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 are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
23 *
24 * Please read Documentation/core-api/workqueue.rst for details.
25 */
26
27#include <linux/export.h>
28#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
38#include <linux/hardirq.h>
39#include <linux/mempolicy.h>
40#include <linux/freezer.h>
41#include <linux/debug_locks.h>
42#include <linux/lockdep.h>
43#include <linux/idr.h>
44#include <linux/jhash.h>
45#include <linux/hashtable.h>
46#include <linux/rculist.h>
47#include <linux/nodemask.h>
48#include <linux/moduleparam.h>
49#include <linux/uaccess.h>
50#include <linux/sched/isolation.h>
51#include <linux/nmi.h>
52
53#include "workqueue_internal.h"
54
55enum {
56 /*
57 * worker_pool flags
58 *
59 * A bound pool is either associated or disassociated with its CPU.
60 * While associated (!DISASSOCIATED), all workers are bound to the
61 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * is in effect.
63 *
64 * While DISASSOCIATED, the cpu may be offline and all workers have
65 * %WORKER_UNBOUND set and concurrency management disabled, and may
66 * be executing on any CPU. The pool behaves as an unbound one.
67 *
68 * Note that DISASSOCIATED should be flipped only while holding
69 * wq_pool_attach_mutex to avoid changing binding state while
70 * worker_attach_to_pool() is in progress.
71 */
72 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
73 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74
75 /* worker flags */
76 WORKER_DIE = 1 << 1, /* die die die */
77 WORKER_IDLE = 1 << 2, /* is idle */
78 WORKER_PREP = 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
82
83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84 WORKER_UNBOUND | WORKER_REBOUND,
85
86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
87
88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
90
91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
93
94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95 /* call for help after 10ms
96 (min two ticks) */
97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99
100 /*
101 * Rescue workers are used only on emergencies and shared by
102 * all cpus. Give MIN_NICE.
103 */
104 RESCUER_NICE_LEVEL = MIN_NICE,
105 HIGHPRI_NICE_LEVEL = MIN_NICE,
106
107 WQ_NAME_LEN = 24,
108};
109
110/*
111 * Structure fields follow one of the following exclusion rules.
112 *
113 * I: Modifiable by initialization/destruction paths and read-only for
114 * everyone else.
115 *
116 * P: Preemption protected. Disabling preemption is enough and should
117 * only be modified and accessed from the local cpu.
118 *
119 * L: pool->lock protected. Access with pool->lock held.
120 *
121 * X: During normal operation, modification requires pool->lock and should
122 * be done only from local cpu. Either disabling preemption on local
123 * cpu or grabbing pool->lock is enough for read access. If
124 * POOL_DISASSOCIATED is set, it's identical to L.
125 *
126 * A: wq_pool_attach_mutex protected.
127 *
128 * PL: wq_pool_mutex protected.
129 *
130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131 *
132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
133 *
134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
135 * sched-RCU for reads.
136 *
137 * WQ: wq->mutex protected.
138 *
139 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
140 *
141 * MD: wq_mayday_lock protected.
142 */
143
144/* struct worker is defined in workqueue_internal.h */
145
146struct worker_pool {
147 spinlock_t lock; /* the pool lock */
148 int cpu; /* I: the associated cpu */
149 int node; /* I: the associated node ID */
150 int id; /* I: pool ID */
151 unsigned int flags; /* X: flags */
152
153 unsigned long watchdog_ts; /* L: watchdog timestamp */
154
155 struct list_head worklist; /* L: list of pending works */
156
157 int nr_workers; /* L: total number of workers */
158 int nr_idle; /* L: currently idle workers */
159
160 struct list_head idle_list; /* X: list of idle workers */
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for workers */
163
164 /* a workers is either on busy_hash or idle_list, or the manager */
165 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
166 /* L: hash of busy workers */
167
168 struct worker *manager; /* L: purely informational */
169 struct list_head workers; /* A: attached workers */
170 struct completion *detach_completion; /* all workers detached */
171
172 struct ida worker_ida; /* worker IDs for task name */
173
174 struct workqueue_attrs *attrs; /* I: worker attributes */
175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
176 int refcnt; /* PL: refcnt for unbound pools */
177
178 /*
179 * The current concurrency level. As it's likely to be accessed
180 * from other CPUs during try_to_wake_up(), put it in a separate
181 * cacheline.
182 */
183 atomic_t nr_running ____cacheline_aligned_in_smp;
184
185 /*
186 * Destruction of pool is sched-RCU protected to allow dereferences
187 * from get_work_pool().
188 */
189 struct rcu_head rcu;
190} ____cacheline_aligned_in_smp;
191
192/*
193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
194 * of work_struct->data are used for flags and the remaining high bits
195 * point to the pwq; thus, pwqs need to be aligned at two's power of the
196 * number of flag bits.
197 */
198struct pool_workqueue {
199 struct worker_pool *pool; /* I: the associated pool */
200 struct workqueue_struct *wq; /* I: the owning workqueue */
201 int work_color; /* L: current color */
202 int flush_color; /* L: flushing color */
203 int refcnt; /* L: reference count */
204 int nr_in_flight[WORK_NR_COLORS];
205 /* L: nr of in_flight works */
206 int nr_active; /* L: nr of active works */
207 int max_active; /* L: max active works */
208 struct list_head delayed_works; /* L: delayed works */
209 struct list_head pwqs_node; /* WR: node on wq->pwqs */
210 struct list_head mayday_node; /* MD: node on wq->maydays */
211
212 /*
213 * Release of unbound pwq is punted to system_wq. See put_pwq()
214 * and pwq_unbound_release_workfn() for details. pool_workqueue
215 * itself is also sched-RCU protected so that the first pwq can be
216 * determined without grabbing wq->mutex.
217 */
218 struct work_struct unbound_release_work;
219 struct rcu_head rcu;
220} __aligned(1 << WORK_STRUCT_FLAG_BITS);
221
222/*
223 * Structure used to wait for workqueue flush.
224 */
225struct wq_flusher {
226 struct list_head list; /* WQ: list of flushers */
227 int flush_color; /* WQ: flush color waiting for */
228 struct completion done; /* flush completion */
229};
230
231struct wq_device;
232
233/*
234 * The externally visible workqueue. It relays the issued work items to
235 * the appropriate worker_pool through its pool_workqueues.
236 */
237struct workqueue_struct {
238 struct list_head pwqs; /* WR: all pwqs of this wq */
239 struct list_head list; /* PR: list of all workqueues */
240
241 struct mutex mutex; /* protects this wq */
242 int work_color; /* WQ: current work color */
243 int flush_color; /* WQ: current flush color */
244 atomic_t nr_pwqs_to_flush; /* flush in progress */
245 struct wq_flusher *first_flusher; /* WQ: first flusher */
246 struct list_head flusher_queue; /* WQ: flush waiters */
247 struct list_head flusher_overflow; /* WQ: flush overflow list */
248
249 struct list_head maydays; /* MD: pwqs requesting rescue */
250 struct worker *rescuer; /* I: rescue worker */
251
252 int nr_drainers; /* WQ: drain in progress */
253 int saved_max_active; /* WQ: saved pwq max_active */
254
255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
257
258#ifdef CONFIG_SYSFS
259 struct wq_device *wq_dev; /* I: for sysfs interface */
260#endif
261#ifdef CONFIG_LOCKDEP
262 struct lockdep_map lockdep_map;
263#endif
264 char name[WQ_NAME_LEN]; /* I: workqueue name */
265
266 /*
267 * Destruction of workqueue_struct is sched-RCU protected to allow
268 * walking the workqueues list without grabbing wq_pool_mutex.
269 * This is used to dump all workqueues from sysrq.
270 */
271 struct rcu_head rcu;
272
273 /* hot fields used during command issue, aligned to cacheline */
274 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
275 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
276 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
277};
278
279static struct kmem_cache *pwq_cache;
280
281static cpumask_var_t *wq_numa_possible_cpumask;
282 /* possible CPUs of each node */
283
284static bool wq_disable_numa;
285module_param_named(disable_numa, wq_disable_numa, bool, 0444);
286
287/* see the comment above the definition of WQ_POWER_EFFICIENT */
288static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
289module_param_named(power_efficient, wq_power_efficient, bool, 0444);
290
291static bool wq_online; /* can kworkers be created yet? */
292
293static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
294
295/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
297
298static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
300static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
301static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
302
303static LIST_HEAD(workqueues); /* PR: list of all workqueues */
304static bool workqueue_freezing; /* PL: have wqs started freezing? */
305
306/* PL: allowable cpus for unbound wqs and work items */
307static cpumask_var_t wq_unbound_cpumask;
308
309/* CPU where unbound work was last round robin scheduled from this CPU */
310static DEFINE_PER_CPU(int, wq_rr_cpu_last);
311
312/*
313 * Local execution of unbound work items is no longer guaranteed. The
314 * following always forces round-robin CPU selection on unbound work items
315 * to uncover usages which depend on it.
316 */
317#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
318static bool wq_debug_force_rr_cpu = true;
319#else
320static bool wq_debug_force_rr_cpu = false;
321#endif
322module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
323
324/* the per-cpu worker pools */
325static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
326
327static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
328
329/* PL: hash of all unbound pools keyed by pool->attrs */
330static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
331
332/* I: attributes used when instantiating standard unbound pools on demand */
333static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
334
335/* I: attributes used when instantiating ordered pools on demand */
336static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
337
338struct workqueue_struct *system_wq __read_mostly;
339EXPORT_SYMBOL(system_wq);
340struct workqueue_struct *system_highpri_wq __read_mostly;
341EXPORT_SYMBOL_GPL(system_highpri_wq);
342struct workqueue_struct *system_long_wq __read_mostly;
343EXPORT_SYMBOL_GPL(system_long_wq);
344struct workqueue_struct *system_unbound_wq __read_mostly;
345EXPORT_SYMBOL_GPL(system_unbound_wq);
346struct workqueue_struct *system_freezable_wq __read_mostly;
347EXPORT_SYMBOL_GPL(system_freezable_wq);
348struct workqueue_struct *system_power_efficient_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_power_efficient_wq);
350struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
351EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
352
353static int worker_thread(void *__worker);
354static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
355
356#define CREATE_TRACE_POINTS
357#include <trace/events/workqueue.h>
358
359#define assert_rcu_or_pool_mutex() \
360 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
361 !lockdep_is_held(&wq_pool_mutex), \
362 "sched RCU or wq_pool_mutex should be held")
363
364#define assert_rcu_or_wq_mutex(wq) \
365 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
366 !lockdep_is_held(&wq->mutex), \
367 "sched RCU or wq->mutex should be held")
368
369#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
370 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
371 !lockdep_is_held(&wq->mutex) && \
372 !lockdep_is_held(&wq_pool_mutex), \
373 "sched 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 sched 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 sched 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 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
427 else
428
429#ifdef CONFIG_DEBUG_OBJECTS_WORK
430
431static struct debug_obj_descr work_debug_descr;
432
433static void *work_debug_hint(void *addr)
434{
435 return ((struct work_struct *) addr)->func;
436}
437
438static bool work_is_static_object(void *addr)
439{
440 struct work_struct *work = addr;
441
442 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
443}
444
445/*
446 * fixup_init is called when:
447 * - an active object is initialized
448 */
449static bool work_fixup_init(void *addr, enum debug_obj_state state)
450{
451 struct work_struct *work = addr;
452
453 switch (state) {
454 case ODEBUG_STATE_ACTIVE:
455 cancel_work_sync(work);
456 debug_object_init(work, &work_debug_descr);
457 return true;
458 default:
459 return false;
460 }
461}
462
463/*
464 * fixup_free is called when:
465 * - an active object is freed
466 */
467static bool work_fixup_free(void *addr, enum debug_obj_state state)
468{
469 struct work_struct *work = addr;
470
471 switch (state) {
472 case ODEBUG_STATE_ACTIVE:
473 cancel_work_sync(work);
474 debug_object_free(work, &work_debug_descr);
475 return true;
476 default:
477 return false;
478 }
479}
480
481static struct debug_obj_descr work_debug_descr = {
482 .name = "work_struct",
483 .debug_hint = work_debug_hint,
484 .is_static_object = work_is_static_object,
485 .fixup_init = work_fixup_init,
486 .fixup_free = work_fixup_free,
487};
488
489static inline void debug_work_activate(struct work_struct *work)
490{
491 debug_object_activate(work, &work_debug_descr);
492}
493
494static inline void debug_work_deactivate(struct work_struct *work)
495{
496 debug_object_deactivate(work, &work_debug_descr);
497}
498
499void __init_work(struct work_struct *work, int onstack)
500{
501 if (onstack)
502 debug_object_init_on_stack(work, &work_debug_descr);
503 else
504 debug_object_init(work, &work_debug_descr);
505}
506EXPORT_SYMBOL_GPL(__init_work);
507
508void destroy_work_on_stack(struct work_struct *work)
509{
510 debug_object_free(work, &work_debug_descr);
511}
512EXPORT_SYMBOL_GPL(destroy_work_on_stack);
513
514void destroy_delayed_work_on_stack(struct delayed_work *work)
515{
516 destroy_timer_on_stack(&work->timer);
517 debug_object_free(&work->work, &work_debug_descr);
518}
519EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
520
521#else
522static inline void debug_work_activate(struct work_struct *work) { }
523static inline void debug_work_deactivate(struct work_struct *work) { }
524#endif
525
526/**
527 * worker_pool_assign_id - allocate ID and assing it to @pool
528 * @pool: the pool pointer of interest
529 *
530 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
531 * successfully, -errno on failure.
532 */
533static int worker_pool_assign_id(struct worker_pool *pool)
534{
535 int ret;
536
537 lockdep_assert_held(&wq_pool_mutex);
538
539 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
540 GFP_KERNEL);
541 if (ret >= 0) {
542 pool->id = ret;
543 return 0;
544 }
545 return ret;
546}
547
548/**
549 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
550 * @wq: the target workqueue
551 * @node: the node ID
552 *
553 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
554 * read locked.
555 * If the pwq needs to be used beyond the locking in effect, the caller is
556 * responsible for guaranteeing that the pwq stays online.
557 *
558 * Return: The unbound pool_workqueue for @node.
559 */
560static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
561 int node)
562{
563 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
564
565 /*
566 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
567 * delayed item is pending. The plan is to keep CPU -> NODE
568 * mapping valid and stable across CPU on/offlines. Once that
569 * happens, this workaround can be removed.
570 */
571 if (unlikely(node == NUMA_NO_NODE))
572 return wq->dfl_pwq;
573
574 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
575}
576
577static unsigned int work_color_to_flags(int color)
578{
579 return color << WORK_STRUCT_COLOR_SHIFT;
580}
581
582static int get_work_color(struct work_struct *work)
583{
584 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
585 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
586}
587
588static int work_next_color(int color)
589{
590 return (color + 1) % WORK_NR_COLORS;
591}
592
593/*
594 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
595 * contain the pointer to the queued pwq. Once execution starts, the flag
596 * is cleared and the high bits contain OFFQ flags and pool ID.
597 *
598 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
599 * and clear_work_data() can be used to set the pwq, pool or clear
600 * work->data. These functions should only be called while the work is
601 * owned - ie. while the PENDING bit is set.
602 *
603 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
604 * corresponding to a work. Pool is available once the work has been
605 * queued anywhere after initialization until it is sync canceled. pwq is
606 * available only while the work item is queued.
607 *
608 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
609 * canceled. While being canceled, a work item may have its PENDING set
610 * but stay off timer and worklist for arbitrarily long and nobody should
611 * try to steal the PENDING bit.
612 */
613static inline void set_work_data(struct work_struct *work, unsigned long data,
614 unsigned long flags)
615{
616 WARN_ON_ONCE(!work_pending(work));
617 atomic_long_set(&work->data, data | flags | work_static(work));
618}
619
620static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
621 unsigned long extra_flags)
622{
623 set_work_data(work, (unsigned long)pwq,
624 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
625}
626
627static void set_work_pool_and_keep_pending(struct work_struct *work,
628 int pool_id)
629{
630 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
631 WORK_STRUCT_PENDING);
632}
633
634static void set_work_pool_and_clear_pending(struct work_struct *work,
635 int pool_id)
636{
637 /*
638 * The following wmb is paired with the implied mb in
639 * test_and_set_bit(PENDING) and ensures all updates to @work made
640 * here are visible to and precede any updates by the next PENDING
641 * owner.
642 */
643 smp_wmb();
644 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
645 /*
646 * The following mb guarantees that previous clear of a PENDING bit
647 * will not be reordered with any speculative LOADS or STORES from
648 * work->current_func, which is executed afterwards. This possible
649 * reordering can lead to a missed execution on attempt to qeueue
650 * the same @work. E.g. consider this case:
651 *
652 * CPU#0 CPU#1
653 * ---------------------------- --------------------------------
654 *
655 * 1 STORE event_indicated
656 * 2 queue_work_on() {
657 * 3 test_and_set_bit(PENDING)
658 * 4 } set_..._and_clear_pending() {
659 * 5 set_work_data() # clear bit
660 * 6 smp_mb()
661 * 7 work->current_func() {
662 * 8 LOAD event_indicated
663 * }
664 *
665 * Without an explicit full barrier speculative LOAD on line 8 can
666 * be executed before CPU#0 does STORE on line 1. If that happens,
667 * CPU#0 observes the PENDING bit is still set and new execution of
668 * a @work is not queued in a hope, that CPU#1 will eventually
669 * finish the queued @work. Meanwhile CPU#1 does not see
670 * event_indicated is set, because speculative LOAD was executed
671 * before actual STORE.
672 */
673 smp_mb();
674}
675
676static void clear_work_data(struct work_struct *work)
677{
678 smp_wmb(); /* see set_work_pool_and_clear_pending() */
679 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
680}
681
682static struct pool_workqueue *get_work_pwq(struct work_struct *work)
683{
684 unsigned long data = atomic_long_read(&work->data);
685
686 if (data & WORK_STRUCT_PWQ)
687 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
688 else
689 return NULL;
690}
691
692/**
693 * get_work_pool - return the worker_pool a given work was associated with
694 * @work: the work item of interest
695 *
696 * Pools are created and destroyed under wq_pool_mutex, and allows read
697 * access under sched-RCU read lock. As such, this function should be
698 * called under wq_pool_mutex or with preemption disabled.
699 *
700 * All fields of the returned pool are accessible as long as the above
701 * mentioned locking is in effect. If the returned pool needs to be used
702 * beyond the critical section, the caller is responsible for ensuring the
703 * returned pool is and stays online.
704 *
705 * Return: The worker_pool @work was last associated with. %NULL if none.
706 */
707static struct worker_pool *get_work_pool(struct work_struct *work)
708{
709 unsigned long data = atomic_long_read(&work->data);
710 int pool_id;
711
712 assert_rcu_or_pool_mutex();
713
714 if (data & WORK_STRUCT_PWQ)
715 return ((struct pool_workqueue *)
716 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
717
718 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
719 if (pool_id == WORK_OFFQ_POOL_NONE)
720 return NULL;
721
722 return idr_find(&worker_pool_idr, pool_id);
723}
724
725/**
726 * get_work_pool_id - return the worker pool ID a given work is associated with
727 * @work: the work item of interest
728 *
729 * Return: The worker_pool ID @work was last associated with.
730 * %WORK_OFFQ_POOL_NONE if none.
731 */
732static int get_work_pool_id(struct work_struct *work)
733{
734 unsigned long data = atomic_long_read(&work->data);
735
736 if (data & WORK_STRUCT_PWQ)
737 return ((struct pool_workqueue *)
738 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
739
740 return data >> WORK_OFFQ_POOL_SHIFT;
741}
742
743static void mark_work_canceling(struct work_struct *work)
744{
745 unsigned long pool_id = get_work_pool_id(work);
746
747 pool_id <<= WORK_OFFQ_POOL_SHIFT;
748 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
749}
750
751static bool work_is_canceling(struct work_struct *work)
752{
753 unsigned long data = atomic_long_read(&work->data);
754
755 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
756}
757
758/*
759 * Policy functions. These define the policies on how the global worker
760 * pools are managed. Unless noted otherwise, these functions assume that
761 * they're being called with pool->lock held.
762 */
763
764static bool __need_more_worker(struct worker_pool *pool)
765{
766 return !atomic_read(&pool->nr_running);
767}
768
769/*
770 * Need to wake up a worker? Called from anything but currently
771 * running workers.
772 *
773 * Note that, because unbound workers never contribute to nr_running, this
774 * function will always return %true for unbound pools as long as the
775 * worklist isn't empty.
776 */
777static bool need_more_worker(struct worker_pool *pool)
778{
779 return !list_empty(&pool->worklist) && __need_more_worker(pool);
780}
781
782/* Can I start working? Called from busy but !running workers. */
783static bool may_start_working(struct worker_pool *pool)
784{
785 return pool->nr_idle;
786}
787
788/* Do I need to keep working? Called from currently running workers. */
789static bool keep_working(struct worker_pool *pool)
790{
791 return !list_empty(&pool->worklist) &&
792 atomic_read(&pool->nr_running) <= 1;
793}
794
795/* Do we need a new worker? Called from manager. */
796static bool need_to_create_worker(struct worker_pool *pool)
797{
798 return need_more_worker(pool) && !may_start_working(pool);
799}
800
801/* Do we have too many workers and should some go away? */
802static bool too_many_workers(struct worker_pool *pool)
803{
804 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
805 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
806 int nr_busy = pool->nr_workers - nr_idle;
807
808 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
809}
810
811/*
812 * Wake up functions.
813 */
814
815/* Return the first idle worker. Safe with preemption disabled */
816static struct worker *first_idle_worker(struct worker_pool *pool)
817{
818 if (unlikely(list_empty(&pool->idle_list)))
819 return NULL;
820
821 return list_first_entry(&pool->idle_list, struct worker, entry);
822}
823
824/**
825 * wake_up_worker - wake up an idle worker
826 * @pool: worker pool to wake worker from
827 *
828 * Wake up the first idle worker of @pool.
829 *
830 * CONTEXT:
831 * spin_lock_irq(pool->lock).
832 */
833static void wake_up_worker(struct worker_pool *pool)
834{
835 struct worker *worker = first_idle_worker(pool);
836
837 if (likely(worker))
838 wake_up_process(worker->task);
839}
840
841/**
842 * wq_worker_waking_up - a worker is waking up
843 * @task: task waking up
844 * @cpu: CPU @task is waking up to
845 *
846 * This function is called during try_to_wake_up() when a worker is
847 * being awoken.
848 *
849 * CONTEXT:
850 * spin_lock_irq(rq->lock)
851 */
852void wq_worker_waking_up(struct task_struct *task, int cpu)
853{
854 struct worker *worker = kthread_data(task);
855
856 if (!(worker->flags & WORKER_NOT_RUNNING)) {
857 WARN_ON_ONCE(worker->pool->cpu != cpu);
858 atomic_inc(&worker->pool->nr_running);
859 }
860}
861
862/**
863 * wq_worker_sleeping - a worker is going to sleep
864 * @task: task going to sleep
865 *
866 * This function is called during schedule() when a busy worker is
867 * going to sleep. Worker on the same cpu can be woken up by
868 * returning pointer to its task.
869 *
870 * CONTEXT:
871 * spin_lock_irq(rq->lock)
872 *
873 * Return:
874 * Worker task on @cpu to wake up, %NULL if none.
875 */
876struct task_struct *wq_worker_sleeping(struct task_struct *task)
877{
878 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
879 struct worker_pool *pool;
880
881 /*
882 * Rescuers, which may not have all the fields set up like normal
883 * workers, also reach here, let's not access anything before
884 * checking NOT_RUNNING.
885 */
886 if (worker->flags & WORKER_NOT_RUNNING)
887 return NULL;
888
889 pool = worker->pool;
890
891 /* this can only happen on the local cpu */
892 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
893 return NULL;
894
895 /*
896 * The counterpart of the following dec_and_test, implied mb,
897 * worklist not empty test sequence is in insert_work().
898 * Please read comment there.
899 *
900 * NOT_RUNNING is clear. This means that we're bound to and
901 * running on the local cpu w/ rq lock held and preemption
902 * disabled, which in turn means that none else could be
903 * manipulating idle_list, so dereferencing idle_list without pool
904 * lock is safe.
905 */
906 if (atomic_dec_and_test(&pool->nr_running) &&
907 !list_empty(&pool->worklist))
908 to_wakeup = first_idle_worker(pool);
909 return to_wakeup ? to_wakeup->task : NULL;
910}
911
912/**
913 * worker_set_flags - set worker flags and adjust nr_running accordingly
914 * @worker: self
915 * @flags: flags to set
916 *
917 * Set @flags in @worker->flags and adjust nr_running accordingly.
918 *
919 * CONTEXT:
920 * spin_lock_irq(pool->lock)
921 */
922static inline void worker_set_flags(struct worker *worker, unsigned int flags)
923{
924 struct worker_pool *pool = worker->pool;
925
926 WARN_ON_ONCE(worker->task != current);
927
928 /* If transitioning into NOT_RUNNING, adjust nr_running. */
929 if ((flags & WORKER_NOT_RUNNING) &&
930 !(worker->flags & WORKER_NOT_RUNNING)) {
931 atomic_dec(&pool->nr_running);
932 }
933
934 worker->flags |= flags;
935}
936
937/**
938 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
939 * @worker: self
940 * @flags: flags to clear
941 *
942 * Clear @flags in @worker->flags and adjust nr_running accordingly.
943 *
944 * CONTEXT:
945 * spin_lock_irq(pool->lock)
946 */
947static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
948{
949 struct worker_pool *pool = worker->pool;
950 unsigned int oflags = worker->flags;
951
952 WARN_ON_ONCE(worker->task != current);
953
954 worker->flags &= ~flags;
955
956 /*
957 * If transitioning out of NOT_RUNNING, increment nr_running. Note
958 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
959 * of multiple flags, not a single flag.
960 */
961 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
962 if (!(worker->flags & WORKER_NOT_RUNNING))
963 atomic_inc(&pool->nr_running);
964}
965
966/**
967 * find_worker_executing_work - find worker which is executing a work
968 * @pool: pool of interest
969 * @work: work to find worker for
970 *
971 * Find a worker which is executing @work on @pool by searching
972 * @pool->busy_hash which is keyed by the address of @work. For a worker
973 * to match, its current execution should match the address of @work and
974 * its work function. This is to avoid unwanted dependency between
975 * unrelated work executions through a work item being recycled while still
976 * being executed.
977 *
978 * This is a bit tricky. A work item may be freed once its execution
979 * starts and nothing prevents the freed area from being recycled for
980 * another work item. If the same work item address ends up being reused
981 * before the original execution finishes, workqueue will identify the
982 * recycled work item as currently executing and make it wait until the
983 * current execution finishes, introducing an unwanted dependency.
984 *
985 * This function checks the work item address and work function to avoid
986 * false positives. Note that this isn't complete as one may construct a
987 * work function which can introduce dependency onto itself through a
988 * recycled work item. Well, if somebody wants to shoot oneself in the
989 * foot that badly, there's only so much we can do, and if such deadlock
990 * actually occurs, it should be easy to locate the culprit work function.
991 *
992 * CONTEXT:
993 * spin_lock_irq(pool->lock).
994 *
995 * Return:
996 * Pointer to worker which is executing @work if found, %NULL
997 * otherwise.
998 */
999static struct worker *find_worker_executing_work(struct worker_pool *pool,
1000 struct work_struct *work)
1001{
1002 struct worker *worker;
1003
1004 hash_for_each_possible(pool->busy_hash, worker, hentry,
1005 (unsigned long)work)
1006 if (worker->current_work == work &&
1007 worker->current_func == work->func)
1008 return worker;
1009
1010 return NULL;
1011}
1012
1013/**
1014 * move_linked_works - move linked works to a list
1015 * @work: start of series of works to be scheduled
1016 * @head: target list to append @work to
1017 * @nextp: out parameter for nested worklist walking
1018 *
1019 * Schedule linked works starting from @work to @head. Work series to
1020 * be scheduled starts at @work and includes any consecutive work with
1021 * WORK_STRUCT_LINKED set in its predecessor.
1022 *
1023 * If @nextp is not NULL, it's updated to point to the next work of
1024 * the last scheduled work. This allows move_linked_works() to be
1025 * nested inside outer list_for_each_entry_safe().
1026 *
1027 * CONTEXT:
1028 * spin_lock_irq(pool->lock).
1029 */
1030static void move_linked_works(struct work_struct *work, struct list_head *head,
1031 struct work_struct **nextp)
1032{
1033 struct work_struct *n;
1034
1035 /*
1036 * Linked worklist will always end before the end of the list,
1037 * use NULL for list head.
1038 */
1039 list_for_each_entry_safe_from(work, n, NULL, entry) {
1040 list_move_tail(&work->entry, head);
1041 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1042 break;
1043 }
1044
1045 /*
1046 * If we're already inside safe list traversal and have moved
1047 * multiple works to the scheduled queue, the next position
1048 * needs to be updated.
1049 */
1050 if (nextp)
1051 *nextp = n;
1052}
1053
1054/**
1055 * get_pwq - get an extra reference on the specified pool_workqueue
1056 * @pwq: pool_workqueue to get
1057 *
1058 * Obtain an extra reference on @pwq. The caller should guarantee that
1059 * @pwq has positive refcnt and be holding the matching pool->lock.
1060 */
1061static void get_pwq(struct pool_workqueue *pwq)
1062{
1063 lockdep_assert_held(&pwq->pool->lock);
1064 WARN_ON_ONCE(pwq->refcnt <= 0);
1065 pwq->refcnt++;
1066}
1067
1068/**
1069 * put_pwq - put a pool_workqueue reference
1070 * @pwq: pool_workqueue to put
1071 *
1072 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1073 * destruction. The caller should be holding the matching pool->lock.
1074 */
1075static void put_pwq(struct pool_workqueue *pwq)
1076{
1077 lockdep_assert_held(&pwq->pool->lock);
1078 if (likely(--pwq->refcnt))
1079 return;
1080 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1081 return;
1082 /*
1083 * @pwq can't be released under pool->lock, bounce to
1084 * pwq_unbound_release_workfn(). This never recurses on the same
1085 * pool->lock as this path is taken only for unbound workqueues and
1086 * the release work item is scheduled on a per-cpu workqueue. To
1087 * avoid lockdep warning, unbound pool->locks are given lockdep
1088 * subclass of 1 in get_unbound_pool().
1089 */
1090 schedule_work(&pwq->unbound_release_work);
1091}
1092
1093/**
1094 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1095 * @pwq: pool_workqueue to put (can be %NULL)
1096 *
1097 * put_pwq() with locking. This function also allows %NULL @pwq.
1098 */
1099static void put_pwq_unlocked(struct pool_workqueue *pwq)
1100{
1101 if (pwq) {
1102 /*
1103 * As both pwqs and pools are sched-RCU protected, the
1104 * following lock operations are safe.
1105 */
1106 spin_lock_irq(&pwq->pool->lock);
1107 put_pwq(pwq);
1108 spin_unlock_irq(&pwq->pool->lock);
1109 }
1110}
1111
1112static void pwq_activate_delayed_work(struct work_struct *work)
1113{
1114 struct pool_workqueue *pwq = get_work_pwq(work);
1115
1116 trace_workqueue_activate_work(work);
1117 if (list_empty(&pwq->pool->worklist))
1118 pwq->pool->watchdog_ts = jiffies;
1119 move_linked_works(work, &pwq->pool->worklist, NULL);
1120 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1121 pwq->nr_active++;
1122}
1123
1124static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1125{
1126 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1127 struct work_struct, entry);
1128
1129 pwq_activate_delayed_work(work);
1130}
1131
1132/**
1133 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1134 * @pwq: pwq of interest
1135 * @color: color of work which left the queue
1136 *
1137 * A work either has completed or is removed from pending queue,
1138 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1139 *
1140 * CONTEXT:
1141 * spin_lock_irq(pool->lock).
1142 */
1143static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1144{
1145 /* uncolored work items don't participate in flushing or nr_active */
1146 if (color == WORK_NO_COLOR)
1147 goto out_put;
1148
1149 pwq->nr_in_flight[color]--;
1150
1151 pwq->nr_active--;
1152 if (!list_empty(&pwq->delayed_works)) {
1153 /* one down, submit a delayed one */
1154 if (pwq->nr_active < pwq->max_active)
1155 pwq_activate_first_delayed(pwq);
1156 }
1157
1158 /* is flush in progress and are we at the flushing tip? */
1159 if (likely(pwq->flush_color != color))
1160 goto out_put;
1161
1162 /* are there still in-flight works? */
1163 if (pwq->nr_in_flight[color])
1164 goto out_put;
1165
1166 /* this pwq is done, clear flush_color */
1167 pwq->flush_color = -1;
1168
1169 /*
1170 * If this was the last pwq, wake up the first flusher. It
1171 * will handle the rest.
1172 */
1173 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1174 complete(&pwq->wq->first_flusher->done);
1175out_put:
1176 put_pwq(pwq);
1177}
1178
1179/**
1180 * try_to_grab_pending - steal work item from worklist and disable irq
1181 * @work: work item to steal
1182 * @is_dwork: @work is a delayed_work
1183 * @flags: place to store irq state
1184 *
1185 * Try to grab PENDING bit of @work. This function can handle @work in any
1186 * stable state - idle, on timer or on worklist.
1187 *
1188 * Return:
1189 * 1 if @work was pending and we successfully stole PENDING
1190 * 0 if @work was idle and we claimed PENDING
1191 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1192 * -ENOENT if someone else is canceling @work, this state may persist
1193 * for arbitrarily long
1194 *
1195 * Note:
1196 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1197 * interrupted while holding PENDING and @work off queue, irq must be
1198 * disabled on entry. This, combined with delayed_work->timer being
1199 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1200 *
1201 * On successful return, >= 0, irq is disabled and the caller is
1202 * responsible for releasing it using local_irq_restore(*@flags).
1203 *
1204 * This function is safe to call from any context including IRQ handler.
1205 */
1206static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1207 unsigned long *flags)
1208{
1209 struct worker_pool *pool;
1210 struct pool_workqueue *pwq;
1211
1212 local_irq_save(*flags);
1213
1214 /* try to steal the timer if it exists */
1215 if (is_dwork) {
1216 struct delayed_work *dwork = to_delayed_work(work);
1217
1218 /*
1219 * dwork->timer is irqsafe. If del_timer() fails, it's
1220 * guaranteed that the timer is not queued anywhere and not
1221 * running on the local CPU.
1222 */
1223 if (likely(del_timer(&dwork->timer)))
1224 return 1;
1225 }
1226
1227 /* try to claim PENDING the normal way */
1228 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1229 return 0;
1230
1231 /*
1232 * The queueing is in progress, or it is already queued. Try to
1233 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1234 */
1235 pool = get_work_pool(work);
1236 if (!pool)
1237 goto fail;
1238
1239 spin_lock(&pool->lock);
1240 /*
1241 * work->data is guaranteed to point to pwq only while the work
1242 * item is queued on pwq->wq, and both updating work->data to point
1243 * to pwq on queueing and to pool on dequeueing are done under
1244 * pwq->pool->lock. This in turn guarantees that, if work->data
1245 * points to pwq which is associated with a locked pool, the work
1246 * item is currently queued on that pool.
1247 */
1248 pwq = get_work_pwq(work);
1249 if (pwq && pwq->pool == pool) {
1250 debug_work_deactivate(work);
1251
1252 /*
1253 * A delayed work item cannot be grabbed directly because
1254 * it might have linked NO_COLOR work items which, if left
1255 * on the delayed_list, will confuse pwq->nr_active
1256 * management later on and cause stall. Make sure the work
1257 * item is activated before grabbing.
1258 */
1259 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1260 pwq_activate_delayed_work(work);
1261
1262 list_del_init(&work->entry);
1263 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1264
1265 /* work->data points to pwq iff queued, point to pool */
1266 set_work_pool_and_keep_pending(work, pool->id);
1267
1268 spin_unlock(&pool->lock);
1269 return 1;
1270 }
1271 spin_unlock(&pool->lock);
1272fail:
1273 local_irq_restore(*flags);
1274 if (work_is_canceling(work))
1275 return -ENOENT;
1276 cpu_relax();
1277 return -EAGAIN;
1278}
1279
1280/**
1281 * insert_work - insert a work into a pool
1282 * @pwq: pwq @work belongs to
1283 * @work: work to insert
1284 * @head: insertion point
1285 * @extra_flags: extra WORK_STRUCT_* flags to set
1286 *
1287 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1288 * work_struct flags.
1289 *
1290 * CONTEXT:
1291 * spin_lock_irq(pool->lock).
1292 */
1293static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1294 struct list_head *head, unsigned int extra_flags)
1295{
1296 struct worker_pool *pool = pwq->pool;
1297
1298 /* we own @work, set data and link */
1299 set_work_pwq(work, pwq, extra_flags);
1300 list_add_tail(&work->entry, head);
1301 get_pwq(pwq);
1302
1303 /*
1304 * Ensure either wq_worker_sleeping() sees the above
1305 * list_add_tail() or we see zero nr_running to avoid workers lying
1306 * around lazily while there are works to be processed.
1307 */
1308 smp_mb();
1309
1310 if (__need_more_worker(pool))
1311 wake_up_worker(pool);
1312}
1313
1314/*
1315 * Test whether @work is being queued from another work executing on the
1316 * same workqueue.
1317 */
1318static bool is_chained_work(struct workqueue_struct *wq)
1319{
1320 struct worker *worker;
1321
1322 worker = current_wq_worker();
1323 /*
1324 * Return %true iff I'm a worker execuing a work item on @wq. If
1325 * I'm @worker, it's safe to dereference it without locking.
1326 */
1327 return worker && worker->current_pwq->wq == wq;
1328}
1329
1330/*
1331 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1332 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1333 * avoid perturbing sensitive tasks.
1334 */
1335static int wq_select_unbound_cpu(int cpu)
1336{
1337 static bool printed_dbg_warning;
1338 int new_cpu;
1339
1340 if (likely(!wq_debug_force_rr_cpu)) {
1341 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1342 return cpu;
1343 } else if (!printed_dbg_warning) {
1344 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1345 printed_dbg_warning = true;
1346 }
1347
1348 if (cpumask_empty(wq_unbound_cpumask))
1349 return cpu;
1350
1351 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1352 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1353 if (unlikely(new_cpu >= nr_cpu_ids)) {
1354 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1355 if (unlikely(new_cpu >= nr_cpu_ids))
1356 return cpu;
1357 }
1358 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1359
1360 return new_cpu;
1361}
1362
1363static void __queue_work(int cpu, struct workqueue_struct *wq,
1364 struct work_struct *work)
1365{
1366 struct pool_workqueue *pwq;
1367 struct worker_pool *last_pool;
1368 struct list_head *worklist;
1369 unsigned int work_flags;
1370 unsigned int req_cpu = cpu;
1371
1372 /*
1373 * While a work item is PENDING && off queue, a task trying to
1374 * steal the PENDING will busy-loop waiting for it to either get
1375 * queued or lose PENDING. Grabbing PENDING and queueing should
1376 * happen with IRQ disabled.
1377 */
1378 lockdep_assert_irqs_disabled();
1379
1380 debug_work_activate(work);
1381
1382 /* if draining, only works from the same workqueue are allowed */
1383 if (unlikely(wq->flags & __WQ_DRAINING) &&
1384 WARN_ON_ONCE(!is_chained_work(wq)))
1385 return;
1386retry:
1387 if (req_cpu == WORK_CPU_UNBOUND)
1388 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1389
1390 /* pwq which will be used unless @work is executing elsewhere */
1391 if (!(wq->flags & WQ_UNBOUND))
1392 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1393 else
1394 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1395
1396 /*
1397 * If @work was previously on a different pool, it might still be
1398 * running there, in which case the work needs to be queued on that
1399 * pool to guarantee non-reentrancy.
1400 */
1401 last_pool = get_work_pool(work);
1402 if (last_pool && last_pool != pwq->pool) {
1403 struct worker *worker;
1404
1405 spin_lock(&last_pool->lock);
1406
1407 worker = find_worker_executing_work(last_pool, work);
1408
1409 if (worker && worker->current_pwq->wq == wq) {
1410 pwq = worker->current_pwq;
1411 } else {
1412 /* meh... not running there, queue here */
1413 spin_unlock(&last_pool->lock);
1414 spin_lock(&pwq->pool->lock);
1415 }
1416 } else {
1417 spin_lock(&pwq->pool->lock);
1418 }
1419
1420 /*
1421 * pwq is determined and locked. For unbound pools, we could have
1422 * raced with pwq release and it could already be dead. If its
1423 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1424 * without another pwq replacing it in the numa_pwq_tbl or while
1425 * work items are executing on it, so the retrying is guaranteed to
1426 * make forward-progress.
1427 */
1428 if (unlikely(!pwq->refcnt)) {
1429 if (wq->flags & WQ_UNBOUND) {
1430 spin_unlock(&pwq->pool->lock);
1431 cpu_relax();
1432 goto retry;
1433 }
1434 /* oops */
1435 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1436 wq->name, cpu);
1437 }
1438
1439 /* pwq determined, queue */
1440 trace_workqueue_queue_work(req_cpu, pwq, work);
1441
1442 if (WARN_ON(!list_empty(&work->entry))) {
1443 spin_unlock(&pwq->pool->lock);
1444 return;
1445 }
1446
1447 pwq->nr_in_flight[pwq->work_color]++;
1448 work_flags = work_color_to_flags(pwq->work_color);
1449
1450 if (likely(pwq->nr_active < pwq->max_active)) {
1451 trace_workqueue_activate_work(work);
1452 pwq->nr_active++;
1453 worklist = &pwq->pool->worklist;
1454 if (list_empty(worklist))
1455 pwq->pool->watchdog_ts = jiffies;
1456 } else {
1457 work_flags |= WORK_STRUCT_DELAYED;
1458 worklist = &pwq->delayed_works;
1459 }
1460
1461 insert_work(pwq, work, worklist, work_flags);
1462
1463 spin_unlock(&pwq->pool->lock);
1464}
1465
1466/**
1467 * queue_work_on - queue work on specific cpu
1468 * @cpu: CPU number to execute work on
1469 * @wq: workqueue to use
1470 * @work: work to queue
1471 *
1472 * We queue the work to a specific CPU, the caller must ensure it
1473 * can't go away.
1474 *
1475 * Return: %false if @work was already on a queue, %true otherwise.
1476 */
1477bool queue_work_on(int cpu, struct workqueue_struct *wq,
1478 struct work_struct *work)
1479{
1480 bool ret = false;
1481 unsigned long flags;
1482
1483 local_irq_save(flags);
1484
1485 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1486 __queue_work(cpu, wq, work);
1487 ret = true;
1488 }
1489
1490 local_irq_restore(flags);
1491 return ret;
1492}
1493EXPORT_SYMBOL(queue_work_on);
1494
1495void delayed_work_timer_fn(struct timer_list *t)
1496{
1497 struct delayed_work *dwork = from_timer(dwork, t, timer);
1498
1499 /* should have been called from irqsafe timer with irq already off */
1500 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1501}
1502EXPORT_SYMBOL(delayed_work_timer_fn);
1503
1504static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1505 struct delayed_work *dwork, unsigned long delay)
1506{
1507 struct timer_list *timer = &dwork->timer;
1508 struct work_struct *work = &dwork->work;
1509
1510 WARN_ON_ONCE(!wq);
1511 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1512 WARN_ON_ONCE(timer_pending(timer));
1513 WARN_ON_ONCE(!list_empty(&work->entry));
1514
1515 /*
1516 * If @delay is 0, queue @dwork->work immediately. This is for
1517 * both optimization and correctness. The earliest @timer can
1518 * expire is on the closest next tick and delayed_work users depend
1519 * on that there's no such delay when @delay is 0.
1520 */
1521 if (!delay) {
1522 __queue_work(cpu, wq, &dwork->work);
1523 return;
1524 }
1525
1526 dwork->wq = wq;
1527 dwork->cpu = cpu;
1528 timer->expires = jiffies + delay;
1529
1530 if (unlikely(cpu != WORK_CPU_UNBOUND))
1531 add_timer_on(timer, cpu);
1532 else
1533 add_timer(timer);
1534}
1535
1536/**
1537 * queue_delayed_work_on - queue work on specific CPU after delay
1538 * @cpu: CPU number to execute work on
1539 * @wq: workqueue to use
1540 * @dwork: work to queue
1541 * @delay: number of jiffies to wait before queueing
1542 *
1543 * Return: %false if @work was already on a queue, %true otherwise. If
1544 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1545 * execution.
1546 */
1547bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1548 struct delayed_work *dwork, unsigned long delay)
1549{
1550 struct work_struct *work = &dwork->work;
1551 bool ret = false;
1552 unsigned long flags;
1553
1554 /* read the comment in __queue_work() */
1555 local_irq_save(flags);
1556
1557 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1558 __queue_delayed_work(cpu, wq, dwork, delay);
1559 ret = true;
1560 }
1561
1562 local_irq_restore(flags);
1563 return ret;
1564}
1565EXPORT_SYMBOL(queue_delayed_work_on);
1566
1567/**
1568 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1569 * @cpu: CPU number to execute work on
1570 * @wq: workqueue to use
1571 * @dwork: work to queue
1572 * @delay: number of jiffies to wait before queueing
1573 *
1574 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1575 * modify @dwork's timer so that it expires after @delay. If @delay is
1576 * zero, @work is guaranteed to be scheduled immediately regardless of its
1577 * current state.
1578 *
1579 * Return: %false if @dwork was idle and queued, %true if @dwork was
1580 * pending and its timer was modified.
1581 *
1582 * This function is safe to call from any context including IRQ handler.
1583 * See try_to_grab_pending() for details.
1584 */
1585bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1586 struct delayed_work *dwork, unsigned long delay)
1587{
1588 unsigned long flags;
1589 int ret;
1590
1591 do {
1592 ret = try_to_grab_pending(&dwork->work, true, &flags);
1593 } while (unlikely(ret == -EAGAIN));
1594
1595 if (likely(ret >= 0)) {
1596 __queue_delayed_work(cpu, wq, dwork, delay);
1597 local_irq_restore(flags);
1598 }
1599
1600 /* -ENOENT from try_to_grab_pending() becomes %true */
1601 return ret;
1602}
1603EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1604
1605static void rcu_work_rcufn(struct rcu_head *rcu)
1606{
1607 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1608
1609 /* read the comment in __queue_work() */
1610 local_irq_disable();
1611 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1612 local_irq_enable();
1613}
1614
1615/**
1616 * queue_rcu_work - queue work after a RCU grace period
1617 * @wq: workqueue to use
1618 * @rwork: work to queue
1619 *
1620 * Return: %false if @rwork was already pending, %true otherwise. Note
1621 * that a full RCU grace period is guaranteed only after a %true return.
1622 * While @rwork is guarnateed to be executed after a %false return, the
1623 * execution may happen before a full RCU grace period has passed.
1624 */
1625bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1626{
1627 struct work_struct *work = &rwork->work;
1628
1629 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1630 rwork->wq = wq;
1631 call_rcu(&rwork->rcu, rcu_work_rcufn);
1632 return true;
1633 }
1634
1635 return false;
1636}
1637EXPORT_SYMBOL(queue_rcu_work);
1638
1639/**
1640 * worker_enter_idle - enter idle state
1641 * @worker: worker which is entering idle state
1642 *
1643 * @worker is entering idle state. Update stats and idle timer if
1644 * necessary.
1645 *
1646 * LOCKING:
1647 * spin_lock_irq(pool->lock).
1648 */
1649static void worker_enter_idle(struct worker *worker)
1650{
1651 struct worker_pool *pool = worker->pool;
1652
1653 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1654 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1655 (worker->hentry.next || worker->hentry.pprev)))
1656 return;
1657
1658 /* can't use worker_set_flags(), also called from create_worker() */
1659 worker->flags |= WORKER_IDLE;
1660 pool->nr_idle++;
1661 worker->last_active = jiffies;
1662
1663 /* idle_list is LIFO */
1664 list_add(&worker->entry, &pool->idle_list);
1665
1666 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1667 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1668
1669 /*
1670 * Sanity check nr_running. Because unbind_workers() releases
1671 * pool->lock between setting %WORKER_UNBOUND and zapping
1672 * nr_running, the warning may trigger spuriously. Check iff
1673 * unbind is not in progress.
1674 */
1675 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1676 pool->nr_workers == pool->nr_idle &&
1677 atomic_read(&pool->nr_running));
1678}
1679
1680/**
1681 * worker_leave_idle - leave idle state
1682 * @worker: worker which is leaving idle state
1683 *
1684 * @worker is leaving idle state. Update stats.
1685 *
1686 * LOCKING:
1687 * spin_lock_irq(pool->lock).
1688 */
1689static void worker_leave_idle(struct worker *worker)
1690{
1691 struct worker_pool *pool = worker->pool;
1692
1693 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1694 return;
1695 worker_clr_flags(worker, WORKER_IDLE);
1696 pool->nr_idle--;
1697 list_del_init(&worker->entry);
1698}
1699
1700static struct worker *alloc_worker(int node)
1701{
1702 struct worker *worker;
1703
1704 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1705 if (worker) {
1706 INIT_LIST_HEAD(&worker->entry);
1707 INIT_LIST_HEAD(&worker->scheduled);
1708 INIT_LIST_HEAD(&worker->node);
1709 /* on creation a worker is in !idle && prep state */
1710 worker->flags = WORKER_PREP;
1711 }
1712 return worker;
1713}
1714
1715/**
1716 * worker_attach_to_pool() - attach a worker to a pool
1717 * @worker: worker to be attached
1718 * @pool: the target pool
1719 *
1720 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1721 * cpu-binding of @worker are kept coordinated with the pool across
1722 * cpu-[un]hotplugs.
1723 */
1724static void worker_attach_to_pool(struct worker *worker,
1725 struct worker_pool *pool)
1726{
1727 mutex_lock(&wq_pool_attach_mutex);
1728
1729 /*
1730 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1731 * online CPUs. It'll be re-applied when any of the CPUs come up.
1732 */
1733 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1734
1735 /*
1736 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1737 * stable across this function. See the comments above the flag
1738 * definition for details.
1739 */
1740 if (pool->flags & POOL_DISASSOCIATED)
1741 worker->flags |= WORKER_UNBOUND;
1742
1743 list_add_tail(&worker->node, &pool->workers);
1744 worker->pool = pool;
1745
1746 mutex_unlock(&wq_pool_attach_mutex);
1747}
1748
1749/**
1750 * worker_detach_from_pool() - detach a worker from its pool
1751 * @worker: worker which is attached to its pool
1752 *
1753 * Undo the attaching which had been done in worker_attach_to_pool(). The
1754 * caller worker shouldn't access to the pool after detached except it has
1755 * other reference to the pool.
1756 */
1757static void worker_detach_from_pool(struct worker *worker)
1758{
1759 struct worker_pool *pool = worker->pool;
1760 struct completion *detach_completion = NULL;
1761
1762 mutex_lock(&wq_pool_attach_mutex);
1763
1764 list_del(&worker->node);
1765 worker->pool = NULL;
1766
1767 if (list_empty(&pool->workers))
1768 detach_completion = pool->detach_completion;
1769 mutex_unlock(&wq_pool_attach_mutex);
1770
1771 /* clear leftover flags without pool->lock after it is detached */
1772 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1773
1774 if (detach_completion)
1775 complete(detach_completion);
1776}
1777
1778/**
1779 * create_worker - create a new workqueue worker
1780 * @pool: pool the new worker will belong to
1781 *
1782 * Create and start a new worker which is attached to @pool.
1783 *
1784 * CONTEXT:
1785 * Might sleep. Does GFP_KERNEL allocations.
1786 *
1787 * Return:
1788 * Pointer to the newly created worker.
1789 */
1790static struct worker *create_worker(struct worker_pool *pool)
1791{
1792 struct worker *worker = NULL;
1793 int id = -1;
1794 char id_buf[16];
1795
1796 /* ID is needed to determine kthread name */
1797 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1798 if (id < 0)
1799 goto fail;
1800
1801 worker = alloc_worker(pool->node);
1802 if (!worker)
1803 goto fail;
1804
1805 worker->id = id;
1806
1807 if (pool->cpu >= 0)
1808 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1809 pool->attrs->nice < 0 ? "H" : "");
1810 else
1811 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1812
1813 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1814 "kworker/%s", id_buf);
1815 if (IS_ERR(worker->task))
1816 goto fail;
1817
1818 set_user_nice(worker->task, pool->attrs->nice);
1819 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1820
1821 /* successful, attach the worker to the pool */
1822 worker_attach_to_pool(worker, pool);
1823
1824 /* start the newly created worker */
1825 spin_lock_irq(&pool->lock);
1826 worker->pool->nr_workers++;
1827 worker_enter_idle(worker);
1828 wake_up_process(worker->task);
1829 spin_unlock_irq(&pool->lock);
1830
1831 return worker;
1832
1833fail:
1834 if (id >= 0)
1835 ida_simple_remove(&pool->worker_ida, id);
1836 kfree(worker);
1837 return NULL;
1838}
1839
1840/**
1841 * destroy_worker - destroy a workqueue worker
1842 * @worker: worker to be destroyed
1843 *
1844 * Destroy @worker and adjust @pool stats accordingly. The worker should
1845 * be idle.
1846 *
1847 * CONTEXT:
1848 * spin_lock_irq(pool->lock).
1849 */
1850static void destroy_worker(struct worker *worker)
1851{
1852 struct worker_pool *pool = worker->pool;
1853
1854 lockdep_assert_held(&pool->lock);
1855
1856 /* sanity check frenzy */
1857 if (WARN_ON(worker->current_work) ||
1858 WARN_ON(!list_empty(&worker->scheduled)) ||
1859 WARN_ON(!(worker->flags & WORKER_IDLE)))
1860 return;
1861
1862 pool->nr_workers--;
1863 pool->nr_idle--;
1864
1865 list_del_init(&worker->entry);
1866 worker->flags |= WORKER_DIE;
1867 wake_up_process(worker->task);
1868}
1869
1870static void idle_worker_timeout(struct timer_list *t)
1871{
1872 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1873
1874 spin_lock_irq(&pool->lock);
1875
1876 while (too_many_workers(pool)) {
1877 struct worker *worker;
1878 unsigned long expires;
1879
1880 /* idle_list is kept in LIFO order, check the last one */
1881 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1882 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1883
1884 if (time_before(jiffies, expires)) {
1885 mod_timer(&pool->idle_timer, expires);
1886 break;
1887 }
1888
1889 destroy_worker(worker);
1890 }
1891
1892 spin_unlock_irq(&pool->lock);
1893}
1894
1895static void send_mayday(struct work_struct *work)
1896{
1897 struct pool_workqueue *pwq = get_work_pwq(work);
1898 struct workqueue_struct *wq = pwq->wq;
1899
1900 lockdep_assert_held(&wq_mayday_lock);
1901
1902 if (!wq->rescuer)
1903 return;
1904
1905 /* mayday mayday mayday */
1906 if (list_empty(&pwq->mayday_node)) {
1907 /*
1908 * If @pwq is for an unbound wq, its base ref may be put at
1909 * any time due to an attribute change. Pin @pwq until the
1910 * rescuer is done with it.
1911 */
1912 get_pwq(pwq);
1913 list_add_tail(&pwq->mayday_node, &wq->maydays);
1914 wake_up_process(wq->rescuer->task);
1915 }
1916}
1917
1918static void pool_mayday_timeout(struct timer_list *t)
1919{
1920 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
1921 struct work_struct *work;
1922
1923 spin_lock_irq(&pool->lock);
1924 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1925
1926 if (need_to_create_worker(pool)) {
1927 /*
1928 * We've been trying to create a new worker but
1929 * haven't been successful. We might be hitting an
1930 * allocation deadlock. Send distress signals to
1931 * rescuers.
1932 */
1933 list_for_each_entry(work, &pool->worklist, entry)
1934 send_mayday(work);
1935 }
1936
1937 spin_unlock(&wq_mayday_lock);
1938 spin_unlock_irq(&pool->lock);
1939
1940 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1941}
1942
1943/**
1944 * maybe_create_worker - create a new worker if necessary
1945 * @pool: pool to create a new worker for
1946 *
1947 * Create a new worker for @pool if necessary. @pool is guaranteed to
1948 * have at least one idle worker on return from this function. If
1949 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1950 * sent to all rescuers with works scheduled on @pool to resolve
1951 * possible allocation deadlock.
1952 *
1953 * On return, need_to_create_worker() is guaranteed to be %false and
1954 * may_start_working() %true.
1955 *
1956 * LOCKING:
1957 * spin_lock_irq(pool->lock) which may be released and regrabbed
1958 * multiple times. Does GFP_KERNEL allocations. Called only from
1959 * manager.
1960 */
1961static void maybe_create_worker(struct worker_pool *pool)
1962__releases(&pool->lock)
1963__acquires(&pool->lock)
1964{
1965restart:
1966 spin_unlock_irq(&pool->lock);
1967
1968 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1969 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1970
1971 while (true) {
1972 if (create_worker(pool) || !need_to_create_worker(pool))
1973 break;
1974
1975 schedule_timeout_interruptible(CREATE_COOLDOWN);
1976
1977 if (!need_to_create_worker(pool))
1978 break;
1979 }
1980
1981 del_timer_sync(&pool->mayday_timer);
1982 spin_lock_irq(&pool->lock);
1983 /*
1984 * This is necessary even after a new worker was just successfully
1985 * created as @pool->lock was dropped and the new worker might have
1986 * already become busy.
1987 */
1988 if (need_to_create_worker(pool))
1989 goto restart;
1990}
1991
1992/**
1993 * manage_workers - manage worker pool
1994 * @worker: self
1995 *
1996 * Assume the manager role and manage the worker pool @worker belongs
1997 * to. At any given time, there can be only zero or one manager per
1998 * pool. The exclusion is handled automatically by this function.
1999 *
2000 * The caller can safely start processing works on false return. On
2001 * true return, it's guaranteed that need_to_create_worker() is false
2002 * and may_start_working() is true.
2003 *
2004 * CONTEXT:
2005 * spin_lock_irq(pool->lock) which may be released and regrabbed
2006 * multiple times. Does GFP_KERNEL allocations.
2007 *
2008 * Return:
2009 * %false if the pool doesn't need management and the caller can safely
2010 * start processing works, %true if management function was performed and
2011 * the conditions that the caller verified before calling the function may
2012 * no longer be true.
2013 */
2014static bool manage_workers(struct worker *worker)
2015{
2016 struct worker_pool *pool = worker->pool;
2017
2018 if (pool->flags & POOL_MANAGER_ACTIVE)
2019 return false;
2020
2021 pool->flags |= POOL_MANAGER_ACTIVE;
2022 pool->manager = worker;
2023
2024 maybe_create_worker(pool);
2025
2026 pool->manager = NULL;
2027 pool->flags &= ~POOL_MANAGER_ACTIVE;
2028 wake_up(&wq_manager_wait);
2029 return true;
2030}
2031
2032/**
2033 * process_one_work - process single work
2034 * @worker: self
2035 * @work: work to process
2036 *
2037 * Process @work. This function contains all the logics necessary to
2038 * process a single work including synchronization against and
2039 * interaction with other workers on the same cpu, queueing and
2040 * flushing. As long as context requirement is met, any worker can
2041 * call this function to process a work.
2042 *
2043 * CONTEXT:
2044 * spin_lock_irq(pool->lock) which is released and regrabbed.
2045 */
2046static void process_one_work(struct worker *worker, struct work_struct *work)
2047__releases(&pool->lock)
2048__acquires(&pool->lock)
2049{
2050 struct pool_workqueue *pwq = get_work_pwq(work);
2051 struct worker_pool *pool = worker->pool;
2052 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2053 int work_color;
2054 struct worker *collision;
2055#ifdef CONFIG_LOCKDEP
2056 /*
2057 * It is permissible to free the struct work_struct from
2058 * inside the function that is called from it, this we need to
2059 * take into account for lockdep too. To avoid bogus "held
2060 * lock freed" warnings as well as problems when looking into
2061 * work->lockdep_map, make a copy and use that here.
2062 */
2063 struct lockdep_map lockdep_map;
2064
2065 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2066#endif
2067 /* ensure we're on the correct CPU */
2068 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2069 raw_smp_processor_id() != pool->cpu);
2070
2071 /*
2072 * A single work shouldn't be executed concurrently by
2073 * multiple workers on a single cpu. Check whether anyone is
2074 * already processing the work. If so, defer the work to the
2075 * currently executing one.
2076 */
2077 collision = find_worker_executing_work(pool, work);
2078 if (unlikely(collision)) {
2079 move_linked_works(work, &collision->scheduled, NULL);
2080 return;
2081 }
2082
2083 /* claim and dequeue */
2084 debug_work_deactivate(work);
2085 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2086 worker->current_work = work;
2087 worker->current_func = work->func;
2088 worker->current_pwq = pwq;
2089 work_color = get_work_color(work);
2090
2091 /*
2092 * Record wq name for cmdline and debug reporting, may get
2093 * overridden through set_worker_desc().
2094 */
2095 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2096
2097 list_del_init(&work->entry);
2098
2099 /*
2100 * CPU intensive works don't participate in concurrency management.
2101 * They're the scheduler's responsibility. This takes @worker out
2102 * of concurrency management and the next code block will chain
2103 * execution of the pending work items.
2104 */
2105 if (unlikely(cpu_intensive))
2106 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2107
2108 /*
2109 * Wake up another worker if necessary. The condition is always
2110 * false for normal per-cpu workers since nr_running would always
2111 * be >= 1 at this point. This is used to chain execution of the
2112 * pending work items for WORKER_NOT_RUNNING workers such as the
2113 * UNBOUND and CPU_INTENSIVE ones.
2114 */
2115 if (need_more_worker(pool))
2116 wake_up_worker(pool);
2117
2118 /*
2119 * Record the last pool and clear PENDING which should be the last
2120 * update to @work. Also, do this inside @pool->lock so that
2121 * PENDING and queued state changes happen together while IRQ is
2122 * disabled.
2123 */
2124 set_work_pool_and_clear_pending(work, pool->id);
2125
2126 spin_unlock_irq(&pool->lock);
2127
2128 lock_map_acquire(&pwq->wq->lockdep_map);
2129 lock_map_acquire(&lockdep_map);
2130 /*
2131 * Strictly speaking we should mark the invariant state without holding
2132 * any locks, that is, before these two lock_map_acquire()'s.
2133 *
2134 * However, that would result in:
2135 *
2136 * A(W1)
2137 * WFC(C)
2138 * A(W1)
2139 * C(C)
2140 *
2141 * Which would create W1->C->W1 dependencies, even though there is no
2142 * actual deadlock possible. There are two solutions, using a
2143 * read-recursive acquire on the work(queue) 'locks', but this will then
2144 * hit the lockdep limitation on recursive locks, or simply discard
2145 * these locks.
2146 *
2147 * AFAICT there is no possible deadlock scenario between the
2148 * flush_work() and complete() primitives (except for single-threaded
2149 * workqueues), so hiding them isn't a problem.
2150 */
2151 lockdep_invariant_state(true);
2152 trace_workqueue_execute_start(work);
2153 worker->current_func(work);
2154 /*
2155 * While we must be careful to not use "work" after this, the trace
2156 * point will only record its address.
2157 */
2158 trace_workqueue_execute_end(work);
2159 lock_map_release(&lockdep_map);
2160 lock_map_release(&pwq->wq->lockdep_map);
2161
2162 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2163 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2164 " last function: %pf\n",
2165 current->comm, preempt_count(), task_pid_nr(current),
2166 worker->current_func);
2167 debug_show_held_locks(current);
2168 dump_stack();
2169 }
2170
2171 /*
2172 * The following prevents a kworker from hogging CPU on !PREEMPT
2173 * kernels, where a requeueing work item waiting for something to
2174 * happen could deadlock with stop_machine as such work item could
2175 * indefinitely requeue itself while all other CPUs are trapped in
2176 * stop_machine. At the same time, report a quiescent RCU state so
2177 * the same condition doesn't freeze RCU.
2178 */
2179 cond_resched();
2180
2181 spin_lock_irq(&pool->lock);
2182
2183 /* clear cpu intensive status */
2184 if (unlikely(cpu_intensive))
2185 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2186
2187 /* we're done with it, release */
2188 hash_del(&worker->hentry);
2189 worker->current_work = NULL;
2190 worker->current_func = NULL;
2191 worker->current_pwq = NULL;
2192 pwq_dec_nr_in_flight(pwq, work_color);
2193}
2194
2195/**
2196 * process_scheduled_works - process scheduled works
2197 * @worker: self
2198 *
2199 * Process all scheduled works. Please note that the scheduled list
2200 * may change while processing a work, so this function repeatedly
2201 * fetches a work from the top and executes it.
2202 *
2203 * CONTEXT:
2204 * spin_lock_irq(pool->lock) which may be released and regrabbed
2205 * multiple times.
2206 */
2207static void process_scheduled_works(struct worker *worker)
2208{
2209 while (!list_empty(&worker->scheduled)) {
2210 struct work_struct *work = list_first_entry(&worker->scheduled,
2211 struct work_struct, entry);
2212 process_one_work(worker, work);
2213 }
2214}
2215
2216static void set_pf_worker(bool val)
2217{
2218 mutex_lock(&wq_pool_attach_mutex);
2219 if (val)
2220 current->flags |= PF_WQ_WORKER;
2221 else
2222 current->flags &= ~PF_WQ_WORKER;
2223 mutex_unlock(&wq_pool_attach_mutex);
2224}
2225
2226/**
2227 * worker_thread - the worker thread function
2228 * @__worker: self
2229 *
2230 * The worker thread function. All workers belong to a worker_pool -
2231 * either a per-cpu one or dynamic unbound one. These workers process all
2232 * work items regardless of their specific target workqueue. The only
2233 * exception is work items which belong to workqueues with a rescuer which
2234 * will be explained in rescuer_thread().
2235 *
2236 * Return: 0
2237 */
2238static int worker_thread(void *__worker)
2239{
2240 struct worker *worker = __worker;
2241 struct worker_pool *pool = worker->pool;
2242
2243 /* tell the scheduler that this is a workqueue worker */
2244 set_pf_worker(true);
2245woke_up:
2246 spin_lock_irq(&pool->lock);
2247
2248 /* am I supposed to die? */
2249 if (unlikely(worker->flags & WORKER_DIE)) {
2250 spin_unlock_irq(&pool->lock);
2251 WARN_ON_ONCE(!list_empty(&worker->entry));
2252 set_pf_worker(false);
2253
2254 set_task_comm(worker->task, "kworker/dying");
2255 ida_simple_remove(&pool->worker_ida, worker->id);
2256 worker_detach_from_pool(worker);
2257 kfree(worker);
2258 return 0;
2259 }
2260
2261 worker_leave_idle(worker);
2262recheck:
2263 /* no more worker necessary? */
2264 if (!need_more_worker(pool))
2265 goto sleep;
2266
2267 /* do we need to manage? */
2268 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2269 goto recheck;
2270
2271 /*
2272 * ->scheduled list can only be filled while a worker is
2273 * preparing to process a work or actually processing it.
2274 * Make sure nobody diddled with it while I was sleeping.
2275 */
2276 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2277
2278 /*
2279 * Finish PREP stage. We're guaranteed to have at least one idle
2280 * worker or that someone else has already assumed the manager
2281 * role. This is where @worker starts participating in concurrency
2282 * management if applicable and concurrency management is restored
2283 * after being rebound. See rebind_workers() for details.
2284 */
2285 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2286
2287 do {
2288 struct work_struct *work =
2289 list_first_entry(&pool->worklist,
2290 struct work_struct, entry);
2291
2292 pool->watchdog_ts = jiffies;
2293
2294 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2295 /* optimization path, not strictly necessary */
2296 process_one_work(worker, work);
2297 if (unlikely(!list_empty(&worker->scheduled)))
2298 process_scheduled_works(worker);
2299 } else {
2300 move_linked_works(work, &worker->scheduled, NULL);
2301 process_scheduled_works(worker);
2302 }
2303 } while (keep_working(pool));
2304
2305 worker_set_flags(worker, WORKER_PREP);
2306sleep:
2307 /*
2308 * pool->lock is held and there's no work to process and no need to
2309 * manage, sleep. Workers are woken up only while holding
2310 * pool->lock or from local cpu, so setting the current state
2311 * before releasing pool->lock is enough to prevent losing any
2312 * event.
2313 */
2314 worker_enter_idle(worker);
2315 __set_current_state(TASK_IDLE);
2316 spin_unlock_irq(&pool->lock);
2317 schedule();
2318 goto woke_up;
2319}
2320
2321/**
2322 * rescuer_thread - the rescuer thread function
2323 * @__rescuer: self
2324 *
2325 * Workqueue rescuer thread function. There's one rescuer for each
2326 * workqueue which has WQ_MEM_RECLAIM set.
2327 *
2328 * Regular work processing on a pool may block trying to create a new
2329 * worker which uses GFP_KERNEL allocation which has slight chance of
2330 * developing into deadlock if some works currently on the same queue
2331 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2332 * the problem rescuer solves.
2333 *
2334 * When such condition is possible, the pool summons rescuers of all
2335 * workqueues which have works queued on the pool and let them process
2336 * those works so that forward progress can be guaranteed.
2337 *
2338 * This should happen rarely.
2339 *
2340 * Return: 0
2341 */
2342static int rescuer_thread(void *__rescuer)
2343{
2344 struct worker *rescuer = __rescuer;
2345 struct workqueue_struct *wq = rescuer->rescue_wq;
2346 struct list_head *scheduled = &rescuer->scheduled;
2347 bool should_stop;
2348
2349 set_user_nice(current, RESCUER_NICE_LEVEL);
2350
2351 /*
2352 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2353 * doesn't participate in concurrency management.
2354 */
2355 set_pf_worker(true);
2356repeat:
2357 set_current_state(TASK_IDLE);
2358
2359 /*
2360 * By the time the rescuer is requested to stop, the workqueue
2361 * shouldn't have any work pending, but @wq->maydays may still have
2362 * pwq(s) queued. This can happen by non-rescuer workers consuming
2363 * all the work items before the rescuer got to them. Go through
2364 * @wq->maydays processing before acting on should_stop so that the
2365 * list is always empty on exit.
2366 */
2367 should_stop = kthread_should_stop();
2368
2369 /* see whether any pwq is asking for help */
2370 spin_lock_irq(&wq_mayday_lock);
2371
2372 while (!list_empty(&wq->maydays)) {
2373 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2374 struct pool_workqueue, mayday_node);
2375 struct worker_pool *pool = pwq->pool;
2376 struct work_struct *work, *n;
2377 bool first = true;
2378
2379 __set_current_state(TASK_RUNNING);
2380 list_del_init(&pwq->mayday_node);
2381
2382 spin_unlock_irq(&wq_mayday_lock);
2383
2384 worker_attach_to_pool(rescuer, pool);
2385
2386 spin_lock_irq(&pool->lock);
2387
2388 /*
2389 * Slurp in all works issued via this workqueue and
2390 * process'em.
2391 */
2392 WARN_ON_ONCE(!list_empty(scheduled));
2393 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2394 if (get_work_pwq(work) == pwq) {
2395 if (first)
2396 pool->watchdog_ts = jiffies;
2397 move_linked_works(work, scheduled, &n);
2398 }
2399 first = false;
2400 }
2401
2402 if (!list_empty(scheduled)) {
2403 process_scheduled_works(rescuer);
2404
2405 /*
2406 * The above execution of rescued work items could
2407 * have created more to rescue through
2408 * pwq_activate_first_delayed() or chained
2409 * queueing. Let's put @pwq back on mayday list so
2410 * that such back-to-back work items, which may be
2411 * being used to relieve memory pressure, don't
2412 * incur MAYDAY_INTERVAL delay inbetween.
2413 */
2414 if (need_to_create_worker(pool)) {
2415 spin_lock(&wq_mayday_lock);
2416 get_pwq(pwq);
2417 list_move_tail(&pwq->mayday_node, &wq->maydays);
2418 spin_unlock(&wq_mayday_lock);
2419 }
2420 }
2421
2422 /*
2423 * Put the reference grabbed by send_mayday(). @pool won't
2424 * go away while we're still attached to it.
2425 */
2426 put_pwq(pwq);
2427
2428 /*
2429 * Leave this pool. If need_more_worker() is %true, notify a
2430 * regular worker; otherwise, we end up with 0 concurrency
2431 * and stalling the execution.
2432 */
2433 if (need_more_worker(pool))
2434 wake_up_worker(pool);
2435
2436 spin_unlock_irq(&pool->lock);
2437
2438 worker_detach_from_pool(rescuer);
2439
2440 spin_lock_irq(&wq_mayday_lock);
2441 }
2442
2443 spin_unlock_irq(&wq_mayday_lock);
2444
2445 if (should_stop) {
2446 __set_current_state(TASK_RUNNING);
2447 set_pf_worker(false);
2448 return 0;
2449 }
2450
2451 /* rescuers should never participate in concurrency management */
2452 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2453 schedule();
2454 goto repeat;
2455}
2456
2457/**
2458 * check_flush_dependency - check for flush dependency sanity
2459 * @target_wq: workqueue being flushed
2460 * @target_work: work item being flushed (NULL for workqueue flushes)
2461 *
2462 * %current is trying to flush the whole @target_wq or @target_work on it.
2463 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2464 * reclaiming memory or running on a workqueue which doesn't have
2465 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2466 * a deadlock.
2467 */
2468static void check_flush_dependency(struct workqueue_struct *target_wq,
2469 struct work_struct *target_work)
2470{
2471 work_func_t target_func = target_work ? target_work->func : NULL;
2472 struct worker *worker;
2473
2474 if (target_wq->flags & WQ_MEM_RECLAIM)
2475 return;
2476
2477 worker = current_wq_worker();
2478
2479 WARN_ONCE(current->flags & PF_MEMALLOC,
2480 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2481 current->pid, current->comm, target_wq->name, target_func);
2482 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2483 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2484 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2485 worker->current_pwq->wq->name, worker->current_func,
2486 target_wq->name, target_func);
2487}
2488
2489struct wq_barrier {
2490 struct work_struct work;
2491 struct completion done;
2492 struct task_struct *task; /* purely informational */
2493};
2494
2495static void wq_barrier_func(struct work_struct *work)
2496{
2497 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2498 complete(&barr->done);
2499}
2500
2501/**
2502 * insert_wq_barrier - insert a barrier work
2503 * @pwq: pwq to insert barrier into
2504 * @barr: wq_barrier to insert
2505 * @target: target work to attach @barr to
2506 * @worker: worker currently executing @target, NULL if @target is not executing
2507 *
2508 * @barr is linked to @target such that @barr is completed only after
2509 * @target finishes execution. Please note that the ordering
2510 * guarantee is observed only with respect to @target and on the local
2511 * cpu.
2512 *
2513 * Currently, a queued barrier can't be canceled. This is because
2514 * try_to_grab_pending() can't determine whether the work to be
2515 * grabbed is at the head of the queue and thus can't clear LINKED
2516 * flag of the previous work while there must be a valid next work
2517 * after a work with LINKED flag set.
2518 *
2519 * Note that when @worker is non-NULL, @target may be modified
2520 * underneath us, so we can't reliably determine pwq from @target.
2521 *
2522 * CONTEXT:
2523 * spin_lock_irq(pool->lock).
2524 */
2525static void insert_wq_barrier(struct pool_workqueue *pwq,
2526 struct wq_barrier *barr,
2527 struct work_struct *target, struct worker *worker)
2528{
2529 struct list_head *head;
2530 unsigned int linked = 0;
2531
2532 /*
2533 * debugobject calls are safe here even with pool->lock locked
2534 * as we know for sure that this will not trigger any of the
2535 * checks and call back into the fixup functions where we
2536 * might deadlock.
2537 */
2538 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2539 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2540
2541 init_completion_map(&barr->done, &target->lockdep_map);
2542
2543 barr->task = current;
2544
2545 /*
2546 * If @target is currently being executed, schedule the
2547 * barrier to the worker; otherwise, put it after @target.
2548 */
2549 if (worker)
2550 head = worker->scheduled.next;
2551 else {
2552 unsigned long *bits = work_data_bits(target);
2553
2554 head = target->entry.next;
2555 /* there can already be other linked works, inherit and set */
2556 linked = *bits & WORK_STRUCT_LINKED;
2557 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2558 }
2559
2560 debug_work_activate(&barr->work);
2561 insert_work(pwq, &barr->work, head,
2562 work_color_to_flags(WORK_NO_COLOR) | linked);
2563}
2564
2565/**
2566 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2567 * @wq: workqueue being flushed
2568 * @flush_color: new flush color, < 0 for no-op
2569 * @work_color: new work color, < 0 for no-op
2570 *
2571 * Prepare pwqs for workqueue flushing.
2572 *
2573 * If @flush_color is non-negative, flush_color on all pwqs should be
2574 * -1. If no pwq has in-flight commands at the specified color, all
2575 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2576 * has in flight commands, its pwq->flush_color is set to
2577 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2578 * wakeup logic is armed and %true is returned.
2579 *
2580 * The caller should have initialized @wq->first_flusher prior to
2581 * calling this function with non-negative @flush_color. If
2582 * @flush_color is negative, no flush color update is done and %false
2583 * is returned.
2584 *
2585 * If @work_color is non-negative, all pwqs should have the same
2586 * work_color which is previous to @work_color and all will be
2587 * advanced to @work_color.
2588 *
2589 * CONTEXT:
2590 * mutex_lock(wq->mutex).
2591 *
2592 * Return:
2593 * %true if @flush_color >= 0 and there's something to flush. %false
2594 * otherwise.
2595 */
2596static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2597 int flush_color, int work_color)
2598{
2599 bool wait = false;
2600 struct pool_workqueue *pwq;
2601
2602 if (flush_color >= 0) {
2603 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2604 atomic_set(&wq->nr_pwqs_to_flush, 1);
2605 }
2606
2607 for_each_pwq(pwq, wq) {
2608 struct worker_pool *pool = pwq->pool;
2609
2610 spin_lock_irq(&pool->lock);
2611
2612 if (flush_color >= 0) {
2613 WARN_ON_ONCE(pwq->flush_color != -1);
2614
2615 if (pwq->nr_in_flight[flush_color]) {
2616 pwq->flush_color = flush_color;
2617 atomic_inc(&wq->nr_pwqs_to_flush);
2618 wait = true;
2619 }
2620 }
2621
2622 if (work_color >= 0) {
2623 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2624 pwq->work_color = work_color;
2625 }
2626
2627 spin_unlock_irq(&pool->lock);
2628 }
2629
2630 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2631 complete(&wq->first_flusher->done);
2632
2633 return wait;
2634}
2635
2636/**
2637 * flush_workqueue - ensure that any scheduled work has run to completion.
2638 * @wq: workqueue to flush
2639 *
2640 * This function sleeps until all work items which were queued on entry
2641 * have finished execution, but it is not livelocked by new incoming ones.
2642 */
2643void flush_workqueue(struct workqueue_struct *wq)
2644{
2645 struct wq_flusher this_flusher = {
2646 .list = LIST_HEAD_INIT(this_flusher.list),
2647 .flush_color = -1,
2648 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2649 };
2650 int next_color;
2651
2652 if (WARN_ON(!wq_online))
2653 return;
2654
2655 mutex_lock(&wq->mutex);
2656
2657 /*
2658 * Start-to-wait phase
2659 */
2660 next_color = work_next_color(wq->work_color);
2661
2662 if (next_color != wq->flush_color) {
2663 /*
2664 * Color space is not full. The current work_color
2665 * becomes our flush_color and work_color is advanced
2666 * by one.
2667 */
2668 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2669 this_flusher.flush_color = wq->work_color;
2670 wq->work_color = next_color;
2671
2672 if (!wq->first_flusher) {
2673 /* no flush in progress, become the first flusher */
2674 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2675
2676 wq->first_flusher = &this_flusher;
2677
2678 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2679 wq->work_color)) {
2680 /* nothing to flush, done */
2681 wq->flush_color = next_color;
2682 wq->first_flusher = NULL;
2683 goto out_unlock;
2684 }
2685 } else {
2686 /* wait in queue */
2687 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2688 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2689 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2690 }
2691 } else {
2692 /*
2693 * Oops, color space is full, wait on overflow queue.
2694 * The next flush completion will assign us
2695 * flush_color and transfer to flusher_queue.
2696 */
2697 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2698 }
2699
2700 check_flush_dependency(wq, NULL);
2701
2702 mutex_unlock(&wq->mutex);
2703
2704 wait_for_completion(&this_flusher.done);
2705
2706 /*
2707 * Wake-up-and-cascade phase
2708 *
2709 * First flushers are responsible for cascading flushes and
2710 * handling overflow. Non-first flushers can simply return.
2711 */
2712 if (wq->first_flusher != &this_flusher)
2713 return;
2714
2715 mutex_lock(&wq->mutex);
2716
2717 /* we might have raced, check again with mutex held */
2718 if (wq->first_flusher != &this_flusher)
2719 goto out_unlock;
2720
2721 wq->first_flusher = NULL;
2722
2723 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2724 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2725
2726 while (true) {
2727 struct wq_flusher *next, *tmp;
2728
2729 /* complete all the flushers sharing the current flush color */
2730 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2731 if (next->flush_color != wq->flush_color)
2732 break;
2733 list_del_init(&next->list);
2734 complete(&next->done);
2735 }
2736
2737 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2738 wq->flush_color != work_next_color(wq->work_color));
2739
2740 /* this flush_color is finished, advance by one */
2741 wq->flush_color = work_next_color(wq->flush_color);
2742
2743 /* one color has been freed, handle overflow queue */
2744 if (!list_empty(&wq->flusher_overflow)) {
2745 /*
2746 * Assign the same color to all overflowed
2747 * flushers, advance work_color and append to
2748 * flusher_queue. This is the start-to-wait
2749 * phase for these overflowed flushers.
2750 */
2751 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2752 tmp->flush_color = wq->work_color;
2753
2754 wq->work_color = work_next_color(wq->work_color);
2755
2756 list_splice_tail_init(&wq->flusher_overflow,
2757 &wq->flusher_queue);
2758 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2759 }
2760
2761 if (list_empty(&wq->flusher_queue)) {
2762 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2763 break;
2764 }
2765
2766 /*
2767 * Need to flush more colors. Make the next flusher
2768 * the new first flusher and arm pwqs.
2769 */
2770 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2771 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2772
2773 list_del_init(&next->list);
2774 wq->first_flusher = next;
2775
2776 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2777 break;
2778
2779 /*
2780 * Meh... this color is already done, clear first
2781 * flusher and repeat cascading.
2782 */
2783 wq->first_flusher = NULL;
2784 }
2785
2786out_unlock:
2787 mutex_unlock(&wq->mutex);
2788}
2789EXPORT_SYMBOL(flush_workqueue);
2790
2791/**
2792 * drain_workqueue - drain a workqueue
2793 * @wq: workqueue to drain
2794 *
2795 * Wait until the workqueue becomes empty. While draining is in progress,
2796 * only chain queueing is allowed. IOW, only currently pending or running
2797 * work items on @wq can queue further work items on it. @wq is flushed
2798 * repeatedly until it becomes empty. The number of flushing is determined
2799 * by the depth of chaining and should be relatively short. Whine if it
2800 * takes too long.
2801 */
2802void drain_workqueue(struct workqueue_struct *wq)
2803{
2804 unsigned int flush_cnt = 0;
2805 struct pool_workqueue *pwq;
2806
2807 /*
2808 * __queue_work() needs to test whether there are drainers, is much
2809 * hotter than drain_workqueue() and already looks at @wq->flags.
2810 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2811 */
2812 mutex_lock(&wq->mutex);
2813 if (!wq->nr_drainers++)
2814 wq->flags |= __WQ_DRAINING;
2815 mutex_unlock(&wq->mutex);
2816reflush:
2817 flush_workqueue(wq);
2818
2819 mutex_lock(&wq->mutex);
2820
2821 for_each_pwq(pwq, wq) {
2822 bool drained;
2823
2824 spin_lock_irq(&pwq->pool->lock);
2825 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2826 spin_unlock_irq(&pwq->pool->lock);
2827
2828 if (drained)
2829 continue;
2830
2831 if (++flush_cnt == 10 ||
2832 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2833 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2834 wq->name, flush_cnt);
2835
2836 mutex_unlock(&wq->mutex);
2837 goto reflush;
2838 }
2839
2840 if (!--wq->nr_drainers)
2841 wq->flags &= ~__WQ_DRAINING;
2842 mutex_unlock(&wq->mutex);
2843}
2844EXPORT_SYMBOL_GPL(drain_workqueue);
2845
2846static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2847{
2848 struct worker *worker = NULL;
2849 struct worker_pool *pool;
2850 struct pool_workqueue *pwq;
2851
2852 might_sleep();
2853
2854 local_irq_disable();
2855 pool = get_work_pool(work);
2856 if (!pool) {
2857 local_irq_enable();
2858 return false;
2859 }
2860
2861 spin_lock(&pool->lock);
2862 /* see the comment in try_to_grab_pending() with the same code */
2863 pwq = get_work_pwq(work);
2864 if (pwq) {
2865 if (unlikely(pwq->pool != pool))
2866 goto already_gone;
2867 } else {
2868 worker = find_worker_executing_work(pool, work);
2869 if (!worker)
2870 goto already_gone;
2871 pwq = worker->current_pwq;
2872 }
2873
2874 check_flush_dependency(pwq->wq, work);
2875
2876 insert_wq_barrier(pwq, barr, work, worker);
2877 spin_unlock_irq(&pool->lock);
2878
2879 /*
2880 * Force a lock recursion deadlock when using flush_work() inside a
2881 * single-threaded or rescuer equipped workqueue.
2882 *
2883 * For single threaded workqueues the deadlock happens when the work
2884 * is after the work issuing the flush_work(). For rescuer equipped
2885 * workqueues the deadlock happens when the rescuer stalls, blocking
2886 * forward progress.
2887 */
2888 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
2889 lock_map_acquire(&pwq->wq->lockdep_map);
2890 lock_map_release(&pwq->wq->lockdep_map);
2891 }
2892
2893 return true;
2894already_gone:
2895 spin_unlock_irq(&pool->lock);
2896 return false;
2897}
2898
2899/**
2900 * flush_work - wait for a work to finish executing the last queueing instance
2901 * @work: the work to flush
2902 *
2903 * Wait until @work has finished execution. @work is guaranteed to be idle
2904 * on return if it hasn't been requeued since flush started.
2905 *
2906 * Return:
2907 * %true if flush_work() waited for the work to finish execution,
2908 * %false if it was already idle.
2909 */
2910bool flush_work(struct work_struct *work)
2911{
2912 struct wq_barrier barr;
2913
2914 if (WARN_ON(!wq_online))
2915 return false;
2916
2917 if (start_flush_work(work, &barr)) {
2918 wait_for_completion(&barr.done);
2919 destroy_work_on_stack(&barr.work);
2920 return true;
2921 } else {
2922 return false;
2923 }
2924}
2925EXPORT_SYMBOL_GPL(flush_work);
2926
2927struct cwt_wait {
2928 wait_queue_entry_t wait;
2929 struct work_struct *work;
2930};
2931
2932static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2933{
2934 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2935
2936 if (cwait->work != key)
2937 return 0;
2938 return autoremove_wake_function(wait, mode, sync, key);
2939}
2940
2941static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2942{
2943 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2944 unsigned long flags;
2945 int ret;
2946
2947 do {
2948 ret = try_to_grab_pending(work, is_dwork, &flags);
2949 /*
2950 * If someone else is already canceling, wait for it to
2951 * finish. flush_work() doesn't work for PREEMPT_NONE
2952 * because we may get scheduled between @work's completion
2953 * and the other canceling task resuming and clearing
2954 * CANCELING - flush_work() will return false immediately
2955 * as @work is no longer busy, try_to_grab_pending() will
2956 * return -ENOENT as @work is still being canceled and the
2957 * other canceling task won't be able to clear CANCELING as
2958 * we're hogging the CPU.
2959 *
2960 * Let's wait for completion using a waitqueue. As this
2961 * may lead to the thundering herd problem, use a custom
2962 * wake function which matches @work along with exclusive
2963 * wait and wakeup.
2964 */
2965 if (unlikely(ret == -ENOENT)) {
2966 struct cwt_wait cwait;
2967
2968 init_wait(&cwait.wait);
2969 cwait.wait.func = cwt_wakefn;
2970 cwait.work = work;
2971
2972 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2973 TASK_UNINTERRUPTIBLE);
2974 if (work_is_canceling(work))
2975 schedule();
2976 finish_wait(&cancel_waitq, &cwait.wait);
2977 }
2978 } while (unlikely(ret < 0));
2979
2980 /* tell other tasks trying to grab @work to back off */
2981 mark_work_canceling(work);
2982 local_irq_restore(flags);
2983
2984 /*
2985 * This allows canceling during early boot. We know that @work
2986 * isn't executing.
2987 */
2988 if (wq_online)
2989 flush_work(work);
2990
2991 clear_work_data(work);
2992
2993 /*
2994 * Paired with prepare_to_wait() above so that either
2995 * waitqueue_active() is visible here or !work_is_canceling() is
2996 * visible there.
2997 */
2998 smp_mb();
2999 if (waitqueue_active(&cancel_waitq))
3000 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3001
3002 return ret;
3003}
3004
3005/**
3006 * cancel_work_sync - cancel a work and wait for it to finish
3007 * @work: the work to cancel
3008 *
3009 * Cancel @work and wait for its execution to finish. This function
3010 * can be used even if the work re-queues itself or migrates to
3011 * another workqueue. On return from this function, @work is
3012 * guaranteed to be not pending or executing on any CPU.
3013 *
3014 * cancel_work_sync(&delayed_work->work) must not be used for
3015 * delayed_work's. Use cancel_delayed_work_sync() instead.
3016 *
3017 * The caller must ensure that the workqueue on which @work was last
3018 * queued can't be destroyed before this function returns.
3019 *
3020 * Return:
3021 * %true if @work was pending, %false otherwise.
3022 */
3023bool cancel_work_sync(struct work_struct *work)
3024{
3025 return __cancel_work_timer(work, false);
3026}
3027EXPORT_SYMBOL_GPL(cancel_work_sync);
3028
3029/**
3030 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3031 * @dwork: the delayed work to flush
3032 *
3033 * Delayed timer is cancelled and the pending work is queued for
3034 * immediate execution. Like flush_work(), this function only
3035 * considers the last queueing instance of @dwork.
3036 *
3037 * Return:
3038 * %true if flush_work() waited for the work to finish execution,
3039 * %false if it was already idle.
3040 */
3041bool flush_delayed_work(struct delayed_work *dwork)
3042{
3043 local_irq_disable();
3044 if (del_timer_sync(&dwork->timer))
3045 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3046 local_irq_enable();
3047 return flush_work(&dwork->work);
3048}
3049EXPORT_SYMBOL(flush_delayed_work);
3050
3051/**
3052 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3053 * @rwork: the rcu work to flush
3054 *
3055 * Return:
3056 * %true if flush_rcu_work() waited for the work to finish execution,
3057 * %false if it was already idle.
3058 */
3059bool flush_rcu_work(struct rcu_work *rwork)
3060{
3061 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3062 rcu_barrier();
3063 flush_work(&rwork->work);
3064 return true;
3065 } else {
3066 return flush_work(&rwork->work);
3067 }
3068}
3069EXPORT_SYMBOL(flush_rcu_work);
3070
3071static bool __cancel_work(struct work_struct *work, bool is_dwork)
3072{
3073 unsigned long flags;
3074 int ret;
3075
3076 do {
3077 ret = try_to_grab_pending(work, is_dwork, &flags);
3078 } while (unlikely(ret == -EAGAIN));
3079
3080 if (unlikely(ret < 0))
3081 return false;
3082
3083 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3084 local_irq_restore(flags);
3085 return ret;
3086}
3087
3088/**
3089 * cancel_delayed_work - cancel a delayed work
3090 * @dwork: delayed_work to cancel
3091 *
3092 * Kill off a pending delayed_work.
3093 *
3094 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3095 * pending.
3096 *
3097 * Note:
3098 * The work callback function may still be running on return, unless
3099 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3100 * use cancel_delayed_work_sync() to wait on it.
3101 *
3102 * This function is safe to call from any context including IRQ handler.
3103 */
3104bool cancel_delayed_work(struct delayed_work *dwork)
3105{
3106 return __cancel_work(&dwork->work, true);
3107}
3108EXPORT_SYMBOL(cancel_delayed_work);
3109
3110/**
3111 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3112 * @dwork: the delayed work cancel
3113 *
3114 * This is cancel_work_sync() for delayed works.
3115 *
3116 * Return:
3117 * %true if @dwork was pending, %false otherwise.
3118 */
3119bool cancel_delayed_work_sync(struct delayed_work *dwork)
3120{
3121 return __cancel_work_timer(&dwork->work, true);
3122}
3123EXPORT_SYMBOL(cancel_delayed_work_sync);
3124
3125/**
3126 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3127 * @func: the function to call
3128 *
3129 * schedule_on_each_cpu() executes @func on each online CPU using the
3130 * system workqueue and blocks until all CPUs have completed.
3131 * schedule_on_each_cpu() is very slow.
3132 *
3133 * Return:
3134 * 0 on success, -errno on failure.
3135 */
3136int schedule_on_each_cpu(work_func_t func)
3137{
3138 int cpu;
3139 struct work_struct __percpu *works;
3140
3141 works = alloc_percpu(struct work_struct);
3142 if (!works)
3143 return -ENOMEM;
3144
3145 get_online_cpus();
3146
3147 for_each_online_cpu(cpu) {
3148 struct work_struct *work = per_cpu_ptr(works, cpu);
3149
3150 INIT_WORK(work, func);
3151 schedule_work_on(cpu, work);
3152 }
3153
3154 for_each_online_cpu(cpu)
3155 flush_work(per_cpu_ptr(works, cpu));
3156
3157 put_online_cpus();
3158 free_percpu(works);
3159 return 0;
3160}
3161
3162/**
3163 * execute_in_process_context - reliably execute the routine with user context
3164 * @fn: the function to execute
3165 * @ew: guaranteed storage for the execute work structure (must
3166 * be available when the work executes)
3167 *
3168 * Executes the function immediately if process context is available,
3169 * otherwise schedules the function for delayed execution.
3170 *
3171 * Return: 0 - function was executed
3172 * 1 - function was scheduled for execution
3173 */
3174int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3175{
3176 if (!in_interrupt()) {
3177 fn(&ew->work);
3178 return 0;
3179 }
3180
3181 INIT_WORK(&ew->work, fn);
3182 schedule_work(&ew->work);
3183
3184 return 1;
3185}
3186EXPORT_SYMBOL_GPL(execute_in_process_context);
3187
3188/**
3189 * free_workqueue_attrs - free a workqueue_attrs
3190 * @attrs: workqueue_attrs to free
3191 *
3192 * Undo alloc_workqueue_attrs().
3193 */
3194void free_workqueue_attrs(struct workqueue_attrs *attrs)
3195{
3196 if (attrs) {
3197 free_cpumask_var(attrs->cpumask);
3198 kfree(attrs);
3199 }
3200}
3201
3202/**
3203 * alloc_workqueue_attrs - allocate a workqueue_attrs
3204 * @gfp_mask: allocation mask to use
3205 *
3206 * Allocate a new workqueue_attrs, initialize with default settings and
3207 * return it.
3208 *
3209 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3210 */
3211struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3212{
3213 struct workqueue_attrs *attrs;
3214
3215 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3216 if (!attrs)
3217 goto fail;
3218 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3219 goto fail;
3220
3221 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3222 return attrs;
3223fail:
3224 free_workqueue_attrs(attrs);
3225 return NULL;
3226}
3227
3228static void copy_workqueue_attrs(struct workqueue_attrs *to,
3229 const struct workqueue_attrs *from)
3230{
3231 to->nice = from->nice;
3232 cpumask_copy(to->cpumask, from->cpumask);
3233 /*
3234 * Unlike hash and equality test, this function doesn't ignore
3235 * ->no_numa as it is used for both pool and wq attrs. Instead,
3236 * get_unbound_pool() explicitly clears ->no_numa after copying.
3237 */
3238 to->no_numa = from->no_numa;
3239}
3240
3241/* hash value of the content of @attr */
3242static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3243{
3244 u32 hash = 0;
3245
3246 hash = jhash_1word(attrs->nice, hash);
3247 hash = jhash(cpumask_bits(attrs->cpumask),
3248 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3249 return hash;
3250}
3251
3252/* content equality test */
3253static bool wqattrs_equal(const struct workqueue_attrs *a,
3254 const struct workqueue_attrs *b)
3255{
3256 if (a->nice != b->nice)
3257 return false;
3258 if (!cpumask_equal(a->cpumask, b->cpumask))
3259 return false;
3260 return true;
3261}
3262
3263/**
3264 * init_worker_pool - initialize a newly zalloc'd worker_pool
3265 * @pool: worker_pool to initialize
3266 *
3267 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3268 *
3269 * Return: 0 on success, -errno on failure. Even on failure, all fields
3270 * inside @pool proper are initialized and put_unbound_pool() can be called
3271 * on @pool safely to release it.
3272 */
3273static int init_worker_pool(struct worker_pool *pool)
3274{
3275 spin_lock_init(&pool->lock);
3276 pool->id = -1;
3277 pool->cpu = -1;
3278 pool->node = NUMA_NO_NODE;
3279 pool->flags |= POOL_DISASSOCIATED;
3280 pool->watchdog_ts = jiffies;
3281 INIT_LIST_HEAD(&pool->worklist);
3282 INIT_LIST_HEAD(&pool->idle_list);
3283 hash_init(pool->busy_hash);
3284
3285 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3286
3287 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3288
3289 INIT_LIST_HEAD(&pool->workers);
3290
3291 ida_init(&pool->worker_ida);
3292 INIT_HLIST_NODE(&pool->hash_node);
3293 pool->refcnt = 1;
3294
3295 /* shouldn't fail above this point */
3296 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3297 if (!pool->attrs)
3298 return -ENOMEM;
3299 return 0;
3300}
3301
3302static void rcu_free_wq(struct rcu_head *rcu)
3303{
3304 struct workqueue_struct *wq =
3305 container_of(rcu, struct workqueue_struct, rcu);
3306
3307 if (!(wq->flags & WQ_UNBOUND))
3308 free_percpu(wq->cpu_pwqs);
3309 else
3310 free_workqueue_attrs(wq->unbound_attrs);
3311
3312 kfree(wq->rescuer);
3313 kfree(wq);
3314}
3315
3316static void rcu_free_pool(struct rcu_head *rcu)
3317{
3318 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3319
3320 ida_destroy(&pool->worker_ida);
3321 free_workqueue_attrs(pool->attrs);
3322 kfree(pool);
3323}
3324
3325/**
3326 * put_unbound_pool - put a worker_pool
3327 * @pool: worker_pool to put
3328 *
3329 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3330 * safe manner. get_unbound_pool() calls this function on its failure path
3331 * and this function should be able to release pools which went through,
3332 * successfully or not, init_worker_pool().
3333 *
3334 * Should be called with wq_pool_mutex held.
3335 */
3336static void put_unbound_pool(struct worker_pool *pool)
3337{
3338 DECLARE_COMPLETION_ONSTACK(detach_completion);
3339 struct worker *worker;
3340
3341 lockdep_assert_held(&wq_pool_mutex);
3342
3343 if (--pool->refcnt)
3344 return;
3345
3346 /* sanity checks */
3347 if (WARN_ON(!(pool->cpu < 0)) ||
3348 WARN_ON(!list_empty(&pool->worklist)))
3349 return;
3350
3351 /* release id and unhash */
3352 if (pool->id >= 0)
3353 idr_remove(&worker_pool_idr, pool->id);
3354 hash_del(&pool->hash_node);
3355
3356 /*
3357 * Become the manager and destroy all workers. This prevents
3358 * @pool's workers from blocking on attach_mutex. We're the last
3359 * manager and @pool gets freed with the flag set.
3360 */
3361 spin_lock_irq(&pool->lock);
3362 wait_event_lock_irq(wq_manager_wait,
3363 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3364 pool->flags |= POOL_MANAGER_ACTIVE;
3365
3366 while ((worker = first_idle_worker(pool)))
3367 destroy_worker(worker);
3368 WARN_ON(pool->nr_workers || pool->nr_idle);
3369 spin_unlock_irq(&pool->lock);
3370
3371 mutex_lock(&wq_pool_attach_mutex);
3372 if (!list_empty(&pool->workers))
3373 pool->detach_completion = &detach_completion;
3374 mutex_unlock(&wq_pool_attach_mutex);
3375
3376 if (pool->detach_completion)
3377 wait_for_completion(pool->detach_completion);
3378
3379 /* shut down the timers */
3380 del_timer_sync(&pool->idle_timer);
3381 del_timer_sync(&pool->mayday_timer);
3382
3383 /* sched-RCU protected to allow dereferences from get_work_pool() */
3384 call_rcu_sched(&pool->rcu, rcu_free_pool);
3385}
3386
3387/**
3388 * get_unbound_pool - get a worker_pool with the specified attributes
3389 * @attrs: the attributes of the worker_pool to get
3390 *
3391 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3392 * reference count and return it. If there already is a matching
3393 * worker_pool, it will be used; otherwise, this function attempts to
3394 * create a new one.
3395 *
3396 * Should be called with wq_pool_mutex held.
3397 *
3398 * Return: On success, a worker_pool with the same attributes as @attrs.
3399 * On failure, %NULL.
3400 */
3401static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3402{
3403 u32 hash = wqattrs_hash(attrs);
3404 struct worker_pool *pool;
3405 int node;
3406 int target_node = NUMA_NO_NODE;
3407
3408 lockdep_assert_held(&wq_pool_mutex);
3409
3410 /* do we already have a matching pool? */
3411 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3412 if (wqattrs_equal(pool->attrs, attrs)) {
3413 pool->refcnt++;
3414 return pool;
3415 }
3416 }
3417
3418 /* if cpumask is contained inside a NUMA node, we belong to that node */
3419 if (wq_numa_enabled) {
3420 for_each_node(node) {
3421 if (cpumask_subset(attrs->cpumask,
3422 wq_numa_possible_cpumask[node])) {
3423 target_node = node;
3424 break;
3425 }
3426 }
3427 }
3428
3429 /* nope, create a new one */
3430 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3431 if (!pool || init_worker_pool(pool) < 0)
3432 goto fail;
3433
3434 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3435 copy_workqueue_attrs(pool->attrs, attrs);
3436 pool->node = target_node;
3437
3438 /*
3439 * no_numa isn't a worker_pool attribute, always clear it. See
3440 * 'struct workqueue_attrs' comments for detail.
3441 */
3442 pool->attrs->no_numa = false;
3443
3444 if (worker_pool_assign_id(pool) < 0)
3445 goto fail;
3446
3447 /* create and start the initial worker */
3448 if (wq_online && !create_worker(pool))
3449 goto fail;
3450
3451 /* install */
3452