Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/workqueue.c - generic async execution with shared worker pool
4 *
5 * Copyright (C) 2002 Ingo Molnar
6 *
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
12 *
13 * Made to use alloc_percpu by Christoph Lameter.
14 *
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 *
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
24 *
25 * Please read Documentation/core-api/workqueue.rst for details.
26 */
27
28#include <linux/export.h>
29#include <linux/kernel.h>
30#include <linux/sched.h>
31#include <linux/init.h>
32#include <linux/signal.h>
33#include <linux/completion.h>
34#include <linux/workqueue.h>
35#include <linux/slab.h>
36#include <linux/cpu.h>
37#include <linux/notifier.h>
38#include <linux/kthread.h>
39#include <linux/hardirq.h>
40#include <linux/mempolicy.h>
41#include <linux/freezer.h>
42#include <linux/debug_locks.h>
43#include <linux/lockdep.h>
44#include <linux/idr.h>
45#include <linux/jhash.h>
46#include <linux/hashtable.h>
47#include <linux/rculist.h>
48#include <linux/nodemask.h>
49#include <linux/moduleparam.h>
50#include <linux/uaccess.h>
51#include <linux/sched/isolation.h>
52#include <linux/nmi.h>
53
54#include "workqueue_internal.h"
55
56enum {
57 /*
58 * worker_pool flags
59 *
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * is in effect.
64 *
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
68 *
69 * Note that DISASSOCIATED should be flipped only while holding
70 * wq_pool_attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
72 */
73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
75
76 /* worker flags */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
83
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
86
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
97 (min two ticks) */
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
100
101 /*
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
104 */
105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
107
108 WQ_NAME_LEN = 24,
109};
110
111/*
112 * Structure fields follow one of the following exclusion rules.
113 *
114 * I: Modifiable by initialization/destruction paths and read-only for
115 * everyone else.
116 *
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
119 *
120 * L: pool->lock protected. Access with pool->lock held.
121 *
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
126 *
127 * A: wq_pool_attach_mutex protected.
128 *
129 * PL: wq_pool_mutex protected.
130 *
131 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
132 *
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
134 *
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
136 * RCU for reads.
137 *
138 * WQ: wq->mutex protected.
139 *
140 * WR: wq->mutex protected for writes. RCU protected for reads.
141 *
142 * MD: wq_mayday_lock protected.
143 */
144
145/* struct worker is defined in workqueue_internal.h */
146
147struct worker_pool {
148 raw_spinlock_t lock; /* the pool lock */
149 int cpu; /* I: the associated cpu */
150 int node; /* I: the associated node ID */
151 int id; /* I: pool ID */
152 unsigned int flags; /* X: flags */
153
154 unsigned long watchdog_ts; /* L: watchdog timestamp */
155
156 struct list_head worklist; /* L: list of pending works */
157
158 int nr_workers; /* L: total number of workers */
159 int nr_idle; /* L: currently idle workers */
160
161 struct list_head idle_list; /* X: list of idle workers */
162 struct timer_list idle_timer; /* L: worker idle timeout */
163 struct timer_list mayday_timer; /* L: SOS timer for workers */
164
165 /* a workers is either on busy_hash or idle_list, or the manager */
166 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
167 /* L: hash of busy workers */
168
169 struct worker *manager; /* L: purely informational */
170 struct list_head workers; /* A: attached workers */
171 struct completion *detach_completion; /* all workers detached */
172
173 struct ida worker_ida; /* worker IDs for task name */
174
175 struct workqueue_attrs *attrs; /* I: worker attributes */
176 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
177 int refcnt; /* PL: refcnt for unbound pools */
178
179 /*
180 * The current concurrency level. As it's likely to be accessed
181 * from other CPUs during try_to_wake_up(), put it in a separate
182 * cacheline.
183 */
184 atomic_t nr_running ____cacheline_aligned_in_smp;
185
186 /*
187 * Destruction of pool is RCU protected to allow dereferences
188 * from get_work_pool().
189 */
190 struct rcu_head rcu;
191} ____cacheline_aligned_in_smp;
192
193/*
194 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
195 * of work_struct->data are used for flags and the remaining high bits
196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
197 * number of flag bits.
198 */
199struct pool_workqueue {
200 struct worker_pool *pool; /* I: the associated pool */
201 struct workqueue_struct *wq; /* I: the owning workqueue */
202 int work_color; /* L: current color */
203 int flush_color; /* L: flushing color */
204 int refcnt; /* L: reference count */
205 int nr_in_flight[WORK_NR_COLORS];
206 /* L: nr of in_flight works */
207 int nr_active; /* L: nr of active works */
208 int max_active; /* L: max active works */
209 struct list_head delayed_works; /* L: delayed works */
210 struct list_head pwqs_node; /* WR: node on wq->pwqs */
211 struct list_head mayday_node; /* MD: node on wq->maydays */
212
213 /*
214 * Release of unbound pwq is punted to system_wq. See put_pwq()
215 * and pwq_unbound_release_workfn() for details. pool_workqueue
216 * itself is also RCU protected so that the first pwq can be
217 * determined without grabbing wq->mutex.
218 */
219 struct work_struct unbound_release_work;
220 struct rcu_head rcu;
221} __aligned(1 << WORK_STRUCT_FLAG_BITS);
222
223/*
224 * Structure used to wait for workqueue flush.
225 */
226struct wq_flusher {
227 struct list_head list; /* WQ: list of flushers */
228 int flush_color; /* WQ: flush color waiting for */
229 struct completion done; /* flush completion */
230};
231
232struct wq_device;
233
234/*
235 * The externally visible workqueue. It relays the issued work items to
236 * the appropriate worker_pool through its pool_workqueues.
237 */
238struct workqueue_struct {
239 struct list_head pwqs; /* WR: all pwqs of this wq */
240 struct list_head list; /* PR: list of all workqueues */
241
242 struct mutex mutex; /* protects this wq */
243 int work_color; /* WQ: current work color */
244 int flush_color; /* WQ: current flush color */
245 atomic_t nr_pwqs_to_flush; /* flush in progress */
246 struct wq_flusher *first_flusher; /* WQ: first flusher */
247 struct list_head flusher_queue; /* WQ: flush waiters */
248 struct list_head flusher_overflow; /* WQ: flush overflow list */
249
250 struct list_head maydays; /* MD: pwqs requesting rescue */
251 struct worker *rescuer; /* MD: rescue worker */
252
253 int nr_drainers; /* WQ: drain in progress */
254 int saved_max_active; /* WQ: saved pwq max_active */
255
256 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
257 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
258
259#ifdef CONFIG_SYSFS
260 struct wq_device *wq_dev; /* I: for sysfs interface */
261#endif
262#ifdef CONFIG_LOCKDEP
263 char *lock_name;
264 struct lock_class_key key;
265 struct lockdep_map lockdep_map;
266#endif
267 char name[WQ_NAME_LEN]; /* I: workqueue name */
268
269 /*
270 * Destruction of workqueue_struct is RCU protected to allow walking
271 * the workqueues list without grabbing wq_pool_mutex.
272 * This is used to dump all workqueues from sysrq.
273 */
274 struct rcu_head rcu;
275
276 /* hot fields used during command issue, aligned to cacheline */
277 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
278 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
279 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
280};
281
282static struct kmem_cache *pwq_cache;
283
284static cpumask_var_t *wq_numa_possible_cpumask;
285 /* possible CPUs of each node */
286
287static bool wq_disable_numa;
288module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289
290/* see the comment above the definition of WQ_POWER_EFFICIENT */
291static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
292module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293
294static bool wq_online; /* can kworkers be created yet? */
295
296static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297
298/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300
301static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
302static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
303static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
304/* wait for manager to go away */
305static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
306
307static LIST_HEAD(workqueues); /* PR: list of all workqueues */
308static bool workqueue_freezing; /* PL: have wqs started freezing? */
309
310/* PL: allowable cpus for unbound wqs and work items */
311static cpumask_var_t wq_unbound_cpumask;
312
313/* CPU where unbound work was last round robin scheduled from this CPU */
314static DEFINE_PER_CPU(int, wq_rr_cpu_last);
315
316/*
317 * Local execution of unbound work items is no longer guaranteed. The
318 * following always forces round-robin CPU selection on unbound work items
319 * to uncover usages which depend on it.
320 */
321#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
322static bool wq_debug_force_rr_cpu = true;
323#else
324static bool wq_debug_force_rr_cpu = false;
325#endif
326module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
327
328/* the per-cpu worker pools */
329static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
330
331static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
332
333/* PL: hash of all unbound pools keyed by pool->attrs */
334static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
335
336/* I: attributes used when instantiating standard unbound pools on demand */
337static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
338
339/* I: attributes used when instantiating ordered pools on demand */
340static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
341
342struct workqueue_struct *system_wq __read_mostly;
343EXPORT_SYMBOL(system_wq);
344struct workqueue_struct *system_highpri_wq __read_mostly;
345EXPORT_SYMBOL_GPL(system_highpri_wq);
346struct workqueue_struct *system_long_wq __read_mostly;
347EXPORT_SYMBOL_GPL(system_long_wq);
348struct workqueue_struct *system_unbound_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_unbound_wq);
350struct workqueue_struct *system_freezable_wq __read_mostly;
351EXPORT_SYMBOL_GPL(system_freezable_wq);
352struct workqueue_struct *system_power_efficient_wq __read_mostly;
353EXPORT_SYMBOL_GPL(system_power_efficient_wq);
354struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
355EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
356
357static int worker_thread(void *__worker);
358static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
359static void show_pwq(struct pool_workqueue *pwq);
360
361#define CREATE_TRACE_POINTS
362#include <trace/events/workqueue.h>
363
364#define assert_rcu_or_pool_mutex() \
365 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
366 !lockdep_is_held(&wq_pool_mutex), \
367 "RCU or wq_pool_mutex should be held")
368
369#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
370 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
371 !lockdep_is_held(&wq->mutex) && \
372 !lockdep_is_held(&wq_pool_mutex), \
373 "RCU, wq->mutex or wq_pool_mutex should be held")
374
375#define for_each_cpu_worker_pool(pool, cpu) \
376 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
377 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
378 (pool)++)
379
380/**
381 * for_each_pool - iterate through all worker_pools in the system
382 * @pool: iteration cursor
383 * @pi: integer used for iteration
384 *
385 * This must be called either with wq_pool_mutex held or RCU read
386 * locked. If the pool needs to be used beyond the locking in effect, the
387 * caller is responsible for guaranteeing that the pool stays online.
388 *
389 * The if/else clause exists only for the lockdep assertion and can be
390 * ignored.
391 */
392#define for_each_pool(pool, pi) \
393 idr_for_each_entry(&worker_pool_idr, pool, pi) \
394 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
395 else
396
397/**
398 * for_each_pool_worker - iterate through all workers of a worker_pool
399 * @worker: iteration cursor
400 * @pool: worker_pool to iterate workers of
401 *
402 * This must be called with wq_pool_attach_mutex.
403 *
404 * The if/else clause exists only for the lockdep assertion and can be
405 * ignored.
406 */
407#define for_each_pool_worker(worker, pool) \
408 list_for_each_entry((worker), &(pool)->workers, node) \
409 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
410 else
411
412/**
413 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
414 * @pwq: iteration cursor
415 * @wq: the target workqueue
416 *
417 * This must be called either with wq->mutex held or RCU read locked.
418 * If the pwq needs to be used beyond the locking in effect, the caller is
419 * responsible for guaranteeing that the pwq stays online.
420 *
421 * The if/else clause exists only for the lockdep assertion and can be
422 * ignored.
423 */
424#define for_each_pwq(pwq, wq) \
425 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
426 lockdep_is_held(&(wq->mutex)))
427
428#ifdef CONFIG_DEBUG_OBJECTS_WORK
429
430static struct debug_obj_descr work_debug_descr;
431
432static void *work_debug_hint(void *addr)
433{
434 return ((struct work_struct *) addr)->func;
435}
436
437static bool work_is_static_object(void *addr)
438{
439 struct work_struct *work = addr;
440
441 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
442}
443
444/*
445 * fixup_init is called when:
446 * - an active object is initialized
447 */
448static bool work_fixup_init(void *addr, enum debug_obj_state state)
449{
450 struct work_struct *work = addr;
451
452 switch (state) {
453 case ODEBUG_STATE_ACTIVE:
454 cancel_work_sync(work);
455 debug_object_init(work, &work_debug_descr);
456 return true;
457 default:
458 return false;
459 }
460}
461
462/*
463 * fixup_free is called when:
464 * - an active object is freed
465 */
466static bool work_fixup_free(void *addr, enum debug_obj_state state)
467{
468 struct work_struct *work = addr;
469
470 switch (state) {
471 case ODEBUG_STATE_ACTIVE:
472 cancel_work_sync(work);
473 debug_object_free(work, &work_debug_descr);
474 return true;
475 default:
476 return false;
477 }
478}
479
480static struct debug_obj_descr work_debug_descr = {
481 .name = "work_struct",
482 .debug_hint = work_debug_hint,
483 .is_static_object = work_is_static_object,
484 .fixup_init = work_fixup_init,
485 .fixup_free = work_fixup_free,
486};
487
488static inline void debug_work_activate(struct work_struct *work)
489{
490 debug_object_activate(work, &work_debug_descr);
491}
492
493static inline void debug_work_deactivate(struct work_struct *work)
494{
495 debug_object_deactivate(work, &work_debug_descr);
496}
497
498void __init_work(struct work_struct *work, int onstack)
499{
500 if (onstack)
501 debug_object_init_on_stack(work, &work_debug_descr);
502 else
503 debug_object_init(work, &work_debug_descr);
504}
505EXPORT_SYMBOL_GPL(__init_work);
506
507void destroy_work_on_stack(struct work_struct *work)
508{
509 debug_object_free(work, &work_debug_descr);
510}
511EXPORT_SYMBOL_GPL(destroy_work_on_stack);
512
513void destroy_delayed_work_on_stack(struct delayed_work *work)
514{
515 destroy_timer_on_stack(&work->timer);
516 debug_object_free(&work->work, &work_debug_descr);
517}
518EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
519
520#else
521static inline void debug_work_activate(struct work_struct *work) { }
522static inline void debug_work_deactivate(struct work_struct *work) { }
523#endif
524
525/**
526 * worker_pool_assign_id - allocate ID and assing it to @pool
527 * @pool: the pool pointer of interest
528 *
529 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
530 * successfully, -errno on failure.
531 */
532static int worker_pool_assign_id(struct worker_pool *pool)
533{
534 int ret;
535
536 lockdep_assert_held(&wq_pool_mutex);
537
538 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
539 GFP_KERNEL);
540 if (ret >= 0) {
541 pool->id = ret;
542 return 0;
543 }
544 return ret;
545}
546
547/**
548 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
549 * @wq: the target workqueue
550 * @node: the node ID
551 *
552 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
553 * read locked.
554 * If the pwq needs to be used beyond the locking in effect, the caller is
555 * responsible for guaranteeing that the pwq stays online.
556 *
557 * Return: The unbound pool_workqueue for @node.
558 */
559static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
560 int node)
561{
562 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
563
564 /*
565 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
566 * delayed item is pending. The plan is to keep CPU -> NODE
567 * mapping valid and stable across CPU on/offlines. Once that
568 * happens, this workaround can be removed.
569 */
570 if (unlikely(node == NUMA_NO_NODE))
571 return wq->dfl_pwq;
572
573 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
574}
575
576static unsigned int work_color_to_flags(int color)
577{
578 return color << WORK_STRUCT_COLOR_SHIFT;
579}
580
581static int get_work_color(struct work_struct *work)
582{
583 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
584 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
585}
586
587static int work_next_color(int color)
588{
589 return (color + 1) % WORK_NR_COLORS;
590}
591
592/*
593 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
594 * contain the pointer to the queued pwq. Once execution starts, the flag
595 * is cleared and the high bits contain OFFQ flags and pool ID.
596 *
597 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
598 * and clear_work_data() can be used to set the pwq, pool or clear
599 * work->data. These functions should only be called while the work is
600 * owned - ie. while the PENDING bit is set.
601 *
602 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
603 * corresponding to a work. Pool is available once the work has been
604 * queued anywhere after initialization until it is sync canceled. pwq is
605 * available only while the work item is queued.
606 *
607 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
608 * canceled. While being canceled, a work item may have its PENDING set
609 * but stay off timer and worklist for arbitrarily long and nobody should
610 * try to steal the PENDING bit.
611 */
612static inline void set_work_data(struct work_struct *work, unsigned long data,
613 unsigned long flags)
614{
615 WARN_ON_ONCE(!work_pending(work));
616 atomic_long_set(&work->data, data | flags | work_static(work));
617}
618
619static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
620 unsigned long extra_flags)
621{
622 set_work_data(work, (unsigned long)pwq,
623 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
624}
625
626static void set_work_pool_and_keep_pending(struct work_struct *work,
627 int pool_id)
628{
629 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
630 WORK_STRUCT_PENDING);
631}
632
633static void set_work_pool_and_clear_pending(struct work_struct *work,
634 int pool_id)
635{
636 /*
637 * The following wmb is paired with the implied mb in
638 * test_and_set_bit(PENDING) and ensures all updates to @work made
639 * here are visible to and precede any updates by the next PENDING
640 * owner.
641 */
642 smp_wmb();
643 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
644 /*
645 * The following mb guarantees that previous clear of a PENDING bit
646 * will not be reordered with any speculative LOADS or STORES from
647 * work->current_func, which is executed afterwards. This possible
648 * reordering can lead to a missed execution on attempt to queue
649 * the same @work. E.g. consider this case:
650 *
651 * CPU#0 CPU#1
652 * ---------------------------- --------------------------------
653 *
654 * 1 STORE event_indicated
655 * 2 queue_work_on() {
656 * 3 test_and_set_bit(PENDING)
657 * 4 } set_..._and_clear_pending() {
658 * 5 set_work_data() # clear bit
659 * 6 smp_mb()
660 * 7 work->current_func() {
661 * 8 LOAD event_indicated
662 * }
663 *
664 * Without an explicit full barrier speculative LOAD on line 8 can
665 * be executed before CPU#0 does STORE on line 1. If that happens,
666 * CPU#0 observes the PENDING bit is still set and new execution of
667 * a @work is not queued in a hope, that CPU#1 will eventually
668 * finish the queued @work. Meanwhile CPU#1 does not see
669 * event_indicated is set, because speculative LOAD was executed
670 * before actual STORE.
671 */
672 smp_mb();
673}
674
675static void clear_work_data(struct work_struct *work)
676{
677 smp_wmb(); /* see set_work_pool_and_clear_pending() */
678 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
679}
680
681static struct pool_workqueue *get_work_pwq(struct work_struct *work)
682{
683 unsigned long data = atomic_long_read(&work->data);
684
685 if (data & WORK_STRUCT_PWQ)
686 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
687 else
688 return NULL;
689}
690
691/**
692 * get_work_pool - return the worker_pool a given work was associated with
693 * @work: the work item of interest
694 *
695 * Pools are created and destroyed under wq_pool_mutex, and allows read
696 * access under RCU read lock. As such, this function should be
697 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
698 *
699 * All fields of the returned pool are accessible as long as the above
700 * mentioned locking is in effect. If the returned pool needs to be used
701 * beyond the critical section, the caller is responsible for ensuring the
702 * returned pool is and stays online.
703 *
704 * Return: The worker_pool @work was last associated with. %NULL if none.
705 */
706static struct worker_pool *get_work_pool(struct work_struct *work)
707{
708 unsigned long data = atomic_long_read(&work->data);
709 int pool_id;
710
711 assert_rcu_or_pool_mutex();
712
713 if (data & WORK_STRUCT_PWQ)
714 return ((struct pool_workqueue *)
715 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
716
717 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
718 if (pool_id == WORK_OFFQ_POOL_NONE)
719 return NULL;
720
721 return idr_find(&worker_pool_idr, pool_id);
722}
723
724/**
725 * get_work_pool_id - return the worker pool ID a given work is associated with
726 * @work: the work item of interest
727 *
728 * Return: The worker_pool ID @work was last associated with.
729 * %WORK_OFFQ_POOL_NONE if none.
730 */
731static int get_work_pool_id(struct work_struct *work)
732{
733 unsigned long data = atomic_long_read(&work->data);
734
735 if (data & WORK_STRUCT_PWQ)
736 return ((struct pool_workqueue *)
737 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
738
739 return data >> WORK_OFFQ_POOL_SHIFT;
740}
741
742static void mark_work_canceling(struct work_struct *work)
743{
744 unsigned long pool_id = get_work_pool_id(work);
745
746 pool_id <<= WORK_OFFQ_POOL_SHIFT;
747 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
748}
749
750static bool work_is_canceling(struct work_struct *work)
751{
752 unsigned long data = atomic_long_read(&work->data);
753
754 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
755}
756
757/*
758 * Policy functions. These define the policies on how the global worker
759 * pools are managed. Unless noted otherwise, these functions assume that
760 * they're being called with pool->lock held.
761 */
762
763static bool __need_more_worker(struct worker_pool *pool)
764{
765 return !atomic_read(&pool->nr_running);
766}
767
768/*
769 * Need to wake up a worker? Called from anything but currently
770 * running workers.
771 *
772 * Note that, because unbound workers never contribute to nr_running, this
773 * function will always return %true for unbound pools as long as the
774 * worklist isn't empty.
775 */
776static bool need_more_worker(struct worker_pool *pool)
777{
778 return !list_empty(&pool->worklist) && __need_more_worker(pool);
779}
780
781/* Can I start working? Called from busy but !running workers. */
782static bool may_start_working(struct worker_pool *pool)
783{
784 return pool->nr_idle;
785}
786
787/* Do I need to keep working? Called from currently running workers. */
788static bool keep_working(struct worker_pool *pool)
789{
790 return !list_empty(&pool->worklist) &&
791 atomic_read(&pool->nr_running) <= 1;
792}
793
794/* Do we need a new worker? Called from manager. */
795static bool need_to_create_worker(struct worker_pool *pool)
796{
797 return need_more_worker(pool) && !may_start_working(pool);
798}
799
800/* Do we have too many workers and should some go away? */
801static bool too_many_workers(struct worker_pool *pool)
802{
803 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
804 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
805 int nr_busy = pool->nr_workers - nr_idle;
806
807 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
808}
809
810/*
811 * Wake up functions.
812 */
813
814/* Return the first idle worker. Safe with preemption disabled */
815static struct worker *first_idle_worker(struct worker_pool *pool)
816{
817 if (unlikely(list_empty(&pool->idle_list)))
818 return NULL;
819
820 return list_first_entry(&pool->idle_list, struct worker, entry);
821}
822
823/**
824 * wake_up_worker - wake up an idle worker
825 * @pool: worker pool to wake worker from
826 *
827 * Wake up the first idle worker of @pool.
828 *
829 * CONTEXT:
830 * raw_spin_lock_irq(pool->lock).
831 */
832static void wake_up_worker(struct worker_pool *pool)
833{
834 struct worker *worker = first_idle_worker(pool);
835
836 if (likely(worker))
837 wake_up_process(worker->task);
838}
839
840/**
841 * wq_worker_running - a worker is running again
842 * @task: task waking up
843 *
844 * This function is called when a worker returns from schedule()
845 */
846void wq_worker_running(struct task_struct *task)
847{
848 struct worker *worker = kthread_data(task);
849
850 if (!worker->sleeping)
851 return;
852 if (!(worker->flags & WORKER_NOT_RUNNING))
853 atomic_inc(&worker->pool->nr_running);
854 worker->sleeping = 0;
855}
856
857/**
858 * wq_worker_sleeping - a worker is going to sleep
859 * @task: task going to sleep
860 *
861 * This function is called from schedule() when a busy worker is
862 * going to sleep. Preemption needs to be disabled to protect ->sleeping
863 * assignment.
864 */
865void wq_worker_sleeping(struct task_struct *task)
866{
867 struct worker *next, *worker = kthread_data(task);
868 struct worker_pool *pool;
869
870 /*
871 * Rescuers, which may not have all the fields set up like normal
872 * workers, also reach here, let's not access anything before
873 * checking NOT_RUNNING.
874 */
875 if (worker->flags & WORKER_NOT_RUNNING)
876 return;
877
878 pool = worker->pool;
879
880 /* Return if preempted before wq_worker_running() was reached */
881 if (worker->sleeping)
882 return;
883
884 worker->sleeping = 1;
885 raw_spin_lock_irq(&pool->lock);
886
887 /*
888 * The counterpart of the following dec_and_test, implied mb,
889 * worklist not empty test sequence is in insert_work().
890 * Please read comment there.
891 *
892 * NOT_RUNNING is clear. This means that we're bound to and
893 * running on the local cpu w/ rq lock held and preemption
894 * disabled, which in turn means that none else could be
895 * manipulating idle_list, so dereferencing idle_list without pool
896 * lock is safe.
897 */
898 if (atomic_dec_and_test(&pool->nr_running) &&
899 !list_empty(&pool->worklist)) {
900 next = first_idle_worker(pool);
901 if (next)
902 wake_up_process(next->task);
903 }
904 raw_spin_unlock_irq(&pool->lock);
905}
906
907/**
908 * wq_worker_last_func - retrieve worker's last work function
909 * @task: Task to retrieve last work function of.
910 *
911 * Determine the last function a worker executed. This is called from
912 * the scheduler to get a worker's last known identity.
913 *
914 * CONTEXT:
915 * raw_spin_lock_irq(rq->lock)
916 *
917 * This function is called during schedule() when a kworker is going
918 * to sleep. It's used by psi to identify aggregation workers during
919 * dequeuing, to allow periodic aggregation to shut-off when that
920 * worker is the last task in the system or cgroup to go to sleep.
921 *
922 * As this function doesn't involve any workqueue-related locking, it
923 * only returns stable values when called from inside the scheduler's
924 * queuing and dequeuing paths, when @task, which must be a kworker,
925 * is guaranteed to not be processing any works.
926 *
927 * Return:
928 * The last work function %current executed as a worker, NULL if it
929 * hasn't executed any work yet.
930 */
931work_func_t wq_worker_last_func(struct task_struct *task)
932{
933 struct worker *worker = kthread_data(task);
934
935 return worker->last_func;
936}
937
938/**
939 * worker_set_flags - set worker flags and adjust nr_running accordingly
940 * @worker: self
941 * @flags: flags to set
942 *
943 * Set @flags in @worker->flags and adjust nr_running accordingly.
944 *
945 * CONTEXT:
946 * raw_spin_lock_irq(pool->lock)
947 */
948static inline void worker_set_flags(struct worker *worker, unsigned int flags)
949{
950 struct worker_pool *pool = worker->pool;
951
952 WARN_ON_ONCE(worker->task != current);
953
954 /* If transitioning into NOT_RUNNING, adjust nr_running. */
955 if ((flags & WORKER_NOT_RUNNING) &&
956 !(worker->flags & WORKER_NOT_RUNNING)) {
957 atomic_dec(&pool->nr_running);
958 }
959
960 worker->flags |= flags;
961}
962
963/**
964 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
965 * @worker: self
966 * @flags: flags to clear
967 *
968 * Clear @flags in @worker->flags and adjust nr_running accordingly.
969 *
970 * CONTEXT:
971 * raw_spin_lock_irq(pool->lock)
972 */
973static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
974{
975 struct worker_pool *pool = worker->pool;
976 unsigned int oflags = worker->flags;
977
978 WARN_ON_ONCE(worker->task != current);
979
980 worker->flags &= ~flags;
981
982 /*
983 * If transitioning out of NOT_RUNNING, increment nr_running. Note
984 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
985 * of multiple flags, not a single flag.
986 */
987 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
988 if (!(worker->flags & WORKER_NOT_RUNNING))
989 atomic_inc(&pool->nr_running);
990}
991
992/**
993 * find_worker_executing_work - find worker which is executing a work
994 * @pool: pool of interest
995 * @work: work to find worker for
996 *
997 * Find a worker which is executing @work on @pool by searching
998 * @pool->busy_hash which is keyed by the address of @work. For a worker
999 * to match, its current execution should match the address of @work and
1000 * its work function. This is to avoid unwanted dependency between
1001 * unrelated work executions through a work item being recycled while still
1002 * being executed.
1003 *
1004 * This is a bit tricky. A work item may be freed once its execution
1005 * starts and nothing prevents the freed area from being recycled for
1006 * another work item. If the same work item address ends up being reused
1007 * before the original execution finishes, workqueue will identify the
1008 * recycled work item as currently executing and make it wait until the
1009 * current execution finishes, introducing an unwanted dependency.
1010 *
1011 * This function checks the work item address and work function to avoid
1012 * false positives. Note that this isn't complete as one may construct a
1013 * work function which can introduce dependency onto itself through a
1014 * recycled work item. Well, if somebody wants to shoot oneself in the
1015 * foot that badly, there's only so much we can do, and if such deadlock
1016 * actually occurs, it should be easy to locate the culprit work function.
1017 *
1018 * CONTEXT:
1019 * raw_spin_lock_irq(pool->lock).
1020 *
1021 * Return:
1022 * Pointer to worker which is executing @work if found, %NULL
1023 * otherwise.
1024 */
1025static struct worker *find_worker_executing_work(struct worker_pool *pool,
1026 struct work_struct *work)
1027{
1028 struct worker *worker;
1029
1030 hash_for_each_possible(pool->busy_hash, worker, hentry,
1031 (unsigned long)work)
1032 if (worker->current_work == work &&
1033 worker->current_func == work->func)
1034 return worker;
1035
1036 return NULL;
1037}
1038
1039/**
1040 * move_linked_works - move linked works to a list
1041 * @work: start of series of works to be scheduled
1042 * @head: target list to append @work to
1043 * @nextp: out parameter for nested worklist walking
1044 *
1045 * Schedule linked works starting from @work to @head. Work series to
1046 * be scheduled starts at @work and includes any consecutive work with
1047 * WORK_STRUCT_LINKED set in its predecessor.
1048 *
1049 * If @nextp is not NULL, it's updated to point to the next work of
1050 * the last scheduled work. This allows move_linked_works() to be
1051 * nested inside outer list_for_each_entry_safe().
1052 *
1053 * CONTEXT:
1054 * raw_spin_lock_irq(pool->lock).
1055 */
1056static void move_linked_works(struct work_struct *work, struct list_head *head,
1057 struct work_struct **nextp)
1058{
1059 struct work_struct *n;
1060
1061 /*
1062 * Linked worklist will always end before the end of the list,
1063 * use NULL for list head.
1064 */
1065 list_for_each_entry_safe_from(work, n, NULL, entry) {
1066 list_move_tail(&work->entry, head);
1067 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1068 break;
1069 }
1070
1071 /*
1072 * If we're already inside safe list traversal and have moved
1073 * multiple works to the scheduled queue, the next position
1074 * needs to be updated.
1075 */
1076 if (nextp)
1077 *nextp = n;
1078}
1079
1080/**
1081 * get_pwq - get an extra reference on the specified pool_workqueue
1082 * @pwq: pool_workqueue to get
1083 *
1084 * Obtain an extra reference on @pwq. The caller should guarantee that
1085 * @pwq has positive refcnt and be holding the matching pool->lock.
1086 */
1087static void get_pwq(struct pool_workqueue *pwq)
1088{
1089 lockdep_assert_held(&pwq->pool->lock);
1090 WARN_ON_ONCE(pwq->refcnt <= 0);
1091 pwq->refcnt++;
1092}
1093
1094/**
1095 * put_pwq - put a pool_workqueue reference
1096 * @pwq: pool_workqueue to put
1097 *
1098 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1099 * destruction. The caller should be holding the matching pool->lock.
1100 */
1101static void put_pwq(struct pool_workqueue *pwq)
1102{
1103 lockdep_assert_held(&pwq->pool->lock);
1104 if (likely(--pwq->refcnt))
1105 return;
1106 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1107 return;
1108 /*
1109 * @pwq can't be released under pool->lock, bounce to
1110 * pwq_unbound_release_workfn(). This never recurses on the same
1111 * pool->lock as this path is taken only for unbound workqueues and
1112 * the release work item is scheduled on a per-cpu workqueue. To
1113 * avoid lockdep warning, unbound pool->locks are given lockdep
1114 * subclass of 1 in get_unbound_pool().
1115 */
1116 schedule_work(&pwq->unbound_release_work);
1117}
1118
1119/**
1120 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1121 * @pwq: pool_workqueue to put (can be %NULL)
1122 *
1123 * put_pwq() with locking. This function also allows %NULL @pwq.
1124 */
1125static void put_pwq_unlocked(struct pool_workqueue *pwq)
1126{
1127 if (pwq) {
1128 /*
1129 * As both pwqs and pools are RCU protected, the
1130 * following lock operations are safe.
1131 */
1132 raw_spin_lock_irq(&pwq->pool->lock);
1133 put_pwq(pwq);
1134 raw_spin_unlock_irq(&pwq->pool->lock);
1135 }
1136}
1137
1138static void pwq_activate_delayed_work(struct work_struct *work)
1139{
1140 struct pool_workqueue *pwq = get_work_pwq(work);
1141
1142 trace_workqueue_activate_work(work);
1143 if (list_empty(&pwq->pool->worklist))
1144 pwq->pool->watchdog_ts = jiffies;
1145 move_linked_works(work, &pwq->pool->worklist, NULL);
1146 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1147 pwq->nr_active++;
1148}
1149
1150static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1151{
1152 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1153 struct work_struct, entry);
1154
1155 pwq_activate_delayed_work(work);
1156}
1157
1158/**
1159 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1160 * @pwq: pwq of interest
1161 * @color: color of work which left the queue
1162 *
1163 * A work either has completed or is removed from pending queue,
1164 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1165 *
1166 * CONTEXT:
1167 * raw_spin_lock_irq(pool->lock).
1168 */
1169static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1170{
1171 /* uncolored work items don't participate in flushing or nr_active */
1172 if (color == WORK_NO_COLOR)
1173 goto out_put;
1174
1175 pwq->nr_in_flight[color]--;
1176
1177 pwq->nr_active--;
1178 if (!list_empty(&pwq->delayed_works)) {
1179 /* one down, submit a delayed one */
1180 if (pwq->nr_active < pwq->max_active)
1181 pwq_activate_first_delayed(pwq);
1182 }
1183
1184 /* is flush in progress and are we at the flushing tip? */
1185 if (likely(pwq->flush_color != color))
1186 goto out_put;
1187
1188 /* are there still in-flight works? */
1189 if (pwq->nr_in_flight[color])
1190 goto out_put;
1191
1192 /* this pwq is done, clear flush_color */
1193 pwq->flush_color = -1;
1194
1195 /*
1196 * If this was the last pwq, wake up the first flusher. It
1197 * will handle the rest.
1198 */
1199 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1200 complete(&pwq->wq->first_flusher->done);
1201out_put:
1202 put_pwq(pwq);
1203}
1204
1205/**
1206 * try_to_grab_pending - steal work item from worklist and disable irq
1207 * @work: work item to steal
1208 * @is_dwork: @work is a delayed_work
1209 * @flags: place to store irq state
1210 *
1211 * Try to grab PENDING bit of @work. This function can handle @work in any
1212 * stable state - idle, on timer or on worklist.
1213 *
1214 * Return:
1215 * 1 if @work was pending and we successfully stole PENDING
1216 * 0 if @work was idle and we claimed PENDING
1217 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1218 * -ENOENT if someone else is canceling @work, this state may persist
1219 * for arbitrarily long
1220 *
1221 * Note:
1222 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1223 * interrupted while holding PENDING and @work off queue, irq must be
1224 * disabled on entry. This, combined with delayed_work->timer being
1225 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1226 *
1227 * On successful return, >= 0, irq is disabled and the caller is
1228 * responsible for releasing it using local_irq_restore(*@flags).
1229 *
1230 * This function is safe to call from any context including IRQ handler.
1231 */
1232static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1233 unsigned long *flags)
1234{
1235 struct worker_pool *pool;
1236 struct pool_workqueue *pwq;
1237
1238 local_irq_save(*flags);
1239
1240 /* try to steal the timer if it exists */
1241 if (is_dwork) {
1242 struct delayed_work *dwork = to_delayed_work(work);
1243
1244 /*
1245 * dwork->timer is irqsafe. If del_timer() fails, it's
1246 * guaranteed that the timer is not queued anywhere and not
1247 * running on the local CPU.
1248 */
1249 if (likely(del_timer(&dwork->timer)))
1250 return 1;
1251 }
1252
1253 /* try to claim PENDING the normal way */
1254 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1255 return 0;
1256
1257 rcu_read_lock();
1258 /*
1259 * The queueing is in progress, or it is already queued. Try to
1260 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1261 */
1262 pool = get_work_pool(work);
1263 if (!pool)
1264 goto fail;
1265
1266 raw_spin_lock(&pool->lock);
1267 /*
1268 * work->data is guaranteed to point to pwq only while the work
1269 * item is queued on pwq->wq, and both updating work->data to point
1270 * to pwq on queueing and to pool on dequeueing are done under
1271 * pwq->pool->lock. This in turn guarantees that, if work->data
1272 * points to pwq which is associated with a locked pool, the work
1273 * item is currently queued on that pool.
1274 */
1275 pwq = get_work_pwq(work);
1276 if (pwq && pwq->pool == pool) {
1277 debug_work_deactivate(work);
1278
1279 /*
1280 * A delayed work item cannot be grabbed directly because
1281 * it might have linked NO_COLOR work items which, if left
1282 * on the delayed_list, will confuse pwq->nr_active
1283 * management later on and cause stall. Make sure the work
1284 * item is activated before grabbing.
1285 */
1286 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1287 pwq_activate_delayed_work(work);
1288
1289 list_del_init(&work->entry);
1290 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1291
1292 /* work->data points to pwq iff queued, point to pool */
1293 set_work_pool_and_keep_pending(work, pool->id);
1294
1295 raw_spin_unlock(&pool->lock);
1296 rcu_read_unlock();
1297 return 1;
1298 }
1299 raw_spin_unlock(&pool->lock);
1300fail:
1301 rcu_read_unlock();
1302 local_irq_restore(*flags);
1303 if (work_is_canceling(work))
1304 return -ENOENT;
1305 cpu_relax();
1306 return -EAGAIN;
1307}
1308
1309/**
1310 * insert_work - insert a work into a pool
1311 * @pwq: pwq @work belongs to
1312 * @work: work to insert
1313 * @head: insertion point
1314 * @extra_flags: extra WORK_STRUCT_* flags to set
1315 *
1316 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1317 * work_struct flags.
1318 *
1319 * CONTEXT:
1320 * raw_spin_lock_irq(pool->lock).
1321 */
1322static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1323 struct list_head *head, unsigned int extra_flags)
1324{
1325 struct worker_pool *pool = pwq->pool;
1326
1327 /* we own @work, set data and link */
1328 set_work_pwq(work, pwq, extra_flags);
1329 list_add_tail(&work->entry, head);
1330 get_pwq(pwq);
1331
1332 /*
1333 * Ensure either wq_worker_sleeping() sees the above
1334 * list_add_tail() or we see zero nr_running to avoid workers lying
1335 * around lazily while there are works to be processed.
1336 */
1337 smp_mb();
1338
1339 if (__need_more_worker(pool))
1340 wake_up_worker(pool);
1341}
1342
1343/*
1344 * Test whether @work is being queued from another work executing on the
1345 * same workqueue.
1346 */
1347static bool is_chained_work(struct workqueue_struct *wq)
1348{
1349 struct worker *worker;
1350
1351 worker = current_wq_worker();
1352 /*
1353 * Return %true iff I'm a worker executing a work item on @wq. If
1354 * I'm @worker, it's safe to dereference it without locking.
1355 */
1356 return worker && worker->current_pwq->wq == wq;
1357}
1358
1359/*
1360 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1361 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1362 * avoid perturbing sensitive tasks.
1363 */
1364static int wq_select_unbound_cpu(int cpu)
1365{
1366 static bool printed_dbg_warning;
1367 int new_cpu;
1368
1369 if (likely(!wq_debug_force_rr_cpu)) {
1370 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1371 return cpu;
1372 } else if (!printed_dbg_warning) {
1373 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1374 printed_dbg_warning = true;
1375 }
1376
1377 if (cpumask_empty(wq_unbound_cpumask))
1378 return cpu;
1379
1380 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1381 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1382 if (unlikely(new_cpu >= nr_cpu_ids)) {
1383 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1384 if (unlikely(new_cpu >= nr_cpu_ids))
1385 return cpu;
1386 }
1387 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1388
1389 return new_cpu;
1390}
1391
1392static void __queue_work(int cpu, struct workqueue_struct *wq,
1393 struct work_struct *work)
1394{
1395 struct pool_workqueue *pwq;
1396 struct worker_pool *last_pool;
1397 struct list_head *worklist;
1398 unsigned int work_flags;
1399 unsigned int req_cpu = cpu;
1400
1401 /*
1402 * While a work item is PENDING && off queue, a task trying to
1403 * steal the PENDING will busy-loop waiting for it to either get
1404 * queued or lose PENDING. Grabbing PENDING and queueing should
1405 * happen with IRQ disabled.
1406 */
1407 lockdep_assert_irqs_disabled();
1408
1409 debug_work_activate(work);
1410
1411 /* if draining, only works from the same workqueue are allowed */
1412 if (unlikely(wq->flags & __WQ_DRAINING) &&
1413 WARN_ON_ONCE(!is_chained_work(wq)))
1414 return;
1415 rcu_read_lock();
1416retry:
1417 /* pwq which will be used unless @work is executing elsewhere */
1418 if (wq->flags & WQ_UNBOUND) {
1419 if (req_cpu == WORK_CPU_UNBOUND)
1420 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1421 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1422 } else {
1423 if (req_cpu == WORK_CPU_UNBOUND)
1424 cpu = raw_smp_processor_id();
1425 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1426 }
1427
1428 /*
1429 * If @work was previously on a different pool, it might still be
1430 * running there, in which case the work needs to be queued on that
1431 * pool to guarantee non-reentrancy.
1432 */
1433 last_pool = get_work_pool(work);
1434 if (last_pool && last_pool != pwq->pool) {
1435 struct worker *worker;
1436
1437 raw_spin_lock(&last_pool->lock);
1438
1439 worker = find_worker_executing_work(last_pool, work);
1440
1441 if (worker && worker->current_pwq->wq == wq) {
1442 pwq = worker->current_pwq;
1443 } else {
1444 /* meh... not running there, queue here */
1445 raw_spin_unlock(&last_pool->lock);
1446 raw_spin_lock(&pwq->pool->lock);
1447 }
1448 } else {
1449 raw_spin_lock(&pwq->pool->lock);
1450 }
1451
1452 /*
1453 * pwq is determined and locked. For unbound pools, we could have
1454 * raced with pwq release and it could already be dead. If its
1455 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1456 * without another pwq replacing it in the numa_pwq_tbl or while
1457 * work items are executing on it, so the retrying is guaranteed to
1458 * make forward-progress.
1459 */
1460 if (unlikely(!pwq->refcnt)) {
1461 if (wq->flags & WQ_UNBOUND) {
1462 raw_spin_unlock(&pwq->pool->lock);
1463 cpu_relax();
1464 goto retry;
1465 }
1466 /* oops */
1467 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1468 wq->name, cpu);
1469 }
1470
1471 /* pwq determined, queue */
1472 trace_workqueue_queue_work(req_cpu, pwq, work);
1473
1474 if (WARN_ON(!list_empty(&work->entry)))
1475 goto out;
1476
1477 pwq->nr_in_flight[pwq->work_color]++;
1478 work_flags = work_color_to_flags(pwq->work_color);
1479
1480 if (likely(pwq->nr_active < pwq->max_active)) {
1481 trace_workqueue_activate_work(work);
1482 pwq->nr_active++;
1483 worklist = &pwq->pool->worklist;
1484 if (list_empty(worklist))
1485 pwq->pool->watchdog_ts = jiffies;
1486 } else {
1487 work_flags |= WORK_STRUCT_DELAYED;
1488 worklist = &pwq->delayed_works;
1489 }
1490
1491 insert_work(pwq, work, worklist, work_flags);
1492
1493out:
1494 raw_spin_unlock(&pwq->pool->lock);
1495 rcu_read_unlock();
1496}
1497
1498/**
1499 * queue_work_on - queue work on specific cpu
1500 * @cpu: CPU number to execute work on
1501 * @wq: workqueue to use
1502 * @work: work to queue
1503 *
1504 * We queue the work to a specific CPU, the caller must ensure it
1505 * can't go away.
1506 *
1507 * Return: %false if @work was already on a queue, %true otherwise.
1508 */
1509bool queue_work_on(int cpu, struct workqueue_struct *wq,
1510 struct work_struct *work)
1511{
1512 bool ret = false;
1513 unsigned long flags;
1514
1515 local_irq_save(flags);
1516
1517 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1518 __queue_work(cpu, wq, work);
1519 ret = true;
1520 }
1521
1522 local_irq_restore(flags);
1523 return ret;
1524}
1525EXPORT_SYMBOL(queue_work_on);
1526
1527/**
1528 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1529 * @node: NUMA node ID that we want to select a CPU from
1530 *
1531 * This function will attempt to find a "random" cpu available on a given
1532 * node. If there are no CPUs available on the given node it will return
1533 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1534 * available CPU if we need to schedule this work.
1535 */
1536static int workqueue_select_cpu_near(int node)
1537{
1538 int cpu;
1539
1540 /* No point in doing this if NUMA isn't enabled for workqueues */
1541 if (!wq_numa_enabled)
1542 return WORK_CPU_UNBOUND;
1543
1544 /* Delay binding to CPU if node is not valid or online */
1545 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1546 return WORK_CPU_UNBOUND;
1547
1548 /* Use local node/cpu if we are already there */
1549 cpu = raw_smp_processor_id();
1550 if (node == cpu_to_node(cpu))
1551 return cpu;
1552
1553 /* Use "random" otherwise know as "first" online CPU of node */
1554 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1555
1556 /* If CPU is valid return that, otherwise just defer */
1557 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1558}
1559
1560/**
1561 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1562 * @node: NUMA node that we are targeting the work for
1563 * @wq: workqueue to use
1564 * @work: work to queue
1565 *
1566 * We queue the work to a "random" CPU within a given NUMA node. The basic
1567 * idea here is to provide a way to somehow associate work with a given
1568 * NUMA node.
1569 *
1570 * This function will only make a best effort attempt at getting this onto
1571 * the right NUMA node. If no node is requested or the requested node is
1572 * offline then we just fall back to standard queue_work behavior.
1573 *
1574 * Currently the "random" CPU ends up being the first available CPU in the
1575 * intersection of cpu_online_mask and the cpumask of the node, unless we
1576 * are running on the node. In that case we just use the current CPU.
1577 *
1578 * Return: %false if @work was already on a queue, %true otherwise.
1579 */
1580bool queue_work_node(int node, struct workqueue_struct *wq,
1581 struct work_struct *work)
1582{
1583 unsigned long flags;
1584 bool ret = false;
1585
1586 /*
1587 * This current implementation is specific to unbound workqueues.
1588 * Specifically we only return the first available CPU for a given
1589 * node instead of cycling through individual CPUs within the node.
1590 *
1591 * If this is used with a per-cpu workqueue then the logic in
1592 * workqueue_select_cpu_near would need to be updated to allow for
1593 * some round robin type logic.
1594 */
1595 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1596
1597 local_irq_save(flags);
1598
1599 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1600 int cpu = workqueue_select_cpu_near(node);
1601
1602 __queue_work(cpu, wq, work);
1603 ret = true;
1604 }
1605
1606 local_irq_restore(flags);
1607 return ret;
1608}
1609EXPORT_SYMBOL_GPL(queue_work_node);
1610
1611void delayed_work_timer_fn(struct timer_list *t)
1612{
1613 struct delayed_work *dwork = from_timer(dwork, t, timer);
1614
1615 /* should have been called from irqsafe timer with irq already off */
1616 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1617}
1618EXPORT_SYMBOL(delayed_work_timer_fn);
1619
1620static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1621 struct delayed_work *dwork, unsigned long delay)
1622{
1623 struct timer_list *timer = &dwork->timer;
1624 struct work_struct *work = &dwork->work;
1625
1626 WARN_ON_ONCE(!wq);
1627 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1628 WARN_ON_ONCE(timer_pending(timer));
1629 WARN_ON_ONCE(!list_empty(&work->entry));
1630
1631 /*
1632 * If @delay is 0, queue @dwork->work immediately. This is for
1633 * both optimization and correctness. The earliest @timer can
1634 * expire is on the closest next tick and delayed_work users depend
1635 * on that there's no such delay when @delay is 0.
1636 */
1637 if (!delay) {
1638 __queue_work(cpu, wq, &dwork->work);
1639 return;
1640 }
1641
1642 dwork->wq = wq;
1643 dwork->cpu = cpu;
1644 timer->expires = jiffies + delay;
1645
1646 if (unlikely(cpu != WORK_CPU_UNBOUND))
1647 add_timer_on(timer, cpu);
1648 else
1649 add_timer(timer);
1650}
1651
1652/**
1653 * queue_delayed_work_on - queue work on specific CPU after delay
1654 * @cpu: CPU number to execute work on
1655 * @wq: workqueue to use
1656 * @dwork: work to queue
1657 * @delay: number of jiffies to wait before queueing
1658 *
1659 * Return: %false if @work was already on a queue, %true otherwise. If
1660 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1661 * execution.
1662 */
1663bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1664 struct delayed_work *dwork, unsigned long delay)
1665{
1666 struct work_struct *work = &dwork->work;
1667 bool ret = false;
1668 unsigned long flags;
1669
1670 /* read the comment in __queue_work() */
1671 local_irq_save(flags);
1672
1673 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1674 __queue_delayed_work(cpu, wq, dwork, delay);
1675 ret = true;
1676 }
1677
1678 local_irq_restore(flags);
1679 return ret;
1680}
1681EXPORT_SYMBOL(queue_delayed_work_on);
1682
1683/**
1684 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1685 * @cpu: CPU number to execute work on
1686 * @wq: workqueue to use
1687 * @dwork: work to queue
1688 * @delay: number of jiffies to wait before queueing
1689 *
1690 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1691 * modify @dwork's timer so that it expires after @delay. If @delay is
1692 * zero, @work is guaranteed to be scheduled immediately regardless of its
1693 * current state.
1694 *
1695 * Return: %false if @dwork was idle and queued, %true if @dwork was
1696 * pending and its timer was modified.
1697 *
1698 * This function is safe to call from any context including IRQ handler.
1699 * See try_to_grab_pending() for details.
1700 */
1701bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1702 struct delayed_work *dwork, unsigned long delay)
1703{
1704 unsigned long flags;
1705 int ret;
1706
1707 do {
1708 ret = try_to_grab_pending(&dwork->work, true, &flags);
1709 } while (unlikely(ret == -EAGAIN));
1710
1711 if (likely(ret >= 0)) {
1712 __queue_delayed_work(cpu, wq, dwork, delay);
1713 local_irq_restore(flags);
1714 }
1715
1716 /* -ENOENT from try_to_grab_pending() becomes %true */
1717 return ret;
1718}
1719EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1720
1721static void rcu_work_rcufn(struct rcu_head *rcu)
1722{
1723 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1724
1725 /* read the comment in __queue_work() */
1726 local_irq_disable();
1727 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1728 local_irq_enable();
1729}
1730
1731/**
1732 * queue_rcu_work - queue work after a RCU grace period
1733 * @wq: workqueue to use
1734 * @rwork: work to queue
1735 *
1736 * Return: %false if @rwork was already pending, %true otherwise. Note
1737 * that a full RCU grace period is guaranteed only after a %true return.
1738 * While @rwork is guaranteed to be executed after a %false return, the
1739 * execution may happen before a full RCU grace period has passed.
1740 */
1741bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1742{
1743 struct work_struct *work = &rwork->work;
1744
1745 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1746 rwork->wq = wq;
1747 call_rcu(&rwork->rcu, rcu_work_rcufn);
1748 return true;
1749 }
1750
1751 return false;
1752}
1753EXPORT_SYMBOL(queue_rcu_work);
1754
1755/**
1756 * worker_enter_idle - enter idle state
1757 * @worker: worker which is entering idle state
1758 *
1759 * @worker is entering idle state. Update stats and idle timer if
1760 * necessary.
1761 *
1762 * LOCKING:
1763 * raw_spin_lock_irq(pool->lock).
1764 */
1765static void worker_enter_idle(struct worker *worker)
1766{
1767 struct worker_pool *pool = worker->pool;
1768
1769 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1770 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1771 (worker->hentry.next || worker->hentry.pprev)))
1772 return;
1773
1774 /* can't use worker_set_flags(), also called from create_worker() */
1775 worker->flags |= WORKER_IDLE;
1776 pool->nr_idle++;
1777 worker->last_active = jiffies;
1778
1779 /* idle_list is LIFO */
1780 list_add(&worker->entry, &pool->idle_list);
1781
1782 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1783 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1784
1785 /*
1786 * Sanity check nr_running. Because unbind_workers() releases
1787 * pool->lock between setting %WORKER_UNBOUND and zapping
1788 * nr_running, the warning may trigger spuriously. Check iff
1789 * unbind is not in progress.
1790 */
1791 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1792 pool->nr_workers == pool->nr_idle &&
1793 atomic_read(&pool->nr_running));
1794}
1795
1796/**
1797 * worker_leave_idle - leave idle state
1798 * @worker: worker which is leaving idle state
1799 *
1800 * @worker is leaving idle state. Update stats.
1801 *
1802 * LOCKING:
1803 * raw_spin_lock_irq(pool->lock).
1804 */
1805static void worker_leave_idle(struct worker *worker)
1806{
1807 struct worker_pool *pool = worker->pool;
1808
1809 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1810 return;
1811 worker_clr_flags(worker, WORKER_IDLE);
1812 pool->nr_idle--;
1813 list_del_init(&worker->entry);
1814}
1815
1816static struct worker *alloc_worker(int node)
1817{
1818 struct worker *worker;
1819
1820 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1821 if (worker) {
1822 INIT_LIST_HEAD(&worker->entry);
1823 INIT_LIST_HEAD(&worker->scheduled);
1824 INIT_LIST_HEAD(&worker->node);
1825 /* on creation a worker is in !idle && prep state */
1826 worker->flags = WORKER_PREP;
1827 }
1828 return worker;
1829}
1830
1831/**
1832 * worker_attach_to_pool() - attach a worker to a pool
1833 * @worker: worker to be attached
1834 * @pool: the target pool
1835 *
1836 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1837 * cpu-binding of @worker are kept coordinated with the pool across
1838 * cpu-[un]hotplugs.
1839 */
1840static void worker_attach_to_pool(struct worker *worker,
1841 struct worker_pool *pool)
1842{
1843 mutex_lock(&wq_pool_attach_mutex);
1844
1845 /*
1846 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1847 * online CPUs. It'll be re-applied when any of the CPUs come up.
1848 */
1849 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1850
1851 /*
1852 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1853 * stable across this function. See the comments above the flag
1854 * definition for details.
1855 */
1856 if (pool->flags & POOL_DISASSOCIATED)
1857 worker->flags |= WORKER_UNBOUND;
1858
1859 list_add_tail(&worker->node, &pool->workers);
1860 worker->pool = pool;
1861
1862 mutex_unlock(&wq_pool_attach_mutex);
1863}
1864
1865/**
1866 * worker_detach_from_pool() - detach a worker from its pool
1867 * @worker: worker which is attached to its pool
1868 *
1869 * Undo the attaching which had been done in worker_attach_to_pool(). The
1870 * caller worker shouldn't access to the pool after detached except it has
1871 * other reference to the pool.
1872 */
1873static void worker_detach_from_pool(struct worker *worker)
1874{
1875 struct worker_pool *pool = worker->pool;
1876 struct completion *detach_completion = NULL;
1877
1878 mutex_lock(&wq_pool_attach_mutex);
1879
1880 list_del(&worker->node);
1881 worker->pool = NULL;
1882
1883 if (list_empty(&pool->workers))
1884 detach_completion = pool->detach_completion;
1885 mutex_unlock(&wq_pool_attach_mutex);
1886
1887 /* clear leftover flags without pool->lock after it is detached */
1888 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1889
1890 if (detach_completion)
1891 complete(detach_completion);
1892}
1893
1894/**
1895 * create_worker - create a new workqueue worker
1896 * @pool: pool the new worker will belong to
1897 *
1898 * Create and start a new worker which is attached to @pool.
1899 *
1900 * CONTEXT:
1901 * Might sleep. Does GFP_KERNEL allocations.
1902 *
1903 * Return:
1904 * Pointer to the newly created worker.
1905 */
1906static struct worker *create_worker(struct worker_pool *pool)
1907{
1908 struct worker *worker = NULL;
1909 int id = -1;
1910 char id_buf[16];
1911
1912 /* ID is needed to determine kthread name */
1913 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1914 if (id < 0)
1915 goto fail;
1916
1917 worker = alloc_worker(pool->node);
1918 if (!worker)
1919 goto fail;
1920
1921 worker->id = id;
1922
1923 if (pool->cpu >= 0)
1924 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1925 pool->attrs->nice < 0 ? "H" : "");
1926 else
1927 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1928
1929 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1930 "kworker/%s", id_buf);
1931 if (IS_ERR(worker->task))
1932 goto fail;
1933
1934 set_user_nice(worker->task, pool->attrs->nice);
1935 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1936
1937 /* successful, attach the worker to the pool */
1938 worker_attach_to_pool(worker, pool);
1939
1940 /* start the newly created worker */
1941 raw_spin_lock_irq(&pool->lock);
1942 worker->pool->nr_workers++;
1943 worker_enter_idle(worker);
1944 wake_up_process(worker->task);
1945 raw_spin_unlock_irq(&pool->lock);
1946
1947 return worker;
1948
1949fail:
1950 if (id >= 0)
1951 ida_simple_remove(&pool->worker_ida, id);
1952 kfree(worker);
1953 return NULL;
1954}
1955
1956/**
1957 * destroy_worker - destroy a workqueue worker
1958 * @worker: worker to be destroyed
1959 *
1960 * Destroy @worker and adjust @pool stats accordingly. The worker should
1961 * be idle.
1962 *
1963 * CONTEXT:
1964 * raw_spin_lock_irq(pool->lock).
1965 */
1966static void destroy_worker(struct worker *worker)
1967{
1968 struct worker_pool *pool = worker->pool;
1969
1970 lockdep_assert_held(&pool->lock);
1971
1972 /* sanity check frenzy */
1973 if (WARN_ON(worker->current_work) ||
1974 WARN_ON(!list_empty(&worker->scheduled)) ||
1975 WARN_ON(!(worker->flags & WORKER_IDLE)))
1976 return;
1977
1978 pool->nr_workers--;
1979 pool->nr_idle--;
1980
1981 list_del_init(&worker->entry);
1982 worker->flags |= WORKER_DIE;
1983 wake_up_process(worker->task);
1984}
1985
1986static void idle_worker_timeout(struct timer_list *t)
1987{
1988 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1989
1990 raw_spin_lock_irq(&pool->lock);
1991
1992 while (too_many_workers(pool)) {
1993 struct worker *worker;
1994 unsigned long expires;
1995
1996 /* idle_list is kept in LIFO order, check the last one */
1997 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1998 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1999
2000 if (time_before(jiffies, expires)) {
2001 mod_timer(&pool->idle_timer, expires);
2002 break;
2003 }
2004
2005 destroy_worker(worker);
2006 }
2007
2008 raw_spin_unlock_irq(&pool->lock);
2009}
2010
2011static void send_mayday(struct work_struct *work)
2012{
2013 struct pool_workqueue *pwq = get_work_pwq(work);
2014 struct workqueue_struct *wq = pwq->wq;
2015
2016 lockdep_assert_held(&wq_mayday_lock);
2017
2018 if (!wq->rescuer)
2019 return;
2020
2021 /* mayday mayday mayday */
2022 if (list_empty(&pwq->mayday_node)) {
2023 /*
2024 * If @pwq is for an unbound wq, its base ref may be put at
2025 * any time due to an attribute change. Pin @pwq until the
2026 * rescuer is done with it.
2027 */
2028 get_pwq(pwq);
2029 list_add_tail(&pwq->mayday_node, &wq->maydays);
2030 wake_up_process(wq->rescuer->task);
2031 }
2032}
2033
2034static void pool_mayday_timeout(struct timer_list *t)
2035{
2036 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2037 struct work_struct *work;
2038
2039 raw_spin_lock_irq(&pool->lock);
2040 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2041
2042 if (need_to_create_worker(pool)) {
2043 /*
2044 * We've been trying to create a new worker but
2045 * haven't been successful. We might be hitting an
2046 * allocation deadlock. Send distress signals to
2047 * rescuers.
2048 */
2049 list_for_each_entry(work, &pool->worklist, entry)
2050 send_mayday(work);
2051 }
2052
2053 raw_spin_unlock(&wq_mayday_lock);
2054 raw_spin_unlock_irq(&pool->lock);
2055
2056 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2057}
2058
2059/**
2060 * maybe_create_worker - create a new worker if necessary
2061 * @pool: pool to create a new worker for
2062 *
2063 * Create a new worker for @pool if necessary. @pool is guaranteed to
2064 * have at least one idle worker on return from this function. If
2065 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2066 * sent to all rescuers with works scheduled on @pool to resolve
2067 * possible allocation deadlock.
2068 *
2069 * On return, need_to_create_worker() is guaranteed to be %false and
2070 * may_start_working() %true.
2071 *
2072 * LOCKING:
2073 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2074 * multiple times. Does GFP_KERNEL allocations. Called only from
2075 * manager.
2076 */
2077static void maybe_create_worker(struct worker_pool *pool)
2078__releases(&pool->lock)
2079__acquires(&pool->lock)
2080{
2081restart:
2082 raw_spin_unlock_irq(&pool->lock);
2083
2084 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2085 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2086
2087 while (true) {
2088 if (create_worker(pool) || !need_to_create_worker(pool))
2089 break;
2090
2091 schedule_timeout_interruptible(CREATE_COOLDOWN);
2092
2093 if (!need_to_create_worker(pool))
2094 break;
2095 }
2096
2097 del_timer_sync(&pool->mayday_timer);
2098 raw_spin_lock_irq(&pool->lock);
2099 /*
2100 * This is necessary even after a new worker was just successfully
2101 * created as @pool->lock was dropped and the new worker might have
2102 * already become busy.
2103 */
2104 if (need_to_create_worker(pool))
2105 goto restart;
2106}
2107
2108/**
2109 * manage_workers - manage worker pool
2110 * @worker: self
2111 *
2112 * Assume the manager role and manage the worker pool @worker belongs
2113 * to. At any given time, there can be only zero or one manager per
2114 * pool. The exclusion is handled automatically by this function.
2115 *
2116 * The caller can safely start processing works on false return. On
2117 * true return, it's guaranteed that need_to_create_worker() is false
2118 * and may_start_working() is true.
2119 *
2120 * CONTEXT:
2121 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2122 * multiple times. Does GFP_KERNEL allocations.
2123 *
2124 * Return:
2125 * %false if the pool doesn't need management and the caller can safely
2126 * start processing works, %true if management function was performed and
2127 * the conditions that the caller verified before calling the function may
2128 * no longer be true.
2129 */
2130static bool manage_workers(struct worker *worker)
2131{
2132 struct worker_pool *pool = worker->pool;
2133
2134 if (pool->flags & POOL_MANAGER_ACTIVE)
2135 return false;
2136
2137 pool->flags |= POOL_MANAGER_ACTIVE;
2138 pool->manager = worker;
2139
2140 maybe_create_worker(pool);
2141
2142 pool->manager = NULL;
2143 pool->flags &= ~POOL_MANAGER_ACTIVE;
2144 rcuwait_wake_up(&manager_wait);
2145 return true;
2146}
2147
2148/**
2149 * process_one_work - process single work
2150 * @worker: self
2151 * @work: work to process
2152 *
2153 * Process @work. This function contains all the logics necessary to
2154 * process a single work including synchronization against and
2155 * interaction with other workers on the same cpu, queueing and
2156 * flushing. As long as context requirement is met, any worker can
2157 * call this function to process a work.
2158 *
2159 * CONTEXT:
2160 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2161 */
2162static void process_one_work(struct worker *worker, struct work_struct *work)
2163__releases(&pool->lock)
2164__acquires(&pool->lock)
2165{
2166 struct pool_workqueue *pwq = get_work_pwq(work);
2167 struct worker_pool *pool = worker->pool;
2168 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2169 int work_color;
2170 struct worker *collision;
2171#ifdef CONFIG_LOCKDEP
2172 /*
2173 * It is permissible to free the struct work_struct from
2174 * inside the function that is called from it, this we need to
2175 * take into account for lockdep too. To avoid bogus "held
2176 * lock freed" warnings as well as problems when looking into
2177 * work->lockdep_map, make a copy and use that here.
2178 */
2179 struct lockdep_map lockdep_map;
2180
2181 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2182#endif
2183 /* ensure we're on the correct CPU */
2184 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2185 raw_smp_processor_id() != pool->cpu);
2186
2187 /*
2188 * A single work shouldn't be executed concurrently by
2189 * multiple workers on a single cpu. Check whether anyone is
2190 * already processing the work. If so, defer the work to the
2191 * currently executing one.
2192 */
2193 collision = find_worker_executing_work(pool, work);
2194 if (unlikely(collision)) {
2195 move_linked_works(work, &collision->scheduled, NULL);
2196 return;
2197 }
2198
2199 /* claim and dequeue */
2200 debug_work_deactivate(work);
2201 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2202 worker->current_work = work;
2203 worker->current_func = work->func;
2204 worker->current_pwq = pwq;
2205 work_color = get_work_color(work);
2206
2207 /*
2208 * Record wq name for cmdline and debug reporting, may get
2209 * overridden through set_worker_desc().
2210 */
2211 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2212
2213 list_del_init(&work->entry);
2214
2215 /*
2216 * CPU intensive works don't participate in concurrency management.
2217 * They're the scheduler's responsibility. This takes @worker out
2218 * of concurrency management and the next code block will chain
2219 * execution of the pending work items.
2220 */
2221 if (unlikely(cpu_intensive))
2222 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2223
2224 /*
2225 * Wake up another worker if necessary. The condition is always
2226 * false for normal per-cpu workers since nr_running would always
2227 * be >= 1 at this point. This is used to chain execution of the
2228 * pending work items for WORKER_NOT_RUNNING workers such as the
2229 * UNBOUND and CPU_INTENSIVE ones.
2230 */
2231 if (need_more_worker(pool))
2232 wake_up_worker(pool);
2233
2234 /*
2235 * Record the last pool and clear PENDING which should be the last
2236 * update to @work. Also, do this inside @pool->lock so that
2237 * PENDING and queued state changes happen together while IRQ is
2238 * disabled.
2239 */
2240 set_work_pool_and_clear_pending(work, pool->id);
2241
2242 raw_spin_unlock_irq(&pool->lock);
2243
2244 lock_map_acquire(&pwq->wq->lockdep_map);
2245 lock_map_acquire(&lockdep_map);
2246 /*
2247 * Strictly speaking we should mark the invariant state without holding
2248 * any locks, that is, before these two lock_map_acquire()'s.
2249 *
2250 * However, that would result in:
2251 *
2252 * A(W1)
2253 * WFC(C)
2254 * A(W1)
2255 * C(C)
2256 *
2257 * Which would create W1->C->W1 dependencies, even though there is no
2258 * actual deadlock possible. There are two solutions, using a
2259 * read-recursive acquire on the work(queue) 'locks', but this will then
2260 * hit the lockdep limitation on recursive locks, or simply discard
2261 * these locks.
2262 *
2263 * AFAICT there is no possible deadlock scenario between the
2264 * flush_work() and complete() primitives (except for single-threaded
2265 * workqueues), so hiding them isn't a problem.
2266 */
2267 lockdep_invariant_state(true);
2268 trace_workqueue_execute_start(work);
2269 worker->current_func(work);
2270 /*
2271 * While we must be careful to not use "work" after this, the trace
2272 * point will only record its address.
2273 */
2274 trace_workqueue_execute_end(work, worker->current_func);
2275 lock_map_release(&lockdep_map);
2276 lock_map_release(&pwq->wq->lockdep_map);
2277
2278 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2279 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2280 " last function: %ps\n",
2281 current->comm, preempt_count(), task_pid_nr(current),
2282 worker->current_func);
2283 debug_show_held_locks(current);
2284 dump_stack();
2285 }
2286
2287 /*
2288 * The following prevents a kworker from hogging CPU on !PREEMPTION
2289 * kernels, where a requeueing work item waiting for something to
2290 * happen could deadlock with stop_machine as such work item could
2291 * indefinitely requeue itself while all other CPUs are trapped in
2292 * stop_machine. At the same time, report a quiescent RCU state so
2293 * the same condition doesn't freeze RCU.
2294 */
2295 cond_resched();
2296
2297 raw_spin_lock_irq(&pool->lock);
2298
2299 /* clear cpu intensive status */
2300 if (unlikely(cpu_intensive))
2301 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2302
2303 /* tag the worker for identification in schedule() */
2304 worker->last_func = worker->current_func;
2305
2306 /* we're done with it, release */
2307 hash_del(&worker->hentry);
2308 worker->current_work = NULL;
2309 worker->current_func = NULL;
2310 worker->current_pwq = NULL;
2311 pwq_dec_nr_in_flight(pwq, work_color);
2312}
2313
2314/**
2315 * process_scheduled_works - process scheduled works
2316 * @worker: self
2317 *
2318 * Process all scheduled works. Please note that the scheduled list
2319 * may change while processing a work, so this function repeatedly
2320 * fetches a work from the top and executes it.
2321 *
2322 * CONTEXT:
2323 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2324 * multiple times.
2325 */
2326static void process_scheduled_works(struct worker *worker)
2327{
2328 while (!list_empty(&worker->scheduled)) {
2329 struct work_struct *work = list_first_entry(&worker->scheduled,
2330 struct work_struct, entry);
2331 process_one_work(worker, work);
2332 }
2333}
2334
2335static void set_pf_worker(bool val)
2336{
2337 mutex_lock(&wq_pool_attach_mutex);
2338 if (val)
2339 current->flags |= PF_WQ_WORKER;
2340 else
2341 current->flags &= ~PF_WQ_WORKER;
2342 mutex_unlock(&wq_pool_attach_mutex);
2343}
2344
2345/**
2346 * worker_thread - the worker thread function
2347 * @__worker: self
2348 *
2349 * The worker thread function. All workers belong to a worker_pool -
2350 * either a per-cpu one or dynamic unbound one. These workers process all
2351 * work items regardless of their specific target workqueue. The only
2352 * exception is work items which belong to workqueues with a rescuer which
2353 * will be explained in rescuer_thread().
2354 *
2355 * Return: 0
2356 */
2357static int worker_thread(void *__worker)
2358{
2359 struct worker *worker = __worker;
2360 struct worker_pool *pool = worker->pool;
2361
2362 /* tell the scheduler that this is a workqueue worker */
2363 set_pf_worker(true);
2364woke_up:
2365 raw_spin_lock_irq(&pool->lock);
2366
2367 /* am I supposed to die? */
2368 if (unlikely(worker->flags & WORKER_DIE)) {
2369 raw_spin_unlock_irq(&pool->lock);
2370 WARN_ON_ONCE(!list_empty(&worker->entry));
2371 set_pf_worker(false);
2372
2373 set_task_comm(worker->task, "kworker/dying");
2374 ida_simple_remove(&pool->worker_ida, worker->id);
2375 worker_detach_from_pool(worker);
2376 kfree(worker);
2377 return 0;
2378 }
2379
2380 worker_leave_idle(worker);
2381recheck:
2382 /* no more worker necessary? */
2383 if (!need_more_worker(pool))
2384 goto sleep;
2385
2386 /* do we need to manage? */
2387 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2388 goto recheck;
2389
2390 /*
2391 * ->scheduled list can only be filled while a worker is
2392 * preparing to process a work or actually processing it.
2393 * Make sure nobody diddled with it while I was sleeping.
2394 */
2395 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2396
2397 /*
2398 * Finish PREP stage. We're guaranteed to have at least one idle
2399 * worker or that someone else has already assumed the manager
2400 * role. This is where @worker starts participating in concurrency
2401 * management if applicable and concurrency management is restored
2402 * after being rebound. See rebind_workers() for details.
2403 */
2404 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2405
2406 do {
2407 struct work_struct *work =
2408 list_first_entry(&pool->worklist,
2409 struct work_struct, entry);
2410
2411 pool->watchdog_ts = jiffies;
2412
2413 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2414 /* optimization path, not strictly necessary */
2415 process_one_work(worker, work);
2416 if (unlikely(!list_empty(&worker->scheduled)))
2417 process_scheduled_works(worker);
2418 } else {
2419 move_linked_works(work, &worker->scheduled, NULL);
2420 process_scheduled_works(worker);
2421 }
2422 } while (keep_working(pool));
2423
2424 worker_set_flags(worker, WORKER_PREP);
2425sleep:
2426 /*
2427 * pool->lock is held and there's no work to process and no need to
2428 * manage, sleep. Workers are woken up only while holding
2429 * pool->lock or from local cpu, so setting the current state
2430 * before releasing pool->lock is enough to prevent losing any
2431 * event.
2432 */
2433 worker_enter_idle(worker);
2434 __set_current_state(TASK_IDLE);
2435 raw_spin_unlock_irq(&pool->lock);
2436 schedule();
2437 goto woke_up;
2438}
2439
2440/**
2441 * rescuer_thread - the rescuer thread function
2442 * @__rescuer: self
2443 *
2444 * Workqueue rescuer thread function. There's one rescuer for each
2445 * workqueue which has WQ_MEM_RECLAIM set.
2446 *
2447 * Regular work processing on a pool may block trying to create a new
2448 * worker which uses GFP_KERNEL allocation which has slight chance of
2449 * developing into deadlock if some works currently on the same queue
2450 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2451 * the problem rescuer solves.
2452 *
2453 * When such condition is possible, the pool summons rescuers of all
2454 * workqueues which have works queued on the pool and let them process
2455 * those works so that forward progress can be guaranteed.
2456 *
2457 * This should happen rarely.
2458 *
2459 * Return: 0
2460 */
2461static int rescuer_thread(void *__rescuer)
2462{
2463 struct worker *rescuer = __rescuer;
2464 struct workqueue_struct *wq = rescuer->rescue_wq;
2465 struct list_head *scheduled = &rescuer->scheduled;
2466 bool should_stop;
2467
2468 set_user_nice(current, RESCUER_NICE_LEVEL);
2469
2470 /*
2471 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2472 * doesn't participate in concurrency management.
2473 */
2474 set_pf_worker(true);
2475repeat:
2476 set_current_state(TASK_IDLE);
2477
2478 /*
2479 * By the time the rescuer is requested to stop, the workqueue
2480 * shouldn't have any work pending, but @wq->maydays may still have
2481 * pwq(s) queued. This can happen by non-rescuer workers consuming
2482 * all the work items before the rescuer got to them. Go through
2483 * @wq->maydays processing before acting on should_stop so that the
2484 * list is always empty on exit.
2485 */
2486 should_stop = kthread_should_stop();
2487
2488 /* see whether any pwq is asking for help */
2489 raw_spin_lock_irq(&wq_mayday_lock);
2490
2491 while (!list_empty(&wq->maydays)) {
2492 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2493 struct pool_workqueue, mayday_node);
2494 struct worker_pool *pool = pwq->pool;
2495 struct work_struct *work, *n;
2496 bool first = true;
2497
2498 __set_current_state(TASK_RUNNING);
2499 list_del_init(&pwq->mayday_node);
2500
2501 raw_spin_unlock_irq(&wq_mayday_lock);
2502
2503 worker_attach_to_pool(rescuer, pool);
2504
2505 raw_spin_lock_irq(&pool->lock);
2506
2507 /*
2508 * Slurp in all works issued via this workqueue and
2509 * process'em.
2510 */
2511 WARN_ON_ONCE(!list_empty(scheduled));
2512 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2513 if (get_work_pwq(work) == pwq) {
2514 if (first)
2515 pool->watchdog_ts = jiffies;
2516 move_linked_works(work, scheduled, &n);
2517 }
2518 first = false;
2519 }
2520
2521 if (!list_empty(scheduled)) {
2522 process_scheduled_works(rescuer);
2523
2524 /*
2525 * The above execution of rescued work items could
2526 * have created more to rescue through
2527 * pwq_activate_first_delayed() or chained
2528 * queueing. Let's put @pwq back on mayday list so
2529 * that such back-to-back work items, which may be
2530 * being used to relieve memory pressure, don't
2531 * incur MAYDAY_INTERVAL delay inbetween.
2532 */
2533 if (pwq->nr_active && need_to_create_worker(pool)) {
2534 raw_spin_lock(&wq_mayday_lock);
2535 /*
2536 * Queue iff we aren't racing destruction
2537 * and somebody else hasn't queued it already.
2538 */
2539 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2540 get_pwq(pwq);
2541 list_add_tail(&pwq->mayday_node, &wq->maydays);
2542 }
2543 raw_spin_unlock(&wq_mayday_lock);
2544 }
2545 }
2546
2547 /*
2548 * Put the reference grabbed by send_mayday(). @pool won't
2549 * go away while we're still attached to it.
2550 */
2551 put_pwq(pwq);
2552
2553 /*
2554 * Leave this pool. If need_more_worker() is %true, notify a
2555 * regular worker; otherwise, we end up with 0 concurrency
2556 * and stalling the execution.
2557 */
2558 if (need_more_worker(pool))
2559 wake_up_worker(pool);
2560
2561 raw_spin_unlock_irq(&pool->lock);
2562
2563 worker_detach_from_pool(rescuer);
2564
2565 raw_spin_lock_irq(&wq_mayday_lock);
2566 }
2567
2568 raw_spin_unlock_irq(&wq_mayday_lock);
2569
2570 if (should_stop) {
2571 __set_current_state(TASK_RUNNING);
2572 set_pf_worker(false);
2573 return 0;
2574 }
2575
2576 /* rescuers should never participate in concurrency management */
2577 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2578 schedule();
2579 goto repeat;
2580}
2581
2582/**
2583 * check_flush_dependency - check for flush dependency sanity
2584 * @target_wq: workqueue being flushed
2585 * @target_work: work item being flushed (NULL for workqueue flushes)
2586 *
2587 * %current is trying to flush the whole @target_wq or @target_work on it.
2588 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2589 * reclaiming memory or running on a workqueue which doesn't have
2590 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2591 * a deadlock.
2592 */
2593static void check_flush_dependency(struct workqueue_struct *target_wq,
2594 struct work_struct *target_work)
2595{
2596 work_func_t target_func = target_work ? target_work->func : NULL;
2597 struct worker *worker;
2598
2599 if (target_wq->flags & WQ_MEM_RECLAIM)
2600 return;
2601
2602 worker = current_wq_worker();
2603
2604 WARN_ONCE(current->flags & PF_MEMALLOC,
2605 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2606 current->pid, current->comm, target_wq->name, target_func);
2607 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2608 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2609 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2610 worker->current_pwq->wq->name, worker->current_func,
2611 target_wq->name, target_func);
2612}
2613
2614struct wq_barrier {
2615 struct work_struct work;
2616 struct completion done;
2617 struct task_struct *task; /* purely informational */
2618};
2619
2620static void wq_barrier_func(struct work_struct *work)
2621{
2622 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2623 complete(&barr->done);
2624}
2625
2626/**
2627 * insert_wq_barrier - insert a barrier work
2628 * @pwq: pwq to insert barrier into
2629 * @barr: wq_barrier to insert
2630 * @target: target work to attach @barr to
2631 * @worker: worker currently executing @target, NULL if @target is not executing
2632 *
2633 * @barr is linked to @target such that @barr is completed only after
2634 * @target finishes execution. Please note that the ordering
2635 * guarantee is observed only with respect to @target and on the local
2636 * cpu.
2637 *
2638 * Currently, a queued barrier can't be canceled. This is because
2639 * try_to_grab_pending() can't determine whether the work to be
2640 * grabbed is at the head of the queue and thus can't clear LINKED
2641 * flag of the previous work while there must be a valid next work
2642 * after a work with LINKED flag set.
2643 *
2644 * Note that when @worker is non-NULL, @target may be modified
2645 * underneath us, so we can't reliably determine pwq from @target.
2646 *
2647 * CONTEXT:
2648 * raw_spin_lock_irq(pool->lock).
2649 */
2650static void insert_wq_barrier(struct pool_workqueue *pwq,
2651 struct wq_barrier *barr,
2652 struct work_struct *target, struct worker *worker)
2653{
2654 struct list_head *head;
2655 unsigned int linked = 0;
2656
2657 /*
2658 * debugobject calls are safe here even with pool->lock locked
2659 * as we know for sure that this will not trigger any of the
2660 * checks and call back into the fixup functions where we
2661 * might deadlock.
2662 */
2663 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2664 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2665
2666 init_completion_map(&barr->done, &target->lockdep_map);
2667
2668 barr->task = current;
2669
2670 /*
2671 * If @target is currently being executed, schedule the
2672 * barrier to the worker; otherwise, put it after @target.
2673 */
2674 if (worker)
2675 head = worker->scheduled.next;
2676 else {
2677 unsigned long *bits = work_data_bits(target);
2678
2679 head = target->entry.next;
2680 /* there can already be other linked works, inherit and set */
2681 linked = *bits & WORK_STRUCT_LINKED;
2682 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2683 }
2684
2685 debug_work_activate(&barr->work);
2686 insert_work(pwq, &barr->work, head,
2687 work_color_to_flags(WORK_NO_COLOR) | linked);
2688}
2689
2690/**
2691 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2692 * @wq: workqueue being flushed
2693 * @flush_color: new flush color, < 0 for no-op
2694 * @work_color: new work color, < 0 for no-op
2695 *
2696 * Prepare pwqs for workqueue flushing.
2697 *
2698 * If @flush_color is non-negative, flush_color on all pwqs should be
2699 * -1. If no pwq has in-flight commands at the specified color, all
2700 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2701 * has in flight commands, its pwq->flush_color is set to
2702 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2703 * wakeup logic is armed and %true is returned.
2704 *
2705 * The caller should have initialized @wq->first_flusher prior to
2706 * calling this function with non-negative @flush_color. If
2707 * @flush_color is negative, no flush color update is done and %false
2708 * is returned.
2709 *
2710 * If @work_color is non-negative, all pwqs should have the same
2711 * work_color which is previous to @work_color and all will be
2712 * advanced to @work_color.
2713 *
2714 * CONTEXT:
2715 * mutex_lock(wq->mutex).
2716 *
2717 * Return:
2718 * %true if @flush_color >= 0 and there's something to flush. %false
2719 * otherwise.
2720 */
2721static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2722 int flush_color, int work_color)
2723{
2724 bool wait = false;
2725 struct pool_workqueue *pwq;
2726
2727 if (flush_color >= 0) {
2728 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2729 atomic_set(&wq->nr_pwqs_to_flush, 1);
2730 }
2731
2732 for_each_pwq(pwq, wq) {
2733 struct worker_pool *pool = pwq->pool;
2734
2735 raw_spin_lock_irq(&pool->lock);
2736
2737 if (flush_color >= 0) {
2738 WARN_ON_ONCE(pwq->flush_color != -1);
2739
2740 if (pwq->nr_in_flight[flush_color]) {
2741 pwq->flush_color = flush_color;
2742 atomic_inc(&wq->nr_pwqs_to_flush);
2743 wait = true;
2744 }
2745 }
2746
2747 if (work_color >= 0) {
2748 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2749 pwq->work_color = work_color;
2750 }
2751
2752 raw_spin_unlock_irq(&pool->lock);
2753 }
2754
2755 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2756 complete(&wq->first_flusher->done);
2757
2758 return wait;
2759}
2760
2761/**
2762 * flush_workqueue - ensure that any scheduled work has run to completion.
2763 * @wq: workqueue to flush
2764 *
2765 * This function sleeps until all work items which were queued on entry
2766 * have finished execution, but it is not livelocked by new incoming ones.
2767 */
2768void flush_workqueue(struct workqueue_struct *wq)
2769{
2770 struct wq_flusher this_flusher = {
2771 .list = LIST_HEAD_INIT(this_flusher.list),
2772 .flush_color = -1,
2773 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2774 };
2775 int next_color;
2776
2777 if (WARN_ON(!wq_online))
2778 return;
2779
2780 lock_map_acquire(&wq->lockdep_map);
2781 lock_map_release(&wq->lockdep_map);
2782
2783 mutex_lock(&wq->mutex);
2784
2785 /*
2786 * Start-to-wait phase
2787 */
2788 next_color = work_next_color(wq->work_color);
2789
2790 if (next_color != wq->flush_color) {
2791 /*
2792 * Color space is not full. The current work_color
2793 * becomes our flush_color and work_color is advanced
2794 * by one.
2795 */
2796 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2797 this_flusher.flush_color = wq->work_color;
2798 wq->work_color = next_color;
2799
2800 if (!wq->first_flusher) {
2801 /* no flush in progress, become the first flusher */
2802 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2803
2804 wq->first_flusher = &this_flusher;
2805
2806 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2807 wq->work_color)) {
2808 /* nothing to flush, done */
2809 wq->flush_color = next_color;
2810 wq->first_flusher = NULL;
2811 goto out_unlock;
2812 }
2813 } else {
2814 /* wait in queue */
2815 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2816 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2817 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2818 }
2819 } else {
2820 /*
2821 * Oops, color space is full, wait on overflow queue.
2822 * The next flush completion will assign us
2823 * flush_color and transfer to flusher_queue.
2824 */
2825 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2826 }
2827
2828 check_flush_dependency(wq, NULL);
2829
2830 mutex_unlock(&wq->mutex);
2831
2832 wait_for_completion(&this_flusher.done);
2833
2834 /*
2835 * Wake-up-and-cascade phase
2836 *
2837 * First flushers are responsible for cascading flushes and
2838 * handling overflow. Non-first flushers can simply return.
2839 */
2840 if (READ_ONCE(wq->first_flusher) != &this_flusher)
2841 return;
2842
2843 mutex_lock(&wq->mutex);
2844
2845 /* we might have raced, check again with mutex held */
2846 if (wq->first_flusher != &this_flusher)
2847 goto out_unlock;
2848
2849 WRITE_ONCE(wq->first_flusher, NULL);
2850
2851 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2852 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2853
2854 while (true) {
2855 struct wq_flusher *next, *tmp;
2856
2857 /* complete all the flushers sharing the current flush color */
2858 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2859 if (next->flush_color != wq->flush_color)
2860 break;
2861 list_del_init(&next->list);
2862 complete(&next->done);
2863 }
2864
2865 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2866 wq->flush_color != work_next_color(wq->work_color));
2867
2868 /* this flush_color is finished, advance by one */
2869 wq->flush_color = work_next_color(wq->flush_color);
2870
2871 /* one color has been freed, handle overflow queue */
2872 if (!list_empty(&wq->flusher_overflow)) {
2873 /*
2874 * Assign the same color to all overflowed
2875 * flushers, advance work_color and append to
2876 * flusher_queue. This is the start-to-wait
2877 * phase for these overflowed flushers.
2878 */
2879 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2880 tmp->flush_color = wq->work_color;
2881
2882 wq->work_color = work_next_color(wq->work_color);
2883
2884 list_splice_tail_init(&wq->flusher_overflow,
2885 &wq->flusher_queue);
2886 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2887 }
2888
2889 if (list_empty(&wq->flusher_queue)) {
2890 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2891 break;
2892 }
2893
2894 /*
2895 * Need to flush more colors. Make the next flusher
2896 * the new first flusher and arm pwqs.
2897 */
2898 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2899 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2900
2901 list_del_init(&next->list);
2902 wq->first_flusher = next;
2903
2904 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2905 break;
2906
2907 /*
2908 * Meh... this color is already done, clear first
2909 * flusher and repeat cascading.
2910 */
2911 wq->first_flusher = NULL;
2912 }
2913
2914out_unlock:
2915 mutex_unlock(&wq->mutex);
2916}
2917EXPORT_SYMBOL(flush_workqueue);
2918
2919/**
2920 * drain_workqueue - drain a workqueue
2921 * @wq: workqueue to drain
2922 *
2923 * Wait until the workqueue becomes empty. While draining is in progress,
2924 * only chain queueing is allowed. IOW, only currently pending or running
2925 * work items on @wq can queue further work items on it. @wq is flushed
2926 * repeatedly until it becomes empty. The number of flushing is determined
2927 * by the depth of chaining and should be relatively short. Whine if it
2928 * takes too long.
2929 */
2930void drain_workqueue(struct workqueue_struct *wq)
2931{
2932 unsigned int flush_cnt = 0;
2933 struct pool_workqueue *pwq;
2934
2935 /*
2936 * __queue_work() needs to test whether there are drainers, is much
2937 * hotter than drain_workqueue() and already looks at @wq->flags.
2938 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2939 */
2940 mutex_lock(&wq->mutex);
2941 if (!wq->nr_drainers++)
2942 wq->flags |= __WQ_DRAINING;
2943 mutex_unlock(&wq->mutex);
2944reflush:
2945 flush_workqueue(wq);
2946
2947 mutex_lock(&wq->mutex);
2948
2949 for_each_pwq(pwq, wq) {
2950 bool drained;
2951
2952 raw_spin_lock_irq(&pwq->pool->lock);
2953 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2954 raw_spin_unlock_irq(&pwq->pool->lock);
2955
2956 if (drained)
2957 continue;
2958
2959 if (++flush_cnt == 10 ||
2960 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2961 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2962 wq->name, flush_cnt);
2963
2964 mutex_unlock(&wq->mutex);
2965 goto reflush;
2966 }
2967
2968 if (!--wq->nr_drainers)
2969 wq->flags &= ~__WQ_DRAINING;
2970 mutex_unlock(&wq->mutex);
2971}
2972EXPORT_SYMBOL_GPL(drain_workqueue);
2973
2974static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2975 bool from_cancel)
2976{
2977 struct worker *worker = NULL;
2978 struct worker_pool *pool;
2979 struct pool_workqueue *pwq;
2980
2981 might_sleep();
2982
2983 rcu_read_lock();
2984 pool = get_work_pool(work);
2985 if (!pool) {
2986 rcu_read_unlock();
2987 return false;
2988 }
2989
2990 raw_spin_lock_irq(&pool->lock);
2991 /* see the comment in try_to_grab_pending() with the same code */
2992 pwq = get_work_pwq(work);
2993 if (pwq) {
2994 if (unlikely(pwq->pool != pool))
2995 goto already_gone;
2996 } else {
2997 worker = find_worker_executing_work(pool, work);
2998 if (!worker)
2999 goto already_gone;
3000 pwq = worker->current_pwq;
3001 }
3002
3003 check_flush_dependency(pwq->wq, work);
3004
3005 insert_wq_barrier(pwq, barr, work, worker);
3006 raw_spin_unlock_irq(&pool->lock);
3007
3008 /*
3009 * Force a lock recursion deadlock when using flush_work() inside a
3010 * single-threaded or rescuer equipped workqueue.
3011 *
3012 * For single threaded workqueues the deadlock happens when the work
3013 * is after the work issuing the flush_work(). For rescuer equipped
3014 * workqueues the deadlock happens when the rescuer stalls, blocking
3015 * forward progress.
3016 */
3017 if (!from_cancel &&
3018 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3019 lock_map_acquire(&pwq->wq->lockdep_map);
3020 lock_map_release(&pwq->wq->lockdep_map);
3021 }
3022 rcu_read_unlock();
3023 return true;
3024already_gone:
3025 raw_spin_unlock_irq(&pool->lock);
3026 rcu_read_unlock();
3027 return false;
3028}
3029
3030static bool __flush_work(struct work_struct *work, bool from_cancel)
3031{
3032 struct wq_barrier barr;
3033
3034 if (WARN_ON(!wq_online))
3035 return false;
3036
3037 if (WARN_ON(!work->func))
3038 return false;
3039
3040 if (!from_cancel) {
3041 lock_map_acquire(&work->lockdep_map);
3042 lock_map_release(&work->lockdep_map);
3043 }
3044
3045 if (start_flush_work(work, &barr, from_cancel)) {
3046 wait_for_completion(&barr.done);
3047 destroy_work_on_stack(&barr.work);
3048 return true;
3049 } else {
3050 return false;
3051 }
3052}
3053
3054/**
3055 * flush_work - wait for a work to finish executing the last queueing instance
3056 * @work: the work to flush
3057 *
3058 * Wait until @work has finished execution. @work is guaranteed to be idle
3059 * on return if it hasn't been requeued since flush started.
3060 *
3061 * Return:
3062 * %true if flush_work() waited for the work to finish execution,
3063 * %false if it was already idle.
3064 */
3065bool flush_work(struct work_struct *work)
3066{
3067 return __flush_work(work, false);
3068}
3069EXPORT_SYMBOL_GPL(flush_work);
3070
3071struct cwt_wait {
3072 wait_queue_entry_t wait;
3073 struct work_struct *work;
3074};
3075
3076static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3077{
3078 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3079
3080 if (cwait->work != key)
3081 return 0;
3082 return autoremove_wake_function(wait, mode, sync, key);
3083}
3084
3085static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3086{
3087 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3088 unsigned long flags;
3089 int ret;
3090
3091 do {
3092 ret = try_to_grab_pending(work, is_dwork, &flags);
3093 /*
3094 * If someone else is already canceling, wait for it to
3095 * finish. flush_work() doesn't work for PREEMPT_NONE
3096 * because we may get scheduled between @work's completion
3097 * and the other canceling task resuming and clearing
3098 * CANCELING - flush_work() will return false immediately
3099 * as @work is no longer busy, try_to_grab_pending() will
3100 * return -ENOENT as @work is still being canceled and the
3101 * other canceling task won't be able to clear CANCELING as
3102 * we're hogging the CPU.
3103 *
3104 * Let's wait for completion using a waitqueue. As this
3105 * may lead to the thundering herd problem, use a custom
3106 * wake function which matches @work along with exclusive
3107 * wait and wakeup.
3108 */
3109 if (unlikely(ret == -ENOENT)) {
3110 struct cwt_wait cwait;
3111
3112 init_wait(&cwait.wait);
3113 cwait.wait.func = cwt_wakefn;
3114 cwait.work = work;
3115
3116 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3117 TASK_UNINTERRUPTIBLE);
3118 if (work_is_canceling(work))
3119 schedule();
3120 finish_wait(&cancel_waitq, &cwait.wait);
3121 }
3122 } while (unlikely(ret < 0));
3123
3124 /* tell other tasks trying to grab @work to back off */
3125 mark_work_canceling(work);
3126 local_irq_restore(flags);
3127
3128 /*
3129 * This allows canceling during early boot. We know that @work
3130 * isn't executing.
3131 */
3132 if (wq_online)
3133 __flush_work(work, true);
3134
3135 clear_work_data(work);
3136
3137 /*
3138 * Paired with prepare_to_wait() above so that either
3139 * waitqueue_active() is visible here or !work_is_canceling() is
3140 * visible there.
3141 */
3142 smp_mb();
3143 if (waitqueue_active(&cancel_waitq))
3144 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3145
3146 return ret;
3147}
3148
3149/**
3150 * cancel_work_sync - cancel a work and wait for it to finish
3151 * @work: the work to cancel
3152 *
3153 * Cancel @work and wait for its execution to finish. This function
3154 * can be used even if the work re-queues itself or migrates to
3155 * another workqueue. On return from this function, @work is
3156 * guaranteed to be not pending or executing on any CPU.
3157 *
3158 * cancel_work_sync(&delayed_work->work) must not be used for
3159 * delayed_work's. Use cancel_delayed_work_sync() instead.
3160 *
3161 * The caller must ensure that the workqueue on which @work was last
3162 * queued can't be destroyed before this function returns.
3163 *
3164 * Return:
3165 * %true if @work was pending, %false otherwise.
3166 */
3167bool cancel_work_sync(struct work_struct *work)
3168{
3169 return __cancel_work_timer(work, false);
3170}
3171EXPORT_SYMBOL_GPL(cancel_work_sync);
3172
3173/**
3174 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3175 * @dwork: the delayed work to flush
3176 *
3177 * Delayed timer is cancelled and the pending work is queued for
3178 * immediate execution. Like flush_work(), this function only
3179 * considers the last queueing instance of @dwork.
3180 *
3181 * Return:
3182 * %true if flush_work() waited for the work to finish execution,
3183 * %false if it was already idle.
3184 */
3185bool flush_delayed_work(struct delayed_work *dwork)
3186{
3187 local_irq_disable();
3188 if (del_timer_sync(&dwork->timer))
3189 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3190 local_irq_enable();
3191 return flush_work(&dwork->work);
3192}
3193EXPORT_SYMBOL(flush_delayed_work);
3194
3195/**
3196 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3197 * @rwork: the rcu work to flush
3198 *
3199 * Return:
3200 * %true if flush_rcu_work() waited for the work to finish execution,
3201 * %false if it was already idle.
3202 */
3203bool flush_rcu_work(struct rcu_work *rwork)
3204{
3205 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3206 rcu_barrier();
3207 flush_work(&rwork->work);
3208 return true;
3209 } else {
3210 return flush_work(&rwork->work);
3211 }
3212}
3213EXPORT_SYMBOL(flush_rcu_work);
3214
3215static bool __cancel_work(struct work_struct *work, bool is_dwork)
3216{
3217 unsigned long flags;
3218 int ret;
3219
3220 do {
3221 ret = try_to_grab_pending(work, is_dwork, &flags);
3222 } while (unlikely(ret == -EAGAIN));
3223
3224 if (unlikely(ret < 0))
3225 return false;
3226
3227 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3228 local_irq_restore(flags);
3229 return ret;
3230}
3231
3232/**
3233 * cancel_delayed_work - cancel a delayed work
3234 * @dwork: delayed_work to cancel
3235 *
3236 * Kill off a pending delayed_work.
3237 *
3238 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3239 * pending.
3240 *
3241 * Note:
3242 * The work callback function may still be running on return, unless
3243 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3244 * use cancel_delayed_work_sync() to wait on it.
3245 *
3246 * This function is safe to call from any context including IRQ handler.
3247 */
3248bool cancel_delayed_work(struct delayed_work *dwork)
3249{
3250 return __cancel_work(&dwork->work, true);
3251}
3252EXPORT_SYMBOL(cancel_delayed_work);
3253
3254/**
3255 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3256 * @dwork: the delayed work cancel
3257 *
3258 * This is cancel_work_sync() for delayed works.
3259 *
3260 * Return:
3261 * %true if @dwork was pending, %false otherwise.
3262 */
3263bool cancel_delayed_work_sync(struct delayed_work *dwork)
3264{
3265 return __cancel_work_timer(&dwork->work, true);
3266}
3267EXPORT_SYMBOL(cancel_delayed_work_sync);
3268
3269/**
3270 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3271 * @func: the function to call
3272 *
3273 * schedule_on_each_cpu() executes @func on each online CPU using the
3274 * system workqueue and blocks until all CPUs have completed.
3275 * schedule_on_each_cpu() is very slow.
3276 *
3277 * Return:
3278 * 0 on success, -errno on failure.
3279 */
3280int schedule_on_each_cpu(work_func_t func)
3281{
3282 int cpu;
3283 struct work_struct __percpu *works;
3284
3285 works = alloc_percpu(struct work_struct);
3286 if (!works)
3287 return -ENOMEM;
3288
3289 get_online_cpus();
3290
3291 for_each_online_cpu(cpu) {
3292 struct work_struct *work = per_cpu_ptr(works, cpu);
3293
3294 INIT_WORK(work, func);
3295 schedule_work_on(cpu, work);
3296 }
3297
3298 for_each_online_cpu(cpu)
3299 flush_work(per_cpu_ptr(works, cpu));
3300
3301 put_online_cpus();
3302 free_percpu(works);
3303 return 0;
3304}
3305
3306/**
3307 * execute_in_process_context - reliably execute the routine with user context
3308 * @fn: the function to execute
3309 * @ew: guaranteed storage for the execute work structure (must
3310 * be available when the work executes)
3311 *
3312 * Executes the function immediately if process context is available,
3313 * otherwise schedules the function for delayed execution.
3314 *
3315 * Return: 0 - function was executed
3316 * 1 - function was scheduled for execution
3317 */
3318int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3319{
3320 if (!in_interrupt()) {
3321 fn(&ew->work);
3322 return 0;
3323 }
3324
3325 INIT_WORK(&ew->work, fn);
3326 schedule_work(&ew->work);
3327
3328 return 1;
3329}
3330EXPORT_SYMBOL_GPL(execute_in_process_context);
3331
3332/**
3333 * free_workqueue_attrs - free a workqueue_attrs
3334 * @attrs: workqueue_attrs to free
3335 *
3336 * Undo alloc_workqueue_attrs().
3337 */
3338void free_workqueue_attrs(struct workqueue_attrs *attrs)
3339{
3340 if (attrs) {
3341 free_cpumask_var(attrs->cpumask);
3342 kfree(attrs);
3343 }
3344}
3345
3346/**
3347 * alloc_workqueue_attrs - allocate a workqueue_attrs
3348 *
3349 * Allocate a new workqueue_attrs, initialize with default settings and
3350 * return it.
3351 *
3352 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3353 */
3354struct workqueue_attrs *alloc_workqueue_attrs(void)
3355{
3356 struct workqueue_attrs *attrs;
3357
3358 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3359 if (!attrs)
3360 goto fail;
3361 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3362 goto fail;
3363
3364 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3365 return attrs;
3366fail:
3367 free_workqueue_attrs(attrs);
3368 return NULL;
3369}
3370
3371static void copy_workqueue_attrs(struct workqueue_attrs *to,
3372 const struct workqueue_attrs *from)
3373{
3374 to->nice = from->nice;
3375 cpumask_copy(to->cpumask, from->cpumask);
3376 /*
3377 * Unlike hash and equality test, this function doesn't ignore
3378 * ->no_numa as it is used for both pool and wq attrs. Instead,
3379 * get_unbound_pool() explicitly clears ->no_numa after copying.
3380 */
3381 to->no_numa = from->no_numa;
3382}
3383
3384/* hash value of the content of @attr */
3385static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3386{
3387 u32 hash = 0;
3388
3389 hash = jhash_1word(attrs->nice, hash);
3390 hash = jhash(cpumask_bits(attrs->cpumask),
3391 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3392 return hash;
3393}
3394
3395/* content equality test */
3396static bool wqattrs_equal(const struct workqueue_attrs *a,
3397 const struct workqueue_attrs *b)
3398{
3399 if (a->nice != b->nice)
3400 return false;
3401 if (!cpumask_equal(a->cpumask, b->cpumask))
3402 return false;
3403 return true;
3404}
3405
3406/**
3407 * init_worker_pool - initialize a newly zalloc'd worker_pool
3408 * @pool: worker_pool to initialize
3409 *
3410 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3411 *
3412 * Return: 0 on success, -errno on failure. Even on failure, all fields
3413 * inside @pool proper are initialized and put_unbound_pool() can be called
3414 * on @pool safely to release it.
3415 */
3416static int init_worker_pool(struct worker_pool *pool)
3417{
3418 raw_spin_lock_init(&pool->lock);
3419 pool->id = -1;
3420 pool->cpu = -1;
3421 pool->node = NUMA_NO_NODE;
3422 pool->flags |= POOL_DISASSOCIATED;
3423 pool->watchdog_ts = jiffies;
3424 INIT_LIST_HEAD(&pool->worklist);
3425 INIT_LIST_HEAD(&pool->idle_list);
3426 hash_init(pool->busy_hash);
3427
3428 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3429
3430 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3431
3432 INIT_LIST_HEAD(&pool->workers);
3433
3434 ida_init(&pool->worker_ida);
3435 INIT_HLIST_NODE(&pool->hash_node);
3436 pool->refcnt = 1;
3437
3438 /* shouldn't fail above this point */
3439 pool->attrs = alloc_workqueue_attrs();
3440 if (!pool->attrs)
3441 return -ENOMEM;
3442 return 0;
3443}
3444
3445#ifdef CONFIG_LOCKDEP
3446static void wq_init_lockdep(struct workqueue_struct *wq)
3447{
3448 char *lock_name;
3449
3450 lockdep_register_key(&wq->key);
3451 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3452 if (!lock_name)
3453 lock_name = wq->name;
3454
3455 wq->lock_name = lock_name;
3456 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3457}
3458
3459static void wq_unregister_lockdep(struct workqueue_struct *wq)
3460{
3461 lockdep_unregister_key(&wq->key);
3462}
3463
3464static void wq_free_lockdep(struct workqueue_struct *wq)
3465{
3466 if (wq->lock_name != wq->name)
3467 kfree(wq->lock_name);
3468}
3469#else
3470static void wq_init_lockdep(struct workqueue_struct *wq)
3471{
3472}
3473
3474static void wq_unregister_lockdep(struct workqueue_struct *wq)
3475{
3476}
3477
3478static void wq_free_lockdep(struct workqueue_struct *wq)
3479{
3480}
3481#endif
3482
3483static void rcu_free_wq(struct rcu_head *rcu)
3484{
3485 struct workqueue_struct *wq =
3486 container_of(rcu, struct workqueue_struct, rcu);
3487
3488 wq_free_lockdep(wq);
3489
3490 if (!(wq->flags & WQ_UNBOUND))
3491 free_percpu(wq->cpu_pwqs);
3492 else
3493 free_workqueue_attrs(wq->unbound_attrs);
3494
3495 kfree(wq);
3496}
3497
3498static void rcu_free_pool(struct rcu_head *rcu)
3499{
3500 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3501
3502 ida_destroy(&pool->worker_ida);
3503 free_workqueue_attrs(pool->attrs);
3504 kfree(pool);
3505}
3506
3507/* This returns with the lock held on success (pool manager is inactive). */
3508static bool wq_manager_inactive(struct worker_pool *pool)
3509{
3510 raw_spin_lock_irq(&pool->lock);
3511
3512 if (pool->flags & POOL_MANAGER_ACTIVE) {
3513 raw_spin_unlock_irq(&pool->lock);
3514 return false;
3515 }
3516 return true;
3517}
3518
3519/**
3520 * put_unbound_pool - put a worker_pool
3521 * @pool: worker_pool to put
3522 *
3523 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3524 * safe manner. get_unbound_pool() calls this function on its failure path
3525 * and this function should be able to release pools which went through,
3526 * successfully or not, init_worker_pool().
3527 *
3528 * Should be called with wq_pool_mutex held.
3529 */
3530static void put_unbound_pool(struct worker_pool *pool)
3531{
3532 DECLARE_COMPLETION_ONSTACK(detach_completion);
3533 struct worker *worker;
3534
3535 lockdep_assert_held(&wq_pool_mutex);
3536
3537 if (--pool->refcnt)
3538 return;
3539
3540 /* sanity checks */
3541 if (WARN_ON(!(pool->cpu < 0)) ||
3542 WARN_ON(!list_empty(&pool->worklist)))
3543 return;
3544
3545 /* release id and unhash */
3546 if (pool->id >= 0)
3547 idr_remove(&worker_pool_idr, pool->id);
3548 hash_del(&pool->hash_node);
3549
3550 /*
3551 * Become the manager and destroy all workers. This prevents
3552 * @pool's workers from blocking on attach_mutex. We're the last
3553 * manager and @pool gets freed with the flag set.
3554 * Because of how wq_manager_inactive() works, we will hold the
3555 * spinlock after a successful wait.
3556 */
3557 rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
3558 TASK_UNINTERRUPTIBLE);
3559 pool->flags |= POOL_MANAGER_ACTIVE;
3560
3561 while ((worker = first_idle_worker(pool)))
3562 destroy_worker(worker);
3563 WARN_ON(pool->nr_workers || pool->nr_idle);
3564 raw_spin_unlock_irq(&pool->lock);
3565
3566 mutex_lock(&wq_pool_attach_mutex);
3567 if (!list_empty(&pool->workers))
3568 pool->detach_completion = &detach_completion;
3569 mutex_unlock(&wq_pool_attach_mutex);
3570
3571 if (pool->detach_completion)
3572 wait_for_completion(pool->detach_completion);
3573
3574 /* shut down the timers */
3575 del_timer_sync(&pool->idle_timer);
3576 del_timer_sync(&pool->mayday_timer);
3577
3578 /* RCU protected to allow dereferences from get_work_pool() */
3579 call_rcu(&pool->rcu, rcu_free_pool);
3580}
3581
3582/**
3583 * get_unbound_pool - get a worker_pool with the specified attributes
3584 * @attrs: the attributes of the worker_pool to get
3585 *
3586 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3587 * reference count and return it. If there already is a matching
3588 * worker_pool, it will be used; otherwise, this function attempts to
3589 * create a new one.
3590 *
3591 * Should be called with wq_pool_mutex held.
3592 *
3593 * Return: On success, a worker_pool with the same attributes as @attrs.
3594 * On failure, %NULL.
3595 */
3596static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3597{
3598 u32 hash = wqattrs_hash(attrs);
3599 struct worker_pool *pool;
3600 int node;
3601 int target_node = NUMA_NO_NODE;
3602
3603 lockdep_assert_held(&wq_pool_mutex);
3604
3605 /* do we already have a matching pool? */
3606 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3607 if (wqattrs_equal(pool->attrs, attrs)) {
3608 pool->refcnt++;
3609 return pool;
3610 }
3611 }
3612
3613 /* if cpumask is contained inside a NUMA node, we belong to that node */
3614 if (wq_numa_enabled) {
3615 for_each_node(node) {
3616 if (cpumask_subset(attrs->cpumask,
3617 wq_numa_possible_cpumask[node])) {
3618 target_node = node;
3619 break;
3620 }
3621 }
3622 }
3623
3624 /* nope, create a new one */
3625 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3626 if (!pool || init_worker_pool(pool) < 0)
3627 goto fail;
3628
3629 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3630 copy_workqueue_attrs(pool->attrs, attrs);
3631 pool->node = target_node;
3632
3633 /*
3634 * no_numa isn't a worker_pool attribute, always clear it. See
3635 * 'struct workqueue_attrs' comments for detail.
3636 */
3637 pool->attrs->no_numa = false;
3638
3639 if (worker_pool_assign_id(pool) < 0)
3640 goto fail;
3641
3642 /* create and start the initial worker */
3643 if (wq_online && !create_worker(pool))
3644 goto fail;
3645
3646 /* install */
3647 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3648
3649 return pool;
3650fail:
3651 if (pool)
3652 put_unbound_pool(pool);
3653 return NULL;
3654}
3655
3656static void rcu_free_pwq(struct rcu_head *rcu)
3657{
3658 kmem_cache_free(pwq_cache,
3659 container_of(rcu, struct pool_workqueue, rcu));
3660}
3661
3662/*
3663 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3664 * and needs to be destroyed.
3665 */
3666static void pwq_unbound_release_workfn(struct work_struct *work)
3667{
3668 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3669 unbound_release_work);
3670 struct workqueue_struct *wq = pwq->wq;
3671 struct worker_pool *pool = pwq->pool;
3672 bool is_last;
3673
3674 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3675 return;
3676
3677 mutex_lock(&wq->mutex);
3678 list_del_rcu(&pwq->pwqs_node);
3679 is_last = list_empty(&wq->pwqs);
3680 mutex_unlock(&wq->mutex);
3681
3682 mutex_lock(&wq_pool_mutex);
3683 put_unbound_pool(pool);
3684 mutex_unlock(&wq_pool_mutex);
3685
3686 call_rcu(&pwq->rcu, rcu_free_pwq);
3687
3688 /*
3689 * If we're the last pwq going away, @wq is already dead and no one
3690 * is gonna access it anymore. Schedule RCU free.
3691 */
3692 if (is_last) {
3693 wq_unregister_lockdep(wq);
3694 call_rcu(&wq->rcu, rcu_free_wq);
3695 }
3696}
3697
3698/**
3699 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3700 * @pwq: target pool_workqueue
3701 *
3702 * If @pwq isn't freezing, set @pwq->max_active to the associated
3703 * workqueue's saved_max_active and activate delayed work items
3704 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3705 */
3706static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3707{
3708 struct workqueue_struct *wq = pwq->wq;
3709 bool freezable = wq->flags & WQ_FREEZABLE;
3710 unsigned long flags;
3711
3712 /* for @wq->saved_max_active */
3713 lockdep_assert_held(&wq->mutex);
3714
3715 /* fast exit for non-freezable wqs */
3716 if (!freezable && pwq->max_active == wq->saved_max_active)
3717 return;
3718
3719 /* this function can be called during early boot w/ irq disabled */
3720 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
3721
3722 /*
3723 * During [un]freezing, the caller is responsible for ensuring that
3724 * this function is called at least once after @workqueue_freezing
3725 * is updated and visible.
3726 */
3727 if (!freezable || !workqueue_freezing) {
3728 pwq->max_active = wq->saved_max_active;
3729
3730 while (!list_empty(&pwq->delayed_works) &&
3731 pwq->nr_active < pwq->max_active)
3732 pwq_activate_first_delayed(pwq);
3733
3734 /*
3735 * Need to kick a worker after thawed or an unbound wq's
3736 * max_active is bumped. It's a slow path. Do it always.
3737 */
3738 wake_up_worker(pwq->pool);
3739 } else {
3740 pwq->max_active = 0;
3741 }
3742
3743 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
3744}
3745
3746/* initialize newly alloced @pwq which is associated with @wq and @pool */
3747static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3748 struct worker_pool *pool)
3749{
3750 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3751
3752 memset(pwq, 0, sizeof(*pwq));
3753
3754 pwq->pool = pool;
3755 pwq->wq = wq;
3756 pwq->flush_color = -1;
3757 pwq->refcnt = 1;
3758 INIT_LIST_HEAD(&pwq->delayed_works);
3759 INIT_LIST_HEAD(&pwq->pwqs_node);
3760 INIT_LIST_HEAD(&pwq->mayday_node);
3761 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3762}
3763
3764/* sync @pwq with the current state of its associated wq and link it */
3765static void link_pwq(struct pool_workqueue *pwq)
3766{
3767 struct workqueue_struct *wq = pwq->wq;
3768
3769 lockdep_assert_held(&wq->mutex);
3770
3771 /* may be called multiple times, ignore if already linked */
3772 if (!list_empty(&pwq->pwqs_node))
3773 return;
3774
3775 /* set the matching work_color */
3776 pwq->work_color = wq->work_color;
3777
3778 /* sync max_active to the current setting */
3779 pwq_adjust_max_active(pwq);
3780
3781 /* link in @pwq */
3782 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3783}
3784
3785/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3786static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3787 const struct workqueue_attrs *attrs)
3788{
3789 struct worker_pool *pool;
3790 struct pool_workqueue *pwq;
3791
3792 lockdep_assert_held(&wq_pool_mutex);
3793
3794 pool = get_unbound_pool(attrs);
3795 if (!pool)
3796 return NULL;
3797
3798 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3799 if (!pwq) {
3800 put_unbound_pool(pool);
3801 return NULL;
3802 }
3803
3804 init_pwq(pwq, wq, pool);
3805 return pwq;
3806}
3807
3808/**
3809 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3810 * @attrs: the wq_attrs of the default pwq of the target workqueue
3811 * @node: the target NUMA node
3812 * @cpu_going_down: if >= 0, the CPU to consider as offline
3813 * @cpumask: outarg, the resulting cpumask
3814 *
3815 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3816 * @cpu_going_down is >= 0, that cpu is considered offline during
3817 * calculation. The result is stored in @cpumask.
3818 *
3819 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3820 * enabled and @node has online CPUs requested by @attrs, the returned
3821 * cpumask is the intersection of the possible CPUs of @node and
3822 * @attrs->cpumask.
3823 *
3824 * The caller is responsible for ensuring that the cpumask of @node stays
3825 * stable.
3826 *
3827 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3828 * %false if equal.
3829 */
3830static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3831 int cpu_going_down, cpumask_t *cpumask)
3832{
3833 if (!wq_numa_enabled || attrs->no_numa)
3834 goto use_dfl;
3835
3836 /* does @node have any online CPUs @attrs wants? */
3837 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3838 if (cpu_going_down >= 0)
3839 cpumask_clear_cpu(cpu_going_down, cpumask);
3840
3841 if (cpumask_empty(cpumask))
3842 goto use_dfl;
3843
3844 /* yeap, return possible CPUs in @node that @attrs wants */
3845 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3846
3847 if (cpumask_empty(cpumask)) {
3848 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3849 "possible intersect\n");
3850 return false;
3851 }
3852
3853 return !cpumask_equal(cpumask, attrs->cpumask);
3854
3855use_dfl:
3856 cpumask_copy(cpumask, attrs->cpumask);
3857 return false;
3858}
3859
3860/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3861static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3862 int node,
3863 struct pool_workqueue *pwq)
3864{
3865 struct pool_workqueue *old_pwq;
3866
3867 lockdep_assert_held(&wq_pool_mutex);
3868 lockdep_assert_held(&wq->mutex);
3869
3870 /* link_pwq() can handle duplicate calls */
3871 link_pwq(pwq);
3872
3873 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3874 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3875 return old_pwq;
3876}
3877
3878/* context to store the prepared attrs & pwqs before applying */
3879struct apply_wqattrs_ctx {
3880 struct workqueue_struct *wq; /* target workqueue */
3881 struct workqueue_attrs *attrs; /* attrs to apply */
3882 struct list_head list; /* queued for batching commit */
3883 struct pool_workqueue *dfl_pwq;
3884 struct pool_workqueue *pwq_tbl[];
3885};
3886
3887/* free the resources after success or abort */
3888static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3889{
3890 if (ctx) {
3891 int node;
3892
3893 for_each_node(node)
3894 put_pwq_unlocked(ctx->pwq_tbl[node]);
3895 put_pwq_unlocked(ctx->dfl_pwq);
3896
3897 free_workqueue_attrs(ctx->attrs);
3898
3899 kfree(ctx);
3900 }
3901}
3902
3903/* allocate the attrs and pwqs for later installation */
3904static struct apply_wqattrs_ctx *
3905apply_wqattrs_prepare(struct workqueue_struct *wq,
3906 const struct workqueue_attrs *attrs)
3907{
3908 struct apply_wqattrs_ctx *ctx;
3909 struct workqueue_attrs *new_attrs, *tmp_attrs;
3910 int node;
3911
3912 lockdep_assert_held(&wq_pool_mutex);
3913
3914 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3915
3916 new_attrs = alloc_workqueue_attrs();
3917 tmp_attrs = alloc_workqueue_attrs();
3918 if (!ctx || !new_attrs || !tmp_attrs)
3919 goto out_free;
3920
3921 /*
3922 * Calculate the attrs of the default pwq.
3923 * If the user configured cpumask doesn't overlap with the
3924 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3925 */
3926 copy_workqueue_attrs(new_attrs, attrs);
3927 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3928 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3929 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3930
3931 /*
3932 * We may create multiple pwqs with differing cpumasks. Make a
3933 * copy of @new_attrs which will be modified and used to obtain
3934 * pools.
3935 */
3936 copy_workqueue_attrs(tmp_attrs, new_attrs);
3937
3938 /*
3939 * If something goes wrong during CPU up/down, we'll fall back to
3940 * the default pwq covering whole @attrs->cpumask. Always create
3941 * it even if we don't use it immediately.
3942 */
3943 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3944 if (!ctx->dfl_pwq)
3945 goto out_free;
3946
3947 for_each_node(node) {
3948 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3949 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3950 if (!ctx->pwq_tbl[node])
3951 goto out_free;
3952 } else {
3953 ctx->dfl_pwq->refcnt++;
3954 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3955 }
3956 }
3957
3958 /* save the user configured attrs and sanitize it. */
3959 copy_workqueue_attrs(new_attrs, attrs);
3960 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3961 ctx->attrs = new_attrs;
3962
3963 ctx->wq = wq;
3964 free_workqueue_attrs(tmp_attrs);
3965 return ctx;
3966
3967out_free:
3968 free_workqueue_attrs(tmp_attrs);
3969 free_workqueue_attrs(new_attrs);
3970 apply_wqattrs_cleanup(ctx);
3971 return NULL;
3972}
3973
3974/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3975static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3976{
3977 int node;
3978
3979 /* all pwqs have been created successfully, let's install'em */
3980 mutex_lock(&ctx->wq->mutex);
3981
3982 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3983
3984 /* save the previous pwq and install the new one */
3985 for_each_node(node)
3986 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3987 ctx->pwq_tbl[node]);
3988
3989 /* @dfl_pwq might not have been used, ensure it's linked */
3990 link_pwq(ctx->dfl_pwq);
3991 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3992
3993 mutex_unlock(&ctx->wq->mutex);
3994}
3995
3996static void apply_wqattrs_lock(void)
3997{
3998 /* CPUs should stay stable across pwq creations and installations */
3999 get_online_cpus();
4000 mutex_lock(&wq_pool_mutex);
4001}
4002
4003static void apply_wqattrs_unlock(void)
4004{
4005 mutex_unlock(&wq_pool_mutex);
4006 put_online_cpus();
4007}
4008
4009static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4010 const struct workqueue_attrs *attrs)
4011{
4012 struct apply_wqattrs_ctx *ctx;
4013
4014 /* only unbound workqueues can change attributes */
4015 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4016 return -EINVAL;
4017
4018 /* creating multiple pwqs breaks ordering guarantee */
4019 if (!list_empty(&wq->pwqs)) {
4020 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4021 return -EINVAL;
4022
4023 wq->flags &= ~__WQ_ORDERED;
4024 }
4025
4026 ctx = apply_wqattrs_prepare(wq, attrs);
4027 if (!ctx)
4028 return -ENOMEM;
4029
4030 /* the ctx has been prepared successfully, let's commit it */
4031 apply_wqattrs_commit(ctx);
4032 apply_wqattrs_cleanup(ctx);
4033
4034 return 0;
4035}
4036
4037/**
4038 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4039 * @wq: the target workqueue
4040 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4041 *
4042 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4043 * machines, this function maps a separate pwq to each NUMA node with
4044 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4045 * NUMA node it was issued on. Older pwqs are released as in-flight work
4046 * items finish. Note that a work item which repeatedly requeues itself
4047 * back-to-back will stay on its current pwq.
4048 *
4049 * Performs GFP_KERNEL allocations.
4050 *
4051 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4052 *
4053 * Return: 0 on success and -errno on failure.
4054 */
4055int apply_workqueue_attrs(struct workqueue_struct *wq,
4056 const struct workqueue_attrs *attrs)
4057{
4058 int ret;
4059
4060 lockdep_assert_cpus_held();
4061
4062 mutex_lock(&wq_pool_mutex);
4063 ret = apply_workqueue_attrs_locked(wq, attrs);
4064 mutex_unlock(&wq_pool_mutex);
4065
4066 return ret;
4067}
4068
4069/**
4070 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4071 * @wq: the target workqueue
4072 * @cpu: the CPU coming up or going down
4073 * @online: whether @cpu is coming up or going down
4074 *
4075 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4076 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4077 * @wq accordingly.
4078 *
4079 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4080 * falls back to @wq->dfl_pwq which may not be optimal but is always
4081 * correct.
4082 *
4083 * Note that when the last allowed CPU of a NUMA node goes offline for a
4084 * workqueue with a cpumask spanning multiple nodes, the workers which were
4085 * already executing the work items for the workqueue will lose their CPU
4086 * affinity and may execute on any CPU. This is similar to how per-cpu
4087 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4088 * affinity, it's the user's responsibility to flush the work item from
4089 * CPU_DOWN_PREPARE.
4090 */
4091static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4092 bool online)
4093{
4094 int node = cpu_to_node(cpu);
4095 int cpu_off = online ? -1 : cpu;
4096 struct pool_workqueue *old_pwq = NULL, *pwq;
4097 struct workqueue_attrs *target_attrs;
4098 cpumask_t *cpumask;
4099
4100 lockdep_assert_held(&wq_pool_mutex);
4101
4102 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4103 wq->unbound_attrs->no_numa)
4104 return;
4105
4106 /*
4107 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4108 * Let's use a preallocated one. The following buf is protected by
4109 * CPU hotplug exclusion.
4110 */
4111 target_attrs = wq_update_unbound_numa_attrs_buf;
4112 cpumask = target_attrs->cpumask;
4113
4114 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4115 pwq = unbound_pwq_by_node(wq, node);
4116
4117 /*
4118 * Let's determine what needs to be done. If the target cpumask is
4119 * different from the default pwq's, we need to compare it to @pwq's
4120 * and create a new one if they don't match. If the target cpumask
4121 * equals the default pwq's, the default pwq should be used.
4122 */
4123 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4124 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4125 return;
4126 } else {
4127 goto use_dfl_pwq;
4128 }
4129
4130 /* create a new pwq */
4131 pwq = alloc_unbound_pwq(wq, target_attrs);
4132 if (!pwq) {
4133 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4134 wq->name);
4135 goto use_dfl_pwq;
4136 }
4137
4138 /* Install the new pwq. */
4139 mutex_lock(&wq->mutex);
4140 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4141 goto out_unlock;
4142
4143use_dfl_pwq:
4144 mutex_lock(&wq->mutex);
4145 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4146 get_pwq(wq->dfl_pwq);
4147 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4148 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4149out_unlock:
4150 mutex_unlock(&wq->mutex);
4151 put_pwq_unlocked(old_pwq);
4152}
4153
4154static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4155{
4156 bool highpri = wq->flags & WQ_HIGHPRI;
4157 int cpu, ret;
4158
4159 if (!(wq->flags & WQ_UNBOUND)) {
4160 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4161 if (!wq->cpu_pwqs)
4162 return -ENOMEM;
4163
4164 for_each_possible_cpu(cpu) {
4165 struct pool_workqueue *pwq =
4166 per_cpu_ptr(wq->cpu_pwqs, cpu);
4167 struct worker_pool *cpu_pools =
4168 per_cpu(cpu_worker_pools, cpu);
4169
4170 init_pwq(pwq, wq, &cpu_pools[highpri]);
4171
4172 mutex_lock(&wq->mutex);
4173 link_pwq(pwq);
4174 mutex_unlock(&wq->mutex);
4175 }
4176 return 0;
4177 }
4178
4179 get_online_cpus();
4180 if (wq->flags & __WQ_ORDERED) {
4181 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4182 /* there should only be single pwq for ordering guarantee */
4183 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4184 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4185 "ordering guarantee broken for workqueue %s\n", wq->name);
4186 } else {
4187 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4188 }
4189 put_online_cpus();
4190
4191 return ret;
4192}
4193
4194static int wq_clamp_max_active(int max_active, unsigned int flags,
4195 const char *name)
4196{
4197 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4198
4199 if (max_active < 1 || max_active > lim)
4200 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4201 max_active, name, 1, lim);
4202
4203 return clamp_val(max_active, 1, lim);
4204}
4205
4206/*
4207 * Workqueues which may be used during memory reclaim should have a rescuer
4208 * to guarantee forward progress.
4209 */
4210static int init_rescuer(struct workqueue_struct *wq)
4211{
4212 struct worker *rescuer;
4213 int ret;
4214
4215 if (!(wq->flags & WQ_MEM_RECLAIM))
4216 return 0;
4217
4218 rescuer = alloc_worker(NUMA_NO_NODE);
4219 if (!rescuer)
4220 return -ENOMEM;
4221
4222 rescuer->rescue_wq = wq;
4223 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4224 if (IS_ERR(rescuer->task)) {
4225 ret = PTR_ERR(rescuer->task);
4226 kfree(rescuer);
4227 return ret;
4228 }
4229
4230 wq->rescuer = rescuer;
4231 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4232 wake_up_process(rescuer->task);
4233
4234 return 0;
4235}
4236
4237__printf(1, 4)
4238struct workqueue_struct *alloc_workqueue(const char *fmt,
4239 unsigned int flags,
4240 int max_active, ...)
4241{
4242 size_t tbl_size = 0;
4243 va_list args;
4244 struct workqueue_struct *wq;
4245 struct pool_workqueue *pwq;
4246
4247 /*
4248 * Unbound && max_active == 1 used to imply ordered, which is no
4249 * longer the case on NUMA machines due to per-node pools. While
4250 * alloc_ordered_workqueue() is the right way to create an ordered
4251 * workqueue, keep the previous behavior to avoid subtle breakages
4252 * on NUMA.
4253 */
4254 if ((flags & WQ_UNBOUND) && max_active == 1)
4255 flags |= __WQ_ORDERED;
4256
4257 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4258 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4259 flags |= WQ_UNBOUND;
4260
4261 /* allocate wq and format name */
4262 if (flags & WQ_UNBOUND)
4263 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4264
4265 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4266 if (!wq)
4267 return NULL;
4268
4269 if (flags & WQ_UNBOUND) {
4270 wq->unbound_attrs = alloc_workqueue_attrs();
4271 if (!wq->unbound_attrs)
4272 goto err_free_wq;
4273 }
4274
4275 va_start(args, max_active);
4276 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4277 va_end(args);
4278
4279 max_active = max_active ?: WQ_DFL_ACTIVE;
4280 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4281
4282 /* init wq */
4283 wq->flags = flags;
4284 wq->saved_max_active = max_active;
4285 mutex_init(&wq->mutex);
4286 atomic_set(&wq->nr_pwqs_to_flush, 0);
4287 INIT_LIST_HEAD(&wq->pwqs);
4288 INIT_LIST_HEAD(&wq->flusher_queue);
4289 INIT_LIST_HEAD(&wq->flusher_overflow);
4290 INIT_LIST_HEAD(&wq->maydays);
4291
4292 wq_init_lockdep(wq);
4293 INIT_LIST_HEAD(&wq->list);
4294
4295 if (alloc_and_link_pwqs(wq) < 0)
4296 goto err_unreg_lockdep;
4297
4298 if (wq_online && init_rescuer(wq) < 0)
4299 goto err_destroy;
4300
4301 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4302 goto err_destroy;
4303
4304 /*
4305 * wq_pool_mutex protects global freeze state and workqueues list.
4306 * Grab it, adjust max_active and add the new @wq to workqueues
4307 * list.
4308 */
4309 mutex_lock(&wq_pool_mutex);
4310
4311 mutex_lock(&wq->mutex);
4312 for_each_pwq(pwq, wq)
4313 pwq_adjust_max_active(pwq);
4314 mutex_unlock(&wq->mutex);
4315
4316 list_add_tail_rcu(&wq->list, &workqueues);
4317
4318 mutex_unlock(&wq_pool_mutex);
4319
4320 return wq;
4321
4322err_unreg_lockdep:
4323 wq_unregister_lockdep(wq);
4324 wq_free_lockdep(wq);
4325err_free_wq:
4326 free_workqueue_attrs(wq->unbound_attrs);
4327 kfree(wq);
4328 return NULL;
4329err_destroy:
4330 destroy_workqueue(wq);
4331 return NULL;
4332}
4333EXPORT_SYMBOL_GPL(alloc_workqueue);
4334
4335static bool pwq_busy(struct pool_workqueue *pwq)
4336{
4337 int i;
4338
4339 for (i = 0; i < WORK_NR_COLORS; i++)
4340 if (pwq->nr_in_flight[i])
4341 return true;
4342
4343 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4344 return true;
4345 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4346 return true;
4347
4348 return false;
4349}
4350
4351/**
4352 * destroy_workqueue - safely terminate a workqueue
4353 * @wq: target workqueue
4354 *
4355 * Safely destroy a workqueue. All work currently pending will be done first.
4356 */
4357void destroy_workqueue(struct workqueue_struct *wq)
4358{
4359 struct pool_workqueue *pwq;
4360 int node;
4361
4362 /*
4363 * Remove it from sysfs first so that sanity check failure doesn't
4364 * lead to sysfs name conflicts.
4365 */
4366 workqueue_sysfs_unregister(wq);
4367
4368 /* drain it before proceeding with destruction */
4369 drain_workqueue(wq);
4370
4371 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4372 if (wq->rescuer) {
4373 struct worker *rescuer = wq->rescuer;
4374
4375 /* this prevents new queueing */
4376 raw_spin_lock_irq(&wq_mayday_lock);
4377 wq->rescuer = NULL;
4378 raw_spin_unlock_irq(&wq_mayday_lock);
4379
4380 /* rescuer will empty maydays list before exiting */
4381 kthread_stop(rescuer->task);
4382 kfree(rescuer);
4383 }
4384
4385 /*
4386 * Sanity checks - grab all the locks so that we wait for all
4387 * in-flight operations which may do put_pwq().
4388 */
4389 mutex_lock(&wq_pool_mutex);
4390 mutex_lock(&wq->mutex);
4391 for_each_pwq(pwq, wq) {
4392 raw_spin_lock_irq(&pwq->pool->lock);
4393 if (WARN_ON(pwq_busy(pwq))) {
4394 pr_warn("%s: %s has the following busy pwq\n",
4395 __func__, wq->name);
4396 show_pwq(pwq);
4397 raw_spin_unlock_irq(&pwq->pool->lock);
4398 mutex_unlock(&wq->mutex);
4399 mutex_unlock(&wq_pool_mutex);
4400 show_workqueue_state();
4401 return;
4402 }
4403 raw_spin_unlock_irq(&pwq->pool->lock);
4404 }
4405 mutex_unlock(&wq->mutex);
4406
4407 /*
4408 * wq list is used to freeze wq, remove from list after
4409 * flushing is complete in case freeze races us.
4410 */
4411 list_del_rcu(&wq->list);
4412 mutex_unlock(&wq_pool_mutex);
4413
4414 if (!(wq->flags & WQ_UNBOUND)) {
4415 wq_unregister_lockdep(wq);
4416 /*
4417 * The base ref is never dropped on per-cpu pwqs. Directly
4418 * schedule RCU free.
4419 */
4420 call_rcu(&wq->rcu, rcu_free_wq);
4421 } else {
4422 /*
4423 * We're the sole accessor of @wq at this point. Directly
4424 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4425 * @wq will be freed when the last pwq is released.
4426 */
4427 for_each_node(node) {
4428 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4429 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4430 put_pwq_unlocked(pwq);
4431 }
4432
4433 /*
4434 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4435 * put. Don't access it afterwards.
4436 */
4437 pwq = wq->dfl_pwq;
4438 wq->dfl_pwq = NULL;
4439 put_pwq_unlocked(pwq);
4440 }
4441}
4442EXPORT_SYMBOL_GPL(destroy_workqueue);
4443
4444/**
4445 * workqueue_set_max_active - adjust max_active of a workqueue
4446 * @wq: target workqueue
4447 * @max_active: new max_active value.
4448 *
4449 * Set max_active of @wq to @max_active.
4450 *
4451 * CONTEXT:
4452 * Don't call from IRQ context.
4453 */
4454void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4455{
4456 struct pool_workqueue *pwq;
4457
4458 /* disallow meddling with max_active for ordered workqueues */
4459 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4460 return;
4461
4462 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4463
4464 mutex_lock(&wq->mutex);
4465
4466 wq->flags &= ~__WQ_ORDERED;
4467 wq->saved_max_active = max_active;
4468
4469 for_each_pwq(pwq, wq)
4470 pwq_adjust_max_active(pwq);
4471
4472 mutex_unlock(&wq->mutex);
4473}
4474EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4475
4476/**
4477 * current_work - retrieve %current task's work struct
4478 *
4479 * Determine if %current task is a workqueue worker and what it's working on.
4480 * Useful to find out the context that the %current task is running in.
4481 *
4482 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4483 */
4484struct work_struct *current_work(void)
4485{
4486 struct worker *worker = current_wq_worker();
4487
4488 return worker ? worker->current_work : NULL;
4489}
4490EXPORT_SYMBOL(current_work);
4491
4492/**
4493 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4494 *
4495 * Determine whether %current is a workqueue rescuer. Can be used from
4496 * work functions to determine whether it's being run off the rescuer task.
4497 *
4498 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4499 */
4500bool current_is_workqueue_rescuer(void)
4501{
4502 struct worker *worker = current_wq_worker();
4503
4504 return worker && worker->rescue_wq;
4505}
4506
4507/**
4508 * workqueue_congested - test whether a workqueue is congested
4509 * @cpu: CPU in question
4510 * @wq: target workqueue
4511 *
4512 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4513 * no synchronization around this function and the test result is
4514 * unreliable and only useful as advisory hints or for debugging.
4515 *
4516 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4517 * Note that both per-cpu and unbound workqueues may be associated with
4518 * multiple pool_workqueues which have separate congested states. A
4519 * workqueue being congested on one CPU doesn't mean the workqueue is also
4520 * contested on other CPUs / NUMA nodes.
4521 *
4522 * Return:
4523 * %true if congested, %false otherwise.
4524 */
4525bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4526{
4527 struct pool_workqueue *pwq;
4528 bool ret;
4529
4530 rcu_read_lock();
4531 preempt_disable();
4532
4533 if (cpu == WORK_CPU_UNBOUND)
4534 cpu = smp_processor_id();
4535
4536 if (!(wq->flags & WQ_UNBOUND))
4537 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4538 else
4539 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4540
4541 ret = !list_empty(&pwq->delayed_works);
4542 preempt_enable();
4543 rcu_read_unlock();
4544
4545 return ret;
4546}
4547EXPORT_SYMBOL_GPL(workqueue_congested);
4548
4549/**
4550 * work_busy - test whether a work is currently pending or running
4551 * @work: the work to be tested
4552 *
4553 * Test whether @work is currently pending or running. There is no
4554 * synchronization around this function and the test result is
4555 * unreliable and only useful as advisory hints or for debugging.
4556 *
4557 * Return:
4558 * OR'd bitmask of WORK_BUSY_* bits.
4559 */
4560unsigned int work_busy(struct work_struct *work)
4561{
4562 struct worker_pool *pool;
4563 unsigned long flags;
4564 unsigned int ret = 0;
4565
4566 if (work_pending(work))
4567 ret |= WORK_BUSY_PENDING;
4568
4569 rcu_read_lock();
4570 pool = get_work_pool(work);
4571 if (pool) {
4572 raw_spin_lock_irqsave(&pool->lock, flags);
4573 if (find_worker_executing_work(pool, work))
4574 ret |= WORK_BUSY_RUNNING;
4575 raw_spin_unlock_irqrestore(&pool->lock, flags);
4576 }
4577 rcu_read_unlock();
4578
4579 return ret;
4580}
4581EXPORT_SYMBOL_GPL(work_busy);
4582
4583/**
4584 * set_worker_desc - set description for the current work item
4585 * @fmt: printf-style format string
4586 * @...: arguments for the format string
4587 *
4588 * This function can be called by a running work function to describe what
4589 * the work item is about. If the worker task gets dumped, this
4590 * information will be printed out together to help debugging. The
4591 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4592 */
4593void set_worker_desc(const char *fmt, ...)
4594{
4595 struct worker *worker = current_wq_worker();
4596 va_list args;
4597
4598 if (worker) {
4599 va_start(args, fmt);
4600 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4601 va_end(args);
4602 }
4603}
4604EXPORT_SYMBOL_GPL(set_worker_desc);
4605
4606/**
4607 * print_worker_info - print out worker information and description
4608 * @log_lvl: the log level to use when printing
4609 * @task: target task
4610 *
4611 * If @task is a worker and currently executing a work item, print out the
4612 * name of the workqueue being serviced and worker description set with
4613 * set_worker_desc() by the currently executing work item.
4614 *
4615 * This function can be safely called on any task as long as the
4616 * task_struct itself is accessible. While safe, this function isn't
4617 * synchronized and may print out mixups or garbages of limited length.
4618 */
4619void print_worker_info(const char *log_lvl, struct task_struct *task)
4620{
4621 work_func_t *fn = NULL;
4622 char name[WQ_NAME_LEN] = { };
4623 char desc[WORKER_DESC_LEN] = { };
4624 struct pool_workqueue *pwq = NULL;
4625 struct workqueue_struct *wq = NULL;
4626 struct worker *worker;
4627
4628 if (!(task->flags & PF_WQ_WORKER))
4629 return;
4630
4631 /*
4632 * This function is called without any synchronization and @task
4633 * could be in any state. Be careful with dereferences.
4634 */
4635 worker = kthread_probe_data(task);
4636
4637 /*
4638 * Carefully copy the associated workqueue's workfn, name and desc.
4639 * Keep the original last '\0' in case the original is garbage.
4640 */
4641 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
4642 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
4643 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
4644 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
4645 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
4646
4647 if (fn || name[0] || desc[0]) {
4648 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4649 if (strcmp(name, desc))
4650 pr_cont(" (%s)", desc);
4651 pr_cont("\n");
4652 }
4653}
4654
4655static void pr_cont_pool_info(struct worker_pool *pool)
4656{
4657 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4658 if (pool->node != NUMA_NO_NODE)
4659 pr_cont(" node=%d", pool->node);
4660 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4661}
4662
4663static void pr_cont_work(bool comma, struct work_struct *work)
4664{
4665 if (work->func == wq_barrier_func) {
4666 struct wq_barrier *barr;
4667
4668 barr = container_of(work, struct wq_barrier, work);
4669
4670 pr_cont("%s BAR(%d)", comma ? "," : "",
4671 task_pid_nr(barr->task));
4672 } else {
4673 pr_cont("%s %ps", comma ? "," : "", work->func);
4674 }
4675}
4676
4677static void show_pwq(struct pool_workqueue *pwq)
4678{
4679 struct worker_pool *pool = pwq->pool;
4680 struct work_struct *work;
4681 struct worker *worker;
4682 bool has_in_flight = false, has_pending = false;
4683 int bkt;
4684
4685 pr_info(" pwq %d:", pool->id);
4686 pr_cont_pool_info(pool);
4687
4688 pr_cont(" active=%d/%d refcnt=%d%s\n",
4689 pwq->nr_active, pwq->max_active, pwq->refcnt,
4690 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4691
4692 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4693 if (worker->current_pwq == pwq) {
4694 has_in_flight = true;
4695 break;
4696 }
4697 }
4698 if (has_in_flight) {
4699 bool comma = false;
4700
4701 pr_info(" in-flight:");
4702 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4703 if (worker->current_pwq != pwq)
4704 continue;
4705
4706 pr_cont("%s %d%s:%ps", comma ? "," : "",
4707 task_pid_nr(worker->task),
4708 worker->rescue_wq ? "(RESCUER)" : "",
4709 worker->current_func);
4710 list_for_each_entry(work, &worker->scheduled, entry)
4711 pr_cont_work(false, work);
4712 comma = true;
4713 }
4714 pr_cont("\n");
4715 }
4716
4717 list_for_each_entry(work, &pool->worklist, entry) {
4718 if (get_work_pwq(work) == pwq) {
4719 has_pending = true;
4720 break;
4721 }
4722 }
4723 if (has_pending) {
4724 bool comma = false;
4725
4726 pr_info(" pending:");
4727 list_for_each_entry(work, &pool->worklist, entry) {
4728 if (get_work_pwq(work) != pwq)
4729 continue;
4730
4731 pr_cont_work(comma, work);
4732 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4733 }
4734 pr_cont("\n");
4735 }
4736
4737 if (!list_empty(&pwq->delayed_works)) {
4738 bool comma = false;
4739
4740 pr_info(" delayed:");
4741 list_for_each_entry(work, &pwq->delayed_works, entry) {
4742 pr_cont_work(comma, work);
4743 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4744 }
4745 pr_cont("\n");
4746 }
4747}
4748
4749/**
4750 * show_workqueue_state - dump workqueue state
4751 *
4752 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4753 * all busy workqueues and pools.
4754 */
4755void show_workqueue_state(void)
4756{
4757 struct workqueue_struct *wq;
4758 struct worker_pool *pool;
4759 unsigned long flags;
4760 int pi;
4761
4762 rcu_read_lock();
4763
4764 pr_info("Showing busy workqueues and worker pools:\n");
4765
4766 list_for_each_entry_rcu(wq, &workqueues, list) {
4767 struct pool_workqueue *pwq;
4768 bool idle = true;
4769
4770 for_each_pwq(pwq, wq) {
4771 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4772 idle = false;
4773 break;
4774 }
4775 }
4776 if (idle)
4777 continue;
4778
4779 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4780
4781 for_each_pwq(pwq, wq) {
4782 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4783 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4784 show_pwq(pwq);
4785 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4786 /*
4787 * We could be printing a lot from atomic context, e.g.
4788 * sysrq-t -> show_workqueue_state(). Avoid triggering
4789 * hard lockup.
4790 */
4791 touch_nmi_watchdog();
4792 }
4793 }
4794
4795 for_each_pool(pool, pi) {
4796 struct worker *worker;
4797 bool first = true;
4798
4799 raw_spin_lock_irqsave(&pool->lock, flags);
4800 if (pool->nr_workers == pool->nr_idle)
4801 goto next_pool;
4802
4803 pr_info("pool %d:", pool->id);
4804 pr_cont_pool_info(pool);
4805 pr_cont(" hung=%us workers=%d",
4806 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4807 pool->nr_workers);
4808 if (pool->manager)
4809 pr_cont(" manager: %d",
4810 task_pid_nr(pool->manager->task));
4811 list_for_each_entry(worker, &pool->idle_list, entry) {
4812 pr_cont(" %s%d", first ? "idle: " : "",
4813 task_pid_nr(worker->task));
4814 first = false;
4815 }
4816 pr_cont("\n");
4817 next_pool:
4818 raw_spin_unlock_irqrestore(&pool->lock, flags);
4819 /*
4820 * We could be printing a lot from atomic context, e.g.
4821 * sysrq-t -> show_workqueue_state(). Avoid triggering
4822 * hard lockup.
4823 */
4824 touch_nmi_watchdog();
4825 }
4826
4827 rcu_read_unlock();
4828}
4829
4830/* used to show worker information through /proc/PID/{comm,stat,status} */
4831void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4832{
4833 int off;
4834
4835 /* always show the actual comm */
4836 off = strscpy(buf, task->comm, size);
4837 if (off < 0)
4838 return;
4839
4840 /* stabilize PF_WQ_WORKER and worker pool association */
4841 mutex_lock(&wq_pool_attach_mutex);
4842
4843 if (task->flags & PF_WQ_WORKER) {
4844 struct worker *worker = kthread_data(task);
4845 struct worker_pool *pool = worker->pool;
4846
4847 if (pool) {
4848 raw_spin_lock_irq(&pool->lock);
4849 /*
4850 * ->desc tracks information (wq name or
4851 * set_worker_desc()) for the latest execution. If
4852 * current, prepend '+', otherwise '-'.
4853 */
4854 if (worker->desc[0] != '\0') {
4855 if (worker->current_work)
4856 scnprintf(buf + off, size - off, "+%s",
4857 worker->desc);
4858 else
4859 scnprintf(buf + off, size - off, "-%s",
4860 worker->desc);
4861 }
4862 raw_spin_unlock_irq(&pool->lock);
4863 }
4864 }
4865
4866 mutex_unlock(&wq_pool_attach_mutex);
4867}
4868
4869#ifdef CONFIG_SMP
4870
4871/*
4872 * CPU hotplug.
4873 *
4874 * There are two challenges in supporting CPU hotplug. Firstly, there
4875 * are a lot of assumptions on strong associations among work, pwq and
4876 * pool which make migrating pending and scheduled works very
4877 * difficult to implement without impacting hot paths. Secondly,
4878 * worker pools serve mix of short, long and very long running works making
4879 * blocked draining impractical.
4880 *
4881 * This is solved by allowing the pools to be disassociated from the CPU
4882 * running as an unbound one and allowing it to be reattached later if the
4883 * cpu comes back online.
4884 */
4885
4886static void unbind_workers(int cpu)
4887{
4888 struct worker_pool *pool;
4889 struct worker *worker;
4890
4891 for_each_cpu_worker_pool(pool, cpu) {
4892 mutex_lock(&wq_pool_attach_mutex);
4893 raw_spin_lock_irq(&pool->lock);
4894
4895 /*
4896 * We've blocked all attach/detach operations. Make all workers
4897 * unbound and set DISASSOCIATED. Before this, all workers
4898 * except for the ones which are still executing works from
4899 * before the last CPU down must be on the cpu. After
4900 * this, they may become diasporas.
4901 */
4902 for_each_pool_worker(worker, pool)
4903 worker->flags |= WORKER_UNBOUND;
4904
4905 pool->flags |= POOL_DISASSOCIATED;
4906
4907 raw_spin_unlock_irq(&pool->lock);
4908 mutex_unlock(&wq_pool_attach_mutex);
4909
4910 /*
4911 * Call schedule() so that we cross rq->lock and thus can
4912 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4913 * This is necessary as scheduler callbacks may be invoked
4914 * from other cpus.
4915 */
4916 schedule();
4917
4918 /*
4919 * Sched callbacks are disabled now. Zap nr_running.
4920 * After this, nr_running stays zero and need_more_worker()
4921 * and keep_working() are always true as long as the
4922 * worklist is not empty. This pool now behaves as an
4923 * unbound (in terms of concurrency management) pool which
4924 * are served by workers tied to the pool.
4925 */
4926 atomic_set(&pool->nr_running, 0);
4927
4928 /*
4929 * With concurrency management just turned off, a busy
4930 * worker blocking could lead to lengthy stalls. Kick off
4931 * unbound chain execution of currently pending work items.
4932 */
4933 raw_spin_lock_irq(&pool->lock);
4934 wake_up_worker(pool);
4935 raw_spin_unlock_irq(&pool->lock);
4936 }
4937}
4938
4939/**
4940 * rebind_workers - rebind all workers of a pool to the associated CPU
4941 * @pool: pool of interest
4942 *
4943 * @pool->cpu is coming online. Rebind all workers to the CPU.
4944 */
4945static void rebind_workers(struct worker_pool *pool)
4946{
4947 struct worker *worker;
4948
4949 lockdep_assert_held(&wq_pool_attach_mutex);
4950
4951 /*
4952 * Restore CPU affinity of all workers. As all idle workers should
4953 * be on the run-queue of the associated CPU before any local
4954 * wake-ups for concurrency management happen, restore CPU affinity
4955 * of all workers first and then clear UNBOUND. As we're called
4956 * from CPU_ONLINE, the following shouldn't fail.
4957 */
4958 for_each_pool_worker(worker, pool)
4959 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4960 pool->attrs->cpumask) < 0);
4961
4962 raw_spin_lock_irq(&pool->lock);
4963
4964 pool->flags &= ~POOL_DISASSOCIATED;
4965
4966 for_each_pool_worker(worker, pool) {
4967 unsigned int worker_flags = worker->flags;
4968
4969 /*
4970 * A bound idle worker should actually be on the runqueue
4971 * of the associated CPU for local wake-ups targeting it to
4972 * work. Kick all idle workers so that they migrate to the
4973 * associated CPU. Doing this in the same loop as
4974 * replacing UNBOUND with REBOUND is safe as no worker will
4975 * be bound before @pool->lock is released.
4976 */
4977 if (worker_flags & WORKER_IDLE)
4978 wake_up_process(worker->task);
4979
4980 /*
4981 * We want to clear UNBOUND but can't directly call
4982 * worker_clr_flags() or adjust nr_running. Atomically
4983 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4984 * @worker will clear REBOUND using worker_clr_flags() when
4985 * it initiates the next execution cycle thus restoring
4986 * concurrency management. Note that when or whether
4987 * @worker clears REBOUND doesn't affect correctness.
4988 *
4989 * WRITE_ONCE() is necessary because @worker->flags may be
4990 * tested without holding any lock in
4991 * wq_worker_running(). Without it, NOT_RUNNING test may
4992 * fail incorrectly leading to premature concurrency
4993 * management operations.
4994 */
4995 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4996 worker_flags |= WORKER_REBOUND;
4997 worker_flags &= ~WORKER_UNBOUND;
4998 WRITE_ONCE(worker->flags, worker_flags);
4999 }
5000
5001 raw_spin_unlock_irq(&pool->lock);
5002}
5003
5004/**
5005 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5006 * @pool: unbound pool of interest
5007 * @cpu: the CPU which is coming up
5008 *
5009 * An unbound pool may end up with a cpumask which doesn't have any online
5010 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5011 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5012 * online CPU before, cpus_allowed of all its workers should be restored.
5013 */
5014static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5015{
5016 static cpumask_t cpumask;
5017 struct worker *worker;
5018
5019 lockdep_assert_held(&wq_pool_attach_mutex);
5020
5021 /* is @cpu allowed for @pool? */
5022 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5023 return;
5024
5025 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5026
5027 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5028 for_each_pool_worker(worker, pool)
5029 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5030}
5031
5032int workqueue_prepare_cpu(unsigned int cpu)
5033{
5034 struct worker_pool *pool;
5035
5036 for_each_cpu_worker_pool(pool, cpu) {
5037 if (pool->nr_workers)
5038 continue;
5039 if (!create_worker(pool))
5040 return -ENOMEM;
5041 }
5042 return 0;
5043}
5044
5045int workqueue_online_cpu(unsigned int cpu)
5046{
5047 struct worker_pool *pool;
5048 struct workqueue_struct *wq;
5049 int pi;
5050
5051 mutex_lock(&wq_pool_mutex);
5052
5053 for_each_pool(pool, pi) {
5054 mutex_lock(&wq_pool_attach_mutex);
5055
5056 if (pool->cpu == cpu)
5057 rebind_workers(pool);
5058 else if (pool->cpu < 0)
5059 restore_unbound_workers_cpumask(pool, cpu);
5060
5061 mutex_unlock(&wq_pool_attach_mutex);
5062 }
5063
5064 /* update NUMA affinity of unbound workqueues */
5065 list_for_each_entry(wq, &workqueues, list)
5066 wq_update_unbound_numa(wq, cpu, true);
5067
5068 mutex_unlock(&wq_pool_mutex);
5069 return 0;
5070}
5071
5072int workqueue_offline_cpu(unsigned int cpu)
5073{
5074 struct workqueue_struct *wq;
5075
5076 /* unbinding per-cpu workers should happen on the local CPU */
5077 if (WARN_ON(cpu != smp_processor_id()))
5078 return -1;
5079
5080 unbind_workers(cpu);
5081
5082 /* update NUMA affinity of unbound workqueues */
5083 mutex_lock(&wq_pool_mutex);
5084 list_for_each_entry(wq, &workqueues, list)
5085 wq_update_unbound_numa(wq, cpu, false);
5086 mutex_unlock(&wq_pool_mutex);
5087
5088 return 0;
5089}
5090
5091struct work_for_cpu {
5092 struct work_struct work;
5093 long (*fn)(void *);
5094 void *arg;
5095 long ret;
5096};
5097
5098static void work_for_cpu_fn(struct work_struct *work)
5099{
5100 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5101
5102 wfc->ret = wfc->fn(wfc->arg);
5103}
5104
5105/**
5106 * work_on_cpu - run a function in thread context on a particular cpu
5107 * @cpu: the cpu to run on
5108 * @fn: the function to run
5109 * @arg: the function arg
5110 *
5111 * It is up to the caller to ensure that the cpu doesn't go offline.
5112 * The caller must not hold any locks which would prevent @fn from completing.
5113 *
5114 * Return: The value @fn returns.
5115 */
5116long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5117{
5118 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5119
5120 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5121 schedule_work_on(cpu, &wfc.work);
5122 flush_work(&wfc.work);
5123 destroy_work_on_stack(&wfc.work);
5124 return wfc.ret;
5125}
5126EXPORT_SYMBOL_GPL(work_on_cpu);
5127
5128/**
5129 * work_on_cpu_safe - run a function in thread context on a particular cpu
5130 * @cpu: the cpu to run on
5131 * @fn: the function to run
5132 * @arg: the function argument
5133 *
5134 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5135 * any locks which would prevent @fn from completing.
5136 *
5137 * Return: The value @fn returns.
5138 */
5139long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5140{
5141 long ret = -ENODEV;
5142
5143 get_online_cpus();
5144 if (cpu_online(cpu))
5145 ret = work_on_cpu(cpu, fn, arg);
5146 put_online_cpus();
5147 return ret;
5148}
5149EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5150#endif /* CONFIG_SMP */
5151
5152#ifdef CONFIG_FREEZER
5153
5154/**
5155 * freeze_workqueues_begin - begin freezing workqueues
5156 *
5157 * Start freezing workqueues. After this function returns, all freezable
5158 * workqueues will queue new works to their delayed_works list instead of
5159 * pool->worklist.
5160 *
5161 * CONTEXT:
5162 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5163 */
5164void freeze_workqueues_begin(void)
5165{
5166 struct workqueue_struct *wq;
5167 struct pool_workqueue *pwq;
5168
5169 mutex_lock(&wq_pool_mutex);
5170
5171 WARN_ON_ONCE(workqueue_freezing);
5172 workqueue_freezing = true;
5173
5174 list_for_each_entry(wq, &workqueues, list) {
5175 mutex_lock(&wq->mutex);
5176 for_each_pwq(pwq, wq)
5177 pwq_adjust_max_active(pwq);
5178 mutex_unlock(&wq->mutex);
5179 }
5180
5181 mutex_unlock(&wq_pool_mutex);
5182}
5183
5184/**
5185 * freeze_workqueues_busy - are freezable workqueues still busy?
5186 *
5187 * Check whether freezing is complete. This function must be called
5188 * between freeze_workqueues_begin() and thaw_workqueues().
5189 *
5190 * CONTEXT:
5191 * Grabs and releases wq_pool_mutex.
5192 *
5193 * Return:
5194 * %true if some freezable workqueues are still busy. %false if freezing
5195 * is complete.
5196 */
5197bool freeze_workqueues_busy(void)
5198{
5199 bool busy = false;
5200 struct workqueue_struct *wq;
5201 struct pool_workqueue *pwq;
5202
5203 mutex_lock(&wq_pool_mutex);
5204
5205 WARN_ON_ONCE(!workqueue_freezing);
5206
5207 list_for_each_entry(wq, &workqueues, list) {
5208 if (!(wq->flags & WQ_FREEZABLE))
5209 continue;
5210 /*
5211 * nr_active is monotonically decreasing. It's safe
5212 * to peek without lock.
5213 */
5214 rcu_read_lock();
5215 for_each_pwq(pwq, wq) {
5216 WARN_ON_ONCE(pwq->nr_active < 0);
5217 if (pwq->nr_active) {
5218 busy = true;
5219 rcu_read_unlock();
5220 goto out_unlock;
5221 }
5222 }
5223 rcu_read_unlock();
5224 }
5225out_unlock:
5226 mutex_unlock(&wq_pool_mutex);
5227 return busy;
5228}
5229
5230/**
5231 * thaw_workqueues - thaw workqueues
5232 *
5233 * Thaw workqueues. Normal queueing is restored and all collected
5234 * frozen works are transferred to their respective pool worklists.
5235 *
5236 * CONTEXT:
5237 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5238 */
5239void thaw_workqueues(void)
5240{
5241 struct workqueue_struct *wq;
5242 struct pool_workqueue *pwq;
5243
5244 mutex_lock(&wq_pool_mutex);
5245
5246 if (!workqueue_freezing)
5247 goto out_unlock;
5248
5249 workqueue_freezing = false;
5250
5251 /* restore max_active and repopulate worklist */
5252 list_for_each_entry(wq, &workqueues, list) {
5253 mutex_lock(&wq->mutex);
5254 for_each_pwq(pwq, wq)
5255 pwq_adjust_max_active(pwq);
5256 mutex_unlock(&wq->mutex);
5257 }
5258
5259out_unlock:
5260 mutex_unlock(&wq_pool_mutex);
5261}
5262#endif /* CONFIG_FREEZER */
5263
5264static int workqueue_apply_unbound_cpumask(void)
5265{
5266 LIST_HEAD(ctxs);
5267 int ret = 0;
5268 struct workqueue_struct *wq;
5269 struct apply_wqattrs_ctx *ctx, *n;
5270
5271 lockdep_assert_held(&wq_pool_mutex);
5272
5273 list_for_each_entry(wq, &workqueues, list) {
5274 if (!(wq->flags & WQ_UNBOUND))
5275 continue;
5276 /* creating multiple pwqs breaks ordering guarantee */
5277 if (wq->flags & __WQ_ORDERED)
5278 continue;
5279
5280 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5281 if (!ctx) {
5282 ret = -ENOMEM;
5283 break;
5284 }
5285
5286 list_add_tail(&ctx->list, &ctxs);
5287 }
5288
5289 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5290 if (!ret)
5291 apply_wqattrs_commit(ctx);
5292 apply_wqattrs_cleanup(ctx);
5293 }
5294
5295 return ret;
5296}
5297
5298/**
5299 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5300 * @cpumask: the cpumask to set
5301 *
5302 * The low-level workqueues cpumask is a global cpumask that limits
5303 * the affinity of all unbound workqueues. This function check the @cpumask
5304 * and apply it to all unbound workqueues and updates all pwqs of them.
5305 *
5306 * Retun: 0 - Success
5307 * -EINVAL - Invalid @cpumask
5308 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5309 */
5310int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5311{
5312 int ret = -EINVAL;
5313 cpumask_var_t saved_cpumask;
5314
5315 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5316 return -ENOMEM;
5317
5318 /*
5319 * Not excluding isolated cpus on purpose.
5320 * If the user wishes to include them, we allow that.
5321 */
5322 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5323 if (!cpumask_empty(cpumask)) {
5324 apply_wqattrs_lock();
5325
5326 /* save the old wq_unbound_cpumask. */
5327 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5328
5329 /* update wq_unbound_cpumask at first and apply it to wqs. */
5330 cpumask_copy(wq_unbound_cpumask, cpumask);
5331 ret = workqueue_apply_unbound_cpumask();
5332
5333 /* restore the wq_unbound_cpumask when failed. */
5334 if (ret < 0)
5335 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5336
5337 apply_wqattrs_unlock();
5338 }
5339
5340 free_cpumask_var(saved_cpumask);
5341 return ret;
5342}
5343
5344#ifdef CONFIG_SYSFS
5345/*
5346 * Workqueues with WQ_SYSFS flag set is visible to userland via
5347 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5348 * following attributes.
5349 *
5350 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5351 * max_active RW int : maximum number of in-flight work items
5352 *
5353 * Unbound workqueues have the following extra attributes.
5354 *
5355 * pool_ids RO int : the associated pool IDs for each node
5356 * nice RW int : nice value of the workers
5357 * cpumask RW mask : bitmask of allowed CPUs for the workers
5358 * numa RW bool : whether enable NUMA affinity
5359 */
5360struct wq_device {
5361 struct workqueue_struct *wq;
5362 struct device dev;
5363};
5364
5365static struct workqueue_struct *dev_to_wq(struct device *dev)
5366{
5367 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5368
5369 return wq_dev->wq;
5370}
5371
5372static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5373 char *buf)
5374{
5375 struct workqueue_struct *wq = dev_to_wq(dev);
5376
5377 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5378}
5379static DEVICE_ATTR_RO(per_cpu);
5380
5381static ssize_t max_active_show(struct device *dev,
5382 struct device_attribute *attr, char *buf)
5383{
5384 struct workqueue_struct *wq = dev_to_wq(dev);
5385
5386 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5387}
5388
5389static ssize_t max_active_store(struct device *dev,
5390 struct device_attribute *attr, const char *buf,
5391 size_t count)
5392{
5393 struct workqueue_struct *wq = dev_to_wq(dev);
5394 int val;
5395
5396 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5397 return -EINVAL;
5398
5399 workqueue_set_max_active(wq, val);
5400 return count;
5401}
5402static DEVICE_ATTR_RW(max_active);
5403
5404static struct attribute *wq_sysfs_attrs[] = {
5405 &dev_attr_per_cpu.attr,
5406 &dev_attr_max_active.attr,
5407 NULL,
5408};
5409ATTRIBUTE_GROUPS(wq_sysfs);
5410
5411static ssize_t wq_pool_ids_show(struct device *dev,
5412 struct device_attribute *attr, char *buf)
5413{
5414 struct workqueue_struct *wq = dev_to_wq(dev);
5415 const char *delim = "";
5416 int node, written = 0;
5417
5418 get_online_cpus();
5419 rcu_read_lock();
5420 for_each_node(node) {
5421 written += scnprintf(buf + written, PAGE_SIZE - written,
5422 "%s%d:%d", delim, node,
5423 unbound_pwq_by_node(wq, node)->pool->id);
5424 delim = " ";
5425 }
5426 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5427 rcu_read_unlock();
5428 put_online_cpus();
5429
5430 return written;
5431}
5432
5433static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5434 char *buf)
5435{
5436 struct workqueue_struct *wq = dev_to_wq(dev);
5437 int written;
5438
5439 mutex_lock(&wq->mutex);
5440 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5441 mutex_unlock(&wq->mutex);
5442
5443 return written;
5444}
5445
5446/* prepare workqueue_attrs for sysfs store operations */
5447static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5448{
5449 struct workqueue_attrs *attrs;
5450
5451 lockdep_assert_held(&wq_pool_mutex);
5452
5453 attrs = alloc_workqueue_attrs();
5454 if (!attrs)
5455 return NULL;
5456
5457 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5458 return attrs;
5459}
5460
5461static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5462 const char *buf, size_t count)
5463{
5464 struct workqueue_struct *wq = dev_to_wq(dev);
5465 struct workqueue_attrs *attrs;
5466 int ret = -ENOMEM;
5467
5468 apply_wqattrs_lock();
5469
5470 attrs = wq_sysfs_prep_attrs(wq);
5471 if (!attrs)
5472 goto out_unlock;
5473
5474 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5475 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5476 ret = apply_workqueue_attrs_locked(wq, attrs);
5477 else
5478 ret = -EINVAL;
5479
5480out_unlock:
5481 apply_wqattrs_unlock();
5482 free_workqueue_attrs(attrs);
5483 return ret ?: count;
5484}
5485
5486static ssize_t wq_cpumask_show(struct device *dev,
5487 struct device_attribute *attr, char *buf)
5488{
5489 struct workqueue_struct *wq = dev_to_wq(dev);
5490 int written;
5491
5492 mutex_lock(&wq->mutex);
5493 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5494 cpumask_pr_args(wq->unbound_attrs->cpumask));
5495 mutex_unlock(&wq->mutex);
5496 return written;
5497}
5498
5499static ssize_t wq_cpumask_store(struct device *dev,
5500 struct device_attribute *attr,
5501 const char *buf, size_t count)
5502{
5503 struct workqueue_struct *wq = dev_to_wq(dev);
5504 struct workqueue_attrs *attrs;
5505 int ret = -ENOMEM;
5506
5507 apply_wqattrs_lock();
5508
5509 attrs = wq_sysfs_prep_attrs(wq);
5510 if (!attrs)
5511 goto out_unlock;
5512
5513 ret = cpumask_parse(buf, attrs->cpumask);
5514 if (!ret)
5515 ret = apply_workqueue_attrs_locked(wq, attrs);
5516
5517out_unlock:
5518 apply_wqattrs_unlock();
5519 free_workqueue_attrs(attrs);
5520 return ret ?: count;
5521}
5522
5523static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5524 char *buf)
5525{
5526 struct workqueue_struct *wq = dev_to_wq(dev);
5527 int written;
5528
5529 mutex_lock(&wq->mutex);
5530 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5531 !wq->unbound_attrs->no_numa);
5532 mutex_unlock(&wq->mutex);
5533
5534 return written;
5535}
5536
5537static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5538 const char *buf, size_t count)
5539{
5540 struct workqueue_struct *wq = dev_to_wq(dev);
5541 struct workqueue_attrs *attrs;
5542 int v, ret = -ENOMEM;
5543
5544 apply_wqattrs_lock();
5545
5546 attrs = wq_sysfs_prep_attrs(wq);
5547 if (!attrs)
5548 goto out_unlock;
5549
5550 ret = -EINVAL;
5551 if (sscanf(buf, "%d", &v) == 1) {
5552 attrs->no_numa = !v;
5553 ret = apply_workqueue_attrs_locked(wq, attrs);
5554 }
5555
5556out_unlock:
5557 apply_wqattrs_unlock();
5558 free_workqueue_attrs(attrs);
5559 return ret ?: count;
5560}
5561
5562static struct device_attribute wq_sysfs_unbound_attrs[] = {
5563 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5564 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5565 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5566 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5567 __ATTR_NULL,
5568};
5569
5570static struct bus_type wq_subsys = {
5571 .name = "workqueue",
5572 .dev_groups = wq_sysfs_groups,
5573};
5574
5575static ssize_t wq_unbound_cpumask_show(struct device *dev,
5576 struct device_attribute *attr, char *buf)
5577{
5578 int written;
5579
5580 mutex_lock(&wq_pool_mutex);
5581 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5582 cpumask_pr_args(wq_unbound_cpumask));
5583 mutex_unlock(&wq_pool_mutex);
5584
5585 return written;
5586}
5587
5588static ssize_t wq_unbound_cpumask_store(struct device *dev,
5589 struct device_attribute *attr, const char *buf, size_t count)
5590{
5591 cpumask_var_t cpumask;
5592 int ret;
5593
5594 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5595 return -ENOMEM;
5596
5597 ret = cpumask_parse(buf, cpumask);
5598 if (!ret)
5599 ret = workqueue_set_unbound_cpumask(cpumask);
5600
5601 free_cpumask_var(cpumask);
5602 return ret ? ret : count;
5603}
5604
5605static struct device_attribute wq_sysfs_cpumask_attr =
5606 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5607 wq_unbound_cpumask_store);
5608
5609static int __init wq_sysfs_init(void)
5610{
5611 int err;
5612
5613 err = subsys_virtual_register(&wq_subsys, NULL);
5614 if (err)
5615 return err;
5616
5617 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5618}
5619core_initcall(wq_sysfs_init);
5620
5621static void wq_device_release(struct device *dev)
5622{
5623 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5624
5625 kfree(wq_dev);
5626}
5627
5628/**
5629 * workqueue_sysfs_register - make a workqueue visible in sysfs
5630 * @wq: the workqueue to register
5631 *
5632 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5633 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5634 * which is the preferred method.
5635 *
5636 * Workqueue user should use this function directly iff it wants to apply
5637 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5638 * apply_workqueue_attrs() may race against userland updating the
5639 * attributes.
5640 *
5641 * Return: 0 on success, -errno on failure.
5642 */
5643int workqueue_sysfs_register(struct workqueue_struct *wq)
5644{
5645 struct wq_device *wq_dev;
5646 int ret;
5647
5648 /*
5649 * Adjusting max_active or creating new pwqs by applying
5650 * attributes breaks ordering guarantee. Disallow exposing ordered
5651 * workqueues.
5652 */
5653 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5654 return -EINVAL;
5655
5656 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5657 if (!wq_dev)
5658 return -ENOMEM;
5659
5660 wq_dev->wq = wq;
5661 wq_dev->dev.bus = &wq_subsys;
5662 wq_dev->dev.release = wq_device_release;
5663 dev_set_name(&wq_dev->dev, "%s", wq->name);
5664
5665 /*
5666 * unbound_attrs are created separately. Suppress uevent until
5667 * everything is ready.
5668 */
5669 dev_set_uevent_suppress(&wq_dev->dev, true);
5670
5671 ret = device_register(&wq_dev->dev);
5672 if (ret) {
5673 put_device(&wq_dev->dev);
5674 wq->wq_dev = NULL;
5675 return ret;
5676 }
5677
5678 if (wq->flags & WQ_UNBOUND) {
5679 struct device_attribute *attr;
5680
5681 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5682 ret = device_create_file(&wq_dev->dev, attr);
5683 if (ret) {
5684 device_unregister(&wq_dev->dev);
5685 wq->wq_dev = NULL;
5686 return ret;
5687 }
5688 }
5689 }
5690
5691 dev_set_uevent_suppress(&wq_dev->dev, false);
5692 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5693 return 0;
5694}
5695
5696/**
5697 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5698 * @wq: the workqueue to unregister
5699 *
5700 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5701 */
5702static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5703{
5704 struct wq_device *wq_dev = wq->wq_dev;
5705
5706 if (!wq->wq_dev)
5707 return;
5708
5709 wq->wq_dev = NULL;
5710 device_unregister(&wq_dev->dev);
5711}
5712#else /* CONFIG_SYSFS */
5713static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5714#endif /* CONFIG_SYSFS */
5715
5716/*
5717 * Workqueue watchdog.
5718 *
5719 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5720 * flush dependency, a concurrency managed work item which stays RUNNING
5721 * indefinitely. Workqueue stalls can be very difficult to debug as the
5722 * usual warning mechanisms don't trigger and internal workqueue state is
5723 * largely opaque.
5724 *
5725 * Workqueue watchdog monitors all worker pools periodically and dumps
5726 * state if some pools failed to make forward progress for a while where
5727 * forward progress is defined as the first item on ->worklist changing.
5728 *
5729 * This mechanism is controlled through the kernel parameter
5730 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5731 * corresponding sysfs parameter file.
5732 */
5733#ifdef CONFIG_WQ_WATCHDOG
5734
5735static unsigned long wq_watchdog_thresh = 30;
5736static struct timer_list wq_watchdog_timer;
5737
5738static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5739static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5740
5741static void wq_watchdog_reset_touched(void)
5742{
5743 int cpu;
5744
5745 wq_watchdog_touched = jiffies;
5746 for_each_possible_cpu(cpu)
5747 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5748}
5749
5750static void wq_watchdog_timer_fn(struct timer_list *unused)
5751{
5752 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5753 bool lockup_detected = false;
5754 struct worker_pool *pool;
5755 int pi;
5756
5757 if (!thresh)
5758 return;
5759
5760 rcu_read_lock();
5761
5762 for_each_pool(pool, pi) {
5763 unsigned long pool_ts, touched, ts;
5764
5765 if (list_empty(&pool->worklist))
5766 continue;
5767
5768 /* get the latest of pool and touched timestamps */
5769 pool_ts = READ_ONCE(pool->watchdog_ts);
5770 touched = READ_ONCE(wq_watchdog_touched);
5771
5772 if (time_after(pool_ts, touched))
5773 ts = pool_ts;
5774 else
5775 ts = touched;
5776
5777 if (pool->cpu >= 0) {
5778 unsigned long cpu_touched =
5779 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5780 pool->cpu));
5781 if (time_after(cpu_touched, ts))
5782 ts = cpu_touched;
5783 }
5784
5785 /* did we stall? */
5786 if (time_after(jiffies, ts + thresh)) {
5787 lockup_detected = true;
5788 pr_emerg("BUG: workqueue lockup - pool");
5789 pr_cont_pool_info(pool);
5790 pr_cont(" stuck for %us!\n",
5791 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5792 }
5793 }
5794
5795 rcu_read_unlock();
5796
5797 if (lockup_detected)
5798 show_workqueue_state();
5799
5800 wq_watchdog_reset_touched();
5801 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5802}
5803
5804notrace void wq_watchdog_touch(int cpu)
5805{
5806 if (cpu >= 0)
5807 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5808 else
5809 wq_watchdog_touched = jiffies;
5810}
5811
5812static void wq_watchdog_set_thresh(unsigned long thresh)
5813{
5814 wq_watchdog_thresh = 0;
5815 del_timer_sync(&wq_watchdog_timer);
5816
5817 if (thresh) {
5818 wq_watchdog_thresh = thresh;
5819 wq_watchdog_reset_touched();
5820 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5821 }
5822}
5823
5824static int wq_watchdog_param_set_thresh(const char *val,
5825 const struct kernel_param *kp)
5826{
5827 unsigned long thresh;
5828 int ret;
5829
5830 ret = kstrtoul(val, 0, &thresh);
5831 if (ret)
5832 return ret;
5833
5834 if (system_wq)
5835 wq_watchdog_set_thresh(thresh);
5836 else
5837 wq_watchdog_thresh = thresh;
5838
5839 return 0;
5840}
5841
5842static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5843 .set = wq_watchdog_param_set_thresh,
5844 .get = param_get_ulong,
5845};
5846
5847module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5848 0644);
5849
5850static void wq_watchdog_init(void)
5851{
5852 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5853 wq_watchdog_set_thresh(wq_watchdog_thresh);
5854}
5855
5856#else /* CONFIG_WQ_WATCHDOG */
5857
5858static inline void wq_watchdog_init(void) { }
5859
5860#endif /* CONFIG_WQ_WATCHDOG */
5861
5862static void __init wq_numa_init(void)
5863{
5864 cpumask_var_t *tbl;
5865 int node, cpu;
5866
5867 if (num_possible_nodes() <= 1)
5868 return;
5869
5870 if (wq_disable_numa) {
5871 pr_info("workqueue: NUMA affinity support disabled\n");
5872 return;
5873 }
5874
5875 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5876 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5877
5878 /*
5879 * We want masks of possible CPUs of each node which isn't readily
5880 * available. Build one from cpu_to_node() which should have been
5881 * fully initialized by now.
5882 */
5883 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5884 BUG_ON(!tbl);
5885
5886 for_each_node(node)
5887 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5888 node_online(node) ? node : NUMA_NO_NODE));
5889
5890 for_each_possible_cpu(cpu) {
5891 node = cpu_to_node(cpu);
5892 if (WARN_ON(node == NUMA_NO_NODE)) {
5893 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5894 /* happens iff arch is bonkers, let's just proceed */
5895 return;
5896 }
5897 cpumask_set_cpu(cpu, tbl[node]);
5898 }
5899
5900 wq_numa_possible_cpumask = tbl;
5901 wq_numa_enabled = true;
5902}
5903
5904/**
5905 * workqueue_init_early - early init for workqueue subsystem
5906 *
5907 * This is the first half of two-staged workqueue subsystem initialization
5908 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5909 * idr are up. It sets up all the data structures and system workqueues
5910 * and allows early boot code to create workqueues and queue/cancel work
5911 * items. Actual work item execution starts only after kthreads can be
5912 * created and scheduled right before early initcalls.
5913 */
5914void __init workqueue_init_early(void)
5915{
5916 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5917 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5918 int i, cpu;
5919
5920 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5921
5922 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5923 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5924
5925 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5926
5927 /* initialize CPU pools */
5928 for_each_possible_cpu(cpu) {
5929 struct worker_pool *pool;
5930
5931 i = 0;
5932 for_each_cpu_worker_pool(pool, cpu) {
5933 BUG_ON(init_worker_pool(pool));
5934 pool->cpu = cpu;
5935 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5936 pool->attrs->nice = std_nice[i++];
5937 pool->node = cpu_to_node(cpu);
5938
5939 /* alloc pool ID */
5940 mutex_lock(&wq_pool_mutex);
5941 BUG_ON(worker_pool_assign_id(pool));
5942 mutex_unlock(&wq_pool_mutex);
5943 }
5944 }
5945
5946 /* create default unbound and ordered wq attrs */
5947 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5948 struct workqueue_attrs *attrs;
5949
5950 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5951 attrs->nice = std_nice[i];
5952 unbound_std_wq_attrs[i] = attrs;
5953
5954 /*
5955 * An ordered wq should have only one pwq as ordering is
5956 * guaranteed by max_active which is enforced by pwqs.
5957 * Turn off NUMA so that dfl_pwq is used for all nodes.
5958 */
5959 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5960 attrs->nice = std_nice[i];
5961 attrs->no_numa = true;
5962 ordered_wq_attrs[i] = attrs;
5963 }
5964
5965 system_wq = alloc_workqueue("events", 0, 0);
5966 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5967 system_long_wq = alloc_workqueue("events_long", 0, 0);
5968 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5969 WQ_UNBOUND_MAX_ACTIVE);
5970 system_freezable_wq = alloc_workqueue("events_freezable",
5971 WQ_FREEZABLE, 0);
5972 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5973 WQ_POWER_EFFICIENT, 0);
5974 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5975 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5976 0);
5977 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5978 !system_unbound_wq || !system_freezable_wq ||
5979 !system_power_efficient_wq ||
5980 !system_freezable_power_efficient_wq);
5981}
5982
5983/**
5984 * workqueue_init - bring workqueue subsystem fully online
5985 *
5986 * This is the latter half of two-staged workqueue subsystem initialization
5987 * and invoked as soon as kthreads can be created and scheduled.
5988 * Workqueues have been created and work items queued on them, but there
5989 * are no kworkers executing the work items yet. Populate the worker pools
5990 * with the initial workers and enable future kworker creations.
5991 */
5992void __init workqueue_init(void)
5993{
5994 struct workqueue_struct *wq;
5995 struct worker_pool *pool;
5996 int cpu, bkt;
5997
5998 /*
5999 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6000 * CPU to node mapping may not be available that early on some
6001 * archs such as power and arm64. As per-cpu pools created
6002 * previously could be missing node hint and unbound pools NUMA
6003 * affinity, fix them up.
6004 *
6005 * Also, while iterating workqueues, create rescuers if requested.
6006 */
6007 wq_numa_init();
6008
6009 mutex_lock(&wq_pool_mutex);
6010
6011 for_each_possible_cpu(cpu) {
6012 for_each_cpu_worker_pool(pool, cpu) {
6013 pool->node = cpu_to_node(cpu);
6014 }
6015 }
6016
6017 list_for_each_entry(wq, &workqueues, list) {
6018 wq_update_unbound_numa(wq, smp_processor_id(), true);
6019 WARN(init_rescuer(wq),
6020 "workqueue: failed to create early rescuer for %s",
6021 wq->name);
6022 }
6023
6024 mutex_unlock(&wq_pool_mutex);
6025
6026 /* create the initial workers */
6027 for_each_online_cpu(cpu) {
6028 for_each_cpu_worker_pool(pool, cpu) {
6029 pool->flags &= ~POOL_DISASSOCIATED;
6030 BUG_ON(!create_worker(pool));
6031 }
6032 }
6033
6034 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6035 BUG_ON(!create_worker(pool));
6036
6037 wq_online = true;
6038 wq_watchdog_init();
6039}
1/*
2 * kernel/workqueue.c - generic async execution with shared worker pool
3 *
4 * Copyright (C) 2002 Ingo Molnar
5 *
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
11 *
12 * Made to use alloc_percpu by Christoph Lameter.
13 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
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.
22 *
23 * Please read Documentation/workqueue.txt for details.
24 */
25
26#include <linux/module.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
45#include "workqueue_sched.h"
46
47enum {
48 /* global_cwq flags */
49 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
52 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
53 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
54
55 /* worker flags */
56 WORKER_STARTED = 1 << 0, /* started */
57 WORKER_DIE = 1 << 1, /* die die die */
58 WORKER_IDLE = 1 << 2, /* is idle */
59 WORKER_PREP = 1 << 3, /* preparing to run works */
60 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
61 WORKER_REBIND = 1 << 5, /* mom is home, come back */
62 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
63 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
64
65 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
66 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
67
68 /* gcwq->trustee_state */
69 TRUSTEE_START = 0, /* start */
70 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER = 2, /* butcher workers */
72 TRUSTEE_RELEASE = 3, /* release workers */
73 TRUSTEE_DONE = 4, /* trustee is done */
74
75 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
78
79 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
81
82 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
83 /* call for help after 10ms
84 (min two ticks) */
85 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86 CREATE_COOLDOWN = HZ, /* time to breath after fail */
87 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
88
89 /*
90 * Rescue workers are used only on emergencies and shared by
91 * all cpus. Give -20.
92 */
93 RESCUER_NICE_LEVEL = -20,
94};
95
96/*
97 * Structure fields follow one of the following exclusion rules.
98 *
99 * I: Modifiable by initialization/destruction paths and read-only for
100 * everyone else.
101 *
102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
104 *
105 * L: gcwq->lock protected. Access with gcwq->lock held.
106 *
107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
111 *
112 * F: wq->flush_mutex protected.
113 *
114 * W: workqueue_lock protected.
115 */
116
117struct global_cwq;
118
119/*
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
122 */
123struct worker {
124 /* on idle list while idle, on busy hash table while busy */
125 union {
126 struct list_head entry; /* L: while idle */
127 struct hlist_node hentry; /* L: while busy */
128 };
129
130 struct work_struct *current_work; /* L: work being processed */
131 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
132 struct list_head scheduled; /* L: scheduled works */
133 struct task_struct *task; /* I: worker task */
134 struct global_cwq *gcwq; /* I: the associated gcwq */
135 /* 64 bytes boundary on 64bit, 32 on 32bit */
136 unsigned long last_active; /* L: last active timestamp */
137 unsigned int flags; /* X: flags */
138 int id; /* I: worker id */
139 struct work_struct rebind_work; /* L: rebind worker to cpu */
140};
141
142/*
143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
145 * target workqueues.
146 */
147struct global_cwq {
148 spinlock_t lock; /* the gcwq lock */
149 struct list_head worklist; /* L: list of pending works */
150 unsigned int cpu; /* I: the associated cpu */
151 unsigned int flags; /* L: GCWQ_* flags */
152
153 int nr_workers; /* L: total number of workers */
154 int nr_idle; /* L: currently idle ones */
155
156 /* workers are chained either in the idle_list or busy_hash */
157 struct list_head idle_list; /* X: list of idle workers */
158 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159 /* L: hash of busy workers */
160
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
163
164 struct ida worker_ida; /* L: for worker IDs */
165
166 struct task_struct *trustee; /* L: for gcwq shutdown */
167 unsigned int trustee_state; /* L: trustee state */
168 wait_queue_head_t trustee_wait; /* trustee wait */
169 struct worker *first_idle; /* L: first idle worker */
170} ____cacheline_aligned_in_smp;
171
172/*
173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
176 */
177struct cpu_workqueue_struct {
178 struct global_cwq *gcwq; /* I: the associated gcwq */
179 struct workqueue_struct *wq; /* I: the owning workqueue */
180 int work_color; /* L: current color */
181 int flush_color; /* L: flushing color */
182 int nr_in_flight[WORK_NR_COLORS];
183 /* L: nr of in_flight works */
184 int nr_active; /* L: nr of active works */
185 int max_active; /* L: max active works */
186 struct list_head delayed_works; /* L: delayed works */
187};
188
189/*
190 * Structure used to wait for workqueue flush.
191 */
192struct wq_flusher {
193 struct list_head list; /* F: list of flushers */
194 int flush_color; /* F: flush color waiting for */
195 struct completion done; /* flush completion */
196};
197
198/*
199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
201 */
202#ifdef CONFIG_SMP
203typedef cpumask_var_t mayday_mask_t;
204#define mayday_test_and_set_cpu(cpu, mask) \
205 cpumask_test_and_set_cpu((cpu), (mask))
206#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209#define free_mayday_mask(mask) free_cpumask_var((mask))
210#else
211typedef unsigned long mayday_mask_t;
212#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215#define alloc_mayday_mask(maskp, gfp) true
216#define free_mayday_mask(mask) do { } while (0)
217#endif
218
219/*
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
222 */
223struct workqueue_struct {
224 unsigned int flags; /* W: WQ_* flags */
225 union {
226 struct cpu_workqueue_struct __percpu *pcpu;
227 struct cpu_workqueue_struct *single;
228 unsigned long v;
229 } cpu_wq; /* I: cwq's */
230 struct list_head list; /* W: list of all workqueues */
231
232 struct mutex flush_mutex; /* protects wq flushing */
233 int work_color; /* F: current work color */
234 int flush_color; /* F: current flush color */
235 atomic_t nr_cwqs_to_flush; /* flush in progress */
236 struct wq_flusher *first_flusher; /* F: first flusher */
237 struct list_head flusher_queue; /* F: flush waiters */
238 struct list_head flusher_overflow; /* F: flush overflow list */
239
240 mayday_mask_t mayday_mask; /* cpus requesting rescue */
241 struct worker *rescuer; /* I: rescue worker */
242
243 int nr_drainers; /* W: drain in progress */
244 int saved_max_active; /* W: saved cwq max_active */
245 const char *name; /* I: workqueue name */
246#ifdef CONFIG_LOCKDEP
247 struct lockdep_map lockdep_map;
248#endif
249};
250
251struct workqueue_struct *system_wq __read_mostly;
252struct workqueue_struct *system_long_wq __read_mostly;
253struct workqueue_struct *system_nrt_wq __read_mostly;
254struct workqueue_struct *system_unbound_wq __read_mostly;
255struct workqueue_struct *system_freezable_wq __read_mostly;
256EXPORT_SYMBOL_GPL(system_wq);
257EXPORT_SYMBOL_GPL(system_long_wq);
258EXPORT_SYMBOL_GPL(system_nrt_wq);
259EXPORT_SYMBOL_GPL(system_unbound_wq);
260EXPORT_SYMBOL_GPL(system_freezable_wq);
261
262#define CREATE_TRACE_POINTS
263#include <trace/events/workqueue.h>
264
265#define for_each_busy_worker(worker, i, pos, gcwq) \
266 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
267 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
268
269static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
270 unsigned int sw)
271{
272 if (cpu < nr_cpu_ids) {
273 if (sw & 1) {
274 cpu = cpumask_next(cpu, mask);
275 if (cpu < nr_cpu_ids)
276 return cpu;
277 }
278 if (sw & 2)
279 return WORK_CPU_UNBOUND;
280 }
281 return WORK_CPU_NONE;
282}
283
284static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
285 struct workqueue_struct *wq)
286{
287 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
288}
289
290/*
291 * CPU iterators
292 *
293 * An extra gcwq is defined for an invalid cpu number
294 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
295 * specific CPU. The following iterators are similar to
296 * for_each_*_cpu() iterators but also considers the unbound gcwq.
297 *
298 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
299 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
300 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
301 * WORK_CPU_UNBOUND for unbound workqueues
302 */
303#define for_each_gcwq_cpu(cpu) \
304 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
305 (cpu) < WORK_CPU_NONE; \
306 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
307
308#define for_each_online_gcwq_cpu(cpu) \
309 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
310 (cpu) < WORK_CPU_NONE; \
311 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
312
313#define for_each_cwq_cpu(cpu, wq) \
314 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
315 (cpu) < WORK_CPU_NONE; \
316 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
317
318#ifdef CONFIG_DEBUG_OBJECTS_WORK
319
320static struct debug_obj_descr work_debug_descr;
321
322static void *work_debug_hint(void *addr)
323{
324 return ((struct work_struct *) addr)->func;
325}
326
327/*
328 * fixup_init is called when:
329 * - an active object is initialized
330 */
331static int work_fixup_init(void *addr, enum debug_obj_state state)
332{
333 struct work_struct *work = addr;
334
335 switch (state) {
336 case ODEBUG_STATE_ACTIVE:
337 cancel_work_sync(work);
338 debug_object_init(work, &work_debug_descr);
339 return 1;
340 default:
341 return 0;
342 }
343}
344
345/*
346 * fixup_activate is called when:
347 * - an active object is activated
348 * - an unknown object is activated (might be a statically initialized object)
349 */
350static int work_fixup_activate(void *addr, enum debug_obj_state state)
351{
352 struct work_struct *work = addr;
353
354 switch (state) {
355
356 case ODEBUG_STATE_NOTAVAILABLE:
357 /*
358 * This is not really a fixup. The work struct was
359 * statically initialized. We just make sure that it
360 * is tracked in the object tracker.
361 */
362 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
363 debug_object_init(work, &work_debug_descr);
364 debug_object_activate(work, &work_debug_descr);
365 return 0;
366 }
367 WARN_ON_ONCE(1);
368 return 0;
369
370 case ODEBUG_STATE_ACTIVE:
371 WARN_ON(1);
372
373 default:
374 return 0;
375 }
376}
377
378/*
379 * fixup_free is called when:
380 * - an active object is freed
381 */
382static int work_fixup_free(void *addr, enum debug_obj_state state)
383{
384 struct work_struct *work = addr;
385
386 switch (state) {
387 case ODEBUG_STATE_ACTIVE:
388 cancel_work_sync(work);
389 debug_object_free(work, &work_debug_descr);
390 return 1;
391 default:
392 return 0;
393 }
394}
395
396static struct debug_obj_descr work_debug_descr = {
397 .name = "work_struct",
398 .debug_hint = work_debug_hint,
399 .fixup_init = work_fixup_init,
400 .fixup_activate = work_fixup_activate,
401 .fixup_free = work_fixup_free,
402};
403
404static inline void debug_work_activate(struct work_struct *work)
405{
406 debug_object_activate(work, &work_debug_descr);
407}
408
409static inline void debug_work_deactivate(struct work_struct *work)
410{
411 debug_object_deactivate(work, &work_debug_descr);
412}
413
414void __init_work(struct work_struct *work, int onstack)
415{
416 if (onstack)
417 debug_object_init_on_stack(work, &work_debug_descr);
418 else
419 debug_object_init(work, &work_debug_descr);
420}
421EXPORT_SYMBOL_GPL(__init_work);
422
423void destroy_work_on_stack(struct work_struct *work)
424{
425 debug_object_free(work, &work_debug_descr);
426}
427EXPORT_SYMBOL_GPL(destroy_work_on_stack);
428
429#else
430static inline void debug_work_activate(struct work_struct *work) { }
431static inline void debug_work_deactivate(struct work_struct *work) { }
432#endif
433
434/* Serializes the accesses to the list of workqueues. */
435static DEFINE_SPINLOCK(workqueue_lock);
436static LIST_HEAD(workqueues);
437static bool workqueue_freezing; /* W: have wqs started freezing? */
438
439/*
440 * The almighty global cpu workqueues. nr_running is the only field
441 * which is expected to be used frequently by other cpus via
442 * try_to_wake_up(). Put it in a separate cacheline.
443 */
444static DEFINE_PER_CPU(struct global_cwq, global_cwq);
445static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
446
447/*
448 * Global cpu workqueue and nr_running counter for unbound gcwq. The
449 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
450 * workers have WORKER_UNBOUND set.
451 */
452static struct global_cwq unbound_global_cwq;
453static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
454
455static int worker_thread(void *__worker);
456
457static struct global_cwq *get_gcwq(unsigned int cpu)
458{
459 if (cpu != WORK_CPU_UNBOUND)
460 return &per_cpu(global_cwq, cpu);
461 else
462 return &unbound_global_cwq;
463}
464
465static atomic_t *get_gcwq_nr_running(unsigned int cpu)
466{
467 if (cpu != WORK_CPU_UNBOUND)
468 return &per_cpu(gcwq_nr_running, cpu);
469 else
470 return &unbound_gcwq_nr_running;
471}
472
473static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
474 struct workqueue_struct *wq)
475{
476 if (!(wq->flags & WQ_UNBOUND)) {
477 if (likely(cpu < nr_cpu_ids)) {
478#ifdef CONFIG_SMP
479 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
480#else
481 return wq->cpu_wq.single;
482#endif
483 }
484 } else if (likely(cpu == WORK_CPU_UNBOUND))
485 return wq->cpu_wq.single;
486 return NULL;
487}
488
489static unsigned int work_color_to_flags(int color)
490{
491 return color << WORK_STRUCT_COLOR_SHIFT;
492}
493
494static int get_work_color(struct work_struct *work)
495{
496 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
497 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
498}
499
500static int work_next_color(int color)
501{
502 return (color + 1) % WORK_NR_COLORS;
503}
504
505/*
506 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
507 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
508 * cleared and the work data contains the cpu number it was last on.
509 *
510 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
511 * cwq, cpu or clear work->data. These functions should only be
512 * called while the work is owned - ie. while the PENDING bit is set.
513 *
514 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
515 * corresponding to a work. gcwq is available once the work has been
516 * queued anywhere after initialization. cwq is available only from
517 * queueing until execution starts.
518 */
519static inline void set_work_data(struct work_struct *work, unsigned long data,
520 unsigned long flags)
521{
522 BUG_ON(!work_pending(work));
523 atomic_long_set(&work->data, data | flags | work_static(work));
524}
525
526static void set_work_cwq(struct work_struct *work,
527 struct cpu_workqueue_struct *cwq,
528 unsigned long extra_flags)
529{
530 set_work_data(work, (unsigned long)cwq,
531 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
532}
533
534static void set_work_cpu(struct work_struct *work, unsigned int cpu)
535{
536 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
537}
538
539static void clear_work_data(struct work_struct *work)
540{
541 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
542}
543
544static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
545{
546 unsigned long data = atomic_long_read(&work->data);
547
548 if (data & WORK_STRUCT_CWQ)
549 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
550 else
551 return NULL;
552}
553
554static struct global_cwq *get_work_gcwq(struct work_struct *work)
555{
556 unsigned long data = atomic_long_read(&work->data);
557 unsigned int cpu;
558
559 if (data & WORK_STRUCT_CWQ)
560 return ((struct cpu_workqueue_struct *)
561 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
562
563 cpu = data >> WORK_STRUCT_FLAG_BITS;
564 if (cpu == WORK_CPU_NONE)
565 return NULL;
566
567 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
568 return get_gcwq(cpu);
569}
570
571/*
572 * Policy functions. These define the policies on how the global
573 * worker pool is managed. Unless noted otherwise, these functions
574 * assume that they're being called with gcwq->lock held.
575 */
576
577static bool __need_more_worker(struct global_cwq *gcwq)
578{
579 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
580 gcwq->flags & GCWQ_HIGHPRI_PENDING;
581}
582
583/*
584 * Need to wake up a worker? Called from anything but currently
585 * running workers.
586 */
587static bool need_more_worker(struct global_cwq *gcwq)
588{
589 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
590}
591
592/* Can I start working? Called from busy but !running workers. */
593static bool may_start_working(struct global_cwq *gcwq)
594{
595 return gcwq->nr_idle;
596}
597
598/* Do I need to keep working? Called from currently running workers. */
599static bool keep_working(struct global_cwq *gcwq)
600{
601 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
602
603 return !list_empty(&gcwq->worklist) &&
604 (atomic_read(nr_running) <= 1 ||
605 gcwq->flags & GCWQ_HIGHPRI_PENDING);
606}
607
608/* Do we need a new worker? Called from manager. */
609static bool need_to_create_worker(struct global_cwq *gcwq)
610{
611 return need_more_worker(gcwq) && !may_start_working(gcwq);
612}
613
614/* Do I need to be the manager? */
615static bool need_to_manage_workers(struct global_cwq *gcwq)
616{
617 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
618}
619
620/* Do we have too many workers and should some go away? */
621static bool too_many_workers(struct global_cwq *gcwq)
622{
623 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
624 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
625 int nr_busy = gcwq->nr_workers - nr_idle;
626
627 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
628}
629
630/*
631 * Wake up functions.
632 */
633
634/* Return the first worker. Safe with preemption disabled */
635static struct worker *first_worker(struct global_cwq *gcwq)
636{
637 if (unlikely(list_empty(&gcwq->idle_list)))
638 return NULL;
639
640 return list_first_entry(&gcwq->idle_list, struct worker, entry);
641}
642
643/**
644 * wake_up_worker - wake up an idle worker
645 * @gcwq: gcwq to wake worker for
646 *
647 * Wake up the first idle worker of @gcwq.
648 *
649 * CONTEXT:
650 * spin_lock_irq(gcwq->lock).
651 */
652static void wake_up_worker(struct global_cwq *gcwq)
653{
654 struct worker *worker = first_worker(gcwq);
655
656 if (likely(worker))
657 wake_up_process(worker->task);
658}
659
660/**
661 * wq_worker_waking_up - a worker is waking up
662 * @task: task waking up
663 * @cpu: CPU @task is waking up to
664 *
665 * This function is called during try_to_wake_up() when a worker is
666 * being awoken.
667 *
668 * CONTEXT:
669 * spin_lock_irq(rq->lock)
670 */
671void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
672{
673 struct worker *worker = kthread_data(task);
674
675 if (!(worker->flags & WORKER_NOT_RUNNING))
676 atomic_inc(get_gcwq_nr_running(cpu));
677}
678
679/**
680 * wq_worker_sleeping - a worker is going to sleep
681 * @task: task going to sleep
682 * @cpu: CPU in question, must be the current CPU number
683 *
684 * This function is called during schedule() when a busy worker is
685 * going to sleep. Worker on the same cpu can be woken up by
686 * returning pointer to its task.
687 *
688 * CONTEXT:
689 * spin_lock_irq(rq->lock)
690 *
691 * RETURNS:
692 * Worker task on @cpu to wake up, %NULL if none.
693 */
694struct task_struct *wq_worker_sleeping(struct task_struct *task,
695 unsigned int cpu)
696{
697 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
698 struct global_cwq *gcwq = get_gcwq(cpu);
699 atomic_t *nr_running = get_gcwq_nr_running(cpu);
700
701 if (worker->flags & WORKER_NOT_RUNNING)
702 return NULL;
703
704 /* this can only happen on the local cpu */
705 BUG_ON(cpu != raw_smp_processor_id());
706
707 /*
708 * The counterpart of the following dec_and_test, implied mb,
709 * worklist not empty test sequence is in insert_work().
710 * Please read comment there.
711 *
712 * NOT_RUNNING is clear. This means that trustee is not in
713 * charge and we're running on the local cpu w/ rq lock held
714 * and preemption disabled, which in turn means that none else
715 * could be manipulating idle_list, so dereferencing idle_list
716 * without gcwq lock is safe.
717 */
718 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
719 to_wakeup = first_worker(gcwq);
720 return to_wakeup ? to_wakeup->task : NULL;
721}
722
723/**
724 * worker_set_flags - set worker flags and adjust nr_running accordingly
725 * @worker: self
726 * @flags: flags to set
727 * @wakeup: wakeup an idle worker if necessary
728 *
729 * Set @flags in @worker->flags and adjust nr_running accordingly. If
730 * nr_running becomes zero and @wakeup is %true, an idle worker is
731 * woken up.
732 *
733 * CONTEXT:
734 * spin_lock_irq(gcwq->lock)
735 */
736static inline void worker_set_flags(struct worker *worker, unsigned int flags,
737 bool wakeup)
738{
739 struct global_cwq *gcwq = worker->gcwq;
740
741 WARN_ON_ONCE(worker->task != current);
742
743 /*
744 * If transitioning into NOT_RUNNING, adjust nr_running and
745 * wake up an idle worker as necessary if requested by
746 * @wakeup.
747 */
748 if ((flags & WORKER_NOT_RUNNING) &&
749 !(worker->flags & WORKER_NOT_RUNNING)) {
750 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
751
752 if (wakeup) {
753 if (atomic_dec_and_test(nr_running) &&
754 !list_empty(&gcwq->worklist))
755 wake_up_worker(gcwq);
756 } else
757 atomic_dec(nr_running);
758 }
759
760 worker->flags |= flags;
761}
762
763/**
764 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
765 * @worker: self
766 * @flags: flags to clear
767 *
768 * Clear @flags in @worker->flags and adjust nr_running accordingly.
769 *
770 * CONTEXT:
771 * spin_lock_irq(gcwq->lock)
772 */
773static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
774{
775 struct global_cwq *gcwq = worker->gcwq;
776 unsigned int oflags = worker->flags;
777
778 WARN_ON_ONCE(worker->task != current);
779
780 worker->flags &= ~flags;
781
782 /*
783 * If transitioning out of NOT_RUNNING, increment nr_running. Note
784 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
785 * of multiple flags, not a single flag.
786 */
787 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
788 if (!(worker->flags & WORKER_NOT_RUNNING))
789 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
790}
791
792/**
793 * busy_worker_head - return the busy hash head for a work
794 * @gcwq: gcwq of interest
795 * @work: work to be hashed
796 *
797 * Return hash head of @gcwq for @work.
798 *
799 * CONTEXT:
800 * spin_lock_irq(gcwq->lock).
801 *
802 * RETURNS:
803 * Pointer to the hash head.
804 */
805static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
806 struct work_struct *work)
807{
808 const int base_shift = ilog2(sizeof(struct work_struct));
809 unsigned long v = (unsigned long)work;
810
811 /* simple shift and fold hash, do we need something better? */
812 v >>= base_shift;
813 v += v >> BUSY_WORKER_HASH_ORDER;
814 v &= BUSY_WORKER_HASH_MASK;
815
816 return &gcwq->busy_hash[v];
817}
818
819/**
820 * __find_worker_executing_work - find worker which is executing a work
821 * @gcwq: gcwq of interest
822 * @bwh: hash head as returned by busy_worker_head()
823 * @work: work to find worker for
824 *
825 * Find a worker which is executing @work on @gcwq. @bwh should be
826 * the hash head obtained by calling busy_worker_head() with the same
827 * work.
828 *
829 * CONTEXT:
830 * spin_lock_irq(gcwq->lock).
831 *
832 * RETURNS:
833 * Pointer to worker which is executing @work if found, NULL
834 * otherwise.
835 */
836static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
837 struct hlist_head *bwh,
838 struct work_struct *work)
839{
840 struct worker *worker;
841 struct hlist_node *tmp;
842
843 hlist_for_each_entry(worker, tmp, bwh, hentry)
844 if (worker->current_work == work)
845 return worker;
846 return NULL;
847}
848
849/**
850 * find_worker_executing_work - find worker which is executing a work
851 * @gcwq: gcwq of interest
852 * @work: work to find worker for
853 *
854 * Find a worker which is executing @work on @gcwq. This function is
855 * identical to __find_worker_executing_work() except that this
856 * function calculates @bwh itself.
857 *
858 * CONTEXT:
859 * spin_lock_irq(gcwq->lock).
860 *
861 * RETURNS:
862 * Pointer to worker which is executing @work if found, NULL
863 * otherwise.
864 */
865static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
866 struct work_struct *work)
867{
868 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
869 work);
870}
871
872/**
873 * gcwq_determine_ins_pos - find insertion position
874 * @gcwq: gcwq of interest
875 * @cwq: cwq a work is being queued for
876 *
877 * A work for @cwq is about to be queued on @gcwq, determine insertion
878 * position for the work. If @cwq is for HIGHPRI wq, the work is
879 * queued at the head of the queue but in FIFO order with respect to
880 * other HIGHPRI works; otherwise, at the end of the queue. This
881 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
882 * there are HIGHPRI works pending.
883 *
884 * CONTEXT:
885 * spin_lock_irq(gcwq->lock).
886 *
887 * RETURNS:
888 * Pointer to inserstion position.
889 */
890static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
891 struct cpu_workqueue_struct *cwq)
892{
893 struct work_struct *twork;
894
895 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
896 return &gcwq->worklist;
897
898 list_for_each_entry(twork, &gcwq->worklist, entry) {
899 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
900
901 if (!(tcwq->wq->flags & WQ_HIGHPRI))
902 break;
903 }
904
905 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
906 return &twork->entry;
907}
908
909/**
910 * insert_work - insert a work into gcwq
911 * @cwq: cwq @work belongs to
912 * @work: work to insert
913 * @head: insertion point
914 * @extra_flags: extra WORK_STRUCT_* flags to set
915 *
916 * Insert @work which belongs to @cwq into @gcwq after @head.
917 * @extra_flags is or'd to work_struct flags.
918 *
919 * CONTEXT:
920 * spin_lock_irq(gcwq->lock).
921 */
922static void insert_work(struct cpu_workqueue_struct *cwq,
923 struct work_struct *work, struct list_head *head,
924 unsigned int extra_flags)
925{
926 struct global_cwq *gcwq = cwq->gcwq;
927
928 /* we own @work, set data and link */
929 set_work_cwq(work, cwq, extra_flags);
930
931 /*
932 * Ensure that we get the right work->data if we see the
933 * result of list_add() below, see try_to_grab_pending().
934 */
935 smp_wmb();
936
937 list_add_tail(&work->entry, head);
938
939 /*
940 * Ensure either worker_sched_deactivated() sees the above
941 * list_add_tail() or we see zero nr_running to avoid workers
942 * lying around lazily while there are works to be processed.
943 */
944 smp_mb();
945
946 if (__need_more_worker(gcwq))
947 wake_up_worker(gcwq);
948}
949
950/*
951 * Test whether @work is being queued from another work executing on the
952 * same workqueue. This is rather expensive and should only be used from
953 * cold paths.
954 */
955static bool is_chained_work(struct workqueue_struct *wq)
956{
957 unsigned long flags;
958 unsigned int cpu;
959
960 for_each_gcwq_cpu(cpu) {
961 struct global_cwq *gcwq = get_gcwq(cpu);
962 struct worker *worker;
963 struct hlist_node *pos;
964 int i;
965
966 spin_lock_irqsave(&gcwq->lock, flags);
967 for_each_busy_worker(worker, i, pos, gcwq) {
968 if (worker->task != current)
969 continue;
970 spin_unlock_irqrestore(&gcwq->lock, flags);
971 /*
972 * I'm @worker, no locking necessary. See if @work
973 * is headed to the same workqueue.
974 */
975 return worker->current_cwq->wq == wq;
976 }
977 spin_unlock_irqrestore(&gcwq->lock, flags);
978 }
979 return false;
980}
981
982static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
983 struct work_struct *work)
984{
985 struct global_cwq *gcwq;
986 struct cpu_workqueue_struct *cwq;
987 struct list_head *worklist;
988 unsigned int work_flags;
989 unsigned long flags;
990
991 debug_work_activate(work);
992
993 /* if dying, only works from the same workqueue are allowed */
994 if (unlikely(wq->flags & WQ_DRAINING) &&
995 WARN_ON_ONCE(!is_chained_work(wq)))
996 return;
997
998 /* determine gcwq to use */
999 if (!(wq->flags & WQ_UNBOUND)) {
1000 struct global_cwq *last_gcwq;
1001
1002 if (unlikely(cpu == WORK_CPU_UNBOUND))
1003 cpu = raw_smp_processor_id();
1004
1005 /*
1006 * It's multi cpu. If @wq is non-reentrant and @work
1007 * was previously on a different cpu, it might still
1008 * be running there, in which case the work needs to
1009 * be queued on that cpu to guarantee non-reentrance.
1010 */
1011 gcwq = get_gcwq(cpu);
1012 if (wq->flags & WQ_NON_REENTRANT &&
1013 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1014 struct worker *worker;
1015
1016 spin_lock_irqsave(&last_gcwq->lock, flags);
1017
1018 worker = find_worker_executing_work(last_gcwq, work);
1019
1020 if (worker && worker->current_cwq->wq == wq)
1021 gcwq = last_gcwq;
1022 else {
1023 /* meh... not running there, queue here */
1024 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1025 spin_lock_irqsave(&gcwq->lock, flags);
1026 }
1027 } else
1028 spin_lock_irqsave(&gcwq->lock, flags);
1029 } else {
1030 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1031 spin_lock_irqsave(&gcwq->lock, flags);
1032 }
1033
1034 /* gcwq determined, get cwq and queue */
1035 cwq = get_cwq(gcwq->cpu, wq);
1036 trace_workqueue_queue_work(cpu, cwq, work);
1037
1038 BUG_ON(!list_empty(&work->entry));
1039
1040 cwq->nr_in_flight[cwq->work_color]++;
1041 work_flags = work_color_to_flags(cwq->work_color);
1042
1043 if (likely(cwq->nr_active < cwq->max_active)) {
1044 trace_workqueue_activate_work(work);
1045 cwq->nr_active++;
1046 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1047 } else {
1048 work_flags |= WORK_STRUCT_DELAYED;
1049 worklist = &cwq->delayed_works;
1050 }
1051
1052 insert_work(cwq, work, worklist, work_flags);
1053
1054 spin_unlock_irqrestore(&gcwq->lock, flags);
1055}
1056
1057/**
1058 * queue_work - queue work on a workqueue
1059 * @wq: workqueue to use
1060 * @work: work to queue
1061 *
1062 * Returns 0 if @work was already on a queue, non-zero otherwise.
1063 *
1064 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1065 * it can be processed by another CPU.
1066 */
1067int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1068{
1069 int ret;
1070
1071 ret = queue_work_on(get_cpu(), wq, work);
1072 put_cpu();
1073
1074 return ret;
1075}
1076EXPORT_SYMBOL_GPL(queue_work);
1077
1078/**
1079 * queue_work_on - queue work on specific cpu
1080 * @cpu: CPU number to execute work on
1081 * @wq: workqueue to use
1082 * @work: work to queue
1083 *
1084 * Returns 0 if @work was already on a queue, non-zero otherwise.
1085 *
1086 * We queue the work to a specific CPU, the caller must ensure it
1087 * can't go away.
1088 */
1089int
1090queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1091{
1092 int ret = 0;
1093
1094 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1095 __queue_work(cpu, wq, work);
1096 ret = 1;
1097 }
1098 return ret;
1099}
1100EXPORT_SYMBOL_GPL(queue_work_on);
1101
1102static void delayed_work_timer_fn(unsigned long __data)
1103{
1104 struct delayed_work *dwork = (struct delayed_work *)__data;
1105 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1106
1107 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1108}
1109
1110/**
1111 * queue_delayed_work - queue work on a workqueue after delay
1112 * @wq: workqueue to use
1113 * @dwork: delayable work to queue
1114 * @delay: number of jiffies to wait before queueing
1115 *
1116 * Returns 0 if @work was already on a queue, non-zero otherwise.
1117 */
1118int queue_delayed_work(struct workqueue_struct *wq,
1119 struct delayed_work *dwork, unsigned long delay)
1120{
1121 if (delay == 0)
1122 return queue_work(wq, &dwork->work);
1123
1124 return queue_delayed_work_on(-1, wq, dwork, delay);
1125}
1126EXPORT_SYMBOL_GPL(queue_delayed_work);
1127
1128/**
1129 * queue_delayed_work_on - queue work on specific CPU after delay
1130 * @cpu: CPU number to execute work on
1131 * @wq: workqueue to use
1132 * @dwork: work to queue
1133 * @delay: number of jiffies to wait before queueing
1134 *
1135 * Returns 0 if @work was already on a queue, non-zero otherwise.
1136 */
1137int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1138 struct delayed_work *dwork, unsigned long delay)
1139{
1140 int ret = 0;
1141 struct timer_list *timer = &dwork->timer;
1142 struct work_struct *work = &dwork->work;
1143
1144 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1145 unsigned int lcpu;
1146
1147 BUG_ON(timer_pending(timer));
1148 BUG_ON(!list_empty(&work->entry));
1149
1150 timer_stats_timer_set_start_info(&dwork->timer);
1151
1152 /*
1153 * This stores cwq for the moment, for the timer_fn.
1154 * Note that the work's gcwq is preserved to allow
1155 * reentrance detection for delayed works.
1156 */
1157 if (!(wq->flags & WQ_UNBOUND)) {
1158 struct global_cwq *gcwq = get_work_gcwq(work);
1159
1160 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1161 lcpu = gcwq->cpu;
1162 else
1163 lcpu = raw_smp_processor_id();
1164 } else
1165 lcpu = WORK_CPU_UNBOUND;
1166
1167 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1168
1169 timer->expires = jiffies + delay;
1170 timer->data = (unsigned long)dwork;
1171 timer->function = delayed_work_timer_fn;
1172
1173 if (unlikely(cpu >= 0))
1174 add_timer_on(timer, cpu);
1175 else
1176 add_timer(timer);
1177 ret = 1;
1178 }
1179 return ret;
1180}
1181EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1182
1183/**
1184 * worker_enter_idle - enter idle state
1185 * @worker: worker which is entering idle state
1186 *
1187 * @worker is entering idle state. Update stats and idle timer if
1188 * necessary.
1189 *
1190 * LOCKING:
1191 * spin_lock_irq(gcwq->lock).
1192 */
1193static void worker_enter_idle(struct worker *worker)
1194{
1195 struct global_cwq *gcwq = worker->gcwq;
1196
1197 BUG_ON(worker->flags & WORKER_IDLE);
1198 BUG_ON(!list_empty(&worker->entry) &&
1199 (worker->hentry.next || worker->hentry.pprev));
1200
1201 /* can't use worker_set_flags(), also called from start_worker() */
1202 worker->flags |= WORKER_IDLE;
1203 gcwq->nr_idle++;
1204 worker->last_active = jiffies;
1205
1206 /* idle_list is LIFO */
1207 list_add(&worker->entry, &gcwq->idle_list);
1208
1209 if (likely(!(worker->flags & WORKER_ROGUE))) {
1210 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1211 mod_timer(&gcwq->idle_timer,
1212 jiffies + IDLE_WORKER_TIMEOUT);
1213 } else
1214 wake_up_all(&gcwq->trustee_wait);
1215
1216 /* sanity check nr_running */
1217 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1218 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1219}
1220
1221/**
1222 * worker_leave_idle - leave idle state
1223 * @worker: worker which is leaving idle state
1224 *
1225 * @worker is leaving idle state. Update stats.
1226 *
1227 * LOCKING:
1228 * spin_lock_irq(gcwq->lock).
1229 */
1230static void worker_leave_idle(struct worker *worker)
1231{
1232 struct global_cwq *gcwq = worker->gcwq;
1233
1234 BUG_ON(!(worker->flags & WORKER_IDLE));
1235 worker_clr_flags(worker, WORKER_IDLE);
1236 gcwq->nr_idle--;
1237 list_del_init(&worker->entry);
1238}
1239
1240/**
1241 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1242 * @worker: self
1243 *
1244 * Works which are scheduled while the cpu is online must at least be
1245 * scheduled to a worker which is bound to the cpu so that if they are
1246 * flushed from cpu callbacks while cpu is going down, they are
1247 * guaranteed to execute on the cpu.
1248 *
1249 * This function is to be used by rogue workers and rescuers to bind
1250 * themselves to the target cpu and may race with cpu going down or
1251 * coming online. kthread_bind() can't be used because it may put the
1252 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1253 * verbatim as it's best effort and blocking and gcwq may be
1254 * [dis]associated in the meantime.
1255 *
1256 * This function tries set_cpus_allowed() and locks gcwq and verifies
1257 * the binding against GCWQ_DISASSOCIATED which is set during
1258 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1259 * idle state or fetches works without dropping lock, it can guarantee
1260 * the scheduling requirement described in the first paragraph.
1261 *
1262 * CONTEXT:
1263 * Might sleep. Called without any lock but returns with gcwq->lock
1264 * held.
1265 *
1266 * RETURNS:
1267 * %true if the associated gcwq is online (@worker is successfully
1268 * bound), %false if offline.
1269 */
1270static bool worker_maybe_bind_and_lock(struct worker *worker)
1271__acquires(&gcwq->lock)
1272{
1273 struct global_cwq *gcwq = worker->gcwq;
1274 struct task_struct *task = worker->task;
1275
1276 while (true) {
1277 /*
1278 * The following call may fail, succeed or succeed
1279 * without actually migrating the task to the cpu if
1280 * it races with cpu hotunplug operation. Verify
1281 * against GCWQ_DISASSOCIATED.
1282 */
1283 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1284 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1285
1286 spin_lock_irq(&gcwq->lock);
1287 if (gcwq->flags & GCWQ_DISASSOCIATED)
1288 return false;
1289 if (task_cpu(task) == gcwq->cpu &&
1290 cpumask_equal(¤t->cpus_allowed,
1291 get_cpu_mask(gcwq->cpu)))
1292 return true;
1293 spin_unlock_irq(&gcwq->lock);
1294
1295 /*
1296 * We've raced with CPU hot[un]plug. Give it a breather
1297 * and retry migration. cond_resched() is required here;
1298 * otherwise, we might deadlock against cpu_stop trying to
1299 * bring down the CPU on non-preemptive kernel.
1300 */
1301 cpu_relax();
1302 cond_resched();
1303 }
1304}
1305
1306/*
1307 * Function for worker->rebind_work used to rebind rogue busy workers
1308 * to the associated cpu which is coming back online. This is
1309 * scheduled by cpu up but can race with other cpu hotplug operations
1310 * and may be executed twice without intervening cpu down.
1311 */
1312static void worker_rebind_fn(struct work_struct *work)
1313{
1314 struct worker *worker = container_of(work, struct worker, rebind_work);
1315 struct global_cwq *gcwq = worker->gcwq;
1316
1317 if (worker_maybe_bind_and_lock(worker))
1318 worker_clr_flags(worker, WORKER_REBIND);
1319
1320 spin_unlock_irq(&gcwq->lock);
1321}
1322
1323static struct worker *alloc_worker(void)
1324{
1325 struct worker *worker;
1326
1327 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1328 if (worker) {
1329 INIT_LIST_HEAD(&worker->entry);
1330 INIT_LIST_HEAD(&worker->scheduled);
1331 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1332 /* on creation a worker is in !idle && prep state */
1333 worker->flags = WORKER_PREP;
1334 }
1335 return worker;
1336}
1337
1338/**
1339 * create_worker - create a new workqueue worker
1340 * @gcwq: gcwq the new worker will belong to
1341 * @bind: whether to set affinity to @cpu or not
1342 *
1343 * Create a new worker which is bound to @gcwq. The returned worker
1344 * can be started by calling start_worker() or destroyed using
1345 * destroy_worker().
1346 *
1347 * CONTEXT:
1348 * Might sleep. Does GFP_KERNEL allocations.
1349 *
1350 * RETURNS:
1351 * Pointer to the newly created worker.
1352 */
1353static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1354{
1355 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1356 struct worker *worker = NULL;
1357 int id = -1;
1358
1359 spin_lock_irq(&gcwq->lock);
1360 while (ida_get_new(&gcwq->worker_ida, &id)) {
1361 spin_unlock_irq(&gcwq->lock);
1362 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1363 goto fail;
1364 spin_lock_irq(&gcwq->lock);
1365 }
1366 spin_unlock_irq(&gcwq->lock);
1367
1368 worker = alloc_worker();
1369 if (!worker)
1370 goto fail;
1371
1372 worker->gcwq = gcwq;
1373 worker->id = id;
1374
1375 if (!on_unbound_cpu)
1376 worker->task = kthread_create_on_node(worker_thread,
1377 worker,
1378 cpu_to_node(gcwq->cpu),
1379 "kworker/%u:%d", gcwq->cpu, id);
1380 else
1381 worker->task = kthread_create(worker_thread, worker,
1382 "kworker/u:%d", id);
1383 if (IS_ERR(worker->task))
1384 goto fail;
1385
1386 /*
1387 * A rogue worker will become a regular one if CPU comes
1388 * online later on. Make sure every worker has
1389 * PF_THREAD_BOUND set.
1390 */
1391 if (bind && !on_unbound_cpu)
1392 kthread_bind(worker->task, gcwq->cpu);
1393 else {
1394 worker->task->flags |= PF_THREAD_BOUND;
1395 if (on_unbound_cpu)
1396 worker->flags |= WORKER_UNBOUND;
1397 }
1398
1399 return worker;
1400fail:
1401 if (id >= 0) {
1402 spin_lock_irq(&gcwq->lock);
1403 ida_remove(&gcwq->worker_ida, id);
1404 spin_unlock_irq(&gcwq->lock);
1405 }
1406 kfree(worker);
1407 return NULL;
1408}
1409
1410/**
1411 * start_worker - start a newly created worker
1412 * @worker: worker to start
1413 *
1414 * Make the gcwq aware of @worker and start it.
1415 *
1416 * CONTEXT:
1417 * spin_lock_irq(gcwq->lock).
1418 */
1419static void start_worker(struct worker *worker)
1420{
1421 worker->flags |= WORKER_STARTED;
1422 worker->gcwq->nr_workers++;
1423 worker_enter_idle(worker);
1424 wake_up_process(worker->task);
1425}
1426
1427/**
1428 * destroy_worker - destroy a workqueue worker
1429 * @worker: worker to be destroyed
1430 *
1431 * Destroy @worker and adjust @gcwq stats accordingly.
1432 *
1433 * CONTEXT:
1434 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1435 */
1436static void destroy_worker(struct worker *worker)
1437{
1438 struct global_cwq *gcwq = worker->gcwq;
1439 int id = worker->id;
1440
1441 /* sanity check frenzy */
1442 BUG_ON(worker->current_work);
1443 BUG_ON(!list_empty(&worker->scheduled));
1444
1445 if (worker->flags & WORKER_STARTED)
1446 gcwq->nr_workers--;
1447 if (worker->flags & WORKER_IDLE)
1448 gcwq->nr_idle--;
1449
1450 list_del_init(&worker->entry);
1451 worker->flags |= WORKER_DIE;
1452
1453 spin_unlock_irq(&gcwq->lock);
1454
1455 kthread_stop(worker->task);
1456 kfree(worker);
1457
1458 spin_lock_irq(&gcwq->lock);
1459 ida_remove(&gcwq->worker_ida, id);
1460}
1461
1462static void idle_worker_timeout(unsigned long __gcwq)
1463{
1464 struct global_cwq *gcwq = (void *)__gcwq;
1465
1466 spin_lock_irq(&gcwq->lock);
1467
1468 if (too_many_workers(gcwq)) {
1469 struct worker *worker;
1470 unsigned long expires;
1471
1472 /* idle_list is kept in LIFO order, check the last one */
1473 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1474 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1475
1476 if (time_before(jiffies, expires))
1477 mod_timer(&gcwq->idle_timer, expires);
1478 else {
1479 /* it's been idle for too long, wake up manager */
1480 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1481 wake_up_worker(gcwq);
1482 }
1483 }
1484
1485 spin_unlock_irq(&gcwq->lock);
1486}
1487
1488static bool send_mayday(struct work_struct *work)
1489{
1490 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1491 struct workqueue_struct *wq = cwq->wq;
1492 unsigned int cpu;
1493
1494 if (!(wq->flags & WQ_RESCUER))
1495 return false;
1496
1497 /* mayday mayday mayday */
1498 cpu = cwq->gcwq->cpu;
1499 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1500 if (cpu == WORK_CPU_UNBOUND)
1501 cpu = 0;
1502 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1503 wake_up_process(wq->rescuer->task);
1504 return true;
1505}
1506
1507static void gcwq_mayday_timeout(unsigned long __gcwq)
1508{
1509 struct global_cwq *gcwq = (void *)__gcwq;
1510 struct work_struct *work;
1511
1512 spin_lock_irq(&gcwq->lock);
1513
1514 if (need_to_create_worker(gcwq)) {
1515 /*
1516 * We've been trying to create a new worker but
1517 * haven't been successful. We might be hitting an
1518 * allocation deadlock. Send distress signals to
1519 * rescuers.
1520 */
1521 list_for_each_entry(work, &gcwq->worklist, entry)
1522 send_mayday(work);
1523 }
1524
1525 spin_unlock_irq(&gcwq->lock);
1526
1527 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1528}
1529
1530/**
1531 * maybe_create_worker - create a new worker if necessary
1532 * @gcwq: gcwq to create a new worker for
1533 *
1534 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1535 * have at least one idle worker on return from this function. If
1536 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1537 * sent to all rescuers with works scheduled on @gcwq to resolve
1538 * possible allocation deadlock.
1539 *
1540 * On return, need_to_create_worker() is guaranteed to be false and
1541 * may_start_working() true.
1542 *
1543 * LOCKING:
1544 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1545 * multiple times. Does GFP_KERNEL allocations. Called only from
1546 * manager.
1547 *
1548 * RETURNS:
1549 * false if no action was taken and gcwq->lock stayed locked, true
1550 * otherwise.
1551 */
1552static bool maybe_create_worker(struct global_cwq *gcwq)
1553__releases(&gcwq->lock)
1554__acquires(&gcwq->lock)
1555{
1556 if (!need_to_create_worker(gcwq))
1557 return false;
1558restart:
1559 spin_unlock_irq(&gcwq->lock);
1560
1561 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1562 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1563
1564 while (true) {
1565 struct worker *worker;
1566
1567 worker = create_worker(gcwq, true);
1568 if (worker) {
1569 del_timer_sync(&gcwq->mayday_timer);
1570 spin_lock_irq(&gcwq->lock);
1571 start_worker(worker);
1572 BUG_ON(need_to_create_worker(gcwq));
1573 return true;
1574 }
1575
1576 if (!need_to_create_worker(gcwq))
1577 break;
1578
1579 __set_current_state(TASK_INTERRUPTIBLE);
1580 schedule_timeout(CREATE_COOLDOWN);
1581
1582 if (!need_to_create_worker(gcwq))
1583 break;
1584 }
1585
1586 del_timer_sync(&gcwq->mayday_timer);
1587 spin_lock_irq(&gcwq->lock);
1588 if (need_to_create_worker(gcwq))
1589 goto restart;
1590 return true;
1591}
1592
1593/**
1594 * maybe_destroy_worker - destroy workers which have been idle for a while
1595 * @gcwq: gcwq to destroy workers for
1596 *
1597 * Destroy @gcwq workers which have been idle for longer than
1598 * IDLE_WORKER_TIMEOUT.
1599 *
1600 * LOCKING:
1601 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1602 * multiple times. Called only from manager.
1603 *
1604 * RETURNS:
1605 * false if no action was taken and gcwq->lock stayed locked, true
1606 * otherwise.
1607 */
1608static bool maybe_destroy_workers(struct global_cwq *gcwq)
1609{
1610 bool ret = false;
1611
1612 while (too_many_workers(gcwq)) {
1613 struct worker *worker;
1614 unsigned long expires;
1615
1616 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1617 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1618
1619 if (time_before(jiffies, expires)) {
1620 mod_timer(&gcwq->idle_timer, expires);
1621 break;
1622 }
1623
1624 destroy_worker(worker);
1625 ret = true;
1626 }
1627
1628 return ret;
1629}
1630
1631/**
1632 * manage_workers - manage worker pool
1633 * @worker: self
1634 *
1635 * Assume the manager role and manage gcwq worker pool @worker belongs
1636 * to. At any given time, there can be only zero or one manager per
1637 * gcwq. The exclusion is handled automatically by this function.
1638 *
1639 * The caller can safely start processing works on false return. On
1640 * true return, it's guaranteed that need_to_create_worker() is false
1641 * and may_start_working() is true.
1642 *
1643 * CONTEXT:
1644 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1645 * multiple times. Does GFP_KERNEL allocations.
1646 *
1647 * RETURNS:
1648 * false if no action was taken and gcwq->lock stayed locked, true if
1649 * some action was taken.
1650 */
1651static bool manage_workers(struct worker *worker)
1652{
1653 struct global_cwq *gcwq = worker->gcwq;
1654 bool ret = false;
1655
1656 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1657 return ret;
1658
1659 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1660 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1661
1662 /*
1663 * Destroy and then create so that may_start_working() is true
1664 * on return.
1665 */
1666 ret |= maybe_destroy_workers(gcwq);
1667 ret |= maybe_create_worker(gcwq);
1668
1669 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1670
1671 /*
1672 * The trustee might be waiting to take over the manager
1673 * position, tell it we're done.
1674 */
1675 if (unlikely(gcwq->trustee))
1676 wake_up_all(&gcwq->trustee_wait);
1677
1678 return ret;
1679}
1680
1681/**
1682 * move_linked_works - move linked works to a list
1683 * @work: start of series of works to be scheduled
1684 * @head: target list to append @work to
1685 * @nextp: out paramter for nested worklist walking
1686 *
1687 * Schedule linked works starting from @work to @head. Work series to
1688 * be scheduled starts at @work and includes any consecutive work with
1689 * WORK_STRUCT_LINKED set in its predecessor.
1690 *
1691 * If @nextp is not NULL, it's updated to point to the next work of
1692 * the last scheduled work. This allows move_linked_works() to be
1693 * nested inside outer list_for_each_entry_safe().
1694 *
1695 * CONTEXT:
1696 * spin_lock_irq(gcwq->lock).
1697 */
1698static void move_linked_works(struct work_struct *work, struct list_head *head,
1699 struct work_struct **nextp)
1700{
1701 struct work_struct *n;
1702
1703 /*
1704 * Linked worklist will always end before the end of the list,
1705 * use NULL for list head.
1706 */
1707 list_for_each_entry_safe_from(work, n, NULL, entry) {
1708 list_move_tail(&work->entry, head);
1709 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1710 break;
1711 }
1712
1713 /*
1714 * If we're already inside safe list traversal and have moved
1715 * multiple works to the scheduled queue, the next position
1716 * needs to be updated.
1717 */
1718 if (nextp)
1719 *nextp = n;
1720}
1721
1722static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1723{
1724 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1725 struct work_struct, entry);
1726 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1727
1728 trace_workqueue_activate_work(work);
1729 move_linked_works(work, pos, NULL);
1730 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1731 cwq->nr_active++;
1732}
1733
1734/**
1735 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1736 * @cwq: cwq of interest
1737 * @color: color of work which left the queue
1738 * @delayed: for a delayed work
1739 *
1740 * A work either has completed or is removed from pending queue,
1741 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1742 *
1743 * CONTEXT:
1744 * spin_lock_irq(gcwq->lock).
1745 */
1746static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1747 bool delayed)
1748{
1749 /* ignore uncolored works */
1750 if (color == WORK_NO_COLOR)
1751 return;
1752
1753 cwq->nr_in_flight[color]--;
1754
1755 if (!delayed) {
1756 cwq->nr_active--;
1757 if (!list_empty(&cwq->delayed_works)) {
1758 /* one down, submit a delayed one */
1759 if (cwq->nr_active < cwq->max_active)
1760 cwq_activate_first_delayed(cwq);
1761 }
1762 }
1763
1764 /* is flush in progress and are we at the flushing tip? */
1765 if (likely(cwq->flush_color != color))
1766 return;
1767
1768 /* are there still in-flight works? */
1769 if (cwq->nr_in_flight[color])
1770 return;
1771
1772 /* this cwq is done, clear flush_color */
1773 cwq->flush_color = -1;
1774
1775 /*
1776 * If this was the last cwq, wake up the first flusher. It
1777 * will handle the rest.
1778 */
1779 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1780 complete(&cwq->wq->first_flusher->done);
1781}
1782
1783/**
1784 * process_one_work - process single work
1785 * @worker: self
1786 * @work: work to process
1787 *
1788 * Process @work. This function contains all the logics necessary to
1789 * process a single work including synchronization against and
1790 * interaction with other workers on the same cpu, queueing and
1791 * flushing. As long as context requirement is met, any worker can
1792 * call this function to process a work.
1793 *
1794 * CONTEXT:
1795 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1796 */
1797static void process_one_work(struct worker *worker, struct work_struct *work)
1798__releases(&gcwq->lock)
1799__acquires(&gcwq->lock)
1800{
1801 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1802 struct global_cwq *gcwq = cwq->gcwq;
1803 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1804 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1805 work_func_t f = work->func;
1806 int work_color;
1807 struct worker *collision;
1808#ifdef CONFIG_LOCKDEP
1809 /*
1810 * It is permissible to free the struct work_struct from
1811 * inside the function that is called from it, this we need to
1812 * take into account for lockdep too. To avoid bogus "held
1813 * lock freed" warnings as well as problems when looking into
1814 * work->lockdep_map, make a copy and use that here.
1815 */
1816 struct lockdep_map lockdep_map = work->lockdep_map;
1817#endif
1818 /*
1819 * A single work shouldn't be executed concurrently by
1820 * multiple workers on a single cpu. Check whether anyone is
1821 * already processing the work. If so, defer the work to the
1822 * currently executing one.
1823 */
1824 collision = __find_worker_executing_work(gcwq, bwh, work);
1825 if (unlikely(collision)) {
1826 move_linked_works(work, &collision->scheduled, NULL);
1827 return;
1828 }
1829
1830 /* claim and process */
1831 debug_work_deactivate(work);
1832 hlist_add_head(&worker->hentry, bwh);
1833 worker->current_work = work;
1834 worker->current_cwq = cwq;
1835 work_color = get_work_color(work);
1836
1837 /* record the current cpu number in the work data and dequeue */
1838 set_work_cpu(work, gcwq->cpu);
1839 list_del_init(&work->entry);
1840
1841 /*
1842 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1843 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1844 */
1845 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1846 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1847 struct work_struct, entry);
1848
1849 if (!list_empty(&gcwq->worklist) &&
1850 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1851 wake_up_worker(gcwq);
1852 else
1853 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1854 }
1855
1856 /*
1857 * CPU intensive works don't participate in concurrency
1858 * management. They're the scheduler's responsibility.
1859 */
1860 if (unlikely(cpu_intensive))
1861 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1862
1863 spin_unlock_irq(&gcwq->lock);
1864
1865 work_clear_pending(work);
1866 lock_map_acquire_read(&cwq->wq->lockdep_map);
1867 lock_map_acquire(&lockdep_map);
1868 trace_workqueue_execute_start(work);
1869 f(work);
1870 /*
1871 * While we must be careful to not use "work" after this, the trace
1872 * point will only record its address.
1873 */
1874 trace_workqueue_execute_end(work);
1875 lock_map_release(&lockdep_map);
1876 lock_map_release(&cwq->wq->lockdep_map);
1877
1878 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1879 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1880 "%s/0x%08x/%d\n",
1881 current->comm, preempt_count(), task_pid_nr(current));
1882 printk(KERN_ERR " last function: ");
1883 print_symbol("%s\n", (unsigned long)f);
1884 debug_show_held_locks(current);
1885 dump_stack();
1886 }
1887
1888 spin_lock_irq(&gcwq->lock);
1889
1890 /* clear cpu intensive status */
1891 if (unlikely(cpu_intensive))
1892 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1893
1894 /* we're done with it, release */
1895 hlist_del_init(&worker->hentry);
1896 worker->current_work = NULL;
1897 worker->current_cwq = NULL;
1898 cwq_dec_nr_in_flight(cwq, work_color, false);
1899}
1900
1901/**
1902 * process_scheduled_works - process scheduled works
1903 * @worker: self
1904 *
1905 * Process all scheduled works. Please note that the scheduled list
1906 * may change while processing a work, so this function repeatedly
1907 * fetches a work from the top and executes it.
1908 *
1909 * CONTEXT:
1910 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1911 * multiple times.
1912 */
1913static void process_scheduled_works(struct worker *worker)
1914{
1915 while (!list_empty(&worker->scheduled)) {
1916 struct work_struct *work = list_first_entry(&worker->scheduled,
1917 struct work_struct, entry);
1918 process_one_work(worker, work);
1919 }
1920}
1921
1922/**
1923 * worker_thread - the worker thread function
1924 * @__worker: self
1925 *
1926 * The gcwq worker thread function. There's a single dynamic pool of
1927 * these per each cpu. These workers process all works regardless of
1928 * their specific target workqueue. The only exception is works which
1929 * belong to workqueues with a rescuer which will be explained in
1930 * rescuer_thread().
1931 */
1932static int worker_thread(void *__worker)
1933{
1934 struct worker *worker = __worker;
1935 struct global_cwq *gcwq = worker->gcwq;
1936
1937 /* tell the scheduler that this is a workqueue worker */
1938 worker->task->flags |= PF_WQ_WORKER;
1939woke_up:
1940 spin_lock_irq(&gcwq->lock);
1941
1942 /* DIE can be set only while we're idle, checking here is enough */
1943 if (worker->flags & WORKER_DIE) {
1944 spin_unlock_irq(&gcwq->lock);
1945 worker->task->flags &= ~PF_WQ_WORKER;
1946 return 0;
1947 }
1948
1949 worker_leave_idle(worker);
1950recheck:
1951 /* no more worker necessary? */
1952 if (!need_more_worker(gcwq))
1953 goto sleep;
1954
1955 /* do we need to manage? */
1956 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1957 goto recheck;
1958
1959 /*
1960 * ->scheduled list can only be filled while a worker is
1961 * preparing to process a work or actually processing it.
1962 * Make sure nobody diddled with it while I was sleeping.
1963 */
1964 BUG_ON(!list_empty(&worker->scheduled));
1965
1966 /*
1967 * When control reaches this point, we're guaranteed to have
1968 * at least one idle worker or that someone else has already
1969 * assumed the manager role.
1970 */
1971 worker_clr_flags(worker, WORKER_PREP);
1972
1973 do {
1974 struct work_struct *work =
1975 list_first_entry(&gcwq->worklist,
1976 struct work_struct, entry);
1977
1978 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1979 /* optimization path, not strictly necessary */
1980 process_one_work(worker, work);
1981 if (unlikely(!list_empty(&worker->scheduled)))
1982 process_scheduled_works(worker);
1983 } else {
1984 move_linked_works(work, &worker->scheduled, NULL);
1985 process_scheduled_works(worker);
1986 }
1987 } while (keep_working(gcwq));
1988
1989 worker_set_flags(worker, WORKER_PREP, false);
1990sleep:
1991 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1992 goto recheck;
1993
1994 /*
1995 * gcwq->lock is held and there's no work to process and no
1996 * need to manage, sleep. Workers are woken up only while
1997 * holding gcwq->lock or from local cpu, so setting the
1998 * current state before releasing gcwq->lock is enough to
1999 * prevent losing any event.
2000 */
2001 worker_enter_idle(worker);
2002 __set_current_state(TASK_INTERRUPTIBLE);
2003 spin_unlock_irq(&gcwq->lock);
2004 schedule();
2005 goto woke_up;
2006}
2007
2008/**
2009 * rescuer_thread - the rescuer thread function
2010 * @__wq: the associated workqueue
2011 *
2012 * Workqueue rescuer thread function. There's one rescuer for each
2013 * workqueue which has WQ_RESCUER set.
2014 *
2015 * Regular work processing on a gcwq may block trying to create a new
2016 * worker which uses GFP_KERNEL allocation which has slight chance of
2017 * developing into deadlock if some works currently on the same queue
2018 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2019 * the problem rescuer solves.
2020 *
2021 * When such condition is possible, the gcwq summons rescuers of all
2022 * workqueues which have works queued on the gcwq and let them process
2023 * those works so that forward progress can be guaranteed.
2024 *
2025 * This should happen rarely.
2026 */
2027static int rescuer_thread(void *__wq)
2028{
2029 struct workqueue_struct *wq = __wq;
2030 struct worker *rescuer = wq->rescuer;
2031 struct list_head *scheduled = &rescuer->scheduled;
2032 bool is_unbound = wq->flags & WQ_UNBOUND;
2033 unsigned int cpu;
2034
2035 set_user_nice(current, RESCUER_NICE_LEVEL);
2036repeat:
2037 set_current_state(TASK_INTERRUPTIBLE);
2038
2039 if (kthread_should_stop())
2040 return 0;
2041
2042 /*
2043 * See whether any cpu is asking for help. Unbounded
2044 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2045 */
2046 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2047 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2048 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2049 struct global_cwq *gcwq = cwq->gcwq;
2050 struct work_struct *work, *n;
2051
2052 __set_current_state(TASK_RUNNING);
2053 mayday_clear_cpu(cpu, wq->mayday_mask);
2054
2055 /* migrate to the target cpu if possible */
2056 rescuer->gcwq = gcwq;
2057 worker_maybe_bind_and_lock(rescuer);
2058
2059 /*
2060 * Slurp in all works issued via this workqueue and
2061 * process'em.
2062 */
2063 BUG_ON(!list_empty(&rescuer->scheduled));
2064 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2065 if (get_work_cwq(work) == cwq)
2066 move_linked_works(work, scheduled, &n);
2067
2068 process_scheduled_works(rescuer);
2069
2070 /*
2071 * Leave this gcwq. If keep_working() is %true, notify a
2072 * regular worker; otherwise, we end up with 0 concurrency
2073 * and stalling the execution.
2074 */
2075 if (keep_working(gcwq))
2076 wake_up_worker(gcwq);
2077
2078 spin_unlock_irq(&gcwq->lock);
2079 }
2080
2081 schedule();
2082 goto repeat;
2083}
2084
2085struct wq_barrier {
2086 struct work_struct work;
2087 struct completion done;
2088};
2089
2090static void wq_barrier_func(struct work_struct *work)
2091{
2092 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2093 complete(&barr->done);
2094}
2095
2096/**
2097 * insert_wq_barrier - insert a barrier work
2098 * @cwq: cwq to insert barrier into
2099 * @barr: wq_barrier to insert
2100 * @target: target work to attach @barr to
2101 * @worker: worker currently executing @target, NULL if @target is not executing
2102 *
2103 * @barr is linked to @target such that @barr is completed only after
2104 * @target finishes execution. Please note that the ordering
2105 * guarantee is observed only with respect to @target and on the local
2106 * cpu.
2107 *
2108 * Currently, a queued barrier can't be canceled. This is because
2109 * try_to_grab_pending() can't determine whether the work to be
2110 * grabbed is at the head of the queue and thus can't clear LINKED
2111 * flag of the previous work while there must be a valid next work
2112 * after a work with LINKED flag set.
2113 *
2114 * Note that when @worker is non-NULL, @target may be modified
2115 * underneath us, so we can't reliably determine cwq from @target.
2116 *
2117 * CONTEXT:
2118 * spin_lock_irq(gcwq->lock).
2119 */
2120static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2121 struct wq_barrier *barr,
2122 struct work_struct *target, struct worker *worker)
2123{
2124 struct list_head *head;
2125 unsigned int linked = 0;
2126
2127 /*
2128 * debugobject calls are safe here even with gcwq->lock locked
2129 * as we know for sure that this will not trigger any of the
2130 * checks and call back into the fixup functions where we
2131 * might deadlock.
2132 */
2133 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2134 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2135 init_completion(&barr->done);
2136
2137 /*
2138 * If @target is currently being executed, schedule the
2139 * barrier to the worker; otherwise, put it after @target.
2140 */
2141 if (worker)
2142 head = worker->scheduled.next;
2143 else {
2144 unsigned long *bits = work_data_bits(target);
2145
2146 head = target->entry.next;
2147 /* there can already be other linked works, inherit and set */
2148 linked = *bits & WORK_STRUCT_LINKED;
2149 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2150 }
2151
2152 debug_work_activate(&barr->work);
2153 insert_work(cwq, &barr->work, head,
2154 work_color_to_flags(WORK_NO_COLOR) | linked);
2155}
2156
2157/**
2158 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2159 * @wq: workqueue being flushed
2160 * @flush_color: new flush color, < 0 for no-op
2161 * @work_color: new work color, < 0 for no-op
2162 *
2163 * Prepare cwqs for workqueue flushing.
2164 *
2165 * If @flush_color is non-negative, flush_color on all cwqs should be
2166 * -1. If no cwq has in-flight commands at the specified color, all
2167 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2168 * has in flight commands, its cwq->flush_color is set to
2169 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2170 * wakeup logic is armed and %true is returned.
2171 *
2172 * The caller should have initialized @wq->first_flusher prior to
2173 * calling this function with non-negative @flush_color. If
2174 * @flush_color is negative, no flush color update is done and %false
2175 * is returned.
2176 *
2177 * If @work_color is non-negative, all cwqs should have the same
2178 * work_color which is previous to @work_color and all will be
2179 * advanced to @work_color.
2180 *
2181 * CONTEXT:
2182 * mutex_lock(wq->flush_mutex).
2183 *
2184 * RETURNS:
2185 * %true if @flush_color >= 0 and there's something to flush. %false
2186 * otherwise.
2187 */
2188static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2189 int flush_color, int work_color)
2190{
2191 bool wait = false;
2192 unsigned int cpu;
2193
2194 if (flush_color >= 0) {
2195 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2196 atomic_set(&wq->nr_cwqs_to_flush, 1);
2197 }
2198
2199 for_each_cwq_cpu(cpu, wq) {
2200 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2201 struct global_cwq *gcwq = cwq->gcwq;
2202
2203 spin_lock_irq(&gcwq->lock);
2204
2205 if (flush_color >= 0) {
2206 BUG_ON(cwq->flush_color != -1);
2207
2208 if (cwq->nr_in_flight[flush_color]) {
2209 cwq->flush_color = flush_color;
2210 atomic_inc(&wq->nr_cwqs_to_flush);
2211 wait = true;
2212 }
2213 }
2214
2215 if (work_color >= 0) {
2216 BUG_ON(work_color != work_next_color(cwq->work_color));
2217 cwq->work_color = work_color;
2218 }
2219
2220 spin_unlock_irq(&gcwq->lock);
2221 }
2222
2223 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2224 complete(&wq->first_flusher->done);
2225
2226 return wait;
2227}
2228
2229/**
2230 * flush_workqueue - ensure that any scheduled work has run to completion.
2231 * @wq: workqueue to flush
2232 *
2233 * Forces execution of the workqueue and blocks until its completion.
2234 * This is typically used in driver shutdown handlers.
2235 *
2236 * We sleep until all works which were queued on entry have been handled,
2237 * but we are not livelocked by new incoming ones.
2238 */
2239void flush_workqueue(struct workqueue_struct *wq)
2240{
2241 struct wq_flusher this_flusher = {
2242 .list = LIST_HEAD_INIT(this_flusher.list),
2243 .flush_color = -1,
2244 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2245 };
2246 int next_color;
2247
2248 lock_map_acquire(&wq->lockdep_map);
2249 lock_map_release(&wq->lockdep_map);
2250
2251 mutex_lock(&wq->flush_mutex);
2252
2253 /*
2254 * Start-to-wait phase
2255 */
2256 next_color = work_next_color(wq->work_color);
2257
2258 if (next_color != wq->flush_color) {
2259 /*
2260 * Color space is not full. The current work_color
2261 * becomes our flush_color and work_color is advanced
2262 * by one.
2263 */
2264 BUG_ON(!list_empty(&wq->flusher_overflow));
2265 this_flusher.flush_color = wq->work_color;
2266 wq->work_color = next_color;
2267
2268 if (!wq->first_flusher) {
2269 /* no flush in progress, become the first flusher */
2270 BUG_ON(wq->flush_color != this_flusher.flush_color);
2271
2272 wq->first_flusher = &this_flusher;
2273
2274 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2275 wq->work_color)) {
2276 /* nothing to flush, done */
2277 wq->flush_color = next_color;
2278 wq->first_flusher = NULL;
2279 goto out_unlock;
2280 }
2281 } else {
2282 /* wait in queue */
2283 BUG_ON(wq->flush_color == this_flusher.flush_color);
2284 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2285 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2286 }
2287 } else {
2288 /*
2289 * Oops, color space is full, wait on overflow queue.
2290 * The next flush completion will assign us
2291 * flush_color and transfer to flusher_queue.
2292 */
2293 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2294 }
2295
2296 mutex_unlock(&wq->flush_mutex);
2297
2298 wait_for_completion(&this_flusher.done);
2299
2300 /*
2301 * Wake-up-and-cascade phase
2302 *
2303 * First flushers are responsible for cascading flushes and
2304 * handling overflow. Non-first flushers can simply return.
2305 */
2306 if (wq->first_flusher != &this_flusher)
2307 return;
2308
2309 mutex_lock(&wq->flush_mutex);
2310
2311 /* we might have raced, check again with mutex held */
2312 if (wq->first_flusher != &this_flusher)
2313 goto out_unlock;
2314
2315 wq->first_flusher = NULL;
2316
2317 BUG_ON(!list_empty(&this_flusher.list));
2318 BUG_ON(wq->flush_color != this_flusher.flush_color);
2319
2320 while (true) {
2321 struct wq_flusher *next, *tmp;
2322
2323 /* complete all the flushers sharing the current flush color */
2324 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2325 if (next->flush_color != wq->flush_color)
2326 break;
2327 list_del_init(&next->list);
2328 complete(&next->done);
2329 }
2330
2331 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2332 wq->flush_color != work_next_color(wq->work_color));
2333
2334 /* this flush_color is finished, advance by one */
2335 wq->flush_color = work_next_color(wq->flush_color);
2336
2337 /* one color has been freed, handle overflow queue */
2338 if (!list_empty(&wq->flusher_overflow)) {
2339 /*
2340 * Assign the same color to all overflowed
2341 * flushers, advance work_color and append to
2342 * flusher_queue. This is the start-to-wait
2343 * phase for these overflowed flushers.
2344 */
2345 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2346 tmp->flush_color = wq->work_color;
2347
2348 wq->work_color = work_next_color(wq->work_color);
2349
2350 list_splice_tail_init(&wq->flusher_overflow,
2351 &wq->flusher_queue);
2352 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2353 }
2354
2355 if (list_empty(&wq->flusher_queue)) {
2356 BUG_ON(wq->flush_color != wq->work_color);
2357 break;
2358 }
2359
2360 /*
2361 * Need to flush more colors. Make the next flusher
2362 * the new first flusher and arm cwqs.
2363 */
2364 BUG_ON(wq->flush_color == wq->work_color);
2365 BUG_ON(wq->flush_color != next->flush_color);
2366
2367 list_del_init(&next->list);
2368 wq->first_flusher = next;
2369
2370 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2371 break;
2372
2373 /*
2374 * Meh... this color is already done, clear first
2375 * flusher and repeat cascading.
2376 */
2377 wq->first_flusher = NULL;
2378 }
2379
2380out_unlock:
2381 mutex_unlock(&wq->flush_mutex);
2382}
2383EXPORT_SYMBOL_GPL(flush_workqueue);
2384
2385/**
2386 * drain_workqueue - drain a workqueue
2387 * @wq: workqueue to drain
2388 *
2389 * Wait until the workqueue becomes empty. While draining is in progress,
2390 * only chain queueing is allowed. IOW, only currently pending or running
2391 * work items on @wq can queue further work items on it. @wq is flushed
2392 * repeatedly until it becomes empty. The number of flushing is detemined
2393 * by the depth of chaining and should be relatively short. Whine if it
2394 * takes too long.
2395 */
2396void drain_workqueue(struct workqueue_struct *wq)
2397{
2398 unsigned int flush_cnt = 0;
2399 unsigned int cpu;
2400
2401 /*
2402 * __queue_work() needs to test whether there are drainers, is much
2403 * hotter than drain_workqueue() and already looks at @wq->flags.
2404 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2405 */
2406 spin_lock(&workqueue_lock);
2407 if (!wq->nr_drainers++)
2408 wq->flags |= WQ_DRAINING;
2409 spin_unlock(&workqueue_lock);
2410reflush:
2411 flush_workqueue(wq);
2412
2413 for_each_cwq_cpu(cpu, wq) {
2414 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2415 bool drained;
2416
2417 spin_lock_irq(&cwq->gcwq->lock);
2418 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2419 spin_unlock_irq(&cwq->gcwq->lock);
2420
2421 if (drained)
2422 continue;
2423
2424 if (++flush_cnt == 10 ||
2425 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2426 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2427 wq->name, flush_cnt);
2428 goto reflush;
2429 }
2430
2431 spin_lock(&workqueue_lock);
2432 if (!--wq->nr_drainers)
2433 wq->flags &= ~WQ_DRAINING;
2434 spin_unlock(&workqueue_lock);
2435}
2436EXPORT_SYMBOL_GPL(drain_workqueue);
2437
2438static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2439 bool wait_executing)
2440{
2441 struct worker *worker = NULL;
2442 struct global_cwq *gcwq;
2443 struct cpu_workqueue_struct *cwq;
2444
2445 might_sleep();
2446 gcwq = get_work_gcwq(work);
2447 if (!gcwq)
2448 return false;
2449
2450 spin_lock_irq(&gcwq->lock);
2451 if (!list_empty(&work->entry)) {
2452 /*
2453 * See the comment near try_to_grab_pending()->smp_rmb().
2454 * If it was re-queued to a different gcwq under us, we
2455 * are not going to wait.
2456 */
2457 smp_rmb();
2458 cwq = get_work_cwq(work);
2459 if (unlikely(!cwq || gcwq != cwq->gcwq))
2460 goto already_gone;
2461 } else if (wait_executing) {
2462 worker = find_worker_executing_work(gcwq, work);
2463 if (!worker)
2464 goto already_gone;
2465 cwq = worker->current_cwq;
2466 } else
2467 goto already_gone;
2468
2469 insert_wq_barrier(cwq, barr, work, worker);
2470 spin_unlock_irq(&gcwq->lock);
2471
2472 /*
2473 * If @max_active is 1 or rescuer is in use, flushing another work
2474 * item on the same workqueue may lead to deadlock. Make sure the
2475 * flusher is not running on the same workqueue by verifying write
2476 * access.
2477 */
2478 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2479 lock_map_acquire(&cwq->wq->lockdep_map);
2480 else
2481 lock_map_acquire_read(&cwq->wq->lockdep_map);
2482 lock_map_release(&cwq->wq->lockdep_map);
2483
2484 return true;
2485already_gone:
2486 spin_unlock_irq(&gcwq->lock);
2487 return false;
2488}
2489
2490/**
2491 * flush_work - wait for a work to finish executing the last queueing instance
2492 * @work: the work to flush
2493 *
2494 * Wait until @work has finished execution. This function considers
2495 * only the last queueing instance of @work. If @work has been
2496 * enqueued across different CPUs on a non-reentrant workqueue or on
2497 * multiple workqueues, @work might still be executing on return on
2498 * some of the CPUs from earlier queueing.
2499 *
2500 * If @work was queued only on a non-reentrant, ordered or unbound
2501 * workqueue, @work is guaranteed to be idle on return if it hasn't
2502 * been requeued since flush started.
2503 *
2504 * RETURNS:
2505 * %true if flush_work() waited for the work to finish execution,
2506 * %false if it was already idle.
2507 */
2508bool flush_work(struct work_struct *work)
2509{
2510 struct wq_barrier barr;
2511
2512 if (start_flush_work(work, &barr, true)) {
2513 wait_for_completion(&barr.done);
2514 destroy_work_on_stack(&barr.work);
2515 return true;
2516 } else
2517 return false;
2518}
2519EXPORT_SYMBOL_GPL(flush_work);
2520
2521static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2522{
2523 struct wq_barrier barr;
2524 struct worker *worker;
2525
2526 spin_lock_irq(&gcwq->lock);
2527
2528 worker = find_worker_executing_work(gcwq, work);
2529 if (unlikely(worker))
2530 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2531
2532 spin_unlock_irq(&gcwq->lock);
2533
2534 if (unlikely(worker)) {
2535 wait_for_completion(&barr.done);
2536 destroy_work_on_stack(&barr.work);
2537 return true;
2538 } else
2539 return false;
2540}
2541
2542static bool wait_on_work(struct work_struct *work)
2543{
2544 bool ret = false;
2545 int cpu;
2546
2547 might_sleep();
2548
2549 lock_map_acquire(&work->lockdep_map);
2550 lock_map_release(&work->lockdep_map);
2551
2552 for_each_gcwq_cpu(cpu)
2553 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2554 return ret;
2555}
2556
2557/**
2558 * flush_work_sync - wait until a work has finished execution
2559 * @work: the work to flush
2560 *
2561 * Wait until @work has finished execution. On return, it's
2562 * guaranteed that all queueing instances of @work which happened
2563 * before this function is called are finished. In other words, if
2564 * @work hasn't been requeued since this function was called, @work is
2565 * guaranteed to be idle on return.
2566 *
2567 * RETURNS:
2568 * %true if flush_work_sync() waited for the work to finish execution,
2569 * %false if it was already idle.
2570 */
2571bool flush_work_sync(struct work_struct *work)
2572{
2573 struct wq_barrier barr;
2574 bool pending, waited;
2575
2576 /* we'll wait for executions separately, queue barr only if pending */
2577 pending = start_flush_work(work, &barr, false);
2578
2579 /* wait for executions to finish */
2580 waited = wait_on_work(work);
2581
2582 /* wait for the pending one */
2583 if (pending) {
2584 wait_for_completion(&barr.done);
2585 destroy_work_on_stack(&barr.work);
2586 }
2587
2588 return pending || waited;
2589}
2590EXPORT_SYMBOL_GPL(flush_work_sync);
2591
2592/*
2593 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2594 * so this work can't be re-armed in any way.
2595 */
2596static int try_to_grab_pending(struct work_struct *work)
2597{
2598 struct global_cwq *gcwq;
2599 int ret = -1;
2600
2601 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2602 return 0;
2603
2604 /*
2605 * The queueing is in progress, or it is already queued. Try to
2606 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2607 */
2608 gcwq = get_work_gcwq(work);
2609 if (!gcwq)
2610 return ret;
2611
2612 spin_lock_irq(&gcwq->lock);
2613 if (!list_empty(&work->entry)) {
2614 /*
2615 * This work is queued, but perhaps we locked the wrong gcwq.
2616 * In that case we must see the new value after rmb(), see
2617 * insert_work()->wmb().
2618 */
2619 smp_rmb();
2620 if (gcwq == get_work_gcwq(work)) {
2621 debug_work_deactivate(work);
2622 list_del_init(&work->entry);
2623 cwq_dec_nr_in_flight(get_work_cwq(work),
2624 get_work_color(work),
2625 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2626 ret = 1;
2627 }
2628 }
2629 spin_unlock_irq(&gcwq->lock);
2630
2631 return ret;
2632}
2633
2634static bool __cancel_work_timer(struct work_struct *work,
2635 struct timer_list* timer)
2636{
2637 int ret;
2638
2639 do {
2640 ret = (timer && likely(del_timer(timer)));
2641 if (!ret)
2642 ret = try_to_grab_pending(work);
2643 wait_on_work(work);
2644 } while (unlikely(ret < 0));
2645
2646 clear_work_data(work);
2647 return ret;
2648}
2649
2650/**
2651 * cancel_work_sync - cancel a work and wait for it to finish
2652 * @work: the work to cancel
2653 *
2654 * Cancel @work and wait for its execution to finish. This function
2655 * can be used even if the work re-queues itself or migrates to
2656 * another workqueue. On return from this function, @work is
2657 * guaranteed to be not pending or executing on any CPU.
2658 *
2659 * cancel_work_sync(&delayed_work->work) must not be used for
2660 * delayed_work's. Use cancel_delayed_work_sync() instead.
2661 *
2662 * The caller must ensure that the workqueue on which @work was last
2663 * queued can't be destroyed before this function returns.
2664 *
2665 * RETURNS:
2666 * %true if @work was pending, %false otherwise.
2667 */
2668bool cancel_work_sync(struct work_struct *work)
2669{
2670 return __cancel_work_timer(work, NULL);
2671}
2672EXPORT_SYMBOL_GPL(cancel_work_sync);
2673
2674/**
2675 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2676 * @dwork: the delayed work to flush
2677 *
2678 * Delayed timer is cancelled and the pending work is queued for
2679 * immediate execution. Like flush_work(), this function only
2680 * considers the last queueing instance of @dwork.
2681 *
2682 * RETURNS:
2683 * %true if flush_work() waited for the work to finish execution,
2684 * %false if it was already idle.
2685 */
2686bool flush_delayed_work(struct delayed_work *dwork)
2687{
2688 if (del_timer_sync(&dwork->timer))
2689 __queue_work(raw_smp_processor_id(),
2690 get_work_cwq(&dwork->work)->wq, &dwork->work);
2691 return flush_work(&dwork->work);
2692}
2693EXPORT_SYMBOL(flush_delayed_work);
2694
2695/**
2696 * flush_delayed_work_sync - wait for a dwork to finish
2697 * @dwork: the delayed work to flush
2698 *
2699 * Delayed timer is cancelled and the pending work is queued for
2700 * execution immediately. Other than timer handling, its behavior
2701 * is identical to flush_work_sync().
2702 *
2703 * RETURNS:
2704 * %true if flush_work_sync() waited for the work to finish execution,
2705 * %false if it was already idle.
2706 */
2707bool flush_delayed_work_sync(struct delayed_work *dwork)
2708{
2709 if (del_timer_sync(&dwork->timer))
2710 __queue_work(raw_smp_processor_id(),
2711 get_work_cwq(&dwork->work)->wq, &dwork->work);
2712 return flush_work_sync(&dwork->work);
2713}
2714EXPORT_SYMBOL(flush_delayed_work_sync);
2715
2716/**
2717 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2718 * @dwork: the delayed work cancel
2719 *
2720 * This is cancel_work_sync() for delayed works.
2721 *
2722 * RETURNS:
2723 * %true if @dwork was pending, %false otherwise.
2724 */
2725bool cancel_delayed_work_sync(struct delayed_work *dwork)
2726{
2727 return __cancel_work_timer(&dwork->work, &dwork->timer);
2728}
2729EXPORT_SYMBOL(cancel_delayed_work_sync);
2730
2731/**
2732 * schedule_work - put work task in global workqueue
2733 * @work: job to be done
2734 *
2735 * Returns zero if @work was already on the kernel-global workqueue and
2736 * non-zero otherwise.
2737 *
2738 * This puts a job in the kernel-global workqueue if it was not already
2739 * queued and leaves it in the same position on the kernel-global
2740 * workqueue otherwise.
2741 */
2742int schedule_work(struct work_struct *work)
2743{
2744 return queue_work(system_wq, work);
2745}
2746EXPORT_SYMBOL(schedule_work);
2747
2748/*
2749 * schedule_work_on - put work task on a specific cpu
2750 * @cpu: cpu to put the work task on
2751 * @work: job to be done
2752 *
2753 * This puts a job on a specific cpu
2754 */
2755int schedule_work_on(int cpu, struct work_struct *work)
2756{
2757 return queue_work_on(cpu, system_wq, work);
2758}
2759EXPORT_SYMBOL(schedule_work_on);
2760
2761/**
2762 * schedule_delayed_work - put work task in global workqueue after delay
2763 * @dwork: job to be done
2764 * @delay: number of jiffies to wait or 0 for immediate execution
2765 *
2766 * After waiting for a given time this puts a job in the kernel-global
2767 * workqueue.
2768 */
2769int schedule_delayed_work(struct delayed_work *dwork,
2770 unsigned long delay)
2771{
2772 return queue_delayed_work(system_wq, dwork, delay);
2773}
2774EXPORT_SYMBOL(schedule_delayed_work);
2775
2776/**
2777 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2778 * @cpu: cpu to use
2779 * @dwork: job to be done
2780 * @delay: number of jiffies to wait
2781 *
2782 * After waiting for a given time this puts a job in the kernel-global
2783 * workqueue on the specified CPU.
2784 */
2785int schedule_delayed_work_on(int cpu,
2786 struct delayed_work *dwork, unsigned long delay)
2787{
2788 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2789}
2790EXPORT_SYMBOL(schedule_delayed_work_on);
2791
2792/**
2793 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2794 * @func: the function to call
2795 *
2796 * schedule_on_each_cpu() executes @func on each online CPU using the
2797 * system workqueue and blocks until all CPUs have completed.
2798 * schedule_on_each_cpu() is very slow.
2799 *
2800 * RETURNS:
2801 * 0 on success, -errno on failure.
2802 */
2803int schedule_on_each_cpu(work_func_t func)
2804{
2805 int cpu;
2806 struct work_struct __percpu *works;
2807
2808 works = alloc_percpu(struct work_struct);
2809 if (!works)
2810 return -ENOMEM;
2811
2812 get_online_cpus();
2813
2814 for_each_online_cpu(cpu) {
2815 struct work_struct *work = per_cpu_ptr(works, cpu);
2816
2817 INIT_WORK(work, func);
2818 schedule_work_on(cpu, work);
2819 }
2820
2821 for_each_online_cpu(cpu)
2822 flush_work(per_cpu_ptr(works, cpu));
2823
2824 put_online_cpus();
2825 free_percpu(works);
2826 return 0;
2827}
2828
2829/**
2830 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2831 *
2832 * Forces execution of the kernel-global workqueue and blocks until its
2833 * completion.
2834 *
2835 * Think twice before calling this function! It's very easy to get into
2836 * trouble if you don't take great care. Either of the following situations
2837 * will lead to deadlock:
2838 *
2839 * One of the work items currently on the workqueue needs to acquire
2840 * a lock held by your code or its caller.
2841 *
2842 * Your code is running in the context of a work routine.
2843 *
2844 * They will be detected by lockdep when they occur, but the first might not
2845 * occur very often. It depends on what work items are on the workqueue and
2846 * what locks they need, which you have no control over.
2847 *
2848 * In most situations flushing the entire workqueue is overkill; you merely
2849 * need to know that a particular work item isn't queued and isn't running.
2850 * In such cases you should use cancel_delayed_work_sync() or
2851 * cancel_work_sync() instead.
2852 */
2853void flush_scheduled_work(void)
2854{
2855 flush_workqueue(system_wq);
2856}
2857EXPORT_SYMBOL(flush_scheduled_work);
2858
2859/**
2860 * execute_in_process_context - reliably execute the routine with user context
2861 * @fn: the function to execute
2862 * @ew: guaranteed storage for the execute work structure (must
2863 * be available when the work executes)
2864 *
2865 * Executes the function immediately if process context is available,
2866 * otherwise schedules the function for delayed execution.
2867 *
2868 * Returns: 0 - function was executed
2869 * 1 - function was scheduled for execution
2870 */
2871int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2872{
2873 if (!in_interrupt()) {
2874 fn(&ew->work);
2875 return 0;
2876 }
2877
2878 INIT_WORK(&ew->work, fn);
2879 schedule_work(&ew->work);
2880
2881 return 1;
2882}
2883EXPORT_SYMBOL_GPL(execute_in_process_context);
2884
2885int keventd_up(void)
2886{
2887 return system_wq != NULL;
2888}
2889
2890static int alloc_cwqs(struct workqueue_struct *wq)
2891{
2892 /*
2893 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2894 * Make sure that the alignment isn't lower than that of
2895 * unsigned long long.
2896 */
2897 const size_t size = sizeof(struct cpu_workqueue_struct);
2898 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2899 __alignof__(unsigned long long));
2900#ifdef CONFIG_SMP
2901 bool percpu = !(wq->flags & WQ_UNBOUND);
2902#else
2903 bool percpu = false;
2904#endif
2905
2906 if (percpu)
2907 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2908 else {
2909 void *ptr;
2910
2911 /*
2912 * Allocate enough room to align cwq and put an extra
2913 * pointer at the end pointing back to the originally
2914 * allocated pointer which will be used for free.
2915 */
2916 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2917 if (ptr) {
2918 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2919 *(void **)(wq->cpu_wq.single + 1) = ptr;
2920 }
2921 }
2922
2923 /* just in case, make sure it's actually aligned */
2924 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2925 return wq->cpu_wq.v ? 0 : -ENOMEM;
2926}
2927
2928static void free_cwqs(struct workqueue_struct *wq)
2929{
2930#ifdef CONFIG_SMP
2931 bool percpu = !(wq->flags & WQ_UNBOUND);
2932#else
2933 bool percpu = false;
2934#endif
2935
2936 if (percpu)
2937 free_percpu(wq->cpu_wq.pcpu);
2938 else if (wq->cpu_wq.single) {
2939 /* the pointer to free is stored right after the cwq */
2940 kfree(*(void **)(wq->cpu_wq.single + 1));
2941 }
2942}
2943
2944static int wq_clamp_max_active(int max_active, unsigned int flags,
2945 const char *name)
2946{
2947 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2948
2949 if (max_active < 1 || max_active > lim)
2950 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2951 "is out of range, clamping between %d and %d\n",
2952 max_active, name, 1, lim);
2953
2954 return clamp_val(max_active, 1, lim);
2955}
2956
2957struct workqueue_struct *__alloc_workqueue_key(const char *name,
2958 unsigned int flags,
2959 int max_active,
2960 struct lock_class_key *key,
2961 const char *lock_name)
2962{
2963 struct workqueue_struct *wq;
2964 unsigned int cpu;
2965
2966 /*
2967 * Workqueues which may be used during memory reclaim should
2968 * have a rescuer to guarantee forward progress.
2969 */
2970 if (flags & WQ_MEM_RECLAIM)
2971 flags |= WQ_RESCUER;
2972
2973 /*
2974 * Unbound workqueues aren't concurrency managed and should be
2975 * dispatched to workers immediately.
2976 */
2977 if (flags & WQ_UNBOUND)
2978 flags |= WQ_HIGHPRI;
2979
2980 max_active = max_active ?: WQ_DFL_ACTIVE;
2981 max_active = wq_clamp_max_active(max_active, flags, name);
2982
2983 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2984 if (!wq)
2985 goto err;
2986
2987 wq->flags = flags;
2988 wq->saved_max_active = max_active;
2989 mutex_init(&wq->flush_mutex);
2990 atomic_set(&wq->nr_cwqs_to_flush, 0);
2991 INIT_LIST_HEAD(&wq->flusher_queue);
2992 INIT_LIST_HEAD(&wq->flusher_overflow);
2993
2994 wq->name = name;
2995 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2996 INIT_LIST_HEAD(&wq->list);
2997
2998 if (alloc_cwqs(wq) < 0)
2999 goto err;
3000
3001 for_each_cwq_cpu(cpu, wq) {
3002 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3003 struct global_cwq *gcwq = get_gcwq(cpu);
3004
3005 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3006 cwq->gcwq = gcwq;
3007 cwq->wq = wq;
3008 cwq->flush_color = -1;
3009 cwq->max_active = max_active;
3010 INIT_LIST_HEAD(&cwq->delayed_works);
3011 }
3012
3013 if (flags & WQ_RESCUER) {
3014 struct worker *rescuer;
3015
3016 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3017 goto err;
3018
3019 wq->rescuer = rescuer = alloc_worker();
3020 if (!rescuer)
3021 goto err;
3022
3023 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
3024 if (IS_ERR(rescuer->task))
3025 goto err;
3026
3027 rescuer->task->flags |= PF_THREAD_BOUND;
3028 wake_up_process(rescuer->task);
3029 }
3030
3031 /*
3032 * workqueue_lock protects global freeze state and workqueues
3033 * list. Grab it, set max_active accordingly and add the new
3034 * workqueue to workqueues list.
3035 */
3036 spin_lock(&workqueue_lock);
3037
3038 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3039 for_each_cwq_cpu(cpu, wq)
3040 get_cwq(cpu, wq)->max_active = 0;
3041
3042 list_add(&wq->list, &workqueues);
3043
3044 spin_unlock(&workqueue_lock);
3045
3046 return wq;
3047err:
3048 if (wq) {
3049 free_cwqs(wq);
3050 free_mayday_mask(wq->mayday_mask);
3051 kfree(wq->rescuer);
3052 kfree(wq);
3053 }
3054 return NULL;
3055}
3056EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3057
3058/**
3059 * destroy_workqueue - safely terminate a workqueue
3060 * @wq: target workqueue
3061 *
3062 * Safely destroy a workqueue. All work currently pending will be done first.
3063 */
3064void destroy_workqueue(struct workqueue_struct *wq)
3065{
3066 unsigned int cpu;
3067
3068 /* drain it before proceeding with destruction */
3069 drain_workqueue(wq);
3070
3071 /*
3072 * wq list is used to freeze wq, remove from list after
3073 * flushing is complete in case freeze races us.
3074 */
3075 spin_lock(&workqueue_lock);
3076 list_del(&wq->list);
3077 spin_unlock(&workqueue_lock);
3078
3079 /* sanity check */
3080 for_each_cwq_cpu(cpu, wq) {
3081 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3082 int i;
3083
3084 for (i = 0; i < WORK_NR_COLORS; i++)
3085 BUG_ON(cwq->nr_in_flight[i]);
3086 BUG_ON(cwq->nr_active);
3087 BUG_ON(!list_empty(&cwq->delayed_works));
3088 }
3089
3090 if (wq->flags & WQ_RESCUER) {
3091 kthread_stop(wq->rescuer->task);
3092 free_mayday_mask(wq->mayday_mask);
3093 kfree(wq->rescuer);
3094 }
3095
3096 free_cwqs(wq);
3097 kfree(wq);
3098}
3099EXPORT_SYMBOL_GPL(destroy_workqueue);
3100
3101/**
3102 * workqueue_set_max_active - adjust max_active of a workqueue
3103 * @wq: target workqueue
3104 * @max_active: new max_active value.
3105 *
3106 * Set max_active of @wq to @max_active.
3107 *
3108 * CONTEXT:
3109 * Don't call from IRQ context.
3110 */
3111void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3112{
3113 unsigned int cpu;
3114
3115 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3116
3117 spin_lock(&workqueue_lock);
3118
3119 wq->saved_max_active = max_active;
3120
3121 for_each_cwq_cpu(cpu, wq) {
3122 struct global_cwq *gcwq = get_gcwq(cpu);
3123
3124 spin_lock_irq(&gcwq->lock);
3125
3126 if (!(wq->flags & WQ_FREEZABLE) ||
3127 !(gcwq->flags & GCWQ_FREEZING))
3128 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3129
3130 spin_unlock_irq(&gcwq->lock);
3131 }
3132
3133 spin_unlock(&workqueue_lock);
3134}
3135EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3136
3137/**
3138 * workqueue_congested - test whether a workqueue is congested
3139 * @cpu: CPU in question
3140 * @wq: target workqueue
3141 *
3142 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3143 * no synchronization around this function and the test result is
3144 * unreliable and only useful as advisory hints or for debugging.
3145 *
3146 * RETURNS:
3147 * %true if congested, %false otherwise.
3148 */
3149bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3150{
3151 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3152
3153 return !list_empty(&cwq->delayed_works);
3154}
3155EXPORT_SYMBOL_GPL(workqueue_congested);
3156
3157/**
3158 * work_cpu - return the last known associated cpu for @work
3159 * @work: the work of interest
3160 *
3161 * RETURNS:
3162 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3163 */
3164unsigned int work_cpu(struct work_struct *work)
3165{
3166 struct global_cwq *gcwq = get_work_gcwq(work);
3167
3168 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3169}
3170EXPORT_SYMBOL_GPL(work_cpu);
3171
3172/**
3173 * work_busy - test whether a work is currently pending or running
3174 * @work: the work to be tested
3175 *
3176 * Test whether @work is currently pending or running. There is no
3177 * synchronization around this function and the test result is
3178 * unreliable and only useful as advisory hints or for debugging.
3179 * Especially for reentrant wqs, the pending state might hide the
3180 * running state.
3181 *
3182 * RETURNS:
3183 * OR'd bitmask of WORK_BUSY_* bits.
3184 */
3185unsigned int work_busy(struct work_struct *work)
3186{
3187 struct global_cwq *gcwq = get_work_gcwq(work);
3188 unsigned long flags;
3189 unsigned int ret = 0;
3190
3191 if (!gcwq)
3192 return false;
3193
3194 spin_lock_irqsave(&gcwq->lock, flags);
3195
3196 if (work_pending(work))
3197 ret |= WORK_BUSY_PENDING;
3198 if (find_worker_executing_work(gcwq, work))
3199 ret |= WORK_BUSY_RUNNING;
3200
3201 spin_unlock_irqrestore(&gcwq->lock, flags);
3202
3203 return ret;
3204}
3205EXPORT_SYMBOL_GPL(work_busy);
3206
3207/*
3208 * CPU hotplug.
3209 *
3210 * There are two challenges in supporting CPU hotplug. Firstly, there
3211 * are a lot of assumptions on strong associations among work, cwq and
3212 * gcwq which make migrating pending and scheduled works very
3213 * difficult to implement without impacting hot paths. Secondly,
3214 * gcwqs serve mix of short, long and very long running works making
3215 * blocked draining impractical.
3216 *
3217 * This is solved by allowing a gcwq to be detached from CPU, running
3218 * it with unbound (rogue) workers and allowing it to be reattached
3219 * later if the cpu comes back online. A separate thread is created
3220 * to govern a gcwq in such state and is called the trustee of the
3221 * gcwq.
3222 *
3223 * Trustee states and their descriptions.
3224 *
3225 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3226 * new trustee is started with this state.
3227 *
3228 * IN_CHARGE Once started, trustee will enter this state after
3229 * assuming the manager role and making all existing
3230 * workers rogue. DOWN_PREPARE waits for trustee to
3231 * enter this state. After reaching IN_CHARGE, trustee
3232 * tries to execute the pending worklist until it's empty
3233 * and the state is set to BUTCHER, or the state is set
3234 * to RELEASE.
3235 *
3236 * BUTCHER Command state which is set by the cpu callback after
3237 * the cpu has went down. Once this state is set trustee
3238 * knows that there will be no new works on the worklist
3239 * and once the worklist is empty it can proceed to
3240 * killing idle workers.
3241 *
3242 * RELEASE Command state which is set by the cpu callback if the
3243 * cpu down has been canceled or it has come online
3244 * again. After recognizing this state, trustee stops
3245 * trying to drain or butcher and clears ROGUE, rebinds
3246 * all remaining workers back to the cpu and releases
3247 * manager role.
3248 *
3249 * DONE Trustee will enter this state after BUTCHER or RELEASE
3250 * is complete.
3251 *
3252 * trustee CPU draining
3253 * took over down complete
3254 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3255 * | | ^
3256 * | CPU is back online v return workers |
3257 * ----------------> RELEASE --------------
3258 */
3259
3260/**
3261 * trustee_wait_event_timeout - timed event wait for trustee
3262 * @cond: condition to wait for
3263 * @timeout: timeout in jiffies
3264 *
3265 * wait_event_timeout() for trustee to use. Handles locking and
3266 * checks for RELEASE request.
3267 *
3268 * CONTEXT:
3269 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3270 * multiple times. To be used by trustee.
3271 *
3272 * RETURNS:
3273 * Positive indicating left time if @cond is satisfied, 0 if timed
3274 * out, -1 if canceled.
3275 */
3276#define trustee_wait_event_timeout(cond, timeout) ({ \
3277 long __ret = (timeout); \
3278 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3279 __ret) { \
3280 spin_unlock_irq(&gcwq->lock); \
3281 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3282 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3283 __ret); \
3284 spin_lock_irq(&gcwq->lock); \
3285 } \
3286 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3287})
3288
3289/**
3290 * trustee_wait_event - event wait for trustee
3291 * @cond: condition to wait for
3292 *
3293 * wait_event() for trustee to use. Automatically handles locking and
3294 * checks for CANCEL request.
3295 *
3296 * CONTEXT:
3297 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3298 * multiple times. To be used by trustee.
3299 *
3300 * RETURNS:
3301 * 0 if @cond is satisfied, -1 if canceled.
3302 */
3303#define trustee_wait_event(cond) ({ \
3304 long __ret1; \
3305 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3306 __ret1 < 0 ? -1 : 0; \
3307})
3308
3309static int __cpuinit trustee_thread(void *__gcwq)
3310{
3311 struct global_cwq *gcwq = __gcwq;
3312 struct worker *worker;
3313 struct work_struct *work;
3314 struct hlist_node *pos;
3315 long rc;
3316 int i;
3317
3318 BUG_ON(gcwq->cpu != smp_processor_id());
3319
3320 spin_lock_irq(&gcwq->lock);
3321 /*
3322 * Claim the manager position and make all workers rogue.
3323 * Trustee must be bound to the target cpu and can't be
3324 * cancelled.
3325 */
3326 BUG_ON(gcwq->cpu != smp_processor_id());
3327 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3328 BUG_ON(rc < 0);
3329
3330 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3331
3332 list_for_each_entry(worker, &gcwq->idle_list, entry)
3333 worker->flags |= WORKER_ROGUE;
3334
3335 for_each_busy_worker(worker, i, pos, gcwq)
3336 worker->flags |= WORKER_ROGUE;
3337
3338 /*
3339 * Call schedule() so that we cross rq->lock and thus can
3340 * guarantee sched callbacks see the rogue flag. This is
3341 * necessary as scheduler callbacks may be invoked from other
3342 * cpus.
3343 */
3344 spin_unlock_irq(&gcwq->lock);
3345 schedule();
3346 spin_lock_irq(&gcwq->lock);
3347
3348 /*
3349 * Sched callbacks are disabled now. Zap nr_running. After
3350 * this, nr_running stays zero and need_more_worker() and
3351 * keep_working() are always true as long as the worklist is
3352 * not empty.
3353 */
3354 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3355
3356 spin_unlock_irq(&gcwq->lock);
3357 del_timer_sync(&gcwq->idle_timer);
3358 spin_lock_irq(&gcwq->lock);
3359
3360 /*
3361 * We're now in charge. Notify and proceed to drain. We need
3362 * to keep the gcwq running during the whole CPU down
3363 * procedure as other cpu hotunplug callbacks may need to
3364 * flush currently running tasks.
3365 */
3366 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3367 wake_up_all(&gcwq->trustee_wait);
3368
3369 /*
3370 * The original cpu is in the process of dying and may go away
3371 * anytime now. When that happens, we and all workers would
3372 * be migrated to other cpus. Try draining any left work. We
3373 * want to get it over with ASAP - spam rescuers, wake up as
3374 * many idlers as necessary and create new ones till the
3375 * worklist is empty. Note that if the gcwq is frozen, there
3376 * may be frozen works in freezable cwqs. Don't declare
3377 * completion while frozen.
3378 */
3379 while (gcwq->nr_workers != gcwq->nr_idle ||
3380 gcwq->flags & GCWQ_FREEZING ||
3381 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3382 int nr_works = 0;
3383
3384 list_for_each_entry(work, &gcwq->worklist, entry) {
3385 send_mayday(work);
3386 nr_works++;
3387 }
3388
3389 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3390 if (!nr_works--)
3391 break;
3392 wake_up_process(worker->task);
3393 }
3394
3395 if (need_to_create_worker(gcwq)) {
3396 spin_unlock_irq(&gcwq->lock);
3397 worker = create_worker(gcwq, false);
3398 spin_lock_irq(&gcwq->lock);
3399 if (worker) {
3400 worker->flags |= WORKER_ROGUE;
3401 start_worker(worker);
3402 }
3403 }
3404
3405 /* give a breather */
3406 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3407 break;
3408 }
3409
3410 /*
3411 * Either all works have been scheduled and cpu is down, or
3412 * cpu down has already been canceled. Wait for and butcher
3413 * all workers till we're canceled.
3414 */
3415 do {
3416 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3417 while (!list_empty(&gcwq->idle_list))
3418 destroy_worker(list_first_entry(&gcwq->idle_list,
3419 struct worker, entry));
3420 } while (gcwq->nr_workers && rc >= 0);
3421
3422 /*
3423 * At this point, either draining has completed and no worker
3424 * is left, or cpu down has been canceled or the cpu is being
3425 * brought back up. There shouldn't be any idle one left.
3426 * Tell the remaining busy ones to rebind once it finishes the
3427 * currently scheduled works by scheduling the rebind_work.
3428 */
3429 WARN_ON(!list_empty(&gcwq->idle_list));
3430
3431 for_each_busy_worker(worker, i, pos, gcwq) {
3432 struct work_struct *rebind_work = &worker->rebind_work;
3433
3434 /*
3435 * Rebind_work may race with future cpu hotplug
3436 * operations. Use a separate flag to mark that
3437 * rebinding is scheduled.
3438 */
3439 worker->flags |= WORKER_REBIND;
3440 worker->flags &= ~WORKER_ROGUE;
3441
3442 /* queue rebind_work, wq doesn't matter, use the default one */
3443 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3444 work_data_bits(rebind_work)))
3445 continue;
3446
3447 debug_work_activate(rebind_work);
3448 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3449 worker->scheduled.next,
3450 work_color_to_flags(WORK_NO_COLOR));
3451 }
3452
3453 /* relinquish manager role */
3454 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3455
3456 /* notify completion */
3457 gcwq->trustee = NULL;
3458 gcwq->trustee_state = TRUSTEE_DONE;
3459 wake_up_all(&gcwq->trustee_wait);
3460 spin_unlock_irq(&gcwq->lock);
3461 return 0;
3462}
3463
3464/**
3465 * wait_trustee_state - wait for trustee to enter the specified state
3466 * @gcwq: gcwq the trustee of interest belongs to
3467 * @state: target state to wait for
3468 *
3469 * Wait for the trustee to reach @state. DONE is already matched.
3470 *
3471 * CONTEXT:
3472 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3473 * multiple times. To be used by cpu_callback.
3474 */
3475static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3476__releases(&gcwq->lock)
3477__acquires(&gcwq->lock)
3478{
3479 if (!(gcwq->trustee_state == state ||
3480 gcwq->trustee_state == TRUSTEE_DONE)) {
3481 spin_unlock_irq(&gcwq->lock);
3482 __wait_event(gcwq->trustee_wait,
3483 gcwq->trustee_state == state ||
3484 gcwq->trustee_state == TRUSTEE_DONE);
3485 spin_lock_irq(&gcwq->lock);
3486 }
3487}
3488
3489static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3490 unsigned long action,
3491 void *hcpu)
3492{
3493 unsigned int cpu = (unsigned long)hcpu;
3494 struct global_cwq *gcwq = get_gcwq(cpu);
3495 struct task_struct *new_trustee = NULL;
3496 struct worker *uninitialized_var(new_worker);
3497 unsigned long flags;
3498
3499 action &= ~CPU_TASKS_FROZEN;
3500
3501 switch (action) {
3502 case CPU_DOWN_PREPARE:
3503 new_trustee = kthread_create(trustee_thread, gcwq,
3504 "workqueue_trustee/%d\n", cpu);
3505 if (IS_ERR(new_trustee))
3506 return notifier_from_errno(PTR_ERR(new_trustee));
3507 kthread_bind(new_trustee, cpu);
3508 /* fall through */
3509 case CPU_UP_PREPARE:
3510 BUG_ON(gcwq->first_idle);
3511 new_worker = create_worker(gcwq, false);
3512 if (!new_worker) {
3513 if (new_trustee)
3514 kthread_stop(new_trustee);
3515 return NOTIFY_BAD;
3516 }
3517 }
3518
3519 /* some are called w/ irq disabled, don't disturb irq status */
3520 spin_lock_irqsave(&gcwq->lock, flags);
3521
3522 switch (action) {
3523 case CPU_DOWN_PREPARE:
3524 /* initialize trustee and tell it to acquire the gcwq */
3525 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3526 gcwq->trustee = new_trustee;
3527 gcwq->trustee_state = TRUSTEE_START;
3528 wake_up_process(gcwq->trustee);
3529 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3530 /* fall through */
3531 case CPU_UP_PREPARE:
3532 BUG_ON(gcwq->first_idle);
3533 gcwq->first_idle = new_worker;
3534 break;
3535
3536 case CPU_DYING:
3537 /*
3538 * Before this, the trustee and all workers except for
3539 * the ones which are still executing works from
3540 * before the last CPU down must be on the cpu. After
3541 * this, they'll all be diasporas.
3542 */
3543 gcwq->flags |= GCWQ_DISASSOCIATED;
3544 break;
3545
3546 case CPU_POST_DEAD:
3547 gcwq->trustee_state = TRUSTEE_BUTCHER;
3548 /* fall through */
3549 case CPU_UP_CANCELED:
3550 destroy_worker(gcwq->first_idle);
3551 gcwq->first_idle = NULL;
3552 break;
3553
3554 case CPU_DOWN_FAILED:
3555 case CPU_ONLINE:
3556 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3557 if (gcwq->trustee_state != TRUSTEE_DONE) {
3558 gcwq->trustee_state = TRUSTEE_RELEASE;
3559 wake_up_process(gcwq->trustee);
3560 wait_trustee_state(gcwq, TRUSTEE_DONE);
3561 }
3562
3563 /*
3564 * Trustee is done and there might be no worker left.
3565 * Put the first_idle in and request a real manager to
3566 * take a look.
3567 */
3568 spin_unlock_irq(&gcwq->lock);
3569 kthread_bind(gcwq->first_idle->task, cpu);
3570 spin_lock_irq(&gcwq->lock);
3571 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3572 start_worker(gcwq->first_idle);
3573 gcwq->first_idle = NULL;
3574 break;
3575 }
3576
3577 spin_unlock_irqrestore(&gcwq->lock, flags);
3578
3579 return notifier_from_errno(0);
3580}
3581
3582#ifdef CONFIG_SMP
3583
3584struct work_for_cpu {
3585 struct completion completion;
3586 long (*fn)(void *);
3587 void *arg;
3588 long ret;
3589};
3590
3591static int do_work_for_cpu(void *_wfc)
3592{
3593 struct work_for_cpu *wfc = _wfc;
3594 wfc->ret = wfc->fn(wfc->arg);
3595 complete(&wfc->completion);
3596 return 0;
3597}
3598
3599/**
3600 * work_on_cpu - run a function in user context on a particular cpu
3601 * @cpu: the cpu to run on
3602 * @fn: the function to run
3603 * @arg: the function arg
3604 *
3605 * This will return the value @fn returns.
3606 * It is up to the caller to ensure that the cpu doesn't go offline.
3607 * The caller must not hold any locks which would prevent @fn from completing.
3608 */
3609long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3610{
3611 struct task_struct *sub_thread;
3612 struct work_for_cpu wfc = {
3613 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3614 .fn = fn,
3615 .arg = arg,
3616 };
3617
3618 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3619 if (IS_ERR(sub_thread))
3620 return PTR_ERR(sub_thread);
3621 kthread_bind(sub_thread, cpu);
3622 wake_up_process(sub_thread);
3623 wait_for_completion(&wfc.completion);
3624 return wfc.ret;
3625}
3626EXPORT_SYMBOL_GPL(work_on_cpu);
3627#endif /* CONFIG_SMP */
3628
3629#ifdef CONFIG_FREEZER
3630
3631/**
3632 * freeze_workqueues_begin - begin freezing workqueues
3633 *
3634 * Start freezing workqueues. After this function returns, all freezable
3635 * workqueues will queue new works to their frozen_works list instead of
3636 * gcwq->worklist.
3637 *
3638 * CONTEXT:
3639 * Grabs and releases workqueue_lock and gcwq->lock's.
3640 */
3641void freeze_workqueues_begin(void)
3642{
3643 unsigned int cpu;
3644
3645 spin_lock(&workqueue_lock);
3646
3647 BUG_ON(workqueue_freezing);
3648 workqueue_freezing = true;
3649
3650 for_each_gcwq_cpu(cpu) {
3651 struct global_cwq *gcwq = get_gcwq(cpu);
3652 struct workqueue_struct *wq;
3653
3654 spin_lock_irq(&gcwq->lock);
3655
3656 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3657 gcwq->flags |= GCWQ_FREEZING;
3658
3659 list_for_each_entry(wq, &workqueues, list) {
3660 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3661
3662 if (cwq && wq->flags & WQ_FREEZABLE)
3663 cwq->max_active = 0;
3664 }
3665
3666 spin_unlock_irq(&gcwq->lock);
3667 }
3668
3669 spin_unlock(&workqueue_lock);
3670}
3671
3672/**
3673 * freeze_workqueues_busy - are freezable workqueues still busy?
3674 *
3675 * Check whether freezing is complete. This function must be called
3676 * between freeze_workqueues_begin() and thaw_workqueues().
3677 *
3678 * CONTEXT:
3679 * Grabs and releases workqueue_lock.
3680 *
3681 * RETURNS:
3682 * %true if some freezable workqueues are still busy. %false if freezing
3683 * is complete.
3684 */
3685bool freeze_workqueues_busy(void)
3686{
3687 unsigned int cpu;
3688 bool busy = false;
3689
3690 spin_lock(&workqueue_lock);
3691
3692 BUG_ON(!workqueue_freezing);
3693
3694 for_each_gcwq_cpu(cpu) {
3695 struct workqueue_struct *wq;
3696 /*
3697 * nr_active is monotonically decreasing. It's safe
3698 * to peek without lock.
3699 */
3700 list_for_each_entry(wq, &workqueues, list) {
3701 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3702
3703 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3704 continue;
3705
3706 BUG_ON(cwq->nr_active < 0);
3707 if (cwq->nr_active) {
3708 busy = true;
3709 goto out_unlock;
3710 }
3711 }
3712 }
3713out_unlock:
3714 spin_unlock(&workqueue_lock);
3715 return busy;
3716}
3717
3718/**
3719 * thaw_workqueues - thaw workqueues
3720 *
3721 * Thaw workqueues. Normal queueing is restored and all collected
3722 * frozen works are transferred to their respective gcwq worklists.
3723 *
3724 * CONTEXT:
3725 * Grabs and releases workqueue_lock and gcwq->lock's.
3726 */
3727void thaw_workqueues(void)
3728{
3729 unsigned int cpu;
3730
3731 spin_lock(&workqueue_lock);
3732
3733 if (!workqueue_freezing)
3734 goto out_unlock;
3735
3736 for_each_gcwq_cpu(cpu) {
3737 struct global_cwq *gcwq = get_gcwq(cpu);
3738 struct workqueue_struct *wq;
3739
3740 spin_lock_irq(&gcwq->lock);
3741
3742 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3743 gcwq->flags &= ~GCWQ_FREEZING;
3744
3745 list_for_each_entry(wq, &workqueues, list) {
3746 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3747
3748 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3749 continue;
3750
3751 /* restore max_active and repopulate worklist */
3752 cwq->max_active = wq->saved_max_active;
3753
3754 while (!list_empty(&cwq->delayed_works) &&
3755 cwq->nr_active < cwq->max_active)
3756 cwq_activate_first_delayed(cwq);
3757 }
3758
3759 wake_up_worker(gcwq);
3760
3761 spin_unlock_irq(&gcwq->lock);
3762 }
3763
3764 workqueue_freezing = false;
3765out_unlock:
3766 spin_unlock(&workqueue_lock);
3767}
3768#endif /* CONFIG_FREEZER */
3769
3770static int __init init_workqueues(void)
3771{
3772 unsigned int cpu;
3773 int i;
3774
3775 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3776
3777 /* initialize gcwqs */
3778 for_each_gcwq_cpu(cpu) {
3779 struct global_cwq *gcwq = get_gcwq(cpu);
3780
3781 spin_lock_init(&gcwq->lock);
3782 INIT_LIST_HEAD(&gcwq->worklist);
3783 gcwq->cpu = cpu;
3784 gcwq->flags |= GCWQ_DISASSOCIATED;
3785
3786 INIT_LIST_HEAD(&gcwq->idle_list);
3787 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3788 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3789
3790 init_timer_deferrable(&gcwq->idle_timer);
3791 gcwq->idle_timer.function = idle_worker_timeout;
3792 gcwq->idle_timer.data = (unsigned long)gcwq;
3793
3794 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3795 (unsigned long)gcwq);
3796
3797 ida_init(&gcwq->worker_ida);
3798
3799 gcwq->trustee_state = TRUSTEE_DONE;
3800 init_waitqueue_head(&gcwq->trustee_wait);
3801 }
3802
3803 /* create the initial worker */
3804 for_each_online_gcwq_cpu(cpu) {
3805 struct global_cwq *gcwq = get_gcwq(cpu);
3806 struct worker *worker;
3807
3808 if (cpu != WORK_CPU_UNBOUND)
3809 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3810 worker = create_worker(gcwq, true);
3811 BUG_ON(!worker);
3812 spin_lock_irq(&gcwq->lock);
3813 start_worker(worker);
3814 spin_unlock_irq(&gcwq->lock);
3815 }
3816
3817 system_wq = alloc_workqueue("events", 0, 0);
3818 system_long_wq = alloc_workqueue("events_long", 0, 0);
3819 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3820 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3821 WQ_UNBOUND_MAX_ACTIVE);
3822 system_freezable_wq = alloc_workqueue("events_freezable",
3823 WQ_FREEZABLE, 0);
3824 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3825 !system_unbound_wq || !system_freezable_wq);
3826 return 0;
3827}
3828early_initcall(init_workqueues);