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);
617 * The following mb guarantees that previous clear of a PENDING bit
618 * will not be reordered with any speculative LOADS or STORES from
619 * work->current_func, which is executed afterwards. This possible
620 * reordering can lead to a missed execution on attempt to qeueue
621 * the same @work. E.g. consider this case:
624 * ---------------------------- --------------------------------
626 * 1 STORE event_indicated
627 * 2 queue_work_on() {
628 * 3 test_and_set_bit(PENDING)
629 * 4 } set_..._and_clear_pending() {
630 * 5 set_work_data() # clear bit
632 * 7 work->current_func() {
633 * 8 LOAD event_indicated
636 * Without an explicit full barrier speculative LOAD on line 8 can
637 * be executed before CPU#0 does STORE on line 1. If that happens,
638 * CPU#0 observes the PENDING bit is still set and new execution of
639 * a @work is not queued in a hope, that CPU#1 will eventually
640 * finish the queued @work. Meanwhile CPU#1 does not see
641 * event_indicated is set, because speculative LOAD was executed
642 * before actual STORE.
647 static void clear_work_data(struct work_struct
*work
)
649 smp_wmb(); /* see set_work_pool_and_clear_pending() */
650 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
653 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
655 unsigned long data
= atomic_long_read(&work
->data
);
657 if (data
& WORK_STRUCT_PWQ
)
658 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
664 * get_work_pool - return the worker_pool a given work was associated with
665 * @work: the work item of interest
667 * Return the worker_pool @work was last associated with. %NULL if none.
669 * Pools are created and destroyed under wq_pool_mutex, and allows read
670 * access under sched-RCU read lock. As such, this function should be
671 * called under wq_pool_mutex or with preemption disabled.
673 * All fields of the returned pool are accessible as long as the above
674 * mentioned locking is in effect. If the returned pool needs to be used
675 * beyond the critical section, the caller is responsible for ensuring the
676 * returned pool is and stays online.
678 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
680 unsigned long data
= atomic_long_read(&work
->data
);
683 assert_rcu_or_pool_mutex();
685 if (data
& WORK_STRUCT_PWQ
)
686 return ((struct pool_workqueue
*)
687 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
689 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
690 if (pool_id
== WORK_OFFQ_POOL_NONE
)
693 return idr_find(&worker_pool_idr
, pool_id
);
697 * get_work_pool_id - return the worker pool ID a given work is associated with
698 * @work: the work item of interest
700 * Return the worker_pool ID @work was last associated with.
701 * %WORK_OFFQ_POOL_NONE if none.
703 static int get_work_pool_id(struct work_struct
*work
)
705 unsigned long data
= atomic_long_read(&work
->data
);
707 if (data
& WORK_STRUCT_PWQ
)
708 return ((struct pool_workqueue
*)
709 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
711 return data
>> WORK_OFFQ_POOL_SHIFT
;
714 static void mark_work_canceling(struct work_struct
*work
)
716 unsigned long pool_id
= get_work_pool_id(work
);
718 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
719 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
722 static bool work_is_canceling(struct work_struct
*work
)
724 unsigned long data
= atomic_long_read(&work
->data
);
726 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
730 * Policy functions. These define the policies on how the global worker
731 * pools are managed. Unless noted otherwise, these functions assume that
732 * they're being called with pool->lock held.
735 static bool __need_more_worker(struct worker_pool
*pool
)
737 return !atomic_read(&pool
->nr_running
);
741 * Need to wake up a worker? Called from anything but currently
744 * Note that, because unbound workers never contribute to nr_running, this
745 * function will always return %true for unbound pools as long as the
746 * worklist isn't empty.
748 static bool need_more_worker(struct worker_pool
*pool
)
750 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
753 /* Can I start working? Called from busy but !running workers. */
754 static bool may_start_working(struct worker_pool
*pool
)
756 return pool
->nr_idle
;
759 /* Do I need to keep working? Called from currently running workers. */
760 static bool keep_working(struct worker_pool
*pool
)
762 return !list_empty(&pool
->worklist
) &&
763 atomic_read(&pool
->nr_running
) <= 1;
766 /* Do we need a new worker? Called from manager. */
767 static bool need_to_create_worker(struct worker_pool
*pool
)
769 return need_more_worker(pool
) && !may_start_working(pool
);
772 /* Do I need to be the manager? */
773 static bool need_to_manage_workers(struct worker_pool
*pool
)
775 return need_to_create_worker(pool
) ||
776 (pool
->flags
& POOL_MANAGE_WORKERS
);
779 /* Do we have too many workers and should some go away? */
780 static bool too_many_workers(struct worker_pool
*pool
)
782 bool managing
= mutex_is_locked(&pool
->manager_arb
);
783 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
784 int nr_busy
= pool
->nr_workers
- nr_idle
;
787 * nr_idle and idle_list may disagree if idle rebinding is in
788 * progress. Never return %true if idle_list is empty.
790 if (list_empty(&pool
->idle_list
))
793 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
800 /* Return the first worker. Safe with preemption disabled */
801 static struct worker
*first_worker(struct worker_pool
*pool
)
803 if (unlikely(list_empty(&pool
->idle_list
)))
806 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
810 * wake_up_worker - wake up an idle worker
811 * @pool: worker pool to wake worker from
813 * Wake up the first idle worker of @pool.
816 * spin_lock_irq(pool->lock).
818 static void wake_up_worker(struct worker_pool
*pool
)
820 struct worker
*worker
= first_worker(pool
);
823 wake_up_process(worker
->task
);
827 * wq_worker_waking_up - a worker is waking up
828 * @task: task waking up
829 * @cpu: CPU @task is waking up to
831 * This function is called during try_to_wake_up() when a worker is
835 * spin_lock_irq(rq->lock)
837 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
839 struct worker
*worker
= kthread_data(task
);
841 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
842 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
843 atomic_inc(&worker
->pool
->nr_running
);
848 * wq_worker_sleeping - a worker is going to sleep
849 * @task: task going to sleep
850 * @cpu: CPU in question, must be the current CPU number
852 * This function is called during schedule() when a busy worker is
853 * going to sleep. Worker on the same cpu can be woken up by
854 * returning pointer to its task.
857 * spin_lock_irq(rq->lock)
860 * Worker task on @cpu to wake up, %NULL if none.
862 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
864 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
865 struct worker_pool
*pool
;
868 * Rescuers, which may not have all the fields set up like normal
869 * workers, also reach here, let's not access anything before
870 * checking NOT_RUNNING.
872 if (worker
->flags
& WORKER_NOT_RUNNING
)
877 /* this can only happen on the local cpu */
878 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
882 * The counterpart of the following dec_and_test, implied mb,
883 * worklist not empty test sequence is in insert_work().
884 * Please read comment there.
886 * NOT_RUNNING is clear. This means that we're bound to and
887 * running on the local cpu w/ rq lock held and preemption
888 * disabled, which in turn means that none else could be
889 * manipulating idle_list, so dereferencing idle_list without pool
892 if (atomic_dec_and_test(&pool
->nr_running
) &&
893 !list_empty(&pool
->worklist
))
894 to_wakeup
= first_worker(pool
);
895 return to_wakeup
? to_wakeup
->task
: NULL
;
899 * worker_set_flags - set worker flags and adjust nr_running accordingly
901 * @flags: flags to set
902 * @wakeup: wakeup an idle worker if necessary
904 * Set @flags in @worker->flags and adjust nr_running accordingly. If
905 * nr_running becomes zero and @wakeup is %true, an idle worker is
909 * spin_lock_irq(pool->lock)
911 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
914 struct worker_pool
*pool
= worker
->pool
;
916 WARN_ON_ONCE(worker
->task
!= current
);
919 * If transitioning into NOT_RUNNING, adjust nr_running and
920 * wake up an idle worker as necessary if requested by
923 if ((flags
& WORKER_NOT_RUNNING
) &&
924 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
926 if (atomic_dec_and_test(&pool
->nr_running
) &&
927 !list_empty(&pool
->worklist
))
928 wake_up_worker(pool
);
930 atomic_dec(&pool
->nr_running
);
933 worker
->flags
|= flags
;
937 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
939 * @flags: flags to clear
941 * Clear @flags in @worker->flags and adjust nr_running accordingly.
944 * spin_lock_irq(pool->lock)
946 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
948 struct worker_pool
*pool
= worker
->pool
;
949 unsigned int oflags
= worker
->flags
;
951 WARN_ON_ONCE(worker
->task
!= current
);
953 worker
->flags
&= ~flags
;
956 * If transitioning out of NOT_RUNNING, increment nr_running. Note
957 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
958 * of multiple flags, not a single flag.
960 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
961 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
962 atomic_inc(&pool
->nr_running
);
966 * find_worker_executing_work - find worker which is executing a work
967 * @pool: pool of interest
968 * @work: work to find worker for
970 * Find a worker which is executing @work on @pool by searching
971 * @pool->busy_hash which is keyed by the address of @work. For a worker
972 * to match, its current execution should match the address of @work and
973 * its work function. This is to avoid unwanted dependency between
974 * unrelated work executions through a work item being recycled while still
977 * This is a bit tricky. A work item may be freed once its execution
978 * starts and nothing prevents the freed area from being recycled for
979 * another work item. If the same work item address ends up being reused
980 * before the original execution finishes, workqueue will identify the
981 * recycled work item as currently executing and make it wait until the
982 * current execution finishes, introducing an unwanted dependency.
984 * This function checks the work item address and work function to avoid
985 * false positives. Note that this isn't complete as one may construct a
986 * work function which can introduce dependency onto itself through a
987 * recycled work item. Well, if somebody wants to shoot oneself in the
988 * foot that badly, there's only so much we can do, and if such deadlock
989 * actually occurs, it should be easy to locate the culprit work function.
992 * spin_lock_irq(pool->lock).
995 * Pointer to worker which is executing @work if found, NULL
998 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
999 struct work_struct
*work
)
1001 struct worker
*worker
;
1003 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
1004 (unsigned long)work
)
1005 if (worker
->current_work
== work
&&
1006 worker
->current_func
== work
->func
)
1013 * move_linked_works - move linked works to a list
1014 * @work: start of series of works to be scheduled
1015 * @head: target list to append @work to
1016 * @nextp: out paramter for nested worklist walking
1018 * Schedule linked works starting from @work to @head. Work series to
1019 * be scheduled starts at @work and includes any consecutive work with
1020 * WORK_STRUCT_LINKED set in its predecessor.
1022 * If @nextp is not NULL, it's updated to point to the next work of
1023 * the last scheduled work. This allows move_linked_works() to be
1024 * nested inside outer list_for_each_entry_safe().
1027 * spin_lock_irq(pool->lock).
1029 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1030 struct work_struct
**nextp
)
1032 struct work_struct
*n
;
1035 * Linked worklist will always end before the end of the list,
1036 * use NULL for list head.
1038 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1039 list_move_tail(&work
->entry
, head
);
1040 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1045 * If we're already inside safe list traversal and have moved
1046 * multiple works to the scheduled queue, the next position
1047 * needs to be updated.
1054 * get_pwq - get an extra reference on the specified pool_workqueue
1055 * @pwq: pool_workqueue to get
1057 * Obtain an extra reference on @pwq. The caller should guarantee that
1058 * @pwq has positive refcnt and be holding the matching pool->lock.
1060 static void get_pwq(struct pool_workqueue
*pwq
)
1062 lockdep_assert_held(&pwq
->pool
->lock
);
1063 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1068 * put_pwq - put a pool_workqueue reference
1069 * @pwq: pool_workqueue to put
1071 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1072 * destruction. The caller should be holding the matching pool->lock.
1074 static void put_pwq(struct pool_workqueue
*pwq
)
1076 lockdep_assert_held(&pwq
->pool
->lock
);
1077 if (likely(--pwq
->refcnt
))
1079 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1082 * @pwq can't be released under pool->lock, bounce to
1083 * pwq_unbound_release_workfn(). This never recurses on the same
1084 * pool->lock as this path is taken only for unbound workqueues and
1085 * the release work item is scheduled on a per-cpu workqueue. To
1086 * avoid lockdep warning, unbound pool->locks are given lockdep
1087 * subclass of 1 in get_unbound_pool().
1089 schedule_work(&pwq
->unbound_release_work
);
1093 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1094 * @pwq: pool_workqueue to put (can be %NULL)
1096 * put_pwq() with locking. This function also allows %NULL @pwq.
1098 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1102 * As both pwqs and pools are sched-RCU protected, the
1103 * following lock operations are safe.
1105 spin_lock_irq(&pwq
->pool
->lock
);
1107 spin_unlock_irq(&pwq
->pool
->lock
);
1111 static void pwq_activate_delayed_work(struct work_struct
*work
)
1113 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1115 trace_workqueue_activate_work(work
);
1116 #ifdef CONFIG_MTK_WQ_DEBUG
1117 mttrace_workqueue_activate_work(work
);
1118 #endif //CONFIG_MTK_WQ_DEBUG
1119 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1120 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1124 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1126 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1127 struct work_struct
, entry
);
1129 pwq_activate_delayed_work(work
);
1133 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1134 * @pwq: pwq of interest
1135 * @color: color of work which left the queue
1137 * A work either has completed or is removed from pending queue,
1138 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1141 * spin_lock_irq(pool->lock).
1143 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1145 /* uncolored work items don't participate in flushing or nr_active */
1146 if (color
== WORK_NO_COLOR
)
1149 pwq
->nr_in_flight
[color
]--;
1152 if (!list_empty(&pwq
->delayed_works
)) {
1153 /* one down, submit a delayed one */
1154 if (pwq
->nr_active
< pwq
->max_active
)
1155 pwq_activate_first_delayed(pwq
);
1158 /* is flush in progress and are we at the flushing tip? */
1159 if (likely(pwq
->flush_color
!= color
))
1162 /* are there still in-flight works? */
1163 if (pwq
->nr_in_flight
[color
])
1166 /* this pwq is done, clear flush_color */
1167 pwq
->flush_color
= -1;
1170 * If this was the last pwq, wake up the first flusher. It
1171 * will handle the rest.
1173 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1174 complete(&pwq
->wq
->first_flusher
->done
);
1180 * try_to_grab_pending - steal work item from worklist and disable irq
1181 * @work: work item to steal
1182 * @is_dwork: @work is a delayed_work
1183 * @flags: place to store irq state
1185 * Try to grab PENDING bit of @work. This function can handle @work in any
1186 * stable state - idle, on timer or on worklist. Return values are
1188 * 1 if @work was pending and we successfully stole PENDING
1189 * 0 if @work was idle and we claimed PENDING
1190 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1191 * -ENOENT if someone else is canceling @work, this state may persist
1192 * for arbitrarily long
1194 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1195 * interrupted while holding PENDING and @work off queue, irq must be
1196 * disabled on entry. This, combined with delayed_work->timer being
1197 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1199 * On successful return, >= 0, irq is disabled and the caller is
1200 * responsible for releasing it using local_irq_restore(*@flags).
1202 * This function is safe to call from any context including IRQ handler.
1204 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1205 unsigned long *flags
)
1207 struct worker_pool
*pool
;
1208 struct pool_workqueue
*pwq
;
1210 local_irq_save(*flags
);
1212 /* try to steal the timer if it exists */
1214 struct delayed_work
*dwork
= to_delayed_work(work
);
1217 * dwork->timer is irqsafe. If del_timer() fails, it's
1218 * guaranteed that the timer is not queued anywhere and not
1219 * running on the local CPU.
1221 if (likely(del_timer(&dwork
->timer
)))
1225 /* try to claim PENDING the normal way */
1226 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1230 * The queueing is in progress, or it is already queued. Try to
1231 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1233 pool
= get_work_pool(work
);
1237 spin_lock(&pool
->lock
);
1239 * work->data is guaranteed to point to pwq only while the work
1240 * item is queued on pwq->wq, and both updating work->data to point
1241 * to pwq on queueing and to pool on dequeueing are done under
1242 * pwq->pool->lock. This in turn guarantees that, if work->data
1243 * points to pwq which is associated with a locked pool, the work
1244 * item is currently queued on that pool.
1246 pwq
= get_work_pwq(work
);
1247 if (pwq
&& pwq
->pool
== pool
) {
1248 debug_work_deactivate(work
);
1251 * A delayed work item cannot be grabbed directly because
1252 * it might have linked NO_COLOR work items which, if left
1253 * on the delayed_list, will confuse pwq->nr_active
1254 * management later on and cause stall. Make sure the work
1255 * item is activated before grabbing.
1257 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1258 pwq_activate_delayed_work(work
);
1260 list_del_init(&work
->entry
);
1261 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1263 /* work->data points to pwq iff queued, point to pool */
1264 set_work_pool_and_keep_pending(work
, pool
->id
);
1266 spin_unlock(&pool
->lock
);
1269 spin_unlock(&pool
->lock
);
1271 local_irq_restore(*flags
);
1272 if (work_is_canceling(work
))
1279 * insert_work - insert a work into a pool
1280 * @pwq: pwq @work belongs to
1281 * @work: work to insert
1282 * @head: insertion point
1283 * @extra_flags: extra WORK_STRUCT_* flags to set
1285 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1286 * work_struct flags.
1289 * spin_lock_irq(pool->lock).
1291 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1292 struct list_head
*head
, unsigned int extra_flags
)
1294 struct worker_pool
*pool
= pwq
->pool
;
1296 /* we own @work, set data and link */
1297 set_work_pwq(work
, pwq
, extra_flags
);
1298 list_add_tail(&work
->entry
, head
);
1302 * Ensure either wq_worker_sleeping() sees the above
1303 * list_add_tail() or we see zero nr_running to avoid workers lying
1304 * around lazily while there are works to be processed.
1308 if (__need_more_worker(pool
))
1309 wake_up_worker(pool
);
1313 * Test whether @work is being queued from another work executing on the
1316 static bool is_chained_work(struct workqueue_struct
*wq
)
1318 struct worker
*worker
;
1320 worker
= current_wq_worker();
1322 * Return %true iff I'm a worker execuing a work item on @wq. If
1323 * I'm @worker, it's safe to dereference it without locking.
1325 return worker
&& worker
->current_pwq
->wq
== wq
;
1328 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1329 struct work_struct
*work
)
1331 struct pool_workqueue
*pwq
;
1332 struct worker_pool
*last_pool
;
1333 struct list_head
*worklist
;
1334 unsigned int work_flags
;
1335 unsigned int req_cpu
= cpu
;
1338 * While a work item is PENDING && off queue, a task trying to
1339 * steal the PENDING will busy-loop waiting for it to either get
1340 * queued or lose PENDING. Grabbing PENDING and queueing should
1341 * happen with IRQ disabled.
1343 WARN_ON_ONCE(!irqs_disabled());
1345 debug_work_activate(work
);
1347 /* if dying, only works from the same workqueue are allowed */
1348 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1349 WARN_ON_ONCE(!is_chained_work(wq
)))
1352 if (req_cpu
== WORK_CPU_UNBOUND
)
1353 cpu
= raw_smp_processor_id();
1355 /* pwq which will be used unless @work is executing elsewhere */
1356 if (!(wq
->flags
& WQ_UNBOUND
))
1357 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1359 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1362 * If @work was previously on a different pool, it might still be
1363 * running there, in which case the work needs to be queued on that
1364 * pool to guarantee non-reentrancy.
1366 last_pool
= get_work_pool(work
);
1367 if (last_pool
&& last_pool
!= pwq
->pool
) {
1368 struct worker
*worker
;
1370 spin_lock(&last_pool
->lock
);
1372 worker
= find_worker_executing_work(last_pool
, work
);
1374 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1375 pwq
= worker
->current_pwq
;
1377 /* meh... not running there, queue here */
1378 spin_unlock(&last_pool
->lock
);
1379 spin_lock(&pwq
->pool
->lock
);
1382 spin_lock(&pwq
->pool
->lock
);
1386 * pwq is determined and locked. For unbound pools, we could have
1387 * raced with pwq release and it could already be dead. If its
1388 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1389 * without another pwq replacing it in the numa_pwq_tbl or while
1390 * work items are executing on it, so the retrying is guaranteed to
1391 * make forward-progress.
1393 if (unlikely(!pwq
->refcnt
)) {
1394 if (wq
->flags
& WQ_UNBOUND
) {
1395 spin_unlock(&pwq
->pool
->lock
);
1400 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1404 /* pwq determined, queue */
1405 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1406 #ifdef CONFIG_MTK_WQ_DEBUG
1407 mttrace_workqueue_queue_work(cpu
, work
);
1408 #endif //CONFIG_MTK_WQ_DEBUG
1410 if (WARN_ON(!list_empty(&work
->entry
))) {
1411 spin_unlock(&pwq
->pool
->lock
);
1415 pwq
->nr_in_flight
[pwq
->work_color
]++;
1416 work_flags
= work_color_to_flags(pwq
->work_color
);
1418 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1419 trace_workqueue_activate_work(work
);
1420 #ifdef CONFIG_MTK_WQ_DEBUG
1421 mttrace_workqueue_activate_work(work
);
1422 #endif //CONFIG_MTK_WQ_DEBUG
1424 worklist
= &pwq
->pool
->worklist
;
1426 work_flags
|= WORK_STRUCT_DELAYED
;
1427 worklist
= &pwq
->delayed_works
;
1430 insert_work(pwq
, work
, worklist
, work_flags
);
1432 spin_unlock(&pwq
->pool
->lock
);
1436 * queue_work_on - queue work on specific cpu
1437 * @cpu: CPU number to execute work on
1438 * @wq: workqueue to use
1439 * @work: work to queue
1441 * Returns %false if @work was already on a queue, %true otherwise.
1443 * We queue the work to a specific CPU, the caller must ensure it
1446 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1447 struct work_struct
*work
)
1450 unsigned long flags
;
1452 local_irq_save(flags
);
1454 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1455 __queue_work(cpu
, wq
, work
);
1459 local_irq_restore(flags
);
1462 EXPORT_SYMBOL(queue_work_on
);
1464 void delayed_work_timer_fn(unsigned long __data
)
1466 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1468 /* should have been called from irqsafe timer with irq already off */
1469 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1471 EXPORT_SYMBOL(delayed_work_timer_fn
);
1473 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1474 struct delayed_work
*dwork
, unsigned long delay
)
1476 struct timer_list
*timer
= &dwork
->timer
;
1477 struct work_struct
*work
= &dwork
->work
;
1479 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1480 timer
->data
!= (unsigned long)dwork
);
1481 WARN_ON_ONCE(timer_pending(timer
));
1482 WARN_ON_ONCE(!list_empty(&work
->entry
));
1485 * If @delay is 0, queue @dwork->work immediately. This is for
1486 * both optimization and correctness. The earliest @timer can
1487 * expire is on the closest next tick and delayed_work users depend
1488 * on that there's no such delay when @delay is 0.
1491 __queue_work(cpu
, wq
, &dwork
->work
);
1495 timer_stats_timer_set_start_info(&dwork
->timer
);
1499 timer
->expires
= jiffies
+ delay
;
1501 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1502 add_timer_on(timer
, cpu
);
1508 * queue_delayed_work_on - queue work on specific CPU after delay
1509 * @cpu: CPU number to execute work on
1510 * @wq: workqueue to use
1511 * @dwork: work to queue
1512 * @delay: number of jiffies to wait before queueing
1514 * Returns %false if @work was already on a queue, %true otherwise. If
1515 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1518 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1519 struct delayed_work
*dwork
, unsigned long delay
)
1521 struct work_struct
*work
= &dwork
->work
;
1523 unsigned long flags
;
1525 /* read the comment in __queue_work() */
1526 local_irq_save(flags
);
1528 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1529 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1533 local_irq_restore(flags
);
1536 EXPORT_SYMBOL(queue_delayed_work_on
);
1539 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1540 * @cpu: CPU number to execute work on
1541 * @wq: workqueue to use
1542 * @dwork: work to queue
1543 * @delay: number of jiffies to wait before queueing
1545 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1546 * modify @dwork's timer so that it expires after @delay. If @delay is
1547 * zero, @work is guaranteed to be scheduled immediately regardless of its
1550 * Returns %false if @dwork was idle and queued, %true if @dwork was
1551 * pending and its timer was modified.
1553 * This function is safe to call from any context including IRQ handler.
1554 * See try_to_grab_pending() for details.
1556 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1557 struct delayed_work
*dwork
, unsigned long delay
)
1559 unsigned long flags
;
1563 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1564 } while (unlikely(ret
== -EAGAIN
));
1566 if (likely(ret
>= 0)) {
1567 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1568 local_irq_restore(flags
);
1571 /* -ENOENT from try_to_grab_pending() becomes %true */
1574 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1577 * worker_enter_idle - enter idle state
1578 * @worker: worker which is entering idle state
1580 * @worker is entering idle state. Update stats and idle timer if
1584 * spin_lock_irq(pool->lock).
1586 static void worker_enter_idle(struct worker
*worker
)
1588 struct worker_pool
*pool
= worker
->pool
;
1590 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1591 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1592 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1595 /* can't use worker_set_flags(), also called from start_worker() */
1596 worker
->flags
|= WORKER_IDLE
;
1598 worker
->last_active
= jiffies
;
1600 /* idle_list is LIFO */
1601 list_add(&worker
->entry
, &pool
->idle_list
);
1603 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1604 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1607 * Sanity check nr_running. Because wq_unbind_fn() releases
1608 * pool->lock between setting %WORKER_UNBOUND and zapping
1609 * nr_running, the warning may trigger spuriously. Check iff
1610 * unbind is not in progress.
1612 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1613 pool
->nr_workers
== pool
->nr_idle
&&
1614 atomic_read(&pool
->nr_running
));
1618 * worker_leave_idle - leave idle state
1619 * @worker: worker which is leaving idle state
1621 * @worker is leaving idle state. Update stats.
1624 * spin_lock_irq(pool->lock).
1626 static void worker_leave_idle(struct worker
*worker
)
1628 struct worker_pool
*pool
= worker
->pool
;
1630 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1632 worker_clr_flags(worker
, WORKER_IDLE
);
1634 list_del_init(&worker
->entry
);
1638 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1639 * @pool: target worker_pool
1641 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1643 * Works which are scheduled while the cpu is online must at least be
1644 * scheduled to a worker which is bound to the cpu so that if they are
1645 * flushed from cpu callbacks while cpu is going down, they are
1646 * guaranteed to execute on the cpu.
1648 * This function is to be used by unbound workers and rescuers to bind
1649 * themselves to the target cpu and may race with cpu going down or
1650 * coming online. kthread_bind() can't be used because it may put the
1651 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1652 * verbatim as it's best effort and blocking and pool may be
1653 * [dis]associated in the meantime.
1655 * This function tries set_cpus_allowed() and locks pool and verifies the
1656 * binding against %POOL_DISASSOCIATED which is set during
1657 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1658 * enters idle state or fetches works without dropping lock, it can
1659 * guarantee the scheduling requirement described in the first paragraph.
1662 * Might sleep. Called without any lock but returns with pool->lock
1666 * %true if the associated pool is online (@worker is successfully
1667 * bound), %false if offline.
1669 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1670 __acquires(&pool
->lock
)
1674 * The following call may fail, succeed or succeed
1675 * without actually migrating the task to the cpu if
1676 * it races with cpu hotunplug operation. Verify
1677 * against POOL_DISASSOCIATED.
1679 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1680 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1682 spin_lock_irq(&pool
->lock
);
1683 if (pool
->flags
& POOL_DISASSOCIATED
)
1685 if (task_cpu(current
) == pool
->cpu
&&
1686 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1688 spin_unlock_irq(&pool
->lock
);
1691 * We've raced with CPU hot[un]plug. Give it a breather
1692 * and retry migration. cond_resched() is required here;
1693 * otherwise, we might deadlock against cpu_stop trying to
1694 * bring down the CPU on non-preemptive kernel.
1701 static struct worker
*alloc_worker(void)
1703 struct worker
*worker
;
1705 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1707 INIT_LIST_HEAD(&worker
->entry
);
1708 INIT_LIST_HEAD(&worker
->scheduled
);
1709 /* on creation a worker is in !idle && prep state */
1710 worker
->flags
= WORKER_PREP
;
1716 * create_worker - create a new workqueue worker
1717 * @pool: pool the new worker will belong to
1719 * Create a new worker which is bound to @pool. The returned worker
1720 * can be started by calling start_worker() or destroyed using
1724 * Might sleep. Does GFP_KERNEL allocations.
1727 * Pointer to the newly created worker.
1729 static struct worker
*create_worker(struct worker_pool
*pool
)
1731 struct worker
*worker
= NULL
;
1735 lockdep_assert_held(&pool
->manager_mutex
);
1738 * ID is needed to determine kthread name. Allocate ID first
1739 * without installing the pointer.
1741 idr_preload(GFP_KERNEL
);
1742 spin_lock_irq(&pool
->lock
);
1744 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1746 spin_unlock_irq(&pool
->lock
);
1751 worker
= alloc_worker();
1755 worker
->pool
= pool
;
1759 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1760 pool
->attrs
->nice
< 0 ? "H" : "");
1762 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1764 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1765 "kworker/%s", id_buf
);
1766 if (IS_ERR(worker
->task
))
1770 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1771 * online CPUs. It'll be re-applied when any of the CPUs come up.
1773 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1774 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1776 /* prevent userland from meddling with cpumask of workqueue workers */
1777 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1780 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1781 * remains stable across this function. See the comments above the
1782 * flag definition for details.
1784 if (pool
->flags
& POOL_DISASSOCIATED
)
1785 worker
->flags
|= WORKER_UNBOUND
;
1787 /* successful, commit the pointer to idr */
1788 spin_lock_irq(&pool
->lock
);
1789 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1790 spin_unlock_irq(&pool
->lock
);
1796 spin_lock_irq(&pool
->lock
);
1797 idr_remove(&pool
->worker_idr
, id
);
1798 spin_unlock_irq(&pool
->lock
);
1805 * start_worker - start a newly created worker
1806 * @worker: worker to start
1808 * Make the pool aware of @worker and start it.
1811 * spin_lock_irq(pool->lock).
1813 static void start_worker(struct worker
*worker
)
1815 worker
->flags
|= WORKER_STARTED
;
1816 worker
->pool
->nr_workers
++;
1817 worker_enter_idle(worker
);
1818 wake_up_process(worker
->task
);
1822 * create_and_start_worker - create and start a worker for a pool
1823 * @pool: the target pool
1825 * Grab the managership of @pool and create and start a new worker for it.
1827 static int create_and_start_worker(struct worker_pool
*pool
)
1829 struct worker
*worker
;
1831 mutex_lock(&pool
->manager_mutex
);
1833 worker
= create_worker(pool
);
1835 spin_lock_irq(&pool
->lock
);
1836 start_worker(worker
);
1837 spin_unlock_irq(&pool
->lock
);
1840 mutex_unlock(&pool
->manager_mutex
);
1842 return worker
? 0 : -ENOMEM
;
1846 * destroy_worker - destroy a workqueue worker
1847 * @worker: worker to be destroyed
1849 * Destroy @worker and adjust @pool stats accordingly.
1852 * spin_lock_irq(pool->lock) which is released and regrabbed.
1854 static void destroy_worker(struct worker
*worker
)
1856 struct worker_pool
*pool
= worker
->pool
;
1858 lockdep_assert_held(&pool
->manager_mutex
);
1859 lockdep_assert_held(&pool
->lock
);
1861 /* sanity check frenzy */
1862 if (WARN_ON(worker
->current_work
) ||
1863 WARN_ON(!list_empty(&worker
->scheduled
)))
1866 if (worker
->flags
& WORKER_STARTED
)
1868 if (worker
->flags
& WORKER_IDLE
)
1872 * Once WORKER_DIE is set, the kworker may destroy itself at any
1873 * point. Pin to ensure the task stays until we're done with it.
1875 get_task_struct(worker
->task
);
1877 list_del_init(&worker
->entry
);
1878 worker
->flags
|= WORKER_DIE
;
1880 idr_remove(&pool
->worker_idr
, worker
->id
);
1882 spin_unlock_irq(&pool
->lock
);
1884 kthread_stop(worker
->task
);
1885 put_task_struct(worker
->task
);
1888 spin_lock_irq(&pool
->lock
);
1891 static void idle_worker_timeout(unsigned long __pool
)
1893 struct worker_pool
*pool
= (void *)__pool
;
1895 spin_lock_irq(&pool
->lock
);
1897 if (too_many_workers(pool
)) {
1898 struct worker
*worker
;
1899 unsigned long expires
;
1901 /* idle_list is kept in LIFO order, check the last one */
1902 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1903 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1905 if (time_before(jiffies
, expires
))
1906 mod_timer(&pool
->idle_timer
, expires
);
1908 /* it's been idle for too long, wake up manager */
1909 pool
->flags
|= POOL_MANAGE_WORKERS
;
1910 wake_up_worker(pool
);
1914 spin_unlock_irq(&pool
->lock
);
1917 static void send_mayday(struct work_struct
*work
)
1919 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1920 struct workqueue_struct
*wq
= pwq
->wq
;
1922 lockdep_assert_held(&wq_mayday_lock
);
1927 /* mayday mayday mayday */
1928 if (list_empty(&pwq
->mayday_node
)) {
1930 * If @pwq is for an unbound wq, its base ref may be put at
1931 * any time due to an attribute change. Pin @pwq until the
1932 * rescuer is done with it.
1935 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1936 wake_up_process(wq
->rescuer
->task
);
1940 static void pool_mayday_timeout(unsigned long __pool
)
1942 struct worker_pool
*pool
= (void *)__pool
;
1943 struct work_struct
*work
;
1945 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1946 spin_lock(&pool
->lock
);
1948 if (need_to_create_worker(pool
)) {
1950 * We've been trying to create a new worker but
1951 * haven't been successful. We might be hitting an
1952 * allocation deadlock. Send distress signals to
1955 list_for_each_entry(work
, &pool
->worklist
, entry
)
1959 spin_unlock(&pool
->lock
);
1960 spin_unlock_irq(&wq_mayday_lock
);
1962 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1966 * maybe_create_worker - create a new worker if necessary
1967 * @pool: pool to create a new worker for
1969 * Create a new worker for @pool if necessary. @pool is guaranteed to
1970 * have at least one idle worker on return from this function. If
1971 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1972 * sent to all rescuers with works scheduled on @pool to resolve
1973 * possible allocation deadlock.
1975 * On return, need_to_create_worker() is guaranteed to be %false and
1976 * may_start_working() %true.
1979 * spin_lock_irq(pool->lock) which may be released and regrabbed
1980 * multiple times. Does GFP_KERNEL allocations. Called only from
1983 static void maybe_create_worker(struct worker_pool
*pool
)
1984 __releases(&pool
->lock
)
1985 __acquires(&pool
->lock
)
1987 if (!need_to_create_worker(pool
))
1990 spin_unlock_irq(&pool
->lock
);
1992 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1993 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1996 struct worker
*worker
;
1998 worker
= create_worker(pool
);
2000 del_timer_sync(&pool
->mayday_timer
);
2001 spin_lock_irq(&pool
->lock
);
2002 start_worker(worker
);
2003 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
2008 if (!need_to_create_worker(pool
))
2011 __set_current_state(TASK_INTERRUPTIBLE
);
2012 schedule_timeout(CREATE_COOLDOWN
);
2014 if (!need_to_create_worker(pool
))
2018 del_timer_sync(&pool
->mayday_timer
);
2019 spin_lock_irq(&pool
->lock
);
2020 if (need_to_create_worker(pool
))
2026 * maybe_destroy_worker - destroy workers which have been idle for a while
2027 * @pool: pool to destroy workers for
2029 * Destroy @pool workers which have been idle for longer than
2030 * IDLE_WORKER_TIMEOUT.
2033 * spin_lock_irq(pool->lock) which may be released and regrabbed
2034 * multiple times. Called only from manager.
2036 static void maybe_destroy_workers(struct worker_pool
*pool
)
2038 while (too_many_workers(pool
)) {
2039 struct worker
*worker
;
2040 unsigned long expires
;
2042 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2043 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2045 if (time_before(jiffies
, expires
)) {
2046 mod_timer(&pool
->idle_timer
, expires
);
2050 destroy_worker(worker
);
2055 * manage_workers - manage worker pool
2058 * Assume the manager role and manage the worker pool @worker belongs
2059 * to. At any given time, there can be only zero or one manager per
2060 * pool. The exclusion is handled automatically by this function.
2062 * The caller can safely start processing works on false return. On
2063 * true return, it's guaranteed that need_to_create_worker() is false
2064 * and may_start_working() is true.
2067 * spin_lock_irq(pool->lock) which may be released and regrabbed
2068 * multiple times. Does GFP_KERNEL allocations.
2071 * %false if the pool doesn't need management and the caller can safely
2072 * start processing works, %true if management function was performed and
2073 * the conditions that the caller verified before calling the function may
2074 * no longer be true.
2076 static bool manage_workers(struct worker
*worker
)
2078 struct worker_pool
*pool
= worker
->pool
;
2081 * Managership is governed by two mutexes - manager_arb and
2082 * manager_mutex. manager_arb handles arbitration of manager role.
2083 * Anyone who successfully grabs manager_arb wins the arbitration
2084 * and becomes the manager. mutex_trylock() on pool->manager_arb
2085 * failure while holding pool->lock reliably indicates that someone
2086 * else is managing the pool and the worker which failed trylock
2087 * can proceed to executing work items. This means that anyone
2088 * grabbing manager_arb is responsible for actually performing
2089 * manager duties. If manager_arb is grabbed and released without
2090 * actual management, the pool may stall indefinitely.
2092 * manager_mutex is used for exclusion of actual management
2093 * operations. The holder of manager_mutex can be sure that none
2094 * of management operations, including creation and destruction of
2095 * workers, won't take place until the mutex is released. Because
2096 * manager_mutex doesn't interfere with manager role arbitration,
2097 * it is guaranteed that the pool's management, while may be
2098 * delayed, won't be disturbed by someone else grabbing
2101 if (!mutex_trylock(&pool
->manager_arb
))
2105 * With manager arbitration won, manager_mutex would be free in
2106 * most cases. trylock first without dropping @pool->lock.
2108 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2109 spin_unlock_irq(&pool
->lock
);
2110 mutex_lock(&pool
->manager_mutex
);
2111 spin_lock_irq(&pool
->lock
);
2114 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2117 * Destroy and then create so that may_start_working() is true
2120 maybe_destroy_workers(pool
);
2121 maybe_create_worker(pool
);
2123 mutex_unlock(&pool
->manager_mutex
);
2124 mutex_unlock(&pool
->manager_arb
);
2129 * process_one_work - process single work
2131 * @work: work to process
2133 * Process @work. This function contains all the logics necessary to
2134 * process a single work including synchronization against and
2135 * interaction with other workers on the same cpu, queueing and
2136 * flushing. As long as context requirement is met, any worker can
2137 * call this function to process a work.
2140 * spin_lock_irq(pool->lock) which is released and regrabbed.
2142 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2143 __releases(&pool
->lock
)
2144 __acquires(&pool
->lock
)
2146 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2147 struct worker_pool
*pool
= worker
->pool
;
2148 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2150 struct worker
*collision
;
2151 unsigned long long exec_start
;
2154 #ifdef CONFIG_LOCKDEP
2156 * It is permissible to free the struct work_struct from
2157 * inside the function that is called from it, this we need to
2158 * take into account for lockdep too. To avoid bogus "held
2159 * lock freed" warnings as well as problems when looking into
2160 * work->lockdep_map, make a copy and use that here.
2162 struct lockdep_map lockdep_map
;
2164 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2167 * Ensure we're on the correct CPU. DISASSOCIATED test is
2168 * necessary to avoid spurious warnings from rescuers servicing the
2169 * unbound or a disassociated pool.
2171 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2172 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2173 raw_smp_processor_id() != pool
->cpu
);
2176 * A single work shouldn't be executed concurrently by
2177 * multiple workers on a single cpu. Check whether anyone is
2178 * already processing the work. If so, defer the work to the
2179 * currently executing one.
2181 collision
= find_worker_executing_work(pool
, work
);
2182 if (unlikely(collision
)) {
2183 move_linked_works(work
, &collision
->scheduled
, NULL
);
2187 /* claim and dequeue */
2188 debug_work_deactivate(work
);
2189 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2190 worker
->current_work
= work
;
2191 worker
->current_func
= work
->func
;
2192 worker
->current_pwq
= pwq
;
2193 work_color
= get_work_color(work
);
2195 list_del_init(&work
->entry
);
2198 * CPU intensive works don't participate in concurrency
2199 * management. They're the scheduler's responsibility.
2201 if (unlikely(cpu_intensive
))
2202 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2205 * Unbound pool isn't concurrency managed and work items should be
2206 * executed ASAP. Wake up another worker if necessary.
2208 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2209 wake_up_worker(pool
);
2212 * Record the last pool and clear PENDING which should be the last
2213 * update to @work. Also, do this inside @pool->lock so that
2214 * PENDING and queued state changes happen together while IRQ is
2217 set_work_pool_and_clear_pending(work
, pool
->id
);
2219 spin_unlock_irq(&pool
->lock
);
2221 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2222 lock_map_acquire(&lockdep_map
);
2224 exec_start
= sched_clock();
2225 sprintf(func
, "%pf", work
->func
);
2227 trace_workqueue_execute_start(work
);
2228 #ifdef CONFIG_MTK_WQ_DEBUG
2229 mttrace_workqueue_execute_work(work
);
2230 #endif //CONFIG_MTK_WQ_DEBUG
2232 worker
->current_func(work
);
2235 * While we must be careful to not use "work" after this, the trace
2236 * point will only record its address.
2238 trace_workqueue_execute_end(work
);
2239 #ifdef CONFIG_MTK_WQ_DEBUG
2240 mttrace_workqueue_execute_end(work
);
2241 #endif //CONFIG_MTK_WQ_DEBUG
2243 if ((sched_clock() - exec_start
)> 1000000000) // dump log if execute more than 1 sec
2244 pr_warning("WQ warning! work (%s, %p) execute more than 1 sec, time: %llu ns\n", func
, work
, sched_clock() - exec_start
);
2246 lock_map_release(&lockdep_map
);
2247 lock_map_release(&pwq
->wq
->lockdep_map
);
2249 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2250 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2251 " last function: %pf\n",
2252 current
->comm
, preempt_count(), task_pid_nr(current
),
2253 worker
->current_func
);
2254 debug_show_held_locks(current
);
2259 * The following prevents a kworker from hogging CPU on !PREEMPT
2260 * kernels, where a requeueing work item waiting for something to
2261 * happen could deadlock with stop_machine as such work item could
2262 * indefinitely requeue itself while all other CPUs are trapped in
2267 spin_lock_irq(&pool
->lock
);
2269 /* clear cpu intensive status */
2270 if (unlikely(cpu_intensive
))
2271 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2273 /* we're done with it, release */
2274 hash_del(&worker
->hentry
);
2275 worker
->current_work
= NULL
;
2276 worker
->current_func
= NULL
;
2277 worker
->current_pwq
= NULL
;
2278 worker
->desc_valid
= false;
2279 pwq_dec_nr_in_flight(pwq
, work_color
);
2283 * process_scheduled_works - process scheduled works
2286 * Process all scheduled works. Please note that the scheduled list
2287 * may change while processing a work, so this function repeatedly
2288 * fetches a work from the top and executes it.
2291 * spin_lock_irq(pool->lock) which may be released and regrabbed
2294 static void process_scheduled_works(struct worker
*worker
)
2296 while (!list_empty(&worker
->scheduled
)) {
2297 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2298 struct work_struct
, entry
);
2299 process_one_work(worker
, work
);
2304 * worker_thread - the worker thread function
2307 * The worker thread function. All workers belong to a worker_pool -
2308 * either a per-cpu one or dynamic unbound one. These workers process all
2309 * work items regardless of their specific target workqueue. The only
2310 * exception is work items which belong to workqueues with a rescuer which
2311 * will be explained in rescuer_thread().
2313 static int worker_thread(void *__worker
)
2315 struct worker
*worker
= __worker
;
2316 struct worker_pool
*pool
= worker
->pool
;
2318 /* tell the scheduler that this is a workqueue worker */
2319 worker
->task
->flags
|= PF_WQ_WORKER
;
2321 spin_lock_irq(&pool
->lock
);
2323 /* am I supposed to die? */
2324 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2325 spin_unlock_irq(&pool
->lock
);
2326 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2327 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2331 worker_leave_idle(worker
);
2333 /* no more worker necessary? */
2334 if (!need_more_worker(pool
))
2337 /* do we need to manage? */
2338 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2342 * ->scheduled list can only be filled while a worker is
2343 * preparing to process a work or actually processing it.
2344 * Make sure nobody diddled with it while I was sleeping.
2346 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2349 * Finish PREP stage. We're guaranteed to have at least one idle
2350 * worker or that someone else has already assumed the manager
2351 * role. This is where @worker starts participating in concurrency
2352 * management if applicable and concurrency management is restored
2353 * after being rebound. See rebind_workers() for details.
2355 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2358 struct work_struct
*work
=
2359 list_first_entry(&pool
->worklist
,
2360 struct work_struct
, entry
);
2362 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2363 /* optimization path, not strictly necessary */
2364 process_one_work(worker
, work
);
2365 if (unlikely(!list_empty(&worker
->scheduled
)))
2366 process_scheduled_works(worker
);
2368 move_linked_works(work
, &worker
->scheduled
, NULL
);
2369 process_scheduled_works(worker
);
2371 } while (keep_working(pool
));
2373 worker_set_flags(worker
, WORKER_PREP
, false);
2375 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2379 * pool->lock is held and there's no work to process and no need to
2380 * manage, sleep. Workers are woken up only while holding
2381 * pool->lock or from local cpu, so setting the current state
2382 * before releasing pool->lock is enough to prevent losing any
2385 worker_enter_idle(worker
);
2386 __set_current_state(TASK_INTERRUPTIBLE
);
2387 spin_unlock_irq(&pool
->lock
);
2393 * rescuer_thread - the rescuer thread function
2396 * Workqueue rescuer thread function. There's one rescuer for each
2397 * workqueue which has WQ_MEM_RECLAIM set.
2399 * Regular work processing on a pool may block trying to create a new
2400 * worker which uses GFP_KERNEL allocation which has slight chance of
2401 * developing into deadlock if some works currently on the same queue
2402 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2403 * the problem rescuer solves.
2405 * When such condition is possible, the pool summons rescuers of all
2406 * workqueues which have works queued on the pool and let them process
2407 * those works so that forward progress can be guaranteed.
2409 * This should happen rarely.
2411 static int rescuer_thread(void *__rescuer
)
2413 struct worker
*rescuer
= __rescuer
;
2414 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2415 struct list_head
*scheduled
= &rescuer
->scheduled
;
2418 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2421 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2422 * doesn't participate in concurrency management.
2424 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2426 set_current_state(TASK_INTERRUPTIBLE
);
2429 * By the time the rescuer is requested to stop, the workqueue
2430 * shouldn't have any work pending, but @wq->maydays may still have
2431 * pwq(s) queued. This can happen by non-rescuer workers consuming
2432 * all the work items before the rescuer got to them. Go through
2433 * @wq->maydays processing before acting on should_stop so that the
2434 * list is always empty on exit.
2436 should_stop
= kthread_should_stop();
2438 /* see whether any pwq is asking for help */
2439 spin_lock_irq(&wq_mayday_lock
);
2441 while (!list_empty(&wq
->maydays
)) {
2442 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2443 struct pool_workqueue
, mayday_node
);
2444 struct worker_pool
*pool
= pwq
->pool
;
2445 struct work_struct
*work
, *n
;
2447 __set_current_state(TASK_RUNNING
);
2448 list_del_init(&pwq
->mayday_node
);
2450 spin_unlock_irq(&wq_mayday_lock
);
2452 /* migrate to the target cpu if possible */
2453 worker_maybe_bind_and_lock(pool
);
2454 rescuer
->pool
= pool
;
2457 * Slurp in all works issued via this workqueue and
2460 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2461 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2462 if (get_work_pwq(work
) == pwq
)
2463 move_linked_works(work
, scheduled
, &n
);
2465 process_scheduled_works(rescuer
);
2468 * Put the reference grabbed by send_mayday(). @pool won't
2469 * go away while we're holding its lock.
2474 * Leave this pool. If keep_working() is %true, notify a
2475 * regular worker; otherwise, we end up with 0 concurrency
2476 * and stalling the execution.
2478 if (keep_working(pool
))
2479 wake_up_worker(pool
);
2481 rescuer
->pool
= NULL
;
2482 spin_unlock(&pool
->lock
);
2483 spin_lock(&wq_mayday_lock
);
2486 spin_unlock_irq(&wq_mayday_lock
);
2489 __set_current_state(TASK_RUNNING
);
2490 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2494 /* rescuers should never participate in concurrency management */
2495 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2501 struct work_struct work
;
2502 struct completion done
;
2505 static void wq_barrier_func(struct work_struct
*work
)
2507 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2508 complete(&barr
->done
);
2512 * insert_wq_barrier - insert a barrier work
2513 * @pwq: pwq to insert barrier into
2514 * @barr: wq_barrier to insert
2515 * @target: target work to attach @barr to
2516 * @worker: worker currently executing @target, NULL if @target is not executing
2518 * @barr is linked to @target such that @barr is completed only after
2519 * @target finishes execution. Please note that the ordering
2520 * guarantee is observed only with respect to @target and on the local
2523 * Currently, a queued barrier can't be canceled. This is because
2524 * try_to_grab_pending() can't determine whether the work to be
2525 * grabbed is at the head of the queue and thus can't clear LINKED
2526 * flag of the previous work while there must be a valid next work
2527 * after a work with LINKED flag set.
2529 * Note that when @worker is non-NULL, @target may be modified
2530 * underneath us, so we can't reliably determine pwq from @target.
2533 * spin_lock_irq(pool->lock).
2535 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2536 struct wq_barrier
*barr
,
2537 struct work_struct
*target
, struct worker
*worker
)
2539 struct list_head
*head
;
2540 unsigned int linked
= 0;
2543 * debugobject calls are safe here even with pool->lock locked
2544 * as we know for sure that this will not trigger any of the
2545 * checks and call back into the fixup functions where we
2548 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2549 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2550 init_completion(&barr
->done
);
2553 * If @target is currently being executed, schedule the
2554 * barrier to the worker; otherwise, put it after @target.
2557 head
= worker
->scheduled
.next
;
2559 unsigned long *bits
= work_data_bits(target
);
2561 head
= target
->entry
.next
;
2562 /* there can already be other linked works, inherit and set */
2563 linked
= *bits
& WORK_STRUCT_LINKED
;
2564 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2567 debug_work_activate(&barr
->work
);
2568 insert_work(pwq
, &barr
->work
, head
,
2569 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2573 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2574 * @wq: workqueue being flushed
2575 * @flush_color: new flush color, < 0 for no-op
2576 * @work_color: new work color, < 0 for no-op
2578 * Prepare pwqs for workqueue flushing.
2580 * If @flush_color is non-negative, flush_color on all pwqs should be
2581 * -1. If no pwq has in-flight commands at the specified color, all
2582 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2583 * has in flight commands, its pwq->flush_color is set to
2584 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2585 * wakeup logic is armed and %true is returned.
2587 * The caller should have initialized @wq->first_flusher prior to
2588 * calling this function with non-negative @flush_color. If
2589 * @flush_color is negative, no flush color update is done and %false
2592 * If @work_color is non-negative, all pwqs should have the same
2593 * work_color which is previous to @work_color and all will be
2594 * advanced to @work_color.
2597 * mutex_lock(wq->mutex).
2600 * %true if @flush_color >= 0 and there's something to flush. %false
2603 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2604 int flush_color
, int work_color
)
2607 struct pool_workqueue
*pwq
;
2609 if (flush_color
>= 0) {
2610 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2611 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2614 for_each_pwq(pwq
, wq
) {
2615 struct worker_pool
*pool
= pwq
->pool
;
2617 spin_lock_irq(&pool
->lock
);
2619 if (flush_color
>= 0) {
2620 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2622 if (pwq
->nr_in_flight
[flush_color
]) {
2623 pwq
->flush_color
= flush_color
;
2624 atomic_inc(&wq
->nr_pwqs_to_flush
);
2629 if (work_color
>= 0) {
2630 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2631 pwq
->work_color
= work_color
;
2634 spin_unlock_irq(&pool
->lock
);
2637 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2638 complete(&wq
->first_flusher
->done
);
2644 * flush_workqueue - ensure that any scheduled work has run to completion.
2645 * @wq: workqueue to flush
2647 * This function sleeps until all work items which were queued on entry
2648 * have finished execution, but it is not livelocked by new incoming ones.
2650 void flush_workqueue(struct workqueue_struct
*wq
)
2652 struct wq_flusher this_flusher
= {
2653 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2655 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2659 lock_map_acquire(&wq
->lockdep_map
);
2660 lock_map_release(&wq
->lockdep_map
);
2662 mutex_lock(&wq
->mutex
);
2665 * Start-to-wait phase
2667 next_color
= work_next_color(wq
->work_color
);
2669 if (next_color
!= wq
->flush_color
) {
2671 * Color space is not full. The current work_color
2672 * becomes our flush_color and work_color is advanced
2675 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2676 this_flusher
.flush_color
= wq
->work_color
;
2677 wq
->work_color
= next_color
;
2679 if (!wq
->first_flusher
) {
2680 /* no flush in progress, become the first flusher */
2681 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2683 wq
->first_flusher
= &this_flusher
;
2685 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2687 /* nothing to flush, done */
2688 wq
->flush_color
= next_color
;
2689 wq
->first_flusher
= NULL
;
2694 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2695 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2696 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2700 * Oops, color space is full, wait on overflow queue.
2701 * The next flush completion will assign us
2702 * flush_color and transfer to flusher_queue.
2704 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2707 mutex_unlock(&wq
->mutex
);
2709 wait_for_completion(&this_flusher
.done
);
2712 * Wake-up-and-cascade phase
2714 * First flushers are responsible for cascading flushes and
2715 * handling overflow. Non-first flushers can simply return.
2717 if (wq
->first_flusher
!= &this_flusher
)
2720 mutex_lock(&wq
->mutex
);
2722 /* we might have raced, check again with mutex held */
2723 if (wq
->first_flusher
!= &this_flusher
)
2726 wq
->first_flusher
= NULL
;
2728 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2729 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2732 struct wq_flusher
*next
, *tmp
;
2734 /* complete all the flushers sharing the current flush color */
2735 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2736 if (next
->flush_color
!= wq
->flush_color
)
2738 list_del_init(&next
->list
);
2739 complete(&next
->done
);
2742 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2743 wq
->flush_color
!= work_next_color(wq
->work_color
));
2745 /* this flush_color is finished, advance by one */
2746 wq
->flush_color
= work_next_color(wq
->flush_color
);
2748 /* one color has been freed, handle overflow queue */
2749 if (!list_empty(&wq
->flusher_overflow
)) {
2751 * Assign the same color to all overflowed
2752 * flushers, advance work_color and append to
2753 * flusher_queue. This is the start-to-wait
2754 * phase for these overflowed flushers.
2756 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2757 tmp
->flush_color
= wq
->work_color
;
2759 wq
->work_color
= work_next_color(wq
->work_color
);
2761 list_splice_tail_init(&wq
->flusher_overflow
,
2762 &wq
->flusher_queue
);
2763 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2766 if (list_empty(&wq
->flusher_queue
)) {
2767 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2772 * Need to flush more colors. Make the next flusher
2773 * the new first flusher and arm pwqs.
2775 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2776 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2778 list_del_init(&next
->list
);
2779 wq
->first_flusher
= next
;
2781 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2785 * Meh... this color is already done, clear first
2786 * flusher and repeat cascading.
2788 wq
->first_flusher
= NULL
;
2792 mutex_unlock(&wq
->mutex
);
2794 EXPORT_SYMBOL_GPL(flush_workqueue
);
2797 * drain_workqueue - drain a workqueue
2798 * @wq: workqueue to drain
2800 * Wait until the workqueue becomes empty. While draining is in progress,
2801 * only chain queueing is allowed. IOW, only currently pending or running
2802 * work items on @wq can queue further work items on it. @wq is flushed
2803 * repeatedly until it becomes empty. The number of flushing is detemined
2804 * by the depth of chaining and should be relatively short. Whine if it
2807 void drain_workqueue(struct workqueue_struct
*wq
)
2809 unsigned int flush_cnt
= 0;
2810 struct pool_workqueue
*pwq
;
2813 * __queue_work() needs to test whether there are drainers, is much
2814 * hotter than drain_workqueue() and already looks at @wq->flags.
2815 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2817 mutex_lock(&wq
->mutex
);
2818 if (!wq
->nr_drainers
++)
2819 wq
->flags
|= __WQ_DRAINING
;
2820 mutex_unlock(&wq
->mutex
);
2822 flush_workqueue(wq
);
2824 mutex_lock(&wq
->mutex
);
2826 for_each_pwq(pwq
, wq
) {
2829 spin_lock_irq(&pwq
->pool
->lock
);
2830 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2831 spin_unlock_irq(&pwq
->pool
->lock
);
2836 if (++flush_cnt
== 10 ||
2837 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2838 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2839 wq
->name
, flush_cnt
);
2841 mutex_unlock(&wq
->mutex
);
2845 if (!--wq
->nr_drainers
)
2846 wq
->flags
&= ~__WQ_DRAINING
;
2847 mutex_unlock(&wq
->mutex
);
2849 EXPORT_SYMBOL_GPL(drain_workqueue
);
2851 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2853 struct worker
*worker
= NULL
;
2854 struct worker_pool
*pool
;
2855 struct pool_workqueue
*pwq
;
2859 local_irq_disable();
2860 pool
= get_work_pool(work
);
2866 spin_lock(&pool
->lock
);
2867 /* see the comment in try_to_grab_pending() with the same code */
2868 pwq
= get_work_pwq(work
);
2870 if (unlikely(pwq
->pool
!= pool
))
2873 worker
= find_worker_executing_work(pool
, work
);
2876 pwq
= worker
->current_pwq
;
2879 insert_wq_barrier(pwq
, barr
, work
, worker
);
2880 spin_unlock_irq(&pool
->lock
);
2883 * If @max_active is 1 or rescuer is in use, flushing another work
2884 * item on the same workqueue may lead to deadlock. Make sure the
2885 * flusher is not running on the same workqueue by verifying write
2888 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2889 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2891 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2892 lock_map_release(&pwq
->wq
->lockdep_map
);
2896 spin_unlock_irq(&pool
->lock
);
2901 * flush_work - wait for a work to finish executing the last queueing instance
2902 * @work: the work to flush
2904 * Wait until @work has finished execution. @work is guaranteed to be idle
2905 * on return if it hasn't been requeued since flush started.
2908 * %true if flush_work() waited for the work to finish execution,
2909 * %false if it was already idle.
2911 bool flush_work(struct work_struct
*work
)
2913 struct wq_barrier barr
;
2915 lock_map_acquire(&work
->lockdep_map
);
2916 lock_map_release(&work
->lockdep_map
);
2918 if (start_flush_work(work
, &barr
)) {
2919 wait_for_completion(&barr
.done
);
2920 destroy_work_on_stack(&barr
.work
);
2926 EXPORT_SYMBOL_GPL(flush_work
);
2930 struct work_struct
*work
;
2933 static int cwt_wakefn(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
2935 struct cwt_wait
*cwait
= container_of(wait
, struct cwt_wait
, wait
);
2937 if (cwait
->work
!= key
)
2939 return autoremove_wake_function(wait
, mode
, sync
, key
);
2942 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2944 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq
);
2945 unsigned long flags
;
2949 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2951 * If someone else is already canceling, wait for it to
2952 * finish. flush_work() doesn't work for PREEMPT_NONE
2953 * because we may get scheduled between @work's completion
2954 * and the other canceling task resuming and clearing
2955 * CANCELING - flush_work() will return false immediately
2956 * as @work is no longer busy, try_to_grab_pending() will
2957 * return -ENOENT as @work is still being canceled and the
2958 * other canceling task won't be able to clear CANCELING as
2959 * we're hogging the CPU.
2961 * Let's wait for completion using a waitqueue. As this
2962 * may lead to the thundering herd problem, use a custom
2963 * wake function which matches @work along with exclusive
2966 if (unlikely(ret
== -ENOENT
)) {
2967 struct cwt_wait cwait
;
2969 init_wait(&cwait
.wait
);
2970 cwait
.wait
.func
= cwt_wakefn
;
2973 prepare_to_wait_exclusive(&cancel_waitq
, &cwait
.wait
,
2974 TASK_UNINTERRUPTIBLE
);
2975 if (work_is_canceling(work
))
2977 finish_wait(&cancel_waitq
, &cwait
.wait
);
2979 } while (unlikely(ret
< 0));
2981 /* tell other tasks trying to grab @work to back off */
2982 mark_work_canceling(work
);
2983 local_irq_restore(flags
);
2986 clear_work_data(work
);
2989 * Paired with prepare_to_wait() above so that either
2990 * waitqueue_active() is visible here or !work_is_canceling() is
2994 if (waitqueue_active(&cancel_waitq
))
2995 __wake_up(&cancel_waitq
, TASK_NORMAL
, 1, work
);
3001 * cancel_work_sync - cancel a work and wait for it to finish
3002 * @work: the work to cancel
3004 * Cancel @work and wait for its execution to finish. This function
3005 * can be used even if the work re-queues itself or migrates to
3006 * another workqueue. On return from this function, @work is
3007 * guaranteed to be not pending or executing on any CPU.
3009 * cancel_work_sync(&delayed_work->work) must not be used for
3010 * delayed_work's. Use cancel_delayed_work_sync() instead.
3012 * The caller must ensure that the workqueue on which @work was last
3013 * queued can't be destroyed before this function returns.
3016 * %true if @work was pending, %false otherwise.
3018 bool cancel_work_sync(struct work_struct
*work
)
3020 return __cancel_work_timer(work
, false);
3022 EXPORT_SYMBOL_GPL(cancel_work_sync
);
3025 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3026 * @dwork: the delayed work to flush
3028 * Delayed timer is cancelled and the pending work is queued for
3029 * immediate execution. Like flush_work(), this function only
3030 * considers the last queueing instance of @dwork.
3033 * %true if flush_work() waited for the work to finish execution,
3034 * %false if it was already idle.
3036 bool flush_delayed_work(struct delayed_work
*dwork
)
3038 local_irq_disable();
3039 if (del_timer_sync(&dwork
->timer
))
3040 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
3042 return flush_work(&dwork
->work
);
3044 EXPORT_SYMBOL(flush_delayed_work
);
3047 * cancel_delayed_work - cancel a delayed work
3048 * @dwork: delayed_work to cancel
3050 * Kill off a pending delayed_work. Returns %true if @dwork was pending
3051 * and canceled; %false if wasn't pending. Note that the work callback
3052 * function may still be running on return, unless it returns %true and the
3053 * work doesn't re-arm itself. Explicitly flush or use
3054 * cancel_delayed_work_sync() to wait on it.
3056 * This function is safe to call from any context including IRQ handler.
3058 bool cancel_delayed_work(struct delayed_work
*dwork
)
3060 unsigned long flags
;
3064 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
3065 } while (unlikely(ret
== -EAGAIN
));
3067 if (unlikely(ret
< 0))
3070 set_work_pool_and_clear_pending(&dwork
->work
,
3071 get_work_pool_id(&dwork
->work
));
3072 local_irq_restore(flags
);
3075 EXPORT_SYMBOL(cancel_delayed_work
);
3078 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3079 * @dwork: the delayed work cancel
3081 * This is cancel_work_sync() for delayed works.
3084 * %true if @dwork was pending, %false otherwise.
3086 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3088 return __cancel_work_timer(&dwork
->work
, true);
3090 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3093 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3094 * @func: the function to call
3096 * schedule_on_each_cpu() executes @func on each online CPU using the
3097 * system workqueue and blocks until all CPUs have completed.
3098 * schedule_on_each_cpu() is very slow.
3101 * 0 on success, -errno on failure.
3103 int schedule_on_each_cpu(work_func_t func
)
3106 struct work_struct __percpu
*works
;
3108 works
= alloc_percpu(struct work_struct
);
3114 for_each_online_cpu(cpu
) {
3115 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3117 INIT_WORK(work
, func
);
3118 schedule_work_on(cpu
, work
);
3121 for_each_online_cpu(cpu
)
3122 flush_work(per_cpu_ptr(works
, cpu
));
3130 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3132 * Forces execution of the kernel-global workqueue and blocks until its
3135 * Think twice before calling this function! It's very easy to get into
3136 * trouble if you don't take great care. Either of the following situations
3137 * will lead to deadlock:
3139 * One of the work items currently on the workqueue needs to acquire
3140 * a lock held by your code or its caller.
3142 * Your code is running in the context of a work routine.
3144 * They will be detected by lockdep when they occur, but the first might not
3145 * occur very often. It depends on what work items are on the workqueue and
3146 * what locks they need, which you have no control over.
3148 * In most situations flushing the entire workqueue is overkill; you merely
3149 * need to know that a particular work item isn't queued and isn't running.
3150 * In such cases you should use cancel_delayed_work_sync() or
3151 * cancel_work_sync() instead.
3153 void flush_scheduled_work(void)
3155 flush_workqueue(system_wq
);
3157 EXPORT_SYMBOL(flush_scheduled_work
);
3160 * execute_in_process_context - reliably execute the routine with user context
3161 * @fn: the function to execute
3162 * @ew: guaranteed storage for the execute work structure (must
3163 * be available when the work executes)
3165 * Executes the function immediately if process context is available,
3166 * otherwise schedules the function for delayed execution.
3168 * Returns: 0 - function was executed
3169 * 1 - function was scheduled for execution
3171 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3173 if (!in_interrupt()) {
3178 INIT_WORK(&ew
->work
, fn
);
3179 schedule_work(&ew
->work
);
3183 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3187 * Workqueues with WQ_SYSFS flag set is visible to userland via
3188 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3189 * following attributes.
3191 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3192 * max_active RW int : maximum number of in-flight work items
3194 * Unbound workqueues have the following extra attributes.
3196 * id RO int : the associated pool ID
3197 * nice RW int : nice value of the workers
3198 * cpumask RW mask : bitmask of allowed CPUs for the workers
3201 struct workqueue_struct
*wq
;
3205 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3207 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3212 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3213 struct device_attribute
*attr
, char *buf
)
3215 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3217 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3220 static ssize_t
wq_max_active_show(struct device
*dev
,
3221 struct device_attribute
*attr
, char *buf
)
3223 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3225 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3228 static ssize_t
wq_max_active_store(struct device
*dev
,
3229 struct device_attribute
*attr
,
3230 const char *buf
, size_t count
)
3232 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3235 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3238 workqueue_set_max_active(wq
, val
);
3242 static struct device_attribute wq_sysfs_attrs
[] = {
3243 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3244 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3248 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3249 struct device_attribute
*attr
, char *buf
)
3251 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3252 const char *delim
= "";
3253 int node
, written
= 0;
3255 rcu_read_lock_sched();
3256 for_each_node(node
) {
3257 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3258 "%s%d:%d", delim
, node
,
3259 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3262 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3263 rcu_read_unlock_sched();
3268 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3271 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3274 mutex_lock(&wq
->mutex
);
3275 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3276 mutex_unlock(&wq
->mutex
);
3281 /* prepare workqueue_attrs for sysfs store operations */
3282 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3284 struct workqueue_attrs
*attrs
;
3286 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3290 mutex_lock(&wq
->mutex
);
3291 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3292 mutex_unlock(&wq
->mutex
);
3296 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3297 const char *buf
, size_t count
)
3299 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3300 struct workqueue_attrs
*attrs
;
3303 attrs
= wq_sysfs_prep_attrs(wq
);
3307 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3308 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3309 ret
= apply_workqueue_attrs(wq
, attrs
);
3313 free_workqueue_attrs(attrs
);
3314 return ret
?: count
;
3317 static ssize_t
wq_cpumask_show(struct device
*dev
,
3318 struct device_attribute
*attr
, char *buf
)
3320 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3323 mutex_lock(&wq
->mutex
);
3324 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3325 mutex_unlock(&wq
->mutex
);
3327 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3331 static ssize_t
wq_cpumask_store(struct device
*dev
,
3332 struct device_attribute
*attr
,
3333 const char *buf
, size_t count
)
3335 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3336 struct workqueue_attrs
*attrs
;
3339 attrs
= wq_sysfs_prep_attrs(wq
);
3343 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3345 ret
= apply_workqueue_attrs(wq
, attrs
);
3347 free_workqueue_attrs(attrs
);
3348 return ret
?: count
;
3351 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3354 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3357 mutex_lock(&wq
->mutex
);
3358 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3359 !wq
->unbound_attrs
->no_numa
);
3360 mutex_unlock(&wq
->mutex
);
3365 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3366 const char *buf
, size_t count
)
3368 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3369 struct workqueue_attrs
*attrs
;
3372 attrs
= wq_sysfs_prep_attrs(wq
);
3377 if (sscanf(buf
, "%d", &v
) == 1) {
3378 attrs
->no_numa
= !v
;
3379 ret
= apply_workqueue_attrs(wq
, attrs
);
3382 free_workqueue_attrs(attrs
);
3383 return ret
?: count
;
3386 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3387 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3388 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3389 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3390 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3394 static struct bus_type wq_subsys
= {
3395 .name
= "workqueue",
3396 .dev_attrs
= wq_sysfs_attrs
,
3399 static int __init
wq_sysfs_init(void)
3401 return subsys_virtual_register(&wq_subsys
, NULL
);
3403 core_initcall(wq_sysfs_init
);
3405 static void wq_device_release(struct device
*dev
)
3407 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3413 * workqueue_sysfs_register - make a workqueue visible in sysfs
3414 * @wq: the workqueue to register
3416 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3417 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3418 * which is the preferred method.
3420 * Workqueue user should use this function directly iff it wants to apply
3421 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3422 * apply_workqueue_attrs() may race against userland updating the
3425 * Returns 0 on success, -errno on failure.
3427 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3429 struct wq_device
*wq_dev
;
3433 * Adjusting max_active or creating new pwqs by applyting
3434 * attributes breaks ordering guarantee. Disallow exposing ordered
3437 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
3440 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3445 wq_dev
->dev
.bus
= &wq_subsys
;
3446 wq_dev
->dev
.init_name
= wq
->name
;
3447 wq_dev
->dev
.release
= wq_device_release
;
3450 * unbound_attrs are created separately. Suppress uevent until
3451 * everything is ready.
3453 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3455 ret
= device_register(&wq_dev
->dev
);
3462 if (wq
->flags
& WQ_UNBOUND
) {
3463 struct device_attribute
*attr
;
3465 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3466 ret
= device_create_file(&wq_dev
->dev
, attr
);
3468 device_unregister(&wq_dev
->dev
);
3475 dev_set_uevent_suppress(&wq_dev
->dev
, false);
3476 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3481 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3482 * @wq: the workqueue to unregister
3484 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3486 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3488 struct wq_device
*wq_dev
= wq
->wq_dev
;
3494 device_unregister(&wq_dev
->dev
);
3496 #else /* CONFIG_SYSFS */
3497 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3498 #endif /* CONFIG_SYSFS */
3501 * free_workqueue_attrs - free a workqueue_attrs
3502 * @attrs: workqueue_attrs to free
3504 * Undo alloc_workqueue_attrs().
3506 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3509 free_cpumask_var(attrs
->cpumask
);
3515 * alloc_workqueue_attrs - allocate a workqueue_attrs
3516 * @gfp_mask: allocation mask to use
3518 * Allocate a new workqueue_attrs, initialize with default settings and
3519 * return it. Returns NULL on failure.
3521 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3523 struct workqueue_attrs
*attrs
;
3525 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3528 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3531 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3534 free_workqueue_attrs(attrs
);
3538 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3539 const struct workqueue_attrs
*from
)
3541 to
->nice
= from
->nice
;
3542 cpumask_copy(to
->cpumask
, from
->cpumask
);
3544 * Unlike hash and equality test, this function doesn't ignore
3545 * ->no_numa as it is used for both pool and wq attrs. Instead,
3546 * get_unbound_pool() explicitly clears ->no_numa after copying.
3548 to
->no_numa
= from
->no_numa
;
3551 /* hash value of the content of @attr */
3552 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3556 hash
= jhash_1word(attrs
->nice
, hash
);
3557 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3558 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3562 /* content equality test */
3563 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3564 const struct workqueue_attrs
*b
)
3566 if (a
->nice
!= b
->nice
)
3568 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3574 * init_worker_pool - initialize a newly zalloc'd worker_pool
3575 * @pool: worker_pool to initialize
3577 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3578 * Returns 0 on success, -errno on failure. Even on failure, all fields
3579 * inside @pool proper are initialized and put_unbound_pool() can be called
3580 * on @pool safely to release it.
3582 static int init_worker_pool(struct worker_pool
*pool
)
3584 spin_lock_init(&pool
->lock
);
3587 pool
->node
= NUMA_NO_NODE
;
3588 pool
->flags
|= POOL_DISASSOCIATED
;
3589 INIT_LIST_HEAD(&pool
->worklist
);
3590 INIT_LIST_HEAD(&pool
->idle_list
);
3591 hash_init(pool
->busy_hash
);
3593 init_timer_deferrable(&pool
->idle_timer
);
3594 pool
->idle_timer
.function
= idle_worker_timeout
;
3595 pool
->idle_timer
.data
= (unsigned long)pool
;
3597 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3598 (unsigned long)pool
);
3600 mutex_init(&pool
->manager_arb
);
3601 mutex_init(&pool
->manager_mutex
);
3602 idr_init(&pool
->worker_idr
);
3604 INIT_HLIST_NODE(&pool
->hash_node
);
3607 /* shouldn't fail above this point */
3608 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3614 static void rcu_free_pool(struct rcu_head
*rcu
)
3616 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3618 idr_destroy(&pool
->worker_idr
);
3619 free_workqueue_attrs(pool
->attrs
);
3624 * put_unbound_pool - put a worker_pool
3625 * @pool: worker_pool to put
3627 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3628 * safe manner. get_unbound_pool() calls this function on its failure path
3629 * and this function should be able to release pools which went through,
3630 * successfully or not, init_worker_pool().
3632 * Should be called with wq_pool_mutex held.
3634 static void put_unbound_pool(struct worker_pool
*pool
)
3636 struct worker
*worker
;
3638 lockdep_assert_held(&wq_pool_mutex
);
3644 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3645 WARN_ON(!list_empty(&pool
->worklist
)))
3648 /* release id and unhash */
3650 idr_remove(&worker_pool_idr
, pool
->id
);
3651 hash_del(&pool
->hash_node
);
3654 * Become the manager and destroy all workers. Grabbing
3655 * manager_arb prevents @pool's workers from blocking on
3658 mutex_lock(&pool
->manager_arb
);
3659 mutex_lock(&pool
->manager_mutex
);
3660 spin_lock_irq(&pool
->lock
);
3662 while ((worker
= first_worker(pool
)))
3663 destroy_worker(worker
);
3664 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3666 spin_unlock_irq(&pool
->lock
);
3667 mutex_unlock(&pool
->manager_mutex
);
3668 mutex_unlock(&pool
->manager_arb
);
3670 /* shut down the timers */
3671 del_timer_sync(&pool
->idle_timer
);
3672 del_timer_sync(&pool
->mayday_timer
);
3674 /* sched-RCU protected to allow dereferences from get_work_pool() */
3675 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3679 * get_unbound_pool - get a worker_pool with the specified attributes
3680 * @attrs: the attributes of the worker_pool to get
3682 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3683 * reference count and return it. If there already is a matching
3684 * worker_pool, it will be used; otherwise, this function attempts to
3685 * create a new one. On failure, returns NULL.
3687 * Should be called with wq_pool_mutex held.
3689 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3691 u32 hash
= wqattrs_hash(attrs
);
3692 struct worker_pool
*pool
;
3695 lockdep_assert_held(&wq_pool_mutex
);
3697 /* do we already have a matching pool? */
3698 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3699 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3705 /* nope, create a new one */
3706 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3707 if (!pool
|| init_worker_pool(pool
) < 0)
3710 if (workqueue_freezing
)
3711 pool
->flags
|= POOL_FREEZING
;
3713 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3714 copy_workqueue_attrs(pool
->attrs
, attrs
);
3717 * no_numa isn't a worker_pool attribute, always clear it. See
3718 * 'struct workqueue_attrs' comments for detail.
3720 pool
->attrs
->no_numa
= false;
3722 /* if cpumask is contained inside a NUMA node, we belong to that node */
3723 if (wq_numa_enabled
) {
3724 for_each_node(node
) {
3725 if (cpumask_subset(pool
->attrs
->cpumask
,
3726 wq_numa_possible_cpumask
[node
])) {
3733 if (worker_pool_assign_id(pool
) < 0)
3736 /* create and start the initial worker */
3737 if (create_and_start_worker(pool
) < 0)
3741 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3746 put_unbound_pool(pool
);
3750 static void rcu_free_pwq(struct rcu_head
*rcu
)
3752 kmem_cache_free(pwq_cache
,
3753 container_of(rcu
, struct pool_workqueue
, rcu
));
3757 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3758 * and needs to be destroyed.
3760 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3762 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3763 unbound_release_work
);
3764 struct workqueue_struct
*wq
= pwq
->wq
;
3765 struct worker_pool
*pool
= pwq
->pool
;
3768 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3772 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3773 * necessary on release but do it anyway. It's easier to verify
3774 * and consistent with the linking path.
3776 mutex_lock(&wq
->mutex
);
3777 list_del_rcu(&pwq
->pwqs_node
);
3778 is_last
= list_empty(&wq
->pwqs
);
3779 mutex_unlock(&wq
->mutex
);
3781 mutex_lock(&wq_pool_mutex
);
3782 put_unbound_pool(pool
);
3783 mutex_unlock(&wq_pool_mutex
);
3785 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3788 * If we're the last pwq going away, @wq is already dead and no one
3789 * is gonna access it anymore. Free it.
3792 free_workqueue_attrs(wq
->unbound_attrs
);
3798 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3799 * @pwq: target pool_workqueue
3801 * If @pwq isn't freezing, set @pwq->max_active to the associated
3802 * workqueue's saved_max_active and activate delayed work items
3803 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3805 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3807 struct workqueue_struct
*wq
= pwq
->wq
;
3808 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3810 /* for @wq->saved_max_active */
3811 lockdep_assert_held(&wq
->mutex
);
3813 /* fast exit for non-freezable wqs */
3814 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3817 spin_lock_irq(&pwq
->pool
->lock
);
3819 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3820 pwq
->max_active
= wq
->saved_max_active
;
3822 while (!list_empty(&pwq
->delayed_works
) &&
3823 pwq
->nr_active
< pwq
->max_active
)
3824 pwq_activate_first_delayed(pwq
);
3827 * Need to kick a worker after thawed or an unbound wq's
3828 * max_active is bumped. It's a slow path. Do it always.
3830 wake_up_worker(pwq
->pool
);
3832 pwq
->max_active
= 0;
3835 spin_unlock_irq(&pwq
->pool
->lock
);
3838 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3839 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3840 struct worker_pool
*pool
)
3842 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3844 memset(pwq
, 0, sizeof(*pwq
));
3848 pwq
->flush_color
= -1;
3850 INIT_LIST_HEAD(&pwq
->delayed_works
);
3851 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3852 INIT_LIST_HEAD(&pwq
->mayday_node
);
3853 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3856 /* sync @pwq with the current state of its associated wq and link it */
3857 static void link_pwq(struct pool_workqueue
*pwq
)
3859 struct workqueue_struct
*wq
= pwq
->wq
;
3861 lockdep_assert_held(&wq
->mutex
);
3863 /* may be called multiple times, ignore if already linked */
3864 if (!list_empty(&pwq
->pwqs_node
))
3868 * Set the matching work_color. This is synchronized with
3869 * wq->mutex to avoid confusing flush_workqueue().
3871 pwq
->work_color
= wq
->work_color
;
3873 /* sync max_active to the current setting */
3874 pwq_adjust_max_active(pwq
);
3877 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3880 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3881 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3882 const struct workqueue_attrs
*attrs
)
3884 struct worker_pool
*pool
;
3885 struct pool_workqueue
*pwq
;
3887 lockdep_assert_held(&wq_pool_mutex
);
3889 pool
= get_unbound_pool(attrs
);
3893 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3895 put_unbound_pool(pool
);
3899 init_pwq(pwq
, wq
, pool
);
3903 /* undo alloc_unbound_pwq(), used only in the error path */
3904 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3906 lockdep_assert_held(&wq_pool_mutex
);
3909 put_unbound_pool(pwq
->pool
);
3910 kmem_cache_free(pwq_cache
, pwq
);
3915 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3916 * @attrs: the wq_attrs of interest
3917 * @node: the target NUMA node
3918 * @cpu_going_down: if >= 0, the CPU to consider as offline
3919 * @cpumask: outarg, the resulting cpumask
3921 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3922 * @cpu_going_down is >= 0, that cpu is considered offline during
3923 * calculation. The result is stored in @cpumask. This function returns
3924 * %true if the resulting @cpumask is different from @attrs->cpumask,
3927 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3928 * enabled and @node has online CPUs requested by @attrs, the returned
3929 * cpumask is the intersection of the possible CPUs of @node and
3932 * The caller is responsible for ensuring that the cpumask of @node stays
3935 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3936 int cpu_going_down
, cpumask_t
*cpumask
)
3938 if (!wq_numa_enabled
|| attrs
->no_numa
)
3941 /* does @node have any online CPUs @attrs wants? */
3942 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3943 if (cpu_going_down
>= 0)
3944 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3946 if (cpumask_empty(cpumask
))
3949 /* yeap, return possible CPUs in @node that @attrs wants */
3950 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3951 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3954 cpumask_copy(cpumask
, attrs
->cpumask
);
3958 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3959 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3961 struct pool_workqueue
*pwq
)
3963 struct pool_workqueue
*old_pwq
;
3965 lockdep_assert_held(&wq
->mutex
);
3967 /* link_pwq() can handle duplicate calls */
3970 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3971 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3976 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3977 * @wq: the target workqueue
3978 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3980 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3981 * machines, this function maps a separate pwq to each NUMA node with
3982 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3983 * NUMA node it was issued on. Older pwqs are released as in-flight work
3984 * items finish. Note that a work item which repeatedly requeues itself
3985 * back-to-back will stay on its current pwq.
3987 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3990 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3991 const struct workqueue_attrs
*attrs
)
3993 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3994 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3997 /* only unbound workqueues can change attributes */
3998 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
4001 /* creating multiple pwqs breaks ordering guarantee */
4002 if (!list_empty(&wq
->pwqs
)) {
4003 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4006 wq
->flags
&= ~__WQ_ORDERED
;
4009 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
4010 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4011 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4012 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
4015 /* make a copy of @attrs and sanitize it */
4016 copy_workqueue_attrs(new_attrs
, attrs
);
4017 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
4020 * We may create multiple pwqs with differing cpumasks. Make a
4021 * copy of @new_attrs which will be modified and used to obtain
4024 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
4027 * CPUs should stay stable across pwq creations and installations.
4028 * Pin CPUs, determine the target cpumask for each node and create
4033 mutex_lock(&wq_pool_mutex
);
4036 * If something goes wrong during CPU up/down, we'll fall back to
4037 * the default pwq covering whole @attrs->cpumask. Always create
4038 * it even if we don't use it immediately.
4040 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
4044 for_each_node(node
) {
4045 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
4046 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
4051 pwq_tbl
[node
] = dfl_pwq
;
4055 mutex_unlock(&wq_pool_mutex
);
4057 /* all pwqs have been created successfully, let's install'em */
4058 mutex_lock(&wq
->mutex
);
4060 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
4062 /* save the previous pwq and install the new one */
4064 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
4066 /* @dfl_pwq might not have been used, ensure it's linked */
4068 swap(wq
->dfl_pwq
, dfl_pwq
);
4070 mutex_unlock(&wq
->mutex
);
4072 /* put the old pwqs */
4074 put_pwq_unlocked(pwq_tbl
[node
]);
4075 put_pwq_unlocked(dfl_pwq
);
4081 free_workqueue_attrs(tmp_attrs
);
4082 free_workqueue_attrs(new_attrs
);
4087 free_unbound_pwq(dfl_pwq
);
4089 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
4090 free_unbound_pwq(pwq_tbl
[node
]);
4091 mutex_unlock(&wq_pool_mutex
);
4099 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4100 * @wq: the target workqueue
4101 * @cpu: the CPU coming up or going down
4102 * @online: whether @cpu is coming up or going down
4104 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4105 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4108 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4109 * falls back to @wq->dfl_pwq which may not be optimal but is always
4112 * Note that when the last allowed CPU of a NUMA node goes offline for a
4113 * workqueue with a cpumask spanning multiple nodes, the workers which were
4114 * already executing the work items for the workqueue will lose their CPU
4115 * affinity and may execute on any CPU. This is similar to how per-cpu
4116 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4117 * affinity, it's the user's responsibility to flush the work item from
4120 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
4123 int node
= cpu_to_node(cpu
);
4124 int cpu_off
= online
? -1 : cpu
;
4125 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
4126 struct workqueue_attrs
*target_attrs
;
4129 lockdep_assert_held(&wq_pool_mutex
);
4131 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
4135 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4136 * Let's use a preallocated one. The following buf is protected by
4137 * CPU hotplug exclusion.
4139 target_attrs
= wq_update_unbound_numa_attrs_buf
;
4140 cpumask
= target_attrs
->cpumask
;
4142 mutex_lock(&wq
->mutex
);
4143 if (wq
->unbound_attrs
->no_numa
)
4146 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4147 pwq
= unbound_pwq_by_node(wq
, node
);
4150 * Let's determine what needs to be done. If the target cpumask is
4151 * different from wq's, we need to compare it to @pwq's and create
4152 * a new one if they don't match. If the target cpumask equals
4153 * wq's, the default pwq should be used. If @pwq is already the
4154 * default one, nothing to do; otherwise, install the default one.
4156 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
4157 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4160 if (pwq
== wq
->dfl_pwq
)
4166 mutex_unlock(&wq
->mutex
);
4168 /* create a new pwq */
4169 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4171 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4173 mutex_lock(&wq
->mutex
);
4178 * Install the new pwq. As this function is called only from CPU
4179 * hotplug callbacks and applying a new attrs is wrapped with
4180 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4183 mutex_lock(&wq
->mutex
);
4184 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4188 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4189 get_pwq(wq
->dfl_pwq
);
4190 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4191 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4193 mutex_unlock(&wq
->mutex
);
4194 put_pwq_unlocked(old_pwq
);
4197 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4199 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4202 if (!(wq
->flags
& WQ_UNBOUND
)) {
4203 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4207 for_each_possible_cpu(cpu
) {
4208 struct pool_workqueue
*pwq
=
4209 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4210 struct worker_pool
*cpu_pools
=
4211 per_cpu(cpu_worker_pools
, cpu
);
4213 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4215 mutex_lock(&wq
->mutex
);
4217 mutex_unlock(&wq
->mutex
);
4220 } else if (wq
->flags
& __WQ_ORDERED
) {
4221 ret
= apply_workqueue_attrs(wq
, ordered_wq_attrs
[highpri
]);
4222 /* there should only be single pwq for ordering guarantee */
4223 WARN(!ret
&& (wq
->pwqs
.next
!= &wq
->dfl_pwq
->pwqs_node
||
4224 wq
->pwqs
.prev
!= &wq
->dfl_pwq
->pwqs_node
),
4225 "ordering guarantee broken for workqueue %s\n", wq
->name
);
4228 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4232 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4235 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4237 if (max_active
< 1 || max_active
> lim
)
4238 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4239 max_active
, name
, 1, lim
);
4241 return clamp_val(max_active
, 1, lim
);
4244 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4247 struct lock_class_key
*key
,
4248 const char *lock_name
, ...)
4250 size_t tbl_size
= 0;
4252 struct workqueue_struct
*wq
;
4253 struct pool_workqueue
*pwq
;
4256 * Unbound && max_active == 1 used to imply ordered, which is no
4257 * longer the case on NUMA machines due to per-node pools. While
4258 * alloc_ordered_workqueue() is the right way to create an ordered
4259 * workqueue, keep the previous behavior to avoid subtle breakages
4262 if ((flags
& WQ_UNBOUND
) && max_active
== 1)
4263 flags
|= __WQ_ORDERED
;
4265 /* allocate wq and format name */
4266 if (flags
& WQ_UNBOUND
)
4267 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4269 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4273 if (flags
& WQ_UNBOUND
) {
4274 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4275 if (!wq
->unbound_attrs
)
4279 va_start(args
, lock_name
);
4280 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4283 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4284 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4288 wq
->saved_max_active
= max_active
;
4289 mutex_init(&wq
->mutex
);
4290 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4291 INIT_LIST_HEAD(&wq
->pwqs
);
4292 INIT_LIST_HEAD(&wq
->flusher_queue
);
4293 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4294 INIT_LIST_HEAD(&wq
->maydays
);
4296 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4297 INIT_LIST_HEAD(&wq
->list
);
4299 if (alloc_and_link_pwqs(wq
) < 0)
4303 * Workqueues which may be used during memory reclaim should
4304 * have a rescuer to guarantee forward progress.
4306 if (flags
& WQ_MEM_RECLAIM
) {
4307 struct worker
*rescuer
;
4309 rescuer
= alloc_worker();
4313 rescuer
->rescue_wq
= wq
;
4314 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4316 if (IS_ERR(rescuer
->task
)) {
4321 wq
->rescuer
= rescuer
;
4322 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4323 wake_up_process(rescuer
->task
);
4326 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4330 * wq_pool_mutex protects global freeze state and workqueues list.
4331 * Grab it, adjust max_active and add the new @wq to workqueues
4334 mutex_lock(&wq_pool_mutex
);
4336 mutex_lock(&wq
->mutex
);
4337 for_each_pwq(pwq
, wq
)
4338 pwq_adjust_max_active(pwq
);
4339 mutex_unlock(&wq
->mutex
);
4341 list_add(&wq
->list
, &workqueues
);
4343 mutex_unlock(&wq_pool_mutex
);
4348 free_workqueue_attrs(wq
->unbound_attrs
);
4352 destroy_workqueue(wq
);
4355 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4358 * destroy_workqueue - safely terminate a workqueue
4359 * @wq: target workqueue
4361 * Safely destroy a workqueue. All work currently pending will be done first.
4363 void destroy_workqueue(struct workqueue_struct
*wq
)
4365 struct pool_workqueue
*pwq
;
4368 /* drain it before proceeding with destruction */
4369 drain_workqueue(wq
);
4372 mutex_lock(&wq
->mutex
);
4373 for_each_pwq(pwq
, wq
) {
4376 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4377 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4378 mutex_unlock(&wq
->mutex
);
4383 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4384 WARN_ON(pwq
->nr_active
) ||
4385 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4386 mutex_unlock(&wq
->mutex
);
4390 mutex_unlock(&wq
->mutex
);
4393 * wq list is used to freeze wq, remove from list after
4394 * flushing is complete in case freeze races us.
4396 mutex_lock(&wq_pool_mutex
);
4397 list_del_init(&wq
->list
);
4398 mutex_unlock(&wq_pool_mutex
);
4400 workqueue_sysfs_unregister(wq
);
4403 kthread_stop(wq
->rescuer
->task
);
4408 if (!(wq
->flags
& WQ_UNBOUND
)) {
4410 * The base ref is never dropped on per-cpu pwqs. Directly
4411 * free the pwqs and wq.
4413 free_percpu(wq
->cpu_pwqs
);
4417 * We're the sole accessor of @wq at this point. Directly
4418 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4419 * @wq will be freed when the last pwq is released.
4421 for_each_node(node
) {
4422 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4423 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4424 put_pwq_unlocked(pwq
);
4428 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4429 * put. Don't access it afterwards.
4433 put_pwq_unlocked(pwq
);
4436 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4439 * workqueue_set_max_active - adjust max_active of a workqueue
4440 * @wq: target workqueue
4441 * @max_active: new max_active value.
4443 * Set max_active of @wq to @max_active.
4446 * Don't call from IRQ context.
4448 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4450 struct pool_workqueue
*pwq
;
4452 /* disallow meddling with max_active for ordered workqueues */
4453 if (WARN_ON(wq
->flags
& __WQ_ORDERED_EXPLICIT
))
4456 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4458 mutex_lock(&wq
->mutex
);
4460 wq
->flags
&= ~__WQ_ORDERED
;
4461 wq
->saved_max_active
= max_active
;
4463 for_each_pwq(pwq
, wq
)
4464 pwq_adjust_max_active(pwq
);
4466 mutex_unlock(&wq
->mutex
);
4468 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4471 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4473 * Determine whether %current is a workqueue rescuer. Can be used from
4474 * work functions to determine whether it's being run off the rescuer task.
4476 bool current_is_workqueue_rescuer(void)
4478 struct worker
*worker
= current_wq_worker();
4480 return worker
&& worker
->rescue_wq
;
4484 * workqueue_congested - test whether a workqueue is congested
4485 * @cpu: CPU in question
4486 * @wq: target workqueue
4488 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4489 * no synchronization around this function and the test result is
4490 * unreliable and only useful as advisory hints or for debugging.
4492 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4493 * Note that both per-cpu and unbound workqueues may be associated with
4494 * multiple pool_workqueues which have separate congested states. A
4495 * workqueue being congested on one CPU doesn't mean the workqueue is also
4496 * contested on other CPUs / NUMA nodes.
4499 * %true if congested, %false otherwise.
4501 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4503 struct pool_workqueue
*pwq
;
4506 rcu_read_lock_sched();
4508 if (cpu
== WORK_CPU_UNBOUND
)
4509 cpu
= smp_processor_id();
4511 if (!(wq
->flags
& WQ_UNBOUND
))
4512 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4514 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4516 ret
= !list_empty(&pwq
->delayed_works
);
4517 rcu_read_unlock_sched();
4521 EXPORT_SYMBOL_GPL(workqueue_congested
);
4524 * work_busy - test whether a work is currently pending or running
4525 * @work: the work to be tested
4527 * Test whether @work is currently pending or running. There is no
4528 * synchronization around this function and the test result is
4529 * unreliable and only useful as advisory hints or for debugging.
4532 * OR'd bitmask of WORK_BUSY_* bits.
4534 unsigned int work_busy(struct work_struct
*work
)
4536 struct worker_pool
*pool
;
4537 unsigned long flags
;
4538 unsigned int ret
= 0;
4540 if (work_pending(work
))
4541 ret
|= WORK_BUSY_PENDING
;
4543 local_irq_save(flags
);
4544 pool
= get_work_pool(work
);
4546 spin_lock(&pool
->lock
);
4547 if (find_worker_executing_work(pool
, work
))
4548 ret
|= WORK_BUSY_RUNNING
;
4549 spin_unlock(&pool
->lock
);
4551 local_irq_restore(flags
);
4555 EXPORT_SYMBOL_GPL(work_busy
);
4558 * set_worker_desc - set description for the current work item
4559 * @fmt: printf-style format string
4560 * @...: arguments for the format string
4562 * This function can be called by a running work function to describe what
4563 * the work item is about. If the worker task gets dumped, this
4564 * information will be printed out together to help debugging. The
4565 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4567 void set_worker_desc(const char *fmt
, ...)
4569 struct worker
*worker
= current_wq_worker();
4573 va_start(args
, fmt
);
4574 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4576 worker
->desc_valid
= true;
4581 * print_worker_info - print out worker information and description
4582 * @log_lvl: the log level to use when printing
4583 * @task: target task
4585 * If @task is a worker and currently executing a work item, print out the
4586 * name of the workqueue being serviced and worker description set with
4587 * set_worker_desc() by the currently executing work item.
4589 * This function can be safely called on any task as long as the
4590 * task_struct itself is accessible. While safe, this function isn't
4591 * synchronized and may print out mixups or garbages of limited length.
4593 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4595 work_func_t
*fn
= NULL
;
4596 char name
[WQ_NAME_LEN
] = { };
4597 char desc
[WORKER_DESC_LEN
] = { };
4598 struct pool_workqueue
*pwq
= NULL
;
4599 struct workqueue_struct
*wq
= NULL
;
4600 bool desc_valid
= false;
4601 struct worker
*worker
;
4603 if (!(task
->flags
& PF_WQ_WORKER
))
4607 * This function is called without any synchronization and @task
4608 * could be in any state. Be careful with dereferences.
4610 worker
= probe_kthread_data(task
);
4613 * Carefully copy the associated workqueue's workfn and name. Keep
4614 * the original last '\0' in case the original contains garbage.
4616 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4617 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4618 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4619 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4621 /* copy worker description */
4622 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4624 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4626 if (fn
|| name
[0] || desc
[0]) {
4627 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4629 pr_cont(" (%s)", desc
);
4637 * There are two challenges in supporting CPU hotplug. Firstly, there
4638 * are a lot of assumptions on strong associations among work, pwq and
4639 * pool which make migrating pending and scheduled works very
4640 * difficult to implement without impacting hot paths. Secondly,
4641 * worker pools serve mix of short, long and very long running works making
4642 * blocked draining impractical.
4644 * This is solved by allowing the pools to be disassociated from the CPU
4645 * running as an unbound one and allowing it to be reattached later if the
4646 * cpu comes back online.
4649 static void wq_unbind_fn(struct work_struct
*work
)
4651 int cpu
= smp_processor_id();
4652 struct worker_pool
*pool
;
4653 struct worker
*worker
;
4656 for_each_cpu_worker_pool(pool
, cpu
) {
4657 WARN_ON_ONCE(cpu
!= smp_processor_id());
4659 mutex_lock(&pool
->manager_mutex
);
4660 spin_lock_irq(&pool
->lock
);
4663 * We've blocked all manager operations. Make all workers
4664 * unbound and set DISASSOCIATED. Before this, all workers
4665 * except for the ones which are still executing works from
4666 * before the last CPU down must be on the cpu. After
4667 * this, they may become diasporas.
4669 for_each_pool_worker(worker
, wi
, pool
)
4670 worker
->flags
|= WORKER_UNBOUND
;
4672 pool
->flags
|= POOL_DISASSOCIATED
;
4674 spin_unlock_irq(&pool
->lock
);
4675 mutex_unlock(&pool
->manager_mutex
);
4678 * Call schedule() so that we cross rq->lock and thus can
4679 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4680 * This is necessary as scheduler callbacks may be invoked
4686 * Sched callbacks are disabled now. Zap nr_running.
4687 * After this, nr_running stays zero and need_more_worker()
4688 * and keep_working() are always true as long as the
4689 * worklist is not empty. This pool now behaves as an
4690 * unbound (in terms of concurrency management) pool which
4691 * are served by workers tied to the pool.
4693 atomic_set(&pool
->nr_running
, 0);
4696 * With concurrency management just turned off, a busy
4697 * worker blocking could lead to lengthy stalls. Kick off
4698 * unbound chain execution of currently pending work items.
4700 spin_lock_irq(&pool
->lock
);
4701 wake_up_worker(pool
);
4702 spin_unlock_irq(&pool
->lock
);
4707 * rebind_workers - rebind all workers of a pool to the associated CPU
4708 * @pool: pool of interest
4710 * @pool->cpu is coming online. Rebind all workers to the CPU.
4712 static void rebind_workers(struct worker_pool
*pool
)
4714 struct worker
*worker
;
4717 lockdep_assert_held(&pool
->manager_mutex
);
4720 * Restore CPU affinity of all workers. As all idle workers should
4721 * be on the run-queue of the associated CPU before any local
4722 * wake-ups for concurrency management happen, restore CPU affinty
4723 * of all workers first and then clear UNBOUND. As we're called
4724 * from CPU_ONLINE, the following shouldn't fail.
4726 for_each_pool_worker(worker
, wi
, pool
)
4727 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4728 pool
->attrs
->cpumask
) < 0);
4730 spin_lock_irq(&pool
->lock
);
4732 for_each_pool_worker(worker
, wi
, pool
) {
4733 unsigned int worker_flags
= worker
->flags
;
4736 * A bound idle worker should actually be on the runqueue
4737 * of the associated CPU for local wake-ups targeting it to
4738 * work. Kick all idle workers so that they migrate to the
4739 * associated CPU. Doing this in the same loop as
4740 * replacing UNBOUND with REBOUND is safe as no worker will
4741 * be bound before @pool->lock is released.
4743 if (worker_flags
& WORKER_IDLE
)
4744 wake_up_process(worker
->task
);
4747 * We want to clear UNBOUND but can't directly call
4748 * worker_clr_flags() or adjust nr_running. Atomically
4749 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4750 * @worker will clear REBOUND using worker_clr_flags() when
4751 * it initiates the next execution cycle thus restoring
4752 * concurrency management. Note that when or whether
4753 * @worker clears REBOUND doesn't affect correctness.
4755 * ACCESS_ONCE() is necessary because @worker->flags may be
4756 * tested without holding any lock in
4757 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4758 * fail incorrectly leading to premature concurrency
4759 * management operations.
4761 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4762 worker_flags
|= WORKER_REBOUND
;
4763 worker_flags
&= ~WORKER_UNBOUND
;
4764 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4767 spin_unlock_irq(&pool
->lock
);
4771 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4772 * @pool: unbound pool of interest
4773 * @cpu: the CPU which is coming up
4775 * An unbound pool may end up with a cpumask which doesn't have any online
4776 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4777 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4778 * online CPU before, cpus_allowed of all its workers should be restored.
4780 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4782 static cpumask_t cpumask
;
4783 struct worker
*worker
;
4786 lockdep_assert_held(&pool
->manager_mutex
);
4788 /* is @cpu allowed for @pool? */
4789 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4792 /* is @cpu the only online CPU? */
4793 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4794 if (cpumask_weight(&cpumask
) != 1)
4797 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4798 for_each_pool_worker(worker
, wi
, pool
)
4799 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4800 pool
->attrs
->cpumask
) < 0);
4804 * Workqueues should be brought up before normal priority CPU notifiers.
4805 * This will be registered high priority CPU notifier.
4807 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4808 unsigned long action
,
4811 int cpu
= (unsigned long)hcpu
;
4812 struct worker_pool
*pool
;
4813 struct workqueue_struct
*wq
;
4816 switch (action
& ~CPU_TASKS_FROZEN
) {
4817 case CPU_UP_PREPARE
:
4818 for_each_cpu_worker_pool(pool
, cpu
) {
4819 if (pool
->nr_workers
)
4821 if (create_and_start_worker(pool
) < 0)
4826 case CPU_DOWN_FAILED
:
4828 mutex_lock(&wq_pool_mutex
);
4830 for_each_pool(pool
, pi
) {
4831 mutex_lock(&pool
->manager_mutex
);
4833 if (pool
->cpu
== cpu
) {
4834 spin_lock_irq(&pool
->lock
);
4835 pool
->flags
&= ~POOL_DISASSOCIATED
;
4836 spin_unlock_irq(&pool
->lock
);
4838 rebind_workers(pool
);
4839 } else if (pool
->cpu
< 0) {
4840 restore_unbound_workers_cpumask(pool
, cpu
);
4843 mutex_unlock(&pool
->manager_mutex
);
4846 /* update NUMA affinity of unbound workqueues */
4847 list_for_each_entry(wq
, &workqueues
, list
)
4848 wq_update_unbound_numa(wq
, cpu
, true);
4850 mutex_unlock(&wq_pool_mutex
);
4857 * Workqueues should be brought down after normal priority CPU notifiers.
4858 * This will be registered as low priority CPU notifier.
4860 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4861 unsigned long action
,
4864 int cpu
= (unsigned long)hcpu
;
4865 struct work_struct unbind_work
;
4866 struct workqueue_struct
*wq
;
4868 switch (action
& ~CPU_TASKS_FROZEN
) {
4869 case CPU_DOWN_PREPARE
:
4870 /* unbinding per-cpu workers should happen on the local CPU */
4871 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4872 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4874 /* update NUMA affinity of unbound workqueues */
4875 mutex_lock(&wq_pool_mutex
);
4876 list_for_each_entry(wq
, &workqueues
, list
)
4877 wq_update_unbound_numa(wq
, cpu
, false);
4878 mutex_unlock(&wq_pool_mutex
);
4880 /* wait for per-cpu unbinding to finish */
4881 flush_work(&unbind_work
);
4889 struct work_for_cpu
{
4890 struct work_struct work
;
4896 static void work_for_cpu_fn(struct work_struct
*work
)
4898 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4900 wfc
->ret
= wfc
->fn(wfc
->arg
);
4904 * work_on_cpu - run a function in user context on a particular cpu
4905 * @cpu: the cpu to run on
4906 * @fn: the function to run
4907 * @arg: the function arg
4909 * This will return the value @fn returns.
4910 * It is up to the caller to ensure that the cpu doesn't go offline.
4911 * The caller must not hold any locks which would prevent @fn from completing.
4913 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4915 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4917 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4918 schedule_work_on(cpu
, &wfc
.work
);
4919 flush_work(&wfc
.work
);
4922 EXPORT_SYMBOL_GPL(work_on_cpu
);
4923 #endif /* CONFIG_SMP */
4925 #ifdef CONFIG_FREEZER
4928 * freeze_workqueues_begin - begin freezing workqueues
4930 * Start freezing workqueues. After this function returns, all freezable
4931 * workqueues will queue new works to their delayed_works list instead of
4935 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4937 void freeze_workqueues_begin(void)
4939 struct worker_pool
*pool
;
4940 struct workqueue_struct
*wq
;
4941 struct pool_workqueue
*pwq
;
4944 mutex_lock(&wq_pool_mutex
);
4946 WARN_ON_ONCE(workqueue_freezing
);
4947 workqueue_freezing
= true;
4950 for_each_pool(pool
, pi
) {
4951 spin_lock_irq(&pool
->lock
);
4952 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4953 pool
->flags
|= POOL_FREEZING
;
4954 spin_unlock_irq(&pool
->lock
);
4957 list_for_each_entry(wq
, &workqueues
, list
) {
4958 mutex_lock(&wq
->mutex
);
4959 for_each_pwq(pwq
, wq
)
4960 pwq_adjust_max_active(pwq
);
4961 mutex_unlock(&wq
->mutex
);
4964 mutex_unlock(&wq_pool_mutex
);
4968 * freeze_workqueues_busy - are freezable workqueues still busy?
4970 * Check whether freezing is complete. This function must be called
4971 * between freeze_workqueues_begin() and thaw_workqueues().
4974 * Grabs and releases wq_pool_mutex.
4977 * %true if some freezable workqueues are still busy. %false if freezing
4980 bool freeze_workqueues_busy(void)
4983 struct workqueue_struct
*wq
;
4984 struct pool_workqueue
*pwq
;
4986 mutex_lock(&wq_pool_mutex
);
4988 WARN_ON_ONCE(!workqueue_freezing
);
4990 list_for_each_entry(wq
, &workqueues
, list
) {
4991 if (!(wq
->flags
& WQ_FREEZABLE
))
4994 * nr_active is monotonically decreasing. It's safe
4995 * to peek without lock.
4997 rcu_read_lock_sched();
4998 for_each_pwq(pwq
, wq
) {
4999 WARN_ON_ONCE(pwq
->nr_active
< 0);
5000 if (pwq
->nr_active
) {
5002 rcu_read_unlock_sched();
5006 rcu_read_unlock_sched();
5009 mutex_unlock(&wq_pool_mutex
);
5014 * thaw_workqueues - thaw workqueues
5016 * Thaw workqueues. Normal queueing is restored and all collected
5017 * frozen works are transferred to their respective pool worklists.
5020 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5022 void thaw_workqueues(void)
5024 struct workqueue_struct
*wq
;
5025 struct pool_workqueue
*pwq
;
5026 struct worker_pool
*pool
;
5029 mutex_lock(&wq_pool_mutex
);
5031 if (!workqueue_freezing
)
5034 /* clear FREEZING */
5035 for_each_pool(pool
, pi
) {
5036 spin_lock_irq(&pool
->lock
);
5037 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
5038 pool
->flags
&= ~POOL_FREEZING
;
5039 spin_unlock_irq(&pool
->lock
);
5042 /* restore max_active and repopulate worklist */
5043 list_for_each_entry(wq
, &workqueues
, list
) {
5044 mutex_lock(&wq
->mutex
);
5045 for_each_pwq(pwq
, wq
)
5046 pwq_adjust_max_active(pwq
);
5047 mutex_unlock(&wq
->mutex
);
5050 workqueue_freezing
= false;
5052 mutex_unlock(&wq_pool_mutex
);
5054 #endif /* CONFIG_FREEZER */
5056 static void __init
wq_numa_init(void)
5061 /* determine NUMA pwq table len - highest node id + 1 */
5063 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
5065 if (num_possible_nodes() <= 1)
5068 if (wq_disable_numa
) {
5069 pr_info("workqueue: NUMA affinity support disabled\n");
5073 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
5074 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
5077 * We want masks of possible CPUs of each node which isn't readily
5078 * available. Build one from cpu_to_node() which should have been
5079 * fully initialized by now.
5081 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
5085 BUG_ON(!zalloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
5086 node_online(node
) ? node
: NUMA_NO_NODE
));
5088 for_each_possible_cpu(cpu
) {
5089 node
= cpu_to_node(cpu
);
5090 if (WARN_ON(node
== NUMA_NO_NODE
)) {
5091 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
5092 /* happens iff arch is bonkers, let's just proceed */
5095 cpumask_set_cpu(cpu
, tbl
[node
]);
5098 wq_numa_possible_cpumask
= tbl
;
5099 wq_numa_enabled
= true;
5102 static int __init
init_workqueues(void)
5104 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
5107 /* make sure we have enough bits for OFFQ pool ID */
5108 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
5109 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
5111 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
5113 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
5115 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
5116 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
5120 /* initialize CPU pools */
5121 for_each_possible_cpu(cpu
) {
5122 struct worker_pool
*pool
;
5125 for_each_cpu_worker_pool(pool
, cpu
) {
5126 BUG_ON(init_worker_pool(pool
));
5128 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
5129 pool
->attrs
->nice
= std_nice
[i
++];
5130 pool
->node
= cpu_to_node(cpu
);
5133 mutex_lock(&wq_pool_mutex
);
5134 BUG_ON(worker_pool_assign_id(pool
));
5135 mutex_unlock(&wq_pool_mutex
);
5139 /* create the initial worker */
5140 for_each_online_cpu(cpu
) {
5141 struct worker_pool
*pool
;
5143 for_each_cpu_worker_pool(pool
, cpu
) {
5144 pool
->flags
&= ~POOL_DISASSOCIATED
;
5145 BUG_ON(create_and_start_worker(pool
) < 0);
5149 /* create default unbound and ordered wq attrs */
5150 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
5151 struct workqueue_attrs
*attrs
;
5153 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5154 attrs
->nice
= std_nice
[i
];
5155 unbound_std_wq_attrs
[i
] = attrs
;
5158 * An ordered wq should have only one pwq as ordering is
5159 * guaranteed by max_active which is enforced by pwqs.
5160 * Turn off NUMA so that dfl_pwq is used for all nodes.
5162 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
5163 attrs
->nice
= std_nice
[i
];
5164 attrs
->no_numa
= true;
5165 ordered_wq_attrs
[i
] = attrs
;
5168 system_wq
= alloc_workqueue("events", 0, 0);
5169 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
5170 system_long_wq
= alloc_workqueue("events_long", 0, 0);
5171 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
5172 WQ_UNBOUND_MAX_ACTIVE
);
5173 system_freezable_wq
= alloc_workqueue("events_freezable",
5175 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5176 !system_unbound_wq
|| !system_freezable_wq
);
5179 early_initcall(init_workqueues
);