is released and regrabbed.
1825 static void destroy_worker(struct worker
*worker
)
1827 struct worker_pool
*pool
= worker
->pool
;
1829 lockdep_assert_held(&pool
->manager_mutex
);
1830 lockdep_assert_held(&pool
->lock
);
1832 /* sanity check frenzy */
1833 if (WARN_ON(worker
->current_work
) ||
1834 WARN_ON(!list_empty(&worker
->scheduled
)))
1837 if (worker
->flags
& WORKER_STARTED
)
1839 if (worker
->flags
& WORKER_IDLE
)
1843 * Once WORKER_DIE is set, the kworker may destroy itself at any
1844 * point. Pin to ensure the task stays until we're done with it.
1846 get_task_struct(worker
->task
);
1848 list_del_init(&worker
->entry
);
1849 worker
->flags
|= WORKER_DIE
;
1851 idr_remove(&pool
->worker_idr
, worker
->id
);
1853 spin_unlock_irq(&pool
->lock
);
1855 kthread_stop(worker
->task
);
1856 put_task_struct(worker
->task
);
1859 spin_lock_irq(&pool
->lock
);
1862 static void idle_worker_timeout(unsigned long __pool
)
1864 struct worker_pool
*pool
= (void *)__pool
;
1866 spin_lock_irq(&pool
->lock
);
1868 if (too_many_workers(pool
)) {
1869 struct worker
*worker
;
1870 unsigned long expires
;
1872 /* idle_list is kept in LIFO order, check the last one */
1873 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1874 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1876 if (time_before(jiffies
, expires
))
1877 mod_timer(&pool
->idle_timer
, expires
);
1879 /* it's been idle for too long, wake up manager */
1880 pool
->flags
|= POOL_MANAGE_WORKERS
;
1881 wake_up_worker(pool
);
1885 spin_unlock_irq(&pool
->lock
);
1888 static void send_mayday(struct work_struct
*work
)
1890 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1891 struct workqueue_struct
*wq
= pwq
->wq
;
1893 lockdep_assert_held(&wq_mayday_lock
);
1898 /* mayday mayday mayday */
1899 if (list_empty(&pwq
->mayday_node
)) {
1901 * If @pwq is for an unbound wq, its base ref may be put at
1902 * any time due to an attribute change. Pin @pwq until the
1903 * rescuer is done with it.
1906 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1907 wake_up_process(wq
->rescuer
->task
);
1911 static void pool_mayday_timeout(unsigned long __pool
)
1913 struct worker_pool
*pool
= (void *)__pool
;
1914 struct work_struct
*work
;
1916 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1917 spin_lock(&pool
->lock
);
1919 if (need_to_create_worker(pool
)) {
1921 * We've been trying to create a new worker but
1922 * haven't been successful. We might be hitting an
1923 * allocation deadlock. Send distress signals to
1926 list_for_each_entry(work
, &pool
->worklist
, entry
)
1930 spin_unlock(&pool
->lock
);
1931 spin_unlock_irq(&wq_mayday_lock
);
1933 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1937 * maybe_create_worker - create a new worker if necessary
1938 * @pool: pool to create a new worker for
1940 * Create a new worker for @pool if necessary. @pool is guaranteed to
1941 * have at least one idle worker on return from this function. If
1942 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1943 * sent to all rescuers with works scheduled on @pool to resolve
1944 * possible allocation deadlock.
1946 * On return, need_to_create_worker() is guaranteed to be %false and
1947 * may_start_working() %true.
1950 * spin_lock_irq(pool->lock) which may be released and regrabbed
1951 * multiple times. Does GFP_KERNEL allocations. Called only from
1954 static void maybe_create_worker(struct worker_pool
*pool
)
1955 __releases(&pool
->lock
)
1956 __acquires(&pool
->lock
)
1958 if (!need_to_create_worker(pool
))
1961 spin_unlock_irq(&pool
->lock
);
1963 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1964 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1967 struct worker
*worker
;
1969 worker
= create_worker(pool
);
1971 del_timer_sync(&pool
->mayday_timer
);
1972 spin_lock_irq(&pool
->lock
);
1973 start_worker(worker
);
1974 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1979 if (!need_to_create_worker(pool
))
1982 __set_current_state(TASK_INTERRUPTIBLE
);
1983 schedule_timeout(CREATE_COOLDOWN
);
1985 if (!need_to_create_worker(pool
))
1989 del_timer_sync(&pool
->mayday_timer
);
1990 spin_lock_irq(&pool
->lock
);
1991 if (need_to_create_worker(pool
))
1997 * maybe_destroy_worker - destroy workers which have been idle for a while
1998 * @pool: pool to destroy workers for
2000 * Destroy @pool workers which have been idle for longer than
2001 * IDLE_WORKER_TIMEOUT.
2004 * spin_lock_irq(pool->lock) which may be released and regrabbed
2005 * multiple times. Called only from manager.
2007 static void maybe_destroy_workers(struct worker_pool
*pool
)
2009 while (too_many_workers(pool
)) {
2010 struct worker
*worker
;
2011 unsigned long expires
;
2013 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2014 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2016 if (time_before(jiffies
, expires
)) {
2017 mod_timer(&pool
->idle_timer
, expires
);
2021 destroy_worker(worker
);
2026 * manage_workers - manage worker pool
2029 * Assume the manager role and manage the worker pool @worker belongs
2030 * to. At any given time, there can be only zero or one manager per
2031 * pool. The exclusion is handled automatically by this function.
2033 * The caller can safely start processing works on false return. On
2034 * true return, it's guaranteed that need_to_create_worker() is false
2035 * and may_start_working() is true.
2038 * spin_lock_irq(pool->lock) which may be released and regrabbed
2039 * multiple times. Does GFP_KERNEL allocations.
2042 * %false if the pool doesn't need management and the caller can safely
2043 * start processing works, %true if management function was performed and
2044 * the conditions that the caller verified before calling the function may
2045 * no longer be true.
2047 static bool manage_workers(struct worker
*worker
)
2049 struct worker_pool
*pool
= worker
->pool
;
2052 * Managership is governed by two mutexes - manager_arb and
2053 * manager_mutex. manager_arb handles arbitration of manager role.
2054 * Anyone who successfully grabs manager_arb wins the arbitration
2055 * and becomes the manager. mutex_trylock() on pool->manager_arb
2056 * failure while holding pool->lock reliably indicates that someone
2057 * else is managing the pool and the worker which failed trylock
2058 * can proceed to executing work items. This means that anyone
2059 * grabbing manager_arb is responsible for actually performing
2060 * manager duties. If manager_arb is grabbed and released without
2061 * actual management, the pool may stall indefinitely.
2063 * manager_mutex is used for exclusion of actual management
2064 * operations. The holder of manager_mutex can be sure that none
2065 * of management operations, including creation and destruction of
2066 * workers, won't take place until the mutex is released. Because
2067 * manager_mutex doesn't interfere with manager role arbitration,
2068 * it is guaranteed that the pool's management, while may be
2069 * delayed, won't be disturbed by someone else grabbing
2072 if (!mutex_trylock(&pool
->manager_arb
))
2076 * With manager arbitration won, manager_mutex would be free in
2077 * most cases. trylock first without dropping @pool->lock.
2079 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2080 spin_unlock_irq(&pool
->lock
);
2081 mutex_lock(&pool
->manager_mutex
);
2082 spin_lock_irq(&pool
->lock
);
2085 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2088 * Destroy and then create so that may_start_working() is true
2091 maybe_destroy_workers(pool
);
2092 maybe_create_worker(pool
);
2094 mutex_unlock(&pool
->manager_mutex
);
2095 mutex_unlock(&pool
->manager_arb
);
2100 * process_one_work - process single work
2102 * @work: work to process
2104 * Process @work. This function contains all the logics necessary to
2105 * process a single work including synchronization against and
2106 * interaction with other workers on the same cpu, queueing and
2107 * flushing. As long as context requirement is met, any worker can
2108 * call this function to process a work.
2111 * spin_lock_irq(pool->lock) which is released and regrabbed.
2113 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2114 __releases(&pool
->lock
)
2115 __acquires(&pool
->lock
)
2117 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2118 struct worker_pool
*pool
= worker
->pool
;
2119 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2121 struct worker
*collision
;
2122 unsigned long long exec_start
;
2125 #ifdef CONFIG_LOCKDEP
2127 * It is permissible to free the struct work_struct from
2128 * inside the function that is called from it, this we need to
2129 * take into account for lockdep too. To avoid bogus "held
2130 * lock freed" warnings as well as problems when looking into
2131 * work->lockdep_map, make a copy and use that here.
2133 struct lockdep_map lockdep_map
;
2135 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2138 * Ensure we're on the correct CPU. DISASSOCIATED test is
2139 * necessary to avoid spurious warnings from rescuers servicing the
2140 * unbound or a disassociated pool.
2142 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2143 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2144 raw_smp_processor_id() != pool
->cpu
);
2147 * A single work shouldn't be executed concurrently by
2148 * multiple workers on a single cpu. Check whether anyone is
2149 * already processing the work. If so, defer the work to the
2150 * currently executing one.
2152 collision
= find_worker_executing_work(pool
, work
);
2153 if (unlikely(collision
)) {
2154 move_linked_works(work
, &collision
->scheduled
, NULL
);
2158 /* claim and dequeue */
2159 debug_work_deactivate(work
);
2160 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2161 worker
->current_work
= work
;
2162 worker
->current_func
= work
->func
;
2163 worker
->current_pwq
= pwq
;
2164 work_color
= get_work_color(work
);
2166 list_del_init(&work
->entry
);
2169 * CPU intensive works don't participate in concurrency
2170 * management. They're the scheduler's responsibility.
2172 if (unlikely(cpu_intensive
))
2173 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2176 * Unbound pool isn't concurrency managed and work items should be
2177 * executed ASAP. Wake up another worker if necessary.
2179 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2180 wake_up_worker(pool
);
2183 * Record the last pool and clear PENDING which should be the last
2184 * update to @work. Also, do this inside @pool->lock so that
2185 * PENDING and queued state changes happen together while IRQ is
2188 set_work_pool_and_clear_pending(work
, pool
->id
);
2190 spin_unlock_irq(&pool
->lock
);
2192 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2193 lock_map_acquire(&lockdep_map
);
2195 exec_start
= sched_clock();
2196 sprintf(func
, "%pf", work
->func
);
2198 trace_workqueue_execute_start(work
);
2199 #ifdef CONFIG_MTK_WQ_DEBUG
2200 mttrace_workqueue_execute_work(work
);
2201 #endif //CONFIG_MTK_WQ_DEBUG
2203 worker
->current_func(work
);
2206 * While we must be careful to not use "work" after this, the trace
2207 * point will only record its address.
2209 trace_workqueue_execute_end(work
);
2210 #ifdef CONFIG_MTK_WQ_DEBUG
2211 mttrace_workqueue_execute_end(work
);
2212 #endif //CONFIG_MTK_WQ_DEBUG
2214 if ((sched_clock() - exec_start
)> 1000000000) // dump log if execute more than 1 sec
2215 pr_warning("WQ warning! work (%s, %p) execute more than 1 sec, time: %llu ns\n", func
, work
, sched_clock() - exec_start
);
2217 lock_map_release(&lockdep_map
);
2218 lock_map_release(&pwq
->wq
->lockdep_map
);
2220 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2221 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2222 " last function: %pf\n",
2223 current
->comm
, preempt_count(), task_pid_nr(current
),
2224 worker
->current_func
);
2225 debug_show_held_locks(current
);
2230 * The following prevents a kworker from hogging CPU on !PREEMPT
2231 * kernels, where a requeueing work item waiting for something to
2232 * happen could deadlock with stop_machine as such work item could
2233 * indefinitely requeue itself while all other CPUs are trapped in
2238 spin_lock_irq(&pool
->lock
);
2240 /* clear cpu intensive status */
2241 if (unlikely(cpu_intensive
))
2242 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2244 /* we're done with it, release */
2245 hash_del(&worker
->hentry
);
2246 worker
->current_work
= NULL
;
2247 worker
->current_func
= NULL
;
2248 worker
->current_pwq
= NULL
;
2249 worker
->desc_valid
= false;
2250 pwq_dec_nr_in_flight(pwq
, work_color
);
2254 * process_scheduled_works - process scheduled works
2257 * Process all scheduled works. Please note that the scheduled list
2258 * may change while processing a work, so this function repeatedly
2259 * fetches a work from the top and executes it.
2262 * spin_lock_irq(pool->lock) which may be released and regrabbed
2265 static void process_scheduled_works(struct worker
*worker
)
2267 while (!list_empty(&worker
->scheduled
)) {
2268 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2269 struct work_struct
, entry
);
2270 process_one_work(worker
, work
);
2275 * worker_thread - the worker thread function
2278 * The worker thread function. All workers belong to a worker_pool -
2279 * either a per-cpu one or dynamic unbound one. These workers process all
2280 * work items regardless of their specific target workqueue. The only
2281 * exception is work items which belong to workqueues with a rescuer which
2282 * will be explained in rescuer_thread().
2284 static int worker_thread(void *__worker
)
2286 struct worker
*worker
= __worker
;
2287 struct worker_pool
*pool
= worker
->pool
;
2289 /* tell the scheduler that this is a workqueue worker */
2290 worker
->task
->flags
|= PF_WQ_WORKER
;
2292 spin_lock_irq(&pool
->lock
);
2294 /* am I supposed to die? */
2295 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2296 spin_unlock_irq(&pool
->lock
);
2297 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2298 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2302 worker_leave_idle(worker
);
2304 /* no more worker necessary? */
2305 if (!need_more_worker(pool
))
2308 /* do we need to manage? */
2309 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2313 * ->scheduled list can only be filled while a worker is
2314 * preparing to process a work or actually processing it.
2315 * Make sure nobody diddled with it while I was sleeping.
2317 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2320 * Finish PREP stage. We're guaranteed to have at least one idle
2321 * worker or that someone else has already assumed the manager
2322 * role. This is where @worker starts participating in concurrency
2323 * management if applicable and concurrency management is restored
2324 * after being rebound. See rebind_workers() for details.
2326 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2329 struct work_struct
*work
=
2330 list_first_entry(&pool
->worklist
,
2331 struct work_struct
, entry
);
2333 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2334 /* optimization path, not strictly necessary */
2335 process_one_work(worker
, work
);
2336 if (unlikely(!list_empty(&worker
->scheduled
)))
2337 process_scheduled_works(worker
);
2339 move_linked_works(work
, &worker
->scheduled
, NULL
);
2340 process_scheduled_works(worker
);
2342 } while (keep_working(pool
));
2344 worker_set_flags(worker
, WORKER_PREP
, false);
2346 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2350 * pool->lock is held and there's no work to process and no need to
2351 * manage, sleep. Workers are woken up only while holding
2352 * pool->lock or from local cpu, so setting the current state
2353 * before releasing pool->lock is enough to prevent losing any
2356 worker_enter_idle(worker
);
2357 __set_current_state(TASK_INTERRUPTIBLE
);
2358 spin_unlock_irq(&pool
->lock
);
2364 * rescuer_thread - the rescuer thread function
2367 * Workqueue rescuer thread function. There's one rescuer for each
2368 * workqueue which has WQ_MEM_RECLAIM set.
2370 * Regular work processing on a pool may block trying to create a new
2371 * worker which uses GFP_KERNEL allocation which has slight chance of
2372 * developing into deadlock if some works currently on the same queue
2373 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2374 * the problem rescuer solves.
2376 * When such condition is possible, the pool summons rescuers of all
2377 * workqueues which have works queued on the pool and let them process
2378 * those works so that forward progress can be guaranteed.
2380 * This should happen rarely.
2382 static int rescuer_thread(void *__rescuer
)
2384 struct worker
*rescuer
= __rescuer
;
2385 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2386 struct list_head
*scheduled
= &rescuer
->scheduled
;
2389 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2392 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2393 * doesn't participate in concurrency management.
2395 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2397 set_current_state(TASK_INTERRUPTIBLE
);
2400 * By the time the rescuer is requested to stop, the workqueue
2401 * shouldn't have any work pending, but @wq->maydays may still have
2402 * pwq(s) queued. This can happen by non-rescuer workers consuming
2403 * all the work items before the rescuer got to them. Go through
2404 * @wq->maydays processing before acting on should_stop so that the
2405 * list is always empty on exit.
2407 should_stop
= kthread_should_stop();
2409 /* see whether any pwq is asking for help */
2410 spin_lock_irq(&wq_mayday_lock
);
2412 while (!list_empty(&wq
->maydays
)) {
2413 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2414 struct pool_workqueue
, mayday_node
);
2415 struct worker_pool
*pool
= pwq
->pool
;
2416 struct work_struct
*work
, *n
;
2418 __set_current_state(TASK_RUNNING
);
2419 list_del_init(&pwq
->mayday_node
);
2421 spin_unlock_irq(&wq_mayday_lock
);
2423 /* migrate to the target cpu if possible */
2424 worker_maybe_bind_and_lock(pool
);
2425 rescuer
->pool
= pool
;
2428 * Slurp in all works issued via this workqueue and
2431 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2432 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2433 if (get_work_pwq(work
) == pwq
)
2434 move_linked_works(work
, scheduled
, &n
);
2436 process_scheduled_works(rescuer
);
2439 * Put the reference grabbed by send_mayday(). @pool won't
2440 * go away while we're holding its lock.
2445 * Leave this pool. If keep_working() is %true, notify a
2446 * regular worker; otherwise, we end up with 0 concurrency
2447 * and stalling the execution.
2449 if (keep_working(pool
))
2450 wake_up_worker(pool
);
2452 rescuer
->pool
= NULL
;
2453 spin_unlock(&pool
->lock
);
2454 spin_lock(&wq_mayday_lock
);
2457 spin_unlock_irq(&wq_mayday_lock
);
2460 __set_current_state(TASK_RUNNING
);
2461 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2465 /* rescuers should never participate in concurrency management */
2466 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2472 struct work_struct work
;
2473 struct completion done
;
2476 static void wq_barrier_func(struct work_struct
*work
)
2478 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2479 complete(&barr
->done
);
2483 * insert_wq_barrier - insert a barrier work
2484 * @pwq: pwq to insert barrier into
2485 * @barr: wq_barrier to insert
2486 * @target: target work to attach @barr to
2487 * @worker: worker currently executing @target, NULL if @target is not executing
2489 * @barr is linked to @target such that @barr is completed only after
2490 * @target finishes execution. Please note that the ordering
2491 * guarantee is observed only with respect to @target and on the local
2494 * Currently, a queued barrier can't be canceled. This is because
2495 * try_to_grab_pending() can't determine whether the work to be
2496 * grabbed is at the head of the queue and thus can't clear LINKED
2497 * flag of the previous work while there must be a valid next work
2498 * after a work with LINKED flag set.
2500 * Note that when @worker is non-NULL, @target may be modified
2501 * underneath us, so we can't reliably determine pwq from @target.
2504 * spin_lock_irq(pool->lock).
2506 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2507 struct wq_barrier
*barr
,
2508 struct work_struct
*target
, struct worker
*worker
)
2510 struct list_head
*head
;
2511 unsigned int linked
= 0;
2514 * debugobject calls are safe here even with pool->lock locked
2515 * as we know for sure that this will not trigger any of the
2516 * checks and call back into the fixup functions where we
2519 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2520 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2521 init_completion(&barr
->done
);
2524 * If @target is currently being executed, schedule the
2525 * barrier to the worker; otherwise, put it after @target.
2528 head
= worker
->scheduled
.next
;
2530 unsigned long *bits
= work_data_bits(target
);
2532 head
= target
->entry
.next
;
2533 /* there can already be other linked works, inherit and set */
2534 linked
= *bits
& WORK_STRUCT_LINKED
;
2535 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2538 debug_work_activate(&barr
->work
);
2539 insert_work(pwq
, &barr
->work
, head
,
2540 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2544 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2545 * @wq: workqueue being flushed
2546 * @flush_color: new flush color, < 0 for no-op
2547 * @work_color: new work color, < 0 for no-op
2549 * Prepare pwqs for workqueue flushing.
2551 * If @flush_color is non-negative, flush_color on all pwqs should be
2552 * -1. If no pwq has in-flight commands at the specified color, all
2553 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2554 * has in flight commands, its pwq->flush_color is set to
2555 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2556 * wakeup logic is armed and %true is returned.
2558 * The caller should have initialized @wq->first_flusher prior to
2559 * calling this function with non-negative @flush_color. If
2560 * @flush_color is negative, no flush color update is done and %false
2563 * If @work_color is non-negative, all pwqs should have the same
2564 * work_color which is previous to @work_color and all will be
2565 * advanced to @work_color.
2568 * mutex_lock(wq->mutex).
2571 * %true if @flush_color >= 0 and there's something to flush. %false
2574 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2575 int flush_color
, int work_color
)
2578 struct pool_workqueue
*pwq
;
2580 if (flush_color
>= 0) {
2581 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2582 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2585 for_each_pwq(pwq
, wq
) {
2586 struct worker_pool
*pool
= pwq
->pool
;
2588 spin_lock_irq(&pool
->lock
);
2590 if (flush_color
>= 0) {
2591 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2593 if (pwq
->nr_in_flight
[flush_color
]) {
2594 pwq
->flush_color
= flush_color
;
2595 atomic_inc(&wq
->nr_pwqs_to_flush
);
2600 if (work_color
>= 0) {
2601 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2602 pwq
->work_color
= work_color
;
2605 spin_unlock_irq(&pool
->lock
);
2608 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2609 complete(&wq
->first_flusher
->done
);
2615 * flush_workqueue - ensure that any scheduled work has run to completion.
2616 * @wq: workqueue to flush
2618 * This function sleeps until all work items which were queued on entry
2619 * have finished execution, but it is not livelocked by new incoming ones.
2621 void flush_workqueue(struct workqueue_struct
*wq
)
2623 struct wq_flusher this_flusher
= {
2624 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2626 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2630 lock_map_acquire(&wq
->lockdep_map
);
2631 lock_map_release(&wq
->lockdep_map
);
2633 mutex_lock(&wq
->mutex
);
2636 * Start-to-wait phase
2638 next_color
= work_next_color(wq
->work_color
);
2640 if (next_color
!= wq
->flush_color
) {
2642 * Color space is not full. The current work_color
2643 * becomes our flush_color and work_color is advanced
2646 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2647 this_flusher
.flush_color
= wq
->work_color
;
2648 wq
->work_color
= next_color
;
2650 if (!wq
->first_flusher
) {
2651 /* no flush in progress, become the first flusher */
2652 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2654 wq
->first_flusher
= &this_flusher
;
2656 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2658 /* nothing to flush, done */
2659 wq
->flush_color
= next_color
;
2660 wq
->first_flusher
= NULL
;
2665 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2666 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2667 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2671 * Oops, color space is full, wait on overflow queue.
2672 * The next flush completion will assign us
2673 * flush_color and transfer to flusher_queue.
2675 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2678 mutex_unlock(&wq
->mutex
);
2680 wait_for_completion(&this_flusher
.done
);
2683 * Wake-up-and-cascade phase
2685 * First flushers are responsible for cascading flushes and
2686 * handling overflow. Non-first flushers can simply return.
2688 if (wq
->first_flusher
!= &this_flusher
)
2691 mutex_lock(&wq
->mutex
);
2693 /* we might have raced, check again with mutex held */
2694 if (wq
->first_flusher
!= &this_flusher
)
2697 wq
->first_flusher
= NULL
;
2699 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2700 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2703 struct wq_flusher
*next
, *tmp
;
2705 /* complete all the flushers sharing the current flush color */
2706 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2707 if (next
->flush_color
!= wq
->flush_color
)
2709 list_del_init(&next
->list
);
2710 complete(&next
->done
);
2713 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2714 wq
->flush_color
!= work_next_color(wq
->work_color
));
2716 /* this flush_color is finished, advance by one */
2717 wq
->flush_color
= work_next_color(wq
->flush_color
);
2719 /* one color has been freed, handle overflow queue */
2720 if (!list_empty(&wq
->flusher_overflow
)) {
2722 * Assign the same color to all overflowed
2723 * flushers, advance work_color and append to
2724 * flusher_queue. This is the start-to-wait
2725 * phase for these overflowed flushers.
2727 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2728 tmp
->flush_color
= wq
->work_color
;
2730 wq
->work_color
= work_next_color(wq
->work_color
);
2732 list_splice_tail_init(&wq
->flusher_overflow
,
2733 &wq
->flusher_queue
);
2734 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2737 if (list_empty(&wq
->flusher_queue
)) {
2738 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2743 * Need to flush more colors. Make the next flusher
2744 * the new first flusher and arm pwqs.
2746 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2747 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2749 list_del_init(&next
->list
);
2750 wq
->first_flusher
= next
;
2752 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2756 * Meh... this color is already done, clear first
2757 * flusher and repeat cascading.
2759 wq
->first_flusher
= NULL
;
2763 mutex_unlock(&wq
->mutex
);
2765 EXPORT_SYMBOL_GPL(flush_workqueue
);
2768 * drain_workqueue - drain a workqueue
2769 * @wq: workqueue to drain
2771 * Wait until the workqueue becomes empty. While draining is in progress,
2772 * only chain queueing is allowed. IOW, only currently pending or running
2773 * work items on @wq can queue further work items on it. @wq is flushed
2774 * repeatedly until it becomes empty. The number of flushing is detemined
2775 * by the depth of chaining and should be relatively short. Whine if it
2778 void drain_workqueue(struct workqueue_struct
*wq
)
2780 unsigned int flush_cnt
= 0;
2781 struct pool_workqueue
*pwq
;
2784 * __queue_work() needs to test whether there are drainers, is much
2785 * hotter than drain_workqueue() and already looks at @wq->flags.
2786 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2788 mutex_lock(&wq
->mutex
);
2789 if (!wq
->nr_drainers
++)
2790 wq
->flags
|= __WQ_DRAINING
;
2791 mutex_unlock(&wq
->mutex
);
2793 flush_workqueue(wq
);
2795 mutex_lock(&wq
->mutex
);
2797 for_each_pwq(pwq
, wq
) {
2800 spin_lock_irq(&pwq
->pool
->lock
);
2801 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2802 spin_unlock_irq(&pwq
->pool
->lock
);
2807 if (++flush_cnt
== 10 ||
2808 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2809 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2810 wq
->name
, flush_cnt
);
2812 mutex_unlock(&wq
->mutex
);
2816 if (!--wq
->nr_drainers
)
2817 wq
->flags
&= ~__WQ_DRAINING
;
2818 mutex_unlock(&wq
->mutex
);
2820 EXPORT_SYMBOL_GPL(drain_workqueue
);
2822 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2824 struct worker
*worker
= NULL
;
2825 struct worker_pool
*pool
;
2826 struct pool_workqueue
*pwq
;
2830 local_irq_disable();
2831 pool
= get_work_pool(work
);
2837 spin_lock(&pool
->lock
);
2838 /* see the comment in try_to_grab_pending() with the same code */
2839 pwq
= get_work_pwq(work
);
2841 if (unlikely(pwq
->pool
!= pool
))
2844 worker
= find_worker_executing_work(pool
, work
);
2847 pwq
= worker
->current_pwq
;
2850 insert_wq_barrier(pwq
, barr
, work
, worker
);
2851 spin_unlock_irq(&pool
->lock
);
2854 * If @max_active is 1 or rescuer is in use, flushing another work
2855 * item on the same workqueue may lead to deadlock. Make sure the
2856 * flusher is not running on the same workqueue by verifying write
2859 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2860 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2862 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2863 lock_map_release(&pwq
->wq
->lockdep_map
);
2867 spin_unlock_irq(&pool
->lock
);
2872 * flush_work - wait for a work to finish executing the last queueing instance
2873 * @work: the work to flush
2875 * Wait until @work has finished execution. @work is guaranteed to be idle
2876 * on return if it hasn't been requeued since flush started.
2879 * %true if flush_work() waited for the work to finish execution,
2880 * %false if it was already idle.
2882 bool flush_work(struct work_struct
*work
)
2884 struct wq_barrier barr
;
2886 lock_map_acquire(&work
->lockdep_map
);
2887 lock_map_release(&work
->lockdep_map
);
2889 if (start_flush_work(work
, &barr
)) {
2890 wait_for_completion(&barr
.done
);
2891 destroy_work_on_stack(&barr
.work
);
2897 EXPORT_SYMBOL_GPL(flush_work
);
2899 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2901 unsigned long flags
;
2905 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2907 * If someone else is canceling, wait for the same event it
2908 * would be waiting for before retrying.
2910 if (unlikely(ret
== -ENOENT
))
2912 } while (unlikely(ret
< 0));
2914 /* tell other tasks trying to grab @work to back off */
2915 mark_work_canceling(work
);
2916 local_irq_restore(flags
);
2919 clear_work_data(work
);
2924 * cancel_work_sync - cancel a work and wait for it to finish
2925 * @work: the work to cancel
2927 * Cancel @work and wait for its execution to finish. This function
2928 * can be used even if the work re-queues itself or migrates to
2929 * another workqueue. On return from this function, @work is
2930 * guaranteed to be not pending or executing on any CPU.
2932 * cancel_work_sync(&delayed_work->work) must not be used for
2933 * delayed_work's. Use cancel_delayed_work_sync() instead.
2935 * The caller must ensure that the workqueue on which @work was last
2936 * queued can't be destroyed before this function returns.
2939 * %true if @work was pending, %false otherwise.
2941 bool cancel_work_sync(struct work_struct
*work
)
2943 return __cancel_work_timer(work
, false);
2945 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2948 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2949 * @dwork: the delayed work to flush
2951 * Delayed timer is cancelled and the pending work is queued for
2952 * immediate execution. Like flush_work(), this function only
2953 * considers the last queueing instance of @dwork.
2956 * %true if flush_work() waited for the work to finish execution,
2957 * %false if it was already idle.
2959 bool flush_delayed_work(struct delayed_work
*dwork
)
2961 local_irq_disable();
2962 if (del_timer_sync(&dwork
->timer
))
2963 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2965 return flush_work(&dwork
->work
);
2967 EXPORT_SYMBOL(flush_delayed_work
);
2970 * cancel_delayed_work - cancel a delayed work
2971 * @dwork: delayed_work to cancel
2973 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2974 * and canceled; %false if wasn't pending. Note that the work callback
2975 * function may still be running on return, unless it returns %true and the
2976 * work doesn't re-arm itself. Explicitly flush or use
2977 * cancel_delayed_work_sync() to wait on it.
2979 * This function is safe to call from any context including IRQ handler.
2981 bool cancel_delayed_work(struct delayed_work
*dwork
)
2983 unsigned long flags
;
2987 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2988 } while (unlikely(ret
== -EAGAIN
));
2990 if (unlikely(ret
< 0))
2993 set_work_pool_and_clear_pending(&dwork
->work
,
2994 get_work_pool_id(&dwork
->work
));
2995 local_irq_restore(flags
);
2998 EXPORT_SYMBOL(cancel_delayed_work
);
3001 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3002 * @dwork: the delayed work cancel
3004 * This is cancel_work_sync() for delayed works.
3007 * %true if @dwork was pending, %false otherwise.
3009 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3011 return __cancel_work_timer(&dwork
->work
, true);
3013 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3016 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3017 * @func: the function to call
3019 * schedule_on_each_cpu() executes @func on each online CPU using the
3020 * system workqueue and blocks until all CPUs have completed.
3021 * schedule_on_each_cpu() is very slow.
3024 * 0 on success, -errno on failure.
3026 int schedule_on_each_cpu(work_func_t func
)
3029 struct work_struct __percpu
*works
;
3031 works
= alloc_percpu(struct work_struct
);
3037 for_each_online_cpu(cpu
) {
3038 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3040 INIT_WORK(work
, func
);
3041 schedule_work_on(cpu
, work
);
3044 for_each_online_cpu(cpu
)
3045 flush_work(per_cpu_ptr(works
, cpu
));
3053 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3055 * Forces execution of the kernel-global workqueue and blocks until its
3058 * Think twice before calling this function! It's very easy to get into
3059 * trouble if you don't take great care. Either of the following situations
3060 * will lead to deadlock:
3062 * One of the work items currently on the workqueue needs to acquire
3063 * a lock held by your code or its caller.
3065 * Your code is running in the context of a work routine.
3067 * They will be detected by lockdep when they occur, but the first might not
3068 * occur very often. It depends on what work items are on the workqueue and
3069 * what locks they need, which you have no control over.
3071 * In most situations flushing the entire workqueue is overkill; you merely
3072 * need to know that a particular work item isn't queued and isn't running.
3073 * In such cases you should use cancel_delayed_work_sync() or
3074 * cancel_work_sync() instead.
3076 void flush_scheduled_work(void)
3078 flush_workqueue(system_wq
);
3080 EXPORT_SYMBOL(flush_scheduled_work
);
3083 * execute_in_process_context - reliably execute the routine with user context
3084 * @fn: the function to execute
3085 * @ew: guaranteed storage for the execute work structure (must
3086 * be available when the work executes)
3088 * Executes the function immediately if process context is available,
3089 * otherwise schedules the function for delayed execution.
3091 * Returns: 0 - function was executed
3092 * 1 - function was scheduled for execution
3094 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3096 if (!in_interrupt()) {
3101 INIT_WORK(&ew
->work
, fn
);
3102 schedule_work(&ew
->work
);
3106 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3110 * Workqueues with WQ_SYSFS flag set is visible to userland via
3111 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3112 * following attributes.
3114 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3115 * max_active RW int : maximum number of in-flight work items
3117 * Unbound workqueues have the following extra attributes.
3119 * id RO int : the associated pool ID
3120 * nice RW int : nice value of the workers
3121 * cpumask RW mask : bitmask of allowed CPUs for the workers
3124 struct workqueue_struct
*wq
;
3128 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3130 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3135 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3136 struct device_attribute
*attr
, char *buf
)
3138 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3140 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3143 static ssize_t
wq_max_active_show(struct device
*dev
,
3144 struct device_attribute
*attr
, char *buf
)
3146 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3148 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3151 static ssize_t
wq_max_active_store(struct device
*dev
,
3152 struct device_attribute
*attr
,
3153 const char *buf
, size_t count
)
3155 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3158 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3161 workqueue_set_max_active(wq
, val
);
3165 static struct device_attribute wq_sysfs_attrs
[] = {
3166 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3167 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3171 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3172 struct device_attribute
*attr
, char *buf
)
3174 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3175 const char *delim
= "";
3176 int node
, written
= 0;
3178 rcu_read_lock_sched();
3179 for_each_node(node
) {
3180 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3181 "%s%d:%d", delim
, node
,
3182 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3185 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3186 rcu_read_unlock_sched();
3191 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3194 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3197 mutex_lock(&wq
->mutex
);
3198 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3199 mutex_unlock(&wq
->mutex
);
3204 /* prepare workqueue_attrs for sysfs store operations */
3205 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3207 struct workqueue_attrs
*attrs
;
3209 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3213 mutex_lock(&wq
->mutex
);
3214 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3215 mutex_unlock(&wq
->mutex
);
3219 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3220 const char *buf
, size_t count
)
3222 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3223 struct workqueue_attrs
*attrs
;
3226 attrs
= wq_sysfs_prep_attrs(wq
);
3230 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3231 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3232 ret
= apply_workqueue_attrs(wq
, attrs
);
3236 free_workqueue_attrs(attrs
);
3237 return ret
?: count
;
3240 static ssize_t
wq_cpumask_show(struct device
*dev
,
3241 struct device_attribute
*attr
, char *buf
)
3243 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3246 mutex_lock(&wq
->mutex
);
3247 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3248 mutex_unlock(&wq
->mutex
);
3250 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3254 static ssize_t
wq_cpumask_store(struct device
*dev
,
3255 struct device_attribute
*attr
,
3256 const char *buf
, size_t count
)
3258 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3259 struct workqueue_attrs
*attrs
;
3262 attrs
= wq_sysfs_prep_attrs(wq
);
3266 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3268 ret
= apply_workqueue_attrs(wq
, attrs
);
3270 free_workqueue_attrs(attrs
);
3271 return ret
?: count
;
3274 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3277 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3280 mutex_lock(&wq
->mutex
);
3281 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3282 !wq
->unbound_attrs
->no_numa
);
3283 mutex_unlock(&wq
->mutex
);
3288 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3289 const char *buf
, size_t count
)
3291 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3292 struct workqueue_attrs
*attrs
;
3295 attrs
= wq_sysfs_prep_attrs(wq
);
3300 if (sscanf(buf
, "%d", &v
) == 1) {
3301 attrs
->no_numa
= !v
;
3302 ret
= apply_workqueue_attrs(wq
, attrs
);
3305 free_workqueue_attrs(attrs
);
3306 return ret
?: count
;
3309 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3310 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3311 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3312 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3313 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3317 static struct bus_type wq_subsys
= {
3318 .name
= "workqueue",
3319 .dev_attrs
= wq_sysfs_attrs
,
3322 static int __init
wq_sysfs_init(void)
3324 return subsys_virtual_register(&wq_subsys
, NULL
);
3326 core_initcall(wq_sysfs_init
);
3328 static void wq_device_release(struct device
*dev
)
3330 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3336 * workqueue_sysfs_register - make a workqueue visible in sysfs
3337 * @wq: the workqueue to register
3339 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3340 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3341 * which is the preferred method.
3343 * Workqueue user should use this function directly iff it wants to apply
3344 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3345 * apply_workqueue_attrs() may race against userland updating the
3348 * Returns 0 on success, -errno on failure.
3350 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3352 struct wq_device
*wq_dev
;
3356 * Adjusting max_active or creating new pwqs by applyting
3357 * attributes breaks ordering guarantee. Disallow exposing ordered
3360 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3363 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3368 wq_dev
->dev
.bus
= &wq_subsys
;
3369 wq_dev
->dev
.init_name
= wq
->name
;
3370 wq_dev
->dev
.release
= wq_device_release
;
3373 * unbound_attrs are created separately. Suppress uevent until
3374 * everything is ready.
3376 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3378 ret
= device_register(&wq_dev
->dev
);
3385 if (wq
->flags
& WQ_UNBOUND
) {
3386 struct device_attribute
*attr
;
3388 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3389 ret
= device_create_file(&wq_dev
->dev
, attr
);
3391 device_unregister(&wq_dev
->dev
);
3398 dev_set_uevent_suppress(&wq_dev
->dev
, false);
3399 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3404 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3405 * @wq: the workqueue to unregister
3407 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3409 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3411 struct wq_device
*wq_dev
= wq
->wq_dev
;
3417 device_unregister(&wq_dev
->dev
);
3419 #else /* CONFIG_SYSFS */
3420 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3421 #endif /* CONFIG_SYSFS */
3424 * free_workqueue_attrs - free a workqueue_attrs
3425 * @attrs: workqueue_attrs to free
3427 * Undo alloc_workqueue_attrs().
3429 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3432 free_cpumask_var(attrs
->cpumask
);
3438 * alloc_workqueue_attrs - allocate a workqueue_attrs
3439 * @gfp_mask: allocation mask to use
3441 * Allocate a new workqueue_attrs, initialize with default settings and
3442 * return it. Returns NULL on failure.
3444 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3446 struct workqueue_attrs
*attrs
;
3448 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3451 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3454 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3457 free_workqueue_attrs(attrs
);
3461 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3462 const struct workqueue_attrs
*from
)
3464 to
->nice
= from
->nice
;
3465 cpumask_copy(to
->cpumask
, from
->cpumask
);
3467 * Unlike hash and equality test, this function doesn't ignore
3468 * ->no_numa as it is used for both pool and wq attrs. Instead,
3469 * get_unbound_pool() explicitly clears ->no_numa after copying.
3471 to
->no_numa
= from
->no_numa
;
3474 /* hash value of the content of @attr */
3475 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3479 hash
= jhash_1word(attrs
->nice
, hash
);
3480 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3481 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3485 /* content equality test */
3486 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3487 const struct workqueue_attrs
*b
)
3489 if (a
->nice
!= b
->nice
)
3491 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3497 * init_worker_pool - initialize a newly zalloc'd worker_pool
3498 * @pool: worker_pool to initialize
3500 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3501 * Returns 0 on success, -errno on failure. Even on failure, all fields
3502 * inside @pool proper are initialized and put_unbound_pool() can be called
3503 * on @pool safely to release it.
3505 static int init_worker_pool(struct worker_pool
*pool
)
3507 spin_lock_init(&pool
->lock
);
3510 pool
->node
= NUMA_NO_NODE
;
3511 pool
->flags
|= POOL_DISASSOCIATED
;
3512 INIT_LIST_HEAD(&pool
->worklist
);
3513 INIT_LIST_HEAD(&pool
->idle_list
);
3514 hash_init(pool
->busy_hash
);
3516 init_timer_deferrable(&pool
->idle_timer
);
3517 pool
->idle_timer
.function
= idle_worker_timeout
;
3518 pool
->idle_timer
.data
= (unsigned long)pool
;
3520 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3521 (unsigned long)pool
);
3523 mutex_init(&pool
->manager_arb
);
3524 mutex_init(&pool
->manager_mutex
);
3525 idr_init(&pool
->worker_idr
);
3527 INIT_HLIST_NODE(&pool
->hash_node
);
3530 /* shouldn't fail above this point */
3531 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3537 static void rcu_free_pool(struct rcu_head
*rcu
)
3539 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3541 idr_destroy(&pool
->worker_idr
);
3542 free_workqueue_attrs(pool
->attrs
);
3547 * put_unbound_pool - put a worker_pool
3548 * @pool: worker_pool to put
3550 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3551 * safe manner. get_unbound_pool() calls this function on its failure path
3552 * and this function should be able to release pools which went through,
3553 * successfully or not, init_worker_pool().
3555 * Should be called with wq_pool_mutex held.
3557 static void put_unbound_pool(struct worker_pool
*pool
)
3559 struct worker
*worker
;
3561 lockdep_assert_held(&wq_pool_mutex
);
3567 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3568 WARN_ON(!list_empty(&pool
->worklist
)))
3571 /* release id and unhash */
3573 idr_remove(&worker_pool_idr
, pool
->id
);
3574 hash_del(&pool
->hash_node
);
3577 * Become the manager and destroy all workers. Grabbing
3578 * manager_arb prevents @pool's workers from blocking on
3581 mutex_lock(&pool
->manager_arb
);
3582 mutex_lock(&pool
->manager_mutex
);
3583 spin_lock_irq(&pool
->lock
);
3585 while ((worker
= first_worker(pool
)))
3586 destroy_worker(worker
);
3587 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3589 spin_unlock_irq(&pool
->lock
);
3590 mutex_unlock(&pool
->manager_mutex
);
3591 mutex_unlock(&pool
->manager_arb
);
3593 /* shut down the timers */
3594 del_timer_sync(&pool
->idle_timer
);
3595 del_timer_sync(&pool
->mayday_timer
);
3597 /* sched-RCU protected to allow dereferences from get_work_pool() */
3598 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3602 * get_unbound_pool - get a worker_pool with the specified attributes
3603 * @attrs: the attributes of the worker_pool to get
3605 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3606 * reference count and return it. If there already is a matching
3607 * worker_pool, it will be used; otherwise, this function attempts to
3608 * create a new one. On failure, returns NULL.
3610 * Should be called with wq_pool_mutex held.
3612 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3614 u32 hash
= wqattrs_hash(attrs
);
3615 struct worker_pool
*pool
;
3618 lockdep_assert_held(&wq_pool_mutex
);
3620 /* do we already have a matching pool? */
3621 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3622 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3628 /* nope, create a new one */
3629 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3630 if (!pool
|| init_worker_pool(pool
) < 0)
3633 if (workqueue_freezing
)
3634 pool
->flags
|= POOL_FREEZING
;
3636 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3637 copy_workqueue_attrs(pool
->attrs
, attrs
);
3640 * no_numa isn't a worker_pool attribute, always clear it. See
3641 * 'struct workqueue_attrs' comments for detail.
3643 pool
->attrs
->no_numa
= false;
3645 /* if cpumask is contained inside a NUMA node, we belong to that node */
3646 if (wq_numa_enabled
) {
3647 for_each_node(node
) {
3648 if (cpumask_subset(pool
->attrs
->cpumask
,
3649 wq_numa_possible_cpumask
[node
])) {
3656 if (worker_pool_assign_id(pool
) < 0)
3659 /* create and start the initial worker */
3660 if (create_and_start_worker(pool
) < 0)
3664 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3669 put_unbound_pool(pool
);
3673 static void rcu_free_pwq(struct rcu_head
*rcu
)
3675 kmem_cache_free(pwq_cache
,
3676 container_of(rcu
, struct pool_workqueue
, rcu
));
3680 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3681 * and needs to be destroyed.
3683 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3685 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3686 unbound_release_work
);
3687 struct workqueue_struct
*wq
= pwq
->wq
;
3688 struct worker_pool
*pool
= pwq
->pool
;
3691 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3695 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3696 * necessary on release but do it anyway. It's easier to verify
3697 * and consistent with the linking path.
3699 mutex_lock(&wq
->mutex
);
3700 list_del_rcu(&pwq
->pwqs_node
);
3701 is_last
= list_empty(&wq
->pwqs
);
3702 mutex_unlock(&wq
->mutex
);
3704 mutex_lock(&wq_pool_mutex
);
3705 put_unbound_pool(pool
);
3706 mutex_unlock(&wq_pool_mutex
);
3708 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3711 * If we're the last pwq going away, @wq is already dead and no one
3712 * is gonna access it anymore. Free it.
3715 free_workqueue_attrs(wq
->unbound_attrs
);
3721 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3722 * @pwq: target pool_workqueue
3724 * If @pwq isn't freezing, set @pwq->max_active to the associated
3725 * workqueue's saved_max_active and activate delayed work items
3726 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3728 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3730 struct workqueue_struct
*wq
= pwq
->wq
;
3731 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3733 /* for @wq->saved_max_active */
3734 lockdep_assert_held(&wq
->mutex
);
3736 /* fast exit for non-freezable wqs */
3737 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3740 spin_lock_irq(&pwq
->pool
->lock
);
3742 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3743 pwq
->max_active
= wq
->saved_max_active
;
3745 while (!list_empty(&pwq
->delayed_works
) &&
3746 pwq
->nr_active
< pwq
->max_active
)
3747 pwq_activate_first_delayed(pwq
);
3750 * Need to kick a worker after thawed or an unbound wq's
3751 * max_active is bumped. It's a slow path. Do it always.
3753 wake_up_worker(pwq
->pool
);
3755 pwq
->max_active
= 0;
3758 spin_unlock_irq(&pwq
->pool
->lock
);
3761 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3762 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3763 struct worker_pool
*pool
)
3765 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3767 memset(pwq
, 0, sizeof(*pwq
));
3771 pwq
->flush_color
= -1;
3773 INIT_LIST_HEAD(&pwq
->delayed_works
);
3774 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3775 INIT_LIST_HEAD(&pwq
->mayday_node
);
3776 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3779 /* sync @pwq with the current state of its associated wq and link it */
3780 static void link_pwq(struct pool_workqueue
*pwq
)
3782 struct workqueue_struct
*wq
= pwq
->wq
;
3784 lockdep_assert_held(&wq
->mutex
);
3786 /* may be called multiple times, ignore if already linked */
3787 if (!list_empty(&pwq
->pwqs_node
))
3791 * Set the matching work_color. This is synchronized with
3792 * wq->mutex to avoid confusing flush_workqueue().
3794 pwq
->work_color
= wq
->work_color
;
3796 /* sync max_active to the current setting */
3797 pwq_adjust_max_active(pwq
);
3800 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3803 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3804 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3805 const struct workqueue_attrs
*attrs
)
3807 struct worker_pool
*pool
;
3808 struct pool_workqueue
*pwq
;
3810 lockdep_assert_held(&wq_pool_mutex
);
3812 pool
= get_unbound_pool(attrs
);
3816 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3818 put_unbound_pool(pool
);
3822 init_pwq(pwq
, wq
, pool
);
3826 /* undo alloc_unbound_pwq(), used only in the error path */
3827 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3829 lockdep_assert_held(&wq_pool_mutex
);
3832 put_unbound_pool(pwq
->pool
);
3833 kmem_cache_free(pwq_cache
, pwq
);
3838 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3839 * @attrs: the wq_attrs of interest
3840 * @node: the target NUMA node
3841 * @cpu_going_down: if >= 0, the CPU to consider as offline
3842 * @cpumask: outarg, the resulting cpumask
3844 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3845 * @cpu_going_down is >= 0, that cpu is considered offline during
3846 * calculation. The result is stored in @cpumask. This function returns
3847 * %true if the resulting @cpumask is different from @attrs->cpumask,
3850 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3851 * enabled and @node has online CPUs requested by @attrs, the returned
3852 * cpumask is the intersection of the possible CPUs of @node and
3855 * The caller is responsible for ensuring that the cpumask of @node stays
3858 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3859 int cpu_going_down
, cpumask_t
*cpumask
)
3861 if (!wq_numa_enabled
|| attrs
->no_numa
)
3864 /* does @node have any online CPUs @attrs wants? */
3865 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3866 if (cpu_going_down
>= 0)
3867 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3869 if (cpumask_empty(cpumask
))
3872 /* yeap, return possible CPUs in @node that @attrs wants */
3873 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3874 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3877 cpumask_copy(cpumask
, attrs
->cpumask
);
3881 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3882 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3884 struct pool_workqueue
*pwq
)
3886 struct pool_workqueue
*old_pwq
;
3888 lockdep_assert_held(&wq
->mutex
);
3890 /* link_pwq() can handle duplicate calls */
3893 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3894 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3899 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3900 * @wq: the target workqueue
3901 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3903 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3904 * machines, this function maps a separate pwq to each NUMA node with
3905 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3906 * NUMA node it was issued on. Older pwqs are released as in-flight work
3907 * items finish. Note that a work item which repeatedly requeues itself
3908 * back-to-back will stay on its current pwq.
3910 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3913 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3914 const struct workqueue_attrs
*attrs
)
3916 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3917 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3920 /* only unbound workqueues can change attributes */
3921 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3924 /* creating multiple pwqs breaks ordering guarantee */
3925 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3928 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3929 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3930 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3931 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3934 /* make a copy of @attrs and sanitize it */
3935 copy_workqueue_attrs(new_attrs
, attrs
);
3936 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3939 * We may create multiple pwqs with differing cpumasks. Make a
3940 * copy of @new_attrs which will be modified and used to obtain
3943 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3946 * CPUs should stay stable across pwq creations and installations.
3947 * Pin CPUs, determine the target cpumask for each node and create
3952 mutex_lock(&wq_pool_mutex
);
3955 * If something goes wrong during CPU up/down, we'll fall back to
3956 * the default pwq covering whole @attrs->cpumask. Always create
3957 * it even if we don't use it immediately.
3959 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3963 for_each_node(node
) {
3964 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3965 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3970 pwq_tbl
[node
] = dfl_pwq
;
3974 mutex_unlock(&wq_pool_mutex
);
3976 /* all pwqs have been created successfully, let's install'em */
3977 mutex_lock(&wq
->mutex
);
3979 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3981 /* save the previous pwq and install the new one */
3983 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3985 /* @dfl_pwq might not have been used, ensure it's linked */
3987 swap(wq
->dfl_pwq
, dfl_pwq
);
3989 mutex_unlock(&wq
->mutex
);
3991 /* put the old pwqs */
3993 put_pwq_unlocked(pwq_tbl
[node
]);
3994 put_pwq_unlocked(dfl_pwq
);
4000 free_workqueue_attrs(tmp_attrs
);
4001 free_workqueue_attrs(new_attrs
);
4006 free_unbound_pwq(dfl_pwq
);
4008 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
4009 free_unbound_pwq(pwq_tbl
[node
]);
4010 mutex_unlock(&wq_pool_mutex
);
4018 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4019 * @wq: the target workqueue
4020 * @cpu: the CPU coming up or going down
4021 * @online: whether @cpu is coming up or going down
4023 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4024 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4027 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4028 * falls back to @wq->dfl_pwq which may not be optimal but is always
4031 * Note that when the last allowed CPU of a NUMA node goes offline for a
4032 * workqueue with a cpumask spanning multiple nodes, the workers which were
4033 * already executing the work items for the workqueue will lose their CPU
4034 * affinity and may execute on any CPU. This is similar to how per-cpu
4035 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4036 * affinity, it's the user's responsibility to flush the work item from
4039 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
4042 int node
= cpu_to_node(cpu
);
4043 int cpu_off
= online
? -1 : cpu
;
4044 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
4045 struct workqueue_attrs
*target_attrs
;
4048 lockdep_assert_held(&wq_pool_mutex
);
4050 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
4054 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4055 * Let's use a preallocated one. The following buf is protected by
4056 * CPU hotplug exclusion.
4058 target_attrs
= wq_update_unbound_numa_attrs_buf
;
4059 cpumask
= target_attrs
->cpumask
;
4061 mutex_lock(&wq
->mutex
);
4062 if (wq
->unbound_attrs
->no_numa
)
4065 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4066 pwq
= unbound_pwq_by_node(wq
, node
);
4069 * Let's determine what needs to be done. If the target cpumask is
4070 * different from wq's, we need to compare it to @pwq's and create
4071 * a new one if they don't match. If the target cpumask equals
4072 * wq's, the default pwq should be used. If @pwq is already the
4073 * default one, nothing to do; otherwise, install the default one.
4075 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
4076 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4079 if (pwq
== wq
->dfl_pwq
)
4085 mutex_unlock(&wq
->mutex
);
4087 /* create a new pwq */
4088 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4090 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4092 mutex_lock(&wq
->mutex
);
4097 * Install the new pwq. As this function is called only from CPU
4098 * hotplug callbacks and applying a new attrs is wrapped with
4099 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4102 mutex_lock(&wq
->mutex
);
4103 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4107 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4108 get_pwq(wq
->dfl_pwq
);
4109 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4110 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4112 mutex_unlock(&wq
->mutex
);
4113 put_pwq_unlocked(old_pwq
);
4116 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4118 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4121 if (!(wq
->flags
& WQ_UNBOUND
)) {
4122 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4126 for_each_possible_cpu(cpu
) {
4127 struct pool_workqueue
*pwq
=
4128 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4129 struct worker_pool
*cpu_pools
=
4130 per_cpu(cpu_worker_pools
, cpu
);
4132 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4134 mutex_lock(&wq
->mutex
);
4136 mutex_unlock(&wq
->mutex
);
4139 } else if (wq
->flags
& __WQ_ORDERED
) {
4140 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4141 /* there should only be single pwq for ordering guarantee */
4142 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4143 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4144 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4147 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4151 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4154 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4156 if (max_active
< 1 || max_active
> lim
)
4157 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4158 max_active
, name
, 1, lim
);
4160 return clamp_val(max_active
, 1, lim
);
4163 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4166 struct lock_class_key
*key
,
4167 const char *lock_name
, ...)
4169 size_t tbl_size
= 0;
4171 struct workqueue_struct
*wq
;
4172 struct pool_workqueue
*pwq
;
4174 /* allocate wq and format name */
4175 if (flags
& WQ_UNBOUND
)
4176 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4178 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4182 if (flags
& WQ_UNBOUND
) {
4183 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4184 if (!wq
->unbound_attrs
)
4188 va_start(args
, lock_name
);
4189 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4192 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4193 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4197 wq
->saved_max_active
= max_active
;
4198 mutex_init(&wq
->mutex
);
4199 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4200 INIT_LIST_HEAD(&wq
->pwqs
);
4201 INIT_LIST_HEAD(&wq
->flusher_queue
);
4202 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4203 INIT_LIST_HEAD(&wq
->maydays
);
4205 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4206 INIT_LIST_HEAD(&wq
->list
);
4208 if (alloc_and_link_pwqs(wq
) < 0)
4212 * Workqueues which may be used during memory reclaim should
4213 * have a rescuer to guarantee forward progress.
4215 if (flags
& WQ_MEM_RECLAIM
) {
4216 struct worker
*rescuer
;
4218 rescuer
= alloc_worker();
4222 rescuer
->rescue_wq
= wq
;
4223 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4225 if (IS_ERR(rescuer
->task
)) {
4230 wq
->rescuer
= rescuer
;
4231 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4232 wake_up_process(rescuer
->task
);
4235 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4239 * wq_pool_mutex protects global freeze state and workqueues list.
4240 * Grab it, adjust max_active and add the new @wq to workqueues
4243 mutex_lock(&wq_pool_mutex
);
4245 mutex_lock(&wq
->mutex
);
4246 for_each_pwq(pwq
, wq
)
4247 pwq_adjust_max_active(pwq
);
4248 mutex_unlock(&wq
->mutex
);
4250 list_add(&wq
->list
, &workqueues
);
4252 mutex_unlock(&wq_pool_mutex
);
4257 free_workqueue_attrs(wq
->unbound_attrs
);
4261 destroy_workqueue(wq
);
4264 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4267 * destroy_workqueue - safely terminate a workqueue
4268 * @wq: target workqueue
4270 * Safely destroy a workqueue. All work currently pending will be done first.
4272 void destroy_workqueue(struct workqueue_struct
*wq
)
4274 struct pool_workqueue
*pwq
;
4277 /* drain it before proceeding with destruction */
4278 drain_workqueue(wq
);
4281 mutex_lock(&wq
->mutex
);
4282 for_each_pwq(pwq
, wq
) {
4285 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4286 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4287 mutex_unlock(&wq
->mutex
);
4292 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4293 WARN_ON(pwq
->nr_active
) ||
4294 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4295 mutex_unlock(&wq
->mutex
);
4299 mutex_unlock(&wq
->mutex
);
4302 * wq list is used to freeze wq, remove from list after
4303 * flushing is complete in case freeze races us.
4305 mutex_lock(&wq_pool_mutex
);
4306 list_del_init(&wq
->list
);
4307 mutex_unlock(&wq_pool_mutex
);
4309 workqueue_sysfs_unregister(wq
);
4312 kthread_stop(wq
->rescuer
->task
);
4317 if (!(wq
->flags
& WQ_UNBOUND
)) {
4319 * The base ref is never dropped on per-cpu pwqs. Directly
4320 * free the pwqs and wq.
4322 free_percpu(wq
->cpu_pwqs
);
4326 * We're the sole accessor of @wq at this point. Directly
4327 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4328 * @wq will be freed when the last pwq is released.
4330 for_each_node(node
) {
4331 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4332 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4333 put_pwq_unlocked(pwq
);
4337 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4338 * put. Don't access it afterwards.
4342 put_pwq_unlocked(pwq
);
4345 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4348 * workqueue_set_max_active - adjust max_active of a workqueue
4349 * @wq: target workqueue
4350 * @max_active: new max_active value.
4352 * Set max_active of @wq to @max_active.
4355 * Don't call from IRQ context.
4357 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4359 struct pool_workqueue
*pwq
;
4361 /* disallow meddling with max_active for ordered workqueues */
4362 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4365 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4367 mutex_lock(&wq
->mutex
);
4369 wq
->saved_max_active
= max_active
;
4371 for_each_pwq(pwq
, wq
)
4372 pwq_adjust_max_active(pwq
);
4374 mutex_unlock(&wq
->mutex
);
4376 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4379 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4381 * Determine whether %current is a workqueue rescuer. Can be used from
4382 * work functions to determine whether it's being run off the rescuer task.
4384 bool current_is_workqueue_rescuer(void)
4386 struct worker
*worker
= current_wq_worker();
4388 return worker
&& worker
->rescue_wq
;
4392 * workqueue_congested - test whether a workqueue is congested
4393 * @cpu: CPU in question
4394 * @wq: target workqueue
4396 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4397 * no synchronization around this function and the test result is
4398 * unreliable and only useful as advisory hints or for debugging.
4400 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4401 * Note that both per-cpu and unbound workqueues may be associated with
4402 * multiple pool_workqueues which have separate congested states. A
4403 * workqueue being congested on one CPU doesn't mean the workqueue is also
4404 * contested on other CPUs / NUMA nodes.
4407 * %true if congested, %false otherwise.
4409 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4411 struct pool_workqueue
*pwq
;
4414 rcu_read_lock_sched();
4416 if (cpu
== WORK_CPU_UNBOUND
)
4417 cpu
= smp_processor_id();
4419 if (!(wq
->flags
& WQ_UNBOUND
))
4420 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4422 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4424 ret
= !list_empty(&pwq
->delayed_works
);
4425 rcu_read_unlock_sched();
4429 EXPORT_SYMBOL_GPL(workqueue_congested
);
4432 * work_busy - test whether a work is currently pending or running
4433 * @work: the work to be tested
4435 * Test whether @work is currently pending or running. There is no
4436 * synchronization around this function and the test result is
4437 * unreliable and only useful as advisory hints or for debugging.
4440 * OR'd bitmask of WORK_BUSY_* bits.
4442 unsigned int work_busy(struct work_struct
*work
)
4444 struct worker_pool
*pool
;
4445 unsigned long flags
;
4446 unsigned int ret
= 0;
4448 if (work_pending(work
))
4449 ret
|= WORK_BUSY_PENDING
;
4451 local_irq_save(flags
);
4452 pool
= get_work_pool(work
);
4454 spin_lock(&pool
->lock
);
4455 if (find_worker_executing_work(pool
, work
))
4456 ret
|= WORK_BUSY_RUNNING
;
4457 spin_unlock(&pool
->lock
);
4459 local_irq_restore(flags
);
4463 EXPORT_SYMBOL_GPL(work_busy
);
4466 * set_worker_desc - set description for the current work item
4467 * @fmt: printf-style format string
4468 * @...: arguments for the format string
4470 * This function can be called by a running work function to describe what
4471 * the work item is about. If the worker task gets dumped, this
4472 * information will be printed out together to help debugging. The
4473 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4475 void set_worker_desc(const char *fmt
, ...)
4477 struct worker
*worker
= current_wq_worker();
4481 va_start(args
, fmt
);
4482 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4484 worker
->desc_valid
= true;
4489 * print_worker_info - print out worker information and description
4490 * @log_lvl: the log level to use when printing
4491 * @task: target task
4493 * If @task is a worker and currently executing a work item, print out the
4494 * name of the workqueue being serviced and worker description set with
4495 * set_worker_desc() by the currently executing work item.
4497 * This function can be safely called on any task as long as the
4498 * task_struct itself is accessible. While safe, this function isn't
4499 * synchronized and may print out mixups or garbages of limited length.
4501 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4503 work_func_t
*fn
= NULL
;
4504 char name
[WQ_NAME_LEN
] = { };
4505 char desc
[WORKER_DESC_LEN
] = { };
4506 struct pool_workqueue
*pwq
= NULL
;
4507 struct workqueue_struct
*wq
= NULL
;
4508 bool desc_valid
= false;
4509 struct worker
*worker
;
4511 if (!(task
->flags
& PF_WQ_WORKER
))
4515 * This function is called without any synchronization and @task
4516 * could be in any state. Be careful with dereferences.
4518 worker
= probe_kthread_data(task
);
4521 * Carefully copy the associated workqueue's workfn and name. Keep
4522 * the original last '\0' in case the original contains garbage.
4524 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4525 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4526 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4527 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4529 /* copy worker description */
4530 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4532 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4534 if (fn
|| name
[0] || desc
[0]) {
4535 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4537 pr_cont(" (%s)", desc
);
4545 * There are two challenges in supporting CPU hotplug. Firstly, there
4546 * are a lot of assumptions on strong associations among work, pwq and
4547 * pool which make migrating pending and scheduled works very
4548 * difficult to implement without impacting hot paths. Secondly,
4549 * worker pools serve mix of short, long and very long running works making
4550 * blocked draining impractical.
4552 * This is solved by allowing the pools to be disassociated from the CPU
4553 * running as an unbound one and allowing it to be reattached later if the
4554 * cpu comes back online.
4557 static void wq_unbind_fn(struct work_struct
*work
)
4559 int cpu
= smp_processor_id();
4560 struct worker_pool
*pool
;
4561 struct worker
*worker
;
4564 for_each_cpu_worker_pool(pool
, cpu
) {
4565 WARN_ON_ONCE(cpu
!= smp_processor_id());
4567 mutex_lock(&pool
->manager_mutex
);
4568 spin_lock_irq(&pool
->lock
);
4571 * We've blocked all manager operations. Make all workers
4572 * unbound and set DISASSOCIATED. Before this, all workers
4573 * except for the ones which are still executing works from
4574 * before the last CPU down must be on the cpu. After
4575 * this, they may become diasporas.
4577 for_each_pool_worker(worker
, wi
, pool
)
4578 worker
->flags
|= WORKER_UNBOUND
;
4580 pool
->flags
|= POOL_DISASSOCIATED
;
4582 spin_unlock_irq(&pool
->lock
);
4583 mutex_unlock(&pool
->manager_mutex
);
4586 * Call schedule() so that we cross rq->lock and thus can
4587 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4588 * This is necessary as scheduler callbacks may be invoked
4594 * Sched callbacks are disabled now. Zap nr_running.
4595 * After this, nr_running stays zero and need_more_worker()
4596 * and keep_working() are always true as long as the
4597 * worklist is not empty. This pool now behaves as an
4598 * unbound (in terms of concurrency management) pool which
4599 * are served by workers tied to the pool.
4601 atomic_set(&pool
->nr_running
, 0);
4604 * With concurrency management just turned off, a busy
4605 * worker blocking could lead to lengthy stalls. Kick off
4606 * unbound chain execution of currently pending work items.
4608 spin_lock_irq(&pool
->lock
);
4609 wake_up_worker(pool
);
4610 spin_unlock_irq(&pool
->lock
);
4615 * rebind_workers - rebind all workers of a pool to the associated CPU
4616 * @pool: pool of interest
4618 * @pool->cpu is coming online. Rebind all workers to the CPU.
4620 static void rebind_workers(struct worker_pool
*pool
)
4622 struct worker
*worker
;
4625 lockdep_assert_held(&pool
->manager_mutex
);
4628 * Restore CPU affinity of all workers. As all idle workers should
4629 * be on the run-queue of the associated CPU before any local
4630 * wake-ups for concurrency management happen, restore CPU affinty
4631 * of all workers first and then clear UNBOUND. As we're called
4632 * from CPU_ONLINE, the following shouldn't fail.
4634 for_each_pool_worker(worker
, wi
, pool
)
4635 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4636 pool
->attrs
->cpumask
) < 0);
4638 spin_lock_irq(&pool
->lock
);
4640 for_each_pool_worker(worker
, wi
, pool
) {
4641 unsigned int worker_flags
= worker
->flags
;
4644 * A bound idle worker should actually be on the runqueue
4645 * of the associated CPU for local wake-ups targeting it to
4646 * work. Kick all idle workers so that they migrate to the
4647 * associated CPU. Doing this in the same loop as
4648 * replacing UNBOUND with REBOUND is safe as no worker will
4649 * be bound before @pool->lock is released.
4651 if (worker_flags
& WORKER_IDLE
)
4652 wake_up_process(worker
->task
);
4655 * We want to clear UNBOUND but can't directly call
4656 * worker_clr_flags() or adjust nr_running. Atomically
4657 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4658 * @worker will clear REBOUND using worker_clr_flags() when
4659 * it initiates the next execution cycle thus restoring
4660 * concurrency management. Note that when or whether
4661 * @worker clears REBOUND doesn't affect correctness.
4663 * ACCESS_ONCE() is necessary because @worker->flags may be
4664 * tested without holding any lock in
4665 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4666 * fail incorrectly leading to premature concurrency
4667 * management operations.
4669 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4670 worker_flags
|= WORKER_REBOUND
;
4671 worker_flags
&= ~WORKER_UNBOUND
;
4672 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4675 spin_unlock_irq(&pool
->lock
);
4679 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4680 * @pool: unbound pool of interest
4681 * @cpu: the CPU which is coming up
4683 * An unbound pool may end up with a cpumask which doesn't have any online
4684 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4685 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4686 * online CPU before, cpus_allowed of all its workers should be restored.
4688 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4690 static cpumask_t cpumask
;
4691 struct worker
*worker
;
4694 lockdep_assert_held(&pool
->manager_mutex
);
4696 /* is @cpu allowed for @pool? */
4697 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4700 /* is @cpu the only online CPU? */
4701 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4702 if (cpumask_weight(&cpumask
) != 1)
4705 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4706 for_each_pool_worker(worker
, wi
, pool
)
4707 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4708 pool
->attrs
->cpumask
) < 0);
4712 * Workqueues should be brought up before normal priority CPU notifiers.
4713 * This will be registered high priority CPU notifier.
4715 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4716 unsigned long action
,
4719 int cpu
= (unsigned long)hcpu
;
4720 struct worker_pool
*pool
;
4721 struct workqueue_struct
*wq
;
4724 switch (action
& ~CPU_TASKS_FROZEN
) {
4725 case CPU_UP_PREPARE
:
4726 for_each_cpu_worker_pool(pool
, cpu
) {
4727 if (pool
->nr_workers
)
4729 if (create_and_start_worker(pool
) < 0)
4734 case CPU_DOWN_FAILED
:
4736 mutex_lock(&wq_pool_mutex
);
4738 for_each_pool(pool
, pi
) {
4739 mutex_lock(&pool
->manager_mutex
);
4741 if (pool
->cpu
== cpu
) {
4742 spin_lock_irq(&pool
->lock
);
4743 pool
->flags
&= ~POOL_DISASSOCIATED
;
4744 spin_unlock_irq(&pool
->lock
);
4746 rebind_workers(pool
);
4747 } else if (pool
->cpu
< 0) {
4748 restore_unbound_workers_cpumask(pool
, cpu
);
4751 mutex_unlock(&pool
->manager_mutex
);
4754 /* update NUMA affinity of unbound workqueues */
4755 list_for_each_entry(wq
, &workqueues
, list
)
4756 wq_update_unbound_numa(wq
, cpu
, true);
4758 mutex_unlock(&wq_pool_mutex
);
4765 * Workqueues should be brought down after normal priority CPU notifiers.
4766 * This will be registered as low priority CPU notifier.
4768 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4769 unsigned long action
,
4772 int cpu
= (unsigned long)hcpu
;
4773 struct work_struct unbind_work
;
4774 struct workqueue_struct
*wq
;
4776 switch (action
& ~CPU_TASKS_FROZEN
) {
4777 case CPU_DOWN_PREPARE
:
4778 /* unbinding per-cpu workers should happen on the local CPU */
4779 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4780 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4782 /* update NUMA affinity of unbound workqueues */
4783 mutex_lock(&wq_pool_mutex
);
4784 list_for_each_entry(wq
, &workqueues
, list
)
4785 wq_update_unbound_numa(wq
, cpu
, false);
4786 mutex_unlock(&wq_pool_mutex
);
4788 /* wait for per-cpu unbinding to finish */
4789 flush_work(&unbind_work
);
4797 struct work_for_cpu
{
4798 struct work_struct work
;
4804 static void work_for_cpu_fn(struct work_struct
*work
)
4806 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4808 wfc
->ret
= wfc
->fn(wfc
->arg
);
4812 * work_on_cpu - run a function in user context on a particular cpu
4813 * @cpu: the cpu to run on
4814 * @fn: the function to run
4815 * @arg: the function arg
4817 * This will return the value @fn returns.
4818 * It is up to the caller to ensure that the cpu doesn't go offline.
4819 * The caller must not hold any locks which would prevent @fn from completing.
4821 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4823 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4825 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4826 schedule_work_on(cpu
, &wfc
.work
);
4827 flush_work(&wfc
.work
);
4830 EXPORT_SYMBOL_GPL(work_on_cpu
);
4831 #endif /* CONFIG_SMP */
4833 #ifdef CONFIG_FREEZER
4836 * freeze_workqueues_begin - begin freezing workqueues
4838 * Start freezing workqueues. After this function returns, all freezable
4839 * workqueues will queue new works to their delayed_works list instead of
4843 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4845 void freeze_workqueues_begin(void)
4847 struct worker_pool
*pool
;
4848 struct workqueue_struct
*wq
;
4849 struct pool_workqueue
*pwq
;
4852 mutex_lock(&wq_pool_mutex
);
4854 WARN_ON_ONCE(workqueue_freezing
);
4855 workqueue_freezing
= true;
4858 for_each_pool(pool
, pi
) {
4859 spin_lock_irq(&pool
->lock
);
4860 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4861 pool
->flags
|= POOL_FREEZING
;
4862 spin_unlock_irq(&pool
->lock
);
4865 list_for_each_entry(wq
, &workqueues
, list
) {
4866 mutex_lock(&wq
->mutex
);
4867 for_each_pwq(pwq
, wq
)
4868 pwq_adjust_max_active(pwq
);
4869 mutex_unlock(&wq
->mutex
);
4872 mutex_unlock(&wq_pool_mutex
);
4876 * freeze_workqueues_busy - are freezable workqueues still busy?
4878 * Check whether freezing is complete. This function must be called
4879 * between freeze_workqueues_begin() and thaw_workqueues().
4882 * Grabs and releases wq_pool_mutex.
4885 * %true if some freezable workqueues are still busy. %false if freezing
4888 bool freeze_workqueues_busy(void)
4891 struct workqueue_struct
*wq
;
4892 struct pool_workqueue
*pwq
;
4894 mutex_lock(&wq_pool_mutex
);
4896 WARN_ON_ONCE(!workqueue_freezing
);
4898 list_for_each_entry(wq
, &workqueues
, list
) {
4899 if (!(wq
->flags
& WQ_FREEZABLE
))
4902 * nr_active is monotonically decreasing. It's safe
4903 * to peek without lock.
4905 rcu_read_lock_sched();
4906 for_each_pwq(pwq
, wq
) {
4907 WARN_ON_ONCE(pwq
->nr_active
< 0);
4908 if (pwq
->nr_active
) {
4910 rcu_read_unlock_sched();
4914 rcu_read_unlock_sched();
4917 mutex_unlock(&wq_pool_mutex
);
4922 * thaw_workqueues - thaw workqueues
4924 * Thaw workqueues. Normal queueing is restored and all collected
4925 * frozen works are transferred to their respective pool worklists.
4928 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4930 void thaw_workqueues(void)
4932 struct workqueue_struct
*wq
;
4933 struct pool_workqueue
*pwq
;
4934 struct worker_pool
*pool
;
4937 mutex_lock(&wq_pool_mutex
);
4939 if (!workqueue_freezing
)
4942 /* clear FREEZING */
4943 for_each_pool(pool
, pi
) {
4944 spin_lock_irq(&pool
->lock
);
4945 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4946 pool
->flags
&= ~POOL_FREEZING
;
4947 spin_unlock_irq(&pool
->lock
);
4950 /* restore max_active and repopulate worklist */
4951 list_for_each_entry(wq
, &workqueues
, list
) {
4952 mutex_lock(&wq
->mutex
);
4953 for_each_pwq(pwq
, wq
)
4954 pwq_adjust_max_active(pwq
);
4955 mutex_unlock(&wq
->mutex
);
4958 workqueue_freezing
= false;
4960 mutex_unlock(&wq_pool_mutex
);
4962 #endif /* CONFIG_FREEZER */
4964 static void __init
wq_numa_init(void)
4969 /* determine NUMA pwq table len - highest node id + 1 */
4971 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4973 if (num_possible_nodes() <= 1)
4976 if (wq_disable_numa
) {
4977 pr_info("workqueue: NUMA affinity support disabled\n");
4981 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4982 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4985 * We want masks of possible CPUs of each node which isn't readily
4986 * available. Build one from cpu_to_node() which should have been
4987 * fully initialized by now.
4989 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
4993 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
4994 node_online(node
) ? node
: NUMA_NO_NODE
));
4996 for_each_possible_cpu(cpu
) {
4997 node
= cpu_to_node(cpu
);
4998 if (WARN_ON(node
== NUMA_NO_NODE
)) {
4999 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5000 /* happens iff arch is bonkers, let's just proceed */
5003 cpumask_set_cpu(cpu
, tbl
[node
]);
5006 wq_numa_possible_cpumask
= tbl
;
5007 wq_numa_enabled
= true;
5010 static int __init
init_workqueues(void)
5012 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5015 /* make sure we have enough bits for OFFQ pool ID */
5016 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
5017 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
5019 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5021 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5023 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
5024 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
5028 /* initialize CPU pools */
5029 for_each_possible_cpu(cpu
) {
5030 struct worker_pool
*pool
;
5033 for_each_cpu_worker_pool(pool
, cpu
) {
5034 BUG_ON(init_worker_pool(pool
));
5036 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5037 pool
->attrs
->nice
= std_nice
[i
++];
5038 pool
->node
= cpu_to_node(cpu
);
5041 mutex_lock(&wq_pool_mutex
);
5042 BUG_ON(worker_pool_assign_id(pool
));
5043 mutex_unlock(&wq_pool_mutex
);
5047 /* create the initial worker */
5048 for_each_online_cpu(cpu
) {
5049 struct worker_pool
*pool
;
5051 for_each_cpu_worker_pool(pool
, cpu
) {
5052 pool
->flags
&= ~POOL_DISASSOCIATED
;
5053 BUG_ON(create_and_start_worker(pool
) < 0);
5057 /* create default unbound and ordered wq attrs */
5058 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5059 struct workqueue_attrs
*attrs
;
5061 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5062 attrs
->nice
= std_nice
[i
];
5063 unbound_std_wq_attrs
[i
] = attrs
;
5066 * An ordered wq should have only one pwq as ordering is
5067 * guaranteed by max_active which is enforced by pwqs.
5068 * Turn off NUMA so that dfl_pwq is used for all nodes.
5070 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5071 attrs
->nice
= std_nice
[i
];
5072 attrs
->no_numa
= true;
5073 ordered_wq_attrs
[i
] = attrs
;
5076 system_wq
= alloc_workqueue("events", 0, 0);
5077 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5078 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5079 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5080 WQ_UNBOUND_MAX_ACTIVE
);
5081 system_freezable_wq
= alloc_workqueue("events_freezable",
5083 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5084 !system_unbound_wq
|| !system_freezable_wq
);
5087 early_initcall(init_workqueues
);