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