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 char *lock_name;
263 struct lock_class_key key;
264 struct lockdep_map lockdep_map;
265#endif
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
267
268 /*
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
272 */
273 struct rcu_head rcu;
274
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279};
280
281static struct kmem_cache *pwq_cache;
282
283static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
285
286static bool wq_disable_numa;
287module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
289/* see the comment above the definition of WQ_POWER_EFFICIENT */
290static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
293static bool wq_online; /* can kworkers be created yet? */
294
295static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
296
297/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
299
300static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
301static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
302static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
303static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
304
305static LIST_HEAD(workqueues); /* PR: list of all workqueues */
306static bool workqueue_freezing; /* PL: have wqs started freezing? */
307
308/* PL: allowable cpus for unbound wqs and work items */
309static cpumask_var_t wq_unbound_cpumask;
310
311/* CPU where unbound work was last round robin scheduled from this CPU */
312static DEFINE_PER_CPU(int, wq_rr_cpu_last);
313
314/*
315 * Local execution of unbound work items is no longer guaranteed. The
316 * following always forces round-robin CPU selection on unbound work items
317 * to uncover usages which depend on it.
318 */
319#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
320static bool wq_debug_force_rr_cpu = true;
321#else
322static bool wq_debug_force_rr_cpu = false;
323#endif
324module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
325
326/* the per-cpu worker pools */
327static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
328
329static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
330
331/* PL: hash of all unbound pools keyed by pool->attrs */
332static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
333
334/* I: attributes used when instantiating standard unbound pools on demand */
335static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
336
337/* I: attributes used when instantiating ordered pools on demand */
338static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
339
340struct workqueue_struct *system_wq __read_mostly;
341EXPORT_SYMBOL(system_wq);
342struct workqueue_struct *system_highpri_wq __read_mostly;
343EXPORT_SYMBOL_GPL(system_highpri_wq);
344struct workqueue_struct *system_long_wq __read_mostly;
345EXPORT_SYMBOL_GPL(system_long_wq);
346struct workqueue_struct *system_unbound_wq __read_mostly;
347EXPORT_SYMBOL_GPL(system_unbound_wq);
348struct workqueue_struct *system_freezable_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_freezable_wq);
350struct workqueue_struct *system_power_efficient_wq __read_mostly;
351EXPORT_SYMBOL_GPL(system_power_efficient_wq);
352struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
353EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
354
355static int worker_thread(void *__worker);
356static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
357
358#define CREATE_TRACE_POINTS
359#include <trace/events/workqueue.h>
360
361#define assert_rcu_or_pool_mutex() \
362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
363 !lockdep_is_held(&wq_pool_mutex), \
364 "sched RCU or wq_pool_mutex should be held")
365
366#define assert_rcu_or_wq_mutex(wq) \
367 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
368 !lockdep_is_held(&wq->mutex), \
369 "sched RCU or wq->mutex should be held")
370
371#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
373 !lockdep_is_held(&wq->mutex) && \
374 !lockdep_is_held(&wq_pool_mutex), \
375 "sched RCU, wq->mutex or wq_pool_mutex should be held")
376
377#define for_each_cpu_worker_pool(pool, cpu) \
378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
380 (pool)++)
381
382/**
383 * for_each_pool - iterate through all worker_pools in the system
384 * @pool: iteration cursor
385 * @pi: integer used for iteration
386 *
387 * This must be called either with wq_pool_mutex held or sched RCU read
388 * locked. If the pool needs to be used beyond the locking in effect, the
389 * caller is responsible for guaranteeing that the pool stays online.
390 *
391 * The if/else clause exists only for the lockdep assertion and can be
392 * ignored.
393 */
394#define for_each_pool(pool, pi) \
395 idr_for_each_entry(&worker_pool_idr, pool, pi) \
396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
397 else
398
399/**
400 * for_each_pool_worker - iterate through all workers of a worker_pool
401 * @worker: iteration cursor
402 * @pool: worker_pool to iterate workers of
403 *
404 * This must be called with wq_pool_attach_mutex.
405 *
406 * The if/else clause exists only for the lockdep assertion and can be
407 * ignored.
408 */
409#define for_each_pool_worker(worker, pool) \
410 list_for_each_entry((worker), &(pool)->workers, node) \
411 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
412 else
413
414/**
415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
416 * @pwq: iteration cursor
417 * @wq: the target workqueue
418 *
419 * This must be called either with wq->mutex held or sched RCU read locked.
420 * If the pwq needs to be used beyond the locking in effect, the caller is
421 * responsible for guaranteeing that the pwq stays online.
422 *
423 * The if/else clause exists only for the lockdep assertion and can be
424 * ignored.
425 */
426#define for_each_pwq(pwq, wq) \
427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
429 else
430
431#ifdef CONFIG_DEBUG_OBJECTS_WORK
432
433static struct debug_obj_descr work_debug_descr;
434
435static void *work_debug_hint(void *addr)
436{
437 return ((struct work_struct *) addr)->func;
438}
439
440static bool work_is_static_object(void *addr)
441{
442 struct work_struct *work = addr;
443
444 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
445}
446
447/*
448 * fixup_init is called when:
449 * - an active object is initialized
450 */
451static bool work_fixup_init(void *addr, enum debug_obj_state state)
452{
453 struct work_struct *work = addr;
454
455 switch (state) {
456 case ODEBUG_STATE_ACTIVE:
457 cancel_work_sync(work);
458 debug_object_init(work, &work_debug_descr);
459 return true;
460 default:
461 return false;
462 }
463}
464
465/*
466 * fixup_free is called when:
467 * - an active object is freed
468 */
469static bool work_fixup_free(void *addr, enum debug_obj_state state)
470{
471 struct work_struct *work = addr;
472
473 switch (state) {
474 case ODEBUG_STATE_ACTIVE:
475 cancel_work_sync(work);
476 debug_object_free(work, &work_debug_descr);
477 return true;
478 default:
479 return false;
480 }
481}
482
483static struct debug_obj_descr work_debug_descr = {
484 .name = "work_struct",
485 .debug_hint = work_debug_hint,
486 .is_static_object = work_is_static_object,
487 .fixup_init = work_fixup_init,
488 .fixup_free = work_fixup_free,
489};
490
491static inline void debug_work_activate(struct work_struct *work)
492{
493 debug_object_activate(work, &work_debug_descr);
494}
495
496static inline void debug_work_deactivate(struct work_struct *work)
497{
498 debug_object_deactivate(work, &work_debug_descr);
499}
500
501void __init_work(struct work_struct *work, int onstack)
502{
503 if (onstack)
504 debug_object_init_on_stack(work, &work_debug_descr);
505 else
506 debug_object_init(work, &work_debug_descr);
507}
508EXPORT_SYMBOL_GPL(__init_work);
509
510void destroy_work_on_stack(struct work_struct *work)
511{
512 debug_object_free(work, &work_debug_descr);
513}
514EXPORT_SYMBOL_GPL(destroy_work_on_stack);
515
516void destroy_delayed_work_on_stack(struct delayed_work *work)
517{
518 destroy_timer_on_stack(&work->timer);
519 debug_object_free(&work->work, &work_debug_descr);
520}
521EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
522
523#else
524static inline void debug_work_activate(struct work_struct *work) { }
525static inline void debug_work_deactivate(struct work_struct *work) { }
526#endif
527
528/**
529 * worker_pool_assign_id - allocate ID and assing it to @pool
530 * @pool: the pool pointer of interest
531 *
532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
533 * successfully, -errno on failure.
534 */
535static int worker_pool_assign_id(struct worker_pool *pool)
536{
537 int ret;
538
539 lockdep_assert_held(&wq_pool_mutex);
540
541 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
542 GFP_KERNEL);
543 if (ret >= 0) {
544 pool->id = ret;
545 return 0;
546 }
547 return ret;
548}
549
550/**
551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
552 * @wq: the target workqueue
553 * @node: the node ID
554 *
555 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
556 * read locked.
557 * If the pwq needs to be used beyond the locking in effect, the caller is
558 * responsible for guaranteeing that the pwq stays online.
559 *
560 * Return: The unbound pool_workqueue for @node.
561 */
562static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
563 int node)
564{
565 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
566
567 /*
568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
569 * delayed item is pending. The plan is to keep CPU -> NODE
570 * mapping valid and stable across CPU on/offlines. Once that
571 * happens, this workaround can be removed.
572 */
573 if (unlikely(node == NUMA_NO_NODE))
574 return wq->dfl_pwq;
575
576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
577}
578
579static unsigned int work_color_to_flags(int color)
580{
581 return color << WORK_STRUCT_COLOR_SHIFT;
582}
583
584static int get_work_color(struct work_struct *work)
585{
586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
587 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
588}
589
590static int work_next_color(int color)
591{
592 return (color + 1) % WORK_NR_COLORS;
593}
594
595/*
596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597 * contain the pointer to the queued pwq. Once execution starts, the flag
598 * is cleared and the high bits contain OFFQ flags and pool ID.
599 *
600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601 * and clear_work_data() can be used to set the pwq, pool or clear
602 * work->data. These functions should only be called while the work is
603 * owned - ie. while the PENDING bit is set.
604 *
605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
606 * corresponding to a work. Pool is available once the work has been
607 * queued anywhere after initialization until it is sync canceled. pwq is
608 * available only while the work item is queued.
609 *
610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
611 * canceled. While being canceled, a work item may have its PENDING set
612 * but stay off timer and worklist for arbitrarily long and nobody should
613 * try to steal the PENDING bit.
614 */
615static inline void set_work_data(struct work_struct *work, unsigned long data,
616 unsigned long flags)
617{
618 WARN_ON_ONCE(!work_pending(work));
619 atomic_long_set(&work->data, data | flags | work_static(work));
620}
621
622static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
623 unsigned long extra_flags)
624{
625 set_work_data(work, (unsigned long)pwq,
626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
627}
628
629static void set_work_pool_and_keep_pending(struct work_struct *work,
630 int pool_id)
631{
632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
633 WORK_STRUCT_PENDING);
634}
635
636static void set_work_pool_and_clear_pending(struct work_struct *work,
637 int pool_id)
638{
639 /*
640 * The following wmb is paired with the implied mb in
641 * test_and_set_bit(PENDING) and ensures all updates to @work made
642 * here are visible to and precede any updates by the next PENDING
643 * owner.
644 */
645 smp_wmb();
646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
647 /*
648 * The following mb guarantees that previous clear of a PENDING bit
649 * will not be reordered with any speculative LOADS or STORES from
650 * work->current_func, which is executed afterwards. This possible
651 * reordering can lead to a missed execution on attempt to queue
652 * the same @work. E.g. consider this case:
653 *
654 * CPU#0 CPU#1
655 * ---------------------------- --------------------------------
656 *
657 * 1 STORE event_indicated
658 * 2 queue_work_on() {
659 * 3 test_and_set_bit(PENDING)
660 * 4 } set_..._and_clear_pending() {
661 * 5 set_work_data() # clear bit
662 * 6 smp_mb()
663 * 7 work->current_func() {
664 * 8 LOAD event_indicated
665 * }
666 *
667 * Without an explicit full barrier speculative LOAD on line 8 can
668 * be executed before CPU#0 does STORE on line 1. If that happens,
669 * CPU#0 observes the PENDING bit is still set and new execution of
670 * a @work is not queued in a hope, that CPU#1 will eventually
671 * finish the queued @work. Meanwhile CPU#1 does not see
672 * event_indicated is set, because speculative LOAD was executed
673 * before actual STORE.
674 */
675 smp_mb();
676}
677
678static void clear_work_data(struct work_struct *work)
679{
680 smp_wmb(); /* see set_work_pool_and_clear_pending() */
681 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
682}
683
684static struct pool_workqueue *get_work_pwq(struct work_struct *work)
685{
686 unsigned long data = atomic_long_read(&work->data);
687
688 if (data & WORK_STRUCT_PWQ)
689 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
690 else
691 return NULL;
692}
693
694/**
695 * get_work_pool - return the worker_pool a given work was associated with
696 * @work: the work item of interest
697 *
698 * Pools are created and destroyed under wq_pool_mutex, and allows read
699 * access under sched-RCU read lock. As such, this function should be
700 * called under wq_pool_mutex or with preemption disabled.
701 *
702 * All fields of the returned pool are accessible as long as the above
703 * mentioned locking is in effect. If the returned pool needs to be used
704 * beyond the critical section, the caller is responsible for ensuring the
705 * returned pool is and stays online.
706 *
707 * Return: The worker_pool @work was last associated with. %NULL if none.
708 */
709static struct worker_pool *get_work_pool(struct work_struct *work)
710{
711 unsigned long data = atomic_long_read(&work->data);
712 int pool_id;
713
714 assert_rcu_or_pool_mutex();
715
716 if (data & WORK_STRUCT_PWQ)
717 return ((struct pool_workqueue *)
718 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
719
720 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
721 if (pool_id == WORK_OFFQ_POOL_NONE)
722 return NULL;
723
724 return idr_find(&worker_pool_idr, pool_id);
725}
726
727/**
728 * get_work_pool_id - return the worker pool ID a given work is associated with
729 * @work: the work item of interest
730 *
731 * Return: The worker_pool ID @work was last associated with.
732 * %WORK_OFFQ_POOL_NONE if none.
733 */
734static int get_work_pool_id(struct work_struct *work)
735{
736 unsigned long data = atomic_long_read(&work->data);
737
738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
741
742 return data >> WORK_OFFQ_POOL_SHIFT;
743}
744
745static void mark_work_canceling(struct work_struct *work)
746{
747 unsigned long pool_id = get_work_pool_id(work);
748
749 pool_id <<= WORK_OFFQ_POOL_SHIFT;
750 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
751}
752
753static bool work_is_canceling(struct work_struct *work)
754{
755 unsigned long data = atomic_long_read(&work->data);
756
757 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
758}
759
760/*
761 * Policy functions. These define the policies on how the global worker
762 * pools are managed. Unless noted otherwise, these functions assume that
763 * they're being called with pool->lock held.
764 */
765
766static bool __need_more_worker(struct worker_pool *pool)
767{
768 return !atomic_read(&pool->nr_running);
769}
770
771/*
772 * Need to wake up a worker? Called from anything but currently
773 * running workers.
774 *
775 * Note that, because unbound workers never contribute to nr_running, this
776 * function will always return %true for unbound pools as long as the
777 * worklist isn't empty.
778 */
779static bool need_more_worker(struct worker_pool *pool)
780{
781 return !list_empty(&pool->worklist) && __need_more_worker(pool);
782}
783
784/* Can I start working? Called from busy but !running workers. */
785static bool may_start_working(struct worker_pool *pool)
786{
787 return pool->nr_idle;
788}
789
790/* Do I need to keep working? Called from currently running workers. */
791static bool keep_working(struct worker_pool *pool)
792{
793 return !list_empty(&pool->worklist) &&
794 atomic_read(&pool->nr_running) <= 1;
795}
796
797/* Do we need a new worker? Called from manager. */
798static bool need_to_create_worker(struct worker_pool *pool)
799{
800 return need_more_worker(pool) && !may_start_working(pool);
801}
802
803/* Do we have too many workers and should some go away? */
804static bool too_many_workers(struct worker_pool *pool)
805{
806 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
807 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
808 int nr_busy = pool->nr_workers - nr_idle;
809
810 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
811}
812
813/*
814 * Wake up functions.
815 */
816
817/* Return the first idle worker. Safe with preemption disabled */
818static struct worker *first_idle_worker(struct worker_pool *pool)
819{
820 if (unlikely(list_empty(&pool->idle_list)))
821 return NULL;
822
823 return list_first_entry(&pool->idle_list, struct worker, entry);
824}
825
826/**
827 * wake_up_worker - wake up an idle worker
828 * @pool: worker pool to wake worker from
829 *
830 * Wake up the first idle worker of @pool.
831 *
832 * CONTEXT:
833 * spin_lock_irq(pool->lock).
834 */
835static void wake_up_worker(struct worker_pool *pool)
836{
837 struct worker *worker = first_idle_worker(pool);
838
839 if (likely(worker))
840 wake_up_process(worker->task);
841}
842
843/**
844 * wq_worker_waking_up - a worker is waking up
845 * @task: task waking up
846 * @cpu: CPU @task is waking up to
847 *
848 * This function is called during try_to_wake_up() when a worker is
849 * being awoken.
850 *
851 * CONTEXT:
852 * spin_lock_irq(rq->lock)
853 */
854void wq_worker_waking_up(struct task_struct *task, int cpu)
855{
856 struct worker *worker = kthread_data(task);
857
858 if (!(worker->flags & WORKER_NOT_RUNNING)) {
859 WARN_ON_ONCE(worker->pool->cpu != cpu);
860 atomic_inc(&worker->pool->nr_running);
861 }
862}
863
864/**
865 * wq_worker_sleeping - a worker is going to sleep
866 * @task: task going to sleep
867 *
868 * This function is called during schedule() when a busy worker is
869 * going to sleep. Worker on the same cpu can be woken up by
870 * returning pointer to its task.
871 *
872 * CONTEXT:
873 * spin_lock_irq(rq->lock)
874 *
875 * Return:
876 * Worker task on @cpu to wake up, %NULL if none.
877 */
878struct task_struct *wq_worker_sleeping(struct task_struct *task)
879{
880 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
881 struct worker_pool *pool;
882
883 /*
884 * Rescuers, which may not have all the fields set up like normal
885 * workers, also reach here, let's not access anything before
886 * checking NOT_RUNNING.
887 */
888 if (worker->flags & WORKER_NOT_RUNNING)
889 return NULL;
890
891 pool = worker->pool;
892
893 /* this can only happen on the local cpu */
894 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
895 return NULL;
896
897 /*
898 * The counterpart of the following dec_and_test, implied mb,
899 * worklist not empty test sequence is in insert_work().
900 * Please read comment there.
901 *
902 * NOT_RUNNING is clear. This means that we're bound to and
903 * running on the local cpu w/ rq lock held and preemption
904 * disabled, which in turn means that none else could be
905 * manipulating idle_list, so dereferencing idle_list without pool
906 * lock is safe.
907 */
908 if (atomic_dec_and_test(&pool->nr_running) &&
909 !list_empty(&pool->worklist))
910 to_wakeup = first_idle_worker(pool);
911 return to_wakeup ? to_wakeup->task : NULL;
912}
913
914/**
915 * wq_worker_last_func - retrieve worker's last work function
916 *
917 * Determine the last function a worker executed. This is called from
918 * the scheduler to get a worker's last known identity.
919 *
920 * CONTEXT:
921 * spin_lock_irq(rq->lock)
922 *
923 * This function is called during schedule() when a kworker is going
924 * to sleep. It's used by psi to identify aggregation workers during
925 * dequeuing, to allow periodic aggregation to shut-off when that
926 * worker is the last task in the system or cgroup to go to sleep.
927 *
928 * As this function doesn't involve any workqueue-related locking, it
929 * only returns stable values when called from inside the scheduler's
930 * queuing and dequeuing paths, when @task, which must be a kworker,
931 * is guaranteed to not be processing any works.
932 *
933 * Return:
934 * The last work function %current executed as a worker, NULL if it
935 * hasn't executed any work yet.
936 */
937work_func_t wq_worker_last_func(struct task_struct *task)
938{
939 struct worker *worker = kthread_data(task);
940
941 return worker->last_func;
942}
943
944/**
945 * worker_set_flags - set worker flags and adjust nr_running accordingly
946 * @worker: self
947 * @flags: flags to set
948 *
949 * Set @flags in @worker->flags and adjust nr_running accordingly.
950 *
951 * CONTEXT:
952 * spin_lock_irq(pool->lock)
953 */
954static inline void worker_set_flags(struct worker *worker, unsigned int flags)
955{
956 struct worker_pool *pool = worker->pool;
957
958 WARN_ON_ONCE(worker->task != current);
959
960 /* If transitioning into NOT_RUNNING, adjust nr_running. */
961 if ((flags & WORKER_NOT_RUNNING) &&
962 !(worker->flags & WORKER_NOT_RUNNING)) {
963 atomic_dec(&pool->nr_running);
964 }
965
966 worker->flags |= flags;
967}
968
969/**
970 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
971 * @worker: self
972 * @flags: flags to clear
973 *
974 * Clear @flags in @worker->flags and adjust nr_running accordingly.
975 *
976 * CONTEXT:
977 * spin_lock_irq(pool->lock)
978 */
979static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
980{
981 struct worker_pool *pool = worker->pool;
982 unsigned int oflags = worker->flags;
983
984 WARN_ON_ONCE(worker->task != current);
985
986 worker->flags &= ~flags;
987
988 /*
989 * If transitioning out of NOT_RUNNING, increment nr_running. Note
990 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
991 * of multiple flags, not a single flag.
992 */
993 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
994 if (!(worker->flags & WORKER_NOT_RUNNING))
995 atomic_inc(&pool->nr_running);
996}
997
998/**
999 * find_worker_executing_work - find worker which is executing a work
1000 * @pool: pool of interest
1001 * @work: work to find worker for
1002 *
1003 * Find a worker which is executing @work on @pool by searching
1004 * @pool->busy_hash which is keyed by the address of @work. For a worker
1005 * to match, its current execution should match the address of @work and
1006 * its work function. This is to avoid unwanted dependency between
1007 * unrelated work executions through a work item being recycled while still
1008 * being executed.
1009 *
1010 * This is a bit tricky. A work item may be freed once its execution
1011 * starts and nothing prevents the freed area from being recycled for
1012 * another work item. If the same work item address ends up being reused
1013 * before the original execution finishes, workqueue will identify the
1014 * recycled work item as currently executing and make it wait until the
1015 * current execution finishes, introducing an unwanted dependency.
1016 *
1017 * This function checks the work item address and work function to avoid
1018 * false positives. Note that this isn't complete as one may construct a
1019 * work function which can introduce dependency onto itself through a
1020 * recycled work item. Well, if somebody wants to shoot oneself in the
1021 * foot that badly, there's only so much we can do, and if such deadlock
1022 * actually occurs, it should be easy to locate the culprit work function.
1023 *
1024 * CONTEXT:
1025 * spin_lock_irq(pool->lock).
1026 *
1027 * Return:
1028 * Pointer to worker which is executing @work if found, %NULL
1029 * otherwise.
1030 */
1031static struct worker *find_worker_executing_work(struct worker_pool *pool,
1032 struct work_struct *work)
1033{
1034 struct worker *worker;
1035
1036 hash_for_each_possible(pool->busy_hash, worker, hentry,
1037 (unsigned long)work)
1038 if (worker->current_work == work &&
1039 worker->current_func == work->func)
1040 return worker;
1041
1042 return NULL;
1043}
1044
1045/**
1046 * move_linked_works - move linked works to a list
1047 * @work: start of series of works to be scheduled
1048 * @head: target list to append @work to
1049 * @nextp: out parameter for nested worklist walking
1050 *
1051 * Schedule linked works starting from @work to @head. Work series to
1052 * be scheduled starts at @work and includes any consecutive work with
1053 * WORK_STRUCT_LINKED set in its predecessor.
1054 *
1055 * If @nextp is not NULL, it's updated to point to the next work of
1056 * the last scheduled work. This allows move_linked_works() to be
1057 * nested inside outer list_for_each_entry_safe().
1058 *
1059 * CONTEXT:
1060 * spin_lock_irq(pool->lock).
1061 */
1062static void move_linked_works(struct work_struct *work, struct list_head *head,
1063 struct work_struct **nextp)
1064{
1065 struct work_struct *n;
1066
1067 /*
1068 * Linked worklist will always end before the end of the list,
1069 * use NULL for list head.
1070 */
1071 list_for_each_entry_safe_from(work, n, NULL, entry) {
1072 list_move_tail(&work->entry, head);
1073 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1074 break;
1075 }
1076
1077 /*
1078 * If we're already inside safe list traversal and have moved
1079 * multiple works to the scheduled queue, the next position
1080 * needs to be updated.
1081 */
1082 if (nextp)
1083 *nextp = n;
1084}
1085
1086/**
1087 * get_pwq - get an extra reference on the specified pool_workqueue
1088 * @pwq: pool_workqueue to get
1089 *
1090 * Obtain an extra reference on @pwq. The caller should guarantee that
1091 * @pwq has positive refcnt and be holding the matching pool->lock.
1092 */
1093static void get_pwq(struct pool_workqueue *pwq)
1094{
1095 lockdep_assert_held(&pwq->pool->lock);
1096 WARN_ON_ONCE(pwq->refcnt <= 0);
1097 pwq->refcnt++;
1098}
1099
1100/**
1101 * put_pwq - put a pool_workqueue reference
1102 * @pwq: pool_workqueue to put
1103 *
1104 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1105 * destruction. The caller should be holding the matching pool->lock.
1106 */
1107static void put_pwq(struct pool_workqueue *pwq)
1108{
1109 lockdep_assert_held(&pwq->pool->lock);
1110 if (likely(--pwq->refcnt))
1111 return;
1112 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1113 return;
1114 /*
1115 * @pwq can't be released under pool->lock, bounce to
1116 * pwq_unbound_release_workfn(). This never recurses on the same
1117 * pool->lock as this path is taken only for unbound workqueues and
1118 * the release work item is scheduled on a per-cpu workqueue. To
1119 * avoid lockdep warning, unbound pool->locks are given lockdep
1120 * subclass of 1 in get_unbound_pool().
1121 */
1122 schedule_work(&pwq->unbound_release_work);
1123}
1124
1125/**
1126 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1127 * @pwq: pool_workqueue to put (can be %NULL)
1128 *
1129 * put_pwq() with locking. This function also allows %NULL @pwq.
1130 */
1131static void put_pwq_unlocked(struct pool_workqueue *pwq)
1132{
1133 if (pwq) {
1134 /*
1135 * As both pwqs and pools are sched-RCU protected, the
1136 * following lock operations are safe.
1137 */
1138 spin_lock_irq(&pwq->pool->lock);
1139 put_pwq(pwq);
1140 spin_unlock_irq(&pwq->pool->lock);
1141 }
1142}
1143
1144static void pwq_activate_delayed_work(struct work_struct *work)
1145{
1146 struct pool_workqueue *pwq = get_work_pwq(work);
1147
1148 trace_workqueue_activate_work(work);
1149 if (list_empty(&pwq->pool->worklist))
1150 pwq->pool->watchdog_ts = jiffies;
1151 move_linked_works(work, &pwq->pool->worklist, NULL);
1152 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1153 pwq->nr_active++;
1154}
1155
1156static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1157{
1158 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1159 struct work_struct, entry);
1160
1161 pwq_activate_delayed_work(work);
1162}
1163
1164/**
1165 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1166 * @pwq: pwq of interest
1167 * @color: color of work which left the queue
1168 *
1169 * A work either has completed or is removed from pending queue,
1170 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1171 *
1172 * CONTEXT:
1173 * spin_lock_irq(pool->lock).
1174 */
1175static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1176{
1177 /* uncolored work items don't participate in flushing or nr_active */
1178 if (color == WORK_NO_COLOR)
1179 goto out_put;
1180
1181 pwq->nr_in_flight[color]--;
1182
1183 pwq->nr_active--;
1184 if (!list_empty(&pwq->delayed_works)) {
1185 /* one down, submit a delayed one */
1186 if (pwq->nr_active < pwq->max_active)
1187 pwq_activate_first_delayed(pwq);
1188 }
1189
1190 /* is flush in progress and are we at the flushing tip? */
1191 if (likely(pwq->flush_color != color))
1192 goto out_put;
1193
1194 /* are there still in-flight works? */
1195 if (pwq->nr_in_flight[color])
1196 goto out_put;
1197
1198 /* this pwq is done, clear flush_color */
1199 pwq->flush_color = -1;
1200
1201 /*
1202 * If this was the last pwq, wake up the first flusher. It
1203 * will handle the rest.
1204 */
1205 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1206 complete(&pwq->wq->first_flusher->done);
1207out_put:
1208 put_pwq(pwq);
1209}
1210
1211/**
1212 * try_to_grab_pending - steal work item from worklist and disable irq
1213 * @work: work item to steal
1214 * @is_dwork: @work is a delayed_work
1215 * @flags: place to store irq state
1216 *
1217 * Try to grab PENDING bit of @work. This function can handle @work in any
1218 * stable state - idle, on timer or on worklist.
1219 *
1220 * Return:
1221 * 1 if @work was pending and we successfully stole PENDING
1222 * 0 if @work was idle and we claimed PENDING
1223 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1224 * -ENOENT if someone else is canceling @work, this state may persist
1225 * for arbitrarily long
1226 *
1227 * Note:
1228 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1229 * interrupted while holding PENDING and @work off queue, irq must be
1230 * disabled on entry. This, combined with delayed_work->timer being
1231 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1232 *
1233 * On successful return, >= 0, irq is disabled and the caller is
1234 * responsible for releasing it using local_irq_restore(*@flags).
1235 *
1236 * This function is safe to call from any context including IRQ handler.
1237 */
1238static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1239 unsigned long *flags)
1240{
1241 struct worker_pool *pool;
1242 struct pool_workqueue *pwq;
1243
1244 local_irq_save(*flags);
1245
1246 /* try to steal the timer if it exists */
1247 if (is_dwork) {
1248 struct delayed_work *dwork = to_delayed_work(work);
1249
1250 /*
1251 * dwork->timer is irqsafe. If del_timer() fails, it's
1252 * guaranteed that the timer is not queued anywhere and not
1253 * running on the local CPU.
1254 */
1255 if (likely(del_timer(&dwork->timer)))
1256 return 1;
1257 }
1258
1259 /* try to claim PENDING the normal way */
1260 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1261 return 0;
1262
1263 /*
1264 * The queueing is in progress, or it is already queued. Try to
1265 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1266 */
1267 pool = get_work_pool(work);
1268 if (!pool)
1269 goto fail;
1270
1271 spin_lock(&pool->lock);
1272 /*
1273 * work->data is guaranteed to point to pwq only while the work
1274 * item is queued on pwq->wq, and both updating work->data to point
1275 * to pwq on queueing and to pool on dequeueing are done under
1276 * pwq->pool->lock. This in turn guarantees that, if work->data
1277 * points to pwq which is associated with a locked pool, the work
1278 * item is currently queued on that pool.
1279 */
1280 pwq = get_work_pwq(work);
1281 if (pwq && pwq->pool == pool) {
1282 debug_work_deactivate(work);
1283
1284 /*
1285 * A delayed work item cannot be grabbed directly because
1286 * it might have linked NO_COLOR work items which, if left
1287 * on the delayed_list, will confuse pwq->nr_active
1288 * management later on and cause stall. Make sure the work
1289 * item is activated before grabbing.
1290 */
1291 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1292 pwq_activate_delayed_work(work);
1293
1294 list_del_init(&work->entry);
1295 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1296
1297 /* work->data points to pwq iff queued, point to pool */
1298 set_work_pool_and_keep_pending(work, pool->id);
1299
1300 spin_unlock(&pool->lock);
1301 return 1;
1302 }
1303 spin_unlock(&pool->lock);
1304fail:
1305 local_irq_restore(*flags);
1306 if (work_is_canceling(work))
1307 return -ENOENT;
1308 cpu_relax();
1309 return -EAGAIN;
1310}
1311
1312/**
1313 * insert_work - insert a work into a pool
1314 * @pwq: pwq @work belongs to
1315 * @work: work to insert
1316 * @head: insertion point
1317 * @extra_flags: extra WORK_STRUCT_* flags to set
1318 *
1319 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1320 * work_struct flags.
1321 *
1322 * CONTEXT:
1323 * spin_lock_irq(pool->lock).
1324 */
1325static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1326 struct list_head *head, unsigned int extra_flags)
1327{
1328 struct worker_pool *pool = pwq->pool;
1329
1330 /* we own @work, set data and link */
1331 set_work_pwq(work, pwq, extra_flags);
1332 list_add_tail(&work->entry, head);
1333 get_pwq(pwq);
1334
1335 /*
1336 * Ensure either wq_worker_sleeping() sees the above
1337 * list_add_tail() or we see zero nr_running to avoid workers lying
1338 * around lazily while there are works to be processed.
1339 */
1340 smp_mb();
1341
1342 if (__need_more_worker(pool))
1343 wake_up_worker(pool);
1344}
1345
1346/*
1347 * Test whether @work is being queued from another work executing on the
1348 * same workqueue.
1349 */
1350static bool is_chained_work(struct workqueue_struct *wq)
1351{
1352 struct worker *worker;
1353
1354 worker = current_wq_worker();
1355 /*
1356 * Return %true iff I'm a worker executing a work item on @wq. If
1357 * I'm @worker, it's safe to dereference it without locking.
1358 */
1359 return worker && worker->current_pwq->wq == wq;
1360}
1361
1362/*
1363 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1364 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1365 * avoid perturbing sensitive tasks.
1366 */
1367static int wq_select_unbound_cpu(int cpu)
1368{
1369 static bool printed_dbg_warning;
1370 int new_cpu;
1371
1372 if (likely(!wq_debug_force_rr_cpu)) {
1373 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1374 return cpu;
1375 } else if (!printed_dbg_warning) {
1376 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1377 printed_dbg_warning = true;
1378 }
1379
1380 if (cpumask_empty(wq_unbound_cpumask))
1381 return cpu;
1382
1383 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1384 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1385 if (unlikely(new_cpu >= nr_cpu_ids)) {
1386 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1387 if (unlikely(new_cpu >= nr_cpu_ids))
1388 return cpu;
1389 }
1390 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1391
1392 return new_cpu;
1393}
1394
1395static void __queue_work(int cpu, struct workqueue_struct *wq,
1396 struct work_struct *work)
1397{
1398 struct pool_workqueue *pwq;
1399 struct worker_pool *last_pool;
1400 struct list_head *worklist;
1401 unsigned int work_flags;
1402 unsigned int req_cpu = cpu;
1403
1404 /*
1405 * While a work item is PENDING && off queue, a task trying to
1406 * steal the PENDING will busy-loop waiting for it to either get
1407 * queued or lose PENDING. Grabbing PENDING and queueing should
1408 * happen with IRQ disabled.
1409 */
1410 lockdep_assert_irqs_disabled();
1411
1412 debug_work_activate(work);
1413
1414 /* if draining, only works from the same workqueue are allowed */
1415 if (unlikely(wq->flags & __WQ_DRAINING) &&
1416 WARN_ON_ONCE(!is_chained_work(wq)))
1417 return;
1418retry:
1419 if (req_cpu == WORK_CPU_UNBOUND)
1420 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1421
1422 /* pwq which will be used unless @work is executing elsewhere */
1423 if (!(wq->flags & WQ_UNBOUND))
1424 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1425 else
1426 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1427
1428 /*
1429 * If @work was previously on a different pool, it might still be
1430 * running there, in which case the work needs to be queued on that
1431 * pool to guarantee non-reentrancy.
1432 */
1433 last_pool = get_work_pool(work);
1434 if (last_pool && last_pool != pwq->pool) {
1435 struct worker *worker;
1436
1437 spin_lock(&last_pool->lock);
1438
1439 worker = find_worker_executing_work(last_pool, work);
1440
1441 if (worker && worker->current_pwq->wq == wq) {
1442 pwq = worker->current_pwq;
1443 } else {
1444 /* meh... not running there, queue here */
1445 spin_unlock(&last_pool->lock);
1446 spin_lock(&pwq->pool->lock);
1447 }
1448 } else {
1449 spin_lock(&pwq->pool->lock);
1450 }
1451
1452 /*
1453 * pwq is determined and locked. For unbound pools, we could have
1454 * raced with pwq release and it could already be dead. If its
1455 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1456 * without another pwq replacing it in the numa_pwq_tbl or while
1457 * work items are executing on it, so the retrying is guaranteed to
1458 * make forward-progress.
1459 */
1460 if (unlikely(!pwq->refcnt)) {
1461 if (wq->flags & WQ_UNBOUND) {
1462 spin_unlock(&pwq->pool->lock);
1463 cpu_relax();
1464 goto retry;
1465 }
1466 /* oops */
1467 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1468 wq->name, cpu);
1469 }
1470
1471 /* pwq determined, queue */
1472 trace_workqueue_queue_work(req_cpu, pwq, work);
1473
1474 if (WARN_ON(!list_empty(&work->entry))) {
1475 spin_unlock(&pwq->pool->lock);
1476 return;
1477 }
1478
1479 pwq->nr_in_flight[pwq->work_color]++;
1480 work_flags = work_color_to_flags(pwq->work_color);
1481
1482 if (likely(pwq->nr_active < pwq->max_active)) {
1483 trace_workqueue_activate_work(work);
1484 pwq->nr_active++;
1485 worklist = &pwq->pool->worklist;
1486 if (list_empty(worklist))
1487 pwq->pool->watchdog_ts = jiffies;
1488 } else {
1489 work_flags |= WORK_STRUCT_DELAYED;
1490 worklist = &pwq->delayed_works;
1491 }
1492
1493 insert_work(pwq, work, worklist, work_flags);
1494
1495 spin_unlock(&pwq->pool->lock);
1496}
1497
1498/**
1499 * queue_work_on - queue work on specific cpu
1500 * @cpu: CPU number to execute work on
1501 * @wq: workqueue to use
1502 * @work: work to queue
1503 *
1504 * We queue the work to a specific CPU, the caller must ensure it
1505 * can't go away.
1506 *
1507 * Return: %false if @work was already on a queue, %true otherwise.
1508 */
1509bool queue_work_on(int cpu, struct workqueue_struct *wq,
1510 struct work_struct *work)
1511{
1512 bool ret = false;
1513 unsigned long flags;
1514
1515 local_irq_save(flags);
1516
1517 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1518 __queue_work(cpu, wq, work);
1519 ret = true;
1520 }
1521
1522 local_irq_restore(flags);
1523 return ret;
1524}
1525EXPORT_SYMBOL(queue_work_on);
1526
1527/**
1528 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1529 * @node: NUMA node ID that we want to select a CPU from
1530 *
1531 * This function will attempt to find a "random" cpu available on a given
1532 * node. If there are no CPUs available on the given node it will return
1533 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1534 * available CPU if we need to schedule this work.
1535 */
1536static int workqueue_select_cpu_near(int node)
1537{
1538 int cpu;
1539
1540 /* No point in doing this if NUMA isn't enabled for workqueues */
1541 if (!wq_numa_enabled)
1542 return WORK_CPU_UNBOUND;
1543
1544 /* Delay binding to CPU if node is not valid or online */
1545 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1546 return WORK_CPU_UNBOUND;
1547
1548 /* Use local node/cpu if we are already there */
1549 cpu = raw_smp_processor_id();
1550 if (node == cpu_to_node(cpu))
1551 return cpu;
1552
1553 /* Use "random" otherwise know as "first" online CPU of node */
1554 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1555
1556 /* If CPU is valid return that, otherwise just defer */
1557 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1558}
1559
1560/**
1561 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1562 * @node: NUMA node that we are targeting the work for
1563 * @wq: workqueue to use
1564 * @work: work to queue
1565 *
1566 * We queue the work to a "random" CPU within a given NUMA node. The basic
1567 * idea here is to provide a way to somehow associate work with a given
1568 * NUMA node.
1569 *
1570 * This function will only make a best effort attempt at getting this onto
1571 * the right NUMA node. If no node is requested or the requested node is
1572 * offline then we just fall back to standard queue_work behavior.
1573 *
1574 * Currently the "random" CPU ends up being the first available CPU in the
1575 * intersection of cpu_online_mask and the cpumask of the node, unless we
1576 * are running on the node. In that case we just use the current CPU.
1577 *
1578 * Return: %false if @work was already on a queue, %true otherwise.
1579 */
1580bool queue_work_node(int node, struct workqueue_struct *wq,
1581 struct work_struct *work)
1582{
1583 unsigned long flags;
1584 bool ret = false;
1585
1586 /*
1587 * This current implementation is specific to unbound workqueues.
1588 * Specifically we only return the first available CPU for a given
1589 * node instead of cycling through individual CPUs within the node.
1590 *
1591 * If this is used with a per-cpu workqueue then the logic in
1592 * workqueue_select_cpu_near would need to be updated to allow for
1593 * some round robin type logic.
1594 */
1595 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1596
1597 local_irq_save(flags);
1598
1599 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1600 int cpu = workqueue_select_cpu_near(node);
1601
1602 __queue_work(cpu, wq, work);
1603 ret = true;
1604 }
1605
1606 local_irq_restore(flags);
1607 return ret;
1608}
1609EXPORT_SYMBOL_GPL(queue_work_node);
1610
1611void delayed_work_timer_fn(struct timer_list *t)
1612{
1613 struct delayed_work *dwork = from_timer(dwork, t, timer);
1614
1615 /* should have been called from irqsafe timer with irq already off */
1616 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1617}
1618EXPORT_SYMBOL(delayed_work_timer_fn);
1619
1620static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1621 struct delayed_work *dwork, unsigned long delay)
1622{
1623 struct timer_list *timer = &dwork->timer;
1624 struct work_struct *work = &dwork->work;
1625
1626 WARN_ON_ONCE(!wq);
1627 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1628 WARN_ON_ONCE(timer_pending(timer));
1629 WARN_ON_ONCE(!list_empty(&work->entry));
1630
1631 /*
1632 * If @delay is 0, queue @dwork->work immediately. This is for
1633 * both optimization and correctness. The earliest @timer can
1634 * expire is on the closest next tick and delayed_work users depend
1635 * on that there's no such delay when @delay is 0.
1636 */
1637 if (!delay) {
1638 __queue_work(cpu, wq, &dwork->work);
1639 return;
1640 }
1641
1642 dwork->wq = wq;
1643 dwork->cpu = cpu;
1644 timer->expires = jiffies + delay;
1645
1646 if (unlikely(cpu != WORK_CPU_UNBOUND))
1647 add_timer_on(timer, cpu);
1648 else
1649 add_timer(timer);
1650}
1651
1652/**
1653 * queue_delayed_work_on - queue work on specific CPU after delay
1654 * @cpu: CPU number to execute work on
1655 * @wq: workqueue to use
1656 * @dwork: work to queue
1657 * @delay: number of jiffies to wait before queueing
1658 *
1659 * Return: %false if @work was already on a queue, %true otherwise. If
1660 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1661 * execution.
1662 */
1663bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1664 struct delayed_work *dwork, unsigned long delay)
1665{
1666 struct work_struct *work = &dwork->work;
1667 bool ret = false;
1668 unsigned long flags;
1669
1670 /* read the comment in __queue_work() */
1671 local_irq_save(flags);
1672
1673 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1674 __queue_delayed_work(cpu, wq, dwork, delay);
1675 ret = true;
1676 }
1677
1678 local_irq_restore(flags);
1679 return ret;
1680}
1681EXPORT_SYMBOL(queue_delayed_work_on);
1682
1683/**
1684 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1685 * @cpu: CPU number to execute work on
1686 * @wq: workqueue to use
1687 * @dwork: work to queue
1688 * @delay: number of jiffies to wait before queueing
1689 *
1690 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1691 * modify @dwork's timer so that it expires after @delay. If @delay is
1692 * zero, @work is guaranteed to be scheduled immediately regardless of its
1693 * current state.
1694 *
1695 * Return: %false if @dwork was idle and queued, %true if @dwork was
1696 * pending and its timer was modified.
1697 *
1698 * This function is safe to call from any context including IRQ handler.
1699 * See try_to_grab_pending() for details.
1700 */
1701bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1702 struct delayed_work *dwork, unsigned long delay)
1703{
1704 unsigned long flags;
1705 int ret;
1706
1707 do {
1708 ret = try_to_grab_pending(&dwork->work, true, &flags);
1709 } while (unlikely(ret == -EAGAIN));
1710
1711 if (likely(ret >= 0)) {
1712 __queue_delayed_work(cpu, wq, dwork, delay);
1713 local_irq_restore(flags);
1714 }
1715
1716 /* -ENOENT from try_to_grab_pending() becomes %true */
1717 return ret;
1718}
1719EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1720
1721static void rcu_work_rcufn(struct rcu_head *rcu)
1722{
1723 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1724
1725 /* read the comment in __queue_work() */
1726 local_irq_disable();
1727 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1728 local_irq_enable();
1729}
1730
1731/**
1732 * queue_rcu_work - queue work after a RCU grace period
1733 * @wq: workqueue to use
1734 * @rwork: work to queue
1735 *
1736 * Return: %false if @rwork was already pending, %true otherwise. Note
1737 * that a full RCU grace period is guaranteed only after a %true return.
1738 * While @rwork is guaranteed to be executed after a %false return, the
1739 * execution may happen before a full RCU grace period has passed.
1740 */
1741bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1742{
1743 struct work_struct *work = &rwork->work;
1744
1745 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1746 rwork->wq = wq;
1747 call_rcu(&rwork->rcu, rcu_work_rcufn);
1748 return true;
1749 }
1750
1751 return false;
1752}
1753EXPORT_SYMBOL(queue_rcu_work);
1754
1755/**
1756 * worker_enter_idle - enter idle state
1757 * @worker: worker which is entering idle state
1758 *
1759 * @worker is entering idle state. Update stats and idle timer if
1760 * necessary.
1761 *
1762 * LOCKING:
1763 * spin_lock_irq(pool->lock).
1764 */
1765static void worker_enter_idle(struct worker *worker)
1766{
1767 struct worker_pool *pool = worker->pool;
1768
1769 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1770 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1771 (worker->hentry.next || worker->hentry.pprev)))
1772 return;
1773
1774 /* can't use worker_set_flags(), also called from create_worker() */
1775 worker->flags |= WORKER_IDLE;
1776 pool->nr_idle++;
1777 worker->last_active = jiffies;
1778
1779 /* idle_list is LIFO */
1780 list_add(&worker->entry, &pool->idle_list);
1781
1782 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1783 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1784
1785 /*
1786 * Sanity check nr_running. Because unbind_workers() releases
1787 * pool->lock between setting %WORKER_UNBOUND and zapping
1788 * nr_running, the warning may trigger spuriously. Check iff
1789 * unbind is not in progress.
1790 */
1791 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1792 pool->nr_workers == pool->nr_idle &&
1793 atomic_read(&pool->nr_running));
1794}
1795
1796/**
1797 * worker_leave_idle - leave idle state
1798 * @worker: worker which is leaving idle state
1799 *
1800 * @worker is leaving idle state. Update stats.
1801 *
1802 * LOCKING:
1803 * spin_lock_irq(pool->lock).
1804 */
1805static void worker_leave_idle(struct worker *worker)
1806{
1807 struct worker_pool *pool = worker->pool;
1808
1809 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1810 return;
1811 worker_clr_flags(worker, WORKER_IDLE);
1812 pool->nr_idle--;
1813 list_del_init(&worker->entry);
1814}
1815
1816static struct worker *alloc_worker(int node)
1817{
1818 struct worker *worker;
1819
1820 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1821 if (worker) {
1822 INIT_LIST_HEAD(&worker->entry);
1823 INIT_LIST_HEAD(&worker->scheduled);
1824 INIT_LIST_HEAD(&worker->node);
1825 /* on creation a worker is in !idle && prep state */
1826 worker->flags = WORKER_PREP;
1827 }
1828 return worker;
1829}
1830
1831/**
1832 * worker_attach_to_pool() - attach a worker to a pool
1833 * @worker: worker to be attached
1834 * @pool: the target pool
1835 *
1836 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1837 * cpu-binding of @worker are kept coordinated with the pool across
1838 * cpu-[un]hotplugs.
1839 */
1840static void worker_attach_to_pool(struct worker *worker,
1841 struct worker_pool *pool)
1842{
1843 mutex_lock(&wq_pool_attach_mutex);
1844
1845 /*
1846 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1847 * online CPUs. It'll be re-applied when any of the CPUs come up.
1848 */
1849 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1850
1851 /*
1852 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1853 * stable across this function. See the comments above the flag
1854 * definition for details.
1855 */
1856 if (pool->flags & POOL_DISASSOCIATED)
1857 worker->flags |= WORKER_UNBOUND;
1858
1859 list_add_tail(&worker->node, &pool->workers);
1860 worker->pool = pool;
1861
1862 mutex_unlock(&wq_pool_attach_mutex);
1863}
1864
1865/**
1866 * worker_detach_from_pool() - detach a worker from its pool
1867 * @worker: worker which is attached to its pool
1868 *
1869 * Undo the attaching which had been done in worker_attach_to_pool(). The
1870 * caller worker shouldn't access to the pool after detached except it has
1871 * other reference to the pool.
1872 */
1873static void worker_detach_from_pool(struct worker *worker)
1874{
1875 struct worker_pool *pool = worker->pool;
1876 struct completion *detach_completion = NULL;
1877
1878 mutex_lock(&wq_pool_attach_mutex);
1879
1880 list_del(&worker->node);
1881 worker->pool = NULL;
1882
1883 if (list_empty(&pool->workers))
1884 detach_completion = pool->detach_completion;
1885 mutex_unlock(&wq_pool_attach_mutex);
1886
1887 /* clear leftover flags without pool->lock after it is detached */
1888 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1889
1890 if (detach_completion)
1891 complete(detach_completion);
1892}
1893
1894/**
1895 * create_worker - create a new workqueue worker
1896 * @pool: pool the new worker will belong to
1897 *
1898 * Create and start a new worker which is attached to @pool.
1899 *
1900 * CONTEXT:
1901 * Might sleep. Does GFP_KERNEL allocations.
1902 *
1903 * Return:
1904 * Pointer to the newly created worker.
1905 */
1906static struct worker *create_worker(struct worker_pool *pool)
1907{
1908 struct worker *worker = NULL;
1909 int id = -1;
1910 char id_buf[16];
1911
1912 /* ID is needed to determine kthread name */
1913 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1914 if (id < 0)
1915 goto fail;
1916
1917 worker = alloc_worker(pool->node);
1918 if (!worker)
1919 goto fail;
1920
1921 worker->id = id;
1922
1923 if (pool->cpu >= 0)
1924 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1925 pool->attrs->nice < 0 ? "H" : "");
1926 else
1927 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1928
1929 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1930 "kworker/%s", id_buf);
1931 if (IS_ERR(worker->task))
1932 goto fail;
1933
1934 set_user_nice(worker->task, pool->attrs->nice);
1935 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1936
1937 /* successful, attach the worker to the pool */
1938 worker_attach_to_pool(worker, pool);
1939
1940 /* start the newly created worker */
1941 spin_lock_irq(&pool->lock);
1942 worker->pool->nr_workers++;
1943 worker_enter_idle(worker);
1944 wake_up_process(worker->task);
1945 spin_unlock_irq(&pool->lock);
1946
1947 return worker;
1948
1949fail:
1950 if (id >= 0)
1951 ida_simple_remove(&pool->worker_ida, id);
1952 kfree(worker);
1953 return NULL;
1954}
1955
1956/**
1957 * destroy_worker - destroy a workqueue worker
1958 * @worker: worker to be destroyed
1959 *
1960 * Destroy @worker and adjust @pool stats accordingly. The worker should
1961 * be idle.
1962 *
1963 * CONTEXT:
1964 * spin_lock_irq(pool->lock).
1965 */
1966static void destroy_worker(struct worker *worker)
1967{
1968 struct worker_pool *pool = worker->pool;
1969
1970 lockdep_assert_held(&pool->lock);
1971
1972 /* sanity check frenzy */
1973 if (WARN_ON(worker->current_work) ||
1974 WARN_ON(!list_empty(&worker->scheduled)) ||
1975 WARN_ON(!(worker->flags & WORKER_IDLE)))
1976 return;
1977
1978 pool->nr_workers--;
1979 pool->nr_idle--;
1980
1981 list_del_init(&worker->entry);
1982 worker->flags |= WORKER_DIE;
1983 wake_up_process(worker->task);
1984}
1985
1986static void idle_worker_timeout(struct timer_list *t)
1987{
1988 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1989
1990 spin_lock_irq(&pool->lock);
1991
1992 while (too_many_workers(pool)) {
1993 struct worker *worker;
1994 unsigned long expires;
1995
1996 /* idle_list is kept in LIFO order, check the last one */
1997 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1998 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1999
2000 if (time_before(jiffies, expires)) {
2001 mod_timer(&pool->idle_timer, expires);
2002 break;
2003 }
2004
2005 destroy_worker(worker);
2006 }
2007
2008 spin_unlock_irq(&pool->lock);
2009}
2010
2011static void send_mayday(struct work_struct *work)
2012{
2013 struct pool_workqueue *pwq = get_work_pwq(work);
2014 struct workqueue_struct *wq = pwq->wq;
2015
2016 lockdep_assert_held(&wq_mayday_lock);
2017
2018 if (!wq->rescuer)
2019 return;
2020
2021 /* mayday mayday mayday */
2022 if (list_empty(&pwq->mayday_node)) {
2023 /*
2024 * If @pwq is for an unbound wq, its base ref may be put at
2025 * any time due to an attribute change. Pin @pwq until the
2026 * rescuer is done with it.
2027 */
2028 get_pwq(pwq);
2029 list_add_tail(&pwq->mayday_node, &wq->maydays);
2030 wake_up_process(wq->rescuer->task);
2031 }
2032}
2033
2034static void pool_mayday_timeout(struct timer_list *t)
2035{
2036 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2037 struct work_struct *work;
2038
2039 spin_lock_irq(&pool->lock);
2040 spin_lock(&wq_mayday_lock); /* for wq->maydays */
2041
2042 if (need_to_create_worker(pool)) {
2043 /*
2044 * We've been trying to create a new worker but
2045 * haven't been successful. We might be hitting an
2046 * allocation deadlock. Send distress signals to
2047 * rescuers.
2048 */
2049 list_for_each_entry(work, &pool->worklist, entry)
2050 send_mayday(work);
2051 }
2052
2053 spin_unlock(&wq_mayday_lock);
2054 spin_unlock_irq(&pool->lock);
2055
2056 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2057}
2058
2059/**
2060 * maybe_create_worker - create a new worker if necessary
2061 * @pool: pool to create a new worker for
2062 *
2063 * Create a new worker for @pool if necessary. @pool is guaranteed to
2064 * have at least one idle worker on return from this function. If
2065 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2066 * sent to all rescuers with works scheduled on @pool to resolve
2067 * possible allocation deadlock.
2068 *
2069 * On return, need_to_create_worker() is guaranteed to be %false and
2070 * may_start_working() %true.
2071 *
2072 * LOCKING:
2073 * spin_lock_irq(pool->lock) which may be released and regrabbed
2074 * multiple times. Does GFP_KERNEL allocations. Called only from
2075 * manager.
2076 */
2077static void maybe_create_worker(struct worker_pool *pool)
2078__releases(&pool->lock)
2079__acquires(&pool->lock)
2080{
2081restart:
2082 spin_unlock_irq(&pool->lock);
2083
2084 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2085 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2086
2087 while (true) {
2088 if (create_worker(pool) || !need_to_create_worker(pool))
2089 break;
2090
2091 schedule_timeout_interruptible(CREATE_COOLDOWN);
2092
2093 if (!need_to_create_worker(pool))
2094 break;
2095 }
2096
2097 del_timer_sync(&pool->mayday_timer);
2098 spin_lock_irq(&pool->lock);
2099 /*
2100 * This is necessary even after a new worker was just successfully
2101 * created as @pool->lock was dropped and the new worker might have
2102 * already become busy.
2103 */
2104 if (need_to_create_worker(pool))
2105 goto restart;
2106}
2107
2108/**
2109 * manage_workers - manage worker pool
2110 * @worker: self
2111 *
2112 * Assume the manager role and manage the worker pool @worker belongs
2113 * to. At any given time, there can be only zero or one manager per
2114 * pool. The exclusion is handled automatically by this function.
2115 *
2116 * The caller can safely start processing works on false return. On
2117 * true return, it's guaranteed that need_to_create_worker() is false
2118 * and may_start_working() is true.
2119 *
2120 * CONTEXT:
2121 * spin_lock_irq(pool->lock) which may be released and regrabbed
2122 * multiple times. Does GFP_KERNEL allocations.
2123 *
2124 * Return:
2125 * %false if the pool doesn't need management and the caller can safely
2126 * start processing works, %true if management function was performed and
2127 * the conditions that the caller verified before calling the function may
2128 * no longer be true.
2129 */
2130static bool manage_workers(struct worker *worker)
2131{
2132 struct worker_pool *pool = worker->pool;
2133
2134 if (pool->flags & POOL_MANAGER_ACTIVE)
2135 return false;
2136
2137 pool->flags |= POOL_MANAGER_ACTIVE;
2138 pool->manager = worker;
2139
2140 maybe_create_worker(pool);
2141
2142 pool->manager = NULL;
2143 pool->flags &= ~POOL_MANAGER_ACTIVE;
2144 wake_up(&wq_manager_wait);
2145 return true;
2146}
2147
2148/**
2149 * process_one_work - process single work
2150 * @worker: self
2151 * @work: work to process
2152 *
2153 * Process @work. This function contains all the logics necessary to
2154 * process a single work including synchronization against and
2155 * interaction with other workers on the same cpu, queueing and
2156 * flushing. As long as context requirement is met, any worker can
2157 * call this function to process a work.
2158 *
2159 * CONTEXT:
2160 * spin_lock_irq(pool->lock) which is released and regrabbed.
2161 */
2162static void process_one_work(struct worker *worker, struct work_struct *work)
2163__releases(&pool->lock)
2164__acquires(&pool->lock)
2165{
2166 struct pool_workqueue *pwq = get_work_pwq(work);
2167 struct worker_pool *pool = worker->pool;
2168 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2169 int work_color;
2170 struct worker *collision;
2171#ifdef CONFIG_LOCKDEP
2172 /*
2173 * It is permissible to free the struct work_struct from
2174 * inside the function that is called from it, this we need to
2175 * take into account for lockdep too. To avoid bogus "held
2176 * lock freed" warnings as well as problems when looking into
2177 * work->lockdep_map, make a copy and use that here.
2178 */
2179 struct lockdep_map lockdep_map;
2180
2181 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2182#endif
2183 /* ensure we're on the correct CPU */
2184 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2185 raw_smp_processor_id() != pool->cpu);
2186
2187 /*
2188 * A single work shouldn't be executed concurrently by
2189 * multiple workers on a single cpu. Check whether anyone is
2190 * already processing the work. If so, defer the work to the
2191 * currently executing one.
2192 */
2193 collision = find_worker_executing_work(pool, work);
2194 if (unlikely(collision)) {
2195 move_linked_works(work, &collision->scheduled, NULL);
2196 return;
2197 }
2198
2199 /* claim and dequeue */
2200 debug_work_deactivate(work);
2201 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2202 worker->current_work = work;
2203 worker->current_func = work->func;
2204 worker->current_pwq = pwq;
2205 work_color = get_work_color(work);
2206
2207 /*
2208 * Record wq name for cmdline and debug reporting, may get
2209 * overridden through set_worker_desc().
2210 */
2211 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2212
2213 list_del_init(&work->entry);
2214
2215 /*
2216 * CPU intensive works don't participate in concurrency management.
2217 * They're the scheduler's responsibility. This takes @worker out
2218 * of concurrency management and the next code block will chain
2219 * execution of the pending work items.
2220 */
2221 if (unlikely(cpu_intensive))
2222 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2223
2224 /*
2225 * Wake up another worker if necessary. The condition is always
2226 * false for normal per-cpu workers since nr_running would always
2227 * be >= 1 at this point. This is used to chain execution of the
2228 * pending work items for WORKER_NOT_RUNNING workers such as the
2229 * UNBOUND and CPU_INTENSIVE ones.
2230 */
2231 if (need_more_worker(pool))
2232 wake_up_worker(pool);
2233
2234 /*
2235 * Record the last pool and clear PENDING which should be the last
2236 * update to @work. Also, do this inside @pool->lock so that
2237 * PENDING and queued state changes happen together while IRQ is
2238 * disabled.
2239 */
2240 set_work_pool_and_clear_pending(work, pool->id);
2241
2242 spin_unlock_irq(&pool->lock);
2243
2244 lock_map_acquire(&pwq->wq->lockdep_map);
2245 lock_map_acquire(&lockdep_map);
2246 /*
2247 * Strictly speaking we should mark the invariant state without holding
2248 * any locks, that is, before these two lock_map_acquire()'s.
2249 *
2250 * However, that would result in:
2251 *
2252 * A(W1)
2253 * WFC(C)
2254 * A(W1)
2255 * C(C)
2256 *
2257 * Which would create W1->C->W1 dependencies, even though there is no
2258 * actual deadlock possible. There are two solutions, using a
2259 * read-recursive acquire on the work(queue) 'locks', but this will then
2260 * hit the lockdep limitation on recursive locks, or simply discard
2261 * these locks.
2262 *
2263 * AFAICT there is no possible deadlock scenario between the
2264 * flush_work() and complete() primitives (except for single-threaded
2265 * workqueues), so hiding them isn't a problem.
2266 */
2267 lockdep_invariant_state(true);
2268 trace_workqueue_execute_start(work);
2269 worker->current_func(work);
2270 /*
2271 * While we must be careful to not use "work" after this, the trace
2272 * point will only record its address.
2273 */
2274 trace_workqueue_execute_end(work);
2275 lock_map_release(&lockdep_map);
2276 lock_map_release(&pwq->wq->lockdep_map);
2277
2278 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2279 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2280 " last function: %pf\n",
2281 current->comm, preempt_count(), task_pid_nr(current),
2282 worker->current_func);
2283 debug_show_held_locks(current);
2284 dump_stack();
2285 }
2286
2287 /*
2288 * The following prevents a kworker from hogging CPU on !PREEMPT
2289 * kernels, where a requeueing work item waiting for something to
2290 * happen could deadlock with stop_machine as such work item could
2291 * indefinitely requeue itself while all other CPUs are trapped in
2292 * stop_machine. At the same time, report a quiescent RCU state so
2293 * the same condition doesn't freeze RCU.
2294 */
2295 cond_resched();
2296
2297 spin_lock_irq(&pool->lock);
2298
2299 /* clear cpu intensive status */
2300 if (unlikely(cpu_intensive))
2301 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2302
2303 /* tag the worker for identification in schedule() */
2304 worker->last_func = worker->current_func;
2305
2306 /* we're done with it, release */
2307 hash_del(&worker->hentry);
2308 worker->current_work = NULL;
2309 worker->current_func = NULL;
2310 worker->current_pwq = NULL;
2311 pwq_dec_nr_in_flight(pwq, work_color);
2312}
2313
2314/**
2315 * process_scheduled_works - process scheduled works
2316 * @worker: self
2317 *
2318 * Process all scheduled works. Please note that the scheduled list
2319 * may change while processing a work, so this function repeatedly
2320 * fetches a work from the top and executes it.
2321 *
2322 * CONTEXT:
2323 * spin_lock_irq(pool->lock) which may be released and regrabbed
2324 * multiple times.
2325 */
2326static void process_scheduled_works(struct worker *worker)
2327{
2328 while (!list_empty(&worker->scheduled)) {
2329 struct work_struct *work = list_first_entry(&worker->scheduled,
2330 struct work_struct, entry);
2331 process_one_work(worker, work);
2332 }
2333}
2334
2335static void set_pf_worker(bool val)
2336{
2337 mutex_lock(&wq_pool_attach_mutex);
2338 if (val)
2339 current->flags |= PF_WQ_WORKER;
2340 else
2341 current->flags &= ~PF_WQ_WORKER;
2342 mutex_unlock(&wq_pool_attach_mutex);
2343}
2344
2345/**
2346 * worker_thread - the worker thread function
2347 * @__worker: self
2348 *
2349 * The worker thread function. All workers belong to a worker_pool -
2350 * either a per-cpu one or dynamic unbound one. These workers process all
2351 * work items regardless of their specific target workqueue. The only
2352 * exception is work items which belong to workqueues with a rescuer which
2353 * will be explained in rescuer_thread().
2354 *
2355 * Return: 0
2356 */
2357static int worker_thread(void *__worker)
2358{
2359 struct worker *worker = __worker;
2360 struct worker_pool *pool = worker->pool;
2361
2362 /* tell the scheduler that this is a workqueue worker */
2363 set_pf_worker(true);
2364woke_up:
2365 spin_lock_irq(&pool->lock);
2366
2367 /* am I supposed to die? */
2368 if (unlikely(worker->flags & WORKER_DIE)) {
2369 spin_unlock_irq(&pool->lock);
2370 WARN_ON_ONCE(!list_empty(&worker->entry));
2371 set_pf_worker(false);
2372
2373 set_task_comm(worker->task, "kworker/dying");
2374 ida_simple_remove(&pool->worker_ida, worker->id);
2375 worker_detach_from_pool(worker);
2376 kfree(worker);
2377 return 0;
2378 }
2379
2380 worker_leave_idle(worker);
2381recheck:
2382 /* no more worker necessary? */
2383 if (!need_more_worker(pool))
2384 goto sleep;
2385
2386 /* do we need to manage? */
2387 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2388 goto recheck;
2389
2390 /*
2391 * ->scheduled list can only be filled while a worker is
2392 * preparing to process a work or actually processing it.
2393 * Make sure nobody diddled with it while I was sleeping.
2394 */
2395 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2396
2397 /*
2398 * Finish PREP stage. We're guaranteed to have at least one idle
2399 * worker or that someone else has already assumed the manager
2400 * role. This is where @worker starts participating in concurrency
2401 * management if applicable and concurrency management is restored
2402 * after being rebound. See rebind_workers() for details.
2403 */
2404 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2405
2406 do {
2407 struct work_struct *work =
2408 list_first_entry(&pool->worklist,
2409 struct work_struct, entry);
2410
2411 pool->watchdog_ts = jiffies;
2412
2413 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2414 /* optimization path, not strictly necessary */
2415 process_one_work(worker, work);
2416 if (unlikely(!list_empty(&worker->scheduled)))
2417 process_scheduled_works(worker);
2418 } else {
2419 move_linked_works(work, &worker->scheduled, NULL);
2420 process_scheduled_works(worker);
2421 }
2422 } while (keep_working(pool));
2423
2424 worker_set_flags(worker, WORKER_PREP);
2425sleep:
2426 /*
2427 * pool->lock is held and there's no work to process and no need to
2428 * manage, sleep. Workers are woken up only while holding
2429 * pool->lock or from local cpu, so setting the current state
2430 * before releasing pool->lock is enough to prevent losing any
2431 * event.
2432 */
2433 worker_enter_idle(worker);
2434 __set_current_state(TASK_IDLE);
2435 spin_unlock_irq(&pool->lock);
2436 schedule();
2437 goto woke_up;
2438}
2439
2440/**
2441 * rescuer_thread - the rescuer thread function
2442 * @__rescuer: self
2443 *
2444 * Workqueue rescuer thread function. There's one rescuer for each
2445 * workqueue which has WQ_MEM_RECLAIM set.
2446 *
2447 * Regular work processing on a pool may block trying to create a new
2448 * worker which uses GFP_KERNEL allocation which has slight chance of
2449 * developing into deadlock if some works currently on the same queue
2450 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2451 * the problem rescuer solves.
2452 *
2453 * When such condition is possible, the pool summons rescuers of all
2454 * workqueues which have works queued on the pool and let them process
2455 * those works so that forward progress can be guaranteed.
2456 *
2457 * This should happen rarely.
2458 *
2459 * Return: 0
2460 */
2461static int rescuer_thread(void *__rescuer)
2462{
2463 struct worker *rescuer = __rescuer;
2464 struct workqueue_struct *wq = rescuer->rescue_wq;
2465 struct list_head *scheduled = &rescuer->scheduled;
2466 bool should_stop;
2467
2468 set_user_nice(current, RESCUER_NICE_LEVEL);
2469
2470 /*
2471 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2472 * doesn't participate in concurrency management.
2473 */
2474 set_pf_worker(true);
2475repeat:
2476 set_current_state(TASK_IDLE);
2477
2478 /*
2479 * By the time the rescuer is requested to stop, the workqueue
2480 * shouldn't have any work pending, but @wq->maydays may still have
2481 * pwq(s) queued. This can happen by non-rescuer workers consuming
2482 * all the work items before the rescuer got to them. Go through
2483 * @wq->maydays processing before acting on should_stop so that the
2484 * list is always empty on exit.
2485 */
2486 should_stop = kthread_should_stop();
2487
2488 /* see whether any pwq is asking for help */
2489 spin_lock_irq(&wq_mayday_lock);
2490
2491 while (!list_empty(&wq->maydays)) {
2492 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2493 struct pool_workqueue, mayday_node);
2494 struct worker_pool *pool = pwq->pool;
2495 struct work_struct *work, *n;
2496 bool first = true;
2497
2498 __set_current_state(TASK_RUNNING);
2499 list_del_init(&pwq->mayday_node);
2500
2501 spin_unlock_irq(&wq_mayday_lock);
2502
2503 worker_attach_to_pool(rescuer, pool);
2504
2505 spin_lock_irq(&pool->lock);
2506
2507 /*
2508 * Slurp in all works issued via this workqueue and
2509 * process'em.
2510 */
2511 WARN_ON_ONCE(!list_empty(scheduled));
2512 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2513 if (get_work_pwq(work) == pwq) {
2514 if (first)
2515 pool->watchdog_ts = jiffies;
2516 move_linked_works(work, scheduled, &n);
2517 }
2518 first = false;
2519 }
2520
2521 if (!list_empty(scheduled)) {
2522 process_scheduled_works(rescuer);
2523
2524 /*
2525 * The above execution of rescued work items could
2526 * have created more to rescue through
2527 * pwq_activate_first_delayed() or chained
2528 * queueing. Let's put @pwq back on mayday list so
2529 * that such back-to-back work items, which may be
2530 * being used to relieve memory pressure, don't
2531 * incur MAYDAY_INTERVAL delay inbetween.
2532 */
2533 if (need_to_create_worker(pool)) {
2534 spin_lock(&wq_mayday_lock);
2535 get_pwq(pwq);
2536 list_move_tail(&pwq->mayday_node, &wq->maydays);
2537 spin_unlock(&wq_mayday_lock);
2538 }
2539 }
2540
2541 /*
2542 * Put the reference grabbed by send_mayday(). @pool won't
2543 * go away while we're still attached to it.
2544 */
2545 put_pwq(pwq);
2546
2547 /*
2548 * Leave this pool. If need_more_worker() is %true, notify a
2549 * regular worker; otherwise, we end up with 0 concurrency
2550 * and stalling the execution.
2551 */
2552 if (need_more_worker(pool))
2553 wake_up_worker(pool);
2554
2555 spin_unlock_irq(&pool->lock);
2556
2557 worker_detach_from_pool(rescuer);
2558
2559 spin_lock_irq(&wq_mayday_lock);
2560 }
2561
2562 spin_unlock_irq(&wq_mayday_lock);
2563
2564 if (should_stop) {
2565 __set_current_state(TASK_RUNNING);
2566 set_pf_worker(false);
2567 return 0;
2568 }
2569
2570 /* rescuers should never participate in concurrency management */
2571 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2572 schedule();
2573 goto repeat;
2574}
2575
2576/**
2577 * check_flush_dependency - check for flush dependency sanity
2578 * @target_wq: workqueue being flushed
2579 * @target_work: work item being flushed (NULL for workqueue flushes)
2580 *
2581 * %current is trying to flush the whole @target_wq or @target_work on it.
2582 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2583 * reclaiming memory or running on a workqueue which doesn't have
2584 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2585 * a deadlock.
2586 */
2587static void check_flush_dependency(struct workqueue_struct *target_wq,
2588 struct work_struct *target_work)
2589{
2590 work_func_t target_func = target_work ? target_work->func : NULL;
2591 struct worker *worker;
2592
2593 if (target_wq->flags & WQ_MEM_RECLAIM)
2594 return;
2595
2596 worker = current_wq_worker();
2597
2598 WARN_ONCE(current->flags & PF_MEMALLOC,
2599 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2600 current->pid, current->comm, target_wq->name, target_func);
2601 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2602 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2603 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2604 worker->current_pwq->wq->name, worker->current_func,
2605 target_wq->name, target_func);
2606}
2607
2608struct wq_barrier {
2609 struct work_struct work;
2610 struct completion done;
2611 struct task_struct *task; /* purely informational */
2612};
2613
2614static void wq_barrier_func(struct work_struct *work)
2615{
2616 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2617 complete(&barr->done);
2618}
2619
2620/**
2621 * insert_wq_barrier - insert a barrier work
2622 * @pwq: pwq to insert barrier into
2623 * @barr: wq_barrier to insert
2624 * @target: target work to attach @barr to
2625 * @worker: worker currently executing @target, NULL if @target is not executing
2626 *
2627 * @barr is linked to @target such that @barr is completed only after
2628 * @target finishes execution. Please note that the ordering
2629 * guarantee is observed only with respect to @target and on the local
2630 * cpu.
2631 *
2632 * Currently, a queued barrier can't be canceled. This is because
2633 * try_to_grab_pending() can't determine whether the work to be
2634 * grabbed is at the head of the queue and thus can't clear LINKED
2635 * flag of the previous work while there must be a valid next work
2636 * after a work with LINKED flag set.
2637 *
2638 * Note that when @worker is non-NULL, @target may be modified
2639 * underneath us, so we can't reliably determine pwq from @target.
2640 *
2641 * CONTEXT:
2642 * spin_lock_irq(pool->lock).
2643 */
2644static void insert_wq_barrier(struct pool_workqueue *pwq,
2645 struct wq_barrier *barr,
2646 struct work_struct *target, struct worker *worker)
2647{
2648 struct list_head *head;
2649 unsigned int linked = 0;
2650
2651 /*
2652 * debugobject calls are safe here even with pool->lock locked
2653 * as we know for sure that this will not trigger any of the
2654 * checks and call back into the fixup functions where we
2655 * might deadlock.
2656 */
2657 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2658 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2659
2660 init_completion_map(&barr->done, &target->lockdep_map);
2661
2662 barr->task = current;
2663
2664 /*
2665 * If @target is currently being executed, schedule the
2666 * barrier to the worker; otherwise, put it after @target.
2667 */
2668 if (worker)
2669 head = worker->scheduled.next;
2670 else {
2671 unsigned long *bits = work_data_bits(target);
2672
2673 head = target->entry.next;
2674 /* there can already be other linked works, inherit and set */
2675 linked = *bits & WORK_STRUCT_LINKED;
2676 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2677 }
2678
2679 debug_work_activate(&barr->work);
2680 insert_work(pwq, &barr->work, head,
2681 work_color_to_flags(WORK_NO_COLOR) | linked);
2682}
2683
2684/**
2685 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2686 * @wq: workqueue being flushed
2687 * @flush_color: new flush color, < 0 for no-op
2688 * @work_color: new work color, < 0 for no-op
2689 *
2690 * Prepare pwqs for workqueue flushing.
2691 *
2692 * If @flush_color is non-negative, flush_color on all pwqs should be
2693 * -1. If no pwq has in-flight commands at the specified color, all
2694 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2695 * has in flight commands, its pwq->flush_color is set to
2696 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2697 * wakeup logic is armed and %true is returned.
2698 *
2699 * The caller should have initialized @wq->first_flusher prior to
2700 * calling this function with non-negative @flush_color. If
2701 * @flush_color is negative, no flush color update is done and %false
2702 * is returned.
2703 *
2704 * If @work_color is non-negative, all pwqs should have the same
2705 * work_color which is previous to @work_color and all will be
2706 * advanced to @work_color.
2707 *
2708 * CONTEXT:
2709 * mutex_lock(wq->mutex).
2710 *
2711 * Return:
2712 * %true if @flush_color >= 0 and there's something to flush. %false
2713 * otherwise.
2714 */
2715static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2716 int flush_color, int work_color)
2717{
2718 bool wait = false;
2719 struct pool_workqueue *pwq;
2720
2721 if (flush_color >= 0) {
2722 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2723 atomic_set(&wq->nr_pwqs_to_flush, 1);
2724 }
2725
2726 for_each_pwq(pwq, wq) {
2727 struct worker_pool *pool = pwq->pool;
2728
2729 spin_lock_irq(&pool->lock);
2730
2731 if (flush_color >= 0) {
2732 WARN_ON_ONCE(pwq->flush_color != -1);
2733
2734 if (pwq->nr_in_flight[flush_color]) {
2735 pwq->flush_color = flush_color;
2736 atomic_inc(&wq->nr_pwqs_to_flush);
2737 wait = true;
2738 }
2739 }
2740
2741 if (work_color >= 0) {
2742 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2743 pwq->work_color = work_color;
2744 }
2745
2746 spin_unlock_irq(&pool->lock);
2747 }
2748
2749 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2750 complete(&wq->first_flusher->done);
2751
2752 return wait;
2753}
2754
2755/**
2756 * flush_workqueue - ensure that any scheduled work has run to completion.
2757 * @wq: workqueue to flush
2758 *
2759 * This function sleeps until all work items which were queued on entry
2760 * have finished execution, but it is not livelocked by new incoming ones.
2761 */
2762void flush_workqueue(struct workqueue_struct *wq)
2763{
2764 struct wq_flusher this_flusher = {
2765 .list = LIST_HEAD_INIT(this_flusher.list),
2766 .flush_color = -1,
2767 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2768 };
2769 int next_color;
2770
2771 if (WARN_ON(!wq_online))
2772 return;
2773
2774 lock_map_acquire(&wq->lockdep_map);
2775 lock_map_release(&wq->lockdep_map);
2776
2777 mutex_lock(&wq->mutex);
2778
2779 /*
2780 * Start-to-wait phase
2781 */
2782 next_color = work_next_color(wq->work_color);
2783
2784 if (next_color != wq->flush_color) {
2785 /*
2786 * Color space is not full. The current work_color
2787 * becomes our flush_color and work_color is advanced
2788 * by one.
2789 */
2790 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2791 this_flusher.flush_color = wq->work_color;
2792 wq->work_color = next_color;
2793
2794 if (!wq->first_flusher) {
2795 /* no flush in progress, become the first flusher */
2796 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2797
2798 wq->first_flusher = &this_flusher;
2799
2800 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2801 wq->work_color)) {
2802 /* nothing to flush, done */
2803 wq->flush_color = next_color;
2804 wq->first_flusher = NULL;
2805 goto out_unlock;
2806 }
2807 } else {
2808 /* wait in queue */
2809 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2810 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2811 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2812 }
2813 } else {
2814 /*
2815 * Oops, color space is full, wait on overflow queue.
2816 * The next flush completion will assign us
2817 * flush_color and transfer to flusher_queue.
2818 */
2819 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2820 }
2821
2822 check_flush_dependency(wq, NULL);
2823
2824 mutex_unlock(&wq->mutex);
2825
2826 wait_for_completion(&this_flusher.done);
2827
2828 /*
2829 * Wake-up-and-cascade phase
2830 *
2831 * First flushers are responsible for cascading flushes and
2832 * handling overflow. Non-first flushers can simply return.
2833 */
2834 if (wq->first_flusher != &this_flusher)
2835 return;
2836
2837 mutex_lock(&wq->mutex);
2838
2839 /* we might have raced, check again with mutex held */
2840 if (wq->first_flusher != &this_flusher)
2841 goto out_unlock;
2842
2843 wq->first_flusher = NULL;
2844
2845 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2846 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2847
2848 while (true) {
2849 struct wq_flusher *next, *tmp;
2850
2851 /* complete all the flushers sharing the current flush color */
2852 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2853 if (next->flush_color != wq->flush_color)
2854 break;
2855 list_del_init(&next->list);
2856 complete(&next->done);
2857 }
2858
2859 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2860 wq->flush_color != work_next_color(wq->work_color));
2861
2862 /* this flush_color is finished, advance by one */
2863 wq->flush_color = work_next_color(wq->flush_color);
2864
2865 /* one color has been freed, handle overflow queue */
2866 if (!list_empty(&wq->flusher_overflow)) {
2867 /*
2868 * Assign the same color to all overflowed
2869 * flushers, advance work_color and append to
2870 * flusher_queue. This is the start-to-wait
2871 * phase for these overflowed flushers.
2872 */
2873 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2874 tmp->flush_color = wq->work_color;
2875
2876 wq->work_color = work_next_color(wq->work_color);
2877
2878 list_splice_tail_init(&wq->flusher_overflow,
2879 &wq->flusher_queue);
2880 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2881 }
2882
2883 if (list_empty(&wq->flusher_queue)) {
2884 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2885 break;
2886 }
2887
2888 /*
2889 * Need to flush more colors. Make the next flusher
2890 * the new first flusher and arm pwqs.
2891 */
2892 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2893 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2894
2895 list_del_init(&next->list);
2896 wq->first_flusher = next;
2897
2898 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2899 break;
2900
2901 /*
2902 * Meh... this color is already done, clear first
2903 * flusher and repeat cascading.
2904 */
2905 wq->first_flusher = NULL;
2906 }
2907
2908out_unlock:
2909 mutex_unlock(&wq->mutex);
2910}
2911EXPORT_SYMBOL(flush_workqueue);
2912
2913/**
2914 * drain_workqueue - drain a workqueue
2915 * @wq: workqueue to drain
2916 *
2917 * Wait until the workqueue becomes empty. While draining is in progress,
2918 * only chain queueing is allowed. IOW, only currently pending or running
2919 * work items on @wq can queue further work items on it. @wq is flushed
2920 * repeatedly until it becomes empty. The number of flushing is determined
2921 * by the depth of chaining and should be relatively short. Whine if it
2922 * takes too long.
2923 */
2924void drain_workqueue(struct workqueue_struct *wq)
2925{
2926 unsigned int flush_cnt = 0;
2927 struct pool_workqueue *pwq;
2928
2929 /*
2930 * __queue_work() needs to test whether there are drainers, is much
2931 * hotter than drain_workqueue() and already looks at @wq->flags.
2932 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2933 */
2934 mutex_lock(&wq->mutex);
2935 if (!wq->nr_drainers++)
2936 wq->flags |= __WQ_DRAINING;
2937 mutex_unlock(&wq->mutex);
2938reflush:
2939 flush_workqueue(wq);
2940
2941 mutex_lock(&wq->mutex);
2942
2943 for_each_pwq(pwq, wq) {
2944 bool drained;
2945
2946 spin_lock_irq(&pwq->pool->lock);
2947 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2948 spin_unlock_irq(&pwq->pool->lock);
2949
2950 if (drained)
2951 continue;
2952
2953 if (++flush_cnt == 10 ||
2954 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2955 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2956 wq->name, flush_cnt);
2957
2958 mutex_unlock(&wq->mutex);
2959 goto reflush;
2960 }
2961
2962 if (!--wq->nr_drainers)
2963 wq->flags &= ~__WQ_DRAINING;
2964 mutex_unlock(&wq->mutex);
2965}
2966EXPORT_SYMBOL_GPL(drain_workqueue);
2967
2968static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2969 bool from_cancel)
2970{
2971 struct worker *worker = NULL;
2972 struct worker_pool *pool;
2973 struct pool_workqueue *pwq;
2974
2975 might_sleep();
2976
2977 local_irq_disable();
2978 pool = get_work_pool(work);
2979 if (!pool) {
2980 local_irq_enable();
2981 return false;
2982 }
2983
2984 spin_lock(&pool->lock);
2985 /* see the comment in try_to_grab_pending() with the same code */
2986 pwq = get_work_pwq(work);
2987 if (pwq) {
2988 if (unlikely(pwq->pool != pool))
2989 goto already_gone;
2990 } else {
2991 worker = find_worker_executing_work(pool, work);
2992 if (!worker)
2993 goto already_gone;
2994 pwq = worker->current_pwq;
2995 }
2996
2997 check_flush_dependency(pwq->wq, work);
2998
2999 insert_wq_barrier(pwq, barr, work, worker);
3000 spin_unlock_irq(&pool->lock);
3001
3002 /*
3003 * Force a lock recursion deadlock when using flush_work() inside a
3004 * single-threaded or rescuer equipped workqueue.
3005 *
3006 * For single threaded workqueues the deadlock happens when the work
3007 * is after the work issuing the flush_work(). For rescuer equipped
3008 * workqueues the deadlock happens when the rescuer stalls, blocking
3009 * forward progress.
3010 */
3011 if (!from_cancel &&
3012 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3013 lock_map_acquire(&pwq->wq->lockdep_map);
3014 lock_map_release(&pwq->wq->lockdep_map);
3015 }
3016
3017 return true;
3018already_gone:
3019 spin_unlock_irq(&pool->lock);
3020 return false;
3021}
3022
3023static bool __flush_work(struct work_struct *work, bool from_cancel)
3024{
3025 struct wq_barrier barr;
3026
3027 if (WARN_ON(!wq_online))
3028 return false;
3029
3030 if (WARN_ON(!work->func))
3031 return false;
3032
3033 if (!from_cancel) {
3034 lock_map_acquire(&work->lockdep_map);
3035 lock_map_release(&work->lockdep_map);
3036 }
3037
3038 if (start_flush_work(work, &barr, from_cancel)) {
3039 wait_for_completion(&barr.done);
3040 destroy_work_on_stack(&barr.work);
3041 return true;
3042 } else {
3043 return false;
3044 }
3045}
3046
3047/**
3048 * flush_work - wait for a work to finish executing the last queueing instance
3049 * @work: the work to flush
3050 *
3051 * Wait until @work has finished execution. @work is guaranteed to be idle
3052 * on return if it hasn't been requeued since flush started.
3053 *
3054 * Return:
3055 * %true if flush_work() waited for the work to finish execution,
3056 * %false if it was already idle.
3057 */
3058bool flush_work(struct work_struct *work)
3059{
3060 return __flush_work(work, false);
3061}
3062EXPORT_SYMBOL_GPL(flush_work);
3063
3064struct cwt_wait {
3065 wait_queue_entry_t wait;
3066 struct work_struct *work;
3067};
3068
3069static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3070{
3071 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3072
3073 if (cwait->work != key)
3074 return 0;
3075 return autoremove_wake_function(wait, mode, sync, key);
3076}
3077
3078static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3079{
3080 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3081 unsigned long flags;
3082 int ret;
3083
3084 do {
3085 ret = try_to_grab_pending(work, is_dwork, &flags);
3086 /*
3087 * If someone else is already canceling, wait for it to
3088 * finish. flush_work() doesn't work for PREEMPT_NONE
3089 * because we may get scheduled between @work's completion
3090 * and the other canceling task resuming and clearing
3091 * CANCELING - flush_work() will return false immediately
3092 * as @work is no longer busy, try_to_grab_pending() will
3093 * return -ENOENT as @work is still being canceled and the
3094 * other canceling task won't be able to clear CANCELING as
3095 * we're hogging the CPU.
3096 *
3097 * Let's wait for completion using a waitqueue. As this
3098 * may lead to the thundering herd problem, use a custom
3099 * wake function which matches @work along with exclusive
3100 * wait and wakeup.
3101 */
3102 if (unlikely(ret == -ENOENT)) {
3103 struct cwt_wait cwait;
3104
3105 init_wait(&cwait.wait);
3106 cwait.wait.func = cwt_wakefn;
3107 cwait.work = work;
3108
3109 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3110 TASK_UNINTERRUPTIBLE);
3111 if (work_is_canceling(work))
3112 schedule();
3113 finish_wait(&cancel_waitq, &cwait.wait);
3114 }
3115 } while (unlikely(ret < 0));
3116
3117 /* tell other tasks trying to grab @work to back off */
3118 mark_work_canceling(work);
3119 local_irq_restore(flags);
3120
3121 /*
3122 * This allows canceling during early boot. We know that @work
3123 * isn't executing.
3124 */
3125 if (wq_online)
3126 __flush_work(work, true);
3127
3128 clear_work_data(work);
3129
3130 /*
3131 * Paired with prepare_to_wait() above so that either
3132 * waitqueue_active() is visible here or !work_is_canceling() is
3133 * visible there.
3134 */
3135 smp_mb();
3136 if (waitqueue_active(&cancel_waitq))
3137 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3138
3139 return ret;
3140}
3141
3142/**
3143 * cancel_work_sync - cancel a work and wait for it to finish
3144 * @work: the work to cancel
3145 *
3146 * Cancel @work and wait for its execution to finish. This function
3147 * can be used even if the work re-queues itself or migrates to
3148 * another workqueue. On return from this function, @work is
3149 * guaranteed to be not pending or executing on any CPU.
3150 *
3151 * cancel_work_sync(&delayed_work->work) must not be used for
3152 * delayed_work's. Use cancel_delayed_work_sync() instead.
3153 *
3154 * The caller must ensure that the workqueue on which @work was last
3155 * queued can't be destroyed before this function returns.
3156 *
3157 * Return:
3158 * %true if @work was pending, %false otherwise.
3159 */
3160bool cancel_work_sync(struct work_struct *work)
3161{
3162 return __cancel_work_timer(work, false);
3163}
3164EXPORT_SYMBOL_GPL(cancel_work_sync);
3165
3166/**
3167 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3168 * @dwork: the delayed work to flush
3169 *
3170 * Delayed timer is cancelled and the pending work is queued for
3171 * immediate execution. Like flush_work(), this function only
3172 * considers the last queueing instance of @dwork.
3173 *
3174 * Return:
3175 * %true if flush_work() waited for the work to finish execution,
3176 * %false if it was already idle.
3177 */
3178bool flush_delayed_work(struct delayed_work *dwork)
3179{
3180 local_irq_disable();
3181 if (del_timer_sync(&dwork->timer))
3182 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3183 local_irq_enable();
3184 return flush_work(&dwork->work);
3185}
3186EXPORT_SYMBOL(flush_delayed_work);
3187
3188/**
3189 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3190 * @rwork: the rcu work to flush
3191 *
3192 * Return:
3193 * %true if flush_rcu_work() waited for the work to finish execution,
3194 * %false if it was already idle.
3195 */
3196bool flush_rcu_work(struct rcu_work *rwork)
3197{
3198 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3199 rcu_barrier();
3200 flush_work(&rwork->work);
3201 return true;
3202 } else {
3203 return flush_work(&rwork->work);
3204 }
3205}
3206EXPORT_SYMBOL(flush_rcu_work);
3207
3208static bool __cancel_work(struct work_struct *work, bool is_dwork)
3209{
3210 unsigned long flags;
3211 int ret;
3212
3213 do {
3214 ret = try_to_grab_pending(work, is_dwork, &flags);
3215 } while (unlikely(ret == -EAGAIN));
3216
3217 if (unlikely(ret < 0))
3218 return false;
3219
3220 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3221 local_irq_restore(flags);
3222 return ret;
3223}
3224
3225/**
3226 * cancel_delayed_work - cancel a delayed work
3227 * @dwork: delayed_work to cancel
3228 *
3229 * Kill off a pending delayed_work.
3230 *
3231 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3232 * pending.
3233 *
3234 * Note:
3235 * The work callback function may still be running on return, unless
3236 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3237 * use cancel_delayed_work_sync() to wait on it.
3238 *
3239 * This function is safe to call from any context including IRQ handler.
3240 */
3241bool cancel_delayed_work(struct delayed_work *dwork)
3242{
3243 return __cancel_work(&dwork->work, true);
3244}
3245EXPORT_SYMBOL(cancel_delayed_work);
3246
3247/**
3248 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3249 * @dwork: the delayed work cancel
3250 *
3251 * This is cancel_work_sync() for delayed works.
3252 *
3253 * Return:
3254 * %true if @dwork was pending, %false otherwise.
3255 */
3256bool cancel_delayed_work_sync(struct delayed_work *dwork)
3257{
3258 return __cancel_work_timer(&dwork->work, true);
3259}
3260EXPORT_SYMBOL(cancel_delayed_work_sync);
3261
3262/**
3263 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3264 * @func: the function to call
3265 *
3266 * schedule_on_each_cpu() executes @func on each online CPU using the
3267 * system workqueue and blocks until all CPUs have completed.
3268 * schedule_on_each_cpu() is very slow.
3269 *
3270 * Return:
3271 * 0 on success, -errno on failure.
3272 */
3273int schedule_on_each_cpu(work_func_t func)
3274{
3275 int cpu;
3276 struct work_struct __percpu *works;
3277
3278 works = alloc_percpu(struct work_struct);
3279 if (!works)
3280 return -ENOMEM;
3281
3282 get_online_cpus();
3283
3284 for_each_online_cpu(cpu) {
3285 struct work_struct *work = per_cpu_ptr(works, cpu);
3286
3287 INIT_WORK(work, func);
3288 schedule_work_on(cpu, work);
3289 }
3290
3291 for_each_online_cpu(cpu)
3292 flush_work(per_cpu_ptr(works, cpu));
3293
3294 put_online_cpus();
3295 free_percpu(works);
3296 return 0;
3297}
3298
3299/**
3300 * execute_in_process_context - reliably execute the routine with user context
3301 * @fn: the function to execute
3302 * @ew: guaranteed storage for the execute work structure (must
3303 * be available when the work executes)
3304 *
3305 * Executes the function immediately if process context is available,
3306 * otherwise schedules the function for delayed execution.
3307 *
3308 * Return: 0 - function was executed
3309 * 1 - function was scheduled for execution
3310 */
3311int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3312{
3313 if (!in_interrupt()) {
3314 fn(&ew->work);
3315 return 0;
3316 }
3317
3318 INIT_WORK(&ew->work, fn);
3319 schedule_work(&ew->work);
3320
3321 return 1;
3322}
3323EXPORT_SYMBOL_GPL(execute_in_process_context);
3324
3325/**
3326 * free_workqueue_attrs - free a workqueue_attrs
3327 * @attrs: workqueue_attrs to free
3328 *
3329 * Undo alloc_workqueue_attrs().
3330 */
3331void free_workqueue_attrs(struct workqueue_attrs *attrs)
3332{
3333 if (attrs) {
3334 free_cpumask_var(attrs->cpumask);
3335 kfree(attrs);
3336 }
3337}
3338
3339/**
3340 * alloc_workqueue_attrs - allocate a workqueue_attrs
3341 * @gfp_mask: allocation mask to use
3342 *
3343 * Allocate a new workqueue_attrs, initialize with default settings and
3344 * return it.
3345 *
3346 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3347 */
3348struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3349{
3350 struct workqueue_attrs *attrs;
3351
3352 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3353 if (!attrs)
3354 goto fail;
3355 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3356 goto fail;
3357
3358 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3359 return attrs;
3360fail:
3361 free_workqueue_attrs(attrs);
3362 return NULL;
3363}
3364
3365static void copy_workqueue_attrs(struct workqueue_attrs *to,
3366 const struct workqueue_attrs *from)
3367{
3368 to->nice = from->nice;
3369 cpumask_copy(to->cpumask, from->cpumask);
3370 /*
3371 * Unlike hash and equality test, this function doesn't ignore
3372 * ->no_numa as it is used for both pool and wq attrs. Instead,
3373 * get_unbound_pool() explicitly clears ->no_numa after copying.
3374 */
3375 to->no_numa = from->no_numa;
3376}
3377
3378/* hash value of the content of @attr */
3379static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3380{
3381 u32 hash = 0;
3382
3383 hash = jhash_1word(attrs->nice, hash);
3384 hash = jhash(cpumask_bits(attrs->cpumask),
3385 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3386 return hash;
3387}
3388
3389/* content equality test */
3390static bool wqattrs_equal(const struct workqueue_attrs *a,
3391 const struct workqueue_attrs *b)
3392{
3393 if (a->nice != b->nice)
3394 return false;
3395 if (!cpumask_equal(a->cpumask, b->cpumask))
3396 return false;
3397 return true;
3398}
3399
3400/**
3401 * init_worker_pool - initialize a newly zalloc'd worker_pool
3402 * @pool: worker_pool to initialize
3403 *
3404 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3405 *
3406 * Return: 0 on success, -errno on failure. Even on failure, all fields
3407 * inside @pool proper are initialized and put_unbound_pool() can be called
3408 * on @pool safely to release it.
3409 */
3410static int init_worker_pool(struct worker_pool *pool)
3411{
3412 spin_lock_init(&pool->lock);
3413 pool->id = -1;
3414 pool->cpu = -1;
3415 pool->node = NUMA_NO_NODE;
3416 pool->flags |= POOL_DISASSOCIATED;
3417 pool->watchdog_ts = jiffies;
3418 INIT_LIST_HEAD(&pool->worklist);
3419 INIT_LIST_HEAD(&pool->idle_list);
3420 hash_init(pool->busy_hash);
3421
3422 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3423
3424 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3425
3426 INIT_LIST_HEAD(&pool->workers);
3427
3428 ida_init(&pool->worker_ida);
3429 INIT_HLIST_NODE(&pool->hash_node);
3430 pool->refcnt = 1;
3431
3432 /* shouldn't fail above this point */
3433 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3434 if (!pool->attrs)
3435 return -ENOMEM;
3436 return 0;
3437}
3438
3439#ifdef CONFIG_LOCKDEP
3440static void wq_init_lockdep(struct workqueue_struct *wq)
3441{
3442 char *lock_name;
3443
3444 lockdep_register_key(&wq->key);
3445 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3446 if (!lock_name)
3447 lock_name = wq->name;
3448
3449 wq->lock_name = lock_name;
3450 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3451}
3452
3453static void wq_unregister_lockdep(struct workqueue_struct *wq)
3454{
3455 lockdep_unregister_key(&wq->key);
3456}
3457
3458static void wq_free_lockdep(struct workqueue_struct *wq)
3459{
3460 if (wq->lock_name != wq->name)
3461 kfree(wq->lock_name);
3462}
3463#else
3464static void wq_init_lockdep(struct workqueue_struct *wq)
3465{
3466}
3467
3468static void wq_unregister_lockdep(struct workqueue_struct *wq)
3469{
3470}
3471
3472static void wq_free_lockdep(struct workqueue_struct *wq)
3473{
3474}
3475#endif
3476
3477static void rcu_free_wq(struct rcu_head *rcu)
3478{
3479 struct workqueue_struct *wq =
3480 container_of(rcu, struct workqueue_struct, rcu);
3481
3482 wq_free_lockdep(wq);
3483
3484 if (!(wq->flags & WQ_UNBOUND))
3485 free_percpu(wq->cpu_pwqs);
3486 else
3487 free_workqueue_attrs(wq->unbound_attrs);
3488
3489 kfree(wq->rescuer);
3490 kfree(wq);
3491}
3492
3493static void rcu_free_pool(struct rcu_head *rcu)
3494{
3495 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3496
3497 ida_destroy(&pool->worker_ida);
3498 free_workqueue_attrs(pool->attrs);
3499 kfree(pool);
3500}
3501
3502/**
3503 * put_unbound_pool - put a worker_pool
3504 * @pool: worker_pool to put
3505 *
3506 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3507 * safe manner. get_unbound_pool() calls this function on its failure path
3508 * and this function should be able to release pools which went through,
3509 * successfully or not, init_worker_pool().
3510 *
3511 * Should be called with wq_pool_mutex held.
3512 */
3513static void put_unbound_pool(struct worker_pool *pool)
3514{
3515 DECLARE_COMPLETION_ONSTACK(detach_completion);
3516 struct worker *worker;
3517
3518 lockdep_assert_held(&wq_pool_mutex);
3519
3520 if (--pool->refcnt)
3521 return;
3522
3523 /* sanity checks */
3524 if (WARN_ON(!(pool->cpu < 0)) ||
3525 WARN_ON(!list_empty(&pool->worklist)))
3526 return;
3527
3528 /* release id and unhash */
3529 if (pool->id >= 0)
3530 idr_remove(&worker_pool_idr, pool->id);
3531 hash_del(&pool->hash_node);
3532
3533 /*
3534 * Become the manager and destroy all workers. This prevents
3535 * @pool's workers from blocking on attach_mutex. We're the last
3536 * manager and @pool gets freed with the flag set.
3537 */
3538 spin_lock_irq(&pool->lock);
3539 wait_event_lock_irq(wq_manager_wait,
3540 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3541 pool->flags |= POOL_MANAGER_ACTIVE;
3542
3543 while ((worker = first_idle_worker(pool)))
3544 destroy_worker(worker);
3545 WARN_ON(pool->nr_workers || pool->nr_idle);
3546 spin_unlock_irq(&pool->lock);
3547
3548 mutex_lock(&wq_pool_attach_mutex);
3549 if (!list_empty(&pool->workers))
3550 pool->detach_completion = &detach_completion;
3551 mutex_unlock(&wq_pool_attach_mutex);
3552
3553 if (pool->detach_completion)
3554 wait_for_completion(pool->detach_completion);
3555
3556 /* shut down the timers */
3557 del_timer_sync(&pool->idle_timer);
3558 del_timer_sync(&pool->mayday_timer);
3559
3560 /* sched-RCU protected to allow dereferences from get_work_pool() */
3561 call_rcu(&pool->rcu, rcu_free_pool);
3562}
3563
3564/**
3565 * get_unbound_pool - get a worker_pool with the specified attributes
3566 * @attrs: the attributes of the worker_pool to get
3567 *
3568 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3569 * reference count and return it. If there already is a matching
3570 * worker_pool, it will be used; otherwise, this function attempts to
3571 * create a new one.
3572 *
3573 * Should be called with wq_pool_mutex held.
3574 *
3575 * Return: On success, a worker_pool with the same attributes as @attrs.
3576 * On failure, %NULL.
3577 */
3578static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3579{
3580 u32 hash = wqattrs_hash(attrs);
3581 struct worker_pool *pool;
3582 int node;
3583 int target_node = NUMA_NO_NODE;
3584
3585 lockdep_assert_held(&wq_pool_mutex);
3586
3587 /* do we already have a matching pool? */
3588