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 struct workqueue_struct
*system_wq __read_mostly
;
299 EXPORT_SYMBOL(system_wq
);
300 struct workqueue_struct
*system_highpri_wq __read_mostly
;
301 EXPORT_SYMBOL_GPL(system_highpri_wq
);
302 struct workqueue_struct
*system_long_wq __read_mostly
;
303 EXPORT_SYMBOL_GPL(system_long_wq
);
304 struct workqueue_struct
*system_unbound_wq __read_mostly
;
305 EXPORT_SYMBOL_GPL(system_unbound_wq
);
306 struct workqueue_struct
*system_freezable_wq __read_mostly
;
307 EXPORT_SYMBOL_GPL(system_freezable_wq
);
309 static int worker_thread(void *__worker
);
310 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
311 const struct workqueue_attrs
*from
);
313 #define CREATE_TRACE_POINTS
314 #include <trace/events/workqueue.h>
316 #define assert_rcu_or_pool_mutex() \
317 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
318 lockdep_is_held(&wq_pool_mutex), \
319 "sched RCU or wq_pool_mutex should be held")
321 #define assert_rcu_or_wq_mutex(wq) \
322 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
323 lockdep_is_held(&wq->mutex), \
324 "sched RCU or wq->mutex should be held")
326 #ifdef CONFIG_LOCKDEP
327 #define assert_manager_or_pool_lock(pool) \
328 WARN_ONCE(debug_locks && \
329 !lockdep_is_held(&(pool)->manager_mutex) && \
330 !lockdep_is_held(&(pool)->lock), \
331 "pool->manager_mutex or ->lock should be held")
333 #define assert_manager_or_pool_lock(pool) do { } while (0)
336 #define for_each_cpu_worker_pool(pool, cpu) \
337 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
338 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
342 * for_each_pool - iterate through all worker_pools in the system
343 * @pool: iteration cursor
344 * @pi: integer used for iteration
346 * This must be called either with wq_pool_mutex held or sched RCU read
347 * locked. If the pool needs to be used beyond the locking in effect, the
348 * caller is responsible for guaranteeing that the pool stays online.
350 * The if/else clause exists only for the lockdep assertion and can be
353 #define for_each_pool(pool, pi) \
354 idr_for_each_entry(&worker_pool_idr, pool, pi) \
355 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
359 * for_each_pool_worker - iterate through all workers of a worker_pool
360 * @worker: iteration cursor
361 * @wi: integer used for iteration
362 * @pool: worker_pool to iterate workers of
364 * This must be called with either @pool->manager_mutex or ->lock held.
366 * The if/else clause exists only for the lockdep assertion and can be
369 #define for_each_pool_worker(worker, wi, pool) \
370 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
371 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
375 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
376 * @pwq: iteration cursor
377 * @wq: the target workqueue
379 * This must be called either with wq->mutex held or sched RCU read locked.
380 * If the pwq needs to be used beyond the locking in effect, the caller is
381 * responsible for guaranteeing that the pwq stays online.
383 * The if/else clause exists only for the lockdep assertion and can be
386 #define for_each_pwq(pwq, wq) \
387 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
388 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
391 #ifdef CONFIG_DEBUG_OBJECTS_WORK
393 static struct debug_obj_descr work_debug_descr
;
395 static void *work_debug_hint(void *addr
)
397 return ((struct work_struct
*) addr
)->func
;
401 * fixup_init is called when:
402 * - an active object is initialized
404 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
406 struct work_struct
*work
= addr
;
409 case ODEBUG_STATE_ACTIVE
:
410 cancel_work_sync(work
);
411 debug_object_init(work
, &work_debug_descr
);
419 * fixup_activate is called when:
420 * - an active object is activated
421 * - an unknown object is activated (might be a statically initialized object)
423 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
425 struct work_struct
*work
= addr
;
429 case ODEBUG_STATE_NOTAVAILABLE
:
431 * This is not really a fixup. The work struct was
432 * statically initialized. We just make sure that it
433 * is tracked in the object tracker.
435 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
436 debug_object_init(work
, &work_debug_descr
);
437 debug_object_activate(work
, &work_debug_descr
);
443 case ODEBUG_STATE_ACTIVE
:
452 * fixup_free is called when:
453 * - an active object is freed
455 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
457 struct work_struct
*work
= addr
;
460 case ODEBUG_STATE_ACTIVE
:
461 cancel_work_sync(work
);
462 debug_object_free(work
, &work_debug_descr
);
469 static struct debug_obj_descr work_debug_descr
= {
470 .name
= "work_struct",
471 .debug_hint
= work_debug_hint
,
472 .fixup_init
= work_fixup_init
,
473 .fixup_activate
= work_fixup_activate
,
474 .fixup_free
= work_fixup_free
,
477 static inline void debug_work_activate(struct work_struct
*work
)
479 debug_object_activate(work
, &work_debug_descr
);
482 static inline void debug_work_deactivate(struct work_struct
*work
)
484 debug_object_deactivate(work
, &work_debug_descr
);
487 void __init_work(struct work_struct
*work
, int onstack
)
490 debug_object_init_on_stack(work
, &work_debug_descr
);
492 debug_object_init(work
, &work_debug_descr
);
494 EXPORT_SYMBOL_GPL(__init_work
);
496 void destroy_work_on_stack(struct work_struct
*work
)
498 debug_object_free(work
, &work_debug_descr
);
500 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
503 static inline void debug_work_activate(struct work_struct
*work
) { }
504 static inline void debug_work_deactivate(struct work_struct
*work
) { }
507 /* allocate ID and assign it to @pool */
508 static int worker_pool_assign_id(struct worker_pool
*pool
)
512 lockdep_assert_held(&wq_pool_mutex
);
514 ret
= idr_alloc(&worker_pool_idr
, pool
, 0, 0, GFP_KERNEL
);
523 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
524 * @wq: the target workqueue
527 * This must be called either with pwq_lock held or sched RCU read locked.
528 * If the pwq needs to be used beyond the locking in effect, the caller is
529 * responsible for guaranteeing that the pwq stays online.
531 static struct pool_workqueue
*unbound_pwq_by_node(struct workqueue_struct
*wq
,
534 assert_rcu_or_wq_mutex(wq
);
535 return rcu_dereference_raw(wq
->numa_pwq_tbl
[node
]);
538 static unsigned int work_color_to_flags(int color
)
540 return color
<< WORK_STRUCT_COLOR_SHIFT
;
543 static int get_work_color(struct work_struct
*work
)
545 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
546 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
549 static int work_next_color(int color
)
551 return (color
+ 1) % WORK_NR_COLORS
;
555 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
556 * contain the pointer to the queued pwq. Once execution starts, the flag
557 * is cleared and the high bits contain OFFQ flags and pool ID.
559 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
560 * and clear_work_data() can be used to set the pwq, pool or clear
561 * work->data. These functions should only be called while the work is
562 * owned - ie. while the PENDING bit is set.
564 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
565 * corresponding to a work. Pool is available once the work has been
566 * queued anywhere after initialization until it is sync canceled. pwq is
567 * available only while the work item is queued.
569 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
570 * canceled. While being canceled, a work item may have its PENDING set
571 * but stay off timer and worklist for arbitrarily long and nobody should
572 * try to steal the PENDING bit.
574 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
577 WARN_ON_ONCE(!work_pending(work
));
578 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
581 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
582 unsigned long extra_flags
)
584 set_work_data(work
, (unsigned long)pwq
,
585 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
588 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
591 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
592 WORK_STRUCT_PENDING
);
595 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
599 * The following wmb is paired with the implied mb in
600 * test_and_set_bit(PENDING) and ensures all updates to @work made
601 * here are visible to and precede any updates by the next PENDING
605 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
608 static void clear_work_data(struct work_struct
*work
)
610 smp_wmb(); /* see set_work_pool_and_clear_pending() */
611 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
614 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
616 unsigned long data
= atomic_long_read(&work
->data
);
618 if (data
& WORK_STRUCT_PWQ
)
619 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
625 * get_work_pool - return the worker_pool a given work was associated with
626 * @work: the work item of interest
628 * Return the worker_pool @work was last associated with. %NULL if none.
630 * Pools are created and destroyed under wq_pool_mutex, and allows read
631 * access under sched-RCU read lock. As such, this function should be
632 * called under wq_pool_mutex or with preemption disabled.
634 * All fields of the returned pool are accessible as long as the above
635 * mentioned locking is in effect. If the returned pool needs to be used
636 * beyond the critical section, the caller is responsible for ensuring the
637 * returned pool is and stays online.
639 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
641 unsigned long data
= atomic_long_read(&work
->data
);
644 assert_rcu_or_pool_mutex();
646 if (data
& WORK_STRUCT_PWQ
)
647 return ((struct pool_workqueue
*)
648 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
650 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
651 if (pool_id
== WORK_OFFQ_POOL_NONE
)
654 return idr_find(&worker_pool_idr
, pool_id
);
658 * get_work_pool_id - return the worker pool ID a given work is associated with
659 * @work: the work item of interest
661 * Return the worker_pool ID @work was last associated with.
662 * %WORK_OFFQ_POOL_NONE if none.
664 static int get_work_pool_id(struct work_struct
*work
)
666 unsigned long data
= atomic_long_read(&work
->data
);
668 if (data
& WORK_STRUCT_PWQ
)
669 return ((struct pool_workqueue
*)
670 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
672 return data
>> WORK_OFFQ_POOL_SHIFT
;
675 static void mark_work_canceling(struct work_struct
*work
)
677 unsigned long pool_id
= get_work_pool_id(work
);
679 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
680 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
683 static bool work_is_canceling(struct work_struct
*work
)
685 unsigned long data
= atomic_long_read(&work
->data
);
687 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
691 * Policy functions. These define the policies on how the global worker
692 * pools are managed. Unless noted otherwise, these functions assume that
693 * they're being called with pool->lock held.
696 static bool __need_more_worker(struct worker_pool
*pool
)
698 return !atomic_read(&pool
->nr_running
);
702 * Need to wake up a worker? Called from anything but currently
705 * Note that, because unbound workers never contribute to nr_running, this
706 * function will always return %true for unbound pools as long as the
707 * worklist isn't empty.
709 static bool need_more_worker(struct worker_pool
*pool
)
711 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
714 /* Can I start working? Called from busy but !running workers. */
715 static bool may_start_working(struct worker_pool
*pool
)
717 return pool
->nr_idle
;
720 /* Do I need to keep working? Called from currently running workers. */
721 static bool keep_working(struct worker_pool
*pool
)
723 return !list_empty(&pool
->worklist
) &&
724 atomic_read(&pool
->nr_running
) <= 1;
727 /* Do we need a new worker? Called from manager. */
728 static bool need_to_create_worker(struct worker_pool
*pool
)
730 return need_more_worker(pool
) && !may_start_working(pool
);
733 /* Do I need to be the manager? */
734 static bool need_to_manage_workers(struct worker_pool
*pool
)
736 return need_to_create_worker(pool
) ||
737 (pool
->flags
& POOL_MANAGE_WORKERS
);
740 /* Do we have too many workers and should some go away? */
741 static bool too_many_workers(struct worker_pool
*pool
)
743 bool managing
= mutex_is_locked(&pool
->manager_arb
);
744 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
745 int nr_busy
= pool
->nr_workers
- nr_idle
;
748 * nr_idle and idle_list may disagree if idle rebinding is in
749 * progress. Never return %true if idle_list is empty.
751 if (list_empty(&pool
->idle_list
))
754 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
761 /* Return the first worker. Safe with preemption disabled */
762 static struct worker
*first_worker(struct worker_pool
*pool
)
764 if (unlikely(list_empty(&pool
->idle_list
)))
767 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
771 * wake_up_worker - wake up an idle worker
772 * @pool: worker pool to wake worker from
774 * Wake up the first idle worker of @pool.
777 * spin_lock_irq(pool->lock).
779 static void wake_up_worker(struct worker_pool
*pool
)
781 struct worker
*worker
= first_worker(pool
);
784 wake_up_process(worker
->task
);
788 * wq_worker_waking_up - a worker is waking up
789 * @task: task waking up
790 * @cpu: CPU @task is waking up to
792 * This function is called during try_to_wake_up() when a worker is
796 * spin_lock_irq(rq->lock)
798 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
800 struct worker
*worker
= kthread_data(task
);
802 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
803 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
804 atomic_inc(&worker
->pool
->nr_running
);
809 * wq_worker_sleeping - a worker is going to sleep
810 * @task: task going to sleep
811 * @cpu: CPU in question, must be the current CPU number
813 * This function is called during schedule() when a busy worker is
814 * going to sleep. Worker on the same cpu can be woken up by
815 * returning pointer to its task.
818 * spin_lock_irq(rq->lock)
821 * Worker task on @cpu to wake up, %NULL if none.
823 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
825 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
826 struct worker_pool
*pool
;
829 * Rescuers, which may not have all the fields set up like normal
830 * workers, also reach here, let's not access anything before
831 * checking NOT_RUNNING.
833 if (worker
->flags
& WORKER_NOT_RUNNING
)
838 /* this can only happen on the local cpu */
839 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
843 * The counterpart of the following dec_and_test, implied mb,
844 * worklist not empty test sequence is in insert_work().
845 * Please read comment there.
847 * NOT_RUNNING is clear. This means that we're bound to and
848 * running on the local cpu w/ rq lock held and preemption
849 * disabled, which in turn means that none else could be
850 * manipulating idle_list, so dereferencing idle_list without pool
853 if (atomic_dec_and_test(&pool
->nr_running
) &&
854 !list_empty(&pool
->worklist
))
855 to_wakeup
= first_worker(pool
);
856 return to_wakeup
? to_wakeup
->task
: NULL
;
860 * worker_set_flags - set worker flags and adjust nr_running accordingly
862 * @flags: flags to set
863 * @wakeup: wakeup an idle worker if necessary
865 * Set @flags in @worker->flags and adjust nr_running accordingly. If
866 * nr_running becomes zero and @wakeup is %true, an idle worker is
870 * spin_lock_irq(pool->lock)
872 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
875 struct worker_pool
*pool
= worker
->pool
;
877 WARN_ON_ONCE(worker
->task
!= current
);
880 * If transitioning into NOT_RUNNING, adjust nr_running and
881 * wake up an idle worker as necessary if requested by
884 if ((flags
& WORKER_NOT_RUNNING
) &&
885 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
887 if (atomic_dec_and_test(&pool
->nr_running
) &&
888 !list_empty(&pool
->worklist
))
889 wake_up_worker(pool
);
891 atomic_dec(&pool
->nr_running
);
894 worker
->flags
|= flags
;
898 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
900 * @flags: flags to clear
902 * Clear @flags in @worker->flags and adjust nr_running accordingly.
905 * spin_lock_irq(pool->lock)
907 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
909 struct worker_pool
*pool
= worker
->pool
;
910 unsigned int oflags
= worker
->flags
;
912 WARN_ON_ONCE(worker
->task
!= current
);
914 worker
->flags
&= ~flags
;
917 * If transitioning out of NOT_RUNNING, increment nr_running. Note
918 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
919 * of multiple flags, not a single flag.
921 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
922 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
923 atomic_inc(&pool
->nr_running
);
927 * find_worker_executing_work - find worker which is executing a work
928 * @pool: pool of interest
929 * @work: work to find worker for
931 * Find a worker which is executing @work on @pool by searching
932 * @pool->busy_hash which is keyed by the address of @work. For a worker
933 * to match, its current execution should match the address of @work and
934 * its work function. This is to avoid unwanted dependency between
935 * unrelated work executions through a work item being recycled while still
938 * This is a bit tricky. A work item may be freed once its execution
939 * starts and nothing prevents the freed area from being recycled for
940 * another work item. If the same work item address ends up being reused
941 * before the original execution finishes, workqueue will identify the
942 * recycled work item as currently executing and make it wait until the
943 * current execution finishes, introducing an unwanted dependency.
945 * This function checks the work item address and work function to avoid
946 * false positives. Note that this isn't complete as one may construct a
947 * work function which can introduce dependency onto itself through a
948 * recycled work item. Well, if somebody wants to shoot oneself in the
949 * foot that badly, there's only so much we can do, and if such deadlock
950 * actually occurs, it should be easy to locate the culprit work function.
953 * spin_lock_irq(pool->lock).
956 * Pointer to worker which is executing @work if found, NULL
959 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
960 struct work_struct
*work
)
962 struct worker
*worker
;
964 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
966 if (worker
->current_work
== work
&&
967 worker
->current_func
== work
->func
)
974 * move_linked_works - move linked works to a list
975 * @work: start of series of works to be scheduled
976 * @head: target list to append @work to
977 * @nextp: out paramter for nested worklist walking
979 * Schedule linked works starting from @work to @head. Work series to
980 * be scheduled starts at @work and includes any consecutive work with
981 * WORK_STRUCT_LINKED set in its predecessor.
983 * If @nextp is not NULL, it's updated to point to the next work of
984 * the last scheduled work. This allows move_linked_works() to be
985 * nested inside outer list_for_each_entry_safe().
988 * spin_lock_irq(pool->lock).
990 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
991 struct work_struct
**nextp
)
993 struct work_struct
*n
;
996 * Linked worklist will always end before the end of the list,
997 * use NULL for list head.
999 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1000 list_move_tail(&work
->entry
, head
);
1001 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1006 * If we're already inside safe list traversal and have moved
1007 * multiple works to the scheduled queue, the next position
1008 * needs to be updated.
1015 * get_pwq - get an extra reference on the specified pool_workqueue
1016 * @pwq: pool_workqueue to get
1018 * Obtain an extra reference on @pwq. The caller should guarantee that
1019 * @pwq has positive refcnt and be holding the matching pool->lock.
1021 static void get_pwq(struct pool_workqueue
*pwq
)
1023 lockdep_assert_held(&pwq
->pool
->lock
);
1024 WARN_ON_ONCE(pwq
->refcnt
<= 0);
1029 * put_pwq - put a pool_workqueue reference
1030 * @pwq: pool_workqueue to put
1032 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1033 * destruction. The caller should be holding the matching pool->lock.
1035 static void put_pwq(struct pool_workqueue
*pwq
)
1037 lockdep_assert_held(&pwq
->pool
->lock
);
1038 if (likely(--pwq
->refcnt
))
1040 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
1043 * @pwq can't be released under pool->lock, bounce to
1044 * pwq_unbound_release_workfn(). This never recurses on the same
1045 * pool->lock as this path is taken only for unbound workqueues and
1046 * the release work item is scheduled on a per-cpu workqueue. To
1047 * avoid lockdep warning, unbound pool->locks are given lockdep
1048 * subclass of 1 in get_unbound_pool().
1050 schedule_work(&pwq
->unbound_release_work
);
1054 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1055 * @pwq: pool_workqueue to put (can be %NULL)
1057 * put_pwq() with locking. This function also allows %NULL @pwq.
1059 static void put_pwq_unlocked(struct pool_workqueue
*pwq
)
1063 * As both pwqs and pools are sched-RCU protected, the
1064 * following lock operations are safe.
1066 spin_lock_irq(&pwq
->pool
->lock
);
1068 spin_unlock_irq(&pwq
->pool
->lock
);
1072 static void pwq_activate_delayed_work(struct work_struct
*work
)
1074 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1076 trace_workqueue_activate_work(work
);
1077 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1078 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1082 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1084 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1085 struct work_struct
, entry
);
1087 pwq_activate_delayed_work(work
);
1091 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1092 * @pwq: pwq of interest
1093 * @color: color of work which left the queue
1095 * A work either has completed or is removed from pending queue,
1096 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1099 * spin_lock_irq(pool->lock).
1101 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1103 /* uncolored work items don't participate in flushing or nr_active */
1104 if (color
== WORK_NO_COLOR
)
1107 pwq
->nr_in_flight
[color
]--;
1110 if (!list_empty(&pwq
->delayed_works
)) {
1111 /* one down, submit a delayed one */
1112 if (pwq
->nr_active
< pwq
->max_active
)
1113 pwq_activate_first_delayed(pwq
);
1116 /* is flush in progress and are we at the flushing tip? */
1117 if (likely(pwq
->flush_color
!= color
))
1120 /* are there still in-flight works? */
1121 if (pwq
->nr_in_flight
[color
])
1124 /* this pwq is done, clear flush_color */
1125 pwq
->flush_color
= -1;
1128 * If this was the last pwq, wake up the first flusher. It
1129 * will handle the rest.
1131 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1132 complete(&pwq
->wq
->first_flusher
->done
);
1138 * try_to_grab_pending - steal work item from worklist and disable irq
1139 * @work: work item to steal
1140 * @is_dwork: @work is a delayed_work
1141 * @flags: place to store irq state
1143 * Try to grab PENDING bit of @work. This function can handle @work in any
1144 * stable state - idle, on timer or on worklist. Return values are
1146 * 1 if @work was pending and we successfully stole PENDING
1147 * 0 if @work was idle and we claimed PENDING
1148 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1149 * -ENOENT if someone else is canceling @work, this state may persist
1150 * for arbitrarily long
1152 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1153 * interrupted while holding PENDING and @work off queue, irq must be
1154 * disabled on entry. This, combined with delayed_work->timer being
1155 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1157 * On successful return, >= 0, irq is disabled and the caller is
1158 * responsible for releasing it using local_irq_restore(*@flags).
1160 * This function is safe to call from any context including IRQ handler.
1162 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1163 unsigned long *flags
)
1165 struct worker_pool
*pool
;
1166 struct pool_workqueue
*pwq
;
1168 local_irq_save(*flags
);
1170 /* try to steal the timer if it exists */
1172 struct delayed_work
*dwork
= to_delayed_work(work
);
1175 * dwork->timer is irqsafe. If del_timer() fails, it's
1176 * guaranteed that the timer is not queued anywhere and not
1177 * running on the local CPU.
1179 if (likely(del_timer(&dwork
->timer
)))
1183 /* try to claim PENDING the normal way */
1184 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1188 * The queueing is in progress, or it is already queued. Try to
1189 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1191 pool
= get_work_pool(work
);
1195 spin_lock(&pool
->lock
);
1197 * work->data is guaranteed to point to pwq only while the work
1198 * item is queued on pwq->wq, and both updating work->data to point
1199 * to pwq on queueing and to pool on dequeueing are done under
1200 * pwq->pool->lock. This in turn guarantees that, if work->data
1201 * points to pwq which is associated with a locked pool, the work
1202 * item is currently queued on that pool.
1204 pwq
= get_work_pwq(work
);
1205 if (pwq
&& pwq
->pool
== pool
) {
1206 debug_work_deactivate(work
);
1209 * A delayed work item cannot be grabbed directly because
1210 * it might have linked NO_COLOR work items which, if left
1211 * on the delayed_list, will confuse pwq->nr_active
1212 * management later on and cause stall. Make sure the work
1213 * item is activated before grabbing.
1215 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1216 pwq_activate_delayed_work(work
);
1218 list_del_init(&work
->entry
);
1219 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1221 /* work->data points to pwq iff queued, point to pool */
1222 set_work_pool_and_keep_pending(work
, pool
->id
);
1224 spin_unlock(&pool
->lock
);
1227 spin_unlock(&pool
->lock
);
1229 local_irq_restore(*flags
);
1230 if (work_is_canceling(work
))
1237 * insert_work - insert a work into a pool
1238 * @pwq: pwq @work belongs to
1239 * @work: work to insert
1240 * @head: insertion point
1241 * @extra_flags: extra WORK_STRUCT_* flags to set
1243 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1244 * work_struct flags.
1247 * spin_lock_irq(pool->lock).
1249 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1250 struct list_head
*head
, unsigned int extra_flags
)
1252 struct worker_pool
*pool
= pwq
->pool
;
1254 /* we own @work, set data and link */
1255 set_work_pwq(work
, pwq
, extra_flags
);
1256 list_add_tail(&work
->entry
, head
);
1260 * Ensure either wq_worker_sleeping() sees the above
1261 * list_add_tail() or we see zero nr_running to avoid workers lying
1262 * around lazily while there are works to be processed.
1266 if (__need_more_worker(pool
))
1267 wake_up_worker(pool
);
1271 * Test whether @work is being queued from another work executing on the
1274 static bool is_chained_work(struct workqueue_struct
*wq
)
1276 struct worker
*worker
;
1278 worker
= current_wq_worker();
1280 * Return %true iff I'm a worker execuing a work item on @wq. If
1281 * I'm @worker, it's safe to dereference it without locking.
1283 return worker
&& worker
->current_pwq
->wq
== wq
;
1286 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1287 struct work_struct
*work
)
1289 struct pool_workqueue
*pwq
;
1290 struct worker_pool
*last_pool
;
1291 struct list_head
*worklist
;
1292 unsigned int work_flags
;
1293 unsigned int req_cpu
= cpu
;
1296 * While a work item is PENDING && off queue, a task trying to
1297 * steal the PENDING will busy-loop waiting for it to either get
1298 * queued or lose PENDING. Grabbing PENDING and queueing should
1299 * happen with IRQ disabled.
1301 WARN_ON_ONCE(!irqs_disabled());
1303 debug_work_activate(work
);
1305 /* if dying, only works from the same workqueue are allowed */
1306 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1307 WARN_ON_ONCE(!is_chained_work(wq
)))
1310 if (req_cpu
== WORK_CPU_UNBOUND
)
1311 cpu
= raw_smp_processor_id();
1313 /* pwq which will be used unless @work is executing elsewhere */
1314 if (!(wq
->flags
& WQ_UNBOUND
))
1315 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1317 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
1320 * If @work was previously on a different pool, it might still be
1321 * running there, in which case the work needs to be queued on that
1322 * pool to guarantee non-reentrancy.
1324 last_pool
= get_work_pool(work
);
1325 if (last_pool
&& last_pool
!= pwq
->pool
) {
1326 struct worker
*worker
;
1328 spin_lock(&last_pool
->lock
);
1330 worker
= find_worker_executing_work(last_pool
, work
);
1332 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1333 pwq
= worker
->current_pwq
;
1335 /* meh... not running there, queue here */
1336 spin_unlock(&last_pool
->lock
);
1337 spin_lock(&pwq
->pool
->lock
);
1340 spin_lock(&pwq
->pool
->lock
);
1344 * pwq is determined and locked. For unbound pools, we could have
1345 * raced with pwq release and it could already be dead. If its
1346 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1347 * without another pwq replacing it in the numa_pwq_tbl or while
1348 * work items are executing on it, so the retrying is guaranteed to
1349 * make forward-progress.
1351 if (unlikely(!pwq
->refcnt
)) {
1352 if (wq
->flags
& WQ_UNBOUND
) {
1353 spin_unlock(&pwq
->pool
->lock
);
1358 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1362 /* pwq determined, queue */
1363 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1365 if (WARN_ON(!list_empty(&work
->entry
))) {
1366 spin_unlock(&pwq
->pool
->lock
);
1370 pwq
->nr_in_flight
[pwq
->work_color
]++;
1371 work_flags
= work_color_to_flags(pwq
->work_color
);
1373 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1374 trace_workqueue_activate_work(work
);
1376 worklist
= &pwq
->pool
->worklist
;
1378 work_flags
|= WORK_STRUCT_DELAYED
;
1379 worklist
= &pwq
->delayed_works
;
1382 insert_work(pwq
, work
, worklist
, work_flags
);
1384 spin_unlock(&pwq
->pool
->lock
);
1388 * queue_work_on - queue work on specific cpu
1389 * @cpu: CPU number to execute work on
1390 * @wq: workqueue to use
1391 * @work: work to queue
1393 * Returns %false if @work was already on a queue, %true otherwise.
1395 * We queue the work to a specific CPU, the caller must ensure it
1398 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1399 struct work_struct
*work
)
1402 unsigned long flags
;
1404 local_irq_save(flags
);
1406 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1407 __queue_work(cpu
, wq
, work
);
1411 local_irq_restore(flags
);
1414 EXPORT_SYMBOL(queue_work_on
);
1416 void delayed_work_timer_fn(unsigned long __data
)
1418 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1420 /* should have been called from irqsafe timer with irq already off */
1421 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1423 EXPORT_SYMBOL(delayed_work_timer_fn
);
1425 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1426 struct delayed_work
*dwork
, unsigned long delay
)
1428 struct timer_list
*timer
= &dwork
->timer
;
1429 struct work_struct
*work
= &dwork
->work
;
1431 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1432 timer
->data
!= (unsigned long)dwork
);
1433 WARN_ON_ONCE(timer_pending(timer
));
1434 WARN_ON_ONCE(!list_empty(&work
->entry
));
1437 * If @delay is 0, queue @dwork->work immediately. This is for
1438 * both optimization and correctness. The earliest @timer can
1439 * expire is on the closest next tick and delayed_work users depend
1440 * on that there's no such delay when @delay is 0.
1443 __queue_work(cpu
, wq
, &dwork
->work
);
1447 timer_stats_timer_set_start_info(&dwork
->timer
);
1451 timer
->expires
= jiffies
+ delay
;
1453 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1454 add_timer_on(timer
, cpu
);
1460 * queue_delayed_work_on - queue work on specific CPU after delay
1461 * @cpu: CPU number to execute work on
1462 * @wq: workqueue to use
1463 * @dwork: work to queue
1464 * @delay: number of jiffies to wait before queueing
1466 * Returns %false if @work was already on a queue, %true otherwise. If
1467 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1470 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1471 struct delayed_work
*dwork
, unsigned long delay
)
1473 struct work_struct
*work
= &dwork
->work
;
1475 unsigned long flags
;
1477 /* read the comment in __queue_work() */
1478 local_irq_save(flags
);
1480 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1481 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1485 local_irq_restore(flags
);
1488 EXPORT_SYMBOL(queue_delayed_work_on
);
1491 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1492 * @cpu: CPU number to execute work on
1493 * @wq: workqueue to use
1494 * @dwork: work to queue
1495 * @delay: number of jiffies to wait before queueing
1497 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1498 * modify @dwork's timer so that it expires after @delay. If @delay is
1499 * zero, @work is guaranteed to be scheduled immediately regardless of its
1502 * Returns %false if @dwork was idle and queued, %true if @dwork was
1503 * pending and its timer was modified.
1505 * This function is safe to call from any context including IRQ handler.
1506 * See try_to_grab_pending() for details.
1508 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1509 struct delayed_work
*dwork
, unsigned long delay
)
1511 unsigned long flags
;
1515 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1516 } while (unlikely(ret
== -EAGAIN
));
1518 if (likely(ret
>= 0)) {
1519 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1520 local_irq_restore(flags
);
1523 /* -ENOENT from try_to_grab_pending() becomes %true */
1526 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1529 * worker_enter_idle - enter idle state
1530 * @worker: worker which is entering idle state
1532 * @worker is entering idle state. Update stats and idle timer if
1536 * spin_lock_irq(pool->lock).
1538 static void worker_enter_idle(struct worker
*worker
)
1540 struct worker_pool
*pool
= worker
->pool
;
1542 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1543 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1544 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1547 /* can't use worker_set_flags(), also called from start_worker() */
1548 worker
->flags
|= WORKER_IDLE
;
1550 worker
->last_active
= jiffies
;
1552 /* idle_list is LIFO */
1553 list_add(&worker
->entry
, &pool
->idle_list
);
1555 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1556 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1559 * Sanity check nr_running. Because wq_unbind_fn() releases
1560 * pool->lock between setting %WORKER_UNBOUND and zapping
1561 * nr_running, the warning may trigger spuriously. Check iff
1562 * unbind is not in progress.
1564 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1565 pool
->nr_workers
== pool
->nr_idle
&&
1566 atomic_read(&pool
->nr_running
));
1570 * worker_leave_idle - leave idle state
1571 * @worker: worker which is leaving idle state
1573 * @worker is leaving idle state. Update stats.
1576 * spin_lock_irq(pool->lock).
1578 static void worker_leave_idle(struct worker
*worker
)
1580 struct worker_pool
*pool
= worker
->pool
;
1582 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1584 worker_clr_flags(worker
, WORKER_IDLE
);
1586 list_del_init(&worker
->entry
);
1590 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1591 * @pool: target worker_pool
1593 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1595 * Works which are scheduled while the cpu is online must at least be
1596 * scheduled to a worker which is bound to the cpu so that if they are
1597 * flushed from cpu callbacks while cpu is going down, they are
1598 * guaranteed to execute on the cpu.
1600 * This function is to be used by unbound workers and rescuers to bind
1601 * themselves to the target cpu and may race with cpu going down or
1602 * coming online. kthread_bind() can't be used because it may put the
1603 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1604 * verbatim as it's best effort and blocking and pool may be
1605 * [dis]associated in the meantime.
1607 * This function tries set_cpus_allowed() and locks pool and verifies the
1608 * binding against %POOL_DISASSOCIATED which is set during
1609 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1610 * enters idle state or fetches works without dropping lock, it can
1611 * guarantee the scheduling requirement described in the first paragraph.
1614 * Might sleep. Called without any lock but returns with pool->lock
1618 * %true if the associated pool is online (@worker is successfully
1619 * bound), %false if offline.
1621 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1622 __acquires(&pool
->lock
)
1626 * The following call may fail, succeed or succeed
1627 * without actually migrating the task to the cpu if
1628 * it races with cpu hotunplug operation. Verify
1629 * against POOL_DISASSOCIATED.
1631 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1632 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1634 spin_lock_irq(&pool
->lock
);
1635 if (pool
->flags
& POOL_DISASSOCIATED
)
1637 if (task_cpu(current
) == pool
->cpu
&&
1638 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1640 spin_unlock_irq(&pool
->lock
);
1643 * We've raced with CPU hot[un]plug. Give it a breather
1644 * and retry migration. cond_resched() is required here;
1645 * otherwise, we might deadlock against cpu_stop trying to
1646 * bring down the CPU on non-preemptive kernel.
1653 static struct worker
*alloc_worker(void)
1655 struct worker
*worker
;
1657 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1659 INIT_LIST_HEAD(&worker
->entry
);
1660 INIT_LIST_HEAD(&worker
->scheduled
);
1661 /* on creation a worker is in !idle && prep state */
1662 worker
->flags
= WORKER_PREP
;
1668 * create_worker - create a new workqueue worker
1669 * @pool: pool the new worker will belong to
1671 * Create a new worker which is bound to @pool. The returned worker
1672 * can be started by calling start_worker() or destroyed using
1676 * Might sleep. Does GFP_KERNEL allocations.
1679 * Pointer to the newly created worker.
1681 static struct worker
*create_worker(struct worker_pool
*pool
)
1683 struct worker
*worker
= NULL
;
1687 lockdep_assert_held(&pool
->manager_mutex
);
1690 * ID is needed to determine kthread name. Allocate ID first
1691 * without installing the pointer.
1693 idr_preload(GFP_KERNEL
);
1694 spin_lock_irq(&pool
->lock
);
1696 id
= idr_alloc(&pool
->worker_idr
, NULL
, 0, 0, GFP_NOWAIT
);
1698 spin_unlock_irq(&pool
->lock
);
1703 worker
= alloc_worker();
1707 worker
->pool
= pool
;
1711 snprintf(id_buf
, sizeof(id_buf
), "%d:%d%s", pool
->cpu
, id
,
1712 pool
->attrs
->nice
< 0 ? "H" : "");
1714 snprintf(id_buf
, sizeof(id_buf
), "u%d:%d", pool
->id
, id
);
1716 worker
->task
= kthread_create_on_node(worker_thread
, worker
, pool
->node
,
1717 "kworker/%s", id_buf
);
1718 if (IS_ERR(worker
->task
))
1722 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723 * online CPUs. It'll be re-applied when any of the CPUs come up.
1725 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1726 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1728 /* prevent userland from meddling with cpumask of workqueue workers */
1729 worker
->task
->flags
|= PF_NO_SETAFFINITY
;
1732 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1733 * remains stable across this function. See the comments above the
1734 * flag definition for details.
1736 if (pool
->flags
& POOL_DISASSOCIATED
)
1737 worker
->flags
|= WORKER_UNBOUND
;
1739 /* successful, commit the pointer to idr */
1740 spin_lock_irq(&pool
->lock
);
1741 idr_replace(&pool
->worker_idr
, worker
, worker
->id
);
1742 spin_unlock_irq(&pool
->lock
);
1748 spin_lock_irq(&pool
->lock
);
1749 idr_remove(&pool
->worker_idr
, id
);
1750 spin_unlock_irq(&pool
->lock
);
1757 * start_worker - start a newly created worker
1758 * @worker: worker to start
1760 * Make the pool aware of @worker and start it.
1763 * spin_lock_irq(pool->lock).
1765 static void start_worker(struct worker
*worker
)
1767 worker
->flags
|= WORKER_STARTED
;
1768 worker
->pool
->nr_workers
++;
1769 worker_enter_idle(worker
);
1770 wake_up_process(worker
->task
);
1774 * create_and_start_worker - create and start a worker for a pool
1775 * @pool: the target pool
1777 * Grab the managership of @pool and create and start a new worker for it.
1779 static int create_and_start_worker(struct worker_pool
*pool
)
1781 struct worker
*worker
;
1783 mutex_lock(&pool
->manager_mutex
);
1785 worker
= create_worker(pool
);
1787 spin_lock_irq(&pool
->lock
);
1788 start_worker(worker
);
1789 spin_unlock_irq(&pool
->lock
);
1792 mutex_unlock(&pool
->manager_mutex
);
1794 return worker
? 0 : -ENOMEM
;
1798 * destroy_worker - destroy a workqueue worker
1799 * @worker: worker to be destroyed
1801 * Destroy @worker and adjust @pool stats accordingly.
1804 * spin_lock_irq(pool->lock) which is released and regrabbed.
1806 static void destroy_worker(struct worker
*worker
)
1808 struct worker_pool
*pool
= worker
->pool
;
1810 lockdep_assert_held(&pool
->manager_mutex
);
1811 lockdep_assert_held(&pool
->lock
);
1813 /* sanity check frenzy */
1814 if (WARN_ON(worker
->current_work
) ||
1815 WARN_ON(!list_empty(&worker
->scheduled
)))
1818 if (worker
->flags
& WORKER_STARTED
)
1820 if (worker
->flags
& WORKER_IDLE
)
1823 list_del_init(&worker
->entry
);
1824 worker
->flags
|= WORKER_DIE
;
1826 idr_remove(&pool
->worker_idr
, worker
->id
);
1828 spin_unlock_irq(&pool
->lock
);
1830 kthread_stop(worker
->task
);
1833 spin_lock_irq(&pool
->lock
);
1836 static void idle_worker_timeout(unsigned long __pool
)
1838 struct worker_pool
*pool
= (void *)__pool
;
1840 spin_lock_irq(&pool
->lock
);
1842 if (too_many_workers(pool
)) {
1843 struct worker
*worker
;
1844 unsigned long expires
;
1846 /* idle_list is kept in LIFO order, check the last one */
1847 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1848 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1850 if (time_before(jiffies
, expires
))
1851 mod_timer(&pool
->idle_timer
, expires
);
1853 /* it's been idle for too long, wake up manager */
1854 pool
->flags
|= POOL_MANAGE_WORKERS
;
1855 wake_up_worker(pool
);
1859 spin_unlock_irq(&pool
->lock
);
1862 static void send_mayday(struct work_struct
*work
)
1864 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1865 struct workqueue_struct
*wq
= pwq
->wq
;
1867 lockdep_assert_held(&wq_mayday_lock
);
1872 /* mayday mayday mayday */
1873 if (list_empty(&pwq
->mayday_node
)) {
1874 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1875 wake_up_process(wq
->rescuer
->task
);
1879 static void pool_mayday_timeout(unsigned long __pool
)
1881 struct worker_pool
*pool
= (void *)__pool
;
1882 struct work_struct
*work
;
1884 spin_lock_irq(&wq_mayday_lock
); /* for wq->maydays */
1885 spin_lock(&pool
->lock
);
1887 if (need_to_create_worker(pool
)) {
1889 * We've been trying to create a new worker but
1890 * haven't been successful. We might be hitting an
1891 * allocation deadlock. Send distress signals to
1894 list_for_each_entry(work
, &pool
->worklist
, entry
)
1898 spin_unlock(&pool
->lock
);
1899 spin_unlock_irq(&wq_mayday_lock
);
1901 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1905 * maybe_create_worker - create a new worker if necessary
1906 * @pool: pool to create a new worker for
1908 * Create a new worker for @pool if necessary. @pool is guaranteed to
1909 * have at least one idle worker on return from this function. If
1910 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1911 * sent to all rescuers with works scheduled on @pool to resolve
1912 * possible allocation deadlock.
1914 * On return, need_to_create_worker() is guaranteed to be %false and
1915 * may_start_working() %true.
1918 * spin_lock_irq(pool->lock) which may be released and regrabbed
1919 * multiple times. Does GFP_KERNEL allocations. Called only from
1923 * %false if no action was taken and pool->lock stayed locked, %true
1926 static bool maybe_create_worker(struct worker_pool
*pool
)
1927 __releases(&pool
->lock
)
1928 __acquires(&pool
->lock
)
1930 if (!need_to_create_worker(pool
))
1933 spin_unlock_irq(&pool
->lock
);
1935 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1936 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1939 struct worker
*worker
;
1941 worker
= create_worker(pool
);
1943 del_timer_sync(&pool
->mayday_timer
);
1944 spin_lock_irq(&pool
->lock
);
1945 start_worker(worker
);
1946 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1951 if (!need_to_create_worker(pool
))
1954 __set_current_state(TASK_INTERRUPTIBLE
);
1955 schedule_timeout(CREATE_COOLDOWN
);
1957 if (!need_to_create_worker(pool
))
1961 del_timer_sync(&pool
->mayday_timer
);
1962 spin_lock_irq(&pool
->lock
);
1963 if (need_to_create_worker(pool
))
1969 * maybe_destroy_worker - destroy workers which have been idle for a while
1970 * @pool: pool to destroy workers for
1972 * Destroy @pool workers which have been idle for longer than
1973 * IDLE_WORKER_TIMEOUT.
1976 * spin_lock_irq(pool->lock) which may be released and regrabbed
1977 * multiple times. Called only from manager.
1980 * %false if no action was taken and pool->lock stayed locked, %true
1983 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1987 while (too_many_workers(pool
)) {
1988 struct worker
*worker
;
1989 unsigned long expires
;
1991 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1992 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1994 if (time_before(jiffies
, expires
)) {
1995 mod_timer(&pool
->idle_timer
, expires
);
1999 destroy_worker(worker
);
2007 * manage_workers - manage worker pool
2010 * Assume the manager role and manage the worker pool @worker belongs
2011 * to. At any given time, there can be only zero or one manager per
2012 * pool. The exclusion is handled automatically by this function.
2014 * The caller can safely start processing works on false return. On
2015 * true return, it's guaranteed that need_to_create_worker() is false
2016 * and may_start_working() is true.
2019 * spin_lock_irq(pool->lock) which may be released and regrabbed
2020 * multiple times. Does GFP_KERNEL allocations.
2023 * spin_lock_irq(pool->lock) which may be released and regrabbed
2024 * multiple times. Does GFP_KERNEL allocations.
2026 static bool manage_workers(struct worker
*worker
)
2028 struct worker_pool
*pool
= worker
->pool
;
2032 * Managership is governed by two mutexes - manager_arb and
2033 * manager_mutex. manager_arb handles arbitration of manager role.
2034 * Anyone who successfully grabs manager_arb wins the arbitration
2035 * and becomes the manager. mutex_trylock() on pool->manager_arb
2036 * failure while holding pool->lock reliably indicates that someone
2037 * else is managing the pool and the worker which failed trylock
2038 * can proceed to executing work items. This means that anyone
2039 * grabbing manager_arb is responsible for actually performing
2040 * manager duties. If manager_arb is grabbed and released without
2041 * actual management, the pool may stall indefinitely.
2043 * manager_mutex is used for exclusion of actual management
2044 * operations. The holder of manager_mutex can be sure that none
2045 * of management operations, including creation and destruction of
2046 * workers, won't take place until the mutex is released. Because
2047 * manager_mutex doesn't interfere with manager role arbitration,
2048 * it is guaranteed that the pool's management, while may be
2049 * delayed, won't be disturbed by someone else grabbing
2052 if (!mutex_trylock(&pool
->manager_arb
))
2056 * With manager arbitration won, manager_mutex would be free in
2057 * most cases. trylock first without dropping @pool->lock.
2059 if (unlikely(!mutex_trylock(&pool
->manager_mutex
))) {
2060 spin_unlock_irq(&pool
->lock
);
2061 mutex_lock(&pool
->manager_mutex
);
2062 spin_lock_irq(&pool
->lock
);
2066 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2069 * Destroy and then create so that may_start_working() is true
2072 ret
|= maybe_destroy_workers(pool
);
2073 ret
|= maybe_create_worker(pool
);
2075 mutex_unlock(&pool
->manager_mutex
);
2076 mutex_unlock(&pool
->manager_arb
);
2081 * process_one_work - process single work
2083 * @work: work to process
2085 * Process @work. This function contains all the logics necessary to
2086 * process a single work including synchronization against and
2087 * interaction with other workers on the same cpu, queueing and
2088 * flushing. As long as context requirement is met, any worker can
2089 * call this function to process a work.
2092 * spin_lock_irq(pool->lock) which is released and regrabbed.
2094 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2095 __releases(&pool
->lock
)
2096 __acquires(&pool
->lock
)
2098 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2099 struct worker_pool
*pool
= worker
->pool
;
2100 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2102 struct worker
*collision
;
2103 #ifdef CONFIG_LOCKDEP
2105 * It is permissible to free the struct work_struct from
2106 * inside the function that is called from it, this we need to
2107 * take into account for lockdep too. To avoid bogus "held
2108 * lock freed" warnings as well as problems when looking into
2109 * work->lockdep_map, make a copy and use that here.
2111 struct lockdep_map lockdep_map
;
2113 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2116 * Ensure we're on the correct CPU. DISASSOCIATED test is
2117 * necessary to avoid spurious warnings from rescuers servicing the
2118 * unbound or a disassociated pool.
2120 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2121 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2122 raw_smp_processor_id() != pool
->cpu
);
2125 * A single work shouldn't be executed concurrently by
2126 * multiple workers on a single cpu. Check whether anyone is
2127 * already processing the work. If so, defer the work to the
2128 * currently executing one.
2130 collision
= find_worker_executing_work(pool
, work
);
2131 if (unlikely(collision
)) {
2132 move_linked_works(work
, &collision
->scheduled
, NULL
);
2136 /* claim and dequeue */
2137 debug_work_deactivate(work
);
2138 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2139 worker
->current_work
= work
;
2140 worker
->current_func
= work
->func
;
2141 worker
->current_pwq
= pwq
;
2142 work_color
= get_work_color(work
);
2144 list_del_init(&work
->entry
);
2147 * CPU intensive works don't participate in concurrency
2148 * management. They're the scheduler's responsibility.
2150 if (unlikely(cpu_intensive
))
2151 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2154 * Unbound pool isn't concurrency managed and work items should be
2155 * executed ASAP. Wake up another worker if necessary.
2157 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2158 wake_up_worker(pool
);
2161 * Record the last pool and clear PENDING which should be the last
2162 * update to @work. Also, do this inside @pool->lock so that
2163 * PENDING and queued state changes happen together while IRQ is
2166 set_work_pool_and_clear_pending(work
, pool
->id
);
2168 spin_unlock_irq(&pool
->lock
);
2170 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2171 lock_map_acquire(&lockdep_map
);
2172 trace_workqueue_execute_start(work
);
2173 worker
->current_func(work
);
2175 * While we must be careful to not use "work" after this, the trace
2176 * point will only record its address.
2178 trace_workqueue_execute_end(work
);
2179 lock_map_release(&lockdep_map
);
2180 lock_map_release(&pwq
->wq
->lockdep_map
);
2182 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2183 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2184 " last function: %pf\n",
2185 current
->comm
, preempt_count(), task_pid_nr(current
),
2186 worker
->current_func
);
2187 debug_show_held_locks(current
);
2191 spin_lock_irq(&pool
->lock
);
2193 /* clear cpu intensive status */
2194 if (unlikely(cpu_intensive
))
2195 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2197 /* we're done with it, release */
2198 hash_del(&worker
->hentry
);
2199 worker
->current_work
= NULL
;
2200 worker
->current_func
= NULL
;
2201 worker
->current_pwq
= NULL
;
2202 worker
->desc_valid
= false;
2203 pwq_dec_nr_in_flight(pwq
, work_color
);
2207 * process_scheduled_works - process scheduled works
2210 * Process all scheduled works. Please note that the scheduled list
2211 * may change while processing a work, so this function repeatedly
2212 * fetches a work from the top and executes it.
2215 * spin_lock_irq(pool->lock) which may be released and regrabbed
2218 static void process_scheduled_works(struct worker
*worker
)
2220 while (!list_empty(&worker
->scheduled
)) {
2221 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2222 struct work_struct
, entry
);
2223 process_one_work(worker
, work
);
2228 * worker_thread - the worker thread function
2231 * The worker thread function. All workers belong to a worker_pool -
2232 * either a per-cpu one or dynamic unbound one. These workers process all
2233 * work items regardless of their specific target workqueue. The only
2234 * exception is work items which belong to workqueues with a rescuer which
2235 * will be explained in rescuer_thread().
2237 static int worker_thread(void *__worker
)
2239 struct worker
*worker
= __worker
;
2240 struct worker_pool
*pool
= worker
->pool
;
2242 /* tell the scheduler that this is a workqueue worker */
2243 worker
->task
->flags
|= PF_WQ_WORKER
;
2245 spin_lock_irq(&pool
->lock
);
2247 /* am I supposed to die? */
2248 if (unlikely(worker
->flags
& WORKER_DIE
)) {
2249 spin_unlock_irq(&pool
->lock
);
2250 WARN_ON_ONCE(!list_empty(&worker
->entry
));
2251 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2255 worker_leave_idle(worker
);
2257 /* no more worker necessary? */
2258 if (!need_more_worker(pool
))
2261 /* do we need to manage? */
2262 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2266 * ->scheduled list can only be filled while a worker is
2267 * preparing to process a work or actually processing it.
2268 * Make sure nobody diddled with it while I was sleeping.
2270 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2273 * Finish PREP stage. We're guaranteed to have at least one idle
2274 * worker or that someone else has already assumed the manager
2275 * role. This is where @worker starts participating in concurrency
2276 * management if applicable and concurrency management is restored
2277 * after being rebound. See rebind_workers() for details.
2279 worker_clr_flags(worker
, WORKER_PREP
| WORKER_REBOUND
);
2282 struct work_struct
*work
=
2283 list_first_entry(&pool
->worklist
,
2284 struct work_struct
, entry
);
2286 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2287 /* optimization path, not strictly necessary */
2288 process_one_work(worker
, work
);
2289 if (unlikely(!list_empty(&worker
->scheduled
)))
2290 process_scheduled_works(worker
);
2292 move_linked_works(work
, &worker
->scheduled
, NULL
);
2293 process_scheduled_works(worker
);
2295 } while (keep_working(pool
));
2297 worker_set_flags(worker
, WORKER_PREP
, false);
2299 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2303 * pool->lock is held and there's no work to process and no need to
2304 * manage, sleep. Workers are woken up only while holding
2305 * pool->lock or from local cpu, so setting the current state
2306 * before releasing pool->lock is enough to prevent losing any
2309 worker_enter_idle(worker
);
2310 __set_current_state(TASK_INTERRUPTIBLE
);
2311 spin_unlock_irq(&pool
->lock
);
2317 * rescuer_thread - the rescuer thread function
2320 * Workqueue rescuer thread function. There's one rescuer for each
2321 * workqueue which has WQ_MEM_RECLAIM set.
2323 * Regular work processing on a pool may block trying to create a new
2324 * worker which uses GFP_KERNEL allocation which has slight chance of
2325 * developing into deadlock if some works currently on the same queue
2326 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2327 * the problem rescuer solves.
2329 * When such condition is possible, the pool summons rescuers of all
2330 * workqueues which have works queued on the pool and let them process
2331 * those works so that forward progress can be guaranteed.
2333 * This should happen rarely.
2335 static int rescuer_thread(void *__rescuer
)
2337 struct worker
*rescuer
= __rescuer
;
2338 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2339 struct list_head
*scheduled
= &rescuer
->scheduled
;
2341 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2344 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2345 * doesn't participate in concurrency management.
2347 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2349 set_current_state(TASK_INTERRUPTIBLE
);
2351 if (kthread_should_stop()) {
2352 __set_current_state(TASK_RUNNING
);
2353 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2357 /* see whether any pwq is asking for help */
2358 spin_lock_irq(&wq_mayday_lock
);
2360 while (!list_empty(&wq
->maydays
)) {
2361 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2362 struct pool_workqueue
, mayday_node
);
2363 struct worker_pool
*pool
= pwq
->pool
;
2364 struct work_struct
*work
, *n
;
2366 __set_current_state(TASK_RUNNING
);
2367 list_del_init(&pwq
->mayday_node
);
2369 spin_unlock_irq(&wq_mayday_lock
);
2371 /* migrate to the target cpu if possible */
2372 worker_maybe_bind_and_lock(pool
);
2373 rescuer
->pool
= pool
;
2376 * Slurp in all works issued via this workqueue and
2379 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2380 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2381 if (get_work_pwq(work
) == pwq
)
2382 move_linked_works(work
, scheduled
, &n
);
2384 process_scheduled_works(rescuer
);
2387 * Leave this pool. If keep_working() is %true, notify a
2388 * regular worker; otherwise, we end up with 0 concurrency
2389 * and stalling the execution.
2391 if (keep_working(pool
))
2392 wake_up_worker(pool
);
2394 rescuer
->pool
= NULL
;
2395 spin_unlock(&pool
->lock
);
2396 spin_lock(&wq_mayday_lock
);
2399 spin_unlock_irq(&wq_mayday_lock
);
2401 /* rescuers should never participate in concurrency management */
2402 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2408 struct work_struct work
;
2409 struct completion done
;
2412 static void wq_barrier_func(struct work_struct
*work
)
2414 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2415 complete(&barr
->done
);
2419 * insert_wq_barrier - insert a barrier work
2420 * @pwq: pwq to insert barrier into
2421 * @barr: wq_barrier to insert
2422 * @target: target work to attach @barr to
2423 * @worker: worker currently executing @target, NULL if @target is not executing
2425 * @barr is linked to @target such that @barr is completed only after
2426 * @target finishes execution. Please note that the ordering
2427 * guarantee is observed only with respect to @target and on the local
2430 * Currently, a queued barrier can't be canceled. This is because
2431 * try_to_grab_pending() can't determine whether the work to be
2432 * grabbed is at the head of the queue and thus can't clear LINKED
2433 * flag of the previous work while there must be a valid next work
2434 * after a work with LINKED flag set.
2436 * Note that when @worker is non-NULL, @target may be modified
2437 * underneath us, so we can't reliably determine pwq from @target.
2440 * spin_lock_irq(pool->lock).
2442 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2443 struct wq_barrier
*barr
,
2444 struct work_struct
*target
, struct worker
*worker
)
2446 struct list_head
*head
;
2447 unsigned int linked
= 0;
2450 * debugobject calls are safe here even with pool->lock locked
2451 * as we know for sure that this will not trigger any of the
2452 * checks and call back into the fixup functions where we
2455 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2456 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2457 init_completion(&barr
->done
);
2460 * If @target is currently being executed, schedule the
2461 * barrier to the worker; otherwise, put it after @target.
2464 head
= worker
->scheduled
.next
;
2466 unsigned long *bits
= work_data_bits(target
);
2468 head
= target
->entry
.next
;
2469 /* there can already be other linked works, inherit and set */
2470 linked
= *bits
& WORK_STRUCT_LINKED
;
2471 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2474 debug_work_activate(&barr
->work
);
2475 insert_work(pwq
, &barr
->work
, head
,
2476 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2480 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2481 * @wq: workqueue being flushed
2482 * @flush_color: new flush color, < 0 for no-op
2483 * @work_color: new work color, < 0 for no-op
2485 * Prepare pwqs for workqueue flushing.
2487 * If @flush_color is non-negative, flush_color on all pwqs should be
2488 * -1. If no pwq has in-flight commands at the specified color, all
2489 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2490 * has in flight commands, its pwq->flush_color is set to
2491 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2492 * wakeup logic is armed and %true is returned.
2494 * The caller should have initialized @wq->first_flusher prior to
2495 * calling this function with non-negative @flush_color. If
2496 * @flush_color is negative, no flush color update is done and %false
2499 * If @work_color is non-negative, all pwqs should have the same
2500 * work_color which is previous to @work_color and all will be
2501 * advanced to @work_color.
2504 * mutex_lock(wq->mutex).
2507 * %true if @flush_color >= 0 and there's something to flush. %false
2510 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2511 int flush_color
, int work_color
)
2514 struct pool_workqueue
*pwq
;
2516 if (flush_color
>= 0) {
2517 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2518 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2521 for_each_pwq(pwq
, wq
) {
2522 struct worker_pool
*pool
= pwq
->pool
;
2524 spin_lock_irq(&pool
->lock
);
2526 if (flush_color
>= 0) {
2527 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2529 if (pwq
->nr_in_flight
[flush_color
]) {
2530 pwq
->flush_color
= flush_color
;
2531 atomic_inc(&wq
->nr_pwqs_to_flush
);
2536 if (work_color
>= 0) {
2537 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2538 pwq
->work_color
= work_color
;
2541 spin_unlock_irq(&pool
->lock
);
2544 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2545 complete(&wq
->first_flusher
->done
);
2551 * flush_workqueue - ensure that any scheduled work has run to completion.
2552 * @wq: workqueue to flush
2554 * This function sleeps until all work items which were queued on entry
2555 * have finished execution, but it is not livelocked by new incoming ones.
2557 void flush_workqueue(struct workqueue_struct
*wq
)
2559 struct wq_flusher this_flusher
= {
2560 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2562 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2566 lock_map_acquire(&wq
->lockdep_map
);
2567 lock_map_release(&wq
->lockdep_map
);
2569 mutex_lock(&wq
->mutex
);
2572 * Start-to-wait phase
2574 next_color
= work_next_color(wq
->work_color
);
2576 if (next_color
!= wq
->flush_color
) {
2578 * Color space is not full. The current work_color
2579 * becomes our flush_color and work_color is advanced
2582 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2583 this_flusher
.flush_color
= wq
->work_color
;
2584 wq
->work_color
= next_color
;
2586 if (!wq
->first_flusher
) {
2587 /* no flush in progress, become the first flusher */
2588 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2590 wq
->first_flusher
= &this_flusher
;
2592 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2594 /* nothing to flush, done */
2595 wq
->flush_color
= next_color
;
2596 wq
->first_flusher
= NULL
;
2601 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2602 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2603 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2607 * Oops, color space is full, wait on overflow queue.
2608 * The next flush completion will assign us
2609 * flush_color and transfer to flusher_queue.
2611 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2614 mutex_unlock(&wq
->mutex
);
2616 wait_for_completion(&this_flusher
.done
);
2619 * Wake-up-and-cascade phase
2621 * First flushers are responsible for cascading flushes and
2622 * handling overflow. Non-first flushers can simply return.
2624 if (wq
->first_flusher
!= &this_flusher
)
2627 mutex_lock(&wq
->mutex
);
2629 /* we might have raced, check again with mutex held */
2630 if (wq
->first_flusher
!= &this_flusher
)
2633 wq
->first_flusher
= NULL
;
2635 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2636 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2639 struct wq_flusher
*next
, *tmp
;
2641 /* complete all the flushers sharing the current flush color */
2642 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2643 if (next
->flush_color
!= wq
->flush_color
)
2645 list_del_init(&next
->list
);
2646 complete(&next
->done
);
2649 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2650 wq
->flush_color
!= work_next_color(wq
->work_color
));
2652 /* this flush_color is finished, advance by one */
2653 wq
->flush_color
= work_next_color(wq
->flush_color
);
2655 /* one color has been freed, handle overflow queue */
2656 if (!list_empty(&wq
->flusher_overflow
)) {
2658 * Assign the same color to all overflowed
2659 * flushers, advance work_color and append to
2660 * flusher_queue. This is the start-to-wait
2661 * phase for these overflowed flushers.
2663 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2664 tmp
->flush_color
= wq
->work_color
;
2666 wq
->work_color
= work_next_color(wq
->work_color
);
2668 list_splice_tail_init(&wq
->flusher_overflow
,
2669 &wq
->flusher_queue
);
2670 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2673 if (list_empty(&wq
->flusher_queue
)) {
2674 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2679 * Need to flush more colors. Make the next flusher
2680 * the new first flusher and arm pwqs.
2682 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2683 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2685 list_del_init(&next
->list
);
2686 wq
->first_flusher
= next
;
2688 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2692 * Meh... this color is already done, clear first
2693 * flusher and repeat cascading.
2695 wq
->first_flusher
= NULL
;
2699 mutex_unlock(&wq
->mutex
);
2701 EXPORT_SYMBOL_GPL(flush_workqueue
);
2704 * drain_workqueue - drain a workqueue
2705 * @wq: workqueue to drain
2707 * Wait until the workqueue becomes empty. While draining is in progress,
2708 * only chain queueing is allowed. IOW, only currently pending or running
2709 * work items on @wq can queue further work items on it. @wq is flushed
2710 * repeatedly until it becomes empty. The number of flushing is detemined
2711 * by the depth of chaining and should be relatively short. Whine if it
2714 void drain_workqueue(struct workqueue_struct
*wq
)
2716 unsigned int flush_cnt
= 0;
2717 struct pool_workqueue
*pwq
;
2720 * __queue_work() needs to test whether there are drainers, is much
2721 * hotter than drain_workqueue() and already looks at @wq->flags.
2722 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2724 mutex_lock(&wq
->mutex
);
2725 if (!wq
->nr_drainers
++)
2726 wq
->flags
|= __WQ_DRAINING
;
2727 mutex_unlock(&wq
->mutex
);
2729 flush_workqueue(wq
);
2731 mutex_lock(&wq
->mutex
);
2733 for_each_pwq(pwq
, wq
) {
2736 spin_lock_irq(&pwq
->pool
->lock
);
2737 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2738 spin_unlock_irq(&pwq
->pool
->lock
);
2743 if (++flush_cnt
== 10 ||
2744 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2745 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2746 wq
->name
, flush_cnt
);
2748 mutex_unlock(&wq
->mutex
);
2752 if (!--wq
->nr_drainers
)
2753 wq
->flags
&= ~__WQ_DRAINING
;
2754 mutex_unlock(&wq
->mutex
);
2756 EXPORT_SYMBOL_GPL(drain_workqueue
);
2758 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2760 struct worker
*worker
= NULL
;
2761 struct worker_pool
*pool
;
2762 struct pool_workqueue
*pwq
;
2766 local_irq_disable();
2767 pool
= get_work_pool(work
);
2773 spin_lock(&pool
->lock
);
2774 /* see the comment in try_to_grab_pending() with the same code */
2775 pwq
= get_work_pwq(work
);
2777 if (unlikely(pwq
->pool
!= pool
))
2780 worker
= find_worker_executing_work(pool
, work
);
2783 pwq
= worker
->current_pwq
;
2786 insert_wq_barrier(pwq
, barr
, work
, worker
);
2787 spin_unlock_irq(&pool
->lock
);
2790 * If @max_active is 1 or rescuer is in use, flushing another work
2791 * item on the same workqueue may lead to deadlock. Make sure the
2792 * flusher is not running on the same workqueue by verifying write
2795 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2796 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2798 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2799 lock_map_release(&pwq
->wq
->lockdep_map
);
2803 spin_unlock_irq(&pool
->lock
);
2808 * flush_work - wait for a work to finish executing the last queueing instance
2809 * @work: the work to flush
2811 * Wait until @work has finished execution. @work is guaranteed to be idle
2812 * on return if it hasn't been requeued since flush started.
2815 * %true if flush_work() waited for the work to finish execution,
2816 * %false if it was already idle.
2818 bool flush_work(struct work_struct
*work
)
2820 struct wq_barrier barr
;
2822 lock_map_acquire(&work
->lockdep_map
);
2823 lock_map_release(&work
->lockdep_map
);
2825 if (start_flush_work(work
, &barr
)) {
2826 wait_for_completion(&barr
.done
);
2827 destroy_work_on_stack(&barr
.work
);
2833 EXPORT_SYMBOL_GPL(flush_work
);
2835 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2837 unsigned long flags
;
2841 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2843 * If someone else is canceling, wait for the same event it
2844 * would be waiting for before retrying.
2846 if (unlikely(ret
== -ENOENT
))
2848 } while (unlikely(ret
< 0));
2850 /* tell other tasks trying to grab @work to back off */
2851 mark_work_canceling(work
);
2852 local_irq_restore(flags
);
2855 clear_work_data(work
);
2860 * cancel_work_sync - cancel a work and wait for it to finish
2861 * @work: the work to cancel
2863 * Cancel @work and wait for its execution to finish. This function
2864 * can be used even if the work re-queues itself or migrates to
2865 * another workqueue. On return from this function, @work is
2866 * guaranteed to be not pending or executing on any CPU.
2868 * cancel_work_sync(&delayed_work->work) must not be used for
2869 * delayed_work's. Use cancel_delayed_work_sync() instead.
2871 * The caller must ensure that the workqueue on which @work was last
2872 * queued can't be destroyed before this function returns.
2875 * %true if @work was pending, %false otherwise.
2877 bool cancel_work_sync(struct work_struct
*work
)
2879 return __cancel_work_timer(work
, false);
2881 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2884 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2885 * @dwork: the delayed work to flush
2887 * Delayed timer is cancelled and the pending work is queued for
2888 * immediate execution. Like flush_work(), this function only
2889 * considers the last queueing instance of @dwork.
2892 * %true if flush_work() waited for the work to finish execution,
2893 * %false if it was already idle.
2895 bool flush_delayed_work(struct delayed_work
*dwork
)
2897 local_irq_disable();
2898 if (del_timer_sync(&dwork
->timer
))
2899 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2901 return flush_work(&dwork
->work
);
2903 EXPORT_SYMBOL(flush_delayed_work
);
2906 * cancel_delayed_work - cancel a delayed work
2907 * @dwork: delayed_work to cancel
2909 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2910 * and canceled; %false if wasn't pending. Note that the work callback
2911 * function may still be running on return, unless it returns %true and the
2912 * work doesn't re-arm itself. Explicitly flush or use
2913 * cancel_delayed_work_sync() to wait on it.
2915 * This function is safe to call from any context including IRQ handler.
2917 bool cancel_delayed_work(struct delayed_work
*dwork
)
2919 unsigned long flags
;
2923 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2924 } while (unlikely(ret
== -EAGAIN
));
2926 if (unlikely(ret
< 0))
2929 set_work_pool_and_clear_pending(&dwork
->work
,
2930 get_work_pool_id(&dwork
->work
));
2931 local_irq_restore(flags
);
2934 EXPORT_SYMBOL(cancel_delayed_work
);
2937 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2938 * @dwork: the delayed work cancel
2940 * This is cancel_work_sync() for delayed works.
2943 * %true if @dwork was pending, %false otherwise.
2945 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2947 return __cancel_work_timer(&dwork
->work
, true);
2949 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2952 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2953 * @func: the function to call
2955 * schedule_on_each_cpu() executes @func on each online CPU using the
2956 * system workqueue and blocks until all CPUs have completed.
2957 * schedule_on_each_cpu() is very slow.
2960 * 0 on success, -errno on failure.
2962 int schedule_on_each_cpu(work_func_t func
)
2965 struct work_struct __percpu
*works
;
2967 works
= alloc_percpu(struct work_struct
);
2973 for_each_online_cpu(cpu
) {
2974 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2976 INIT_WORK(work
, func
);
2977 schedule_work_on(cpu
, work
);
2980 for_each_online_cpu(cpu
)
2981 flush_work(per_cpu_ptr(works
, cpu
));
2989 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2991 * Forces execution of the kernel-global workqueue and blocks until its
2994 * Think twice before calling this function! It's very easy to get into
2995 * trouble if you don't take great care. Either of the following situations
2996 * will lead to deadlock:
2998 * One of the work items currently on the workqueue needs to acquire
2999 * a lock held by your code or its caller.
3001 * Your code is running in the context of a work routine.
3003 * They will be detected by lockdep when they occur, but the first might not
3004 * occur very often. It depends on what work items are on the workqueue and
3005 * what locks they need, which you have no control over.
3007 * In most situations flushing the entire workqueue is overkill; you merely
3008 * need to know that a particular work item isn't queued and isn't running.
3009 * In such cases you should use cancel_delayed_work_sync() or
3010 * cancel_work_sync() instead.
3012 void flush_scheduled_work(void)
3014 flush_workqueue(system_wq
);
3016 EXPORT_SYMBOL(flush_scheduled_work
);
3019 * execute_in_process_context - reliably execute the routine with user context
3020 * @fn: the function to execute
3021 * @ew: guaranteed storage for the execute work structure (must
3022 * be available when the work executes)
3024 * Executes the function immediately if process context is available,
3025 * otherwise schedules the function for delayed execution.
3027 * Returns: 0 - function was executed
3028 * 1 - function was scheduled for execution
3030 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3032 if (!in_interrupt()) {
3037 INIT_WORK(&ew
->work
, fn
);
3038 schedule_work(&ew
->work
);
3042 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3046 * Workqueues with WQ_SYSFS flag set is visible to userland via
3047 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3048 * following attributes.
3050 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3051 * max_active RW int : maximum number of in-flight work items
3053 * Unbound workqueues have the following extra attributes.
3055 * id RO int : the associated pool ID
3056 * nice RW int : nice value of the workers
3057 * cpumask RW mask : bitmask of allowed CPUs for the workers
3060 struct workqueue_struct
*wq
;
3064 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3066 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3071 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3072 struct device_attribute
*attr
, char *buf
)
3074 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3076 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3079 static ssize_t
wq_max_active_show(struct device
*dev
,
3080 struct device_attribute
*attr
, char *buf
)
3082 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3084 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3087 static ssize_t
wq_max_active_store(struct device
*dev
,
3088 struct device_attribute
*attr
,
3089 const char *buf
, size_t count
)
3091 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3094 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3097 workqueue_set_max_active(wq
, val
);
3101 static struct device_attribute wq_sysfs_attrs
[] = {
3102 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3103 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3107 static ssize_t
wq_pool_ids_show(struct device
*dev
,
3108 struct device_attribute
*attr
, char *buf
)
3110 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3111 const char *delim
= "";
3112 int node
, written
= 0;
3114 rcu_read_lock_sched();
3115 for_each_node(node
) {
3116 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
,
3117 "%s%d:%d", delim
, node
,
3118 unbound_pwq_by_node(wq
, node
)->pool
->id
);
3121 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3122 rcu_read_unlock_sched();
3127 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3130 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3133 mutex_lock(&wq
->mutex
);
3134 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->unbound_attrs
->nice
);
3135 mutex_unlock(&wq
->mutex
);
3140 /* prepare workqueue_attrs for sysfs store operations */
3141 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3143 struct workqueue_attrs
*attrs
;
3145 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3149 mutex_lock(&wq
->mutex
);
3150 copy_workqueue_attrs(attrs
, wq
->unbound_attrs
);
3151 mutex_unlock(&wq
->mutex
);
3155 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3156 const char *buf
, size_t count
)
3158 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3159 struct workqueue_attrs
*attrs
;
3162 attrs
= wq_sysfs_prep_attrs(wq
);
3166 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3167 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3168 ret
= apply_workqueue_attrs(wq
, attrs
);
3172 free_workqueue_attrs(attrs
);
3173 return ret
?: count
;
3176 static ssize_t
wq_cpumask_show(struct device
*dev
,
3177 struct device_attribute
*attr
, char *buf
)
3179 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3182 mutex_lock(&wq
->mutex
);
3183 written
= cpumask_scnprintf(buf
, PAGE_SIZE
, wq
->unbound_attrs
->cpumask
);
3184 mutex_unlock(&wq
->mutex
);
3186 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3190 static ssize_t
wq_cpumask_store(struct device
*dev
,
3191 struct device_attribute
*attr
,
3192 const char *buf
, size_t count
)
3194 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3195 struct workqueue_attrs
*attrs
;
3198 attrs
= wq_sysfs_prep_attrs(wq
);
3202 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3204 ret
= apply_workqueue_attrs(wq
, attrs
);
3206 free_workqueue_attrs(attrs
);
3207 return ret
?: count
;
3210 static ssize_t
wq_numa_show(struct device
*dev
, struct device_attribute
*attr
,
3213 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3216 mutex_lock(&wq
->mutex
);
3217 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3218 !wq
->unbound_attrs
->no_numa
);
3219 mutex_unlock(&wq
->mutex
);
3224 static ssize_t
wq_numa_store(struct device
*dev
, struct device_attribute
*attr
,
3225 const char *buf
, size_t count
)
3227 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3228 struct workqueue_attrs
*attrs
;
3231 attrs
= wq_sysfs_prep_attrs(wq
);
3236 if (sscanf(buf
, "%d", &v
) == 1) {
3237 attrs
->no_numa
= !v
;
3238 ret
= apply_workqueue_attrs(wq
, attrs
);
3241 free_workqueue_attrs(attrs
);
3242 return ret
?: count
;
3245 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3246 __ATTR(pool_ids
, 0444, wq_pool_ids_show
, NULL
),
3247 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3248 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3249 __ATTR(numa
, 0644, wq_numa_show
, wq_numa_store
),
3253 static struct bus_type wq_subsys
= {
3254 .name
= "workqueue",
3255 .dev_attrs
= wq_sysfs_attrs
,
3258 static int __init
wq_sysfs_init(void)
3260 return subsys_virtual_register(&wq_subsys
, NULL
);
3262 core_initcall(wq_sysfs_init
);
3264 static void wq_device_release(struct device
*dev
)
3266 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3272 * workqueue_sysfs_register - make a workqueue visible in sysfs
3273 * @wq: the workqueue to register
3275 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3276 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3277 * which is the preferred method.
3279 * Workqueue user should use this function directly iff it wants to apply
3280 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3281 * apply_workqueue_attrs() may race against userland updating the
3284 * Returns 0 on success, -errno on failure.
3286 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3288 struct wq_device
*wq_dev
;
3292 * Adjusting max_active or creating new pwqs by applyting
3293 * attributes breaks ordering guarantee. Disallow exposing ordered
3296 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3299 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3304 wq_dev
->dev
.bus
= &wq_subsys
;
3305 wq_dev
->dev
.init_name
= wq
->name
;
3306 wq_dev
->dev
.release
= wq_device_release
;
3309 * unbound_attrs are created separately. Suppress uevent until
3310 * everything is ready.
3312 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3314 ret
= device_register(&wq_dev
->dev
);
3321 if (wq
->flags
& WQ_UNBOUND
) {
3322 struct device_attribute
*attr
;
3324 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3325 ret
= device_create_file(&wq_dev
->dev
, attr
);
3327 device_unregister(&wq_dev
->dev
);
3334 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3339 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3340 * @wq: the workqueue to unregister
3342 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3344 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3346 struct wq_device
*wq_dev
= wq
->wq_dev
;
3352 device_unregister(&wq_dev
->dev
);
3354 #else /* CONFIG_SYSFS */
3355 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3356 #endif /* CONFIG_SYSFS */
3359 * free_workqueue_attrs - free a workqueue_attrs
3360 * @attrs: workqueue_attrs to free
3362 * Undo alloc_workqueue_attrs().
3364 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3367 free_cpumask_var(attrs
->cpumask
);
3373 * alloc_workqueue_attrs - allocate a workqueue_attrs
3374 * @gfp_mask: allocation mask to use
3376 * Allocate a new workqueue_attrs, initialize with default settings and
3377 * return it. Returns NULL on failure.
3379 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3381 struct workqueue_attrs
*attrs
;
3383 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3386 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3389 cpumask_copy(attrs
->cpumask
, cpu_possible_mask
);
3392 free_workqueue_attrs(attrs
);
3396 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3397 const struct workqueue_attrs
*from
)
3399 to
->nice
= from
->nice
;
3400 cpumask_copy(to
->cpumask
, from
->cpumask
);
3402 * Unlike hash and equality test, this function doesn't ignore
3403 * ->no_numa as it is used for both pool and wq attrs. Instead,
3404 * get_unbound_pool() explicitly clears ->no_numa after copying.
3406 to
->no_numa
= from
->no_numa
;
3409 /* hash value of the content of @attr */
3410 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3414 hash
= jhash_1word(attrs
->nice
, hash
);
3415 hash
= jhash(cpumask_bits(attrs
->cpumask
),
3416 BITS_TO_LONGS(nr_cpumask_bits
) * sizeof(long), hash
);
3420 /* content equality test */
3421 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3422 const struct workqueue_attrs
*b
)
3424 if (a
->nice
!= b
->nice
)
3426 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3432 * init_worker_pool - initialize a newly zalloc'd worker_pool
3433 * @pool: worker_pool to initialize
3435 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3436 * Returns 0 on success, -errno on failure. Even on failure, all fields
3437 * inside @pool proper are initialized and put_unbound_pool() can be called
3438 * on @pool safely to release it.
3440 static int init_worker_pool(struct worker_pool
*pool
)
3442 spin_lock_init(&pool
->lock
);
3445 pool
->node
= NUMA_NO_NODE
;
3446 pool
->flags
|= POOL_DISASSOCIATED
;
3447 INIT_LIST_HEAD(&pool
->worklist
);
3448 INIT_LIST_HEAD(&pool
->idle_list
);
3449 hash_init(pool
->busy_hash
);
3451 init_timer_deferrable(&pool
->idle_timer
);
3452 pool
->idle_timer
.function
= idle_worker_timeout
;
3453 pool
->idle_timer
.data
= (unsigned long)pool
;
3455 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3456 (unsigned long)pool
);
3458 mutex_init(&pool
->manager_arb
);
3459 mutex_init(&pool
->manager_mutex
);
3460 idr_init(&pool
->worker_idr
);
3462 INIT_HLIST_NODE(&pool
->hash_node
);
3465 /* shouldn't fail above this point */
3466 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3472 static void rcu_free_pool(struct rcu_head
*rcu
)
3474 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3476 idr_destroy(&pool
->worker_idr
);
3477 free_workqueue_attrs(pool
->attrs
);
3482 * put_unbound_pool - put a worker_pool
3483 * @pool: worker_pool to put
3485 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3486 * safe manner. get_unbound_pool() calls this function on its failure path
3487 * and this function should be able to release pools which went through,
3488 * successfully or not, init_worker_pool().
3490 * Should be called with wq_pool_mutex held.
3492 static void put_unbound_pool(struct worker_pool
*pool
)
3494 struct worker
*worker
;
3496 lockdep_assert_held(&wq_pool_mutex
);
3502 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3503 WARN_ON(!list_empty(&pool
->worklist
)))
3506 /* release id and unhash */
3508 idr_remove(&worker_pool_idr
, pool
->id
);
3509 hash_del(&pool
->hash_node
);
3512 * Become the manager and destroy all workers. Grabbing
3513 * manager_arb prevents @pool's workers from blocking on
3516 mutex_lock(&pool
->manager_arb
);
3517 mutex_lock(&pool
->manager_mutex
);
3518 spin_lock_irq(&pool
->lock
);
3520 while ((worker
= first_worker(pool
)))
3521 destroy_worker(worker
);
3522 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3524 spin_unlock_irq(&pool
->lock
);
3525 mutex_unlock(&pool
->manager_mutex
);
3526 mutex_unlock(&pool
->manager_arb
);
3528 /* shut down the timers */
3529 del_timer_sync(&pool
->idle_timer
);
3530 del_timer_sync(&pool
->mayday_timer
);
3532 /* sched-RCU protected to allow dereferences from get_work_pool() */
3533 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3537 * get_unbound_pool - get a worker_pool with the specified attributes
3538 * @attrs: the attributes of the worker_pool to get
3540 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3541 * reference count and return it. If there already is a matching
3542 * worker_pool, it will be used; otherwise, this function attempts to
3543 * create a new one. On failure, returns NULL.
3545 * Should be called with wq_pool_mutex held.
3547 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3549 u32 hash
= wqattrs_hash(attrs
);
3550 struct worker_pool
*pool
;
3553 lockdep_assert_held(&wq_pool_mutex
);
3555 /* do we already have a matching pool? */
3556 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3557 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3563 /* nope, create a new one */
3564 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3565 if (!pool
|| init_worker_pool(pool
) < 0)
3568 if (workqueue_freezing
)
3569 pool
->flags
|= POOL_FREEZING
;
3571 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3572 copy_workqueue_attrs(pool
->attrs
, attrs
);
3575 * no_numa isn't a worker_pool attribute, always clear it. See
3576 * 'struct workqueue_attrs' comments for detail.
3578 pool
->attrs
->no_numa
= false;
3580 /* if cpumask is contained inside a NUMA node, we belong to that node */
3581 if (wq_numa_enabled
) {
3582 for_each_node(node
) {
3583 if (cpumask_subset(pool
->attrs
->cpumask
,
3584 wq_numa_possible_cpumask
[node
])) {
3591 if (worker_pool_assign_id(pool
) < 0)
3594 /* create and start the initial worker */
3595 if (create_and_start_worker(pool
) < 0)
3599 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3604 put_unbound_pool(pool
);
3608 static void rcu_free_pwq(struct rcu_head
*rcu
)
3610 kmem_cache_free(pwq_cache
,
3611 container_of(rcu
, struct pool_workqueue
, rcu
));
3615 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3616 * and needs to be destroyed.
3618 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3620 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3621 unbound_release_work
);
3622 struct workqueue_struct
*wq
= pwq
->wq
;
3623 struct worker_pool
*pool
= pwq
->pool
;
3626 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3630 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3631 * necessary on release but do it anyway. It's easier to verify
3632 * and consistent with the linking path.
3634 mutex_lock(&wq
->mutex
);
3635 list_del_rcu(&pwq
->pwqs_node
);
3636 is_last
= list_empty(&wq
->pwqs
);
3637 mutex_unlock(&wq
->mutex
);
3639 mutex_lock(&wq_pool_mutex
);
3640 put_unbound_pool(pool
);
3641 mutex_unlock(&wq_pool_mutex
);
3643 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3646 * If we're the last pwq going away, @wq is already dead and no one
3647 * is gonna access it anymore. Free it.
3650 free_workqueue_attrs(wq
->unbound_attrs
);
3656 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3657 * @pwq: target pool_workqueue
3659 * If @pwq isn't freezing, set @pwq->max_active to the associated
3660 * workqueue's saved_max_active and activate delayed work items
3661 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3663 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3665 struct workqueue_struct
*wq
= pwq
->wq
;
3666 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3668 /* for @wq->saved_max_active */
3669 lockdep_assert_held(&wq
->mutex
);
3671 /* fast exit for non-freezable wqs */
3672 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3675 spin_lock_irq(&pwq
->pool
->lock
);
3677 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3678 pwq
->max_active
= wq
->saved_max_active
;
3680 while (!list_empty(&pwq
->delayed_works
) &&
3681 pwq
->nr_active
< pwq
->max_active
)
3682 pwq_activate_first_delayed(pwq
);
3685 * Need to kick a worker after thawed or an unbound wq's
3686 * max_active is bumped. It's a slow path. Do it always.
3688 wake_up_worker(pwq
->pool
);
3690 pwq
->max_active
= 0;
3693 spin_unlock_irq(&pwq
->pool
->lock
);
3696 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3697 static void init_pwq(struct pool_workqueue
*pwq
, struct workqueue_struct
*wq
,
3698 struct worker_pool
*pool
)
3700 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3702 memset(pwq
, 0, sizeof(*pwq
));
3706 pwq
->flush_color
= -1;
3708 INIT_LIST_HEAD(&pwq
->delayed_works
);
3709 INIT_LIST_HEAD(&pwq
->pwqs_node
);
3710 INIT_LIST_HEAD(&pwq
->mayday_node
);
3711 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3714 /* sync @pwq with the current state of its associated wq and link it */
3715 static void link_pwq(struct pool_workqueue
*pwq
)
3717 struct workqueue_struct
*wq
= pwq
->wq
;
3719 lockdep_assert_held(&wq
->mutex
);
3721 /* may be called multiple times, ignore if already linked */
3722 if (!list_empty(&pwq
->pwqs_node
))
3726 * Set the matching work_color. This is synchronized with
3727 * wq->mutex to avoid confusing flush_workqueue().
3729 pwq
->work_color
= wq
->work_color
;
3731 /* sync max_active to the current setting */
3732 pwq_adjust_max_active(pwq
);
3735 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3738 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3739 static struct pool_workqueue
*alloc_unbound_pwq(struct workqueue_struct
*wq
,
3740 const struct workqueue_attrs
*attrs
)
3742 struct worker_pool
*pool
;
3743 struct pool_workqueue
*pwq
;
3745 lockdep_assert_held(&wq_pool_mutex
);
3747 pool
= get_unbound_pool(attrs
);
3751 pwq
= kmem_cache_alloc_node(pwq_cache
, GFP_KERNEL
, pool
->node
);
3753 put_unbound_pool(pool
);
3757 init_pwq(pwq
, wq
, pool
);
3761 /* undo alloc_unbound_pwq(), used only in the error path */
3762 static void free_unbound_pwq(struct pool_workqueue
*pwq
)
3764 lockdep_assert_held(&wq_pool_mutex
);
3767 put_unbound_pool(pwq
->pool
);
3768 kmem_cache_free(pwq_cache
, pwq
);
3773 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3774 * @attrs: the wq_attrs of interest
3775 * @node: the target NUMA node
3776 * @cpu_going_down: if >= 0, the CPU to consider as offline
3777 * @cpumask: outarg, the resulting cpumask
3779 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3780 * @cpu_going_down is >= 0, that cpu is considered offline during
3781 * calculation. The result is stored in @cpumask. This function returns
3782 * %true if the resulting @cpumask is different from @attrs->cpumask,
3785 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3786 * enabled and @node has online CPUs requested by @attrs, the returned
3787 * cpumask is the intersection of the possible CPUs of @node and
3790 * The caller is responsible for ensuring that the cpumask of @node stays
3793 static bool wq_calc_node_cpumask(const struct workqueue_attrs
*attrs
, int node
,
3794 int cpu_going_down
, cpumask_t
*cpumask
)
3796 if (!wq_numa_enabled
|| attrs
->no_numa
)
3799 /* does @node have any online CPUs @attrs wants? */
3800 cpumask_and(cpumask
, cpumask_of_node(node
), attrs
->cpumask
);
3801 if (cpu_going_down
>= 0)
3802 cpumask_clear_cpu(cpu_going_down
, cpumask
);
3804 if (cpumask_empty(cpumask
))
3807 /* yeap, return possible CPUs in @node that @attrs wants */
3808 cpumask_and(cpumask
, attrs
->cpumask
, wq_numa_possible_cpumask
[node
]);
3809 return !cpumask_equal(cpumask
, attrs
->cpumask
);
3812 cpumask_copy(cpumask
, attrs
->cpumask
);
3816 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3817 static struct pool_workqueue
*numa_pwq_tbl_install(struct workqueue_struct
*wq
,
3819 struct pool_workqueue
*pwq
)
3821 struct pool_workqueue
*old_pwq
;
3823 lockdep_assert_held(&wq
->mutex
);
3825 /* link_pwq() can handle duplicate calls */
3828 old_pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
3829 rcu_assign_pointer(wq
->numa_pwq_tbl
[node
], pwq
);
3834 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3835 * @wq: the target workqueue
3836 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3838 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3839 * machines, this function maps a separate pwq to each NUMA node with
3840 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3841 * NUMA node it was issued on. Older pwqs are released as in-flight work
3842 * items finish. Note that a work item which repeatedly requeues itself
3843 * back-to-back will stay on its current pwq.
3845 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3848 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3849 const struct workqueue_attrs
*attrs
)
3851 struct workqueue_attrs
*new_attrs
, *tmp_attrs
;
3852 struct pool_workqueue
**pwq_tbl
, *dfl_pwq
;
3855 /* only unbound workqueues can change attributes */
3856 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3859 /* creating multiple pwqs breaks ordering guarantee */
3860 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3863 pwq_tbl
= kzalloc(wq_numa_tbl_len
* sizeof(pwq_tbl
[0]), GFP_KERNEL
);
3864 new_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3865 tmp_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3866 if (!pwq_tbl
|| !new_attrs
|| !tmp_attrs
)
3869 /* make a copy of @attrs and sanitize it */
3870 copy_workqueue_attrs(new_attrs
, attrs
);
3871 cpumask_and(new_attrs
->cpumask
, new_attrs
->cpumask
, cpu_possible_mask
);
3874 * We may create multiple pwqs with differing cpumasks. Make a
3875 * copy of @new_attrs which will be modified and used to obtain
3878 copy_workqueue_attrs(tmp_attrs
, new_attrs
);
3881 * CPUs should stay stable across pwq creations and installations.
3882 * Pin CPUs, determine the target cpumask for each node and create
3887 mutex_lock(&wq_pool_mutex
);
3890 * If something goes wrong during CPU up/down, we'll fall back to
3891 * the default pwq covering whole @attrs->cpumask. Always create
3892 * it even if we don't use it immediately.
3894 dfl_pwq
= alloc_unbound_pwq(wq
, new_attrs
);
3898 for_each_node(node
) {
3899 if (wq_calc_node_cpumask(attrs
, node
, -1, tmp_attrs
->cpumask
)) {
3900 pwq_tbl
[node
] = alloc_unbound_pwq(wq
, tmp_attrs
);
3905 pwq_tbl
[node
] = dfl_pwq
;
3909 mutex_unlock(&wq_pool_mutex
);
3911 /* all pwqs have been created successfully, let's install'em */
3912 mutex_lock(&wq
->mutex
);
3914 copy_workqueue_attrs(wq
->unbound_attrs
, new_attrs
);
3916 /* save the previous pwq and install the new one */
3918 pwq_tbl
[node
] = numa_pwq_tbl_install(wq
, node
, pwq_tbl
[node
]);
3920 /* @dfl_pwq might not have been used, ensure it's linked */
3922 swap(wq
->dfl_pwq
, dfl_pwq
);
3924 mutex_unlock(&wq
->mutex
);
3926 /* put the old pwqs */
3928 put_pwq_unlocked(pwq_tbl
[node
]);
3929 put_pwq_unlocked(dfl_pwq
);
3935 free_workqueue_attrs(tmp_attrs
);
3936 free_workqueue_attrs(new_attrs
);
3941 free_unbound_pwq(dfl_pwq
);
3943 if (pwq_tbl
&& pwq_tbl
[node
] != dfl_pwq
)
3944 free_unbound_pwq(pwq_tbl
[node
]);
3945 mutex_unlock(&wq_pool_mutex
);
3953 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3954 * @wq: the target workqueue
3955 * @cpu: the CPU coming up or going down
3956 * @online: whether @cpu is coming up or going down
3958 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3959 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3962 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3963 * falls back to @wq->dfl_pwq which may not be optimal but is always
3966 * Note that when the last allowed CPU of a NUMA node goes offline for a
3967 * workqueue with a cpumask spanning multiple nodes, the workers which were
3968 * already executing the work items for the workqueue will lose their CPU
3969 * affinity and may execute on any CPU. This is similar to how per-cpu
3970 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3971 * affinity, it's the user's responsibility to flush the work item from
3974 static void wq_update_unbound_numa(struct workqueue_struct
*wq
, int cpu
,
3977 int node
= cpu_to_node(cpu
);
3978 int cpu_off
= online
? -1 : cpu
;
3979 struct pool_workqueue
*old_pwq
= NULL
, *pwq
;
3980 struct workqueue_attrs
*target_attrs
;
3983 lockdep_assert_held(&wq_pool_mutex
);
3985 if (!wq_numa_enabled
|| !(wq
->flags
& WQ_UNBOUND
))
3989 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3990 * Let's use a preallocated one. The following buf is protected by
3991 * CPU hotplug exclusion.
3993 target_attrs
= wq_update_unbound_numa_attrs_buf
;
3994 cpumask
= target_attrs
->cpumask
;
3996 mutex_lock(&wq
->mutex
);
3997 if (wq
->unbound_attrs
->no_numa
)
4000 copy_workqueue_attrs(target_attrs
, wq
->unbound_attrs
);
4001 pwq
= unbound_pwq_by_node(wq
, node
);
4004 * Let's determine what needs to be done. If the target cpumask is
4005 * different from wq's, we need to compare it to @pwq's and create
4006 * a new one if they don't match. If the target cpumask equals
4007 * wq's, the default pwq should be used. If @pwq is already the
4008 * default one, nothing to do; otherwise, install the default one.
4010 if (wq_calc_node_cpumask(wq
->unbound_attrs
, node
, cpu_off
, cpumask
)) {
4011 if (cpumask_equal(cpumask
, pwq
->pool
->attrs
->cpumask
))
4014 if (pwq
== wq
->dfl_pwq
)
4020 mutex_unlock(&wq
->mutex
);
4022 /* create a new pwq */
4023 pwq
= alloc_unbound_pwq(wq
, target_attrs
);
4025 pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4031 * Install the new pwq. As this function is called only from CPU
4032 * hotplug callbacks and applying a new attrs is wrapped with
4033 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4036 mutex_lock(&wq
->mutex
);
4037 old_pwq
= numa_pwq_tbl_install(wq
, node
, pwq
);
4041 spin_lock_irq(&wq
->dfl_pwq
->pool
->lock
);
4042 get_pwq(wq
->dfl_pwq
);
4043 spin_unlock_irq(&wq
->dfl_pwq
->pool
->lock
);
4044 old_pwq
= numa_pwq_tbl_install(wq
, node
, wq
->dfl_pwq
);
4046 mutex_unlock(&wq
->mutex
);
4047 put_pwq_unlocked(old_pwq
);
4050 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
4052 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
4055 if (!(wq
->flags
& WQ_UNBOUND
)) {
4056 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
4060 for_each_possible_cpu(cpu
) {
4061 struct pool_workqueue
*pwq
=
4062 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4063 struct worker_pool
*cpu_pools
=
4064 per_cpu(cpu_worker_pools
, cpu
);
4066 init_pwq(pwq
, wq
, &cpu_pools
[highpri
]);
4068 mutex_lock(&wq
->mutex
);
4070 mutex_unlock(&wq
->mutex
);
4074 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
4078 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
4081 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
4083 if (max_active
< 1 || max_active
> lim
)
4084 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4085 max_active
, name
, 1, lim
);
4087 return clamp_val(max_active
, 1, lim
);
4090 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
4093 struct lock_class_key
*key
,
4094 const char *lock_name
, ...)
4096 size_t tbl_size
= 0;
4098 struct workqueue_struct
*wq
;
4099 struct pool_workqueue
*pwq
;
4101 /* allocate wq and format name */
4102 if (flags
& WQ_UNBOUND
)
4103 tbl_size
= wq_numa_tbl_len
* sizeof(wq
->numa_pwq_tbl
[0]);
4105 wq
= kzalloc(sizeof(*wq
) + tbl_size
, GFP_KERNEL
);
4109 if (flags
& WQ_UNBOUND
) {
4110 wq
->unbound_attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
4111 if (!wq
->unbound_attrs
)
4115 va_start(args
, lock_name
);
4116 vsnprintf(wq
->name
, sizeof(wq
->name
), fmt
, args
);
4119 max_active
= max_active
?: WQ_DFL_ACTIVE
;
4120 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
4124 wq
->saved_max_active
= max_active
;
4125 mutex_init(&wq
->mutex
);
4126 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
4127 INIT_LIST_HEAD(&wq
->pwqs
);
4128 INIT_LIST_HEAD(&wq
->flusher_queue
);
4129 INIT_LIST_HEAD(&wq
->flusher_overflow
);
4130 INIT_LIST_HEAD(&wq
->maydays
);
4132 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
4133 INIT_LIST_HEAD(&wq
->list
);
4135 if (alloc_and_link_pwqs(wq
) < 0)
4139 * Workqueues which may be used during memory reclaim should
4140 * have a rescuer to guarantee forward progress.
4142 if (flags
& WQ_MEM_RECLAIM
) {
4143 struct worker
*rescuer
;
4145 rescuer
= alloc_worker();
4149 rescuer
->rescue_wq
= wq
;
4150 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
4152 if (IS_ERR(rescuer
->task
)) {
4157 wq
->rescuer
= rescuer
;
4158 rescuer
->task
->flags
|= PF_NO_SETAFFINITY
;
4159 wake_up_process(rescuer
->task
);
4162 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
4166 * wq_pool_mutex protects global freeze state and workqueues list.
4167 * Grab it, adjust max_active and add the new @wq to workqueues
4170 mutex_lock(&wq_pool_mutex
);
4172 mutex_lock(&wq
->mutex
);
4173 for_each_pwq(pwq
, wq
)
4174 pwq_adjust_max_active(pwq
);
4175 mutex_unlock(&wq
->mutex
);
4177 list_add(&wq
->list
, &workqueues
);
4179 mutex_unlock(&wq_pool_mutex
);
4184 free_workqueue_attrs(wq
->unbound_attrs
);
4188 destroy_workqueue(wq
);
4191 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
4194 * destroy_workqueue - safely terminate a workqueue
4195 * @wq: target workqueue
4197 * Safely destroy a workqueue. All work currently pending will be done first.
4199 void destroy_workqueue(struct workqueue_struct
*wq
)
4201 struct pool_workqueue
*pwq
;
4204 /* drain it before proceeding with destruction */
4205 drain_workqueue(wq
);
4208 mutex_lock(&wq
->mutex
);
4209 for_each_pwq(pwq
, wq
) {
4212 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
4213 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
4214 mutex_unlock(&wq
->mutex
);
4219 if (WARN_ON((pwq
!= wq
->dfl_pwq
) && (pwq
->refcnt
> 1)) ||
4220 WARN_ON(pwq
->nr_active
) ||
4221 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4222 mutex_unlock(&wq
->mutex
);
4226 mutex_unlock(&wq
->mutex
);
4229 * wq list is used to freeze wq, remove from list after
4230 * flushing is complete in case freeze races us.
4232 mutex_lock(&wq_pool_mutex
);
4233 list_del_init(&wq
->list
);
4234 mutex_unlock(&wq_pool_mutex
);
4236 workqueue_sysfs_unregister(wq
);
4239 kthread_stop(wq
->rescuer
->task
);
4244 if (!(wq
->flags
& WQ_UNBOUND
)) {
4246 * The base ref is never dropped on per-cpu pwqs. Directly
4247 * free the pwqs and wq.
4249 free_percpu(wq
->cpu_pwqs
);
4253 * We're the sole accessor of @wq at this point. Directly
4254 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4255 * @wq will be freed when the last pwq is released.
4257 for_each_node(node
) {
4258 pwq
= rcu_access_pointer(wq
->numa_pwq_tbl
[node
]);
4259 RCU_INIT_POINTER(wq
->numa_pwq_tbl
[node
], NULL
);
4260 put_pwq_unlocked(pwq
);
4264 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4265 * put. Don't access it afterwards.
4269 put_pwq_unlocked(pwq
);
4272 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4275 * workqueue_set_max_active - adjust max_active of a workqueue
4276 * @wq: target workqueue
4277 * @max_active: new max_active value.
4279 * Set max_active of @wq to @max_active.
4282 * Don't call from IRQ context.
4284 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4286 struct pool_workqueue
*pwq
;
4288 /* disallow meddling with max_active for ordered workqueues */
4289 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4292 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4294 mutex_lock(&wq
->mutex
);
4296 wq
->saved_max_active
= max_active
;
4298 for_each_pwq(pwq
, wq
)
4299 pwq_adjust_max_active(pwq
);
4301 mutex_unlock(&wq
->mutex
);
4303 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4306 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4308 * Determine whether %current is a workqueue rescuer. Can be used from
4309 * work functions to determine whether it's being run off the rescuer task.
4311 bool current_is_workqueue_rescuer(void)
4313 struct worker
*worker
= current_wq_worker();
4315 return worker
&& worker
->rescue_wq
;
4319 * workqueue_congested - test whether a workqueue is congested
4320 * @cpu: CPU in question
4321 * @wq: target workqueue
4323 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4324 * no synchronization around this function and the test result is
4325 * unreliable and only useful as advisory hints or for debugging.
4327 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4328 * Note that both per-cpu and unbound workqueues may be associated with
4329 * multiple pool_workqueues which have separate congested states. A
4330 * workqueue being congested on one CPU doesn't mean the workqueue is also
4331 * contested on other CPUs / NUMA nodes.
4334 * %true if congested, %false otherwise.
4336 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4338 struct pool_workqueue
*pwq
;
4341 rcu_read_lock_sched();
4343 if (cpu
== WORK_CPU_UNBOUND
)
4344 cpu
= smp_processor_id();
4346 if (!(wq
->flags
& WQ_UNBOUND
))
4347 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4349 pwq
= unbound_pwq_by_node(wq
, cpu_to_node(cpu
));
4351 ret
= !list_empty(&pwq
->delayed_works
);
4352 rcu_read_unlock_sched();
4356 EXPORT_SYMBOL_GPL(workqueue_congested
);
4359 * work_busy - test whether a work is currently pending or running
4360 * @work: the work to be tested
4362 * Test whether @work is currently pending or running. There is no
4363 * synchronization around this function and the test result is
4364 * unreliable and only useful as advisory hints or for debugging.
4367 * OR'd bitmask of WORK_BUSY_* bits.
4369 unsigned int work_busy(struct work_struct
*work
)
4371 struct worker_pool
*pool
;
4372 unsigned long flags
;
4373 unsigned int ret
= 0;
4375 if (work_pending(work
))
4376 ret
|= WORK_BUSY_PENDING
;
4378 local_irq_save(flags
);
4379 pool
= get_work_pool(work
);
4381 spin_lock(&pool
->lock
);
4382 if (find_worker_executing_work(pool
, work
))
4383 ret
|= WORK_BUSY_RUNNING
;
4384 spin_unlock(&pool
->lock
);
4386 local_irq_restore(flags
);
4390 EXPORT_SYMBOL_GPL(work_busy
);
4393 * set_worker_desc - set description for the current work item
4394 * @fmt: printf-style format string
4395 * @...: arguments for the format string
4397 * This function can be called by a running work function to describe what
4398 * the work item is about. If the worker task gets dumped, this
4399 * information will be printed out together to help debugging. The
4400 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4402 void set_worker_desc(const char *fmt
, ...)
4404 struct worker
*worker
= current_wq_worker();
4408 va_start(args
, fmt
);
4409 vsnprintf(worker
->desc
, sizeof(worker
->desc
), fmt
, args
);
4411 worker
->desc_valid
= true;
4416 * print_worker_info - print out worker information and description
4417 * @log_lvl: the log level to use when printing
4418 * @task: target task
4420 * If @task is a worker and currently executing a work item, print out the
4421 * name of the workqueue being serviced and worker description set with
4422 * set_worker_desc() by the currently executing work item.
4424 * This function can be safely called on any task as long as the
4425 * task_struct itself is accessible. While safe, this function isn't
4426 * synchronized and may print out mixups or garbages of limited length.
4428 void print_worker_info(const char *log_lvl
, struct task_struct
*task
)
4430 work_func_t
*fn
= NULL
;
4431 char name
[WQ_NAME_LEN
] = { };
4432 char desc
[WORKER_DESC_LEN
] = { };
4433 struct pool_workqueue
*pwq
= NULL
;
4434 struct workqueue_struct
*wq
= NULL
;
4435 bool desc_valid
= false;
4436 struct worker
*worker
;
4438 if (!(task
->flags
& PF_WQ_WORKER
))
4442 * This function is called without any synchronization and @task
4443 * could be in any state. Be careful with dereferences.
4445 worker
= probe_kthread_data(task
);
4448 * Carefully copy the associated workqueue's workfn and name. Keep
4449 * the original last '\0' in case the original contains garbage.
4451 probe_kernel_read(&fn
, &worker
->current_func
, sizeof(fn
));
4452 probe_kernel_read(&pwq
, &worker
->current_pwq
, sizeof(pwq
));
4453 probe_kernel_read(&wq
, &pwq
->wq
, sizeof(wq
));
4454 probe_kernel_read(name
, wq
->name
, sizeof(name
) - 1);
4456 /* copy worker description */
4457 probe_kernel_read(&desc_valid
, &worker
->desc_valid
, sizeof(desc_valid
));
4459 probe_kernel_read(desc
, worker
->desc
, sizeof(desc
) - 1);
4461 if (fn
|| name
[0] || desc
[0]) {
4462 printk("%sWorkqueue: %s %pf", log_lvl
, name
, fn
);
4464 pr_cont(" (%s)", desc
);
4472 * There are two challenges in supporting CPU hotplug. Firstly, there
4473 * are a lot of assumptions on strong associations among work, pwq and
4474 * pool which make migrating pending and scheduled works very
4475 * difficult to implement without impacting hot paths. Secondly,
4476 * worker pools serve mix of short, long and very long running works making
4477 * blocked draining impractical.
4479 * This is solved by allowing the pools to be disassociated from the CPU
4480 * running as an unbound one and allowing it to be reattached later if the
4481 * cpu comes back online.
4484 static void wq_unbind_fn(struct work_struct
*work
)
4486 int cpu
= smp_processor_id();
4487 struct worker_pool
*pool
;
4488 struct worker
*worker
;
4491 for_each_cpu_worker_pool(pool
, cpu
) {
4492 WARN_ON_ONCE(cpu
!= smp_processor_id());
4494 mutex_lock(&pool
->manager_mutex
);
4495 spin_lock_irq(&pool
->lock
);
4498 * We've blocked all manager operations. Make all workers
4499 * unbound and set DISASSOCIATED. Before this, all workers
4500 * except for the ones which are still executing works from
4501 * before the last CPU down must be on the cpu. After
4502 * this, they may become diasporas.
4504 for_each_pool_worker(worker
, wi
, pool
)
4505 worker
->flags
|= WORKER_UNBOUND
;
4507 pool
->flags
|= POOL_DISASSOCIATED
;
4509 spin_unlock_irq(&pool
->lock
);
4510 mutex_unlock(&pool
->manager_mutex
);
4513 * Call schedule() so that we cross rq->lock and thus can
4514 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4515 * This is necessary as scheduler callbacks may be invoked
4521 * Sched callbacks are disabled now. Zap nr_running.
4522 * After this, nr_running stays zero and need_more_worker()
4523 * and keep_working() are always true as long as the
4524 * worklist is not empty. This pool now behaves as an
4525 * unbound (in terms of concurrency management) pool which
4526 * are served by workers tied to the pool.
4528 atomic_set(&pool
->nr_running
, 0);
4531 * With concurrency management just turned off, a busy
4532 * worker blocking could lead to lengthy stalls. Kick off
4533 * unbound chain execution of currently pending work items.
4535 spin_lock_irq(&pool
->lock
);
4536 wake_up_worker(pool
);
4537 spin_unlock_irq(&pool
->lock
);
4542 * rebind_workers - rebind all workers of a pool to the associated CPU
4543 * @pool: pool of interest
4545 * @pool->cpu is coming online. Rebind all workers to the CPU.
4547 static void rebind_workers(struct worker_pool
*pool
)
4549 struct worker
*worker
;
4552 lockdep_assert_held(&pool
->manager_mutex
);
4555 * Restore CPU affinity of all workers. As all idle workers should
4556 * be on the run-queue of the associated CPU before any local
4557 * wake-ups for concurrency management happen, restore CPU affinty
4558 * of all workers first and then clear UNBOUND. As we're called
4559 * from CPU_ONLINE, the following shouldn't fail.
4561 for_each_pool_worker(worker
, wi
, pool
)
4562 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4563 pool
->attrs
->cpumask
) < 0);
4565 spin_lock_irq(&pool
->lock
);
4567 for_each_pool_worker(worker
, wi
, pool
) {
4568 unsigned int worker_flags
= worker
->flags
;
4571 * A bound idle worker should actually be on the runqueue
4572 * of the associated CPU for local wake-ups targeting it to
4573 * work. Kick all idle workers so that they migrate to the
4574 * associated CPU. Doing this in the same loop as
4575 * replacing UNBOUND with REBOUND is safe as no worker will
4576 * be bound before @pool->lock is released.
4578 if (worker_flags
& WORKER_IDLE
)
4579 wake_up_process(worker
->task
);
4582 * We want to clear UNBOUND but can't directly call
4583 * worker_clr_flags() or adjust nr_running. Atomically
4584 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4585 * @worker will clear REBOUND using worker_clr_flags() when
4586 * it initiates the next execution cycle thus restoring
4587 * concurrency management. Note that when or whether
4588 * @worker clears REBOUND doesn't affect correctness.
4590 * ACCESS_ONCE() is necessary because @worker->flags may be
4591 * tested without holding any lock in
4592 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4593 * fail incorrectly leading to premature concurrency
4594 * management operations.
4596 WARN_ON_ONCE(!(worker_flags
& WORKER_UNBOUND
));
4597 worker_flags
|= WORKER_REBOUND
;
4598 worker_flags
&= ~WORKER_UNBOUND
;
4599 ACCESS_ONCE(worker
->flags
) = worker_flags
;
4602 spin_unlock_irq(&pool
->lock
);
4606 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4607 * @pool: unbound pool of interest
4608 * @cpu: the CPU which is coming up
4610 * An unbound pool may end up with a cpumask which doesn't have any online
4611 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4612 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4613 * online CPU before, cpus_allowed of all its workers should be restored.
4615 static void restore_unbound_workers_cpumask(struct worker_pool
*pool
, int cpu
)
4617 static cpumask_t cpumask
;
4618 struct worker
*worker
;
4621 lockdep_assert_held(&pool
->manager_mutex
);
4623 /* is @cpu allowed for @pool? */
4624 if (!cpumask_test_cpu(cpu
, pool
->attrs
->cpumask
))
4627 /* is @cpu the only online CPU? */
4628 cpumask_and(&cpumask
, pool
->attrs
->cpumask
, cpu_online_mask
);
4629 if (cpumask_weight(&cpumask
) != 1)
4632 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4633 for_each_pool_worker(worker
, wi
, pool
)
4634 WARN_ON_ONCE(set_cpus_allowed_ptr(worker
->task
,
4635 pool
->attrs
->cpumask
) < 0);
4639 * Workqueues should be brought up before normal priority CPU notifiers.
4640 * This will be registered high priority CPU notifier.
4642 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4643 unsigned long action
,
4646 int cpu
= (unsigned long)hcpu
;
4647 struct worker_pool
*pool
;
4648 struct workqueue_struct
*wq
;
4651 switch (action
& ~CPU_TASKS_FROZEN
) {
4652 case CPU_UP_PREPARE
:
4653 for_each_cpu_worker_pool(pool
, cpu
) {
4654 if (pool
->nr_workers
)
4656 if (create_and_start_worker(pool
) < 0)
4661 case CPU_DOWN_FAILED
:
4663 mutex_lock(&wq_pool_mutex
);
4665 for_each_pool(pool
, pi
) {
4666 mutex_lock(&pool
->manager_mutex
);
4668 if (pool
->cpu
== cpu
) {
4669 spin_lock_irq(&pool
->lock
);
4670 pool
->flags
&= ~POOL_DISASSOCIATED
;
4671 spin_unlock_irq(&pool
->lock
);
4673 rebind_workers(pool
);
4674 } else if (pool
->cpu
< 0) {
4675 restore_unbound_workers_cpumask(pool
, cpu
);
4678 mutex_unlock(&pool
->manager_mutex
);
4681 /* update NUMA affinity of unbound workqueues */
4682 list_for_each_entry(wq
, &workqueues
, list
)
4683 wq_update_unbound_numa(wq
, cpu
, true);
4685 mutex_unlock(&wq_pool_mutex
);
4692 * Workqueues should be brought down after normal priority CPU notifiers.
4693 * This will be registered as low priority CPU notifier.
4695 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4696 unsigned long action
,
4699 int cpu
= (unsigned long)hcpu
;
4700 struct work_struct unbind_work
;
4701 struct workqueue_struct
*wq
;
4703 switch (action
& ~CPU_TASKS_FROZEN
) {
4704 case CPU_DOWN_PREPARE
:
4705 /* unbinding per-cpu workers should happen on the local CPU */
4706 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4707 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4709 /* update NUMA affinity of unbound workqueues */
4710 mutex_lock(&wq_pool_mutex
);
4711 list_for_each_entry(wq
, &workqueues
, list
)
4712 wq_update_unbound_numa(wq
, cpu
, false);
4713 mutex_unlock(&wq_pool_mutex
);
4715 /* wait for per-cpu unbinding to finish */
4716 flush_work(&unbind_work
);
4724 struct work_for_cpu
{
4725 struct work_struct work
;
4731 static void work_for_cpu_fn(struct work_struct
*work
)
4733 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4735 wfc
->ret
= wfc
->fn(wfc
->arg
);
4739 * work_on_cpu - run a function in user context on a particular cpu
4740 * @cpu: the cpu to run on
4741 * @fn: the function to run
4742 * @arg: the function arg
4744 * This will return the value @fn returns.
4745 * It is up to the caller to ensure that the cpu doesn't go offline.
4746 * The caller must not hold any locks which would prevent @fn from completing.
4748 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4750 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4752 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4753 schedule_work_on(cpu
, &wfc
.work
);
4754 flush_work(&wfc
.work
);
4757 EXPORT_SYMBOL_GPL(work_on_cpu
);
4758 #endif /* CONFIG_SMP */
4760 #ifdef CONFIG_FREEZER
4763 * freeze_workqueues_begin - begin freezing workqueues
4765 * Start freezing workqueues. After this function returns, all freezable
4766 * workqueues will queue new works to their delayed_works list instead of
4770 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4772 void freeze_workqueues_begin(void)
4774 struct worker_pool
*pool
;
4775 struct workqueue_struct
*wq
;
4776 struct pool_workqueue
*pwq
;
4779 mutex_lock(&wq_pool_mutex
);
4781 WARN_ON_ONCE(workqueue_freezing
);
4782 workqueue_freezing
= true;
4785 for_each_pool(pool
, pi
) {
4786 spin_lock_irq(&pool
->lock
);
4787 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4788 pool
->flags
|= POOL_FREEZING
;
4789 spin_unlock_irq(&pool
->lock
);
4792 list_for_each_entry(wq
, &workqueues
, list
) {
4793 mutex_lock(&wq
->mutex
);
4794 for_each_pwq(pwq
, wq
)
4795 pwq_adjust_max_active(pwq
);
4796 mutex_unlock(&wq
->mutex
);
4799 mutex_unlock(&wq_pool_mutex
);
4803 * freeze_workqueues_busy - are freezable workqueues still busy?
4805 * Check whether freezing is complete. This function must be called
4806 * between freeze_workqueues_begin() and thaw_workqueues().
4809 * Grabs and releases wq_pool_mutex.
4812 * %true if some freezable workqueues are still busy. %false if freezing
4815 bool freeze_workqueues_busy(void)
4818 struct workqueue_struct
*wq
;
4819 struct pool_workqueue
*pwq
;
4821 mutex_lock(&wq_pool_mutex
);
4823 WARN_ON_ONCE(!workqueue_freezing
);
4825 list_for_each_entry(wq
, &workqueues
, list
) {
4826 if (!(wq
->flags
& WQ_FREEZABLE
))
4829 * nr_active is monotonically decreasing. It's safe
4830 * to peek without lock.
4832 rcu_read_lock_sched();
4833 for_each_pwq(pwq
, wq
) {
4834 WARN_ON_ONCE(pwq
->nr_active
< 0);
4835 if (pwq
->nr_active
) {
4837 rcu_read_unlock_sched();
4841 rcu_read_unlock_sched();
4844 mutex_unlock(&wq_pool_mutex
);
4849 * thaw_workqueues - thaw workqueues
4851 * Thaw workqueues. Normal queueing is restored and all collected
4852 * frozen works are transferred to their respective pool worklists.
4855 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4857 void thaw_workqueues(void)
4859 struct workqueue_struct
*wq
;
4860 struct pool_workqueue
*pwq
;
4861 struct worker_pool
*pool
;
4864 mutex_lock(&wq_pool_mutex
);
4866 if (!workqueue_freezing
)
4869 /* clear FREEZING */
4870 for_each_pool(pool
, pi
) {
4871 spin_lock_irq(&pool
->lock
);
4872 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4873 pool
->flags
&= ~POOL_FREEZING
;
4874 spin_unlock_irq(&pool
->lock
);
4877 /* restore max_active and repopulate worklist */
4878 list_for_each_entry(wq
, &workqueues
, list
) {
4879 mutex_lock(&wq
->mutex
);
4880 for_each_pwq(pwq
, wq
)
4881 pwq_adjust_max_active(pwq
);
4882 mutex_unlock(&wq
->mutex
);
4885 workqueue_freezing
= false;
4887 mutex_unlock(&wq_pool_mutex
);
4889 #endif /* CONFIG_FREEZER */
4891 static void __init
wq_numa_init(void)
4896 /* determine NUMA pwq table len - highest node id + 1 */
4898 wq_numa_tbl_len
= max(wq_numa_tbl_len
, node
+ 1);
4900 if (num_possible_nodes() <= 1)
4903 if (wq_disable_numa
) {
4904 pr_info("workqueue: NUMA affinity support disabled\n");
4908 wq_update_unbound_numa_attrs_buf
= alloc_workqueue_attrs(GFP_KERNEL
);
4909 BUG_ON(!wq_update_unbound_numa_attrs_buf
);
4912 * We want masks of possible CPUs of each node which isn't readily
4913 * available. Build one from cpu_to_node() which should have been
4914 * fully initialized by now.
4916 tbl
= kzalloc(wq_numa_tbl_len
* sizeof(tbl
[0]), GFP_KERNEL
);
4920 BUG_ON(!alloc_cpumask_var_node(&tbl
[node
], GFP_KERNEL
,
4921 node_online(node
) ? node
: NUMA_NO_NODE
));
4923 for_each_possible_cpu(cpu
) {
4924 node
= cpu_to_node(cpu
);
4925 if (WARN_ON(node
== NUMA_NO_NODE
)) {
4926 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu
);
4927 /* happens iff arch is bonkers, let's just proceed */
4930 cpumask_set_cpu(cpu
, tbl
[node
]);
4933 wq_numa_possible_cpumask
= tbl
;
4934 wq_numa_enabled
= true;
4937 static int __init
init_workqueues(void)
4939 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4942 /* make sure we have enough bits for OFFQ pool ID */
4943 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
4944 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
4946 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4948 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4950 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4951 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4955 /* initialize CPU pools */
4956 for_each_possible_cpu(cpu
) {
4957 struct worker_pool
*pool
;
4960 for_each_cpu_worker_pool(pool
, cpu
) {
4961 BUG_ON(init_worker_pool(pool
));
4963 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4964 pool
->attrs
->nice
= std_nice
[i
++];
4965 pool
->node
= cpu_to_node(cpu
);
4968 mutex_lock(&wq_pool_mutex
);
4969 BUG_ON(worker_pool_assign_id(pool
));
4970 mutex_unlock(&wq_pool_mutex
);
4974 /* create the initial worker */
4975 for_each_online_cpu(cpu
) {
4976 struct worker_pool
*pool
;
4978 for_each_cpu_worker_pool(pool
, cpu
) {
4979 pool
->flags
&= ~POOL_DISASSOCIATED
;
4980 BUG_ON(create_and_start_worker(pool
) < 0);
4984 /* create default unbound wq attrs */
4985 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4986 struct workqueue_attrs
*attrs
;
4988 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4989 attrs
->nice
= std_nice
[i
];
4990 unbound_std_wq_attrs
[i
] = attrs
;
4993 system_wq
= alloc_workqueue("events", 0, 0);
4994 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4995 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4996 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4997 WQ_UNBOUND_MAX_ACTIVE
);
4998 system_freezable_wq
= alloc_workqueue("events_freezable",
5000 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
5001 !system_unbound_wq
|| !system_freezable_wq
);
5004 early_initcall(init_workqueues
);